drivers/infiniband/hw/hfi1/user_exp_rcv.c

   1 /*
   2  * Copyright(c) 2015, 2016 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
  49 #include "user_exp_rcv.h"
  50 #include "trace.h"
  51 #include "mmu_rb.h"
  52
  53 struct tid_group {
  54         struct list_head list;
  55         unsigned base;
  56         u8 size;
  57         u8 used;
  58         u8 map;
  59 };
  60
  61 struct tid_rb_node {
  62         struct mmu_rb_node mmu;
  63         unsigned long phys;
  64         struct tid_group *grp;
  65         u32 rcventry;
  66         dma_addr_t dma_addr;
  67         bool freed;
  68         unsigned npages;
  69         struct page *pages[0];
  70 };
  71
  72 struct tid_pageset {
  73         u16 idx;
  74         u16 count;
  75 };
  76
  77 #define EXP_TID_SET_EMPTY(set) (set.count == 0 && list_empty(&set.list))
  78
  79 #define num_user_pages(vaddr, len)                                     \
  80         (1 + (((((unsigned long)(vaddr) +                              \
  81                  (unsigned long)(len) - 1) & PAGE_MASK) -              \
  82                ((unsigned long)vaddr & PAGE_MASK)) >> PAGE_SHIFT))
  83
  84 static void unlock_exp_tids(struct hfi1_ctxtdata *, struct exp_tid_set *,
  85                             struct rb_root *);
  86 static u32 find_phys_blocks(struct page **, unsigned, struct tid_pageset *);
  87 static int set_rcvarray_entry(struct file *, unsigned long, u32,
  88                               struct tid_group *, struct page **, unsigned);
  89 static int mmu_rb_insert(struct rb_root *, struct mmu_rb_node *);
  90 static void mmu_rb_remove(struct rb_root *, struct mmu_rb_node *,
  91                           struct mm_struct *);
  92 static int mmu_rb_invalidate(struct rb_root *, struct mmu_rb_node *);
  93 static int program_rcvarray(struct file *, unsigned long, struct tid_group *,
  94                             struct tid_pageset *, unsigned, u16, struct page **,
  95                             u32 *, unsigned *, unsigned *);
  96 static int unprogram_rcvarray(struct file *, u32, struct tid_group **);
  97 static void clear_tid_node(struct hfi1_filedata *, u16, struct tid_rb_node *);
  98
  99 static struct mmu_rb_ops tid_rb_ops = {
 100         .insert = mmu_rb_insert,
 101         .remove = mmu_rb_remove,
 102         .invalidate = mmu_rb_invalidate
 103 };
 104
 105 static inline u32 rcventry2tidinfo(u32 rcventry)
 106 {
 107         u32 pair = rcventry & ~0x1;
 108
 109         return EXP_TID_SET(IDX, pair >> 1) |
 110                 EXP_TID_SET(CTRL, 1 << (rcventry - pair));
 111 }
 112
 113 static inline void exp_tid_group_init(struct exp_tid_set *set)
 114 {
 115         INIT_LIST_HEAD(&set->list);
 116         set->count = 0;
 117 }
 118
 119 static inline void tid_group_remove(struct tid_group *grp,
 120                                     struct exp_tid_set *set)
 121 {
 122         list_del_init(&grp->list);
 123         set->count--;
 124 }
 125
 126 static inline void tid_group_add_tail(struct tid_group *grp,
 127                                       struct exp_tid_set *set)
 128 {
 129         list_add_tail(&grp->list, &set->list);
 130         set->count++;
 131 }
 132
 133 static inline struct tid_group *tid_group_pop(struct exp_tid_set *set)
 134 {
 135         struct tid_group *grp =
 136                 list_first_entry(&set->list, struct tid_group, list);
 137         list_del_init(&grp->list);
 138         set->count--;
 139         return grp;
 140 }
 141
 142 static inline void tid_group_move(struct tid_group *group,
 143                                   struct exp_tid_set *s1,
 144                                   struct exp_tid_set *s2)
 145 {
 146         tid_group_remove(group, s1);
 147         tid_group_add_tail(group, s2);
 148 }
 149
 150 /*
 151  * Initialize context and file private data needed for Expected
 152  * receive caching. This needs to be done after the context has
 153  * been configured with the eager/expected RcvEntry counts.
 154  */
 155 int hfi1_user_exp_rcv_init(struct file *fp)
 156 {
 157         struct hfi1_filedata *fd = fp->private_data;
 158         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 159         struct hfi1_devdata *dd = uctxt->dd;
 160         unsigned tidbase;
 161         int i, ret = 0;
 162
 163         spin_lock_init(&fd->tid_lock);
 164         spin_lock_init(&fd->invalid_lock);
 165         fd->tid_rb_root = RB_ROOT;
 166
 167         if (!uctxt->subctxt_cnt || !fd->subctxt) {
 168                 exp_tid_group_init(&uctxt->tid_group_list);
 169                 exp_tid_group_init(&uctxt->tid_used_list);
 170                 exp_tid_group_init(&uctxt->tid_full_list);
 171
 172                 tidbase = uctxt->expected_base;
 173                 for (i = 0; i < uctxt->expected_count /
 174                              dd->rcv_entries.group_size; i++) {
 175                         struct tid_group *grp;
 176
 177                         grp = kzalloc(sizeof(*grp), GFP_KERNEL);
 178                         if (!grp) {
 179                                 /*
 180                                  * If we fail here, the groups already
 181                                  * allocated will be freed by the close
 182                                  * call.
