drivers/dma/ioat/dma.c

   1 /*
   2  * Intel I/OAT DMA Linux driver
   3  * Copyright(c) 2004 - 2015 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  * The full GNU General Public License is included in this distribution in
  15  * the file called "COPYING".
  16  *
  17  */
  18
  19 /*
  20  * This driver supports an Intel I/OAT DMA engine, which does asynchronous
  21  * copy operations.
  22  */
  23
  24 #include <linux/init.h>
  25 #include <linux/module.h>
  26 #include <linux/slab.h>
  27 #include <linux/pci.h>
  28 #include <linux/interrupt.h>
  29 #include <linux/dmaengine.h>
  30 #include <linux/delay.h>
  31 #include <linux/dma-mapping.h>
  32 #include <linux/workqueue.h>
  33 #include <linux/prefetch.h>
  34 #include <linux/i7300_idle.h>
  35 #include "dma.h"
  36 #include "registers.h"
  37 #include "hw.h"
  38
  39 #include "../dmaengine.h"
  40
  41 int ioat_pending_level = 4;
  42 module_param(ioat_pending_level, int, 0644);
  43 MODULE_PARM_DESC(ioat_pending_level,
  44                  "high-water mark for pushing ioat descriptors (default: 4)");
  45
  46 /**
  47  * ioat_dma_do_interrupt - handler used for single vector interrupt mode
  48  * @irq: interrupt id
  49  * @data: interrupt data
  50  */
  51 static irqreturn_t ioat_dma_do_interrupt(int irq, void *data)
  52 {
  53         struct ioatdma_device *instance = data;
  54         struct ioat_chan_common *chan;
  55         unsigned long attnstatus;
  56         int bit;
  57         u8 intrctrl;
  58
  59         intrctrl = readb(instance->reg_base + IOAT_INTRCTRL_OFFSET);
  60
  61         if (!(intrctrl & IOAT_INTRCTRL_MASTER_INT_EN))
  62                 return IRQ_NONE;
  63
  64         if (!(intrctrl & IOAT_INTRCTRL_INT_STATUS)) {
  65                 writeb(intrctrl, instance->reg_base + IOAT_INTRCTRL_OFFSET);
  66                 return IRQ_NONE;
  67         }
  68
  69         attnstatus = readl(instance->reg_base + IOAT_ATTNSTATUS_OFFSET);
  70         for_each_set_bit(bit, &attnstatus, BITS_PER_LONG) {
  71                 chan = ioat_chan_by_index(instance, bit);
  72                 if (test_bit(IOAT_RUN, &chan->state))
  73                         tasklet_schedule(&chan->cleanup_task);
  74         }
  75
  76         writeb(intrctrl, instance->reg_base + IOAT_INTRCTRL_OFFSET);
  77         return IRQ_HANDLED;
  78 }
  79
  80 /**
  81  * ioat_dma_do_interrupt_msix - handler used for vector-per-channel interrupt mode
  82  * @irq: interrupt id
  83  * @data: interrupt data
  84  */
  85 static irqreturn_t ioat_dma_do_interrupt_msix(int irq, void *data)
  86 {
  87         struct ioat_chan_common *chan = data;
  88
  89         if (test_bit(IOAT_RUN, &chan->state))
  90                 tasklet_schedule(&chan->cleanup_task);
  91
  92         return IRQ_HANDLED;
  93 }
  94
  95 /* common channel initialization */
  96 void ioat_init_channel(struct ioatdma_device *device, struct ioat_chan_common *chan, int idx)
  97 {
  98         struct dma_device *dma = &device->common;
  99         struct dma_chan *c = &chan->common;
 100         unsigned long data = (unsigned long) c;
 101
 102         chan->device = device;
 103         chan->reg_base = device->reg_base + (0x80 * (idx + 1));
 104         spin_lock_init(&chan->cleanup_lock);
 105         chan->common.device = dma;
 106         dma_cookie_init(&chan->common);
 107         list_add_tail(&chan->common.device_node, &dma->channels);
 108         device->idx[idx] = chan;
 109         init_timer(&chan->timer);
 110         chan->timer.function = device->timer_fn;
