drivers/infiniband/hw/hfi1/user_exp_rcv.c

   1 /*
   2  * Copyright(c) 2015-2017 Intel Corporation.
   3  *
   4  * This file is provided under a dual BSD/GPLv2 license.  When using or
   5  * redistributing this file, you may do so under either license.
   6  *
   7  * GPL LICENSE SUMMARY
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of version 2 of the GNU General Public License as
  11  * published by the Free Software Foundation.
  12  *
  13  * This program is distributed in the hope that it will be useful, but
  14  * WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * General Public License for more details.
  17  *
  18  * BSD LICENSE
  19  *
  20  * Redistribution and use in source and binary forms, with or without
  21  * modification, are permitted provided that the following conditions
  22  * are met:
  23  *
  24  *  - Redistributions of source code must retain the above copyright
  25  *    notice, this list of conditions and the following disclaimer.
  26  *  - Redistributions in binary form must reproduce the above copyright
  27  *    notice, this list of conditions and the following disclaimer in
  28  *    the documentation and/or other materials provided with the
  29  *    distribution.
  30  *  - Neither the name of Intel Corporation nor the names of its
  31  *    contributors may be used to endorse or promote products derived
  32  *    from this software without specific prior written permission.
  33  *
  34  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  35  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  36  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  37  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  38  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  39  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  40  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  41  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  42  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  43  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  44  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  45  *
  46  */
  47 #include <asm/page.h>
  48 #include <linux/string.h>
  49
  50 #include "user_exp_rcv.h"
  51 #include "trace.h"
  52 #include "mmu_rb.h"
  53
  54 struct tid_group {
  55         struct list_head list;
  56         unsigned base;
  57         u8 size;
  58         u8 used;
  59         u8 map;
  60 };
  61
  62 struct tid_rb_node {
  63         struct mmu_rb_node mmu;
  64         unsigned long phys;
  65         struct tid_group *grp;
  66         u32 rcventry;
  67         dma_addr_t dma_addr;
  68         bool freed;
  69         unsigned npages;
  70         struct page *pages[0];
  71 };
  72
  73 struct tid_pageset {
  74         u16 idx;
  75         u16 count;
  76 };
  77
  78 #define EXP_TID_SET_EMPTY(set) (set.count == 0 && list_empty(&set.list))
  79
  80 #define num_user_pages(vaddr, len)                                     \
  81         (1 + (((((unsigned long)(vaddr) +                              \
  82                  (unsigned long)(len) - 1) & PAGE_MASK) -              \
  83                ((unsigned long)vaddr & PAGE_MASK)) >> PAGE_SHIFT))
  84
  85 static void unlock_exp_tids(struct hfi1_ctxtdata *uctxt,
  86                             struct exp_tid_set *set,
  87                             struct hfi1_filedata *fd);
  88 static u32 find_phys_blocks(struct page **pages, unsigned npages,
  89                             struct tid_pageset *list);
  90 static int set_rcvarray_entry(struct hfi1_filedata *fd, unsigned long vaddr,
  91                               u32 rcventry, struct tid_group *grp,
  92                               struct page **pages, unsigned npages);
  93 static int tid_rb_insert(void *arg, struct mmu_rb_node *node);
  94 static void cacheless_tid_rb_remove(struct hfi1_filedata *fdata,
  95                                     struct tid_rb_node *tnode);
  96 static void tid_rb_remove(void *arg, struct mmu_rb_node *node);
  97 static int tid_rb_invalidate(void *arg, struct mmu_rb_node *mnode);
  98 static int program_rcvarray(struct hfi1_filedata *fd, unsigned long vaddr,
  99                             struct tid_group *grp, struct tid_pageset *sets,
 100                             unsigned start, u16 count, struct page **pages,
 101                             u32 *tidlist, unsigned *tididx, unsigned *pmapped);
 102 static int unprogram_rcvarray(struct hfi1_filedata *fd, u32 tidinfo,
 103                               struct tid_group **grp);
 104 static void clear_tid_node(struct hfi1_filedata *fd, struct tid_rb_node *node);
 105
 106 static struct mmu_rb_ops tid_rb_ops = {
 107         .insert = tid_rb_insert,
 108         .remove = tid_rb_remove,
 109         .invalidate = tid_rb_invalidate
 110 };
 111
 112 static inline u32 rcventry2tidinfo(u32 rcventry)
 113 {
 114         u32 pair = rcventry & ~0x1;
 115
 116         return EXP_TID_SET(IDX, pair >> 1) |
