drivers/staging/rdma/hfi1/user_exp_rcv.c

   1 /*
   2  * Copyright(c) 2015, 2016 Intel Corporation.
   3  *
   4  * This file is provided under a dual BSD/GPLv2 license.  When using or
   5  * redistributing this file, you may do so under either license.
   6  *
   7  * GPL LICENSE SUMMARY
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of version 2 of the GNU General Public License as
  11  * published by the Free Software Foundation.
  12  *
  13  * This program is distributed in the hope that it will be useful, but
  14  * WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * General Public License for more details.
  17  *
  18  * BSD LICENSE
  19  *
  20  * Redistribution and use in source and binary forms, with or without
  21  * modification, are permitted provided that the following conditions
  22  * are met:
  23  *
  24  *  - Redistributions of source code must retain the above copyright
  25  *    notice, this list of conditions and the following disclaimer.
  26  *  - Redistributions in binary form must reproduce the above copyright
  27  *    notice, this list of conditions and the following disclaimer in
  28  *    the documentation and/or other materials provided with the
  29  *    distribution.
  30  *  - Neither the name of Intel Corporation nor the names of its
  31  *    contributors may be used to endorse or promote products derived
  32  *    from this software without specific prior written permission.
  33  *
  34  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  35  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  36  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  37  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  38  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  39  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  40  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  41  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  42  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  43  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  44  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  45  *
  46  */
  47 #include <asm/page.h>
  48
  49 #include "user_exp_rcv.h"
  50 #include "trace.h"
  51 #include "mmu_rb.h"
  52
  53 struct tid_group {
  54         struct list_head list;
  55         unsigned base;
  56         u8 size;
  57         u8 used;
  58         u8 map;
  59 };
  60
  61 struct tid_rb_node {
  62         struct mmu_rb_node mmu;
  63         unsigned long phys;
  64         struct tid_group *grp;
  65         u32 rcventry;
  66         dma_addr_t dma_addr;
  67         bool freed;
  68         unsigned npages;
  69         struct page *pages[0];
  70 };
  71
  72 struct tid_pageset {
  73         u16 idx;
  74         u16 count;
  75 };
  76
  77 #define EXP_TID_SET_EMPTY(set) (set.count == 0 && list_empty(&set.list))
  78
  79 #define num_user_pages(vaddr, len)                                     \
  80         (1 + (((((unsigned long)(vaddr) +                              \
  81                  (unsigned long)(len) - 1) & PAGE_MASK) -              \
  82                ((unsigned long)vaddr & PAGE_MASK)) >> PAGE_SHIFT))
  83
  84 static void unlock_exp_tids(struct hfi1_ctxtdata *, struct exp_tid_set *,
  85                             struct rb_root *);
  86 static u32 find_phys_blocks(struct page **, unsigned, struct tid_pageset *);
  87 static int set_rcvarray_entry(struct file *, unsigned long, u32,
  88                               struct tid_group *, struct page **, unsigned);
  89 static int mmu_rb_insert(struct rb_root *, struct mmu_rb_node *);
  90 static void mmu_rb_remove(struct rb_root *, struct mmu_rb_node *,
  91                           struct mm_struct *);
  92 static int mmu_rb_invalidate(struct rb_root *, struct mmu_rb_node *);
  93 static int program_rcvarray(struct file *, unsigned long, struct tid_group *,
  94                             struct tid_pageset *, unsigned, u16, struct page **,
  95                             u32 *, unsigned *, unsigned *);
  96 static int unprogram_rcvarray(struct file *, u32, struct tid_group **);
  97 static void clear_tid_node(struct hfi1_filedata *, u16, struct tid_rb_node *);
  98
  99 static struct mmu_rb_ops tid_rb_ops = {
 100         .insert = mmu_rb_insert,
 101         .remove = mmu_rb_remove,
 102         .invalidate = mmu_rb_invalidate
 103 };
 104
 105 static inline u32 rcventry2tidinfo(u32 rcventry)
 106 {
 107         u32 pair = rcventry & ~0x1;
 108
 109         return EXP_TID_SET(IDX, pair >> 1) |
 110                 EXP_TID_SET(CTRL, 1 << (rcventry - pair));
 111 }
 112
 113 static inline void exp_tid_group_init(struct exp_tid_set *set)
 114 {
 115         INIT_LIST_HEAD(&set->list);
 116         set->count = 0;
 117 }
 118
 119 static inline void tid_group_remove(struct tid_group *grp,
 120                                     struct exp_tid_set *set)
 121 {
 122         list_del_init(&grp->list);
 123         set->count--;
 124 }
 125
 126 static inline void tid_group_add_tail(struct tid_group *grp,
 127                                       struct exp_tid_set *set)
 128 {
 129         list_add_tail(&grp->list, &set->list);
 130         set->count++;
 131 }
 132
 133 static inline struct tid_group *tid_group_pop(struct exp_tid_set *set)
 134 {
 135         struct tid_group *grp =
 136                 list_first_entry(&set->list, struct tid_group, list);
 137         list_del_init(&grp->list);
 138         set->count--;
 139         return grp;
 140 }
 141
 142 static inline void tid_group_move(struct tid_group *group,
 143                                   struct exp_tid_set *s1,
 144                                   struct exp_tid_set *s2)
 145 {
 146         tid_group_remove(group, s1);
 147         tid_group_add_tail(group, s2);
 148 }
 149
 150 /*
 151  * Initialize context and file private data needed for Expected
 152  * receive caching. This needs to be done after the context has
 153  * been configured with the eager/expected RcvEntry counts.
