git.oblomov.eu Git - linux-2.6/blob - 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 "ioatdma.h"
  36 #include "ioatdma_registers.h"
  37 #include "ioatdma_hw.h"
  38
  39 #define INITIAL_IOAT_DESC_COUNT 128
  40
  41 #define to_ioat_chan(chan) container_of(chan, struct ioat_dma_chan, common)
  42 #define to_ioatdma_device(dev) container_of(dev, struct ioatdma_device, common)
  43 #define to_ioat_desc(lh) container_of(lh, struct ioat_desc_sw, node)
  44 #define tx_to_ioat_desc(tx) container_of(tx, struct ioat_desc_sw, async_tx)
  45
  46 /* internal functions */
  47 static void ioat_dma_start_null_desc(struct ioat_dma_chan *ioat_chan);
  48 static void ioat_dma_memcpy_cleanup(struct ioat_dma_chan *ioat_chan);
  49 static struct ioat_desc_sw *
  50 ioat_dma_get_next_descriptor(struct ioat_dma_chan *ioat_chan);
  51
  52 static inline struct ioat_dma_chan *ioat_lookup_chan_by_index(
  53                                                 struct ioatdma_device *device,
  54                                                 int index)
  55 {
  56         return device->idx[index];
  57 }
  58
  59 /**
  60  * ioat_dma_do_interrupt - handler used for single vector interrupt mode
  61  * @irq: interrupt id
  62  * @data: interrupt data
  63  */
  64 static irqreturn_t ioat_dma_do_interrupt(int irq, void *data)
  65 {
  66         struct ioatdma_device *instance = data;
  67         struct ioat_dma_chan *ioat_chan;
  68         unsigned long attnstatus;
  69         int bit;
  70         u8 intrctrl;
  71
  72         intrctrl = readb(instance->reg_base + IOAT_INTRCTRL_OFFSET);
  73
  74         if (!(intrctrl & IOAT_INTRCTRL_MASTER_INT_EN))
  75                 return IRQ_NONE;
  76
  77         if (!(intrctrl & IOAT_INTRCTRL_INT_STATUS)) {
  78                 writeb(intrctrl, instance->reg_base + IOAT_INTRCTRL_OFFSET);
  79                 return IRQ_NONE;
  80         }
  81
  82         attnstatus = readl(instance->reg_base + IOAT_ATTNSTATUS_OFFSET);
  83         for_each_bit(bit, &attnstatus, BITS_PER_LONG) {
  84                 ioat_chan = ioat_lookup_chan_by_index(instance, bit);
  85                 tasklet_schedule(&ioat_chan->cleanup_task);
  86         }
  87
  88         writeb(intrctrl, instance->reg_base + IOAT_INTRCTRL_OFFSET);
  89         return IRQ_HANDLED;
  90 }
  91
  92 /**
  93  * ioat_dma_do_interrupt_msix - handler used for vector-per-channel interrupt mode
  94  * @irq: interrupt id
  95  * @data: interrupt data
  96  */
  97 static irqreturn_t ioat_dma_do_interrupt_msix(int irq, void *data)
  98 {
  99         struct ioat_dma_chan *ioat_chan = data;
 100
 101         tasklet_schedule(&ioat_chan->cleanup_task);
 102
 103         return IRQ_HANDLED;
 104 }
 105
 106 static void ioat_dma_cleanup_tasklet(unsigned long data);
 107
 108 /**
 109  * ioat_dma_enumerate_channels - find and initialize the device's channels
 110  * @device: the device to be enumerated
 111  */
 112 static int ioat_dma_enumerate_channels(struct ioatdma_device *device)
 113 {
 114         u8 xfercap_scale;
 115         u32 xfercap;
 116         int i;
 117         struct ioat_dma_chan *ioat_chan;
 118
 119         device->common.chancnt = readb(device->reg_base + IOAT_CHANCNT_OFFSET);
 120         xfercap_scale = readb(device->reg_base + IOAT_XFERCAP_OFFSET);
 121         xfercap = (xfercap_scale == 0 ? -1 : (1UL << xfercap_scale));
 122
 123         for (i = 0; i < device->common.chancnt; i++) {
 124                 ioat_chan = kzalloc(sizeof(*ioat_chan), GFP_KERNEL);
 125                 if (!ioat_chan) {
 126                         device->common.chancnt = i;
 127                         break;
 128                 }
 129
 130                 ioat_chan->device = device;
 131                 ioat_chan->reg_base = device->reg_base + (0x80 * (i + 1));
 132                 ioat_chan->xfercap = xfercap;
 133                 spin_lock_init(&ioat_chan->cleanup_lock);
 134                 spin_lock_init(&ioat_chan->desc_lock);
 135                 INIT_LIST_HEAD(&ioat_chan->free_desc);
 136                 INIT_LIST_HEAD(&ioat_chan->used_desc);
 137                 /* This should be made common somewhere in dmaengine.c */
 138                 ioat_chan->common.device = &device->common;
 139                 list_add_tail(&ioat_chan->common.device_node,
 140                               &device->common.channels);
 141                 device->idx[i] = ioat_chan;
 142                 tasklet_init(&ioat_chan->cleanup_task,
 143                              ioat_dma_cleanup_tasklet,
