drivers/infiniband/hw/hfi1/user_exp_rcv.c

   1 /*
   2  * Copyright(c) 2015-2017 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                 if (!EXP_TID_SET_EMPTY(uctxt->tid_full_list))
 166                         unlock_exp_tids(uctxt, &uctxt->tid_full_list, fd);
 167                 if (!EXP_TID_SET_EMPTY(uctxt->tid_used_list))
 168                         unlock_exp_tids(uctxt, &uctxt->tid_used_list, fd);
 169         }
 170
 171         kfree(fd->invalid_tids);
 172         fd->invalid_tids = NULL;
 173
 174         kfree(fd->entry_to_rb);
 175         fd->entry_to_rb = NULL;
 176 }
 177
 178 /**
 179  * Release pinned receive buffer pages.
 180  *
 181  * @mapped - true if the pages have been DMA mapped. false otherwise.
 182  * @idx - Index of the first page to unpin.
 183  * @npages - No of pages to unpin.
 184  *
 185  * If the pages have been DMA mapped (indicated by mapped parameter), their
 186  * info will be passed via a struct tid_rb_node. If they haven't been mapped,
 187  * their info will be passed via a struct tid_user_buf.
 188  */
 189 static void unpin_rcv_pages(struct hfi1_filedata *fd,
 190                             struct tid_user_buf *tidbuf,
 191                             struct tid_rb_node *node,
 192                             unsigned int idx,
 193                             unsigned int npages,
 194                             bool mapped)
 195 {
 196         struct page **pages;
 197         struct hfi1_devdata *dd = fd->uctxt->dd;
 198
 199         if (mapped) {
 200                 pci_unmap_single(dd->pcidev, node->dma_addr,
 201                                  node->mmu.len, PCI_DMA_FROMDEVICE);
 202                 pages = &node->pages[idx];
 203         } else {
 204                 pages = &tidbuf->pages[idx];
 205         }
 206         hfi1_release_user_pages(fd->mm, pages, npages, mapped);
 207         fd->tid_n_pinned -= npages;
 208 }
 209
 210 /**
 211  * Pin receive buffer pages.
 212  */
 213 static int pin_rcv_pages(struct hfi1_filedata *fd, struct tid_user_buf *tidbuf)
 214 {
 215         int pinned;
 216         unsigned int npages = tidbuf->npages;
 217         unsigned long vaddr = tidbuf->vaddr;
 218         struct page **pages = NULL;
 219         struct hfi1_devdata *dd = fd->uctxt->dd;
 220
 221         if (npages > fd->uctxt->expected_count) {
 222                 dd_dev_err(dd, "Expected buffer too big\n");
 223                 return -EINVAL;
 224         }
 225
 226         /* Verify that access is OK for the user buffer */
 227         if (!access_ok(VERIFY_WRITE, (void __user *)vaddr,
 228                        npages * PAGE_SIZE)) {
 229                 dd_dev_err(dd, "Fail vaddr %p, %u pages, !access_ok\n",
 230                            (void *)vaddr, npages);
 231                 return -EFAULT;
 232         }
 233         /* Allocate the array of struct page pointers needed for pinning */
 234         pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL);
 235         if (!pages)
 236                 return -ENOMEM;
 237
 238         /*
 239          * Pin all the pages of the user buffer. If we can't pin all the
 240          * pages, accept the amount pinned so far and program only that.
 241          * User space knows how to deal with partially programmed buffers.
 242          */
 243         if (!hfi1_can_pin_pages(dd, fd->mm, fd->tid_n_pinned, npages)) {
 244                 kfree(pages);
 245                 return -ENOMEM;
 246         }
 247
 248         pinned = hfi1_acquire_user_pages(fd->mm, vaddr, npages, true, pages);
 249         if (pinned <= 0) {
 250                 kfree(pages);
 251                 return pinned;
 252         }
 253         tidbuf->pages = pages;
 254         fd->tid_n_pinned += pinned;
 255         return pinned;
 256 }
 257
 258 /*
 259  * RcvArray entry allocation for Expected Receives is done by the
 260  * following algorithm:
 261  *
 262  * The context keeps 3 lists of groups of RcvArray entries:
 263  *   1. List of empty groups - tid_group_list
 264  *      This list is created during user context creation and
 265  *      contains elements which describe sets (of 8) of empty
 266  *      RcvArray entries.