 183                                  */
 184                                 ret = -ENOMEM;
 185                                 goto done;
 186                         }
 187                         grp->size = dd->rcv_entries.group_size;
 188                         grp->base = tidbase;
 189                         tid_group_add_tail(grp, &uctxt->tid_group_list);
 190                         tidbase += dd->rcv_entries.group_size;
 191                 }
 192         }
 193
 194         fd->entry_to_rb = kcalloc(uctxt->expected_count,
 195                                      sizeof(struct rb_node *),
 196                                      GFP_KERNEL);
 197         if (!fd->entry_to_rb)
 198                 return -ENOMEM;
 199
 200         if (!HFI1_CAP_IS_USET(TID_UNMAP)) {
 201                 fd->invalid_tid_idx = 0;
 202                 fd->invalid_tids = kzalloc(uctxt->expected_count *
 203                                            sizeof(u32), GFP_KERNEL);
 204                 if (!fd->invalid_tids) {
 205                         ret = -ENOMEM;
 206                         goto done;
 207                 }
 208
 209                 /*
 210                  * Register MMU notifier callbacks. If the registration
 211                  * fails, continue but turn off the TID caching for
 212                  * all user contexts.
 213                  */
 214                 ret = hfi1_mmu_rb_register(&fd->tid_rb_root, &tid_rb_ops);
 215                 if (ret) {
 216                         dd_dev_info(dd,
 217                                     "Failed MMU notifier registration %d\n",
 218                                     ret);
 219                         HFI1_CAP_USET(TID_UNMAP);
 220                         ret = 0;
 221                 }
 222         }
 223
 224         /*
 225          * PSM does not have a good way to separate, count, and
 226          * effectively enforce a limit on RcvArray entries used by
 227          * subctxts (when context sharing is used) when TID caching
 228          * is enabled. To help with that, we calculate a per-process
 229          * RcvArray entry share and enforce that.
 230          * If TID caching is not in use, PSM deals with usage on its
 231          * own. In that case, we allow any subctxt to take all of the
 232          * entries.
 233          *
 234          * Make sure that we set the tid counts only after successful
 235          * init.
 236          */
 237         spin_lock(&fd->tid_lock);
 238         if (uctxt->subctxt_cnt && !HFI1_CAP_IS_USET(TID_UNMAP)) {
 239                 u16 remainder;
 240
 241                 fd->tid_limit = uctxt->expected_count / uctxt->subctxt_cnt;
 242                 remainder = uctxt->expected_count % uctxt->subctxt_cnt;
 243                 if (remainder && fd->subctxt < remainder)
 244                         fd->tid_limit++;
 245         } else {
 246                 fd->tid_limit = uctxt->expected_count;
 247         }
 248         spin_unlock(&fd->tid_lock);
 249 done:
 250         return ret;
 251 }
 252
 253 int hfi1_user_exp_rcv_free(struct hfi1_filedata *fd)
 254 {
 255         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 256         struct tid_group *grp, *gptr;
 257
 258         if (!test_bit(HFI1_CTXT_SETUP_DONE, &uctxt->event_flags))
 259                 return 0;
 260         /*
 261          * The notifier would have been removed when the process'es mm
 262          * was freed.
 263          */
 264         if (!HFI1_CAP_IS_USET(TID_UNMAP))
 265                 hfi1_mmu_rb_unregister(&fd->tid_rb_root);
 266
 267         kfree(fd->invalid_tids);
 268
 269         if (!uctxt->cnt) {
 270                 if (!EXP_TID_SET_EMPTY(uctxt->tid_full_list))
 271                         unlock_exp_tids(uctxt, &uctxt->tid_full_list,
 272                                         &fd->tid_rb_root);
 273                 if (!EXP_TID_SET_EMPTY(uctxt->tid_used_list))
 274                         unlock_exp_tids(uctxt, &uctxt->tid_used_list,
 275                                         &fd->tid_rb_root);
 276                 list_for_each_entry_safe(grp, gptr, &uctxt->tid_group_list.list,
 277                                          list) {
 278                         list_del_init(&grp->list);
 279                         kfree(grp);
 280                 }
 281                 hfi1_clear_tids(uctxt);
 282         }
 283
 284         kfree(fd->entry_to_rb);
 285         return 0;
 286 }
 287
 288 /*
 289  * Write an "empty" RcvArray entry.
 290  * This function exists so the TID registaration code can use it
 291  * to write to unused/unneeded entries and still take advantage
 292  * of the WC performance improvements. The HFI will ignore this
 293  * write to the RcvArray entry.
 294  */
 295 static inline void rcv_array_wc_fill(struct hfi1_devdata *dd, u32 index)
 296 {
 297         /*
 298          * Doing the WC fill writes only makes sense if the device is
 299          * present and the RcvArray has been mapped as WC memory.
