drivers/staging/rdma/hfi1/user_exp_rcv.c

   1 /*
   2  * Copyright(c) 2015, 2016 Intel Corporation.
   3  *
   4  * This file is provided under a dual BSD/GPLv2 license.  When using or
   5  * redistributing this file, you may do so under either license.
   6  *
   7  * GPL LICENSE SUMMARY
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of version 2 of the GNU General Public License as
  11  * published by the Free Software Foundation.
  12  *
  13  * This program is distributed in the hope that it will be useful, but
  14  * WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * General Public License for more details.
  17  *
  18  * BSD LICENSE
  19  *
  20  * Redistribution and use in source and binary forms, with or without
  21  * modification, are permitted provided that the following conditions
  22  * are met:
  23  *
  24  *  - Redistributions of source code must retain the above copyright
  25  *    notice, this list of conditions and the following disclaimer.
  26  *  - Redistributions in binary form must reproduce the above copyright
  27  *    notice, this list of conditions and the following disclaimer in
  28  *    the documentation and/or other materials provided with the
  29  *    distribution.
  30  *  - Neither the name of Intel Corporation nor the names of its
  31  *    contributors may be used to endorse or promote products derived
  32  *    from this software without specific prior written permission.
  33  *
  34  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  35  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  36  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  37  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  38  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  39  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  40  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  41  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  42  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  43  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  44  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  45  *
  46  */
  47 #include <asm/page.h>
  48
  49 #include "user_exp_rcv.h"
  50 #include "trace.h"
  51 #include "mmu_rb.h"
  52
  53 struct tid_group {
  54         struct list_head list;
  55         unsigned base;
  56         u8 size;
  57         u8 used;
  58         u8 map;
  59 };
  60
  61 struct tid_rb_node {
  62         struct mmu_rb_node mmu;
  63         unsigned long phys;
  64         struct tid_group *grp;
  65         u32 rcventry;
  66         dma_addr_t dma_addr;
  67         bool freed;
  68         unsigned npages;
  69         struct page *pages[0];
  70 };
  71
  72 struct tid_pageset {
  73         u16 idx;
  74         u16 count;
  75 };
  76
  77 #define EXP_TID_SET_EMPTY(set) (set.count == 0 && list_empty(&set.list))
  78
  79 #define num_user_pages(vaddr, len)                                     \
  80         (1 + (((((unsigned long)(vaddr) +                              \
  81                  (unsigned long)(len) - 1) & PAGE_MASK) -              \
  82                ((unsigned long)vaddr & PAGE_MASK)) >> PAGE_SHIFT))
  83
  84 static void unlock_exp_tids(struct hfi1_ctxtdata *, struct exp_tid_set *,
  85                             struct rb_root *);
  86 static u32 find_phys_blocks(struct page **, unsigned, struct tid_pageset *);
  87 static int set_rcvarray_entry(struct file *, unsigned long, u32,
  88                               struct tid_group *, struct page **, unsigned);
  89 static int mmu_rb_insert(struct rb_root *, struct mmu_rb_node *);
  90 static void mmu_rb_remove(struct rb_root *, struct mmu_rb_node *, bool);
  91 static int mmu_rb_invalidate(struct rb_root *, struct mmu_rb_node *);
  92 static int program_rcvarray(struct file *, unsigned long, struct tid_group *,
  93                             struct tid_pageset *, unsigned, u16, struct page **,
  94                             u32 *, unsigned *, unsigned *);
  95 static int unprogram_rcvarray(struct file *, u32, struct tid_group **);
  96 static void clear_tid_node(struct hfi1_filedata *, u16, struct tid_rb_node *);
  97
  98 static struct mmu_rb_ops tid_rb_ops = {
  99         .insert = mmu_rb_insert,
 100         .remove = mmu_rb_remove,
 101         .invalidate = mmu_rb_invalidate
 102 };
 103
 104 static inline u32 rcventry2tidinfo(u32 rcventry)
 105 {
 106         u32 pair = rcventry & ~0x1;
 107
 108         return EXP_TID_SET(IDX, pair >> 1) |
 109                 EXP_TID_SET(CTRL, 1 << (rcventry - pair));
 110 }
 111
 112 static inline void exp_tid_group_init(struct exp_tid_set *set)
 113 {
 114         INIT_LIST_HEAD(&set->list);
 115         set->count = 0;
 116 }
 117
 118 static inline void tid_group_remove(struct tid_group *grp,
 119                                     struct exp_tid_set *set)
 120 {
 121         list_del_init(&grp->list);
 122         set->count--;
 123 }
 124
 125 static inline void tid_group_add_tail(struct tid_group *grp,
 126                                       struct exp_tid_set *set)
 127 {
 128         list_add_tail(&grp->list, &set->list);
 129         set->count++;
 130 }
 131
 132 static inline struct tid_group *tid_group_pop(struct exp_tid_set *set)
 133 {
 134         struct tid_group *grp =
 135                 list_first_entry(&set->list, struct tid_group, list);
 136         list_del_init(&grp->list);
 137         set->count--;
 138         return grp;
 139 }
 140
 141 static inline void tid_group_move(struct tid_group *group,
 142                                   struct exp_tid_set *s1,
 143                                   struct exp_tid_set *s2)
 144 {
 145         tid_group_remove(group, s1);
 146         tid_group_add_tail(group, s2);
 147 }
 148
 149 /*
 150  * Initialize context and file private data needed for Expected
 151  * receive caching. This needs to be done after the context has
 152  * been configured with the eager/expected RcvEntry counts.