 117                 EXP_TID_SET(CTRL, 1 << (rcventry - pair));
 118 }
 119
 120 static inline void exp_tid_group_init(struct exp_tid_set *set)
 121 {
 122         INIT_LIST_HEAD(&set->list);
 123         set->count = 0;
 124 }
 125
 126 static inline void tid_group_remove(struct tid_group *grp,
 127                                     struct exp_tid_set *set)
 128 {
 129         list_del_init(&grp->list);
 130         set->count--;
 131 }
 132
 133 static inline void tid_group_add_tail(struct tid_group *grp,
 134                                       struct exp_tid_set *set)
 135 {
 136         list_add_tail(&grp->list, &set->list);
 137         set->count++;
 138 }
 139
 140 static inline struct tid_group *tid_group_pop(struct exp_tid_set *set)
 141 {
 142         struct tid_group *grp =
 143                 list_first_entry(&set->list, struct tid_group, list);
 144         list_del_init(&grp->list);
 145         set->count--;
 146         return grp;
 147 }
 148
 149 static inline void tid_group_move(struct tid_group *group,
 150                                   struct exp_tid_set *s1,
 151                                   struct exp_tid_set *s2)
 152 {
 153         tid_group_remove(group, s1);
 154         tid_group_add_tail(group, s2);
 155 }
 156
 157 /*
 158  * Initialize context and file private data needed for Expected
 159  * receive caching. This needs to be done after the context has
 160  * been configured with the eager/expected RcvEntry counts.
 161  */
 162 int hfi1_user_exp_rcv_init(struct hfi1_filedata *fd)
 163 {
 164         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 165         struct hfi1_devdata *dd = uctxt->dd;
 166         unsigned tidbase;
 167         int i, ret = 0;
 168
 169         spin_lock_init(&fd->tid_lock);
 170         spin_lock_init(&fd->invalid_lock);
 171
 172         if (!uctxt->subctxt_cnt || !fd->subctxt) {
 173                 exp_tid_group_init(&uctxt->tid_group_list);
 174                 exp_tid_group_init(&uctxt->tid_used_list);
 175                 exp_tid_group_init(&uctxt->tid_full_list);
 176
 177                 tidbase = uctxt->expected_base;
 178                 for (i = 0; i < uctxt->expected_count /
 179                              dd->rcv_entries.group_size; i++) {
 180                         struct tid_group *grp;
 181
 182                         grp = kzalloc(sizeof(*grp), GFP_KERNEL);
 183                         if (!grp) {
 184                                 /*
 185                                  * If we fail here, the groups already
 186                                  * allocated will be freed by the close
 187                                  * call.
 188                                  */
 189                                 ret = -ENOMEM;
 190                                 goto done;
 191                         }
 192                         grp->size = dd->rcv_entries.group_size;
 193                         grp->base = tidbase;
 194                         tid_group_add_tail(grp, &uctxt->tid_group_list);
 195                         tidbase += dd->rcv_entries.group_size;
 196                 }
 197         }
 198
 199         fd->entry_to_rb = kcalloc(uctxt->expected_count,
 200                                      sizeof(struct rb_node *),
 201                                      GFP_KERNEL);
 202         if (!fd->entry_to_rb)
 203                 return -ENOMEM;
 204
 205         if (!HFI1_CAP_UGET_MASK(uctxt->flags, TID_UNMAP)) {
 206                 fd->invalid_tid_idx = 0;
 207                 fd->invalid_tids = kcalloc(uctxt->expected_count,
 208                                            sizeof(*fd->invalid_tids),
 209                                            GFP_KERNEL);
 210                 if (!fd->invalid_tids) {
 211                         ret = -ENOMEM;
 212                         goto done;
 213                 }
 214
 215                 /*
 216                  * Register MMU notifier callbacks. If the registration
 217                  * fails, continue without TID caching for this context.
 218                  */
 219                 ret = hfi1_mmu_rb_register(fd, fd->mm, &tid_rb_ops,
 220                                            dd->pport->hfi1_wq,
 221                                            &fd->handler);
 222                 if (ret) {
 223                         dd_dev_info(dd,
 224                                     "Failed MMU notifier registration %d\n",
 225                                     ret);
 226                         ret = 0;
 227                 }
 228         }
 229
 230         /*
 231          * PSM does not have a good way to separate, count, and
 232          * effectively enforce a limit on RcvArray entries used by
 233          * subctxts (when context sharing is used) when TID caching
 234          * is enabled. To help with that, we calculate a per-process
 235          * RcvArray entry share and enforce that.
