1 /* 2 * Copyright(c) 2015-2018 Intel Corporation. 3 * 4 * This file is provided under a dual BSD/GPLv2 license. When using or 5 * redistributing this file, you may do so under either license. 6 * 7 * GPL LICENSE SUMMARY 8 * 9 * This program is free software; you can redistribute it and/or modify 10 * it under the terms of version 2 of the GNU General Public License as 11 * published by the Free Software Foundation. 12 * 13 * This program is distributed in the hope that it will be useful, but 14 * WITHOUT ANY WARRANTY; without even the implied warranty of 15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 16 * General Public License for more details. 17 * 18 * BSD LICENSE 19 * 20 * Redistribution and use in source and binary forms, with or without 21 * modification, are permitted provided that the following conditions 22 * are met: 23 * 24 * - Redistributions of source code must retain the above copyright 25 * notice, this list of conditions and the following disclaimer. 26 * - Redistributions in binary form must reproduce the above copyright 27 * notice, this list of conditions and the following disclaimer in 28 * the documentation and/or other materials provided with the 29 * distribution. 30 * - Neither the name of Intel Corporation nor the names of its 31 * contributors may be used to endorse or promote products derived 32 * from this software without specific prior written permission. 33 * 34 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 35 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 36 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 37 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 38 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 39 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 40 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 41 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 42 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 43 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 44 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 45 * 46 */ 47 #include <asm/page.h> 48 #include <linux/string.h> 49 50 #include "mmu_rb.h" 51 #include "user_exp_rcv.h" 52 #include "trace.h" 53 54 static void unlock_exp_tids(struct hfi1_ctxtdata *uctxt, 55 struct exp_tid_set *set, 56 struct hfi1_filedata *fd); 57 static u32 find_phys_blocks(struct tid_user_buf *tidbuf, unsigned int npages); 58 static int set_rcvarray_entry(struct hfi1_filedata *fd, 59 struct tid_user_buf *tbuf, 60 u32 rcventry, struct tid_group *grp, 61 u16 pageidx, unsigned int npages); 62 static int tid_rb_insert(void *arg, struct mmu_rb_node *node); 63 static void cacheless_tid_rb_remove(struct hfi1_filedata *fdata, 64 struct tid_rb_node *tnode); 65 static void tid_rb_remove(void *arg, struct mmu_rb_node *node); 66 static int tid_rb_invalidate(void *arg, struct mmu_rb_node *mnode); 67 static int program_rcvarray(struct hfi1_filedata *fd, struct tid_user_buf *, 68 struct tid_group *grp, 69 unsigned int start, u16 count, 70 u32 *tidlist, unsigned int *tididx, 71 unsigned int *pmapped); 72 static int unprogram_rcvarray(struct hfi1_filedata *fd, u32 tidinfo, 73 struct tid_group **grp); 74 static void clear_tid_node(struct hfi1_filedata *fd, struct tid_rb_node *node); 75 76 static struct mmu_rb_ops tid_rb_ops = { 77 .insert = tid_rb_insert, 78 .remove = tid_rb_remove, 79 .invalidate = tid_rb_invalidate 80 }; 81 82 /* 83 * Initialize context and file private data needed for Expected 84 * receive caching. This needs to be done after the context has 85 * been configured with the eager/expected RcvEntry counts. 