drivers/block/mtip32xx/mtip32xx.c

   1 /*
   2  * Driver for the Micron P320 SSD
   3  *   Copyright (C) 2011 Micron Technology, Inc.
   4  *
   5  * Portions of this code were derived from works subjected to the
   6  * following copyright:
   7  *    Copyright (C) 2009 Integrated Device Technology, Inc.
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License as published by
  11  * the Free Software Foundation; either version 2 of the License, or
  12  * (at your option) any later version.
  13  *
  14  * This program is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  * GNU General Public License for more details.
  18  *
  19  */
  20
  21 #include <linux/pci.h>
  22 #include <linux/interrupt.h>
  23 #include <linux/ata.h>
  24 #include <linux/delay.h>
  25 #include <linux/hdreg.h>
  26 #include <linux/uaccess.h>
  27 #include <linux/random.h>
  28 #include <linux/smp.h>
  29 #include <linux/compat.h>
  30 #include <linux/fs.h>
  31 #include <linux/module.h>
  32 #include <linux/genhd.h>
  33 #include <linux/blkdev.h>
  34 #include <linux/blk-mq.h>
  35 #include <linux/bio.h>
  36 #include <linux/dma-mapping.h>
  37 #include <linux/idr.h>
  38 #include <linux/kthread.h>
  39 #include <../drivers/ata/ahci.h>
  40 #include <linux/export.h>
  41 #include <linux/debugfs.h>
  42 #include <linux/prefetch.h>
  43 #include "mtip32xx.h"
  44
  45 #define HW_CMD_SLOT_SZ          (MTIP_MAX_COMMAND_SLOTS * 32)
  46
  47 /* DMA region containing RX Fis, Identify, RLE10, and SMART buffers */
  48 #define AHCI_RX_FIS_SZ          0x100
  49 #define AHCI_RX_FIS_OFFSET      0x0
  50 #define AHCI_IDFY_SZ            ATA_SECT_SIZE
  51 #define AHCI_IDFY_OFFSET        0x400
  52 #define AHCI_SECTBUF_SZ         ATA_SECT_SIZE
  53 #define AHCI_SECTBUF_OFFSET     0x800
  54 #define AHCI_SMARTBUF_SZ        ATA_SECT_SIZE
  55 #define AHCI_SMARTBUF_OFFSET    0xC00
  56 /* 0x100 + 0x200 + 0x200 + 0x200 is smaller than 4k but we pad it out */
  57 #define BLOCK_DMA_ALLOC_SZ      4096
  58
  59 /* DMA region containing command table (should be 8192 bytes) */
  60 #define AHCI_CMD_SLOT_SZ        sizeof(struct mtip_cmd_hdr)
  61 #define AHCI_CMD_TBL_SZ         (MTIP_MAX_COMMAND_SLOTS * AHCI_CMD_SLOT_SZ)
  62 #define AHCI_CMD_TBL_OFFSET     0x0
  63
  64 /* DMA region per command (contains header and SGL) */
  65 #define AHCI_CMD_TBL_HDR_SZ     0x80
  66 #define AHCI_CMD_TBL_HDR_OFFSET 0x0
  67 #define AHCI_CMD_TBL_SGL_SZ     (MTIP_MAX_SG * sizeof(struct mtip_cmd_sg))
  68 #define AHCI_CMD_TBL_SGL_OFFSET AHCI_CMD_TBL_HDR_SZ
  69 #define CMD_DMA_ALLOC_SZ        (AHCI_CMD_TBL_SGL_SZ + AHCI_CMD_TBL_HDR_SZ)
  70
  71
  72 #define HOST_CAP_NZDMA          (1 << 19)
  73 #define HOST_HSORG              0xFC
  74 #define HSORG_DISABLE_SLOTGRP_INTR (1<<24)
  75 #define HSORG_DISABLE_SLOTGRP_PXIS (1<<16)
  76 #define HSORG_HWREV             0xFF00
  77 #define HSORG_STYLE             0x8
  78 #define HSORG_SLOTGROUPS        0x7
  79
  80 #define PORT_COMMAND_ISSUE      0x38
  81 #define PORT_SDBV               0x7C
  82
  83 #define PORT_OFFSET             0x100
  84 #define PORT_MEM_SIZE           0x80
  85
  86 #define PORT_IRQ_ERR \
  87         (PORT_IRQ_HBUS_ERR | PORT_IRQ_IF_ERR | PORT_IRQ_CONNECT | \
  88          PORT_IRQ_PHYRDY | PORT_IRQ_UNK_FIS | PORT_IRQ_BAD_PMP | \
  89          PORT_IRQ_TF_ERR | PORT_IRQ_HBUS_DATA_ERR | PORT_IRQ_IF_NONFATAL | \
  90          PORT_IRQ_OVERFLOW)
  91 #define PORT_IRQ_LEGACY \
  92         (PORT_IRQ_PIOS_FIS | PORT_IRQ_D2H_REG_FIS)
  93 #define PORT_IRQ_HANDLED \
  94         (PORT_IRQ_SDB_FIS | PORT_IRQ_LEGACY | \
  95          PORT_IRQ_TF_ERR | PORT_IRQ_IF_ERR | \
  96          PORT_IRQ_CONNECT | PORT_IRQ_PHYRDY)
  97 #define DEF_PORT_IRQ \
  98         (PORT_IRQ_ERR | PORT_IRQ_LEGACY | PORT_IRQ_SDB_FIS)
  99
 100 /* product numbers */
 101 #define MTIP_PRODUCT_UNKNOWN    0x00
 102 #define MTIP_PRODUCT_ASICFPGA   0x11
 103
 104 /* Device instance number, incremented each time a device is probed. */
 105 static int instance;
 106
 107 static struct list_head online_list;
 108 static struct list_head removing_list;
 109 static spinlock_t dev_lock;
 110
 111 /*
 112  * Global variable used to hold the major block device number
 113  * allocated in mtip_init().
 114  */
 115 static int mtip_major;
 116 static struct dentry *dfs_parent;
 117 static struct dentry *dfs_device_status;
 118
 119 static u32 cpu_use[NR_CPUS];
 120
 121 static DEFINE_SPINLOCK(rssd_index_lock);
 122 static DEFINE_IDA(rssd_index_ida);
 123
 124 static int mtip_block_initialize(struct driver_data *dd);
 125
 126 #ifdef CONFIG_COMPAT
 127 struct mtip_compat_ide_task_request_s {
 128         __u8            io_ports[8];
 129         __u8            hob_ports[8];
 130         ide_reg_valid_t out_flags;
 131         ide_reg_valid_t in_flags;
 132         int             data_phase;
 133         int             req_cmd;
 134         compat_ulong_t  out_size;
 135         compat_ulong_t  in_size;
 136 };
 137 #endif
 138
 139 /*
 140  * This function check_for_surprise_removal is called
 141  * while card is removed from the system and it will
 142  * read the vendor id from the configration space
 143  *
 144  * @pdev Pointer to the pci_dev structure.
 145  *
 146  * return value
 147  *       true if device removed, else false
 148  */
 149 static bool mtip_check_surprise_removal(struct pci_dev *pdev)
 150 {
 151         u16 vendor_id = 0;
 152         struct driver_data *dd = pci_get_drvdata(pdev);
 153
 154         if (dd->sr)
 155                 return true;
 156
 157        /* Read the vendorID from the configuration space */
 158         pci_read_config_word(pdev, 0x00, &vendor_id);
 159         if (vendor_id == 0xFFFF) {
 160                 dd->sr = true;
 161                 if (dd->queue)
 162                         set_bit(QUEUE_FLAG_DEAD, &dd->queue->queue_flags);
 163                 else
 164                         dev_warn(&dd->pdev->dev,
 165                                 "%s: dd->queue is NULL\n", __func__);
 166                 return true; /* device removed */
 167         }
 168
 169         return false; /* device present */
 170 }
 171
 172 /* we have to use runtime tag to setup command header */
 173 static void mtip_init_cmd_header(struct request *rq)
 174 {
 175         struct driver_data *dd = rq->q->queuedata;
 176         struct mtip_cmd *cmd = blk_mq_rq_to_pdu(rq);
 177         u32 host_cap_64 = readl(dd->mmio + HOST_CAP) & HOST_CAP_64;
 178
 179         /* Point the command headers at the command tables. */
 180         cmd->command_header = dd->port->command_list +
 181                                 (sizeof(struct mtip_cmd_hdr) * rq->tag);
 182         cmd->command_header_dma = dd->port->command_list_dma +
 183                                 (sizeof(struct mtip_cmd_hdr) * rq->tag);
 184
 185         if (host_cap_64)
 186                 cmd->command_header->ctbau = __force_bit2int cpu_to_le32((cmd->command_dma >> 16) >> 16);
 187
 188         cmd->command_header->ctba = __force_bit2int cpu_to_le32(cmd->command_dma & 0xFFFFFFFF);
 189 }
 190
 191 static struct mtip_cmd *mtip_get_int_command(struct driver_data *dd)
 192 {
 193         struct request *rq;
 194
 195         if (mtip_check_surprise_removal(dd->pdev))
 196                 return NULL;
 197
 198         rq = blk_mq_alloc_request(dd->queue, REQ_OP_DRV_IN, BLK_MQ_REQ_RESERVED);
 199         if (IS_ERR(rq))
 200                 return NULL;
 201
 202         /* Internal cmd isn't submitted via .queue_rq */
 203         mtip_init_cmd_header(rq);
 204
 205         return blk_mq_rq_to_pdu(rq);
 206 }
 207
 208 static void mtip_put_int_command(struct driver_data *dd, struct mtip_cmd *cmd)
 209 {
 210         blk_put_request(blk_mq_rq_from_pdu(cmd));
 211 }
 212
 213 /*
 214  * Once we add support for one hctx per mtip group, this will change a bit
 215  */
 216 static struct request *mtip_rq_from_tag(struct driver_data *dd,
 217                                         unsigned int tag)
 218 {
 219         struct blk_mq_hw_ctx *hctx = dd->queue->queue_hw_ctx[0];
 220
 221         return blk_mq_tag_to_rq(hctx->tags, tag);
 222 }
 223
 224 static struct mtip_cmd *mtip_cmd_from_tag(struct driver_data *dd,
 225                                           unsigned int tag)
 226 {
 227         struct request *rq = mtip_rq_from_tag(dd, tag);
 228
 229         return blk_mq_rq_to_pdu(rq);
 230 }
 231
 232 /*
 233  * IO completion function.
 234  *
 235  * This completion function is called by the driver ISR when a
 236  * command that was issued by the kernel completes. It first calls the
 237  * asynchronous completion function which normally calls back into the block
 238  * layer passing the asynchronous callback data, then unmaps the
 239  * scatter list associated with the completed command, and finally
 240  * clears the allocated bit associated with the completed command.
 241  *
 242  * @port   Pointer to the port data structure.
 243  * @tag    Tag of the command.
 244  * @data   Pointer to driver_data.
 245  * @status Completion status.
 246  *
 247  * return value
 248  *      None
 249  */
 250 static void mtip_async_complete(struct mtip_port *port,
 251                                 int tag, struct mtip_cmd *cmd, int status)
 252 {
 253         struct driver_data *dd = port->dd;
 254         struct request *rq;
 255
 256         if (unlikely(!dd) || unlikely(!port))
 257                 return;
 258
 259         if (unlikely(status == PORT_IRQ_TF_ERR)) {
 260                 dev_warn(&port->dd->pdev->dev,
 261                         "Command tag %d failed due to TFE\n", tag);
 262         }
 263
 264         rq = mtip_rq_from_tag(dd, tag);
 265
 266         cmd->status = status;
 267         blk_mq_complete_request(rq);
 268 }
 269
 270 /*
 271  * Reset the HBA (without sleeping)
 272  *
 273  * @dd Pointer to the driver data structure.
 274  *
 275  * return value
 276  *      0       The reset was successful.
 277  *      -1      The HBA Reset bit did not clear.
 278  */
 279 static int mtip_hba_reset(struct driver_data *dd)
 280 {
 281         unsigned long timeout;
 282
 283         /* Set the reset bit */
 284         writel(HOST_RESET, dd->mmio + HOST_CTL);
 285
 286         /* Flush */
 287         readl(dd->mmio + HOST_CTL);
 288
 289         /*
 290          * Spin for up to 10 seconds waiting for reset acknowledgement. Spec
 291          * is 1 sec but in LUN failure conditions, up to 10 secs are required
 292          */
 293         timeout = jiffies + msecs_to_jiffies(10000);
 294         do {
 295                 mdelay(10);
 296                 if (test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &dd->dd_flag))
 297                         return -1;
 298
 299         } while ((readl(dd->mmio + HOST_CTL) & HOST_RESET)
 300                  && time_before(jiffies, timeout));
 301
 302         if (readl(dd->mmio + HOST_CTL) & HOST_RESET)
 303                 return -1;
 304
 305         return 0;
 306 }
 307
 308 /*
 309  * Issue a command to the hardware.
 310  *
 311  * Set the appropriate bit in the s_active and Command Issue hardware
 312  * registers, causing hardware command processing to begin.
 313  *
 314  * @port Pointer to the port structure.
 315  * @tag  The tag of the command to be issued.
 316  *
 317  * return value
 318  *      None
 319  */
 320 static inline void mtip_issue_ncq_command(struct mtip_port *port, int tag)
 321 {
 322         int group = tag >> 5;
 323
 324         /* guard SACT and CI registers */
 325         spin_lock(&port->cmd_issue_lock[group]);
 326         writel((1 << MTIP_TAG_BIT(tag)),
 327                         port->s_active[MTIP_TAG_INDEX(tag)]);
 328         writel((1 << MTIP_TAG_BIT(tag)),
 329                         port->cmd_issue[MTIP_TAG_INDEX(tag)]);
 330         spin_unlock(&port->cmd_issue_lock[group]);
 331 }
 332
 333 /*
 334  * Enable/disable the reception of FIS
 335  *
 336  * @port   Pointer to the port data structure
 337  * @enable 1 to enable, 0 to disable
 338  *
 339  * return value
 340  *      Previous state: 1 enabled, 0 disabled
 341  */
 342 static int mtip_enable_fis(struct mtip_port *port, int enable)
 343 {
 344         u32 tmp;
 345
 346         /* enable FIS reception */
 347         tmp = readl(port->mmio + PORT_CMD);
 348         if (enable)
 349                 writel(tmp | PORT_CMD_FIS_RX, port->mmio + PORT_CMD);
 350         else
 351                 writel(tmp & ~PORT_CMD_FIS_RX, port->mmio + PORT_CMD);
 352
 353         /* Flush */
 354         readl(port->mmio + PORT_CMD);
 355
 356         return (((tmp & PORT_CMD_FIS_RX) == PORT_CMD_FIS_RX));
 357 }
 358
 359 /*
 360  * Enable/disable the DMA engine
 361  *
 362  * @port   Pointer to the port data structure
 363  * @enable 1 to enable, 0 to disable
 364  *
 365  * return value
 366  *      Previous state: 1 enabled, 0 disabled.
 367  */
 368 static int mtip_enable_engine(struct mtip_port *port, int enable)
 369 {
 370         u32 tmp;
 371
 372         /* enable FIS reception */
 373         tmp = readl(port->mmio + PORT_CMD);
 374         if (enable)
 375                 writel(tmp | PORT_CMD_START, port->mmio + PORT_CMD);
 376         else
 377                 writel(tmp & ~PORT_CMD_START, port->mmio + PORT_CMD);
 378
 379         readl(port->mmio + PORT_CMD);
 380         return (((tmp & PORT_CMD_START) == PORT_CMD_START));
 381 }
 382
 383 /*
 384  * Enables the port DMA engine and FIS reception.
 385  *
 386  * return value
 387  *      None
 388  */
 389 static inline void mtip_start_port(struct mtip_port *port)
 390 {
 391         /* Enable FIS reception */
 392         mtip_enable_fis(port, 1);
 393
 394         /* Enable the DMA engine */
 395         mtip_enable_engine(port, 1);
 396 }
 397
 398 /*
 399  * Deinitialize a port by disabling port interrupts, the DMA engine,
 400  * and FIS reception.
 401  *
 402  * @port Pointer to the port structure
 403  *
 404  * return value
 405  *      None
 406  */
 407 static inline void mtip_deinit_port(struct mtip_port *port)
 408 {
 409         /* Disable interrupts on this port */
 410         writel(0, port->mmio + PORT_IRQ_MASK);
 411
 412         /* Disable the DMA engine */
 413         mtip_enable_engine(port, 0);
 414
 415         /* Disable FIS reception */
 416         mtip_enable_fis(port, 0);
 417 }
 418
 419 /*
 420  * Initialize a port.
 421  *
 422  * This function deinitializes the port by calling mtip_deinit_port() and
 423  * then initializes it by setting the command header and RX FIS addresses,
 424  * clearing the SError register and any pending port interrupts before
 425  * re-enabling the default set of port interrupts.
 426  *
 427  * @port Pointer to the port structure.
 428  *
 429  * return value
 430  *      None
 431  */
 432 static void mtip_init_port(struct mtip_port *port)
 433 {
 434         int i;
 435         mtip_deinit_port(port);
 436
 437         /* Program the command list base and FIS base addresses */
 438         if (readl(port->dd->mmio + HOST_CAP) & HOST_CAP_64) {
 439                 writel((port->command_list_dma >> 16) >> 16,
 440                          port->mmio + PORT_LST_ADDR_HI);
 441                 writel((port->rxfis_dma >> 16) >> 16,
 442                          port->mmio + PORT_FIS_ADDR_HI);
 443         }
 444
 445         writel(port->command_list_dma & 0xFFFFFFFF,
 446                         port->mmio + PORT_LST_ADDR);
 447         writel(port->rxfis_dma & 0xFFFFFFFF, port->mmio + PORT_FIS_ADDR);
 448
 449         /* Clear SError */
 450         writel(readl(port->mmio + PORT_SCR_ERR), port->mmio + PORT_SCR_ERR);
 451
 452         /* reset the completed registers.*/
 453         for (i = 0; i < port->dd->slot_groups; i++)
 454                 writel(0xFFFFFFFF, port->completed[i]);
 455
 456         /* Clear any pending interrupts for this port */
 457         writel(readl(port->mmio + PORT_IRQ_STAT), port->mmio + PORT_IRQ_STAT);
 458
 459         /* Clear any pending interrupts on the HBA. */
 460         writel(readl(port->dd->mmio + HOST_IRQ_STAT),
 461                                         port->dd->mmio + HOST_IRQ_STAT);
 462
 463         /* Enable port interrupts */
 464         writel(DEF_PORT_IRQ, port->mmio + PORT_IRQ_MASK);
 465 }
 466
 467 /*
 468  * Restart a port
 469  *
 470  * @port Pointer to the port data structure.
 471  *
 472  * return value
 473  *      None
 474  */
 475 static void mtip_restart_port(struct mtip_port *port)
 476 {
 477         unsigned long timeout;
 478
 479         /* Disable the DMA engine */
 480         mtip_enable_engine(port, 0);
 481
 482         /* Chip quirk: wait up to 500ms for PxCMD.CR == 0 */
 483         timeout = jiffies + msecs_to_jiffies(500);
 484         while ((readl(port->mmio + PORT_CMD) & PORT_CMD_LIST_ON)
 485                  && time_before(jiffies, timeout))
 486                 ;
 487
 488         if (test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &port->dd->dd_flag))
 489                 return;
 490
 491         /*
 492          * Chip quirk: escalate to hba reset if
 493          * PxCMD.CR not clear after 500 ms
 494          */
 495         if (readl(port->mmio + PORT_CMD) & PORT_CMD_LIST_ON) {
 496                 dev_warn(&port->dd->pdev->dev,
 497                         "PxCMD.CR not clear, escalating reset\n");
 498
 499                 if (mtip_hba_reset(port->dd))
 500                         dev_err(&port->dd->pdev->dev,
 501                                 "HBA reset escalation failed.\n");
 502
 503                 /* 30 ms delay before com reset to quiesce chip */
 504                 mdelay(30);
 505         }
 506
 507         dev_warn(&port->dd->pdev->dev, "Issuing COM reset\n");
 508
 509         /* Set PxSCTL.DET */
 510         writel(readl(port->mmio + PORT_SCR_CTL) |
 511                          1, port->mmio + PORT_SCR_CTL);
 512         readl(port->mmio + PORT_SCR_CTL);
 513
 514         /* Wait 1 ms to quiesce chip function */
 515         timeout = jiffies + msecs_to_jiffies(1);
 516         while (time_before(jiffies, timeout))
 517                 ;
 518
 519         if (test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &port->dd->dd_flag))
 520                 return;
 521
 522         /* Clear PxSCTL.DET */
 523         writel(readl(port->mmio + PORT_SCR_CTL) & ~1,
 524                          port->mmio + PORT_SCR_CTL);
 525         readl(port->mmio + PORT_SCR_CTL);
 526
 527         /* Wait 500 ms for bit 0 of PORT_SCR_STS to be set */
 528         timeout = jiffies + msecs_to_jiffies(500);
 529         while (((readl(port->mmio + PORT_SCR_STAT) & 0x01) == 0)
 530                          && time_before(jiffies, timeout))
 531                 ;
 532
 533         if (test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &port->dd->dd_flag))
 534                 return;
 535
 536         if ((readl(port->mmio + PORT_SCR_STAT) & 0x01) == 0)
 537                 dev_warn(&port->dd->pdev->dev,
 538                         "COM reset failed\n");
 539
 540         mtip_init_port(port);
 541         mtip_start_port(port);
 542
 543 }
 544
 545 static int mtip_device_reset(struct driver_data *dd)
 546 {
 547         int rv = 0;
 548
 549         if (mtip_check_surprise_removal(dd->pdev))
 550                 return 0;
 551
 552         if (mtip_hba_reset(dd) < 0)
 553                 rv = -EFAULT;
 554
 555         mdelay(1);
 556         mtip_init_port(dd->port);
 557         mtip_start_port(dd->port);
 558
 559         /* Enable interrupts on the HBA. */
 560         writel(readl(dd->mmio + HOST_CTL) | HOST_IRQ_EN,
 561                                         dd->mmio + HOST_CTL);
 562         return rv;
 563 }
 564
 565 /*
 566  * Helper function for tag logging
 567  */
 568 static void print_tags(struct driver_data *dd,
 569                         char *msg,
 570                         unsigned long *tagbits,
 571                         int cnt)
 572 {
 573         unsigned char tagmap[128];
 574         int group, tagmap_len = 0;
 575
 576         memset(tagmap, 0, sizeof(tagmap));
 577         for (group = SLOTBITS_IN_LONGS; group > 0; group--)
 578                 tagmap_len += sprintf(tagmap + tagmap_len, "%016lX ",
 579                                                 tagbits[group-1]);
 580         dev_warn(&dd->pdev->dev,
 581                         "%d command(s) %s: tagmap [%s]", cnt, msg, tagmap);
 582 }
 583
 584 /*
 585  * Internal command completion callback function.
 586  *
 587  * This function is normally called by the driver ISR when an internal
 588  * command completed. This function signals the command completion by
 589  * calling complete().
 590  *
 591  * @port   Pointer to the port data structure.
 592  * @tag    Tag of the command that has completed.