 153  */
 154 int hfi1_user_exp_rcv_init(struct file *fp)
 155 {
 156         struct hfi1_filedata *fd = fp->private_data;
 157         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 158         struct hfi1_devdata *dd = uctxt->dd;
 159         unsigned tidbase;
 160         int i, ret = 0;
 161
 162         spin_lock_init(&fd->tid_lock);
 163         spin_lock_init(&fd->invalid_lock);
 164         fd->tid_rb_root = RB_ROOT;
 165
 166         if (!uctxt->subctxt_cnt || !fd->subctxt) {
 167                 exp_tid_group_init(&uctxt->tid_group_list);
 168                 exp_tid_group_init(&uctxt->tid_used_list);
 169                 exp_tid_group_init(&uctxt->tid_full_list);
 170
 171                 tidbase = uctxt->expected_base;
 172                 for (i = 0; i < uctxt->expected_count /
 173                              dd->rcv_entries.group_size; i++) {
 174                         struct tid_group *grp;
 175
 176                         grp = kzalloc(sizeof(*grp), GFP_KERNEL);
 177                         if (!grp) {
 178                                 /*
 179                                  * If we fail here, the groups already
 180                                  * allocated will be freed by the close
 181                                  * call.
 182                                  */
 183                                 ret = -ENOMEM;
 184                                 goto done;
 185                         }
 186                         grp->size = dd->rcv_entries.group_size;
 187                         grp->base = tidbase;
 188                         tid_group_add_tail(grp, &uctxt->tid_group_list);
 189                         tidbase += dd->rcv_entries.group_size;
 190                 }
 191         }
 192
 193         fd->entry_to_rb = kcalloc(uctxt->expected_count,
 194                                      sizeof(struct rb_node *),
 195                                      GFP_KERNEL);
 196         if (!fd->entry_to_rb)
 197                 return -ENOMEM;
 198
 199         if (!HFI1_CAP_IS_USET(TID_UNMAP)) {
 200                 fd->invalid_tid_idx = 0;
 201                 fd->invalid_tids = kzalloc(uctxt->expected_count *
 202                                            sizeof(u32), GFP_KERNEL);
 203                 if (!fd->invalid_tids) {
 204                         ret = -ENOMEM;
 205                         goto done;
 206                 }
 207
 208                 /*
 209                  * Register MMU notifier callbacks. If the registration
 210                  * fails, continue but turn off the TID caching for
 211                  * all user contexts.
 212                  */
 213                 ret = hfi1_mmu_rb_register(&fd->tid_rb_root, &tid_rb_ops);
 214                 if (ret) {
 215                         dd_dev_info(dd,
 216                                     "Failed MMU notifier registration %d\n",
 217                                     ret);
 218                         HFI1_CAP_USET(TID_UNMAP);
 219                         ret = 0;
 220                 }
 221         }
 222
 223         /*
 224          * PSM does not have a good way to separate, count, and
 225          * effectively enforce a limit on RcvArray entries used by
 226          * subctxts (when context sharing is used) when TID caching
 227          * is enabled. To help with that, we calculate a per-process
 228          * RcvArray entry share and enforce that.
 229          * If TID caching is not in use, PSM deals with usage on its
 230          * own. In that case, we allow any subctxt to take all of the
 231          * entries.
 232          *
 233          * Make sure that we set the tid counts only after successful
 234          * init.
 235          */
 236         spin_lock(&fd->tid_lock);
 237         if (uctxt->subctxt_cnt && !HFI1_CAP_IS_USET(TID_UNMAP)) {
 238                 u16 remainder;
 239
 240                 fd->tid_limit = uctxt->expected_count / uctxt->subctxt_cnt;
 241                 remainder = uctxt->expected_count % uctxt->subctxt_cnt;
 242                 if (remainder && fd->subctxt < remainder)
 243                         fd->tid_limit++;
 244         } else {
 245                 fd->tid_limit = uctxt->expected_count;
 246         }
 247         spin_unlock(&fd->tid_lock);
 248 done:
 249         return ret;
 250 }
 251
 252 int hfi1_user_exp_rcv_free(struct hfi1_filedata *fd)
 253 {
 254         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 255         struct tid_group *grp, *gptr;
 256
 257         /*
 258          * The notifier would have been removed when the process'es mm
 259          * was freed.
 260          */
 261         if (!HFI1_CAP_IS_USET(TID_UNMAP))
 262                 hfi1_mmu_rb_unregister(&fd->tid_rb_root);
 263
 264         kfree(fd->invalid_tids);
 265
 266         if (!uctxt->cnt) {
 267                 if (!EXP_TID_SET_EMPTY(uctxt->tid_full_list))
 268                         unlock_exp_tids(uctxt, &uctxt->tid_full_list,
 269                                         &fd->tid_rb_root);
 270                 if (!EXP_TID_SET_EMPTY(uctxt->tid_used_list))
 271                         unlock_exp_tids(uctxt, &uctxt->tid_used_list,
 272                                         &fd->tid_rb_root);
 273                 list_for_each_entry_safe(grp, gptr, &uctxt->tid_group_list.list,
 274                                          list) {
 275                         list_del_init(&grp->list);
 276                         kfree(grp);
 277                 }
 278                 hfi1_clear_tids(uctxt);
 279         }
 280
 281         kfree(fd->entry_to_rb);
 282         return 0;
 283 }
 284
 285 /*
 286  * Write an "empty" RcvArray entry.
