drivers/infiniband/hw/hfi1/user_exp_rcv.c

   1 /*
   2  * Copyright(c) 2015-2017 Intel Corporation.
   3  *
   4  * This file is provided under a dual BSD/GPLv2 license.  When using or
   5  * redistributing this file, you may do so under either license.
   6  *
   7  * GPL LICENSE SUMMARY
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of version 2 of the GNU General Public License as
  11  * published by the Free Software Foundation.
  12  *
  13  * This program is distributed in the hope that it will be useful, but
  14  * WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * General Public License for more details.
  17  *
  18  * BSD LICENSE
  19  *
  20  * Redistribution and use in source and binary forms, with or without
  21  * modification, are permitted provided that the following conditions
  22  * are met:
  23  *
  24  *  - Redistributions of source code must retain the above copyright
  25  *    notice, this list of conditions and the following disclaimer.
  26  *  - Redistributions in binary form must reproduce the above copyright
  27  *    notice, this list of conditions and the following disclaimer in
  28  *    the documentation and/or other materials provided with the
  29  *    distribution.
  30  *  - Neither the name of Intel Corporation nor the names of its
  31  *    contributors may be used to endorse or promote products derived
  32  *    from this software without specific prior written permission.
  33  *
  34  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  35  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  36  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  37  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  38  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  39  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  40  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  41  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  42  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  43  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  44  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  45  *
  46  */
  47 #include <asm/page.h>
  48 #include <linux/string.h>
  49
  50 #include "mmu_rb.h"
  51 #include "user_exp_rcv.h"
  52 #include "trace.h"
  53
  54 static void unlock_exp_tids(struct hfi1_ctxtdata *uctxt,
  55                             struct exp_tid_set *set,
  56                             struct hfi1_filedata *fd);
  57 static u32 find_phys_blocks(struct tid_user_buf *tidbuf, unsigned int npages);
  58 static int set_rcvarray_entry(struct hfi1_filedata *fd,
  59                               struct tid_user_buf *tbuf,
  60                               u32 rcventry, struct tid_group *grp,
  61                               u16 pageidx, unsigned int npages);
  62 static int tid_rb_insert(void *arg, struct mmu_rb_node *node);
  63 static void cacheless_tid_rb_remove(struct hfi1_filedata *fdata,
  64                                     struct tid_rb_node *tnode);
  65 static void tid_rb_remove(void *arg, struct mmu_rb_node *node);
  66 static int tid_rb_invalidate(void *arg, struct mmu_rb_node *mnode);
  67 static int program_rcvarray(struct hfi1_filedata *fd, struct tid_user_buf *,
  68                             struct tid_group *grp,
  69                             unsigned int start, u16 count,
  70                             u32 *tidlist, unsigned int *tididx,
  71                             unsigned int *pmapped);
  72 static int unprogram_rcvarray(struct hfi1_filedata *fd, u32 tidinfo,
  73                               struct tid_group **grp);
  74 static void clear_tid_node(struct hfi1_filedata *fd, struct tid_rb_node *node);
  75
  76 static struct mmu_rb_ops tid_rb_ops = {
  77         .insert = tid_rb_insert,
  78         .remove = tid_rb_remove,
  79         .invalidate = tid_rb_invalidate
  80 };
  81
  82 /*
  83  * Initialize context and file private data needed for Expected
  84  * receive caching. This needs to be done after the context has
  85  * been configured with the eager/expected RcvEntry counts.
  86  */
  87 int hfi1_user_exp_rcv_init(struct hfi1_filedata *fd,
  88                            struct hfi1_ctxtdata *uctxt)
  89 {
  90         struct hfi1_devdata *dd = uctxt->dd;
  91         int ret = 0;
  92
  93         fd->entry_to_rb = kcalloc(uctxt->expected_count,
  94                                   sizeof(struct rb_node *),
  95                                   GFP_KERNEL);
  96         if (!fd->entry_to_rb)
  97                 return -ENOMEM;
  98
  99         if (!HFI1_CAP_UGET_MASK(uctxt->flags, TID_UNMAP)) {
 100                 fd->invalid_tid_idx = 0;
 101                 fd->invalid_tids = kcalloc(uctxt->expected_count,
 102                                            sizeof(*fd->invalid_tids),
 103                                            GFP_KERNEL);
 104                 if (!fd->invalid_tids) {
 105                         kfree(fd->entry_to_rb);
 106                         fd->entry_to_rb = NULL;
 107                         return -ENOMEM;
 108                 }
 109
 110                 /*
 111                  * Register MMU notifier callbacks. If the registration
 112                  * fails, continue without TID caching for this context.
 113                  */
 114                 ret = hfi1_mmu_rb_register(fd, fd->mm, &tid_rb_ops,
 115                                            dd->pport->hfi1_wq,
 116                                            &fd->handler);
 117                 if (ret) {
 118                         dd_dev_info(dd,
 119                                     "Failed MMU notifier registration %d\n",
 120                                     ret);
 121                         ret = 0;
 122                 }
 123         }
 124
 125         /*
 126          * PSM does not have a good way to separate, count, and
 127          * effectively enforce a limit on RcvArray entries used by
 128          * subctxts (when context sharing is used) when TID caching
 129          * is enabled. To help with that, we calculate a per-process
 130          * RcvArray entry share and enforce that.
