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 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"
  46                  " (default: 8 max: 16)");
  47 static int ioat_ring_max_alloc_order = IOAT_MAX_ORDER;
  48 module_param(ioat_ring_max_alloc_order, int, 0644);
  49 MODULE_PARM_DESC(ioat_ring_max_alloc_order,
  50                  "ioat2+: upper limit for ring size (default: 16)");
  51
  52 void __ioat2_issue_pending(struct ioat2_dma_chan *ioat)
  53 {
  54         void * __iomem reg_base = ioat->base.reg_base;
  55
  56         ioat->pending = 0;
  57         ioat->dmacount += ioat2_ring_pending(ioat);
  58         ioat->issued = ioat->head;
  59         /* make descriptor updates globally visible before notifying channel */
  60         wmb();
  61         writew(ioat->dmacount, reg_base + IOAT_CHAN_DMACOUNT_OFFSET);
  62         dev_dbg(to_dev(&ioat->base),
  63                 "%s: head: %#x tail: %#x issued: %#x count: %#x\n",
  64                 __func__, ioat->head, ioat->tail, ioat->issued, ioat->dmacount);
  65 }
  66
  67 void ioat2_issue_pending(struct dma_chan *chan)
  68 {
  69         struct ioat2_dma_chan *ioat = to_ioat2_chan(chan);
  70
  71         spin_lock_bh(&ioat->ring_lock);
  72         if (ioat->pending == 1)
  73                 __ioat2_issue_pending(ioat);
  74         spin_unlock_bh(&ioat->ring_lock);
  75 }
  76
  77 /**
  78  * ioat2_update_pending - log pending descriptors
  79  * @ioat: ioat2+ channel
  80  *
  81  * set pending to '1' unless pending is already set to '2', pending == 2
  82  * indicates that submission is temporarily blocked due to an in-flight
  83  * reset.  If we are already above the ioat_pending_level threshold then
  84  * just issue pending.
  85  *
  86  * called with ring_lock held
  87  */
  88 static void ioat2_update_pending(struct ioat2_dma_chan *ioat)
  89 {
  90         if (unlikely(ioat->pending == 2))
  91                 return;
  92         else if (ioat2_ring_pending(ioat) > ioat_pending_level)
  93                 __ioat2_issue_pending(ioat);
  94         else
  95                 ioat->pending = 1;
  96 }
  97
  98 static void __ioat2_start_null_desc(struct ioat2_dma_chan *ioat)
  99 {
 100         struct ioat_ring_ent *desc;
 101         struct ioat_dma_descriptor *hw;
 102         int idx;
 103
 104         if (ioat2_ring_space(ioat) < 1) {
 105                 dev_err(to_dev(&ioat->base),
 106                         "Unable to start null desc - ring full\n");
 107                 return;
 108         }
 109
 110         dev_dbg(to_dev(&ioat->base), "%s: head: %#x tail: %#x issued: %#x\n",
 111                 __func__, ioat->head, ioat->tail, ioat->issued);
 112         idx = ioat2_desc_alloc(ioat, 1);
 113         desc = ioat2_get_ring_ent(ioat, idx);
 114
 115         hw = desc->hw;
 116         hw->ctl = 0;
 117         hw->ctl_f.null = 1;
 118         hw->ctl_f.int_en = 1;
 119         hw->ctl_f.compl_write = 1;
 120         /* set size to non-zero value (channel returns error when size is 0) */
 121         hw->size = NULL_DESC_BUFFER_SIZE;
 122         hw->src_addr = 0;
 123         hw->dst_addr = 0;
 124         async_tx_ack(&desc->txd);
 125         ioat2_set_chainaddr(ioat, desc->txd.phys);
 126         dump_desc_dbg(ioat, desc);
 127         __ioat2_issue_pending(ioat);
 128 }
 129
 130 static void ioat2_start_null_desc(struct ioat2_dma_chan *ioat)
 131 {
 132         spin_lock_bh(&ioat->ring_lock);
 133         __ioat2_start_null_desc(ioat);
 134         spin_unlock_bh(&ioat->ring_lock);
 135 }
 136
 137 static void __cleanup(struct ioat2_dma_chan *ioat, unsigned long phys_complete)
 138 {
 139         struct ioat_chan_common *chan = &ioat->base;
 140         struct dma_async_tx_descriptor *tx;
 141         struct ioat_ring_ent *desc;
 142         bool seen_current = false;
 143         u16 active;
 144         int i;
 145
 146         dev_dbg(to_dev(chan), "%s: head: %#x tail: %#x issued: %#x\n",