86 */ 87 int hfi1_user_exp_rcv_init(struct hfi1_filedata *fd, 88 struct hfi1_ctxtdata *uctxt) 89 { 90 struct hfi1_devdata *dd = uctxt->dd; 91 int ret = 0; 92 93 spin_lock_init(&fd->tid_lock); 94 spin_lock_init(&fd->invalid_lock); 95 96 fd->entry_to_rb = kcalloc(uctxt->expected_count, 97 sizeof(struct rb_node *), 98 GFP_KERNEL); 99 if (!fd->entry_to_rb) 100 return -ENOMEM; 101 102 if (!HFI1_CAP_UGET_MASK(uctxt->flags, TID_UNMAP)) { 103 fd->invalid_tid_idx = 0; 104 fd->invalid_tids = kcalloc(uctxt->expected_count, 105 sizeof(*fd->invalid_tids), 106 GFP_KERNEL); 107 if (!fd->invalid_tids) { 108 kfree(fd->entry_to_rb); 109 fd->entry_to_rb = NULL; 110 return -ENOMEM; 111 } 112 113 /* 114 * Register MMU notifier callbacks. If the registration 115 * fails, continue without TID caching for this context. 116 */ 117 ret = hfi1_mmu_rb_register(fd, fd->mm, &tid_rb_ops, 118 dd->pport->hfi1_wq, 119 &fd->handler); 120 if (ret) { 121 dd_dev_info(dd, 122 "Failed MMU notifier registration %d\n", 123 ret); 124 ret = 0; 125 } 126 } 127 128 /* 129 * PSM does not have a good way to separate, count, and 130 * effectively enforce a limit on RcvArray entries used by 131 * subctxts (when context sharing is used) when TID caching 132 * is enabled. To help with that, we calculate a per-process 133 * RcvArray entry share and enforce that. 134 * If TID caching is not in use, PSM deals with usage on its 135 * own. In that case, we allow any subctxt to take all of the 136 * entries. 137 * 138 * Make sure that we set the tid counts only after successful 139 * init. 140 */ 141 spin_lock(&fd->tid_lock); 142 if (uctxt->subctxt_cnt && fd->handler) { 143 u16 remainder; 144 145 fd->tid_limit = uctxt->expected_count / uctxt->subctxt_cnt; 146 remainder = uctxt->expected_count % uctxt->subctxt_cnt; 147 if (remainder && fd->subctxt < remainder) 148 fd->tid_limit++; 149 } else { 150 fd->tid_limit = uctxt->expected_count; 151 } 152 spin_unlock(&fd->tid_lock); 153 154 return ret; 155 } 156 157 void hfi1_user_exp_rcv_free(struct hfi1_filedata *fd) 158 { 159 struct hfi1_ctxtdata *uctxt = fd->uctxt; 160 161 /* 162 * The notifier would have been removed when the process'es mm 163 * was freed. 164 */ 165 if (fd->handler) { 166 hfi1_mmu_rb_unregister(fd->handler); 167 } else { 168 if (!EXP_TID_SET_EMPTY(uctxt->tid_full_list)) 169 unlock_exp_tids(uctxt, &uctxt->tid_full_list, fd); 170 if (!EXP_TID_SET_EMPTY(uctxt->tid_used_list)) 171 unlock_exp_tids(uctxt, &uctxt->tid_used_list, fd); 172 } 173 174 kfree(fd->invalid_tids); 175 fd->invalid_tids = NULL; 176 177 kfree(fd->entry_to_rb); 178 fd->entry_to_rb = NULL; 179 } 180 181 /** 182 * Release pinned receive buffer pages. 183 * 184 * @mapped - true if the pages have been DMA mapped. false otherwise. 185 * @idx - Index of the first page to unpin. 186 * @npages - No of pages to unpin. 187 * 188 * If the pages have been DMA mapped (indicated by mapped parameter), their 189 * info will be passed via a struct tid_rb_node. If they haven't been mapped, 190 * their info will be passed via a struct tid_user_buf. 191 */ 192 static void unpin_rcv_pages(struct hfi1_filedata *fd, 193 struct tid_user_buf *tidbuf, 194 struct tid_rb_node *node, 195 unsigned int idx, 196 unsigned int npages, 197 bool mapped) 198 { 199 struct page **pages; 200 struct hfi1_devdata *dd = fd->uctxt->dd; 201 202 if (mapped) { 203 pci_unmap_single(dd->pcidev, node->dma_addr, 204 node->mmu.len, PCI_DMA_FROMDEVICE); 205 pages = &node->pages[idx]; 206 } else { 207 pages = &tidbuf->pages[idx]; 208 } 209 hfi1_release_user_pages(fd->mm, pages, npages, mapped); 210 fd->tid_n_pinned -= npages; 211 } 212 213 /** 214 * Pin receive buffer pages. 