 131          * If TID caching is not in use, PSM deals with usage on its
 132          * own. In that case, we allow any subctxt to take all of the
 133          * entries.
 134          *
 135          * Make sure that we set the tid counts only after successful
 136          * init.
 137          */
 138         spin_lock(&fd->tid_lock);
 139         if (uctxt->subctxt_cnt && fd->handler) {
 140                 u16 remainder;
 141
 142                 fd->tid_limit = uctxt->expected_count / uctxt->subctxt_cnt;
 143                 remainder = uctxt->expected_count % uctxt->subctxt_cnt;
 144                 if (remainder && fd->subctxt < remainder)
 145                         fd->tid_limit++;
 146         } else {
 147                 fd->tid_limit = uctxt->expected_count;
 148         }
 149         spin_unlock(&fd->tid_lock);
 150
 151         return ret;
 152 }
 153
 154 void hfi1_user_exp_rcv_free(struct hfi1_filedata *fd)
 155 {
 156         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 157
 158         /*
 159          * The notifier would have been removed when the process'es mm
 160          * was freed.
 161          */
 162         if (fd->handler) {
 163                 hfi1_mmu_rb_unregister(fd->handler);
 164         } else {
 165                 if (!EXP_TID_SET_EMPTY(uctxt->tid_full_list))
 166                         unlock_exp_tids(uctxt, &uctxt->tid_full_list, fd);
 167                 if (!EXP_TID_SET_EMPTY(uctxt->tid_used_list))
 168                         unlock_exp_tids(uctxt, &uctxt->tid_used_list, fd);
 169         }
 170
 171         kfree(fd->invalid_tids);
 172         fd->invalid_tids = NULL;
 173
 174         kfree(fd->entry_to_rb);
 175         fd->entry_to_rb = NULL;
 176 }
 177
 178 /**
 179  * Release pinned receive buffer pages.
 180  *
 181  * @mapped - true if the pages have been DMA mapped. false otherwise.
 182  * @idx - Index of the first page to unpin.
 183  * @npages - No of pages to unpin.
 184  *
 185  * If the pages have been DMA mapped (indicated by mapped parameter), their
 186  * info will be passed via a struct tid_rb_node. If they haven't been mapped,
 187  * their info will be passed via a struct tid_user_buf.
 188  */
 189 static void unpin_rcv_pages(struct hfi1_filedata *fd,
 190                             struct tid_user_buf *tidbuf,
 191                             struct tid_rb_node *node,
 192                             unsigned int idx,
 193                             unsigned int npages,
 194                             bool mapped)
 195 {
 196         struct page **pages;
 197         struct hfi1_devdata *dd = fd->uctxt->dd;
 198
 199         if (mapped) {
 200                 pci_unmap_single(dd->pcidev, node->dma_addr,
 201                                  node->mmu.len, PCI_DMA_FROMDEVICE);
 202                 pages = &node->pages[idx];
 203         } else {
 204                 pages = &tidbuf->pages[idx];
 205         }
 206         hfi1_release_user_pages(fd->mm, pages, npages, mapped);
 207         fd->tid_n_pinned -= npages;
 208 }
 209
 210 /**
 211  * Pin receive buffer pages.
 212  */
 213 static int pin_rcv_pages(struct hfi1_filedata *fd, struct tid_user_buf *tidbuf)
 214 {
 215         int pinned;
 216         unsigned int npages;
 217         unsigned long vaddr = tidbuf->vaddr;
 218         struct page **pages = NULL;
 219         struct hfi1_devdata *dd = fd->uctxt->dd;
 220
 221         /* Get the number of pages the user buffer spans */
 222         npages = num_user_pages(vaddr, tidbuf->length);
 223         if (!npages)
 224                 return -EINVAL;
 225
 226         if (npages > fd->uctxt->expected_count) {
 227                 dd_dev_err(dd, "Expected buffer too big\n");
 228                 return -EINVAL;
 229         }
 230
 231         /* Verify that access is OK for the user buffer */
 232         if (!access_ok(VERIFY_WRITE, (void __user *)vaddr,
 233                        npages * PAGE_SIZE)) {
 234                 dd_dev_err(dd, "Fail vaddr %p, %u pages, !access_ok\n",
 235                            (void *)vaddr, npages);
 236                 return -EFAULT;
 237         }
 238         /* Allocate the array of struct page pointers needed for pinning */
 239         pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL);
 240         if (!pages)
 241                 return -ENOMEM;
 242
 243         /*
 244          * Pin all the pages of the user buffer. If we can't pin all the
 245          * pages, accept the amount pinned so far and program only that.
 246          * User space knows how to deal with partially programmed buffers.