 144                              (unsigned long) ioat_chan);
 145                 tasklet_disable(&ioat_chan->cleanup_task);
 146         }
 147         return device->common.chancnt;
 148 }
 149
 150 static void ioat_set_src(dma_addr_t addr,
 151                          struct dma_async_tx_descriptor *tx,
 152                          int index)
 153 {
 154         tx_to_ioat_desc(tx)->src = addr;
 155 }
 156
 157 static void ioat_set_dest(dma_addr_t addr,
 158                           struct dma_async_tx_descriptor *tx,
 159                           int index)
 160 {
 161         tx_to_ioat_desc(tx)->dst = addr;
 162 }
 163
 164 static dma_cookie_t ioat_tx_submit(struct dma_async_tx_descriptor *tx)
 165 {
 166         struct ioat_dma_chan *ioat_chan = to_ioat_chan(tx->chan);
 167         struct ioat_desc_sw *first = tx_to_ioat_desc(tx);
 168         struct ioat_desc_sw *prev, *new;
 169         struct ioat_dma_descriptor *hw;
 170         int append = 0;
 171         dma_cookie_t cookie;
 172         LIST_HEAD(new_chain);
 173         u32 copy;
 174         size_t len;
 175         dma_addr_t src, dst;
 176         int orig_ack;
 177         unsigned int desc_count = 0;
 178
 179         /* src and dest and len are stored in the initial descriptor */
 180         len = first->len;
 181         src = first->src;
 182         dst = first->dst;
 183         orig_ack = first->async_tx.ack;
 184         new = first;
 185
 186         spin_lock_bh(&ioat_chan->desc_lock);
 187         prev = to_ioat_desc(ioat_chan->used_desc.prev);
 188         prefetch(prev->hw);
 189         do {
 190                 copy = min((u32) len, ioat_chan->xfercap);
 191
 192                 new->async_tx.ack = 1;
 193
 194                 hw = new->hw;
 195                 hw->size = copy;
 196                 hw->ctl = 0;
 197                 hw->src_addr = src;
 198                 hw->dst_addr = dst;
 199                 hw->next = 0;
 200
 201                 /* chain together the physical address list for the HW */
 202                 wmb();
 203                 prev->hw->next = (u64) new->async_tx.phys;
 204
 205                 len -= copy;
 206                 dst += copy;
 207                 src += copy;
 208
 209                 list_add_tail(&new->node, &new_chain);
 210                 desc_count++;
 211                 prev = new;
 212         } while (len && (new = ioat_dma_get_next_descriptor(ioat_chan)));
 213
 214         hw->ctl = IOAT_DMA_DESCRIPTOR_CTL_CP_STS;
 215         if (new->async_tx.callback) {
 216                 hw->ctl |= IOAT_DMA_DESCRIPTOR_CTL_INT_GN;
 217                 if (first != new) {
 218                         /* move callback into to last desc */
 219                         new->async_tx.callback = first->async_tx.callback;
 220                         new->async_tx.callback_param
 221                                         = first->async_tx.callback_param;
 222                         first->async_tx.callback = NULL;
 223                         first->async_tx.callback_param = NULL;
 224                 }
 225         }
 226
 227         new->tx_cnt = desc_count;
 228         new->async_tx.ack = orig_ack; /* client is in control of this ack */
 229
 230         /* store the original values for use in later cleanup */
 231         if (new != first) {
 232                 new->src = first->src;
 233                 new->dst = first->dst;
 234                 new->len = first->len;
 235         }
 236
 237         /* cookie incr and addition to used_list must be atomic */
 238         cookie = ioat_chan->common.cookie;
 239         cookie++;
 240         if (cookie < 0)
 241                 cookie = 1;
 242         ioat_chan->common.cookie = new->async_tx.cookie = cookie;
 243
 244         /* write address into NextDescriptor field of last desc in chain */
 245         to_ioat_desc(ioat_chan->used_desc.prev)->hw->next =
 246                                                         first->async_tx.phys;
 247         __list_splice(&new_chain, ioat_chan->used_desc.prev);
 248
 249         ioat_chan->pending += desc_count;
 250         if (ioat_chan->pending >= 4) {
 251                 append = 1;
 252                 ioat_chan->pending = 0;
 253         }
 254         spin_unlock_bh(&ioat_chan->desc_lock);
 255
 256         if (append)
 257                 writeb(IOAT_CHANCMD_APPEND,
 258                         ioat_chan->reg_base + IOAT_CHANCMD_OFFSET);
 259
 260         return cookie;
 261 }
 262
 263 static struct ioat_desc_sw *ioat_dma_alloc_descriptor(
