drivers/infiniband/hw/hfi1/user_exp_rcv.c

   1 // SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause
   2 /*
   3  * Copyright(c) 2020 Cornelis Networks, Inc.
   4  * Copyright(c) 2015-2018 Intel Corporation.
   5  */
   6 #include <asm/page.h>
   7 #include <linux/string.h>
   8
   9 #include "mmu_rb.h"
  10 #include "user_exp_rcv.h"
  11 #include "trace.h"
  12
  13 static void unlock_exp_tids(struct hfi1_ctxtdata *uctxt,
  14                             struct exp_tid_set *set,
  15                             struct hfi1_filedata *fd);
  16 static u32 find_phys_blocks(struct tid_user_buf *tidbuf, unsigned int npages);
  17 static int set_rcvarray_entry(struct hfi1_filedata *fd,
  18                               struct tid_user_buf *tbuf,
  19                               u32 rcventry, struct tid_group *grp,
  20                               u16 pageidx, unsigned int npages);
  21 static void cacheless_tid_rb_remove(struct hfi1_filedata *fdata,
  22                                     struct tid_rb_node *tnode);
  23 static bool tid_rb_invalidate(struct mmu_interval_notifier *mni,
  24                               const struct mmu_notifier_range *range,
  25                               unsigned long cur_seq);
  26 static int program_rcvarray(struct hfi1_filedata *fd, struct tid_user_buf *,
  27                             struct tid_group *grp,
  28                             unsigned int start, u16 count,
  29                             u32 *tidlist, unsigned int *tididx,
  30                             unsigned int *pmapped);
  31 static int unprogram_rcvarray(struct hfi1_filedata *fd, u32 tidinfo,
  32                               struct tid_group **grp);
  33 static void clear_tid_node(struct hfi1_filedata *fd, struct tid_rb_node *node);
  34
  35 static const struct mmu_interval_notifier_ops tid_mn_ops = {
  36         .invalidate = tid_rb_invalidate,
  37 };
  38
  39 /*
  40  * Initialize context and file private data needed for Expected
  41  * receive caching. This needs to be done after the context has
  42  * been configured with the eager/expected RcvEntry counts.
  43  */
  44 int hfi1_user_exp_rcv_init(struct hfi1_filedata *fd,
  45                            struct hfi1_ctxtdata *uctxt)
  46 {
  47         int ret = 0;
  48
  49         fd->entry_to_rb = kcalloc(uctxt->expected_count,
  50                                   sizeof(struct rb_node *),
  51                                   GFP_KERNEL);
  52         if (!fd->entry_to_rb)
  53                 return -ENOMEM;
  54
  55         if (!HFI1_CAP_UGET_MASK(uctxt->flags, TID_UNMAP)) {
  56                 fd->invalid_tid_idx = 0;
  57                 fd->invalid_tids = kcalloc(uctxt->expected_count,
  58                                            sizeof(*fd->invalid_tids),
  59                                            GFP_KERNEL);
  60                 if (!fd->invalid_tids) {
  61                         kfree(fd->entry_to_rb);
  62                         fd->entry_to_rb = NULL;
  63                         return -ENOMEM;
  64                 }
  65                 fd->use_mn = true;
  66         }
  67
  68         /*
  69          * PSM does not have a good way to separate, count, and
  70          * effectively enforce a limit on RcvArray entries used by
  71          * subctxts (when context sharing is used) when TID caching
  72          * is enabled. To help with that, we calculate a per-process
  73          * RcvArray entry share and enforce that.
  74          * If TID caching is not in use, PSM deals with usage on its
  75          * own. In that case, we allow any subctxt to take all of the
  76          * entries.
  77          *
  78          * Make sure that we set the tid counts only after successful
  79          * init.
  80          */
  81         spin_lock(&fd->tid_lock);
  82         if (uctxt->subctxt_cnt && fd->use_mn) {
  83                 u16 remainder;
  84
  85                 fd->tid_limit = uctxt->expected_count / uctxt->subctxt_cnt;
  86                 remainder = uctxt->expected_count % uctxt->subctxt_cnt;
  87                 if (remainder && fd->subctxt < remainder)
  88                         fd->tid_limit++;
  89         } else {
  90                 fd->tid_limit = uctxt->expected_count;
  91         }
  92         spin_unlock(&fd->tid_lock);
  93
  94         return ret;
  95 }
  96
  97 void hfi1_user_exp_rcv_free(struct hfi1_filedata *fd)
  98 {
  99         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 100
 101         mutex_lock(&uctxt->exp_mutex);
 102         if (!EXP_TID_SET_EMPTY(uctxt->tid_full_list))
 103                 unlock_exp_tids(uctxt, &uctxt->tid_full_list, fd);
 104         if (!EXP_TID_SET_EMPTY(uctxt->tid_used_list))
 105                 unlock_exp_tids(uctxt, &uctxt->tid_used_list, fd);
 106         mutex_unlock(&uctxt->exp_mutex);
 107
 108         kfree(fd->invalid_tids);
 109         fd->invalid_tids = NULL;
 110
 111         kfree(fd->entry_to_rb);
 112         fd->entry_to_rb = NULL;
 113 }
 114
 115 /*
 116  * Release pinned receive buffer pages.
