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 = tidbuf->npages;
 219         unsigned long vaddr = tidbuf->vaddr;
 220         struct page **pages = NULL;
 221         struct hfi1_devdata *dd = fd->uctxt->dd;
 222
 223         if (npages > fd->uctxt->expected_count) {
 224                 dd_dev_err(dd, "Expected buffer too big\n");
 225                 return -EINVAL;
 226         }
 227
 228         /* Verify that access is OK for the user buffer */
 229         if (!access_ok(VERIFY_WRITE, (void __user *)vaddr,
 230                        npages * PAGE_SIZE)) {
 231                 dd_dev_err(dd, "Fail vaddr %p, %u pages, !access_ok\n",
 232                            (void *)vaddr, npages);
 233                 return -EFAULT;
 234         }
 235         /* Allocate the array of struct page pointers needed for pinning */
 236         pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL);
 237         if (!pages)
 238                 return -ENOMEM;
 239
 240         /*
 241          * Pin all the pages of the user buffer. If we can't pin all the
 242          * pages, accept the amount pinned so far and program only that.
 243          * User space knows how to deal with partially programmed buffers.
 244          */
 245         if (!hfi1_can_pin_pages(dd, fd->mm, fd->tid_n_pinned, npages)) {
 246                 kfree(pages);
 247                 return -ENOMEM;
 248         }
 249
 250         pinned = hfi1_acquire_user_pages(fd->mm, vaddr, npages, true, pages);
 251         if (pinned <= 0) {
 252                 kfree(pages);
 253                 return pinned;
 254         }
 255         tidbuf->pages = pages;
 256         fd->tid_n_pinned += pinned;
 257         return pinned;
 258 }
 259
 260 /*
 261  * RcvArray entry allocation for Expected Receives is done by the
 262  * following algorithm:
 263  *
 264  * The context keeps 3 lists of groups of RcvArray entries:
 265  *   1. List of empty groups - tid_group_list
 266  *      This list is created during user context creation and
 267  *      contains elements which describe sets (of 8) of empty
 268  *      RcvArray entries.
 269  *   2. List of partially used groups - tid_used_list
 270  *      This list contains sets of RcvArray entries which are
 271  *      not completely used up. Another mapping request could
 272  *      use some of all of the remaining entries.
 273  *   3. List of full groups - tid_full_list
 274  *      This is the list where sets that are completely used
 275  *      up go.
 276  *
 277  * An attempt to optimize the usage of RcvArray entries is
 278  * made by finding all sets of physically contiguous pages in a
 279  * user's buffer.
 280  * These physically contiguous sets are further split into
 281  * sizes supported by the receive engine of the HFI. The
 282  * resulting sets of pages are stored in struct tid_pageset,
 283  * which describes the sets as:
 284  *    * .count - number of pages in this set
 285  *    * .idx - starting index into struct page ** array
 286  *                    of this set
 287  *
 288  * From this point on, the algorithm deals with the page sets
 289  * described above. The number of pagesets is divided by the
 290  * RcvArray group size to produce the number of full groups
 291  * needed.
 292  *
 293  * Groups from the 3 lists are manipulated using the following
 294  * rules:
 295  *   1. For each set of 8 pagesets, a complete group from
 296  *      tid_group_list is taken, programmed, and moved to
 297  *      the tid_full_list list.
 298  *   2. For all remaining pagesets:
 299  *      2.1 If the tid_used_list is empty and the tid_group_list
 300  *          is empty, stop processing pageset and return only
 301  *          what has been programmed up to this point.
 302  *      2.2 If the tid_used_list is empty and the tid_group_list
 303  *          is not empty, move a group from tid_group_list to
 304  *          tid_used_list.
 305  *      2.3 For each group is tid_used_group, program as much as
 306  *          can fit into the group. If the group becomes fully
 307  *          used, move it to tid_full_list.