 264                                         struct ioat_dma_chan *ioat_chan,
 265                                         gfp_t flags)
 266 {
 267         struct ioat_dma_descriptor *desc;
 268         struct ioat_desc_sw *desc_sw;
 269         struct ioatdma_device *ioatdma_device;
 270         dma_addr_t phys;
 271
 272         ioatdma_device = to_ioatdma_device(ioat_chan->common.device);
 273         desc = pci_pool_alloc(ioatdma_device->dma_pool, flags, &phys);
 274         if (unlikely(!desc))
 275                 return NULL;
 276
 277         desc_sw = kzalloc(sizeof(*desc_sw), flags);
 278         if (unlikely(!desc_sw)) {
 279                 pci_pool_free(ioatdma_device->dma_pool, desc, phys);
 280                 return NULL;
 281         }
 282
 283         memset(desc, 0, sizeof(*desc));
 284         dma_async_tx_descriptor_init(&desc_sw->async_tx, &ioat_chan->common);
 285         desc_sw->async_tx.tx_set_src = ioat_set_src;
 286         desc_sw->async_tx.tx_set_dest = ioat_set_dest;
 287         desc_sw->async_tx.tx_submit = ioat_tx_submit;
 288         INIT_LIST_HEAD(&desc_sw->async_tx.tx_list);
 289         desc_sw->hw = desc;
 290         desc_sw->async_tx.phys = phys;
 291
 292         return desc_sw;
 293 }
 294
 295 /* returns the actual number of allocated descriptors */
 296 static int ioat_dma_alloc_chan_resources(struct dma_chan *chan)
 297 {
 298         struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
 299         struct ioat_desc_sw *desc = NULL;
 300         u16 chanctrl;
 301         u32 chanerr;
 302         int i;
 303         LIST_HEAD(tmp_list);
 304
 305         /* have we already been set up? */
 306         if (!list_empty(&ioat_chan->free_desc))
 307                 return INITIAL_IOAT_DESC_COUNT;
 308
 309         /* Setup register to interrupt and write completion status on error */
 310         chanctrl = IOAT_CHANCTRL_ERR_INT_EN |
 311                 IOAT_CHANCTRL_ANY_ERR_ABORT_EN |
 312                 IOAT_CHANCTRL_ERR_COMPLETION_EN;
 313         writew(chanctrl, ioat_chan->reg_base + IOAT_CHANCTRL_OFFSET);
 314
 315         chanerr = readl(ioat_chan->reg_base + IOAT_CHANERR_OFFSET);
 316         if (chanerr) {
 317                 dev_err(&ioat_chan->device->pdev->dev,
 318                         "CHANERR = %x, clearing\n", chanerr);
 319                 writel(chanerr, ioat_chan->reg_base + IOAT_CHANERR_OFFSET);
 320         }
 321
 322         /* Allocate descriptors */
 323         for (i = 0; i < INITIAL_IOAT_DESC_COUNT; i++) {
 324                 desc = ioat_dma_alloc_descriptor(ioat_chan, GFP_KERNEL);
 325                 if (!desc) {
 326                         dev_err(&ioat_chan->device->pdev->dev,
 327                                 "Only %d initial descriptors\n", i);
 328                         break;
 329                 }
 330                 list_add_tail(&desc->node, &tmp_list);
 331         }
 332         spin_lock_bh(&ioat_chan->desc_lock);
 333         list_splice(&tmp_list, &ioat_chan->free_desc);
 334         spin_unlock_bh(&ioat_chan->desc_lock);
 335
 336         /* allocate a completion writeback area */
 337         /* doing 2 32bit writes to mmio since 1 64b write doesn't work */
 338         ioat_chan->completion_virt =
 339                 pci_pool_alloc(ioat_chan->device->completion_pool,
 340                                GFP_KERNEL,
 341                                &ioat_chan->completion_addr);
 342         memset(ioat_chan->completion_virt, 0,
 343                sizeof(*ioat_chan->completion_virt));
 344         writel(((u64) ioat_chan->completion_addr) & 0x00000000FFFFFFFF,
 345                ioat_chan->reg_base + IOAT_CHANCMP_OFFSET_LOW);
 346         writel(((u64) ioat_chan->completion_addr) >> 32,
 347                ioat_chan->reg_base + IOAT_CHANCMP_OFFSET_HIGH);
 348
 349         tasklet_enable(&ioat_chan->cleanup_task);
 350         ioat_dma_start_null_desc(ioat_chan);
 351         return i;
 352 }
 353
 354 static void ioat_dma_free_chan_resources(struct dma_chan *chan)
 355 {
 356         struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
 357         struct ioatdma_device *ioatdma_device = to_ioatdma_device(chan->device);
 358         struct ioat_desc_sw *desc, *_desc;
 359         int in_use_descs = 0;
 360
 361         tasklet_disable(&ioat_chan->cleanup_task);
 362         ioat_dma_memcpy_cleanup(ioat_chan);
 363
 364         /* Delay 100ms after reset to allow internal DMA logic to quiesce
 365          * before removing DMA descriptor resources.