 287  * This function exists so the TID registaration code can use it
 288  * to write to unused/unneeded entries and still take advantage
 289  * of the WC performance improvements. The HFI will ignore this
 290  * write to the RcvArray entry.
 291  */
 292 static inline void rcv_array_wc_fill(struct hfi1_devdata *dd, u32 index)
 293 {
 294         /*
 295          * Doing the WC fill writes only makes sense if the device is
 296          * present and the RcvArray has been mapped as WC memory.
 297          */
 298         if ((dd->flags & HFI1_PRESENT) && dd->rcvarray_wc)
 299                 writeq(0, dd->rcvarray_wc + (index * 8));
 300 }
 301
 302 /*
 303  * RcvArray entry allocation for Expected Receives is done by the
 304  * following algorithm:
 305  *
 306  * The context keeps 3 lists of groups of RcvArray entries:
 307  *   1. List of empty groups - tid_group_list
 308  *      This list is created during user context creation and
 309  *      contains elements which describe sets (of 8) of empty
 310  *      RcvArray entries.
 311  *   2. List of partially used groups - tid_used_list
 312  *      This list contains sets of RcvArray entries which are
 313  *      not completely used up. Another mapping request could
 314  *      use some of all of the remaining entries.
 315  *   3. List of full groups - tid_full_list
 316  *      This is the list where sets that are completely used
 317  *      up go.
 318  *
 319  * An attempt to optimize the usage of RcvArray entries is
 320  * made by finding all sets of physically contiguous pages in a
 321  * user's buffer.
 322  * These physically contiguous sets are further split into
 323  * sizes supported by the receive engine of the HFI. The
 324  * resulting sets of pages are stored in struct tid_pageset,
 325  * which describes the sets as:
 326  *    * .count - number of pages in this set
 327  *    * .idx - starting index into struct page ** array
 328  *                    of this set
 329  *
 330  * From this point on, the algorithm deals with the page sets
 331  * described above. The number of pagesets is divided by the
 332  * RcvArray group size to produce the number of full groups
 333  * needed.
 334  *
 335  * Groups from the 3 lists are manipulated using the following
 336  * rules:
 337  *   1. For each set of 8 pagesets, a complete group from
 338  *      tid_group_list is taken, programmed, and moved to
 339  *      the tid_full_list list.
 340  *   2. For all remaining pagesets:
 341  *      2.1 If the tid_used_list is empty and the tid_group_list
 342  *          is empty, stop processing pageset and return only
 343  *          what has been programmed up to this point.
 344  *      2.2 If the tid_used_list is empty and the tid_group_list
 345  *          is not empty, move a group from tid_group_list to
 346  *          tid_used_list.
 347  *      2.3 For each group is tid_used_group, program as much as
 348  *          can fit into the group. If the group becomes fully
 349  *          used, move it to tid_full_list.
 350  */
 351 int hfi1_user_exp_rcv_setup(struct file *fp, struct hfi1_tid_info *tinfo)
 352 {
 353         int ret = 0, need_group = 0, pinned;
 354         struct hfi1_filedata *fd = fp->private_data;
 355         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 356         struct hfi1_devdata *dd = uctxt->dd;
 357         unsigned npages, ngroups, pageidx = 0, pageset_count, npagesets,
 358                 tididx = 0, mapped, mapped_pages = 0;
 359         unsigned long vaddr = tinfo->vaddr;
 360         struct page **pages = NULL;
 361         u32 *tidlist = NULL;
 362         struct tid_pageset *pagesets = NULL;
 363
 364         /* Get the number of pages the user buffer spans */
 365         npages = num_user_pages(vaddr, tinfo->length);
 366         if (!npages)
 367                 return -EINVAL;
 368
 369         if (npages > uctxt->expected_count) {
 370                 dd_dev_err(dd, "Expected buffer too big\n");
 371                 return -EINVAL;
 372         }
 373
 374         /* Verify that access is OK for the user buffer */
 375         if (!access_ok(VERIFY_WRITE, (void __user *)vaddr,
 376                        npages * PAGE_SIZE)) {
 377                 dd_dev_err(dd, "Fail vaddr %p, %u pages, !access_ok\n",
 378                            (void *)vaddr, npages);
 379                 return -EFAULT;
 380         }
 381
 382         pagesets = kcalloc(uctxt->expected_count, sizeof(*pagesets),
 383                            GFP_KERNEL);
 384         if (!pagesets)
 385                 return -ENOMEM;
 386
 387         /* Allocate the array of struct page pointers needed for pinning */
 388         pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL);
 389         if (!pages) {
 390                 ret = -ENOMEM;
 391                 goto bail;
 392         }
 393
 394         /*
 395          * Pin all the pages of the user buffer. If we can't pin all the
 396          * pages, accept the amount pinned so far and program only that.
 397          * User space knows how to deal with partially programmed buffers.