 154  */
 155 int hfi1_user_exp_rcv_init(struct file *fp)
 156 {
 157         struct hfi1_filedata *fd = fp->private_data;
 158         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 159         struct hfi1_devdata *dd = uctxt->dd;
 160         unsigned tidbase;
 161         int i, ret = 0;
 162
 163         spin_lock_init(&fd->tid_lock);
 164         spin_lock_init(&fd->invalid_lock);
 165         fd->tid_rb_root = RB_ROOT;
 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_IS_USET(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 but turn off the TID caching for
 212                  * all user contexts.
 213                  */
 214                 ret = hfi1_mmu_rb_register(&fd->tid_rb_root, &tid_rb_ops);
 215                 if (ret) {
 216                         dd_dev_info(dd,
 217                                     "Failed MMU notifier registration %d\n",
 218                                     ret);
 219                         HFI1_CAP_USET(TID_UNMAP);
 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 && !HFI1_CAP_IS_USET(TID_UNMAP)) {
 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 (!HFI1_CAP_IS_USET(TID_UNMAP))
 265                 hfi1_mmu_rb_unregister(&fd->tid_rb_root);
 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,
 272                                         &fd->tid_rb_root);
 273                 if (!EXP_TID_SET_EMPTY(uctxt->tid_used_list))
 274                         unlock_exp_tids(uctxt, &uctxt->tid_used_list,
 275                                         &fd->tid_rb_root);
 276                 list_for_each_entry_safe(grp, gptr, &uctxt->tid_group_list.list,
 277                                          list) {
 278                         list_del_init(&grp->list);
 279                         kfree(grp);
 280                 }
 281                 hfi1_clear_tids(uctxt);
 282         }
 283
 284         kfree(fd->entry_to_rb);
 285         return 0;
 286 }
 287
 288 /*
 289  * Write an "empty" RcvArray entry.
 290  * This function exists so the TID registaration code can use it
 291  * to write to unused/unneeded entries and still take advantage
 292  * of the WC performance improvements. The HFI will ignore this
 293  * write to the RcvArray entry.
 294  */
 295 static inline void rcv_array_wc_fill(struct hfi1_devdata *dd, u32 index)
 296 {
 297         /*
 298          * Doing the WC fill writes only makes sense if the device is
 299          * present and the RcvArray has been mapped as WC memory.
 300          */
 301         if ((dd->flags & HFI1_PRESENT) && dd->rcvarray_wc)
 302                 writeq(0, dd->rcvarray_wc + (index * 8));
 303 }
 304
 305 /*
 306  * RcvArray entry allocation for Expected Receives is done by the
 307  * following algorithm:
 308  *
 309  * The context keeps 3 lists of groups of RcvArray entries:
 310  *   1. List of empty groups - tid_group_list
 311  *      This list is created during user context creation and
 312  *      contains elements which describe sets (of 8) of empty
 313  *      RcvArray entries.
 314  *   2. List of partially used groups - tid_used_list
 315  *      This list contains sets of RcvArray entries which are
 316  *      not completely used up. Another mapping request could
 317  *      use some of all of the remaining entries.
 318  *   3. List of full groups - tid_full_list
 319  *      This is the list where sets that are completely used
 320  *      up go.