215 */ 216 static int pin_rcv_pages(struct hfi1_filedata *fd, struct tid_user_buf *tidbuf) 217 { 218 int pinned; 219 unsigned int npages; 220 unsigned long vaddr = tidbuf->vaddr; 221 struct page **pages = NULL; 222 struct hfi1_devdata *dd = fd->uctxt->dd; 223 224 /* Get the number of pages the user buffer spans */ 225 npages = num_user_pages(vaddr, tidbuf->length); 226 if (!npages) 227 return -EINVAL; 228 229 if (npages > fd->uctxt->expected_count) { 230 dd_dev_err(dd, "Expected buffer too big\n"); 231 return -EINVAL; 232 } 233 234 /* Verify that access is OK for the user buffer */ 235 if (!access_ok((void __user *)vaddr, 236 npages * PAGE_SIZE)) { 237 dd_dev_err(dd, "Fail vaddr %p, %u pages, !access_ok\n", 238 (void *)vaddr, npages); 239 return -EFAULT; 240 } 241 /* Allocate the array of struct page pointers needed for pinning */ 242 pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL); 243 if (!pages) 244 return -ENOMEM; 245 246 /* 247 * Pin all the pages of the user buffer. If we can't pin all the 248 * pages, accept the amount pinned so far and program only that. 249 * User space knows how to deal with partially programmed buffers. 250 */ 251 if (!hfi1_can_pin_pages(dd, fd->mm, fd->tid_n_pinned, npages)) { 252 kfree(pages); 253 return -ENOMEM; 254 } 255 256 pinned = hfi1_acquire_user_pages(fd->mm, vaddr, npages, true, pages); 257 if (pinned <= 0) { 258 kfree(pages); 259 return pinned; 260 } 261 tidbuf->pages = pages; 262 tidbuf->npages = npages; 263 fd->tid_n_pinned += pinned; 264 return pinned; 265 } 266 267 /* 268 * RcvArray entry allocation for Expected Receives is done by the 269 * following algorithm: 270 * 271 * The context keeps 3 lists of groups of RcvArray entries: 272 * 1. List of empty groups - tid_group_list 273 * This list is created during user context creation and 274 * contains elements which describe sets (of 8) of empty 275 * RcvArray entries. 276 * 2. List of partially used groups - tid_used_list 277 * This list contains sets of RcvArray entries which are 278 * not completely used up. Another mapping request could 279 * use some of all of the remaining entries. 280 * 3. List of full groups - tid_full_list 281 * This is the list where sets that are completely used 282 * up go. 283 * 284 * An attempt to optimize the usage of RcvArray entries is 285 * made by finding all sets of physically contiguous pages in a 286 * user's buffer. 287 * These physically contiguous sets are further split into 288 * sizes supported by the receive engine of the HFI. The 289 * resulting sets of pages are stored in struct tid_pageset, 290 * which describes the sets as: 291 * * .count - number of pages in this set 292 * * .idx - starting index into struct page ** array 293 * of this set 294 * 295 * From this point on, the algorithm deals with the page sets 296 * described above. The number of pagesets is divided by the 297 * RcvArray group size to produce the number of full groups 298 * needed. 299 * 300 * Groups from the 3 lists are manipulated using the following 301 * rules: 302 * 1. For each set of 8 pagesets, a complete group from 303 * tid_group_list is taken, programmed, and moved to 304 * the tid_full_list list. 305 * 2. For all remaining pagesets: 306 * 2.1 If the tid_used_list is empty and the tid_group_list 307 * is empty, stop processing pageset and return only 308 * what has been programmed up to this point. 309 * 2.2 If the tid_used_list is empty and the tid_group_list 310 * is not empty, move a group from tid_group_list to 311 * tid_used_list. 312 * 2.3 For each group is tid_used_group, program as much as 313 * can fit into the group. If the group becomes fully 314 * used, move it to tid_full_list. 315 */ 316 int hfi1_user_exp_rcv_setup(struct hfi1_filedata *fd, 317 struct hfi1_tid_info *tinfo) 318 { 319 int ret = 0, need_group = 0, pinned; 320 struct hfi1_ctxtdata *uctxt = fd->uctxt; 321 struct hfi1_devdata *dd = uctxt->dd; 322 unsigned int ngroups, pageidx = 0, pageset_count, 323 tididx = 0, mapped, mapped_pages = 0; 324 u32 *tidlist = NULL; 325 struct tid_user_buf *tidbuf; 326 327 if (!PAGE_ALIGNED(tinfo->vaddr)) 328 return -EINVAL; 329 330 tidbuf = kzalloc(sizeof(*tidbuf), GFP_KERNEL); 331 if (!tidbuf) 332 return -ENOMEM; 333 