drivers/infiniband/hw/hfi1/user_exp_rcv.c

   1 /*
   2  * Copyright(c) 2015, 2016 Intel Corporation.
   3  *
   4  * This file is provided under a dual BSD/GPLv2 license.  When using or
   5  * redistributing this file, you may do so under either license.
   6  *
   7  * GPL LICENSE SUMMARY
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of version 2 of the GNU General Public License as
  11  * published by the Free Software Foundation.
  12  *
  13  * This program is distributed in the hope that it will be useful, but
  14  * WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * General Public License for more details.
  17  *
  18  * BSD LICENSE
  19  *
  20  * Redistribution and use in source and binary forms, with or without
  21  * modification, are permitted provided that the following conditions
  22  * are met:
  23  *
  24  *  - Redistributions of source code must retain the above copyright
  25  *    notice, this list of conditions and the following disclaimer.
  26  *  - Redistributions in binary form must reproduce the above copyright
  27  *    notice, this list of conditions and the following disclaimer in
  28  *    the documentation and/or other materials provided with the
  29  *    distribution.
  30  *  - Neither the name of Intel Corporation nor the names of its
  31  *    contributors may be used to endorse or promote products derived
  32  *    from this software without specific prior written permission.
  33  *
  34  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  35  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  36  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  37  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  38  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  39  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  40  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  41  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  42  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  43  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  44  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  45  *
  46  */
  47 #include <asm/page.h>
  48
  49 #include "user_exp_rcv.h"
  50 #include "trace.h"
  51 #include "mmu_rb.h"
  52
  53 struct tid_group {
  54         struct list_head list;
  55         unsigned base;
  56         u8 size;
  57         u8 used;
  58         u8 map;
  59 };
  60
  61 struct tid_rb_node {
  62         struct mmu_rb_node mmu;
  63         unsigned long phys;
  64         struct tid_group *grp;
  65         u32 rcventry;
  66         dma_addr_t dma_addr;
  67         bool freed;
  68         unsigned npages;
  69         struct page *pages[0];
  70 };
  71
  72 struct tid_pageset {
  73         u16 idx;
  74         u16 count;
  75 };
  76
  77 #define EXP_TID_SET_EMPTY(set) (set.count == 0 && list_empty(&set.list))
  78
  79 #define num_user_pages(vaddr, len)                                     \
  80         (1 + (((((unsigned long)(vaddr) +                              \
  81                  (unsigned long)(len) - 1) & PAGE_MASK) -              \
  82                ((unsigned long)vaddr & PAGE_MASK)) >> PAGE_SHIFT))
  83
  84 static void unlock_exp_tids(struct hfi1_ctxtdata *, struct exp_tid_set *,
  85                             struct hfi1_filedata *);
  86 static u32 find_phys_blocks(struct page **, unsigned, struct tid_pageset *);
  87 static int set_rcvarray_entry(struct file *, unsigned long, u32,
  88                               struct tid_group *, struct page **, unsigned);
  89 static int tid_rb_insert(void *, struct mmu_rb_node *);
  90 static void cacheless_tid_rb_remove(struct hfi1_filedata *fdata,
  91                                     struct tid_rb_node *tnode);
  92 static void tid_rb_remove(void *, struct mmu_rb_node *);
  93 static int tid_rb_invalidate(void *, struct mmu_rb_node *);
  94 static int program_rcvarray(struct file *, unsigned long, struct tid_group *,
  95                             struct tid_pageset *, unsigned, u16, struct page **,
  96                             u32 *, unsigned *, unsigned *);
  97 static int unprogram_rcvarray(struct file *, u32, struct tid_group **);
  98 static void clear_tid_node(struct hfi1_filedata *fd, struct tid_rb_node *node);
  99
 100 static struct mmu_rb_ops tid_rb_ops = {
 101         .insert = tid_rb_insert,
 102         .remove = tid_rb_remove,
 103         .invalidate = tid_rb_invalidate
 104 };
 105
 106 static inline u32 rcventry2tidinfo(u32 rcventry)
 107 {
 108         u32 pair = rcventry & ~0x1;
 109
 110         return EXP_TID_SET(IDX, pair >> 1) |
 111                 EXP_TID_SET(CTRL, 1 << (rcventry - pair));
 112 }
 113
 114 static inline void exp_tid_group_init(struct exp_tid_set *set)
 115 {
 116         INIT_LIST_HEAD(&set->list);
 117         set->count = 0;
 118 }
 119
 120 static inline void tid_group_remove(struct tid_group *grp,
 121                                     struct exp_tid_set *set)
 122 {
 123         list_del_init(&grp->list);
 124         set->count--;
 125 }
 126
 127 static inline void tid_group_add_tail(struct tid_group *grp,
 128                                       struct exp_tid_set *set)
 129 {
 130         list_add_tail(&grp->list, &set->list);
 131         set->count++;
 132 }
 133
 134 static inline struct tid_group *tid_group_pop(struct exp_tid_set *set)
 135 {
 136         struct tid_group *grp =
 137                 list_first_entry(&set->list, struct tid_group, list);
 138         list_del_init(&grp->list);
 139         set->count--;
 140         return grp;
 141 }
 142
 143 static inline void tid_group_move(struct tid_group *group,
 144                                   struct exp_tid_set *s1,
 145                                   struct exp_tid_set *s2)
 146 {
 147         tid_group_remove(group, s1);
 148         tid_group_add_tail(group, s2);
 149 }
 150
 151 /*
 152  * Initialize context and file private data needed for Expected
 153  * receive caching. This needs to be done after the context has
 154  * been configured with the eager/expected RcvEntry counts.