 593  * @data   Pointer to a completion structure.
 594  * @status Completion status.
 595  *
 596  * return value
 597  *      None
 598  */
 599 static void mtip_completion(struct mtip_port *port,
 600                             int tag, struct mtip_cmd *command, int status)
 601 {
 602         struct completion *waiting = command->comp_data;
 603         if (unlikely(status == PORT_IRQ_TF_ERR))
 604                 dev_warn(&port->dd->pdev->dev,
 605                         "Internal command %d completed with TFE\n", tag);
 606
 607         command->comp_func = NULL;
 608         command->comp_data = NULL;
 609         complete(waiting);
 610 }
 611
 612 static int mtip_read_log_page(struct mtip_port *port, u8 page, u16 *buffer,
 613                                 dma_addr_t buffer_dma, unsigned int sectors);
 614 static int mtip_get_smart_attr(struct mtip_port *port, unsigned int id,
 615                                                 struct smart_attr *attrib);
 616 /*
 617  * Handle an error.
 618  *
 619  * @dd Pointer to the DRIVER_DATA structure.
 620  *
 621  * return value
 622  *      None
 623  */
 624 static void mtip_handle_tfe(struct driver_data *dd)
 625 {
 626         int group, tag, bit, reissue, rv;
 627         struct mtip_port *port;
 628         struct mtip_cmd  *cmd;
 629         u32 completed;
 630         struct host_to_dev_fis *fis;
 631         unsigned long tagaccum[SLOTBITS_IN_LONGS];
 632         unsigned int cmd_cnt = 0;
 633         unsigned char *buf;
 634         char *fail_reason = NULL;
 635         int fail_all_ncq_write = 0, fail_all_ncq_cmds = 0;
 636
 637         dev_warn(&dd->pdev->dev, "Taskfile error\n");
 638
 639         port = dd->port;
 640
 641         if (test_bit(MTIP_PF_IC_ACTIVE_BIT, &port->flags)) {
 642                 cmd = mtip_cmd_from_tag(dd, MTIP_TAG_INTERNAL);
 643                 dbg_printk(MTIP_DRV_NAME " TFE for the internal command\n");
 644
 645                 if (cmd->comp_data && cmd->comp_func) {
 646                         cmd->comp_func(port, MTIP_TAG_INTERNAL,
 647                                         cmd, PORT_IRQ_TF_ERR);
 648                 }
 649                 return;
 650         }
 651
 652         /* clear the tag accumulator */
 653         memset(tagaccum, 0, SLOTBITS_IN_LONGS * sizeof(long));
 654
 655         /* Loop through all the groups */
 656         for (group = 0; group < dd->slot_groups; group++) {
 657                 completed = readl(port->completed[group]);
 658
 659                 dev_warn(&dd->pdev->dev, "g=%u, comp=%x\n", group, completed);
 660
 661                 /* clear completed status register in the hardware.*/
 662                 writel(completed, port->completed[group]);
 663
 664                 /* Process successfully completed commands */
 665                 for (bit = 0; bit < 32 && completed; bit++) {
 666                         if (!(completed & (1<<bit)))
 667                                 continue;
 668                         tag = (group << 5) + bit;
 669
 670                         /* Skip the internal command slot */
 671                         if (tag == MTIP_TAG_INTERNAL)
 672                                 continue;
 673
 674                         cmd = mtip_cmd_from_tag(dd, tag);
 675                         if (likely(cmd->comp_func)) {
 676                                 set_bit(tag, tagaccum);
 677                                 cmd_cnt++;
 678                                 cmd->comp_func(port, tag, cmd, 0);
 679                         } else {
 680                                 dev_err(&port->dd->pdev->dev,
 681                                         "Missing completion func for tag %d",
 682                                         tag);
 683                                 if (mtip_check_surprise_removal(dd->pdev)) {
 684                                         /* don't proceed further */
 685                                         return;
 686                                 }
 687                         }
 688                 }
 689         }
 690
 691         print_tags(dd, "completed (TFE)", tagaccum, cmd_cnt);
 692
 693         /* Restart the port */
 694         mdelay(20);
 695         mtip_restart_port(port);
 696
 697         /* Trying to determine the cause of the error */
 698         rv = mtip_read_log_page(dd->port, ATA_LOG_SATA_NCQ,
 699                                 dd->port->log_buf,
 700                                 dd->port->log_buf_dma, 1);
 701         if (rv) {
 702                 dev_warn(&dd->pdev->dev,
 703                         "Error in READ LOG EXT (10h) command\n");
 704                 /* non-critical error, don't fail the load */
 705         } else {
 706                 buf = (unsigned char *)dd->port->log_buf;
 707                 if (buf[259] & 0x1) {
 708                         dev_info(&dd->pdev->dev,
 709                                 "Write protect bit is set.\n");
 710                         set_bit(MTIP_DDF_WRITE_PROTECT_BIT, &dd->dd_flag);
 711                         fail_all_ncq_write = 1;
 712                         fail_reason = "write protect";
 713                 }
 714                 if (buf[288] == 0xF7) {
 715                         dev_info(&dd->pdev->dev,
 716                                 "Exceeded Tmax, drive in thermal shutdown.\n");
 717                         set_bit(MTIP_DDF_OVER_TEMP_BIT, &dd->dd_flag);
 718                         fail_all_ncq_cmds = 1;
 719                         fail_reason = "thermal shutdown";
 720                 }
 721                 if (buf[288] == 0xBF) {
 722                         set_bit(MTIP_DDF_REBUILD_FAILED_BIT, &dd->dd_flag);
 723                         dev_info(&dd->pdev->dev,
 724                                 "Drive indicates rebuild has failed. Secure erase required.\n");
 725                         fail_all_ncq_cmds = 1;
 726                         fail_reason = "rebuild failed";
 727                 }
 728         }
 729
 730         /* clear the tag accumulator */
 731         memset(tagaccum, 0, SLOTBITS_IN_LONGS * sizeof(long));
 732
 733         /* Loop through all the groups */
 734         for (group = 0; group < dd->slot_groups; group++) {
 735                 for (bit = 0; bit < 32; bit++) {
 736                         reissue = 1;
 737                         tag = (group << 5) + bit;
 738                         cmd = mtip_cmd_from_tag(dd, tag);
 739
 740                         fis = (struct host_to_dev_fis *)cmd->command;
 741
 742                         /* Should re-issue? */
 743                         if (tag == MTIP_TAG_INTERNAL ||
 744                             fis->command == ATA_CMD_SET_FEATURES)
 745                                 reissue = 0;
 746                         else {
 747                                 if (fail_all_ncq_cmds ||
 748                                         (fail_all_ncq_write &&
 749                                         fis->command == ATA_CMD_FPDMA_WRITE)) {
 750                                         dev_warn(&dd->pdev->dev,
 751                                         "  Fail: %s w/tag %d [%s].\n",
 752                                         fis->command == ATA_CMD_FPDMA_WRITE ?
 753                                                 "write" : "read",
 754                                         tag,
 755                                         fail_reason != NULL ?
 756                                                 fail_reason : "unknown");
 757                                         if (cmd->comp_func) {
 758                                                 cmd->comp_func(port, tag,
 759                                                         cmd, -ENODATA);
 760                                         }
 761                                         continue;
 762                                 }
 763                         }
 764
 765                         /*
 766                          * First check if this command has
 767                          *  exceeded its retries.
 768                          */
 769                         if (reissue && (cmd->retries-- > 0)) {
 770
 771                                 set_bit(tag, tagaccum);
 772
 773                                 /* Re-issue the command. */
 774                                 mtip_issue_ncq_command(port, tag);
 775
 776                                 continue;
 777                         }
 778
 779                         /* Retire a command that will not be reissued */
 780                         dev_warn(&port->dd->pdev->dev,
 781                                 "retiring tag %d\n", tag);
 782
 783                         if (cmd->comp_func)
 784                                 cmd->comp_func(port, tag, cmd, PORT_IRQ_TF_ERR);
 785                         else
 786                                 dev_warn(&port->dd->pdev->dev,
 787                                         "Bad completion for tag %d\n",
 788                                         tag);
 789                 }
 790         }
 791         print_tags(dd, "reissued (TFE)", tagaccum, cmd_cnt);
 792 }
 793
 794 /*
 795  * Handle a set device bits interrupt
 796  */
 797 static inline void mtip_workq_sdbfx(struct mtip_port *port, int group,
 798                                                         u32 completed)
 799 {
 800         struct driver_data *dd = port->dd;
 801         int tag, bit;
 802         struct mtip_cmd *command;
 803
 804         if (!completed) {
 805                 WARN_ON_ONCE(!completed);
 806                 return;
 807         }
 808         /* clear completed status register in the hardware.*/
 809         writel(completed, port->completed[group]);
 810
 811         /* Process completed commands. */
 812         for (bit = 0; (bit < 32) && completed; bit++) {
 813                 if (completed & 0x01) {
 814                         tag = (group << 5) | bit;
 815
 816                         /* skip internal command slot. */
 817                         if (unlikely(tag == MTIP_TAG_INTERNAL))
 818                                 continue;
 819
 820                         command = mtip_cmd_from_tag(dd, tag);
 821                         if (likely(command->comp_func))
 822                                 command->comp_func(port, tag, command, 0);
 823                         else {
 824                                 dev_dbg(&dd->pdev->dev,
 825                                         "Null completion for tag %d",
 826                                         tag);
 827
 828                                 if (mtip_check_surprise_removal(
 829                                         dd->pdev)) {
 830                                         return;
 831                                 }
 832                         }
 833                 }
 834                 completed >>= 1;
 835         }
 836
 837         /* If last, re-enable interrupts */
 838         if (atomic_dec_return(&dd->irq_workers_active) == 0)
 839                 writel(0xffffffff, dd->mmio + HOST_IRQ_STAT);
 840 }
 841
 842 /*
 843  * Process legacy pio and d2h interrupts
 844  */
 845 static inline void mtip_process_legacy(struct driver_data *dd, u32 port_stat)
 846 {
 847         struct mtip_port *port = dd->port;
 848         struct mtip_cmd *cmd = mtip_cmd_from_tag(dd, MTIP_TAG_INTERNAL);
 849
 850         if (test_bit(MTIP_PF_IC_ACTIVE_BIT, &port->flags) &&
 851             (cmd != NULL) && !(readl(port->cmd_issue[MTIP_TAG_INTERNAL])
 852                 & (1 << MTIP_TAG_INTERNAL))) {
 853                 if (cmd->comp_func) {
 854                         cmd->comp_func(port, MTIP_TAG_INTERNAL, cmd, 0);
 855                         return;
 856                 }
 857         }
 858
 859         return;
 860 }
 861
 862 /*
 863  * Demux and handle errors
 864  */
 865 static inline void mtip_process_errors(struct driver_data *dd, u32 port_stat)
 866 {
 867
 868         if (unlikely(port_stat & PORT_IRQ_CONNECT)) {
 869                 dev_warn(&dd->pdev->dev,
 870                         "Clearing PxSERR.DIAG.x\n");
 871                 writel((1 << 26), dd->port->mmio + PORT_SCR_ERR);
 872         }
 873
 874         if (unlikely(port_stat & PORT_IRQ_PHYRDY)) {
 875                 dev_warn(&dd->pdev->dev,
 876                         "Clearing PxSERR.DIAG.n\n");
 877                 writel((1 << 16), dd->port->mmio + PORT_SCR_ERR);
 878         }
 879
 880         if (unlikely(port_stat & ~PORT_IRQ_HANDLED)) {
 881                 dev_warn(&dd->pdev->dev,
 882                         "Port stat errors %x unhandled\n",
 883                         (port_stat & ~PORT_IRQ_HANDLED));
 884                 if (mtip_check_surprise_removal(dd->pdev))
 885                         return;
 886         }
 887         if (likely(port_stat & (PORT_IRQ_TF_ERR | PORT_IRQ_IF_ERR))) {
 888                 set_bit(MTIP_PF_EH_ACTIVE_BIT, &dd->port->flags);
 889                 wake_up_interruptible(&dd->port->svc_wait);
 890         }
 891 }
 892
 893 static inline irqreturn_t mtip_handle_irq(struct driver_data *data)
 894 {
 895         struct driver_data *dd = (struct driver_data *) data;
 896         struct mtip_port *port = dd->port;
 897         u32 hba_stat, port_stat;
 898         int rv = IRQ_NONE;
 899         int do_irq_enable = 1, i, workers;
 900         struct mtip_work *twork;
 901
 902         hba_stat = readl(dd->mmio + HOST_IRQ_STAT);
 903         if (hba_stat) {
 904                 rv = IRQ_HANDLED;
 905
 906                 /* Acknowledge the interrupt status on the port.*/
 907                 port_stat = readl(port->mmio + PORT_IRQ_STAT);
 908                 if (unlikely(port_stat == 0xFFFFFFFF)) {
 909                         mtip_check_surprise_removal(dd->pdev);
 910                         return IRQ_HANDLED;
 911                 }
 912                 writel(port_stat, port->mmio + PORT_IRQ_STAT);
 913
 914                 /* Demux port status */
 915                 if (likely(port_stat & PORT_IRQ_SDB_FIS)) {
 916                         do_irq_enable = 0;
 917                         WARN_ON_ONCE(atomic_read(&dd->irq_workers_active) != 0);
 918
 919                         /* Start at 1: group zero is always local? */
 920                         for (i = 0, workers = 0; i < MTIP_MAX_SLOT_GROUPS;
 921                                                                         i++) {
 922                                 twork = &dd->work[i];
 923                                 twork->completed = readl(port->completed[i]);
 924                                 if (twork->completed)
 925                                         workers++;
 926                         }
 927
 928                         atomic_set(&dd->irq_workers_active, workers);
 929                         if (workers) {
 930                                 for (i = 1; i < MTIP_MAX_SLOT_GROUPS; i++) {
 931                                         twork = &dd->work[i];
 932                                         if (twork->completed)
 933                                                 queue_work_on(
 934                                                         twork->cpu_binding,
 935                                                         dd->isr_workq,
 936                                                         &twork->work);
 937                                 }
 938
 939                                 if (likely(dd->work[0].completed))
 940                                         mtip_workq_sdbfx(port, 0,
 941                                                         dd->work[0].completed);
 942
 943                         } else {
 944                                 /*
 945                                  * Chip quirk: SDB interrupt but nothing
 946                                  * to complete
 947                                  */
 948                                 do_irq_enable = 1;
 949                         }
 950                 }
 951
 952                 if (unlikely(port_stat & PORT_IRQ_ERR)) {
 953                         if (unlikely(mtip_check_surprise_removal(dd->pdev))) {
 954                                 /* don't proceed further */
 955                                 return IRQ_HANDLED;
 956                         }
 957                         if (test_bit(MTIP_DDF_REMOVE_PENDING_BIT,
 958                                                         &dd->dd_flag))
 959                                 return rv;
 960
 961                         mtip_process_errors(dd, port_stat & PORT_IRQ_ERR);
 962                 }
 963
 964                 if (unlikely(port_stat & PORT_IRQ_LEGACY))
 965                         mtip_process_legacy(dd, port_stat & PORT_IRQ_LEGACY);
 966         }
 967
 968         /* acknowledge interrupt */
 969         if (unlikely(do_irq_enable))
 970                 writel(hba_stat, dd->mmio + HOST_IRQ_STAT);
 971
 972         return rv;
 973 }
 974
 975 /*
 976  * HBA interrupt subroutine.
 977  *
 978  * @irq         IRQ number.
 979  * @instance    Pointer to the driver data structure.
 980  *
 981  * return value
 982  *      IRQ_HANDLED     A HBA interrupt was pending and handled.
 983  *      IRQ_NONE        This interrupt was not for the HBA.
 984  */
 985 static irqreturn_t mtip_irq_handler(int irq, void *instance)
 986 {
 987         struct driver_data *dd = instance;
 988
 989         return mtip_handle_irq(dd);
 990 }
 991
 992 static void mtip_issue_non_ncq_command(struct mtip_port *port, int tag)
 993 {
 994         writel(1 << MTIP_TAG_BIT(tag),
 995                 port->cmd_issue[MTIP_TAG_INDEX(tag)]);
 996 }
 997
 998 static bool mtip_pause_ncq(struct mtip_port *port,
 999                                 struct host_to_dev_fis *fis)
1000 {
1001         struct host_to_dev_fis *reply;
1002         unsigned long task_file_data;
1003
1004         reply = port->rxfis + RX_FIS_D2H_REG;
1005         task_file_data = readl(port->mmio+PORT_TFDATA);
1006
1007         if ((task_file_data & 1))
1008                 return false;
1009
1010         if (fis->command == ATA_CMD_SEC_ERASE_PREP) {
1011                 port->ic_pause_timer = jiffies;
1012                 return true;
1013         } else if ((fis->command == ATA_CMD_DOWNLOAD_MICRO) &&
1014                                         (fis->features == 0x03)) {
1015                 set_bit(MTIP_PF_DM_ACTIVE_BIT, &port->flags);
1016                 port->ic_pause_timer = jiffies;
1017                 return true;
1018         } else if ((fis->command == ATA_CMD_SEC_ERASE_UNIT) ||
1019                 ((fis->command == 0xFC) &&
1020                         (fis->features == 0x27 || fis->features == 0x72 ||
1021                          fis->features == 0x62 || fis->features == 0x26))) {
1022                 clear_bit(MTIP_DDF_SEC_LOCK_BIT, &port->dd->dd_flag);
1023                 clear_bit(MTIP_DDF_REBUILD_FAILED_BIT, &port->dd->dd_flag);
1024                 /* Com reset after secure erase or lowlevel format */
1025                 mtip_restart_port(port);
1026                 clear_bit(MTIP_PF_SE_ACTIVE_BIT, &port->flags);
1027                 return false;
1028         }
1029
1030         return false;
1031 }
1032
1033 static bool mtip_commands_active(struct mtip_port *port)
1034 {
1035         unsigned int active;
1036         unsigned int n;
1037
1038         /*
1039          * Ignore s_active bit 0 of array element 0.
1040          * This bit will always be set
1041          */
1042         active = readl(port->s_active[0]) & 0xFFFFFFFE;
1043         for (n = 1; n < port->dd->slot_groups; n++)
1044                 active |= readl(port->s_active[n]);
1045
1046         return active != 0;
1047 }
1048
1049 /*
1050  * Wait for port to quiesce
1051  *
1052  * @port    Pointer to port data structure
1053  * @timeout Max duration to wait (ms)
1054  *
1055  * return value
1056  *      0       Success
1057  *      -EBUSY  Commands still active
1058  */
1059 static int mtip_quiesce_io(struct mtip_port *port, unsigned long timeout)
1060 {
1061         unsigned long to;
1062         bool active = true;
1063
1064         blk_mq_stop_hw_queues(port->dd->queue);
1065
1066         to = jiffies + msecs_to_jiffies(timeout);
1067         do {
1068                 if (test_bit(MTIP_PF_SVC_THD_ACTIVE_BIT, &port->flags) &&
1069                         test_bit(MTIP_PF_ISSUE_CMDS_BIT, &port->flags)) {
1070                         msleep(20);
1071                         continue; /* svc thd is actively issuing commands */
1072                 }
1073
1074                 msleep(100);
1075
1076                 if (mtip_check_surprise_removal(port->dd->pdev))
1077                         goto err_fault;
1078
1079                 active = mtip_commands_active(port);
1080                 if (!active)
1081                         break;
1082         } while (time_before(jiffies, to));
1083
1084         blk_mq_start_stopped_hw_queues(port->dd->queue, true);
1085         return active ? -EBUSY : 0;
1086 err_fault:
1087         blk_mq_start_stopped_hw_queues(port->dd->queue, true);
1088         return -EFAULT;
1089 }
1090
1091 struct mtip_int_cmd {
1092         int fis_len;
1093         dma_addr_t buffer;
1094         int buf_len;
1095         u32 opts;
1096 };
1097
1098 /*
1099  * Execute an internal command and wait for the completion.
1100  *
1101  * @port    Pointer to the port data structure.
1102  * @fis     Pointer to the FIS that describes the command.
1103  * @fis_len  Length in WORDS of the FIS.
1104  * @buffer  DMA accessible for command data.
1105  * @buf_len  Length, in bytes, of the data buffer.
1106  * @opts    Command header options, excluding the FIS length
1107  *             and the number of PRD entries.
1108  * @timeout Time in ms to wait for the command to complete.
1109  *
1110  * return value
1111  *      0        Command completed successfully.
1112  *      -EFAULT  The buffer address is not correctly aligned.
1113  *      -EBUSY   Internal command or other IO in progress.
1114  *      -EAGAIN  Time out waiting for command to complete.
