drivers/dma/ioat/dma_v2.c

   1 /*
   2  * Intel I/OAT DMA Linux driver
   3  * Copyright(c) 2004 - 2009 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 (versions >= 2), which
  25  * does asynchronous data movement and checksumming 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 <linux/i7300_idle.h>
  37 #include "dma.h"
  38 #include "dma_v2.h"
  39 #include "registers.h"
  40 #include "hw.h"
  41
  42 static int ioat_ring_alloc_order = 8;
  43 module_param(ioat_ring_alloc_order, int, 0644);
  44 MODULE_PARM_DESC(ioat_ring_alloc_order,
  45                  "ioat2+: allocate 2^n descriptors per channel (default: n=8)");
  46
  47 static void __ioat2_issue_pending(struct ioat2_dma_chan *ioat)
  48 {
  49         void * __iomem reg_base = ioat->base.reg_base;
  50
  51         ioat->pending = 0;
  52         ioat->dmacount += ioat2_ring_pending(ioat);
  53         ioat->issued = ioat->head;
  54         /* make descriptor updates globally visible before notifying channel */
  55         wmb();
  56         writew(ioat->dmacount, reg_base + IOAT_CHAN_DMACOUNT_OFFSET);
  57         dev_dbg(to_dev(&ioat->base),
  58                 "%s: head: %#x tail: %#x issued: %#x count: %#x\n",
  59                 __func__, ioat->head, ioat->tail, ioat->issued, ioat->dmacount);
  60 }
  61
  62 static void ioat2_issue_pending(struct dma_chan *chan)
  63 {
  64         struct ioat2_dma_chan *ioat = to_ioat2_chan(chan);
  65
  66         spin_lock_bh(&ioat->ring_lock);
  67         if (ioat->pending == 1)
  68                 __ioat2_issue_pending(ioat);
  69         spin_unlock_bh(&ioat->ring_lock);
  70 }
  71
  72 /**
  73  * ioat2_update_pending - log pending descriptors
  74  * @ioat: ioat2+ channel
  75  *
  76  * set pending to '1' unless pending is already set to '2', pending == 2
  77  * indicates that submission is temporarily blocked due to an in-flight
  78  * reset.  If we are already above the ioat_pending_level threshold then
  79  * just issue pending.
  80  *
  81  * called with ring_lock held
  82  */
  83 static void ioat2_update_pending(struct ioat2_dma_chan *ioat)
  84 {
  85         if (unlikely(ioat->pending == 2))
  86                 return;
  87         else if (ioat2_ring_pending(ioat) > ioat_pending_level)
  88                 __ioat2_issue_pending(ioat);
  89         else
  90                 ioat->pending = 1;
  91 }
  92
  93 static void __ioat2_start_null_desc(struct ioat2_dma_chan *ioat)
  94 {
  95         void __iomem *reg_base = ioat->base.reg_base;
  96         struct ioat_ring_ent *desc;
  97         struct ioat_dma_descriptor *hw;
  98         int idx;
  99
 100         if (ioat2_ring_space(ioat) < 1) {
 101                 dev_err(to_dev(&ioat->base),
 102                         "Unable to start null desc - ring full\n");
 103                 return;
 104         }
 105
 106         dev_dbg(to_dev(&ioat->base), "%s: head: %#x tail: %#x issued: %#x\n",
 107                 __func__, ioat->head, ioat->tail, ioat->issued);
 108         idx = ioat2_desc_alloc(ioat, 1);
 109         desc = ioat2_get_ring_ent(ioat, idx);
 110
 111         hw = desc->hw;
 112         hw->ctl = 0;
 113         hw->ctl_f.null = 1;
 114         hw->ctl_f.int_en = 1;
 115         hw->ctl_f.compl_write = 1;
 116         /* set size to non-zero value (channel returns error when size is 0) */
 117         hw->size = NULL_DESC_BUFFER_SIZE;
 118         hw->src_addr = 0;
 119         hw->dst_addr = 0;
