drivers/staging/lustre/lustre/llite/vvp_page.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2015, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * Implementation of cl_page for VVP layer.
  33  *
  34  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  35  *   Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_LLITE
  39
  40 #include <linux/atomic.h>
  41 #include <linux/bitops.h>
  42 #include <linux/mm.h>
  43 #include <linux/mutex.h>
  44 #include <linux/page-flags.h>
  45 #include <linux/pagemap.h>
  46
  47 #include "../include/lustre_lite.h"
  48
  49 #include "llite_internal.h"
  50 #include "vvp_internal.h"
  51
  52 /*****************************************************************************
  53  *
  54  * Page operations.
  55  *
  56  */
  57
  58 static void vvp_page_fini_common(struct vvp_page *vpg)
  59 {
  60         struct page *vmpage = vpg->vpg_page;
  61
  62         LASSERT(vmpage);
  63         put_page(vmpage);
  64 }
  65
  66 static void vvp_page_fini(const struct lu_env *env,
  67                           struct cl_page_slice *slice)
  68 {
  69         struct vvp_page *vpg     = cl2vvp_page(slice);
  70         struct page     *vmpage  = vpg->vpg_page;
  71
  72         /*
  73          * vmpage->private was already cleared when page was moved into
  74          * VPG_FREEING state.
  75          */
  76         LASSERT((struct cl_page *)vmpage->private != slice->cpl_page);
  77         vvp_page_fini_common(vpg);
  78 }
  79
  80 static int vvp_page_own(const struct lu_env *env,
  81                         const struct cl_page_slice *slice, struct cl_io *io,
  82                         int nonblock)
  83 {
  84         struct vvp_page *vpg    = cl2vvp_page(slice);
  85         struct page     *vmpage = vpg->vpg_page;
  86
  87         LASSERT(vmpage);
  88         if (nonblock) {
  89                 if (!trylock_page(vmpage))
  90                         return -EAGAIN;
  91
  92                 if (unlikely(PageWriteback(vmpage))) {
  93                         unlock_page(vmpage);
  94                         return -EAGAIN;
  95                 }
  96
  97                 return 0;
  98         }
  99
 100         lock_page(vmpage);
 101         wait_on_page_writeback(vmpage);
 102
 103         return 0;
 104 }
 105
 106 static void vvp_page_assume(const struct lu_env *env,
 107                             const struct cl_page_slice *slice,
 108                             struct cl_io *unused)
 109 {
 110         struct page *vmpage = cl2vm_page(slice);
 111
 112         LASSERT(vmpage);
 113         LASSERT(PageLocked(vmpage));
 114         wait_on_page_writeback(vmpage);
 115 }
 116
 117 static void vvp_page_unassume(const struct lu_env *env,
 118                               const struct cl_page_slice *slice,
 119                               struct cl_io *unused)
 120 {
 121         struct page *vmpage = cl2vm_page(slice);
 122
 123         LASSERT(vmpage);
 124         LASSERT(PageLocked(vmpage));
 125 }
 126
 127 static void vvp_page_disown(const struct lu_env *env,
 128                             const struct cl_page_slice *slice, struct cl_io *io)
 129 {
 130         struct page *vmpage = cl2vm_page(slice);
 131
 132         LASSERT(vmpage);
 133         LASSERT(PageLocked(vmpage));
 134
 135         unlock_page(cl2vm_page(slice));
 136 }
 137
 138 static void vvp_page_discard(const struct lu_env *env,
 139                              const struct cl_page_slice *slice,
 140                              struct cl_io *unused)
 141 {
 142         struct page        *vmpage  = cl2vm_page(slice);
 143         struct vvp_page      *vpg     = cl2vvp_page(slice);
 144
 145         LASSERT(vmpage);
 146         LASSERT(PageLocked(vmpage));
 147
 148         if (vpg->vpg_defer_uptodate && !vpg->vpg_ra_used)
 149                 ll_ra_stats_inc(vmpage->mapping->host, RA_STAT_DISCARDED);
 150
 151         ll_invalidate_page(vmpage);
 152 }
 153
 154 static void vvp_page_delete(const struct lu_env *env,
 155                             const struct cl_page_slice *slice)
 156 {
 157         struct page       *vmpage = cl2vm_page(slice);
 158         struct inode     *inode  = vmpage->mapping->host;
 159         struct cl_object *obj    = slice->cpl_obj;
 160         struct cl_page   *page   = slice->cpl_page;
 161         int refc;
 162
 163         LASSERT(PageLocked(vmpage));
 164         LASSERT((struct cl_page *)vmpage->private == page);
 165         LASSERT(inode == vvp_object_inode(obj));
 166
 167         vvp_write_complete(cl2vvp(obj), cl2vvp_page(slice));
 168
 169         /* Drop the reference count held in vvp_page_init */
 170         refc = atomic_dec_return(&page->cp_ref);
 171         LASSERTF(refc >= 1, "page = %p, refc = %d\n", page, refc);
 172
 173         ClearPageUptodate(vmpage);
 174         ClearPagePrivate(vmpage);
 175         vmpage->private = 0;
 176         /*
 177          * Reference from vmpage to cl_page is removed, but the reference back
 178          * is still here. It is removed later in vvp_page_fini().
