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
  45 #include <obd.h>
  46 #include "llite_internal.h"
  47 #include "vvp_internal.h"
  48
  49 /*****************************************************************************
  50  *
  51  * Page operations.
  52  *
  53  */
  54
  55 static void vvp_page_fini_common(struct ccc_page *cp)
  56 {
  57         struct page *vmpage = cp->cpg_page;
  58
  59         LASSERT(vmpage != NULL);
  60         page_cache_release(vmpage);
  61 }
  62
  63 static void vvp_page_fini(const struct lu_env *env,
  64                           struct cl_page_slice *slice)
  65 {
  66         struct ccc_page *cp = cl2ccc_page(slice);
  67         struct page *vmpage  = cp->cpg_page;
  68
  69         /*
  70          * vmpage->private was already cleared when page was moved into
  71          * VPG_FREEING state.
  72          */
  73         LASSERT((struct cl_page *)vmpage->private != slice->cpl_page);
  74         vvp_page_fini_common(cp);
  75 }
  76
  77 static int vvp_page_own(const struct lu_env *env,
  78                         const struct cl_page_slice *slice, struct cl_io *io,
  79                         int nonblock)
  80 {
  81         struct ccc_page *vpg    = cl2ccc_page(slice);
  82         struct page      *vmpage = vpg->cpg_page;
  83
  84         LASSERT(vmpage != NULL);
  85         if (nonblock) {
  86                 if (!trylock_page(vmpage))
  87                         return -EAGAIN;
  88
  89                 if (unlikely(PageWriteback(vmpage))) {
  90                         unlock_page(vmpage);
  91                         return -EAGAIN;
  92                 }
  93
  94                 return 0;
  95         }
  96
  97         lock_page(vmpage);
  98         wait_on_page_writeback(vmpage);
  99         return 0;
 100 }
 101
 102 static void vvp_page_assume(const struct lu_env *env,
 103                             const struct cl_page_slice *slice,
 104                             struct cl_io *unused)
 105 {
 106         struct page *vmpage = cl2vm_page(slice);
 107
 108         LASSERT(vmpage != NULL);
 109         LASSERT(PageLocked(vmpage));
 110         wait_on_page_writeback(vmpage);
 111 }
 112
 113 static void vvp_page_unassume(const struct lu_env *env,
 114                               const struct cl_page_slice *slice,
 115                               struct cl_io *unused)
 116 {
 117         struct page *vmpage = cl2vm_page(slice);
 118
 119         LASSERT(vmpage != NULL);
 120         LASSERT(PageLocked(vmpage));
 121 }
 122
 123 static void vvp_page_disown(const struct lu_env *env,
 124                             const struct cl_page_slice *slice, struct cl_io *io)
 125 {
 126         struct page *vmpage = cl2vm_page(slice);
 127
 128         LASSERT(vmpage != NULL);
 129         LASSERT(PageLocked(vmpage));
 130
 131         unlock_page(cl2vm_page(slice));
 132 }
 133
 134 static void vvp_page_discard(const struct lu_env *env,
 135                              const struct cl_page_slice *slice,
 136                              struct cl_io *unused)
 137 {
 138         struct page     *vmpage = cl2vm_page(slice);
 139         struct ccc_page *cpg    = cl2ccc_page(slice);
 140
 141         LASSERT(vmpage != NULL);
 142         LASSERT(PageLocked(vmpage));
 143
 144         if (cpg->cpg_defer_uptodate && !cpg->cpg_ra_used)
 145                 ll_ra_stats_inc(vmpage->mapping->host, RA_STAT_DISCARDED);
 146
 147         ll_invalidate_page(vmpage);
 148 }
 149
 150 static void vvp_page_delete(const struct lu_env *env,
 151                             const struct cl_page_slice *slice)
 152 {
 153         struct page      *vmpage = cl2vm_page(slice);
 154         struct inode     *inode  = vmpage->mapping->host;
 155         struct cl_object *obj    = slice->cpl_obj;
 156         struct cl_page   *page   = slice->cpl_page;
 157         int refc;
 158
 159         LASSERT(PageLocked(vmpage));
 160         LASSERT((struct cl_page *)vmpage->private == page);
 161         LASSERT(inode == ccc_object_inode(obj));
 162
 163         vvp_write_complete(cl2ccc(obj), cl2ccc_page(slice));
 164
 165         /* Drop the reference count held in vvp_page_init */
 166         refc = atomic_dec_return(&page->cp_ref);
 167         LASSERTF(refc >= 1, "page = %p, refc = %d\n", page, refc);
 168
 169         ClearPageUptodate(vmpage);
 170         ClearPagePrivate(vmpage);
 171         vmpage->private = 0;
 172         /*
 173          * Reference from vmpage to cl_page is removed, but the reference back
 174          * is still here. It is removed later in vvp_page_fini().
