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, 2012, 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         cfs_page_t *vmpage = cp->cpg_page;
  62
  63         LASSERT(vmpage != NULL);
  64         page_cache_release(vmpage);
  65         OBD_SLAB_FREE_PTR(cp, vvp_page_kmem);
  66 }
  67
  68 static void vvp_page_fini(const struct lu_env *env,
  69                           struct cl_page_slice *slice)
  70 {
  71         struct ccc_page *cp = cl2ccc_page(slice);
  72         cfs_page_t *vmpage  = cp->cpg_page;
  73
  74         /*
  75          * vmpage->private was already cleared when page was moved into
  76          * VPG_FREEING state.
  77          */
  78         LASSERT((struct cl_page *)vmpage->private != slice->cpl_page);
  79         vvp_page_fini_common(cp);
  80 }
  81
  82 static int vvp_page_own(const struct lu_env *env,
  83                         const struct cl_page_slice *slice, struct cl_io *io,
  84                         int nonblock)
  85 {
  86         struct ccc_page *vpg    = cl2ccc_page(slice);
  87         cfs_page_t      *vmpage = vpg->cpg_page;
  88
  89         LASSERT(vmpage != NULL);
  90         if (nonblock) {
  91                 if (!trylock_page(vmpage))
  92                         return -EAGAIN;
  93
  94                 if (unlikely(PageWriteback(vmpage))) {
  95                         unlock_page(vmpage);
  96                         return -EAGAIN;
  97                 }
  98
  99                 return 0;
 100         }
 101
 102         lock_page(vmpage);
 103         wait_on_page_writeback(vmpage);
 104         return 0;
 105 }
 106
 107 static void vvp_page_assume(const struct lu_env *env,
 108                             const struct cl_page_slice *slice,
 109                             struct cl_io *unused)
 110 {
 111         cfs_page_t *vmpage = cl2vm_page(slice);
 112
 113         LASSERT(vmpage != NULL);
 114         LASSERT(PageLocked(vmpage));
 115         wait_on_page_writeback(vmpage);
 116 }
 117
 118 static void vvp_page_unassume(const struct lu_env *env,
 119                               const struct cl_page_slice *slice,
 120                               struct cl_io *unused)
 121 {
 122         cfs_page_t *vmpage = cl2vm_page(slice);
 123
 124         LASSERT(vmpage != NULL);
 125         LASSERT(PageLocked(vmpage));
 126 }
 127
 128 static void vvp_page_disown(const struct lu_env *env,
 129                             const struct cl_page_slice *slice, struct cl_io *io)
 130 {
 131         cfs_page_t *vmpage = cl2vm_page(slice);
 132
 133         LASSERT(vmpage != NULL);
 134         LASSERT(PageLocked(vmpage));
 135
 136         unlock_page(cl2vm_page(slice));
 137 }
 138
 139 static void vvp_page_discard(const struct lu_env *env,
 140                              const struct cl_page_slice *slice,
 141                              struct cl_io *unused)
 142 {
 143         cfs_page_t           *vmpage  = cl2vm_page(slice);
 144         struct address_space *mapping;
 145         struct ccc_page      *cpg     = cl2ccc_page(slice);
 146
 147         LASSERT(vmpage != NULL);
 148         LASSERT(PageLocked(vmpage));
 149
 150         mapping = vmpage->mapping;
 151
 152         if (cpg->cpg_defer_uptodate && !cpg->cpg_ra_used)
 153                 ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
 154
 155         /*
 156          * truncate_complete_page() calls
 157          * a_ops->invalidatepage()->cl_page_delete()->vvp_page_delete().
 158          */
 159         truncate_complete_page(mapping, vmpage);
 160 }
 161
 162 static int vvp_page_unmap(const struct lu_env *env,
 163                           const struct cl_page_slice *slice,
 164                           struct cl_io *unused)
 165 {
 166         cfs_page_t *vmpage = cl2vm_page(slice);
 167         __u64       offset;
 168
 169         LASSERT(vmpage != NULL);
 170         LASSERT(PageLocked(vmpage));
 171
 172         offset = vmpage->index << CFS_PAGE_SHIFT;
 173
 174         /*
 175          * XXX is it safe to call this with the page lock held?