 155  */
 156 int hfi1_user_exp_rcv_init(struct file *fp)
 157 {
 158         struct hfi1_filedata *fd = fp->private_data;
 159         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 160         struct hfi1_devdata *dd = uctxt->dd;
 161         unsigned tidbase;
 162         int i, ret = 0;
 163
 164         spin_lock_init(&fd->tid_lock);
 165         spin_lock_init(&fd->invalid_lock);
 166
 167         if (!uctxt->subctxt_cnt || !fd->subctxt) {
 168                 exp_tid_group_init(&uctxt->tid_group_list);
 169                 exp_tid_group_init(&uctxt->tid_used_list);
 170                 exp_tid_group_init(&uctxt->tid_full_list);
 171
 172                 tidbase = uctxt->expected_base;
 173                 for (i = 0; i < uctxt->expected_count /
 174                              dd->rcv_entries.group_size; i++) {
 175                         struct tid_group *grp;
 176
 177                         grp = kzalloc(sizeof(*grp), GFP_KERNEL);
 178                         if (!grp) {
 179                                 /*
 180                                  * If we fail here, the groups already
 181                                  * allocated will be freed by the close
 182                                  * call.
 183                                  */
 184                                 ret = -ENOMEM;
 185                                 goto done;
 186                         }
 187                         grp->size = dd->rcv_entries.group_size;
 188                         grp->base = tidbase;
 189                         tid_group_add_tail(grp, &uctxt->tid_group_list);
 190                         tidbase += dd->rcv_entries.group_size;
 191                 }
 192         }
 193
 194         fd->entry_to_rb = kcalloc(uctxt->expected_count,
 195                                      sizeof(struct rb_node *),
 196                                      GFP_KERNEL);
 197         if (!fd->entry_to_rb)
 198                 return -ENOMEM;
 199
 200         if (!HFI1_CAP_UGET_MASK(uctxt->flags, TID_UNMAP)) {
 201                 fd->invalid_tid_idx = 0;
 202                 fd->invalid_tids = kzalloc(uctxt->expected_count *
 203                                            sizeof(u32), GFP_KERNEL);
 204                 if (!fd->invalid_tids) {
 205                         ret = -ENOMEM;
 206                         goto done;
 207                 }
 208
 209                 /*
 210                  * Register MMU notifier callbacks. If the registration
 211                  * fails, continue without TID caching for this context.
 212                  */
 213                 ret = hfi1_mmu_rb_register(fd, fd->mm, &tid_rb_ops,
 214                                            dd->pport->hfi1_wq,
 215                                            &fd->handler);
 216                 if (ret) {
 217                         dd_dev_info(dd,
 218                                     "Failed MMU notifier registration %d\n",
 219                                     ret);
 220                         ret = 0;
 221                 }
 222         }
 223
 224         /*
 225          * PSM does not have a good way to separate, count, and
 226          * effectively enforce a limit on RcvArray entries used by
 227          * subctxts (when context sharing is used) when TID caching
 228          * is enabled. To help with that, we calculate a per-process
 229          * RcvArray entry share and enforce that.
 230          * If TID caching is not in use, PSM deals with usage on its
 231          * own. In that case, we allow any subctxt to take all of the
 232          * entries.
 233          *
 234          * Make sure that we set the tid counts only after successful
 235          * init.
 236          */
 237         spin_lock(&fd->tid_lock);
 238         if (uctxt->subctxt_cnt && fd->handler) {
 239                 u16 remainder;
 240
 241                 fd->tid_limit = uctxt->expected_count / uctxt->subctxt_cnt;
 242                 remainder = uctxt->expected_count % uctxt->subctxt_cnt;
 243                 if (remainder && fd->subctxt < remainder)
 244                         fd->tid_limit++;
 245         } else {
 246                 fd->tid_limit = uctxt->expected_count;
 247         }
 248         spin_unlock(&fd->tid_lock);
 249 done:
 250         return ret;
 251 }
 252
 253 int hfi1_user_exp_rcv_free(struct hfi1_filedata *fd)
 254 {
 255         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 256         struct tid_group *grp, *gptr;
 257
 258         if (!test_bit(HFI1_CTXT_SETUP_DONE, &uctxt->event_flags))
 259                 return 0;
 260         /*
 261          * The notifier would have been removed when the process'es mm
 262          * was freed.
 263          */
 264         if (fd->handler)
 265                 hfi1_mmu_rb_unregister(fd->handler);
 266
 267         kfree(fd->invalid_tids);
 268
 269         if (!uctxt->cnt) {
 270                 if (!EXP_TID_SET_EMPTY(uctxt->tid_full_list))
 271                         unlock_exp_tids(uctxt, &uctxt->tid_full_list, fd);
 272                 if (!EXP_TID_SET_EMPTY(uctxt->tid_used_list))
 273                         unlock_exp_tids(uctxt, &uctxt->tid_used_list, fd);
 274                 list_for_each_entry_safe(grp, gptr, &uctxt->tid_group_list.list,
 275                                          list) {
 276                         list_del_init(&grp->list);
 277                         kfree(grp);
 278                 }
 279                 hfi1_clear_tids(uctxt);
 280         }
 281
 282         kfree(fd->entry_to_rb);
 283         return 0;
 284 }
 285
 286 /*
 287  * Write an "empty" RcvArray entry.
