drivers/infiniband/hw/hfi1/user_exp_rcv.c

   1 /*
   2  * Copyright(c) 2015-2017 Intel Corporation.
   3  *
   4  * This file is provided under a dual BSD/GPLv2 license.  When using or
   5  * redistributing this file, you may do so under either license.
   6  *
   7  * GPL LICENSE SUMMARY
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of version 2 of the GNU General Public License as
  11  * published by the Free Software Foundation.
  12  *
  13  * This program is distributed in the hope that it will be useful, but
  14  * WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * General Public License for more details.
  17  *
  18  * BSD LICENSE
  19  *
  20  * Redistribution and use in source and binary forms, with or without
  21  * modification, are permitted provided that the following conditions
  22  * are met:
  23  *
  24  *  - Redistributions of source code must retain the above copyright
  25  *    notice, this list of conditions and the following disclaimer.
  26  *  - Redistributions in binary form must reproduce the above copyright
  27  *    notice, this list of conditions and the following disclaimer in
  28  *    the documentation and/or other materials provided with the
  29  *    distribution.
  30  *  - Neither the name of Intel Corporation nor the names of its
  31  *    contributors may be used to endorse or promote products derived
  32  *    from this software without specific prior written permission.
  33  *
  34  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  35  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  36  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  37  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  38  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  39  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  40  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  41  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  42  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  43  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  44  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  45  *
  46  */
  47 #include <asm/page.h>
  48 #include <linux/string.h>
  49
  50 #include "user_exp_rcv.h"
  51 #include "trace.h"
  52 #include "mmu_rb.h"
  53
  54 struct tid_rb_node {
  55         struct mmu_rb_node mmu;
  56         unsigned long phys;
  57         struct tid_group *grp;
  58         u32 rcventry;
  59         dma_addr_t dma_addr;
  60         bool freed;
  61         unsigned npages;
  62         struct page *pages[0];
  63 };
  64
  65 struct tid_pageset {
  66         u16 idx;
  67         u16 count;
  68 };
  69
  70 static void unlock_exp_tids(struct hfi1_ctxtdata *uctxt,
  71                             struct exp_tid_set *set,
  72                             struct hfi1_filedata *fd);
  73 static u32 find_phys_blocks(struct tid_user_buf *tidbuf, unsigned int npages);
  74 static int set_rcvarray_entry(struct hfi1_filedata *fd,
  75                               struct tid_user_buf *tbuf,
  76                               u32 rcventry, struct tid_group *grp,
  77                               u16 pageidx, unsigned int npages);
  78 static int tid_rb_insert(void *arg, struct mmu_rb_node *node);
  79 static void cacheless_tid_rb_remove(struct hfi1_filedata *fdata,
  80                                     struct tid_rb_node *tnode);
  81 static void tid_rb_remove(void *arg, struct mmu_rb_node *node);
  82 static int tid_rb_invalidate(void *arg, struct mmu_rb_node *mnode);
  83 static int program_rcvarray(struct hfi1_filedata *fd, struct tid_user_buf *,
  84                             struct tid_group *grp,
  85                             unsigned int start, u16 count,
  86                             u32 *tidlist, unsigned int *tididx,
  87                             unsigned int *pmapped);
  88 static int unprogram_rcvarray(struct hfi1_filedata *fd, u32 tidinfo,
  89                               struct tid_group **grp);
  90 static void clear_tid_node(struct hfi1_filedata *fd, struct tid_rb_node *node);
  91
  92 static struct mmu_rb_ops tid_rb_ops = {
  93         .insert = tid_rb_insert,
  94         .remove = tid_rb_remove,
  95         .invalidate = tid_rb_invalidate
  96 };
  97
  98 /*
  99  * Initialize context and file private data needed for Expected
 100  * receive caching. This needs to be done after the context has
 101  * been configured with the eager/expected RcvEntry counts.
 102  */
 103 int hfi1_user_exp_rcv_init(struct hfi1_filedata *fd,
 104                            struct hfi1_ctxtdata *uctxt)
 105 {
 106         struct hfi1_devdata *dd = uctxt->dd;
 107         int ret = 0;
 108
 109         spin_lock_init(&fd->tid_lock);
 110         spin_lock_init(&fd->invalid_lock);
 111
 112         fd->entry_to_rb = kcalloc(uctxt->expected_count,
 113                                   sizeof(struct rb_node *),
 114                                   GFP_KERNEL);
 115         if (!fd->entry_to_rb)
 116                 return -ENOMEM;
 117
 118         if (!HFI1_CAP_UGET_MASK(uctxt->flags, TID_UNMAP)) {
 119                 fd->invalid_tid_idx = 0;
 120                 fd->invalid_tids = kcalloc(uctxt->expected_count,
 121                                            sizeof(*fd->invalid_tids),
 122                                            GFP_KERNEL);
 123                 if (!fd->invalid_tids) {
 124                         kfree(fd->entry_to_rb);
 125                         fd->entry_to_rb = NULL;
 126                         return -ENOMEM;
 127                 }
 128
 129                 /*
 130                  * Register MMU notifier callbacks. If the registration
 131                  * fails, continue without TID caching for this context.
