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, which does asynchronous 25 * copy operations. 26 */ 27 28 #include <linux/init.h> 29 #include <linux/module.h> 30 #include <linux/slab.h> 31 #include <linux/pci.h> 32 #include <linux/interrupt.h> 33 #include <linux/dmaengine.h> 34 #include <linux/delay.h> 35 #include <linux/dma-mapping.h> 36 #include <linux/workqueue.h> 37 #include <linux/i7300_idle.h> 38 #include "dma.h" 39 #include "registers.h" 40 #include "hw.h" 41 42 int ioat_pending_level = 4; 43 module_param(ioat_pending_level, int, 0644); 44 MODULE_PARM_DESC(ioat_pending_level, 45 "high-water mark for pushing ioat descriptors (default: 4)"); 46 47 /* internal functions */ 48 static void ioat1_cleanup(struct ioat_dma_chan *ioat); 49 static void ioat1_dma_start_null_desc(struct ioat_dma_chan *ioat); 50 51 /** 52 * ioat_dma_do_interrupt - handler used for single vector interrupt mode 53 * @irq: interrupt id 54 * @data: interrupt data 55 */ 56 static irqreturn_t ioat_dma_do_interrupt(int irq, void *data) 57 { 58 struct ioatdma_device *instance = data; 59 struct ioat_chan_common *chan; 60 unsigned long attnstatus; 61 int bit; 62 u8 intrctrl; 63 64 intrctrl = readb(instance->reg_base + IOAT_INTRCTRL_OFFSET); 65 66 if (!(intrctrl & IOAT_INTRCTRL_MASTER_INT_EN)) 67 return IRQ_NONE; 68 69 if (!(intrctrl & IOAT_INTRCTRL_INT_STATUS)) { 70 writeb(intrctrl, instance->reg_base + IOAT_INTRCTRL_OFFSET); 71 return IRQ_NONE; 72 } 73 74 attnstatus = readl(instance->reg_base + IOAT_ATTNSTATUS_OFFSET); 75 for_each_set_bit(bit, &attnstatus, BITS_PER_LONG) { 76 chan = ioat_chan_by_index(instance, bit); 77 tasklet_schedule(&chan->cleanup_task); 78 } 79 80 writeb(intrctrl, instance->reg_base + IOAT_INTRCTRL_OFFSET); 81 return IRQ_HANDLED; 82 } 83 84 /** 85 * ioat_dma_do_interrupt_msix - handler used for vector-per-channel interrupt mode 86 * @irq: interrupt id 87 * @data: interrupt data 88 */ 89 static irqreturn_t ioat_dma_do_interrupt_msix(int irq, void *data) 90 { 91 struct ioat_chan_common *chan = data; 92 93 tasklet_schedule(&chan->cleanup_task); 94 95 return IRQ_HANDLED; 96 } 97 98 /* common channel initialization */ 99 void ioat_init_channel(struct ioatdma_device *device, struct ioat_chan_common *chan, int idx) 100 { 101 struct dma_device *dma = &device->common; 102 struct dma_chan *c = &chan->common; 103 unsigned long data = (unsigned long) c; 104 105 chan->device = device; 106 chan->reg_base = device->reg_base + (0x80 * (idx + 1)); 107 spin_lock_init(&chan->cleanup_lock); 108 chan->common.device = dma; 109 list_add_tail(&chan->common.device_node, &dma->channels); 110 device->idx[idx] = chan; 111 init_timer(&chan->timer); 112 chan->timer.function = device->timer_fn; 113 chan->timer.data = data; 114 tasklet_init(&chan->cleanup_task, device->cleanup_fn, data); 115 tasklet_disable(&chan->cleanup_task); 116 } 117 118 /** 119 * ioat1_dma_enumerate_channels - find and initialize the device's channels 120 * @device: the device to be enumerated 121 */ 122 static int ioat1_enumerate_channels(struct ioatdma_device *device) 123 { 124 u8 xfercap_scale; 125 u32 xfercap; 126 int i; 127 struct ioat_dma_chan *ioat; 128 struct device *dev = &device->pdev->dev; 129 struct dma_device *dma = &device->common; 130 131 INIT_LIST_HEAD(&dma->channels); 132 dma->chancnt = readb(device->reg_base + IOAT_CHANCNT_OFFSET); 133 dma->chancnt &= 0x1f; /* bits [4:0] valid */ 134 if (dma->chancnt > ARRAY_SIZE(device->idx)) { 135 dev_warn(dev, "(%d) exceeds max supported channels (%zu)\n", 136 dma->chancnt, ARRAY_SIZE(device->idx)); 137 dma->chancnt = ARRAY_SIZE(device->idx); 138 } 139 xfercap_scale = readb(device->reg_base + IOAT_XFERCAP_OFFSET); 140 xfercap_scale &= 0x1f; /* bits [4:0] valid */ 141 xfercap = (xfercap_scale == 0 ? -1 : (1UL << xfercap_scale)); 142 dev_dbg(dev, "%s: xfercap = %d\n", __func__, xfercap); 143 144 #ifdef CONFIG_I7300_IDLE_IOAT_CHANNEL 145 if (i7300_idle_platform_probe(NULL, NULL, 1) == 0) 146 dma->chancnt--; 147 #endif 148 for (i = 0; i < dma->chancnt; i++) { 149 ioat = devm_kzalloc(dev, sizeof(*ioat), GFP_KERNEL); 150 if (!ioat) 151 break; 152 153 ioat_init_channel(device, &ioat->base, i); 154 ioat->xfercap = xfercap; 155 spin_lock_init(&ioat->desc_lock); 156 INIT_LIST_HEAD(&ioat->free_desc); 157 INIT_LIST_HEAD(&ioat->used_desc); 158 } 159 dma->chancnt = i; 160 return i; 161 } 162 163 /** 164 * ioat_dma_memcpy_issue_pending - push potentially unrecognized appended 165 * descriptors to hw 166 * @chan: DMA channel handle 