 288  * This function exists so the TID registaration code can use it
 289  * to write to unused/unneeded entries and still take advantage
 290  * of the WC performance improvements. The HFI will ignore this
 291  * write to the RcvArray entry.
 292  */
 293 static inline void rcv_array_wc_fill(struct hfi1_devdata *dd, u32 index)
 294 {
 295         /*
 296          * Doing the WC fill writes only makes sense if the device is
 297          * present and the RcvArray has been mapped as WC memory.
 298          */
 299         if ((dd->flags & HFI1_PRESENT) && dd->rcvarray_wc)
 300                 writeq(0, dd->rcvarray_wc + (index * 8));
 301 }
 302
 303 /*
 304  * RcvArray entry allocation for Expected Receives is done by the
 305  * following algorithm:
 306  *
 307  * The context keeps 3 lists of groups of RcvArray entries:
 308  *   1. List of empty groups - tid_group_list
 309  *      This list is created during user context creation and
 310  *      contains elements which describe sets (of 8) of empty
 311  *      RcvArray entries.
 312  *   2. List of partially used groups - tid_used_list
 313  *      This list contains sets of RcvArray entries which are
 314  *      not completely used up. Another mapping request could
 315  *      use some of all of the remaining entries.
 316  *   3. List of full groups - tid_full_list
 317  *      This is the list where sets that are completely used
 318  *      up go.
 319  *
 320  * An attempt to optimize the usage of RcvArray entries is
 321  * made by finding all sets of physically contiguous pages in a
 322  * user's buffer.
 323  * These physically contiguous sets are further split into
 324  * sizes supported by the receive engine of the HFI. The
 325  * resulting sets of pages are stored in struct tid_pageset,
 326  * which describes the sets as:
 327  *    * .count - number of pages in this set
 328  *    * .idx - starting index into struct page ** array
 329  *                    of this set
 330  *
 331  * From this point on, the algorithm deals with the page sets
 332  * described above. The number of pagesets is divided by the
 333  * RcvArray group size to produce the number of full groups
 334  * needed.
 335  *
 336  * Groups from the 3 lists are manipulated using the following
 337  * rules:
 338  *   1. For each set of 8 pagesets, a complete group from
 339  *      tid_group_list is taken, programmed, and moved to
 340  *      the tid_full_list list.
 341  *   2. For all remaining pagesets:
 342  *      2.1 If the tid_used_list is empty and the tid_group_list
 343  *          is empty, stop processing pageset and return only
 344  *          what has been programmed up to this point.
 345  *      2.2 If the tid_used_list is empty and the tid_group_list
 346  *          is not empty, move a group from tid_group_list to
 347  *          tid_used_list.
 348  *      2.3 For each group is tid_used_group, program as much as
 349  *          can fit into the group. If the group becomes fully
 350  *          used, move it to tid_full_list.
 351  */
 352 int hfi1_user_exp_rcv_setup(struct file *fp, struct hfi1_tid_info *tinfo)
 353 {
 354         int ret = 0, need_group = 0, pinned;
 355         struct hfi1_filedata *fd = fp->private_data;
 356         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 357         struct hfi1_devdata *dd = uctxt->dd;
 358         unsigned npages, ngroups, pageidx = 0, pageset_count, npagesets,
 359                 tididx = 0, mapped, mapped_pages = 0;
 360         unsigned long vaddr = tinfo->vaddr;
 361         struct page **pages = NULL;
 362         u32 *tidlist = NULL;
 363         struct tid_pageset *pagesets = NULL;
 364
 365         /* Get the number of pages the user buffer spans */
 366         npages = num_user_pages(vaddr, tinfo->length);
 367         if (!npages)
 368                 return -EINVAL;
 369
 370         if (npages > uctxt->expected_count) {
 371                 dd_dev_err(dd, "Expected buffer too big\n");
 372                 return -EINVAL;
 373         }
 374
 375         /* Verify that access is OK for the user buffer */
 376         if (!access_ok(VERIFY_WRITE, (void __user *)vaddr,
 377                        npages * PAGE_SIZE)) {
 378                 dd_dev_err(dd, "Fail vaddr %p, %u pages, !access_ok\n",
 379                            (void *)vaddr, npages);
 380                 return -EFAULT;
 381         }
 382
 383         pagesets = kcalloc(uctxt->expected_count, sizeof(*pagesets),
 384                            GFP_KERNEL);
 385         if (!pagesets)
 386                 return -ENOMEM;
 387
 388         /* Allocate the array of struct page pointers needed for pinning */
 389         pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL);
 390         if (!pages) {
 391                 ret = -ENOMEM;
 392                 goto bail;
 393         }
 394
 395         /*
 396          * Pin all the pages of the user buffer. If we can't pin all the
 397          * pages, accept the amount pinned so far and program only that.
 398          * User space knows how to deal with partially programmed buffers.