 308  */
 309 int hfi1_user_exp_rcv_setup(struct hfi1_filedata *fd,
 310                             struct hfi1_tid_info *tinfo)
 311 {
 312         int ret = 0, need_group = 0, pinned;
 313         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 314         struct hfi1_devdata *dd = uctxt->dd;
 315         unsigned int ngroups, pageidx = 0, pageset_count,
 316                 tididx = 0, mapped, mapped_pages = 0;
 317         u32 *tidlist = NULL;
 318         struct tid_user_buf *tidbuf;
 319
 320         if (!PAGE_ALIGNED(tinfo->vaddr))
 321                 return -EINVAL;
 322         if (tinfo->length == 0)
 323                 return -EINVAL;
 324
 325         tidbuf = kzalloc(sizeof(*tidbuf), GFP_KERNEL);
 326         if (!tidbuf)
 327                 return -ENOMEM;
 328
 329         tidbuf->vaddr = tinfo->vaddr;
 330         tidbuf->length = tinfo->length;
 331         tidbuf->npages = num_user_pages(tidbuf->vaddr, tidbuf->length);
 332         tidbuf->psets = kcalloc(uctxt->expected_count, sizeof(*tidbuf->psets),
 333                                 GFP_KERNEL);
 334         if (!tidbuf->psets) {
 335                 ret = -ENOMEM;
 336                 goto fail_release_mem;
 337         }
 338
 339         pinned = pin_rcv_pages(fd, tidbuf);
 340         if (pinned <= 0) {
 341                 ret = (pinned < 0) ? pinned : -ENOSPC;
 342                 goto fail_unpin;
 343         }
 344
 345         /* Find sets of physically contiguous pages */
 346         tidbuf->n_psets = find_phys_blocks(tidbuf, pinned);
 347
 348         /* Reserve the number of expected tids to be used. */
 349         spin_lock(&fd->tid_lock);
 350         if (fd->tid_used + tidbuf->n_psets > fd->tid_limit)
 351                 pageset_count = fd->tid_limit - fd->tid_used;
 352         else
 353                 pageset_count = tidbuf->n_psets;
 354         fd->tid_used += pageset_count;
 355         spin_unlock(&fd->tid_lock);
 356
 357         if (!pageset_count) {
 358                 ret = -ENOSPC;
 359                 goto fail_unreserve;
 360         }
 361
 362         ngroups = pageset_count / dd->rcv_entries.group_size;
 363         tidlist = kcalloc(pageset_count, sizeof(*tidlist), GFP_KERNEL);
 364         if (!tidlist) {
 365                 ret = -ENOMEM;
 366                 goto fail_unreserve;
 367         }
 368
 369         tididx = 0;
 370
 371         /*
 372          * From this point on, we are going to be using shared (between master
 373          * and subcontexts) context resources. We need to take the lock.
 374          */
 375         mutex_lock(&uctxt->exp_mutex);
 376         /*
 377          * The first step is to program the RcvArray entries which are complete
 378          * groups.
 379          */
 380         while (ngroups && uctxt->tid_group_list.count) {
 381                 struct tid_group *grp =
 382                         tid_group_pop(&uctxt->tid_group_list);
 383
 384                 ret = program_rcvarray(fd, tidbuf, grp,
 385                                        pageidx, dd->rcv_entries.group_size,
 386                                        tidlist, &tididx, &mapped);
 387                 /*
 388                  * If there was a failure to program the RcvArray
 389                  * entries for the entire group, reset the grp fields
 390                  * and add the grp back to the free group list.
 391                  */
 392                 if (ret <= 0) {
 393                         tid_group_add_tail(grp, &uctxt->tid_group_list);
 394                         hfi1_cdbg(TID,
 395                                   "Failed to program RcvArray group %d", ret);
 396                         goto unlock;
 397                 }
 398
 399                 tid_group_add_tail(grp, &uctxt->tid_full_list);
 400                 ngroups--;
 401                 pageidx += ret;
 402                 mapped_pages += mapped;
 403         }
 404
 405         while (pageidx < pageset_count) {
 406                 struct tid_group *grp, *ptr;
 407                 /*
 408                  * If we don't have any partially used tid groups, check
 409                  * if we have empty groups. If so, take one from there and
 410                  * put in the partially used list.