 132                  */
 133                 ret = hfi1_mmu_rb_register(fd, fd->mm, &tid_rb_ops,
 134                                            dd->pport->hfi1_wq,
 135                                            &fd->handler);
 136                 if (ret) {
 137                         dd_dev_info(dd,
 138                                     "Failed MMU notifier registration %d\n",
 139                                     ret);
 140                         ret = 0;
 141                 }
 142         }
 143
 144         /*
 145          * PSM does not have a good way to separate, count, and
 146          * effectively enforce a limit on RcvArray entries used by
 147          * subctxts (when context sharing is used) when TID caching
 148          * is enabled. To help with that, we calculate a per-process
 149          * RcvArray entry share and enforce that.
 150          * If TID caching is not in use, PSM deals with usage on its
 151          * own. In that case, we allow any subctxt to take all of the
 152          * entries.
 153          *
 154          * Make sure that we set the tid counts only after successful
 155          * init.
 156          */
 157         spin_lock(&fd->tid_lock);
 158         if (uctxt->subctxt_cnt && fd->handler) {
 159                 u16 remainder;
 160
 161                 fd->tid_limit = uctxt->expected_count / uctxt->subctxt_cnt;
 162                 remainder = uctxt->expected_count % uctxt->subctxt_cnt;
 163                 if (remainder && fd->subctxt < remainder)
 164                         fd->tid_limit++;
 165         } else {
 166                 fd->tid_limit = uctxt->expected_count;
 167         }
 168         spin_unlock(&fd->tid_lock);
 169
 170         return ret;
 171 }
 172
 173 void hfi1_user_exp_rcv_free(struct hfi1_filedata *fd)
 174 {
 175         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 176
 177         /*
 178          * The notifier would have been removed when the process'es mm
 179          * was freed.
 180          */
 181         if (fd->handler) {
 182                 hfi1_mmu_rb_unregister(fd->handler);
 183         } else {
 184                 if (!EXP_TID_SET_EMPTY(uctxt->tid_full_list))
 185                         unlock_exp_tids(uctxt, &uctxt->tid_full_list, fd);
 186                 if (!EXP_TID_SET_EMPTY(uctxt->tid_used_list))
 187                         unlock_exp_tids(uctxt, &uctxt->tid_used_list, fd);
 188         }
 189
 190         kfree(fd->invalid_tids);
 191         fd->invalid_tids = NULL;
 192
 193         kfree(fd->entry_to_rb);
 194         fd->entry_to_rb = NULL;
 195 }
 196
 197 /**
 198  * Release pinned receive buffer pages.
 199  *
 200  * @mapped - true if the pages have been DMA mapped. false otherwise.
 201  * @idx - Index of the first page to unpin.
 202  * @npages - No of pages to unpin.
 203  *
 204  * If the pages have been DMA mapped (indicated by mapped parameter), their
 205  * info will be passed via a struct tid_rb_node. If they haven't been mapped,
 206  * their info will be passed via a struct tid_user_buf.
 207  */
 208 static void unpin_rcv_pages(struct hfi1_filedata *fd,
 209                             struct tid_user_buf *tidbuf,
 210                             struct tid_rb_node *node,
 211                             unsigned int idx,
 212                             unsigned int npages,
 213                             bool mapped)
 214 {
 215         struct page **pages;
 216         struct hfi1_devdata *dd = fd->uctxt->dd;
 217
 218         if (mapped) {
 219                 pci_unmap_single(dd->pcidev, node->dma_addr,
 220                                  node->mmu.len, PCI_DMA_FROMDEVICE);
 221                 pages = &node->pages[idx];
 222         } else {
 223                 pages = &tidbuf->pages[idx];
 224         }
 225         hfi1_release_user_pages(fd->mm, pages, npages, mapped);
 226         fd->tid_n_pinned -= npages;
 227 }
 228
 229 /**
 230  * Pin receive buffer pages.
 231  */
 232 static int pin_rcv_pages(struct hfi1_filedata *fd, struct tid_user_buf *tidbuf)
 233 {
 234         int pinned;
 235         unsigned int npages;
 236         unsigned long vaddr = tidbuf->vaddr;
 237         struct page **pages = NULL;
 238         struct hfi1_devdata *dd = fd->uctxt->dd;
 239
 240         /* Get the number of pages the user buffer spans */
 241         npages = num_user_pages(vaddr, tidbuf->length);
 242         if (!npages)
 243                 return -EINVAL;
 244
 245         if (npages > fd->uctxt->expected_count) {
 246                 dd_dev_err(dd, "Expected buffer too big\n");
 247                 return -EINVAL;
 248         }
 249
 250         /* Verify that access is OK for the user buffer */
 251         if (!access_ok(VERIFY_WRITE, (void __user *)vaddr,
 252                        npages * PAGE_SIZE)) {
 253                 dd_dev_err(dd, "Fail vaddr %p, %u pages, !access_ok\n",
 254                            (void *)vaddr, npages);
 255                 return -EFAULT;
 256         }
 257         /* Allocate the array of struct page pointers needed for pinning */
 258         pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL);
 259         if (!pages)
 260                 return -ENOMEM;
 261
 262         /*
 263          * Pin all the pages of the user buffer. If we can't pin all the
 264          * pages, accept the amount pinned so far and program only that.