 321  *
 322  * An attempt to optimize the usage of RcvArray entries is
 323  * made by finding all sets of physically contiguous pages in a
 324  * user's buffer.
 325  * These physically contiguous sets are further split into
 326  * sizes supported by the receive engine of the HFI. The
 327  * resulting sets of pages are stored in struct tid_pageset,
 328  * which describes the sets as:
 329  *    * .count - number of pages in this set
 330  *    * .idx - starting index into struct page ** array
 331  *                    of this set
 332  *
 333  * From this point on, the algorithm deals with the page sets
 334  * described above. The number of pagesets is divided by the
 335  * RcvArray group size to produce the number of full groups
 336  * needed.
 337  *
 338  * Groups from the 3 lists are manipulated using the following
 339  * rules:
 340  *   1. For each set of 8 pagesets, a complete group from
 341  *      tid_group_list is taken, programmed, and moved to
 342  *      the tid_full_list list.
 343  *   2. For all remaining pagesets:
 344  *      2.1 If the tid_used_list is empty and the tid_group_list
 345  *          is empty, stop processing pageset and return only
 346  *          what has been programmed up to this point.
 347  *      2.2 If the tid_used_list is empty and the tid_group_list
 348  *          is not empty, move a group from tid_group_list to
 349  *          tid_used_list.
 350  *      2.3 For each group is tid_used_group, program as much as
 351  *          can fit into the group. If the group becomes fully
 352  *          used, move it to tid_full_list.
 353  */
 354 int hfi1_user_exp_rcv_setup(struct file *fp, struct hfi1_tid_info *tinfo)
 355 {
 356         int ret = 0, need_group = 0, pinned;
 357         struct hfi1_filedata *fd = fp->private_data;
 358         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 359         struct hfi1_devdata *dd = uctxt->dd;
 360         unsigned npages, ngroups, pageidx = 0, pageset_count, npagesets,
 361                 tididx = 0, mapped, mapped_pages = 0;
 362         unsigned long vaddr = tinfo->vaddr;
 363         struct page **pages = NULL;
 364         u32 *tidlist = NULL;
 365         struct tid_pageset *pagesets = NULL;
 366
 367         /* Get the number of pages the user buffer spans */
 368         npages = num_user_pages(vaddr, tinfo->length);
 369         if (!npages)
 370                 return -EINVAL;
 371
 372         if (npages > uctxt->expected_count) {
 373                 dd_dev_err(dd, "Expected buffer too big\n");
 374                 return -EINVAL;
 375         }
 376
 377         /* Verify that access is OK for the user buffer */
 378         if (!access_ok(VERIFY_WRITE, (void __user *)vaddr,
 379                        npages * PAGE_SIZE)) {
 380                 dd_dev_err(dd, "Fail vaddr %p, %u pages, !access_ok\n",
 381                            (void *)vaddr, npages);
 382                 return -EFAULT;
 383         }
 384
 385         pagesets = kcalloc(uctxt->expected_count, sizeof(*pagesets),
 386                            GFP_KERNEL);
 387         if (!pagesets)
 388                 return -ENOMEM;
 389
 390         /* Allocate the array of struct page pointers needed for pinning */
 391         pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL);
 392         if (!pages) {
 393                 ret = -ENOMEM;
 394                 goto bail;
 395         }
 396
 397         /*
 398          * Pin all the pages of the user buffer. If we can't pin all the
 399          * pages, accept the amount pinned so far and program only that.
 400          * User space knows how to deal with partially programmed buffers.
 401          */
 402         if (!hfi1_can_pin_pages(dd, fd->tid_n_pinned, npages))
 403                 return -ENOMEM;
 404         pinned = hfi1_acquire_user_pages(vaddr, npages, true, pages);
 405         if (pinned <= 0) {
 406                 ret = pinned;
 407                 goto bail;
 408         }
 409         fd->tid_n_pinned += npages;
 410
 411         /* Find sets of physically contiguous pages */
 412         npagesets = find_phys_blocks(pages, pinned, pagesets);
 413
 414         /*
 415          * We don't need to access this under a lock since tid_used is per
 416          * process and the same process cannot be in hfi1_user_exp_rcv_clear()
 417          * and hfi1_user_exp_rcv_setup() at the same time.