 411                  */
 412                 if (!uctxt->tid_used_list.count || need_group) {
 413                         if (!uctxt->tid_group_list.count)
 414                                 goto unlock;
 415
 416                         grp = tid_group_pop(&uctxt->tid_group_list);
 417                         tid_group_add_tail(grp, &uctxt->tid_used_list);
 418                         need_group = 0;
 419                 }
 420                 /*
 421                  * There is an optimization opportunity here - instead of
 422                  * fitting as many page sets as we can, check for a group
 423                  * later on in the list that could fit all of them.
 424                  */
 425                 list_for_each_entry_safe(grp, ptr, &uctxt->tid_used_list.list,
 426                                          list) {
 427                         unsigned use = min_t(unsigned, pageset_count - pageidx,
 428                                              grp->size - grp->used);
 429
 430                         ret = program_rcvarray(fd, tidbuf, grp,
 431                                                pageidx, use, tidlist,
 432                                                &tididx, &mapped);
 433                         if (ret < 0) {
 434                                 hfi1_cdbg(TID,
 435                                           "Failed to program RcvArray entries %d",
 436                                           ret);
 437                                 goto unlock;
 438                         } else if (ret > 0) {
 439                                 if (grp->used == grp->size)
 440                                         tid_group_move(grp,
 441                                                        &uctxt->tid_used_list,
 442                                                        &uctxt->tid_full_list);
 443                                 pageidx += ret;
 444                                 mapped_pages += mapped;
 445                                 need_group = 0;
 446                                 /* Check if we are done so we break out early */
 447                                 if (pageidx >= pageset_count)
 448                                         break;
 449                         } else if (WARN_ON(ret == 0)) {
 450                                 /*
 451                                  * If ret is 0, we did not program any entries
 452                                  * into this group, which can only happen if
 453                                  * we've screwed up the accounting somewhere.
 454                                  * Warn and try to continue.
 455                                  */
 456                                 need_group = 1;
 457                         }
 458                 }
 459         }
 460 unlock:
 461         mutex_unlock(&uctxt->exp_mutex);
 462         hfi1_cdbg(TID, "total mapped: tidpairs:%u pages:%u (%d)", tididx,
 463                   mapped_pages, ret);
 464
 465         /* fail if nothing was programmed, set error if none provided */
 466         if (tididx == 0) {
 467                 if (ret >= 0)
 468                         ret = -ENOSPC;
 469                 goto fail_unreserve;
 470         }
 471
 472         /* adjust reserved tid_used to actual count */
 473         spin_lock(&fd->tid_lock);
 474         fd->tid_used -= pageset_count - tididx;
 475         spin_unlock(&fd->tid_lock);
 476
 477         /* unpin all pages not covered by a TID */
 478         unpin_rcv_pages(fd, tidbuf, NULL, mapped_pages, pinned - mapped_pages,
 479                         false);
 480
 481         tinfo->tidcnt = tididx;
 482         tinfo->length = mapped_pages * PAGE_SIZE;
 483
 484         if (copy_to_user(u64_to_user_ptr(tinfo->tidlist),
 485                          tidlist, sizeof(tidlist[0]) * tididx)) {
 486                 ret = -EFAULT;
 487                 goto fail_unprogram;
 488         }
 489
 490         kfree(tidbuf->pages);
 491         kfree(tidbuf->psets);
 492         kfree(tidbuf);
 493         kfree(tidlist);
 494         return 0;
 495
 496 fail_unprogram:
 497         /* unprogram, unmap, and unpin all allocated TIDs */
 498         tinfo->tidlist = (unsigned long)tidlist;
 499         hfi1_user_exp_rcv_clear(fd, tinfo);