1115  */
1116 static int mtip_exec_internal_command(struct mtip_port *port,
1117                                         struct host_to_dev_fis *fis,
1118                                         int fis_len,
1119                                         dma_addr_t buffer,
1120                                         int buf_len,
1121                                         u32 opts,
1122                                         unsigned long timeout)
1123 {
1124         DECLARE_COMPLETION_ONSTACK(wait);
1125         struct mtip_cmd *int_cmd;
1126         struct driver_data *dd = port->dd;
1127         struct request *rq;
1128         struct mtip_int_cmd icmd = {
1129                 .fis_len = fis_len,
1130                 .buffer = buffer,
1131                 .buf_len = buf_len,
1132                 .opts = opts
1133         };
1134         int rv = 0;
1135         unsigned long start;
1136
1137         /* Make sure the buffer is 8 byte aligned. This is asic specific. */
1138         if (buffer & 0x00000007) {
1139                 dev_err(&dd->pdev->dev, "SG buffer is not 8 byte aligned\n");
1140                 return -EFAULT;
1141         }
1142
1143         int_cmd = mtip_get_int_command(dd);
1144         if (!int_cmd) {
1145                 dbg_printk(MTIP_DRV_NAME "Unable to allocate tag for PIO cmd\n");
1146                 return -EFAULT;
1147         }
1148         rq = blk_mq_rq_from_pdu(int_cmd);
1149         rq->end_io_data = &icmd;
1150
1151         set_bit(MTIP_PF_IC_ACTIVE_BIT, &port->flags);
1152
1153         if (fis->command == ATA_CMD_SEC_ERASE_PREP)
1154                 set_bit(MTIP_PF_SE_ACTIVE_BIT, &port->flags);
1155
1156         clear_bit(MTIP_PF_DM_ACTIVE_BIT, &port->flags);
1157
1158         if (fis->command != ATA_CMD_STANDBYNOW1) {
1159                 /* wait for io to complete if non atomic */
1160                 if (mtip_quiesce_io(port, MTIP_QUIESCE_IO_TIMEOUT_MS) < 0) {
1161                         dev_warn(&dd->pdev->dev, "Failed to quiesce IO\n");
1162                         mtip_put_int_command(dd, int_cmd);
1163                         clear_bit(MTIP_PF_IC_ACTIVE_BIT, &port->flags);
1164                         wake_up_interruptible(&port->svc_wait);
1165                         return -EBUSY;
1166                 }
1167         }
1168
1169         /* Set the completion function and data for the command. */
1170         int_cmd->comp_data = &wait;
1171         int_cmd->comp_func = mtip_completion;
1172
1173         /* Copy the command to the command table */
1174         memcpy(int_cmd->command, fis, fis_len*4);
1175
1176         start = jiffies;
1177         rq->timeout = timeout;
1178
1179         /* insert request and run queue */
1180         blk_execute_rq_nowait(rq->q, NULL, rq, true, NULL);
1181
1182         wait_for_completion(&wait);
1183         rv = int_cmd->status;
1184
1185         if (rv < 0) {
1186                 if (rv == -ERESTARTSYS) { /* interrupted */
1187                         dev_err(&dd->pdev->dev,
1188                                 "Internal command [%02X] was interrupted after %u ms\n",
1189                                 fis->command,
1190                                 jiffies_to_msecs(jiffies - start));
1191                         rv = -EINTR;
1192                         goto exec_ic_exit;
1193                 } else if (rv == 0) /* timeout */
1194                         dev_err(&dd->pdev->dev,
1195                                 "Internal command did not complete [%02X] within timeout of  %lu ms\n",
1196                                 fis->command, timeout);
1197                 else
1198                         dev_err(&dd->pdev->dev,
1199                                 "Internal command [%02X] wait returned code [%d] after %lu ms - unhandled\n",
1200                                 fis->command, rv, timeout);
1201
1202                 if (mtip_check_surprise_removal(dd->pdev) ||
1203                         test_bit(MTIP_DDF_REMOVE_PENDING_BIT,
1204                                         &dd->dd_flag)) {
1205                         dev_err(&dd->pdev->dev,
1206                                 "Internal command [%02X] wait returned due to SR\n",
1207                                 fis->command);
1208                         rv = -ENXIO;
1209                         goto exec_ic_exit;
1210                 }
1211                 mtip_device_reset(dd); /* recover from timeout issue */
1212                 rv = -EAGAIN;
1213                 goto exec_ic_exit;
1214         }
1215
1216         if (readl(port->cmd_issue[MTIP_TAG_INTERNAL])
1217                         & (1 << MTIP_TAG_INTERNAL)) {
1218                 rv = -ENXIO;
1219                 if (!test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &dd->dd_flag)) {
1220                         mtip_device_reset(dd);
1221                         rv = -EAGAIN;
1222                 }
1223         }
1224 exec_ic_exit:
1225         /* Clear the allocated and active bits for the internal command. */
1226         mtip_put_int_command(dd, int_cmd);
1227         clear_bit(MTIP_PF_IC_ACTIVE_BIT, &port->flags);
1228         if (rv >= 0 && mtip_pause_ncq(port, fis)) {
1229                 /* NCQ paused */
1230                 return rv;
1231         }
1232         wake_up_interruptible(&port->svc_wait);
1233
1234         return rv;
1235 }
1236
1237 /*
1238  * Byte-swap ATA ID strings.
1239  *
1240  * ATA identify data contains strings in byte-swapped 16-bit words.
1241  * They must be swapped (on all architectures) to be usable as C strings.
1242  * This function swaps bytes in-place.
1243  *
1244  * @buf The buffer location of the string
1245  * @len The number of bytes to swap
1246  *
1247  * return value
1248  *      None
1249  */
1250 static inline void ata_swap_string(u16 *buf, unsigned int len)
1251 {
1252         int i;
1253         for (i = 0; i < (len/2); i++)
1254                 be16_to_cpus(&buf[i]);
1255 }
1256
1257 static void mtip_set_timeout(struct driver_data *dd,
1258                                         struct host_to_dev_fis *fis,
1259                                         unsigned int *timeout, u8 erasemode)
1260 {
1261         switch (fis->command) {
1262         case ATA_CMD_DOWNLOAD_MICRO:
1263                 *timeout = 120000; /* 2 minutes */
1264                 break;
1265         case ATA_CMD_SEC_ERASE_UNIT:
1266         case 0xFC:
1267                 if (erasemode)
1268                         *timeout = ((*(dd->port->identify + 90) * 2) * 60000);
1269                 else
1270                         *timeout = ((*(dd->port->identify + 89) * 2) * 60000);
1271                 break;
1272         case ATA_CMD_STANDBYNOW1:
1273                 *timeout = 120000;  /* 2 minutes */
1274                 break;
1275         case 0xF7:
1276         case 0xFA:
1277                 *timeout = 60000;  /* 60 seconds */
1278                 break;
1279         case ATA_CMD_SMART:
1280                 *timeout = 15000;  /* 15 seconds */
1281                 break;
1282         default:
1283                 *timeout = MTIP_IOCTL_CMD_TIMEOUT_MS;
1284                 break;
1285         }
1286 }
1287
1288 /*
1289  * Request the device identity information.
1290  *
1291  * If a user space buffer is not specified, i.e. is NULL, the
1292  * identify information is still read from the drive and placed
1293  * into the identify data buffer (@e port->identify) in the
1294  * port data structure.
1295  * When the identify buffer contains valid identify information @e
1296  * port->identify_valid is non-zero.
1297  *
1298  * @port         Pointer to the port structure.
1299  * @user_buffer  A user space buffer where the identify data should be
1300  *                    copied.
1301  *
1302  * return value
1303  *      0       Command completed successfully.
1304  *      -EFAULT An error occurred while coping data to the user buffer.
1305  *      -1      Command failed.
1306  */
1307 static int mtip_get_identify(struct mtip_port *port, void __user *user_buffer)
1308 {
1309         int rv = 0;
1310         struct host_to_dev_fis fis;
1311
1312         if (test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &port->dd->dd_flag))
1313                 return -EFAULT;
1314
1315         /* Build the FIS. */
1316         memset(&fis, 0, sizeof(struct host_to_dev_fis));
1317         fis.type        = 0x27;
1318         fis.opts        = 1 << 7;
1319         fis.command     = ATA_CMD_ID_ATA;
1320
1321         /* Set the identify information as invalid. */
1322         port->identify_valid = 0;
1323
1324         /* Clear the identify information. */
1325         memset(port->identify, 0, sizeof(u16) * ATA_ID_WORDS);
1326
1327         /* Execute the command. */
1328         if (mtip_exec_internal_command(port,
1329                                 &fis,
1330                                 5,
1331                                 port->identify_dma,
1332                                 sizeof(u16) * ATA_ID_WORDS,
1333                                 0,
1334                                 MTIP_INT_CMD_TIMEOUT_MS)
1335                                 < 0) {
1336                 rv = -1;
1337                 goto out;
1338         }
1339
1340         /*
1341          * Perform any necessary byte-swapping.  Yes, the kernel does in fact
1342          * perform field-sensitive swapping on the string fields.
1343          * See the kernel use of ata_id_string() for proof of this.
1344          */
1345 #ifdef __LITTLE_ENDIAN
1346         ata_swap_string(port->identify + 27, 40);  /* model string*/
1347         ata_swap_string(port->identify + 23, 8);   /* firmware string*/
1348         ata_swap_string(port->identify + 10, 20);  /* serial# string*/
1349 #else
1350         {
1351                 int i;
1352                 for (i = 0; i < ATA_ID_WORDS; i++)
1353                         port->identify[i] = le16_to_cpu(port->identify[i]);
1354         }
1355 #endif
1356
1357         /* Check security locked state */
1358         if (port->identify[128] & 0x4)
1359                 set_bit(MTIP_DDF_SEC_LOCK_BIT, &port->dd->dd_flag);
1360         else
1361                 clear_bit(MTIP_DDF_SEC_LOCK_BIT, &port->dd->dd_flag);
1362
1363 #ifdef MTIP_TRIM /* Disabling TRIM support temporarily */
1364         /* Demux ID.DRAT & ID.RZAT to determine trim support */
1365         if (port->identify[69] & (1 << 14) && port->identify[69] & (1 << 5))
1366                 port->dd->trim_supp = true;
1367         else
1368 #endif
1369                 port->dd->trim_supp = false;
1370
1371         /* Set the identify buffer as valid. */
1372         port->identify_valid = 1;
1373
1374         if (user_buffer) {
1375                 if (copy_to_user(
1376                         user_buffer,
1377                         port->identify,
1378                         ATA_ID_WORDS * sizeof(u16))) {
1379                         rv = -EFAULT;
1380                         goto out;
1381                 }
1382         }
1383
1384 out:
1385         return rv;
1386 }
1387
1388 /*
1389  * Issue a standby immediate command to the device.
1390  *
1391  * @port Pointer to the port structure.
1392  *
1393  * return value
1394  *      0       Command was executed successfully.
1395  *      -1      An error occurred while executing the command.
1396  */
1397 static int mtip_standby_immediate(struct mtip_port *port)
1398 {
1399         int rv;
1400         struct host_to_dev_fis  fis;
1401         unsigned long start;
1402         unsigned int timeout;
1403
1404         /* Build the FIS. */
1405         memset(&fis, 0, sizeof(struct host_to_dev_fis));
1406         fis.type        = 0x27;
1407         fis.opts        = 1 << 7;
1408         fis.command     = ATA_CMD_STANDBYNOW1;
1409
1410         mtip_set_timeout(port->dd, &fis, &timeout, 0);
1411
1412         start = jiffies;
1413         rv = mtip_exec_internal_command(port,
1414                                         &fis,
1415                                         5,
1416                                         0,
1417                                         0,
1418                                         0,
1419                                         timeout);
1420         dbg_printk(MTIP_DRV_NAME "Time taken to complete standby cmd: %d ms\n",
1421                         jiffies_to_msecs(jiffies - start));
1422         if (rv)
1423                 dev_warn(&port->dd->pdev->dev,
1424                         "STANDBY IMMEDIATE command failed.\n");
1425
1426         return rv;
1427 }
1428
1429 /*
1430  * Issue a READ LOG EXT command to the device.
1431  *
1432  * @port        pointer to the port structure.
1433  * @page        page number to fetch
1434  * @buffer      pointer to buffer
1435  * @buffer_dma  dma address corresponding to @buffer
1436  * @sectors     page length to fetch, in sectors
1437  *
1438  * return value
1439  *      @rv     return value from mtip_exec_internal_command()
1440  */
1441 static int mtip_read_log_page(struct mtip_port *port, u8 page, u16 *buffer,
1442                                 dma_addr_t buffer_dma, unsigned int sectors)
1443 {
1444         struct host_to_dev_fis fis;
1445
1446         memset(&fis, 0, sizeof(struct host_to_dev_fis));
1447         fis.type        = 0x27;
1448         fis.opts        = 1 << 7;
1449         fis.command     = ATA_CMD_READ_LOG_EXT;
1450         fis.sect_count  = sectors & 0xFF;
1451         fis.sect_cnt_ex = (sectors >> 8) & 0xFF;
1452         fis.lba_low     = page;
1453         fis.lba_mid     = 0;
1454         fis.device      = ATA_DEVICE_OBS;
1455
1456         memset(buffer, 0, sectors * ATA_SECT_SIZE);
1457
1458         return mtip_exec_internal_command(port,
1459                                         &fis,
1460                                         5,
1461                                         buffer_dma,
1462                                         sectors * ATA_SECT_SIZE,
1463                                         0,
1464                                         MTIP_INT_CMD_TIMEOUT_MS);
1465 }
1466
1467 /*
1468  * Issue a SMART READ DATA command to the device.
1469  *
1470  * @port        pointer to the port structure.
1471  * @buffer      pointer to buffer
1472  * @buffer_dma  dma address corresponding to @buffer
1473  *
1474  * return value
1475  *      @rv     return value from mtip_exec_internal_command()
1476  */
1477 static int mtip_get_smart_data(struct mtip_port *port, u8 *buffer,
1478                                         dma_addr_t buffer_dma)
1479 {
1480         struct host_to_dev_fis fis;
1481
1482         memset(&fis, 0, sizeof(struct host_to_dev_fis));
1483         fis.type        = 0x27;
1484         fis.opts        = 1 << 7;
1485         fis.command     = ATA_CMD_SMART;
1486         fis.features    = 0xD0;
1487         fis.sect_count  = 1;
1488         fis.lba_mid     = 0x4F;
1489         fis.lba_hi      = 0xC2;
1490         fis.device      = ATA_DEVICE_OBS;
1491
1492         return mtip_exec_internal_command(port,
1493                                         &fis,
1494                                         5,
1495                                         buffer_dma,
1496                                         ATA_SECT_SIZE,
1497                                         0,
1498                                         15000);
1499 }
1500
1501 /*
1502  * Get the value of a smart attribute
1503  *
1504  * @port        pointer to the port structure
1505  * @id          attribute number
1506  * @attrib      pointer to return attrib information corresponding to @id
1507  *
1508  * return value
1509  *      -EINVAL NULL buffer passed or unsupported attribute @id.
1510  *      -EPERM  Identify data not valid, SMART not supported or not enabled
1511  */
1512 static int mtip_get_smart_attr(struct mtip_port *port, unsigned int id,
1513                                                 struct smart_attr *attrib)
1514 {
1515         int rv, i;
1516         struct smart_attr *pattr;
1517
1518         if (!attrib)
1519                 return -EINVAL;
1520
1521         if (!port->identify_valid) {
1522                 dev_warn(&port->dd->pdev->dev, "IDENTIFY DATA not valid\n");
1523                 return -EPERM;
1524         }
1525         if (!(port->identify[82] & 0x1)) {
1526                 dev_warn(&port->dd->pdev->dev, "SMART not supported\n");
1527                 return -EPERM;
1528         }
1529         if (!(port->identify[85] & 0x1)) {
1530                 dev_warn(&port->dd->pdev->dev, "SMART not enabled\n");
1531                 return -EPERM;
1532         }
1533
1534         memset(port->smart_buf, 0, ATA_SECT_SIZE);
1535         rv = mtip_get_smart_data(port, port->smart_buf, port->smart_buf_dma);
1536         if (rv) {
1537                 dev_warn(&port->dd->pdev->dev, "Failed to ge SMART data\n");
1538                 return rv;
1539         }
1540
1541         pattr = (struct smart_attr *)(port->smart_buf + 2);
1542         for (i = 0; i < 29; i++, pattr++)
1543                 if (pattr->attr_id == id) {
1544                         memcpy(attrib, pattr, sizeof(struct smart_attr));
1545                         break;
1546                 }
1547
1548         if (i == 29) {
1549                 dev_warn(&port->dd->pdev->dev,
1550                         "Query for invalid SMART attribute ID\n");
1551                 rv = -EINVAL;
1552         }
1553
1554         return rv;
1555 }
1556
1557 /*
1558  * Trim unused sectors
1559  *
1560  * @dd          pointer to driver_data structure
1561  * @lba         starting lba
1562  * @len         # of 512b sectors to trim
1563  *
1564  * return value
1565  *      -ENOMEM         Out of dma memory
1566  *      -EINVAL         Invalid parameters passed in, trim not supported
1567  *      -EIO            Error submitting trim request to hw
1568  */
1569 static int mtip_send_trim(struct driver_data *dd, unsigned int lba,
1570                                 unsigned int len)
1571 {
1572         int i, rv = 0;
1573         u64 tlba, tlen, sect_left;
1574         struct mtip_trim_entry *buf;
1575         dma_addr_t dma_addr;
1576         struct host_to_dev_fis fis;
1577
1578         if (!len || dd->trim_supp == false)
1579                 return -EINVAL;
1580
1581         /* Trim request too big */
1582         WARN_ON(len > (MTIP_MAX_TRIM_ENTRY_LEN * MTIP_MAX_TRIM_ENTRIES));
1583
1584         /* Trim request not aligned on 4k boundary */
1585         WARN_ON(len % 8 != 0);
1586
1587         /* Warn if vu_trim structure is too big */
1588         WARN_ON(sizeof(struct mtip_trim) > ATA_SECT_SIZE);
1589
1590         /* Allocate a DMA buffer for the trim structure */
1591         buf = dmam_alloc_coherent(&dd->pdev->dev, ATA_SECT_SIZE, &dma_addr,
1592                                                                 GFP_KERNEL);
1593         if (!buf)
1594                 return -ENOMEM;
1595         memset(buf, 0, ATA_SECT_SIZE);
1596
1597         for (i = 0, sect_left = len, tlba = lba;
1598                         i < MTIP_MAX_TRIM_ENTRIES && sect_left;
1599                         i++) {
1600                 tlen = (sect_left >= MTIP_MAX_TRIM_ENTRY_LEN ?
1601                                         MTIP_MAX_TRIM_ENTRY_LEN :
1602                                         sect_left);
1603                 buf[i].lba = __force_bit2int cpu_to_le32(tlba);
1604                 buf[i].range = __force_bit2int cpu_to_le16(tlen);
1605                 tlba += tlen;
1606                 sect_left -= tlen;
1607         }
1608         WARN_ON(sect_left != 0);
1609
1610         /* Build the fis */
1611         memset(&fis, 0, sizeof(struct host_to_dev_fis));
1612         fis.type       = 0x27;
1613         fis.opts       = 1 << 7;
1614         fis.command    = 0xfb;
1615         fis.features   = 0x60;
1616         fis.sect_count = 1;
1617         fis.device     = ATA_DEVICE_OBS;
1618
1619         if (mtip_exec_internal_command(dd->port,
1620                                         &fis,
1621                                         5,
1622                                         dma_addr,
1623                                         ATA_SECT_SIZE,
1624                                         0,
1625                                         MTIP_TRIM_TIMEOUT_MS) < 0)
1626                 rv = -EIO;
1627
1628         dmam_free_coherent(&dd->pdev->dev, ATA_SECT_SIZE, buf, dma_addr);
1629         return rv;
1630 }
1631
1632 /*
1633  * Get the drive capacity.
1634  *
1635  * @dd      Pointer to the device data structure.
1636  * @sectors Pointer to the variable that will receive the sector count.
1637  *
1638  * return value
1639  *      1 Capacity was returned successfully.
1640  *      0 The identify information is invalid.
1641  */
1642 static bool mtip_hw_get_capacity(struct driver_data *dd, sector_t *sectors)
1643 {
1644         struct mtip_port *port = dd->port;
1645         u64 total, raw0, raw1, raw2, raw3;
1646         raw0 = port->identify[100];
1647         raw1 = port->identify[101];
1648         raw2 = port->identify[102];
1649         raw3 = port->identify[103];
1650         total = raw0 | raw1<<16 | raw2<<32 | raw3<<48;
1651         *sectors = total;
1652         return (bool) !!port->identify_valid;
1653 }
1654
1655 /*
1656  * Display the identify command data.
1657  *
1658  * @port Pointer to the port data structure.
1659  *
1660  * return value
1661  *      None
1662  */
1663 static void mtip_dump_identify(struct mtip_port *port)
1664 {
1665         sector_t sectors;
1666         unsigned short revid;
1667         char cbuf[42];
1668
1669         if (!port->identify_valid)
1670                 return;
1671
1672         strlcpy(cbuf, (char *)(port->identify+10), 21);
1673         dev_info(&port->dd->pdev->dev,
1674                 "Serial No.: %s\n", cbuf);
1675
1676         strlcpy(cbuf, (char *)(port->identify+23), 9);
1677         dev_info(&port->dd->pdev->dev,
1678                 "Firmware Ver.: %s\n", cbuf);
1679
1680         strlcpy(cbuf, (char *)(port->identify+27), 41);
1681         dev_info(&port->dd->pdev->dev, "Model: %s\n", cbuf);
1682
1683         dev_info(&port->dd->pdev->dev, "Security: %04x %s\n",
1684                 port->identify[128],
1685                 port->identify[128] & 0x4 ? "(LOCKED)" : "");
1686
1687         if (mtip_hw_get_capacity(port->dd, &sectors))
1688                 dev_info(&port->dd->pdev->dev,
1689                         "Capacity: %llu sectors (%llu MB)\n",
1690                          (u64)sectors,
1691                          ((u64)sectors) * ATA_SECT_SIZE >> 20);
1692
1693         pci_read_config_word(port->dd->pdev, PCI_REVISION_ID, &revid);
1694         switch (revid & 0xFF) {
1695         case 0x1:
1696                 strlcpy(cbuf, "A0", 3);
1697                 break;
1698         case 0x3:
1699                 strlcpy(cbuf, "A2", 3);
1700                 break;
1701         default:
1702                 strlcpy(cbuf, "?", 2);
1703                 break;
1704         }
1705         dev_info(&port->dd->pdev->dev,
1706                 "Card Type: %s\n", cbuf);
1707 }
1708
1709 /*
1710  * Map the commands scatter list into the command table.
1711  *
1712  * @command Pointer to the command.
1713  * @nents Number of scatter list entries.
1714  *
1715  * return value
1716  *      None
1717  */
1718 static inline void fill_command_sg(struct driver_data *dd,
1719                                 struct mtip_cmd *command,
1720                                 int nents)
1721 {
1722         int n;
1723         unsigned int dma_len;
1724         struct mtip_cmd_sg *command_sg;
1725         struct scatterlist *sg = command->sg;
1726
1727         command_sg = command->command + AHCI_CMD_TBL_HDR_SZ;
1728
1729         for (n = 0; n < nents; n++) {
1730                 dma_len = sg_dma_len(sg);
1731                 if (dma_len > 0x400000)
1732                         dev_err(&dd->pdev->dev,
1733                                 "DMA segment length truncated\n");
1734                 command_sg->info = __force_bit2int
1735                         cpu_to_le32((dma_len-1) & 0x3FFFFF);
1736                 command_sg->dba = __force_bit2int
1737                         cpu_to_le32(sg_dma_address(sg));
1738                 command_sg->dba_upper = __force_bit2int
1739                         cpu_to_le32((sg_dma_address(sg) >> 16) >> 16);
1740                 command_sg++;
1741                 sg++;
1742         }
1743 }
1744
1745 /*
1746  * @brief Execute a drive command.
1747  *
1748  * return value 0 The command completed successfully.
1749  * return value -1 An error occurred while executing the command.
1750  */
1751 static int exec_drive_task(struct mtip_port *port, u8 *command)
1752 {
1753         struct host_to_dev_fis  fis;
1754         struct host_to_dev_fis *reply = (port->rxfis + RX_FIS_D2H_REG);
1755         unsigned int to;
1756
1757         /* Build the FIS. */
1758         memset(&fis, 0, sizeof(struct host_to_dev_fis));
1759         fis.type        = 0x27;
1760         fis.opts        = 1 << 7;
1761         fis.command     = command[0];
1762         fis.features    = command[1];
1763         fis.sect_count  = command[2];
1764         fis.sector      = command[3];
1765         fis.cyl_low     = command[4];
1766         fis.cyl_hi      = command[5];
1767         fis.device      = command[6] & ~0x10; /* Clear the dev bit*/
1768
1769         mtip_set_timeout(port->dd, &fis, &to, 0);
1770
1771         dbg_printk(MTIP_DRV_NAME " %s: User Command: cmd %x, feat %x, nsect %x, sect %x, lcyl %x, hcyl %x, sel %x\n",
1772                 __func__,
1773                 command[0],
1774                 command[1],
1775                 command[2],
1776                 command[3],
1777                 command[4],
1778                 command[5],
1779                 command[6]);
1780
1781         /* Execute the command. */
1782         if (mtip_exec_internal_command(port,
1783                                  &fis,
1784                                  5,
1785                                  0,
1786                                  0,
1787                                  0,
1788                                  to) < 0) {
1789                 return -1;
1790         }
1791
1792         command[0] = reply->command; /* Status*/
1793         command[1] = reply->features; /* Error*/
1794         command[4] = reply->cyl_low;
1795         command[5] = reply->cyl_hi;
1796
1797         dbg_printk(MTIP_DRV_NAME " %s: Completion Status: stat %x, err %x , cyl_lo %x cyl_hi %x\n",
1798                 __func__,
1799                 command[0],
1800                 command[1],
1801                 command[4],
1802                 command[5]);
1803
1804         return 0;
1805 }
1806
1807 /*
1808  * @brief Execute a drive command.