 147                 __func__, ioat->head, ioat->tail, ioat->issued);
 148
 149         active = ioat2_ring_active(ioat);
 150         for (i = 0; i < active && !seen_current; i++) {
 151                 prefetch(ioat2_get_ring_ent(ioat, ioat->tail + i + 1));
 152                 desc = ioat2_get_ring_ent(ioat, ioat->tail + i);
 153                 tx = &desc->txd;
 154                 dump_desc_dbg(ioat, desc);
 155                 if (tx->cookie) {
 156                         ioat_dma_unmap(chan, tx->flags, desc->len, desc->hw);
 157                         chan->completed_cookie = tx->cookie;
 158                         tx->cookie = 0;
 159                         if (tx->callback) {
 160                                 tx->callback(tx->callback_param);
 161                                 tx->callback = NULL;
 162                         }
 163                 }
 164
 165                 if (tx->phys == phys_complete)
 166                         seen_current = true;
 167         }
 168         ioat->tail += i;
 169         BUG_ON(!seen_current); /* no active descs have written a completion? */
 170
 171         chan->last_completion = phys_complete;
 172         if (ioat->head == ioat->tail) {
 173                 dev_dbg(to_dev(chan), "%s: cancel completion timeout\n",
 174                         __func__);
 175                 clear_bit(IOAT_COMPLETION_PENDING, &chan->state);
 176                 mod_timer(&chan->timer, jiffies + IDLE_TIMEOUT);
 177         }
 178 }
 179
 180 /**
 181  * ioat2_cleanup - clean finished descriptors (advance tail pointer)
 182  * @chan: ioat channel to be cleaned up
 183  */
 184 static void ioat2_cleanup(struct ioat2_dma_chan *ioat)
 185 {
 186         struct ioat_chan_common *chan = &ioat->base;
 187         unsigned long phys_complete;
 188
 189         prefetch(chan->completion);
 190
 191         if (!spin_trylock_bh(&chan->cleanup_lock))
 192                 return;
 193
 194         if (!ioat_cleanup_preamble(chan, &phys_complete)) {
 195                 spin_unlock_bh(&chan->cleanup_lock);
 196                 return;
 197         }
 198
 199         if (!spin_trylock_bh(&ioat->ring_lock)) {
 200                 spin_unlock_bh(&chan->cleanup_lock);
 201                 return;
 202         }
 203
 204         __cleanup(ioat, phys_complete);
 205
 206         spin_unlock_bh(&ioat->ring_lock);
 207         spin_unlock_bh(&chan->cleanup_lock);
 208 }
 209
 210 void ioat2_cleanup_tasklet(unsigned long data)
 211 {
 212         struct ioat2_dma_chan *ioat = (void *) data;
 213
 214         ioat2_cleanup(ioat);
 215         writew(IOAT_CHANCTRL_RUN, ioat->base.reg_base + IOAT_CHANCTRL_OFFSET);
 216 }
 217
 218 void __ioat2_restart_chan(struct ioat2_dma_chan *ioat)
 219 {
 220         struct ioat_chan_common *chan = &ioat->base;
 221
 222         /* set the tail to be re-issued */
 223         ioat->issued = ioat->tail;
 224         ioat->dmacount = 0;
 225         set_bit(IOAT_COMPLETION_PENDING, &chan->state);
 226         mod_timer(&chan->timer, jiffies + COMPLETION_TIMEOUT);
 227
 228         dev_dbg(to_dev(chan),
 229                 "%s: head: %#x tail: %#x issued: %#x count: %#x\n",
 230                 __func__, ioat->head, ioat->tail, ioat->issued, ioat->dmacount);
 231
 232         if (ioat2_ring_pending(ioat)) {
 233                 struct ioat_ring_ent *desc;
 234
 235                 desc = ioat2_get_ring_ent(ioat, ioat->tail);
 236                 ioat2_set_chainaddr(ioat, desc->txd.phys);
 237                 __ioat2_issue_pending(ioat);
 238         } else
 239                 __ioat2_start_null_desc(ioat);
 240 }
 241
 242 static void ioat2_restart_channel(struct ioat2_dma_chan *ioat)
 243 {
 244         struct ioat_chan_common *chan = &ioat->base;
 245         unsigned long phys_complete;
 246         u32 status;
 247
 248         status = ioat_chansts(chan);
 249         if (is_ioat_active(status) || is_ioat_idle(status))
