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.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2013, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * Implementation of cl_page for VVP layer.
  37  *
  38  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  39  *   Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_LLITE
  43
  44 #ifndef __KERNEL__
  45 # error This file is kernel only.
  46 #endif
  47
  48 #include <obd.h>
  49 #include <lustre_lite.h>
  50
  51 #include "vvp_internal.h"
  52
  53 /*****************************************************************************
  54  *
  55  * Page operations.
  56  *
  57  */
  58
  59 static void vvp_page_fini_common(struct ccc_page *cp)
  60 {
  61         struct page *vmpage = cp->cpg_page;
  62
  63         LASSERT(vmpage != NULL);
  64         page_cache_release(vmpage);
  65 }
  66
  67 static void vvp_page_fini(const struct lu_env *env,
  68                           struct cl_page_slice *slice)
  69 {
  70         struct ccc_page *cp = cl2ccc_page(slice);
  71         struct page *vmpage  = cp->cpg_page;
  72
  73         /*
  74          * vmpage->private was already cleared when page was moved into
  75          * VPG_FREEING state.
  76          */
  77         LASSERT((struct cl_page *)vmpage->private != slice->cpl_page);
  78         vvp_page_fini_common(cp);
  79 }
  80
  81 static int vvp_page_own(const struct lu_env *env,
  82                         const struct cl_page_slice *slice, struct cl_io *io,
  83                         int nonblock)
  84 {
  85         struct ccc_page *vpg    = cl2ccc_page(slice);
  86         struct page      *vmpage = vpg->cpg_page;
  87
  88         LASSERT(vmpage != NULL);
  89         if (nonblock) {
  90                 if (!trylock_page(vmpage))
  91                         return -EAGAIN;
  92
  93                 if (unlikely(PageWriteback(vmpage))) {
  94                         unlock_page(vmpage);
  95                         return -EAGAIN;
  96                 }
  97
  98                 return 0;
  99         }
 100
 101         lock_page(vmpage);
 102         wait_on_page_writeback(vmpage);
 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 != NULL);
 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 != NULL);
 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 != NULL);
 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 ccc_page *cpg    = cl2ccc_page(slice);
 144
 145         LASSERT(vmpage != NULL);
 146         LASSERT(PageLocked(vmpage));
 147
 148         if (cpg->cpg_defer_uptodate && !cpg->cpg_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 == ccc_object_inode(obj));
 166
 167         vvp_write_complete(cl2ccc(obj), cl2ccc_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 != NULL);
 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         ENTRY;
 207         /* Skip the page already marked as PG_uptodate. */
 208         RETURN(PageUptodate(cl2vm_page(slice)) ? -EALREADY : 0);
 209 }
 210
 211 static int vvp_page_prep_write(const struct lu_env *env,
 212                                const struct cl_page_slice *slice,
 213                                struct cl_io *unused)
 214 {
 215         struct page *vmpage = cl2vm_page(slice);
 216
 217         LASSERT(PageLocked(vmpage));
 218         LASSERT(!PageDirty(vmpage));
 219
 220         set_page_writeback(vmpage);
 221         vvp_write_pending(cl2ccc(slice->cpl_obj), cl2ccc_page(slice));
 222
 223         return 0;
 224 }
 225
 226 /**
 227  * Handles page transfer errors at VM level.
 228  *
 229  * This takes inode as a separate argument, because inode on which error is to
 230  * be set can be different from \a vmpage inode in case of direct-io.
