drivers/dma/ioat_dma.c

   1 /*
   2  * Intel I/OAT DMA Linux driver
   3  * Copyright(c) 2004 - 2007 Intel Corporation.
   4  *
   5  * This program is free software; you can redistribute it and/or modify it
   6  * under the terms and conditions of the GNU General Public License,
   7  * version 2, as published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  12  * more details.
  13  *
  14  * You should have received a copy of the GNU General Public License along with
  15  * this program; if not, write to the Free Software Foundation, Inc.,
  16  * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
  17  *
  18  * The full GNU General Public License is included in this distribution in
  19  * the file called "COPYING".
  20  *
  21  */
  22
  23 /*
  24  * This driver supports an Intel I/OAT DMA engine, which does asynchronous
  25  * copy operations.
  26  */
  27
  28 #include <linux/init.h>
  29 #include <linux/module.h>
  30 #include <linux/pci.h>
  31 #include <linux/interrupt.h>
  32 #include <linux/dmaengine.h>
  33 #include <linux/delay.h>
  34 #include <linux/dma-mapping.h>
  35 #include <linux/workqueue.h>
  36 #include "ioatdma.h"
  37 #include "ioatdma_registers.h"
  38 #include "ioatdma_hw.h"
  39
  40 #define to_ioat_chan(chan) container_of(chan, struct ioat_dma_chan, common)
  41 #define to_ioatdma_device(dev) container_of(dev, struct ioatdma_device, common)
  42 #define to_ioat_desc(lh) container_of(lh, struct ioat_desc_sw, node)
  43 #define tx_to_ioat_desc(tx) container_of(tx, struct ioat_desc_sw, async_tx)
  44
  45 #define chan_num(ch) ((int)((ch)->reg_base - (ch)->device->reg_base) / 0x80)
  46 static int ioat_pending_level = 4;
  47 module_param(ioat_pending_level, int, 0644);
  48 MODULE_PARM_DESC(ioat_pending_level,
  49                  "high-water mark for pushing ioat descriptors (default: 4)");
  50
  51 #define RESET_DELAY  msecs_to_jiffies(100)
  52 #define WATCHDOG_DELAY  round_jiffies(msecs_to_jiffies(2000))
  53 static void ioat_dma_chan_reset_part2(struct work_struct *work);
  54 static void ioat_dma_chan_watchdog(struct work_struct *work);
  55
  56 /*
  57  * workaround for IOAT ver.3.0 null descriptor issue
  58  * (channel returns error when size is 0)
  59  */
  60 #define NULL_DESC_BUFFER_SIZE 1
  61
  62 /* internal functions */
  63 static void ioat_dma_start_null_desc(struct ioat_dma_chan *ioat_chan);
  64 static void ioat_dma_memcpy_cleanup(struct ioat_dma_chan *ioat_chan);
  65
  66 static struct ioat_desc_sw *
  67 ioat1_dma_get_next_descriptor(struct ioat_dma_chan *ioat_chan);
  68 static struct ioat_desc_sw *
  69 ioat2_dma_get_next_descriptor(struct ioat_dma_chan *ioat_chan);
  70
  71 static inline struct ioat_dma_chan *ioat_lookup_chan_by_index(
  72                                                 struct ioatdma_device *device,
  73                                                 int index)
  74 {
  75         return device->idx[index];
  76 }
  77
  78 /**
  79  * ioat_dma_do_interrupt - handler used for single vector interrupt mode
  80  * @irq: interrupt id
  81  * @data: interrupt data
  82  */
  83 static irqreturn_t ioat_dma_do_interrupt(int irq, void *data)
  84 {
  85         struct ioatdma_device *instance = data;
  86         struct ioat_dma_chan *ioat_chan;
  87         unsigned long attnstatus;
  88         int bit;
  89         u8 intrctrl;
  90
  91         intrctrl = readb(instance->reg_base + IOAT_INTRCTRL_OFFSET);
  92
  93         if (!(intrctrl & IOAT_INTRCTRL_MASTER_INT_EN))
  94                 return IRQ_NONE;
  95
  96         if (!(intrctrl & IOAT_INTRCTRL_INT_STATUS)) {
  97                 writeb(intrctrl, instance->reg_base + IOAT_INTRCTRL_OFFSET);
  98                 return IRQ_NONE;
  99         }
 100
 101         attnstatus = readl(instance->reg_base + IOAT_ATTNSTATUS_OFFSET);
 102         for_each_bit(bit, &attnstatus, BITS_PER_LONG) {
 103                 ioat_chan = ioat_lookup_chan_by_index(instance, bit);
 104                 tasklet_schedule(&ioat_chan->cleanup_task);
 105         }
 106
 107         writeb(intrctrl, instance->reg_base + IOAT_INTRCTRL_OFFSET);
 108         return IRQ_HANDLED;
 109 }
 110
 111 /**
 112  * ioat_dma_do_interrupt_msix - handler used for vector-per-channel interrupt mode
 113  * @irq: interrupt id
 114  * @data: interrupt data
 115  */
 116 static irqreturn_t ioat_dma_do_interrupt_msix(int irq, void *data)
 117 {
 118         struct ioat_dma_chan *ioat_chan = data;
 119
 120         tasklet_schedule(&ioat_chan->cleanup_task);
 121
 122         return IRQ_HANDLED;
 123 }
 124
 125 static void ioat_dma_cleanup_tasklet(unsigned long data);
 126
 127 /**
 128  * ioat_dma_enumerate_channels - find and initialize the device's channels
 129  * @device: the device to be enumerated
 130  */
 131 static int ioat_dma_enumerate_channels(struct ioatdma_device *device)
 132 {
 133         u8 xfercap_scale;
 134         u32 xfercap;
 135         int i;
 136         struct ioat_dma_chan *ioat_chan;
 137
 138         /*
 139          * IOAT ver.3 workarounds
 140          */
 141         if (device->version == IOAT_VER_3_0) {
 142                 u32 chan_err_mask;
 143                 u16 dev_id;
 144                 u32 dmauncerrsts;
 145
 146                 /*
 147                  * Write CHANERRMSK_INT with 3E07h to mask out the errors
 148                  * that can cause stability issues for IOAT ver.3
 149                  */
 150                 chan_err_mask = 0x3E07;
 151                 pci_write_config_dword(device->pdev,
 152                         IOAT_PCI_CHANERRMASK_INT_OFFSET,
 153                         chan_err_mask);
 154
 155                 /*
 156                  * Clear DMAUNCERRSTS Cfg-Reg Parity Error status bit
 157                  * (workaround for spurious config parity error after restart)
 158                  */
 159                 pci_read_config_word(device->pdev,
 160                         IOAT_PCI_DEVICE_ID_OFFSET,
 161                         &dev_id);
 162                 if (dev_id == PCI_DEVICE_ID_INTEL_IOAT_TBG0) {
 163                         dmauncerrsts = 0x10;
 164                         pci_write_config_dword(device->pdev,
 165                                 IOAT_PCI_DMAUNCERRSTS_OFFSET,
 166                                 dmauncerrsts);
 167                 }
 168         }
 169
 170         device->common.chancnt = readb(device->reg_base + IOAT_CHANCNT_OFFSET);
 171         xfercap_scale = readb(device->reg_base + IOAT_XFERCAP_OFFSET);
 172         xfercap = (xfercap_scale == 0 ? -1 : (1UL << xfercap_scale));
 173
 174         for (i = 0; i < device->common.chancnt; i++) {
 175                 ioat_chan = kzalloc(sizeof(*ioat_chan), GFP_KERNEL);
 176                 if (!ioat_chan) {
 177                         device->common.chancnt = i;
 178                         break;
 179                 }
 180
 181                 ioat_chan->device = device;
 182                 ioat_chan->reg_base = device->reg_base + (0x80 * (i + 1));
 183                 ioat_chan->xfercap = xfercap;
 184                 ioat_chan->desccount = 0;
 185                 INIT_DELAYED_WORK(&ioat_chan->work, ioat_dma_chan_reset_part2);
 186                 if (ioat_chan->device->version != IOAT_VER_1_2) {
 187                         writel(IOAT_DCACTRL_CMPL_WRITE_ENABLE
 188                                         | IOAT_DMA_DCA_ANY_CPU,
 189                                 ioat_chan->reg_base + IOAT_DCACTRL_OFFSET);
 190                 }
 191                 spin_lock_init(&ioat_chan->cleanup_lock);
 192                 spin_lock_init(&ioat_chan->desc_lock);
 193                 INIT_LIST_HEAD(&ioat_chan->free_desc);
 194                 INIT_LIST_HEAD(&ioat_chan->used_desc);
 195                 /* This should be made common somewhere in dmaengine.c */
 196                 ioat_chan->common.device = &device->common;
 197                 list_add_tail(&ioat_chan->common.device_node,
 198                               &device->common.channels);
 199                 device->idx[i] = ioat_chan;
 200                 tasklet_init(&ioat_chan->cleanup_task,
 201                              ioat_dma_cleanup_tasklet,
 202                              (unsigned long) ioat_chan);
 203                 tasklet_disable(&ioat_chan->cleanup_task);
 204         }
 205         return device->common.chancnt;
 206 }
 207
 208 /**
