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 "user_exp_rcv.h"
  51 #include "trace.h"
  52 #include "mmu_rb.h"
  53
  54 struct tid_group {
  55         struct list_head list;
  56         u32 base;
  57         u8 size;
  58         u8 used;
  59         u8 map;
  60 };
  61
  62 struct tid_rb_node {
  63         struct mmu_rb_node mmu;
  64         unsigned long phys;
  65         struct tid_group *grp;
  66         u32 rcventry;
  67         dma_addr_t dma_addr;
  68         bool freed;
  69         unsigned npages;
  70         struct page *pages[0];
  71 };
  72
  73 struct tid_pageset {
  74         u16 idx;
  75         u16 count;
  76 };
  77
  78 #define EXP_TID_SET_EMPTY(set) (set.count == 0 && list_empty(&set.list))
  79
  80 #define num_user_pages(vaddr, len)                                     \
  81         (1 + (((((unsigned long)(vaddr) +                              \
  82                  (unsigned long)(len) - 1) & PAGE_MASK) -              \
  83                ((unsigned long)vaddr & PAGE_MASK)) >> PAGE_SHIFT))
  84
  85 static void unlock_exp_tids(struct hfi1_ctxtdata *uctxt,
  86                             struct exp_tid_set *set,
  87                             struct hfi1_filedata *fd);
  88 static u32 find_phys_blocks(struct page **pages, unsigned npages,
  89                             struct tid_pageset *list);
  90 static int set_rcvarray_entry(struct hfi1_filedata *fd, unsigned long vaddr,
  91                               u32 rcventry, struct tid_group *grp,
  92                               struct page **pages, unsigned npages);
  93 static int tid_rb_insert(void *arg, struct mmu_rb_node *node);
  94 static void cacheless_tid_rb_remove(struct hfi1_filedata *fdata,
  95                                     struct tid_rb_node *tnode);
  96 static void tid_rb_remove(void *arg, struct mmu_rb_node *node);
  97 static int tid_rb_invalidate(void *arg, struct mmu_rb_node *mnode);
  98 static int program_rcvarray(struct hfi1_filedata *fd, unsigned long vaddr,
  99                             struct tid_group *grp, struct tid_pageset *sets,
 100                             unsigned start, u16 count, struct page **pages,
 101                             u32 *tidlist, unsigned *tididx, unsigned *pmapped);
 102 static int unprogram_rcvarray(struct hfi1_filedata *fd, u32 tidinfo,
 103                               struct tid_group **grp);
 104 static void clear_tid_node(struct hfi1_filedata *fd, struct tid_rb_node *node);
 105
 106 static struct mmu_rb_ops tid_rb_ops = {
 107         .insert = tid_rb_insert,
 108         .remove = tid_rb_remove,
 109         .invalidate = tid_rb_invalidate
 110 };
 111
 112 static inline u32 rcventry2tidinfo(u32 rcventry)
 113 {
 114         u32 pair = rcventry & ~0x1;
 115
 116         return EXP_TID_SET(IDX, pair >> 1) |
 117                 EXP_TID_SET(CTRL, 1 << (rcventry - pair));
 118 }
 119
 120 static inline void exp_tid_group_init(struct exp_tid_set *set)
 121 {
 122         INIT_LIST_HEAD(&set->list);
 123         set->count = 0;
 124 }
 125
 126 static inline void tid_group_remove(struct tid_group *grp,
 127                                     struct exp_tid_set *set)
 128 {
 129         list_del_init(&grp->list);
 130         set->count--;
 131 }
 132
 133 static inline void tid_group_add_tail(struct tid_group *grp,
 134                                       struct exp_tid_set *set)
 135 {
 136         list_add_tail(&grp->list, &set->list);
 137         set->count++;
 138 }
 139
 140 static inline struct tid_group *tid_group_pop(struct exp_tid_set *set)
 141 {
 142         struct tid_group *grp =
 143                 list_first_entry(&set->list, struct tid_group, list);
 144         list_del_init(&grp->list);
 145         set->count--;
 146         return grp;
 147 }
 148
 149 static inline void tid_group_move(struct tid_group *group,
 150                                   struct exp_tid_set *s1,
 151                                   struct exp_tid_set *s2)
 152 {
 153         tid_group_remove(group, s1);
 154         tid_group_add_tail(group, s2);
 155 }
 156
 157 int hfi1_user_exp_rcv_grp_init(struct hfi1_filedata *fd)
 158 {
 159         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 160         struct hfi1_devdata *dd = fd->dd;
 161         u32 tidbase;
 162         u32 i;
 163         struct tid_group *grp, *gptr;
 164
 165         exp_tid_group_init(&uctxt->tid_group_list);
 166         exp_tid_group_init(&uctxt->tid_used_list);
 167         exp_tid_group_init(&uctxt->tid_full_list);
 168
 169         tidbase = uctxt->expected_base;
 170         for (i = 0; i < uctxt->expected_count /
 171                      dd->rcv_entries.group_size; i++) {
 172                 grp = kzalloc(sizeof(*grp), GFP_KERNEL);
 173                 if (!grp)
 174                         goto grp_failed;
 175
 176                 grp->size = dd->rcv_entries.group_size;
 177                 grp->base = tidbase;
 178                 tid_group_add_tail(grp, &uctxt->tid_group_list);
 179                 tidbase += dd->rcv_entries.group_size;
 180         }
 181
 182         return 0;
 183
 184 grp_failed:
 185         list_for_each_entry_safe(grp, gptr, &uctxt->tid_group_list.list,
 186                                  list) {
 187                 list_del_init(&grp->list);
 188                 kfree(grp);
 189         }
 190
 191         return -ENOMEM;
 192 }
 193
 194 /*
 195  * Initialize context and file private data needed for Expected
 196  * receive caching. This needs to be done after the context has
 197  * been configured with the eager/expected RcvEntry counts.