 399          */
 400         if (!hfi1_can_pin_pages(dd, fd->mm, fd->tid_n_pinned, npages)) {
 401                 ret = -ENOMEM;
 402                 goto bail;
 403         }
 404
 405         pinned = hfi1_acquire_user_pages(fd->mm, vaddr, npages, true, pages);
 406         if (pinned <= 0) {
 407                 ret = pinned;
 408                 goto bail;
 409         }
 410         fd->tid_n_pinned += npages;
 411
 412         /* Find sets of physically contiguous pages */
 413         npagesets = find_phys_blocks(pages, pinned, pagesets);
 414
 415         /*
 416          * We don't need to access this under a lock since tid_used is per
 417          * process and the same process cannot be in hfi1_user_exp_rcv_clear()
 418          * and hfi1_user_exp_rcv_setup() at the same time.
 419          */
 420         spin_lock(&fd->tid_lock);
 421         if (fd->tid_used + npagesets > fd->tid_limit)
 422                 pageset_count = fd->tid_limit - fd->tid_used;
 423         else
 424                 pageset_count = npagesets;
 425         spin_unlock(&fd->tid_lock);
 426
 427         if (!pageset_count)
 428                 goto bail;
 429
 430         ngroups = pageset_count / dd->rcv_entries.group_size;
 431         tidlist = kcalloc(pageset_count, sizeof(*tidlist), GFP_KERNEL);
 432         if (!tidlist) {
 433                 ret = -ENOMEM;
 434                 goto nomem;
 435         }
 436
 437         tididx = 0;
 438
 439         /*
 440          * From this point on, we are going to be using shared (between master
 441          * and subcontexts) context resources. We need to take the lock.
 442          */
 443         mutex_lock(&uctxt->exp_lock);
 444         /*
 445          * The first step is to program the RcvArray entries which are complete
 446          * groups.
 447          */
 448         while (ngroups && uctxt->tid_group_list.count) {
 449                 struct tid_group *grp =
 450                         tid_group_pop(&uctxt->tid_group_list);
 451
 452                 ret = program_rcvarray(fp, vaddr, grp, pagesets,
 453                                        pageidx, dd->rcv_entries.group_size,
 454                                        pages, tidlist, &tididx, &mapped);
 455                 /*
 456                  * If there was a failure to program the RcvArray
 457                  * entries for the entire group, reset the grp fields
 458                  * and add the grp back to the free group list.
 459                  */
 460                 if (ret <= 0) {
 461                         tid_group_add_tail(grp, &uctxt->tid_group_list);
 462                         hfi1_cdbg(TID,
 463                                   "Failed to program RcvArray group %d", ret);
 464                         goto unlock;
 465                 }
 466
 467                 tid_group_add_tail(grp, &uctxt->tid_full_list);
 468                 ngroups--;
 469                 pageidx += ret;
 470                 mapped_pages += mapped;
 471         }
 472
 473         while (pageidx < pageset_count) {
 474                 struct tid_group *grp, *ptr;
 475                 /*
 476                  * If we don't have any partially used tid groups, check
 477                  * if we have empty groups. If so, take one from there and
 478                  * put in the partially used list.
 479                  */
 480                 if (!uctxt->tid_used_list.count || need_group) {
 481                         if (!uctxt->tid_group_list.count)
 482                                 goto unlock;
 483
 484                         grp = tid_group_pop(&uctxt->tid_group_list);
 485                         tid_group_add_tail(grp, &uctxt->tid_used_list);
 486                         need_group = 0;
 487                 }
 488                 /*
 489                  * There is an optimization opportunity here - instead of
 490                  * fitting as many page sets as we can, check for a group
 491                  * later on in the list that could fit all of them.
 492                  */
 493                 list_for_each_entry_safe(grp, ptr, &uctxt->tid_used_list.list,
 494                                          list) {
 495                         unsigned use = min_t(unsigned, pageset_count - pageidx,
 496                                              grp->size - grp->used);
 497
 498                         ret = program_rcvarray(fp, vaddr, grp, pagesets,
 499                                                pageidx, use, pages, tidlist,
 500                                                &tididx, &mapped);
 501                         if (ret < 0) {
 502                                 hfi1_cdbg(TID,
 503                                           "Failed to program RcvArray entries %d",
 504                                           ret);
 505                                 ret = -EFAULT;
 506                                 goto unlock;
 507                         } else if (ret > 0) {
 508                                 if (grp->used == grp->size)
 509                                         tid_group_move(grp,
 510                                                        &uctxt->tid_used_list,
 511                                                        &uctxt->tid_full_list);
 512                                 pageidx += ret;
 513                                 mapped_pages += mapped;
 514                                 need_group = 0;
 515                                 /* Check if we are done so we break out early */
 516                                 if (pageidx >= pageset_count)
 517                                         break;
 518                         } else if (WARN_ON(ret == 0)) {
 519                                 /*
 520                                  * If ret is 0, we did not program any entries
 521                                  * into this group, which can only happen if
 522                                  * we've screwed up the accounting somewhere.
 523                                  * Warn and try to continue.
 524                                  */
 525                                 need_group = 1;
 526                         }
 527                 }
 528         }
 529 unlock:
 530         mutex_unlock(&uctxt->exp_lock);
 531 nomem:
 532         hfi1_cdbg(TID, "total mapped: tidpairs:%u pages:%u (%d)", tididx,
 533                   mapped_pages, ret);
 534         if (tididx) {
 535                 spin_lock(&fd->tid_lock);
 536                 fd->tid_used += tididx;
 537                 spin_unlock(&fd->tid_lock);
 538                 tinfo->tidcnt = tididx;
 539                 tinfo->length = mapped_pages * PAGE_SIZE;
 540
 541                 if (copy_to_user((void __user *)(unsigned long)tinfo->tidlist,
 542                                  tidlist, sizeof(tidlist[0]) * tididx)) {
 543                         /*
 544                          * On failure to copy to the user level, we need to undo
 545                          * everything done so far so we don't leak resources.