 418          */
 419         spin_lock(&fd->tid_lock);
 420         if (fd->tid_used + npagesets > fd->tid_limit)
 421                 pageset_count = fd->tid_limit - fd->tid_used;
 422         else
 423                 pageset_count = npagesets;
 424         spin_unlock(&fd->tid_lock);
 425
 426         if (!pageset_count)
 427                 goto bail;
 428
 429         ngroups = pageset_count / dd->rcv_entries.group_size;
 430         tidlist = kcalloc(pageset_count, sizeof(*tidlist), GFP_KERNEL);
 431         if (!tidlist) {
 432                 ret = -ENOMEM;
 433                 goto nomem;
 434         }
 435
 436         tididx = 0;
 437
 438         /*
 439          * From this point on, we are going to be using shared (between master
 440          * and subcontexts) context resources. We need to take the lock.
 441          */
 442         mutex_lock(&uctxt->exp_lock);
 443         /*
 444          * The first step is to program the RcvArray entries which are complete
 445          * groups.
 446          */
 447         while (ngroups && uctxt->tid_group_list.count) {
 448                 struct tid_group *grp =
 449                         tid_group_pop(&uctxt->tid_group_list);
 450
 451                 ret = program_rcvarray(fp, vaddr, grp, pagesets,
 452                                        pageidx, dd->rcv_entries.group_size,
 453                                        pages, tidlist, &tididx, &mapped);
 454                 /*
 455                  * If there was a failure to program the RcvArray
 456                  * entries for the entire group, reset the grp fields
 457                  * and add the grp back to the free group list.
 458                  */
 459                 if (ret <= 0) {
 460                         tid_group_add_tail(grp, &uctxt->tid_group_list);
 461                         hfi1_cdbg(TID,
 462                                   "Failed to program RcvArray group %d", ret);
 463                         goto unlock;
 464                 }
 465
 466                 tid_group_add_tail(grp, &uctxt->tid_full_list);
 467                 ngroups--;
 468                 pageidx += ret;
 469                 mapped_pages += mapped;
 470         }
 471
 472         while (pageidx < pageset_count) {
 473                 struct tid_group *grp, *ptr;
 474                 /*
 475                  * If we don't have any partially used tid groups, check
 476                  * if we have empty groups. If so, take one from there and
 477                  * put in the partially used list.
 478                  */
 479                 if (!uctxt->tid_used_list.count || need_group) {
 480                         if (!uctxt->tid_group_list.count)
 481                                 goto unlock;
 482
 483                         grp = tid_group_pop(&uctxt->tid_group_list);
 484                         tid_group_add_tail(grp, &uctxt->tid_used_list);
 485                         need_group = 0;
 486                 }
 487                 /*
 488                  * There is an optimization opportunity here - instead of
 489                  * fitting as many page sets as we can, check for a group
 490                  * later on in the list that could fit all of them.
 491                  */
 492                 list_for_each_entry_safe(grp, ptr, &uctxt->tid_used_list.list,
 493                                          list) {
 494                         unsigned use = min_t(unsigned, pageset_count - pageidx,
 495                                              grp->size - grp->used);
 496
 497                         ret = program_rcvarray(fp, vaddr, grp, pagesets,
 498                                                pageidx, use, pages, tidlist,
 499                                                &tididx, &mapped);
 500                         if (ret < 0) {
 501                                 hfi1_cdbg(TID,
 502                                           "Failed to program RcvArray entries %d",
 503                                           ret);
 504                                 ret = -EFAULT;
 505                                 goto unlock;
 506                         } else if (ret > 0) {
 507                                 if (grp->used == grp->size)
 508                                         tid_group_move(grp,
 509                                                        &uctxt->tid_used_list,
 510                                                        &uctxt->tid_full_list);
 511                                 pageidx += ret;
 512                                 mapped_pages += mapped;
 513                                 need_group = 0;
 514                                 /* Check if we are done so we break out early */
 515                                 if (pageidx >= pageset_count)
 516                                         break;
 517                         } else if (WARN_ON(ret == 0)) {
 518                                 /*
 519                                  * If ret is 0, we did not program any entries
 520                                  * into this group, which can only happen if
 521                                  * we've screwed up the accounting somewhere.
 522                                  * Warn and try to continue.