 267  *   2. List of partially used groups - tid_used_list
 268  *      This list contains sets of RcvArray entries which are
 269  *      not completely used up. Another mapping request could
 270  *      use some of all of the remaining entries.
 271  *   3. List of full groups - tid_full_list
 272  *      This is the list where sets that are completely used
 273  *      up go.
 274  *
 275  * An attempt to optimize the usage of RcvArray entries is
 276  * made by finding all sets of physically contiguous pages in a
 277  * user's buffer.
 278  * These physically contiguous sets are further split into
 279  * sizes supported by the receive engine of the HFI. The
 280  * resulting sets of pages are stored in struct tid_pageset,
 281  * which describes the sets as:
 282  *    * .count - number of pages in this set
 283  *    * .idx - starting index into struct page ** array
 284  *                    of this set
 285  *
 286  * From this point on, the algorithm deals with the page sets
 287  * described above. The number of pagesets is divided by the
 288  * RcvArray group size to produce the number of full groups
 289  * needed.
 290  *
 291  * Groups from the 3 lists are manipulated using the following
 292  * rules:
 293  *   1. For each set of 8 pagesets, a complete group from
 294  *      tid_group_list is taken, programmed, and moved to
 295  *      the tid_full_list list.
 296  *   2. For all remaining pagesets:
 297  *      2.1 If the tid_used_list is empty and the tid_group_list
 298  *          is empty, stop processing pageset and return only
 299  *          what has been programmed up to this point.
 300  *      2.2 If the tid_used_list is empty and the tid_group_list
 301  *          is not empty, move a group from tid_group_list to
 302  *          tid_used_list.
 303  *      2.3 For each group is tid_used_group, program as much as
 304  *          can fit into the group. If the group becomes fully
 305  *          used, move it to tid_full_list.
 306  */
 307 int hfi1_user_exp_rcv_setup(struct hfi1_filedata *fd,
 308                             struct hfi1_tid_info *tinfo)
 309 {
 310         int ret = 0, need_group = 0, pinned;
 311         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 312         struct hfi1_devdata *dd = uctxt->dd;
 313         unsigned int ngroups, pageidx = 0, pageset_count,
 314                 tididx = 0, mapped, mapped_pages = 0;
 315         u32 *tidlist = NULL;
 316         struct tid_user_buf *tidbuf;
 317
 318         if (!PAGE_ALIGNED(tinfo->vaddr))
 319                 return -EINVAL;
 320         if (tinfo->length == 0)
 321                 return -EINVAL;
 322
 323         tidbuf = kzalloc(sizeof(*tidbuf), GFP_KERNEL);
 324         if (!tidbuf)
 325                 return -ENOMEM;
 326
 327         tidbuf->vaddr = tinfo->vaddr;
 328         tidbuf->length = tinfo->length;
 329         tidbuf->npages = num_user_pages(tidbuf->vaddr, tidbuf->length);
 330         tidbuf->psets = kcalloc(uctxt->expected_count, sizeof(*tidbuf->psets),
 331                                 GFP_KERNEL);
 332         if (!tidbuf->psets) {
 333                 kfree(tidbuf);
 334                 return -ENOMEM;
 335         }
 336
 337         pinned = pin_rcv_pages(fd, tidbuf);
 338         if (pinned <= 0) {
 339                 kfree(tidbuf->psets);
 340                 kfree(tidbuf);
 341                 return pinned;
 342         }
 343
 344         /* Find sets of physically contiguous pages */
 345         tidbuf->n_psets = find_phys_blocks(tidbuf, pinned);
 346
 347         /* Reserve the number of expected tids to be used. */
 348         spin_lock(&fd->tid_lock);
 349         if (fd->tid_used + tidbuf->n_psets > fd->tid_limit)
 350                 pageset_count = fd->tid_limit - fd->tid_used;
 351         else
 352                 pageset_count = tidbuf->n_psets;
 353         fd->tid_used += pageset_count;
 354         spin_unlock(&fd->tid_lock);
 355
 356         if (!pageset_count)
 357                 goto bail;
 358
 359         ngroups = pageset_count / dd->rcv_entries.group_size;
 360         tidlist = kcalloc(pageset_count, sizeof(*tidlist), GFP_KERNEL);
 361         if (!tidlist) {
 362                 ret = -ENOMEM;
 363                 goto nomem;
 364         }
 365
 366         tididx = 0;
 367
 368         /*
 369          * From this point on, we are going to be using shared (between master
 370          * and subcontexts) context resources. We need to take the lock.