 111         chan->timer.data = data;
 112         tasklet_init(&chan->cleanup_task, device->cleanup_fn, data);
 113 }
 114
 115 void ioat_stop(struct ioat_chan_common *chan)
 116 {
 117         struct ioatdma_device *device = chan->device;
 118         struct pci_dev *pdev = device->pdev;
 119         int chan_id = chan_num(chan);
 120         struct msix_entry *msix;
 121
 122         /* 1/ stop irq from firing tasklets
 123          * 2/ stop the tasklet from re-arming irqs
 124          */
 125         clear_bit(IOAT_RUN, &chan->state);
 126
 127         /* flush inflight interrupts */
 128         switch (device->irq_mode) {
 129         case IOAT_MSIX:
 130                 msix = &device->msix_entries[chan_id];
 131                 synchronize_irq(msix->vector);
 132                 break;
 133         case IOAT_MSI:
 134         case IOAT_INTX:
 135                 synchronize_irq(pdev->irq);
 136                 break;
 137         default:
 138                 break;
 139         }
 140
 141         /* flush inflight timers */
 142         del_timer_sync(&chan->timer);
 143
 144         /* flush inflight tasklet runs */
 145         tasklet_kill(&chan->cleanup_task);
 146
 147         /* final cleanup now that everything is quiesced and can't re-arm */
 148         device->cleanup_fn((unsigned long) &chan->common);
 149 }
 150
 151 dma_addr_t ioat_get_current_completion(struct ioat_chan_common *chan)
 152 {
 153         dma_addr_t phys_complete;
 154         u64 completion;
 155
 156         completion = *chan->completion;
 157         phys_complete = ioat_chansts_to_addr(completion);
 158
 159         dev_dbg(to_dev(chan), "%s: phys_complete: %#llx\n", __func__,
 160                 (unsigned long long) phys_complete);
 161
 162         if (is_ioat_halted(completion)) {
 163                 u32 chanerr = readl(chan->reg_base + IOAT_CHANERR_OFFSET);
 164                 dev_err(to_dev(chan), "Channel halted, chanerr = %x\n",
 165                         chanerr);
 166
 167                 /* TODO do something to salvage the situation */
 168         }
 169
 170         return phys_complete;
 171 }
 172
 173 bool ioat_cleanup_preamble(struct ioat_chan_common *chan,
 174                            dma_addr_t *phys_complete)
 175 {
 176         *phys_complete = ioat_get_current_completion(chan);
 177         if (*phys_complete == chan->last_completion)
 178                 return false;
 179         clear_bit(IOAT_COMPLETION_ACK, &chan->state);
 180         mod_timer(&chan->timer, jiffies + COMPLETION_TIMEOUT);
 181
 182         return true;
 183 }
 184
 185 enum dma_status
 186 ioat_dma_tx_status(struct dma_chan *c, dma_cookie_t cookie,
 187                    struct dma_tx_state *txstate)
 188 {
 189         struct ioat_chan_common *chan = to_chan_common(c);
 190         struct ioatdma_device *device = chan->device;
 191         enum dma_status ret;
 192
 193         ret = dma_cookie_status(c, cookie, txstate);
 194         if (ret == DMA_COMPLETE)
 195                 return ret;
 196
 197         device->cleanup_fn((unsigned long) c);
 198
 199         return dma_cookie_status(c, cookie, txstate);
 200 }
 201
 202 /*
 203  * Perform a IOAT transaction to verify the HW works.
 204  */
 205 #define IOAT_TEST_SIZE 2000
 206
 207 static void ioat_dma_test_callback(void *dma_async_param)
 208 {
 209         struct completion *cmp = dma_async_param;
 210
 211         complete(cmp);
 212 }
 213
 214 /**
 215  * ioat_dma_self_test - Perform a IOAT transaction to verify the HW works.