 250                 ioat_suspend(chan);
 251         while (is_ioat_active(status) || is_ioat_idle(status)) {
 252                 status = ioat_chansts(chan);
 253                 cpu_relax();
 254         }
 255
 256         if (ioat_cleanup_preamble(chan, &phys_complete))
 257                 __cleanup(ioat, phys_complete);
 258
 259         __ioat2_restart_chan(ioat);
 260 }
 261
 262 void ioat2_timer_event(unsigned long data)
 263 {
 264         struct ioat2_dma_chan *ioat = (void *) data;
 265         struct ioat_chan_common *chan = &ioat->base;
 266
 267         spin_lock_bh(&chan->cleanup_lock);
 268         if (test_bit(IOAT_COMPLETION_PENDING, &chan->state)) {
 269                 unsigned long phys_complete;
 270                 u64 status;
 271
 272                 spin_lock_bh(&ioat->ring_lock);
 273                 status = ioat_chansts(chan);
 274
 275                 /* when halted due to errors check for channel
 276                  * programming errors before advancing the completion state
 277                  */
 278                 if (is_ioat_halted(status)) {
 279                         u32 chanerr;
 280
 281                         chanerr = readl(chan->reg_base + IOAT_CHANERR_OFFSET);
 282                         BUG_ON(is_ioat_bug(chanerr));
 283                 }
 284
 285                 /* if we haven't made progress and we have already
 286                  * acknowledged a pending completion once, then be more
 287                  * forceful with a restart
 288                  */
 289                 if (ioat_cleanup_preamble(chan, &phys_complete))
 290                         __cleanup(ioat, phys_complete);
 291                 else if (test_bit(IOAT_COMPLETION_ACK, &chan->state))
 292                         ioat2_restart_channel(ioat);
 293                 else {
 294                         set_bit(IOAT_COMPLETION_ACK, &chan->state);
 295                         mod_timer(&chan->timer, jiffies + COMPLETION_TIMEOUT);
 296                 }
 297                 spin_unlock_bh(&ioat->ring_lock);
 298         } else {
 299                 u16 active;
 300
 301                 /* if the ring is idle, empty, and oversized try to step
 302                  * down the size
 303                  */
 304                 spin_lock_bh(&ioat->ring_lock);
 305                 active = ioat2_ring_active(ioat);
 306                 if (active == 0 && ioat->alloc_order > ioat_get_alloc_order())
 307                         reshape_ring(ioat, ioat->alloc_order-1);
 308                 spin_unlock_bh(&ioat->ring_lock);
 309
 310                 /* keep shrinking until we get back to our minimum
 311                  * default size
 312                  */
 313                 if (ioat->alloc_order > ioat_get_alloc_order())
 314                         mod_timer(&chan->timer, jiffies + IDLE_TIMEOUT);
 315         }
 316         spin_unlock_bh(&chan->cleanup_lock);
 317 }
 318
 319 /**
 320  * ioat2_enumerate_channels - find and initialize the device's channels
 321  * @device: the device to be enumerated
 322  */
 323 int ioat2_enumerate_channels(struct ioatdma_device *device)
 324 {
 325         struct ioat2_dma_chan *ioat;
 326         struct device *dev = &device->pdev->dev;
 327         struct dma_device *dma = &device->common;
 328         u8 xfercap_log;
 329         int i;
 330
 331         INIT_LIST_HEAD(&dma->channels);
 332         dma->chancnt = readb(device->reg_base + IOAT_CHANCNT_OFFSET);
 333         dma->chancnt &= 0x1f; /* bits [4:0] valid */
 334         if (dma->chancnt > ARRAY_SIZE(device->idx)) {
 335                 dev_warn(dev, "(%d) exceeds max supported channels (%zu)\n",
 336                          dma->chancnt, ARRAY_SIZE(device->idx));
 337                 dma->chancnt = ARRAY_SIZE(device->idx);
 338         }
 339         xfercap_log = readb(device->reg_base + IOAT_XFERCAP_OFFSET);
 340         xfercap_log &= 0x1f; /* bits [4:0] valid */