1809  *
1810  * @param port Pointer to the port data structure.
1811  * @param command Pointer to the user specified command parameters.
1812  * @param user_buffer Pointer to the user space buffer where read sector
1813  *                   data should be copied.
1814  *
1815  * return value 0 The command completed successfully.
1816  * return value -EFAULT An error occurred while copying the completion
1817  *                 data to the user space buffer.
1818  * return value -1 An error occurred while executing the command.
1819  */
1820 static int exec_drive_command(struct mtip_port *port, u8 *command,
1821                                 void __user *user_buffer)
1822 {
1823         struct host_to_dev_fis  fis;
1824         struct host_to_dev_fis *reply;
1825         u8 *buf = NULL;
1826         dma_addr_t dma_addr = 0;
1827         int rv = 0, xfer_sz = command[3];
1828         unsigned int to;
1829
1830         if (xfer_sz) {
1831                 if (!user_buffer)
1832                         return -EFAULT;
1833
1834                 buf = dmam_alloc_coherent(&port->dd->pdev->dev,
1835                                 ATA_SECT_SIZE * xfer_sz,
1836                                 &dma_addr,
1837                                 GFP_KERNEL);
1838                 if (!buf) {
1839                         dev_err(&port->dd->pdev->dev,
1840                                 "Memory allocation failed (%d bytes)\n",
1841                                 ATA_SECT_SIZE * xfer_sz);
1842                         return -ENOMEM;
1843                 }
1844                 memset(buf, 0, ATA_SECT_SIZE * xfer_sz);
1845         }
1846
1847         /* Build the FIS. */
1848         memset(&fis, 0, sizeof(struct host_to_dev_fis));
1849         fis.type        = 0x27;
1850         fis.opts        = 1 << 7;
1851         fis.command     = command[0];
1852         fis.features    = command[2];
1853         fis.sect_count  = command[3];
1854         if (fis.command == ATA_CMD_SMART) {
1855                 fis.sector      = command[1];
1856                 fis.cyl_low     = 0x4F;
1857                 fis.cyl_hi      = 0xC2;
1858         }
1859
1860         mtip_set_timeout(port->dd, &fis, &to, 0);
1861
1862         if (xfer_sz)
1863                 reply = (port->rxfis + RX_FIS_PIO_SETUP);
1864         else
1865                 reply = (port->rxfis + RX_FIS_D2H_REG);
1866
1867         dbg_printk(MTIP_DRV_NAME
1868                 " %s: User Command: cmd %x, sect %x, "
1869                 "feat %x, sectcnt %x\n",
1870                 __func__,
1871                 command[0],
1872                 command[1],
1873                 command[2],
1874                 command[3]);
1875
1876         /* Execute the command. */
1877         if (mtip_exec_internal_command(port,
1878                                 &fis,
1879                                  5,
1880                                  (xfer_sz ? dma_addr : 0),
1881                                  (xfer_sz ? ATA_SECT_SIZE * xfer_sz : 0),
1882                                  0,
1883                                  to)
1884                                  < 0) {
1885                 rv = -EFAULT;
1886                 goto exit_drive_command;
1887         }
1888
1889         /* Collect the completion status. */
1890         command[0] = reply->command; /* Status*/
1891         command[1] = reply->features; /* Error*/
1892         command[2] = reply->sect_count;
1893
1894         dbg_printk(MTIP_DRV_NAME
1895                 " %s: Completion Status: stat %x, "
1896                 "err %x, nsect %x\n",
1897                 __func__,
1898                 command[0],
1899                 command[1],
1900                 command[2]);
1901
1902         if (xfer_sz) {
1903                 if (copy_to_user(user_buffer,
1904                                  buf,
1905                                  ATA_SECT_SIZE * command[3])) {
1906                         rv = -EFAULT;
1907                         goto exit_drive_command;
1908                 }
1909         }
1910 exit_drive_command:
1911         if (buf)
1912                 dmam_free_coherent(&port->dd->pdev->dev,
1913                                 ATA_SECT_SIZE * xfer_sz, buf, dma_addr);
1914         return rv;
1915 }
1916
1917 /*
1918  *  Indicates whether a command has a single sector payload.
1919  *
1920  *  @command passed to the device to perform the certain event.
1921  *  @features passed to the device to perform the certain event.
1922  *
1923  *  return value
1924  *      1       command is one that always has a single sector payload,
1925  *              regardless of the value in the Sector Count field.
1926  *      0       otherwise
1927  *
1928  */
1929 static unsigned int implicit_sector(unsigned char command,
1930                                     unsigned char features)
1931 {
1932         unsigned int rv = 0;
1933
1934         /* list of commands that have an implicit sector count of 1 */
1935         switch (command) {
1936         case ATA_CMD_SEC_SET_PASS:
1937         case ATA_CMD_SEC_UNLOCK:
1938         case ATA_CMD_SEC_ERASE_PREP:
1939         case ATA_CMD_SEC_ERASE_UNIT:
1940         case ATA_CMD_SEC_FREEZE_LOCK:
1941         case ATA_CMD_SEC_DISABLE_PASS:
1942         case ATA_CMD_PMP_READ:
1943         case ATA_CMD_PMP_WRITE:
1944                 rv = 1;
1945                 break;
1946         case ATA_CMD_SET_MAX:
1947                 if (features == ATA_SET_MAX_UNLOCK)
1948                         rv = 1;
1949                 break;
1950         case ATA_CMD_SMART:
1951                 if ((features == ATA_SMART_READ_VALUES) ||
1952                                 (features == ATA_SMART_READ_THRESHOLDS))
1953                         rv = 1;
1954                 break;
1955         case ATA_CMD_CONF_OVERLAY:
1956                 if ((features == ATA_DCO_IDENTIFY) ||
1957                                 (features == ATA_DCO_SET))
1958                         rv = 1;
1959                 break;
1960         }
1961         return rv;
1962 }
1963
1964 /*
1965  * Executes a taskfile
1966  * See ide_taskfile_ioctl() for derivation
1967  */
1968 static int exec_drive_taskfile(struct driver_data *dd,
1969                                void __user *buf,
1970                                ide_task_request_t *req_task,
1971                                int outtotal)
1972 {
1973         struct host_to_dev_fis  fis;
1974         struct host_to_dev_fis *reply;
1975         u8 *outbuf = NULL;
1976         u8 *inbuf = NULL;
1977         dma_addr_t outbuf_dma = 0;
1978         dma_addr_t inbuf_dma = 0;
1979         dma_addr_t dma_buffer = 0;
1980         int err = 0;
1981         unsigned int taskin = 0;
1982         unsigned int taskout = 0;
1983         u8 nsect = 0;
1984         unsigned int timeout;
1985         unsigned int force_single_sector;
1986         unsigned int transfer_size;
1987         unsigned long task_file_data;
1988         int intotal = outtotal + req_task->out_size;
1989         int erasemode = 0;
1990
1991         taskout = req_task->out_size;
1992         taskin = req_task->in_size;
1993         /* 130560 = 512 * 0xFF*/
1994         if (taskin > 130560 || taskout > 130560)
1995                 return -EINVAL;
1996
1997         if (taskout) {
1998                 outbuf = memdup_user(buf + outtotal, taskout);
1999                 if (IS_ERR(outbuf))
2000                         return PTR_ERR(outbuf);
2001
2002                 outbuf_dma = pci_map_single(dd->pdev,
2003                                          outbuf,
2004                                          taskout,
2005                                          DMA_TO_DEVICE);
2006                 if (pci_dma_mapping_error(dd->pdev, outbuf_dma)) {
2007                         err = -ENOMEM;
2008                         goto abort;
2009                 }
2010                 dma_buffer = outbuf_dma;
2011         }
2012
2013         if (taskin) {
2014                 inbuf = memdup_user(buf + intotal, taskin);
2015                 if (IS_ERR(inbuf)) {
2016                         err = PTR_ERR(inbuf);
2017                         inbuf = NULL;
2018                         goto abort;
2019                 }
2020                 inbuf_dma = pci_map_single(dd->pdev,
2021                                          inbuf,
2022                                          taskin, DMA_FROM_DEVICE);
2023                 if (pci_dma_mapping_error(dd->pdev, inbuf_dma)) {
2024                         err = -ENOMEM;
2025                         goto abort;
2026                 }
2027                 dma_buffer = inbuf_dma;
2028         }
2029
2030         /* only supports PIO and non-data commands from this ioctl. */
2031         switch (req_task->data_phase) {
2032         case TASKFILE_OUT:
2033                 nsect = taskout / ATA_SECT_SIZE;
2034                 reply = (dd->port->rxfis + RX_FIS_PIO_SETUP);
2035                 break;
2036         case TASKFILE_IN:
2037                 reply = (dd->port->rxfis + RX_FIS_PIO_SETUP);
2038                 break;
2039         case TASKFILE_NO_DATA:
2040                 reply = (dd->port->rxfis + RX_FIS_D2H_REG);
2041                 break;
2042         default:
2043                 err = -EINVAL;
2044                 goto abort;
2045         }
2046
2047         /* Build the FIS. */
2048         memset(&fis, 0, sizeof(struct host_to_dev_fis));
2049
2050         fis.type        = 0x27;
2051         fis.opts        = 1 << 7;
2052         fis.command     = req_task->io_ports[7];
2053         fis.features    = req_task->io_ports[1];
2054         fis.sect_count  = req_task->io_ports[2];
2055         fis.lba_low     = req_task->io_ports[3];
2056         fis.lba_mid     = req_task->io_ports[4];
2057         fis.lba_hi      = req_task->io_ports[5];
2058          /* Clear the dev bit*/
2059         fis.device      = req_task->io_ports[6] & ~0x10;
2060
2061         if ((req_task->in_flags.all == 0) && (req_task->out_flags.all & 1)) {
2062                 req_task->in_flags.all  =
2063                         IDE_TASKFILE_STD_IN_FLAGS |
2064                         (IDE_HOB_STD_IN_FLAGS << 8);
2065                 fis.lba_low_ex          = req_task->hob_ports[3];
2066                 fis.lba_mid_ex          = req_task->hob_ports[4];
2067                 fis.lba_hi_ex           = req_task->hob_ports[5];
2068                 fis.features_ex         = req_task->hob_ports[1];
2069                 fis.sect_cnt_ex         = req_task->hob_ports[2];
2070
2071         } else {
2072                 req_task->in_flags.all = IDE_TASKFILE_STD_IN_FLAGS;
2073         }
2074
2075         force_single_sector = implicit_sector(fis.command, fis.features);
2076
2077         if ((taskin || taskout) && (!fis.sect_count)) {
2078                 if (nsect)
2079                         fis.sect_count = nsect;
2080                 else {
2081                         if (!force_single_sector) {
2082                                 dev_warn(&dd->pdev->dev,
2083                                         "data movement but "
2084                                         "sect_count is 0\n");
2085                                         err = -EINVAL;
2086                                         goto abort;
2087                         }
2088                 }
2089         }
2090
2091         dbg_printk(MTIP_DRV_NAME
2092                 " %s: cmd %x, feat %x, nsect %x,"
2093                 " sect/lbal %x, lcyl/lbam %x, hcyl/lbah %x,"
2094                 " head/dev %x\n",
2095                 __func__,
2096                 fis.command,
2097                 fis.features,
2098                 fis.sect_count,
2099                 fis.lba_low,
2100                 fis.lba_mid,
2101                 fis.lba_hi,
2102                 fis.device);
2103
2104         /* check for erase mode support during secure erase.*/
2105         if ((fis.command == ATA_CMD_SEC_ERASE_UNIT) && outbuf &&
2106                                         (outbuf[0] & MTIP_SEC_ERASE_MODE)) {
2107                 erasemode = 1;
2108         }
2109
2110         mtip_set_timeout(dd, &fis, &timeout, erasemode);
2111
2112         /* Determine the correct transfer size.*/
2113         if (force_single_sector)
2114                 transfer_size = ATA_SECT_SIZE;
2115         else
2116                 transfer_size = ATA_SECT_SIZE * fis.sect_count;
2117
2118         /* Execute the command.*/
2119         if (mtip_exec_internal_command(dd->port,
2120                                  &fis,
2121                                  5,
2122                                  dma_buffer,
2123                                  transfer_size,
2124                                  0,
2125                                  timeout) < 0) {
2126                 err = -EIO;
2127                 goto abort;
2128         }
2129
2130         task_file_data = readl(dd->port->mmio+PORT_TFDATA);
2131
2132         if ((req_task->data_phase == TASKFILE_IN) && !(task_file_data & 1)) {
2133                 reply = dd->port->rxfis + RX_FIS_PIO_SETUP;
2134                 req_task->io_ports[7] = reply->control;
2135         } else {
2136                 reply = dd->port->rxfis + RX_FIS_D2H_REG;
2137                 req_task->io_ports[7] = reply->command;
2138         }
2139
2140         /* reclaim the DMA buffers.*/
2141         if (inbuf_dma)
2142                 pci_unmap_single(dd->pdev, inbuf_dma,
2143                         taskin, DMA_FROM_DEVICE);
2144         if (outbuf_dma)
2145                 pci_unmap_single(dd->pdev, outbuf_dma,
2146                         taskout, DMA_TO_DEVICE);
2147         inbuf_dma  = 0;
2148         outbuf_dma = 0;
2149
2150         /* return the ATA registers to the caller.*/
2151         req_task->io_ports[1] = reply->features;
2152         req_task->io_ports[2] = reply->sect_count;
2153         req_task->io_ports[3] = reply->lba_low;
2154         req_task->io_ports[4] = reply->lba_mid;
2155         req_task->io_ports[5] = reply->lba_hi;
2156         req_task->io_ports[6] = reply->device;
2157
2158         if (req_task->out_flags.all & 1)  {
2159
2160                 req_task->hob_ports[3] = reply->lba_low_ex;
2161                 req_task->hob_ports[4] = reply->lba_mid_ex;
2162                 req_task->hob_ports[5] = reply->lba_hi_ex;
2163                 req_task->hob_ports[1] = reply->features_ex;
2164                 req_task->hob_ports[2] = reply->sect_cnt_ex;
2165         }
2166         dbg_printk(MTIP_DRV_NAME
2167                 " %s: Completion: stat %x,"
2168                 "err %x, sect_cnt %x, lbalo %x,"
2169                 "lbamid %x, lbahi %x, dev %x\n",
2170                 __func__,
2171                 req_task->io_ports[7],
2172                 req_task->io_ports[1],
2173                 req_task->io_ports[2],
2174                 req_task->io_ports[3],
2175                 req_task->io_ports[4],
2176                 req_task->io_ports[5],
2177                 req_task->io_ports[6]);
2178
2179         if (taskout) {
2180                 if (copy_to_user(buf + outtotal, outbuf, taskout)) {
2181                         err = -EFAULT;
2182                         goto abort;
2183                 }
2184         }
2185         if (taskin) {
2186                 if (copy_to_user(buf + intotal, inbuf, taskin)) {
2187                         err = -EFAULT;
2188                         goto abort;
2189                 }
2190         }
2191 abort:
2192         if (inbuf_dma)
2193                 pci_unmap_single(dd->pdev, inbuf_dma,
2194                                         taskin, DMA_FROM_DEVICE);
2195         if (outbuf_dma)
2196                 pci_unmap_single(dd->pdev, outbuf_dma,
2197                                         taskout, DMA_TO_DEVICE);
2198         kfree(outbuf);
2199         kfree(inbuf);
2200
2201         return err;
2202 }
2203
2204 /*
2205  * Handle IOCTL calls from the Block Layer.
2206  *
2207  * This function is called by the Block Layer when it receives an IOCTL
2208  * command that it does not understand. If the IOCTL command is not supported
2209  * this function returns -ENOTTY.
2210  *
2211  * @dd  Pointer to the driver data structure.
2212  * @cmd IOCTL command passed from the Block Layer.
2213  * @arg IOCTL argument passed from the Block Layer.
2214  *
2215  * return value
2216  *      0       The IOCTL completed successfully.
2217  *      -ENOTTY The specified command is not supported.
2218  *      -EFAULT An error occurred copying data to a user space buffer.
2219  *      -EIO    An error occurred while executing the command.
2220  */
2221 static int mtip_hw_ioctl(struct driver_data *dd, unsigned int cmd,
2222                          unsigned long arg)
2223 {
2224         switch (cmd) {
2225         case HDIO_GET_IDENTITY:
2226         {
2227                 if (copy_to_user((void __user *)arg, dd->port->identify,
2228                                                 sizeof(u16) * ATA_ID_WORDS))
2229                         return -EFAULT;
2230                 break;
2231         }
2232         case HDIO_DRIVE_CMD:
2233         {
2234                 u8 drive_command[4];
2235
2236                 /* Copy the user command info to our buffer. */
2237                 if (copy_from_user(drive_command,
2238                                          (void __user *) arg,
2239                                          sizeof(drive_command)))
2240                         return -EFAULT;
2241
2242                 /* Execute the drive command. */
2243                 if (exec_drive_command(dd->port,
2244                                          drive_command,
2245                                          (void __user *) (arg+4)))
2246                         return -EIO;
2247
2248                 /* Copy the status back to the users buffer. */
2249                 if (copy_to_user((void __user *) arg,
2250                                          drive_command,
2251                                          sizeof(drive_command)))
2252                         return -EFAULT;
2253
2254                 break;
2255         }
2256         case HDIO_DRIVE_TASK:
2257         {
2258                 u8 drive_command[7];
2259
2260                 /* Copy the user command info to our buffer. */
2261                 if (copy_from_user(drive_command,
2262                                          (void __user *) arg,
2263                                          sizeof(drive_command)))
2264                         return -EFAULT;
2265
2266                 /* Execute the drive command. */
2267                 if (exec_drive_task(dd->port, drive_command))
2268                         return -EIO;
2269
2270                 /* Copy the status back to the users buffer. */
2271                 if (copy_to_user((void __user *) arg,
2272                                          drive_command,
2273                                          sizeof(drive_command)))
2274                         return -EFAULT;
2275
2276                 break;
2277         }
2278         case HDIO_DRIVE_TASKFILE: {
2279                 ide_task_request_t req_task;
2280                 int ret, outtotal;
2281
2282                 if (copy_from_user(&req_task, (void __user *) arg,
2283                                         sizeof(req_task)))
2284                         return -EFAULT;
2285
2286                 outtotal = sizeof(req_task);
2287
2288                 ret = exec_drive_taskfile(dd, (void __user *) arg,
2289                                                 &req_task, outtotal);
2290
2291                 if (copy_to_user((void __user *) arg, &req_task,
2292                                                         sizeof(req_task)))
2293                         return -EFAULT;
2294
2295                 return ret;
2296         }
2297
2298         default:
2299                 return -EINVAL;
2300         }
2301         return 0;
2302 }
2303
2304 /*
2305  * Submit an IO to the hw
2306  *
2307  * This function is called by the block layer to issue an io
2308  * to the device. Upon completion, the callback function will
2309  * be called with the data parameter passed as the callback data.
2310  *
2311  * @dd       Pointer to the driver data structure.
2312  * @start    First sector to read.
2313  * @nsect    Number of sectors to read.
2314  * @nents    Number of entries in scatter list for the read command.
2315  * @tag      The tag of this read command.
2316  * @callback Pointer to the function that should be called
2317  *           when the read completes.
2318  * @data     Callback data passed to the callback function
2319  *           when the read completes.
2320  * @dir      Direction (read or write)
2321  *
2322  * return value
2323  *      None
2324  */
2325 static void mtip_hw_submit_io(struct driver_data *dd, struct request *rq,
2326                               struct mtip_cmd *command, int nents,
2327                               struct blk_mq_hw_ctx *hctx)
2328 {
2329         struct host_to_dev_fis  *fis;
2330         struct mtip_port *port = dd->port;
2331         int dma_dir = rq_data_dir(rq) == READ ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
2332         u64 start = blk_rq_pos(rq);
2333         unsigned int nsect = blk_rq_sectors(rq);
2334
2335         /* Map the scatter list for DMA access */
2336         nents = dma_map_sg(&dd->pdev->dev, command->sg, nents, dma_dir);
2337
2338         prefetch(&port->flags);
2339
2340         command->scatter_ents = nents;
2341
2342         /*
2343          * The number of retries for this command before it is
2344          * reported as a failure to the upper layers.
2345          */
2346         command->retries = MTIP_MAX_RETRIES;
2347
2348         /* Fill out fis */
2349         fis = command->command;
2350         fis->type        = 0x27;
2351         fis->opts        = 1 << 7;
2352         if (dma_dir == DMA_FROM_DEVICE)
2353                 fis->command = ATA_CMD_FPDMA_READ;
2354         else
2355                 fis->command = ATA_CMD_FPDMA_WRITE;
2356         fis->lba_low     = start & 0xFF;
2357         fis->lba_mid     = (start >> 8) & 0xFF;
2358         fis->lba_hi      = (start >> 16) & 0xFF;
2359         fis->lba_low_ex  = (start >> 24) & 0xFF;
2360         fis->lba_mid_ex  = (start >> 32) & 0xFF;
2361         fis->lba_hi_ex   = (start >> 40) & 0xFF;
2362         fis->device      = 1 << 6;
2363         fis->features    = nsect & 0xFF;
2364         fis->features_ex = (nsect >> 8) & 0xFF;
2365         fis->sect_count  = ((rq->tag << 3) | (rq->tag >> 5));
2366         fis->sect_cnt_ex = 0;
2367         fis->control     = 0;
2368         fis->res2        = 0;
2369         fis->res3        = 0;
2370         fill_command_sg(dd, command, nents);
2371
2372         if (unlikely(command->unaligned))
2373                 fis->device |= 1 << 7;
2374
2375         /* Populate the command header */
2376         command->command_header->opts =
2377                         __force_bit2int cpu_to_le32(
2378                                 (nents << 16) | 5 | AHCI_CMD_PREFETCH);
2379         command->command_header->byte_count = 0;
2380
2381         /*
2382          * Set the completion function and data for the command
2383          * within this layer.
2384          */
2385         command->comp_data = dd;
2386         command->comp_func = mtip_async_complete;
2387         command->direction = dma_dir;
2388
2389         /*
2390          * To prevent this command from being issued
2391          * if an internal command is in progress or error handling is active.
2392          */
2393         if (unlikely(port->flags & MTIP_PF_PAUSE_IO)) {
2394                 set_bit(rq->tag, port->cmds_to_issue);
2395                 set_bit(MTIP_PF_ISSUE_CMDS_BIT, &port->flags);
2396                 return;
2397         }
2398
2399         /* Issue the command to the hardware */
2400         mtip_issue_ncq_command(port, rq->tag);
2401 }
2402
2403 /*
2404  * Sysfs status dump.
2405  *
2406  * @dev  Pointer to the device structure, passed by the kernrel.