334 tidbuf->vaddr = tinfo->vaddr; 335 tidbuf->length = tinfo->length; 336 tidbuf->psets = kcalloc(uctxt->expected_count, sizeof(*tidbuf->psets), 337 GFP_KERNEL); 338 if (!tidbuf->psets) { 339 kfree(tidbuf); 340 return -ENOMEM; 341 } 342 343 pinned = pin_rcv_pages(fd, tidbuf); 344 if (pinned <= 0) { 345 kfree(tidbuf->psets); 346 kfree(tidbuf); 347 return pinned; 348 } 349 350 /* Find sets of physically contiguous pages */ 351 tidbuf->n_psets = find_phys_blocks(tidbuf, pinned); 352 353 /* 354 * We don't need to access this under a lock since tid_used is per 355 * process and the same process cannot be in hfi1_user_exp_rcv_clear() 356 * and hfi1_user_exp_rcv_setup() at the same time. 357 */ 358 spin_lock(&fd->tid_lock); 359 if (fd->tid_used + tidbuf->n_psets > fd->tid_limit) 360 pageset_count = fd->tid_limit - fd->tid_used; 361 else 362 pageset_count = tidbuf->n_psets; 363 spin_unlock(&fd->tid_lock); 364 365 if (!pageset_count) 366 goto bail; 367 368 ngroups = pageset_count / dd->rcv_entries.group_size; 369 tidlist = kcalloc(pageset_count, sizeof(*tidlist), GFP_KERNEL); 370 if (!tidlist) { 371 ret = -ENOMEM; 372 goto nomem; 373 } 374 375 tididx = 0; 376 377 /* 378 * From this point on, we are going to be using shared (between master 379 * and subcontexts) context resources. We need to take the lock. 380 */ 381 mutex_lock(&uctxt->exp_mutex); 382 /* 383 * The first step is to program the RcvArray entries which are complete 384 * groups. 385 */ 386 while (ngroups && uctxt->tid_group_list.count) { 387 struct tid_group *grp = 388 tid_group_pop(&uctxt->tid_group_list); 389 390 ret = program_rcvarray(fd, tidbuf, grp, 391 pageidx, dd->rcv_entries.group_size, 392 tidlist, &tididx, &mapped); 393 /* 394 * If there was a failure to program the RcvArray 395 * entries for the entire group, reset the grp fields 396 * and add the grp back to the free group list. 397 */ 398 if (ret <= 0) { 399 tid_group_add_tail(grp, &uctxt->tid_group_list); 400 hfi1_cdbg(TID, 401 "Failed to program RcvArray group %d", ret); 402 goto unlock; 403 } 404 405 tid_group_add_tail(grp, &uctxt->tid_full_list); 406 ngroups--; 407 pageidx += ret; 408 mapped_pages += mapped; 409 } 410 411 while (pageidx < pageset_count) { 412 struct tid_group *grp, *ptr; 413 /* 414 * If we don't have any partially used tid groups, check 415 * if we have empty groups. If so, take one from there and 416 * put in the partially used list. 417 */ 418 if (!uctxt->tid_used_list.count || need_group) { 419 if (!uctxt->tid_group_list.count) 420 goto unlock; 421 422 grp = tid_group_pop(&uctxt->tid_group_list); 423 tid_group_add_tail(grp, &uctxt->tid_used_list); 424 need_group = 0; 425 } 426 /* 427 * There is an optimization opportunity here - instead of 428 * fitting as many page sets as we can, check for a group 429 * later on in the list that could fit all of them. 430 */ 431 list_for_each_entry_safe(grp, ptr, &uctxt->tid_used_list.list, 432 list) { 433 unsigned use = min_t(unsigned, pageset_count - pageidx, 434 grp->size - grp->used); 435 436 ret = program_rcvarray(fd, tidbuf, grp, 437 pageidx, use, tidlist, 438 &tididx, &mapped); 439 if (ret < 0) { 440 hfi1_cdbg(TID, 441 "Failed to program RcvArray entries %d", 442 ret); 443 goto unlock; 444 } else if (ret > 0) { 445 if (grp->used == grp->size) 446 tid_group_move(grp, 447 &uctxt->tid_used_list, 448 &uctxt->tid_full_list); 449 pageidx += ret; 450 mapped_pages += mapped; 451 need_group = 0; 452 /* Check if we are done so we break out early */ 453 if (pageidx >= pageset_count) 454 break; 455 } else if (WARN_ON(ret == 0)) { 456 /* 457 * If ret is 0, we did not program any entries 458 * into this group, which can only happen if 459 * we've screwed up the accounting somewhere. 460 * Warn and try to continue. 