 216  * @device: device to be tested
 217  */
 218 int ioat_dma_self_test(struct ioatdma_device *device)
 219 {
 220         int i;
 221         u8 *src;
 222         u8 *dest;
 223         struct dma_device *dma = &device->common;
 224         struct device *dev = &device->pdev->dev;
 225         struct dma_chan *dma_chan;
 226         struct dma_async_tx_descriptor *tx;
 227         dma_addr_t dma_dest, dma_src;
 228         dma_cookie_t cookie;
 229         int err = 0;
 230         struct completion cmp;
 231         unsigned long tmo;
 232         unsigned long flags;
 233
 234         src = kzalloc(sizeof(u8) * IOAT_TEST_SIZE, GFP_KERNEL);
 235         if (!src)
 236                 return -ENOMEM;
 237         dest = kzalloc(sizeof(u8) * IOAT_TEST_SIZE, GFP_KERNEL);
 238         if (!dest) {
 239                 kfree(src);
 240                 return -ENOMEM;
 241         }
 242
 243         /* Fill in src buffer */
 244         for (i = 0; i < IOAT_TEST_SIZE; i++)
 245                 src[i] = (u8)i;
 246
 247         /* Start copy, using first DMA channel */
 248         dma_chan = container_of(dma->channels.next, struct dma_chan,
 249                                 device_node);
 250         if (dma->device_alloc_chan_resources(dma_chan) < 1) {
 251                 dev_err(dev, "selftest cannot allocate chan resource\n");
 252                 err = -ENODEV;
 253                 goto out;
 254         }
 255
 256         dma_src = dma_map_single(dev, src, IOAT_TEST_SIZE, DMA_TO_DEVICE);
 257         if (dma_mapping_error(dev, dma_src)) {
 258                 dev_err(dev, "mapping src buffer failed\n");
 259                 goto free_resources;
 260         }
 261         dma_dest = dma_map_single(dev, dest, IOAT_TEST_SIZE, DMA_FROM_DEVICE);
 262         if (dma_mapping_error(dev, dma_dest)) {
 263                 dev_err(dev, "mapping dest buffer failed\n");
 264                 goto unmap_src;
 265         }
 266         flags = DMA_PREP_INTERRUPT;
 267         tx = device->common.device_prep_dma_memcpy(dma_chan, dma_dest, dma_src,
 268                                                    IOAT_TEST_SIZE, flags);
 269         if (!tx) {
 270                 dev_err(dev, "Self-test prep failed, disabling\n");
 271                 err = -ENODEV;
 272                 goto unmap_dma;
 273         }
 274
 275         async_tx_ack(tx);
 276         init_completion(&cmp);
 277         tx->callback = ioat_dma_test_callback;
 278         tx->callback_param = &cmp;
 279         cookie = tx->tx_submit(tx);
 280         if (cookie < 0) {
 281                 dev_err(dev, "Self-test setup failed, disabling\n");
 282                 err = -ENODEV;
 283                 goto unmap_dma;
 284         }
 285         dma->device_issue_pending(dma_chan);
 286
 287         tmo = wait_for_completion_timeout(&cmp, msecs_to_jiffies(3000));
 288
 289         if (tmo == 0 ||
 290             dma->device_tx_status(dma_chan, cookie, NULL)
 291                                         != DMA_COMPLETE) {
 292                 dev_err(dev, "Self-test copy timed out, disabling\n");
 293                 err = -ENODEV;
 294                 goto unmap_dma;
 295         }
 296         if (memcmp(src, dest, IOAT_TEST_SIZE)) {
 297                 dev_err(dev, "Self-test copy failed compare, disabling\n");
 298                 err = -ENODEV;
 299                 goto free_resources;
 300         }
 301
 302 unmap_dma:
 303         dma_unmap_single(dev, dma_dest, IOAT_TEST_SIZE, DMA_FROM_DEVICE);
 304 unmap_src:
 305         dma_unmap_single(dev, dma_src, IOAT_TEST_SIZE, DMA_TO_DEVICE);
 306 free_resources:
 307         dma->device_free_chan_resources(dma_chan);
 308 out:
 309         kfree(src);
 310         kfree(dest);
 311         return err;
 312 }
 313
 314 static char ioat_interrupt_style[32] = "msix";
 315 module_param_string(ioat_interrupt_style, ioat_interrupt_style,
 316                     sizeof(ioat_interrupt_style), 0644);
 317 MODULE_PARM_DESC(ioat_interrupt_style,
 318                  "set ioat interrupt style: msix (default), msi, intx");
 319
 320 /**
 321  * ioat_dma_setup_interrupts - setup interrupt handler
 322  * @device: ioat device
 323  */
 324 int ioat_dma_setup_interrupts(struct ioatdma_device *device)
 325 {
 326         struct ioat_chan_common *chan;
 327         struct pci_dev *pdev = device->pdev;
 328         struct device *dev = &pdev->dev;
 329         struct msix_entry *msix;
 330         int i, j, msixcnt;