 265          * User space knows how to deal with partially programmed buffers.
 266          */
 267         if (!hfi1_can_pin_pages(dd, fd->mm, fd->tid_n_pinned, npages)) {
 268                 kfree(pages);
 269                 return -ENOMEM;
 270         }
 271
 272         pinned = hfi1_acquire_user_pages(fd->mm, vaddr, npages, true, pages);
 273         if (pinned <= 0) {
 274                 kfree(pages);
 275                 return pinned;
 276         }
 277         tidbuf->pages = pages;
 278         tidbuf->npages = npages;
 279         fd->tid_n_pinned += pinned;
 280         return pinned;
 281 }
 282
 283 /*
 284  * RcvArray entry allocation for Expected Receives is done by the
 285  * following algorithm:
 286  *
 287  * The context keeps 3 lists of groups of RcvArray entries:
 288  *   1. List of empty groups - tid_group_list
 289  *      This list is created during user context creation and
 290  *      contains elements which describe sets (of 8) of empty
 291  *      RcvArray entries.
 292  *   2. List of partially used groups - tid_used_list
 293  *      This list contains sets of RcvArray entries which are
 294  *      not completely used up. Another mapping request could
 295  *      use some of all of the remaining entries.
 296  *   3. List of full groups - tid_full_list
 297  *      This is the list where sets that are completely used
 298  *      up go.
 299  *
 300  * An attempt to optimize the usage of RcvArray entries is
 301  * made by finding all sets of physically contiguous pages in a
 302  * user's buffer.
 303  * These physically contiguous sets are further split into
 304  * sizes supported by the receive engine of the HFI. The
 305  * resulting sets of pages are stored in struct tid_pageset,
 306  * which describes the sets as:
 307  *    * .count - number of pages in this set
 308  *    * .idx - starting index into struct page ** array
 309  *                    of this set
 310  *
 311  * From this point on, the algorithm deals with the page sets
 312  * described above. The number of pagesets is divided by the
 313  * RcvArray group size to produce the number of full groups
 314  * needed.
 315  *
 316  * Groups from the 3 lists are manipulated using the following
 317  * rules:
 318  *   1. For each set of 8 pagesets, a complete group from
 319  *      tid_group_list is taken, programmed, and moved to
 320  *      the tid_full_list list.
 321  *   2. For all remaining pagesets:
 322  *      2.1 If the tid_used_list is empty and the tid_group_list
 323  *          is empty, stop processing pageset and return only
 324  *          what has been programmed up to this point.
 325  *      2.2 If the tid_used_list is empty and the tid_group_list
 326  *          is not empty, move a group from tid_group_list to
 327  *          tid_used_list.
 328  *      2.3 For each group is tid_used_group, program as much as
 329  *          can fit into the group. If the group becomes fully
 330  *          used, move it to tid_full_list.
 331  */
 332 int hfi1_user_exp_rcv_setup(struct hfi1_filedata *fd,
 333                             struct hfi1_tid_info *tinfo)
 334 {
 335         int ret = 0, need_group = 0, pinned;
 336         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 337         struct hfi1_devdata *dd = uctxt->dd;
 338         unsigned int ngroups, pageidx = 0, pageset_count,
 339                 tididx = 0, mapped, mapped_pages = 0;
 340         u32 *tidlist = NULL;
 341         struct tid_user_buf *tidbuf;
 342
 343         tidbuf = kzalloc(sizeof(*tidbuf), GFP_KERNEL);
 344         if (!tidbuf)
 345                 return -ENOMEM;
 346
 347         tidbuf->vaddr = tinfo->vaddr;
 348         tidbuf->length = tinfo->length;
 349         tidbuf->psets = kcalloc(uctxt->expected_count, sizeof(*tidbuf->psets),
 350                                 GFP_KERNEL);
 351         if (!tidbuf->psets) {
 352                 kfree(tidbuf);
 353                 return -ENOMEM;
 354         }
 355
 356         pinned = pin_rcv_pages(fd, tidbuf);
 357         if (pinned <= 0) {
 358                 kfree(tidbuf->psets);
 359                 kfree(tidbuf);
 360                 return pinned;
 361         }
 362
 363         /* Find sets of physically contiguous pages */
 364         tidbuf->n_psets = find_phys_blocks(tidbuf, pinned);
 365
 366         /*
 367          * We don't need to access this under a lock since tid_used is per
 368          * process and the same process cannot be in hfi1_user_exp_rcv_clear()
 369          * and hfi1_user_exp_rcv_setup() at the same time.