 500         tinfo->tidlist = 0;
 501         pinned = 0;             /* nothing left to unpin */
 502         pageset_count = 0;      /* nothing left reserved */
 503 fail_unreserve:
 504         spin_lock(&fd->tid_lock);
 505         fd->tid_used -= pageset_count;
 506         spin_unlock(&fd->tid_lock);
 507 fail_unpin:
 508         if (pinned > 0)
 509                 unpin_rcv_pages(fd, tidbuf, NULL, 0, pinned, false);
 510 fail_release_mem:
 511         kfree(tidbuf->pages);
 512         kfree(tidbuf->psets);
 513         kfree(tidbuf);
 514         kfree(tidlist);
 515         return ret;
 516 }
 517
 518 int hfi1_user_exp_rcv_clear(struct hfi1_filedata *fd,
 519                             struct hfi1_tid_info *tinfo)
 520 {
 521         int ret = 0;
 522         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 523         u32 *tidinfo;
 524         unsigned tididx;
 525
 526         if (unlikely(tinfo->tidcnt > fd->tid_used))
 527                 return -EINVAL;
 528
 529         tidinfo = memdup_user(u64_to_user_ptr(tinfo->tidlist),
 530                               sizeof(tidinfo[0]) * tinfo->tidcnt);
 531         if (IS_ERR(tidinfo))
 532                 return PTR_ERR(tidinfo);
 533
 534         mutex_lock(&uctxt->exp_mutex);
 535         for (tididx = 0; tididx < tinfo->tidcnt; tididx++) {
 536                 ret = unprogram_rcvarray(fd, tidinfo[tididx], NULL);
 537                 if (ret) {
 538                         hfi1_cdbg(TID, "Failed to unprogram rcv array %d",
 539                                   ret);
 540                         break;
 541                 }
 542         }
 543         spin_lock(&fd->tid_lock);
 544         fd->tid_used -= tididx;
 545         spin_unlock(&fd->tid_lock);
 546         tinfo->tidcnt = tididx;
 547         mutex_unlock(&uctxt->exp_mutex);
 548
 549         kfree(tidinfo);
 550         return ret;
 551 }
 552
 553 int hfi1_user_exp_rcv_invalid(struct hfi1_filedata *fd,
 554                               struct hfi1_tid_info *tinfo)
 555 {
 556         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 557         unsigned long *ev = uctxt->dd->events +
 558                 (uctxt_offset(uctxt) + fd->subctxt);
 559         u32 *array;
 560         int ret = 0;
 561
 562         /*
 563          * copy_to_user() can sleep, which will leave the invalid_lock
 564          * locked and cause the MMU notifier to be blocked on the lock
 565          * for a long time.
 566          * Copy the data to a local buffer so we can release the lock.
 567          */
 568         array = kcalloc(uctxt->expected_count, sizeof(*array), GFP_KERNEL);
 569         if (!array)
 570                 return -EFAULT;
 571
 572         spin_lock(&fd->invalid_lock);
 573         if (fd->invalid_tid_idx) {
 574                 memcpy(array, fd->invalid_tids, sizeof(*array) *
 575                        fd->invalid_tid_idx);
 576                 memset(fd->invalid_tids, 0, sizeof(*fd->invalid_tids) *
 577                        fd->invalid_tid_idx);
 578                 tinfo->tidcnt = fd->invalid_tid_idx;
 579                 fd->invalid_tid_idx = 0;
 580                 /*
 581                  * Reset the user flag while still holding the lock.
 582                  * Otherwise, PSM can miss events.
 583                  */
 584                 clear_bit(_HFI1_EVENT_TID_MMU_NOTIFY_BIT, ev);
 585         } else {
 586                 tinfo->tidcnt = 0;
 587         }
 588         spin_unlock(&fd->invalid_lock);
 589
 590         if (tinfo->tidcnt) {
 591                 if (copy_to_user((void __user *)tinfo->tidlist,
 592                                  array, sizeof(*array) * tinfo->tidcnt))
 593                         ret = -EFAULT;
 594         }
 595         kfree(array);
 596
 597         return ret;
 598 }
 599
 600 static u32 find_phys_blocks(struct tid_user_buf *tidbuf, unsigned int npages)
 601 {
 602         unsigned pagecount, pageidx, setcount = 0, i;
 603         unsigned long pfn, this_pfn;
 604         struct page **pages = tidbuf->pages;
 605         struct tid_pageset *list = tidbuf->psets;
 606
 607         if (!npages)
 608                 return 0;
 609
 610         /*
 611          * Look for sets of physically contiguous pages in the user buffer.