 209  * ioat_dma_memcpy_issue_pending - push potentially unrecognized appended
 210  *                                 descriptors to hw
 211  * @chan: DMA channel handle
 212  */
 213 static inline void __ioat1_dma_memcpy_issue_pending(
 214                                                 struct ioat_dma_chan *ioat_chan)
 215 {
 216         ioat_chan->pending = 0;
 217         writeb(IOAT_CHANCMD_APPEND, ioat_chan->reg_base + IOAT1_CHANCMD_OFFSET);
 218 }
 219
 220 static void ioat1_dma_memcpy_issue_pending(struct dma_chan *chan)
 221 {
 222         struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
 223
 224         if (ioat_chan->pending > 0) {
 225                 spin_lock_bh(&ioat_chan->desc_lock);
 226                 __ioat1_dma_memcpy_issue_pending(ioat_chan);
 227                 spin_unlock_bh(&ioat_chan->desc_lock);
 228         }
 229 }
 230
 231 static inline void __ioat2_dma_memcpy_issue_pending(
 232                                                 struct ioat_dma_chan *ioat_chan)
 233 {
 234         ioat_chan->pending = 0;
 235         writew(ioat_chan->dmacount,
 236                ioat_chan->reg_base + IOAT_CHAN_DMACOUNT_OFFSET);
 237 }
 238
 239 static void ioat2_dma_memcpy_issue_pending(struct dma_chan *chan)
 240 {
 241         struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
 242
 243         if (ioat_chan->pending > 0) {
 244                 spin_lock_bh(&ioat_chan->desc_lock);
 245                 __ioat2_dma_memcpy_issue_pending(ioat_chan);
 246                 spin_unlock_bh(&ioat_chan->desc_lock);
 247         }
 248 }
 249
 250
 251 /**
 252  * ioat_dma_chan_reset_part2 - reinit the channel after a reset
 253  */
 254 static void ioat_dma_chan_reset_part2(struct work_struct *work)
 255 {
 256         struct ioat_dma_chan *ioat_chan =
 257                 container_of(work, struct ioat_dma_chan, work.work);
 258         struct ioat_desc_sw *desc;
 259
 260         spin_lock_bh(&ioat_chan->cleanup_lock);
 261         spin_lock_bh(&ioat_chan->desc_lock);
 262
 263         ioat_chan->completion_virt->low = 0;
 264         ioat_chan->completion_virt->high = 0;
 265         ioat_chan->pending = 0;
 266
 267         /*
 268          * count the descriptors waiting, and be sure to do it
 269          * right for both the CB1 line and the CB2 ring
 270          */
 271         ioat_chan->dmacount = 0;
 272         if (ioat_chan->used_desc.prev) {
 273                 desc = to_ioat_desc(ioat_chan->used_desc.prev);
 274                 do {
 275                         ioat_chan->dmacount++;
 276                         desc = to_ioat_desc(desc->node.next);
 277                 } while (&desc->node != ioat_chan->used_desc.next);
 278         }
 279
 280         /*
 281          * write the new starting descriptor address
 282          * this puts channel engine into ARMED state
 283          */
 284         desc = to_ioat_desc(ioat_chan->used_desc.prev);
 285         switch (ioat_chan->device->version) {
 286         case IOAT_VER_1_2:
 287                 writel(((u64) desc->async_tx.phys) & 0x00000000FFFFFFFF,
 288                        ioat_chan->reg_base + IOAT1_CHAINADDR_OFFSET_LOW);
 289                 writel(((u64) desc->async_tx.phys) >> 32,
 290                        ioat_chan->reg_base + IOAT1_CHAINADDR_OFFSET_HIGH);
 291
 292                 writeb(IOAT_CHANCMD_START, ioat_chan->reg_base
 293                         + IOAT_CHANCMD_OFFSET(ioat_chan->device->version));
 294                 break;
 295         case IOAT_VER_2_0:
 296                 writel(((u64) desc->async_tx.phys) & 0x00000000FFFFFFFF,
 297                        ioat_chan->reg_base + IOAT2_CHAINADDR_OFFSET_LOW);
 298                 writel(((u64) desc->async_tx.phys) >> 32,
 299                        ioat_chan->reg_base + IOAT2_CHAINADDR_OFFSET_HIGH);
 300
 301                 /* tell the engine to go with what's left to be done */
 302                 writew(ioat_chan->dmacount,
 303                        ioat_chan->reg_base + IOAT_CHAN_DMACOUNT_OFFSET);
 304
 305                 break;
 306         }
 307         dev_err(&ioat_chan->device->pdev->dev,
 308                 "chan%d reset - %d descs waiting, %d total desc\n",
 309                 chan_num(ioat_chan), ioat_chan->dmacount, ioat_chan->desccount);
 310
 311         spin_unlock_bh(&ioat_chan->desc_lock);
 312         spin_unlock_bh(&ioat_chan->cleanup_lock);
 313 }
 314
 315 /**
 316  * ioat_dma_reset_channel - restart a channel
 317  * @ioat_chan: IOAT DMA channel handle
 318  */
 319 static void ioat_dma_reset_channel(struct ioat_dma_chan *ioat_chan)
 320 {
 321         u32 chansts, chanerr;
 322
 323         if (!ioat_chan->used_desc.prev)
 324                 return;
 325
 326         chanerr = readl(ioat_chan->reg_base + IOAT_CHANERR_OFFSET);
 327         chansts = (ioat_chan->completion_virt->low
 328                                         & IOAT_CHANSTS_DMA_TRANSFER_STATUS);
 329         if (chanerr) {
 330                 dev_err(&ioat_chan->device->pdev->dev,
 331                         "chan%d, CHANSTS = 0x%08x CHANERR = 0x%04x, clearing\n",
 332                         chan_num(ioat_chan), chansts, chanerr);
 333                 writel(chanerr, ioat_chan->reg_base + IOAT_CHANERR_OFFSET);
 334         }
 335
 336         /*
 337          * whack it upside the head with a reset
 338          * and wait for things to settle out.
 339          * force the pending count to a really big negative
 340          * to make sure no one forces an issue_pending
 341          * while we're waiting.
 342          */
 343
 344         spin_lock_bh(&ioat_chan->desc_lock);
 345         ioat_chan->pending = INT_MIN;
 346         writeb(IOAT_CHANCMD_RESET,
 347                ioat_chan->reg_base
 348                + IOAT_CHANCMD_OFFSET(ioat_chan->device->version));
 349         spin_unlock_bh(&ioat_chan->desc_lock);
 350
 351         /* schedule the 2nd half instead of sleeping a long time */
 352         schedule_delayed_work(&ioat_chan->work, RESET_DELAY);
 353 }
 354
 355 /**
 356  * ioat_dma_chan_watchdog - watch for stuck channels
 357  */
 358 static void ioat_dma_chan_watchdog(struct work_struct *work)
 359 {
 360         struct ioatdma_device *device =
 361                 container_of(work, struct ioatdma_device, work.work);
 362         struct ioat_dma_chan *ioat_chan;
 363         int i;
 364
 365         union {
 366                 u64 full;
 367                 struct {
 368                         u32 low;
 369                         u32 high;
 370                 };
 371         } completion_hw;
 372         unsigned long compl_desc_addr_hw;
 373
 374         for (i = 0; i < device->common.chancnt; i++) {
 375                 ioat_chan = ioat_lookup_chan_by_index(device, i);
 376
 377                 if (ioat_chan->device->version == IOAT_VER_1_2
 378                         /* have we started processing anything yet */
 379                     && ioat_chan->last_completion
 380                         /* have we completed any since last watchdog cycle? */
 381                     && (ioat_chan->last_completion ==
 382                                 ioat_chan->watchdog_completion)
 383                         /* has TCP stuck on one cookie since last watchdog? */
 384                     && (ioat_chan->watchdog_tcp_cookie ==
 385                                 ioat_chan->watchdog_last_tcp_cookie)
 386                     && (ioat_chan->watchdog_tcp_cookie !=
 387                                 ioat_chan->completed_cookie)
 388                         /* is there something in the chain to be processed? */
 389                         /* CB1 chain always has at least the last one processed */
 390                     && (ioat_chan->used_desc.prev != ioat_chan->used_desc.next)
 391                     && ioat_chan->pending == 0) {
 392
 393                         /*
 394                          * check CHANSTS register for completed
 395                          * descriptor address.
 396                          * if it is different than completion writeback,
 397                          * it is not zero
 398                          * and it has changed since the last watchdog
 399                          *     we can assume that channel
 400                          *     is still working correctly
 401                          *     and the problem is in completion writeback.