2407  * @attr Pointer to the device_attribute structure passed by the kernel.
2408  * @buf  Pointer to the char buffer that will receive the stats info.
2409  *
2410  * return value
2411  *      The size, in bytes, of the data copied into buf.
2412  */
2413 static ssize_t mtip_hw_show_status(struct device *dev,
2414                                 struct device_attribute *attr,
2415                                 char *buf)
2416 {
2417         struct driver_data *dd = dev_to_disk(dev)->private_data;
2418         int size = 0;
2419
2420         if (test_bit(MTIP_DDF_OVER_TEMP_BIT, &dd->dd_flag))
2421                 size += sprintf(buf, "%s", "thermal_shutdown\n");
2422         else if (test_bit(MTIP_DDF_WRITE_PROTECT_BIT, &dd->dd_flag))
2423                 size += sprintf(buf, "%s", "write_protect\n");
2424         else
2425                 size += sprintf(buf, "%s", "online\n");
2426
2427         return size;
2428 }
2429
2430 static DEVICE_ATTR(status, S_IRUGO, mtip_hw_show_status, NULL);
2431
2432 /* debugsfs entries */
2433
2434 static ssize_t show_device_status(struct device_driver *drv, char *buf)
2435 {
2436         int size = 0;
2437         struct driver_data *dd, *tmp;
2438         unsigned long flags;
2439         char id_buf[42];
2440         u16 status = 0;
2441
2442         spin_lock_irqsave(&dev_lock, flags);
2443         size += sprintf(&buf[size], "Devices Present:\n");
2444         list_for_each_entry_safe(dd, tmp, &online_list, online_list) {
2445                 if (dd->pdev) {
2446                         if (dd->port &&
2447                             dd->port->identify &&
2448                             dd->port->identify_valid) {
2449                                 strlcpy(id_buf,
2450                                         (char *) (dd->port->identify + 10), 21);
2451                                 status = *(dd->port->identify + 141);
2452                         } else {
2453                                 memset(id_buf, 0, 42);
2454                                 status = 0;
2455                         }
2456
2457                         if (dd->port &&
2458                             test_bit(MTIP_PF_REBUILD_BIT, &dd->port->flags)) {
2459                                 size += sprintf(&buf[size],
2460                                         " device %s %s (ftl rebuild %d %%)\n",
2461                                         dev_name(&dd->pdev->dev),
2462                                         id_buf,
2463                                         status);
2464                         } else {
2465                                 size += sprintf(&buf[size],
2466                                         " device %s %s\n",
2467                                         dev_name(&dd->pdev->dev),
2468                                         id_buf);
2469                         }
2470                 }
2471         }
2472
2473         size += sprintf(&buf[size], "Devices Being Removed:\n");
2474         list_for_each_entry_safe(dd, tmp, &removing_list, remove_list) {
2475                 if (dd->pdev) {
2476                         if (dd->port &&
2477                             dd->port->identify &&
2478                             dd->port->identify_valid) {
2479                                 strlcpy(id_buf,
2480                                         (char *) (dd->port->identify+10), 21);
2481                                 status = *(dd->port->identify + 141);
2482                         } else {
2483                                 memset(id_buf, 0, 42);
2484                                 status = 0;
2485                         }
2486
2487                         if (dd->port &&
2488                             test_bit(MTIP_PF_REBUILD_BIT, &dd->port->flags)) {
2489                                 size += sprintf(&buf[size],
2490                                         " device %s %s (ftl rebuild %d %%)\n",
2491                                         dev_name(&dd->pdev->dev),
2492                                         id_buf,
2493                                         status);
2494                         } else {
2495                                 size += sprintf(&buf[size],
2496                                         " device %s %s\n",
2497                                         dev_name(&dd->pdev->dev),
2498                                         id_buf);
2499                         }
2500                 }
2501         }
2502         spin_unlock_irqrestore(&dev_lock, flags);
2503
2504         return size;
2505 }
2506
2507 static ssize_t mtip_hw_read_device_status(struct file *f, char __user *ubuf,
2508                                                 size_t len, loff_t *offset)
2509 {
2510         struct driver_data *dd =  (struct driver_data *)f->private_data;
2511         int size = *offset;
2512         char *buf;
2513         int rv = 0;
2514
2515         if (!len || *offset)
2516                 return 0;
2517
2518         buf = kzalloc(MTIP_DFS_MAX_BUF_SIZE, GFP_KERNEL);
2519         if (!buf) {
2520                 dev_err(&dd->pdev->dev,
2521                         "Memory allocation: status buffer\n");
2522                 return -ENOMEM;
2523         }
2524
2525         size += show_device_status(NULL, buf);
2526
2527         *offset = size <= len ? size : len;
2528         size = copy_to_user(ubuf, buf, *offset);
2529         if (size)
2530                 rv = -EFAULT;
2531
2532         kfree(buf);
2533         return rv ? rv : *offset;
2534 }
2535
2536 static ssize_t mtip_hw_read_registers(struct file *f, char __user *ubuf,
2537                                   size_t len, loff_t *offset)
2538 {
2539         struct driver_data *dd =  (struct driver_data *)f->private_data;
2540         char *buf;
2541         u32 group_allocated;
2542         int size = *offset;
2543         int n, rv = 0;
2544
2545         if (!len || size)
2546                 return 0;
2547
2548         buf = kzalloc(MTIP_DFS_MAX_BUF_SIZE, GFP_KERNEL);
2549         if (!buf) {
2550                 dev_err(&dd->pdev->dev,
2551                         "Memory allocation: register buffer\n");
2552                 return -ENOMEM;
2553         }
2554
2555         size += sprintf(&buf[size], "H/ S ACTive      : [ 0x");
2556
2557         for (n = dd->slot_groups-1; n >= 0; n--)
2558                 size += sprintf(&buf[size], "%08X ",
2559                                          readl(dd->port->s_active[n]));
2560
2561         size += sprintf(&buf[size], "]\n");
2562         size += sprintf(&buf[size], "H/ Command Issue : [ 0x");
2563
2564         for (n = dd->slot_groups-1; n >= 0; n--)
2565                 size += sprintf(&buf[size], "%08X ",
2566                                         readl(dd->port->cmd_issue[n]));
2567
2568         size += sprintf(&buf[size], "]\n");
2569         size += sprintf(&buf[size], "H/ Completed     : [ 0x");
2570
2571         for (n = dd->slot_groups-1; n >= 0; n--)
2572                 size += sprintf(&buf[size], "%08X ",
2573                                 readl(dd->port->completed[n]));
2574
2575         size += sprintf(&buf[size], "]\n");
2576         size += sprintf(&buf[size], "H/ PORT IRQ STAT : [ 0x%08X ]\n",
2577                                 readl(dd->port->mmio + PORT_IRQ_STAT));
2578         size += sprintf(&buf[size], "H/ HOST IRQ STAT : [ 0x%08X ]\n",
2579                                 readl(dd->mmio + HOST_IRQ_STAT));
2580         size += sprintf(&buf[size], "\n");
2581
2582         size += sprintf(&buf[size], "L/ Commands in Q : [ 0x");
2583
2584         for (n = dd->slot_groups-1; n >= 0; n--) {
2585                 if (sizeof(long) > sizeof(u32))
2586                         group_allocated =
2587                                 dd->port->cmds_to_issue[n/2] >> (32*(n&1));
2588                 else
2589                         group_allocated = dd->port->cmds_to_issue[n];
2590                 size += sprintf(&buf[size], "%08X ", group_allocated);
2591         }
2592         size += sprintf(&buf[size], "]\n");
2593
2594         *offset = size <= len ? size : len;
2595         size = copy_to_user(ubuf, buf, *offset);
2596         if (size)
2597                 rv = -EFAULT;
2598
2599         kfree(buf);
2600         return rv ? rv : *offset;
2601 }
2602
2603 static ssize_t mtip_hw_read_flags(struct file *f, char __user *ubuf,
2604                                   size_t len, loff_t *offset)
2605 {
2606         struct driver_data *dd =  (struct driver_data *)f->private_data;
2607         char *buf;
2608         int size = *offset;
2609         int rv = 0;
2610
2611         if (!len || size)
2612                 return 0;
2613
2614         buf = kzalloc(MTIP_DFS_MAX_BUF_SIZE, GFP_KERNEL);
2615         if (!buf) {
2616                 dev_err(&dd->pdev->dev,
2617                         "Memory allocation: flag buffer\n");
2618                 return -ENOMEM;
2619         }
2620
2621         size += sprintf(&buf[size], "Flag-port : [ %08lX ]\n",
2622                                                         dd->port->flags);
2623         size += sprintf(&buf[size], "Flag-dd   : [ %08lX ]\n",
2624                                                         dd->dd_flag);
2625
2626         *offset = size <= len ? size : len;
2627         size = copy_to_user(ubuf, buf, *offset);
2628         if (size)
2629                 rv = -EFAULT;
2630
2631         kfree(buf);
2632         return rv ? rv : *offset;
2633 }
2634
2635 static const struct file_operations mtip_device_status_fops = {
2636         .owner  = THIS_MODULE,
2637         .open   = simple_open,
2638         .read   = mtip_hw_read_device_status,
2639         .llseek = no_llseek,
2640 };
2641
2642 static const struct file_operations mtip_regs_fops = {
2643         .owner  = THIS_MODULE,
2644         .open   = simple_open,
2645         .read   = mtip_hw_read_registers,
2646         .llseek = no_llseek,
2647 };
2648
2649 static const struct file_operations mtip_flags_fops = {
2650         .owner  = THIS_MODULE,
2651         .open   = simple_open,
2652         .read   = mtip_hw_read_flags,
2653         .llseek = no_llseek,
2654 };
2655
2656 /*
2657  * Create the sysfs related attributes.
2658  *
2659  * @dd   Pointer to the driver data structure.
2660  * @kobj Pointer to the kobj for the block device.
2661  *
2662  * return value
2663  *      0       Operation completed successfully.
2664  *      -EINVAL Invalid parameter.
2665  */
2666 static int mtip_hw_sysfs_init(struct driver_data *dd, struct kobject *kobj)
2667 {
2668         if (!kobj || !dd)
2669                 return -EINVAL;
2670
2671         if (sysfs_create_file(kobj, &dev_attr_status.attr))
2672                 dev_warn(&dd->pdev->dev,
2673                         "Error creating 'status' sysfs entry\n");
2674         return 0;
2675 }
2676
2677 /*
2678  * Remove the sysfs related attributes.
2679  *
2680  * @dd   Pointer to the driver data structure.
2681  * @kobj Pointer to the kobj for the block device.
2682  *
2683  * return value
2684  *      0       Operation completed successfully.
2685  *      -EINVAL Invalid parameter.
2686  */
2687 static int mtip_hw_sysfs_exit(struct driver_data *dd, struct kobject *kobj)
2688 {
2689         if (!kobj || !dd)
2690                 return -EINVAL;
2691
2692         sysfs_remove_file(kobj, &dev_attr_status.attr);
2693
2694         return 0;
2695 }
2696
2697 static int mtip_hw_debugfs_init(struct driver_data *dd)
2698 {
2699         if (!dfs_parent)
2700                 return -1;
2701
2702         dd->dfs_node = debugfs_create_dir(dd->disk->disk_name, dfs_parent);
2703         if (IS_ERR_OR_NULL(dd->dfs_node)) {
2704                 dev_warn(&dd->pdev->dev,
2705                         "Error creating node %s under debugfs\n",
2706                                                 dd->disk->disk_name);
2707                 dd->dfs_node = NULL;
2708                 return -1;
2709         }
2710
2711         debugfs_create_file("flags", S_IRUGO, dd->dfs_node, dd,
2712                                                         &mtip_flags_fops);
2713         debugfs_create_file("registers", S_IRUGO, dd->dfs_node, dd,
2714                                                         &mtip_regs_fops);
2715
2716         return 0;
2717 }
2718
2719 static void mtip_hw_debugfs_exit(struct driver_data *dd)
2720 {
2721         if (dd->dfs_node)
2722                 debugfs_remove_recursive(dd->dfs_node);
2723 }
2724
2725 /*
2726  * Perform any init/resume time hardware setup
2727  *
2728  * @dd Pointer to the driver data structure.
2729  *
2730  * return value
2731  *      None
2732  */
2733 static inline void hba_setup(struct driver_data *dd)
2734 {
2735         u32 hwdata;
2736         hwdata = readl(dd->mmio + HOST_HSORG);
2737
2738         /* interrupt bug workaround: use only 1 IS bit.*/
2739         writel(hwdata |
2740                 HSORG_DISABLE_SLOTGRP_INTR |
2741                 HSORG_DISABLE_SLOTGRP_PXIS,
2742                 dd->mmio + HOST_HSORG);
2743 }
2744
2745 static int mtip_device_unaligned_constrained(struct driver_data *dd)
2746 {
2747         return (dd->pdev->device == P420M_DEVICE_ID ? 1 : 0);
2748 }
2749
2750 /*
2751  * Detect the details of the product, and store anything needed
2752  * into the driver data structure.  This includes product type and
2753  * version and number of slot groups.
2754  *
2755  * @dd Pointer to the driver data structure.
2756  *
2757  * return value
2758  *      None
2759  */
2760 static void mtip_detect_product(struct driver_data *dd)
2761 {
2762         u32 hwdata;
2763         unsigned int rev, slotgroups;
2764
2765         /*
2766          * HBA base + 0xFC [15:0] - vendor-specific hardware interface
2767          * info register:
2768          * [15:8] hardware/software interface rev#
2769          * [   3] asic-style interface
2770          * [ 2:0] number of slot groups, minus 1 (only valid for asic-style).
2771          */
2772         hwdata = readl(dd->mmio + HOST_HSORG);
2773
2774         dd->product_type = MTIP_PRODUCT_UNKNOWN;
2775         dd->slot_groups = 1;
2776
2777         if (hwdata & 0x8) {
2778                 dd->product_type = MTIP_PRODUCT_ASICFPGA;
2779                 rev = (hwdata & HSORG_HWREV) >> 8;
2780                 slotgroups = (hwdata & HSORG_SLOTGROUPS) + 1;
2781                 dev_info(&dd->pdev->dev,
2782                         "ASIC-FPGA design, HS rev 0x%x, "
2783                         "%i slot groups [%i slots]\n",
2784                          rev,
2785                          slotgroups,
2786                          slotgroups * 32);
2787
2788                 if (slotgroups > MTIP_MAX_SLOT_GROUPS) {
2789                         dev_warn(&dd->pdev->dev,
2790                                 "Warning: driver only supports "
2791                                 "%i slot groups.\n", MTIP_MAX_SLOT_GROUPS);
2792                         slotgroups = MTIP_MAX_SLOT_GROUPS;
2793                 }
2794                 dd->slot_groups = slotgroups;
2795                 return;
2796         }
2797
2798         dev_warn(&dd->pdev->dev, "Unrecognized product id\n");
2799 }
2800
2801 /*
2802  * Blocking wait for FTL rebuild to complete
2803  *
2804  * @dd Pointer to the DRIVER_DATA structure.
2805  *
2806  * return value
2807  *      0       FTL rebuild completed successfully
2808  *      -EFAULT FTL rebuild error/timeout/interruption
2809  */
2810 static int mtip_ftl_rebuild_poll(struct driver_data *dd)
2811 {
2812         unsigned long timeout, cnt = 0, start;
2813
2814         dev_warn(&dd->pdev->dev,
2815                 "FTL rebuild in progress. Polling for completion.\n");
2816
2817         start = jiffies;
2818         timeout = jiffies + msecs_to_jiffies(MTIP_FTL_REBUILD_TIMEOUT_MS);
2819
2820         do {
2821                 if (unlikely(test_bit(MTIP_DDF_REMOVE_PENDING_BIT,
2822                                 &dd->dd_flag)))
2823                         return -EFAULT;
2824                 if (mtip_check_surprise_removal(dd->pdev))
2825                         return -EFAULT;
2826
2827                 if (mtip_get_identify(dd->port, NULL) < 0)
2828                         return -EFAULT;
2829
2830                 if (*(dd->port->identify + MTIP_FTL_REBUILD_OFFSET) ==
2831                         MTIP_FTL_REBUILD_MAGIC) {
2832                         ssleep(1);
2833                         /* Print message every 3 minutes */
2834                         if (cnt++ >= 180) {
2835                                 dev_warn(&dd->pdev->dev,
2836                                 "FTL rebuild in progress (%d secs).\n",
2837                                 jiffies_to_msecs(jiffies - start) / 1000);
2838                                 cnt = 0;
2839                         }
2840                 } else {
2841                         dev_warn(&dd->pdev->dev,
2842                                 "FTL rebuild complete (%d secs).\n",
2843                         jiffies_to_msecs(jiffies - start) / 1000);
2844                         mtip_block_initialize(dd);
2845                         return 0;
2846                 }
2847         } while (time_before(jiffies, timeout));
2848
2849         /* Check for timeout */
2850         dev_err(&dd->pdev->dev,
2851                 "Timed out waiting for FTL rebuild to complete (%d secs).\n",
2852                 jiffies_to_msecs(jiffies - start) / 1000);
2853         return -EFAULT;
2854 }
2855
2856 static void mtip_softirq_done_fn(struct request *rq)
2857 {
2858         struct mtip_cmd *cmd = blk_mq_rq_to_pdu(rq);
2859         struct driver_data *dd = rq->q->queuedata;
2860
2861         /* Unmap the DMA scatter list entries */
2862         dma_unmap_sg(&dd->pdev->dev, cmd->sg, cmd->scatter_ents,
2863                                                         cmd->direction);
2864
2865         if (unlikely(cmd->unaligned))
2866                 up(&dd->port->cmd_slot_unal);
2867
2868         blk_mq_end_request(rq, cmd->status);
2869 }
2870
2871 static void mtip_abort_cmd(struct request *req, void *data,
2872                                                         bool reserved)
2873 {
2874         struct mtip_cmd *cmd = blk_mq_rq_to_pdu(req);
2875         struct driver_data *dd = data;
2876
2877         dbg_printk(MTIP_DRV_NAME " Aborting request, tag = %d\n", req->tag);
2878
2879         clear_bit(req->tag, dd->port->cmds_to_issue);
2880         cmd->status = -EIO;
2881         mtip_softirq_done_fn(req);
2882 }
2883
2884 static void mtip_queue_cmd(struct request *req, void *data,
2885                                                         bool reserved)
2886 {
2887         struct driver_data *dd = data;
2888
2889         set_bit(req->tag, dd->port->cmds_to_issue);
2890         blk_abort_request(req);
2891 }
2892
2893 /*
2894  * service thread to issue queued commands
2895  *
2896  * @data Pointer to the driver data structure.