 398          */
 399         if (!hfi1_can_pin_pages(dd, fd->tid_n_pinned, npages))
 400                 return -ENOMEM;
 401         pinned = hfi1_acquire_user_pages(vaddr, npages, true, pages);
 402         if (pinned <= 0) {
 403                 ret = pinned;
 404                 goto bail;
 405         }
 406         fd->tid_n_pinned += npages;
 407
 408         /* Find sets of physically contiguous pages */
 409         npagesets = find_phys_blocks(pages, pinned, pagesets);
 410
 411         /*
 412          * We don't need to access this under a lock since tid_used is per
 413          * process and the same process cannot be in hfi1_user_exp_rcv_clear()
 414          * and hfi1_user_exp_rcv_setup() at the same time.
 415          */
 416         spin_lock(&fd->tid_lock);
 417         if (fd->tid_used + npagesets > fd->tid_limit)
 418                 pageset_count = fd->tid_limit - fd->tid_used;
 419         else
 420                 pageset_count = npagesets;
 421         spin_unlock(&fd->tid_lock);
 422
 423         if (!pageset_count)
 424                 goto bail;
 425
 426         ngroups = pageset_count / dd->rcv_entries.group_size;
 427         tidlist = kcalloc(pageset_count, sizeof(*tidlist), GFP_KERNEL);
 428         if (!tidlist) {
 429                 ret = -ENOMEM;
 430                 goto nomem;
 431         }
 432
 433         tididx = 0;
 434
 435         /*
 436          * From this point on, we are going to be using shared (between master
 437          * and subcontexts) context resources. We need to take the lock.
 438          */
 439         mutex_lock(&uctxt->exp_lock);
 440         /*
 441          * The first step is to program the RcvArray entries which are complete
 442          * groups.
 443          */
 444         while (ngroups && uctxt->tid_group_list.count) {
 445                 struct tid_group *grp =
 446                         tid_group_pop(&uctxt->tid_group_list);
 447
 448                 ret = program_rcvarray(fp, vaddr, grp, pagesets,
 449                                        pageidx, dd->rcv_entries.group_size,
 450                                        pages, tidlist, &tididx, &mapped);
 451                 /*
 452                  * If there was a failure to program the RcvArray
 453                  * entries for the entire group, reset the grp fields
 454                  * and add the grp back to the free group list.
 455                  */
 456                 if (ret <= 0) {
 457                         tid_group_add_tail(grp, &uctxt->tid_group_list);
 458                         hfi1_cdbg(TID,
 459                                   "Failed to program RcvArray group %d", ret);
 460                         goto unlock;
 461                 }
 462
 463                 tid_group_add_tail(grp, &uctxt->tid_full_list);
 464                 ngroups--;
 465                 pageidx += ret;
 466                 mapped_pages += mapped;
 467         }
 468
 469         while (pageidx < pageset_count) {
 470                 struct tid_group *grp, *ptr;
 471                 /*
 472                  * If we don't have any partially used tid groups, check
 473                  * if we have empty groups. If so, take one from there and
 474                  * put in the partially used list.
 475                  */
 476                 if (!uctxt->tid_used_list.count || need_group) {
 477                         if (!uctxt->tid_group_list.count)
 478                                 goto unlock;
 479
 480                         grp = tid_group_pop(&uctxt->tid_group_list);
 481                         tid_group_add_tail(grp, &uctxt->tid_used_list);
 482                         need_group = 0;
 483                 }
 484                 /*
 485                  * There is an optimization opportunity here - instead of
 486                  * fitting as many page sets as we can, check for a group
 487                  * later on in the list that could fit all of them.
 488                  */
 489                 list_for_each_entry_safe(grp, ptr, &uctxt->tid_used_list.list,
 490                                          list) {
 491                         unsigned use = min_t(unsigned, pageset_count - pageidx,
 492                                              grp->size - grp->used);
 493
 494                         ret = program_rcvarray(fp, vaddr, grp, pagesets,
 495                                                pageidx, use, pages, tidlist,
 496                                                &tididx, &mapped);
 497                         if (ret < 0) {
 498                                 hfi1_cdbg(TID,
 499                                           "Failed to program RcvArray entries %d",
 500                                           ret);
 501                                 ret = -EFAULT;
 502                                 goto unlock;
 503                         } else if (ret > 0) {
 504                                 if (grp->used == grp->size)
 505                                         tid_group_move(grp,
 506                                                        &uctxt->tid_used_list,
 507                                                        &uctxt->tid_full_list);
 508                                 pageidx += ret;
 509                                 mapped_pages += mapped;
 510                                 need_group = 0;
 511                                 /* Check if we are done so we break out early */
 512                                 if (pageidx >= pageset_count)
 513                                         break;
 514                         } else if (WARN_ON(ret == 0)) {
 515                                 /*
 516                                  * If ret is 0, we did not program any entries
 517                                  * into this group, which can only happen if
 518                                  * we've screwed up the accounting somewhere.
 519                                  * Warn and try to continue.
 520                                  */
 521                                 need_group = 1;
 522                         }
 523                 }
 524         }
 525 unlock:
 526         mutex_unlock(&uctxt->exp_lock);
 527 nomem:
 528         hfi1_cdbg(TID, "total mapped: tidpairs:%u pages:%u (%d)", tididx,
 529                   mapped_pages, ret);
 530         if (tididx) {
 531                 spin_lock(&fd->tid_lock);
 532                 fd->tid_used += tididx;
 533                 spin_unlock(&fd->tid_lock);
 534                 tinfo->tidcnt = tididx;
 535                 tinfo->length = mapped_pages * PAGE_SIZE;
 536
 537                 if (copy_to_user((void __user *)(unsigned long)tinfo->tidlist,
 538                                  tidlist, sizeof(tidlist[0]) * tididx)) {
 539                         /*
 540                          * On failure to copy to the user level, we need to undo
 541                          * everything done so far so we don't leak resources.