 175          */
 176 }
 177
 178 static void vvp_page_export(const struct lu_env *env,
 179                             const struct cl_page_slice *slice,
 180                             int uptodate)
 181 {
 182         struct page *vmpage = cl2vm_page(slice);
 183
 184         LASSERT(vmpage != NULL);
 185         LASSERT(PageLocked(vmpage));
 186         if (uptodate)
 187                 SetPageUptodate(vmpage);
 188         else
 189                 ClearPageUptodate(vmpage);
 190 }
 191
 192 static int vvp_page_is_vmlocked(const struct lu_env *env,
 193                                 const struct cl_page_slice *slice)
 194 {
 195         return PageLocked(cl2vm_page(slice)) ? -EBUSY : -ENODATA;
 196 }
 197
 198 static int vvp_page_prep_read(const struct lu_env *env,
 199                               const struct cl_page_slice *slice,
 200                               struct cl_io *unused)
 201 {
 202         ENTRY;
 203         /* Skip the page already marked as PG_uptodate. */
 204         RETURN(PageUptodate(cl2vm_page(slice)) ? -EALREADY : 0);
 205 }
 206
 207 static int vvp_page_prep_write(const struct lu_env *env,
 208                                const struct cl_page_slice *slice,
 209                                struct cl_io *unused)
 210 {
 211         struct page *vmpage = cl2vm_page(slice);
 212
 213         LASSERT(PageLocked(vmpage));
 214         LASSERT(!PageDirty(vmpage));
 215
 216         set_page_writeback(vmpage);
 217         vvp_write_pending(cl2ccc(slice->cpl_obj), cl2ccc_page(slice));
 218
 219         return 0;
 220 }
 221
 222 /**
 223  * Handles page transfer errors at VM level.
 224  *
 225  * This takes inode as a separate argument, because inode on which error is to
 226  * be set can be different from \a vmpage inode in case of direct-io.
 227  */
 228 static void vvp_vmpage_error(struct inode *inode, struct page *vmpage, int ioret)
 229 {
 230         struct ccc_object *obj = cl_inode2ccc(inode);
 231
 232         if (ioret == 0) {
 233                 ClearPageError(vmpage);
 234                 obj->cob_discard_page_warned = 0;
 235         } else {
 236                 SetPageError(vmpage);
 237                 if (ioret == -ENOSPC)
 238                         set_bit(AS_ENOSPC, &inode->i_mapping->flags);
 239                 else
 240                         set_bit(AS_EIO, &inode->i_mapping->flags);
 241
 242                 if ((ioret == -ESHUTDOWN || ioret == -EINTR) &&
 243                      obj->cob_discard_page_warned == 0) {
 244                         obj->cob_discard_page_warned = 1;
 245                         ll_dirty_page_discard_warn(vmpage, ioret);
 246                 }
 247         }
 248 }
 249
 250 static void vvp_page_completion_read(const struct lu_env *env,
 251                                      const struct cl_page_slice *slice,
 252                                      int ioret)
 253 {
 254         struct ccc_page *cp     = cl2ccc_page(slice);