 120         async_tx_ack(&desc->txd);
 121         writel(((u64) desc->txd.phys) & 0x00000000FFFFFFFF,
 122                reg_base + IOAT2_CHAINADDR_OFFSET_LOW);
 123         writel(((u64) desc->txd.phys) >> 32,
 124                reg_base + IOAT2_CHAINADDR_OFFSET_HIGH);
 125         dump_desc_dbg(ioat, desc);
 126         __ioat2_issue_pending(ioat);
 127 }
 128
 129 static void ioat2_start_null_desc(struct ioat2_dma_chan *ioat)
 130 {
 131         spin_lock_bh(&ioat->ring_lock);
 132         __ioat2_start_null_desc(ioat);
 133         spin_unlock_bh(&ioat->ring_lock);
 134 }
 135
 136 static void ioat2_cleanup(struct ioat2_dma_chan *ioat);
 137
 138 /**
 139  * ioat2_reset_part2 - reinit the channel after a reset
 140  */
 141 static void ioat2_reset_part2(struct work_struct *work)
 142 {
 143         struct ioat_chan_common *chan;
 144         struct ioat2_dma_chan *ioat;
 145
 146         chan = container_of(work, struct ioat_chan_common, work.work);
 147         ioat = container_of(chan, struct ioat2_dma_chan, base);
 148
 149         /* ensure that ->tail points to the stalled descriptor
 150          * (ioat->pending is set to 2 at this point so no new
 151          * descriptors will be issued while we perform this cleanup)
 152          */
 153         ioat2_cleanup(ioat);
 154
 155         spin_lock_bh(&chan->cleanup_lock);
 156         spin_lock_bh(&ioat->ring_lock);
 157
 158         /* set the tail to be re-issued */
 159         ioat->issued = ioat->tail;
 160         ioat->dmacount = 0;
 161
 162         dev_dbg(to_dev(&ioat->base),
 163                 "%s: head: %#x tail: %#x issued: %#x count: %#x\n",
 164                 __func__, ioat->head, ioat->tail, ioat->issued, ioat->dmacount);
 165
 166         if (ioat2_ring_pending(ioat)) {
 167                 struct ioat_ring_ent *desc;
 168
 169                 desc = ioat2_get_ring_ent(ioat, ioat->tail);
 170                 writel(((u64) desc->txd.phys) & 0x00000000FFFFFFFF,
 171                        chan->reg_base + IOAT2_CHAINADDR_OFFSET_LOW);
 172                 writel(((u64) desc->txd.phys) >> 32,
 173                        chan->reg_base + IOAT2_CHAINADDR_OFFSET_HIGH);
 174                 __ioat2_issue_pending(ioat);
 175         } else
 176                 __ioat2_start_null_desc(ioat);
 177
 178         spin_unlock_bh(&ioat->ring_lock);
 179         spin_unlock_bh(&chan->cleanup_lock);
 180
 181         dev_info(to_dev(chan),
 182                  "chan%d reset - %d descs waiting, %d total desc\n",
 183                  chan_num(chan), ioat->dmacount, 1 << ioat->alloc_order);
 184 }
 185
 186 /**
 187  * ioat2_reset_channel - restart a channel
 188  * @ioat: IOAT DMA channel handle
 189  */
 190 static void ioat2_reset_channel(struct ioat2_dma_chan *ioat)
 191 {
 192         u32 chansts, chanerr;
 193         struct ioat_chan_common *chan = &ioat->base;
 194         u16 active;
 195
 196         spin_lock_bh(&ioat->ring_lock);
 197         active = ioat2_ring_active(ioat);
 198         spin_unlock_bh(&ioat->ring_lock);
 199         if (!active)
 200                 return;
 201
 202         chanerr = readl(chan->reg_base + IOAT_CHANERR_OFFSET);
 203         chansts = *chan->completion & IOAT_CHANSTS_DMA_TRANSFER_STATUS;
 204         if (chanerr) {
 205                 dev_err(to_dev(chan),
 206                         "chan%d, CHANSTS = 0x%08x CHANERR = 0x%04x, clearing\n",
 207                         chan_num(chan), chansts, chanerr);