 402                          *     update completion writeback
 403                          *     with actual CHANSTS value
 404                          * else
 405                          *     try resetting the channel
 406                          */
 407
 408                         completion_hw.low = readl(ioat_chan->reg_base +
 409                                 IOAT_CHANSTS_OFFSET_LOW(ioat_chan->device->version));
 410                         completion_hw.high = readl(ioat_chan->reg_base +
 411                                 IOAT_CHANSTS_OFFSET_HIGH(ioat_chan->device->version));
 412 #if (BITS_PER_LONG == 64)
 413                         compl_desc_addr_hw =
 414                                 completion_hw.full
 415                                 & IOAT_CHANSTS_COMPLETED_DESCRIPTOR_ADDR;
 416 #else
 417                         compl_desc_addr_hw =
 418                                 completion_hw.low & IOAT_LOW_COMPLETION_MASK;
 419 #endif
 420
 421                         if ((compl_desc_addr_hw != 0)
 422                            && (compl_desc_addr_hw != ioat_chan->watchdog_completion)
 423                            && (compl_desc_addr_hw != ioat_chan->last_compl_desc_addr_hw)) {
 424                                 ioat_chan->last_compl_desc_addr_hw = compl_desc_addr_hw;
 425                                 ioat_chan->completion_virt->low = completion_hw.low;
 426                                 ioat_chan->completion_virt->high = completion_hw.high;
 427                         } else {
 428                                 ioat_dma_reset_channel(ioat_chan);
 429                                 ioat_chan->watchdog_completion = 0;
 430                                 ioat_chan->last_compl_desc_addr_hw = 0;
 431                         }
 432
 433                 /*
 434                  * for version 2.0 if there are descriptors yet to be processed
 435                  * and the last completed hasn't changed since the last watchdog
 436                  *      if they haven't hit the pending level
 437                  *          issue the pending to push them through
 438                  *      else
 439                  *          try resetting the channel
 440                  */
 441                 } else if (ioat_chan->device->version == IOAT_VER_2_0
 442                     && ioat_chan->used_desc.prev
 443                     && ioat_chan->last_completion
 444                     && ioat_chan->last_completion == ioat_chan->watchdog_completion) {
 445
 446                         if (ioat_chan->pending < ioat_pending_level)
 447                                 ioat2_dma_memcpy_issue_pending(&ioat_chan->common);
 448                         else {
 449                                 ioat_dma_reset_channel(ioat_chan);
 450                                 ioat_chan->watchdog_completion = 0;
 451                         }
 452                 } else {
 453                         ioat_chan->last_compl_desc_addr_hw = 0;
 454                         ioat_chan->watchdog_completion
 455                                         = ioat_chan->last_completion;
 456                 }
 457
 458                 ioat_chan->watchdog_last_tcp_cookie =
 459                         ioat_chan->watchdog_tcp_cookie;
 460         }
 461
 462         schedule_delayed_work(&device->work, WATCHDOG_DELAY);
 463 }
 464
 465 static dma_cookie_t ioat1_tx_submit(struct dma_async_tx_descriptor *tx)
 466 {
 467         struct ioat_dma_chan *ioat_chan = to_ioat_chan(tx->chan);
 468         struct ioat_desc_sw *first = tx_to_ioat_desc(tx);
 469         struct ioat_desc_sw *prev, *new;
 470         struct ioat_dma_descriptor *hw;
 471         dma_cookie_t cookie;
 472         LIST_HEAD(new_chain);
 473         u32 copy;
 474         size_t len;
 475         dma_addr_t src, dst;
 476         unsigned long orig_flags;
 477         unsigned int desc_count = 0;
 478
 479         /* src and dest and len are stored in the initial descriptor */
 480         len = first->len;
 481         src = first->src;
 482         dst = first->dst;
 483         orig_flags = first->async_tx.flags;
 484         new = first;
 485
 486         spin_lock_bh(&ioat_chan->desc_lock);
 487         prev = to_ioat_desc(ioat_chan->used_desc.prev);
 488         prefetch(prev->hw);
 489         do {
 490                 copy = min_t(size_t, len, ioat_chan->xfercap);
 491
 492                 async_tx_ack(&new->async_tx);
 493
 494                 hw = new->hw;
 495                 hw->size = copy;
 496                 hw->ctl = 0;
 497                 hw->src_addr = src;
 498                 hw->dst_addr = dst;
 499                 hw->next = 0;
 500
 501                 /* chain together the physical address list for the HW */
 502                 wmb();
 503                 prev->hw->next = (u64) new->async_tx.phys;
 504
 505                 len -= copy;
 506                 dst += copy;
 507                 src += copy;
 508
 509                 list_add_tail(&new->node, &new_chain);
 510                 desc_count++;
 511                 prev = new;
 512         } while (len && (new = ioat1_dma_get_next_descriptor(ioat_chan)));
 513
 514         if (!new) {
 515                 dev_err(&ioat_chan->device->pdev->dev,
 516                         "tx submit failed\n");
 517                 spin_unlock_bh(&ioat_chan->desc_lock);
 518                 return -ENOMEM;
 519         }
 520
 521         hw->ctl = IOAT_DMA_DESCRIPTOR_CTL_CP_STS;
 522         if (new->async_tx.callback) {
 523                 hw->ctl |= IOAT_DMA_DESCRIPTOR_CTL_INT_GN;
 524                 if (first != new) {
 525                         /* move callback into to last desc */
 526                         new->async_tx.callback = first->async_tx.callback;
 527                         new->async_tx.callback_param
 528                                         = first->async_tx.callback_param;
 529                         first->async_tx.callback = NULL;
 530                         first->async_tx.callback_param = NULL;
 531                 }
 532         }
 533
 534         new->tx_cnt = desc_count;
 535         new->async_tx.flags = orig_flags; /* client is in control of this ack */
 536
 537         /* store the original values for use in later cleanup */
 538         if (new != first) {
 539                 new->src = first->src;
 540                 new->dst = first->dst;
 541                 new->len = first->len;
 542         }
 543
 544         /* cookie incr and addition to used_list must be atomic */
 545         cookie = ioat_chan->common.cookie;
 546         cookie++;
 547         if (cookie < 0)
 548                 cookie = 1;
 549         ioat_chan->common.cookie = new->async_tx.cookie = cookie;
 550
 551         /* write address into NextDescriptor field of last desc in chain */
 552         to_ioat_desc(ioat_chan->used_desc.prev)->hw->next =
 553                                                         first->async_tx.phys;
 554         list_splice_tail(&new_chain, &ioat_chan->used_desc);
 555
 556         ioat_chan->dmacount += desc_count;
 557         ioat_chan->pending += desc_count;
 558         if (ioat_chan->pending >= ioat_pending_level)
 559                 __ioat1_dma_memcpy_issue_pending(ioat_chan);
 560         spin_unlock_bh(&ioat_chan->desc_lock);
 561
 562         return cookie;
 563 }
 564
 565 static dma_cookie_t ioat2_tx_submit(struct dma_async_tx_descriptor *tx)
 566 {
 567         struct ioat_dma_chan *ioat_chan = to_ioat_chan(tx->chan);
 568         struct ioat_desc_sw *first = tx_to_ioat_desc(tx);
 569         struct ioat_desc_sw *new;
 570         struct ioat_dma_descriptor *hw;
 571         dma_cookie_t cookie;
 572         u32 copy;
 573         size_t len;
 574         dma_addr_t src, dst;
 575         unsigned long orig_flags;
 576         unsigned int desc_count = 0;
 577
 578         /* src and dest and len are stored in the initial descriptor */
 579         len = first->len;
 580         src = first->src;
 581         dst = first->dst;
 582         orig_flags = first->async_tx.flags;
 583         new = first;
 584
 585         /*
 586          * ioat_chan->desc_lock is still in force in version 2 path
 587          * it gets unlocked at end of this function
 588          */
 589         do {
 590                 copy = min_t(size_t, len, ioat_chan->xfercap);
 591
 592                 async_tx_ack(&new->async_tx);
 593
 594                 hw = new->hw;
 595                 hw->size = copy;
 596                 hw->ctl = 0;
 597                 hw->src_addr = src;
 598                 hw->dst_addr = dst;
 599
 600                 len -= copy;
 601                 dst += copy;
 602                 src += copy;
 603                 desc_count++;
 604         } while (len && (new = ioat2_dma_get_next_descriptor(ioat_chan)));
 605
 606         if (!new) {