 523                                  */
 524                                 need_group = 1;
 525                         }
 526                 }
 527         }
 528 unlock:
 529         mutex_unlock(&uctxt->exp_lock);
 530 nomem:
 531         hfi1_cdbg(TID, "total mapped: tidpairs:%u pages:%u (%d)", tididx,
 532                   mapped_pages, ret);
 533         if (tididx) {
 534                 spin_lock(&fd->tid_lock);
 535                 fd->tid_used += tididx;
 536                 spin_unlock(&fd->tid_lock);
 537                 tinfo->tidcnt = tididx;
 538                 tinfo->length = mapped_pages * PAGE_SIZE;
 539
 540                 if (copy_to_user((void __user *)(unsigned long)tinfo->tidlist,
 541                                  tidlist, sizeof(tidlist[0]) * tididx)) {
 542                         /*
 543                          * On failure to copy to the user level, we need to undo
 544                          * everything done so far so we don't leak resources.
 545                          */
 546                         tinfo->tidlist = (unsigned long)&tidlist;
 547                         hfi1_user_exp_rcv_clear(fp, tinfo);
 548                         tinfo->tidlist = 0;
 549                         ret = -EFAULT;
 550                         goto bail;
 551                 }
 552         }
 553
 554         /*
 555          * If not everything was mapped (due to insufficient RcvArray entries,
 556          * for example), unpin all unmapped pages so we can pin them nex time.
 557          */
 558         if (mapped_pages != pinned) {
 559                 hfi1_release_user_pages(current->mm, &pages[mapped_pages],
 560                                         pinned - mapped_pages,
 561                                         false);
 562                 fd->tid_n_pinned -= pinned - mapped_pages;
 563         }
 564 bail:
 565         kfree(pagesets);
 566         kfree(pages);
 567         kfree(tidlist);
 568         return ret > 0 ? 0 : ret;
 569 }
 570
 571 int hfi1_user_exp_rcv_clear(struct file *fp, struct hfi1_tid_info *tinfo)
 572 {
 573         int ret = 0;
 574         struct hfi1_filedata *fd = fp->private_data;
 575         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 576         u32 *tidinfo;
 577         unsigned tididx;
 578
 579         tidinfo = kcalloc(tinfo->tidcnt, sizeof(*tidinfo), GFP_KERNEL);
 580         if (!tidinfo)
 581                 return -ENOMEM;
 582
 583         if (copy_from_user(tidinfo, (void __user *)(unsigned long)
 584                            tinfo->tidlist, sizeof(tidinfo[0]) *
 585                            tinfo->tidcnt)) {
 586                 ret = -EFAULT;
 587                 goto done;
 588         }
 589
 590         mutex_lock(&uctxt->exp_lock);
 591         for (tididx = 0; tididx < tinfo->tidcnt; tididx++) {
 592                 ret = unprogram_rcvarray(fp, tidinfo[tididx], NULL);
 593                 if (ret) {
 594                         hfi1_cdbg(TID, "Failed to unprogram rcv array %d",
 595                                   ret);
 596                         break;
 597                 }
 598         }
 599         spin_lock(&fd->tid_lock);
 600         fd->tid_used -= tididx;
 601         spin_unlock(&fd->tid_lock);
 602         tinfo->tidcnt = tididx;
 603         mutex_unlock(&uctxt->exp_lock);
 604 done:
 605         kfree(tidinfo);
 606         return ret;
 607 }
 608
 609 int hfi1_user_exp_rcv_invalid(struct file *fp, struct hfi1_tid_info *tinfo)
 610 {
 611         struct hfi1_filedata *fd = fp->private_data;
 612         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 613         unsigned long *ev = uctxt->dd->events +
 614                 (((uctxt->ctxt - uctxt->dd->first_user_ctxt) *
 615                   HFI1_MAX_SHARED_CTXTS) + fd->subctxt);
 616         u32 *array;
 617         int ret = 0;
 618
 619         if (!fd->invalid_tids)
 620                 return -EINVAL;
 621
 622         /*
 623          * copy_to_user() can sleep, which will leave the invalid_lock
 624          * locked and cause the MMU notifier to be blocked on the lock
 625          * for a long time.
 626          * Copy the data to a local buffer so we can release the lock.
 627          */
 628         array = kcalloc(uctxt->expected_count, sizeof(*array), GFP_KERNEL);
 629         if (!array)
 630                 return -EFAULT;
 631
 632         spin_lock(&fd->invalid_lock);
 633         if (fd->invalid_tid_idx) {
 634                 memcpy(array, fd->invalid_tids, sizeof(*array) *
 635                        fd->invalid_tid_idx);
 636                 memset(fd->invalid_tids, 0, sizeof(*fd->invalid_tids) *
 637                        fd->invalid_tid_idx);
 638                 tinfo->tidcnt = fd->invalid_tid_idx;
 639                 fd->invalid_tid_idx = 0;
 640                 /*
 641                  * Reset the user flag while still holding the lock.