 208                 writel(chanerr, chan->reg_base + IOAT_CHANERR_OFFSET);
 209         }
 210
 211         spin_lock_bh(&ioat->ring_lock);
 212         ioat->pending = 2;
 213         writeb(IOAT_CHANCMD_RESET,
 214                chan->reg_base
 215                + IOAT_CHANCMD_OFFSET(chan->device->version));
 216         spin_unlock_bh(&ioat->ring_lock);
 217         schedule_delayed_work(&chan->work, RESET_DELAY);
 218 }
 219
 220 /**
 221  * ioat2_chan_watchdog - watch for stuck channels
 222  */
 223 static void ioat2_chan_watchdog(struct work_struct *work)
 224 {
 225         struct ioatdma_device *device =
 226                 container_of(work, struct ioatdma_device, work.work);
 227         struct ioat2_dma_chan *ioat;
 228         struct ioat_chan_common *chan;
 229         u16 active;
 230         int i;
 231
 232         dev_dbg(&device->pdev->dev, "%s\n", __func__);
 233
 234         for (i = 0; i < device->common.chancnt; i++) {
 235                 chan = ioat_chan_by_index(device, i);
 236                 ioat = container_of(chan, struct ioat2_dma_chan, base);
 237
 238                 /*
 239                  * for version 2.0 if there are descriptors yet to be processed
 240                  * and the last completed hasn't changed since the last watchdog
 241                  *      if they haven't hit the pending level
 242                  *          issue the pending to push them through
 243                  *      else
 244                  *          try resetting the channel
 245                  */
 246                 spin_lock_bh(&ioat->ring_lock);
 247                 active = ioat2_ring_active(ioat);
 248                 spin_unlock_bh(&ioat->ring_lock);
 249
 250                 if (active &&
 251                     chan->last_completion &&
 252                     chan->last_completion == chan->watchdog_completion) {
 253
 254                         if (ioat->pending == 1)
 255                                 ioat2_issue_pending(&chan->common);
 256                         else {
 257                                 ioat2_reset_channel(ioat);
 258                                 chan->watchdog_completion = 0;
 259                         }
 260                 } else {
 261                         chan->last_compl_desc_addr_hw = 0;
 262                         chan->watchdog_completion = chan->last_completion;
 263                 }
 264                 chan->watchdog_last_tcp_cookie = chan->watchdog_tcp_cookie;
 265         }
 266         schedule_delayed_work(&device->work, WATCHDOG_DELAY);
 267 }
 268
 269 /**
 270  * ioat2_cleanup - clean finished descriptors (advance tail pointer)
 271  * @chan: ioat channel to be cleaned up
 272  */
 273 static void ioat2_cleanup(struct ioat2_dma_chan *ioat)
 274 {
 275         struct ioat_chan_common *chan = &ioat->base;
 276         unsigned long phys_complete;
 277         struct ioat_ring_ent *desc;
 278         bool seen_current = false;
 279         u16 active;
 280         int i;
 281         struct dma_async_tx_descriptor *tx;
 282
 283         prefetch(chan->completion);
 284
 285         spin_lock_bh(&chan->cleanup_lock);
 286         phys_complete = ioat_get_current_completion(chan);
 287         if (phys_complete == chan->last_completion) {
 288                 spin_unlock_bh(&chan->cleanup_lock);
 289                 /*
 290                  * perhaps we're stuck so hard that the watchdog can't go off?