 300          */
 301         if ((dd->flags & HFI1_PRESENT) && dd->rcvarray_wc)
 302                 writeq(0, dd->rcvarray_wc + (index * 8));
 303 }
 304
 305 /*
 306  * RcvArray entry allocation for Expected Receives is done by the
 307  * following algorithm:
 308  *
 309  * The context keeps 3 lists of groups of RcvArray entries:
 310  *   1. List of empty groups - tid_group_list
 311  *      This list is created during user context creation and
 312  *      contains elements which describe sets (of 8) of empty
 313  *      RcvArray entries.
 314  *   2. List of partially used groups - tid_used_list
 315  *      This list contains sets of RcvArray entries which are
 316  *      not completely used up. Another mapping request could
 317  *      use some of all of the remaining entries.
 318  *   3. List of full groups - tid_full_list
 319  *      This is the list where sets that are completely used
 320  *      up go.
 321  *
 322  * An attempt to optimize the usage of RcvArray entries is
 323  * made by finding all sets of physically contiguous pages in a
 324  * user's buffer.
 325  * These physically contiguous sets are further split into
 326  * sizes supported by the receive engine of the HFI. The
 327  * resulting sets of pages are stored in struct tid_pageset,
 328  * which describes the sets as:
 329  *    * .count - number of pages in this set
 330  *    * .idx - starting index into struct page ** array
 331  *                    of this set
 332  *
 333  * From this point on, the algorithm deals with the page sets
 334  * described above. The number of pagesets is divided by the
 335  * RcvArray group size to produce the number of full groups
 336  * needed.
 337  *
 338  * Groups from the 3 lists are manipulated using the following
 339  * rules:
 340  *   1. For each set of 8 pagesets, a complete group from
 341  *      tid_group_list is taken, programmed, and moved to
 342  *      the tid_full_list list.
 343  *   2. For all remaining pagesets:
 344  *      2.1 If the tid_used_list is empty and the tid_group_list
 345  *          is empty, stop processing pageset and return only
 346  *          what has been programmed up to this point.
 347  *      2.2 If the tid_used_list is empty and the tid_group_list
 348  *          is not empty, move a group from tid_group_list to
 349  *          tid_used_list.
 350  *      2.3 For each group is tid_used_group, program as much as
 351  *          can fit into the group. If the group becomes fully
 352  *          used, move it to tid_full_list.
 353  */
 354 int hfi1_user_exp_rcv_setup(struct file *fp, struct hfi1_tid_info *tinfo)
 355 {
 356         int ret = 0, need_group = 0, pinned;
 357         struct hfi1_filedata *fd = fp->private_data;
 358         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 359         struct hfi1_devdata *dd = uctxt->dd;
 360         unsigned npages, ngroups, pageidx = 0, pageset_count, npagesets,
 361                 tididx = 0, mapped, mapped_pages = 0;
 362         unsigned long vaddr = tinfo->vaddr;
 363         struct page **pages = NULL;
 364         u32 *tidlist = NULL;
 365         struct tid_pageset *pagesets = NULL;
 366
 367         /* Get the number of pages the user buffer spans */
 368         npages = num_user_pages(vaddr, tinfo->length);
 369         if (!npages)
 370                 return -EINVAL;
 371
 372         if (npages > uctxt->expected_count) {
 373                 dd_dev_err(dd, "Expected buffer too big\n");
 374                 return -EINVAL;
 375         }
 376
 377         /* Verify that access is OK for the user buffer */
 378         if (!access_ok(VERIFY_WRITE, (void __user *)vaddr,
 379                        npages * PAGE_SIZE)) {
 380                 dd_dev_err(dd, "Fail vaddr %p, %u pages, !access_ok\n",
 381                            (void *)vaddr, npages);
 382                 return -EFAULT;
 383         }
 384
 385         pagesets = kcalloc(uctxt->expected_count, sizeof(*pagesets),
 386                            GFP_KERNEL);
 387         if (!pagesets)
 388                 return -ENOMEM;
 389
 390         /* Allocate the array of struct page pointers needed for pinning */
 391         pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL);
 392         if (!pages) {
 393                 ret = -ENOMEM;
 394                 goto bail;
 395         }
 396
 397         /*
 398          * Pin all the pages of the user buffer. If we can't pin all the
 399          * pages, accept the amount pinned so far and program only that.
 400          * User space knows how to deal with partially programmed buffers.
 401          */
 402         if (!hfi1_can_pin_pages(dd, fd->tid_n_pinned, npages)) {
 403                 ret = -ENOMEM;
 404                 goto bail;
 405         }
 406
 407         pinned = hfi1_acquire_user_pages(vaddr, npages, true, pages);
 408         if (pinned <= 0) {
 409                 ret = pinned;
 410                 goto bail;
 411         }
 412         fd->tid_n_pinned += npages;
 413
 414         /* Find sets of physically contiguous pages */
 415         npagesets = find_phys_blocks(pages, pinned, pagesets);
 416
 417         /*
 418          * We don't need to access this under a lock since tid_used is per
 419          * process and the same process cannot be in hfi1_user_exp_rcv_clear()
 420          * and hfi1_user_exp_rcv_setup() at the same time.