 370          */
 371         spin_lock(&fd->tid_lock);
 372         if (fd->tid_used + tidbuf->n_psets > fd->tid_limit)
 373                 pageset_count = fd->tid_limit - fd->tid_used;
 374         else
 375                 pageset_count = tidbuf->n_psets;
 376         spin_unlock(&fd->tid_lock);
 377
 378         if (!pageset_count)
 379                 goto bail;
 380
 381         ngroups = pageset_count / dd->rcv_entries.group_size;
 382         tidlist = kcalloc(pageset_count, sizeof(*tidlist), GFP_KERNEL);
 383         if (!tidlist) {
 384                 ret = -ENOMEM;
 385                 goto nomem;
 386         }
 387
 388         tididx = 0;
 389
 390         /*
 391          * From this point on, we are going to be using shared (between master
 392          * and subcontexts) context resources. We need to take the lock.
 393          */
 394         mutex_lock(&uctxt->exp_lock);
 395         /*
 396          * The first step is to program the RcvArray entries which are complete
 397          * groups.
 398          */
 399         while (ngroups && uctxt->tid_group_list.count) {
 400                 struct tid_group *grp =
 401                         tid_group_pop(&uctxt->tid_group_list);
 402
 403                 ret = program_rcvarray(fd, tidbuf, grp,
 404                                        pageidx, dd->rcv_entries.group_size,
 405                                        tidlist, &tididx, &mapped);
 406                 /*
 407                  * If there was a failure to program the RcvArray
 408                  * entries for the entire group, reset the grp fields
 409                  * and add the grp back to the free group list.
 410                  */
 411                 if (ret <= 0) {
 412                         tid_group_add_tail(grp, &uctxt->tid_group_list);
 413                         hfi1_cdbg(TID,
 414                                   "Failed to program RcvArray group %d", ret);
 415                         goto unlock;
 416                 }
 417
 418                 tid_group_add_tail(grp, &uctxt->tid_full_list);
 419                 ngroups--;
 420                 pageidx += ret;
 421                 mapped_pages += mapped;
 422         }
 423
 424         while (pageidx < pageset_count) {
 425                 struct tid_group *grp, *ptr;
 426                 /*
 427                  * If we don't have any partially used tid groups, check
 428                  * if we have empty groups. If so, take one from there and
 429                  * put in the partially used list.
 430                  */
 431                 if (!uctxt->tid_used_list.count || need_group) {
 432                         if (!uctxt->tid_group_list.count)
 433                                 goto unlock;
 434
 435                         grp = tid_group_pop(&uctxt->tid_group_list);
 436                         tid_group_add_tail(grp, &uctxt->tid_used_list);
 437                         need_group = 0;
 438                 }
 439                 /*
 440                  * There is an optimization opportunity here - instead of
 441                  * fitting as many page sets as we can, check for a group
 442                  * later on in the list that could fit all of them.
 443                  */
 444                 list_for_each_entry_safe(grp, ptr, &uctxt->tid_used_list.list,
 445                                          list) {
 446                         unsigned use = min_t(unsigned, pageset_count - pageidx,
 447                                              grp->size - grp->used);
 448
 449                         ret = program_rcvarray(fd, tidbuf, grp,
 450                                                pageidx, use, tidlist,
 451                                                &tididx, &mapped);
 452                         if (ret < 0) {
 453                                 hfi1_cdbg(TID,
 454                                           "Failed to program RcvArray entries %d",
 455                                           ret);
 456                                 ret = -EFAULT;
 457                                 goto unlock;
 458                         } else if (ret > 0) {
 459                                 if (grp->used == grp->size)
 460                                         tid_group_move(grp,
 461                                                        &uctxt->tid_used_list,
 462                                                        &uctxt->tid_full_list);
 463                                 pageidx += ret;
 464                                 mapped_pages += mapped;
 465                                 need_group = 0;
 466                                 /* Check if we are done so we break out early */
 467                                 if (pageidx >= pageset_count)
 468                                         break;
 469                         } else if (WARN_ON(ret == 0)) {
 470                                 /*
 471                                  * If ret is 0, we did not program any entries
 472                                  * into this group, which can only happen if
 473                                  * we've screwed up the accounting somewhere.
 474                                  * Warn and try to continue.
 475                                  */
 476                                 need_group = 1;
 477                         }
 478                 }
 479         }
 480 unlock:
 481         mutex_unlock(&uctxt->exp_lock);
 482 nomem:
 483         hfi1_cdbg(TID, "total mapped: tidpairs:%u pages:%u (%d)", tididx,
 484                   mapped_pages, ret);
 485         if (tididx) {
 486                 spin_lock(&fd->tid_lock);
 487                 fd->tid_used += tididx;
 488                 spin_unlock(&fd->tid_lock);
 489                 tinfo->tidcnt = tididx;
 490                 tinfo->length = mapped_pages * PAGE_SIZE;
 491
 492                 if (copy_to_user((void __user *)(unsigned long)tinfo->tidlist,
 493                                  tidlist, sizeof(tidlist[0]) * tididx)) {
 494                         /*
 495                          * On failure to copy to the user level, we need to undo
 496                          * everything done so far so we don't leak resources.
 497                          */
 498                         tinfo->tidlist = (unsigned long)&tidlist;
 499                         hfi1_user_exp_rcv_clear(fd, tinfo);
 500                         tinfo->tidlist = 0;
 501                         ret = -EFAULT;
 502                         goto bail;
 503                 }
 504         }
 505
 506         /*
 507          * If not everything was mapped (due to insufficient RcvArray entries,
 508          * for example), unpin all unmapped pages so we can pin them nex time.