 176          */
 177         ll_teardown_mmaps(vmpage->mapping, offset, offset + CFS_PAGE_SIZE);
 178         return 0;
 179 }
 180
 181 static void vvp_page_delete(const struct lu_env *env,
 182                             const struct cl_page_slice *slice)
 183 {
 184         cfs_page_t       *vmpage = cl2vm_page(slice);
 185         struct inode     *inode  = vmpage->mapping->host;
 186         struct cl_object *obj    = slice->cpl_obj;
 187
 188         LASSERT(PageLocked(vmpage));
 189         LASSERT((struct cl_page *)vmpage->private == slice->cpl_page);
 190         LASSERT(inode == ccc_object_inode(obj));
 191
 192         vvp_write_complete(cl2ccc(obj), cl2ccc_page(slice));
 193         ClearPagePrivate(vmpage);
 194         vmpage->private = 0;
 195         /*
 196          * Reference from vmpage to cl_page is removed, but the reference back
 197          * is still here. It is removed later in vvp_page_fini().
 198          */
 199 }
 200
 201 static void vvp_page_export(const struct lu_env *env,
 202                             const struct cl_page_slice *slice,
 203                             int uptodate)
 204 {
 205         cfs_page_t *vmpage = cl2vm_page(slice);
 206
 207         LASSERT(vmpage != NULL);
 208         LASSERT(PageLocked(vmpage));
 209         if (uptodate)
 210                 SetPageUptodate(vmpage);
 211         else
 212                 ClearPageUptodate(vmpage);
 213 }
 214
 215 static int vvp_page_is_vmlocked(const struct lu_env *env,
 216                                 const struct cl_page_slice *slice)
 217 {
 218         return PageLocked(cl2vm_page(slice)) ? -EBUSY : -ENODATA;
 219 }
 220
 221 static int vvp_page_prep_read(const struct lu_env *env,
 222                               const struct cl_page_slice *slice,
 223                               struct cl_io *unused)
 224 {
 225         ENTRY;
 226         /* Skip the page already marked as PG_uptodate. */
 227         RETURN(PageUptodate(cl2vm_page(slice)) ? -EALREADY : 0);
 228 }
 229
 230 static int vvp_page_prep_write(const struct lu_env *env,
 231                                const struct cl_page_slice *slice,
 232                                struct cl_io *unused)
 233 {
 234         cfs_page_t *vmpage = cl2vm_page(slice);
 235
 236         LASSERT(PageLocked(vmpage));
 237         LASSERT(!PageDirty(vmpage));
 238
 239         set_page_writeback(vmpage);
 240         vvp_write_pending(cl2ccc(slice->cpl_obj), cl2ccc_page(slice));
 241
 242         return 0;
 243 }
 244
 245 /**
 246  * Handles page transfer errors at VM level.
 247  *
 248  * This takes inode as a separate argument, because inode on which error is to
 249  * be set can be different from \a vmpage inode in case of direct-io.
 250  */
 251 static void vvp_vmpage_error(struct inode *inode, cfs_page_t *vmpage, int ioret)
 252 {
 253         struct ccc_object *obj = cl_inode2ccc(inode);
 254
 255         if (ioret == 0) {
 256                 ClearPageError(vmpage);
 257                 obj->cob_discard_page_warned = 0;
 258         } else {
 259                 SetPageError(vmpage);
 260                 if (ioret == -ENOSPC)
 261                         set_bit(AS_ENOSPC, &inode->i_mapping->flags);
 262                 else
 263                         set_bit(AS_EIO, &inode->i_mapping->flags);
 264
 265                 if ((ioret == -ESHUTDOWN || ioret == -EINTR) &&
 266                      obj->cob_discard_page_warned == 0) {
 267                         obj->cob_discard_page_warned = 1;
 268                         ll_dirty_page_discard_warn(vmpage, ioret);
 269                 }
 270         }
 271 }
 272
 273 static void vvp_page_completion_read(const struct lu_env *env,
 274                                      const struct cl_page_slice *slice,
 275                                      int ioret)
 276 {
 277         struct ccc_page *cp     = cl2ccc_page(slice);
 278         cfs_page_t      *vmpage = cp->cpg_page;
 279         struct cl_page  *page   = cl_page_top(slice->cpl_page);
 280         struct inode    *inode  = ccc_object_inode(page->cp_obj);
 281         ENTRY;
 282
 283         LASSERT(PageLocked(vmpage));
 284         CL_PAGE_HEADER(D_PAGE, env, page, "completing READ with %d\n", ioret);