 612          * This will allow us to optimize Expected RcvArray entry usage by
 613          * using the bigger supported sizes.
 614          */
 615         pfn = page_to_pfn(pages[0]);
 616         for (pageidx = 0, pagecount = 1, i = 1; i <= npages; i++) {
 617                 this_pfn = i < npages ? page_to_pfn(pages[i]) : 0;
 618
 619                 /*
 620                  * If the pfn's are not sequential, pages are not physically
 621                  * contiguous.
 622                  */
 623                 if (this_pfn != ++pfn) {
 624                         /*
 625                          * At this point we have to loop over the set of
 626                          * physically contiguous pages and break them down it
 627                          * sizes supported by the HW.
 628                          * There are two main constraints:
 629                          *     1. The max buffer size is MAX_EXPECTED_BUFFER.
 630                          *        If the total set size is bigger than that
 631                          *        program only a MAX_EXPECTED_BUFFER chunk.
 632                          *     2. The buffer size has to be a power of two. If
 633                          *        it is not, round down to the closes power of
 634                          *        2 and program that size.
 635                          */
 636                         while (pagecount) {
 637                                 int maxpages = pagecount;
 638                                 u32 bufsize = pagecount * PAGE_SIZE;
 639
 640                                 if (bufsize > MAX_EXPECTED_BUFFER)
 641                                         maxpages =
 642                                                 MAX_EXPECTED_BUFFER >>
 643                                                 PAGE_SHIFT;
 644                                 else if (!is_power_of_2(bufsize))
 645                                         maxpages =
 646                                                 rounddown_pow_of_two(bufsize) >>
 647                                                 PAGE_SHIFT;
 648
 649                                 list[setcount].idx = pageidx;
 650                                 list[setcount].count = maxpages;
 651                                 pagecount -= maxpages;
 652                                 pageidx += maxpages;
 653                                 setcount++;
 654                         }
 655                         pageidx = i;
 656                         pagecount = 1;
 657                         pfn = this_pfn;
 658                 } else {
 659                         pagecount++;
 660                 }
 661         }
 662         return setcount;
 663 }
 664
 665 /**
 666  * program_rcvarray() - program an RcvArray group with receive buffers
 667  * @fd: filedata pointer
 668  * @tbuf: pointer to struct tid_user_buf that has the user buffer starting
 669  *        virtual address, buffer length, page pointers, pagesets (array of
 670  *        struct tid_pageset holding information on physically contiguous
 671  *        chunks from the user buffer), and other fields.
 672  * @grp: RcvArray group
 673  * @start: starting index into sets array
 674  * @count: number of struct tid_pageset's to program
 675  * @tidlist: the array of u32 elements when the information about the
 676  *           programmed RcvArray entries is to be encoded.
 677  * @tididx: starting offset into tidlist
 678  * @pmapped: (output parameter) number of pages programmed into the RcvArray
 679  *           entries.
 680  *
 681  * This function will program up to 'count' number of RcvArray entries from the
 682  * group 'grp'. To make best use of write-combining writes, the function will
 683  * perform writes to the unused RcvArray entries which will be ignored by the
 684  * HW. Each RcvArray entry will be programmed with a physically contiguous
 685  * buffer chunk from the user's virtual buffer.
 686  *
 687  * Return:
 688  * -EINVAL if the requested count is larger than the size of the group,
 689  * -ENOMEM or -EFAULT on error from set_rcvarray_entry(), or
 690  * number of RcvArray entries programmed.