 542                          */
 543                         tinfo->tidlist = (unsigned long)&tidlist;
 544                         hfi1_user_exp_rcv_clear(fp, tinfo);
 545                         tinfo->tidlist = 0;
 546                         ret = -EFAULT;
 547                         goto bail;
 548                 }
 549         }
 550
 551         /*
 552          * If not everything was mapped (due to insufficient RcvArray entries,
 553          * for example), unpin all unmapped pages so we can pin them nex time.
 554          */
 555         if (mapped_pages != pinned) {
 556                 hfi1_release_user_pages(current->mm, &pages[mapped_pages],
 557                                         pinned - mapped_pages,
 558                                         false);
 559                 fd->tid_n_pinned -= pinned - mapped_pages;
 560         }
 561 bail:
 562         kfree(pagesets);
 563         kfree(pages);
 564         kfree(tidlist);
 565         return ret > 0 ? 0 : ret;
 566 }
 567
 568 int hfi1_user_exp_rcv_clear(struct file *fp, struct hfi1_tid_info *tinfo)
 569 {
 570         int ret = 0;
 571         struct hfi1_filedata *fd = fp->private_data;
 572         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 573         u32 *tidinfo;
 574         unsigned tididx;
 575
 576         tidinfo = kcalloc(tinfo->tidcnt, sizeof(*tidinfo), GFP_KERNEL);
 577         if (!tidinfo)
 578                 return -ENOMEM;
 579
 580         if (copy_from_user(tidinfo, (void __user *)(unsigned long)
 581                            tinfo->tidlist, sizeof(tidinfo[0]) *
 582                            tinfo->tidcnt)) {
 583                 ret = -EFAULT;
 584                 goto done;
 585         }
 586
 587         mutex_lock(&uctxt->exp_lock);
 588         for (tididx = 0; tididx < tinfo->tidcnt; tididx++) {
 589                 ret = unprogram_rcvarray(fp, tidinfo[tididx], NULL);
 590                 if (ret) {
 591                         hfi1_cdbg(TID, "Failed to unprogram rcv array %d",
 592                                   ret);
 593                         break;
 594                 }
 595         }
 596         spin_lock(&fd->tid_lock);
 597         fd->tid_used -= tididx;
 598         spin_unlock(&fd->tid_lock);
 599         tinfo->tidcnt = tididx;
 600         mutex_unlock(&uctxt->exp_lock);
 601 done:
 602         kfree(tidinfo);
 603         return ret;
 604 }
 605
 606 int hfi1_user_exp_rcv_invalid(struct file *fp, struct hfi1_tid_info *tinfo)
 607 {
 608         struct hfi1_filedata *fd = fp->private_data;
 609         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 610         unsigned long *ev = uctxt->dd->events +
 611                 (((uctxt->ctxt - uctxt->dd->first_user_ctxt) *
 612                   HFI1_MAX_SHARED_CTXTS) + fd->subctxt);
 613         u32 *array;
 614         int ret = 0;
 615
 616         if (!fd->invalid_tids)
 617                 return -EINVAL;
 618
 619         /*
 620          * copy_to_user() can sleep, which will leave the invalid_lock
 621          * locked and cause the MMU notifier to be blocked on the lock
 622          * for a long time.
 623          * Copy the data to a local buffer so we can release the lock.
 624          */
 625         array = kcalloc(uctxt->expected_count, sizeof(*array), GFP_KERNEL);
 626         if (!array)
 627                 return -EFAULT;
 628
 629         spin_lock(&fd->invalid_lock);
 630         if (fd->invalid_tid_idx) {
 631                 memcpy(array, fd->invalid_tids, sizeof(*array) *
 632                        fd->invalid_tid_idx);
 633                 memset(fd->invalid_tids, 0, sizeof(*fd->invalid_tids) *
 634                        fd->invalid_tid_idx);
 635                 tinfo->tidcnt = fd->invalid_tid_idx;
 636                 fd->invalid_tid_idx = 0;
 637                 /*
 638                  * Reset the user flag while still holding the lock.
 639                  * Otherwise, PSM can miss events.
 640                  */
 641                 clear_bit(_HFI1_EVENT_TID_MMU_NOTIFY_BIT, ev);
 642         } else {
 643                 tinfo->tidcnt = 0;
 644         }
 645         spin_unlock(&fd->invalid_lock);
 646
 647         if (tinfo->tidcnt) {
 648                 if (copy_to_user((void __user *)tinfo->tidlist,
 649                                  array, sizeof(*array) * tinfo->tidcnt))
 650                         ret = -EFAULT;
 651         }
 652         kfree(array);
 653
 654         return ret;
 655 }
 656
 657 static u32 find_phys_blocks(struct page **pages, unsigned npages,
 658                             struct tid_pageset *list)
 659 {
 660         unsigned pagecount, pageidx, setcount = 0, i;
 661         unsigned long pfn, this_pfn;
 662
 663         if (!npages)
 664                 return 0;
 665
 666         /*
 667          * Look for sets of physically contiguous pages in the user buffer.
 668          * This will allow us to optimize Expected RcvArray entry usage by
 669          * using the bigger supported sizes.