 366          */
 367         writeb(IOAT_CHANCMD_RESET, ioat_chan->reg_base + IOAT_CHANCMD_OFFSET);
 368         mdelay(100);
 369
 370         spin_lock_bh(&ioat_chan->desc_lock);
 371         list_for_each_entry_safe(desc, _desc, &ioat_chan->used_desc, node) {
 372                 in_use_descs++;
 373                 list_del(&desc->node);
 374                 pci_pool_free(ioatdma_device->dma_pool, desc->hw,
 375                               desc->async_tx.phys);
 376                 kfree(desc);
 377         }
 378         list_for_each_entry_safe(desc, _desc, &ioat_chan->free_desc, node) {
 379                 list_del(&desc->node);
 380                 pci_pool_free(ioatdma_device->dma_pool, desc->hw,
 381                               desc->async_tx.phys);
 382                 kfree(desc);
 383         }
 384         spin_unlock_bh(&ioat_chan->desc_lock);
 385
 386         pci_pool_free(ioatdma_device->completion_pool,
 387                       ioat_chan->completion_virt,
 388                       ioat_chan->completion_addr);
 389
 390         /* one is ok since we left it on there on purpose */
 391         if (in_use_descs > 1)
 392                 dev_err(&ioat_chan->device->pdev->dev,
 393                         "Freeing %d in use descriptors!\n",
 394                         in_use_descs - 1);
 395
 396         ioat_chan->last_completion = ioat_chan->completion_addr = 0;
 397         ioat_chan->pending = 0;
 398 }
 399
 400 /**
 401  * ioat_dma_get_next_descriptor - return the next available descriptor
 402  * @ioat_chan: IOAT DMA channel handle
 403  *
 404  * Gets the next descriptor from the chain, and must be called with the
 405  * channel's desc_lock held.  Allocates more descriptors if the channel
 406  * has run out.
 407  */
 408 static struct ioat_desc_sw *
 409 ioat_dma_get_next_descriptor(struct ioat_dma_chan *ioat_chan)
 410 {
 411         struct ioat_desc_sw *new = NULL;
 412
 413         if (!list_empty(&ioat_chan->free_desc)) {
 414                 new = to_ioat_desc(ioat_chan->free_desc.next);
 415                 list_del(&new->node);
 416         } else {
 417                 /* try to get another desc */
 418                 new = ioat_dma_alloc_descriptor(ioat_chan, GFP_ATOMIC);
 419                 /* will this ever happen? */
 420                 /* TODO add upper limit on these */
 421                 BUG_ON(!new);
 422         }
 423
 424         prefetch(new->hw);
 425         return new;
 426 }
 427
 428 static struct dma_async_tx_descriptor *ioat_dma_prep_memcpy(
 429                                                 struct dma_chan *chan,
 430                                                 size_t len,
 431                                                 int int_en)
 432 {
 433         struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
 434         struct ioat_desc_sw *new;
 435
 436         spin_lock_bh(&ioat_chan->desc_lock);
 437         new = ioat_dma_get_next_descriptor(ioat_chan);
 438         new->len = len;
 439         spin_unlock_bh(&ioat_chan->desc_lock);
 440
 441         return new ? &new->async_tx : NULL;
 442 }
 443
 444 /**
 445  * ioat_dma_memcpy_issue_pending - push potentially unrecognized appended
 446  *                                 descriptors to hw
 447  * @chan: DMA channel handle
 448  */
 449 static void ioat_dma_memcpy_issue_pending(struct dma_chan *chan)
 450 {
 451         struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
 452
 453         if (ioat_chan->pending != 0) {
 454                 ioat_chan->pending = 0;
 455                 writeb(IOAT_CHANCMD_APPEND,
 456                        ioat_chan->reg_base + IOAT_CHANCMD_OFFSET);
 457         }
 458 }
 459
 460 static void ioat_dma_cleanup_tasklet(unsigned long data)
 461 {
 462         struct ioat_dma_chan *chan = (void *)data;
 463         ioat_dma_memcpy_cleanup(chan);
 464         writew(IOAT_CHANCTRL_INT_DISABLE,
 465                chan->reg_base + IOAT_CHANCTRL_OFFSET);
 466 }
 467
 468 static void ioat_dma_memcpy_cleanup(struct ioat_dma_chan *ioat_chan)
 469 {
 470         unsigned long phys_complete;
 471         struct ioat_desc_sw *desc, *_desc;
 472         dma_cookie_t cookie = 0;
 473
 474         prefetch(ioat_chan->completion_virt);
 475
 476         if (!spin_trylock_bh(&ioat_chan->cleanup_lock))
 477                 return;
 478
 479         /* The completion writeback can happen at any time,
 480            so reads by the driver need to be atomic operations
 481            The descriptor physical addresses are limited to 32-bits
 482            when the CPU can only do a 32-bit mov */
 483
 484 #if (BITS_PER_LONG == 64)
 485         phys_complete =
 486                 ioat_chan->completion_virt->full
 487                 & IOAT_CHANSTS_COMPLETED_DESCRIPTOR_ADDR;
 488 #else
 489         phys_complete =
 490                 ioat_chan->completion_virt->low & IOAT_LOW_COMPLETION_MASK;
 491 #endif
 492
 493         if ((ioat_chan->completion_virt->full
 494                 & IOAT_CHANSTS_DMA_TRANSFER_STATUS) ==
 495                                 IOAT_CHANSTS_DMA_TRANSFER_STATUS_HALTED) {
 496                 dev_err(&ioat_chan->device->pdev->dev,
 497                         "Channel halted, chanerr = %x\n",
 498                         readl(ioat_chan->reg_base + IOAT_CHANERR_OFFSET));
 499
 500                 /* TODO do something to salvage the situation */
 501         }
 502
 503         if (phys_complete == ioat_chan->last_completion) {
 504                 spin_unlock_bh(&ioat_chan->cleanup_lock);
 505                 return;
 506         }
 507
 508         cookie = 0;
 509         spin_lock_bh(&ioat_chan->desc_lock);
 510         list_for_each_entry_safe(desc, _desc, &ioat_chan->used_desc, node) {
 511
 512                 /*
 513                  * Incoming DMA requests may use multiple descriptors, due to
 514                  * exceeding xfercap, perhaps. If so, only the last one will
 515                  * have a cookie, and require unmapping.