 179          */
 180 }
 181
 182 static void vvp_page_export(const struct lu_env *env,
 183                             const struct cl_page_slice *slice,
 184                             int uptodate)
 185 {
 186         struct page *vmpage = cl2vm_page(slice);
 187
 188         LASSERT(vmpage);
 189         LASSERT(PageLocked(vmpage));
 190         if (uptodate)
 191                 SetPageUptodate(vmpage);
 192         else
 193                 ClearPageUptodate(vmpage);
 194 }
 195
 196 static int vvp_page_is_vmlocked(const struct lu_env *env,
 197                                 const struct cl_page_slice *slice)
 198 {
 199         return PageLocked(cl2vm_page(slice)) ? -EBUSY : -ENODATA;
 200 }
 201
 202 static int vvp_page_prep_read(const struct lu_env *env,
 203                               const struct cl_page_slice *slice,
 204                               struct cl_io *unused)
 205 {
 206         /* Skip the page already marked as PG_uptodate. */
 207         return PageUptodate(cl2vm_page(slice)) ? -EALREADY : 0;
 208 }
 209
 210 static int vvp_page_prep_write(const struct lu_env *env,
 211                                const struct cl_page_slice *slice,
 212                                struct cl_io *unused)
 213 {
 214         struct page *vmpage = cl2vm_page(slice);
 215         struct cl_page *pg = slice->cpl_page;
 216
 217         LASSERT(PageLocked(vmpage));
 218         LASSERT(!PageDirty(vmpage));
 219
 220         /* ll_writepage path is not a sync write, so need to set page writeback
 221          * flag
 222          */
 223         if (!pg->cp_sync_io)
 224                 set_page_writeback(vmpage);
 225
 226         vvp_write_pending(cl2vvp(slice->cpl_obj), cl2vvp_page(slice));
 227
 228         return 0;
 229 }
 230
 231 /**
 232  * Handles page transfer errors at VM level.
 233  *
 234  * This takes inode as a separate argument, because inode on which error is to
 235  * be set can be different from \a vmpage inode in case of direct-io.