 255         struct page     *vmpage = cp->cpg_page;
 256         struct cl_page  *page   = slice->cpl_page;
 257         struct inode    *inode  = ccc_object_inode(page->cp_obj);
 258         ENTRY;
 259
 260         LASSERT(PageLocked(vmpage));
 261         CL_PAGE_HEADER(D_PAGE, env, page, "completing READ with %d\n", ioret);
 262
 263         if (cp->cpg_defer_uptodate)
 264                 ll_ra_count_put(ll_i2sbi(inode), 1);
 265
 266         if (ioret == 0)  {
 267                 if (!cp->cpg_defer_uptodate)
 268                         cl_page_export(env, page, 1);
 269         } else
 270                 cp->cpg_defer_uptodate = 0;
 271
 272         if (page->cp_sync_io == NULL)
 273                 unlock_page(vmpage);
 274
 275         EXIT;
 276 }
 277
 278 static void vvp_page_completion_write(const struct lu_env *env,
 279                                       const struct cl_page_slice *slice,
 280                                       int ioret)
 281 {
 282         struct ccc_page *cp     = cl2ccc_page(slice);
 283         struct cl_page  *pg     = slice->cpl_page;
 284         struct page      *vmpage = cp->cpg_page;
 285         ENTRY;
 286
 287         LASSERT(ergo(pg->cp_sync_io != NULL, PageLocked(vmpage)));
 288         LASSERT(PageWriteback(vmpage));
 289
 290         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 291
 292         /*
 293          * TODO: Actually it makes sense to add the page into oap pending
 294          * list again and so that we don't need to take the page out from
 295          * SoM write pending list, if we just meet a recoverable error,
 296          * -ENOMEM, etc.
 297          * To implement this, we just need to return a non zero value in
 298          * ->cpo_completion method. The underlying transfer should be notified
 299          * and then re-add the page into pending transfer queue.  -jay
 300          */
 301
 302         cp->cpg_write_queued = 0;
 303         vvp_write_complete(cl2ccc(slice->cpl_obj), cp);
 304
 305         /*
 306          * Only mark the page error only when it's an async write because
 307          * applications won't wait for IO to finish.
 308          */
 309         if (pg->cp_sync_io == NULL)
 310                 vvp_vmpage_error(ccc_object_inode(pg->cp_obj), vmpage, ioret);
 311
 312         end_page_writeback(vmpage);
 313         EXIT;
 314 }
 315
 316 /**
 317  * Implements cl_page_operations::cpo_make_ready() method.
 318  *
 319  * This is called to yank a page from the transfer cache and to send it out as
 320  * a part of transfer. This function try-locks the page. If try-lock failed,
 321  * page is owned by some concurrent IO, and should be skipped (this is bad,
 322  * but hopefully rare situation, as it usually results in transfer being
 323  * shorter than possible).
 324  *
 325  * \retval 0      success, page can be placed into transfer
 326  *
 327  * \retval -EAGAIN page is either used by concurrent IO has been
 328  * truncated. Skip it.