 691  */
 692 static int program_rcvarray(struct hfi1_filedata *fd, struct tid_user_buf *tbuf,
 693                             struct tid_group *grp,
 694                             unsigned int start, u16 count,
 695                             u32 *tidlist, unsigned int *tididx,
 696                             unsigned int *pmapped)
 697 {
 698         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 699         struct hfi1_devdata *dd = uctxt->dd;
 700         u16 idx;
 701         u32 tidinfo = 0, rcventry, useidx = 0;
 702         int mapped = 0;
 703
 704         /* Count should never be larger than the group size */
 705         if (count > grp->size)
 706                 return -EINVAL;
 707
 708         /* Find the first unused entry in the group */
 709         for (idx = 0; idx < grp->size; idx++) {
 710                 if (!(grp->map & (1 << idx))) {
 711                         useidx = idx;
 712                         break;
 713                 }
 714                 rcv_array_wc_fill(dd, grp->base + idx);
 715         }
 716
 717         idx = 0;
 718         while (idx < count) {
 719                 u16 npages, pageidx, setidx = start + idx;
 720                 int ret = 0;
 721
 722                 /*
 723                  * If this entry in the group is used, move to the next one.
 724                  * If we go past the end of the group, exit the loop.
 725                  */
 726                 if (useidx >= grp->size) {
 727                         break;
 728                 } else if (grp->map & (1 << useidx)) {
 729                         rcv_array_wc_fill(dd, grp->base + useidx);
 730                         useidx++;
 731                         continue;
 732                 }
 733
 734                 rcventry = grp->base + useidx;
 735                 npages = tbuf->psets[setidx].count;
 736                 pageidx = tbuf->psets[setidx].idx;
 737
 738                 ret = set_rcvarray_entry(fd, tbuf,
 739                                          rcventry, grp, pageidx,
 740                                          npages);
 741                 if (ret)
 742                         return ret;
 743                 mapped += npages;
 744
 745                 tidinfo = rcventry2tidinfo(rcventry - uctxt->expected_base) |
 746                         EXP_TID_SET(LEN, npages);
 747                 tidlist[(*tididx)++] = tidinfo;
 748                 grp->used++;
 749                 grp->map |= 1 << useidx++;
 750                 idx++;
 751         }
 752
 753         /* Fill the rest of the group with "blank" writes */
 754         for (; useidx < grp->size; useidx++)
 755                 rcv_array_wc_fill(dd, grp->base + useidx);
 756         *pmapped = mapped;
 757         return idx;
 758 }
 759
 760 static int set_rcvarray_entry(struct hfi1_filedata *fd,
 761                               struct tid_user_buf *tbuf,
 762                               u32 rcventry, struct tid_group *grp,
 763                               u16 pageidx, unsigned int npages)
 764 {
 765         int ret;
 766         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 767         struct tid_rb_node *node;
 768         struct hfi1_devdata *dd = uctxt->dd;
 769         dma_addr_t phys;
 770         struct page **pages = tbuf->pages + pageidx;
 771
 772         /*
 773          * Allocate the node first so we can handle a potential
 774          * failure before we've programmed anything.
 775          */
 776         node = kzalloc(sizeof(*node) + (sizeof(struct page *) * npages),
 777                        GFP_KERNEL);
 778         if (!node)
 779                 return -ENOMEM;
 780
 781         phys = pci_map_single(dd->pcidev,
 782                               __va(page_to_phys(pages[0])),
 783                               npages * PAGE_SIZE, PCI_DMA_FROMDEVICE);
 784         if (dma_mapping_error(&dd->pcidev->dev, phys)) {
 785                 dd_dev_err(dd, "Failed to DMA map Exp Rcv pages 0x%llx\n",