 236  */
 237 static void vvp_vmpage_error(struct inode *inode, struct page *vmpage, int ioret)
 238 {
 239         struct vvp_object *obj = cl_inode2vvp(inode);
 240
 241         if (ioret == 0) {
 242                 ClearPageError(vmpage);
 243                 obj->vob_discard_page_warned = 0;
 244         } else {
 245                 SetPageError(vmpage);
 246                 if (ioret == -ENOSPC)
 247                         set_bit(AS_ENOSPC, &inode->i_mapping->flags);
 248                 else
 249                         set_bit(AS_EIO, &inode->i_mapping->flags);
 250
 251                 if ((ioret == -ESHUTDOWN || ioret == -EINTR) &&
 252                      obj->vob_discard_page_warned == 0) {
 253                         obj->vob_discard_page_warned = 1;
 254                         ll_dirty_page_discard_warn(vmpage, ioret);
 255                 }
 256         }
 257 }
 258
 259 static void vvp_page_completion_read(const struct lu_env *env,
 260                                      const struct cl_page_slice *slice,
 261                                      int ioret)
 262 {
 263         struct vvp_page *vpg    = cl2vvp_page(slice);
 264         struct page     *vmpage = vpg->vpg_page;
 265         struct cl_page  *page   = slice->cpl_page;
 266         struct inode    *inode  = vvp_object_inode(page->cp_obj);
 267
 268         LASSERT(PageLocked(vmpage));
 269         CL_PAGE_HEADER(D_PAGE, env, page, "completing READ with %d\n", ioret);
 270
 271         if (vpg->vpg_defer_uptodate)
 272                 ll_ra_count_put(ll_i2sbi(inode), 1);
 273
 274         if (ioret == 0)  {
 275                 if (!vpg->vpg_defer_uptodate)
 276                         cl_page_export(env, page, 1);
 277         } else {
 278                 vpg->vpg_defer_uptodate = 0;
 279         }
 280
 281         if (!page->cp_sync_io)
 282                 unlock_page(vmpage);
 283 }
 284
 285 static void vvp_page_completion_write(const struct lu_env *env,
 286                                       const struct cl_page_slice *slice,
 287                                       int ioret)
 288 {
 289         struct vvp_page *vpg     = cl2vvp_page(slice);
 290         struct cl_page  *pg     = slice->cpl_page;
 291         struct page     *vmpage = vpg->vpg_page;
 292
 293         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 294
 295         /*
 296          * TODO: Actually it makes sense to add the page into oap pending
 297          * list again and so that we don't need to take the page out from
 298          * SoM write pending list, if we just meet a recoverable error,
 299          * -ENOMEM, etc.
 300          * To implement this, we just need to return a non zero value in
 301          * ->cpo_completion method. The underlying transfer should be notified
 302          * and then re-add the page into pending transfer queue.  -jay
 303          */
 304
 305         vpg->vpg_write_queued = 0;
 306         vvp_write_complete(cl2vvp(slice->cpl_obj), vpg);
 307
 308         if (pg->cp_sync_io) {
 309                 LASSERT(PageLocked(vmpage));
 310                 LASSERT(!PageWriteback(vmpage));
 311         } else {
 312                 LASSERT(PageWriteback(vmpage));
 313                 /*
 314                  * Only mark the page error only when it's an async write
 315                  * because applications won't wait for IO to finish.
 316                  */
 317                 vvp_vmpage_error(vvp_object_inode(pg->cp_obj), vmpage, ioret);
 318
 319                 end_page_writeback(vmpage);
 320         }
 321 }
 322
 323 /**
 324  * Implements cl_page_operations::cpo_make_ready() method.
 325  *
 326  * This is called to yank a page from the transfer cache and to send it out as
 327  * a part of transfer. This function try-locks the page. If try-lock failed,
 328  * page is owned by some concurrent IO, and should be skipped (this is bad,
 329  * but hopefully rare situation, as it usually results in transfer being
 330  * shorter than possible).
 331  *
 332  * \retval 0      success, page can be placed into transfer
 333  *
 334  * \retval -EAGAIN page is either used by concurrent IO has been
 335  * truncated. Skip it.
 336  */
 337 static int vvp_page_make_ready(const struct lu_env *env,
 338                                const struct cl_page_slice *slice)
 339 {
 340         struct page *vmpage = cl2vm_page(slice);
 341         struct cl_page *pg = slice->cpl_page;
 342         int result = 0;
 343
 344         lock_page(vmpage);
 345         if (clear_page_dirty_for_io(vmpage)) {
 346                 LASSERT(pg->cp_state == CPS_CACHED);
 347                 /* This actually clears the dirty bit in the radix tree. */
 348                 set_page_writeback(vmpage);
 349                 vvp_write_pending(cl2vvp(slice->cpl_obj), cl2vvp_page(slice));
 350                 CL_PAGE_HEADER(D_PAGE, env, pg, "readied\n");
 351         } else if (pg->cp_state == CPS_PAGEOUT) {
 352                 /* is it possible for osc_flush_async_page() to already
 353                  * make it ready?