 291                  * try to catch it after WATCHDOG_DELAY seconds
 292                  */
 293                 if (chan->device->version < IOAT_VER_3_0) {
 294                         unsigned long tmo;
 295
 296                         tmo = chan->last_completion_time + HZ*WATCHDOG_DELAY;
 297                         if (time_after(jiffies, tmo)) {
 298                                 ioat2_chan_watchdog(&(chan->device->work.work));
 299                                 chan->last_completion_time = jiffies;
 300                         }
 301                 }
 302                 return;
 303         }
 304         chan->last_completion_time = jiffies;
 305
 306         spin_lock_bh(&ioat->ring_lock);
 307
 308         dev_dbg(to_dev(chan), "%s: head: %#x tail: %#x issued: %#x\n",
 309                 __func__, ioat->head, ioat->tail, ioat->issued);
 310
 311         active = ioat2_ring_active(ioat);
 312         for (i = 0; i < active && !seen_current; i++) {
 313                 prefetch(ioat2_get_ring_ent(ioat, ioat->tail + i + 1));
 314                 desc = ioat2_get_ring_ent(ioat, ioat->tail + i);
 315                 tx = &desc->txd;
 316                 dump_desc_dbg(ioat, desc);
 317                 if (tx->cookie) {
 318                         ioat_dma_unmap(chan, tx->flags, desc->len, desc->hw);
 319                         chan->completed_cookie = tx->cookie;
 320                         tx->cookie = 0;
 321                         if (tx->callback) {
 322                                 tx->callback(tx->callback_param);
 323                                 tx->callback = NULL;
 324                         }
 325                 }
 326
 327                 if (tx->phys == phys_complete)
 328                         seen_current = true;
 329         }
 330         ioat->tail += i;
 331         BUG_ON(!seen_current); /* no active descs have written a completion? */
 332         spin_unlock_bh(&ioat->ring_lock);
 333
 334         chan->last_completion = phys_complete;
 335
 336         spin_unlock_bh(&chan->cleanup_lock);
 337 }
 338
 339 static void ioat2_cleanup_tasklet(unsigned long data)
 340 {
 341         struct ioat2_dma_chan *ioat = (void *) data;
 342
 343         ioat2_cleanup(ioat);
 344         writew(IOAT_CHANCTRL_INT_DISABLE,
 345                ioat->base.reg_base + IOAT_CHANCTRL_OFFSET);
 346 }
 347
 348 /**
 349  * ioat2_enumerate_channels - find and initialize the device's channels
 350  * @device: the device to be enumerated
 351  */
 352 static int ioat2_enumerate_channels(struct ioatdma_device *device)
 353 {
 354         struct ioat2_dma_chan *ioat;
 355         struct device *dev = &device->pdev->dev;
 356         struct dma_device *dma = &device->common;
 357         u8 xfercap_log;
 358         int i;
 359
 360         INIT_LIST_HEAD(&dma->channels);
 361         dma->chancnt = readb(device->reg_base + IOAT_CHANCNT_OFFSET);
 362         xfercap_log = readb(device->reg_base + IOAT_XFERCAP_OFFSET);
 363         if (xfercap_log == 0)
 364                 return 0;
 365         dev_dbg(dev, "%s: xfercap = %d\n", __func__, 1 << xfercap_log);
 366
 367         /* FIXME which i/oat version is i7300? */
 368 #ifdef CONFIG_I7300_IDLE_IOAT_CHANNEL
 369         if (i7300_idle_platform_probe(NULL, NULL, 1) == 0)
 370                 dma->chancnt--;
 371 #endif
 372         for (i = 0; i < dma->chancnt; i++) {
 373                 ioat = devm_kzalloc(dev, sizeof(*ioat), GFP_KERNEL);
 374                 if (!ioat)
 375                         break;
 376
 377                 ioat_init_channel(device, &ioat->base, i,
 378                                   ioat2_reset_part2,
 379                                   ioat2_cleanup_tasklet,