2897  *
2898  * return value
2899  *      0
2900  */
2901
2902 static int mtip_service_thread(void *data)
2903 {
2904         struct driver_data *dd = (struct driver_data *)data;
2905         unsigned long slot, slot_start, slot_wrap, to;
2906         unsigned int num_cmd_slots = dd->slot_groups * 32;
2907         struct mtip_port *port = dd->port;
2908
2909         while (1) {
2910                 if (kthread_should_stop() ||
2911                         test_bit(MTIP_PF_SVC_THD_STOP_BIT, &port->flags))
2912                         goto st_out;
2913                 clear_bit(MTIP_PF_SVC_THD_ACTIVE_BIT, &port->flags);
2914
2915                 /*
2916                  * the condition is to check neither an internal command is
2917                  * is in progress nor error handling is active
2918                  */
2919                 wait_event_interruptible(port->svc_wait, (port->flags) &&
2920                         (port->flags & MTIP_PF_SVC_THD_WORK));
2921
2922                 if (kthread_should_stop() ||
2923                         test_bit(MTIP_PF_SVC_THD_STOP_BIT, &port->flags))
2924                         goto st_out;
2925
2926                 if (unlikely(test_bit(MTIP_DDF_REMOVE_PENDING_BIT,
2927                                 &dd->dd_flag)))
2928                         goto st_out;
2929
2930                 set_bit(MTIP_PF_SVC_THD_ACTIVE_BIT, &port->flags);
2931
2932 restart_eh:
2933                 /* Demux bits: start with error handling */
2934                 if (test_bit(MTIP_PF_EH_ACTIVE_BIT, &port->flags)) {
2935                         mtip_handle_tfe(dd);
2936                         clear_bit(MTIP_PF_EH_ACTIVE_BIT, &port->flags);
2937                 }
2938
2939                 if (test_bit(MTIP_PF_EH_ACTIVE_BIT, &port->flags))
2940                         goto restart_eh;
2941
2942                 if (test_bit(MTIP_PF_TO_ACTIVE_BIT, &port->flags)) {
2943                         to = jiffies + msecs_to_jiffies(5000);
2944
2945                         do {
2946                                 mdelay(100);
2947                         } while (atomic_read(&dd->irq_workers_active) != 0 &&
2948                                 time_before(jiffies, to));
2949
2950                         if (atomic_read(&dd->irq_workers_active) != 0)
2951                                 dev_warn(&dd->pdev->dev,
2952                                         "Completion workers still active!");
2953
2954                         spin_lock(dd->queue->queue_lock);
2955                         blk_mq_tagset_busy_iter(&dd->tags,
2956                                                         mtip_queue_cmd, dd);
2957                         spin_unlock(dd->queue->queue_lock);
2958
2959                         set_bit(MTIP_PF_ISSUE_CMDS_BIT, &dd->port->flags);
2960
2961                         if (mtip_device_reset(dd))
2962                                 blk_mq_tagset_busy_iter(&dd->tags,
2963                                                         mtip_abort_cmd, dd);
2964
2965                         clear_bit(MTIP_PF_TO_ACTIVE_BIT, &dd->port->flags);
2966                 }
2967
2968                 if (test_bit(MTIP_PF_ISSUE_CMDS_BIT, &port->flags)) {
2969                         slot = 1;
2970                         /* used to restrict the loop to one iteration */
2971                         slot_start = num_cmd_slots;
2972                         slot_wrap = 0;
2973                         while (1) {
2974                                 slot = find_next_bit(port->cmds_to_issue,
2975                                                 num_cmd_slots, slot);
2976                                 if (slot_wrap == 1) {
2977                                         if ((slot_start >= slot) ||
2978                                                 (slot >= num_cmd_slots))
2979                                                 break;
2980                                 }
2981                                 if (unlikely(slot_start == num_cmd_slots))
2982                                         slot_start = slot;
2983
2984                                 if (unlikely(slot == num_cmd_slots)) {
2985                                         slot = 1;
2986                                         slot_wrap = 1;
2987                                         continue;
2988                                 }
2989
2990                                 /* Issue the command to the hardware */
2991                                 mtip_issue_ncq_command(port, slot);
2992
2993                                 clear_bit(slot, port->cmds_to_issue);
2994                         }
2995
2996                         clear_bit(MTIP_PF_ISSUE_CMDS_BIT, &port->flags);
2997                 }
2998
2999                 if (test_bit(MTIP_PF_REBUILD_BIT, &port->flags)) {
3000                         if (mtip_ftl_rebuild_poll(dd) == 0)
3001                                 clear_bit(MTIP_PF_REBUILD_BIT, &port->flags);
3002                 }
3003         }
3004
3005 st_out:
3006         return 0;
3007 }
3008
3009 /*
3010  * DMA region teardown
3011  *
3012  * @dd Pointer to driver_data structure
3013  *
3014  * return value
3015  *      None
3016  */
3017 static void mtip_dma_free(struct driver_data *dd)
3018 {
3019         struct mtip_port *port = dd->port;
3020
3021         if (port->block1)
3022                 dmam_free_coherent(&dd->pdev->dev, BLOCK_DMA_ALLOC_SZ,
3023                                         port->block1, port->block1_dma);
3024
3025         if (port->command_list) {
3026                 dmam_free_coherent(&dd->pdev->dev, AHCI_CMD_TBL_SZ,
3027                                 port->command_list, port->command_list_dma);
3028         }
3029 }
3030
3031 /*
3032  * DMA region setup
3033  *
3034  * @dd Pointer to driver_data structure
3035  *
3036  * return value
3037  *      -ENOMEM Not enough free DMA region space to initialize driver
3038  */
3039 static int mtip_dma_alloc(struct driver_data *dd)
3040 {
3041         struct mtip_port *port = dd->port;
3042
3043         /* Allocate dma memory for RX Fis, Identify, and Sector Bufffer */
3044         port->block1 =
3045                 dmam_alloc_coherent(&dd->pdev->dev, BLOCK_DMA_ALLOC_SZ,
3046                                         &port->block1_dma, GFP_KERNEL);
3047         if (!port->block1)
3048                 return -ENOMEM;
3049         memset(port->block1, 0, BLOCK_DMA_ALLOC_SZ);
3050
3051         /* Allocate dma memory for command list */
3052         port->command_list =
3053                 dmam_alloc_coherent(&dd->pdev->dev, AHCI_CMD_TBL_SZ,
3054                                         &port->command_list_dma, GFP_KERNEL);
3055         if (!port->command_list) {
3056                 dmam_free_coherent(&dd->pdev->dev, BLOCK_DMA_ALLOC_SZ,
3057                                         port->block1, port->block1_dma);
3058                 port->block1 = NULL;
3059                 port->block1_dma = 0;
3060                 return -ENOMEM;
3061         }
3062         memset(port->command_list, 0, AHCI_CMD_TBL_SZ);
3063
3064         /* Setup all pointers into first DMA region */
3065         port->rxfis         = port->block1 + AHCI_RX_FIS_OFFSET;
3066         port->rxfis_dma     = port->block1_dma + AHCI_RX_FIS_OFFSET;
3067         port->identify      = port->block1 + AHCI_IDFY_OFFSET;
3068         port->identify_dma  = port->block1_dma + AHCI_IDFY_OFFSET;
3069         port->log_buf       = port->block1 + AHCI_SECTBUF_OFFSET;
3070         port->log_buf_dma   = port->block1_dma + AHCI_SECTBUF_OFFSET;
3071         port->smart_buf     = port->block1 + AHCI_SMARTBUF_OFFSET;
3072         port->smart_buf_dma = port->block1_dma + AHCI_SMARTBUF_OFFSET;
3073
3074         return 0;
3075 }
3076
3077 static int mtip_hw_get_identify(struct driver_data *dd)
3078 {
3079         struct smart_attr attr242;
3080         unsigned char *buf;
3081         int rv;
3082
3083         if (mtip_get_identify(dd->port, NULL) < 0)
3084                 return -EFAULT;
3085
3086         if (*(dd->port->identify + MTIP_FTL_REBUILD_OFFSET) ==
3087                 MTIP_FTL_REBUILD_MAGIC) {
3088                 set_bit(MTIP_PF_REBUILD_BIT, &dd->port->flags);
3089                 return MTIP_FTL_REBUILD_MAGIC;
3090         }
3091         mtip_dump_identify(dd->port);
3092
3093         /* check write protect, over temp and rebuild statuses */
3094         rv = mtip_read_log_page(dd->port, ATA_LOG_SATA_NCQ,
3095                                 dd->port->log_buf,
3096                                 dd->port->log_buf_dma, 1);
3097         if (rv) {
3098                 dev_warn(&dd->pdev->dev,
3099                         "Error in READ LOG EXT (10h) command\n");
3100                 /* non-critical error, don't fail the load */
3101         } else {
3102                 buf = (unsigned char *)dd->port->log_buf;
3103                 if (buf[259] & 0x1) {
3104                         dev_info(&dd->pdev->dev,
3105                                 "Write protect bit is set.\n");
3106                         set_bit(MTIP_DDF_WRITE_PROTECT_BIT, &dd->dd_flag);
3107                 }
3108                 if (buf[288] == 0xF7) {
3109                         dev_info(&dd->pdev->dev,
3110                                 "Exceeded Tmax, drive in thermal shutdown.\n");
3111                         set_bit(MTIP_DDF_OVER_TEMP_BIT, &dd->dd_flag);
3112                 }
3113                 if (buf[288] == 0xBF) {
3114                         dev_info(&dd->pdev->dev,
3115                                 "Drive indicates rebuild has failed.\n");
3116                         set_bit(MTIP_DDF_REBUILD_FAILED_BIT, &dd->dd_flag);
3117                 }
3118         }
3119
3120         /* get write protect progess */
3121         memset(&attr242, 0, sizeof(struct smart_attr));
3122         if (mtip_get_smart_attr(dd->port, 242, &attr242))
3123                 dev_warn(&dd->pdev->dev,
3124                                 "Unable to check write protect progress\n");
3125         else
3126                 dev_info(&dd->pdev->dev,
3127                                 "Write protect progress: %u%% (%u blocks)\n",
3128                                 attr242.cur, le32_to_cpu(attr242.data));
3129
3130         return rv;
3131 }
3132
3133 /*
3134  * Called once for each card.
3135  *
3136  * @dd Pointer to the driver data structure.
3137  *
3138  * return value
3139  *      0 on success, else an error code.
3140  */
3141 static int mtip_hw_init(struct driver_data *dd)
3142 {
3143         int i;
3144         int rv;
3145         unsigned int num_command_slots;
3146         unsigned long timeout, timetaken;
3147
3148         dd->mmio = pcim_iomap_table(dd->pdev)[MTIP_ABAR];
3149
3150         mtip_detect_product(dd);
3151         if (dd->product_type == MTIP_PRODUCT_UNKNOWN) {
3152                 rv = -EIO;
3153                 goto out1;
3154         }
3155         num_command_slots = dd->slot_groups * 32;
3156
3157         hba_setup(dd);
3158
3159         dd->port = kzalloc_node(sizeof(struct mtip_port), GFP_KERNEL,
3160                                 dd->numa_node);
3161         if (!dd->port) {
3162                 dev_err(&dd->pdev->dev,
3163                         "Memory allocation: port structure\n");
3164                 return -ENOMEM;
3165         }
3166
3167         /* Continue workqueue setup */
3168         for (i = 0; i < MTIP_MAX_SLOT_GROUPS; i++)
3169                 dd->work[i].port = dd->port;
3170
3171         /* Enable unaligned IO constraints for some devices */
3172         if (mtip_device_unaligned_constrained(dd))
3173                 dd->unal_qdepth = MTIP_MAX_UNALIGNED_SLOTS;
3174         else
3175                 dd->unal_qdepth = 0;
3176
3177         sema_init(&dd->port->cmd_slot_unal, dd->unal_qdepth);
3178
3179         /* Spinlock to prevent concurrent issue */
3180         for (i = 0; i < MTIP_MAX_SLOT_GROUPS; i++)
3181                 spin_lock_init(&dd->port->cmd_issue_lock[i]);
3182
3183         /* Set the port mmio base address. */
3184         dd->port->mmio  = dd->mmio + PORT_OFFSET;
3185         dd->port->dd    = dd;
3186
3187         /* DMA allocations */
3188         rv = mtip_dma_alloc(dd);
3189         if (rv < 0)
3190                 goto out1;
3191
3192         /* Setup the pointers to the extended s_active and CI registers. */
3193         for (i = 0; i < dd->slot_groups; i++) {
3194                 dd->port->s_active[i] =
3195                         dd->port->mmio + i*0x80 + PORT_SCR_ACT;
3196                 dd->port->cmd_issue[i] =
3197                         dd->port->mmio + i*0x80 + PORT_COMMAND_ISSUE;
3198                 dd->port->completed[i] =
3199                         dd->port->mmio + i*0x80 + PORT_SDBV;
3200         }
3201
3202         timetaken = jiffies;
3203         timeout = jiffies + msecs_to_jiffies(30000);
3204         while (((readl(dd->port->mmio + PORT_SCR_STAT) & 0x0F) != 0x03) &&
3205                  time_before(jiffies, timeout)) {
3206                 mdelay(100);
3207         }
3208         if (unlikely(mtip_check_surprise_removal(dd->pdev))) {
3209                 timetaken = jiffies - timetaken;
3210                 dev_warn(&dd->pdev->dev,
3211                         "Surprise removal detected at %u ms\n",
3212                         jiffies_to_msecs(timetaken));
3213                 rv = -ENODEV;
3214                 goto out2 ;
3215         }
3216         if (unlikely(test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &dd->dd_flag))) {
3217                 timetaken = jiffies - timetaken;
3218                 dev_warn(&dd->pdev->dev,
3219                         "Removal detected at %u ms\n",
3220                         jiffies_to_msecs(timetaken));
3221                 rv = -EFAULT;
3222                 goto out2;
3223         }
3224
3225         /* Conditionally reset the HBA. */
3226         if (!(readl(dd->mmio + HOST_CAP) & HOST_CAP_NZDMA)) {
3227                 if (mtip_hba_reset(dd) < 0) {
3228                         dev_err(&dd->pdev->dev,
3229                                 "Card did not reset within timeout\n");
3230                         rv = -EIO;
3231                         goto out2;
3232                 }
3233         } else {
3234                 /* Clear any pending interrupts on the HBA */
3235                 writel(readl(dd->mmio + HOST_IRQ_STAT),
3236                         dd->mmio + HOST_IRQ_STAT);
3237         }
3238
3239         mtip_init_port(dd->port);
3240         mtip_start_port(dd->port);
3241
3242         /* Setup the ISR and enable interrupts. */
3243         rv = devm_request_irq(&dd->pdev->dev,
3244                                 dd->pdev->irq,
3245                                 mtip_irq_handler,
3246                                 IRQF_SHARED,
3247                                 dev_driver_string(&dd->pdev->dev),
3248                                 dd);
3249
3250         if (rv) {
3251                 dev_err(&dd->pdev->dev,
3252                         "Unable to allocate IRQ %d\n", dd->pdev->irq);
3253                 goto out2;
3254         }
3255         irq_set_affinity_hint(dd->pdev->irq, get_cpu_mask(dd->isr_binding));
3256
3257         /* Enable interrupts on the HBA. */
3258         writel(readl(dd->mmio + HOST_CTL) | HOST_IRQ_EN,
3259                                         dd->mmio + HOST_CTL);
3260
3261         init_waitqueue_head(&dd->port->svc_wait);
3262
3263         if (test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &dd->dd_flag)) {
3264                 rv = -EFAULT;
3265                 goto out3;
3266         }
3267
3268         return rv;
3269
3270 out3:
3271         /* Disable interrupts on the HBA. */
3272         writel(readl(dd->mmio + HOST_CTL) & ~HOST_IRQ_EN,
3273                         dd->mmio + HOST_CTL);
3274
3275         /* Release the IRQ. */
3276         irq_set_affinity_hint(dd->pdev->irq, NULL);
3277         devm_free_irq(&dd->pdev->dev, dd->pdev->irq, dd);
3278
3279 out2:
3280         mtip_deinit_port(dd->port);
3281         mtip_dma_free(dd);
3282
3283 out1:
3284         /* Free the memory allocated for the for structure. */
3285         kfree(dd->port);
3286
3287         return rv;
3288 }
3289
3290 static int mtip_standby_drive(struct driver_data *dd)
3291 {
3292         int rv = 0;
3293
3294         if (dd->sr || !dd->port)
3295                 return -ENODEV;
3296         /*
3297          * Send standby immediate (E0h) to the drive so that it
3298          * saves its state.
3299          */
3300         if (!test_bit(MTIP_PF_REBUILD_BIT, &dd->port->flags) &&
3301             !test_bit(MTIP_DDF_REBUILD_FAILED_BIT, &dd->dd_flag) &&
3302             !test_bit(MTIP_DDF_SEC_LOCK_BIT, &dd->dd_flag)) {
3303                 rv = mtip_standby_immediate(dd->port);
3304                 if (rv)
3305                         dev_warn(&dd->pdev->dev,
3306                                 "STANDBY IMMEDIATE failed\n");
3307         }
3308         return rv;
3309 }
3310
3311 /*
3312  * Called to deinitialize an interface.
3313  *
3314  * @dd Pointer to the driver data structure.
3315  *
3316  * return value
3317  *      0
3318  */
3319 static int mtip_hw_exit(struct driver_data *dd)
3320 {
3321         if (!dd->sr) {
3322                 /* de-initialize the port. */
3323                 mtip_deinit_port(dd->port);
3324
3325                 /* Disable interrupts on the HBA. */
3326                 writel(readl(dd->mmio + HOST_CTL) & ~HOST_IRQ_EN,
3327                                 dd->mmio + HOST_CTL);
3328         }
3329
3330         /* Release the IRQ. */
3331         irq_set_affinity_hint(dd->pdev->irq, NULL);
3332         devm_free_irq(&dd->pdev->dev, dd->pdev->irq, dd);
3333         msleep(1000);
3334
3335         /* Free dma regions */
3336         mtip_dma_free(dd);
3337
3338         /* Free the memory allocated for the for structure. */
3339         kfree(dd->port);
3340         dd->port = NULL;
3341
3342         return 0;
3343 }
3344
3345 /*
3346  * Issue a Standby Immediate command to the device.
3347  *
3348  * This function is called by the Block Layer just before the
3349  * system powers off during a shutdown.
3350  *
3351  * @dd Pointer to the driver data structure.
3352  *
3353  * return value
3354  *      0
3355  */
3356 static int mtip_hw_shutdown(struct driver_data *dd)
3357 {
3358         /*
3359          * Send standby immediate (E0h) to the drive so that it
3360          * saves its state.
3361          */
3362         mtip_standby_drive(dd);
3363
3364         return 0;
3365 }
3366
3367 /*
3368  * Suspend function
3369  *
3370  * This function is called by the Block Layer just before the
3371  * system hibernates.
3372  *
3373  * @dd Pointer to the driver data structure.
3374  *
3375  * return value
3376  *      0       Suspend was successful
3377  *      -EFAULT Suspend was not successful
3378  */
3379 static int mtip_hw_suspend(struct driver_data *dd)
3380 {
3381         /*
3382          * Send standby immediate (E0h) to the drive
3383          * so that it saves its state.
3384          */
3385         if (mtip_standby_drive(dd) != 0) {
3386                 dev_err(&dd->pdev->dev,
3387                         "Failed standby-immediate command\n");
3388                 return -EFAULT;
3389         }
3390
3391         /* Disable interrupts on the HBA.*/
3392         writel(readl(dd->mmio + HOST_CTL) & ~HOST_IRQ_EN,
3393                         dd->mmio + HOST_CTL);
3394         mtip_deinit_port(dd->port);
3395
3396         return 0;
3397 }
3398
3399 /*
3400  * Resume function
3401  *
3402  * This function is called by the Block Layer as the
3403  * system resumes.
3404  *
3405  * @dd Pointer to the driver data structure.
3406  *
3407  * return value
3408  *      0       Resume was successful
3409  *      -EFAULT Resume was not successful
3410  */
3411 static int mtip_hw_resume(struct driver_data *dd)
3412 {
3413         /* Perform any needed hardware setup steps */
3414         hba_setup(dd);
3415
3416         /* Reset the HBA */
3417         if (mtip_hba_reset(dd) != 0) {
3418                 dev_err(&dd->pdev->dev,
3419                         "Unable to reset the HBA\n");
3420                 return -EFAULT;
3421         }
3422
3423         /*
3424          * Enable the port, DMA engine, and FIS reception specific
3425          * h/w in controller.
3426          */
3427         mtip_init_port(dd->port);
3428         mtip_start_port(dd->port);
3429
3430         /* Enable interrupts on the HBA.*/
3431         writel(readl(dd->mmio + HOST_CTL) | HOST_IRQ_EN,
3432                         dd->mmio + HOST_CTL);
3433
3434         return 0;
3435 }
3436
3437 /*
3438  * Helper function for reusing disk name
3439  * upon hot insertion.
3440  */
3441 static int rssd_disk_name_format(char *prefix,
3442                                  int index,
3443                                  char *buf,
3444                                  int buflen)
3445 {
3446         const int base = 'z' - 'a' + 1;
3447         char *begin = buf + strlen(prefix);
3448         char *end = buf + buflen;
3449         char *p;
3450         int unit;
3451
3452         p = end - 1;
3453         *p = '\0';
3454         unit = base;
3455         do {
3456                 if (p == begin)
3457                         return -EINVAL;
3458                 *--p = 'a' + (index % unit);
3459                 index = (index / unit) - 1;
3460         } while (index >= 0);
3461
3462         memmove(begin, p, end - p);
3463         memcpy(buf, prefix, strlen(prefix));
3464
3465         return 0;
3466 }
3467
3468 /*
3469  * Block layer IOCTL handler.
3470  *
3471  * @dev Pointer to the block_device structure.
3472  * @mode ignored
3473  * @cmd IOCTL command passed from the user application.
3474  * @arg Argument passed from the user application.
3475  *
3476  * return value
3477  *      0        IOCTL completed successfully.
3478  *      -ENOTTY  IOCTL not supported or invalid driver data
3479  *                 structure pointer.
3480  */
3481 static int mtip_block_ioctl(struct block_device *dev,
3482                             fmode_t mode,
3483                             unsigned cmd,
3484                             unsigned long arg)
3485 {
3486         struct driver_data *dd = dev->bd_disk->private_data;
3487
3488         if (!capable(CAP_SYS_ADMIN))
3489                 return -EACCES;
3490
3491         if (!dd)
3492                 return -ENOTTY;
3493
3494         if (unlikely(test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &dd->dd_flag)))
3495                 return -ENOTTY;
3496
3497         switch (cmd) {
3498         case BLKFLSBUF:
3499                 return -ENOTTY;
3500         default:
3501                 return mtip_hw_ioctl(dd, cmd, arg);
3502         }
3503 }
3504
3505 #ifdef CONFIG_COMPAT
3506 /*
3507  * Block layer compat IOCTL handler.
3508  *
3509  * @dev Pointer to the block_device structure.
3510  * @mode ignored
3511  * @cmd IOCTL command passed from the user application.
3512  * @arg Argument passed from the user application.
3513  *
3514  * return value
3515  *      0        IOCTL completed successfully.
3516  *      -ENOTTY  IOCTL not supported or invalid driver data
3517  *                 structure pointer.
3518  */
3519 static int mtip_block_compat_ioctl(struct block_device *dev,
3520                             fmode_t mode,
3521                             unsigned cmd,
3522                             unsigned long arg)
3523 {
3524         struct driver_data *dd = dev->bd_disk->private_data;
3525
3526         if (!capable(CAP_SYS_ADMIN))
3527                 return -EACCES;
3528
3529         if (!dd)
3530                 return -ENOTTY;
3531
3532         if (unlikely(test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &dd->dd_flag)))
3533                 return -ENOTTY;
3534
3535         switch (cmd) {
3536         case BLKFLSBUF:
3537                 return -ENOTTY;
3538         case HDIO_DRIVE_TASKFILE: {
3539                 struct mtip_compat_ide_task_request_s __user *compat_req_task;
3540                 ide_task_request_t req_task;
3541                 int compat_tasksize, outtotal, ret;
3542
3543                 compat_tasksize =
3544                         sizeof(struct mtip_compat_ide_task_request_s);
3545
3546                 compat_req_task =
3547                         (struct mtip_compat_ide_task_request_s __user *) arg;
3548
3549                 if (copy_from_user(&req_task, (void __user *) arg,
3550                         compat_tasksize - (2 * sizeof(compat_long_t))))
3551                         return -EFAULT;
3552
3553                 if (get_user(req_task.out_size, &compat_req_task->out_size))
3554                         return -EFAULT;
3555
3556                 if (get_user(req_task.in_size, &compat_req_task->in_size))
3557                         return -EFAULT;
3558
3559                 outtotal = sizeof(struct mtip_compat_ide_task_request_s);
3560
3561                 ret = exec_drive_taskfile(dd, (void __user *) arg,
3562                                                 &req_task, outtotal);
3563
3564                 if (copy_to_user((void __user *) arg, &req_task,
3565                                 compat_tasksize -
3566                                 (2 * sizeof(compat_long_t))))
3567                         return -EFAULT;
3568
3569                 if (put_user(req_task.out_size, &compat_req_task->out_size))
3570                         return -EFAULT;
3571
3572                 if (put_user(req_task.in_size, &compat_req_task->in_size))
3573                         return -EFAULT;
3574
3575                 return ret;
3576         }
3577         default:
3578                 return mtip_hw_ioctl(dd, cmd, arg);
3579         }
3580 }
3581 #endif
3582
3583 /*
3584  * Obtain the geometry of the device.
3585  *
3586  * You may think that this function is obsolete, but some applications,
3587  * fdisk for example still used CHS values. This function describes the
3588  * device as having 224 heads and 56 sectors per cylinder. These values are
3589  * chosen so that each cylinder is aligned on a 4KB boundary. Since a
3590  * partition is described in terms of a start and end cylinder this means
3591  * that each partition is also 4KB aligned. Non-aligned partitions adversely
3592  * affects performance.
3593  *
3594  * @dev Pointer to the block_device strucutre.
3595  * @geo Pointer to a hd_geometry structure.
3596  *
3597  * return value
3598  *      0       Operation completed successfully.
3599  *      -ENOTTY An error occurred while reading the drive capacity.
3600  */
3601 static int mtip_block_getgeo(struct block_device *dev,
3602                                 struct hd_geometry *geo)
3603 {
3604         struct driver_data *dd = dev->bd_disk->private_data;
3605         sector_t capacity;
3606
3607         if (!dd)
3608                 return -ENOTTY;
3609
3610         if (!(mtip_hw_get_capacity(dd, &capacity))) {
3611                 dev_warn(&dd->pdev->dev,
3612                         "Could not get drive capacity.\n");
3613                 return -ENOTTY;
3614         }
3615
3616         geo->heads = 224;
3617         geo->sectors = 56;
3618         sector_div(capacity, (geo->heads * geo->sectors));
3619         geo->cylinders = capacity;
3620         return 0;
3621 }
3622
3623 static int mtip_block_open(struct block_device *dev, fmode_t mode)
3624 {
3625         struct driver_data *dd;
3626
3627         if (dev && dev->bd_disk) {
3628                 dd = (struct driver_data *) dev->bd_disk->private_data;
3629
3630                 if (dd) {
3631                         if (test_bit(MTIP_DDF_REMOVAL_BIT,
3632                                                         &dd->dd_flag)) {
3633                                 return -ENODEV;
3634                         }
3635                         return 0;
3636                 }
3637         }
3638         return -ENODEV;
3639 }
3640
3641 static void mtip_block_release(struct gendisk *disk, fmode_t mode)
3642 {
3643 }
3644
3645 /*
3646  * Block device operation function.