 117  *
 118  * @mapped: true if the pages have been DMA mapped. false otherwise.
 119  * @idx: Index of the first page to unpin.
 120  * @npages: No of pages to unpin.
 121  *
 122  * If the pages have been DMA mapped (indicated by mapped parameter), their
 123  * info will be passed via a struct tid_rb_node. If they haven't been mapped,
 124  * their info will be passed via a struct tid_user_buf.
 125  */
 126 static void unpin_rcv_pages(struct hfi1_filedata *fd,
 127                             struct tid_user_buf *tidbuf,
 128                             struct tid_rb_node *node,
 129                             unsigned int idx,
 130                             unsigned int npages,
 131                             bool mapped)
 132 {
 133         struct page **pages;
 134         struct hfi1_devdata *dd = fd->uctxt->dd;
 135         struct mm_struct *mm;
 136
 137         if (mapped) {
 138                 dma_unmap_single(&dd->pcidev->dev, node->dma_addr,
 139                                  node->npages * PAGE_SIZE, DMA_FROM_DEVICE);
 140                 pages = &node->pages[idx];
 141                 mm = mm_from_tid_node(node);
 142         } else {
 143                 pages = &tidbuf->pages[idx];
 144                 mm = current->mm;
 145         }
 146         hfi1_release_user_pages(mm, pages, npages, mapped);
 147         fd->tid_n_pinned -= npages;
 148 }
 149
 150 /*
 151  * Pin receive buffer pages.
 152  */
 153 static int pin_rcv_pages(struct hfi1_filedata *fd, struct tid_user_buf *tidbuf)
 154 {
 155         int pinned;
 156         unsigned int npages;
 157         unsigned long vaddr = tidbuf->vaddr;
 158         struct page **pages = NULL;
 159         struct hfi1_devdata *dd = fd->uctxt->dd;
 160
 161         /* Get the number of pages the user buffer spans */
 162         npages = num_user_pages(vaddr, tidbuf->length);
 163         if (!npages)
 164                 return -EINVAL;
 165
 166         if (npages > fd->uctxt->expected_count) {
 167                 dd_dev_err(dd, "Expected buffer too big\n");
 168                 return -EINVAL;
 169         }
 170
 171         /* Allocate the array of struct page pointers needed for pinning */
 172         pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL);
 173         if (!pages)
 174                 return -ENOMEM;
 175
 176         /*
 177          * Pin all the pages of the user buffer. If we can't pin all the
 178          * pages, accept the amount pinned so far and program only that.
 179          * User space knows how to deal with partially programmed buffers.
 180          */
 181         if (!hfi1_can_pin_pages(dd, current->mm, fd->tid_n_pinned, npages)) {
 182                 kfree(pages);
 183                 return -ENOMEM;
 184         }
 185
 186         pinned = hfi1_acquire_user_pages(current->mm, vaddr, npages, true, pages);
 187         if (pinned <= 0) {
 188                 kfree(pages);
 189                 return pinned;
 190         }
 191         tidbuf->pages = pages;
 192         tidbuf->npages = npages;
 193         fd->tid_n_pinned += pinned;
 194         return pinned;
 195 }
 196
 197 /*
 198  * RcvArray entry allocation for Expected Receives is done by the
 199  * following algorithm:
 200  *
 201  * The context keeps 3 lists of groups of RcvArray entries:
 202  *   1. List of empty groups - tid_group_list
 203  *      This list is created during user context creation and
 204  *      contains elements which describe sets (of 8) of empty
 205  *      RcvArray entries.
 206  *   2. List of partially used groups - tid_used_list
 207  *      This list contains sets of RcvArray entries which are
 208  *      not completely used up. Another mapping request could
 209  *      use some of all of the remaining entries.
 210  *   3. List of full groups - tid_full_list
 211  *      This is the list where sets that are completely used
 212  *      up go.
 213  *
 214  * An attempt to optimize the usage of RcvArray entries is
 215  * made by finding all sets of physically contiguous pages in a
 216  * user's buffer.