461 */ 462 need_group = 1; 463 } 464 } 465 } 466 unlock: 467 mutex_unlock(&uctxt->exp_mutex); 468 nomem: 469 hfi1_cdbg(TID, "total mapped: tidpairs:%u pages:%u (%d)", tididx, 470 mapped_pages, ret); 471 if (tididx) { 472 spin_lock(&fd->tid_lock); 473 fd->tid_used += tididx; 474 spin_unlock(&fd->tid_lock); 475 tinfo->tidcnt = tididx; 476 tinfo->length = mapped_pages * PAGE_SIZE; 477 478 if (copy_to_user(u64_to_user_ptr(tinfo->tidlist), 479 tidlist, sizeof(tidlist[0]) * tididx)) { 480 /* 481 * On failure to copy to the user level, we need to undo 482 * everything done so far so we don't leak resources. 483 */ 484 tinfo->tidlist = (unsigned long)&tidlist; 485 hfi1_user_exp_rcv_clear(fd, tinfo); 486 tinfo->tidlist = 0; 487 ret = -EFAULT; 488 goto bail; 489 } 490 } 491 492 /* 493 * If not everything was mapped (due to insufficient RcvArray entries, 494 * for example), unpin all unmapped pages so we can pin them nex time. 495 */ 496 if (mapped_pages != pinned) 497 unpin_rcv_pages(fd, tidbuf, NULL, mapped_pages, 498 (pinned - mapped_pages), false); 499 bail: 500 kfree(tidbuf->psets); 501 kfree(tidlist); 502 kfree(tidbuf->pages); 503 kfree(tidbuf); 504 return ret > 0 ? 0 : ret; 505 } 506 507 int hfi1_user_exp_rcv_clear(struct hfi1_filedata *fd, 508 struct hfi1_tid_info *tinfo) 509 { 510 int ret = 0; 511 struct hfi1_ctxtdata *uctxt = fd->uctxt; 512 u32 *tidinfo; 513 unsigned tididx; 514 515 if (unlikely(tinfo->tidcnt > fd->tid_used)) 516 return -EINVAL; 517 518 tidinfo = memdup_user(u64_to_user_ptr(tinfo->tidlist), 519 sizeof(tidinfo[0]) * tinfo->tidcnt); 520 if (IS_ERR(tidinfo)) 521 return PTR_ERR(tidinfo); 522 523 mutex_lock(&uctxt->exp_mutex); 524 for (tididx = 0; tididx < tinfo->tidcnt; tididx++) { 525 ret = unprogram_rcvarray(fd, tidinfo[tididx], NULL); 526 if (ret) { 527 hfi1_cdbg(TID, "Failed to unprogram rcv array %d", 528 ret); 529 break; 530 } 531 } 532 spin_lock(&fd->tid_lock); 533 fd->tid_used -= tididx; 534 spin_unlock(&fd->tid_lock); 535 tinfo->tidcnt = tididx; 536 mutex_unlock(&uctxt->exp_mutex); 537 538 kfree(tidinfo); 539 return ret; 540 } 541 542 int hfi1_user_exp_rcv_invalid(struct hfi1_filedata *fd, 543 struct hfi1_tid_info *tinfo) 544 { 545 struct hfi1_ctxtdata *uctxt = fd->uctxt; 546 unsigned long *ev = uctxt->dd->events + 547 (uctxt_offset(uctxt) + fd->subctxt); 548 u32 *array; 549 int ret = 0; 550 551 /* 552 * copy_to_user() can sleep, which will leave the invalid_lock 553 * locked and cause the MMU notifier to be blocked on the lock 554 * for a long time. 555 * Copy the data to a local buffer so we can release the lock. 556 */ 557 array = kcalloc(uctxt->expected_count, sizeof(*array), GFP_KERNEL); 558 if (!array) 559 return -EFAULT; 560 561 spin_lock(&fd->invalid_lock); 562 if (fd->invalid_tid_idx) { 563 memcpy(array, fd->invalid_tids, sizeof(*array) * 564 fd->invalid_tid_idx); 565 memset(fd->invalid_tids, 0, sizeof(*fd->invalid_tids) * 566 fd->invalid_tid_idx); 567 tinfo->tidcnt = fd->invalid_tid_idx; 568 fd->invalid_tid_idx = 0; 569 /* 570 * Reset the user flag while still holding the lock. 571 * Otherwise, PSM can miss events. 572 */ 573 clear_bit(_HFI1_EVENT_TID_MMU_NOTIFY_BIT, ev); 574 } else { 575 tinfo->tidcnt = 0; 576 } 577 spin_unlock(&fd->invalid_lock); 578 579 if (tinfo->tidcnt) { 580 if (copy_to_user((void __user *)tinfo->tidlist, 581 array, sizeof(*array) * tinfo->tidcnt)) 582 ret = -EFAULT; 583 } 584 kfree(array); 585 586 return ret; 587 } 588 589 static u32 find_phys_blocks(struct tid_user_buf *tidbuf, unsigned int npages) 590 { 591 unsigned pagecount, pageidx, setcount = 0, i; 592 unsigned long pfn, this_pfn; 593 struct page **pages = tidbuf->pages; 594 struct tid_pageset *list = tidbuf->psets; 595 596 if (!npages) 597 return 0; 598 599 /* 600 * Look for sets of physically contiguous pages in the user buffer. 601 * This will allow us to optimize Expected RcvArray entry usage by 602 * using the bigger supported sizes. 