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 <lustre_lite.h>
  47
  48 #include "vvp_internal.h"
  49
  50 /*****************************************************************************
  51  *
  52  * Page operations.
  53  *
  54  */
  55
  56 static void vvp_page_fini_common(struct ccc_page *cp)
  57 {
  58         struct page *vmpage = cp->cpg_page;
  59
  60         LASSERT(vmpage != NULL);
  61         page_cache_release(vmpage);
  62 }
  63
  64 static void vvp_page_fini(const struct lu_env *env,
  65                           struct cl_page_slice *slice)
  66 {
  67         struct ccc_page *cp = cl2ccc_page(slice);
  68         struct page *vmpage  = cp->cpg_page;
  69
  70         /*
  71          * vmpage->private was already cleared when page was moved into
  72          * VPG_FREEING state.
  73          */
  74         LASSERT((struct cl_page *)vmpage->private != slice->cpl_page);
  75         vvp_page_fini_common(cp);
  76 }
  77
  78 static int vvp_page_own(const struct lu_env *env,
  79                         const struct cl_page_slice *slice, struct cl_io *io,
  80                         int nonblock)
  81 {
  82         struct ccc_page *vpg    = cl2ccc_page(slice);
  83         struct page      *vmpage = vpg->cpg_page;
  84
  85         LASSERT(vmpage != NULL);
  86         if (nonblock) {
  87                 if (!trylock_page(vmpage))
  88                         return -EAGAIN;
  89
  90                 if (unlikely(PageWriteback(vmpage))) {
  91                         unlock_page(vmpage);
  92                         return -EAGAIN;
  93                 }
  94
  95                 return 0;
  96         }
  97
  98         lock_page(vmpage);
  99         wait_on_page_writeback(vmpage);
 100         return 0;
 101 }
 102
 103 static void vvp_page_assume(const struct lu_env *env,
 104                             const struct cl_page_slice *slice,
 105                             struct cl_io *unused)
 106 {
 107         struct page *vmpage = cl2vm_page(slice);
 108
 109         LASSERT(vmpage != NULL);
 110         LASSERT(PageLocked(vmpage));
 111         wait_on_page_writeback(vmpage);
 112 }
 113
 114 static void vvp_page_unassume(const struct lu_env *env,
 115                               const struct cl_page_slice *slice,
 116                               struct cl_io *unused)
 117 {
 118         struct page *vmpage = cl2vm_page(slice);
 119
 120         LASSERT(vmpage != NULL);
 121         LASSERT(PageLocked(vmpage));
 122 }
 123
 124 static void vvp_page_disown(const struct lu_env *env,
 125                             const struct cl_page_slice *slice, struct cl_io *io)
 126 {
 127         struct page *vmpage = cl2vm_page(slice);
 128
 129         LASSERT(vmpage != NULL);
 130         LASSERT(PageLocked(vmpage));
 131
 132         unlock_page(cl2vm_page(slice));
 133 }
 134
 135 static void vvp_page_discard(const struct lu_env *env,
 136                              const struct cl_page_slice *slice,
 137                              struct cl_io *unused)
 138 {
 139         struct page     *vmpage = cl2vm_page(slice);
 140         struct ccc_page *cpg    = cl2ccc_page(slice);
 141
 142         LASSERT(vmpage != NULL);
 143         LASSERT(PageLocked(vmpage));
 144
 145         if (cpg->cpg_defer_uptodate && !cpg->cpg_ra_used)
 146                 ll_ra_stats_inc(vmpage->mapping->host, RA_STAT_DISCARDED);
 147
 148         ll_invalidate_page(vmpage);
 149 }
 150
 151 static void vvp_page_delete(const struct lu_env *env,
 152                             const struct cl_page_slice *slice)
 153 {
 154         struct page      *vmpage = cl2vm_page(slice);
 155         struct inode     *inode  = vmpage->mapping->host;
 156         struct cl_object *obj    = slice->cpl_obj;
 157         struct cl_page   *page   = slice->cpl_page;
 158         int refc;
 159
 160         LASSERT(PageLocked(vmpage));
 161         LASSERT((struct cl_page *)vmpage->private == page);
 162         LASSERT(inode == ccc_object_inode(obj));
 163
 164         vvp_write_complete(cl2ccc(obj), cl2ccc_page(slice));
 165
 166         /* Drop the reference count held in vvp_page_init */
 167         refc = atomic_dec_return(&page->cp_ref);
 168         LASSERTF(refc >= 1, "page = %p, refc = %d\n", page, refc);
 169
 170         ClearPageUptodate(vmpage);
 171         ClearPagePrivate(vmpage);
 172         vmpage->private = 0;
 173         /*
 174          * Reference from vmpage to cl_page is removed, but the reference back
 175          * is still here. It is removed later in vvp_page_fini().