 217  * These physically contiguous sets are further split into
 218  * sizes supported by the receive engine of the HFI. The
 219  * resulting sets of pages are stored in struct tid_pageset,
 220  * which describes the sets as:
 221  *    * .count - number of pages in this set
 222  *    * .idx - starting index into struct page ** array
 223  *                    of this set
 224  *
 225  * From this point on, the algorithm deals with the page sets
 226  * described above. The number of pagesets is divided by the
 227  * RcvArray group size to produce the number of full groups
 228  * needed.
 229  *
 230  * Groups from the 3 lists are manipulated using the following
 231  * rules:
 232  *   1. For each set of 8 pagesets, a complete group from
 233  *      tid_group_list is taken, programmed, and moved to
 234  *      the tid_full_list list.
 235  *   2. For all remaining pagesets:
 236  *      2.1 If the tid_used_list is empty and the tid_group_list
 237  *          is empty, stop processing pageset and return only
 238  *          what has been programmed up to this point.
 239  *      2.2 If the tid_used_list is empty and the tid_group_list
 240  *          is not empty, move a group from tid_group_list to
 241  *          tid_used_list.
 242  *      2.3 For each group is tid_used_group, program as much as
 243  *          can fit into the group. If the group becomes fully
 244  *          used, move it to tid_full_list.
 245  */
 246 int hfi1_user_exp_rcv_setup(struct hfi1_filedata *fd,
 247                             struct hfi1_tid_info *tinfo)
 248 {
 249         int ret = 0, need_group = 0, pinned;
 250         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 251         struct hfi1_devdata *dd = uctxt->dd;
 252         unsigned int ngroups, pageidx = 0, pageset_count,
 253                 tididx = 0, mapped, mapped_pages = 0;
 254         u32 *tidlist = NULL;
 255         struct tid_user_buf *tidbuf;
 256
 257         if (!PAGE_ALIGNED(tinfo->vaddr))
 258                 return -EINVAL;
 259         if (tinfo->length == 0)
 260                 return -EINVAL;
 261
 262         tidbuf = kzalloc(sizeof(*tidbuf), GFP_KERNEL);
 263         if (!tidbuf)
 264                 return -ENOMEM;
 265
 266         tidbuf->vaddr = tinfo->vaddr;
 267         tidbuf->length = tinfo->length;
 268         tidbuf->psets = kcalloc(uctxt->expected_count, sizeof(*tidbuf->psets),
 269                                 GFP_KERNEL);
 270         if (!tidbuf->psets) {
 271                 kfree(tidbuf);
 272                 return -ENOMEM;
 273         }
 274
 275         pinned = pin_rcv_pages(fd, tidbuf);
 276         if (pinned <= 0) {
 277                 kfree(tidbuf->psets);
 278                 kfree(tidbuf);
 279                 return pinned;
 280         }
 281
 282         /* Find sets of physically contiguous pages */
 283         tidbuf->n_psets = find_phys_blocks(tidbuf, pinned);
 284
 285         /*
 286          * We don't need to access this under a lock since tid_used is per
 287          * process and the same process cannot be in hfi1_user_exp_rcv_clear()
 288          * and hfi1_user_exp_rcv_setup() at the same time.
 289          */
 290         spin_lock(&fd->tid_lock);
 291         if (fd->tid_used + tidbuf->n_psets > fd->tid_limit)
 292                 pageset_count = fd->tid_limit - fd->tid_used;
 293         else
 294                 pageset_count = tidbuf->n_psets;
 295         spin_unlock(&fd->tid_lock);
 296
 297         if (!pageset_count)
 298                 goto bail;
 299
 300         ngroups = pageset_count / dd->rcv_entries.group_size;
 301         tidlist = kcalloc(pageset_count, sizeof(*tidlist), GFP_KERNEL);
 302         if (!tidlist) {
 303                 ret = -ENOMEM;
 304                 goto nomem;
 305         }
 306
 307         tididx = 0;
 308
 309         /*
 310          * From this point on, we are going to be using shared (between master
 311          * and subcontexts) context resources. We need to take the lock.
 312          */
 313         mutex_lock(&uctxt->exp_mutex);
 314         /*
 315          * The first step is to program the RcvArray entries which are complete
 316          * groups.
 317          */
 318         while (ngroups && uctxt->tid_group_list.count) {
 319                 struct tid_group *grp =
 320                         tid_group_pop(&uctxt->tid_group_list);
 321
 322                 ret = program_rcvarray(fd, tidbuf, grp,
 323                                        pageidx, dd->rcv_entries.group_size,
 324                                        tidlist, &tididx, &mapped);
 325                 /*
 326                  * If there was a failure to program the RcvArray
 327                  * entries for the entire group, reset the grp fields
 328                  * and add the grp back to the free group list.