 607                 dev_err(&ioat_chan->device->pdev->dev,
 608                         "tx submit failed\n");
 609                 spin_unlock_bh(&ioat_chan->desc_lock);
 610                 return -ENOMEM;
 611         }
 612
 613         hw->ctl |= IOAT_DMA_DESCRIPTOR_CTL_CP_STS;
 614         if (new->async_tx.callback) {
 615                 hw->ctl |= IOAT_DMA_DESCRIPTOR_CTL_INT_GN;
 616                 if (first != new) {
 617                         /* move callback into to last desc */
 618                         new->async_tx.callback = first->async_tx.callback;
 619                         new->async_tx.callback_param
 620                                         = first->async_tx.callback_param;
 621                         first->async_tx.callback = NULL;
 622                         first->async_tx.callback_param = NULL;
 623                 }
 624         }
 625
 626         new->tx_cnt = desc_count;
 627         new->async_tx.flags = orig_flags; /* client is in control of this ack */
 628
 629         /* store the original values for use in later cleanup */
 630         if (new != first) {
 631                 new->src = first->src;
 632                 new->dst = first->dst;
 633                 new->len = first->len;
 634         }
 635
 636         /* cookie incr and addition to used_list must be atomic */
 637         cookie = ioat_chan->common.cookie;
 638         cookie++;
 639         if (cookie < 0)
 640                 cookie = 1;
 641         ioat_chan->common.cookie = new->async_tx.cookie = cookie;
 642
 643         ioat_chan->dmacount += desc_count;
 644         ioat_chan->pending += desc_count;
 645         if (ioat_chan->pending >= ioat_pending_level)
 646                 __ioat2_dma_memcpy_issue_pending(ioat_chan);
 647         spin_unlock_bh(&ioat_chan->desc_lock);
 648
 649         return cookie;
 650 }
 651
 652 /**
 653  * ioat_dma_alloc_descriptor - allocate and return a sw and hw descriptor pair
 654  * @ioat_chan: the channel supplying the memory pool for the descriptors
 655  * @flags: allocation flags
 656  */
 657 static struct ioat_desc_sw *ioat_dma_alloc_descriptor(
 658                                         struct ioat_dma_chan *ioat_chan,
 659                                         gfp_t flags)
 660 {
 661         struct ioat_dma_descriptor *desc;
 662         struct ioat_desc_sw *desc_sw;
 663         struct ioatdma_device *ioatdma_device;
 664         dma_addr_t phys;
 665
 666         ioatdma_device = to_ioatdma_device(ioat_chan->common.device);
 667         desc = pci_pool_alloc(ioatdma_device->dma_pool, flags, &phys);
 668         if (unlikely(!desc))
 669                 return NULL;
 670
 671         desc_sw = kzalloc(sizeof(*desc_sw), flags);
 672         if (unlikely(!desc_sw)) {
 673                 pci_pool_free(ioatdma_device->dma_pool, desc, phys);
 674                 return NULL;
 675         }
 676
 677         memset(desc, 0, sizeof(*desc));
 678         dma_async_tx_descriptor_init(&desc_sw->async_tx, &ioat_chan->common);
 679         switch (ioat_chan->device->version) {
 680         case IOAT_VER_1_2:
 681                 desc_sw->async_tx.tx_submit = ioat1_tx_submit;
 682                 break;
 683         case IOAT_VER_2_0:
 684         case IOAT_VER_3_0:
 685                 desc_sw->async_tx.tx_submit = ioat2_tx_submit;
 686                 break;
 687         }
 688         INIT_LIST_HEAD(&desc_sw->async_tx.tx_list);
 689
 690         desc_sw->hw = desc;
 691         desc_sw->async_tx.phys = phys;
 692
 693         return desc_sw;
 694 }
 695
 696 static int ioat_initial_desc_count = 256;
 697 module_param(ioat_initial_desc_count, int, 0644);
 698 MODULE_PARM_DESC(ioat_initial_desc_count,
 699                  "initial descriptors per channel (default: 256)");
 700
 701 /**
 702  * ioat2_dma_massage_chan_desc - link the descriptors into a circle
 703  * @ioat_chan: the channel to be massaged
 704  */
 705 static void ioat2_dma_massage_chan_desc(struct ioat_dma_chan *ioat_chan)
 706 {
 707         struct ioat_desc_sw *desc, *_desc;
 708
 709         /* setup used_desc */
 710         ioat_chan->used_desc.next = ioat_chan->free_desc.next;
 711         ioat_chan->used_desc.prev = NULL;
 712
 713         /* pull free_desc out of the circle so that every node is a hw
 714          * descriptor, but leave it pointing to the list
 715          */
 716         ioat_chan->free_desc.prev->next = ioat_chan->free_desc.next;
 717         ioat_chan->free_desc.next->prev = ioat_chan->free_desc.prev;
 718
 719         /* circle link the hw descriptors */
 720         desc = to_ioat_desc(ioat_chan->free_desc.next);
 721         desc->hw->next = to_ioat_desc(desc->node.next)->async_tx.phys;
 722         list_for_each_entry_safe(desc, _desc, ioat_chan->free_desc.next, node) {
 723                 desc->hw->next = to_ioat_desc(desc->node.next)->async_tx.phys;
 724         }
 725 }
 726
 727 /**
 728  * ioat_dma_alloc_chan_resources - returns the number of allocated descriptors
 729  * @chan: the channel to be filled out
 730  */
 731 static int ioat_dma_alloc_chan_resources(struct dma_chan *chan,
 732                                          struct dma_client *client)
 733 {
 734         struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
 735         struct ioat_desc_sw *desc;
 736         u16 chanctrl;
 737         u32 chanerr;
 738         int i;
 739         LIST_HEAD(tmp_list);
 740
 741         /* have we already been set up? */
 742         if (!list_empty(&ioat_chan->free_desc))
 743                 return ioat_chan->desccount;
 744
 745         /* Setup register to interrupt and write completion status on error */
 746         chanctrl = IOAT_CHANCTRL_ERR_INT_EN |
 747                 IOAT_CHANCTRL_ANY_ERR_ABORT_EN |
 748                 IOAT_CHANCTRL_ERR_COMPLETION_EN;
 749         writew(chanctrl, ioat_chan->reg_base + IOAT_CHANCTRL_OFFSET);
 750
 751         chanerr = readl(ioat_chan->reg_base + IOAT_CHANERR_OFFSET);
 752         if (chanerr) {
 753                 dev_err(&ioat_chan->device->pdev->dev,
 754                         "CHANERR = %x, clearing\n", chanerr);
 755                 writel(chanerr, ioat_chan->reg_base + IOAT_CHANERR_OFFSET);
 756         }
 757
 758         /* Allocate descriptors */
 759         for (i = 0; i < ioat_initial_desc_count; i++) {
 760                 desc = ioat_dma_alloc_descriptor(ioat_chan, GFP_KERNEL);
 761                 if (!desc) {
 762                         dev_err(&ioat_chan->device->pdev->dev,
 763                                 "Only %d initial descriptors\n", i);
 764                         break;
 765                 }
 766                 list_add_tail(&desc->node, &tmp_list);
 767         }
 768         spin_lock_bh(&ioat_chan->desc_lock);
 769         ioat_chan->desccount = i;
 770         list_splice(&tmp_list, &ioat_chan->free_desc);
 771         if (ioat_chan->device->version != IOAT_VER_1_2)
 772                 ioat2_dma_massage_chan_desc(ioat_chan);
 773         spin_unlock_bh(&ioat_chan->desc_lock);
 774
 775         /* allocate a completion writeback area */
 776         /* doing 2 32bit writes to mmio since 1 64b write doesn't work */
 777         ioat_chan->completion_virt =
 778                 pci_pool_alloc(ioat_chan->device->completion_pool,
 779                                GFP_KERNEL,
 780                                &ioat_chan->completion_addr);
 781         memset(ioat_chan->completion_virt, 0,
 782                sizeof(*ioat_chan->completion_virt));
 783         writel(((u64) ioat_chan->completion_addr) & 0x00000000FFFFFFFF,
 784                ioat_chan->reg_base + IOAT_CHANCMP_OFFSET_LOW);
 785         writel(((u64) ioat_chan->completion_addr) >> 32,
 786                ioat_chan->reg_base + IOAT_CHANCMP_OFFSET_HIGH);
 787
 788         tasklet_enable(&ioat_chan->cleanup_task);
 789         ioat_dma_start_null_desc(ioat_chan);  /* give chain to dma device */
 790         return ioat_chan->desccount;
 791 }
 792
 793 /**
 794  * ioat_dma_free_chan_resources - release all the descriptors
 795  * @chan: the channel to be cleaned
 796  */
 797 static void ioat_dma_free_chan_resources(struct dma_chan *chan)
 798 {
 799         struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
 800         struct ioatdma_device *ioatdma_device = to_ioatdma_device(chan->device);
 801         struct ioat_desc_sw *desc, *_desc;
 802         int in_use_descs = 0;
 803
 804         /* Before freeing channel resources first check
 805          * if they have been previously allocated for this channel.
 806          */
 807         if (ioat_chan->desccount == 0)
 808                 return;
 809
 810         tasklet_disable(&ioat_chan->cleanup_task);
 811         ioat_dma_memcpy_cleanup(ioat_chan);
 812
 813         /* Delay 100ms after reset to allow internal DMA logic to quiesce
 814          * before removing DMA descriptor resources.
