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