 371          */
 372         mutex_lock(&uctxt->exp_lock);
 373         /*
 374          * The first step is to program the RcvArray entries which are complete
 375          * groups.
 376          */
 377         while (ngroups && uctxt->tid_group_list.count) {
 378                 struct tid_group *grp =
 379                         tid_group_pop(&uctxt->tid_group_list);
 380
 381                 ret = program_rcvarray(fd, tidbuf, grp,
 382                                        pageidx, dd->rcv_entries.group_size,
 383                                        tidlist, &tididx, &mapped);
 384                 /*
 385                  * If there was a failure to program the RcvArray
 386                  * entries for the entire group, reset the grp fields
 387                  * and add the grp back to the free group list.
 388                  */
 389                 if (ret <= 0) {
 390                         tid_group_add_tail(grp, &uctxt->tid_group_list);
 391                         hfi1_cdbg(TID,
 392                                   "Failed to program RcvArray group %d", ret);
 393                         goto unlock;
 394                 }
 395
 396                 tid_group_add_tail(grp, &uctxt->tid_full_list);
 397                 ngroups--;
 398                 pageidx += ret;
 399                 mapped_pages += mapped;
 400         }
 401
 402         while (pageidx < pageset_count) {
 403                 struct tid_group *grp, *ptr;
 404                 /*
 405                  * If we don't have any partially used tid groups, check
 406                  * if we have empty groups. If so, take one from there and
 407                  * put in the partially used list.
 408                  */
 409                 if (!uctxt->tid_used_list.count || need_group) {
 410                         if (!uctxt->tid_group_list.count)
 411                                 goto unlock;
 412
 413                         grp = tid_group_pop(&uctxt->tid_group_list);
 414                         tid_group_add_tail(grp, &uctxt->tid_used_list);
 415                         need_group = 0;
 416                 }
 417                 /*
 418                  * There is an optimization opportunity here - instead of
 419                  * fitting as many page sets as we can, check for a group
 420                  * later on in the list that could fit all of them.
 421                  */
 422                 list_for_each_entry_safe(grp, ptr, &uctxt->tid_used_list.list,
 423                                          list) {
 424                         unsigned use = min_t(unsigned, pageset_count - pageidx,
 425                                              grp->size - grp->used);
 426
 427                         ret = program_rcvarray(fd, tidbuf, grp,
 428                                                pageidx, use, tidlist,
 429                                                &tididx, &mapped);
 430                         if (ret < 0) {
 431                                 hfi1_cdbg(TID,
 432                                           "Failed to program RcvArray entries %d",
 433                                           ret);
 434                                 ret = -EFAULT;
 435                                 goto unlock;
 436                         } else if (ret > 0) {
 437                                 if (grp->used == grp->size)
 438                                         tid_group_move(grp,
 439                                                        &uctxt->tid_used_list,
 440                                                        &uctxt->tid_full_list);
 441                                 pageidx += ret;
 442                                 mapped_pages += mapped;
 443                                 need_group = 0;
 444                                 /* Check if we are done so we break out early */
 445                                 if (pageidx >= pageset_count)
 446                                         break;
 447                         } else if (WARN_ON(ret == 0)) {
 448                                 /*
 449                                  * If ret is 0, we did not program any entries
 450                                  * into this group, which can only happen if
 451                                  * we've screwed up the accounting somewhere.
 452                                  * Warn and try to continue.
 453                                  */
 454                                 need_group = 1;
 455                         }
 456                 }
 457         }
 458 unlock:
 459         mutex_unlock(&uctxt->exp_lock);
 460 nomem:
 461         hfi1_cdbg(TID, "total mapped: tidpairs:%u pages:%u (%d)", tididx,
 462                   mapped_pages, ret);
 463         /* adjust reserved tid_used to actual count */
 464         spin_lock(&fd->tid_lock);
 465         fd->tid_used -= pageset_count - tididx;
 466         spin_unlock(&fd->tid_lock);
 467         if (tididx) {
 468                 tinfo->tidcnt = tididx;
 469                 tinfo->length = mapped_pages * PAGE_SIZE;
 470
 471                 if (copy_to_user((void __user *)(unsigned long)tinfo->tidlist,
 472                                  tidlist, sizeof(tidlist[0]) * tididx)) {
 473                         /*
 474                          * On failure to copy to the user level, we need to undo
 475                          * everything done so far so we don't leak resources.