 786                            phys);
 787                 kfree(node);
 788                 return -EFAULT;
 789         }
 790
 791         node->mmu.addr = tbuf->vaddr + (pageidx * PAGE_SIZE);
 792         node->mmu.len = npages * PAGE_SIZE;
 793         node->phys = page_to_phys(pages[0]);
 794         node->npages = npages;
 795         node->rcventry = rcventry;
 796         node->dma_addr = phys;
 797         node->grp = grp;
 798         node->freed = false;
 799         memcpy(node->pages, pages, sizeof(struct page *) * npages);
 800
 801         if (!fd->handler)
 802                 ret = tid_rb_insert(fd, &node->mmu);
 803         else
 804                 ret = hfi1_mmu_rb_insert(fd->handler, &node->mmu);
 805
 806         if (ret) {
 807                 hfi1_cdbg(TID, "Failed to insert RB node %u 0x%lx, 0x%lx %d",
 808                           node->rcventry, node->mmu.addr, node->phys, ret);
 809                 pci_unmap_single(dd->pcidev, phys, npages * PAGE_SIZE,
 810                                  PCI_DMA_FROMDEVICE);
 811                 kfree(node);
 812                 return -EFAULT;
 813         }
 814         hfi1_put_tid(dd, rcventry, PT_EXPECTED, phys, ilog2(npages) + 1);
 815         trace_hfi1_exp_tid_reg(uctxt->ctxt, fd->subctxt, rcventry, npages,
 816                                node->mmu.addr, node->phys, phys);
 817         return 0;
 818 }
 819
 820 static int unprogram_rcvarray(struct hfi1_filedata *fd, u32 tidinfo,
 821                               struct tid_group **grp)
 822 {
 823         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 824         struct hfi1_devdata *dd = uctxt->dd;
 825         struct tid_rb_node *node;
 826         u8 tidctrl = EXP_TID_GET(tidinfo, CTRL);
 827         u32 tididx = EXP_TID_GET(tidinfo, IDX) << 1, rcventry;
 828
 829         if (tididx >= uctxt->expected_count) {
 830                 dd_dev_err(dd, "Invalid RcvArray entry (%u) index for ctxt %u\n",
 831                            tididx, uctxt->ctxt);
 832                 return -EINVAL;
 833         }
 834
 835         if (tidctrl == 0x3)
 836                 return -EINVAL;
 837
 838         rcventry = tididx + (tidctrl - 1);
 839
 840         node = fd->entry_to_rb[rcventry];
 841         if (!node || node->rcventry != (uctxt->expected_base + rcventry))
 842                 return -EBADF;
 843
 844         if (grp)
 845                 *grp = node->grp;
 846
 847         if (!fd->handler)
 848                 cacheless_tid_rb_remove(fd, node);
 849         else
 850                 hfi1_mmu_rb_remove(fd->handler, &node->mmu);
 851
 852         return 0;
 853 }
 854
 855 static void clear_tid_node(struct hfi1_filedata *fd, struct tid_rb_node *node)
 856 {
 857         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 858         struct hfi1_devdata *dd = uctxt->dd;
 859
 860         trace_hfi1_exp_tid_unreg(uctxt->ctxt, fd->subctxt, node->rcventry,
 861                                  node->npages, node->mmu.addr, node->phys,
 862                                  node->dma_addr);
 863
 864         /*
 865          * Make sure device has seen the write before we unpin the
 866          * pages.
 867          */
 868         hfi1_put_tid(dd, node->rcventry, PT_INVALID_FLUSH, 0, 0);
 869
 870         unpin_rcv_pages(fd, NULL, node, 0, node->npages, true);
 871
 872         node->grp->used--;
 873         node->grp->map &= ~(1 << (node->rcventry - node->grp->base));
 874
 875         if (node->grp->used == node->grp->size - 1)
 876                 tid_group_move(node->grp, &uctxt->tid_full_list,
 877                                &uctxt->tid_used_list);
 878         else if (!node->grp->used)
 879                 tid_group_move(node->grp, &uctxt->tid_used_list,
 880                                &uctxt->tid_group_list);
 881         kfree(node);
 882 }
 883
 884 /*
 885  * As a simple helper for hfi1_user_exp_rcv_free, this function deals with
 886  * clearing nodes in the non-cached case.