 341         if (xfercap_log == 0)
 342                 return 0;
 343         dev_dbg(dev, "%s: xfercap = %d\n", __func__, 1 << xfercap_log);
 344
 345         /* FIXME which i/oat version is i7300? */
 346 #ifdef CONFIG_I7300_IDLE_IOAT_CHANNEL
 347         if (i7300_idle_platform_probe(NULL, NULL, 1) == 0)
 348                 dma->chancnt--;
 349 #endif
 350         for (i = 0; i < dma->chancnt; i++) {
 351                 ioat = devm_kzalloc(dev, sizeof(*ioat), GFP_KERNEL);
 352                 if (!ioat)
 353                         break;
 354
 355                 ioat_init_channel(device, &ioat->base, i,
 356                                   device->timer_fn,
 357                                   device->cleanup_tasklet,
 358                                   (unsigned long) ioat);
 359                 ioat->xfercap_log = xfercap_log;
 360                 spin_lock_init(&ioat->ring_lock);
 361         }
 362         dma->chancnt = i;
 363         return i;
 364 }
 365
 366 static dma_cookie_t ioat2_tx_submit_unlock(struct dma_async_tx_descriptor *tx)
 367 {
 368         struct dma_chan *c = tx->chan;
 369         struct ioat2_dma_chan *ioat = to_ioat2_chan(c);
 370         struct ioat_chan_common *chan = &ioat->base;
 371         dma_cookie_t cookie = c->cookie;
 372
 373         cookie++;
 374         if (cookie < 0)
 375                 cookie = 1;
 376         tx->cookie = cookie;
 377         c->cookie = cookie;
 378         dev_dbg(to_dev(&ioat->base), "%s: cookie: %d\n", __func__, cookie);
 379
 380         if (!test_and_set_bit(IOAT_COMPLETION_PENDING, &chan->state))
 381                 mod_timer(&chan->timer, jiffies + COMPLETION_TIMEOUT);
 382         ioat2_update_pending(ioat);
 383         spin_unlock_bh(&ioat->ring_lock);
 384
 385         return cookie;
 386 }
 387
 388 static struct ioat_ring_ent *ioat2_alloc_ring_ent(struct dma_chan *chan, gfp_t flags)
 389 {
 390         struct ioat_dma_descriptor *hw;
 391         struct ioat_ring_ent *desc;
 392         struct ioatdma_device *dma;
 393         dma_addr_t phys;
 394
 395         dma = to_ioatdma_device(chan->device);
 396         hw = pci_pool_alloc(dma->dma_pool, flags, &phys);
 397         if (!hw)
 398                 return NULL;
 399         memset(hw, 0, sizeof(*hw));
 400
 401         desc = kmem_cache_alloc(ioat2_cache, flags);
 402         if (!desc) {
 403                 pci_pool_free(dma->dma_pool, hw, phys);
 404                 return NULL;
 405         }
 406         memset(desc, 0, sizeof(*desc));
 407
 408         dma_async_tx_descriptor_init(&desc->txd, chan);
 409         desc->txd.tx_submit = ioat2_tx_submit_unlock;
 410         desc->hw = hw;
 411         desc->txd.phys = phys;
 412         return desc;
 413 }
 414
 415 static void ioat2_free_ring_ent(struct ioat_ring_ent *desc, struct dma_chan *chan)
 416 {
 417         struct ioatdma_device *dma;
 418
 419         dma = to_ioatdma_device(chan->device);
 420         pci_pool_free(dma->dma_pool, desc->hw, desc->txd.phys);
 421         kmem_cache_free(ioat2_cache, desc);
 422 }
 423
 424 static struct ioat_ring_ent **ioat2_alloc_ring(struct dma_chan *c, int order, gfp_t flags)
 425 {
 426         struct ioat_ring_ent **ring;
 427         int descs = 1 << order;
 428         int i;
 429
 430         if (order > ioat_get_max_alloc_order())
 431                 return NULL;
 432
 433         /* allocate the array to hold the software ring */
 434         ring = kcalloc(descs, sizeof(*ring), flags);
 435         if (!ring)
 436                 return NULL;
 437         for (i = 0; i < descs; i++) {
 438                 ring[i] = ioat2_alloc_ring_ent(c, flags);
 439                 if (!ring[i]) {
 440                         while (i--)
 441                                 ioat2_free_ring_ent(ring[i], c);
 442                         kfree(ring);
 443                         return NULL;
 444                 }
 445                 set_desc_id(ring[i], i);