 642                  * Otherwise, PSM can miss events.
 643                  */
 644                 clear_bit(_HFI1_EVENT_TID_MMU_NOTIFY_BIT, ev);
 645         } else {
 646                 tinfo->tidcnt = 0;
 647         }
 648         spin_unlock(&fd->invalid_lock);
 649
 650         if (tinfo->tidcnt) {
 651                 if (copy_to_user((void __user *)tinfo->tidlist,
 652                                  array, sizeof(*array) * tinfo->tidcnt))
 653                         ret = -EFAULT;
 654         }
 655         kfree(array);
 656
 657         return ret;
 658 }
 659
 660 static u32 find_phys_blocks(struct page **pages, unsigned npages,
 661                             struct tid_pageset *list)
 662 {
 663         unsigned pagecount, pageidx, setcount = 0, i;
 664         unsigned long pfn, this_pfn;
 665
 666         if (!npages)
 667                 return 0;
 668
 669         /*
 670          * Look for sets of physically contiguous pages in the user buffer.
 671          * This will allow us to optimize Expected RcvArray entry usage by
 672          * using the bigger supported sizes.
 673          */
 674         pfn = page_to_pfn(pages[0]);
 675         for (pageidx = 0, pagecount = 1, i = 1; i <= npages; i++) {
 676                 this_pfn = i < npages ? page_to_pfn(pages[i]) : 0;
 677
 678                 /*
 679                  * If the pfn's are not sequential, pages are not physically
 680                  * contiguous.
 681                  */
 682                 if (this_pfn != ++pfn) {
 683                         /*
 684                          * At this point we have to loop over the set of
 685                          * physically contiguous pages and break them down it
 686                          * sizes supported by the HW.
 687                          * There are two main constraints:
 688                          *     1. The max buffer size is MAX_EXPECTED_BUFFER.
 689                          *        If the total set size is bigger than that
 690                          *        program only a MAX_EXPECTED_BUFFER chunk.
 691                          *     2. The buffer size has to be a power of two. If
 692                          *        it is not, round down to the closes power of
 693                          *        2 and program that size.
 694                          */
 695                         while (pagecount) {
 696                                 int maxpages = pagecount;
 697                                 u32 bufsize = pagecount * PAGE_SIZE;
 698
 699                                 if (bufsize > MAX_EXPECTED_BUFFER)
 700                                         maxpages =
 701                                                 MAX_EXPECTED_BUFFER >>
 702                                                 PAGE_SHIFT;
 703                                 else if (!is_power_of_2(bufsize))
 704                                         maxpages =
 705                                                 rounddown_pow_of_two(bufsize) >>
 706                                                 PAGE_SHIFT;
 707
 708                                 list[setcount].idx = pageidx;
 709                                 list[setcount].count = maxpages;
 710                                 pagecount -= maxpages;
 711                                 pageidx += maxpages;
 712                                 setcount++;
 713                         }
 714                         pageidx = i;
 715                         pagecount = 1;
 716                         pfn = this_pfn;
 717                 } else {
 718                         pagecount++;
 719                 }
 720         }
 721         return setcount;
 722 }
 723
 724 /**
 725  * program_rcvarray() - program an RcvArray group with receive buffers
 726  * @fp: file pointer
 727  * @vaddr: starting user virtual address
 728  * @grp: RcvArray group
 729  * @sets: array of struct tid_pageset holding information on physically
 730  *        contiguous chunks from the user buffer
 731  * @start: starting index into sets array
 732  * @count: number of struct tid_pageset's to program
 733  * @pages: an array of struct page * for the user buffer
 734  * @tidlist: the array of u32 elements when the information about the
 735  *           programmed RcvArray entries is to be encoded.
 736  * @tididx: starting offset into tidlist
 737  * @pmapped: (output parameter) number of pages programmed into the RcvArray
 738  *           entries.
 739  *
 740  * This function will program up to 'count' number of RcvArray entries from the
 741  * group 'grp'. To make best use of write-combining writes, the function will
 742  * perform writes to the unused RcvArray entries which will be ignored by the
 743  * HW. Each RcvArray entry will be programmed with a physically contiguous
 744  * buffer chunk from the user's virtual buffer.