 546                          */
 547                         tinfo->tidlist = (unsigned long)&tidlist;
 548                         hfi1_user_exp_rcv_clear(fp, tinfo);
 549                         tinfo->tidlist = 0;
 550                         ret = -EFAULT;
 551                         goto bail;
 552                 }
 553         }
 554
 555         /*
 556          * If not everything was mapped (due to insufficient RcvArray entries,
 557          * for example), unpin all unmapped pages so we can pin them nex time.
 558          */
 559         if (mapped_pages != pinned) {
 560                 hfi1_release_user_pages(fd->mm, &pages[mapped_pages],
 561                                         pinned - mapped_pages,
 562                                         false);
 563                 fd->tid_n_pinned -= pinned - mapped_pages;
 564         }
 565 bail:
 566         kfree(pagesets);
 567         kfree(pages);
 568         kfree(tidlist);
 569         return ret > 0 ? 0 : ret;
 570 }
 571
 572 int hfi1_user_exp_rcv_clear(struct file *fp, struct hfi1_tid_info *tinfo)
 573 {
 574         int ret = 0;
 575         struct hfi1_filedata *fd = fp->private_data;
 576         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 577         u32 *tidinfo;
 578         unsigned tididx;
 579
 580         tidinfo = kcalloc(tinfo->tidcnt, sizeof(*tidinfo), GFP_KERNEL);
 581         if (!tidinfo)
 582                 return -ENOMEM;
 583
 584         if (copy_from_user(tidinfo, (void __user *)(unsigned long)
 585                            tinfo->tidlist, sizeof(tidinfo[0]) *
 586                            tinfo->tidcnt)) {
 587                 ret = -EFAULT;
 588                 goto done;
 589         }
 590
 591         mutex_lock(&uctxt->exp_lock);
 592         for (tididx = 0; tididx < tinfo->tidcnt; tididx++) {
 593                 ret = unprogram_rcvarray(fp, tidinfo[tididx], NULL);
 594                 if (ret) {
 595                         hfi1_cdbg(TID, "Failed to unprogram rcv array %d",
 596                                   ret);
 597                         break;
 598                 }
 599         }
 600         spin_lock(&fd->tid_lock);
 601         fd->tid_used -= tididx;
 602         spin_unlock(&fd->tid_lock);
 603         tinfo->tidcnt = tididx;
 604         mutex_unlock(&uctxt->exp_lock);
 605 done:
 606         kfree(tidinfo);
 607         return ret;
 608 }
 609
 610 int hfi1_user_exp_rcv_invalid(struct file *fp, struct hfi1_tid_info *tinfo)
 611 {
 612         struct hfi1_filedata *fd = fp->private_data;
 613         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 614         unsigned long *ev = uctxt->dd->events +
 615                 (((uctxt->ctxt - uctxt->dd->first_user_ctxt) *
 616                   HFI1_MAX_SHARED_CTXTS) + fd->subctxt);
 617         u32 *array;
 618         int ret = 0;
 619
 620         if (!fd->invalid_tids)
 621                 return -EINVAL;
 622
 623         /*
 624          * copy_to_user() can sleep, which will leave the invalid_lock
 625          * locked and cause the MMU notifier to be blocked on the lock
 626          * for a long time.
 627          * Copy the data to a local buffer so we can release the lock.
 628          */
 629         array = kcalloc(uctxt->expected_count, sizeof(*array), GFP_KERNEL);
 630         if (!array)
 631                 return -EFAULT;
 632
 633         spin_lock(&fd->invalid_lock);
 634         if (fd->invalid_tid_idx) {
 635                 memcpy(array, fd->invalid_tids, sizeof(*array) *
 636                        fd->invalid_tid_idx);
 637                 memset(fd->invalid_tids, 0, sizeof(*fd->invalid_tids) *
 638                        fd->invalid_tid_idx);
 639                 tinfo->tidcnt = fd->invalid_tid_idx;
 640                 fd->invalid_tid_idx = 0;
 641                 /*
 642                  * Reset the user flag while still holding the lock.
 643                  * Otherwise, PSM can miss events.
 644                  */
 645                 clear_bit(_HFI1_EVENT_TID_MMU_NOTIFY_BIT, ev);
 646         } else {
 647                 tinfo->tidcnt = 0;
 648         }
 649         spin_unlock(&fd->invalid_lock);
 650
 651         if (tinfo->tidcnt) {
 652                 if (copy_to_user((void __user *)tinfo->tidlist,
 653                                  array, sizeof(*array) * tinfo->tidcnt))
 654                         ret = -EFAULT;
 655         }
 656         kfree(array);
 657
 658         return ret;
 659 }
 660
 661 static u32 find_phys_blocks(struct page **pages, unsigned npages,
 662                             struct tid_pageset *list)
 663 {
 664         unsigned pagecount, pageidx, setcount = 0, i;
 665         unsigned long pfn, this_pfn;
 666
 667         if (!npages)
 668                 return 0;
 669
 670         /*
 671          * Look for sets of physically contiguous pages in the user buffer.