 476                          */
 477                         tinfo->tidlist = (unsigned long)&tidlist;
 478                         hfi1_user_exp_rcv_clear(fd, tinfo);
 479                         tinfo->tidlist = 0;
 480                         ret = -EFAULT;
 481                         goto bail;
 482                 }
 483         }
 484
 485         /*
 486          * If not everything was mapped (due to insufficient RcvArray entries,
 487          * for example), unpin all unmapped pages so we can pin them nex time.
 488          */
 489         if (mapped_pages != pinned)
 490                 unpin_rcv_pages(fd, tidbuf, NULL, mapped_pages,
 491                                 (pinned - mapped_pages), false);
 492 bail:
 493         kfree(tidbuf->psets);
 494         kfree(tidlist);
 495         kfree(tidbuf->pages);
 496         kfree(tidbuf);
 497         return ret > 0 ? 0 : ret;
 498 }
 499
 500 int hfi1_user_exp_rcv_clear(struct hfi1_filedata *fd,
 501                             struct hfi1_tid_info *tinfo)
 502 {
 503         int ret = 0;
 504         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 505         u32 *tidinfo;
 506         unsigned tididx;
 507
 508         if (unlikely(tinfo->tidcnt > fd->tid_used))
 509                 return -EINVAL;
 510
 511         tidinfo = memdup_user((void __user *)(unsigned long)tinfo->tidlist,
 512                               sizeof(tidinfo[0]) * tinfo->tidcnt);
 513         if (IS_ERR(tidinfo))
 514                 return PTR_ERR(tidinfo);
 515
 516         mutex_lock(&uctxt->exp_lock);
 517         for (tididx = 0; tididx < tinfo->tidcnt; tididx++) {
 518                 ret = unprogram_rcvarray(fd, tidinfo[tididx], NULL);
 519                 if (ret) {
 520                         hfi1_cdbg(TID, "Failed to unprogram rcv array %d",
 521                                   ret);
 522                         break;
 523                 }
 524         }
 525         spin_lock(&fd->tid_lock);
 526         fd->tid_used -= tididx;
 527         spin_unlock(&fd->tid_lock);
 528         tinfo->tidcnt = tididx;
 529         mutex_unlock(&uctxt->exp_lock);
 530
 531         kfree(tidinfo);
 532         return ret;
 533 }
 534
 535 int hfi1_user_exp_rcv_invalid(struct hfi1_filedata *fd,
 536                               struct hfi1_tid_info *tinfo)
 537 {
 538         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 539         unsigned long *ev = uctxt->dd->events +
 540                 (((uctxt->ctxt - uctxt->dd->first_dyn_alloc_ctxt) *
 541                   HFI1_MAX_SHARED_CTXTS) + fd->subctxt);
 542         u32 *array;
 543         int ret = 0;
 544
 545         if (!fd->invalid_tids)
 546                 return -EINVAL;
 547
 548         /*
 549          * copy_to_user() can sleep, which will leave the invalid_lock
 550          * locked and cause the MMU notifier to be blocked on the lock
 551          * for a long time.
 552          * Copy the data to a local buffer so we can release the lock.
 553          */
 554         array = kcalloc(uctxt->expected_count, sizeof(*array), GFP_KERNEL);
 555         if (!array)
 556                 return -EFAULT;
 557
 558         spin_lock(&fd->invalid_lock);
 559         if (fd->invalid_tid_idx) {
 560                 memcpy(array, fd->invalid_tids, sizeof(*array) *
 561                        fd->invalid_tid_idx);
 562                 memset(fd->invalid_tids, 0, sizeof(*fd->invalid_tids) *
 563                        fd->invalid_tid_idx);
 564                 tinfo->tidcnt = fd->invalid_tid_idx;
 565                 fd->invalid_tid_idx = 0;
 566                 /*
 567                  * Reset the user flag while still holding the lock.