 331         int err = -EINVAL;
 332         u8 intrctrl = 0;
 333
 334         if (!strcmp(ioat_interrupt_style, "msix"))
 335                 goto msix;
 336         if (!strcmp(ioat_interrupt_style, "msi"))
 337                 goto msi;
 338         if (!strcmp(ioat_interrupt_style, "intx"))
 339                 goto intx;
 340         dev_err(dev, "invalid ioat_interrupt_style %s\n", ioat_interrupt_style);
 341         goto err_no_irq;
 342
 343 msix:
 344         /* The number of MSI-X vectors should equal the number of channels */
 345         msixcnt = device->common.chancnt;
 346         for (i = 0; i < msixcnt; i++)
 347                 device->msix_entries[i].entry = i;
 348
 349         err = pci_enable_msix_exact(pdev, device->msix_entries, msixcnt);
 350         if (err)
 351                 goto msi;
 352
 353         for (i = 0; i < msixcnt; i++) {
 354                 msix = &device->msix_entries[i];
 355                 chan = ioat_chan_by_index(device, i);
 356                 err = devm_request_irq(dev, msix->vector,
 357                                        ioat_dma_do_interrupt_msix, 0,
 358                                        "ioat-msix", chan);
 359                 if (err) {
 360                         for (j = 0; j < i; j++) {
 361                                 msix = &device->msix_entries[j];
 362                                 chan = ioat_chan_by_index(device, j);
 363                                 devm_free_irq(dev, msix->vector, chan);
 364                         }
 365                         goto msi;
 366                 }
 367         }
 368         intrctrl |= IOAT_INTRCTRL_MSIX_VECTOR_CONTROL;
 369         device->irq_mode = IOAT_MSIX;
 370         goto done;
 371
 372 msi:
 373         err = pci_enable_msi(pdev);
 374         if (err)
 375                 goto intx;
 376
 377         err = devm_request_irq(dev, pdev->irq, ioat_dma_do_interrupt, 0,
 378                                "ioat-msi", device);
 379         if (err) {
 380                 pci_disable_msi(pdev);
 381                 goto intx;
 382         }
 383         device->irq_mode = IOAT_MSI;
 384         goto done;
 385
 386 intx:
 387         err = devm_request_irq(dev, pdev->irq, ioat_dma_do_interrupt,
 388                                IRQF_SHARED, "ioat-intx", device);
 389         if (err)
 390                 goto err_no_irq;
 391
 392         device->irq_mode = IOAT_INTX;
 393 done:
 394         if (device->intr_quirk)
 395                 device->intr_quirk(device);
 396         intrctrl |= IOAT_INTRCTRL_MASTER_INT_EN;
 397         writeb(intrctrl, device->reg_base + IOAT_INTRCTRL_OFFSET);
 398         return 0;
 399
 400 err_no_irq:
 401         /* Disable all interrupt generation */
 402         writeb(0, device->reg_base + IOAT_INTRCTRL_OFFSET);
 403         device->irq_mode = IOAT_NOIRQ;
 404         dev_err(dev, "no usable interrupts\n");
 405         return err;
 406 }
 407 EXPORT_SYMBOL(ioat_dma_setup_interrupts);
 408
 409 static void ioat_disable_interrupts(struct ioatdma_device *device)
 410 {
 411         /* Disable all interrupt generation */
 412         writeb(0, device->reg_base + IOAT_INTRCTRL_OFFSET);
 413 }
 414
 415 int ioat_probe(struct ioatdma_device *device)
 416 {
 417         int err = -ENODEV;
 418         struct dma_device *dma = &device->common;
 419         struct pci_dev *pdev = device->pdev;
 420         struct device *dev = &pdev->dev;
 421
 422         /* DMA coherent memory pool for DMA descriptor allocations */
 423         device->dma_pool = pci_pool_create("dma_desc_pool", pdev,
 424                                            sizeof(struct ioat_dma_descriptor),
 425                                            64, 0);
 426         if (!device->dma_pool) {
 427                 err = -ENOMEM;
 428                 goto err_dma_pool;
 429         }
 430
 431         device->completion_pool = pci_pool_create("completion_pool", pdev,
 432                                                   sizeof(u64), SMP_CACHE_BYTES,
 433                                                   SMP_CACHE_BYTES);
 434
 435         if (!device->completion_pool) {
 436                 err = -ENOMEM;
 437                 goto err_completion_pool;
 438         }
 439
 440         device->enumerate_channels(device);
 441
 442         dma_cap_set(DMA_MEMCPY, dma->cap_mask);
 443         dma->dev = &pdev->dev;
 444
 445         if (!dma->chancnt) {
 446                 dev_err(dev, "channel enumeration error\n");