 516                  */
 517                 if (desc->async_tx.cookie) {
 518                         cookie = desc->async_tx.cookie;
 519
 520                         /*
 521                          * yes we are unmapping both _page and _single alloc'd
 522                          * regions with unmap_page. Is this *really* that bad?
 523                          */
 524                         pci_unmap_page(ioat_chan->device->pdev,
 525                                         pci_unmap_addr(desc, dst),
 526                                         pci_unmap_len(desc, len),
 527                                         PCI_DMA_FROMDEVICE);
 528                         pci_unmap_page(ioat_chan->device->pdev,
 529                                         pci_unmap_addr(desc, src),
 530                                         pci_unmap_len(desc, len),
 531                                         PCI_DMA_TODEVICE);
 532                         if (desc->async_tx.callback) {
 533                                 desc->async_tx.callback(
 534                                                 desc->async_tx.callback_param);
 535                                 desc->async_tx.callback = NULL;
 536                         }
 537                 }
 538
 539                 if (desc->async_tx.phys != phys_complete) {
 540                         /*
 541                          * a completed entry, but not the last, so cleanup
 542                          * if the client is done with the descriptor
 543                          */
 544                         if (desc->async_tx.ack) {
 545                                 list_del(&desc->node);
 546                                 list_add_tail(&desc->node,
 547                                               &ioat_chan->free_desc);
 548                         } else
 549                                 desc->async_tx.cookie = 0;
 550                 } else {
 551                         /*
 552                          * last used desc. Do not remove, so we can append from
 553                          * it, but don't look at it next time, either
 554                          */
 555                         desc->async_tx.cookie = 0;
 556
 557                         /* TODO check status bits? */
 558                         break;
 559                 }
 560         }
 561
 562         spin_unlock_bh(&ioat_chan->desc_lock);
 563
 564         ioat_chan->last_completion = phys_complete;
 565         if (cookie != 0)
 566                 ioat_chan->completed_cookie = cookie;
 567
 568         spin_unlock_bh(&ioat_chan->cleanup_lock);
 569 }
 570
 571 static void ioat_dma_dependency_added(struct dma_chan *chan)
 572 {
 573         struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
 574         spin_lock_bh(&ioat_chan->desc_lock);
 575         if (ioat_chan->pending == 0) {
 576                 spin_unlock_bh(&ioat_chan->desc_lock);
 577                 ioat_dma_memcpy_cleanup(ioat_chan);
 578         } else
 579                 spin_unlock_bh(&ioat_chan->desc_lock);
 580 }
 581
 582 /**
 583  * ioat_dma_is_complete - poll the status of a IOAT DMA transaction
 584  * @chan: IOAT DMA channel handle
 585  * @cookie: DMA transaction identifier
 586  * @done: if not %NULL, updated with last completed transaction
 587  * @used: if not %NULL, updated with last used transaction
 588  */
 589 static enum dma_status ioat_dma_is_complete(struct dma_chan *chan,
 590                                             dma_cookie_t cookie,
 591                                             dma_cookie_t *done,
 592                                             dma_cookie_t *used)
 593 {
 594         struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
 595         dma_cookie_t last_used;
 596         dma_cookie_t last_complete;
 597         enum dma_status ret;
 598
 599         last_used = chan->cookie;
 600         last_complete = ioat_chan->completed_cookie;
 601
 602         if (done)
 603                 *done = last_complete;
 604         if (used)
 605                 *used = last_used;
 606
 607         ret = dma_async_is_complete(cookie, last_complete, last_used);
 608         if (ret == DMA_SUCCESS)
 609                 return ret;
 610
 611         ioat_dma_memcpy_cleanup(ioat_chan);
 612
 613         last_used = chan->cookie;
 614         last_complete = ioat_chan->completed_cookie;
 615
 616         if (done)
 617                 *done = last_complete;
 618         if (used)
 619                 *used = last_used;
 620
 621         return dma_async_is_complete(cookie, last_complete, last_used);
 622 }
 623
 624 /* PCI API */
 625
 626 static void ioat_dma_start_null_desc(struct ioat_dma_chan *ioat_chan)
 627 {
 628         struct ioat_desc_sw *desc;
 629
 630         spin_lock_bh(&ioat_chan->desc_lock);
 631
 632         desc = ioat_dma_get_next_descriptor(ioat_chan);
 633         desc->hw->ctl = IOAT_DMA_DESCRIPTOR_NUL
 634                                 | IOAT_DMA_DESCRIPTOR_CTL_INT_GN
 635                                 | IOAT_DMA_DESCRIPTOR_CTL_CP_STS;
 636         desc->hw->next = 0;
 637         desc->hw->size = 0;
 638         desc->hw->src_addr = 0;
 639         desc->hw->dst_addr = 0;
 640         desc->async_tx.ack = 1;
 641
 642         list_add_tail(&desc->node, &ioat_chan->used_desc);
 643         spin_unlock_bh(&ioat_chan->desc_lock);
 644
 645         writel(((u64) desc->async_tx.phys) & 0x00000000FFFFFFFF,
 646                ioat_chan->reg_base + IOAT_CHAINADDR_OFFSET_LOW);
 647         writel(((u64) desc->async_tx.phys) >> 32,
 648                ioat_chan->reg_base + IOAT_CHAINADDR_OFFSET_HIGH);
 649
 650         writeb(IOAT_CHANCMD_START, ioat_chan->reg_base + IOAT_CHANCMD_OFFSET);
 651 }
 652
 653 /*
 654  * Perform a IOAT transaction to verify the HW works.