 329  */
 330 static int vvp_page_make_ready(const struct lu_env *env,
 331                                const struct cl_page_slice *slice)
 332 {
 333         struct page *vmpage = cl2vm_page(slice);
 334         struct cl_page *pg = slice->cpl_page;
 335         int result = 0;
 336
 337         lock_page(vmpage);
 338         if (clear_page_dirty_for_io(vmpage)) {
 339                 LASSERT(pg->cp_state == CPS_CACHED);
 340                 /* This actually clears the dirty bit in the radix
 341                  * tree. */
 342                 set_page_writeback(vmpage);
 343                 vvp_write_pending(cl2ccc(slice->cpl_obj),
 344                                 cl2ccc_page(slice));
 345                 CL_PAGE_HEADER(D_PAGE, env, pg, "readied\n");
 346         } else if (pg->cp_state == CPS_PAGEOUT) {
 347                 /* is it possible for osc_flush_async_page() to already
 348                  * make it ready? */
 349                 result = -EALREADY;
 350         } else {
 351                 CL_PAGE_DEBUG(D_ERROR, env, pg, "Unexpecting page state %d.\n",
 352                               pg->cp_state);
 353                 LBUG();
 354         }
 355         unlock_page(vmpage);
 356         RETURN(result);
 357 }
 358
 359 static int vvp_page_is_under_lock(const struct lu_env *env,
 360                                   const struct cl_page_slice *slice,
 361                                   struct cl_io *io, pgoff_t *max_index)
 362 {
 363         ENTRY;
 364
 365         if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE ||
 366             io->ci_type == CIT_FAULT) {
 367                 struct ccc_io *cio = ccc_env_io(env);
 368
 369                 if (unlikely(cio->cui_fd->fd_flags & LL_FILE_GROUP_LOCKED))
 370                         *max_index = CL_PAGE_EOF;
 371         }
 372         RETURN(0);
 373 }
 374
 375
 376 static int vvp_page_print(const struct lu_env *env,
 377                           const struct cl_page_slice *slice,
 378                           void *cookie, lu_printer_t printer)
 379 {
 380         struct ccc_page *vp = cl2ccc_page(slice);
 381         struct page      *vmpage = vp->cpg_page;
 382
 383         (*printer)(env, cookie, LUSTRE_VVP_NAME"-page@%p(%d:%d:%d) "
 384                    "vm@%p ",
 385                    vp, vp->cpg_defer_uptodate, vp->cpg_ra_used,
 386                    vp->cpg_write_queued, vmpage);
 387         if (vmpage != NULL) {
 388                 (*printer)(env, cookie, "%lx %d:%d %lx %lu %slru",
 389                            (long)vmpage->flags, page_count(vmpage),
 390                            page_mapcount(vmpage), vmpage->private,
 391                            page_index(vmpage),
 392                            list_empty(&vmpage->lru) ? "not-" : "");
 393         }
 394         (*printer)(env, cookie, "\n");
 395         return 0;
 396 }
 397
 398 static const struct cl_page_operations vvp_page_ops = {
 399         .cpo_own           = vvp_page_own,
 400         .cpo_assume        = vvp_page_assume,
 401         .cpo_unassume      = vvp_page_unassume,
 402         .cpo_disown        = vvp_page_disown,
 403         .cpo_discard       = vvp_page_discard,
 404         .cpo_delete        = vvp_page_delete,
 405         .cpo_export        = vvp_page_export,
 406         .cpo_is_vmlocked   = vvp_page_is_vmlocked,
 407         .cpo_fini          = vvp_page_fini,
 408         .cpo_print         = vvp_page_print,
 409         .cpo_is_under_lock = vvp_page_is_under_lock,
 410         .io = {
 411                 [CRT_READ] = {
 412                         .cpo_prep        = vvp_page_prep_read,
 413                         .cpo_completion  = vvp_page_completion_read,
 414                         .cpo_make_ready  = ccc_fail,
 415                 },
 416                 [CRT_WRITE] = {
 417                         .cpo_prep        = vvp_page_prep_write,
 418                         .cpo_completion  = vvp_page_completion_write,
 419                         .cpo_make_ready  = vvp_page_make_ready,
 420                 }
 421         }
 422 };
 423
 424 static void vvp_transient_page_verify(const struct cl_page *page)
 425 {
 426 }
 427
 428 static int vvp_transient_page_own(const struct lu_env *env,
 429                                   const struct cl_page_slice *slice,
 430                                   struct cl_io *unused, int nonblock)
 431 {
 432         vvp_transient_page_verify(slice->cpl_page);
 433         return 0;
 434 }
 435
 436 static void vvp_transient_page_assume(const struct lu_env *env,
 437                                       const struct cl_page_slice *slice,
 438                                       struct cl_io *unused)
 439 {
 440         vvp_transient_page_verify(slice->cpl_page);
 441 }
 442
 443 static void vvp_transient_page_unassume(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_disown(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_discard(const struct lu_env *env,
 458                                        const struct cl_page_slice *slice,
 459                                        struct cl_io *unused)
 460 {
 461         struct cl_page *page = slice->cpl_page;
 462
 463         vvp_transient_page_verify(slice->cpl_page);
 464
 465         /*
 466          * For transient pages, remove it from the radix tree.