603 */ 604 pfn = page_to_pfn(pages[0]); 605 for (pageidx = 0, pagecount = 1, i = 1; i <= npages; i++) { 606 this_pfn = i < npages ? page_to_pfn(pages[i]) : 0; 607 608 /* 609 * If the pfn's are not sequential, pages are not physically 610 * contiguous. 611 */ 612 if (this_pfn != ++pfn) { 613 /* 614 * At this point we have to loop over the set of 615 * physically contiguous pages and break them down it 616 * sizes supported by the HW. 617 * There are two main constraints: 618 * 1. The max buffer size is MAX_EXPECTED_BUFFER. 619 * If the total set size is bigger than that 620 * program only a MAX_EXPECTED_BUFFER chunk. 621 * 2. The buffer size has to be a power of two. If 622 * it is not, round down to the closes power of 623 * 2 and program that size. 624 */ 625 while (pagecount) { 626 int maxpages = pagecount; 627 u32 bufsize = pagecount * PAGE_SIZE; 628 629 if (bufsize > MAX_EXPECTED_BUFFER) 630 maxpages = 631 MAX_EXPECTED_BUFFER >> 632 PAGE_SHIFT; 633 else if (!is_power_of_2(bufsize)) 634 maxpages = 635 rounddown_pow_of_two(bufsize) >> 636 PAGE_SHIFT; 637 638 list[setcount].idx = pageidx; 639 list[setcount].count = maxpages; 640 pagecount -= maxpages; 641 pageidx += maxpages; 642 setcount++; 643 } 644 pageidx = i; 645 pagecount = 1; 646 pfn = this_pfn; 647 } else { 648 pagecount++; 649 } 650 } 651 return setcount; 652 } 653 654 /** 655 * program_rcvarray() - program an RcvArray group with receive buffers 656 * @fd: filedata pointer 657 * @tbuf: pointer to struct tid_user_buf that has the user buffer starting 658 * virtual address, buffer length, page pointers, pagesets (array of 659 * struct tid_pageset holding information on physically contiguous 660 * chunks from the user buffer), and other fields. 661 * @grp: RcvArray group 662 * @start: starting index into sets array 663 * @count: number of struct tid_pageset's to program 664 * @tidlist: the array of u32 elements when the information about the 665 * programmed RcvArray entries is to be encoded. 666 * @tididx: starting offset into tidlist 667 * @pmapped: (output parameter) number of pages programmed into the RcvArray 668 * entries. 669 * 670 * This function will program up to 'count' number of RcvArray entries from the 671 * group 'grp'. To make best use of write-combining writes, the function will 672 * perform writes to the unused RcvArray entries which will be ignored by the 673 * HW. Each RcvArray entry will be programmed with a physically contiguous 674 * buffer chunk from the user's virtual buffer. 675 * 676 * Return: 677 * -EINVAL if the requested count is larger than the size of the group, 678 * -ENOMEM or -EFAULT on error from set_rcvarray_entry(), or 679 * number of RcvArray entries programmed. 680 */ 681 static int program_rcvarray(struct hfi1_filedata *fd, struct tid_user_buf *tbuf, 682 struct tid_group *grp, 683 unsigned int start, u16 count, 684 u32 *tidlist, unsigned int *tididx, 685 unsigned int *pmapped) 686 { 687 struct hfi1_ctxtdata *uctxt = fd->uctxt; 688 struct hfi1_devdata *dd = uctxt->dd; 689 u16 idx; 690 u32 tidinfo = 0, rcventry, useidx = 0; 691 int mapped = 0; 692 693 /* Count should never be larger than the group size */ 694 if (count > grp->size) 695 return -EINVAL; 696 697 /* Find the first unused entry in the group */ 698 for (idx = 0; idx < grp->size; idx++) { 699 if (!