 329                  */
 330                 if (ret <= 0) {
 331                         tid_group_add_tail(grp, &uctxt->tid_group_list);
 332                         hfi1_cdbg(TID,
 333                                   "Failed to program RcvArray group %d", ret);
 334                         goto unlock;
 335                 }
 336
 337                 tid_group_add_tail(grp, &uctxt->tid_full_list);
 338                 ngroups--;
 339                 pageidx += ret;
 340                 mapped_pages += mapped;
 341         }
 342
 343         while (pageidx < pageset_count) {
 344                 struct tid_group *grp, *ptr;
 345                 /*
 346                  * If we don't have any partially used tid groups, check
 347                  * if we have empty groups. If so, take one from there and
 348                  * put in the partially used list.
 349                  */
 350                 if (!uctxt->tid_used_list.count || need_group) {
 351                         if (!uctxt->tid_group_list.count)
 352                                 goto unlock;
 353
 354                         grp = tid_group_pop(&uctxt->tid_group_list);
 355                         tid_group_add_tail(grp, &uctxt->tid_used_list);
 356                         need_group = 0;
 357                 }
 358                 /*
 359                  * There is an optimization opportunity here - instead of
 360                  * fitting as many page sets as we can, check for a group
 361                  * later on in the list that could fit all of them.
 362                  */
 363                 list_for_each_entry_safe(grp, ptr, &uctxt->tid_used_list.list,
 364                                          list) {
 365                         unsigned use = min_t(unsigned, pageset_count - pageidx,
 366                                              grp->size - grp->used);
 367
 368                         ret = program_rcvarray(fd, tidbuf, grp,
 369                                                pageidx, use, tidlist,
 370                                                &tididx, &mapped);
 371                         if (ret < 0) {
 372                                 hfi1_cdbg(TID,
 373                                           "Failed to program RcvArray entries %d",
 374                                           ret);
 375                                 goto unlock;
 376                         } else if (ret > 0) {
 377                                 if (grp->used == grp->size)
 378                                         tid_group_move(grp,
 379                                                        &uctxt->tid_used_list,
 380                                                        &uctxt->tid_full_list);
 381                                 pageidx += ret;
 382                                 mapped_pages += mapped;
 383                                 need_group = 0;
 384                                 /* Check if we are done so we break out early */
 385                                 if (pageidx >= pageset_count)
 386                                         break;
 387                         } else if (WARN_ON(ret == 0)) {
 388                                 /*
 389                                  * If ret is 0, we did not program any entries
 390                                  * into this group, which can only happen if
 391                                  * we've screwed up the accounting somewhere.
 392                                  * Warn and try to continue.
 393                                  */
 394                                 need_group = 1;
 395                         }
 396                 }
 397         }
 398 unlock:
 399         mutex_unlock(&uctxt->exp_mutex);
 400 nomem:
 401         hfi1_cdbg(TID, "total mapped: tidpairs:%u pages:%u (%d)", tididx,
 402                   mapped_pages, ret);
 403         if (tididx) {
 404                 spin_lock(&fd->tid_lock);
 405                 fd->tid_used += tididx;
 406                 spin_unlock(&fd->tid_lock);
 407                 tinfo->tidcnt = tididx;
 408                 tinfo->length = mapped_pages * PAGE_SIZE;
 409
 410                 if (copy_to_user(u64_to_user_ptr(tinfo->tidlist),
 411                                  tidlist, sizeof(tidlist[0]) * tididx)) {
 412                         /*
 413                          * On failure to copy to the user level, we need to undo
 414                          * everything done so far so we don't leak resources.
 415                          */
 416                         tinfo->tidlist = (unsigned long)&tidlist;
 417                         hfi1_user_exp_rcv_clear(fd, tinfo);
 418                         tinfo->tidlist = 0;
 419                         ret = -EFAULT;
 420                         goto bail;
 421                 }
 422         }
 423
 424         /*
 425          * If not everything was mapped (due to insufficient RcvArray entries,
 426          * for example), unpin all unmapped pages so we can pin them nex time.