 198  */
 199 int hfi1_user_exp_rcv_init(struct hfi1_filedata *fd)
 200 {
 201         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 202         struct hfi1_devdata *dd = uctxt->dd;
 203         int ret = 0;
 204
 205         spin_lock_init(&fd->tid_lock);
 206         spin_lock_init(&fd->invalid_lock);
 207
 208         fd->entry_to_rb = kcalloc(uctxt->expected_count,
 209                                   sizeof(struct rb_node *),
 210                                   GFP_KERNEL);
 211         if (!fd->entry_to_rb)
 212                 return -ENOMEM;
 213
 214         if (!HFI1_CAP_UGET_MASK(uctxt->flags, TID_UNMAP)) {
 215                 fd->invalid_tid_idx = 0;
 216                 fd->invalid_tids = kcalloc(uctxt->expected_count,
 217                                            sizeof(*fd->invalid_tids),
 218                                            GFP_KERNEL);
 219                 if (!fd->invalid_tids) {
 220                         kfree(fd->entry_to_rb);
 221                         fd->entry_to_rb = NULL;
 222                         return -ENOMEM;
 223                 }
 224
 225                 /*
 226                  * Register MMU notifier callbacks. If the registration
 227                  * fails, continue without TID caching for this context.
 228                  */
 229                 ret = hfi1_mmu_rb_register(fd, fd->mm, &tid_rb_ops,
 230                                            dd->pport->hfi1_wq,
 231                                            &fd->handler);
 232                 if (ret) {
 233                         dd_dev_info(dd,
 234                                     "Failed MMU notifier registration %d\n",
 235                                     ret);
 236                         ret = 0;
 237                 }
 238         }
 239
 240         /*
 241          * PSM does not have a good way to separate, count, and
 242          * effectively enforce a limit on RcvArray entries used by
 243          * subctxts (when context sharing is used) when TID caching
 244          * is enabled. To help with that, we calculate a per-process
 245          * RcvArray entry share and enforce that.
 246          * If TID caching is not in use, PSM deals with usage on its
 247          * own. In that case, we allow any subctxt to take all of the
 248          * entries.
 249          *
 250          * Make sure that we set the tid counts only after successful
 251          * init.
 252          */
 253         spin_lock(&fd->tid_lock);
 254         if (uctxt->subctxt_cnt && fd->handler) {
 255                 u16 remainder;
 256
 257                 fd->tid_limit = uctxt->expected_count / uctxt->subctxt_cnt;
 258                 remainder = uctxt->expected_count % uctxt->subctxt_cnt;
 259                 if (remainder && fd->subctxt < remainder)
 260                         fd->tid_limit++;
 261         } else {
 262                 fd->tid_limit = uctxt->expected_count;
 263         }
 264         spin_unlock(&fd->tid_lock);
 265
 266         return ret;
 267 }
 268
 269 void hfi1_user_exp_rcv_grp_free(struct hfi1_ctxtdata *uctxt)
 270 {
 271         struct tid_group *grp, *gptr;
 272
 273         list_for_each_entry_safe(grp, gptr, &uctxt->tid_group_list.list,
 274                                  list) {
 275                 list_del_init(&grp->list);
 276                 kfree(grp);
 277         }
 278         hfi1_clear_tids(uctxt);
 279 }
 280
 281 void hfi1_user_exp_rcv_free(struct hfi1_filedata *fd)
 282 {
 283         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 284
 285         /*
 286          * The notifier would have been removed when the process'es mm
 287          * was freed.
 288          */
 289         if (fd->handler) {
 290                 hfi1_mmu_rb_unregister(fd->handler);
 291         } else {
 292                 if (!EXP_TID_SET_EMPTY(uctxt->tid_full_list))
 293                         unlock_exp_tids(uctxt, &uctxt->tid_full_list, fd);
 294                 if (!EXP_TID_SET_EMPTY(uctxt->tid_used_list))
 295                         unlock_exp_tids(uctxt, &uctxt->tid_used_list, fd);
 296         }
 297
 298         kfree(fd->invalid_tids);
 299         fd->invalid_tids = NULL;
 300
 301         kfree(fd->entry_to_rb);
 302         fd->entry_to_rb = NULL;
 303 }
 304
 305 /*
 306  * Write an "empty" RcvArray entry.
 307  * This function exists so the TID registaration code can use it
 308  * to write to unused/unneeded entries and still take advantage
 309  * of the WC performance improvements. The HFI will ignore this
 310  * write to the RcvArray entry.