 655  */
 656 #define IOAT_TEST_SIZE 2000
 657
 658 static void ioat_dma_test_callback(void *dma_async_param)
 659 {
 660         printk(KERN_ERR "ioatdma: ioat_dma_test_callback(%p)\n",
 661                         dma_async_param);
 662 }
 663
 664 /**
 665  * ioat_dma_self_test - Perform a IOAT transaction to verify the HW works.
 666  * @device: device to be tested
 667  */
 668 static int ioat_dma_self_test(struct ioatdma_device *device)
 669 {
 670         int i;
 671         u8 *src;
 672         u8 *dest;
 673         struct dma_chan *dma_chan;
 674         struct dma_async_tx_descriptor *tx = NULL;
 675         dma_addr_t addr;
 676         dma_cookie_t cookie;
 677         int err = 0;
 678
 679         src = kzalloc(sizeof(u8) * IOAT_TEST_SIZE, GFP_KERNEL);
 680         if (!src)
 681                 return -ENOMEM;
 682         dest = kzalloc(sizeof(u8) * IOAT_TEST_SIZE, GFP_KERNEL);
 683         if (!dest) {
 684                 kfree(src);
 685                 return -ENOMEM;
 686         }
 687
 688         /* Fill in src buffer */
 689         for (i = 0; i < IOAT_TEST_SIZE; i++)
 690                 src[i] = (u8)i;
 691
 692         /* Start copy, using first DMA channel */
 693         dma_chan = container_of(device->common.channels.next,
 694                                 struct dma_chan,
 695                                 device_node);
 696         if (ioat_dma_alloc_chan_resources(dma_chan) < 1) {
 697                 dev_err(&device->pdev->dev,
 698                         "selftest cannot allocate chan resource\n");
 699                 err = -ENODEV;
 700                 goto out;
 701         }
 702
 703         tx = ioat_dma_prep_memcpy(dma_chan, IOAT_TEST_SIZE, 0);
 704         if (!tx) {
 705                 dev_err(&device->pdev->dev,
 706                         "Self-test prep failed, disabling\n");
 707                 err = -ENODEV;
 708                 goto free_resources;
 709         }
 710
 711         async_tx_ack(tx);
 712         addr = dma_map_single(dma_chan->device->dev, src, IOAT_TEST_SIZE,
 713                         DMA_TO_DEVICE);
 714         ioat_set_src(addr, tx, 0);
 715         addr = dma_map_single(dma_chan->device->dev, dest, IOAT_TEST_SIZE,
 716                         DMA_FROM_DEVICE);
 717         ioat_set_dest(addr, tx, 0);
 718         tx->callback = ioat_dma_test_callback;
 719         tx->callback_param = (void *)0x8086;
 720         cookie = ioat_tx_submit(tx);
 721         if (cookie < 0) {
 722                 dev_err(&device->pdev->dev,
 723                         "Self-test setup failed, disabling\n");
 724                 err = -ENODEV;
 725                 goto free_resources;
 726         }
 727         ioat_dma_memcpy_issue_pending(dma_chan);
 728         msleep(1);
 729
 730         if (ioat_dma_is_complete(dma_chan, cookie, NULL, NULL) != DMA_SUCCESS) {
 731                 dev_err(&device->pdev->dev,
 732                         "Self-test copy timed out, disabling\n");
 733                 err = -ENODEV;
 734                 goto free_resources;
 735         }
 736         if (memcmp(src, dest, IOAT_TEST_SIZE)) {
 737                 dev_err(&device->pdev->dev,
 738                         "Self-test copy failed compare, disabling\n");
 739                 err = -ENODEV;
 740                 goto free_resources;
 741         }
 742
 743 free_resources:
 744         ioat_dma_free_chan_resources(dma_chan);
 745 out:
 746         kfree(src);
 747         kfree(dest);
 748         return err;
 749 }
 750
 751 static char ioat_interrupt_style[32] = "msix";
 752 module_param_string(ioat_interrupt_style, ioat_interrupt_style,
 753                     sizeof(ioat_interrupt_style), 0644);
 754 MODULE_PARM_DESC(ioat_interrupt_style,
 755                  "set ioat interrupt style: msix (default), "
 756                  "msix-single-vector, msi, intx)");
 757
 758 /**
 759  * ioat_dma_setup_interrupts - setup interrupt handler
 760  * @device: ioat device
 761  */
 762 static int ioat_dma_setup_interrupts(struct ioatdma_device *device)
 763 {
 764         struct ioat_dma_chan *ioat_chan;
 765         int err, i, j, msixcnt;
 766         u8 intrctrl = 0;
 767
 768         if (!strcmp(ioat_interrupt_style, "msix"))
 769                 goto msix;
 770         if (!strcmp(ioat_interrupt_style, "msix-single-vector"))
 771                 goto msix_single_vector;
 772         if (!strcmp(ioat_interrupt_style, "msi"))
 773                 goto msi;
 774         if (!strcmp(ioat_interrupt_style, "intx"))
 775                 goto intx;
 776         dev_err(&device->pdev->dev, "invalid ioat_interrupt_style %s\n",