 670          */
 671         pfn = page_to_pfn(pages[0]);
 672         for (pageidx = 0, pagecount = 1, i = 1; i <= npages; i++) {
 673                 this_pfn = i < npages ? page_to_pfn(pages[i]) : 0;
 674
 675                 /*
 676                  * If the pfn's are not sequential, pages are not physically
 677                  * contiguous.
 678                  */
 679                 if (this_pfn != ++pfn) {
 680                         /*
 681                          * At this point we have to loop over the set of
 682                          * physically contiguous pages and break them down it
 683                          * sizes supported by the HW.
 684                          * There are two main constraints:
 685                          *     1. The max buffer size is MAX_EXPECTED_BUFFER.
 686                          *        If the total set size is bigger than that
 687                          *        program only a MAX_EXPECTED_BUFFER chunk.
 688                          *     2. The buffer size has to be a power of two. If
 689                          *        it is not, round down to the closes power of
 690                          *        2 and program that size.
 691                          */
 692                         while (pagecount) {
 693                                 int maxpages = pagecount;
 694                                 u32 bufsize = pagecount * PAGE_SIZE;
 695
 696                                 if (bufsize > MAX_EXPECTED_BUFFER)
 697                                         maxpages =
 698                                                 MAX_EXPECTED_BUFFER >>
 699                                                 PAGE_SHIFT;
 700                                 else if (!is_power_of_2(bufsize))
 701                                         maxpages =
 702                                                 rounddown_pow_of_two(bufsize) >>
 703                                                 PAGE_SHIFT;
 704
 705                                 list[setcount].idx = pageidx;
 706                                 list[setcount].count = maxpages;
 707                                 pagecount -= maxpages;
 708                                 pageidx += maxpages;
 709                                 setcount++;
 710                         }
 711                         pageidx = i;
 712                         pagecount = 1;
 713                         pfn = this_pfn;
 714                 } else {
 715                         pagecount++;
 716                 }
 717         }
 718         return setcount;
 719 }
 720
 721 /**
 722  * program_rcvarray() - program an RcvArray group with receive buffers
 723  * @fp: file pointer
 724  * @vaddr: starting user virtual address
 725  * @grp: RcvArray group
 726  * @sets: array of struct tid_pageset holding information on physically
 727  *        contiguous chunks from the user buffer
 728  * @start: starting index into sets array
 729  * @count: number of struct tid_pageset's to program
 730  * @pages: an array of struct page * for the user buffer
 731  * @tidlist: the array of u32 elements when the information about the
 732  *           programmed RcvArray entries is to be encoded.
 733  * @tididx: starting offset into tidlist
 734  * @pmapped: (output parameter) number of pages programmed into the RcvArray
 735  *           entries.
 736  *
 737  * This function will program up to 'count' number of RcvArray entries from the
 738  * group 'grp'. To make best use of write-combining writes, the function will
 739  * perform writes to the unused RcvArray entries which will be ignored by the
 740  * HW. Each RcvArray entry will be programmed with a physically contiguous
 741  * buffer chunk from the user's virtual buffer.
 742  *
 743  * Return:
 744  * -EINVAL if the requested count is larger than the size of the group,
 745  * -ENOMEM or -EFAULT on error from set_rcvarray_entry(), or
 746  * number of RcvArray entries programmed.
 747  */
 748 static int program_rcvarray(struct file *fp, unsigned long vaddr,
 749                             struct tid_group *grp,
 750                             struct tid_pageset *sets,
 751                             unsigned start, u16 count, struct page **pages,
 752                             u32 *tidlist, unsigned *tididx, unsigned *pmapped)
 753 {
 754         struct hfi1_filedata *fd = fp->private_data;
 755         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 756         struct hfi1_devdata *dd = uctxt->dd;
 757         u16 idx;
 758         u32 tidinfo = 0, rcventry, useidx = 0;
 759         int mapped = 0;
 760
 761         /* Count should never be larger than the group size */
 762         if (count > grp->size)
 763                 return -EINVAL;
 764
 765         /* Find the first unused entry in the group */
 766         for (idx = 0; idx < grp->size; idx++) {
 767                 if (!(grp->map & (1 << idx))) {
 768                         useidx = idx;
 769                         break;
 770                 }
 771                 rcv_array_wc_fill(dd, grp->base + idx);
 772         }
 773
 774         idx = 0;
 775         while (idx < count) {
 776                 u16 npages, pageidx, setidx = start + idx;
 777                 int ret = 0;
 778
 779                 /*
 780                  * If this entry in the group is used, move to the next one.
 781                  * If we go past the end of the group, exit the loop.