 236          * If TID caching is not in use, PSM deals with usage on its
 237          * own. In that case, we allow any subctxt to take all of the
 238          * entries.
 239          *
 240          * Make sure that we set the tid counts only after successful
 241          * init.
 242          */
 243         spin_lock(&fd->tid_lock);
 244         if (uctxt->subctxt_cnt && fd->handler) {
 245                 u16 remainder;
 246
 247                 fd->tid_limit = uctxt->expected_count / uctxt->subctxt_cnt;
 248                 remainder = uctxt->expected_count % uctxt->subctxt_cnt;
 249                 if (remainder && fd->subctxt < remainder)
 250                         fd->tid_limit++;
 251         } else {
 252                 fd->tid_limit = uctxt->expected_count;
 253         }
 254         spin_unlock(&fd->tid_lock);
 255 done:
 256         return ret;
 257 }
 258
 259 void hfi1_user_exp_rcv_grp_free(struct hfi1_ctxtdata *uctxt)
 260 {
 261         struct tid_group *grp, *gptr;
 262
 263         list_for_each_entry_safe(grp, gptr, &uctxt->tid_group_list.list,
 264                                  list) {
 265                 list_del_init(&grp->list);
 266                 kfree(grp);
 267         }
 268         hfi1_clear_tids(uctxt);
 269 }
 270
 271 int hfi1_user_exp_rcv_free(struct hfi1_filedata *fd)
 272 {
 273         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 274
 275         /*
 276          * The notifier would have been removed when the process'es mm
 277          * was freed.
 278          */
 279         if (fd->handler) {
 280                 hfi1_mmu_rb_unregister(fd->handler);
 281         } else {
 282                 if (!EXP_TID_SET_EMPTY(uctxt->tid_full_list))
 283                         unlock_exp_tids(uctxt, &uctxt->tid_full_list, fd);
 284                 if (!EXP_TID_SET_EMPTY(uctxt->tid_used_list))
 285                         unlock_exp_tids(uctxt, &uctxt->tid_used_list, fd);
 286         }
 287
 288         kfree(fd->invalid_tids);
 289         fd->invalid_tids = NULL;
 290
 291         kfree(fd->entry_to_rb);
 292         fd->entry_to_rb = NULL;
 293         return 0;
 294 }
 295
 296 /*
 297  * Write an "empty" RcvArray entry.
 298  * This function exists so the TID registaration code can use it
 299  * to write to unused/unneeded entries and still take advantage
 300  * of the WC performance improvements. The HFI will ignore this
 301  * write to the RcvArray entry.
 302  */
 303 static inline void rcv_array_wc_fill(struct hfi1_devdata *dd, u32 index)
 304 {
 305         /*
 306          * Doing the WC fill writes only makes sense if the device is
 307          * present and the RcvArray has been mapped as WC memory.
 308          */
 309         if ((dd->flags & HFI1_PRESENT) && dd->rcvarray_wc)
 310                 writeq(0, dd->rcvarray_wc + (index * 8));
 311 }
 312
 313 /*
 314  * RcvArray entry allocation for Expected Receives is done by the
 315  * following algorithm:
 316  *
 317  * The context keeps 3 lists of groups of RcvArray entries:
 318  *   1. List of empty groups - tid_group_list
 319  *      This list is created during user context creation and
 320  *      contains elements which describe sets (of 8) of empty
 321  *      RcvArray entries.
 322  *   2. List of partially used groups - tid_used_list
 323  *      This list contains sets of RcvArray entries which are
 324  *      not completely used up. Another mapping request could
 325  *      use some of all of the remaining entries.
 326  *   3. List of full groups - tid_full_list
 327  *      This is the list where sets that are completely used
 328  *      up go.
 329  *
 330  * An attempt to optimize the usage of RcvArray entries is
 331  * made by finding all sets of physically contiguous pages in a
 332  * user's buffer.
 333  * These physically contiguous sets are further split into
 334  * sizes supported by the receive engine of the HFI. The
 335  * resulting sets of pages are stored in struct tid_pageset,
 336  * which describes the sets as:
 337  *    * .count - number of pages in this set
 338  *    * .idx - starting index into struct page ** array
 339  *                    of this set
 340  *
 341  * From this point on, the algorithm deals with the page sets
 342  * described above. The number of pagesets is divided by the
 343  * RcvArray group size to produce the number of full groups
 344  * needed.