 777                 ioat_interrupt_style);
 778         goto err_no_irq;
 779
 780 msix:
 781         /* The number of MSI-X vectors should equal the number of channels */
 782         msixcnt = device->common.chancnt;
 783         for (i = 0; i < msixcnt; i++)
 784                 device->msix_entries[i].entry = i;
 785
 786         err = pci_enable_msix(device->pdev, device->msix_entries, msixcnt);
 787         if (err < 0)
 788                 goto msi;
 789         if (err > 0)
 790                 goto msix_single_vector;
 791
 792         for (i = 0; i < msixcnt; i++) {
 793                 ioat_chan = ioat_lookup_chan_by_index(device, i);
 794                 err = request_irq(device->msix_entries[i].vector,
 795                                   ioat_dma_do_interrupt_msix,
 796                                   0, "ioat-msix", ioat_chan);
 797                 if (err) {
 798                         for (j = 0; j < i; j++) {
 799                                 ioat_chan =
 800                                         ioat_lookup_chan_by_index(device, j);
 801                                 free_irq(device->msix_entries[j].vector,
 802                                          ioat_chan);
 803                         }
 804                         goto msix_single_vector;
 805                 }
 806         }
 807         intrctrl |= IOAT_INTRCTRL_MSIX_VECTOR_CONTROL;
 808         device->irq_mode = msix_multi_vector;
 809         goto done;
 810
 811 msix_single_vector:
 812         device->msix_entries[0].entry = 0;
 813         err = pci_enable_msix(device->pdev, device->msix_entries, 1);
 814         if (err)
 815                 goto msi;
 816
 817         err = request_irq(device->msix_entries[0].vector, ioat_dma_do_interrupt,
 818                           0, "ioat-msix", device);
 819         if (err) {
 820                 pci_disable_msix(device->pdev);
 821                 goto msi;
 822         }
 823         device->irq_mode = msix_single_vector;
 824         goto done;
 825
 826 msi:
 827         err = pci_enable_msi(device->pdev);
 828         if (err)
 829                 goto intx;
 830
 831         err = request_irq(device->pdev->irq, ioat_dma_do_interrupt,
 832                           0, "ioat-msi", device);
 833         if (err) {
 834                 pci_disable_msi(device->pdev);
 835                 goto intx;
 836         }
 837         /*
 838          * CB 1.2 devices need a bit set in configuration space to enable MSI
 839          */
 840         if (device->version == IOAT_VER_1_2) {
 841                 u32 dmactrl;
 842                 pci_read_config_dword(device->pdev,
 843                                       IOAT_PCI_DMACTRL_OFFSET, &dmactrl);
 844                 dmactrl |= IOAT_PCI_DMACTRL_MSI_EN;
 845                 pci_write_config_dword(device->pdev,
 846                                        IOAT_PCI_DMACTRL_OFFSET, dmactrl);
 847         }
 848         device->irq_mode = msi;
 849         goto done;
 850
 851 intx:
 852         err = request_irq(device->pdev->irq, ioat_dma_do_interrupt,
 853                           IRQF_SHARED, "ioat-intx", device);
 854         if (err)
 855                 goto err_no_irq;
 856         device->irq_mode = intx;
 857
 858 done:
 859         intrctrl |= IOAT_INTRCTRL_MASTER_INT_EN;
 860         writeb(intrctrl, device->reg_base + IOAT_INTRCTRL_OFFSET);
 861         return 0;
 862
 863 err_no_irq:
 864         /* Disable all interrupt generation */
 865         writeb(0, device->reg_base + IOAT_INTRCTRL_OFFSET);
 866         dev_err(&device->pdev->dev, "no usable interrupts\n");
 867         device->irq_mode = none;
 868         return -1;
 869 }
 870
 871 /**
 872  * ioat_dma_remove_interrupts - remove whatever interrupts were set
 873  * @device: ioat device
 874  */
 875 static void ioat_dma_remove_interrupts(struct ioatdma_device *device)
 876 {
 877         struct ioat_dma_chan *ioat_chan;
 878         int i;
 879
 880         /* Disable all interrupt generation */
 881         writeb(0, device->reg_base + IOAT_INTRCTRL_OFFSET);
 882
 883         switch (device->irq_mode) {
 884         case msix_multi_vector:
 885                 for (i = 0; i < device->common.chancnt; i++) {
 886                         ioat_chan = ioat_lookup_chan_by_index(device, i);
 887                         free_irq(device->msix_entries[i].vector, ioat_chan);
 888                 }
 889                 pci_disable_msix(device->pdev);
 890                 break;
 891         case msix_single_vector:
 892                 free_irq(device->msix_entries[0].vector, device);