 247          */
 248         if (!hfi1_can_pin_pages(dd, fd->mm, fd->tid_n_pinned, npages)) {
 249                 kfree(pages);
 250                 return -ENOMEM;
 251         }
 252
 253         pinned = hfi1_acquire_user_pages(fd->mm, vaddr, npages, true, pages);
 254         if (pinned <= 0) {
 255                 kfree(pages);
 256                 return pinned;
 257         }
 258         tidbuf->pages = pages;
 259         tidbuf->npages = npages;
 260         fd->tid_n_pinned += pinned;
 261         return pinned;
 262 }
 263
 264 /*
 265  * RcvArray entry allocation for Expected Receives is done by the
 266  * following algorithm:
 267  *
 268  * The context keeps 3 lists of groups of RcvArray entries:
 269  *   1. List of empty groups - tid_group_list
 270  *      This list is created during user context creation and
 271  *      contains elements which describe sets (of 8) of empty
 272  *      RcvArray entries.
 273  *   2. List of partially used groups - tid_used_list
 274  *      This list contains sets of RcvArray entries which are
 275  *      not completely used up. Another mapping request could
 276  *      use some of all of the remaining entries.
 277  *   3. List of full groups - tid_full_list
 278  *      This is the list where sets that are completely used
 279  *      up go.
 280  *
 281  * An attempt to optimize the usage of RcvArray entries is
 282  * made by finding all sets of physically contiguous pages in a
 283  * user's buffer.
 284  * These physically contiguous sets are further split into
 285  * sizes supported by the receive engine of the HFI. The
 286  * resulting sets of pages are stored in struct tid_pageset,
 287  * which describes the sets as:
 288  *    * .count - number of pages in this set
 289  *    * .idx - starting index into struct page ** array
 290  *                    of this set
 291  *
 292  * From this point on, the algorithm deals with the page sets
 293  * described above. The number of pagesets is divided by the
 294  * RcvArray group size to produce the number of full groups
 295  * needed.
 296  *
 297  * Groups from the 3 lists are manipulated using the following
 298  * rules:
 299  *   1. For each set of 8 pagesets, a complete group from
 300  *      tid_group_list is taken, programmed, and moved to
 301  *      the tid_full_list list.
 302  *   2. For all remaining pagesets:
 303  *      2.1 If the tid_used_list is empty and the tid_group_list
 304  *          is empty, stop processing pageset and return only
 305  *          what has been programmed up to this point.
 306  *      2.2 If the tid_used_list is empty and the tid_group_list
 307  *          is not empty, move a group from tid_group_list to
 308  *          tid_used_list.
 309  *      2.3 For each group is tid_used_group, program as much as
 310  *          can fit into the group. If the group becomes fully
 311  *          used, move it to tid_full_list.
 312  */
 313 int hfi1_user_exp_rcv_setup(struct hfi1_filedata *fd,
 314                             struct hfi1_tid_info *tinfo)
 315 {
 316         int ret = 0, need_group = 0, pinned;
 317         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 318         struct hfi1_devdata *dd = uctxt->dd;
 319         unsigned int ngroups, pageidx = 0, pageset_count,
 320                 tididx = 0, mapped, mapped_pages = 0;
 321         u32 *tidlist = NULL;
 322         struct tid_user_buf *tidbuf;
 323
 324         if (!PAGE_ALIGNED(tinfo->vaddr))
 325                 return -EINVAL;
 326
 327         tidbuf = kzalloc(sizeof(*tidbuf), GFP_KERNEL);
 328         if (!tidbuf)
 329                 return -ENOMEM;
 330
 331         tidbuf->vaddr = tinfo->vaddr;
 332         tidbuf->length = tinfo->length;
 333         tidbuf->psets = kcalloc(uctxt->expected_count, sizeof(*tidbuf->psets),
 334                                 GFP_KERNEL);
 335         if (!tidbuf->psets) {
 336                 kfree(tidbuf);
 337                 return -ENOMEM;
 338         }
 339
 340         pinned = pin_rcv_pages(fd, tidbuf);
 341         if (pinned <= 0) {
 342                 kfree(tidbuf->psets);
 343                 kfree(tidbuf);
 344                 return pinned;
 345         }
 346
 347         /* Find sets of physically contiguous pages */
 348         tidbuf->n_psets = find_phys_blocks(tidbuf, pinned);
 349
 350         /*
 351          * We don't need to access this under a lock since tid_used is per
 352          * process and the same process cannot be in hfi1_user_exp_rcv_clear()
 353          * and hfi1_user_exp_rcv_setup() at the same time.
 354          */
 355         spin_lock(&fd->tid_lock);
 356         if (fd->tid_used + tidbuf->n_psets > fd->tid_limit)
 357                 pageset_count = fd->tid_limit - fd->tid_used;
 358         else
 359                 pageset_count = tidbuf->n_psets;
 360         spin_unlock(&fd->tid_lock);
 361
 362         if (!pageset_count)
 363                 goto bail;
 364
 365         ngroups = pageset_count / dd->rcv_entries.group_size;
 366         tidlist = kcalloc(pageset_count, sizeof(*tidlist), GFP_KERNEL);
 367         if (!tidlist) {
 368                 ret = -ENOMEM;
 369                 goto nomem;
 370         }
 371
 372         tididx = 0;
 373
 374         /*
 375          * From this point on, we are going to be using shared (between master
 376          * and subcontexts) context resources. We need to take the lock.