167 */ 168 static inline void 169 __ioat1_dma_memcpy_issue_pending(struct ioat_dma_chan *ioat) 170 { 171 void __iomem *reg_base = ioat->base.reg_base; 172 173 dev_dbg(to_dev(&ioat->base), "%s: pending: %d\n", 174 __func__, ioat->pending); 175 ioat->pending = 0; 176 writeb(IOAT_CHANCMD_APPEND, reg_base + IOAT1_CHANCMD_OFFSET); 177 } 178 179 static void ioat1_dma_memcpy_issue_pending(struct dma_chan *chan) 180 { 181 struct ioat_dma_chan *ioat = to_ioat_chan(chan); 182 183 if (ioat->pending > 0) { 184 spin_lock_bh(&ioat->desc_lock); 185 __ioat1_dma_memcpy_issue_pending(ioat); 186 spin_unlock_bh(&ioat->desc_lock); 187 } 188 } 189 190 /** 191 * ioat1_reset_channel - restart a channel 192 * @ioat: IOAT DMA channel handle 193 */ 194 static void ioat1_reset_channel(struct ioat_dma_chan *ioat) 195 { 196 struct ioat_chan_common *chan = &ioat->base; 197 void __iomem *reg_base = chan->reg_base; 198 u32 chansts, chanerr; 199 200 dev_warn(to_dev(chan), "reset\n"); 201 chanerr = readl(reg_base + IOAT_CHANERR_OFFSET); 202 chansts = *chan->completion & IOAT_CHANSTS_STATUS; 203 if (chanerr) { 204 dev_err(to_dev(chan), 205 "chan%d, CHANSTS = 0x%08x CHANERR = 0x%04x, clearing\n", 206 chan_num(chan), chansts, chanerr); 207 writel(chanerr, reg_base + IOAT_CHANERR_OFFSET); 208 } 209 210 /* 211 * whack it upside the head with a reset 212 * and wait for things to settle out. 213 * force the pending count to a really big negative 214 * to make sure no one forces an issue_pending 215 * while we're waiting. 216 */ 217 218 ioat->pending = INT_MIN; 219 writeb(IOAT_CHANCMD_RESET, 220 reg_base + IOAT_CHANCMD_OFFSET(chan->device->version)); 221 set_bit(IOAT_RESET_PENDING, &chan->state); 222 mod_timer(&chan->timer, jiffies + RESET_DELAY); 223 } 224 225 static dma_cookie_t ioat1_tx_submit(struct dma_async_tx_descriptor *tx) 226 { 227 struct dma_chan *c = tx->chan; 228 struct ioat_dma_chan *ioat = to_ioat_chan(c); 229 struct ioat_desc_sw *desc = tx_to_ioat_desc(tx); 230 struct ioat_chan_common *chan = &ioat->base; 231 struct ioat_desc_sw *first; 232 struct ioat_desc_sw *chain_tail; 233 dma_cookie_t cookie; 234 235 spin_lock_bh(&ioat->desc_lock); 236 /* cookie incr and addition to used_list must be atomic */ 237 cookie = c->cookie; 238 cookie++; 239 if (cookie < 0) 240 cookie = 1; 241 c->cookie = cookie; 242 tx->cookie = cookie; 243 dev_dbg(to_dev(&ioat->base), "%s: cookie: %d\n", __func__, cookie); 244 245 /* write address into NextDescriptor field of last desc in chain */ 246 first = to_ioat_desc(desc->tx_list.next); 247 chain_tail = to_ioat_desc(ioat->used_desc.prev); 248 /* make descriptor updates globally visible before chaining */ 249 wmb(); 250 chain_tail->hw->next = first->txd.phys; 251 list_splice_tail_init(&desc->tx_list, &ioat->used_desc); 252 dump_desc_dbg(ioat, chain_tail); 253 dump_desc_dbg(ioat, first); 254 255 if (!test_and_set_bit(IOAT_COMPLETION_PENDING, &chan->state)) 256 mod_timer(&chan->timer, jiffies + COMPLETION_TIMEOUT); 257 258 ioat->active += desc->hw->tx_cnt; 259 ioat->pending += desc->hw->tx_cnt; 260 if (ioat->pending >= ioat_pending_level) 261 __ioat1_dma_memcpy_issue_pending(ioat); 262 spin_unlock_bh(&ioat->desc_lock); 263 264 return cookie; 265 } 266 267 /** 268 * ioat_dma_alloc_descriptor - allocate and return a sw and hw descriptor pair 269 * @ioat: the channel supplying the memory pool for the descriptors 270 * @flags: allocation flags 271 */ 272 static struct ioat_desc_sw * 273 ioat_dma_alloc_descriptor(struct ioat_dma_chan *ioat, gfp_t flags) 274 { 275 struct ioat_dma_descriptor *desc; 276 struct ioat_desc_sw *desc_sw; 277 struct ioatdma_device *ioatdma_device; 278 dma_addr_t phys; 279 280 ioatdma_device = ioat->base.device; 281 desc = pci_pool_alloc(ioatdma_device->dma_pool, flags, &phys); 282 if (unlikely(!desc)) 283 return NULL; 284 285 desc_sw = kzalloc(sizeof(*desc_sw), flags); 286 if (unlikely(!desc_sw)) { 287 pci_pool_free(ioatdma_device->dma_pool, desc, phys); 288 return NULL; 289 } 290 291 memset(desc, 0, sizeof(*desc)); 292 293 INIT_LIST_HEAD(&desc_sw->tx_list); 294 dma_async_tx_descriptor_init(&desc_sw->txd, &ioat->base.common); 295 desc_sw->txd.tx_submit = ioat1_tx_submit; 296 desc_sw->hw = desc; 297 desc_sw->txd.phys = phys; 298 set_desc_id(desc_sw, -1); 299 300 return desc_sw; 301 } 302 303 static int ioat_initial_desc_count = 256; 304 module_param(ioat_initial_desc_count, int, 0644); 305 MODULE_PARM_DESC(ioat_initial_desc_count, 306 "ioat1: initial descriptors per channel (default: 256)"); 307 /** 308 * ioat1_dma_alloc_chan_resources - returns the number of allocated descriptors 309 * @chan: the channel to be filled out 310 */ 311 static int ioat1_dma_alloc_chan_resources(struct dma_chan *c) 312 { 313 struct ioat_dma_chan *ioat = to_ioat_chan(c); 314 struct ioat_chan_common *chan = &ioat->base; 315 struct ioat_desc_sw *desc; 316 u32 chanerr; 317 int i; 318 LIST_HEAD(tmp_list); 319 320 /* have we already been set up? */ 321 if (!list_empty(&ioat->free_desc)) 322 return ioat->desccount; 323 324 /* Setup register to interrupt and write completion status on error */ 325 writew(IOAT_CHANCTRL_RUN, chan->reg_base + IOAT_CHANCTRL_OFFSET); 326 327 chanerr = readl(chan->reg_base + IOAT_CHANERR_OFFSET); 328 if (chanerr) { 329 dev_err(to_dev(chan), "CHANERR = %x, clearing\n", chanerr); 330 writel(chanerr, chan->reg_base + IOAT_CHANERR_OFFSET); 331 } 332 333 /* Allocate descriptors */ 334 for (i = 0; i < ioat_initial_desc_count; i++) { 335 desc = ioat_dma_alloc_descriptor(ioat, GFP_KERNEL); 336 if (!desc) { 337 dev_err(to_dev(chan), "Only %d initial descriptors\n", i); 338 break; 339 } 340 set_desc_id(desc, i); 341 list_add_tail(&desc->node, &tmp_list); 342 } 343 spin_lock_bh(&ioat->desc_lock); 344 ioat->desccount = i; 345 list_splice(&tmp_list, &ioat->free_desc); 346 spin_unlock_bh(&ioat->desc_lock); 347 348 /* allocate a completion writeback area */ 349 /* doing 2 32bit writes to mmio since 1 64b write doesn't work */ 350 chan->completion = pci_pool_alloc(chan->device->completion_pool, 351 GFP_KERNEL, &chan->completion_dma); 352 memset(chan->completion, 0, sizeof(*chan->completion)); 353 writel(((u64) chan->completion_dma) & 0x00000000FFFFFFFF, 354 chan->reg_base + IOAT_CHANCMP_OFFSET_LOW); 355 writel(((u64) chan->completion_dma) >> 32, 356 chan->reg_base + IOAT_CHANCMP_OFFSET_HIGH); 357 358 tasklet_enable(&chan->cleanup_task); 359 ioat1_dma_start_null_desc(ioat); /* give chain to dma device */ 360 dev_dbg(to_dev(chan), "%s: allocated %d descriptors\n", 361 __func__, ioat->desccount); 362 return ioat->desccount; 363 } 364 365 /** 366 * ioat1_dma_free_chan_resources - release all the descriptors 367 * @chan: the channel to be cleaned 368 */ 369 static void ioat1_dma_free_chan_resources(struct dma_chan *c) 370 { 371 struct ioat_dma_chan *ioat = to_ioat_chan(c); 372 struct ioat_chan_common *chan = &ioat->base; 373 struct ioatdma_device *ioatdma_device = chan->device; 374 struct ioat_desc_sw *desc, *_desc; 375 int in_use_descs = 0; 376 377 /* Before freeing channel resources first check 378 * if they have been previously allocated for this channel. 379 */ 380 if (ioat->desccount == 0) 381 return; 382 383 tasklet_disable(&chan->cleanup_task); 384 del_timer_sync(&chan->timer); 385 ioat1_cleanup(ioat); 386 387 /* Delay 100ms after reset to allow internal DMA logic to quiesce 388 * before removing DMA descriptor resources. 389 */ 390 writeb(IOAT_CHANCMD_RESET, 391 chan->reg_base + IOAT_CHANCMD_OFFSET(chan->device->version)); 392 mdelay(100); 393 394 spin_lock_bh(&ioat->desc_lock); 395 list_for_each_entry_safe(desc, _desc, &ioat->used_desc, node) { 396 dev_dbg(to_dev(chan), "%s: freeing %d from used list\n", 397 __func__, desc_id(desc)); 398 dump_desc_dbg(ioat, desc); 399 in_use_descs++; 400 list_del(&desc->node); 401 pci_pool_free(ioatdma_device->dma_pool, desc->hw, 402 desc->txd.phys); 403 kfree(desc); 404 } 405 list_for_each_entry_safe(desc, _desc, 406 &ioat->free_desc, node) { 407 list_del(&desc->node); 408 pci_pool_free(ioatdma_device->dma_pool, desc->hw, 409 desc->txd.phys); 410 kfree(desc); 411 } 412 spin_unlock_bh(&ioat->desc_lock); 413 414 pci_pool_free(ioatdma_device->completion_pool, 415 chan->completion, 416 chan->completion_dma); 417 418 /* one is ok since we left it on there on purpose */ 419 if (in_use_descs > 1) 420 dev_err(to_dev(chan), "Freeing %d in use descriptors!