 345  *
 346  * Groups from the 3 lists are manipulated using the following
 347  * rules:
 348  *   1. For each set of 8 pagesets, a complete group from
 349  *      tid_group_list is taken, programmed, and moved to
 350  *      the tid_full_list list.
 351  *   2. For all remaining pagesets:
 352  *      2.1 If the tid_used_list is empty and the tid_group_list
 353  *          is empty, stop processing pageset and return only
 354  *          what has been programmed up to this point.
 355  *      2.2 If the tid_used_list is empty and the tid_group_list
 356  *          is not empty, move a group from tid_group_list to
 357  *          tid_used_list.
 358  *      2.3 For each group is tid_used_group, program as much as
 359  *          can fit into the group. If the group becomes fully
 360  *          used, move it to tid_full_list.
 361  */
 362 int hfi1_user_exp_rcv_setup(struct hfi1_filedata *fd,
 363                             struct hfi1_tid_info *tinfo)
 364 {
 365         int ret = 0, need_group = 0, pinned;
 366         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 367         struct hfi1_devdata *dd = uctxt->dd;
 368         unsigned npages, ngroups, pageidx = 0, pageset_count, npagesets,
 369                 tididx = 0, mapped, mapped_pages = 0;
 370         unsigned long vaddr = tinfo->vaddr;
 371         struct page **pages = NULL;
 372         u32 *tidlist = NULL;
 373         struct tid_pageset *pagesets = NULL;
 374
 375         /* Get the number of pages the user buffer spans */
 376         npages = num_user_pages(vaddr, tinfo->length);
 377         if (!npages)
 378                 return -EINVAL;
 379
 380         if (npages > uctxt->expected_count) {
 381                 dd_dev_err(dd, "Expected buffer too big\n");
 382                 return -EINVAL;
 383         }
 384
 385         /* Verify that access is OK for the user buffer */
 386         if (!access_ok(VERIFY_WRITE, (void __user *)vaddr,
 387                        npages * PAGE_SIZE)) {
 388                 dd_dev_err(dd, "Fail vaddr %p, %u pages, !access_ok\n",
 389                            (void *)vaddr, npages);
 390                 return -EFAULT;
 391         }
 392
 393         pagesets = kcalloc(uctxt->expected_count, sizeof(*pagesets),
 394                            GFP_KERNEL);
 395         if (!pagesets)
 396                 return -ENOMEM;
 397
 398         /* Allocate the array of struct page pointers needed for pinning */
 399         pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL);
 400         if (!pages) {
 401                 ret = -ENOMEM;
 402                 goto bail;
 403         }
 404
 405         /*
 406          * Pin all the pages of the user buffer. If we can't pin all the
 407          * pages, accept the amount pinned so far and program only that.
 408          * User space knows how to deal with partially programmed buffers.
 409          */
 410         if (!hfi1_can_pin_pages(dd, fd->mm, fd->tid_n_pinned, npages)) {
 411                 ret = -ENOMEM;
 412                 goto bail;
 413         }
 414
 415         pinned = hfi1_acquire_user_pages(fd->mm, vaddr, npages, true, pages);
 416         if (pinned <= 0) {
 417                 ret = pinned;
 418                 goto bail;
 419         }
 420         fd->tid_n_pinned += npages;
 421
 422         /* Find sets of physically contiguous pages */
 423         npagesets = find_phys_blocks(pages, pinned, pagesets);
 424
 425         /*
 426          * We don't need to access this under a lock since tid_used is per
 427          * process and the same process cannot be in hfi1_user_exp_rcv_clear()
 428          * and hfi1_user_exp_rcv_setup() at the same time.
 429          */
 430         spin_lock(&fd->tid_lock);
 431         if (fd->tid_used + npagesets > fd->tid_limit)
 432                 pageset_count = fd->tid_limit - fd->tid_used;
 433         else
 434                 pageset_count = npagesets;
 435         spin_unlock(&fd->tid_lock);
 436
 437         if (!pageset_count)
 438                 goto bail;
 439
 440         ngroups = pageset_count / dd->rcv_entries.group_size;
 441         tidlist = kcalloc(pageset_count, sizeof(*tidlist), GFP_KERNEL);
 442         if (!tidlist) {
 443                 ret = -ENOMEM;
 444                 goto nomem;
 445         }
 446
 447         tididx = 0;
 448
 449         /*
 450          * From this point on, we are going to be using shared (between master
 451          * and subcontexts) context resources. We need to take the lock.