 311  */
 312 static inline void rcv_array_wc_fill(struct hfi1_devdata *dd, u32 index)
 313 {
 314         /*
 315          * Doing the WC fill writes only makes sense if the device is
 316          * present and the RcvArray has been mapped as WC memory.
 317          */
 318         if ((dd->flags & HFI1_PRESENT) && dd->rcvarray_wc)
 319                 writeq(0, dd->rcvarray_wc + (index * 8));
 320 }
 321
 322 /*
 323  * RcvArray entry allocation for Expected Receives is done by the
 324  * following algorithm:
 325  *
 326  * The context keeps 3 lists of groups of RcvArray entries:
 327  *   1. List of empty groups - tid_group_list
 328  *      This list is created during user context creation and
 329  *      contains elements which describe sets (of 8) of empty
 330  *      RcvArray entries.
 331  *   2. List of partially used groups - tid_used_list
 332  *      This list contains sets of RcvArray entries which are
 333  *      not completely used up. Another mapping request could
 334  *      use some of all of the remaining entries.
 335  *   3. List of full groups - tid_full_list
 336  *      This is the list where sets that are completely used
 337  *      up go.
 338  *
 339  * An attempt to optimize the usage of RcvArray entries is
 340  * made by finding all sets of physically contiguous pages in a
 341  * user's buffer.
 342  * These physically contiguous sets are further split into
 343  * sizes supported by the receive engine of the HFI. The
 344  * resulting sets of pages are stored in struct tid_pageset,
 345  * which describes the sets as:
 346  *    * .count - number of pages in this set
 347  *    * .idx - starting index into struct page ** array
 348  *                    of this set
 349  *
 350  * From this point on, the algorithm deals with the page sets
 351  * described above. The number of pagesets is divided by the
 352  * RcvArray group size to produce the number of full groups
 353  * needed.
 354  *
 355  * Groups from the 3 lists are manipulated using the following
 356  * rules:
 357  *   1. For each set of 8 pagesets, a complete group from
 358  *      tid_group_list is taken, programmed, and moved to
 359  *      the tid_full_list list.
 360  *   2. For all remaining pagesets:
 361  *      2.1 If the tid_used_list is empty and the tid_group_list
 362  *          is empty, stop processing pageset and return only
 363  *          what has been programmed up to this point.
 364  *      2.2 If the tid_used_list is empty and the tid_group_list
 365  *          is not empty, move a group from tid_group_list to
 366  *          tid_used_list.
 367  *      2.3 For each group is tid_used_group, program as much as
 368  *          can fit into the group. If the group becomes fully
 369  *          used, move it to tid_full_list.
 370  */
 371 int hfi1_user_exp_rcv_setup(struct hfi1_filedata *fd,
 372                             struct hfi1_tid_info *tinfo)
 373 {
 374         int ret = 0, need_group = 0, pinned;
 375         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 376         struct hfi1_devdata *dd = uctxt->dd;
 377         unsigned npages, ngroups, pageidx = 0, pageset_count, npagesets,
 378                 tididx = 0, mapped, mapped_pages = 0;
 379         unsigned long vaddr = tinfo->vaddr;
 380         struct page **pages = NULL;
 381         u32 *tidlist = NULL;
 382         struct tid_pageset *pagesets = NULL;
 383
 384         /* Get the number of pages the user buffer spans */
 385         npages = num_user_pages(vaddr, tinfo->length);
 386         if (!npages)
 387                 return -EINVAL;
 388
 389         if (npages > uctxt->expected_count) {
 390                 dd_dev_err(dd, "Expected buffer too big\n");
 391                 return -EINVAL;
 392         }
 393
 394         /* Verify that access is OK for the user buffer */
 395         if (!access_ok(VERIFY_WRITE, (void __user *)vaddr,
 396                        npages * PAGE_SIZE)) {
 397                 dd_dev_err(dd, "Fail vaddr %p, %u pages, !access_ok\n",
 398                            (void *)vaddr, npages);
 399                 return -EFAULT;
 400         }
 401
 402         pagesets = kcalloc(uctxt->expected_count, sizeof(*pagesets),
 403                            GFP_KERNEL);
 404         if (!pagesets)
 405                 return -ENOMEM;
 406
 407         /* Allocate the array of struct page pointers needed for pinning */
 408         pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL);
 409         if (!pages) {
 410                 ret = -ENOMEM;
 411                 goto bail;
 412         }
 413
 414         /*
 415          * Pin all the pages of the user buffer. If we can't pin all the
 416          * pages, accept the amount pinned so far and program only that.
 417          * User space knows how to deal with partially programmed buffers.
 418          */
 419         if (!hfi1_can_pin_pages(dd, fd->mm, fd->tid_n_pinned, npages)) {
 420                 ret = -ENOMEM;
 421                 goto bail;
 422         }
 423
 424         pinned = hfi1_acquire_user_pages(fd->mm, vaddr, npages, true, pages);
 425         if (pinned <= 0) {
 426                 ret = pinned;
 427                 goto bail;
 428         }
 429         fd->tid_n_pinned += npages;
 430
 431         /* Find sets of physically contiguous pages */
 432         npagesets = find_phys_blocks(pages, pinned, pagesets);
 433
 434         /*
 435          * We don't need to access this under a lock since tid_used is per
 436          * process and the same process cannot be in hfi1_user_exp_rcv_clear()
 437          * and hfi1_user_exp_rcv_setup() at the same time.