 568                  * Otherwise, PSM can miss events.
 569                  */
 570                 clear_bit(_HFI1_EVENT_TID_MMU_NOTIFY_BIT, ev);
 571         } else {
 572                 tinfo->tidcnt = 0;
 573         }
 574         spin_unlock(&fd->invalid_lock);
 575
 576         if (tinfo->tidcnt) {
 577                 if (copy_to_user((void __user *)tinfo->tidlist,
 578                                  array, sizeof(*array) * tinfo->tidcnt))
 579                         ret = -EFAULT;
 580         }
 581         kfree(array);
 582
 583         return ret;
 584 }
 585
 586 static u32 find_phys_blocks(struct tid_user_buf *tidbuf, unsigned int npages)
 587 {
 588         unsigned pagecount, pageidx, setcount = 0, i;
 589         unsigned long pfn, this_pfn;
 590         struct page **pages = tidbuf->pages;
 591         struct tid_pageset *list = tidbuf->psets;
 592
 593         if (!npages)
 594                 return 0;
 595
 596         /*
 597          * Look for sets of physically contiguous pages in the user buffer.
 598          * This will allow us to optimize Expected RcvArray entry usage by
 599          * using the bigger supported sizes.
 600          */
 601         pfn = page_to_pfn(pages[0]);
 602         for (pageidx = 0, pagecount = 1, i = 1; i <= npages; i++) {
 603                 this_pfn = i < npages ? page_to_pfn(pages[i]) : 0;
 604
 605                 /*
 606                  * If the pfn's are not sequential, pages are not physically
 607                  * contiguous.
 608                  */
 609                 if (this_pfn != ++pfn) {
 610                         /*
 611                          * At this point we have to loop over the set of
 612                          * physically contiguous pages and break them down it
 613                          * sizes supported by the HW.
 614                          * There are two main constraints:
 615                          *     1. The max buffer size is MAX_EXPECTED_BUFFER.
 616                          *        If the total set size is bigger than that
 617                          *        program only a MAX_EXPECTED_BUFFER chunk.
 618                          *     2. The buffer size has to be a power of two. If
 619                          *        it is not, round down to the closes power of
 620                          *        2 and program that size.
 621                          */
 622                         while (pagecount) {
 623                                 int maxpages = pagecount;
 624                                 u32 bufsize = pagecount * PAGE_SIZE;
 625
 626                                 if (bufsize > MAX_EXPECTED_BUFFER)
 627                                         maxpages =
 628                                                 MAX_EXPECTED_BUFFER >>
 629                                                 PAGE_SHIFT;
 630                                 else if (!is_power_of_2(bufsize))
 631                                         maxpages =
 632                                                 rounddown_pow_of_two(bufsize) >>
 633                                                 PAGE_SHIFT;
 634
 635                                 list[setcount].idx = pageidx;
 636                                 list[setcount].count = maxpages;
 637                                 pagecount -= maxpages;
 638                                 pageidx += maxpages;
 639                                 setcount++;
 640                         }
 641                         pageidx = i;
 642                         pagecount = 1;
 643                         pfn = this_pfn;
 644                 } else {
 645                         pagecount++;
 646                 }
 647         }
 648         return setcount;
 649 }
 650
 651 /**
 652  * program_rcvarray() - program an RcvArray group with receive buffers
 653  * @fd: filedata pointer
 654  * @tbuf: pointer to struct tid_user_buf that has the user buffer starting
 655  *        virtual address, buffer length, page pointers, pagesets (array of
 656  *        struct tid_pageset holding information on physically contiguous
 657  *        chunks from the user buffer), and other fields.
 658  * @grp: RcvArray group
 659  * @start: starting index into sets array
 660  * @count: number of struct tid_pageset's to program
 661  * @tidlist: the array of u32 elements when the information about the
 662  *           programmed RcvArray entries is to be encoded.
 663  * @tididx: starting offset into tidlist
 664  * @pmapped: (output parameter) number of pages programmed into the RcvArray
 665  *           entries.
 666  *
 667  * This function will program up to 'count' number of RcvArray entries from the
 668  * group 'grp'. To make best use of write-combining writes, the function will
 669  * perform writes to the unused RcvArray entries which will be ignored by the
 670  * HW. Each RcvArray entry will be programmed with a physically contiguous
 671  * buffer chunk from the user's virtual buffer.