 452          */
 453         mutex_lock(&uctxt->exp_lock);
 454         /*
 455          * The first step is to program the RcvArray entries which are complete
 456          * groups.
 457          */
 458         while (ngroups && uctxt->tid_group_list.count) {
 459                 struct tid_group *grp =
 460                         tid_group_pop(&uctxt->tid_group_list);
 461
 462                 ret = program_rcvarray(fd, vaddr, grp, pagesets,
 463                                        pageidx, dd->rcv_entries.group_size,
 464                                        pages, tidlist, &tididx, &mapped);
 465                 /*
 466                  * If there was a failure to program the RcvArray
 467                  * entries for the entire group, reset the grp fields
 468                  * and add the grp back to the free group list.
 469                  */
 470                 if (ret <= 0) {
 471                         tid_group_add_tail(grp, &uctxt->tid_group_list);
 472                         hfi1_cdbg(TID,
 473                                   "Failed to program RcvArray group %d", ret);
 474                         goto unlock;
 475                 }
 476
 477                 tid_group_add_tail(grp, &uctxt->tid_full_list);
 478                 ngroups--;
 479                 pageidx += ret;
 480                 mapped_pages += mapped;
 481         }
 482
 483         while (pageidx < pageset_count) {
 484                 struct tid_group *grp, *ptr;
 485                 /*
 486                  * If we don't have any partially used tid groups, check
 487                  * if we have empty groups. If so, take one from there and
 488                  * put in the partially used list.
 489                  */
 490                 if (!uctxt->tid_used_list.count || need_group) {
 491                         if (!uctxt->tid_group_list.count)
 492                                 goto unlock;
 493
 494                         grp = tid_group_pop(&uctxt->tid_group_list);
 495                         tid_group_add_tail(grp, &uctxt->tid_used_list);
 496                         need_group = 0;
 497                 }
 498                 /*
 499                  * There is an optimization opportunity here - instead of
 500                  * fitting as many page sets as we can, check for a group
 501                  * later on in the list that could fit all of them.
 502                  */
 503                 list_for_each_entry_safe(grp, ptr, &uctxt->tid_used_list.list,
 504                                          list) {
 505                         unsigned use = min_t(unsigned, pageset_count - pageidx,
 506                                              grp->size - grp->used);
 507
 508                         ret = program_rcvarray(fd, vaddr, grp, pagesets,
 509                                                pageidx, use, pages, tidlist,
 510                                                &tididx, &mapped);
 511                         if (ret < 0) {
 512                                 hfi1_cdbg(TID,
 513                                           "Failed to program RcvArray entries %d",
 514                                           ret);
 515                                 ret = -EFAULT;
 516                                 goto unlock;
 517                         } else if (ret > 0) {
 518                                 if (grp->used == grp->size)
 519                                         tid_group_move(grp,
 520                                                        &uctxt->tid_used_list,
 521                                                        &uctxt->tid_full_list);
 522                                 pageidx += ret;
 523                                 mapped_pages += mapped;
 524                                 need_group = 0;
 525                                 /* Check if we are done so we break out early */
 526                                 if (pageidx >= pageset_count)
 527                                         break;
 528                         } else if (WARN_ON(ret == 0)) {
 529                                 /*
 530                                  * If ret is 0, we did not program any entries
 531                                  * into this group, which can only happen if
 532                                  * we've screwed up the accounting somewhere.
 533                                  * Warn and try to continue.
 534                                  */
 535                                 need_group = 1;
 536                         }
 537                 }
 538         }
 539 unlock:
 540         mutex_unlock(&uctxt->exp_lock);
 541 nomem:
 542         hfi1_cdbg(TID, "total mapped: tidpairs:%u pages:%u (%d)", tididx,
 543                   mapped_pages, ret);
 544         if (tididx) {
 545                 spin_lock(&fd->tid_lock);
 546                 fd->tid_used += tididx;
 547                 spin_unlock(&fd->tid_lock);
 548                 tinfo->tidcnt = tididx;
 549                 tinfo->length = mapped_pages * PAGE_SIZE;
 550
 551                 if (copy_to_user((void __user *)(unsigned long)tinfo->tidlist,
 552                                  tidlist, sizeof(tidlist[0]) * tididx)) {
 553                         /*
 554                          * On failure to copy to the user level, we need to undo
 555                          * everything done so far so we don't leak resources.