 377          */
 378         mutex_lock(&uctxt->exp_lock);
 379         /*
 380          * The first step is to program the RcvArray entries which are complete
 381          * groups.
 382          */
 383         while (ngroups && uctxt->tid_group_list.count) {
 384                 struct tid_group *grp =
 385                         tid_group_pop(&uctxt->tid_group_list);
 386
 387                 ret = program_rcvarray(fd, tidbuf, grp,
 388                                        pageidx, dd->rcv_entries.group_size,
 389                                        tidlist, &tididx, &mapped);
 390                 /*
 391                  * If there was a failure to program the RcvArray
 392                  * entries for the entire group, reset the grp fields
 393                  * and add the grp back to the free group list.
 394                  */
 395                 if (ret <= 0) {
 396                         tid_group_add_tail(grp, &uctxt->tid_group_list);
 397                         hfi1_cdbg(TID,
 398                                   "Failed to program RcvArray group %d", ret);
 399                         goto unlock;
 400                 }
 401
 402                 tid_group_add_tail(grp, &uctxt->tid_full_list);
 403                 ngroups--;
 404                 pageidx += ret;
 405                 mapped_pages += mapped;
 406         }
 407
 408         while (pageidx < pageset_count) {
 409                 struct tid_group *grp, *ptr;
 410                 /*
 411                  * If we don't have any partially used tid groups, check
 412                  * if we have empty groups. If so, take one from there and
 413                  * put in the partially used list.
 414                  */
 415                 if (!uctxt->tid_used_list.count || need_group) {
 416                         if (!uctxt->tid_group_list.count)
 417                                 goto unlock;
 418
 419                         grp = tid_group_pop(&uctxt->tid_group_list);
 420                         tid_group_add_tail(grp, &uctxt->tid_used_list);
 421                         need_group = 0;
 422                 }
 423                 /*
 424                  * There is an optimization opportunity here - instead of
 425                  * fitting as many page sets as we can, check for a group
 426                  * later on in the list that could fit all of them.
 427                  */
 428                 list_for_each_entry_safe(grp, ptr, &uctxt->tid_used_list.list,
 429                                          list) {
 430                         unsigned use = min_t(unsigned, pageset_count - pageidx,
 431                                              grp->size - grp->used);
 432
 433                         ret = program_rcvarray(fd, tidbuf, grp,
 434                                                pageidx, use, tidlist,
 435                                                &tididx, &mapped);
 436                         if (ret < 0) {
 437                                 hfi1_cdbg(TID,
 438                                           "Failed to program RcvArray entries %d",
 439                                           ret);
 440                                 ret = -EFAULT;
 441                                 goto unlock;
 442                         } else if (ret > 0) {
 443                                 if (grp->used == grp->size)
 444                                         tid_group_move(grp,
 445                                                        &uctxt->tid_used_list,
 446                                                        &uctxt->tid_full_list);
 447                                 pageidx += ret;
 448                                 mapped_pages += mapped;
 449                                 need_group = 0;
 450                                 /* Check if we are done so we break out early */
 451                                 if (pageidx >= pageset_count)
 452                                         break;
 453                         } else if (WARN_ON(ret == 0)) {
 454                                 /*
 455                                  * If ret is 0, we did not program any entries
 456                                  * into this group, which can only happen if
 457                                  * we've screwed up the accounting somewhere.
 458                                  * Warn and try to continue.
 459                                  */
 460                                 need_group = 1;
 461                         }
 462                 }
 463         }
 464 unlock:
 465         mutex_unlock(&uctxt->exp_lock);
 466 nomem:
 467         hfi1_cdbg(TID, "total mapped: tidpairs:%u pages:%u (%d)", tididx,
 468                   mapped_pages, ret);
 469         if (tididx) {
 470                 spin_lock(&fd->tid_lock);
 471                 fd->tid_used += tididx;
 472                 spin_unlock(&fd->tid_lock);
 473                 tinfo->tidcnt = tididx;
 474                 tinfo->length = mapped_pages * PAGE_SIZE;
 475
 476                 if (copy_to_user((void __user *)(unsigned long)tinfo->tidlist,
 477                                  tidlist, sizeof(tidlist[0]) * tididx)) {
 478                         /*
 479                          * On failure to copy to the user level, we need to undo
 480                          * everything done so far so we don't leak resources.
 481                          */
 482                         tinfo->tidlist = (unsigned long)&tidlist;
 483                         hfi1_user_exp_rcv_clear(fd, tinfo);
 484                         tinfo->tidlist = 0;
 485                         ret = -EFAULT;
 486                         goto bail;
 487                 }
 488         }
 489
 490         /*
 491          * If not everything was mapped (due to insufficient RcvArray entries,
 492          * for example), unpin all unmapped pages so we can pin them nex time.
