drivers/infiniband/hw/hfi1/user_exp_rcv.c

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