 427          */
 428         if (mapped_pages != pinned)
 429                 unpin_rcv_pages(fd, tidbuf, NULL, mapped_pages,
 430                                 (pinned - mapped_pages), false);
 431 bail:
 432         kfree(tidbuf->psets);
 433         kfree(tidlist);
 434         kfree(tidbuf->pages);
 435         kfree(tidbuf);
 436         return ret > 0 ? 0 : ret;
 437 }
 438
 439 int hfi1_user_exp_rcv_clear(struct hfi1_filedata *fd,
 440                             struct hfi1_tid_info *tinfo)
 441 {
 442         int ret = 0;
 443         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 444         u32 *tidinfo;
 445         unsigned tididx;
 446
 447         if (unlikely(tinfo->tidcnt > fd->tid_used))
 448                 return -EINVAL;
 449
 450         tidinfo = memdup_user(u64_to_user_ptr(tinfo->tidlist),
 451                               sizeof(tidinfo[0]) * tinfo->tidcnt);
 452         if (IS_ERR(tidinfo))
 453                 return PTR_ERR(tidinfo);
 454
 455         mutex_lock(&uctxt->exp_mutex);
 456         for (tididx = 0; tididx < tinfo->tidcnt; tididx++) {
 457                 ret = unprogram_rcvarray(fd, tidinfo[tididx], NULL);
 458                 if (ret) {
 459                         hfi1_cdbg(TID, "Failed to unprogram rcv array %d",
 460                                   ret);
 461                         break;
 462                 }
 463         }
 464         spin_lock(&fd->tid_lock);
 465         fd->tid_used -= tididx;
 466         spin_unlock(&fd->tid_lock);
 467         tinfo->tidcnt = tididx;
 468         mutex_unlock(&uctxt->exp_mutex);
 469
 470         kfree(tidinfo);
 471         return ret;
 472 }
 473
 474 int hfi1_user_exp_rcv_invalid(struct hfi1_filedata *fd,
 475                               struct hfi1_tid_info *tinfo)
 476 {
 477         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 478         unsigned long *ev = uctxt->dd->events +
 479                 (uctxt_offset(uctxt) + fd->subctxt);
 480         u32 *array;
 481         int ret = 0;
 482
 483         /*
 484          * copy_to_user() can sleep, which will leave the invalid_lock
 485          * locked and cause the MMU notifier to be blocked on the lock
 486          * for a long time.
 487          * Copy the data to a local buffer so we can release the lock.
 488          */
 489         array = kcalloc(uctxt->expected_count, sizeof(*array), GFP_KERNEL);
 490         if (!array)
 491                 return -EFAULT;
 492
 493         spin_lock(&fd->invalid_lock);
 494         if (fd->invalid_tid_idx) {
 495                 memcpy(array, fd->invalid_tids, sizeof(*array) *
 496                        fd->invalid_tid_idx);
 497                 memset(fd->invalid_tids, 0, sizeof(*fd->invalid_tids) *
 498                        fd->invalid_tid_idx);
 499                 tinfo->tidcnt = fd->invalid_tid_idx;
 500                 fd->invalid_tid_idx = 0;
 501                 /*
 502                  * Reset the user flag while still holding the lock.
 503                  * Otherwise, PSM can miss events.
 504                  */
 505                 clear_bit(_HFI1_EVENT_TID_MMU_NOTIFY_BIT, ev);
 506         } else {
 507                 tinfo->tidcnt = 0;
 508         }
 509         spin_unlock(&fd->invalid_lock);
 510
 511         if (tinfo->tidcnt) {
 512                 if (copy_to_user((void __user *)tinfo->tidlist,
 513                                  array, sizeof(*array) * tinfo->tidcnt))
 514                         ret = -EFAULT;
 515         }
 516         kfree(array);
 517
 518         return ret;
 519 }
 520
 521 static u32 find_phys_blocks(struct tid_user_buf *tidbuf, unsigned int npages)
 522 {
 523         unsigned pagecount, pageidx, setcount = 0, i;
 524         unsigned long pfn, this_pfn;
 525         struct page **pages = tidbuf->pages;
 526         struct tid_pageset *list = tidbuf->psets;
 527
 528         if (!npages)
 529                 return 0;
 530
 531         /*
 532          * Look for sets of physically contiguous pages in the user buffer.
 533          * This will allow us to optimize Expected RcvArray entry usage by
 534          * using the bigger supported sizes.
 535          */
 536         pfn = page_to_pfn(pages[0]);
 537         for (pageidx = 0, pagecount = 1, i = 1; i <= npages; i++) {
 538                 this_pfn = i < npages ? page_to_pfn(pages[i]) : 0;
 539
 540                 /*
 541                  * If the pfn's are not sequential, pages are not physically
 542                  * contiguous.
 543                  */
 544                 if (this_pfn != ++pfn) {
 545                         /*
 546                          * At this point we have to loop over the set of
 547                          * physically contiguous pages and break them down it
 548                          * sizes supported by the HW.
 549                          * There are two main constraints:
 550                          *     1. The max buffer size is MAX_EXPECTED_BUFFER.
 551                          *        If the total set size is bigger than that
 552                          *        program only a MAX_EXPECTED_BUFFER chunk.
 553                          *     2. The buffer size has to be a power of two. If
 554                          *        it is not, round down to the closes power of
 555                          *        2 and program that size.