 672  *
 673  * Return:
 674  * -EINVAL if the requested count is larger than the size of the group,
 675  * -ENOMEM or -EFAULT on error from set_rcvarray_entry(), or
 676  * number of RcvArray entries programmed.
 677  */
 678 static int program_rcvarray(struct hfi1_filedata *fd, struct tid_user_buf *tbuf,
 679                             struct tid_group *grp,
 680                             unsigned int start, u16 count,
 681                             u32 *tidlist, unsigned int *tididx,
 682                             unsigned int *pmapped)
 683 {
 684         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 685         struct hfi1_devdata *dd = uctxt->dd;
 686         u16 idx;
 687         u32 tidinfo = 0, rcventry, useidx = 0;
 688         int mapped = 0;
 689
 690         /* Count should never be larger than the group size */
 691         if (count > grp->size)
 692                 return -EINVAL;
 693
 694         /* Find the first unused entry in the group */
 695         for (idx = 0; idx < grp->size; idx++) {
 696                 if (!(grp->map & (1 << idx))) {
 697                         useidx = idx;
 698                         break;
 699                 }
 700                 rcv_array_wc_fill(dd, grp->base + idx);
 701         }
 702
 703         idx = 0;
 704         while (idx < count) {
 705                 u16 npages, pageidx, setidx = start + idx;
 706                 int ret = 0;
 707
 708                 /*
 709                  * If this entry in the group is used, move to the next one.
 710                  * If we go past the end of the group, exit the loop.
 711                  */
 712                 if (useidx >= grp->size) {
 713                         break;
 714                 } else if (grp->map & (1 << useidx)) {
 715                         rcv_array_wc_fill(dd, grp->base + useidx);
 716                         useidx++;
 717                         continue;
 718                 }
 719
 720                 rcventry = grp->base + useidx;
 721                 npages = tbuf->psets[setidx].count;
 722                 pageidx = tbuf->psets[setidx].idx;
 723
 724                 ret = set_rcvarray_entry(fd, tbuf,
 725                                          rcventry, grp, pageidx,
 726                                          npages);
 727                 if (ret)
 728                         return ret;
 729                 mapped += npages;
 730
 731                 tidinfo = rcventry2tidinfo(rcventry - uctxt->expected_base) |
 732                         EXP_TID_SET(LEN, npages);
 733                 tidlist[(*tididx)++] = tidinfo;
 734                 grp->used++;
 735                 grp->map |= 1 << useidx++;
 736                 idx++;
 737         }
 738
 739         /* Fill the rest of the group with "blank" writes */
 740         for (; useidx < grp->size; useidx++)
 741                 rcv_array_wc_fill(dd, grp->base + useidx);
 742         *pmapped = mapped;
 743         return idx;
 744 }
 745
 746 static int set_rcvarray_entry(struct hfi1_filedata *fd,
 747                               struct tid_user_buf *tbuf,
 748                               u32 rcventry, struct tid_group *grp,
 749                               u16 pageidx, unsigned int npages)
 750 {
 751         int ret;
 752         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 753         struct tid_rb_node *node;
 754         struct hfi1_devdata *dd = uctxt->dd;
 755         dma_addr_t phys;
 756         struct page **pages = tbuf->pages + pageidx;
 757
 758         /*
 759          * Allocate the node first so we can handle a potential
 760          * failure before we've programmed anything.