 176          */
 177 }
 178
 179 static void vvp_page_export(const struct lu_env *env,
 180                             const struct cl_page_slice *slice,
 181                             int uptodate)
 182 {
 183         struct page *vmpage = cl2vm_page(slice);
 184
 185         LASSERT(vmpage != NULL);
 186         LASSERT(PageLocked(vmpage));
 187         if (uptodate)
 188                 SetPageUptodate(vmpage);
 189         else
 190                 ClearPageUptodate(vmpage);
 191 }
 192
 193 static int vvp_page_is_vmlocked(const struct lu_env *env,
 194                                 const struct cl_page_slice *slice)
 195 {
 196         return PageLocked(cl2vm_page(slice)) ? -EBUSY : -ENODATA;
 197 }
 198
 199 static int vvp_page_prep_read(const struct lu_env *env,
 200                               const struct cl_page_slice *slice,
 201                               struct cl_io *unused)
 202 {
 203         ENTRY;
 204         /* Skip the page already marked as PG_uptodate. */
 205         RETURN(PageUptodate(cl2vm_page(slice)) ? -EALREADY : 0);
 206 }
 207
 208 static int vvp_page_prep_write(const struct lu_env *env,
 209                                const struct cl_page_slice *slice,
 210                                struct cl_io *unused)
 211 {
 212         struct page *vmpage = cl2vm_page(slice);
 213
 214         LASSERT(PageLocked(vmpage));
 215         LASSERT(!PageDirty(vmpage));
 216
 217         set_page_writeback(vmpage);
 218         vvp_write_pending(cl2ccc(slice->cpl_obj), cl2ccc_page(slice));
 219
 220         return 0;
 221 }
 222
 223 /**
 224  * Handles page transfer errors at VM level.
 225  *
 226  * This takes inode as a separate argument, because inode on which error is to
 227  * be set can be different from \a vmpage inode in case of direct-io.
 228  */
 229 static void vvp_vmpage_error(struct inode *inode, struct page *vmpage, int ioret)
 230 {
 231         struct ccc_object *obj = cl_inode2ccc(inode);
 232
 233         if (ioret == 0) {
 234                 ClearPageError(vmpage);
 235                 obj->cob_discard_page_warned = 0;
 236         } else {
 237                 SetPageError(vmpage);
 238                 if (ioret == -ENOSPC)
 239                         set_bit(AS_ENOSPC, &inode->i_mapping->flags);
 240                 else
 241                         set_bit(AS_EIO, &inode->i_mapping->flags);
 242
 243                 if ((ioret == -ESHUTDOWN || ioret == -EINTR) &&
 244                      obj->cob_discard_page_warned == 0) {
 245                         obj->cob_discard_page_warned = 1;
 246                         ll_dirty_page_discard_warn(vmpage, ioret);
 247                 }
 248         }
 249 }
 250
 251 static void vvp_page_completion_read(const struct lu_env *env,
 252                                      const struct cl_page_slice *slice,
 253                                      int ioret)
 254 {
 255         struct ccc_page *cp     = cl2ccc_page(slice);
 256         struct page     *vmpage = cp->cpg_page;
 257         struct cl_page  *page   = slice->cpl_page;
 258         struct inode    *inode  = ccc_object_inode(page->cp_obj);
 259         ENTRY;
 260
 261         LASSERT(PageLocked(vmpage));
 262         CL_PAGE_HEADER(D_PAGE, env, page, "completing READ with %d\n", ioret);
 263
 264         if (cp->cpg_defer_uptodate)
 265                 ll_ra_count_put(ll_i2sbi(inode), 1);
 266
 267         if (ioret == 0)  {
 268                 if (!cp->cpg_defer_uptodate)
 269                         cl_page_export(env, page, 1);
 270         } else
 271                 cp->cpg_defer_uptodate = 0;
 272
 273         if (page->cp_sync_io == NULL)
 274                 unlock_page(vmpage);
 275
 276         EXIT;
 277 }
 278
 279 static void vvp_page_completion_write(const struct lu_env *env,
 280                                       const struct cl_page_slice *slice,
 281                                       int ioret)
 282 {