 493          */
 494         if (mapped_pages != pinned)
 495                 unpin_rcv_pages(fd, tidbuf, NULL, mapped_pages,
 496                                 (pinned - mapped_pages), false);
 497 bail:
 498         kfree(tidbuf->psets);
 499         kfree(tidlist);
 500         kfree(tidbuf->pages);
 501         kfree(tidbuf);
 502         return ret > 0 ? 0 : ret;
 503 }
 504
 505 int hfi1_user_exp_rcv_clear(struct hfi1_filedata *fd,
 506                             struct hfi1_tid_info *tinfo)
 507 {
 508         int ret = 0;
 509         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 510         u32 *tidinfo;
 511         unsigned tididx;
 512
 513         if (unlikely(tinfo->tidcnt > fd->tid_used))
 514                 return -EINVAL;
 515
 516         tidinfo = memdup_user((void __user *)(unsigned long)tinfo->tidlist,
 517                               sizeof(tidinfo[0]) * tinfo->tidcnt);
 518         if (IS_ERR(tidinfo))
 519                 return PTR_ERR(tidinfo);
 520
 521         mutex_lock(&uctxt->exp_lock);
 522         for (tididx = 0; tididx < tinfo->tidcnt; tididx++) {
 523                 ret = unprogram_rcvarray(fd, tidinfo[tididx], NULL);
 524                 if (ret) {
 525                         hfi1_cdbg(TID, "Failed to unprogram rcv array %d",
 526                                   ret);
 527                         break;
 528                 }
 529         }
 530         spin_lock(&fd->tid_lock);
 531         fd->tid_used -= tididx;
 532         spin_unlock(&fd->tid_lock);
 533         tinfo->tidcnt = tididx;
 534         mutex_unlock(&uctxt->exp_lock);
 535
 536         kfree(tidinfo);
 537         return ret;
 538 }
 539
 540 int hfi1_user_exp_rcv_invalid(struct hfi1_filedata *fd,
 541                               struct hfi1_tid_info *tinfo)
 542 {
 543         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 544         unsigned long *ev = uctxt->dd->events +
 545                 (((uctxt->ctxt - uctxt->dd->first_dyn_alloc_ctxt) *
 546                   HFI1_MAX_SHARED_CTXTS) + fd->subctxt);
 547         u32 *array;
 548         int ret = 0;
 549
 550         if (!fd->invalid_tids)
 551                 return -EINVAL;
 552
 553         /*
 554          * copy_to_user() can sleep, which will leave the invalid_lock
 555          * locked and cause the MMU notifier to be blocked on the lock
 556          * for a long time.
 557          * Copy the data to a local buffer so we can release the lock.
 558          */
 559         array = kcalloc(uctxt->expected_count, sizeof(*array), GFP_KERNEL);
 560         if (!array)
 561                 return -EFAULT;
 562
 563         spin_lock(&fd->invalid_lock);
 564         if (fd->invalid_tid_idx) {
 565                 memcpy(array, fd->invalid_tids, sizeof(*array) *
 566                        fd->invalid_tid_idx);
 567                 memset(fd->invalid_tids, 0, sizeof(*fd->invalid_tids) *
 568                        fd->invalid_tid_idx);
 569                 tinfo->tidcnt = fd->invalid_tid_idx;
 570                 fd->invalid_tid_idx = 0;
 571                 /*
 572                  * Reset the user flag while still holding the lock.
 573                  * Otherwise, PSM can miss events.
 574                  */
 575                 clear_bit(_HFI1_EVENT_TID_MMU_NOTIFY_BIT, ev);
 576         } else {
 577                 tinfo->tidcnt = 0;
 578         }
 579         spin_unlock(&fd->invalid_lock);
 580
 581         if (tinfo->tidcnt) {
 582                 if (copy_to_user((void __user *)tinfo->tidlist,
 583                                  array, sizeof(*array) * tinfo->tidcnt))
 584                         ret = -EFAULT;
 585         }
 586         kfree(array);
 587
 588         return ret;
 589 }
 590
 591 static u32 find_phys_blocks(struct tid_user_buf *tidbuf, unsigned int npages)
 592 {
 593         unsigned pagecount, pageidx, setcount = 0, i;
 594         unsigned long pfn, this_pfn;
 595         struct page **pages = tidbuf->pages;
 596         struct tid_pageset *list = tidbuf->psets;
 597
 598         if (!npages)
 599                 return 0;
 600
 601         /*
 602          * Look for sets of physically contiguous pages in the user buffer.
 603          * This will allow us to optimize Expected RcvArray entry usage by
 604          * using the bigger supported sizes.
 605          */
 606         pfn = page_to_pfn(pages[0]);
 607         for (pageidx = 0, pagecount = 1, i = 1; i <= npages; i++) {
 608                 this_pfn = i < npages ? page_to_pfn(pages[i]) : 0;
 609
 610                 /*
 611                  * If the pfn's are not sequential, pages are not physically
 612                  * contiguous.
 613                  */
 614                 if (this_pfn != ++pfn) {
 615                         /*
 616                          * At this point we have to loop over the set of
 617                          * physically contiguous pages and break them down it
 618                          * sizes supported by the HW.