\n", 421 in_use_descs - 1); 422 423 chan->last_completion = 0; 424 chan->completion_dma = 0; 425 ioat->pending = 0; 426 ioat->desccount = 0; 427 } 428 429 /** 430 * ioat1_dma_get_next_descriptor - return the next available descriptor 431 * @ioat: IOAT DMA channel handle 432 * 433 * Gets the next descriptor from the chain, and must be called with the 434 * channel's desc_lock held. Allocates more descriptors if the channel 435 * has run out. 436 */ 437 static struct ioat_desc_sw * 438 ioat1_dma_get_next_descriptor(struct ioat_dma_chan *ioat) 439 { 440 struct ioat_desc_sw *new; 441 442 if (!list_empty(&ioat->free_desc)) { 443 new = to_ioat_desc(ioat->free_desc.next); 444 list_del(&new->node); 445 } else { 446 /* try to get another desc */ 447 new = ioat_dma_alloc_descriptor(ioat, GFP_ATOMIC); 448 if (!new) { 449 dev_err(to_dev(&ioat->base), "alloc failed\n"); 450 return NULL; 451 } 452 } 453 dev_dbg(to_dev(&ioat->base), "%s: allocated: %d\n", 454 __func__, desc_id(new)); 455 prefetch(new->hw); 456 return new; 457 } 458 459 static struct dma_async_tx_descriptor * 460 ioat1_dma_prep_memcpy(struct dma_chan *c, dma_addr_t dma_dest, 461 dma_addr_t dma_src, size_t len, unsigned long flags) 462 { 463 struct ioat_dma_chan *ioat = to_ioat_chan(c); 464 struct ioat_desc_sw *desc; 465 size_t copy; 466 LIST_HEAD(chain); 467 dma_addr_t src = dma_src; 468 dma_addr_t dest = dma_dest; 469 size_t total_len = len; 470 struct ioat_dma_descriptor *hw = NULL; 471 int tx_cnt = 0; 472 473 spin_lock_bh(&ioat->desc_lock); 474 desc = ioat1_dma_get_next_descriptor(ioat); 475 do { 476 if (!desc) 477 break; 478 479 tx_cnt++; 480 copy = min_t(size_t, len, ioat->xfercap); 481 482 hw = desc->hw; 483 hw->size = copy; 484 hw->ctl = 0; 485 hw->src_addr = src; 486 hw->dst_addr = dest; 487 488 list_add_tail(&desc->node, &chain); 489 490 len -= copy; 491 dest += copy; 492 src += copy; 493 if (len) { 494 struct ioat_desc_sw *next; 495 496 async_tx_ack(&desc->txd); 497 next = ioat1_dma_get_next_descriptor(ioat); 498 hw->next = next ? next->txd.phys : 0; 499 dump_desc_dbg(ioat, desc); 500 desc = next; 501 } else 502 hw->next = 0; 503 } while (len); 504 505 if (!desc) { 506 struct ioat_chan_common *chan = &ioat->base; 507 508 dev_err(to_dev(chan), 509 "chan%d - get_next_desc failed\n", chan_num(chan)); 510 list_splice(&chain, &ioat->free_desc); 511 spin_unlock_bh(&ioat->desc_lock); 512 return NULL; 513 } 514 spin_unlock_bh(&ioat->desc_lock); 515 516 desc->txd.flags = flags; 517 desc->len = total_len; 518 list_splice(&chain, &desc->tx_list); 519 hw->ctl_f.int_en = !!(flags & DMA_PREP_INTERRUPT); 520 hw->ctl_f.compl_write = 1; 521 hw->tx_cnt = tx_cnt; 522 dump_desc_dbg(ioat, desc); 523 524 return &desc->txd; 525 } 526 527 static void ioat1_cleanup_event(unsigned long data) 528 { 529 struct ioat_dma_chan *ioat = to_ioat_chan((void *) data); 530 531 ioat1_cleanup(ioat); 532 writew(IOAT_CHANCTRL_RUN, ioat->base.reg_base + IOAT_CHANCTRL_OFFSET); 533 } 534 535 void ioat_dma_unmap(struct ioat_chan_common *chan, enum dma_ctrl_flags flags, 536 size_t len, struct ioat_dma_descriptor *hw) 537 { 538 struct pci_dev *pdev = chan->device->pdev; 539 size_t offset = len - hw->size; 540 541 if (!(flags & DMA_COMPL_SKIP_DEST_UNMAP)) 542 ioat_unmap(pdev, hw->dst_addr - offset, len, 543 PCI_DMA_FROMDEVICE, flags, 1); 544 545 if (!(flags & DMA_COMPL_SKIP_SRC_UNMAP)) 546 ioat_unmap(pdev, hw->src_addr - offset, len, 547 PCI_DMA_TODEVICE, flags, 0); 548 } 549 550 unsigned long ioat_get_current_completion(struct ioat_chan_common *chan) 551 { 552 unsigned long phys_complete; 553 u64 completion; 554 555 completion = *chan->completion; 556 phys_complete = ioat_chansts_to_addr(completion); 557 558 dev_dbg(to_dev(chan), "%s: phys_complete: %#llx\n", __func__, 559 (unsigned long long) phys_complete); 560 561 if (is_ioat_halted(completion)) { 562 u32 chanerr = readl(chan->reg_base + IOAT_CHANERR_OFFSET); 563 dev_err(to_dev(chan), "Channel halted, chanerr = %x\n", 564 chanerr); 565 566 /* TODO do something to salvage the situation */ 567 } 568 569 return phys_complete; 570 } 571 572 bool ioat_cleanup_preamble(struct ioat_chan_common *chan, 573 unsigned long *phys_complete) 574 { 575 *phys_complete = ioat_get_current_completion(chan); 576 if (*phys_complete == chan->last_completion) 577 return false; 578 clear_bit(IOAT_COMPLETION_ACK, &chan->state); 579 mod_timer(&chan->timer, jiffies + COMPLETION_TIMEOUT); 580 581 return true; 582 } 583 584 static void __cleanup(struct ioat_dma_chan *ioat, unsigned long phys_complete) 585 { 586 struct ioat_chan_common *chan = &ioat->base; 587 struct list_head *_desc, *n; 588 struct dma_async_tx_descriptor *tx; 589 590 dev_dbg(to_dev(chan), "%s: phys_complete: %lx\n", 591 __func__, phys_complete); 592 list_for_each_safe(_desc, n, &ioat->used_desc) { 593 struct ioat_desc_sw *desc; 594 595 prefetch(n); 596 desc = list_entry(_desc, typeof(*desc), node); 597 tx = &desc->txd; 598 /* 599 * Incoming DMA requests may use multiple descriptors, 600 * due to exceeding xfercap, perhaps. If so, only the 601 * last one will have a cookie, and require unmapping. 