 815          */
 816         writeb(IOAT_CHANCMD_RESET,
 817                ioat_chan->reg_base
 818                         + IOAT_CHANCMD_OFFSET(ioat_chan->device->version));
 819         mdelay(100);
 820
 821         spin_lock_bh(&ioat_chan->desc_lock);
 822         switch (ioat_chan->device->version) {
 823         case IOAT_VER_1_2:
 824                 list_for_each_entry_safe(desc, _desc,
 825                                          &ioat_chan->used_desc, node) {
 826                         in_use_descs++;
 827                         list_del(&desc->node);
 828                         pci_pool_free(ioatdma_device->dma_pool, desc->hw,
 829                                       desc->async_tx.phys);
 830                         kfree(desc);
 831                 }
 832                 list_for_each_entry_safe(desc, _desc,
 833                                          &ioat_chan->free_desc, node) {
 834                         list_del(&desc->node);
 835                         pci_pool_free(ioatdma_device->dma_pool, desc->hw,
 836                                       desc->async_tx.phys);
 837                         kfree(desc);
 838                 }
 839                 break;
 840         case IOAT_VER_2_0:
 841         case IOAT_VER_3_0:
 842                 list_for_each_entry_safe(desc, _desc,
 843                                          ioat_chan->free_desc.next, node) {
 844                         list_del(&desc->node);
 845                         pci_pool_free(ioatdma_device->dma_pool, desc->hw,
 846                                       desc->async_tx.phys);
 847                         kfree(desc);
 848                 }
 849                 desc = to_ioat_desc(ioat_chan->free_desc.next);
 850                 pci_pool_free(ioatdma_device->dma_pool, desc->hw,
 851                               desc->async_tx.phys);
 852                 kfree(desc);
 853                 INIT_LIST_HEAD(&ioat_chan->free_desc);
 854                 INIT_LIST_HEAD(&ioat_chan->used_desc);
 855                 break;
 856         }
 857         spin_unlock_bh(&ioat_chan->desc_lock);
 858
 859         pci_pool_free(ioatdma_device->completion_pool,
 860                       ioat_chan->completion_virt,
 861                       ioat_chan->completion_addr);
 862
 863         /* one is ok since we left it on there on purpose */
 864         if (in_use_descs > 1)
 865                 dev_err(&ioat_chan->device->pdev->dev,
 866                         "Freeing %d in use descriptors!\n",
 867                         in_use_descs - 1);
 868
 869         ioat_chan->last_completion = ioat_chan->completion_addr = 0;
 870         ioat_chan->pending = 0;
 871         ioat_chan->dmacount = 0;
 872         ioat_chan->desccount = 0;
 873         ioat_chan->watchdog_completion = 0;
 874         ioat_chan->last_compl_desc_addr_hw = 0;
 875         ioat_chan->watchdog_tcp_cookie =
 876                 ioat_chan->watchdog_last_tcp_cookie = 0;
 877 }
 878
 879 /**
 880  * ioat_dma_get_next_descriptor - return the next available descriptor
 881  * @ioat_chan: IOAT DMA channel handle
 882  *
 883  * Gets the next descriptor from the chain, and must be called with the
 884  * channel's desc_lock held.  Allocates more descriptors if the channel
 885  * has run out.
 886  */
 887 static struct ioat_desc_sw *
 888 ioat1_dma_get_next_descriptor(struct ioat_dma_chan *ioat_chan)
 889 {
 890         struct ioat_desc_sw *new;
 891
 892         if (!list_empty(&ioat_chan->free_desc)) {
 893                 new = to_ioat_desc(ioat_chan->free_desc.next);
 894                 list_del(&new->node);
 895         } else {
 896                 /* try to get another desc */
 897                 new = ioat_dma_alloc_descriptor(ioat_chan, GFP_ATOMIC);
 898                 if (!new) {
 899                         dev_err(&ioat_chan->device->pdev->dev,
 900                                 "alloc failed\n");
 901                         return NULL;
 902                 }
 903         }
 904
 905         prefetch(new->hw);
 906         return new;
 907 }
 908
 909 static struct ioat_desc_sw *
 910 ioat2_dma_get_next_descriptor(struct ioat_dma_chan *ioat_chan)
 911 {
 912         struct ioat_desc_sw *new;
 913
 914         /*
 915          * used.prev points to where to start processing
 916          * used.next points to next free descriptor
 917          * if used.prev == NULL, there are none waiting to be processed
 918          * if used.next == used.prev.prev, there is only one free descriptor,
 919          *      and we need to use it to as a noop descriptor before
 920          *      linking in a new set of descriptors, since the device
 921          *      has probably already read the pointer to it
 922          */
 923         if (ioat_chan->used_desc.prev &&
 924             ioat_chan->used_desc.next == ioat_chan->used_desc.prev->prev) {
 925
 926                 struct ioat_desc_sw *desc;
 927                 struct ioat_desc_sw *noop_desc;
 928                 int i;
 929
 930                 /* set up the noop descriptor */
 931                 noop_desc = to_ioat_desc(ioat_chan->used_desc.next);
 932                 /* set size to non-zero value (channel returns error when size is 0) */
 933                 noop_desc->hw->size = NULL_DESC_BUFFER_SIZE;
 934                 noop_desc->hw->ctl = IOAT_DMA_DESCRIPTOR_NUL;
 935                 noop_desc->hw->src_addr = 0;
 936                 noop_desc->hw->dst_addr = 0;
 937
 938                 ioat_chan->used_desc.next = ioat_chan->used_desc.next->next;
 939                 ioat_chan->pending++;
 940                 ioat_chan->dmacount++;
 941
 942                 /* try to get a few more descriptors */
 943                 for (i = 16; i; i--) {
 944                         desc = ioat_dma_alloc_descriptor(ioat_chan, GFP_ATOMIC);
 945                         if (!desc) {
 946                                 dev_err(&ioat_chan->device->pdev->dev,
 947                                         "alloc failed\n");
 948                                 break;
 949                         }
 950                         list_add_tail(&desc->node, ioat_chan->used_desc.next);
 951
 952                         desc->hw->next
 953                                 = to_ioat_desc(desc->node.next)->async_tx.phys;
 954                         to_ioat_desc(desc->node.prev)->hw->next
 955                                 = desc->async_tx.phys;
 956                         ioat_chan->desccount++;
 957                 }
 958
 959                 ioat_chan->used_desc.next = noop_desc->node.next;
 960         }
 961         new = to_ioat_desc(ioat_chan->used_desc.next);
 962         prefetch(new);
 963         ioat_chan->used_desc.next = new->node.next;
 964
 965         if (ioat_chan->used_desc.prev == NULL)
 966                 ioat_chan->used_desc.prev = &new->node;
 967
 968         prefetch(new->hw);
 969         return new;
 970 }
 971
 972 static struct ioat_desc_sw *ioat_dma_get_next_descriptor(
 973                                                 struct ioat_dma_chan *ioat_chan)
 974 {
 975         if (!ioat_chan)
 976                 return NULL;
 977
 978         switch (ioat_chan->device->version) {
 979         case IOAT_VER_1_2:
 980                 return ioat1_dma_get_next_descriptor(ioat_chan);
 981                 break;
 982         case IOAT_VER_2_0:
 983         case IOAT_VER_3_0:
 984                 return ioat2_dma_get_next_descriptor(ioat_chan);
 985                 break;
 986         }
 987         return NULL;
 988 }
 989
 990 static struct dma_async_tx_descriptor *ioat1_dma_prep_memcpy(
 991                                                 struct dma_chan *chan,
 992                                                 dma_addr_t dma_dest,
 993                                                 dma_addr_t dma_src,
 994                                                 size_t len,
 995                                                 unsigned long flags)
 996 {
 997         struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
 998         struct ioat_desc_sw *new;
 999
1000         spin_lock_bh(&ioat_chan->desc_lock);
1001         new = ioat_dma_get_next_descriptor(ioat_chan);
1002         spin_unlock_bh(&ioat_chan->desc_lock);
1003
1004         if (new) {
1005                 new->len = len;
1006                 new->dst = dma_dest;
1007                 new->src = dma_src;
1008                 new->async_tx.flags = flags;
1009                 return &new->async_tx;
1010         } else {
1011                 dev_err(&ioat_chan->device->pdev->dev,
1012                         "chan%d - get_next_desc failed: %d descs waiting, %d total desc\n",
1013                         chan_num(ioat_chan), ioat_chan->dmacount, ioat_chan->desccount);
1014                 return NULL;
1015         }
1016 }
1017
1018 static struct dma_async_tx_descriptor *ioat2_dma_prep_memcpy(
1019                                                 struct dma_chan *chan,
1020                                                 dma_addr_t dma_dest,
1021                                                 dma_addr_t dma_src,
1022                                                 size_t len,
1023                                                 unsigned long flags)
1024 {
1025         struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
1026         struct ioat_desc_sw *new;
1027
1028         spin_lock_bh(&ioat_chan->desc_lock);
1029         new = ioat2_dma_get_next_descriptor(ioat_chan);
1030
1031         /*
1032          * leave ioat_chan->desc_lock set in ioat 2 path
1033          * it will get unlocked at end of tx_submit
1034          */
1035
1036         if (new) {
1037                 new->len = len;
1038                 new->dst = dma_dest;
1039                 new->src = dma_src;
1040                 new->async_tx.flags = flags;
1041                 return &new->async_tx;
1042         } else {
1043                 spin_unlock_bh(&ioat_chan->desc_lock);
1044                 dev_err(&ioat_chan->device->pdev->dev,
1045                         "chan%d - get_next_desc failed: %d descs waiting, %d total desc\n",
1046                         chan_num(ioat_chan), ioat_chan->dmacount, ioat_chan->desccount);
1047                 return NULL;
1048         }
1049 }
1050
1051 static void ioat_dma_cleanup_tasklet(unsigned long data)
1052 {
1053         struct ioat_dma_chan *chan = (void *)data;
1054         ioat_dma_memcpy_cleanup(chan);
1055         writew(IOAT_CHANCTRL_INT_DISABLE,
1056                chan->reg_base + IOAT_CHANCTRL_OFFSET);
1057 }
1058
1059 static void
1060 ioat_dma_unmap(struct ioat_dma_chan *ioat_chan, struct ioat_desc_sw *desc)
1061 {
1062         /*
1063          * yes we are unmapping both _page and _single
1064          * alloc'd regions with unmap_page. Is this
1065          * *really* that bad?