 893                 pci_disable_msix(device->pdev);
 894                 break;
 895         case msi:
 896                 free_irq(device->pdev->irq, device);
 897                 pci_disable_msi(device->pdev);
 898                 break;
 899         case intx:
 900                 free_irq(device->pdev->irq, device);
 901                 break;
 902         case none:
 903                 dev_warn(&device->pdev->dev,
 904                          "call to %s without interrupts setup\n", __func__);
 905         }
 906         device->irq_mode = none;
 907 }
 908
 909 struct ioatdma_device *ioat_dma_probe(struct pci_dev *pdev,
 910                                       void __iomem *iobase)
 911 {
 912         int err;
 913         struct ioatdma_device *device;
 914
 915         device = kzalloc(sizeof(*device), GFP_KERNEL);
 916         if (!device) {
 917                 err = -ENOMEM;
 918                 goto err_kzalloc;
 919         }
 920         device->pdev = pdev;
 921         device->reg_base = iobase;
 922         device->version = readb(device->reg_base + IOAT_VER_OFFSET);
 923
 924         /* DMA coherent memory pool for DMA descriptor allocations */
 925         device->dma_pool = pci_pool_create("dma_desc_pool", pdev,
 926                                            sizeof(struct ioat_dma_descriptor),
 927                                            64, 0);
 928         if (!device->dma_pool) {
 929                 err = -ENOMEM;
 930                 goto err_dma_pool;
 931         }
 932
 933         device->completion_pool = pci_pool_create("completion_pool", pdev,
 934                                                   sizeof(u64), SMP_CACHE_BYTES,
 935                                                   SMP_CACHE_BYTES);
 936         if (!device->completion_pool) {
 937                 err = -ENOMEM;
 938                 goto err_completion_pool;
 939         }
 940
 941         INIT_LIST_HEAD(&device->common.channels);
 942         ioat_dma_enumerate_channels(device);
 943
 944         dma_cap_set(DMA_MEMCPY, device->common.cap_mask);
 945         device->common.device_alloc_chan_resources =
 946                                                 ioat_dma_alloc_chan_resources;
 947         device->common.device_free_chan_resources =
 948                                                 ioat_dma_free_chan_resources;
 949         device->common.device_prep_dma_memcpy = ioat_dma_prep_memcpy;
 950         device->common.device_is_tx_complete = ioat_dma_is_complete;
 951         device->common.device_issue_pending = ioat_dma_memcpy_issue_pending;
 952         device->common.device_dependency_added = ioat_dma_dependency_added;
 953         device->common.dev = &pdev->dev;
 954         dev_err(&device->pdev->dev,
 955                 "Intel(R) I/OAT DMA Engine found,"
 956                 " %d channels, device version 0x%02x, driver version %s\n",
 957                 device->common.chancnt, device->version, IOAT_DMA_VERSION);
 958
 959         err = ioat_dma_setup_interrupts(device);
 960         if (err)
 961                 goto err_setup_interrupts;
 962
 963         err = ioat_dma_self_test(device);
 964         if (err)
 965                 goto err_self_test;
 966
 967         dma_async_device_register(&device->common);
 968
 969         return device;
 970
 971 err_self_test:
 972         ioat_dma_remove_interrupts(device);
 973 err_setup_interrupts:
 974         pci_pool_destroy(device->completion_pool);
 975 err_completion_pool:
 976         pci_pool_destroy(device->dma_pool);
 977 err_dma_pool:
 978         kfree(device);
 979 err_kzalloc:
 980         dev_err(&device->pdev->dev,
 981                 "Intel(R) I/OAT DMA Engine initialization failed\n");
 982         return NULL;
 983 }
 984
 985 void ioat_dma_remove(struct ioatdma_device *device)
 986 {
 987         struct dma_chan *chan, *_chan;
 988         struct ioat_dma_chan *ioat_chan;
 989
 990         ioat_dma_remove_interrupts(device);
 991
 992         dma_async_device_unregister(&device->common);
 993
 994         pci_pool_destroy(device->dma_pool);
 995         pci_pool_destroy(device->completion_pool);
 996
 997         iounmap(device->reg_base);
 998         pci_release_regions(device->pdev);
 999         pci_disable_device(device->pdev);
1000
1001         list_for_each_entry_safe(chan, _chan,
1002                                  &device->common.channels, device_node) {
1003                 ioat_chan = to_ioat_chan(chan);
1004                 list_del(&chan->device_node);
1005                 kfree(ioat_chan);
1006         }
1007         kfree(device);
1008 }
1009