 354                  */
 355                 result = -EALREADY;
 356         } else {
 357                 CL_PAGE_DEBUG(D_ERROR, env, pg, "Unexpecting page state %d.\n",
 358                               pg->cp_state);
 359                 LBUG();
 360         }
 361         unlock_page(vmpage);
 362         return result;
 363 }
 364
 365 static int vvp_page_is_under_lock(const struct lu_env *env,
 366                                   const struct cl_page_slice *slice,
 367                                   struct cl_io *io, pgoff_t *max_index)
 368 {
 369         if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE ||
 370             io->ci_type == CIT_FAULT) {
 371                 struct vvp_io *vio = vvp_env_io(env);
 372
 373                 if (unlikely(vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED))
 374                         *max_index = CL_PAGE_EOF;
 375         }
 376         return 0;
 377 }
 378
 379 static int vvp_page_print(const struct lu_env *env,
 380                           const struct cl_page_slice *slice,
 381                           void *cookie, lu_printer_t printer)
 382 {
 383         struct vvp_page *vpg = cl2vvp_page(slice);
 384         struct page     *vmpage = vpg->vpg_page;
 385
 386         (*printer)(env, cookie, LUSTRE_VVP_NAME "-page@%p(%d:%d:%d) vm@%p ",
 387                    vpg, vpg->vpg_defer_uptodate, vpg->vpg_ra_used,
 388                    vpg->vpg_write_queued, vmpage);
 389         if (vmpage) {
 390                 (*printer)(env, cookie, "%lx %d:%d %lx %lu %slru",
 391                            (long)vmpage->flags, page_count(vmpage),
 392                            page_mapcount(vmpage), vmpage->private,
 393                            vmpage->index,
 394                            list_empty(&vmpage->lru) ? "not-" : "");
 395         }
 396
 397         (*printer)(env, cookie, "\n");
 398
 399         return 0;
 400 }
 401
 402 static int vvp_page_fail(const struct lu_env *env,
 403                          const struct cl_page_slice *slice)
 404 {
 405         /*
 406          * Cached read?
 407          */
 408         LBUG();
 409
 410         return 0;
 411 }
 412
 413 static const struct cl_page_operations vvp_page_ops = {
 414         .cpo_own           = vvp_page_own,
 415         .cpo_assume     = vvp_page_assume,
 416         .cpo_unassume      = vvp_page_unassume,
 417         .cpo_disown     = vvp_page_disown,
 418         .cpo_discard       = vvp_page_discard,
 419         .cpo_delete     = vvp_page_delete,
 420         .cpo_export     = vvp_page_export,
 421         .cpo_is_vmlocked   = vvp_page_is_vmlocked,
 422         .cpo_fini         = vvp_page_fini,
 423         .cpo_print       = vvp_page_print,
 424         .cpo_is_under_lock = vvp_page_is_under_lock,
 425         .io = {
 426                 [CRT_READ] = {
 427                         .cpo_prep       = vvp_page_prep_read,
 428                         .cpo_completion  = vvp_page_completion_read,
 429                         .cpo_make_ready = vvp_page_fail,
 430                 },
 431                 [CRT_WRITE] = {
 432                         .cpo_prep       = vvp_page_prep_write,
 433                         .cpo_completion  = vvp_page_completion_write,
 434                         .cpo_make_ready  = vvp_page_make_ready,
 435                 },
 436         },
 437 };
 438
 439 static int vvp_transient_page_prep(const struct lu_env *env,
 440                                    const struct cl_page_slice *slice,
 441                                    struct cl_io *unused)
 442 {
 443         /* transient page should always be sent. */
 444         return 0;
 445 }
 446
 447 static void vvp_transient_page_verify(const struct cl_page *page)
 448 {
 449         struct inode *inode = vvp_object_inode(page->cp_obj);
 450
 451         LASSERT(!inode_trylock(inode));
 452 }
 453
 454 static int vvp_transient_page_own(const struct lu_env *env,
 455                                   const struct cl_page_slice *slice,
 456                                   struct cl_io *unused, int nonblock)
 457 {
 458         vvp_transient_page_verify(slice->cpl_page);
 459         return 0;
 460 }
 461
 462 static void vvp_transient_page_assume(const struct lu_env *env,
 463                                       const struct cl_page_slice *slice,
 464                                       struct cl_io *unused)
 465 {
 466         vvp_transient_page_verify(slice->cpl_page);
 467 }
 468
 469 static void vvp_transient_page_unassume(const struct lu_env *env,
 470                                         const struct cl_page_slice *slice,
 471                                         struct cl_io *unused)
 472 {
 473         vvp_transient_page_verify(slice->cpl_page);
 474 }
 475
 476 static void vvp_transient_page_disown(const struct lu_env *env,
 477                                       const struct cl_page_slice *slice,
 478                                       struct cl_io *unused)
 479 {
 480         vvp_transient_page_verify(slice->cpl_page);
 481 }
 482
 483 static void vvp_transient_page_discard(const struct lu_env *env,
 484                                        const struct cl_page_slice *slice,
 485                                        struct cl_io *unused)
 486 {
 487         struct cl_page *page = slice->cpl_page;
 488
 489         vvp_transient_page_verify(slice->cpl_page);
 490
 491         /*
 492          * For transient pages, remove it from the radix tree.