3647  *
3648  * This structure contains pointers to the functions required by the block
3649  * layer.
3650  */
3651 static const struct block_device_operations mtip_block_ops = {
3652         .open           = mtip_block_open,
3653         .release        = mtip_block_release,
3654         .ioctl          = mtip_block_ioctl,
3655 #ifdef CONFIG_COMPAT
3656         .compat_ioctl   = mtip_block_compat_ioctl,
3657 #endif
3658         .getgeo         = mtip_block_getgeo,
3659         .owner          = THIS_MODULE
3660 };
3661
3662 static inline bool is_se_active(struct driver_data *dd)
3663 {
3664         if (unlikely(test_bit(MTIP_PF_SE_ACTIVE_BIT, &dd->port->flags))) {
3665                 if (dd->port->ic_pause_timer) {
3666                         unsigned long to = dd->port->ic_pause_timer +
3667                                                         msecs_to_jiffies(1000);
3668                         if (time_after(jiffies, to)) {
3669                                 clear_bit(MTIP_PF_SE_ACTIVE_BIT,
3670                                                         &dd->port->flags);
3671                                 clear_bit(MTIP_DDF_SEC_LOCK_BIT, &dd->dd_flag);
3672                                 dd->port->ic_pause_timer = 0;
3673                                 wake_up_interruptible(&dd->port->svc_wait);
3674                                 return false;
3675                         }
3676                 }
3677                 return true;
3678         }
3679         return false;
3680 }
3681
3682 /*
3683  * Block layer make request function.
3684  *
3685  * This function is called by the kernel to process a BIO for
3686  * the P320 device.
3687  *
3688  * @queue Pointer to the request queue. Unused other than to obtain
3689  *              the driver data structure.
3690  * @rq    Pointer to the request.
3691  *
3692  */
3693 static int mtip_submit_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
3694 {
3695         struct driver_data *dd = hctx->queue->queuedata;
3696         struct mtip_cmd *cmd = blk_mq_rq_to_pdu(rq);
3697         unsigned int nents;
3698
3699         if (is_se_active(dd))
3700                 return -ENODATA;
3701
3702         if (unlikely(dd->dd_flag & MTIP_DDF_STOP_IO)) {
3703                 if (unlikely(test_bit(MTIP_DDF_REMOVE_PENDING_BIT,
3704                                                         &dd->dd_flag))) {
3705                         return -ENXIO;
3706                 }
3707                 if (unlikely(test_bit(MTIP_DDF_OVER_TEMP_BIT, &dd->dd_flag))) {
3708                         return -ENODATA;
3709                 }
3710                 if (unlikely(test_bit(MTIP_DDF_WRITE_PROTECT_BIT,
3711                                                         &dd->dd_flag) &&
3712                                 rq_data_dir(rq))) {
3713                         return -ENODATA;
3714                 }
3715                 if (unlikely(test_bit(MTIP_DDF_SEC_LOCK_BIT, &dd->dd_flag) ||
3716                         test_bit(MTIP_DDF_REBUILD_FAILED_BIT, &dd->dd_flag)))
3717                         return -ENODATA;
3718         }
3719
3720         if (req_op(rq) == REQ_OP_DISCARD) {
3721                 int err;
3722
3723                 err = mtip_send_trim(dd, blk_rq_pos(rq), blk_rq_sectors(rq));
3724                 blk_mq_end_request(rq, err);
3725                 return 0;
3726         }
3727
3728         /* Create the scatter list for this request. */
3729         nents = blk_rq_map_sg(hctx->queue, rq, cmd->sg);
3730
3731         /* Issue the read/write. */
3732         mtip_hw_submit_io(dd, rq, cmd, nents, hctx);
3733         return 0;
3734 }
3735
3736 static bool mtip_check_unal_depth(struct blk_mq_hw_ctx *hctx,
3737                                   struct request *rq)
3738 {
3739         struct driver_data *dd = hctx->queue->queuedata;
3740         struct mtip_cmd *cmd = blk_mq_rq_to_pdu(rq);
3741
3742         if (rq_data_dir(rq) == READ || !dd->unal_qdepth)
3743                 return false;
3744
3745         /*
3746          * If unaligned depth must be limited on this controller, mark it
3747          * as unaligned if the IO isn't on a 4k boundary (start of length).
3748          */
3749         if (blk_rq_sectors(rq) <= 64) {
3750                 if ((blk_rq_pos(rq) & 7) || (blk_rq_sectors(rq) & 7))
3751                         cmd->unaligned = 1;
3752         }
3753
3754         if (cmd->unaligned && down_trylock(&dd->port->cmd_slot_unal))
3755                 return true;
3756
3757         return false;
3758 }
3759
3760 static int mtip_issue_reserved_cmd(struct blk_mq_hw_ctx *hctx,
3761                                    struct request *rq)
3762 {
3763         struct driver_data *dd = hctx->queue->queuedata;
3764         struct mtip_int_cmd *icmd = rq->end_io_data;
3765         struct mtip_cmd *cmd = blk_mq_rq_to_pdu(rq);
3766         struct mtip_cmd_sg *command_sg;
3767
3768         if (mtip_commands_active(dd->port))
3769                 return BLK_MQ_RQ_QUEUE_BUSY;
3770
3771         rq->end_io_data = NULL;
3772
3773         /* Populate the SG list */
3774         cmd->command_header->opts =
3775                  __force_bit2int cpu_to_le32(icmd->opts | icmd->fis_len);
3776         if (icmd->buf_len) {
3777                 command_sg = cmd->command + AHCI_CMD_TBL_HDR_SZ;
3778
3779                 command_sg->info =
3780                         __force_bit2int cpu_to_le32((icmd->buf_len-1) & 0x3FFFFF);
3781                 command_sg->dba =
3782                         __force_bit2int cpu_to_le32(icmd->buffer & 0xFFFFFFFF);
3783                 command_sg->dba_upper =
3784                         __force_bit2int cpu_to_le32((icmd->buffer >> 16) >> 16);
3785
3786                 cmd->command_header->opts |=
3787                         __force_bit2int cpu_to_le32((1 << 16));
3788         }
3789
3790         /* Populate the command header */
3791         cmd->command_header->byte_count = 0;
3792
3793         blk_mq_start_request(rq);
3794         mtip_issue_non_ncq_command(dd->port, rq->tag);
3795         return BLK_MQ_RQ_QUEUE_OK;
3796 }
3797
3798 static int mtip_queue_rq(struct blk_mq_hw_ctx *hctx,
3799                          const struct blk_mq_queue_data *bd)
3800 {
3801         struct request *rq = bd->rq;
3802         int ret;
3803
3804         mtip_init_cmd_header(rq);
3805
3806         if (blk_rq_is_passthrough(rq))
3807                 return mtip_issue_reserved_cmd(hctx, rq);
3808
3809         if (unlikely(mtip_check_unal_depth(hctx, rq)))
3810                 return BLK_MQ_RQ_QUEUE_BUSY;
3811
3812         blk_mq_start_request(rq);
3813
3814         ret = mtip_submit_request(hctx, rq);
3815         if (likely(!ret))
3816                 return BLK_MQ_RQ_QUEUE_OK;
3817
3818         return BLK_MQ_RQ_QUEUE_ERROR;
3819 }
3820
3821 static void mtip_free_cmd(void *data, struct request *rq,
3822                           unsigned int hctx_idx, unsigned int request_idx)
3823 {
3824         struct driver_data *dd = data;
3825         struct mtip_cmd *cmd = blk_mq_rq_to_pdu(rq);
3826
3827         if (!cmd->command)
3828                 return;
3829
3830         dmam_free_coherent(&dd->pdev->dev, CMD_DMA_ALLOC_SZ,
3831                                 cmd->command, cmd->command_dma);
3832 }
3833
3834 static int mtip_init_cmd(void *data, struct request *rq, unsigned int hctx_idx,
3835                          unsigned int request_idx, unsigned int numa_node)
3836 {
3837         struct driver_data *dd = data;
3838         struct mtip_cmd *cmd = blk_mq_rq_to_pdu(rq);
3839
3840         /*
3841          * For flush requests, request_idx starts at the end of the
3842          * tag space.  Since we don't support FLUSH/FUA, simply return
3843          * 0 as there's nothing to be done.
3844          */
3845         if (request_idx >= MTIP_MAX_COMMAND_SLOTS)
3846                 return 0;
3847
3848         cmd->command = dmam_alloc_coherent(&dd->pdev->dev, CMD_DMA_ALLOC_SZ,
3849                         &cmd->command_dma, GFP_KERNEL);
3850         if (!cmd->command)
3851                 return -ENOMEM;
3852
3853         memset(cmd->command, 0, CMD_DMA_ALLOC_SZ);
3854
3855         sg_init_table(cmd->sg, MTIP_MAX_SG);
3856         return 0;
3857 }
3858
3859 static enum blk_eh_timer_return mtip_cmd_timeout(struct request *req,
3860                                                                 bool reserved)
3861 {
3862         struct driver_data *dd = req->q->queuedata;
3863
3864         if (reserved) {
3865                 struct mtip_cmd *cmd = blk_mq_rq_to_pdu(req);
3866
3867                 cmd->status = -ETIME;
3868                 if (cmd->comp_func)
3869                         cmd->comp_func(dd->port, MTIP_TAG_INTERNAL, cmd, -ETIME);
3870                 goto exit_handler;
3871         }
3872
3873         if (test_bit(req->tag, dd->port->cmds_to_issue))
3874                 goto exit_handler;
3875
3876         if (test_and_set_bit(MTIP_PF_TO_ACTIVE_BIT, &dd->port->flags))
3877                 goto exit_handler;
3878
3879         wake_up_interruptible(&dd->port->svc_wait);
3880 exit_handler:
3881         return BLK_EH_RESET_TIMER;
3882 }
3883
3884 static const struct blk_mq_ops mtip_mq_ops = {
3885         .queue_rq       = mtip_queue_rq,
3886         .init_request   = mtip_init_cmd,
3887         .exit_request   = mtip_free_cmd,
3888         .complete       = mtip_softirq_done_fn,
3889         .timeout        = mtip_cmd_timeout,
3890 };
3891
3892 /*
3893  * Block layer initialization function.
3894  *
3895  * This function is called once by the PCI layer for each P320
3896  * device that is connected to the system.
3897  *
3898  * @dd Pointer to the driver data structure.
3899  *
3900  * return value
3901  *      0 on success else an error code.
3902  */
3903 static int mtip_block_initialize(struct driver_data *dd)
3904 {
3905         int rv = 0, wait_for_rebuild = 0;
3906         sector_t capacity;
3907         unsigned int index = 0;
3908         struct kobject *kobj;
3909
3910         if (dd->disk)
3911                 goto skip_create_disk; /* hw init done, before rebuild */
3912
3913         if (mtip_hw_init(dd)) {
3914                 rv = -EINVAL;
3915                 goto protocol_init_error;
3916         }
3917
3918         dd->disk = alloc_disk_node(MTIP_MAX_MINORS, dd->numa_node);
3919         if (dd->disk  == NULL) {
3920                 dev_err(&dd->pdev->dev,
3921                         "Unable to allocate gendisk structure\n");
3922                 rv = -EINVAL;
3923                 goto alloc_disk_error;
3924         }
3925
3926         /* Generate the disk name, implemented same as in sd.c */
3927         do {
3928                 if (!ida_pre_get(&rssd_index_ida, GFP_KERNEL)) {
3929                         rv = -ENOMEM;
3930                         goto ida_get_error;
3931                 }
3932
3933                 spin_lock(&rssd_index_lock);
3934                 rv = ida_get_new(&rssd_index_ida, &index);
3935                 spin_unlock(&rssd_index_lock);
3936         } while (rv == -EAGAIN);
3937
3938         if (rv)
3939                 goto ida_get_error;
3940
3941         rv = rssd_disk_name_format("rssd",
3942                                 index,
3943                                 dd->disk->disk_name,
3944                                 DISK_NAME_LEN);
3945         if (rv)
3946                 goto disk_index_error;
3947
3948         dd->disk->major         = dd->major;
3949         dd->disk->first_minor   = index * MTIP_MAX_MINORS;
3950         dd->disk->minors        = MTIP_MAX_MINORS;
3951         dd->disk->fops          = &mtip_block_ops;
3952         dd->disk->private_data  = dd;
3953         dd->index               = index;
3954
3955         mtip_hw_debugfs_init(dd);
3956
3957         memset(&dd->tags, 0, sizeof(dd->tags));
3958         dd->tags.ops = &mtip_mq_ops;
3959         dd->tags.nr_hw_queues = 1;
3960         dd->tags.queue_depth = MTIP_MAX_COMMAND_SLOTS;
3961         dd->tags.reserved_tags = 1;
3962         dd->tags.cmd_size = sizeof(struct mtip_cmd);
3963         dd->tags.numa_node = dd->numa_node;
3964         dd->tags.flags = BLK_MQ_F_SHOULD_MERGE;
3965         dd->tags.driver_data = dd;
3966         dd->tags.timeout = MTIP_NCQ_CMD_TIMEOUT_MS;
3967
3968         rv = blk_mq_alloc_tag_set(&dd->tags);
3969         if (rv) {
3970                 dev_err(&dd->pdev->dev,
3971                         "Unable to allocate request queue\n");
3972                 goto block_queue_alloc_tag_error;
3973         }
3974
3975         /* Allocate the request queue. */
3976         dd->queue = blk_mq_init_queue(&dd->tags);
3977         if (IS_ERR(dd->queue)) {
3978                 dev_err(&dd->pdev->dev,
3979                         "Unable to allocate request queue\n");
3980                 rv = -ENOMEM;
3981                 goto block_queue_alloc_init_error;
3982         }
3983
3984         dd->disk->queue         = dd->queue;
3985         dd->queue->queuedata    = dd;
3986
3987 skip_create_disk:
3988         /* Initialize the protocol layer. */
3989         wait_for_rebuild = mtip_hw_get_identify(dd);
3990         if (wait_for_rebuild < 0) {
3991                 dev_err(&dd->pdev->dev,
3992                         "Protocol layer initialization failed\n");
3993                 rv = -EINVAL;
3994                 goto init_hw_cmds_error;
3995         }
3996
3997         /*
3998          * if rebuild pending, start the service thread, and delay the block
3999          * queue creation and device_add_disk()
4000          */
4001         if (wait_for_rebuild == MTIP_FTL_REBUILD_MAGIC)
4002                 goto start_service_thread;
4003
4004         /* Set device limits. */
4005         set_bit(QUEUE_FLAG_NONROT, &dd->queue->queue_flags);
4006         clear_bit(QUEUE_FLAG_ADD_RANDOM, &dd->queue->queue_flags);
4007         blk_queue_max_segments(dd->queue, MTIP_MAX_SG);
4008         blk_queue_physical_block_size(dd->queue, 4096);
4009         blk_queue_max_hw_sectors(dd->queue, 0xffff);
4010         blk_queue_max_segment_size(dd->queue, 0x400000);
4011         blk_queue_io_min(dd->queue, 4096);
4012         blk_queue_bounce_limit(dd->queue, dd->pdev->dma_mask);
4013
4014         /* Signal trim support */
4015         if (dd->trim_supp == true) {
4016                 set_bit(QUEUE_FLAG_DISCARD, &dd->queue->queue_flags);
4017                 dd->queue->limits.discard_granularity = 4096;
4018                 blk_queue_max_discard_sectors(dd->queue,
4019                         MTIP_MAX_TRIM_ENTRY_LEN * MTIP_MAX_TRIM_ENTRIES);
4020         }
4021
4022         /* Set the capacity of the device in 512 byte sectors. */
4023         if (!(mtip_hw_get_capacity(dd, &capacity))) {
4024                 dev_warn(&dd->pdev->dev,
4025                         "Could not read drive capacity\n");
4026                 rv = -EIO;
4027                 goto read_capacity_error;
4028         }
4029         set_capacity(dd->disk, capacity);
4030
4031         /* Enable the block device and add it to /dev */
4032         device_add_disk(&dd->pdev->dev, dd->disk);
4033
4034         dd->bdev = bdget_disk(dd->disk, 0);
4035         /*
4036          * Now that the disk is active, initialize any sysfs attributes
4037          * managed by the protocol layer.
4038          */
4039         kobj = kobject_get(&disk_to_dev(dd->disk)->kobj);
4040         if (kobj) {
4041                 mtip_hw_sysfs_init(dd, kobj);
4042                 kobject_put(kobj);
4043         }
4044
4045         if (dd->mtip_svc_handler) {
4046                 set_bit(MTIP_DDF_INIT_DONE_BIT, &dd->dd_flag);
4047                 return rv; /* service thread created for handling rebuild */
4048         }
4049
4050 start_service_thread:
4051         dd->mtip_svc_handler = kthread_create_on_node(mtip_service_thread,
4052                                                 dd, dd->numa_node,
4053                                                 "mtip_svc_thd_%02d", index);
4054
4055         if (IS_ERR(dd->mtip_svc_handler)) {
4056                 dev_err(&dd->pdev->dev, "service thread failed to start\n");
4057                 dd->mtip_svc_handler = NULL;
4058                 rv = -EFAULT;
4059                 goto kthread_run_error;
4060         }
4061         wake_up_process(dd->mtip_svc_handler);
4062         if (wait_for_rebuild == MTIP_FTL_REBUILD_MAGIC)
4063                 rv = wait_for_rebuild;
4064
4065         return rv;
4066
4067 kthread_run_error:
4068         bdput(dd->bdev);
4069         dd->bdev = NULL;
4070
4071         /* Delete our gendisk. This also removes the device from /dev */
4072         del_gendisk(dd->disk);
4073
4074 read_capacity_error:
4075 init_hw_cmds_error:
4076         blk_cleanup_queue(dd->queue);
4077 block_queue_alloc_init_error:
4078         blk_mq_free_tag_set(&dd->tags);
4079 block_queue_alloc_tag_error:
4080         mtip_hw_debugfs_exit(dd);
4081 disk_index_error:
4082         spin_lock(&rssd_index_lock);
4083         ida_remove(&rssd_index_ida, index);
4084         spin_unlock(&rssd_index_lock);
4085
4086 ida_get_error:
4087         put_disk(dd->disk);
4088
4089 alloc_disk_error:
4090         mtip_hw_exit(dd); /* De-initialize the protocol layer. */
4091
4092 protocol_init_error:
4093         return rv;
4094 }
4095
4096 static void mtip_no_dev_cleanup(struct request *rq, void *data, bool reserv)
4097 {
4098         struct driver_data *dd = (struct driver_data *)data;
4099         struct mtip_cmd *cmd;
4100
4101         if (likely(!reserv)) {
4102                 cmd = blk_mq_rq_to_pdu(rq);
4103                 cmd->status = -ENODEV;
4104                 blk_mq_complete_request(rq);
4105         } else if (test_bit(MTIP_PF_IC_ACTIVE_BIT, &dd->port->flags)) {
4106
4107                 cmd = mtip_cmd_from_tag(dd, MTIP_TAG_INTERNAL);
4108                 cmd->status = -ENODEV;
4109                 if (cmd->comp_func)
4110                         cmd->comp_func(dd->port, MTIP_TAG_INTERNAL,
4111                                         cmd, -ENODEV);
4112         }
4113 }
4114
4115 /*
4116  * Block layer deinitialization function.
4117  *
4118  * Called by the PCI layer as each P320 device is removed.
4119  *
4120  * @dd Pointer to the driver data structure.
4121  *
4122  * return value
4123  *      0
4124  */
4125 static int mtip_block_remove(struct driver_data *dd)
4126 {
4127         struct kobject *kobj;
4128
4129         mtip_hw_debugfs_exit(dd);
4130
4131         if (dd->mtip_svc_handler) {
4132                 set_bit(MTIP_PF_SVC_THD_STOP_BIT, &dd->port->flags);
4133                 wake_up_interruptible(&dd->port->svc_wait);
4134                 kthread_stop(dd->mtip_svc_handler);
4135         }
4136
4137         /* Clean up the sysfs attributes, if created */
4138         if (test_bit(MTIP_DDF_INIT_DONE_BIT, &dd->dd_flag)) {
4139                 kobj = kobject_get(&disk_to_dev(dd->disk)->kobj);
4140                 if (kobj) {
4141                         mtip_hw_sysfs_exit(dd, kobj);
4142                         kobject_put(kobj);
4143                 }
4144         }
4145
4146         if (!dd->sr) {
4147                 /*
4148                  * Explicitly wait here for IOs to quiesce,
4149                  * as mtip_standby_drive usually won't wait for IOs.
4150                  */
4151                 if (!mtip_quiesce_io(dd->port, MTIP_QUIESCE_IO_TIMEOUT_MS))
4152                         mtip_standby_drive(dd);
4153         }
4154         else
4155                 dev_info(&dd->pdev->dev, "device %s surprise removal\n",
4156                                                 dd->disk->disk_name);
4157
4158         blk_freeze_queue_start(dd->queue);
4159         blk_mq_stop_hw_queues(dd->queue);
4160         blk_mq_tagset_busy_iter(&dd->tags, mtip_no_dev_cleanup, dd);
4161
4162         /*
4163          * Delete our gendisk structure. This also removes the device
4164          * from /dev
4165          */
4166         if (dd->bdev) {
4167                 bdput(dd->bdev);
4168                 dd->bdev = NULL;
4169         }
4170         if (dd->disk) {
4171                 if (test_bit(MTIP_DDF_INIT_DONE_BIT, &dd->dd_flag))
4172                         del_gendisk(dd->disk);
4173                 if (dd->disk->queue) {
4174                         blk_cleanup_queue(dd->queue);
4175                         blk_mq_free_tag_set(&dd->tags);
4176                         dd->queue = NULL;
4177                 }
4178                 put_disk(dd->disk);
4179         }
4180         dd->disk  = NULL;
4181
4182         spin_lock(&rssd_index_lock);
4183         ida_remove(&rssd_index_ida, dd->index);
4184         spin_unlock(&rssd_index_lock);
4185
4186         /* De-initialize the protocol layer. */
4187         mtip_hw_exit(dd);
4188
4189         return 0;
4190 }
4191
4192 /*
4193  * Function called by the PCI layer when just before the
4194  * machine shuts down.
4195  *
4196  * If a protocol layer shutdown function is present it will be called
4197  * by this function.
4198  *
4199  * @dd Pointer to the driver data structure.