 231  */
 232 static void vvp_vmpage_error(struct inode *inode, struct page *vmpage, int ioret)
 233 {
 234         struct ccc_object *obj = cl_inode2ccc(inode);
 235
 236         if (ioret == 0) {
 237                 ClearPageError(vmpage);
 238                 obj->cob_discard_page_warned = 0;
 239         } else {
 240                 SetPageError(vmpage);
 241                 if (ioret == -ENOSPC)
 242                         set_bit(AS_ENOSPC, &inode->i_mapping->flags);
 243                 else
 244                         set_bit(AS_EIO, &inode->i_mapping->flags);
 245
 246                 if ((ioret == -ESHUTDOWN || ioret == -EINTR) &&
 247                      obj->cob_discard_page_warned == 0) {
 248                         obj->cob_discard_page_warned = 1;
 249                         ll_dirty_page_discard_warn(vmpage, ioret);
 250                 }
 251         }
 252 }
 253
 254 static void vvp_page_completion_read(const struct lu_env *env,
 255                                      const struct cl_page_slice *slice,
 256                                      int ioret)
 257 {
 258         struct ccc_page *cp     = cl2ccc_page(slice);
 259         struct page     *vmpage = cp->cpg_page;
 260         struct cl_page  *page   = slice->cpl_page;
 261         struct inode    *inode  = ccc_object_inode(page->cp_obj);
 262         ENTRY;
 263
 264         LASSERT(PageLocked(vmpage));
 265         CL_PAGE_HEADER(D_PAGE, env, page, "completing READ with %d\n", ioret);
 266
 267         if (cp->cpg_defer_uptodate)
 268                 ll_ra_count_put(ll_i2sbi(inode), 1);
 269
 270         if (ioret == 0)  {
 271                 if (!cp->cpg_defer_uptodate)
 272                         cl_page_export(env, page, 1);
 273         } else
 274                 cp->cpg_defer_uptodate = 0;
 275
 276         if (page->cp_sync_io == NULL)
 277                 unlock_page(vmpage);
 278
 279         EXIT;
 280 }
 281
 282 static void vvp_page_completion_write(const struct lu_env *env,
 283                                       const struct cl_page_slice *slice,
 284                                       int ioret)
 285 {
 286         struct ccc_page *cp     = cl2ccc_page(slice);
 287         struct cl_page  *pg     = slice->cpl_page;
 288         struct page      *vmpage = cp->cpg_page;
 289         ENTRY;
 290
 291         LASSERT(ergo(pg->cp_sync_io != NULL, PageLocked(vmpage)));
 292         LASSERT(PageWriteback(vmpage));
 293
 294         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 295
 296         /*
 297          * TODO: Actually it makes sense to add the page into oap pending
 298          * list again and so that we don't need to take the page out from
 299          * SoM write pending list, if we just meet a recoverable error,
 300          * -ENOMEM, etc.
 301          * To implement this, we just need to return a non zero value in
 302          * ->cpo_completion method. The underlying transfer should be notified
 303          * and then re-add the page into pending transfer queue.  -jay
 304          */
 305
 306         cp->cpg_write_queued = 0;
 307         vvp_write_complete(cl2ccc(slice->cpl_obj), cp);
 308
 309         /*
 310          * Only mark the page error only when it's an async write because
 311          * applications won't wait for IO to finish.
 312          */
 313         if (pg->cp_sync_io == NULL)
 314                 vvp_vmpage_error(ccc_object_inode(pg->cp_obj), vmpage, ioret);
 315
 316         end_page_writeback(vmpage);
 317         EXIT;
 318 }
 319
 320 /**
 321  * Implements cl_page_operations::cpo_make_ready() method.
 322  *
 323  * This is called to yank a page from the transfer cache and to send it out as
 324  * a part of transfer. This function try-locks the page. If try-lock failed,
 325  * page is owned by some concurrent IO, and should be skipped (this is bad,
 326  * but hopefully rare situation, as it usually results in transfer being
 327  * shorter than possible).
 328  *
 329  * \retval 0      success, page can be placed into transfer
 330  *
 331  * \retval -EAGAIN page is either used by concurrent IO has been
 332  * truncated. Skip it.