 438          */
 439         spin_lock(&fd->tid_lock);
 440         if (fd->tid_used + npagesets > fd->tid_limit)
 441                 pageset_count = fd->tid_limit - fd->tid_used;
 442         else
 443                 pageset_count = npagesets;
 444         spin_unlock(&fd->tid_lock);
 445
 446         if (!pageset_count)
 447                 goto bail;
 448
 449         ngroups = pageset_count / dd->rcv_entries.group_size;
 450         tidlist = kcalloc(pageset_count, sizeof(*tidlist), GFP_KERNEL);
 451         if (!tidlist) {
 452                 ret = -ENOMEM;
 453                 goto nomem;
 454         }
 455
 456         tididx = 0;
 457
 458         /*
 459          * From this point on, we are going to be using shared (between master
 460          * and subcontexts) context resources. We need to take the lock.
 461          */
 462         mutex_lock(&uctxt->exp_lock);
 463         /*
 464          * The first step is to program the RcvArray entries which are complete
 465          * groups.
 466          */
 467         while (ngroups && uctxt->tid_group_list.count) {
 468                 struct tid_group *grp =
 469                         tid_group_pop(&uctxt->tid_group_list);
 470
 471                 ret = program_rcvarray(fd, vaddr, grp, pagesets,
 472                                        pageidx, dd->rcv_entries.group_size,
 473                                        pages, tidlist, &tididx, &mapped);
 474                 /*
 475                  * If there was a failure to program the RcvArray
 476                  * entries for the entire group, reset the grp fields
 477                  * and add the grp back to the free group list.
 478                  */
 479                 if (ret <= 0) {
 480                         tid_group_add_tail(grp, &uctxt->tid_group_list);
 481                         hfi1_cdbg(TID,
 482                                   "Failed to program RcvArray group %d", ret);
 483                         goto unlock;
 484                 }
 485
 486                 tid_group_add_tail(grp, &uctxt->tid_full_list);
 487                 ngroups--;
 488                 pageidx += ret;
 489                 mapped_pages += mapped;
 490         }
 491
 492         while (pageidx < pageset_count) {
 493                 struct tid_group *grp, *ptr;
 494                 /*
 495                  * If we don't have any partially used tid groups, check
 496                  * if we have empty groups. If so, take one from there and
 497                  * put in the partially used list.
 498                  */
 499                 if (!uctxt->tid_used_list.count || need_group) {
 500                         if (!uctxt->tid_group_list.count)
 501                                 goto unlock;
 502
 503                         grp = tid_group_pop(&uctxt->tid_group_list);
 504                         tid_group_add_tail(grp, &uctxt->tid_used_list);
 505                         need_group = 0;
 506                 }
 507                 /*
 508                  * There is an optimization opportunity here - instead of
 509                  * fitting as many page sets as we can, check for a group
 510                  * later on in the list that could fit all of them.
 511                  */
 512                 list_for_each_entry_safe(grp, ptr, &uctxt->tid_used_list.list,
 513                                          list) {
 514                         unsigned use = min_t(unsigned, pageset_count - pageidx,
 515                                              grp->size - grp->used);
 516
 517                         ret = program_rcvarray(fd, vaddr, grp, pagesets,
 518                                                pageidx, use, pages, tidlist,
 519                                                &tididx, &mapped);
 520                         if (ret < 0) {
 521                                 hfi1_cdbg(TID,
 522                                           "Failed to program RcvArray entries %d",
 523                                           ret);
 524                                 ret = -EFAULT;
 525                                 goto unlock;
 526                         } else if (ret > 0) {
 527                                 if (grp->used == grp->size)
 528                                         tid_group_move(grp,
 529                                                        &uctxt->tid_used_list,
 530                                                        &uctxt->tid_full_list);
 531                                 pageidx += ret;
 532                                 mapped_pages += mapped;
 533                                 need_group = 0;
 534                                 /* Check if we are done so we break out early */
 535                                 if (pageidx >= pageset_count)
 536                                         break;
 537                         } else if (WARN_ON(ret == 0)) {
 538                                 /*
 539                                  * If ret is 0, we did not program any entries
 540                                  * into this group, which can only happen if
 541                                  * we've screwed up the accounting somewhere.
 542                                  * Warn and try to continue.
 543                                  */
 544                                 need_group = 1;
 545                         }
 546                 }
 547         }
 548 unlock:
 549         mutex_unlock(&uctxt->exp_lock);
 550 nomem:
 551         hfi1_cdbg(TID, "total mapped: tidpairs:%u pages:%u (%d)", tididx,
 552                   mapped_pages, ret);
 553         if (tididx) {
 554                 spin_lock(&fd->tid_lock);
 555                 fd->tid_used += tididx;
 556                 spin_unlock(&fd->tid_lock);
 557                 tinfo->tidcnt = tididx;
 558                 tinfo->length = mapped_pages * PAGE_SIZE;
 559
 560                 if (copy_to_user((void __user *)(unsigned long)tinfo->tidlist,
 561                                  tidlist, sizeof(tidlist[0]) * tididx)) {
 562                         /*
 563                          * On failure to copy to the user level, we need to undo
 564                          * everything done so far so we don't leak resources.