 556                          */
 557                         tinfo->tidlist = (unsigned long)&tidlist;
 558                         hfi1_user_exp_rcv_clear(fd, tinfo);
 559                         tinfo->tidlist = 0;
 560                         ret = -EFAULT;
 561                         goto bail;
 562                 }
 563         }
 564
 565         /*
 566          * If not everything was mapped (due to insufficient RcvArray entries,
 567          * for example), unpin all unmapped pages so we can pin them nex time.
 568          */
 569         if (mapped_pages != pinned) {
 570                 hfi1_release_user_pages(fd->mm, &pages[mapped_pages],
 571                                         pinned - mapped_pages,
 572                                         false);
 573                 fd->tid_n_pinned -= pinned - mapped_pages;
 574         }
 575 bail:
 576         kfree(pagesets);
 577         kfree(pages);
 578         kfree(tidlist);
 579         return ret > 0 ? 0 : ret;
 580 }
 581
 582 int hfi1_user_exp_rcv_clear(struct hfi1_filedata *fd,
 583                             struct hfi1_tid_info *tinfo)
 584 {
 585         int ret = 0;
 586         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 587         u32 *tidinfo;
 588         unsigned tididx;
 589
 590         if (unlikely(tinfo->tidcnt > fd->tid_used))
 591                 return -EINVAL;
 592
 593         tidinfo = memdup_user((void __user *)(unsigned long)tinfo->tidlist,
 594                               sizeof(tidinfo[0]) * tinfo->tidcnt);
 595         if (IS_ERR(tidinfo))
 596                 return PTR_ERR(tidinfo);
 597
 598         mutex_lock(&uctxt->exp_lock);
 599         for (tididx = 0; tididx < tinfo->tidcnt; tididx++) {
 600                 ret = unprogram_rcvarray(fd, tidinfo[tididx], NULL);
 601                 if (ret) {
 602                         hfi1_cdbg(TID, "Failed to unprogram rcv array %d",
 603                                   ret);
 604                         break;
 605                 }
 606         }
 607         spin_lock(&fd->tid_lock);
 608         fd->tid_used -= tididx;
 609         spin_unlock(&fd->tid_lock);
 610         tinfo->tidcnt = tididx;
 611         mutex_unlock(&uctxt->exp_lock);
 612
 613         kfree(tidinfo);
 614         return ret;
 615 }
 616
 617 int hfi1_user_exp_rcv_invalid(struct hfi1_filedata *fd,
 618                               struct hfi1_tid_info *tinfo)
 619 {
 620         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 621         unsigned long *ev = uctxt->dd->events +
 622                 (((uctxt->ctxt - uctxt->dd->first_dyn_alloc_ctxt) *
 623                   HFI1_MAX_SHARED_CTXTS) + fd->subctxt);
 624         u32 *array;
 625         int ret = 0;
 626
 627         if (!fd->invalid_tids)
 628                 return -EINVAL;
 629
 630         /*
 631          * copy_to_user() can sleep, which will leave the invalid_lock
 632          * locked and cause the MMU notifier to be blocked on the lock
 633          * for a long time.
 634          * Copy the data to a local buffer so we can release the lock.
 635          */
 636         array = kcalloc(uctxt->expected_count, sizeof(*array), GFP_KERNEL);
 637         if (!array)
 638                 return -EFAULT;
 639
 640         spin_lock(&fd->invalid_lock);
 641         if (fd->invalid_tid_idx) {
 642                 memcpy(array, fd->invalid_tids, sizeof(*array) *
 643                        fd->invalid_tid_idx);
 644                 memset(fd->invalid_tids, 0, sizeof(*fd->invalid_tids) *
 645                        fd->invalid_tid_idx);
 646                 tinfo->tidcnt = fd->invalid_tid_idx;
 647                 fd->invalid_tid_idx = 0;
 648                 /*
 649                  * Reset the user flag while still holding the lock.
 650                  * Otherwise, PSM can miss events.
 651                  */
 652                 clear_bit(_HFI1_EVENT_TID_MMU_NOTIFY_BIT, ev);
 653         } else {
 654                 tinfo->tidcnt = 0;
 655         }
 656         spin_unlock(&fd->invalid_lock);
 657
 658         if (tinfo->tidcnt) {
 659                 if (copy_to_user((void __user *)tinfo->tidlist,
 660                                  array, sizeof(*array) * tinfo->tidcnt))
 661                         ret = -EFAULT;
 662         }
 663         kfree(array);
 664
 665         return ret;
 666 }
 667
 668 static u32 find_phys_blocks(struct page **pages, unsigned npages,
 669                             struct tid_pageset *list)
 670 {
 671         unsigned pagecount, pageidx, setcount = 0, i;
 672         unsigned long pfn, this_pfn;
 673
 674         if (!npages)
 675                 return 0;
 676
 677         /*
 678          * Look for sets of physically contiguous pages in the user buffer.