(grp->map & (1 << idx))) { 700 useidx = idx; 701 break; 702 } 703 rcv_array_wc_fill(dd, grp->base + idx); 704 } 705 706 idx = 0; 707 while (idx < count) { 708 u16 npages, pageidx, setidx = start + idx; 709 int ret = 0; 710 711 /* 712 * If this entry in the group is used, move to the next one. 713 * If we go past the end of the group, exit the loop. 714 */ 715 if (useidx >= grp->size) { 716 break; 717 } else if (grp->map & (1 << useidx)) { 718 rcv_array_wc_fill(dd, grp->base + useidx); 719 useidx++; 720 continue; 721 } 722 723 rcventry = grp->base + useidx; 724 npages = tbuf->psets[setidx].count; 725 pageidx = tbuf->psets[setidx].idx; 726 727 ret = set_rcvarray_entry(fd, tbuf, 728 rcventry, grp, pageidx, 729 npages); 730 if (ret) 731 return ret; 732 mapped += npages; 733 734 tidinfo = rcventry2tidinfo(rcventry - uctxt->expected_base) | 735 EXP_TID_SET(LEN, npages); 736 tidlist[(*tididx)++] = tidinfo; 737 grp->used++; 738 grp->map |= 1 << useidx++; 739 idx++; 740 } 741 742 /* Fill the rest of the group with "blank" writes */ 743 for (; useidx < grp->size; useidx++) 744 rcv_array_wc_fill(dd, grp->base + useidx); 745 *pmapped = mapped; 746 return idx; 747 } 748 749 static int set_rcvarray_entry(struct hfi1_filedata *fd, 750 struct tid_user_buf *tbuf, 751 u32 rcventry, struct tid_group *grp, 752 u16 pageidx, unsigned int npages) 753 { 754 int ret; 755 struct hfi1_ctxtdata *uctxt = fd->uctxt; 756 struct tid_rb_node *node; 757 struct hfi1_devdata *dd = uctxt->dd; 758 dma_addr_t phys; 759 struct page **pages = tbuf->pages + pageidx; 760 761 /* 762 * Allocate the node first so we can handle a potential 763 * failure before we've programmed anything. 764 */ 765 node = kzalloc(sizeof(*node) + (sizeof(struct page *) * npages), 766 GFP_KERNEL); 767 if (!node) 768 return -ENOMEM; 769 770 phys = pci_map_single(dd->pcidev, 771 __va(page_to_phys(pages[0])), 772 npages * PAGE_SIZE, PCI_DMA_FROMDEVICE); 773 if (dma_mapping_error(&dd->pcidev->dev, phys)) { 774 dd_dev_err(dd, "Failed to DMA map Exp Rcv pages 0x%llx\n", 775 phys); 776 kfree(node); 777 return -EFAULT; 778 } 779 780 node->mmu.addr = tbuf->vaddr + (pageidx * PAGE_SIZE); 781 node->mmu.len = npages * PAGE_SIZE; 782 node->phys = page_to_phys(pages[0]); 783 node->npages = npages; 784 node->rcventry = rcventry; 785 node->dma_addr = phys; 786 node->grp = grp; 787 node->freed = false; 788 memcpy(node->pages, pages, sizeof(struct page *) * npages); 789 790 if (!fd->handler) 791 ret = tid_rb_insert(fd, &node->mmu); 792 else 793 ret = hfi1_mmu_rb_insert(fd->handler, &node->mmu); 794 795 if (ret) { 796 hfi1_cdbg(TID, "Failed to insert RB node %u 0x%lx, 0x%lx %d", 797 node->rcventry, node->mmu.addr, node->phys, ret); 798 pci_unmap_single(dd->pcidev, phys, npages * PAGE_SIZE, 799 PCI_DMA_FROMDEVICE); 800 kfree(node); 801 return -EFAULT; 802 } 803 hfi1_put_tid(dd, rcventry, PT_EXPECTED, phys, ilog2(npages) + 1); 804 trace_hfi1_exp_tid_reg(uctxt->ctxt, fd->subctxt, rcventry, npages, 805 node->mmu.addr, node->phys, phys); 806 return 0; 807 } 808 809 static int unprogram_rcvarray(struct hfi1_filedata *fd, u32 tidinfo, 810 struct tid_group **grp) 811 { 812 struct hfi1_ctxtdata *uctxt = fd->uctxt; 813 struct hfi1_devdata *dd = uctxt->dd; 814 struct tid_rb_node *node; 815 u8 tidctrl = EXP_TID_GET(tidinfo, CTRL); 816 u32 tididx = EXP_TID_GET(tidinfo, IDX) << 1, rcventry; 817 818 if (tididx >= uctxt->expected_count) { 819 dd_dev_err(dd, "Invalid RcvArray entry (%u) index for ctxt %u\n", 820 tididx, uctxt->ctxt); 821 return -EINVAL; 822 } 823 824 if (tidctrl == 0x3) 825 return -EINVAL; 826 827 rcventry = tididx + (tidctrl - 1); 828 829 node = fd->entry_to_rb[rcventry]; 830 if (!node || node->rcventry != (uctxt->expected_base + rcventry)) 831 return -EBADF; 832 833 if (grp) 834 *grp = node->grp; 835 836 if (!fd->handler) 837 cacheless_tid_rb_remove(fd, node); 838 else 839 hfi1_mmu_rb_remove(fd->handler, &node->mmu); 840 841 return 0; 842 } 843 844 static void clear_tid_node(struct hfi1_filedata *fd, struct tid_rb_node *node) 845 { 846 struct hfi1_ctxtdata *uctxt = fd->uctxt; 847 struct hfi1_devdata *dd = uctxt->dd; 848 849 trace_hfi1_exp_tid_unreg(uctxt->ctxt, fd->subctxt, node->rcventry, 850 node->npages, node->mmu.addr, node->phys, 851 node->dma_addr); 852 853 /* 854 * Make sure device has seen the write before we unpin the 855 * pages. 