 446         }
 447
 448         /* link descs */
 449         for (i = 0; i < descs-1; i++) {
 450                 struct ioat_ring_ent *next = ring[i+1];
 451                 struct ioat_dma_descriptor *hw = ring[i]->hw;
 452
 453                 hw->next = next->txd.phys;
 454         }
 455         ring[i]->hw->next = ring[0]->txd.phys;
 456
 457         return ring;
 458 }
 459
 460 /* ioat2_alloc_chan_resources - allocate/initialize ioat2 descriptor ring
 461  * @chan: channel to be initialized
 462  */
 463 int ioat2_alloc_chan_resources(struct dma_chan *c)
 464 {
 465         struct ioat2_dma_chan *ioat = to_ioat2_chan(c);
 466         struct ioat_chan_common *chan = &ioat->base;
 467         struct ioat_ring_ent **ring;
 468         u32 chanerr;
 469         int order;
 470
 471         /* have we already been set up? */
 472         if (ioat->ring)
 473                 return 1 << ioat->alloc_order;
 474
 475         /* Setup register to interrupt and write completion status on error */
 476         writew(IOAT_CHANCTRL_RUN, chan->reg_base + IOAT_CHANCTRL_OFFSET);
 477
 478         chanerr = readl(chan->reg_base + IOAT_CHANERR_OFFSET);
 479         if (chanerr) {
 480                 dev_err(to_dev(chan), "CHANERR = %x, clearing\n", chanerr);
 481                 writel(chanerr, chan->reg_base + IOAT_CHANERR_OFFSET);
 482         }
 483
 484         /* allocate a completion writeback area */
 485         /* doing 2 32bit writes to mmio since 1 64b write doesn't work */
 486         chan->completion = pci_pool_alloc(chan->device->completion_pool,
 487                                           GFP_KERNEL, &chan->completion_dma);
 488         if (!chan->completion)
 489                 return -ENOMEM;
 490
 491         memset(chan->completion, 0, sizeof(*chan->completion));
 492         writel(((u64) chan->completion_dma) & 0x00000000FFFFFFFF,
 493                chan->reg_base + IOAT_CHANCMP_OFFSET_LOW);
 494         writel(((u64) chan->completion_dma) >> 32,
 495                chan->reg_base + IOAT_CHANCMP_OFFSET_HIGH);
 496
 497         order = ioat_get_alloc_order();
 498         ring = ioat2_alloc_ring(c, order, GFP_KERNEL);
 499         if (!ring)
 500                 return -ENOMEM;
 501
 502         spin_lock_bh(&ioat->ring_lock);
 503         ioat->ring = ring;
 504         ioat->head = 0;
 505         ioat->issued = 0;
 506         ioat->tail = 0;
 507         ioat->pending = 0;
 508         ioat->alloc_order = order;
 509         spin_unlock_bh(&ioat->ring_lock);
 510
 511         tasklet_enable(&chan->cleanup_task);
 512         ioat2_start_null_desc(ioat);
 513
 514         return 1 << ioat->alloc_order;
 515 }
 516
 517 bool reshape_ring(struct ioat2_dma_chan *ioat, int order)
 518 {
 519         /* reshape differs from normal ring allocation in that we want
 520          * to allocate a new software ring while only
 521          * extending/truncating the hardware ring
 522          */
 523         struct ioat_chan_common *chan = &ioat->base;
 524         struct dma_chan *c = &chan->common;
 525         const u16 curr_size = ioat2_ring_mask(ioat) + 1;
 526         const u16 active = ioat2_ring_active(ioat);
 527         const u16 new_size = 1 << order;
 528         struct ioat_ring_ent **ring;
 529         u16 i;
 530
 531         if (order > ioat_get_max_alloc_order())
 532                 return false;
 533
 534         /* double check that we have at least 1 free descriptor */
 535         if (active == curr_size)
 536                 return false;
 537
 538         /* when shrinking, verify that we can hold the current active
 539          * set in the new ring
 540          */
 541         if (active >= new_size)
 542                 return false;
 543
 544         /* allocate the array to hold the software ring */
 545         ring = kcalloc(new_size, sizeof(*ring), GFP_NOWAIT);
 546         if (!ring)
 547                 return false;