 509          */
 510         if (mapped_pages != pinned)
 511                 unpin_rcv_pages(fd, tidbuf, NULL, mapped_pages,
 512                                 (pinned - mapped_pages), false);
 513 bail:
 514         kfree(tidbuf->psets);
 515         kfree(tidlist);
 516         kfree(tidbuf->pages);
 517         kfree(tidbuf);
 518         return ret > 0 ? 0 : ret;
 519 }
 520
 521 int hfi1_user_exp_rcv_clear(struct hfi1_filedata *fd,
 522                             struct hfi1_tid_info *tinfo)
 523 {
 524         int ret = 0;
 525         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 526         u32 *tidinfo;
 527         unsigned tididx;
 528
 529         if (unlikely(tinfo->tidcnt > fd->tid_used))
 530                 return -EINVAL;
 531
 532         tidinfo = memdup_user((void __user *)(unsigned long)tinfo->tidlist,
 533                               sizeof(tidinfo[0]) * tinfo->tidcnt);
 534         if (IS_ERR(tidinfo))
 535                 return PTR_ERR(tidinfo);
 536
 537         mutex_lock(&uctxt->exp_lock);
 538         for (tididx = 0; tididx < tinfo->tidcnt; tididx++) {
 539                 ret = unprogram_rcvarray(fd, tidinfo[tididx], NULL);
 540                 if (ret) {
 541                         hfi1_cdbg(TID, "Failed to unprogram rcv array %d",
 542                                   ret);
 543                         break;
 544                 }
 545         }
 546         spin_lock(&fd->tid_lock);
 547         fd->tid_used -= tididx;
 548         spin_unlock(&fd->tid_lock);
 549         tinfo->tidcnt = tididx;
 550         mutex_unlock(&uctxt->exp_lock);
 551
 552         kfree(tidinfo);
 553         return ret;
 554 }
 555
 556 int hfi1_user_exp_rcv_invalid(struct hfi1_filedata *fd,
 557                               struct hfi1_tid_info *tinfo)
 558 {
 559         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 560         unsigned long *ev = uctxt->dd->events +
 561                 (((uctxt->ctxt - uctxt->dd->first_dyn_alloc_ctxt) *
 562                   HFI1_MAX_SHARED_CTXTS) + fd->subctxt);
 563         u32 *array;
 564         int ret = 0;
 565
 566         if (!fd->invalid_tids)
 567                 return -EINVAL;
 568
 569         /*
 570          * copy_to_user() can sleep, which will leave the invalid_lock
 571          * locked and cause the MMU notifier to be blocked on the lock
 572          * for a long time.
 573          * Copy the data to a local buffer so we can release the lock.
 574          */
 575         array = kcalloc(uctxt->expected_count, sizeof(*array), GFP_KERNEL);
 576         if (!array)
 577                 return -EFAULT;
 578
 579         spin_lock(&fd->invalid_lock);
 580         if (fd->invalid_tid_idx) {
 581                 memcpy(array, fd->invalid_tids, sizeof(*array) *
 582                        fd->invalid_tid_idx);
 583                 memset(fd->invalid_tids, 0, sizeof(*fd->invalid_tids) *
 584                        fd->invalid_tid_idx);
 585                 tinfo->tidcnt = fd->invalid_tid_idx;
 586                 fd->invalid_tid_idx = 0;
 587                 /*
 588                  * Reset the user flag while still holding the lock.
 589                  * Otherwise, PSM can miss events.
 590                  */
 591                 clear_bit(_HFI1_EVENT_TID_MMU_NOTIFY_BIT, ev);
 592         } else {
 593                 tinfo->tidcnt = 0;
 594         }
 595         spin_unlock(&fd->invalid_lock);
 596
 597         if (tinfo->tidcnt) {
 598                 if (copy_to_user((void __user *)tinfo->tidlist,
 599                                  array, sizeof(*array) * tinfo->tidcnt))
 600                         ret = -EFAULT;
 601         }
 602         kfree(array);
 603
 604         return ret;
 605 }
 606
 607 static u32 find_phys_blocks(struct tid_user_buf *tidbuf, unsigned int npages)
 608 {
 609         unsigned pagecount, pageidx, setcount = 0, i;
 610         unsigned long pfn, this_pfn;
 611         struct page **pages = tidbuf->pages;
 612         struct tid_pageset *list = tidbuf->psets;
 613
 614         if (!npages)
 615                 return 0;
 616
 617         /*
 618          * Look for sets of physically contiguous pages in the user buffer.
 619          * This will allow us to optimize Expected RcvArray entry usage by
 620          * using the bigger supported sizes.
 621          */
 622         pfn = page_to_pfn(pages[0]);
 623         for (pageidx = 0, pagecount = 1, i = 1; i <= npages; i++) {
 624                 this_pfn = i < npages ? page_to_pfn(pages[i]) : 0;
 625
 626                 /*
 627                  * If the pfn's are not sequential, pages are not physically
 628                  * contiguous.