856 */ 857 hfi1_put_tid(dd, node->rcventry, PT_INVALID_FLUSH, 0, 0); 858 859 unpin_rcv_pages(fd, NULL, node, 0, node->npages, true); 860 861 node->grp->used--; 862 node->grp->map &= ~(1 << (node->rcventry - node->grp->base)); 863 864 if (node->grp->used == node->grp->size - 1) 865 tid_group_move(node->grp, &uctxt->tid_full_list, 866 &uctxt->tid_used_list); 867 else if (!node->grp->used) 868 tid_group_move(node->grp, &uctxt->tid_used_list, 869 &uctxt->tid_group_list); 870 kfree(node); 871 } 872 873 /* 874 * As a simple helper for hfi1_user_exp_rcv_free, this function deals with 875 * clearing nodes in the non-cached case. 876 */ 877 static void unlock_exp_tids(struct hfi1_ctxtdata *uctxt, 878 struct exp_tid_set *set, 879 struct hfi1_filedata *fd) 880 { 881 struct tid_group *grp, *ptr; 882 int i; 883 884 list_for_each_entry_safe(grp, ptr, &set->list, list) { 885 list_del_init(&grp->list); 886 887 for (i = 0; i < grp->size; i++) { 888 if (grp->map & (1 << i)) { 889 u16 rcventry = grp->base + i; 890 struct tid_rb_node *node; 891 892 node = fd->entry_to_rb[rcventry - 893 uctxt->expected_base]; 894 if (!node || node->rcventry != rcventry) 895 continue; 896 897 cacheless_tid_rb_remove(fd, node); 898 } 899 } 900 } 901 } 902 903 /* 904 * Always return 0 from this function. A non-zero return indicates that the 905 * remove operation will be called and that memory should be unpinned. 906 * However, the driver cannot unpin out from under PSM. Instead, retain the 907 * memory (by returning 0) and inform PSM that the memory is going away. PSM 908 * will call back later when it has removed the memory from its list. 909 */ 910 static int tid_rb_invalidate(void *arg, struct mmu_rb_node *mnode) 911 { 912 struct hfi1_filedata *fdata = arg; 913 struct hfi1_ctxtdata *uctxt = fdata->uctxt; 914 struct tid_rb_node *node = 915 container_of(mnode, struct tid_rb_node, mmu); 916 917 if (node->freed) 918 return 0; 919 920 trace_hfi1_exp_tid_inval(uctxt->ctxt, fdata->subctxt, node->mmu.addr, 921 node->rcventry, node->npages, node->dma_addr); 922 node->freed = true; 923 924 spin_lock(&fdata->invalid_lock); 925 if (fdata->invalid_tid_idx < uctxt->expected_count) { 926 fdata->invalid_tids[fdata->invalid_tid_idx] = 927 rcventry2tidinfo(node->rcventry - uctxt->expected_base); 928 fdata->invalid_tids[fdata->invalid_tid_idx] |= 929 EXP_TID_SET(LEN, node->npages); 930 if (!fdata->invalid_tid_idx) { 931 unsigned long *ev; 932 933 /* 934 * hfi1_set_uevent_bits() sets a user event flag 935 * for all processes. Because calling into the 936 * driver to process TID cache invalidations is 937 * expensive and TID cache invalidations are 938 * handled on a per-process basis, we can 939 * optimize this to set the flag only for the 940 * process in question. 941 */ 942 ev = uctxt->dd->events + 943 (uctxt_offset(uctxt) + fdata->subctxt); 944 set_bit(_HFI1_EVENT_TID_MMU_NOTIFY_BIT, ev); 945 } 946 fdata->invalid_tid_idx++; 947 } 948 spin_unlock(&fdata->invalid_lock); 949 return 0; 950 } 951 952 static int tid_rb_insert(void *arg, struct mmu_rb_node *node) 953 { 954 struct hfi1_filedata *fdata = arg; 955 struct tid_rb_node *tnode = 956 container_of(node, struct tid_rb_node, mmu); 957 u32 base = fdata->uctxt->expected_base; 958 959 fdata->entry_to_rb[tnode->rcventry - base] = tnode; 960 return 0; 961 } 962 963 static void cacheless_tid_rb_remove(struct hfi1_filedata *fdata, 964 struct tid_rb_node *tnode) 965 { 966 u32 base = fdata->uctxt->expected_base; 967 968 fdata->entry_to_rb[tnode->rcventry - base] = NULL; 969 clear_tid_node(fdata, tnode); 970 } 971 972 static void tid_rb_remove(void *arg, struct mmu_rb_node *node) 973 { 974 struct hfi1_filedata *fdata = arg; 975 struct tid_rb_node *tnode = 976 container_of(node, struct tid_rb_node, mmu); 977 978 cacheless_tid_rb_remove(fdata, tnode); 979 } 980