 619                          * There are two main constraints:
 620                          *     1. The max buffer size is MAX_EXPECTED_BUFFER.
 621                          *        If the total set size is bigger than that
 622                          *        program only a MAX_EXPECTED_BUFFER chunk.
 623                          *     2. The buffer size has to be a power of two. If
 624                          *        it is not, round down to the closes power of
 625                          *        2 and program that size.
 626                          */
 627                         while (pagecount) {
 628                                 int maxpages = pagecount;
 629                                 u32 bufsize = pagecount * PAGE_SIZE;
 630
 631                                 if (bufsize > MAX_EXPECTED_BUFFER)
 632                                         maxpages =
 633                                                 MAX_EXPECTED_BUFFER >>
 634                                                 PAGE_SHIFT;
 635                                 else if (!is_power_of_2(bufsize))
 636                                         maxpages =
 637                                                 rounddown_pow_of_two(bufsize) >>
 638                                                 PAGE_SHIFT;
 639
 640                                 list[setcount].idx = pageidx;
 641                                 list[setcount].count = maxpages;
 642                                 pagecount -= maxpages;
 643                                 pageidx += maxpages;
 644                                 setcount++;
 645                         }
 646                         pageidx = i;
 647                         pagecount = 1;
 648                         pfn = this_pfn;
 649                 } else {
 650                         pagecount++;
 651                 }
 652         }
 653         return setcount;
 654 }
 655
 656 /**
 657  * program_rcvarray() - program an RcvArray group with receive buffers
 658  * @fd: filedata pointer
 659  * @tbuf: pointer to struct tid_user_buf that has the user buffer starting
 660  *        virtual address, buffer length, page pointers, pagesets (array of
 661  *        struct tid_pageset holding information on physically contiguous
 662  *        chunks from the user buffer), and other fields.
 663  * @grp: RcvArray group
 664  * @start: starting index into sets array
 665  * @count: number of struct tid_pageset's to program
 666  * @tidlist: the array of u32 elements when the information about the
 667  *           programmed RcvArray entries is to be encoded.
 668  * @tididx: starting offset into tidlist
 669  * @pmapped: (output parameter) number of pages programmed into the RcvArray
 670  *           entries.
 671  *
 672  * This function will program up to 'count' number of RcvArray entries from the
 673  * group 'grp'. To make best use of write-combining writes, the function will
 674  * perform writes to the unused RcvArray entries which will be ignored by the
 675  * HW. Each RcvArray entry will be programmed with a physically contiguous
 676  * buffer chunk from the user's virtual buffer.
 677  *
 678  * Return:
 679  * -EINVAL if the requested count is larger than the size of the group,
 680  * -ENOMEM or -EFAULT on error from set_rcvarray_entry(), or
 681  * number of RcvArray entries programmed.
 682  */
 683 static int program_rcvarray(struct hfi1_filedata *fd, struct tid_user_buf *tbuf,
 684                             struct tid_group *grp,
 685                             unsigned int start, u16 count,
 686                             u32 *tidlist, unsigned int *tididx,
 687                             unsigned int *pmapped)
 688 {
 689         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 690         struct hfi1_devdata *dd = uctxt->dd;
 691         u16 idx;
 692         u32 tidinfo = 0, rcventry, useidx = 0;
 693         int mapped = 0;
 694
 695         /* Count should never be larger than the group size */
 696         if (count > grp->size)
 697                 return -EINVAL;
 698
 699         /* Find the first unused entry in the group */
 700         for (idx = 0; idx < grp->size; idx++) {
 701                 if (!(grp->map & (1 << idx))) {
 702                         useidx = idx;
 703                         break;
 704                 }
 705                 rcv_array_wc_fill(dd, grp->base + idx);
 706         }
 707
 708         idx = 0;
 709         while (idx < count) {
 710                 u16 npages, pageidx, setidx = start + idx;
 711                 int ret = 0;
 712
 713                 /*
 714                  * If this entry in the group is used, move to the next one.
 715                  * If we go past the end of the group, exit the loop.
 716                  */
 717                 if (useidx >= grp->size) {
 718                         break;
 719                 } else if (grp->map & (1 << useidx)) {
 720                         rcv_array_wc_fill(dd, grp->base + useidx);
 721                         useidx++;
 722                         continue;
 723                 }
 724
 725                 rcventry = grp->base + useidx;
 726                 npages = tbuf->psets[setidx].count;
 727                 pageidx = tbuf->psets[setidx].idx;
 728
 729                 ret = set_rcvarray_entry(fd, tbuf,
 730                                          rcventry, grp, pageidx,
 731                                          npages);
 732                 if (ret)
 733                         return ret;
 734                 mapped += npages;
 735
 736                 tidinfo = rcventry2tidinfo(rcventry - uctxt->expected_base) |
 737                         EXP_TID_SET(LEN, npages);
 738                 tidlist[(*tididx)++] = tidinfo;
 739                 grp->used++;
 740                 grp->map |= 1 << useidx++;
 741                 idx++;
 742         }
 743
 744         /* Fill the rest of the group with "blank" writes */
 745         for (; useidx < grp->size; useidx++)
 746                 rcv_array_wc_fill(dd, grp->base + useidx);
 747         *pmapped = mapped;
 748         return idx;
 749 }
 750
 751 static int set_rcvarray_entry(struct hfi1_filedata *fd,
 752                               struct tid_user_buf *tbuf,
 753                               u32 rcventry, struct tid_group *grp,
 754                               u16 pageidx, unsigned int npages)
 755 {
 756         int ret;
 757         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 758         struct tid_rb_node *node;
 759         struct hfi1_devdata *dd = uctxt->dd;
 760         dma_addr_t phys;
 761         struct page **pages = tbuf->pages + pageidx;
 762
 763         /*
 764          * Allocate the node first so we can handle a potential
 765          * failure before we've programmed anything.