 380                                   (unsigned long) ioat);
 381                 ioat->xfercap_log = xfercap_log;
 382                 spin_lock_init(&ioat->ring_lock);
 383         }
 384         dma->chancnt = i;
 385         return i;
 386 }
 387
 388 static dma_cookie_t ioat2_tx_submit_unlock(struct dma_async_tx_descriptor *tx)
 389 {
 390         struct dma_chan *c = tx->chan;
 391         struct ioat2_dma_chan *ioat = to_ioat2_chan(c);
 392         dma_cookie_t cookie = c->cookie;
 393
 394         cookie++;
 395         if (cookie < 0)
 396                 cookie = 1;
 397         tx->cookie = cookie;
 398         c->cookie = cookie;
 399         dev_dbg(to_dev(&ioat->base), "%s: cookie: %d\n", __func__, cookie);
 400
 401         ioat2_update_pending(ioat);
 402         spin_unlock_bh(&ioat->ring_lock);
 403
 404         return cookie;
 405 }
 406
 407 static struct ioat_ring_ent *ioat2_alloc_ring_ent(struct dma_chan *chan)
 408 {
 409         struct ioat_dma_descriptor *hw;
 410         struct ioat_ring_ent *desc;
 411         struct ioatdma_device *dma;
 412         dma_addr_t phys;
 413
 414         dma = to_ioatdma_device(chan->device);
 415         hw = pci_pool_alloc(dma->dma_pool, GFP_KERNEL, &phys);
 416         if (!hw)
 417                 return NULL;
 418         memset(hw, 0, sizeof(*hw));
 419
 420         desc = kzalloc(sizeof(*desc), GFP_KERNEL);
 421         if (!desc) {
 422                 pci_pool_free(dma->dma_pool, hw, phys);
 423                 return NULL;
 424         }
 425
 426         dma_async_tx_descriptor_init(&desc->txd, chan);
 427         desc->txd.tx_submit = ioat2_tx_submit_unlock;
 428         desc->hw = hw;
 429         desc->txd.phys = phys;
 430         return desc;
 431 }
 432
 433 static void ioat2_free_ring_ent(struct ioat_ring_ent *desc, struct dma_chan *chan)
 434 {
 435         struct ioatdma_device *dma;
 436
 437         dma = to_ioatdma_device(chan->device);
 438         pci_pool_free(dma->dma_pool, desc->hw, desc->txd.phys);
 439         kfree(desc);
 440 }
 441
 442 /* ioat2_alloc_chan_resources - allocate/initialize ioat2 descriptor ring
 443  * @chan: channel to be initialized
 444  */
 445 static int ioat2_alloc_chan_resources(struct dma_chan *c)
 446 {
 447         struct ioat2_dma_chan *ioat = to_ioat2_chan(c);
 448         struct ioat_chan_common *chan = &ioat->base;
 449         struct ioat_ring_ent **ring;
 450         u16 chanctrl;
 451         u32 chanerr;
 452         int descs;
 453         int i;
 454
 455         /* have we already been set up? */
 456         if (ioat->ring)
 457                 return 1 << ioat->alloc_order;
 458
 459         /* Setup register to interrupt and write completion status on error */
 460         chanctrl = IOAT_CHANCTRL_ERR_INT_EN | IOAT_CHANCTRL_ANY_ERR_ABORT_EN |
 461                    IOAT_CHANCTRL_ERR_COMPLETION_EN;
 462         writew(chanctrl, chan->reg_base + IOAT_CHANCTRL_OFFSET);
 463
 464         chanerr = readl(chan->reg_base + IOAT_CHANERR_OFFSET);
 465         if (chanerr) {
 466                 dev_err(to_dev(chan), "CHANERR = %x, clearing\n", chanerr);
 467                 writel(chanerr, chan->reg_base + IOAT_CHANERR_OFFSET);
 468         }
 469
 470         /* allocate a completion writeback area */
 471         /* doing 2 32bit writes to mmio since 1 64b write doesn't work */
 472         chan->completion = pci_pool_alloc(chan->device->completion_pool,
 473                                           GFP_KERNEL, &chan->completion_dma);
 474         if (!chan->completion)
 475                 return -ENOMEM;
 476
 477         memset(chan->completion, 0, sizeof(*chan->completion));
 478         writel(((u64) chan->completion_dma) & 0x00000000FFFFFFFF,