 283         struct ccc_page *cp     = cl2ccc_page(slice);
 284         struct cl_page  *pg     = slice->cpl_page;
 285         struct page      *vmpage = cp->cpg_page;
 286         ENTRY;
 287
 288         LASSERT(ergo(pg->cp_sync_io != NULL, PageLocked(vmpage)));
 289         LASSERT(PageWriteback(vmpage));
 290
 291         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 292
 293         /*
 294          * TODO: Actually it makes sense to add the page into oap pending
 295          * list again and so that we don't need to take the page out from
 296          * SoM write pending list, if we just meet a recoverable error,
 297          * -ENOMEM, etc.
 298          * To implement this, we just need to return a non zero value in
 299          * ->cpo_completion method. The underlying transfer should be notified
 300          * and then re-add the page into pending transfer queue.  -jay
 301          */
 302
 303         cp->cpg_write_queued = 0;
 304         vvp_write_complete(cl2ccc(slice->cpl_obj), cp);
 305
 306         /*
 307          * Only mark the page error only when it's an async write because
 308          * applications won't wait for IO to finish.
 309          */
 310         if (pg->cp_sync_io == NULL)
 311                 vvp_vmpage_error(ccc_object_inode(pg->cp_obj), vmpage, ioret);
 312
 313         end_page_writeback(vmpage);
 314         EXIT;
 315 }
 316
 317 /**
 318  * Implements cl_page_operations::cpo_make_ready() method.
 319  *
 320  * This is called to yank a page from the transfer cache and to send it out as
 321  * a part of transfer. This function try-locks the page. If try-lock failed,
 322  * page is owned by some concurrent IO, and should be skipped (this is bad,
 323  * but hopefully rare situation, as it usually results in transfer being
 324  * shorter than possible).
 325  *
 326  * \retval 0      success, page can be placed into transfer
 327  *
 328  * \retval -EAGAIN page is either used by concurrent IO has been
 329  * truncated. Skip it.
 330  */
 331 static int vvp_page_make_ready(const struct lu_env *env,
 332                                const struct cl_page_slice *slice)
 333 {
 334         struct page *vmpage = cl2vm_page(slice);
 335         struct cl_page *pg = slice->cpl_page;
 336         int result = 0;
 337
 338         lock_page(vmpage);
 339         if (clear_page_dirty_for_io(vmpage)) {
 340                 LASSERT(pg->cp_state == CPS_CACHED);
 341                 /* This actually clears the dirty bit in the radix
 342                  * tree. */
 343                 set_page_writeback(vmpage);
 344                 vvp_write_pending(cl2ccc(slice->cpl_obj),
 345                                 cl2ccc_page(slice));
 346                 CL_PAGE_HEADER(D_PAGE, env, pg, "readied\n");
 347         } else if (pg->cp_state == CPS_PAGEOUT) {
 348                 /* is it possible for osc_flush_async_page() to already
 349                  * make it ready? */
 350                 result = -EALREADY;
 351         } else {
 352                 CL_PAGE_DEBUG(D_ERROR, env, pg, "Unexpecting page state %d.\n",
 353                               pg->cp_state);
 354                 LBUG();
 355         }
 356         unlock_page(vmpage);
 357         RETURN(result);
 358 }
 359
 360 static int vvp_page_is_under_lock(const struct lu_env *env,
 361                                   const struct cl_page_slice *slice,
 362                                   struct cl_io *io, pgoff_t *max_index)
 363 {
 364         ENTRY;
 365
 366         if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE ||
 367             io->ci_type == CIT_FAULT) {
 368                 struct ccc_io *cio = ccc_env_io(env);
 369
 370                 if (unlikely(cio->cui_fd->fd_flags & LL_FILE_GROUP_LOCKED))
 371                         *max_index = CL_PAGE_EOF;
 372         }
 373         RETURN(0);
 374 }
 375
 376
 377 static int vvp_page_print(const struct lu_env *env,