 766          */
 767         node = kzalloc(sizeof(*node) + (sizeof(struct page *) * npages),
 768                        GFP_KERNEL);
 769         if (!node)
 770                 return -ENOMEM;
 771
 772         phys = pci_map_single(dd->pcidev,
 773                               __va(page_to_phys(pages[0])),
 774                               npages * PAGE_SIZE, PCI_DMA_FROMDEVICE);
 775         if (dma_mapping_error(&dd->pcidev->dev, phys)) {
 776                 dd_dev_err(dd, "Failed to DMA map Exp Rcv pages 0x%llx\n",
 777                            phys);
 778                 kfree(node);
 779                 return -EFAULT;
 780         }
 781
 782         node->mmu.addr = tbuf->vaddr + (pageidx * PAGE_SIZE);
 783         node->mmu.len = npages * PAGE_SIZE;
 784         node->phys = page_to_phys(pages[0]);
 785         node->npages = npages;
 786         node->rcventry = rcventry;
 787         node->dma_addr = phys;
 788         node->grp = grp;
 789         node->freed = false;
 790         memcpy(node->pages, pages, sizeof(struct page *) * npages);
 791
 792         if (!fd->handler)
 793                 ret = tid_rb_insert(fd, &node->mmu);
 794         else
 795                 ret = hfi1_mmu_rb_insert(fd->handler, &node->mmu);
 796
 797         if (ret) {
 798                 hfi1_cdbg(TID, "Failed to insert RB node %u 0x%lx, 0x%lx %d",
 799                           node->rcventry, node->mmu.addr, node->phys, ret);
 800                 pci_unmap_single(dd->pcidev, phys, npages * PAGE_SIZE,
 801                                  PCI_DMA_FROMDEVICE);
 802                 kfree(node);
 803                 return -EFAULT;
 804         }
 805         hfi1_put_tid(dd, rcventry, PT_EXPECTED, phys, ilog2(npages) + 1);
 806         trace_hfi1_exp_tid_reg(uctxt->ctxt, fd->subctxt, rcventry, npages,
 807                                node->mmu.addr, node->phys, phys);
 808         return 0;
 809 }
 810
 811 static int unprogram_rcvarray(struct hfi1_filedata *fd, u32 tidinfo,
 812                               struct tid_group **grp)
 813 {
 814         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 815         struct hfi1_devdata *dd = uctxt->dd;
 816         struct tid_rb_node *node;
 817         u8 tidctrl = EXP_TID_GET(tidinfo, CTRL);
 818         u32 tididx = EXP_TID_GET(tidinfo, IDX) << 1, rcventry;
 819
 820         if (tididx >= uctxt->expected_count) {
 821                 dd_dev_err(dd, "Invalid RcvArray entry (%u) index for ctxt %u\n",
 822                            tididx, uctxt->ctxt);
 823                 return -EINVAL;
 824         }
 825
 826         if (tidctrl == 0x3)
 827                 return -EINVAL;
 828
 829         rcventry = tididx + (tidctrl - 1);
 830
 831         node = fd->entry_to_rb[rcventry];
 832         if (!node || node->rcventry != (uctxt->expected_base + rcventry))
 833                 return -EBADF;
 834
 835         if (grp)
 836                 *grp = node->grp;
 837
 838         if (!fd->handler)
 839                 cacheless_tid_rb_remove(fd, node);
 840         else
 841                 hfi1_mmu_rb_remove(fd->handler, &node->mmu);
 842
 843         return 0;
 844 }
 845
 846 static void clear_tid_node(struct hfi1_filedata *fd, struct tid_rb_node *node)
 847 {
 848         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 849         struct hfi1_devdata *dd = uctxt->dd;
 850
 851         trace_hfi1_exp_tid_unreg(uctxt->ctxt, fd->subctxt, node->rcventry,
 852                                  node->npages, node->mmu.addr, node->phys,
 853                                  node->dma_addr);
 854
 855         /*
 856          * Make sure device has seen the write before we unpin the
 857          * pages.