 548
 549         /* allocate/trim descriptors as needed */
 550         if (new_size > curr_size) {
 551                 /* copy current descriptors to the new ring */
 552                 for (i = 0; i < curr_size; i++) {
 553                         u16 curr_idx = (ioat->tail+i) & (curr_size-1);
 554                         u16 new_idx = (ioat->tail+i) & (new_size-1);
 555
 556                         ring[new_idx] = ioat->ring[curr_idx];
 557                         set_desc_id(ring[new_idx], new_idx);
 558                 }
 559
 560                 /* add new descriptors to the ring */
 561                 for (i = curr_size; i < new_size; i++) {
 562                         u16 new_idx = (ioat->tail+i) & (new_size-1);
 563
 564                         ring[new_idx] = ioat2_alloc_ring_ent(c, GFP_NOWAIT);
 565                         if (!ring[new_idx]) {
 566                                 while (i--) {
 567                                         u16 new_idx = (ioat->tail+i) & (new_size-1);
 568
 569                                         ioat2_free_ring_ent(ring[new_idx], c);
 570                                 }
 571                                 kfree(ring);
 572                                 return false;
 573                         }
 574                         set_desc_id(ring[new_idx], new_idx);
 575                 }
 576
 577                 /* hw link new descriptors */
 578                 for (i = curr_size-1; i < new_size; i++) {
 579                         u16 new_idx = (ioat->tail+i) & (new_size-1);
 580                         struct ioat_ring_ent *next = ring[(new_idx+1) & (new_size-1)];
 581                         struct ioat_dma_descriptor *hw = ring[new_idx]->hw;
 582
 583                         hw->next = next->txd.phys;
 584                 }
 585         } else {
 586                 struct ioat_dma_descriptor *hw;
 587                 struct ioat_ring_ent *next;
 588
 589                 /* copy current descriptors to the new ring, dropping the
 590                  * removed descriptors
 591                  */
 592                 for (i = 0; i < new_size; i++) {
 593                         u16 curr_idx = (ioat->tail+i) & (curr_size-1);
 594                         u16 new_idx = (ioat->tail+i) & (new_size-1);
 595
 596                         ring[new_idx] = ioat->ring[curr_idx];
 597                         set_desc_id(ring[new_idx], new_idx);
 598                 }
 599
 600                 /* free deleted descriptors */
 601                 for (i = new_size; i < curr_size; i++) {
 602                         struct ioat_ring_ent *ent;
 603
 604                         ent = ioat2_get_ring_ent(ioat, ioat->tail+i);
 605                         ioat2_free_ring_ent(ent, c);
 606                 }
 607
 608                 /* fix up hardware ring */
 609                 hw = ring[(ioat->tail+new_size-1) & (new_size-1)]->hw;
 610                 next = ring[(ioat->tail+new_size) & (new_size-1)];
 611                 hw->next = next->txd.phys;
 612         }
 613
 614         dev_dbg(to_dev(chan), "%s: allocated %d descriptors\n",
 615                 __func__, new_size);
 616
 617         kfree(ioat->ring);
 618         ioat->ring = ring;
 619         ioat->alloc_order = order;
 620
 621         return true;
 622 }
 623
 624 /**
 625  * ioat2_alloc_and_lock - common descriptor alloc boilerplate for ioat2,3 ops
 626  * @idx: gets starting descriptor index on successful allocation
 627  * @ioat: ioat2,3 channel (ring) to operate on
 628  * @num_descs: allocation length
 629  */
 630 int ioat2_alloc_and_lock(u16 *idx, struct ioat2_dma_chan *ioat, int num_descs)
 631 {
 632         struct ioat_chan_common *chan = &ioat->base;
 633
 634         spin_lock_bh(&ioat->ring_lock);
 635         /* never allow the last descriptor to be consumed, we need at
 636          * least one free at all times to allow for on-the-fly ring
 637          * resizing.