602 */ 603 dump_desc_dbg(ioat, desc); 604 if (tx->cookie) { 605 chan->completed_cookie = tx->cookie; 606 tx->cookie = 0; 607 ioat_dma_unmap(chan, tx->flags, desc->len, desc->hw); 608 ioat->active -= desc->hw->tx_cnt; 609 if (tx->callback) { 610 tx->callback(tx->callback_param); 611 tx->callback = NULL; 612 } 613 } 614 615 if (tx->phys != phys_complete) { 616 /* 617 * a completed entry, but not the last, so clean 618 * up if the client is done with the descriptor 619 */ 620 if (async_tx_test_ack(tx)) 621 list_move_tail(&desc->node, &ioat->free_desc); 622 } else { 623 /* 624 * last used desc. Do not remove, so we can 625 * append from it. 626 */ 627 628 /* if nothing else is pending, cancel the 629 * completion timeout 630 */ 631 if (n == &ioat->used_desc) { 632 dev_dbg(to_dev(chan), 633 "%s cancel completion timeout\n", 634 __func__); 635 clear_bit(IOAT_COMPLETION_PENDING, &chan->state); 636 } 637 638 /* TODO check status bits? */ 639 break; 640 } 641 } 642 643 chan->last_completion = phys_complete; 644 } 645 646 /** 647 * ioat1_cleanup - cleanup up finished descriptors 648 * @chan: ioat channel to be cleaned up 649 * 650 * To prevent lock contention we defer cleanup when the locks are 651 * contended with a terminal timeout that forces cleanup and catches 652 * completion notification errors. 653 */ 654 static void ioat1_cleanup(struct ioat_dma_chan *ioat) 655 { 656 struct ioat_chan_common *chan = &ioat->base; 657 unsigned long phys_complete; 658 659 prefetch(chan->completion); 660 661 if (!spin_trylock_bh(&chan->cleanup_lock)) 662 return; 663 664 if (!ioat_cleanup_preamble(chan, &phys_complete)) { 665 spin_unlock_bh(&chan->cleanup_lock); 666 return; 667 } 668 669 if (!spin_trylock_bh(&ioat->desc_lock)) { 670 spin_unlock_bh(&chan->cleanup_lock); 671 return; 672 } 673 674 __cleanup(ioat, phys_complete); 675 676 spin_unlock_bh(&ioat->desc_lock); 677 spin_unlock_bh(&chan->cleanup_lock); 678 } 679 680 static void ioat1_timer_event(unsigned long data) 681 { 682 struct ioat_dma_chan *ioat = to_ioat_chan((void *) data); 683 struct ioat_chan_common *chan = &ioat->base; 684 685 dev_dbg(to_dev(chan), "%s: state: %lx\n", __func__, chan->state); 686 687 spin_lock_bh(&chan->cleanup_lock); 688 if (test_and_clear_bit(IOAT_RESET_PENDING, &chan->state)) { 689 struct ioat_desc_sw *desc; 690 691 spin_lock_bh(&ioat->desc_lock); 692 693 /* restart active descriptors */ 694 desc = to_ioat_desc(ioat->used_desc.prev); 695 ioat_set_chainaddr(ioat, desc->txd.phys); 696 ioat_start(chan); 697 698 ioat->pending = 0; 699 set_bit(IOAT_COMPLETION_PENDING, &chan->state); 700 mod_timer(&chan->timer, jiffies + COMPLETION_TIMEOUT); 701 spin_unlock_bh(&ioat->desc_lock); 702 } else if (test_bit(IOAT_COMPLETION_PENDING, &chan->state)) { 703 unsigned long phys_complete; 704 705 spin_lock_bh(&ioat->desc_lock); 706 /* if we haven't made progress and we have already 707 * acknowledged a pending completion once, then be more 708 * forceful with a restart 709 */ 710 if (ioat_cleanup_preamble(chan, &phys_complete)) 711 __cleanup(ioat, phys_complete); 712 else if (test_bit(IOAT_COMPLETION_ACK, &chan->state)) 713 ioat1_reset_channel(ioat); 714 else { 715 u64 status = ioat_chansts(chan); 716 717 /* manually update the last completion address */ 718 if (ioat_chansts_to_addr(status) != 0) 719 *chan->completion = status; 720 721 set_bit(IOAT_COMPLETION_ACK, &chan->state); 722 mod_timer(&chan->timer, jiffies + COMPLETION_TIMEOUT); 723 } 724 spin_unlock_bh(&ioat->desc_lock); 725 } 726 spin_unlock_bh(&chan->cleanup_lock); 727 } 728 729 enum dma_status 730 ioat_dma_tx_status(struct dma_chan *c, dma_cookie_t cookie, 731 struct dma_tx_state *txstate) 732 { 733 struct ioat_chan_common *chan = to_chan_common(c); 734 struct ioatdma_device *device = chan->device; 735 736 if (ioat_tx_status(c, cookie, txstate) == DMA_SUCCESS) 737 return DMA_SUCCESS; 738 739 device->cleanup_fn((unsigned long) c); 740 741 return ioat_tx_status(c, cookie, txstate); 742 } 743 744 static void ioat1_dma_start_null_desc(struct ioat_dma_chan *ioat) 745 { 746 struct ioat_chan_common *chan = &ioat->base; 747 struct ioat_desc_sw *desc; 748 struct ioat_dma_descriptor *hw; 749 750 spin_lock_bh(&ioat->desc_lock); 751 752 desc = ioat1_dma_get_next_descriptor(ioat); 753 754 if (!desc) { 755 dev_err(to_dev(chan), 756 "Unable to start null desc - get next desc failed\n"); 757 spin_unlock_bh(&ioat->desc_lock); 758 return; 759 } 760 761 hw = desc->hw; 762 hw->ctl = 0; 763 hw->ctl_f.null = 1; 764 hw->ctl_f.int_en = 1; 765 hw->ctl_f.compl_write = 1; 766 /* set size to non-zero value (channel returns error when size is 0) */ 767 hw->size = NULL_DESC_BUFFER_SIZE; 768 hw->src_addr = 0; 769 hw->dst_addr = 0; 770 async_tx_ack(&desc->txd); 771 hw->next = 0; 772 list_add_tail(&desc->node, &ioat->used_desc); 773 dump_desc_dbg(ioat, desc); 774 775 ioat_set_chainaddr(ioat, desc->txd.phys); 776 ioat_start(chan); 777 spin_unlock_bh(&ioat->desc_lock); 778 } 779 780 /* 781 * Perform a IOAT transaction to verify the HW works. 