 378                           const struct cl_page_slice *slice,
 379                           void *cookie, lu_printer_t printer)
 380 {
 381         struct ccc_page *vp = cl2ccc_page(slice);
 382         struct page      *vmpage = vp->cpg_page;
 383
 384         (*printer)(env, cookie, LUSTRE_VVP_NAME"-page@%p(%d:%d:%d) "
 385                    "vm@%p ",
 386                    vp, vp->cpg_defer_uptodate, vp->cpg_ra_used,
 387                    vp->cpg_write_queued, vmpage);
 388         if (vmpage != NULL) {
 389                 (*printer)(env, cookie, "%lx %d:%d %lx %lu %slru",
 390                            (long)vmpage->flags, page_count(vmpage),
 391                            page_mapcount(vmpage), vmpage->private,
 392                            page_index(vmpage),
 393                            list_empty(&vmpage->lru) ? "not-" : "");
 394         }
 395         (*printer)(env, cookie, "\n");
 396         return 0;
 397 }
 398
 399 static const struct cl_page_operations vvp_page_ops = {
 400         .cpo_own           = vvp_page_own,
 401         .cpo_assume        = vvp_page_assume,
 402         .cpo_unassume      = vvp_page_unassume,
 403         .cpo_disown        = vvp_page_disown,
 404         .cpo_discard       = vvp_page_discard,
 405         .cpo_delete        = vvp_page_delete,
 406         .cpo_export        = vvp_page_export,
 407         .cpo_is_vmlocked   = vvp_page_is_vmlocked,
 408         .cpo_fini          = vvp_page_fini,
 409         .cpo_print         = vvp_page_print,
 410         .cpo_is_under_lock = vvp_page_is_under_lock,
 411         .io = {
 412                 [CRT_READ] = {
 413                         .cpo_prep        = vvp_page_prep_read,
 414                         .cpo_completion  = vvp_page_completion_read,
 415                         .cpo_make_ready  = ccc_fail,
 416                 },
 417                 [CRT_WRITE] = {
 418                         .cpo_prep        = vvp_page_prep_write,
 419                         .cpo_completion  = vvp_page_completion_write,
 420                         .cpo_make_ready  = vvp_page_make_ready,
 421                 }
 422         }
 423 };
 424
 425 static void vvp_transient_page_verify(const struct cl_page *page)
 426 {
 427         struct inode *inode = ccc_object_inode(page->cp_obj);
 428
 429         LASSERT(!mutex_trylock(&inode->i_mutex));
 430 }
 431
 432 static int vvp_transient_page_own(const struct lu_env *env,
 433                                   const struct cl_page_slice *slice,
 434                                   struct cl_io *unused, int nonblock)
 435 {
 436         vvp_transient_page_verify(slice->cpl_page);
 437         return 0;
 438 }
 439
 440 static void vvp_transient_page_assume(const struct lu_env *env,
 441                                       const struct cl_page_slice *slice,
 442                                       struct cl_io *unused)
 443 {
 444         vvp_transient_page_verify(slice->cpl_page);
 445 }
 446
 447 static void vvp_transient_page_unassume(const struct lu_env *env,
 448                                         const struct cl_page_slice *slice,
 449                                         struct cl_io *unused)
 450 {
 451         vvp_transient_page_verify(slice->cpl_page);
 452 }
 453
 454 static void vvp_transient_page_disown(const struct lu_env *env,
 455                                       const struct cl_page_slice *slice,
 456                                       struct cl_io *unused)
 457 {
 458         vvp_transient_page_verify(slice->cpl_page);
 459 }
 460
 461 static void vvp_transient_page_discard(const struct lu_env *env,
 462                                        const struct cl_page_slice *slice,
 463                                        struct cl_io *unused)
 464 {
 465         struct cl_page *page = slice->cpl_page;
 466
 467         vvp_transient_page_verify(slice->cpl_page);
 468
 469         /*
 470          * For transient pages, remove it from the radix tree.