 629                  */
 630                 if (this_pfn != ++pfn) {
 631                         /*
 632                          * At this point we have to loop over the set of
 633                          * physically contiguous pages and break them down it
 634                          * sizes supported by the HW.
 635                          * There are two main constraints:
 636                          *     1. The max buffer size is MAX_EXPECTED_BUFFER.
 637                          *        If the total set size is bigger than that
 638                          *        program only a MAX_EXPECTED_BUFFER chunk.
 639                          *     2. The buffer size has to be a power of two. If
 640                          *        it is not, round down to the closes power of
 641                          *        2 and program that size.
 642                          */
 643                         while (pagecount) {
 644                                 int maxpages = pagecount;
 645                                 u32 bufsize = pagecount * PAGE_SIZE;
 646
 647                                 if (bufsize > MAX_EXPECTED_BUFFER)
 648                                         maxpages =
 649                                                 MAX_EXPECTED_BUFFER >>
 650                                                 PAGE_SHIFT;
 651                                 else if (!is_power_of_2(bufsize))
 652                                         maxpages =
 653                                                 rounddown_pow_of_two(bufsize) >>
 654                                                 PAGE_SHIFT;
 655
 656                                 list[setcount].idx = pageidx;
 657                                 list[setcount].count = maxpages;
 658                                 pagecount -= maxpages;
 659                                 pageidx += maxpages;
 660                                 setcount++;
 661                         }
 662                         pageidx = i;
 663                         pagecount = 1;
 664                         pfn = this_pfn;
 665                 } else {
 666                         pagecount++;
 667                 }
 668         }
 669         return setcount;
 670 }
 671
 672 /**
 673  * program_rcvarray() - program an RcvArray group with receive buffers
 674  * @fd: filedata pointer
 675  * @tbuf: pointer to struct tid_user_buf that has the user buffer starting
 676  *        virtual address, buffer length, page pointers, pagesets (array of
 677  *        struct tid_pageset holding information on physically contiguous
 678  *        chunks from the user buffer), and other fields.
 679  * @grp: RcvArray group
 680  * @start: starting index into sets array
 681  * @count: number of struct tid_pageset's to program
 682  * @tidlist: the array of u32 elements when the information about the
 683  *           programmed RcvArray entries is to be encoded.
 684  * @tididx: starting offset into tidlist
 685  * @pmapped: (output parameter) number of pages programmed into the RcvArray
 686  *           entries.
 687  *
 688  * This function will program up to 'count' number of RcvArray entries from the
 689  * group 'grp'. To make best use of write-combining writes, the function will
 690  * perform writes to the unused RcvArray entries which will be ignored by the
 691  * HW. Each RcvArray entry will be programmed with a physically contiguous
 692  * buffer chunk from the user's virtual buffer.
 693  *
 694  * Return:
 695  * -EINVAL if the requested count is larger than the size of the group,
 696  * -ENOMEM or -EFAULT on error from set_rcvarray_entry(), or
 697  * number of RcvArray entries programmed.
 698  */
 699 static int program_rcvarray(struct hfi1_filedata *fd, struct tid_user_buf *tbuf,
 700                             struct tid_group *grp,
 701                             unsigned int start, u16 count,
 702                             u32 *tidlist, unsigned int *tididx,
 703                             unsigned int *pmapped)
 704 {
 705         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 706         struct hfi1_devdata *dd = uctxt->dd;
 707         u16 idx;
 708         u32 tidinfo = 0, rcventry, useidx = 0;
 709         int mapped = 0;
 710
 711         /* Count should never be larger than the group size */
 712         if (count > grp->size)
 713                 return -EINVAL;
 714
 715         /* Find the first unused entry in the group */
 716         for (idx = 0; idx < grp->size; idx++) {
 717                 if (!(grp->map & (1 << idx))) {
 718                         useidx = idx;
 719                         break;
 720                 }
 721                 rcv_array_wc_fill(dd, grp->base + idx);
 722         }
 723
 724         idx = 0;
 725         while (idx < count) {
 726                 u16 npages, pageidx, setidx = start + idx;
 727                 int ret = 0;
 728
 729                 /*
 730                  * If this entry in the group is used, move to the next one.
 731                  * If we go past the end of the group, exit the loop.
 732                  */
 733                 if (useidx >= grp->size) {
 734                         break;
 735                 } else if (grp->map & (1 << useidx)) {
 736                         rcv_array_wc_fill(dd, grp->base + useidx);
 737                         useidx++;
 738                         continue;
 739                 }
 740
 741                 rcventry = grp->base + useidx;
 742                 npages = tbuf->psets[setidx].count;
 743                 pageidx = tbuf->psets[setidx].idx;
 744
 745                 ret = set_rcvarray_entry(fd, tbuf,
 746                                          rcventry, grp, pageidx,
 747                                          npages);
 748                 if (ret)
 749                         return ret;
 750                 mapped += npages;
 751
 752                 tidinfo = rcventry2tidinfo(rcventry - uctxt->expected_base) |
 753                         EXP_TID_SET(LEN, npages);
 754                 tidlist[(*tididx)++] = tidinfo;
 755                 grp->used++;
 756                 grp->map |= 1 << useidx++;
 757                 idx++;
 758         }
 759
 760         /* Fill the rest of the group with "blank" writes */
 761         for (; useidx < grp->size; useidx++)
 762                 rcv_array_wc_fill(dd, grp->base + useidx);
 763         *pmapped = mapped;
 764         return idx;
 765 }
 766
 767 static int set_rcvarray_entry(struct hfi1_filedata *fd,
 768                               struct tid_user_buf *tbuf,
 769                               u32 rcventry, struct tid_group *grp,
 770                               u16 pageidx, unsigned int npages)
 771 {
 772         int ret;
 773         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 774         struct tid_rb_node *node;
 775         struct hfi1_devdata *dd = uctxt->dd;
 776         dma_addr_t phys;
 777         struct page **pages = tbuf->pages + pageidx;
 778
 779         /*
 780          * Allocate the node first so we can handle a potential
 781          * failure before we've programmed anything.