782 */ 783 #define IOAT_TEST_SIZE 2000 784 785 static void __devinit ioat_dma_test_callback(void *dma_async_param) 786 { 787 struct completion *cmp = dma_async_param; 788 789 complete(cmp); 790 } 791 792 /** 793 * ioat_dma_self_test - Perform a IOAT transaction to verify the HW works. 794 * @device: device to be tested 795 */ 796 int __devinit ioat_dma_self_test(struct ioatdma_device *device) 797 { 798 int i; 799 u8 *src; 800 u8 *dest; 801 struct dma_device *dma = &device->common; 802 struct device *dev = &device->pdev->dev; 803 struct dma_chan *dma_chan; 804 struct dma_async_tx_descriptor *tx; 805 dma_addr_t dma_dest, dma_src; 806 dma_cookie_t cookie; 807 int err = 0; 808 struct completion cmp; 809 unsigned long tmo; 810 unsigned long flags; 811 812 src = kzalloc(sizeof(u8) * IOAT_TEST_SIZE, GFP_KERNEL); 813 if (!src) 814 return -ENOMEM; 815 dest = kzalloc(sizeof(u8) * IOAT_TEST_SIZE, GFP_KERNEL); 816 if (!dest) { 817 kfree(src); 818 return -ENOMEM; 819 } 820 821 /* Fill in src buffer */ 822 for (i = 0; i < IOAT_TEST_SIZE; i++) 823 src[i] = (u8)i; 824 825 /* Start copy, using first DMA channel */ 826 dma_chan = container_of(dma->channels.next, struct dma_chan, 827 device_node); 828 if (dma->device_alloc_chan_resources(dma_chan) < 1) { 829 dev_err(dev, "selftest cannot allocate chan resource\n"); 830 err = -ENODEV; 831 goto out; 832 } 833 834 dma_src = dma_map_single(dev, src, IOAT_TEST_SIZE, DMA_TO_DEVICE); 835 dma_dest = dma_map_single(dev, dest, IOAT_TEST_SIZE, DMA_FROM_DEVICE); 836 flags = DMA_COMPL_SRC_UNMAP_SINGLE | DMA_COMPL_DEST_UNMAP_SINGLE | 837 DMA_PREP_INTERRUPT; 838 tx = device->common.device_prep_dma_memcpy(dma_chan, dma_dest, dma_src, 839 IOAT_TEST_SIZE, flags); 840 if (!tx) { 841 dev_err(dev, "Self-test prep failed, disabling\n"); 842 err = -ENODEV; 843 goto free_resources; 844 } 845 846 async_tx_ack(tx); 847 init_completion(&cmp); 848 tx->callback = ioat_dma_test_callback; 849 tx->callback_param = &cmp; 850 cookie = tx->tx_submit(tx); 851 if (cookie < 0) { 852 dev_err(dev, "Self-test setup failed, disabling\n"); 853 err = -ENODEV; 854 goto free_resources; 855 } 856 dma->device_issue_pending(dma_chan); 857 858 tmo = wait_for_completion_timeout(&cmp, msecs_to_jiffies(3000)); 859 860 if (tmo == 0 || 861 dma->device_tx_status(dma_chan, cookie, NULL) 862 != DMA_SUCCESS) { 863 dev_err(dev, "Self-test copy timed out, disabling\n"); 864 err = -ENODEV; 865 goto free_resources; 866 } 867 if (memcmp(src, dest, IOAT_TEST_SIZE)) { 868 dev_err(dev, "Self-test copy failed compare, disabling\n"); 869 err = -ENODEV; 870 goto free_resources; 871 } 872 873 free_resources: 874 dma->device_free_chan_resources(dma_chan); 875 out: 876 kfree(src); 877 kfree(dest); 878 return err; 879 } 880 881 static char ioat_interrupt_style[32] = "msix"; 882 module_param_string(ioat_interrupt_style, ioat_interrupt_style, 883 sizeof(ioat_interrupt_style), 0644); 884 MODULE_PARM_DESC(ioat_interrupt_style, 885 "set ioat interrupt style: msix (default), " 886 "msix-single-vector, msi, intx)"); 887 888 /** 889 * ioat_dma_setup_interrupts - setup interrupt handler 890 * @device: ioat device 891 */ 892 static int ioat_dma_setup_interrupts(struct ioatdma_device *device) 893 { 894 struct ioat_chan_common *chan; 895 struct pci_dev *pdev = device->pdev; 896 struct device *dev = &pdev->dev; 897 struct msix_entry *msix; 898 int i, j, msixcnt; 899 int err = -EINVAL; 900 u8 intrctrl = 0; 901 902 if (!strcmp(ioat_interrupt_style, "msix")) 903 goto msix; 904 if (!strcmp(ioat_interrupt_style, "msix-single-vector")) 905 goto msix_single_vector; 906 if (!strcmp(ioat_interrupt_style, "msi")) 907 goto msi; 908 if (!strcmp(ioat_interrupt_style, "intx")) 909 goto intx; 910 dev_err(dev, "invalid ioat_interrupt_style %s\n", ioat_interrupt_style); 911 goto err_no_irq; 912 913 msix: 914 /* The number of MSI-X vectors should equal the number of channels */ 915 msixcnt = device->common.chancnt; 916 for (i = 0; i < msixcnt; i++) 917 device->msix_entries[i].entry = i; 918 919 err = pci_enable_msix(pdev, device->msix_entries, msixcnt); 920 if (err < 0) 921 goto msi; 922 if (err > 0) 923 goto msix_single_vector; 924 925 for (i = 0; i < msixcnt; i++) { 926 msix = &device->msix_entries[i]; 927 chan = ioat_chan_by_index(device, i); 928 err = devm_request_irq(dev, msix->vector, 929 ioat_dma_do_interrupt_msix, 0, 930 "ioat-msix", chan); 931 if (err) { 932 for (j = 0; j < i; j++) { 933 msix = &device->msix_entries[j]; 934 chan = ioat_chan_by_index(device, j); 935 devm_free_irq(dev, msix->vector, chan); 936 } 937 goto msix_single_vector; 938 } 939 } 940 intrctrl |= IOAT_INTRCTRL_MSIX_VECTOR_CONTROL; 941 goto done; 942 943 msix_single_vector: 944 msix = &device->msix_entries[0]; 945 msix->entry = 0; 946 err = pci_enable_msix(pdev, device->msix_entries, 1); 947 if (err) 948 goto msi; 949 950 err = devm_request_irq(dev, msix->vector, ioat_dma_do_interrupt, 0, 951 "ioat-msix", device); 952 if (err) { 953 pci_disable_msix(pdev); 954 goto msi; 955 } 956 goto done; 957 958 msi: 959 err = pci_enable_msi(pdev); 960 if (err) 961 goto intx; 962 963 err = devm_request_irq(dev, pdev->irq, ioat_dma_do_interrupt, 0, 964 "ioat-msi", device); 965 if (err) { 966 pci_disable_msi(pdev); 967 goto intx; 968 } 969 goto done; 970 971 intx: 972 err = devm_request_irq(dev, pdev->irq, ioat_dma_do_interrupt, 973 IRQF_SHARED, "ioat-intx", device); 974 if (err) 975 goto err_no_irq; 976 977 done: 978 if (device->intr_quirk) 979 device->intr_quirk(device); 980 intrctrl |= IOAT_INTRCTRL_MASTER_INT_EN; 981 writeb(intrctrl, device->reg_base + IOAT_INTRCTRL_OFFSET); 982 return 0; 983 984 err_no_irq: 985 /* Disable all interrupt generation */ 986 writeb(0, device->reg_base + IOAT_INTRCTRL_OFFSET); 987 dev_err(dev, "no usable interrupts\n"); 988 return err; 989 } 990 991 static void ioat_disable_interrupts(struct ioatdma_device *device) 992 { 993 /* Disable all interrupt generation */ 994 writeb(0, device->reg_base + IOAT_INTRCTRL_OFFSET); 995 } 996 997 int __devinit ioat_probe(struct ioatdma_device *device) 998 { 999 int err = -ENODEV; 1000 struct dma_device *dma = &device->common; 1001 struct pci_dev *pdev = device->pdev; 1002 struct device *dev = &pdev->dev; 1003 1004 /* DMA coherent memory pool for DMA descriptor allocations */ 1005 device->dma_pool = pci_pool_create("dma_desc_pool", pdev, 1006 sizeof(struct ioat_dma_descriptor), 1007 64, 0); 1008 if (!device->dma_pool) { 1009 err = -ENOMEM; 1010 goto err_dma_pool; 1011 } 1012 1013 device->completion_pool = pci_pool_create("completion_pool", pdev, 1014 sizeof(u64), SMP_CACHE_BYTES, 1015 SMP_CACHE_BYTES); 1016 1017 if (!device->completion_pool) { 1018 err = -ENOMEM; 1019 goto err_completion_pool; 1020 } 1021 1022 device->enumerate_channels(device); 1023 1024 dma_cap_set(DMA_MEMCPY, dma->cap_mask); 1025 dma->dev = &pdev->dev; 1026 1027 if (!dma->chancnt) { 1028 dev_err(dev, "channel enumeration error\n"); 1029 goto err_setup_interrupts; 1030 } 1031 1032 err = ioat_dma_setup_interrupts(device); 1033 if (err) 1034 goto err_setup_interrupts; 1035 1036 err = device->self_test(device); 1037 if (err) 1038 goto err_self_test; 1039 1040 return 0; 1041 1042 err_self_test: 1043 ioat_disable_interrupts(device); 1044 err_setup_interrupts: 1045 pci_pool_destroy(device->completion_pool); 1046 err_completion_pool: 1047 pci_pool_destroy(device->dma_pool); 1048 err_dma_pool: 1049 return err; 1050 } 1051 1052 int __devinit ioat_register(struct ioatdma_device *device) 1053 { 1054 int err = dma_async_device_register(&device->common); 1055 1056 if (err) { 1057 ioat_disable_interrupts(device); 1058 pci_pool_destroy(device->completion_pool); 1059 pci_pool_destroy(device->dma_pool); 1060 } 1061 1062 return err; 1063 } 1064 1065 /* ioat1_intr_quirk - fix up dma ctrl register to enable / disable msi */ 1066 static void ioat1_intr_quirk(struct ioatdma_device *device) 1067 { 1068 struct pci_dev *pdev = device->pdev; 1069 u32 dmactrl; 1070 1071 pci_read_config_dword(pdev, IOAT_PCI_DMACTRL_OFFSET, &dmactrl); 1072 if (pdev->msi_enabled) 1073 dmactrl |= IOAT_PCI_DMACTRL_MSI_EN; 1074 else 1075 dmactrl &= ~IOAT_PCI_DMACTRL_MSI_EN; 1076 pci_write_config_dword(pdev, IOAT_PCI_DMACTRL_OFFSET, dmactrl); 1077 } 1078 1079 static ssize_t ring_size_show(struct dma_chan *c, char *page) 1080 { 1081 struct ioat_dma_chan *ioat = to_ioat_chan(c); 1082 1083 return sprintf(page, "%d\n", ioat->desccount); 1084 } 1085 static struct ioat_sysfs_entry ring_size_attr = __ATTR_RO(ring_size); 1086 1087 static ssize_t ring_active_show(struct