1066          */
1067         if (!(desc->async_tx.flags & DMA_COMPL_SKIP_DEST_UNMAP))
1068                 pci_unmap_page(ioat_chan->device->pdev,
1069                                 pci_unmap_addr(desc, dst),
1070                                 pci_unmap_len(desc, len),
1071                                 PCI_DMA_FROMDEVICE);
1072
1073         if (!(desc->async_tx.flags & DMA_COMPL_SKIP_SRC_UNMAP))
1074                 pci_unmap_page(ioat_chan->device->pdev,
1075                                 pci_unmap_addr(desc, src),
1076                                 pci_unmap_len(desc, len),
1077                                 PCI_DMA_TODEVICE);
1078 }
1079
1080 /**
1081  * ioat_dma_memcpy_cleanup - cleanup up finished descriptors
1082  * @chan: ioat channel to be cleaned up
1083  */
1084 static void ioat_dma_memcpy_cleanup(struct ioat_dma_chan *ioat_chan)
1085 {
1086         unsigned long phys_complete;
1087         struct ioat_desc_sw *desc, *_desc;
1088         dma_cookie_t cookie = 0;
1089         unsigned long desc_phys;
1090         struct ioat_desc_sw *latest_desc;
1091
1092         prefetch(ioat_chan->completion_virt);
1093
1094         if (!spin_trylock_bh(&ioat_chan->cleanup_lock))
1095                 return;
1096
1097         /* The completion writeback can happen at any time,
1098            so reads by the driver need to be atomic operations
1099            The descriptor physical addresses are limited to 32-bits
1100            when the CPU can only do a 32-bit mov */
1101
1102 #if (BITS_PER_LONG == 64)
1103         phys_complete =
1104                 ioat_chan->completion_virt->full
1105                 & IOAT_CHANSTS_COMPLETED_DESCRIPTOR_ADDR;
1106 #else
1107         phys_complete =
1108                 ioat_chan->completion_virt->low & IOAT_LOW_COMPLETION_MASK;
1109 #endif
1110
1111         if ((ioat_chan->completion_virt->full
1112                 & IOAT_CHANSTS_DMA_TRANSFER_STATUS) ==
1113                                 IOAT_CHANSTS_DMA_TRANSFER_STATUS_HALTED) {
1114                 dev_err(&ioat_chan->device->pdev->dev,
1115                         "Channel halted, chanerr = %x\n",
1116                         readl(ioat_chan->reg_base + IOAT_CHANERR_OFFSET));
1117
1118                 /* TODO do something to salvage the situation */
1119         }
1120
1121         if (phys_complete == ioat_chan->last_completion) {
1122                 spin_unlock_bh(&ioat_chan->cleanup_lock);
1123                 /*
1124                  * perhaps we're stuck so hard that the watchdog can't go off?
1125                  * try to catch it after 2 seconds
1126                  */
1127                 if (ioat_chan->device->version != IOAT_VER_3_0) {
1128                         if (time_after(jiffies,
1129                                        ioat_chan->last_completion_time + HZ*WATCHDOG_DELAY)) {
1130                                 ioat_dma_chan_watchdog(&(ioat_chan->device->work.work));
1131                                 ioat_chan->last_completion_time = jiffies;
1132                         }
1133                 }
1134                 return;
1135         }
1136         ioat_chan->last_completion_time = jiffies;
1137
1138         cookie = 0;
1139         if (!spin_trylock_bh(&ioat_chan->desc_lock)) {
1140                 spin_unlock_bh(&ioat_chan->cleanup_lock);
1141                 return;
1142         }
1143
1144         switch (ioat_chan->device->version) {
1145         case IOAT_VER_1_2:
1146                 list_for_each_entry_safe(desc, _desc,
1147                                          &ioat_chan->used_desc, node) {
1148
1149                         /*
1150                          * Incoming DMA requests may use multiple descriptors,
1151                          * due to exceeding xfercap, perhaps. If so, only the
1152                          * last one will have a cookie, and require unmapping.
1153                          */
1154                         if (desc->async_tx.cookie) {
1155                                 cookie = desc->async_tx.cookie;
1156                                 ioat_dma_unmap(ioat_chan, desc);
1157                                 if (desc->async_tx.callback) {
1158                                         desc->async_tx.callback(desc->async_tx.callback_param);
1159                                         desc->async_tx.callback = NULL;
1160                                 }
1161                         }
1162
1163                         if (desc->async_tx.phys != phys_complete) {
1164                                 /*
1165                                  * a completed entry, but not the last, so clean
1166                                  * up if the client is done with the descriptor
1167                                  */
1168                                 if (async_tx_test_ack(&desc->async_tx)) {
1169                                         list_del(&desc->node);
1170                                         list_add_tail(&desc->node,
1171                                                       &ioat_chan->free_desc);
1172                                 } else
1173                                         desc->async_tx.cookie = 0;
1174                         } else {
1175                                 /*
1176                                  * last used desc. Do not remove, so we can
1177                                  * append from it, but don't look at it next
1178                                  * time, either
1179                                  */
1180                                 desc->async_tx.cookie = 0;
1181
1182                                 /* TODO check status bits? */
1183                                 break;
1184                         }
1185                 }
1186                 break;
1187         case IOAT_VER_2_0:
1188         case IOAT_VER_3_0:
1189                 /* has some other thread has already cleaned up? */
1190                 if (ioat_chan->used_desc.prev == NULL)
1191                         break;
1192
1193                 /* work backwards to find latest finished desc */
1194                 desc = to_ioat_desc(ioat_chan->used_desc.next);
1195                 latest_desc = NULL;
1196                 do {
1197                         desc = to_ioat_desc(desc->node.prev);
1198                         desc_phys = (unsigned long)desc->async_tx.phys
1199                                        & IOAT_CHANSTS_COMPLETED_DESCRIPTOR_ADDR;
1200                         if (desc_phys == phys_complete) {
1201                                 latest_desc = desc;
1202                                 break;
1203                         }
1204                 } while (&desc->node != ioat_chan->used_desc.prev);
1205
1206                 if (latest_desc != NULL) {
1207
1208                         /* work forwards to clear finished descriptors */
1209                         for (desc = to_ioat_desc(ioat_chan->used_desc.prev);
1210                              &desc->node != latest_desc->node.next &&
1211                              &desc->node != ioat_chan->used_desc.next;
1212                              desc = to_ioat_desc(desc->node.next)) {
1213                                 if (desc->async_tx.cookie) {
1214                                         cookie = desc->async_tx.cookie;
1215                                         desc->async_tx.cookie = 0;
1216                                         ioat_dma_unmap(ioat_chan, desc);
1217                                         if (desc->async_tx.callback) {
1218                                                 desc->async_tx.callback(desc->async_tx.callback_param);
1219                                                 desc->async_tx.callback = NULL;
1220                                         }
1221                                 }
1222                         }
1223
1224                         /* move used.prev up beyond those that are finished */
1225                         if (&desc->node == ioat_chan->used_desc.next)
1226                                 ioat_chan->used_desc.prev = NULL;
1227                         else
1228                                 ioat_chan->used_desc.prev = &desc->node;
1229                 }
1230                 break;
1231         }
1232
1233         spin_unlock_bh(&ioat_chan->desc_lock);
1234
1235         ioat_chan->last_completion = phys_complete;
1236         if (cookie != 0)
1237                 ioat_chan->completed_cookie = cookie;
1238
1239         spin_unlock_bh(&ioat_chan->cleanup_lock);
1240 }
1241
1242 /**
1243  * ioat_dma_is_complete - poll the status of a IOAT DMA transaction
1244  * @chan: IOAT DMA channel handle
1245  * @cookie: DMA transaction identifier
1246  * @done: if not %NULL, updated with last completed transaction
1247  * @used: if not %NULL, updated with last used transaction
1248  */
1249 static enum dma_status ioat_dma_is_complete(struct dma_chan *chan,
1250                                             dma_cookie_t cookie,
1251                                             dma_cookie_t *done,
1252                                             dma_cookie_t *used)