 672          * This will allow us to optimize Expected RcvArray entry usage by
 673          * using the bigger supported sizes.
 674          */
 675         pfn = page_to_pfn(pages[0]);
 676         for (pageidx = 0, pagecount = 1, i = 1; i <= npages; i++) {
 677                 this_pfn = i < npages ? page_to_pfn(pages[i]) : 0;
 678
 679                 /*
 680                  * If the pfn's are not sequential, pages are not physically
 681                  * contiguous.
 682                  */
 683                 if (this_pfn != ++pfn) {
 684                         /*
 685                          * At this point we have to loop over the set of
 686                          * physically contiguous pages and break them down it
 687                          * sizes supported by the HW.
 688                          * There are two main constraints:
 689                          *     1. The max buffer size is MAX_EXPECTED_BUFFER.
 690                          *        If the total set size is bigger than that
 691                          *        program only a MAX_EXPECTED_BUFFER chunk.
 692                          *     2. The buffer size has to be a power of two. If
 693                          *        it is not, round down to the closes power of
 694                          *        2 and program that size.
 695                          */
 696                         while (pagecount) {
 697                                 int maxpages = pagecount;
 698                                 u32 bufsize = pagecount * PAGE_SIZE;
 699
 700                                 if (bufsize > MAX_EXPECTED_BUFFER)
 701                                         maxpages =
 702                                                 MAX_EXPECTED_BUFFER >>
 703                                                 PAGE_SHIFT;
 704                                 else if (!is_power_of_2(bufsize))
 705                                         maxpages =
 706                                                 rounddown_pow_of_two(bufsize) >>
 707                                                 PAGE_SHIFT;
 708
 709                                 list[setcount].idx = pageidx;
 710                                 list[setcount].count = maxpages;
 711                                 pagecount -= maxpages;
 712                                 pageidx += maxpages;
 713                                 setcount++;
 714                         }
 715                         pageidx = i;
 716                         pagecount = 1;
 717                         pfn = this_pfn;
 718                 } else {
 719                         pagecount++;
 720                 }
 721         }
 722         return setcount;
 723 }
 724
 725 /**
 726  * program_rcvarray() - program an RcvArray group with receive buffers
 727  * @fp: file pointer
 728  * @vaddr: starting user virtual address
 729  * @grp: RcvArray group
 730  * @sets: array of struct tid_pageset holding information on physically
 731  *        contiguous chunks from the user buffer
 732  * @start: starting index into sets array
 733  * @count: number of struct tid_pageset's to program
 734  * @pages: an array of struct page * for the user buffer
 735  * @tidlist: the array of u32 elements when the information about the
 736  *           programmed RcvArray entries is to be encoded.
 737  * @tididx: starting offset into tidlist
 738  * @pmapped: (output parameter) number of pages programmed into the RcvArray
 739  *           entries.
 740  *
 741  * This function will program up to 'count' number of RcvArray entries from the
 742  * group 'grp'. To make best use of write-combining writes, the function will
 743  * perform writes to the unused RcvArray entries which will be ignored by the
 744  * HW. Each RcvArray entry will be programmed with a physically contiguous
 745  * buffer chunk from the user's virtual buffer.
 746  *
 747  * Return:
 748  * -EINVAL if the requested count is larger than the size of the group,
 749  * -ENOMEM or -EFAULT on error from set_rcvarray_entry(), or
 750  * number of RcvArray entries programmed.
 751  */
 752 static int program_rcvarray(struct file *fp, unsigned long vaddr,
 753                             struct tid_group *grp,
 754                             struct tid_pageset *sets,
 755                             unsigned start, u16 count, struct page **pages,
 756                             u32 *tidlist, unsigned *tididx, unsigned *pmapped)
 757 {
 758         struct hfi1_filedata *fd = fp->private_data;
 759         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 760         struct hfi1_devdata *dd = uctxt->dd;
 761         u16 idx;
 762         u32 tidinfo = 0, rcventry, useidx = 0;
 763         int mapped = 0;
 764
 765         /* Count should never be larger than the group size */
 766         if (count > grp->size)
 767                 return -EINVAL;
 768
 769         /* Find the first unused entry in the group */
 770         for (idx = 0; idx < grp->size; idx++) {
 771                 if (!(grp->map & (1 << idx))) {
 772                         useidx = idx;
 773                         break;
 774                 }
 775                 rcv_array_wc_fill(dd, grp->base + idx);
 776         }
 777
 778         idx = 0;
 779         while (idx < count) {
 780                 u16 npages, pageidx, setidx = start + idx;
 781                 int ret = 0;
 782
 783                 /*
 784                  * If this entry in the group is used, move to the next one.
 785                  * If we go past the end of the group, exit the loop.