 421          */
 422         spin_lock(&fd->tid_lock);
 423         if (fd->tid_used + npagesets > fd->tid_limit)
 424                 pageset_count = fd->tid_limit - fd->tid_used;
 425         else
 426                 pageset_count = npagesets;
 427         spin_unlock(&fd->tid_lock);
 428
 429         if (!pageset_count)
 430                 goto bail;
 431
 432         ngroups = pageset_count / dd->rcv_entries.group_size;
 433         tidlist = kcalloc(pageset_count, sizeof(*tidlist), GFP_KERNEL);
 434         if (!tidlist) {
 435                 ret = -ENOMEM;
 436                 goto nomem;
 437         }
 438
 439         tididx = 0;
 440
 441         /*
 442          * From this point on, we are going to be using shared (between master
 443          * and subcontexts) context resources. We need to take the lock.
 444          */
 445         mutex_lock(&uctxt->exp_lock);
 446         /*
 447          * The first step is to program the RcvArray entries which are complete
 448          * groups.
 449          */
 450         while (ngroups && uctxt->tid_group_list.count) {
 451                 struct tid_group *grp =
 452                         tid_group_pop(&uctxt->tid_group_list);
 453
 454                 ret = program_rcvarray(fp, vaddr, grp, pagesets,
 455                                        pageidx, dd->rcv_entries.group_size,
 456                                        pages, tidlist, &tididx, &mapped);
 457                 /*
 458                  * If there was a failure to program the RcvArray
 459                  * entries for the entire group, reset the grp fields
 460                  * and add the grp back to the free group list.
 461                  */
 462                 if (ret <= 0) {
 463                         tid_group_add_tail(grp, &uctxt->tid_group_list);
 464                         hfi1_cdbg(TID,
 465                                   "Failed to program RcvArray group %d", ret);
 466                         goto unlock;
 467                 }
 468
 469                 tid_group_add_tail(grp, &uctxt->tid_full_list);
 470                 ngroups--;
 471                 pageidx += ret;
 472                 mapped_pages += mapped;
 473         }
 474
 475         while (pageidx < pageset_count) {
 476                 struct tid_group *grp, *ptr;
 477                 /*
 478                  * If we don't have any partially used tid groups, check
 479                  * if we have empty groups. If so, take one from there and
 480                  * put in the partially used list.
 481                  */
 482                 if (!uctxt->tid_used_list.count || need_group) {
 483                         if (!uctxt->tid_group_list.count)
 484                                 goto unlock;
 485
 486                         grp = tid_group_pop(&uctxt->tid_group_list);
 487                         tid_group_add_tail(grp, &uctxt->tid_used_list);
 488                         need_group = 0;
 489                 }
 490                 /*
 491                  * There is an optimization opportunity here - instead of
 492                  * fitting as many page sets as we can, check for a group
 493                  * later on in the list that could fit all of them.
 494                  */
 495                 list_for_each_entry_safe(grp, ptr, &uctxt->tid_used_list.list,
 496                                          list) {
 497                         unsigned use = min_t(unsigned, pageset_count - pageidx,
 498                                              grp->size - grp->used);
 499
 500                         ret = program_rcvarray(fp, vaddr, grp, pagesets,
 501                                                pageidx, use, pages, tidlist,
 502                                                &tididx, &mapped);
 503                         if (ret < 0) {
 504                                 hfi1_cdbg(TID,
 505                                           "Failed to program RcvArray entries %d",
 506                                           ret);
 507                                 ret = -EFAULT;
 508                                 goto unlock;
 509                         } else if (ret > 0) {
 510                                 if (grp->used == grp->size)
 511                                         tid_group_move(grp,
 512                                                        &uctxt->tid_used_list,
 513                                                        &uctxt->tid_full_list);
 514                                 pageidx += ret;
 515                                 mapped_pages += mapped;
 516                                 need_group = 0;
 517                                 /* Check if we are done so we break out early */
 518                                 if (pageidx >= pageset_count)
 519                                         break;
 520                         } else if (WARN_ON(ret == 0)) {
 521                                 /*
 522                                  * If ret is 0, we did not program any entries
 523                                  * into this group, which can only happen if
 524                                  * we've screwed up the accounting somewhere.
 525                                  * Warn and try to continue.
 526                                  */
 527                                 need_group = 1;
 528                         }
 529                 }
 530         }
 531 unlock:
 532         mutex_unlock(&uctxt->exp_lock);
 533 nomem:
 534         hfi1_cdbg(TID, "total mapped: tidpairs:%u pages:%u (%d)", tididx,
 535                   mapped_pages, ret);
 536         if (tididx) {
 537                 spin_lock(&fd->tid_lock);
 538                 fd->tid_used += tididx;
 539                 spin_unlock(&fd->tid_lock);
 540                 tinfo->tidcnt = tididx;
 541                 tinfo->length = mapped_pages * PAGE_SIZE;
 542
 543                 if (copy_to_user((void __user *)(unsigned long)tinfo->tidlist,
 544                                  tidlist, sizeof(tidlist[0]) * tididx)) {
 545                         /*
 546                          * On failure to copy to the user level, we need to undo
 547                          * everything done so far so we don't leak resources.