 745  *
 746  * Return:
 747  * -EINVAL if the requested count is larger than the size of the group,
 748  * -ENOMEM or -EFAULT on error from set_rcvarray_entry(), or
 749  * number of RcvArray entries programmed.
 750  */
 751 static int program_rcvarray(struct file *fp, unsigned long vaddr,
 752                             struct tid_group *grp,
 753                             struct tid_pageset *sets,
 754                             unsigned start, u16 count, struct page **pages,
 755                             u32 *tidlist, unsigned *tididx, unsigned *pmapped)
 756 {
 757         struct hfi1_filedata *fd = fp->private_data;
 758         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 759         struct hfi1_devdata *dd = uctxt->dd;
 760         u16 idx;
 761         u32 tidinfo = 0, rcventry, useidx = 0;
 762         int mapped = 0;
 763
 764         /* Count should never be larger than the group size */
 765         if (count > grp->size)
 766                 return -EINVAL;
 767
 768         /* Find the first unused entry in the group */
 769         for (idx = 0; idx < grp->size; idx++) {
 770                 if (!(grp->map & (1 << idx))) {
 771                         useidx = idx;
 772                         break;
 773                 }
 774                 rcv_array_wc_fill(dd, grp->base + idx);
 775         }
 776
 777         idx = 0;
 778         while (idx < count) {
 779                 u16 npages, pageidx, setidx = start + idx;
 780                 int ret = 0;
 781
 782                 /*
 783                  * If this entry in the group is used, move to the next one.
 784                  * If we go past the end of the group, exit the loop.
 785                  */
 786                 if (useidx >= grp->size) {
 787                         break;
 788                 } else if (grp->map & (1 << useidx)) {
 789                         rcv_array_wc_fill(dd, grp->base + useidx);
 790                         useidx++;
 791                         continue;
 792                 }
 793
 794                 rcventry = grp->base + useidx;
 795                 npages = sets[setidx].count;
 796                 pageidx = sets[setidx].idx;
 797
 798                 ret = set_rcvarray_entry(fp, vaddr + (pageidx * PAGE_SIZE),
 799                                          rcventry, grp, pages + pageidx,
 800                                          npages);
 801                 if (ret)
 802                         return ret;
 803                 mapped += npages;
 804
 805                 tidinfo = rcventry2tidinfo(rcventry - uctxt->expected_base) |
 806                         EXP_TID_SET(LEN, npages);
 807                 tidlist[(*tididx)++] = tidinfo;
 808                 grp->used++;
 809                 grp->map |= 1 << useidx++;
 810                 idx++;
 811         }
 812
 813         /* Fill the rest of the group with "blank" writes */
 814         for (; useidx < grp->size; useidx++)
 815                 rcv_array_wc_fill(dd, grp->base + useidx);
 816         *pmapped = mapped;
 817         return idx;
 818 }
 819
 820 static int set_rcvarray_entry(struct file *fp, unsigned long vaddr,
 821                               u32 rcventry, struct tid_group *grp,
 822                               struct page **pages, unsigned npages)
 823 {
 824         int ret;
 825         struct hfi1_filedata *fd = fp->private_data;
 826         struct hfi1_ctxtdata *uctxt = fd->uctxt;
 827         struct tid_rb_node *node;
 828         struct hfi1_devdata *dd = uctxt->dd;
 829         struct rb_root *root = &fd->tid_rb_root;
 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 (HFI1_CAP_IS_USET(TID_UNMAP))
 862                 ret = mmu_rb_insert(root, &node->mmu);
 863         else
 864                 ret = hfi1_mmu_rb_insert(root, &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         if (HFI1_CAP_IS_USET(TID_UNMAP))
 905                 mmu_rb_remove(&fd->tid_rb_root, &node->mmu, NULL);
 906         else
 907                 hfi1_mmu_rb_remove(&fd->tid_rb_root, &node->mmu);
 908
 909         if (grp)
 910                 *grp = node->grp;
 911         clear_tid_node(fd, fd->subctxt, node);
 912         return 0;
 913 }
 914
 915 static void clear_tid_node(struct hfi1_filedata *fd, u16 subctxt,
 916                            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(current->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 static void unlock_exp_tids(struct hfi1_ctxtdata *uctxt,
 950                             struct exp_tid_set *set, struct rb_root *root)
 951 {