 786                  */
 787                 if (useidx >= grp->size) {
 788                         break;
 789                 } else if (grp->map & (1 << useidx)) {
 790                         rcv_array_wc_fill(dd, grp->base + useidx);
 791                         useidx++;
 792                         continue;
 793                 }
 794
 795                 rcventry = grp->base + useidx;
 796                 npages = sets[setidx].count;
 797                 pageidx = sets[setidx].idx;
 798
 799                 ret = set_rcvarray_entry(fp, vaddr + (pageidx * PAGE_SIZE),
 800                                          rcventry, grp, pages + pageidx,
 801                                          npages);
 802                 if (ret)
 803                         return ret;
 804                 mapped += npages;
 805
 806                 tidinfo = rcventry2tidinfo(rcventry - uctxt->expected_base) |
 807                         EXP_TID_SET(LEN, npages);
 808                 tidlist[(*tididx)++] = tidinfo;
 809                 grp->used++;
 810                 grp->map |= 1 << useidx++;
 811                 idx++;
 812         }
 813
 814         /* Fill the rest of the group with "blank" writes */
 815         for (; useidx < grp->size; useidx++)
 816                 rcv_array_wc_fill(dd, grp->base + useidx);
 817         *pmapped = mapped;
 818         return idx;
 819 }
 820
 821 static int set_rcvarray_entry(struct file *fp, unsigned long vaddr,
 822                               u32 rcventry, struct tid_group *grp,
 823                               struct page **pages, unsigned npages)
 824 {
 825         int ret;
 826         struct hfi1_filedata *fd = fp->private_data;
 827         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 828         struct tid_rb_node *node;
 829         struct hfi1_devdata *dd = uctxt->dd;
 830         dma_addr_t phys;
 831
 832         /*
 833          * Allocate the node first so we can handle a potential
 834          * failure before we've programmed anything.
 835          */
 836         node = kzalloc(sizeof(*node) + (sizeof(struct page *) * npages),
 837                        GFP_KERNEL);
 838         if (!node)
 839                 return -ENOMEM;
 840
 841         phys = pci_map_single(dd->pcidev,
 842                               __va(page_to_phys(pages[0])),
 843                               npages * PAGE_SIZE, PCI_DMA_FROMDEVICE);
 844         if (dma_mapping_error(&dd->pcidev->dev, phys)) {
 845                 dd_dev_err(dd, "Failed to DMA map Exp Rcv pages 0x%llx\n",
 846                            phys);
 847                 kfree(node);
 848                 return -EFAULT;
 849         }
 850
 851         node->mmu.addr = vaddr;
 852         node->mmu.len = npages * PAGE_SIZE;
 853         node->phys = page_to_phys(pages[0]);
 854         node->npages = npages;
 855         node->rcventry = rcventry;
 856         node->dma_addr = phys;
 857         node->grp = grp;
 858         node->freed = false;
 859         memcpy(node->pages, pages, sizeof(struct page *) * npages);
 860
 861         if (!fd->handler)
 862                 ret = tid_rb_insert(fd, &node->mmu);
 863         else
 864                 ret = hfi1_mmu_rb_insert(fd->handler, &node->mmu);
 865
 866         if (ret) {
 867                 hfi1_cdbg(TID, "Failed to insert RB node %u 0x%lx, 0x%lx %d",
 868                           node->rcventry, node->mmu.addr, node->phys, ret);
 869                 pci_unmap_single(dd->pcidev, phys, npages * PAGE_SIZE,
 870                                  PCI_DMA_FROMDEVICE);
 871                 kfree(node);
 872                 return -EFAULT;
 873         }
 874         hfi1_put_tid(dd, rcventry, PT_EXPECTED, phys, ilog2(npages) + 1);
 875         trace_hfi1_exp_tid_reg(uctxt->ctxt, fd->subctxt, rcventry, npages,
 876                                node->mmu.addr, node->phys, phys);
 877         return 0;
 878 }
 879
 880 static int unprogram_rcvarray(struct file *fp, u32 tidinfo,
 881                               struct tid_group **grp)
 882 {
 883         struct hfi1_filedata *fd = fp->private_data;
 884         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 885         struct hfi1_devdata *dd = uctxt->dd;
 886         struct tid_rb_node *node;
 887         u8 tidctrl = EXP_TID_GET(tidinfo, CTRL);
 888         u32 tididx = EXP_TID_GET(tidinfo, IDX) << 1, rcventry;
 889
 890         if (tididx >= uctxt->expected_count) {
 891                 dd_dev_err(dd, "Invalid RcvArray entry (%u) index for ctxt %u\n",
 892                            tididx, uctxt->ctxt);
 893                 return -EINVAL;
 894         }
 895
 896         if (tidctrl == 0x3)
 897                 return -EINVAL;
 898
 899         rcventry = tididx + (tidctrl - 1);
 900
 901         node = fd->entry_to_rb[rcventry];
 902         if (!node || node->rcventry != (uctxt->expected_base + rcventry))
 903                 return -EBADF;
 904
 905         if (grp)
 906                 *grp = node->grp;
 907
 908         if (!fd->handler)
 909                 cacheless_tid_rb_remove(fd, node);
 910         else
 911                 hfi1_mmu_rb_remove(fd->handler, &node->mmu);
 912
 913         return 0;
 914 }
 915
 916 static void clear_tid_node(struct hfi1_filedata *fd, struct tid_rb_node *node)
 917 {
 918         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 919         struct hfi1_devdata *dd = uctxt->dd;
 920
 921         trace_hfi1_exp_tid_unreg(uctxt->ctxt, fd->subctxt, node->rcventry,
 922                                  node->npages, node->mmu.addr, node->phys,
 923                                  node->dma_addr);
 924
 925         hfi1_put_tid(dd, node->rcventry, PT_INVALID, 0, 0);
 926         /*
 927          * Make sure device has seen the write before we unpin the
 928          * pages.