 333  */
 334 static int vvp_page_make_ready(const struct lu_env *env,
 335                                const struct cl_page_slice *slice)
 336 {
 337         struct page *vmpage = cl2vm_page(slice);
 338         struct cl_page *pg = slice->cpl_page;
 339         int result = 0;
 340
 341         lock_page(vmpage);
 342         if (clear_page_dirty_for_io(vmpage)) {
 343                 LASSERT(pg->cp_state == CPS_CACHED);
 344                 /* This actually clears the dirty bit in the radix
 345                  * tree. */
 346                 set_page_writeback(vmpage);
 347                 vvp_write_pending(cl2ccc(slice->cpl_obj),
 348                                 cl2ccc_page(slice));
 349                 CL_PAGE_HEADER(D_PAGE, env, pg, "readied\n");
 350         } else if (pg->cp_state == CPS_PAGEOUT) {
 351                 /* is it possible for osc_flush_async_page() to already
 352                  * make it ready? */
 353                 result = -EALREADY;
 354         } else {
 355                 CL_PAGE_DEBUG(D_ERROR, env, pg, "Unexpecting page state %d.\n",
 356                               pg->cp_state);
 357                 LBUG();
 358         }
 359         unlock_page(vmpage);
 360         RETURN(result);
 361 }
 362
 363 static int vvp_page_is_under_lock(const struct lu_env *env,
 364                                   const struct cl_page_slice *slice,
 365                                   struct cl_io *io, pgoff_t *max_index)
 366 {
 367         ENTRY;
 368
 369         if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE ||
 370             io->ci_type == CIT_FAULT) {
 371                 struct ccc_io *cio = ccc_env_io(env);
 372
 373                 if (unlikely(cio->cui_fd->fd_flags & LL_FILE_GROUP_LOCKED))
 374                         *max_index = CL_PAGE_EOF;
 375         }
 376         RETURN(0);
 377 }
 378
 379
 380 static int vvp_page_print(const struct lu_env *env,
 381                           const struct cl_page_slice *slice,
 382                           void *cookie, lu_printer_t printer)
 383 {
 384         struct ccc_page *vp = cl2ccc_page(slice);
 385         struct page      *vmpage = vp->cpg_page;
 386
 387         (*printer)(env, cookie, LUSTRE_VVP_NAME"-page@%p(%d:%d:%d) "
 388                    "vm@%p ",
 389                    vp, vp->cpg_defer_uptodate, vp->cpg_ra_used,
 390                    vp->cpg_write_queued, vmpage);
 391         if (vmpage != NULL) {
 392                 (*printer)(env, cookie, "%lx %d:%d %lx %lu %slru",
 393                            (long)vmpage->flags, page_count(vmpage),
 394                            page_mapcount(vmpage), vmpage->private,
 395                            page_index(vmpage),
 396                            list_empty(&vmpage->lru) ? "not-" : "");
 397         }
 398         (*printer)(env, cookie, "\n");
 399         return 0;
 400 }
 401
 402 static const struct cl_page_operations vvp_page_ops = {
 403         .cpo_own           = vvp_page_own,
 404         .cpo_assume        = vvp_page_assume,
 405         .cpo_unassume      = vvp_page_unassume,
 406         .cpo_disown        = vvp_page_disown,
 407         .cpo_discard       = vvp_page_discard,
 408         .cpo_delete        = vvp_page_delete,
 409         .cpo_export        = vvp_page_export,
 410         .cpo_is_vmlocked   = vvp_page_is_vmlocked,
 411         .cpo_fini          = vvp_page_fini,
 412         .cpo_print         = vvp_page_print,
 413         .cpo_is_under_lock = vvp_page_is_under_lock,
 414         .io = {
 415                 [CRT_READ] = {
 416                         .cpo_prep        = vvp_page_prep_read,
 417                         .cpo_completion  = vvp_page_completion_read,
 418                         .cpo_make_ready  = ccc_fail,
 419                 },
 420                 [CRT_WRITE] = {
 421                         .cpo_prep        = vvp_page_prep_write,
 422                         .cpo_completion  = vvp_page_completion_write,
 423                         .cpo_make_ready  = vvp_page_make_ready,
 424                 }
 425         }
 426 };
 427
 428 static void vvp_transient_page_verify(const struct cl_page *page)
 429 {
 430         struct inode *inode = ccc_object_inode(page->cp_obj);
 431
 432         LASSERT(!mutex_trylock(&inode->i_mutex));
 433 }
 434
 435 static int vvp_transient_page_own(const struct lu_env *env,
 436                                   const struct cl_page_slice *slice,
 437                                   struct cl_io *unused, int nonblock)
 438 {
 439         vvp_transient_page_verify(slice->cpl_page);
 440         return 0;
 441 }
 442
 443 static void vvp_transient_page_assume(const struct lu_env *env,
 444                                       const struct cl_page_slice *slice,
 445                                       struct cl_io *unused)
 446 {
 447         vvp_transient_page_verify(slice->cpl_page);
 448 }
 449
 450 static void vvp_transient_page_unassume(const struct lu_env *env,
 451                                         const struct cl_page_slice *slice,
 452                                         struct cl_io *unused)
 453 {
 454         vvp_transient_page_verify(slice->cpl_page);
 455 }
 456
 457 static void vvp_transient_page_disown(const struct lu_env *env,
 458                                       const struct cl_page_slice *slice,
 459                                       struct cl_io *unused)
 460 {
 461         vvp_transient_page_verify(slice->cpl_page);
 462 }
 463
 464 static void vvp_transient_page_discard(const struct lu_env *env,
 465                                        const struct cl_page_slice *slice,
 466                                        struct cl_io *unused)
 467 {
 468         struct cl_page *page = slice->cpl_page;
 469
 470         vvp_transient_page_verify(slice->cpl_page);
 471
 472         /*
 473          * For transient pages, remove it from the radix tree.