 471          */
 472         cl_page_delete(env, page);
 473 }
 474
 475 static int vvp_transient_page_is_vmlocked(const struct lu_env *env,
 476                                           const struct cl_page_slice *slice)
 477 {
 478         struct inode    *inode = ccc_object_inode(slice->cpl_obj);
 479         int     locked;
 480
 481         locked = !mutex_trylock(&inode->i_mutex);
 482         if (!locked)
 483                 mutex_unlock(&inode->i_mutex);
 484         return locked ? -EBUSY : -ENODATA;
 485 }
 486
 487 static void
 488 vvp_transient_page_completion(const struct lu_env *env,
 489                               const struct cl_page_slice *slice,
 490                               int ioret)
 491 {
 492         vvp_transient_page_verify(slice->cpl_page);
 493 }
 494
 495 static void vvp_transient_page_fini(const struct lu_env *env,
 496                                     struct cl_page_slice *slice)
 497 {
 498         struct ccc_page *cp = cl2ccc_page(slice);
 499         struct cl_page *clp = slice->cpl_page;
 500         struct ccc_object *clobj = cl2ccc(clp->cp_obj);
 501
 502         vvp_page_fini_common(cp);
 503         LASSERT(!mutex_trylock(&clobj->cob_inode->i_mutex));
 504         clobj->cob_transient_pages--;
 505 }
 506
 507 static const struct cl_page_operations vvp_transient_page_ops = {
 508         .cpo_own                = vvp_transient_page_own,
 509         .cpo_assume             = vvp_transient_page_assume,
 510         .cpo_unassume           = vvp_transient_page_unassume,
 511         .cpo_disown             = vvp_transient_page_disown,
 512         .cpo_discard            = vvp_transient_page_discard,
 513         .cpo_fini               = vvp_transient_page_fini,
 514         .cpo_is_vmlocked        = vvp_transient_page_is_vmlocked,
 515         .cpo_print              = vvp_page_print,
 516         .cpo_is_under_lock      = vvp_page_is_under_lock,
 517         .io = {
 518                 [CRT_READ] = {
 519                         .cpo_prep       = ccc_transient_page_prep,
 520                         .cpo_completion = vvp_transient_page_completion,
 521                 },
 522                 [CRT_WRITE] = {
 523                         .cpo_prep       = ccc_transient_page_prep,
 524                         .cpo_completion = vvp_transient_page_completion,
 525                 }
 526         }
 527 };
 528
 529 int vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 530                 struct cl_page *page, pgoff_t index)
 531 {
 532         struct ccc_page *cpg = cl_object_page_slice(obj, page);
 533         struct page     *vmpage = page->cp_vmpage;
 534
 535         CLOBINVRNT(env, obj, ccc_object_invariant(obj));
 536
 537         cpg->cpg_page = vmpage;
 538         page_cache_get(vmpage);
 539
 540         INIT_LIST_HEAD(&cpg->cpg_pending_linkage);
 541         if (page->cp_type == CPT_CACHEABLE) {
 542                 /* in cache, decref in vvp_page_delete */
 543                 atomic_inc(&page->cp_ref);
 544                 SetPagePrivate(vmpage);
 545                 vmpage->private = (unsigned long)page;
 546                 cl_page_slice_add(page, &cpg->cpg_cl, obj, index,
 547                                 &vvp_page_ops);
 548         } else {
 549                 struct ccc_object *clobj = cl2ccc(obj);
 550
 551                 LASSERT(!mutex_trylock(&clobj->cob_inode->i_mutex));
 552                 cl_page_slice_add(page, &cpg->cpg_cl, obj, index,
 553                                 &vvp_transient_page_ops);
 554                 clobj->cob_transient_pages++;
 555         }
 556         return 0;
 557 }
 558