 782                  */
 783                 if (useidx >= grp->size) {
 784                         break;
 785                 } else if (grp->map & (1 << useidx)) {
 786                         rcv_array_wc_fill(dd, grp->base + useidx);
 787                         useidx++;
 788                         continue;
 789                 }
 790
 791                 rcventry = grp->base + useidx;
 792                 npages = sets[setidx].count;
 793                 pageidx = sets[setidx].idx;
 794
 795                 ret = set_rcvarray_entry(fp, vaddr + (pageidx * PAGE_SIZE),
 796                                          rcventry, grp, pages + pageidx,
 797                                          npages);
 798                 if (ret)
 799                         return ret;
 800                 mapped += npages;
 801
 802                 tidinfo = rcventry2tidinfo(rcventry - uctxt->expected_base) |
 803                         EXP_TID_SET(LEN, npages);
 804                 tidlist[(*tididx)++] = tidinfo;
 805                 grp->used++;
 806                 grp->map |= 1 << useidx++;
 807                 idx++;
 808         }
 809
 810         /* Fill the rest of the group with "blank" writes */
 811         for (; useidx < grp->size; useidx++)
 812                 rcv_array_wc_fill(dd, grp->base + useidx);
 813         *pmapped = mapped;
 814         return idx;
 815 }
 816
 817 static int set_rcvarray_entry(struct file *fp, unsigned long vaddr,
 818                               u32 rcventry, struct tid_group *grp,
 819                               struct page **pages, unsigned npages)
 820 {
 821         int ret;
 822         struct hfi1_filedata *fd = fp->private_data;
 823         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 824         struct tid_rb_node *node;
 825         struct hfi1_devdata *dd = uctxt->dd;
 826         struct rb_root *root = &fd->tid_rb_root;
 827         dma_addr_t phys;
 828
 829         /*
 830          * Allocate the node first so we can handle a potential
 831          * failure before we've programmed anything.
 832          */
 833         node = kzalloc(sizeof(*node) + (sizeof(struct page *) * npages),
 834                        GFP_KERNEL);
 835         if (!node)
 836                 return -ENOMEM;
 837
 838         phys = pci_map_single(dd->pcidev,
 839                               __va(page_to_phys(pages[0])),
 840                               npages * PAGE_SIZE, PCI_DMA_FROMDEVICE);
 841         if (dma_mapping_error(&dd->pcidev->dev, phys)) {
 842                 dd_dev_err(dd, "Failed to DMA map Exp Rcv pages 0x%llx\n",
 843                            phys);
 844                 kfree(node);
 845                 return -EFAULT;
 846         }
 847
 848         node->mmu.addr = vaddr;
 849         node->mmu.len = npages * PAGE_SIZE;
 850         node->phys = page_to_phys(pages[0]);
 851         node->npages = npages;
 852         node->rcventry = rcventry;
 853         node->dma_addr = phys;
 854         node->grp = grp;
 855         node->freed = false;
 856         memcpy(node->pages, pages, sizeof(struct page *) * npages);
 857
 858         if (HFI1_CAP_IS_USET(TID_UNMAP))
 859                 ret = mmu_rb_insert(root, &node->mmu);
 860         else
 861                 ret = hfi1_mmu_rb_insert(root, &node->mmu);
 862
 863         if (ret) {
 864                 hfi1_cdbg(TID, "Failed to insert RB node %u 0x%lx, 0x%lx %d",
 865                           node->rcventry, node->mmu.addr, node->phys, ret);
 866                 pci_unmap_single(dd->pcidev, phys, npages * PAGE_SIZE,
 867                                  PCI_DMA_FROMDEVICE);
 868                 kfree(node);
 869                 return -EFAULT;
 870         }
 871         hfi1_put_tid(dd, rcventry, PT_EXPECTED, phys, ilog2(npages) + 1);
 872         trace_hfi1_exp_tid_reg(uctxt->ctxt, fd->subctxt, rcventry, npages,
 873                                node->mmu.addr, node->phys, phys);
 874         return 0;
 875 }
 876
 877 static int unprogram_rcvarray(struct file *fp, u32 tidinfo,
 878                               struct tid_group **grp)
 879 {
 880         struct hfi1_filedata *fd = fp->private_data;
 881         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 882         struct hfi1_devdata *dd = uctxt->dd;
 883         struct tid_rb_node *node;
 884         u8 tidctrl = EXP_TID_GET(tidinfo, CTRL);
 885         u32 tididx = EXP_TID_GET(tidinfo, IDX) << 1, rcventry;
 886
 887         if (tididx >= uctxt->expected_count) {
 888                 dd_dev_err(dd, "Invalid RcvArray entry (%u) index for ctxt %u\n",
 889                            tididx, uctxt->ctxt);
 890                 return -EINVAL;
 891         }
 892
 893         if (tidctrl == 0x3)
 894                 return -EINVAL;
 895
 896         rcventry = tididx + (tidctrl - 1);
 897
 898         node = fd->entry_to_rb[rcventry];
 899         if (!node || node->rcventry != (uctxt->expected_base + rcventry))
 900                 return -EBADF;
 901         if (HFI1_CAP_IS_USET(TID_UNMAP))
 902                 mmu_rb_remove(&fd->tid_rb_root, &node->mmu, false);
 903         else
 904                 hfi1_mmu_rb_remove(&fd->tid_rb_root, &node->mmu);
 905
 906         if (grp)
 907                 *grp = node->grp;
 908         clear_tid_node(fd, fd->subctxt, node);
 909         return 0;
 910 }
 911
 912 static void clear_tid_node(struct hfi1_filedata *fd, u16 subctxt,
 913                            struct tid_rb_node *node)
 914 {
 915         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 916         struct hfi1_devdata *dd = uctxt->dd;
 917
 918         trace_hfi1_exp_tid_unreg(uctxt->ctxt, fd->subctxt, node->rcventry,
 919                                  node->npages, node->mmu.addr, node->phys,
 920                                  node->dma_addr);
 921
 922         hfi1_put_tid(dd, node->rcventry, PT_INVALID, 0, 0);
 923         /*
 924          * Make sure device has seen the write before we unpin the
 925          * pages.