 858          */
 859         hfi1_put_tid(dd, node->rcventry, PT_INVALID_FLUSH, 0, 0);
 860
 861         unpin_rcv_pages(fd, NULL, node, 0, node->npages, true);
 862
 863         node->grp->used--;
 864         node->grp->map &= ~(1 << (node->rcventry - node->grp->base));
 865
 866         if (node->grp->used == node->grp->size - 1)
 867                 tid_group_move(node->grp, &uctxt->tid_full_list,
 868                                &uctxt->tid_used_list);
 869         else if (!node->grp->used)
 870                 tid_group_move(node->grp, &uctxt->tid_used_list,
 871                                &uctxt->tid_group_list);
 872         kfree(node);
 873 }
 874
 875 /*
 876  * As a simple helper for hfi1_user_exp_rcv_free, this function deals with
 877  * clearing nodes in the non-cached case.
 878  */
 879 static void unlock_exp_tids(struct hfi1_ctxtdata *uctxt,
 880                             struct exp_tid_set *set,
 881                             struct hfi1_filedata *fd)
 882 {
 883         struct tid_group *grp, *ptr;
 884         int i;
 885
 886         list_for_each_entry_safe(grp, ptr, &set->list, list) {
 887                 list_del_init(&grp->list);
 888
 889                 for (i = 0; i < grp->size; i++) {
 890                         if (grp->map & (1 << i)) {
 891                                 u16 rcventry = grp->base + i;
 892                                 struct tid_rb_node *node;
 893
 894                                 node = fd->entry_to_rb[rcventry -
 895                                                           uctxt->expected_base];
 896                                 if (!node || node->rcventry != rcventry)
 897                                         continue;
 898
 899                                 cacheless_tid_rb_remove(fd, node);
 900                         }
 901                 }
 902         }
 903 }
 904
 905 /*
 906  * Always return 0 from this function.  A non-zero return indicates that the
 907  * remove operation will be called and that memory should be unpinned.
 908  * However, the driver cannot unpin out from under PSM.  Instead, retain the
 909  * memory (by returning 0) and inform PSM that the memory is going away.  PSM
 910  * will call back later when it has removed the memory from its list.
 911  */
 912 static int tid_rb_invalidate(void *arg, struct mmu_rb_node *mnode)
 913 {
 914         struct hfi1_filedata *fdata = arg;
 915         struct hfi1_ctxtdata *uctxt = fdata->uctxt;
 916         struct tid_rb_node *node =
 917                 container_of(mnode, struct tid_rb_node, mmu);
 918
 919         if (node->freed)
 920                 return 0;
 921
 922         trace_hfi1_exp_tid_inval(uctxt->ctxt, fdata->subctxt, node->mmu.addr,
 923                                  node->rcventry, node->npages, node->dma_addr);
 924         node->freed = true;
 925
 926         spin_lock(&fdata->invalid_lock);
 927         if (fdata->invalid_tid_idx < uctxt->expected_count) {
 928                 fdata->invalid_tids[fdata->invalid_tid_idx] =
 929                         rcventry2tidinfo(node->rcventry - uctxt->expected_base);
 930                 fdata->invalid_tids[fdata->invalid_tid_idx] |=
 931                         EXP_TID_SET(LEN, node->npages);
 932                 if (!fdata->invalid_tid_idx) {
 933                         unsigned long *ev;
 934
 935                         /*
 936                          * hfi1_set_uevent_bits() sets a user event flag
 937                          * for all processes. Because calling into the
 938                          * driver to process TID cache invalidations is
 939                          * expensive and TID cache invalidations are
 940                          * handled on a per-process basis, we can
 941                          * optimize this to set the flag only for the
 942                          * process in question.
 943                          */
 944                         ev = uctxt->dd->events +
 945                           (((uctxt->ctxt - uctxt->dd->first_dyn_alloc_ctxt) *
 946                             HFI1_MAX_SHARED_CTXTS) + fdata->subctxt);
 947                         set_bit(_HFI1_EVENT_TID_MMU_NOTIFY_BIT, ev);
 948                 }
 949                 fdata->invalid_tid_idx++;
 950         }
 951         spin_unlock(&fdata->invalid_lock);
 952         return 0;
 953 }
 954
 955 static int tid_rb_insert(void *arg, struct mmu_rb_node *node)
 956 {
 957         struct hfi1_filedata *fdata = arg;
 958         struct tid_rb_node *tnode =
 959                 container_of(node, struct tid_rb_node, mmu);
 960         u32 base = fdata->uctxt->expected_base;
 961
 962         fdata->entry_to_rb[tnode->rcventry - base] = tnode;
 963         return 0;
 964 }
 965
 966 static void cacheless_tid_rb_remove(struct hfi1_filedata *fdata,
 967                                     struct tid_rb_node *tnode)
 968 {
 969         u32 base = fdata->uctxt->expected_base;
 970
 971         fdata->entry_to_rb[tnode->rcventry - base] = NULL;
 972         clear_tid_node(fdata, tnode);
 973 }
 974
 975 static void tid_rb_remove(void *arg, struct mmu_rb_node *node)
 976 {
 977         struct hfi1_filedata *fdata = arg;
 978         struct tid_rb_node *tnode =
 979                 container_of(node, struct tid_rb_node, mmu);
 980
 981         cacheless_tid_rb_remove(fdata, tnode);
 982 }