 782          */
 783         node = kzalloc(sizeof(*node) + (sizeof(struct page *) * npages),
 784                        GFP_KERNEL);
 785         if (!node)
 786                 return -ENOMEM;
 787
 788         phys = pci_map_single(dd->pcidev,
 789                               __va(page_to_phys(pages[0])),
 790                               npages * PAGE_SIZE, PCI_DMA_FROMDEVICE);
 791         if (dma_mapping_error(&dd->pcidev->dev, phys)) {
 792                 dd_dev_err(dd, "Failed to DMA map Exp Rcv pages 0x%llx\n",
 793                            phys);
 794                 kfree(node);
 795                 return -EFAULT;
 796         }
 797
 798         node->mmu.addr = tbuf->vaddr + (pageidx * PAGE_SIZE);
 799         node->mmu.len = npages * PAGE_SIZE;
 800         node->phys = page_to_phys(pages[0]);
 801         node->npages = npages;
 802         node->rcventry = rcventry;
 803         node->dma_addr = phys;
 804         node->grp = grp;
 805         node->freed = false;
 806         memcpy(node->pages, pages, sizeof(struct page *) * npages);
 807
 808         if (!fd->handler)
 809                 ret = tid_rb_insert(fd, &node->mmu);
 810         else
 811                 ret = hfi1_mmu_rb_insert(fd->handler, &node->mmu);
 812
 813         if (ret) {
 814                 hfi1_cdbg(TID, "Failed to insert RB node %u 0x%lx, 0x%lx %d",
 815                           node->rcventry, node->mmu.addr, node->phys, ret);
 816                 pci_unmap_single(dd->pcidev, phys, npages * PAGE_SIZE,
 817                                  PCI_DMA_FROMDEVICE);
 818                 kfree(node);
 819                 return -EFAULT;
 820         }
 821         hfi1_put_tid(dd, rcventry, PT_EXPECTED, phys, ilog2(npages) + 1);
 822         trace_hfi1_exp_tid_reg(uctxt->ctxt, fd->subctxt, rcventry, npages,
 823                                node->mmu.addr, node->phys, phys);
 824         return 0;
 825 }
 826
 827 static int unprogram_rcvarray(struct hfi1_filedata *fd, u32 tidinfo,
 828                               struct tid_group **grp)
 829 {
 830         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 831         struct hfi1_devdata *dd = uctxt->dd;
 832         struct tid_rb_node *node;
 833         u8 tidctrl = EXP_TID_GET(tidinfo, CTRL);
 834         u32 tididx = EXP_TID_GET(tidinfo, IDX) << 1, rcventry;
 835
 836         if (tididx >= uctxt->expected_count) {
 837                 dd_dev_err(dd, "Invalid RcvArray entry (%u) index for ctxt %u\n",
 838                            tididx, uctxt->ctxt);
 839                 return -EINVAL;
 840         }
 841
 842         if (tidctrl == 0x3)
 843                 return -EINVAL;
 844
 845         rcventry = tididx + (tidctrl - 1);
 846
 847         node = fd->entry_to_rb[rcventry];
 848         if (!node || node->rcventry != (uctxt->expected_base + rcventry))
 849                 return -EBADF;
 850
 851         if (grp)
 852                 *grp = node->grp;
 853
 854         if (!fd->handler)
 855                 cacheless_tid_rb_remove(fd, node);
 856         else
 857                 hfi1_mmu_rb_remove(fd->handler, &node->mmu);
 858
 859         return 0;
 860 }
 861
 862 static void clear_tid_node(struct hfi1_filedata *fd, struct tid_rb_node *node)
 863 {
 864         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 865         struct hfi1_devdata *dd = uctxt->dd;
 866
 867         trace_hfi1_exp_tid_unreg(uctxt->ctxt, fd->subctxt, node->rcventry,
 868                                  node->npages, node->mmu.addr, node->phys,
 869                                  node->dma_addr);
 870
 871         /*
 872          * Make sure device has seen the write before we unpin the
 873          * pages.