1253 {
1254         struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
1255         dma_cookie_t last_used;
1256         dma_cookie_t last_complete;
1257         enum dma_status ret;
1258
1259         last_used = chan->cookie;
1260         last_complete = ioat_chan->completed_cookie;
1261         ioat_chan->watchdog_tcp_cookie = cookie;
1262
1263         if (done)
1264                 *done = last_complete;
1265         if (used)
1266                 *used = last_used;
1267
1268         ret = dma_async_is_complete(cookie, last_complete, last_used);
1269         if (ret == DMA_SUCCESS)
1270                 return ret;
1271
1272         ioat_dma_memcpy_cleanup(ioat_chan);
1273
1274         last_used = chan->cookie;
1275         last_complete = ioat_chan->completed_cookie;
1276
1277         if (done)
1278                 *done = last_complete;
1279         if (used)
1280                 *used = last_used;
1281
1282         return dma_async_is_complete(cookie, last_complete, last_used);
1283 }
1284
1285 static void ioat_dma_start_null_desc(struct ioat_dma_chan *ioat_chan)
1286 {
1287         struct ioat_desc_sw *desc;
1288
1289         spin_lock_bh(&ioat_chan->desc_lock);
1290
1291         desc = ioat_dma_get_next_descriptor(ioat_chan);
1292
1293         if (!desc) {
1294                 dev_err(&ioat_chan->device->pdev->dev,
1295                         "Unable to start null desc - get next desc failed\n");
1296                 spin_unlock_bh(&ioat_chan->desc_lock);
1297                 return;
1298         }
1299
1300         desc->hw->ctl = IOAT_DMA_DESCRIPTOR_NUL
1301                                 | IOAT_DMA_DESCRIPTOR_CTL_INT_GN
1302                                 | IOAT_DMA_DESCRIPTOR_CTL_CP_STS;
1303         /* set size to non-zero value (channel returns error when size is 0) */
1304         desc->hw->size = NULL_DESC_BUFFER_SIZE;
1305         desc->hw->src_addr = 0;
1306         desc->hw->dst_addr = 0;
1307         async_tx_ack(&desc->async_tx);
1308         switch (ioat_chan->device->version) {
1309         case IOAT_VER_1_2:
1310                 desc->hw->next = 0;
1311                 list_add_tail(&desc->node, &ioat_chan->used_desc);
1312
1313                 writel(((u64) desc->async_tx.phys) & 0x00000000FFFFFFFF,
1314                        ioat_chan->reg_base + IOAT1_CHAINADDR_OFFSET_LOW);
1315                 writel(((u64) desc->async_tx.phys) >> 32,
1316                        ioat_chan->reg_base + IOAT1_CHAINADDR_OFFSET_HIGH);
1317
1318                 writeb(IOAT_CHANCMD_START, ioat_chan->reg_base
1319                         + IOAT_CHANCMD_OFFSET(ioat_chan->device->version));
1320                 break;
1321         case IOAT_VER_2_0:
1322         case IOAT_VER_3_0:
1323                 writel(((u64) desc->async_tx.phys) & 0x00000000FFFFFFFF,
1324                        ioat_chan->reg_base + IOAT2_CHAINADDR_OFFSET_LOW);
1325                 writel(((u64) desc->async_tx.phys) >> 32,
1326                        ioat_chan->reg_base + IOAT2_CHAINADDR_OFFSET_HIGH);
1327
1328                 ioat_chan->dmacount++;
1329                 __ioat2_dma_memcpy_issue_pending(ioat_chan);
1330                 break;
1331         }
1332         spin_unlock_bh(&ioat_chan->desc_lock);
1333 }
1334
1335 /*
1336  * Perform a IOAT transaction to verify the HW works.
1337  */
1338 #define IOAT_TEST_SIZE 2000
1339
1340 static void ioat_dma_test_callback(void *dma_async_param)
1341 {
1342         printk(KERN_ERR "ioatdma: ioat_dma_test_callback(%p)\n",
1343                 dma_async_param);
1344 }
1345
1346 /**
1347  * ioat_dma_self_test - Perform a IOAT transaction to verify the HW works.
1348  * @device: device to be tested
1349  */
1350 static int ioat_dma_self_test(struct ioatdma_device *device)
1351 {
1352         int i;
1353         u8 *src;
1354         u8 *dest;
1355         struct dma_chan *dma_chan;
1356         struct dma_async_tx_descriptor *tx;
1357         dma_addr_t dma_dest, dma_src;
1358         dma_cookie_t cookie;
1359         int err = 0;
1360
1361         src = kzalloc(sizeof(u8) * IOAT_TEST_SIZE, GFP_KERNEL);
1362         if (!src)
1363                 return -ENOMEM;
1364         dest = kzalloc(sizeof(u8) * IOAT_TEST_SIZE, GFP_KERNEL);
1365         if (!dest) {
1366                 kfree(src);
1367                 return -ENOMEM;
1368         }
1369
1370         /* Fill in src buffer */
1371         for (i = 0; i < IOAT_TEST_SIZE; i++)
1372                 src[i] = (u8)i;
1373
1374         /* Start copy, using first DMA channel */
1375         dma_chan = container_of(device->common.channels.next,
1376                                 struct dma_chan,
1377                                 device_node);
1378         if (device->common.device_alloc_chan_resources(dma_chan, NULL) < 1) {
1379                 dev_err(&device->pdev->dev,
1380                         "selftest cannot allocate chan resource\n");
1381                 err = -ENODEV;
1382                 goto out;
1383         }
1384
1385         dma_src = dma_map_single(dma_chan->device->dev, src, IOAT_TEST_SIZE,
1386                                  DMA_TO_DEVICE);
1387         dma_dest = dma_map_single(dma_chan->device->dev, dest, IOAT_TEST_SIZE,
1388                                   DMA_FROM_DEVICE);
1389         tx = device->common.device_prep_dma_memcpy(dma_chan, dma_dest, dma_src,
1390                                                    IOAT_TEST_SIZE, 0);
1391         if (!tx) {
1392                 dev_err(&device->pdev->dev,
1393                         "Self-test prep failed, disabling\n");
1394                 err = -ENODEV;
1395                 goto free_resources;
1396         }
1397
1398         async_tx_ack(tx);
1399         tx->callback = ioat_dma_test_callback;
1400         tx->callback_param = (void *)0x8086;
1401         cookie = tx->tx_submit(tx);
1402         if (cookie < 0) {
1403                 dev_err(&device->pdev->dev,
1404                         "Self-test setup failed, disabling\n");
1405                 err = -ENODEV;
1406                 goto free_resources;
1407         }
1408         device->common.device_issue_pending(dma_chan);
1409         msleep(1);
1410
1411         if (device->common.device_is_tx_complete(dma_chan, cookie, NULL, NULL)
1412                                         != DMA_SUCCESS) {
1413                 dev_err(&device->pdev->dev,
1414                         "Self-test copy timed out, disabling\n");
1415                 err = -ENODEV;
1416                 goto free_resources;
1417         }
1418         if (memcmp(src, dest, IOAT_TEST_SIZE)) {
1419                 dev_err(&device->pdev->dev,
1420                         "Self-test copy failed compare, disabling\n");
1421                 err = -ENODEV;
1422                 goto free_resources;
1423         }
1424
1425 free_resources:
1426         device->common.device_free_chan_resources(dma_chan);
1427 out:
1428         kfree(src);
1429         kfree(dest);
1430         return err;
1431 }
1432
1433 static char ioat_interrupt_style[32] = "msix";
1434 module_param_string(ioat_interrupt_style, ioat_interrupt_style,
1435                     sizeof(ioat_interrupt_style), 0644);
1436 MODULE_PARM_DESC(ioat_interrupt_style,
1437                  "set ioat interrupt style: msix (default), "
1438                  "msix-single-vector, msi, intx)");
1439
1440 /**
1441  * ioat_dma_setup_interrupts - setup interrupt handler
1442  * @device: ioat device
1443  */
1444 static int ioat_dma_setup_interrupts(struct ioatdma_device *device)
1445 {
1446         struct ioat_dma_chan *ioat_chan;
1447         int err, i, j, msixcnt;
1448         u8 intrctrl = 0;
1449
1450         if (!strcmp(ioat_interrupt_style, "msix"))
1451                 goto msix;
1452         if (!strcmp(ioat_interrupt_style, "msix-single-vector"))
1453                 goto msix_single_vector;
1454         if (!strcmp(ioat_interrupt_style, "msi"))
1455                 goto msi;
1456         if (!strcmp(ioat_interrupt_style, "intx"))
1457                 goto intx;
1458         dev_err(&device->pdev->dev, "invalid ioat_interrupt_style %s\n",
1459                 ioat_interrupt_style);
1460         goto err_no_irq;
1461
1462 msix:
1463         /* The number of MSI-X vectors should equal the number of channels */
1464         msixcnt = device->common.chancnt;
1465         for (i = 0; i < msixcnt; i++)
1466                 device->msix_entries[i].entry = i;
1467
1468         err = pci_enable_msix(device->pdev, device->msix_entries, msixcnt);
1469         if (err < 0)
1470                 goto msi;
1471         if (err > 0)
1472                 goto msix_single_vector;
1473
1474         for (i = 0; i < msixcnt; i++) {
1475                 ioat_chan = ioat_lookup_chan_by_index(device, i);
1476                 err = request_irq(device->msix_entries[i].vector,
1477                                   ioat_dma_do_interrupt_msix,
1478                                   0, "ioat-msix", ioat_chan);
1479                 if (err) {
1480                         for (j = 0; j < i; j++) {
1481                                 ioat_chan =