 952         struct tid_group *grp, *ptr;
 953         struct hfi1_filedata *fd = container_of(root, struct hfi1_filedata,
 954                                                 tid_rb_root);
 955         int i;
 956
 957         list_for_each_entry_safe(grp, ptr, &set->list, list) {
 958                 list_del_init(&grp->list);
 959
 960                 for (i = 0; i < grp->size; i++) {
 961                         if (grp->map & (1 << i)) {
 962                                 u16 rcventry = grp->base + i;
 963                                 struct tid_rb_node *node;
 964
 965                                 node = fd->entry_to_rb[rcventry -
 966                                                           uctxt->expected_base];
 967                                 if (!node || node->rcventry != rcventry)
 968                                         continue;
 969                                 if (HFI1_CAP_IS_USET(TID_UNMAP))
 970                                         mmu_rb_remove(&fd->tid_rb_root,
 971                                                       &node->mmu, NULL);
 972                                 else
 973                                         hfi1_mmu_rb_remove(&fd->tid_rb_root,
 974                                                            &node->mmu);
 975                                 clear_tid_node(fd, -1, node);
 976                         }
 977                 }
 978         }
 979 }
 980
 981 static int mmu_rb_invalidate(struct rb_root *root, struct mmu_rb_node *mnode)
 982 {
 983         struct hfi1_filedata *fdata =
 984                 container_of(root, struct hfi1_filedata, tid_rb_root);
 985         struct hfi1_ctxtdata *uctxt = fdata->uctxt;
 986         struct tid_rb_node *node =
 987                 container_of(mnode, struct tid_rb_node, mmu);
 988
 989         if (node->freed)
 990                 return 0;
 991
 992         trace_hfi1_exp_tid_inval(uctxt->ctxt, fdata->subctxt, node->mmu.addr,
 993                                  node->rcventry, node->npages, node->dma_addr);
 994         node->freed = true;
 995
 996         spin_lock(&fdata->invalid_lock);
 997         if (fdata->invalid_tid_idx < uctxt->expected_count) {
 998                 fdata->invalid_tids[fdata->invalid_tid_idx] =
 999                         rcventry2tidinfo(node->rcventry - uctxt->expected_base);
1000                 fdata->invalid_tids[fdata->invalid_tid_idx] |=
1001                         EXP_TID_SET(LEN, node->npages);
1002                 if (!fdata->invalid_tid_idx) {
1003                         unsigned long *ev;
1004
1005                         /*
1006                          * hfi1_set_uevent_bits() sets a user event flag
1007                          * for all processes. Because calling into the
1008                          * driver to process TID cache invalidations is
1009                          * expensive and TID cache invalidations are
1010                          * handled on a per-process basis, we can
1011                          * optimize this to set the flag only for the
1012                          * process in question.
1013                          */
1014                         ev = uctxt->dd->events +
1015                                 (((uctxt->ctxt - uctxt->dd->first_user_ctxt) *
1016                                   HFI1_MAX_SHARED_CTXTS) + fdata->subctxt);
1017                         set_bit(_HFI1_EVENT_TID_MMU_NOTIFY_BIT, ev);
1018                 }
1019                 fdata->invalid_tid_idx++;
1020         }
1021         spin_unlock(&fdata->invalid_lock);
1022         return 0;
1023 }
1024
1025 static int mmu_rb_insert(struct rb_root *root, struct mmu_rb_node *node)
1026 {
1027         struct hfi1_filedata *fdata =
1028                 container_of(root, struct hfi1_filedata, tid_rb_root);
1029         struct tid_rb_node *tnode =
1030                 container_of(node, struct tid_rb_node, mmu);
1031         u32 base = fdata->uctxt->expected_base;
1032
1033         fdata->entry_to_rb[tnode->rcventry - base] = tnode;
1034         return 0;
1035 }
1036
1037 static void mmu_rb_remove(struct rb_root *root, struct mmu_rb_node *node,
1038                           struct mm_struct *mm)
1039 {
1040         struct hfi1_filedata *fdata =
1041                 container_of(root, struct hfi1_filedata, tid_rb_root);
1042         struct tid_rb_node *tnode =
1043                 container_of(node, struct tid_rb_node, mmu);
1044         u32 base = fdata->uctxt->expected_base;
1045
1046         fdata->entry_to_rb[tnode->rcventry - base] = NULL;
1047 }