 556                          */
 557                         while (pagecount) {
 558                                 int maxpages = pagecount;
 559                                 u32 bufsize = pagecount * PAGE_SIZE;
 560
 561                                 if (bufsize > MAX_EXPECTED_BUFFER)
 562                                         maxpages =
 563                                                 MAX_EXPECTED_BUFFER >>
 564                                                 PAGE_SHIFT;
 565                                 else if (!is_power_of_2(bufsize))
 566                                         maxpages =
 567                                                 rounddown_pow_of_two(bufsize) >>
 568                                                 PAGE_SHIFT;
 569
 570                                 list[setcount].idx = pageidx;
 571                                 list[setcount].count = maxpages;
 572                                 pagecount -= maxpages;
 573                                 pageidx += maxpages;
 574                                 setcount++;
 575                         }
 576                         pageidx = i;
 577                         pagecount = 1;
 578                         pfn = this_pfn;
 579                 } else {
 580                         pagecount++;
 581                 }
 582         }
 583         return setcount;
 584 }
 585
 586 /**
 587  * program_rcvarray() - program an RcvArray group with receive buffers
 588  * @fd: filedata pointer
 589  * @tbuf: pointer to struct tid_user_buf that has the user buffer starting
 590  *        virtual address, buffer length, page pointers, pagesets (array of
 591  *        struct tid_pageset holding information on physically contiguous
 592  *        chunks from the user buffer), and other fields.
 593  * @grp: RcvArray group
 594  * @start: starting index into sets array
 595  * @count: number of struct tid_pageset's to program
 596  * @tidlist: the array of u32 elements when the information about the
 597  *           programmed RcvArray entries is to be encoded.
 598  * @tididx: starting offset into tidlist
 599  * @pmapped: (output parameter) number of pages programmed into the RcvArray
 600  *           entries.
 601  *
 602  * This function will program up to 'count' number of RcvArray entries from the
 603  * group 'grp'. To make best use of write-combining writes, the function will
 604  * perform writes to the unused RcvArray entries which will be ignored by the
 605  * HW. Each RcvArray entry will be programmed with a physically contiguous
 606  * buffer chunk from the user's virtual buffer.
 607  *
 608  * Return:
 609  * -EINVAL if the requested count is larger than the size of the group,
 610  * -ENOMEM or -EFAULT on error from set_rcvarray_entry(), or
 611  * number of RcvArray entries programmed.
 612  */
 613 static int program_rcvarray(struct hfi1_filedata *fd, struct tid_user_buf *tbuf,
 614                             struct tid_group *grp,
 615                             unsigned int start, u16 count,
 616                             u32 *tidlist, unsigned int *tididx,
 617                             unsigned int *pmapped)
 618 {
 619         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 620         struct hfi1_devdata *dd = uctxt->dd;
 621         u16 idx;
 622         u32 tidinfo = 0, rcventry, useidx = 0;
 623         int mapped = 0;
 624
 625         /* Count should never be larger than the group size */
 626         if (count > grp->size)
 627                 return -EINVAL;
 628
 629         /* Find the first unused entry in the group */
 630         for (idx = 0; idx < grp->size; idx++) {
 631                 if (!(grp->map & (1 << idx))) {
 632                         useidx = idx;
 633                         break;
 634                 }
 635                 rcv_array_wc_fill(dd, grp->base + idx);
 636         }
 637
 638         idx = 0;
 639         while (idx < count) {
 640                 u16 npages, pageidx, setidx = start + idx;
 641                 int ret = 0;
 642
 643                 /*
 644                  * If this entry in the group is used, move to the next one.
 645                  * If we go past the end of the group, exit the loop.
 646                  */
 647                 if (useidx >= grp->size) {
 648                         break;
 649                 } else if (grp->map & (1 << useidx)) {
 650                         rcv_array_wc_fill(dd, grp->base + useidx);
 651                         useidx++;
 652                         continue;
 653                 }
 654
 655                 rcventry = grp->base + useidx;
 656                 npages = tbuf->psets[setidx].count;
 657                 pageidx = tbuf->psets[setidx].idx;
 658
 659                 ret = set_rcvarray_entry(fd, tbuf,
 660                                          rcventry, grp, pageidx,
 661                                          npages);
 662                 if (ret)
 663                         return ret;
 664                 mapped += npages;
 665
 666                 tidinfo = rcventry2tidinfo(rcventry - uctxt->expected_base) |
 667                         EXP_TID_SET(LEN, npages);
 668                 tidlist[(*tididx)++] = tidinfo;
 669                 grp->used++;
 670                 grp->map |= 1 << useidx++;
 671                 idx++;
 672         }
 673
 674         /* Fill the rest of the group with "blank" writes */
 675         for (; useidx < grp->size; useidx++)
 676                 rcv_array_wc_fill(dd, grp->base + useidx);
 677         *pmapped = mapped;
 678         return idx;
 679 }
 680
 681 static int set_rcvarray_entry(struct hfi1_filedata *fd,
 682                               struct tid_user_buf *tbuf,
 683                               u32 rcventry, struct tid_group *grp,
 684                               u16 pageidx, unsigned int npages)
 685 {
 686         int ret;
 687         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 688         struct tid_rb_node *node;
 689         struct hfi1_devdata *dd = uctxt->dd;
 690         dma_addr_t phys;
 691         struct page **pages = tbuf->pages + pageidx;
 692
 693         /*
 694          * Allocate the node first so we can handle a potential
 695          * failure before we've programmed anything.