 874          */
 875         hfi1_put_tid(dd, node->rcventry, PT_INVALID_FLUSH, 0, 0);
 876
 877         unpin_rcv_pages(fd, NULL, node, 0, node->npages, true);
 878
 879         node->grp->used--;
 880         node->grp->map &= ~(1 << (node->rcventry - node->grp->base));
 881
 882         if (node->grp->used == node->grp->size - 1)
 883                 tid_group_move(node->grp, &uctxt->tid_full_list,
 884                                &uctxt->tid_used_list);
 885         else if (!node->grp->used)
 886                 tid_group_move(node->grp, &uctxt->tid_used_list,
 887                                &uctxt->tid_group_list);
 888         kfree(node);
 889 }
 890
 891 /*
 892  * As a simple helper for hfi1_user_exp_rcv_free, this function deals with
 893  * clearing nodes in the non-cached case.
 894  */
 895 static void unlock_exp_tids(struct hfi1_ctxtdata *uctxt,
 896                             struct exp_tid_set *set,
 897                             struct hfi1_filedata *fd)
 898 {
 899         struct tid_group *grp, *ptr;
 900         int i;
 901
 902         list_for_each_entry_safe(grp, ptr, &set->list, list) {
 903                 list_del_init(&grp->list);
 904
 905                 for (i = 0; i < grp->size; i++) {
 906                         if (grp->map & (1 << i)) {
 907                                 u16 rcventry = grp->base + i;
 908                                 struct tid_rb_node *node;
 909
 910                                 node = fd->entry_to_rb[rcventry -
 911                                                           uctxt->expected_base];
 912                                 if (!node || node->rcventry != rcventry)
 913                                         continue;
 914
 915                                 cacheless_tid_rb_remove(fd, node);
 916                         }
 917                 }
 918         }
 919 }
 920
 921 /*
 922  * Always return 0 from this function.  A non-zero return indicates that the
 923  * remove operation will be called and that memory should be unpinned.
 924  * However, the driver cannot unpin out from under PSM.  Instead, retain the
 925  * memory (by returning 0) and inform PSM that the memory is going away.  PSM
 926  * will call back later when it has removed the memory from its list.
 927  */
 928 static int tid_rb_invalidate(void *arg, struct mmu_rb_node *mnode)
 929 {
 930         struct hfi1_filedata *fdata = arg;
 931         struct hfi1_ctxtdata *uctxt = fdata->uctxt;
 932         struct tid_rb_node *node =
 933                 container_of(mnode, struct tid_rb_node, mmu);
 934
 935         if (node->freed)
 936                 return 0;
 937
 938         trace_hfi1_exp_tid_inval(uctxt->ctxt, fdata->subctxt, node->mmu.addr,
 939                                  node->rcventry, node->npages, node->dma_addr);
 940         node->freed = true;
 941
 942         spin_lock(&fdata->invalid_lock);
 943         if (fdata->invalid_tid_idx < uctxt->expected_count) {
 944                 fdata->invalid_tids[fdata->invalid_tid_idx] =
 945                         rcventry2tidinfo(node->rcventry - uctxt->expected_base);
 946                 fdata->invalid_tids[fdata->invalid_tid_idx] |=
 947                         EXP_TID_SET(LEN, node->npages);
 948                 if (!fdata->invalid_tid_idx) {
 949                         unsigned long *ev;
 950
 951                         /*
 952                          * hfi1_set_uevent_bits() sets a user event flag
 953                          * for all processes. Because calling into the
 954                          * driver to process TID cache invalidations is
 955                          * expensive and TID cache invalidations are
 956                          * handled on a per-process basis, we can
 957                          * optimize this to set the flag only for the
 958                          * process in question.
 959                          */
 960                         ev = uctxt->dd->events +
 961                           (((uctxt->ctxt - uctxt->dd->first_dyn_alloc_ctxt) *
 962                             HFI1_MAX_SHARED_CTXTS) + fdata->subctxt);
 963                         set_bit(_HFI1_EVENT_TID_MMU_NOTIFY_BIT, ev);
 964                 }
 965                 fdata->invalid_tid_idx++;
 966         }
 967         spin_unlock(&fdata->invalid_lock);
 968         return 0;
 969 }
 970
 971 static int tid_rb_insert(void *arg, struct mmu_rb_node *node)
 972 {
 973         struct hfi1_filedata *fdata = arg;
 974         struct tid_rb_node *tnode =
 975                 container_of(node, struct tid_rb_node, mmu);
 976         u32 base = fdata->uctxt->expected_base;
 977
 978         fdata->entry_to_rb[tnode->rcventry - base] = tnode;
 979         return 0;
 980 }
 981
 982 static void cacheless_tid_rb_remove(struct hfi1_filedata *fdata,
 983                                     struct tid_rb_node *tnode)
 984 {
 985         u32 base = fdata->uctxt->expected_base;
 986
 987         fdata->entry_to_rb[tnode->rcventry - base] = NULL;
 988         clear_tid_node(fdata, tnode);
 989 }
 990
 991 static void tid_rb_remove(void *arg, struct mmu_rb_node *node)
 992 {
 993         struct hfi1_filedata *fdata = arg;
 994         struct tid_rb_node *tnode =
 995                 container_of(node, struct tid_rb_node, mmu);
 996
 997         cacheless_tid_rb_remove(fdata, tnode);
 998 }