dma_chan *c, char *page) 1088 { 1089 struct ioat_dma_chan *ioat = to_ioat_chan(c); 1090 1091 return sprintf(page, "%d\n", ioat->active); 1092 } 1093 static struct ioat_sysfs_entry ring_active_attr = __ATTR_RO(ring_active); 1094 1095 static ssize_t cap_show(struct dma_chan *c, char *page) 1096 { 1097 struct dma_device *dma = c->device; 1098 1099 return sprintf(page, "copy%s%s%s%s%s%s\n", 1100 dma_has_cap(DMA_PQ, dma->cap_mask) ? " pq" : "", 1101 dma_has_cap(DMA_PQ_VAL, dma->cap_mask) ? " pq_val" : "", 1102 dma_has_cap(DMA_XOR, dma->cap_mask) ? " xor" : "", 1103 dma_has_cap(DMA_XOR_VAL, dma->cap_mask) ? " xor_val" : "", 1104 dma_has_cap(DMA_MEMSET, dma->cap_mask) ? " fill" : "", 1105 dma_has_cap(DMA_INTERRUPT, dma->cap_mask) ? " intr" : ""); 1106 1107 } 1108 struct ioat_sysfs_entry ioat_cap_attr = __ATTR_RO(cap); 1109 1110 static ssize_t version_show(struct dma_chan *c, char *page) 1111 { 1112 struct dma_device *dma = c->device; 1113 struct ioatdma_device *device = to_ioatdma_device(dma); 1114 1115 return sprintf(page, "%d.%d\n", 1116 device->version >> 4, device->version & 0xf); 1117 } 1118 struct ioat_sysfs_entry ioat_version_attr = __ATTR_RO(version); 1119 1120 static struct attribute *ioat1_attrs[] = { 1121 &ring_size_attr.attr, 1122 &ring_active_attr.attr, 1123 &ioat_cap_attr.attr, 1124 &ioat_version_attr.attr, 1125 NULL, 1126 }; 1127 1128 static ssize_t 1129 ioat_attr_show(struct kobject *kobj, struct attribute *attr, char *page) 1130 { 1131 struct ioat_sysfs_entry *entry; 1132 struct ioat_chan_common *chan; 1133 1134 entry = container_of(attr, struct ioat_sysfs_entry, attr); 1135 chan = container_of(kobj, struct ioat_chan_common, kobj); 1136 1137 if (!entry->show) 1138 return -EIO; 1139 return entry->show(&chan->common, page); 1140 } 1141 1142 const struct sysfs_ops ioat_sysfs_ops = { 1143 .show = ioat_attr_show, 1144 }; 1145 1146 static struct kobj_type ioat1_ktype = { 1147 .sysfs_ops = &ioat_sysfs_ops, 1148 .default_attrs = ioat1_attrs, 1149 }; 1150 1151 void ioat_kobject_add(struct ioatdma_device *device, struct kobj_type *type) 1152 { 1153 struct dma_device *dma = &device->common; 1154 struct dma_chan *c; 1155 1156 list_for_each_entry(c, &dma->channels, device_node) { 1157 struct ioat_chan_common *chan = to_chan_common(c); 1158 struct kobject *parent = &c->dev->device.kobj; 1159 int err; 1160 1161 err = kobject_init_and_add(&chan->kobj, type, parent, "quickdata"); 1162 if (err) { 1163 dev_warn(to_dev(chan), 1164 "sysfs init error (%d), continuing...\n", err); 1165 kobject_put(&chan->kobj); 1166 set_bit(IOAT_KOBJ_INIT_FAIL, &chan->state); 1167 } 1168 } 1169 } 1170 1171 void ioat_kobject_del(struct ioatdma_device *device) 1172 { 1173 struct dma_device *dma = &device->common; 1174 struct dma_chan *c; 1175 1176 list_for_each_entry(c, &dma->channels, device_node) { 1177 struct ioat_chan_common *chan = to_chan_common(c); 1178 1179 if (!test_bit(IOAT_KOBJ_INIT_FAIL, &chan->state)) { 1180 kobject_del(&chan->kobj); 1181 kobject_put(&chan->kobj); 1182 } 1183 } 1184 } 1185 1186 int __devinit ioat1_dma_probe(struct ioatdma_device *device, int dca) 1187 { 1188 struct pci_dev *pdev = device->pdev; 1189 struct dma_device *dma; 1190 int err; 1191 1192 device->intr_quirk = ioat1_intr_quirk; 1193 device->enumerate_channels = ioat1_enumerate_channels; 1194 device->self_test = ioat_dma_self_test; 1195 device->timer_fn = ioat1_timer_event; 1196 device->cleanup_fn = ioat1_cleanup_event; 1197 dma = &device->common; 1198 dma->device_prep_dma_memcpy = ioat1_dma_prep_memcpy; 1199 dma->device_issue_pending = ioat1_dma_memcpy_issue_pending; 1200 dma->device_alloc_chan_resources = ioat1_dma_alloc_chan_resources; 1201 dma->device_free_chan_resources = ioat1_dma_free_chan_resources; 1202 dma->device_tx_status = ioat_dma_tx_status; 1203 1204 err = ioat_probe(device); 1205 if (err) 1206 return err; 1207 ioat_set_tcp_copy_break(4096); 1208 err = ioat_register(device); 1209 if (err) 1210 return err; 1211 ioat_kobject_add(device, &ioat1_ktype); 1212 1213 if (dca) 1214 device->dca = ioat_dca_init(pdev, device->reg_base); 1215 1216 return err; 1217 } 1218 1219 void __devexit ioat_dma_remove(struct ioatdma_device *device) 1220 { 1221 struct dma_device *dma = &device->common; 1222 1223 ioat_disable_interrupts(device); 1224 1225 ioat_kobject_del(device); 1226 1227 dma_async_device_unregister(dma); 1228 1229 pci_pool_destroy(device->dma_pool); 1230 pci_pool_destroy(device->completion_pool); 1231 1232 INIT_LIST_HEAD(&dma->channels); 1233 } 1234