 887  */
 888 static void unlock_exp_tids(struct hfi1_ctxtdata *uctxt,
 889                             struct exp_tid_set *set,
 890                             struct hfi1_filedata *fd)
 891 {
 892         struct tid_group *grp, *ptr;
 893         int i;
 894
 895         list_for_each_entry_safe(grp, ptr, &set->list, list) {
 896                 list_del_init(&grp->list);
 897
 898                 for (i = 0; i < grp->size; i++) {
 899                         if (grp->map & (1 << i)) {
 900                                 u16 rcventry = grp->base + i;
 901                                 struct tid_rb_node *node;
 902
 903                                 node = fd->entry_to_rb[rcventry -
 904                                                           uctxt->expected_base];
 905                                 if (!node || node->rcventry != rcventry)
 906                                         continue;
 907
 908                                 cacheless_tid_rb_remove(fd, node);
 909                         }
 910                 }
 911         }
 912 }
 913
 914 /*
 915  * Always return 0 from this function.  A non-zero return indicates that the
 916  * remove operation will be called and that memory should be unpinned.
 917  * However, the driver cannot unpin out from under PSM.  Instead, retain the
 918  * memory (by returning 0) and inform PSM that the memory is going away.  PSM
 919  * will call back later when it has removed the memory from its list.
 920  */
 921 static int tid_rb_invalidate(void *arg, struct mmu_rb_node *mnode)
 922 {
 923         struct hfi1_filedata *fdata = arg;
 924         struct hfi1_ctxtdata *uctxt = fdata->uctxt;
 925         struct tid_rb_node *node =
 926                 container_of(mnode, struct tid_rb_node, mmu);
 927
 928         if (node->freed)
 929                 return 0;
 930
 931         trace_hfi1_exp_tid_inval(uctxt->ctxt, fdata->subctxt, node->mmu.addr,
 932                                  node->rcventry, node->npages, node->dma_addr);
 933         node->freed = true;
 934
 935         spin_lock(&fdata->invalid_lock);
 936         if (fdata->invalid_tid_idx < uctxt->expected_count) {
 937                 fdata->invalid_tids[fdata->invalid_tid_idx] =
 938                         rcventry2tidinfo(node->rcventry - uctxt->expected_base);
 939                 fdata->invalid_tids[fdata->invalid_tid_idx] |=
 940                         EXP_TID_SET(LEN, node->npages);
 941                 if (!fdata->invalid_tid_idx) {
 942                         unsigned long *ev;
 943
 944                         /*
 945                          * hfi1_set_uevent_bits() sets a user event flag
 946                          * for all processes. Because calling into the
 947                          * driver to process TID cache invalidations is
 948                          * expensive and TID cache invalidations are
 949                          * handled on a per-process basis, we can
 950                          * optimize this to set the flag only for the
 951                          * process in question.
 952                          */
 953                         ev = uctxt->dd->events +
 954                                 (uctxt_offset(uctxt) + fdata->subctxt);
 955                         set_bit(_HFI1_EVENT_TID_MMU_NOTIFY_BIT, ev);
 956                 }
 957                 fdata->invalid_tid_idx++;
 958         }
 959         spin_unlock(&fdata->invalid_lock);
 960         return 0;
 961 }
 962
 963 static int tid_rb_insert(void *arg, struct mmu_rb_node *node)
 964 {
 965         struct hfi1_filedata *fdata = arg;
 966         struct tid_rb_node *tnode =
 967                 container_of(node, struct tid_rb_node, mmu);
 968         u32 base = fdata->uctxt->expected_base;
 969
 970         fdata->entry_to_rb[tnode->rcventry - base] = tnode;
 971         return 0;
 972 }
 973
 974 static void cacheless_tid_rb_remove(struct hfi1_filedata *fdata,
 975                                     struct tid_rb_node *tnode)
 976 {
 977         u32 base = fdata->uctxt->expected_base;
 978
 979         fdata->entry_to_rb[tnode->rcventry - base] = NULL;
 980         clear_tid_node(fdata, tnode);
 981 }
 982
 983 static void tid_rb_remove(void *arg, struct mmu_rb_node *node)
 984 {
 985         struct hfi1_filedata *fdata = arg;
 986         struct tid_rb_node *tnode =
 987                 container_of(node, struct tid_rb_node, mmu);
 988
 989         cacheless_tid_rb_remove(fdata, tnode);
 990 }