 493          */
 494         cl_page_delete(env, page);
 495 }
 496
 497 static int vvp_transient_page_is_vmlocked(const struct lu_env *env,
 498                                           const struct cl_page_slice *slice)
 499 {
 500         struct inode    *inode = vvp_object_inode(slice->cpl_obj);
 501         int     locked;
 502
 503         locked = !inode_trylock(inode);
 504         if (!locked)
 505                 inode_unlock(inode);
 506         return locked ? -EBUSY : -ENODATA;
 507 }
 508
 509 static void
 510 vvp_transient_page_completion(const struct lu_env *env,
 511                               const struct cl_page_slice *slice,
 512                               int ioret)
 513 {
 514         vvp_transient_page_verify(slice->cpl_page);
 515 }
 516
 517 static void vvp_transient_page_fini(const struct lu_env *env,
 518                                     struct cl_page_slice *slice)
 519 {
 520         struct vvp_page *vpg = cl2vvp_page(slice);
 521         struct cl_page *clp = slice->cpl_page;
 522         struct vvp_object *clobj = cl2vvp(clp->cp_obj);
 523
 524         vvp_page_fini_common(vpg);
 525         LASSERT(!inode_trylock(clobj->vob_inode));
 526         clobj->vob_transient_pages--;
 527 }
 528
 529 static const struct cl_page_operations vvp_transient_page_ops = {
 530         .cpo_own           = vvp_transient_page_own,
 531         .cpo_assume     = vvp_transient_page_assume,
 532         .cpo_unassume      = vvp_transient_page_unassume,
 533         .cpo_disown     = vvp_transient_page_disown,
 534         .cpo_discard       = vvp_transient_page_discard,
 535         .cpo_fini         = vvp_transient_page_fini,
 536         .cpo_is_vmlocked   = vvp_transient_page_is_vmlocked,
 537         .cpo_print       = vvp_page_print,
 538         .cpo_is_under_lock      = vvp_page_is_under_lock,
 539         .io = {
 540                 [CRT_READ] = {
 541                         .cpo_prep       = vvp_transient_page_prep,
 542                         .cpo_completion  = vvp_transient_page_completion,
 543                 },
 544                 [CRT_WRITE] = {
 545                         .cpo_prep       = vvp_transient_page_prep,
 546                         .cpo_completion  = vvp_transient_page_completion,
 547                 }
 548         }
 549 };
 550
 551 int vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 552                 struct cl_page *page, pgoff_t index)
 553 {
 554         struct vvp_page *vpg = cl_object_page_slice(obj, page);
 555         struct page     *vmpage = page->cp_vmpage;
 556
 557         CLOBINVRNT(env, obj, vvp_object_invariant(obj));
 558
 559         vpg->vpg_page = vmpage;
 560         get_page(vmpage);
 561
 562         INIT_LIST_HEAD(&vpg->vpg_pending_linkage);
 563         if (page->cp_type == CPT_CACHEABLE) {
 564                 /* in cache, decref in vvp_page_delete */
 565                 atomic_inc(&page->cp_ref);
 566                 SetPagePrivate(vmpage);
 567                 vmpage->private = (unsigned long)page;
 568                 cl_page_slice_add(page, &vpg->vpg_cl, obj, index,
 569                                   &vvp_page_ops);
 570         } else {
 571                 struct vvp_object *clobj = cl2vvp(obj);
 572
 573                 LASSERT(!inode_trylock(clobj->vob_inode));
 574                 cl_page_slice_add(page, &vpg->vpg_cl, obj, index,
 575                                   &vvp_transient_page_ops);
 576                 clobj->vob_transient_pages++;
 577         }
 578         return 0;
 579 }