 479                chan->reg_base + IOAT_CHANCMP_OFFSET_LOW);
 480         writel(((u64) chan->completion_dma) >> 32,
 481                chan->reg_base + IOAT_CHANCMP_OFFSET_HIGH);
 482
 483         ioat->alloc_order = ioat_get_alloc_order();
 484         descs = 1 << ioat->alloc_order;
 485
 486         /* allocate the array to hold the software ring */
 487         ring = kcalloc(descs, sizeof(*ring), GFP_KERNEL);
 488         if (!ring)
 489                 return -ENOMEM;
 490         for (i = 0; i < descs; i++) {
 491                 ring[i] = ioat2_alloc_ring_ent(c);
 492                 if (!ring[i]) {
 493                         while (i--)
 494                                 ioat2_free_ring_ent(ring[i], c);
 495                         kfree(ring);
 496                         return -ENOMEM;
 497                 }
 498                 set_desc_id(ring[i], i);
 499         }
 500
 501         /* link descs */
 502         for (i = 0; i < descs-1; i++) {
 503                 struct ioat_ring_ent *next = ring[i+1];
 504                 struct ioat_dma_descriptor *hw = ring[i]->hw;
 505
 506                 hw->next = next->txd.phys;
 507         }
 508         ring[i]->hw->next = ring[0]->txd.phys;
 509
 510         spin_lock_bh(&ioat->ring_lock);
 511         ioat->ring = ring;
 512         ioat->head = 0;
 513         ioat->issued = 0;
 514         ioat->tail = 0;
 515         ioat->pending = 0;
 516         spin_unlock_bh(&ioat->ring_lock);
 517
 518         tasklet_enable(&chan->cleanup_task);
 519         ioat2_start_null_desc(ioat);
 520
 521         return descs;
 522 }
 523
 524 /**
 525  * ioat2_alloc_and_lock - common descriptor alloc boilerplate for ioat2,3 ops
 526  * @idx: gets starting descriptor index on successful allocation
 527  * @ioat: ioat2,3 channel (ring) to operate on
 528  * @num_descs: allocation length
 529  */
 530 static int ioat2_alloc_and_lock(u16 *idx, struct ioat2_dma_chan *ioat, int num_descs)
 531 {
 532         struct ioat_chan_common *chan = &ioat->base;
 533
 534         spin_lock_bh(&ioat->ring_lock);
 535         if (unlikely(ioat2_ring_space(ioat) < num_descs)) {
 536                 if (printk_ratelimit())
 537                         dev_dbg(to_dev(chan),
 538                                 "%s: ring full! num_descs: %d (%x:%x:%x)\n",
 539                                 __func__, num_descs, ioat->head, ioat->tail,
 540                                 ioat->issued);
 541                 spin_unlock_bh(&ioat->ring_lock);
 542
 543                 /* do direct reclaim in the allocation failure case */
 544                 ioat2_cleanup(ioat);
 545
 546                 return -ENOMEM;
 547         }
 548
 549         dev_dbg(to_dev(chan), "%s: num_descs: %d (%x:%x:%x)\n",
 550                 __func__, num_descs, ioat->head, ioat->tail, ioat->issued);
 551
 552         *idx = ioat2_desc_alloc(ioat, num_descs);
 553         return 0;  /* with ioat->ring_lock held */
 554 }
 555
 556 static struct dma_async_tx_descriptor *
 557 ioat2_dma_prep_memcpy_lock(struct dma_chan *c, dma_addr_t dma_dest,
 558                            dma_addr_t dma_src, size_t len, unsigned long flags)
 559 {
 560         struct ioat2_dma_chan *ioat = to_ioat2_chan(c);
 561         struct ioat_dma_descriptor *hw;
 562         struct ioat_ring_ent *desc;
 563         dma_addr_t dst = dma_dest;
 564         dma_addr_t src = dma_src;
 565         size_t total_len = len;
 566         int num_descs;
 567         u16 idx;
 568         int i;
 569
 570         num_descs = ioat2_xferlen_to_descs(ioat, len);
 571         if (likely(num_descs) &&
 572             ioat2_alloc_and_lock(&idx, ioat, num_descs) == 0)
 573                 /* pass */;
 574         else
 575                 return NULL;