 447                 goto err_setup_interrupts;
 448         }
 449
 450         err = ioat_dma_setup_interrupts(device);
 451         if (err)
 452                 goto err_setup_interrupts;
 453
 454         err = device->self_test(device);
 455         if (err)
 456                 goto err_self_test;
 457
 458         return 0;
 459
 460 err_self_test:
 461         ioat_disable_interrupts(device);
 462 err_setup_interrupts:
 463         pci_pool_destroy(device->completion_pool);
 464 err_completion_pool:
 465         pci_pool_destroy(device->dma_pool);
 466 err_dma_pool:
 467         return err;
 468 }
 469
 470 int ioat_register(struct ioatdma_device *device)
 471 {
 472         int err = dma_async_device_register(&device->common);
 473
 474         if (err) {
 475                 ioat_disable_interrupts(device);
 476                 pci_pool_destroy(device->completion_pool);
 477                 pci_pool_destroy(device->dma_pool);
 478         }
 479
 480         return err;
 481 }
 482
 483 static ssize_t cap_show(struct dma_chan *c, char *page)
 484 {
 485         struct dma_device *dma = c->device;
 486
 487         return sprintf(page, "copy%s%s%s%s%s\n",
 488                        dma_has_cap(DMA_PQ, dma->cap_mask) ? " pq" : "",
 489                        dma_has_cap(DMA_PQ_VAL, dma->cap_mask) ? " pq_val" : "",
 490                        dma_has_cap(DMA_XOR, dma->cap_mask) ? " xor" : "",
 491                        dma_has_cap(DMA_XOR_VAL, dma->cap_mask) ? " xor_val" : "",
 492                        dma_has_cap(DMA_INTERRUPT, dma->cap_mask) ? " intr" : "");
 493
 494 }
 495 struct ioat_sysfs_entry ioat_cap_attr = __ATTR_RO(cap);
 496
 497 static ssize_t version_show(struct dma_chan *c, char *page)
 498 {
 499         struct dma_device *dma = c->device;
 500         struct ioatdma_device *device = to_ioatdma_device(dma);
 501
 502         return sprintf(page, "%d.%d\n",
 503                        device->version >> 4, device->version & 0xf);
 504 }
 505 struct ioat_sysfs_entry ioat_version_attr = __ATTR_RO(version);
 506
 507 static ssize_t
 508 ioat_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
 509 {
 510         struct ioat_sysfs_entry *entry;
 511         struct ioat_chan_common *chan;
 512
 513         entry = container_of(attr, struct ioat_sysfs_entry, attr);
 514         chan = container_of(kobj, struct ioat_chan_common, kobj);
 515
 516         if (!entry->show)
 517                 return -EIO;
 518         return entry->show(&chan->common, page);
 519 }
 520
 521 const struct sysfs_ops ioat_sysfs_ops = {
 522         .show   = ioat_attr_show,
 523 };
 524
 525 void ioat_kobject_add(struct ioatdma_device *device, struct kobj_type *type)
 526 {
 527         struct dma_device *dma = &device->common;
 528         struct dma_chan *c;
 529
 530         list_for_each_entry(c, &dma->channels, device_node) {
 531                 struct ioat_chan_common *chan = to_chan_common(c);
 532                 struct kobject *parent = &c->dev->device.kobj;
 533                 int err;
 534
 535                 err = kobject_init_and_add(&chan->kobj, type, parent, "quickdata");
 536                 if (err) {
 537                         dev_warn(to_dev(chan),
 538                                  "sysfs init error (%d), continuing...\n", err);
 539                         kobject_put(&chan->kobj);
 540                         set_bit(IOAT_KOBJ_INIT_FAIL, &chan->state);
 541                 }
 542         }
 543 }
 544
 545 void ioat_kobject_del(struct ioatdma_device *device)
 546 {
 547         struct dma_device *dma = &device->common;
 548         struct dma_chan *c;
 549
 550         list_for_each_entry(c, &dma->channels, device_node) {
 551                 struct ioat_chan_common *chan = to_chan_common(c);
 552
 553                 if (!test_bit(IOAT_KOBJ_INIT_FAIL, &chan->state)) {
 554                         kobject_del(&chan->kobj);
 555                         kobject_put(&chan->kobj);
 556                 }
 557         }
 558 }
 559
 560 void ioat_dma_remove(struct ioatdma_device *device)
 561 {
 562         struct dma_device *dma = &device->common;
 563
 564         ioat_disable_interrupts(device);
 565
 566         ioat_kobject_del(device);
 567
 568         dma_async_device_unregister(dma);
 569
 570         pci_pool_destroy(device->dma_pool);
 571         pci_pool_destroy(device->completion_pool);
 572
 573         INIT_LIST_HEAD(&dma->channels);
 574 }