 761          */
 762         node = kzalloc(sizeof(*node) + (sizeof(struct page *) * npages),
 763                        GFP_KERNEL);
 764         if (!node)
 765                 return -ENOMEM;
 766
 767         phys = pci_map_single(dd->pcidev,
 768                               __va(page_to_phys(pages[0])),
 769                               npages * PAGE_SIZE, PCI_DMA_FROMDEVICE);
 770         if (dma_mapping_error(&dd->pcidev->dev, phys)) {
 771                 dd_dev_err(dd, "Failed to DMA map Exp Rcv pages 0x%llx\n",
 772                            phys);
 773                 kfree(node);
 774                 return -EFAULT;
 775         }
 776
 777         node->mmu.addr = tbuf->vaddr + (pageidx * PAGE_SIZE);
 778         node->mmu.len = npages * PAGE_SIZE;
 779         node->phys = page_to_phys(pages[0]);
 780         node->npages = npages;
 781         node->rcventry = rcventry;
 782         node->dma_addr = phys;
 783         node->grp = grp;
 784         node->freed = false;
 785         memcpy(node->pages, pages, sizeof(struct page *) * npages);
 786
 787         if (!fd->handler)
 788                 ret = tid_rb_insert(fd, &node->mmu);
 789         else
 790                 ret = hfi1_mmu_rb_insert(fd->handler, &node->mmu);
 791
 792         if (ret) {
 793                 hfi1_cdbg(TID, "Failed to insert RB node %u 0x%lx, 0x%lx %d",
 794                           node->rcventry, node->mmu.addr, node->phys, ret);
 795                 pci_unmap_single(dd->pcidev, phys, npages * PAGE_SIZE,
 796                                  PCI_DMA_FROMDEVICE);
 797                 kfree(node);
 798                 return -EFAULT;
 799         }
 800         hfi1_put_tid(dd, rcventry, PT_EXPECTED, phys, ilog2(npages) + 1);
 801         trace_hfi1_exp_tid_reg(uctxt->ctxt, fd->subctxt, rcventry, npages,
 802                                node->mmu.addr, node->phys, phys);
 803         return 0;
 804 }
 805
 806 static int unprogram_rcvarray(struct hfi1_filedata *fd, u32 tidinfo,
 807                               struct tid_group **grp)
 808 {
 809         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 810         struct hfi1_devdata *dd = uctxt->dd;
 811         struct tid_rb_node *node;
 812         u8 tidctrl = EXP_TID_GET(tidinfo, CTRL);
 813         u32 tididx = EXP_TID_GET(tidinfo, IDX) << 1, rcventry;
 814
 815         if (tididx >= uctxt->expected_count) {
 816                 dd_dev_err(dd, "Invalid RcvArray entry (%u) index for ctxt %u\n",
 817                            tididx, uctxt->ctxt);
 818                 return -EINVAL;
 819         }
 820
 821         if (tidctrl == 0x3)
 822                 return -EINVAL;
 823
 824         rcventry = tididx + (tidctrl - 1);
 825
 826         node = fd->entry_to_rb[rcventry];
 827         if (!node || node->rcventry != (uctxt->expected_base + rcventry))
 828                 return -EBADF;
 829
 830         if (grp)
 831                 *grp = node->grp;
 832
 833         if (!fd->handler)
 834                 cacheless_tid_rb_remove(fd, node);
 835         else
 836                 hfi1_mmu_rb_remove(fd->handler, &node->mmu);
 837
 838         return 0;
 839 }
 840
 841 static void clear_tid_node(struct hfi1_filedata *fd, struct tid_rb_node *node)
 842 {
 843         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 844         struct hfi1_devdata *dd = uctxt->dd;
 845
 846         trace_hfi1_exp_tid_unreg(uctxt->ctxt, fd->subctxt, node->rcventry,
 847                                  node->npages, node->mmu.addr, node->phys,
 848                                  node->dma_addr);
 849
 850         /*
 851          * Make sure device has seen the write before we unpin the
 852          * pages.
 853          */
 854         hfi1_put_tid(dd, node->rcventry, PT_INVALID_FLUSH, 0, 0);
 855
 856         unpin_rcv_pages(fd, NULL, node, 0, node->npages, true);
 857
 858         node->grp->used--;
 859         node->grp->map &= ~(1 << (node->rcventry - node->grp->base));
 860
 861         if (node->grp->used == node->grp->size - 1)
 862                 tid_group_move(node->grp, &uctxt->tid_full_list,
 863                                &uctxt->tid_used_list);
 864         else if (!node->grp->used)
 865                 tid_group_move(node->grp, &uctxt->tid_used_list,
 866                                &uctxt->tid_group_list);
 867         kfree(node);
 868 }
 869
 870 /*
 871  * As a simple helper for hfi1_user_exp_rcv_free, this function deals with
 872  * clearing nodes in the non-cached case.