4200  *
4201  * return value
4202  *      0
4203  */
4204 static int mtip_block_shutdown(struct driver_data *dd)
4205 {
4206         mtip_hw_shutdown(dd);
4207
4208         /* Delete our gendisk structure, and cleanup the blk queue. */
4209         if (dd->disk) {
4210                 dev_info(&dd->pdev->dev,
4211                         "Shutting down %s ...\n", dd->disk->disk_name);
4212
4213                 if (test_bit(MTIP_DDF_INIT_DONE_BIT, &dd->dd_flag))
4214                         del_gendisk(dd->disk);
4215                 if (dd->disk->queue) {
4216                         blk_cleanup_queue(dd->queue);
4217                         blk_mq_free_tag_set(&dd->tags);
4218                 }
4219                 put_disk(dd->disk);
4220                 dd->disk  = NULL;
4221                 dd->queue = NULL;
4222         }
4223
4224         spin_lock(&rssd_index_lock);
4225         ida_remove(&rssd_index_ida, dd->index);
4226         spin_unlock(&rssd_index_lock);
4227         return 0;
4228 }
4229
4230 static int mtip_block_suspend(struct driver_data *dd)
4231 {
4232         dev_info(&dd->pdev->dev,
4233                 "Suspending %s ...\n", dd->disk->disk_name);
4234         mtip_hw_suspend(dd);
4235         return 0;
4236 }
4237
4238 static int mtip_block_resume(struct driver_data *dd)
4239 {
4240         dev_info(&dd->pdev->dev, "Resuming %s ...\n",
4241                 dd->disk->disk_name);
4242         mtip_hw_resume(dd);
4243         return 0;
4244 }
4245
4246 static void drop_cpu(int cpu)
4247 {
4248         cpu_use[cpu]--;
4249 }
4250
4251 static int get_least_used_cpu_on_node(int node)
4252 {
4253         int cpu, least_used_cpu, least_cnt;
4254         const struct cpumask *node_mask;
4255
4256         node_mask = cpumask_of_node(node);
4257         least_used_cpu = cpumask_first(node_mask);
4258         least_cnt = cpu_use[least_used_cpu];
4259         cpu = least_used_cpu;
4260
4261         for_each_cpu(cpu, node_mask) {
4262                 if (cpu_use[cpu] < least_cnt) {
4263                         least_used_cpu = cpu;
4264                         least_cnt = cpu_use[cpu];
4265                 }
4266         }
4267         cpu_use[least_used_cpu]++;
4268         return least_used_cpu;
4269 }
4270
4271 /* Helper for selecting a node in round robin mode */
4272 static inline int mtip_get_next_rr_node(void)
4273 {
4274         static int next_node = -1;
4275
4276         if (next_node == -1) {
4277                 next_node = first_online_node;
4278                 return next_node;
4279         }
4280
4281         next_node = next_online_node(next_node);
4282         if (next_node == MAX_NUMNODES)
4283                 next_node = first_online_node;
4284         return next_node;
4285 }
4286
4287 static DEFINE_HANDLER(0);
4288 static DEFINE_HANDLER(1);
4289 static DEFINE_HANDLER(2);
4290 static DEFINE_HANDLER(3);
4291 static DEFINE_HANDLER(4);
4292 static DEFINE_HANDLER(5);
4293 static DEFINE_HANDLER(6);
4294 static DEFINE_HANDLER(7);
4295
4296 static void mtip_disable_link_opts(struct driver_data *dd, struct pci_dev *pdev)
4297 {
4298         int pos;
4299         unsigned short pcie_dev_ctrl;
4300
4301         pos = pci_find_capability(pdev, PCI_CAP_ID_EXP);
4302         if (pos) {
4303                 pci_read_config_word(pdev,
4304                         pos + PCI_EXP_DEVCTL,
4305                         &pcie_dev_ctrl);
4306                 if (pcie_dev_ctrl & (1 << 11) ||
4307                     pcie_dev_ctrl & (1 << 4)) {
4308                         dev_info(&dd->pdev->dev,
4309                                 "Disabling ERO/No-Snoop on bridge device %04x:%04x\n",
4310                                         pdev->vendor, pdev->device);
4311                         pcie_dev_ctrl &= ~(PCI_EXP_DEVCTL_NOSNOOP_EN |
4312                                                 PCI_EXP_DEVCTL_RELAX_EN);
4313                         pci_write_config_word(pdev,
4314                                 pos + PCI_EXP_DEVCTL,
4315                                 pcie_dev_ctrl);
4316                 }
4317         }
4318 }
4319
4320 static void mtip_fix_ero_nosnoop(struct driver_data *dd, struct pci_dev *pdev)
4321 {
4322         /*
4323          * This workaround is specific to AMD/ATI chipset with a PCI upstream
4324          * device with device id 0x5aXX
4325          */
4326         if (pdev->bus && pdev->bus->self) {
4327                 if (pdev->bus->self->vendor == PCI_VENDOR_ID_ATI &&
4328                     ((pdev->bus->self->device & 0xff00) == 0x5a00)) {
4329                         mtip_disable_link_opts(dd, pdev->bus->self);
4330                 } else {
4331                         /* Check further up the topology */
4332                         struct pci_dev *parent_dev = pdev->bus->self;
4333                         if (parent_dev->bus &&
4334                                 parent_dev->bus->parent &&
4335                                 parent_dev->bus->parent->self &&
4336                                 parent_dev->bus->parent->self->vendor ==
4337                                          PCI_VENDOR_ID_ATI &&
4338                                 (parent_dev->bus->parent->self->device &
4339                                         0xff00) == 0x5a00) {
4340                                 mtip_disable_link_opts(dd,
4341                                         parent_dev->bus->parent->self);
4342                         }
4343                 }
4344         }
4345 }
4346
4347 /*
4348  * Called for each supported PCI device detected.
4349  *
4350  * This function allocates the private data structure, enables the
4351  * PCI device and then calls the block layer initialization function.
4352  *
4353  * return value
4354  *      0 on success else an error code.
4355  */
4356 static int mtip_pci_probe(struct pci_dev *pdev,
4357                         const struct pci_device_id *ent)
4358 {
4359         int rv = 0;
4360         struct driver_data *dd = NULL;
4361         char cpu_list[256];
4362         const struct cpumask *node_mask;
4363         int cpu, i = 0, j = 0;
4364         int my_node = NUMA_NO_NODE;
4365         unsigned long flags;
4366
4367         /* Allocate memory for this devices private data. */
4368         my_node = pcibus_to_node(pdev->bus);
4369         if (my_node != NUMA_NO_NODE) {
4370                 if (!node_online(my_node))
4371                         my_node = mtip_get_next_rr_node();
4372         } else {
4373                 dev_info(&pdev->dev, "Kernel not reporting proximity, choosing a node\n");
4374                 my_node = mtip_get_next_rr_node();
4375         }
4376         dev_info(&pdev->dev, "NUMA node %d (closest: %d,%d, probe on %d:%d)\n",
4377                 my_node, pcibus_to_node(pdev->bus), dev_to_node(&pdev->dev),
4378                 cpu_to_node(raw_smp_processor_id()), raw_smp_processor_id());
4379
4380         dd = kzalloc_node(sizeof(struct driver_data), GFP_KERNEL, my_node);
4381         if (dd == NULL) {
4382                 dev_err(&pdev->dev,
4383                         "Unable to allocate memory for driver data\n");
4384                 return -ENOMEM;
4385         }
4386
4387         /* Attach the private data to this PCI device.  */
4388         pci_set_drvdata(pdev, dd);
4389
4390         rv = pcim_enable_device(pdev);
4391         if (rv < 0) {
4392                 dev_err(&pdev->dev, "Unable to enable device\n");
4393                 goto iomap_err;
4394         }
4395
4396         /* Map BAR5 to memory. */
4397         rv = pcim_iomap_regions(pdev, 1 << MTIP_ABAR, MTIP_DRV_NAME);
4398         if (rv < 0) {
4399                 dev_err(&pdev->dev, "Unable to map regions\n");
4400                 goto iomap_err;
4401         }
4402
4403         if (!pci_set_dma_mask(pdev, DMA_BIT_MASK(64))) {
4404                 rv = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(64));
4405
4406                 if (rv) {
4407                         rv = pci_set_consistent_dma_mask(pdev,
4408                                                 DMA_BIT_MASK(32));
4409                         if (rv) {
4410                                 dev_warn(&pdev->dev,
4411                                         "64-bit DMA enable failed\n");
4412                                 goto setmask_err;
4413                         }
4414                 }
4415         }
4416
4417         /* Copy the info we may need later into the private data structure. */
4418         dd->major       = mtip_major;
4419         dd->instance    = instance;
4420         dd->pdev        = pdev;
4421         dd->numa_node   = my_node;
4422
4423         INIT_LIST_HEAD(&dd->online_list);
4424         INIT_LIST_HEAD(&dd->remove_list);
4425
4426         memset(dd->workq_name, 0, 32);
4427         snprintf(dd->workq_name, 31, "mtipq%d", dd->instance);
4428
4429         dd->isr_workq = create_workqueue(dd->workq_name);
4430         if (!dd->isr_workq) {
4431                 dev_warn(&pdev->dev, "Can't create wq %d\n", dd->instance);
4432                 rv = -ENOMEM;
4433                 goto block_initialize_err;
4434         }
4435
4436         memset(cpu_list, 0, sizeof(cpu_list));
4437
4438         node_mask = cpumask_of_node(dd->numa_node);
4439         if (!cpumask_empty(node_mask)) {
4440                 for_each_cpu(cpu, node_mask)
4441                 {
4442                         snprintf(&cpu_list[j], 256 - j, "%d ", cpu);
4443                         j = strlen(cpu_list);
4444                 }
4445
4446                 dev_info(&pdev->dev, "Node %d on package %d has %d cpu(s): %s\n",
4447                         dd->numa_node,
4448                         topology_physical_package_id(cpumask_first(node_mask)),
4449                         nr_cpus_node(dd->numa_node),
4450                         cpu_list);
4451         } else
4452                 dev_dbg(&pdev->dev, "mtip32xx: node_mask empty\n");
4453
4454         dd->isr_binding = get_least_used_cpu_on_node(dd->numa_node);
4455         dev_info(&pdev->dev, "Initial IRQ binding node:cpu %d:%d\n",
4456                 cpu_to_node(dd->isr_binding), dd->isr_binding);
4457
4458         /* first worker context always runs in ISR */
4459         dd->work[0].cpu_binding = dd->isr_binding;
4460         dd->work[1].cpu_binding = get_least_used_cpu_on_node(dd->numa_node);
4461         dd->work[2].cpu_binding = get_least_used_cpu_on_node(dd->numa_node);
4462         dd->work[3].cpu_binding = dd->work[0].cpu_binding;
4463         dd->work[4].cpu_binding = dd->work[1].cpu_binding;
4464         dd->work[5].cpu_binding = dd->work[2].cpu_binding;
4465         dd->work[6].cpu_binding = dd->work[2].cpu_binding;
4466         dd->work[7].cpu_binding = dd->work[1].cpu_binding;
4467
4468         /* Log the bindings */
4469         for_each_present_cpu(cpu) {
4470                 memset(cpu_list, 0, sizeof(cpu_list));
4471                 for (i = 0, j = 0; i < MTIP_MAX_SLOT_GROUPS; i++) {
4472                         if (dd->work[i].cpu_binding == cpu) {
4473                                 snprintf(&cpu_list[j], 256 - j, "%d ", i);
4474                                 j = strlen(cpu_list);
4475                         }
4476                 }
4477                 if (j)
4478                         dev_info(&pdev->dev, "CPU %d: WQs %s\n", cpu, cpu_list);
4479         }
4480
4481         INIT_WORK(&dd->work[0].work, mtip_workq_sdbf0);
4482         INIT_WORK(&dd->work[1].work, mtip_workq_sdbf1);
4483         INIT_WORK(&dd->work[2].work, mtip_workq_sdbf2);
4484         INIT_WORK(&dd->work[3].work, mtip_workq_sdbf3);
4485         INIT_WORK(&dd->work[4].work, mtip_workq_sdbf4);
4486         INIT_WORK(&dd->work[5].work, mtip_workq_sdbf5);
4487         INIT_WORK(&dd->work[6].work, mtip_workq_sdbf6);
4488         INIT_WORK(&dd->work[7].work, mtip_workq_sdbf7);
4489
4490         pci_set_master(pdev);
4491         rv = pci_enable_msi(pdev);
4492         if (rv) {
4493                 dev_warn(&pdev->dev,
4494                         "Unable to enable MSI interrupt.\n");
4495                 goto msi_initialize_err;
4496         }
4497
4498         mtip_fix_ero_nosnoop(dd, pdev);
4499
4500         /* Initialize the block layer. */
4501         rv = mtip_block_initialize(dd);
4502         if (rv < 0) {
4503                 dev_err(&pdev->dev,
4504                         "Unable to initialize block layer\n");
4505                 goto block_initialize_err;
4506         }
4507
4508         /*
4509          * Increment the instance count so that each device has a unique
4510          * instance number.
4511          */
4512         instance++;
4513         if (rv != MTIP_FTL_REBUILD_MAGIC)
4514                 set_bit(MTIP_DDF_INIT_DONE_BIT, &dd->dd_flag);
4515         else
4516                 rv = 0; /* device in rebuild state, return 0 from probe */
4517
4518         /* Add to online list even if in ftl rebuild */
4519         spin_lock_irqsave(&dev_lock, flags);
4520         list_add(&dd->online_list, &online_list);
4521         spin_unlock_irqrestore(&dev_lock, flags);
4522
4523         goto done;
4524
4525 block_initialize_err:
4526         pci_disable_msi(pdev);
4527
4528 msi_initialize_err:
4529         if (dd->isr_workq) {
4530                 flush_workqueue(dd->isr_workq);
4531                 destroy_workqueue(dd->isr_workq);
4532                 drop_cpu(dd->work[0].cpu_binding);
4533                 drop_cpu(dd->work[1].cpu_binding);
4534                 drop_cpu(dd->work[2].cpu_binding);
4535         }
4536 setmask_err:
4537         pcim_iounmap_regions(pdev, 1 << MTIP_ABAR);
4538
4539 iomap_err:
4540         kfree(dd);
4541         pci_set_drvdata(pdev, NULL);
4542         return rv;
4543 done:
4544         return rv;
4545 }
4546
4547 /*
4548  * Called for each probed device when the device is removed or the
4549  * driver is unloaded.
4550  *
4551  * return value
4552  *      None
4553  */
4554 static void mtip_pci_remove(struct pci_dev *pdev)
4555 {
4556         struct driver_data *dd = pci_get_drvdata(pdev);
4557         unsigned long flags, to;
4558
4559         set_bit(MTIP_DDF_REMOVAL_BIT, &dd->dd_flag);
4560
4561         spin_lock_irqsave(&dev_lock, flags);
4562         list_del_init(&dd->online_list);
4563         list_add(&dd->remove_list, &removing_list);
4564         spin_unlock_irqrestore(&dev_lock, flags);
4565
4566         mtip_check_surprise_removal(pdev);
4567         synchronize_irq(dd->pdev->irq);
4568
4569         /* Spin until workers are done */
4570         to = jiffies + msecs_to_jiffies(4000);
4571         do {
4572                 msleep(20);
4573         } while (atomic_read(&dd->irq_workers_active) != 0 &&
4574                 time_before(jiffies, to));
4575
4576         if (!dd->sr)
4577                 fsync_bdev(dd->bdev);
4578
4579         if (atomic_read(&dd->irq_workers_active) != 0) {
4580                 dev_warn(&dd->pdev->dev,
4581                         "Completion workers still active!\n");
4582         }
4583
4584         blk_set_queue_dying(dd->queue);
4585         set_bit(MTIP_DDF_REMOVE_PENDING_BIT, &dd->dd_flag);
4586
4587         /* Clean up the block layer. */
4588         mtip_block_remove(dd);
4589
4590         if (dd->isr_workq) {
4591                 flush_workqueue(dd->isr_workq);
4592                 destroy_workqueue(dd->isr_workq);
4593                 drop_cpu(dd->work[0].cpu_binding);
4594                 drop_cpu(dd->work[1].cpu_binding);
4595                 drop_cpu(dd->work[2].cpu_binding);
4596         }
4597
4598         pci_disable_msi(pdev);
4599
4600         spin_lock_irqsave(&dev_lock, flags);
4601         list_del_init(&dd->remove_list);
4602         spin_unlock_irqrestore(&dev_lock, flags);
4603
4604         kfree(dd);
4605
4606         pcim_iounmap_regions(pdev, 1 << MTIP_ABAR);
4607         pci_set_drvdata(pdev, NULL);
4608 }
4609
4610 /*
4611  * Called for each probed device when the device is suspended.
4612  *
4613  * return value
4614  *      0  Success
4615  *      <0 Error
4616  */
4617 static int mtip_pci_suspend(struct pci_dev *pdev, pm_message_t mesg)
4618 {
4619         int rv = 0;
4620         struct driver_data *dd = pci_get_drvdata(pdev);
4621
4622         if (!dd) {
4623                 dev_err(&pdev->dev,
4624                         "Driver private datastructure is NULL\n");
4625                 return -EFAULT;
4626         }
4627
4628         set_bit(MTIP_DDF_RESUME_BIT, &dd->dd_flag);
4629
4630         /* Disable ports & interrupts then send standby immediate */
4631         rv = mtip_block_suspend(dd);
4632         if (rv < 0) {
4633                 dev_err(&pdev->dev,
4634                         "Failed to suspend controller\n");
4635                 return rv;
4636         }
4637
4638         /*
4639          * Save the pci config space to pdev structure &
4640          * disable the device
4641          */
4642         pci_save_state(pdev);
4643         pci_disable_device(pdev);
4644
4645         /* Move to Low power state*/
4646         pci_set_power_state(pdev, PCI_D3hot);
4647
4648         return rv;
4649 }
4650
4651 /*
4652  * Called for each probed device when the device is resumed.
4653  *
4654  * return value
4655  *      0  Success
4656  *      <0 Error
4657  */
4658 static int mtip_pci_resume(struct pci_dev *pdev)
4659 {
4660         int rv = 0;
4661         struct driver_data *dd;
4662
4663         dd = pci_get_drvdata(pdev);
4664         if (!dd) {
4665                 dev_err(&pdev->dev,
4666                         "Driver private datastructure is NULL\n");
4667                 return -EFAULT;
4668         }
4669
4670         /* Move the device to active State */
4671         pci_set_power_state(pdev, PCI_D0);
4672
4673         /* Restore PCI configuration space */
4674         pci_restore_state(pdev);
4675
4676         /* Enable the PCI device*/
4677         rv = pcim_enable_device(pdev);
4678         if (rv < 0) {
4679                 dev_err(&pdev->dev,
4680                         "Failed to enable card during resume\n");
4681                 goto err;
4682         }
4683         pci_set_master(pdev);
4684
4685         /*
4686          * Calls hbaReset, initPort, & startPort function
4687          * then enables interrupts
4688          */
4689         rv = mtip_block_resume(dd);
4690         if (rv < 0)
4691                 dev_err(&pdev->dev, "Unable to resume\n");
4692
4693 err:
4694         clear_bit(MTIP_DDF_RESUME_BIT, &dd->dd_flag);
4695
4696         return rv;
4697 }
4698
4699 /*
4700  * Shutdown routine
4701  *
4702  * return value
4703  *      None
4704  */
4705 static void mtip_pci_shutdown(struct pci_dev *pdev)
4706 {
4707         struct driver_data *dd = pci_get_drvdata(pdev);
4708         if (dd)
4709                 mtip_block_shutdown(dd);
4710 }
4711
4712 /* Table of device ids supported by this driver. */
4713 static const struct pci_device_id mtip_pci_tbl[] = {
4714         { PCI_DEVICE(PCI_VENDOR_ID_MICRON, P320H_DEVICE_ID) },
4715         { PCI_DEVICE(PCI_VENDOR_ID_MICRON, P320M_DEVICE_ID) },
4716         { PCI_DEVICE(PCI_VENDOR_ID_MICRON, P320S_DEVICE_ID) },
4717         { PCI_DEVICE(PCI_VENDOR_ID_MICRON, P325M_DEVICE_ID) },
4718         { PCI_DEVICE(PCI_VENDOR_ID_MICRON, P420H_DEVICE_ID) },
4719         { PCI_DEVICE(PCI_VENDOR_ID_MICRON, P420M_DEVICE_ID) },
4720         { PCI_DEVICE(PCI_VENDOR_ID_MICRON, P425M_DEVICE_ID) },
4721         { 0 }
4722 };
4723
4724 /* Structure that describes the PCI driver functions. */
4725 static struct pci_driver mtip_pci_driver = {
4726         .name                   = MTIP_DRV_NAME,
4727         .id_table               = mtip_pci_tbl,
4728         .probe                  = mtip_pci_probe,
4729         .remove                 = mtip_pci_remove,
4730         .suspend                = mtip_pci_suspend,
4731         .resume                 = mtip_pci_resume,
4732         .shutdown               = mtip_pci_shutdown,
4733 };
4734
4735 MODULE_DEVICE_TABLE(pci, mtip_pci_tbl);
4736
4737 /*
4738  * Module initialization function.
4739  *
4740  * Called once when the module is loaded. This function allocates a major
4741  * block device number to the Cyclone devices and registers the PCI layer
4742  * of the driver.
4743  *
4744  * Return value
4745  *      0 on success else error code.
4746  */
4747 static int __init mtip_init(void)
4748 {
4749         int error;
4750
4751         pr_info(MTIP_DRV_NAME " Version " MTIP_DRV_VERSION "\n");
4752
4753         spin_lock_init(&dev_lock);
4754
4755         INIT_LIST_HEAD(&online_list);
4756         INIT_LIST_HEAD(&removing_list);
4757
4758         /* Allocate a major block device number to use with this driver. */
4759         error = register_blkdev(0, MTIP_DRV_NAME);
4760         if (error <= 0) {
4761                 pr_err("Unable to register block device (%d)\n",
4762                 error);
4763                 return -EBUSY;
4764         }
4765         mtip_major = error;
4766
4767         dfs_parent = debugfs_create_dir("rssd", NULL);
4768         if (IS_ERR_OR_NULL(dfs_parent)) {
4769                 pr_warn("Error creating debugfs parent\n");
4770                 dfs_parent = NULL;
4771         }
4772         if (dfs_parent) {
4773                 dfs_device_status = debugfs_create_file("device_status",
4774                                         S_IRUGO, dfs_parent, NULL,
4775                                         &mtip_device_status_fops);
4776                 if (IS_ERR_OR_NULL(dfs_device_status)) {
4777                         pr_err("Error creating device_status node\n");
4778                         dfs_device_status = NULL;
4779                 }
4780         }
4781
4782         /* Register our PCI operations. */
4783         error = pci_register_driver(&mtip_pci_driver);
4784         if (error) {
4785                 debugfs_remove(dfs_parent);
4786                 unregister_blkdev(mtip_major, MTIP_DRV_NAME);
4787         }
4788
4789         return error;
4790 }
4791
4792 /*
4793  * Module de-initialization function.
4794  *
4795  * Called once when the module is unloaded. This function deallocates
4796  * the major block device number allocated by mtip_init() and
4797  * unregisters the PCI layer of the driver.
4798  *
4799  * Return value
4800  *      none
4801  */
4802 static void __exit mtip_exit(void)
4803 {
4804         /* Release the allocated major block device number. */
4805         unregister_blkdev(mtip_major, MTIP_DRV_NAME);
4806
4807         /* Unregister the PCI driver. */
4808         pci_unregister_driver(&mtip_pci_driver);
4809
4810         debugfs_remove_recursive(dfs_parent);
4811 }
4812
4813 MODULE_AUTHOR("Micron Technology, Inc");
4814 MODULE_DESCRIPTION("Micron RealSSD PCIe Block Driver");
4815 MODULE_LICENSE("GPL");
4816 MODULE_VERSION(MTIP_DRV_VERSION);
4817
4818 module_init(mtip_init);
4819 module_exit(mtip_exit);