 565                          */
 566                         tinfo->tidlist = (unsigned long)&tidlist;
 567                         hfi1_user_exp_rcv_clear(fd, tinfo);
 568                         tinfo->tidlist = 0;
 569                         ret = -EFAULT;
 570                         goto bail;
 571                 }
 572         }
 573
 574         /*
 575          * If not everything was mapped (due to insufficient RcvArray entries,
 576          * for example), unpin all unmapped pages so we can pin them nex time.
 577          */
 578         if (mapped_pages != pinned) {
 579                 hfi1_release_user_pages(fd->mm, &pages[mapped_pages],
 580                                         pinned - mapped_pages,
 581                                         false);
 582                 fd->tid_n_pinned -= pinned - mapped_pages;
 583         }
 584 bail:
 585         kfree(pagesets);
 586         kfree(pages);
 587         kfree(tidlist);
 588         return ret > 0 ? 0 : ret;
 589 }
 590
 591 int hfi1_user_exp_rcv_clear(struct hfi1_filedata *fd,
 592                             struct hfi1_tid_info *tinfo)
 593 {
 594         int ret = 0;
 595         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 596         u32 *tidinfo;
 597         unsigned tididx;
 598
 599         if (unlikely(tinfo->tidcnt > fd->tid_used))
 600                 return -EINVAL;
 601
 602         tidinfo = memdup_user((void __user *)(unsigned long)tinfo->tidlist,
 603                               sizeof(tidinfo[0]) * tinfo->tidcnt);
 604         if (IS_ERR(tidinfo))
 605                 return PTR_ERR(tidinfo);
 606
 607         mutex_lock(&uctxt->exp_lock);
 608         for (tididx = 0; tididx < tinfo->tidcnt; tididx++) {
 609                 ret = unprogram_rcvarray(fd, tidinfo[tididx], NULL);
 610                 if (ret) {
 611                         hfi1_cdbg(TID, "Failed to unprogram rcv array %d",
 612                                   ret);
 613                         break;
 614                 }
 615         }
 616         spin_lock(&fd->tid_lock);
 617         fd->tid_used -= tididx;
 618         spin_unlock(&fd->tid_lock);
 619         tinfo->tidcnt = tididx;
 620         mutex_unlock(&uctxt->exp_lock);
 621
 622         kfree(tidinfo);
 623         return ret;
 624 }
 625
 626 int hfi1_user_exp_rcv_invalid(struct hfi1_filedata *fd,
 627                               struct hfi1_tid_info *tinfo)
 628 {
 629         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 630         unsigned long *ev = uctxt->dd->events +
 631                 (((uctxt->ctxt - uctxt->dd->first_dyn_alloc_ctxt) *
 632                   HFI1_MAX_SHARED_CTXTS) + fd->subctxt);
 633         u32 *array;
 634         int ret = 0;
 635
 636         if (!fd->invalid_tids)
 637                 return -EINVAL;
 638
 639         /*
 640          * copy_to_user() can sleep, which will leave the invalid_lock
 641          * locked and cause the MMU notifier to be blocked on the lock
 642          * for a long time.
 643          * Copy the data to a local buffer so we can release the lock.
 644          */
 645         array = kcalloc(uctxt->expected_count, sizeof(*array), GFP_KERNEL);
 646         if (!array)
 647                 return -EFAULT;
 648
 649         spin_lock(&fd->invalid_lock);
 650         if (fd->invalid_tid_idx) {
 651                 memcpy(array, fd->invalid_tids, sizeof(*array) *
 652                        fd->invalid_tid_idx);
 653                 memset(fd->invalid_tids, 0, sizeof(*fd->invalid_tids) *
 654                        fd->invalid_tid_idx);
 655                 tinfo->tidcnt = fd->invalid_tid_idx;
 656                 fd->invalid_tid_idx = 0;
 657                 /*
 658                  * Reset the user flag while still holding the lock.
 659                  * Otherwise, PSM can miss events.
 660                  */
 661                 clear_bit(_HFI1_EVENT_TID_MMU_NOTIFY_BIT, ev);
 662         } else {
 663                 tinfo->tidcnt = 0;
 664         }
 665         spin_unlock(&fd->invalid_lock);
 666
 667         if (tinfo->tidcnt) {
 668                 if (copy_to_user((void __user *)tinfo->tidlist,
 669                                  array, sizeof(*array) * tinfo->tidcnt))
 670                         ret = -EFAULT;
 671         }
 672         kfree(array);
 673
 674         return ret;
 675 }
 676
 677 static u32 find_phys_blocks(struct page **pages, unsigned npages,
 678                             struct tid_pageset *list)
 679 {
 680         unsigned pagecount, pageidx, setcount = 0, i;
 681         unsigned long pfn, this_pfn;
 682
 683         if (!npages)
 684                 return 0;
 685
 686         /*
 687          * Look for sets of physically contiguous pages in the user buffer.
 688          * This will allow us to optimize Expected RcvArray entry usage by
 689          * using the bigger supported sizes.