 873  */
 874 static void unlock_exp_tids(struct hfi1_ctxtdata *uctxt,
 875                             struct exp_tid_set *set,
 876                             struct hfi1_filedata *fd)
 877 {
 878         struct tid_group *grp, *ptr;
 879         int i;
 880
 881         list_for_each_entry_safe(grp, ptr, &set->list, list) {
 882                 list_del_init(&grp->list);
 883
 884                 for (i = 0; i < grp->size; i++) {
 885                         if (grp->map & (1 << i)) {
 886                                 u16 rcventry = grp->base + i;
 887                                 struct tid_rb_node *node;
 888
 889                                 node = fd->entry_to_rb[rcventry -
 890                                                           uctxt->expected_base];
 891                                 if (!node || node->rcventry != rcventry)
 892                                         continue;
 893
 894                                 cacheless_tid_rb_remove(fd, node);
 895                         }
 896                 }
 897         }
 898 }
 899
 900 /*
 901  * Always return 0 from this function.  A non-zero return indicates that the
 902  * remove operation will be called and that memory should be unpinned.
 903  * However, the driver cannot unpin out from under PSM.  Instead, retain the
 904  * memory (by returning 0) and inform PSM that the memory is going away.  PSM
 905  * will call back later when it has removed the memory from its list.
 906  */
 907 static int tid_rb_invalidate(void *arg, struct mmu_rb_node *mnode)
 908 {
 909         struct hfi1_filedata *fdata = arg;
 910         struct hfi1_ctxtdata *uctxt = fdata->uctxt;
 911         struct tid_rb_node *node =
 912                 container_of(mnode, struct tid_rb_node, mmu);
 913
 914         if (node->freed)
 915                 return 0;
 916
 917         trace_hfi1_exp_tid_inval(uctxt->ctxt, fdata->subctxt, node->mmu.addr,
 918                                  node->rcventry, node->npages, node->dma_addr);
 919         node->freed = true;
 920
 921         spin_lock(&fdata->invalid_lock);
 922         if (fdata->invalid_tid_idx < uctxt->expected_count) {
 923                 fdata->invalid_tids[fdata->invalid_tid_idx] =
 924                         rcventry2tidinfo(node->rcventry - uctxt->expected_base);
 925                 fdata->invalid_tids[fdata->invalid_tid_idx] |=
 926                         EXP_TID_SET(LEN, node->npages);
 927                 if (!fdata->invalid_tid_idx) {
 928                         unsigned long *ev;
 929
 930                         /*
 931                          * hfi1_set_uevent_bits() sets a user event flag
 932                          * for all processes. Because calling into the
 933                          * driver to process TID cache invalidations is
 934                          * expensive and TID cache invalidations are
 935                          * handled on a per-process basis, we can
 936                          * optimize this to set the flag only for the
 937                          * process in question.
 938                          */
 939                         ev = uctxt->dd->events +
 940                           (((uctxt->ctxt - uctxt->dd->first_dyn_alloc_ctxt) *
 941                             HFI1_MAX_SHARED_CTXTS) + fdata->subctxt);
 942                         set_bit(_HFI1_EVENT_TID_MMU_NOTIFY_BIT, ev);
 943                 }
 944                 fdata->invalid_tid_idx++;
 945         }
 946         spin_unlock(&fdata->invalid_lock);
 947         return 0;
 948 }
 949
 950 static int tid_rb_insert(void *arg, struct mmu_rb_node *node)
 951 {
 952         struct hfi1_filedata *fdata = arg;
 953         struct tid_rb_node *tnode =
 954                 container_of(node, struct tid_rb_node, mmu);
 955         u32 base = fdata->uctxt->expected_base;
 956
 957         fdata->entry_to_rb[tnode->rcventry - base] = tnode;
 958         return 0;
 959 }
 960
 961 static void cacheless_tid_rb_remove(struct hfi1_filedata *fdata,
 962                                     struct tid_rb_node *tnode)
 963 {
 964         u32 base = fdata->uctxt->expected_base;
 965
 966         fdata->entry_to_rb[tnode->rcventry - base] = NULL;
 967         clear_tid_node(fdata, tnode);
 968 }
 969
 970 static void tid_rb_remove(void *arg, struct mmu_rb_node *node)
 971 {
 972         struct hfi1_filedata *fdata = arg;
 973         struct tid_rb_node *tnode =
 974                 container_of(node, struct tid_rb_node, mmu);
 975
 976         cacheless_tid_rb_remove(fdata, tnode);
 977 }