 548                          */
 549                         tinfo->tidlist = (unsigned long)&tidlist;
 550                         hfi1_user_exp_rcv_clear(fp, tinfo);
 551                         tinfo->tidlist = 0;
 552                         ret = -EFAULT;
 553                         goto bail;
 554                 }
 555         }
 556
 557         /*
 558          * If not everything was mapped (due to insufficient RcvArray entries,
 559          * for example), unpin all unmapped pages so we can pin them nex time.
 560          */
 561         if (mapped_pages != pinned) {
 562                 hfi1_release_user_pages(current->mm, &pages[mapped_pages],
 563                                         pinned - mapped_pages,
 564                                         false);
 565                 fd->tid_n_pinned -= pinned - mapped_pages;
 566         }
 567 bail:
 568         kfree(pagesets);
 569         kfree(pages);
 570         kfree(tidlist);
 571         return ret > 0 ? 0 : ret;
 572 }
 573
 574 int hfi1_user_exp_rcv_clear(struct file *fp, struct hfi1_tid_info *tinfo)
 575 {
 576         int ret = 0;
 577         struct hfi1_filedata *fd = fp->private_data;
 578         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 579         u32 *tidinfo;
 580         unsigned tididx;
 581
 582         tidinfo = kcalloc(tinfo->tidcnt, sizeof(*tidinfo), GFP_KERNEL);
 583         if (!tidinfo)
 584                 return -ENOMEM;
 585
 586         if (copy_from_user(tidinfo, (void __user *)(unsigned long)
 587                            tinfo->tidlist, sizeof(tidinfo[0]) *
 588                            tinfo->tidcnt)) {
 589                 ret = -EFAULT;
 590                 goto done;
 591         }
 592
 593         mutex_lock(&uctxt->exp_lock);
 594         for (tididx = 0; tididx < tinfo->tidcnt; tididx++) {
 595                 ret = unprogram_rcvarray(fp, tidinfo[tididx], NULL);
 596                 if (ret) {
 597                         hfi1_cdbg(TID, "Failed to unprogram rcv array %d",
 598                                   ret);
 599                         break;
 600                 }
 601         }
 602         spin_lock(&fd->tid_lock);
 603         fd->tid_used -= tididx;
 604         spin_unlock(&fd->tid_lock);
 605         tinfo->tidcnt = tididx;
 606         mutex_unlock(&uctxt->exp_lock);
 607 done:
 608         kfree(tidinfo);
 609         return ret;
 610 }
 611
 612 int hfi1_user_exp_rcv_invalid(struct file *fp, struct hfi1_tid_info *tinfo)
 613 {
 614         struct hfi1_filedata *fd = fp->private_data;
 615         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 616         unsigned long *ev = uctxt->dd->events +
 617                 (((uctxt->ctxt - uctxt->dd->first_user_ctxt) *
 618                   HFI1_MAX_SHARED_CTXTS) + fd->subctxt);
 619         u32 *array;
 620         int ret = 0;
 621
 622         if (!fd->invalid_tids)
 623                 return -EINVAL;
 624
 625         /*
 626          * copy_to_user() can sleep, which will leave the invalid_lock
 627          * locked and cause the MMU notifier to be blocked on the lock
 628          * for a long time.
 629          * Copy the data to a local buffer so we can release the lock.
 630          */
 631         array = kcalloc(uctxt->expected_count, sizeof(*array), GFP_KERNEL);
 632         if (!array)
 633                 return -EFAULT;
 634
 635         spin_lock(&fd->invalid_lock);
 636         if (fd->invalid_tid_idx) {
 637                 memcpy(array, fd->invalid_tids, sizeof(*array) *
 638                        fd->invalid_tid_idx);
 639                 memset(fd->invalid_tids, 0, sizeof(*fd->invalid_tids) *
 640                        fd->invalid_tid_idx);
 641                 tinfo->tidcnt = fd->invalid_tid_idx;
 642                 fd->invalid_tid_idx = 0;
 643                 /*
 644                  * Reset the user flag while still holding the lock.
 645                  * Otherwise, PSM can miss events.
 646                  */
 647                 clear_bit(_HFI1_EVENT_TID_MMU_NOTIFY_BIT, ev);
 648         } else {
 649                 tinfo->tidcnt = 0;
 650         }
 651         spin_unlock(&fd->invalid_lock);
 652
 653         if (tinfo->tidcnt) {
 654                 if (copy_to_user((void __user *)tinfo->tidlist,
 655                                  array, sizeof(*array) * tinfo->tidcnt))
 656                         ret = -EFAULT;
 657         }
 658         kfree(array);
 659
 660         return ret;
 661 }
 662
 663 static u32 find_phys_blocks(struct page **pages, unsigned npages,
 664                             struct tid_pageset *list)
 665 {
 666         unsigned pagecount, pageidx, setcount = 0, i;
 667         unsigned long pfn, this_pfn;
 668
 669         if (!npages)
 670                 return 0;
 671
 672         /*
 673          * Look for sets of physically contiguous pages in the user buffer.
 674          * This will allow us to optimize Expected RcvArray entry usage by
 675          * using the bigger supported sizes.
 676          */
 677         pfn = page_to_pfn(pages[0]);
 678         for (pageidx = 0, pagecount = 1, i = 1; i <= npages; i++) {
 679                 this_pfn = i < npages ? page_to_pfn(pages[i]) : 0;
 680
 681                 /*
 682                  * If the pfn's are not sequential, pages are not physically
 683                  * contiguous.
