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