 926          */
 927         flush_wc();
 928
 929         pci_unmap_single(dd->pcidev, node->dma_addr, node->mmu.len,
 930                          PCI_DMA_FROMDEVICE);
 931         hfi1_release_user_pages(current->mm, node->pages, node->npages, true);
 932         fd->tid_n_pinned -= node->npages;
 933
 934         node->grp->used--;
 935         node->grp->map &= ~(1 << (node->rcventry - node->grp->base));
 936
 937         if (node->grp->used == node->grp->size - 1)
 938                 tid_group_move(node->grp, &uctxt->tid_full_list,
 939                                &uctxt->tid_used_list);
 940         else if (!node->grp->used)
 941                 tid_group_move(node->grp, &uctxt->tid_used_list,
 942                                &uctxt->tid_group_list);
 943         kfree(node);
 944 }
 945
 946 static void unlock_exp_tids(struct hfi1_ctxtdata *uctxt,
 947                             struct exp_tid_set *set, struct rb_root *root)
 948 {
 949         struct tid_group *grp, *ptr;
 950         struct hfi1_filedata *fd = container_of(root, struct hfi1_filedata,
 951                                                 tid_rb_root);
 952         int i;
 953
 954         list_for_each_entry_safe(grp, ptr, &set->list, list) {
 955                 list_del_init(&grp->list);
 956
 957                 for (i = 0; i < grp->size; i++) {
 958                         if (grp->map & (1 << i)) {
 959                                 u16 rcventry = grp->base + i;
 960                                 struct tid_rb_node *node;
 961
 962                                 node = fd->entry_to_rb[rcventry -
 963                                                           uctxt->expected_base];
 964                                 if (!node || node->rcventry != rcventry)
 965                                         continue;
 966                                 if (HFI1_CAP_IS_USET(TID_UNMAP))
 967                                         mmu_rb_remove(&fd->tid_rb_root,
 968                                                       &node->mmu, false);
 969                                 else
 970                                         hfi1_mmu_rb_remove(&fd->tid_rb_root,
 971                                                            &node->mmu);
 972                                 clear_tid_node(fd, -1, node);
 973                         }
 974                 }
 975         }
 976 }
 977
 978 static int mmu_rb_invalidate(struct rb_root *root, struct mmu_rb_node *mnode)
 979 {
 980         struct hfi1_filedata *fdata =
 981                 container_of(root, struct hfi1_filedata, tid_rb_root);
 982         struct hfi1_ctxtdata *uctxt = fdata->uctxt;
 983         struct tid_rb_node *node =
 984                 container_of(mnode, struct tid_rb_node, mmu);
 985
 986         if (node->freed)
 987                 return 0;
 988
 989         trace_hfi1_exp_tid_inval(uctxt->ctxt, fdata->subctxt, node->mmu.addr,
 990                                  node->rcventry, node->npages, node->dma_addr);
 991         node->freed = true;
 992
 993         spin_lock(&fdata->invalid_lock);
 994         if (fdata->invalid_tid_idx < uctxt->expected_count) {
 995                 fdata->invalid_tids[fdata->invalid_tid_idx] =
 996                         rcventry2tidinfo(node->rcventry - uctxt->expected_base);
 997                 fdata->invalid_tids[fdata->invalid_tid_idx] |=
 998                         EXP_TID_SET(LEN, node->npages);
 999                 if (!fdata->invalid_tid_idx) {
1000                         unsigned long *ev;
1001
1002                         /*
1003                          * hfi1_set_uevent_bits() sets a user event flag
1004                          * for all processes. Because calling into the
1005                          * driver to process TID cache invalidations is
1006                          * expensive and TID cache invalidations are
1007                          * handled on a per-process basis, we can
1008                          * optimize this to set the flag only for the
1009                          * process in question.
1010                          */
1011                         ev = uctxt->dd->events +
1012                                 (((uctxt->ctxt - uctxt->dd->first_user_ctxt) *
1013                                   HFI1_MAX_SHARED_CTXTS) + fdata->subctxt);
1014                         set_bit(_HFI1_EVENT_TID_MMU_NOTIFY_BIT, ev);
1015                 }
1016                 fdata->invalid_tid_idx++;
1017         }
1018         spin_unlock(&fdata->invalid_lock);
1019         return 0;
1020 }
1021
1022 static int mmu_rb_insert(struct rb_root *root, struct mmu_rb_node *node)
1023 {
1024         struct hfi1_filedata *fdata =
1025                 container_of(root, struct hfi1_filedata, tid_rb_root);
1026         struct tid_rb_node *tnode =
1027                 container_of(node, struct tid_rb_node, mmu);
1028         u32 base = fdata->uctxt->expected_base;
1029
1030         fdata->entry_to_rb[tnode->rcventry - base] = tnode;
1031         return 0;
1032 }
1033
1034 static void mmu_rb_remove(struct rb_root *root, struct mmu_rb_node *node,
1035                           bool notifier)
1036 {
1037         struct hfi1_filedata *fdata =
1038                 container_of(root, struct hfi1_filedata, tid_rb_root);
1039         struct tid_rb_node *tnode =
1040                 container_of(node, struct tid_rb_node, mmu);
1041         u32 base = fdata->uctxt->expected_base;
1042
1043         fdata->entry_to_rb[tnode->rcventry - base] = NULL;
1044 }