 684                  */
 685                 if (this_pfn != ++pfn) {
 686                         /*
 687                          * At this point we have to loop over the set of
 688                          * physically contiguous pages and break them down it
 689                          * sizes supported by the HW.
 690                          * There are two main constraints:
 691                          *     1. The max buffer size is MAX_EXPECTED_BUFFER.
 692                          *        If the total set size is bigger than that
 693                          *        program only a MAX_EXPECTED_BUFFER chunk.
 694                          *     2. The buffer size has to be a power of two. If
 695                          *        it is not, round down to the closes power of
 696                          *        2 and program that size.
 697                          */
 698                         while (pagecount) {
 699                                 int maxpages = pagecount;
 700                                 u32 bufsize = pagecount * PAGE_SIZE;
 701
 702                                 if (bufsize > MAX_EXPECTED_BUFFER)
 703                                         maxpages =
 704                                                 MAX_EXPECTED_BUFFER >>
 705                                                 PAGE_SHIFT;
 706                                 else if (!is_power_of_2(bufsize))
 707                                         maxpages =
 708                                                 rounddown_pow_of_two(bufsize) >>
 709                                                 PAGE_SHIFT;
 710
 711                                 list[setcount].idx = pageidx;
 712                                 list[setcount].count = maxpages;
 713                                 pagecount -= maxpages;
 714                                 pageidx += maxpages;
 715                                 setcount++;
 716                         }
 717                         pageidx = i;
 718                         pagecount = 1;
 719                         pfn = this_pfn;
 720                 } else {
 721                         pagecount++;
 722                 }
 723         }
 724         return setcount;
 725 }
 726
 727 /**
 728  * program_rcvarray() - program an RcvArray group with receive buffers
 729  * @fp: file pointer
 730  * @vaddr: starting user virtual address
 731  * @grp: RcvArray group
 732  * @sets: array of struct tid_pageset holding information on physically
 733  *        contiguous chunks from the user buffer
 734  * @start: starting index into sets array
 735  * @count: number of struct tid_pageset's to program
 736  * @pages: an array of struct page * for the user buffer
 737  * @tidlist: the array of u32 elements when the information about the
 738  *           programmed RcvArray entries is to be encoded.
 739  * @tididx: starting offset into tidlist
 740  * @pmapped: (output parameter) number of pages programmed into the RcvArray
 741  *           entries.
 742  *
 743  * This function will program up to 'count' number of RcvArray entries from the
 744  * group 'grp'. To make best use of write-combining writes, the function will
 745  * perform writes to the unused RcvArray entries which will be ignored by the
 746  * HW. Each RcvArray entry will be programmed with a physically contiguous
 747  * buffer chunk from the user's virtual buffer.
 748  *
 749  * Return:
 750  * -EINVAL if the requested count is larger than the size of the group,
 751  * -ENOMEM or -EFAULT on error from set_rcvarray_entry(), or
 752  * number of RcvArray entries programmed.
 753  */
 754 static int program_rcvarray(struct file *fp, unsigned long vaddr,
 755                             struct tid_group *grp,
 756                             struct tid_pageset *sets,
 757                             unsigned start, u16 count, struct page **pages,
 758                             u32 *tidlist, unsigned *tididx, unsigned *pmapped)
 759 {
 760         struct hfi1_filedata *fd = fp->private_data;
 761         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 762         struct hfi1_devdata *dd = uctxt->dd;
 763         u16 idx;
 764         u32 tidinfo = 0, rcventry, useidx = 0;
 765         int mapped = 0;
 766
 767         /* Count should never be larger than the group size */
 768         if (count > grp->size)
 769                 return -EINVAL;
 770
 771         /* Find the first unused entry in the group */
 772         for (idx = 0; idx < grp->size; idx++) {
 773                 if (!(grp->map & (1 << idx))) {
 774                         useidx = idx;
 775                         break;
 776                 }
 777                 rcv_array_wc_fill(dd, grp->base + idx);
 778         }
 779
 780         idx = 0;
 781         while (idx < count) {
 782                 u16 npages, pageidx, setidx = start + idx;
 783                 int ret = 0;
 784
 785                 /*
 786                  * If this entry in the group is used, move to the next one.
 787                  * If we go past the end of the group, exit the loop.