 679          * This will allow us to optimize Expected RcvArray entry usage by
 680          * using the bigger supported sizes.
 681          */
 682         pfn = page_to_pfn(pages[0]);
 683         for (pageidx = 0, pagecount = 1, i = 1; i <= npages; i++) {
 684                 this_pfn = i < npages ? page_to_pfn(pages[i]) : 0;
 685
 686                 /*
 687                  * If the pfn's are not sequential, pages are not physically
 688                  * contiguous.
 689                  */
 690                 if (this_pfn != ++pfn) {
 691                         /*
 692                          * At this point we have to loop over the set of
 693                          * physically contiguous pages and break them down it
 694                          * sizes supported by the HW.
 695                          * There are two main constraints:
 696                          *     1. The max buffer size is MAX_EXPECTED_BUFFER.
 697                          *        If the total set size is bigger than that
 698                          *        program only a MAX_EXPECTED_BUFFER chunk.
 699                          *     2. The buffer size has to be a power of two. If
 700                          *        it is not, round down to the closes power of
 701                          *        2 and program that size.
 702                          */
 703                         while (pagecount) {
 704                                 int maxpages = pagecount;
 705                                 u32 bufsize = pagecount * PAGE_SIZE;
 706
 707                                 if (bufsize > MAX_EXPECTED_BUFFER)
 708                                         maxpages =
 709                                                 MAX_EXPECTED_BUFFER >>
 710                                                 PAGE_SHIFT;
 711                                 else if (!is_power_of_2(bufsize))
 712                                         maxpages =
 713                                                 rounddown_pow_of_two(bufsize) >>
 714                                                 PAGE_SHIFT;
 715
 716                                 list[setcount].idx = pageidx;
 717                                 list[setcount].count = maxpages;
 718                                 pagecount -= maxpages;
 719                                 pageidx += maxpages;
 720                                 setcount++;
 721                         }
 722                         pageidx = i;
 723                         pagecount = 1;
 724                         pfn = this_pfn;
 725                 } else {
 726                         pagecount++;
 727                 }
 728         }
 729         return setcount;
 730 }
 731
 732 /**
 733  * program_rcvarray() - program an RcvArray group with receive buffers
 734  * @fd: filedata pointer
 735  * @vaddr: starting user virtual address
 736  * @grp: RcvArray group
 737  * @sets: array of struct tid_pageset holding information on physically
 738  *        contiguous chunks from the user buffer
 739  * @start: starting index into sets array
 740  * @count: number of struct tid_pageset's to program
 741  * @pages: an array of struct page * for the user buffer
 742  * @tidlist: the array of u32 elements when the information about the
 743  *           programmed RcvArray entries is to be encoded.
 744  * @tididx: starting offset into tidlist
 745  * @pmapped: (output parameter) number of pages programmed into the RcvArray
 746  *           entries.
 747  *
 748  * This function will program up to 'count' number of RcvArray entries from the
 749  * group 'grp'. To make best use of write-combining writes, the function will
 750  * perform writes to the unused RcvArray entries which will be ignored by the
 751  * HW. Each RcvArray entry will be programmed with a physically contiguous
 752  * buffer chunk from the user's virtual buffer.
 753  *
 754  * Return:
 755  * -EINVAL if the requested count is larger than the size of the group,
 756  * -ENOMEM or -EFAULT on error from set_rcvarray_entry(), or
 757  * number of RcvArray entries programmed.
 758  */
 759 static int program_rcvarray(struct hfi1_filedata *fd, unsigned long vaddr,
 760                             struct tid_group *grp,
 761                             struct tid_pageset *sets,
 762                             unsigned start, u16 count, struct page **pages,
 763                             u32 *tidlist, unsigned *tididx, unsigned *pmapped)
 764 {
 765         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 766         struct hfi1_devdata *dd = uctxt->dd;
 767         u16 idx;
 768         u32 tidinfo = 0, rcventry, useidx = 0;
 769         int mapped = 0;
 770
 771         /* Count should never be larger than the group size */
 772         if (count > grp->size)
 773                 return -EINVAL;
 774
 775         /* Find the first unused entry in the group */
 776         for (idx = 0; idx < grp->size; idx++) {
 777                 if (!(grp->map & (1 << idx))) {
 778                         useidx = idx;
 779                         break;
 780                 }
 781                 rcv_array_wc_fill(dd, grp->base + idx);
 782         }
 783
 784         idx = 0;
 785         while (idx < count) {
 786                 u16 npages, pageidx, setidx = start + idx;
 787                 int ret = 0;
 788
 789                 /*
 790                  * If this entry in the group is used, move to the next one.