1482                                         ioat_lookup_chan_by_index(device, j);
1483                                 free_irq(device->msix_entries[j].vector,
1484                                          ioat_chan);
1485                         }
1486                         goto msix_single_vector;
1487                 }
1488         }
1489         intrctrl |= IOAT_INTRCTRL_MSIX_VECTOR_CONTROL;
1490         device->irq_mode = msix_multi_vector;
1491         goto done;
1492
1493 msix_single_vector:
1494         device->msix_entries[0].entry = 0;
1495         err = pci_enable_msix(device->pdev, device->msix_entries, 1);
1496         if (err)
1497                 goto msi;
1498
1499         err = request_irq(device->msix_entries[0].vector, ioat_dma_do_interrupt,
1500                           0, "ioat-msix", device);
1501         if (err) {
1502                 pci_disable_msix(device->pdev);
1503                 goto msi;
1504         }
1505         device->irq_mode = msix_single_vector;
1506         goto done;
1507
1508 msi:
1509         err = pci_enable_msi(device->pdev);
1510         if (err)
1511                 goto intx;
1512
1513         err = request_irq(device->pdev->irq, ioat_dma_do_interrupt,
1514                           0, "ioat-msi", device);
1515         if (err) {
1516                 pci_disable_msi(device->pdev);
1517                 goto intx;
1518         }
1519         /*
1520          * CB 1.2 devices need a bit set in configuration space to enable MSI
1521          */
1522         if (device->version == IOAT_VER_1_2) {
1523                 u32 dmactrl;
1524                 pci_read_config_dword(device->pdev,
1525                                       IOAT_PCI_DMACTRL_OFFSET, &dmactrl);
1526                 dmactrl |= IOAT_PCI_DMACTRL_MSI_EN;
1527                 pci_write_config_dword(device->pdev,
1528                                        IOAT_PCI_DMACTRL_OFFSET, dmactrl);
1529         }
1530         device->irq_mode = msi;
1531         goto done;
1532
1533 intx:
1534         err = request_irq(device->pdev->irq, ioat_dma_do_interrupt,
1535                           IRQF_SHARED, "ioat-intx", device);
1536         if (err)
1537                 goto err_no_irq;
1538         device->irq_mode = intx;
1539
1540 done:
1541         intrctrl |= IOAT_INTRCTRL_MASTER_INT_EN;
1542         writeb(intrctrl, device->reg_base + IOAT_INTRCTRL_OFFSET);
1543         return 0;
1544
1545 err_no_irq:
1546         /* Disable all interrupt generation */
1547         writeb(0, device->reg_base + IOAT_INTRCTRL_OFFSET);
1548         dev_err(&device->pdev->dev, "no usable interrupts\n");
1549         device->irq_mode = none;
1550         return -1;
1551 }
1552
1553 /**
1554  * ioat_dma_remove_interrupts - remove whatever interrupts were set
1555  * @device: ioat device
1556  */
1557 static void ioat_dma_remove_interrupts(struct ioatdma_device *device)
1558 {
1559         struct ioat_dma_chan *ioat_chan;
1560         int i;
1561
1562         /* Disable all interrupt generation */
1563         writeb(0, device->reg_base + IOAT_INTRCTRL_OFFSET);
1564
1565         switch (device->irq_mode) {
1566         case msix_multi_vector:
1567                 for (i = 0; i < device->common.chancnt; i++) {
1568                         ioat_chan = ioat_lookup_chan_by_index(device, i);
1569                         free_irq(device->msix_entries[i].vector, ioat_chan);
1570                 }
1571                 pci_disable_msix(device->pdev);
1572                 break;
1573         case msix_single_vector:
1574                 free_irq(device->msix_entries[0].vector, device);
1575                 pci_disable_msix(device->pdev);
1576                 break;
1577         case msi:
1578                 free_irq(device->pdev->irq, device);
1579                 pci_disable_msi(device->pdev);
1580                 break;
1581         case intx:
1582                 free_irq(device->pdev->irq, device);
1583                 break;
1584         case none:
1585                 dev_warn(&device->pdev->dev,
1586                          "call to %s without interrupts setup\n", __func__);
1587         }
1588         device->irq_mode = none;
1589 }
1590
1591 struct ioatdma_device *ioat_dma_probe(struct pci_dev *pdev,
1592                                       void __iomem *iobase)
1593 {
1594         int err;
1595         struct ioatdma_device *device;
1596
1597         device = kzalloc(sizeof(*device), GFP_KERNEL);
1598         if (!device) {
1599                 err = -ENOMEM;
1600                 goto err_kzalloc;
1601         }
1602         device->pdev = pdev;
1603         device->reg_base = iobase;
1604         device->version = readb(device->reg_base + IOAT_VER_OFFSET);
1605
1606         /* DMA coherent memory pool for DMA descriptor allocations */
1607         device->dma_pool = pci_pool_create("dma_desc_pool", pdev,
1608                                            sizeof(struct ioat_dma_descriptor),
1609                                            64, 0);
1610         if (!device->dma_pool) {
1611                 err = -ENOMEM;
1612                 goto err_dma_pool;
1613         }
1614
1615         device->completion_pool = pci_pool_create("completion_pool", pdev,
1616                                                   sizeof(u64), SMP_CACHE_BYTES,
1617                                                   SMP_CACHE_BYTES);
1618         if (!device->completion_pool) {
1619                 err = -ENOMEM;
1620                 goto err_completion_pool;
1621         }
1622
1623         INIT_LIST_HEAD(&device->common.channels);
1624         ioat_dma_enumerate_channels(device);
1625
1626         device->common.device_alloc_chan_resources =
1627                                                 ioat_dma_alloc_chan_resources;
1628         device->common.device_free_chan_resources =
1629                                                 ioat_dma_free_chan_resources;
1630         device->common.dev = &pdev->dev;
1631
1632         dma_cap_set(DMA_MEMCPY, device->common.cap_mask);
1633         device->common.device_is_tx_complete = ioat_dma_is_complete;
1634         switch (device->version) {
1635         case IOAT_VER_1_2:
1636                 device->common.device_prep_dma_memcpy = ioat1_dma_prep_memcpy;
1637                 device->common.device_issue_pending =
1638                                                 ioat1_dma_memcpy_issue_pending;
1639                 break;
1640         case IOAT_VER_2_0:
1641         case IOAT_VER_3_0:
1642                 device->common.device_prep_dma_memcpy = ioat2_dma_prep_memcpy;
1643                 device->common.device_issue_pending =
1644                                                 ioat2_dma_memcpy_issue_pending;
1645                 break;
1646         }
1647
1648         dev_err(&device->pdev->dev,
1649                 "Intel(R) I/OAT DMA Engine found,"
1650                 " %d channels, device version 0x%02x, driver version %s\n",
1651                 device->common.chancnt, device->version, IOAT_DMA_VERSION);
1652
1653         err = ioat_dma_setup_interrupts(device);
1654         if (err)
1655                 goto err_setup_interrupts;
1656
1657         err = ioat_dma_self_test(device);
1658         if (err)
1659                 goto err_self_test;
1660
1661         ioat_set_tcp_copy_break(device);
1662
1663         dma_async_device_register(&device->common);
1664
1665         if (device->version != IOAT_VER_3_0) {
1666                 INIT_DELAYED_WORK(&device->work, ioat_dma_chan_watchdog);
1667                 schedule_delayed_work(&device->work,
1668                                       WATCHDOG_DELAY);
1669         }
1670
1671         return device;
1672
1673 err_self_test:
1674         ioat_dma_remove_interrupts(device);
1675 err_setup_interrupts:
1676         pci_pool_destroy(device->completion_pool);
1677 err_completion_pool:
1678         pci_pool_destroy(device->dma_pool);
1679 err_dma_pool:
1680         kfree(device);
1681 err_kzalloc:
1682         dev_err(&pdev->dev,
1683                 "Intel(R) I/OAT DMA Engine initialization failed\n");
1684         return NULL;
1685 }
1686
1687 void ioat_dma_remove(struct ioatdma_device *device)
1688 {
1689         struct dma_chan *chan, *_chan;
1690         struct ioat_dma_chan *ioat_chan;
1691
1692         ioat_dma_remove_interrupts(device);
1693
1694         dma_async_device_unregister(&device->common);
1695
1696         pci_pool_destroy(device->dma_pool);
1697         pci_pool_destroy(device->completion_pool);
1698
1699         iounmap(device->reg_base);
1700         pci_release_regions(device->pdev);
1701         pci_disable_device(device->pdev);
1702
1703         if (device->version != IOAT_VER_3_0) {
1704                 cancel_delayed_work(&device->work);
1705         }
1706
1707         list_for_each_entry_safe(chan, _chan,
1708                                  &device->common.channels, device_node) {
1709                 ioat_chan = to_ioat_chan(chan);
1710                 list_del(&chan->device_node);
1711                 kfree(ioat_chan);
1712         }
1713         kfree(device);
1714 }
1715