 788                  */
 789                 if (useidx >= grp->size) {
 790                         break;
 791                 } else if (grp->map & (1 << useidx)) {
 792                         rcv_array_wc_fill(dd, grp->base + useidx);
 793                         useidx++;
 794                         continue;
 795                 }
 796
 797                 rcventry = grp->base + useidx;
 798                 npages = sets[setidx].count;
 799                 pageidx = sets[setidx].idx;
 800
 801                 ret = set_rcvarray_entry(fp, vaddr + (pageidx * PAGE_SIZE),
 802                                          rcventry, grp, pages + pageidx,
 803                                          npages);
 804                 if (ret)
 805                         return ret;
 806                 mapped += npages;
 807
 808                 tidinfo = rcventry2tidinfo(rcventry - uctxt->expected_base) |
 809                         EXP_TID_SET(LEN, npages);
 810                 tidlist[(*tididx)++] = tidinfo;
 811                 grp->used++;
 812                 grp->map |= 1 << useidx++;
 813                 idx++;
 814         }
 815
 816         /* Fill the rest of the group with "blank" writes */
 817         for (; useidx < grp->size; useidx++)
 818                 rcv_array_wc_fill(dd, grp->base + useidx);
 819         *pmapped = mapped;
 820         return idx;
 821 }
 822
 823 static int set_rcvarray_entry(struct file *fp, unsigned long vaddr,
 824                               u32 rcventry, struct tid_group *grp,
 825                               struct page **pages, unsigned npages)
 826 {
 827         int ret;
 828         struct hfi1_filedata *fd = fp->private_data;
 829         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 830         struct tid_rb_node *node;
 831         struct hfi1_devdata *dd = uctxt->dd;
 832         struct rb_root *root = &fd->tid_rb_root;
 833         dma_addr_t phys;
 834
 835         /*
 836          * Allocate the node first so we can handle a potential
 837          * failure before we've programmed anything.
 838          */
 839         node = kzalloc(sizeof(*node) + (sizeof(struct page *) * npages),
 840                        GFP_KERNEL);
 841         if (!node)
 842                 return -ENOMEM;
 843
 844         phys = pci_map_single(dd->pcidev,
 845                               __va(page_to_phys(pages[0])),
 846                               npages * PAGE_SIZE, PCI_DMA_FROMDEVICE);
 847         if (dma_mapping_error(&dd->pcidev->dev, phys)) {
 848                 dd_dev_err(dd, "Failed to DMA map Exp Rcv pages 0x%llx\n",
 849                            phys);
 850                 kfree(node);
 851                 return -EFAULT;
 852         }
 853
 854         node->mmu.addr = vaddr;
 855         node->mmu.len = npages * PAGE_SIZE;
 856         node->phys = page_to_phys(pages[0]);
 857         node->npages = npages;
 858         node->rcventry = rcventry;
 859         node->dma_addr = phys;
 860         node->grp = grp;
 861         node->freed = false;
 862         memcpy(node->pages, pages, sizeof(struct page *) * npages);
 863
 864         if (HFI1_CAP_IS_USET(TID_UNMAP))
 865                 ret = mmu_rb_insert(root, &node->mmu);
 866         else
 867                 ret = hfi1_mmu_rb_insert(root, &node->mmu);
 868
 869         if (ret) {
 870                 hfi1_cdbg(TID, "Failed to insert RB node %u 0x%lx, 0x%lx %d",
 871                           node->rcventry, node->mmu.addr, node->phys, ret);
 872                 pci_unmap_single(dd->pcidev, phys, npages * PAGE_SIZE,
 873                                  PCI_DMA_FROMDEVICE);
 874                 kfree(node);
 875                 return -EFAULT;
 876         }
 877         hfi1_put_tid(dd, rcventry, PT_EXPECTED, phys, ilog2(npages) + 1);
 878         trace_hfi1_exp_tid_reg(uctxt->ctxt, fd->subctxt, rcventry, npages,
 879                                node->mmu.addr, node->phys, phys);
 880         return 0;
 881 }
 882
 883 static int unprogram_rcvarray(struct file *fp, u32 tidinfo,
 884                               struct tid_group **grp)
 885 {
 886         struct hfi1_filedata *fd = fp->private_data;
 887         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 888         struct hfi1_devdata *dd = uctxt->dd;
 889         struct tid_rb_node *node;
 890         u8 tidctrl = EXP_TID_GET(tidinfo, CTRL);
 891         u32 tididx = EXP_TID_GET(tidinfo, IDX) << 1, rcventry;
 892
 893         if (tididx >= uctxt->expected_count) {
 894                 dd_dev_err(dd, "Invalid RcvArray entry (%u) index for ctxt %u\n",
 895                            tididx, uctxt->ctxt);
 896                 return -EINVAL;
 897         }
 898
 899         if (tidctrl == 0x3)
 900                 return -EINVAL;
 901
 902         rcventry = tididx + (tidctrl - 1);
 903
 904         node = fd->entry_to_rb[rcventry];
 905         if (!node || node->rcventry != (uctxt->expected_base + rcventry))
 906                 return -EBADF;
 907         if (HFI1_CAP_IS_USET(TID_UNMAP))
 908                 mmu_rb_remove(&fd->tid_rb_root, &node->mmu, NULL);
 909         else
 910                 hfi1_mmu_rb_remove(&fd->tid_rb_root, &node->mmu);
 911
 912         if (grp)
 913                 *grp = node->grp;
 914         clear_tid_node(fd, fd->subctxt, node);
 915         return 0;
 916 }
 917
 918 static void clear_tid_node(struct hfi1_filedata *fd, u16 subctxt,
 919                            struct tid_rb_node *node)
 920 {
 921         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 922         struct hfi1_devdata *dd = uctxt->dd;
 923
 924         trace_hfi1_exp_tid_unreg(uctxt->ctxt, fd->subctxt, node->rcventry,
 925                                  node->npages, node->mmu.addr, node->phys,
 926                                  node->dma_addr);
 927
 928         hfi1_put_tid(dd, node->rcventry, PT_INVALID, 0, 0);
 929         /*
 930          * Make sure device has seen the write before we unpin the
 931          * pages.