 696          */
 697         node = kzalloc(sizeof(*node) + (sizeof(struct page *) * npages),
 698                        GFP_KERNEL);
 699         if (!node)
 700                 return -ENOMEM;
 701
 702         phys = dma_map_single(&dd->pcidev->dev, __va(page_to_phys(pages[0])),
 703                               npages * PAGE_SIZE, DMA_FROM_DEVICE);
 704         if (dma_mapping_error(&dd->pcidev->dev, phys)) {
 705                 dd_dev_err(dd, "Failed to DMA map Exp Rcv pages 0x%llx\n",
 706                            phys);
 707                 kfree(node);
 708                 return -EFAULT;
 709         }
 710
 711         node->fdata = fd;
 712         node->phys = page_to_phys(pages[0]);
 713         node->npages = npages;
 714         node->rcventry = rcventry;
 715         node->dma_addr = phys;
 716         node->grp = grp;
 717         node->freed = false;
 718         memcpy(node->pages, pages, sizeof(struct page *) * npages);
 719
 720         if (fd->use_mn) {
 721                 ret = mmu_interval_notifier_insert(
 722                         &node->notifier, current->mm,
 723                         tbuf->vaddr + (pageidx * PAGE_SIZE), npages * PAGE_SIZE,
 724                         &tid_mn_ops);
 725                 if (ret)
 726                         goto out_unmap;
 727                 /*
 728                  * FIXME: This is in the wrong order, the notifier should be
 729                  * established before the pages are pinned by pin_rcv_pages.
 730                  */
 731                 mmu_interval_read_begin(&node->notifier);
 732         }
 733         fd->entry_to_rb[node->rcventry - uctxt->expected_base] = node;
 734
 735         hfi1_put_tid(dd, rcventry, PT_EXPECTED, phys, ilog2(npages) + 1);
 736         trace_hfi1_exp_tid_reg(uctxt->ctxt, fd->subctxt, rcventry, npages,
 737                                node->notifier.interval_tree.start, node->phys,
 738                                phys);
 739         return 0;
 740
 741 out_unmap:
 742         hfi1_cdbg(TID, "Failed to insert RB node %u 0x%lx, 0x%lx %d",
 743                   node->rcventry, node->notifier.interval_tree.start,
 744                   node->phys, ret);
 745         dma_unmap_single(&dd->pcidev->dev, phys, npages * PAGE_SIZE,
 746                          DMA_FROM_DEVICE);
 747         kfree(node);
 748         return -EFAULT;
 749 }
 750
 751 static int unprogram_rcvarray(struct hfi1_filedata *fd, u32 tidinfo,
 752                               struct tid_group **grp)
 753 {
 754         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 755         struct hfi1_devdata *dd = uctxt->dd;
 756         struct tid_rb_node *node;
 757         u8 tidctrl = EXP_TID_GET(tidinfo, CTRL);
 758         u32 tididx = EXP_TID_GET(tidinfo, IDX) << 1, rcventry;
 759
 760         if (tididx >= uctxt->expected_count) {
 761                 dd_dev_err(dd, "Invalid RcvArray entry (%u) index for ctxt %u\n",
 762                            tididx, uctxt->ctxt);
 763                 return -EINVAL;
 764         }
 765
 766         if (tidctrl == 0x3)
 767                 return -EINVAL;
 768
 769         rcventry = tididx + (tidctrl - 1);
 770
 771         node = fd->entry_to_rb[rcventry];
 772         if (!node || node->rcventry != (uctxt->expected_base + rcventry))
 773                 return -EBADF;
 774
 775         if (grp)
 776                 *grp = node->grp;
 777
 778         if (fd->use_mn)
 779                 mmu_interval_notifier_remove(&node->notifier);
 780         cacheless_tid_rb_remove(fd, node);
 781
 782         return 0;
 783 }
 784
 785 static void clear_tid_node(struct hfi1_filedata *fd, struct tid_rb_node *node)
 786 {
 787         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 788         struct hfi1_devdata *dd = uctxt->dd;
 789
 790         trace_hfi1_exp_tid_unreg(uctxt->ctxt, fd->subctxt, node->rcventry,
 791                                  node->npages,
 792                                  node->notifier.interval_tree.start, node->phys,
 793                                  node->dma_addr);
 794
 795         /*
 796          * Make sure device has seen the write before we unpin the
 797          * pages.