 576         for (i = 0; i < num_descs; i++) {
 577                 size_t copy = min_t(size_t, len, 1 << ioat->xfercap_log);
 578
 579                 desc = ioat2_get_ring_ent(ioat, idx + i);
 580                 hw = desc->hw;
 581
 582                 hw->size = copy;
 583                 hw->ctl = 0;
 584                 hw->src_addr = src;
 585                 hw->dst_addr = dst;
 586
 587                 len -= copy;
 588                 dst += copy;
 589                 src += copy;
 590                 dump_desc_dbg(ioat, desc);
 591         }
 592
 593         desc->txd.flags = flags;
 594         desc->len = total_len;
 595         hw->ctl_f.int_en = !!(flags & DMA_PREP_INTERRUPT);
 596         hw->ctl_f.compl_write = 1;
 597         dump_desc_dbg(ioat, desc);
 598         /* we leave the channel locked to ensure in order submission */
 599
 600         return &desc->txd;
 601 }
 602
 603 /**
 604  * ioat2_free_chan_resources - release all the descriptors
 605  * @chan: the channel to be cleaned
 606  */
 607 static void ioat2_free_chan_resources(struct dma_chan *c)
 608 {
 609         struct ioat2_dma_chan *ioat = to_ioat2_chan(c);
 610         struct ioat_chan_common *chan = &ioat->base;
 611         struct ioatdma_device *ioatdma_device = chan->device;
 612         struct ioat_ring_ent *desc;
 613         const u16 total_descs = 1 << ioat->alloc_order;
 614         int descs;
 615         int i;
 616
 617         /* Before freeing channel resources first check
 618          * if they have been previously allocated for this channel.
 619          */
 620         if (!ioat->ring)
 621                 return;
 622
 623         tasklet_disable(&chan->cleanup_task);
 624         ioat2_cleanup(ioat);
 625
 626         /* Delay 100ms after reset to allow internal DMA logic to quiesce
 627          * before removing DMA descriptor resources.
 628          */
 629         writeb(IOAT_CHANCMD_RESET,
 630                chan->reg_base + IOAT_CHANCMD_OFFSET(chan->device->version));
 631         mdelay(100);
 632
 633         spin_lock_bh(&ioat->ring_lock);
 634         descs = ioat2_ring_space(ioat);
 635         dev_dbg(to_dev(chan), "freeing %d idle descriptors\n", descs);
 636         for (i = 0; i < descs; i++) {
 637                 desc = ioat2_get_ring_ent(ioat, ioat->head + i);
 638                 ioat2_free_ring_ent(desc, c);
 639         }
 640
 641         if (descs < total_descs)
 642                 dev_err(to_dev(chan), "Freeing %d in use descriptors!\n",
 643                         total_descs - descs);
 644
 645         for (i = 0; i < total_descs - descs; i++) {
 646                 desc = ioat2_get_ring_ent(ioat, ioat->tail + i);
 647                 dump_desc_dbg(ioat, desc);
 648                 ioat2_free_ring_ent(desc, c);
 649         }
 650
 651         kfree(ioat->ring);
 652         ioat->ring = NULL;
 653         ioat->alloc_order = 0;
 654         pci_pool_free(ioatdma_device->completion_pool,
 655                       chan->completion,
 656                       chan->completion_dma);
 657         spin_unlock_bh(&ioat->ring_lock);
 658
 659         chan->last_completion = 0;
 660         chan->completion_dma = 0;
 661         ioat->pending = 0;
 662         ioat->dmacount = 0;
 663         chan->watchdog_completion = 0;
 664         chan->last_compl_desc_addr_hw = 0;
 665         chan->watchdog_tcp_cookie = 0;
 666         chan->watchdog_last_tcp_cookie = 0;
 667 }
 668
 669 static enum dma_status
 670 ioat2_is_complete(struct dma_chan *c, dma_cookie_t cookie,
 671                      dma_cookie_t *done, dma_cookie_t *used)
 672 {
 673         struct ioat2_dma_chan *ioat = to_ioat2_chan(c);
 674
 675         if (ioat_is_complete(c, cookie, done, used) == DMA_SUCCESS)