 474          */
 475         cl_page_delete(env, page);
 476 }
 477
 478 static int vvp_transient_page_is_vmlocked(const struct lu_env *env,
 479                                           const struct cl_page_slice *slice)
 480 {
 481         struct inode    *inode = ccc_object_inode(slice->cpl_obj);
 482         int     locked;
 483
 484         locked = !mutex_trylock(&inode->i_mutex);
 485         if (!locked)
 486                 mutex_unlock(&inode->i_mutex);
 487         return locked ? -EBUSY : -ENODATA;
 488 }
 489
 490 static void
 491 vvp_transient_page_completion(const struct lu_env *env,
 492                               const struct cl_page_slice *slice,
 493                               int ioret)
 494 {
 495         vvp_transient_page_verify(slice->cpl_page);
 496 }
 497
 498 static void vvp_transient_page_fini(const struct lu_env *env,
 499                                     struct cl_page_slice *slice)
 500 {
 501         struct ccc_page *cp = cl2ccc_page(slice);
 502         struct cl_page *clp = slice->cpl_page;
 503         struct ccc_object *clobj = cl2ccc(clp->cp_obj);
 504
 505         vvp_page_fini_common(cp);
 506         LASSERT(!mutex_trylock(&clobj->cob_inode->i_mutex));
 507         clobj->cob_transient_pages--;
 508 }
 509
 510 static const struct cl_page_operations vvp_transient_page_ops = {
 511         .cpo_own                = vvp_transient_page_own,
 512         .cpo_assume             = vvp_transient_page_assume,
 513         .cpo_unassume           = vvp_transient_page_unassume,
 514         .cpo_disown             = vvp_transient_page_disown,
 515         .cpo_discard            = vvp_transient_page_discard,
 516         .cpo_fini               = vvp_transient_page_fini,
 517         .cpo_is_vmlocked        = vvp_transient_page_is_vmlocked,
 518         .cpo_print              = vvp_page_print,
 519         .cpo_is_under_lock      = vvp_page_is_under_lock,
 520         .io = {
 521                 [CRT_READ] = {
 522                         .cpo_prep       = ccc_transient_page_prep,
 523                         .cpo_completion = vvp_transient_page_completion,
 524                 },
 525                 [CRT_WRITE] = {
 526                         .cpo_prep       = ccc_transient_page_prep,
 527                         .cpo_completion = vvp_transient_page_completion,
 528                 }
 529         }
 530 };
 531
 532 int vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 533                 struct cl_page *page, pgoff_t index)
 534 {
 535         struct ccc_page *cpg = cl_object_page_slice(obj, page);
 536         struct page     *vmpage = page->cp_vmpage;
 537
 538         CLOBINVRNT(env, obj, ccc_object_invariant(obj));
 539
 540         cpg->cpg_page = vmpage;
 541         page_cache_get(vmpage);
 542
 543         CFS_INIT_LIST_HEAD(&cpg->cpg_pending_linkage);
 544         if (page->cp_type == CPT_CACHEABLE) {
 545                 /* in cache, decref in vvp_page_delete */
 546                 atomic_inc(&page->cp_ref);
 547                 SetPagePrivate(vmpage);
 548                 vmpage->private = (unsigned long)page;
 549                 cl_page_slice_add(page, &cpg->cpg_cl, obj, index,
 550                                 &vvp_page_ops);
 551         } else {
 552                 struct ccc_object *clobj = cl2ccc(obj);
 553
 554                 LASSERT(!mutex_trylock(&clobj->cob_inode->i_mutex));
 555                 cl_page_slice_add(page, &cpg->cpg_cl, obj, index,
 556                                 &vvp_transient_page_ops);
 557                 clobj->cob_transient_pages++;
 558         }
 559         return 0;
 560 }
 561