 932          */
 933         flush_wc();
 934
 935         pci_unmap_single(dd->pcidev, node->dma_addr, node->mmu.len,
 936                          PCI_DMA_FROMDEVICE);
 937         hfi1_release_user_pages(current->mm, node->pages, node->npages, true);
 938         fd->tid_n_pinned -= node->npages;
 939
 940         node->grp->used--;
 941         node->grp->map &= ~(1 << (node->rcventry - node->grp->base));
 942
 943         if (node->grp->used == node->grp->size - 1)
 944                 tid_group_move(node->grp, &uctxt->tid_full_list,
 945                                &uctxt->tid_used_list);
 946         else if (!node->grp->used)
 947                 tid_group_move(node->grp, &uctxt->tid_used_list,
 948                                &uctxt->tid_group_list);
 949         kfree(node);
 950 }
 951
 952 static void unlock_exp_tids(struct hfi1_ctxtdata *uctxt,
 953                             struct exp_tid_set *set, struct rb_root *root)
 954 {
 955         struct tid_group *grp, *ptr;
 956         struct hfi1_filedata *fd = container_of(root, struct hfi1_filedata,
 957                                                 tid_rb_root);
 958         int i;
 959
 960         list_for_each_entry_safe(grp, ptr, &set->list, list) {
 961                 list_del_init(&grp->list);
 962
 963                 for (i = 0; i < grp->size; i++) {
 964                         if (grp->map & (1 << i)) {
 965                                 u16 rcventry = grp->base + i;
 966                                 struct tid_rb_node *node;
 967
 968                                 node = fd->entry_to_rb[rcventry -
 969                                                           uctxt->expected_base];
 970                                 if (!node || node->rcventry != rcventry)
 971                                         continue;
 972                                 if (HFI1_CAP_IS_USET(TID_UNMAP))
 973                                         mmu_rb_remove(&fd->tid_rb_root,
 974                                                       &node->mmu, NULL);
 975                                 else
 976                                         hfi1_mmu_rb_remove(&fd->tid_rb_root,
 977                                                            &node->mmu);
 978                                 clear_tid_node(fd, -1, node);
 979                         }
 980                 }
 981         }
 982 }
 983
 984 static int mmu_rb_invalidate(struct rb_root *root, struct mmu_rb_node *mnode)
 985 {
 986         struct hfi1_filedata *fdata =
 987                 container_of(root, struct hfi1_filedata, tid_rb_root);
 988         struct hfi1_ctxtdata *uctxt = fdata->uctxt;
 989         struct tid_rb_node *node =
 990                 container_of(mnode, struct tid_rb_node, mmu);
 991
 992         if (node->freed)
 993                 return 0;
 994
 995         trace_hfi1_exp_tid_inval(uctxt->ctxt, fdata->subctxt, node->mmu.addr,
 996                                  node->rcventry, node->npages, node->dma_addr);
 997         node->freed = true;
 998
 999         spin_lock(&fdata->invalid_lock);
1000         if (fdata->invalid_tid_idx < uctxt->expected_count) {
1001                 fdata->invalid_tids[fdata->invalid_tid_idx] =
1002                         rcventry2tidinfo(node->rcventry - uctxt->expected_base);
1003                 fdata->invalid_tids[fdata->invalid_tid_idx] |=
1004                         EXP_TID_SET(LEN, node->npages);
1005                 if (!fdata->invalid_tid_idx) {
1006                         unsigned long *ev;
1007
1008                         /*
1009                          * hfi1_set_uevent_bits() sets a user event flag
1010                          * for all processes. Because calling into the
1011                          * driver to process TID cache invalidations is
1012                          * expensive and TID cache invalidations are
1013                          * handled on a per-process basis, we can
1014                          * optimize this to set the flag only for the
1015                          * process in question.
1016                          */
1017                         ev = uctxt->dd->events +
1018                                 (((uctxt->ctxt - uctxt->dd->first_user_ctxt) *
1019                                   HFI1_MAX_SHARED_CTXTS) + fdata->subctxt);
1020                         set_bit(_HFI1_EVENT_TID_MMU_NOTIFY_BIT, ev);
1021                 }
1022                 fdata->invalid_tid_idx++;
1023         }
1024         spin_unlock(&fdata->invalid_lock);
1025         return 0;
1026 }
1027
1028 static int mmu_rb_insert(struct rb_root *root, struct mmu_rb_node *node)
1029 {
1030         struct hfi1_filedata *fdata =
1031                 container_of(root, struct hfi1_filedata, tid_rb_root);
1032         struct tid_rb_node *tnode =
1033                 container_of(node, struct tid_rb_node, mmu);
1034         u32 base = fdata->uctxt->expected_base;
1035
1036         fdata->entry_to_rb[tnode->rcventry - base] = tnode;
1037         return 0;
1038 }
1039
1040 static void mmu_rb_remove(struct rb_root *root, struct mmu_rb_node *node,
1041                           struct mm_struct *mm)
1042 {
1043         struct hfi1_filedata *fdata =
1044                 container_of(root, struct hfi1_filedata, tid_rb_root);
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] = NULL;
1050 }