 676                 return DMA_SUCCESS;
 677
 678         ioat2_cleanup(ioat);
 679
 680         return ioat_is_complete(c, cookie, done, used);
 681 }
 682
 683 int ioat2_dma_probe(struct ioatdma_device *device, int dca)
 684 {
 685         struct pci_dev *pdev = device->pdev;
 686         struct dma_device *dma;
 687         struct dma_chan *c;
 688         struct ioat_chan_common *chan;
 689         int err;
 690
 691         device->enumerate_channels = ioat2_enumerate_channels;
 692         dma = &device->common;
 693         dma->device_prep_dma_memcpy = ioat2_dma_prep_memcpy_lock;
 694         dma->device_issue_pending = ioat2_issue_pending;
 695         dma->device_alloc_chan_resources = ioat2_alloc_chan_resources;
 696         dma->device_free_chan_resources = ioat2_free_chan_resources;
 697         dma->device_is_tx_complete = ioat2_is_complete;
 698
 699         err = ioat_probe(device);
 700         if (err)
 701                 return err;
 702         ioat_set_tcp_copy_break(2048);
 703
 704         list_for_each_entry(c, &dma->channels, device_node) {
 705                 chan = to_chan_common(c);
 706                 writel(IOAT_DCACTRL_CMPL_WRITE_ENABLE | IOAT_DMA_DCA_ANY_CPU,
 707                        chan->reg_base + IOAT_DCACTRL_OFFSET);
 708         }
 709
 710         err = ioat_register(device);
 711         if (err)
 712                 return err;
 713         if (dca)
 714                 device->dca = ioat2_dca_init(pdev, device->reg_base);
 715
 716         INIT_DELAYED_WORK(&device->work, ioat2_chan_watchdog);
 717         schedule_delayed_work(&device->work, WATCHDOG_DELAY);
 718
 719         return err;
 720 }
 721
 722 int ioat3_dma_probe(struct ioatdma_device *device, int dca)
 723 {
 724         struct pci_dev *pdev = device->pdev;
 725         struct dma_device *dma;
 726         struct dma_chan *c;
 727         struct ioat_chan_common *chan;
 728         int err;
 729         u16 dev_id;
 730
 731         device->enumerate_channels = ioat2_enumerate_channels;
 732         dma = &device->common;
 733         dma->device_prep_dma_memcpy = ioat2_dma_prep_memcpy_lock;
 734         dma->device_issue_pending = ioat2_issue_pending;
 735         dma->device_alloc_chan_resources = ioat2_alloc_chan_resources;
 736         dma->device_free_chan_resources = ioat2_free_chan_resources;
 737         dma->device_is_tx_complete = ioat2_is_complete;
 738
 739         /* -= IOAT ver.3 workarounds =- */
 740         /* Write CHANERRMSK_INT with 3E07h to mask out the errors
 741          * that can cause stability issues for IOAT ver.3
 742          */
 743         pci_write_config_dword(pdev, IOAT_PCI_CHANERRMASK_INT_OFFSET, 0x3e07);
 744
 745         /* Clear DMAUNCERRSTS Cfg-Reg Parity Error status bit
 746          * (workaround for spurious config parity error after restart)
 747          */
 748         pci_read_config_word(pdev, IOAT_PCI_DEVICE_ID_OFFSET, &dev_id);
 749         if (dev_id == PCI_DEVICE_ID_INTEL_IOAT_TBG0)
 750                 pci_write_config_dword(pdev, IOAT_PCI_DMAUNCERRSTS_OFFSET, 0x10);
 751
 752         err = ioat_probe(device);
 753         if (err)
 754                 return err;
 755         ioat_set_tcp_copy_break(262144);
 756
 757         list_for_each_entry(c, &dma->channels, device_node) {
 758                 chan = to_chan_common(c);
 759                 writel(IOAT_DMA_DCA_ANY_CPU,
 760                        chan->reg_base + IOAT_DCACTRL_OFFSET);
 761         }
 762
 763         err = ioat_register(device);
 764         if (err)
 765                 return err;
 766         if (dca)
 767                 device->dca = ioat3_dca_init(pdev, device->reg_base);
 768
 769         return err;
 770 }