 690          */
 691         pfn = page_to_pfn(pages[0]);
 692         for (pageidx = 0, pagecount = 1, i = 1; i <= npages; i++) {
 693                 this_pfn = i < npages ? page_to_pfn(pages[i]) : 0;
 694
 695                 /*
 696                  * If the pfn's are not sequential, pages are not physically
 697                  * contiguous.
 698                  */
 699                 if (this_pfn != ++pfn) {
 700                         /*
 701                          * At this point we have to loop over the set of
 702                          * physically contiguous pages and break them down it
 703                          * sizes supported by the HW.
 704                          * There are two main constraints:
 705                          *     1. The max buffer size is MAX_EXPECTED_BUFFER.
 706                          *        If the total set size is bigger than that
 707                          *        program only a MAX_EXPECTED_BUFFER chunk.
 708                          *     2. The buffer size has to be a power of two. If
 709                          *        it is not, round down to the closes power of
 710                          *        2 and program that size.
 711                          */
 712                         while (pagecount) {
 713                                 int maxpages = pagecount;
 714                                 u32 bufsize = pagecount * PAGE_SIZE;
 715
 716                                 if (bufsize > MAX_EXPECTED_BUFFER)
 717                                         maxpages =
 718                                                 MAX_EXPECTED_BUFFER >>
 719                                                 PAGE_SHIFT;
 720                                 else if (!is_power_of_2(bufsize))
 721                                         maxpages =
 722                                                 rounddown_pow_of_two(bufsize) >>
 723                                                 PAGE_SHIFT;
 724
 725                                 list[setcount].idx = pageidx;
 726                                 list[setcount].count = maxpages;
 727                                 pagecount -= maxpages;
 728                                 pageidx += maxpages;
 729                                 setcount++;
 730                         }
 731                         pageidx = i;
 732                         pagecount = 1;
 733                         pfn = this_pfn;
 734                 } else {
 735                         pagecount++;
 736                 }
 737         }
 738         return setcount;
 739 }
 740
 741 /**
 742  * program_rcvarray() - program an RcvArray group with receive buffers
 743  * @fd: filedata pointer
 744  * @vaddr: starting user virtual address
 745  * @grp: RcvArray group
 746  * @sets: array of struct tid_pageset holding information on physically
 747  *        contiguous chunks from the user buffer
 748  * @start: starting index into sets array
 749  * @count: number of struct tid_pageset's to program
 750  * @pages: an array of struct page * for the user buffer
 751  * @tidlist: the array of u32 elements when the information about the
 752  *           programmed RcvArray entries is to be encoded.
 753  * @tididx: starting offset into tidlist
 754  * @pmapped: (output parameter) number of pages programmed into the RcvArray
 755  *           entries.
 756  *
 757  * This function will program up to 'count' number of RcvArray entries from the
 758  * group 'grp'. To make best use of write-combining writes, the function will
 759  * perform writes to the unused RcvArray entries which will be ignored by the
 760  * HW. Each RcvArray entry will be programmed with a physically contiguous
 761  * buffer chunk from the user's virtual buffer.
 762  *
 763  * Return:
 764  * -EINVAL if the requested count is larger than the size of the group,
 765  * -ENOMEM or -EFAULT on error from set_rcvarray_entry(), or
 766  * number of RcvArray entries programmed.
 767  */
 768 static int program_rcvarray(struct hfi1_filedata *fd, unsigned long vaddr,
 769                             struct tid_group *grp,
 770                             struct tid_pageset *sets,
 771                             unsigned start, u16 count, struct page **pages,
 772                             u32 *tidlist, unsigned *tididx, unsigned *pmapped)
 773 {
 774         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 775         struct hfi1_devdata *dd = uctxt->dd;
 776         u16 idx;
 777         u32 tidinfo = 0, rcventry, useidx = 0;
 778         int mapped = 0;
 779
 780         /* Count should never be larger than the group size */
 781         if (count > grp->size)
 782                 return -EINVAL;
 783
 784         /* Find the first unused entry in the group */
 785         for (idx = 0; idx < grp->size; idx++) {
 786                 if (!(grp->map & (1 << idx))) {
 787                         useidx = idx;
 788                         break;
 789                 }
 790                 rcv_array_wc_fill(dd, grp->base + idx);
 791         }
 792
 793         idx = 0;
 794         while (idx < count) {
 795                 u16 npages, pageidx, setidx = start + idx;
 796                 int ret = 0;
 797
 798                 /*
 799                  * If this entry in the group is used, move to the next one.
 800                  * If we go past the end of the group, exit the loop.