 929          */
 930         flush_wc();
 931
 932         pci_unmap_single(dd->pcidev, node->dma_addr, node->mmu.len,
 933                          PCI_DMA_FROMDEVICE);
 934         hfi1_release_user_pages(fd->mm, node->pages, node->npages, true);
 935         fd->tid_n_pinned -= node->npages;
 936
 937         node->grp->used--;
 938         node->grp->map &= ~(1 << (node->rcventry - node->grp->base));
 939
 940         if (node->grp->used == node->grp->size - 1)
 941                 tid_group_move(node->grp, &uctxt->tid_full_list,
 942                                &uctxt->tid_used_list);
 943         else if (!node->grp->used)
 944                 tid_group_move(node->grp, &uctxt->tid_used_list,
 945                                &uctxt->tid_group_list);
 946         kfree(node);
 947 }
 948
 949 /*
 950  * As a simple helper for hfi1_user_exp_rcv_free, this function deals with
 951  * clearing nodes in the non-cached case.
 952  */
 953 static void unlock_exp_tids(struct hfi1_ctxtdata *uctxt,
 954                             struct exp_tid_set *set,
 955                             struct hfi1_filedata *fd)
 956 {
 957         struct tid_group *grp, *ptr;
 958         int i;
 959
 960         list_for_each_entry_safe(grp, ptr, &set->list, list) {
 961                 list_del_init(&grp->list);
 962
 963                 for (i = 0; i < grp->size; i++) {
 964                         if (grp->map & (1 << i)) {
 965                                 u16 rcventry = grp->base + i;
 966                                 struct tid_rb_node *node;
 967
 968                                 node = fd->entry_to_rb[rcventry -
 969                                                           uctxt->expected_base];
 970                                 if (!node || node->rcventry != rcventry)
 971                                         continue;
 972
 973                                 cacheless_tid_rb_remove(fd, node);
 974                         }
 975                 }
 976         }
 977 }
 978
 979 /*
 980  * Always return 0 from this function.  A non-zero return indicates that the
 981  * remove operation will be called and that memory should be unpinned.
 982  * However, the driver cannot unpin out from under PSM.  Instead, retain the
 983  * memory (by returning 0) and inform PSM that the memory is going away.  PSM
 984  * will call back later when it has removed the memory from its list.
 985  */
 986 static int tid_rb_invalidate(void *arg, struct mmu_rb_node *mnode)
 987 {
 988         struct hfi1_filedata *fdata = arg;
 989         struct hfi1_ctxtdata *uctxt = fdata->uctxt;
 990         struct tid_rb_node *node =
 991                 container_of(mnode, struct tid_rb_node, mmu);
 992
 993         if (node->freed)
 994                 return 0;
 995
 996         trace_hfi1_exp_tid_inval(uctxt->ctxt, fdata->subctxt, node->mmu.addr,
 997                                  node->rcventry, node->npages, node->dma_addr);
 998         node->freed = true;
 999
1000         spin_lock(&fdata->invalid_lock);
1001         if (fdata->invalid_tid_idx < uctxt->expected_count) {
1002                 fdata->invalid_tids[fdata->invalid_tid_idx] =
1003                         rcventry2tidinfo(node->rcventry - uctxt->expected_base);
1004                 fdata->invalid_tids[fdata->invalid_tid_idx] |=
1005                         EXP_TID_SET(LEN, node->npages);
1006                 if (!fdata->invalid_tid_idx) {
1007                         unsigned long *ev;
1008
1009                         /*
1010                          * hfi1_set_uevent_bits() sets a user event flag
1011                          * for all processes. Because calling into the
1012                          * driver to process TID cache invalidations is
1013                          * expensive and TID cache invalidations are
1014                          * handled on a per-process basis, we can
1015                          * optimize this to set the flag only for the
1016                          * process in question.
1017                          */
1018                         ev = uctxt->dd->events +
1019                                 (((uctxt->ctxt - uctxt->dd->first_user_ctxt) *
1020                                   HFI1_MAX_SHARED_CTXTS) + fdata->subctxt);
1021                         set_bit(_HFI1_EVENT_TID_MMU_NOTIFY_BIT, ev);
1022                 }
1023                 fdata->invalid_tid_idx++;
1024         }
1025         spin_unlock(&fdata->invalid_lock);
1026         return 0;
1027 }
1028
1029 static int tid_rb_insert(void *arg, struct mmu_rb_node *node)
1030 {
1031         struct hfi1_filedata *fdata = arg;
1032         struct tid_rb_node *tnode =
1033                 container_of(node, struct tid_rb_node, mmu);
1034         u32 base = fdata->uctxt->expected_base;
1035
1036         fdata->entry_to_rb[tnode->rcventry - base] = tnode;
1037         return 0;
1038 }
1039
1040 static void cacheless_tid_rb_remove(struct hfi1_filedata *fdata,
1041                                     struct tid_rb_node *tnode)
1042 {
1043         u32 base = fdata->uctxt->expected_base;
1044
1045         fdata->entry_to_rb[tnode->rcventry - base] = NULL;
1046         clear_tid_node(fdata, tnode);
1047 }
1048
1049 static void tid_rb_remove(void *arg, struct mmu_rb_node *node)
1050 {
1051         struct hfi1_filedata *fdata = arg;
1052         struct tid_rb_node *tnode =
1053                 container_of(node, struct tid_rb_node, mmu);
1054
1055         cacheless_tid_rb_remove(fdata, tnode);
1056 }