 638          */
 639         while (unlikely(ioat2_ring_space(ioat) <= num_descs)) {
 640                 if (reshape_ring(ioat, ioat->alloc_order + 1) &&
 641                     ioat2_ring_space(ioat) > num_descs)
 642                                 break;
 643
 644                 if (printk_ratelimit())
 645                         dev_dbg(to_dev(chan),
 646                                 "%s: ring full! num_descs: %d (%x:%x:%x)\n",
 647                                 __func__, num_descs, ioat->head, ioat->tail,
 648                                 ioat->issued);
 649                 spin_unlock_bh(&ioat->ring_lock);
 650
 651                 /* progress reclaim in the allocation failure case we
 652                  * may be called under bh_disabled so we need to trigger
 653                  * the timer event directly
 654                  */
 655                 spin_lock_bh(&chan->cleanup_lock);
 656                 if (jiffies > chan->timer.expires &&
 657                     timer_pending(&chan->timer)) {
 658                         struct ioatdma_device *device = chan->device;
 659
 660                         mod_timer(&chan->timer, jiffies + COMPLETION_TIMEOUT);
 661                         spin_unlock_bh(&chan->cleanup_lock);
 662                         device->timer_fn((unsigned long) ioat);
 663                 } else
 664                         spin_unlock_bh(&chan->cleanup_lock);
 665                 return -ENOMEM;
 666         }
 667
 668         dev_dbg(to_dev(chan), "%s: num_descs: %d (%x:%x:%x)\n",
 669                 __func__, num_descs, ioat->head, ioat->tail, ioat->issued);
 670
 671         *idx = ioat2_desc_alloc(ioat, num_descs);
 672         return 0;  /* with ioat->ring_lock held */
 673 }
 674
 675 struct dma_async_tx_descriptor *
 676 ioat2_dma_prep_memcpy_lock(struct dma_chan *c, dma_addr_t dma_dest,
 677                            dma_addr_t dma_src, size_t len, unsigned long flags)
 678 {
 679         struct ioat2_dma_chan *ioat = to_ioat2_chan(c);
 680         struct ioat_dma_descriptor *hw;
 681         struct ioat_ring_ent *desc;
 682         dma_addr_t dst = dma_dest;
 683         dma_addr_t src = dma_src;
 684         size_t total_len = len;
 685         int num_descs;
 686         u16 idx;
 687         int i;
 688
 689         num_descs = ioat2_xferlen_to_descs(ioat, len);
 690         if (likely(num_descs) &&
 691             ioat2_alloc_and_lock(&idx, ioat, num_descs) == 0)
 692                 /* pass */;
 693         else
 694                 return NULL;
 695         i = 0;
 696         do {
 697                 size_t copy = min_t(size_t, len, 1 << ioat->xfercap_log);
 698
 699                 desc = ioat2_get_ring_ent(ioat, idx + i);
 700                 hw = desc->hw;
 701
 702                 hw->size = copy;
 703                 hw->ctl = 0;
 704                 hw->src_addr = src;
 705                 hw->dst_addr = dst;
 706
 707                 len -= copy;
 708                 dst += copy;
 709                 src += copy;
 710                 dump_desc_dbg(ioat, desc);
 711         } while (++i < num_descs);
 712
 713         desc->txd.flags = flags;
 714         desc->len = total_len;
 715         hw->ctl_f.int_en = !!(flags & DMA_PREP_INTERRUPT);
 716         hw->ctl_f.fence = !!(flags & DMA_PREP_FENCE);
 717         hw->ctl_f.compl_write = 1;
 718         dump_desc_dbg(ioat, desc);
 719         /* we leave the channel locked to ensure in order submission */
 720
 721         return &desc->txd;
 722 }
 723
 724 /**
 725  * ioat2_free_chan_resources - release all the descriptors
 726  * @chan: the channel to be cleaned
 727  */
 728 void ioat2_free_chan_resources(struct dma_chan *c)
 729 {
 730         struct ioat2_dma_chan *ioat = to_ioat2_chan(c);
 731         struct ioat_chan_common *chan = &ioat->base;
 732         struct ioatdma_device *device = chan->device;
 733         struct ioat_ring_ent *desc;
 734         const u16 total_descs = 1 << ioat->alloc_order;
 735         int descs;
 736         int i;
 737
 738         /* Before freeing channel resources first check
 739          * if they have been previously allocated for this channel.
 740          */
 741         if (!ioat->ring)
 742                 return;
 743
 744         tasklet_disable(&chan->cleanup_task);
 745         del_timer_sync(&chan->timer);
 746         device->cleanup_tasklet((unsigned long) ioat);
 747
 748         /* Delay 100ms after reset to allow internal DMA logic to quiesce
 749          * before removing DMA descriptor resources.