 801                  */
 802                 if (useidx >= grp->size) {
 803                         break;
 804                 } else if (grp->map & (1 << useidx)) {
 805                         rcv_array_wc_fill(dd, grp->base + useidx);
 806                         useidx++;
 807                         continue;
 808                 }
 809
 810                 rcventry = grp->base + useidx;
 811                 npages = sets[setidx].count;
 812                 pageidx = sets[setidx].idx;
 813
 814                 ret = set_rcvarray_entry(fd, vaddr + (pageidx * PAGE_SIZE),
 815                                          rcventry, grp, pages + pageidx,
 816                                          npages);
 817                 if (ret)
 818                         return ret;
 819                 mapped += npages;
 820
 821                 tidinfo = rcventry2tidinfo(rcventry - uctxt->expected_base) |
 822                         EXP_TID_SET(LEN, npages);
 823                 tidlist[(*tididx)++] = tidinfo;
 824                 grp->used++;
 825                 grp->map |= 1 << useidx++;
 826                 idx++;
 827         }
 828
 829         /* Fill the rest of the group with "blank" writes */
 830         for (; useidx < grp->size; useidx++)
 831                 rcv_array_wc_fill(dd, grp->base + useidx);
 832         *pmapped = mapped;
 833         return idx;
 834 }
 835
 836 static int set_rcvarray_entry(struct hfi1_filedata *fd, unsigned long vaddr,
 837                               u32 rcventry, struct tid_group *grp,
 838                               struct page **pages, unsigned npages)
 839 {
 840         int ret;
 841         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 842         struct tid_rb_node *node;
 843         struct hfi1_devdata *dd = uctxt->dd;
 844         dma_addr_t phys;
 845
 846         /*
 847          * Allocate the node first so we can handle a potential
 848          * failure before we've programmed anything.
 849          */
 850         node = kzalloc(sizeof(*node) + (sizeof(struct page *) * npages),
 851                        GFP_KERNEL);
 852         if (!node)
 853                 return -ENOMEM;
 854
 855         phys = pci_map_single(dd->pcidev,
 856                               __va(page_to_phys(pages[0])),
 857                               npages * PAGE_SIZE, PCI_DMA_FROMDEVICE);
 858         if (dma_mapping_error(&dd->pcidev->dev, phys)) {
 859                 dd_dev_err(dd, "Failed to DMA map Exp Rcv pages 0x%llx\n",
 860                            phys);
 861                 kfree(node);
 862                 return -EFAULT;
 863         }
 864
 865         node->mmu.addr = vaddr;
 866         node->mmu.len = npages * PAGE_SIZE;
 867         node->phys = page_to_phys(pages[0]);
 868         node->npages = npages;
 869         node->rcventry = rcventry;
 870         node->dma_addr = phys;
 871         node->grp = grp;
 872         node->freed = false;
 873         memcpy(node->pages, pages, sizeof(struct page *) * npages);
 874
 875         if (!fd->handler)
 876                 ret = tid_rb_insert(fd, &node->mmu);
 877         else
 878                 ret = hfi1_mmu_rb_insert(fd->handler, &node->mmu);
 879
 880         if (ret) {
 881                 hfi1_cdbg(TID, "Failed to insert RB node %u 0x%lx, 0x%lx %d",
 882                           node->rcventry, node->mmu.addr, node->phys, ret);
 883                 pci_unmap_single(dd->pcidev, phys, npages * PAGE_SIZE,
 884                                  PCI_DMA_FROMDEVICE);
 885                 kfree(node);
 886                 return -EFAULT;
 887         }
 888         hfi1_put_tid(dd, rcventry, PT_EXPECTED, phys, ilog2(npages) + 1);
 889         trace_hfi1_exp_tid_reg(uctxt->ctxt, fd->subctxt, rcventry, npages,
 890                                node->mmu.addr, node->phys, phys);
 891         return 0;
 892 }
 893
 894 static int unprogram_rcvarray(struct hfi1_filedata *fd, u32 tidinfo,
 895                               struct tid_group **grp)
 896 {
 897         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 898         struct hfi1_devdata *dd = uctxt->dd;
 899         struct tid_rb_node *node;
 900         u8 tidctrl = EXP_TID_GET(tidinfo, CTRL);
 901         u32 tididx = EXP_TID_GET(tidinfo, IDX) << 1, rcventry;
 902
 903         if (tididx >= uctxt->expected_count) {
 904                 dd_dev_err(dd, "Invalid RcvArray entry (%u) index for ctxt %u\n",
 905                            tididx, uctxt->ctxt);
 906                 return -EINVAL;
 907         }
 908
 909         if (tidctrl == 0x3)
 910                 return -EINVAL;
 911
 912         rcventry = tididx + (tidctrl - 1);
 913
 914         node = fd->entry_to_rb[rcventry];
 915         if (!node || node->rcventry != (uctxt->expected_base + rcventry))
 916                 return -EBADF;
 917
 918         if (grp)
 919                 *grp = node->grp;
 920
 921         if (!fd->handler)
 922                 cacheless_tid_rb_remove(fd, node);
 923         else
 924                 hfi1_mmu_rb_remove(fd->handler, &node->mmu);
 925
 926         return 0;
 927 }
 928
 929 static void clear_tid_node(struct hfi1_filedata *fd, struct tid_rb_node *node)
 930 {
 931         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 932         struct hfi1_devdata *dd = uctxt->dd;
 933
 934         trace_hfi1_exp_tid_unreg(uctxt->ctxt, fd->subctxt, node->rcventry,
 935                                  node->npages, node->mmu.addr, node->phys,
 936                                  node->dma_addr);
 937
 938         hfi1_put_tid(dd, node->rcventry, PT_INVALID, 0, 0);
 939         /*
 940          * Make sure device has seen the write before we unpin the
 941          * pages.