 467          */
 468         cl_page_delete(env, page);
 469 }
 470
 471 static int vvp_transient_page_is_vmlocked(const struct lu_env *env,
 472                                           const struct cl_page_slice *slice)
 473 {
 474         struct inode    *inode = ccc_object_inode(slice->cpl_obj);
 475         int     locked;
 476
 477         locked = !mutex_trylock(&inode->i_mutex);
 478         if (!locked)
 479                 mutex_unlock(&inode->i_mutex);
 480         return locked ? -EBUSY : -ENODATA;
 481 }
 482
 483 static void
 484 vvp_transient_page_completion(const struct lu_env *env,
 485                               const struct cl_page_slice *slice,
 486                               int ioret)
 487 {
 488         vvp_transient_page_verify(slice->cpl_page);
 489 }
 490
 491 static void vvp_transient_page_fini(const struct lu_env *env,
 492                                     struct cl_page_slice *slice)
 493 {
 494         struct ccc_page *cp = cl2ccc_page(slice);
 495         struct cl_page *clp = slice->cpl_page;
 496         struct ccc_object *clobj = cl2ccc(clp->cp_obj);
 497
 498         vvp_page_fini_common(cp);
 499         atomic_dec(&clobj->cob_transient_pages);
 500 }
 501
 502 static const struct cl_page_operations vvp_transient_page_ops = {
 503         .cpo_own                = vvp_transient_page_own,
 504         .cpo_assume             = vvp_transient_page_assume,
 505         .cpo_unassume           = vvp_transient_page_unassume,
 506         .cpo_disown             = vvp_transient_page_disown,
 507         .cpo_discard            = vvp_transient_page_discard,
 508         .cpo_fini               = vvp_transient_page_fini,
 509         .cpo_is_vmlocked        = vvp_transient_page_is_vmlocked,
 510         .cpo_print              = vvp_page_print,
 511         .cpo_is_under_lock      = vvp_page_is_under_lock,
 512         .io = {
 513                 [CRT_READ] = {
 514                         .cpo_prep       = ccc_transient_page_prep,
 515                         .cpo_completion = vvp_transient_page_completion,
 516                 },
 517                 [CRT_WRITE] = {
 518                         .cpo_prep       = ccc_transient_page_prep,
 519                         .cpo_completion = vvp_transient_page_completion,
 520                 }
 521         }
 522 };
 523
 524 int vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 525                 struct cl_page *page, pgoff_t index)
 526 {
 527         struct ccc_page *cpg = cl_object_page_slice(obj, page);
 528         struct page     *vmpage = page->cp_vmpage;
 529
 530         CLOBINVRNT(env, obj, ccc_object_invariant(obj));
 531
 532         cpg->cpg_page = vmpage;
 533         page_cache_get(vmpage);
 534
 535         INIT_LIST_HEAD(&cpg->cpg_pending_linkage);
 536         if (page->cp_type == CPT_CACHEABLE) {
 537                 /* in cache, decref in vvp_page_delete */
 538                 atomic_inc(&page->cp_ref);
 539                 SetPagePrivate(vmpage);
 540                 vmpage->private = (unsigned long)page;
 541                 cl_page_slice_add(page, &cpg->cpg_cl, obj, index,
 542                                 &vvp_page_ops);
 543         } else {
 544                 struct ccc_object *clobj = cl2ccc(obj);
 545
 546                 cl_page_slice_add(page, &cpg->cpg_cl, obj, index,
 547                                 &vvp_transient_page_ops);
 548                 atomic_inc(&clobj->cob_transient_pages);
 549         }
 550         return 0;
 551 }
 552