 750          */
 751         writeb(IOAT_CHANCMD_RESET,
 752                chan->reg_base + IOAT_CHANCMD_OFFSET(chan->device->version));
 753         mdelay(100);
 754
 755         spin_lock_bh(&ioat->ring_lock);
 756         descs = ioat2_ring_space(ioat);
 757         dev_dbg(to_dev(chan), "freeing %d idle descriptors\n", descs);
 758         for (i = 0; i < descs; i++) {
 759                 desc = ioat2_get_ring_ent(ioat, ioat->head + i);
 760                 ioat2_free_ring_ent(desc, c);
 761         }
 762
 763         if (descs < total_descs)
 764                 dev_err(to_dev(chan), "Freeing %d in use descriptors!\n",
 765                         total_descs - descs);
 766
 767         for (i = 0; i < total_descs - descs; i++) {
 768                 desc = ioat2_get_ring_ent(ioat, ioat->tail + i);
 769                 dump_desc_dbg(ioat, desc);
 770                 ioat2_free_ring_ent(desc, c);
 771         }
 772
 773         kfree(ioat->ring);
 774         ioat->ring = NULL;
 775         ioat->alloc_order = 0;
 776         pci_pool_free(device->completion_pool, chan->completion,
 777                       chan->completion_dma);
 778         spin_unlock_bh(&ioat->ring_lock);
 779
 780         chan->last_completion = 0;
 781         chan->completion_dma = 0;
 782         ioat->pending = 0;
 783         ioat->dmacount = 0;
 784 }
 785
 786 enum dma_status
 787 ioat2_is_complete(struct dma_chan *c, dma_cookie_t cookie,
 788                      dma_cookie_t *done, dma_cookie_t *used)
 789 {
 790         struct ioat2_dma_chan *ioat = to_ioat2_chan(c);
 791         struct ioatdma_device *device = ioat->base.device;
 792
 793         if (ioat_is_complete(c, cookie, done, used) == DMA_SUCCESS)
 794                 return DMA_SUCCESS;
 795
 796         device->cleanup_tasklet((unsigned long) ioat);
 797
 798         return ioat_is_complete(c, cookie, done, used);
 799 }
 800
 801 static ssize_t ring_size_show(struct dma_chan *c, char *page)
 802 {
 803         struct ioat2_dma_chan *ioat = to_ioat2_chan(c);
 804
 805         return sprintf(page, "%d\n", (1 << ioat->alloc_order) & ~1);
 806 }
 807 static struct ioat_sysfs_entry ring_size_attr = __ATTR_RO(ring_size);
 808
 809 static ssize_t ring_active_show(struct dma_chan *c, char *page)
 810 {
 811         struct ioat2_dma_chan *ioat = to_ioat2_chan(c);
 812
 813         /* ...taken outside the lock, no need to be precise */
 814         return sprintf(page, "%d\n", ioat2_ring_active(ioat));
 815 }
 816 static struct ioat_sysfs_entry ring_active_attr = __ATTR_RO(ring_active);
 817
 818 static struct attribute *ioat2_attrs[] = {
 819         &ring_size_attr.attr,
 820         &ring_active_attr.attr,
 821         &ioat_cap_attr.attr,
 822         &ioat_version_attr.attr,
 823         NULL,
 824 };
 825
 826 struct kobj_type ioat2_ktype = {
 827         .sysfs_ops = &ioat_sysfs_ops,
 828         .default_attrs = ioat2_attrs,
 829 };
 830
 831 int __devinit ioat2_dma_probe(struct ioatdma_device *device, int dca)
 832 {
 833         struct pci_dev *pdev = device->pdev;
 834         struct dma_device *dma;
 835         struct dma_chan *c;
 836         struct ioat_chan_common *chan;
 837         int err;
 838
 839         device->enumerate_channels = ioat2_enumerate_channels;
 840         device->cleanup_tasklet = ioat2_cleanup_tasklet;
 841         device->timer_fn = ioat2_timer_event;
 842         device->self_test = ioat_dma_self_test;
 843         dma = &device->common;
 844         dma->device_prep_dma_memcpy = ioat2_dma_prep_memcpy_lock;
 845         dma->device_issue_pending = ioat2_issue_pending;
 846         dma->device_alloc_chan_resources = ioat2_alloc_chan_resources;
 847         dma->device_free_chan_resources = ioat2_free_chan_resources;
 848         dma->device_is_tx_complete = ioat2_is_complete;
 849
 850         err = ioat_probe(device);
 851         if (err)
 852                 return err;
 853         ioat_set_tcp_copy_break(2048);
 854
 855         list_for_each_entry(c, &dma->channels, device_node) {
 856                 chan = to_chan_common(c);
 857                 writel(IOAT_DCACTRL_CMPL_WRITE_ENABLE | IOAT_DMA_DCA_ANY_CPU,
 858                        chan->reg_base + IOAT_DCACTRL_OFFSET);
 859         }
 860
 861         err = ioat_register(device);
 862         if (err)
 863                 return err;
 864
 865         ioat_kobject_add(device, &ioat2_ktype);
 866
 867         if (dca)
 868                 device->dca = ioat2_dca_init(pdev, device->reg_base);
 869
 870         return err;
 871 }