 791                  * If we go past the end of the group, exit the loop.
 792                  */
 793                 if (useidx >= grp->size) {
 794                         break;
 795                 } else if (grp->map & (1 << useidx)) {
 796                         rcv_array_wc_fill(dd, grp->base + useidx);
 797                         useidx++;
 798                         continue;
 799                 }
 800
 801                 rcventry = grp->base + useidx;
 802                 npages = sets[setidx].count;
 803                 pageidx = sets[setidx].idx;
 804
 805                 ret = set_rcvarray_entry(fd, vaddr + (pageidx * PAGE_SIZE),
 806                                          rcventry, grp, pages + pageidx,
 807                                          npages);
 808                 if (ret)
 809                         return ret;
 810                 mapped += npages;
 811
 812                 tidinfo = rcventry2tidinfo(rcventry - uctxt->expected_base) |
 813                         EXP_TID_SET(LEN, npages);
 814                 tidlist[(*tididx)++] = tidinfo;
 815                 grp->used++;
 816                 grp->map |= 1 << useidx++;
 817                 idx++;
 818         }
 819
 820         /* Fill the rest of the group with "blank" writes */
 821         for (; useidx < grp->size; useidx++)
 822                 rcv_array_wc_fill(dd, grp->base + useidx);
 823         *pmapped = mapped;
 824         return idx;
 825 }
 826
 827 static int set_rcvarray_entry(struct hfi1_filedata *fd, unsigned long vaddr,
 828                               u32 rcventry, struct tid_group *grp,
 829                               struct page **pages, unsigned npages)
 830 {
 831         int ret;
 832         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 833         struct tid_rb_node *node;
 834         struct hfi1_devdata *dd = uctxt->dd;
 835         dma_addr_t phys;
 836
 837         /*
 838          * Allocate the node first so we can handle a potential
 839          * failure before we've programmed anything.
 840          */
 841         node = kzalloc(sizeof(*node) + (sizeof(struct page *) * npages),
 842                        GFP_KERNEL);
 843         if (!node)
 844                 return -ENOMEM;
 845
 846         phys = pci_map_single(dd->pcidev,
 847                               __va(page_to_phys(pages[0])),
 848                               npages * PAGE_SIZE, PCI_DMA_FROMDEVICE);
 849         if (dma_mapping_error(&dd->pcidev->dev, phys)) {
 850                 dd_dev_err(dd, "Failed to DMA map Exp Rcv pages 0x%llx\n",
 851                            phys);
 852                 kfree(node);
 853                 return -EFAULT;
 854         }
 855
 856         node->mmu.addr = vaddr;
 857         node->mmu.len = npages * PAGE_SIZE;
 858         node->phys = page_to_phys(pages[0]);
 859         node->npages = npages;
 860         node->rcventry = rcventry;
 861         node->dma_addr = phys;
 862         node->grp = grp;
 863         node->freed = false;
 864         memcpy(node->pages, pages, sizeof(struct page *) * npages);
 865
 866         if (!fd->handler)
 867                 ret = tid_rb_insert(fd, &node->mmu);
 868         else
 869                 ret = hfi1_mmu_rb_insert(fd->handler, &node->mmu);
 870
 871         if (ret) {
 872                 hfi1_cdbg(TID, "Failed to insert RB node %u 0x%lx, 0x%lx %d",
 873                           node->rcventry, node->mmu.addr, node->phys, ret);
 874                 pci_unmap_single(dd->pcidev, phys, npages * PAGE_SIZE,
 875                                  PCI_DMA_FROMDEVICE);
 876                 kfree(node);
 877                 return -EFAULT;
 878         }
 879         hfi1_put_tid(dd, rcventry, PT_EXPECTED, phys, ilog2(npages) + 1);
 880         trace_hfi1_exp_tid_reg(uctxt->ctxt, fd->subctxt, rcventry, npages,
 881                                node->mmu.addr, node->phys, phys);
 882         return 0;
 883 }
 884
 885 static int unprogram_rcvarray(struct hfi1_filedata *fd, u32 tidinfo,
 886                               struct tid_group **grp)
 887 {
 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_dyn_alloc_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 }