 798          */
 799         hfi1_put_tid(dd, node->rcventry, PT_INVALID_FLUSH, 0, 0);
 800
 801         unpin_rcv_pages(fd, NULL, node, 0, node->npages, true);
 802
 803         node->grp->used--;
 804         node->grp->map &= ~(1 << (node->rcventry - node->grp->base));
 805
 806         if (node->grp->used == node->grp->size - 1)
 807                 tid_group_move(node->grp, &uctxt->tid_full_list,
 808                                &uctxt->tid_used_list);
 809         else if (!node->grp->used)
 810                 tid_group_move(node->grp, &uctxt->tid_used_list,
 811                                &uctxt->tid_group_list);
 812         kfree(node);
 813 }
 814
 815 /*
 816  * As a simple helper for hfi1_user_exp_rcv_free, this function deals with
 817  * clearing nodes in the non-cached case.
 818  */
 819 static void unlock_exp_tids(struct hfi1_ctxtdata *uctxt,
 820                             struct exp_tid_set *set,
 821                             struct hfi1_filedata *fd)
 822 {
 823         struct tid_group *grp, *ptr;
 824         int i;
 825
 826         list_for_each_entry_safe(grp, ptr, &set->list, list) {
 827                 list_del_init(&grp->list);
 828
 829                 for (i = 0; i < grp->size; i++) {
 830                         if (grp->map & (1 << i)) {
 831                                 u16 rcventry = grp->base + i;
 832                                 struct tid_rb_node *node;
 833
 834                                 node = fd->entry_to_rb[rcventry -
 835                                                           uctxt->expected_base];
 836                                 if (!node || node->rcventry != rcventry)
 837                                         continue;
 838
 839                                 if (fd->use_mn)
 840                                         mmu_interval_notifier_remove(
 841                                                 &node->notifier);
 842                                 cacheless_tid_rb_remove(fd, node);
 843                         }
 844                 }
 845         }
 846 }
 847
 848 static bool tid_rb_invalidate(struct mmu_interval_notifier *mni,
 849                               const struct mmu_notifier_range *range,
 850                               unsigned long cur_seq)
 851 {
 852         struct tid_rb_node *node =
 853                 container_of(mni, struct tid_rb_node, notifier);
 854         struct hfi1_filedata *fdata = node->fdata;
 855         struct hfi1_ctxtdata *uctxt = fdata->uctxt;
 856
 857         if (node->freed)
 858                 return true;
 859
 860         trace_hfi1_exp_tid_inval(uctxt->ctxt, fdata->subctxt,
 861                                  node->notifier.interval_tree.start,
 862                                  node->rcventry, node->npages, node->dma_addr);
 863         node->freed = true;
 864
 865         spin_lock(&fdata->invalid_lock);
 866         if (fdata->invalid_tid_idx < uctxt->expected_count) {
 867                 fdata->invalid_tids[fdata->invalid_tid_idx] =
 868                         rcventry2tidinfo(node->rcventry - uctxt->expected_base);
 869                 fdata->invalid_tids[fdata->invalid_tid_idx] |=
 870                         EXP_TID_SET(LEN, node->npages);
 871                 if (!fdata->invalid_tid_idx) {
 872                         unsigned long *ev;
 873
 874                         /*
 875                          * hfi1_set_uevent_bits() sets a user event flag
 876                          * for all processes. Because calling into the
 877                          * driver to process TID cache invalidations is
 878                          * expensive and TID cache invalidations are
 879                          * handled on a per-process basis, we can
 880                          * optimize this to set the flag only for the
 881                          * process in question.
 882                          */
 883                         ev = uctxt->dd->events +
 884                                 (uctxt_offset(uctxt) + fdata->subctxt);
 885                         set_bit(_HFI1_EVENT_TID_MMU_NOTIFY_BIT, ev);
 886                 }
 887                 fdata->invalid_tid_idx++;
 888         }
 889         spin_unlock(&fdata->invalid_lock);
 890         return true;
 891 }
 892
 893 static void cacheless_tid_rb_remove(struct hfi1_filedata *fdata,
 894                                     struct tid_rb_node *tnode)
 895 {
 896         u32 base = fdata->uctxt->expected_base;
 897
 898         fdata->entry_to_rb[tnode->rcventry - base] = NULL;
 899         clear_tid_node(fdata, tnode);
 900 }