 942          */
 943         flush_wc();
 944
 945         pci_unmap_single(dd->pcidev, node->dma_addr, node->mmu.len,
 946                          PCI_DMA_FROMDEVICE);
 947         hfi1_release_user_pages(fd->mm, node->pages, node->npages, true);
 948         fd->tid_n_pinned -= node->npages;
 949
 950         node->grp->used--;
 951         node->grp->map &= ~(1 << (node->rcventry - node->grp->base));
 952
 953         if (node->grp->used == node->grp->size - 1)
 954                 tid_group_move(node->grp, &uctxt->tid_full_list,
 955                                &uctxt->tid_used_list);
 956         else if (!node->grp->used)
 957                 tid_group_move(node->grp, &uctxt->tid_used_list,
 958                                &uctxt->tid_group_list);
 959         kfree(node);
 960 }
 961
 962 /*
 963  * As a simple helper for hfi1_user_exp_rcv_free, this function deals with
 964  * clearing nodes in the non-cached case.
 965  */
 966 static void unlock_exp_tids(struct hfi1_ctxtdata *uctxt,
 967                             struct exp_tid_set *set,
 968                             struct hfi1_filedata *fd)
 969 {
 970         struct tid_group *grp, *ptr;
 971         int i;
 972
 973         list_for_each_entry_safe(grp, ptr, &set->list, list) {
 974                 list_del_init(&grp->list);
 975
 976                 for (i = 0; i < grp->size; i++) {
 977                         if (grp->map & (1 << i)) {
 978                                 u16 rcventry = grp->base + i;
 979                                 struct tid_rb_node *node;
 980
 981                                 node = fd->entry_to_rb[rcventry -
 982                                                           uctxt->expected_base];
 983                                 if (!node || node->rcventry != rcventry)
 984                                         continue;
 985
 986                                 cacheless_tid_rb_remove(fd, node);
 987                         }
 988                 }
 989         }
 990 }
 991
 992 /*
 993  * Always return 0 from this function.  A non-zero return indicates that the
 994  * remove operation will be called and that memory should be unpinned.
 995  * However, the driver cannot unpin out from under PSM.  Instead, retain the
 996  * memory (by returning 0) and inform PSM that the memory is going away.  PSM
 997  * will call back later when it has removed the memory from its list.
 998  */
 999 static int tid_rb_invalidate(void *arg, struct mmu_rb_node *mnode)
1000 {
1001         struct hfi1_filedata *fdata = arg;
1002         struct hfi1_ctxtdata *uctxt = fdata->uctxt;
1003         struct tid_rb_node *node =
1004                 container_of(mnode, struct tid_rb_node, mmu);
1005
1006         if (node->freed)
1007                 return 0;
1008
1009         trace_hfi1_exp_tid_inval(uctxt->ctxt, fdata->subctxt, node->mmu.addr,
1010                                  node->rcventry, node->npages, node->dma_addr);
1011         node->freed = true;
1012
1013         spin_lock(&fdata->invalid_lock);
1014         if (fdata->invalid_tid_idx < uctxt->expected_count) {
1015                 fdata->invalid_tids[fdata->invalid_tid_idx] =
1016                         rcventry2tidinfo(node->rcventry - uctxt->expected_base);
1017                 fdata->invalid_tids[fdata->invalid_tid_idx] |=
1018                         EXP_TID_SET(LEN, node->npages);
1019                 if (!fdata->invalid_tid_idx) {
1020                         unsigned long *ev;
1021
1022                         /*
1023                          * hfi1_set_uevent_bits() sets a user event flag
1024                          * for all processes. Because calling into the
1025                          * driver to process TID cache invalidations is
1026                          * expensive and TID cache invalidations are
1027                          * handled on a per-process basis, we can
1028                          * optimize this to set the flag only for the
1029                          * process in question.
1030                          */
1031                         ev = uctxt->dd->events +
1032                           (((uctxt->ctxt - uctxt->dd->first_dyn_alloc_ctxt) *
1033                             HFI1_MAX_SHARED_CTXTS) + fdata->subctxt);
1034                         set_bit(_HFI1_EVENT_TID_MMU_NOTIFY_BIT, ev);
1035                 }
1036                 fdata->invalid_tid_idx++;
1037         }
1038         spin_unlock(&fdata->invalid_lock);
1039         return 0;
1040 }
1041
1042 static int tid_rb_insert(void *arg, struct mmu_rb_node *node)
1043 {
1044         struct hfi1_filedata *fdata = arg;
1045         struct tid_rb_node *tnode =
1046                 container_of(node, struct tid_rb_node, mmu);
1047         u32 base = fdata->uctxt->expected_base;
1048
1049         fdata->entry_to_rb[tnode->rcventry - base] = tnode;
1050         return 0;
1051 }
1052
1053 static void cacheless_tid_rb_remove(struct hfi1_filedata *fdata,
1054                                     struct tid_rb_node *tnode)
1055 {
1056         u32 base = fdata->uctxt->expected_base;
1057
1058         fdata->entry_to_rb[tnode->rcventry - base] = NULL;
1059         clear_tid_node(fdata, tnode);
1060 }
1061
1062 static void tid_rb_remove(void *arg, struct mmu_rb_node *node)
1063 {
1064         struct hfi1_filedata *fdata = arg;
1065         struct tid_rb_node *tnode =
1066                 container_of(node, struct tid_rb_node, mmu);
1067
1068         cacheless_tid_rb_remove(fdata, tnode);
1069 }