 285
 286         if (cp->cpg_defer_uptodate)
 287                 ll_ra_count_put(ll_i2sbi(inode), 1);
 288
 289         if (ioret == 0)  {
 290                 if (!cp->cpg_defer_uptodate)
 291                         cl_page_export(env, page, 1);
 292         } else
 293                 cp->cpg_defer_uptodate = 0;
 294
 295         if (page->cp_sync_io == NULL)
 296                 unlock_page(vmpage);
 297
 298         EXIT;
 299 }
 300
 301 static void vvp_page_completion_write(const struct lu_env *env,
 302                                       const struct cl_page_slice *slice,
 303                                       int ioret)
 304 {
 305         struct ccc_page *cp     = cl2ccc_page(slice);
 306         struct cl_page  *pg     = slice->cpl_page;
 307         cfs_page_t      *vmpage = cp->cpg_page;
 308         ENTRY;
 309
 310         LASSERT(ergo(pg->cp_sync_io != NULL, PageLocked(vmpage)));
 311         LASSERT(PageWriteback(vmpage));
 312
 313         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 314
 315         /*
 316          * TODO: Actually it makes sense to add the page into oap pending
 317          * list again and so that we don't need to take the page out from
 318          * SoM write pending list, if we just meet a recoverable error,
 319          * -ENOMEM, etc.
 320          * To implement this, we just need to return a non zero value in
 321          * ->cpo_completion method. The underlying transfer should be notified
 322          * and then re-add the page into pending transfer queue.  -jay
 323          */
 324
 325         cp->cpg_write_queued = 0;
 326         vvp_write_complete(cl2ccc(slice->cpl_obj), cp);
 327
 328         /*
 329          * Only mark the page error only when it's an async write because
 330          * applications won't wait for IO to finish.
 331          */
 332         if (pg->cp_sync_io == NULL)
 333                 vvp_vmpage_error(ccc_object_inode(pg->cp_obj), vmpage, ioret);
 334
 335         end_page_writeback(vmpage);
 336         EXIT;
 337 }
 338
 339 /**
 340  * Implements cl_page_operations::cpo_make_ready() method.
 341  *
 342  * This is called to yank a page from the transfer cache and to send it out as
 343  * a part of transfer. This function try-locks the page. If try-lock failed,
 344  * page is owned by some concurrent IO, and should be skipped (this is bad,
 345  * but hopefully rare situation, as it usually results in transfer being
 346  * shorter than possible).
 347  *
 348  * \retval 0      success, page can be placed into transfer
 349  *
 350  * \retval -EAGAIN page is either used by concurrent IO has been
 351  * truncated. Skip it.
 352  */
 353 static int vvp_page_make_ready(const struct lu_env *env,
 354                                const struct cl_page_slice *slice)
 355 {
 356         cfs_page_t *vmpage = cl2vm_page(slice);
 357         struct cl_page *pg = slice->cpl_page;
 358         int result = 0;
 359
 360         lock_page(vmpage);
 361         if (clear_page_dirty_for_io(vmpage)) {
 362                 LASSERT(pg->cp_state == CPS_CACHED);
 363                 /* This actually clears the dirty bit in the radix
 364                  * tree. */
 365                 set_page_writeback(vmpage);
 366                 vvp_write_pending(cl2ccc(slice->cpl_obj),
 367                                 cl2ccc_page(slice));
 368                 CL_PAGE_HEADER(D_PAGE, env, pg, "readied\n");
 369         } else if (pg->cp_state == CPS_PAGEOUT) {
 370                 /* is it possible for osc_flush_async_page() to already
 371                  * make it ready? */
 372                 result = -EALREADY;
 373         } else {
 374                 CL_PAGE_DEBUG(D_ERROR, env, pg, "Unexpecting page state %d.\n",
 375                               pg->cp_state);
 376                 LBUG();
 377         }
 378         unlock_page(vmpage);
 379         RETURN(result);
 380 }
 381
 382 static int vvp_page_print(const struct lu_env *env,
 383                           const struct cl_page_slice *slice,
 384                           void *cookie, lu_printer_t printer)
 385 {
 386         struct ccc_page *vp = cl2ccc_page(slice);
 387         cfs_page_t      *vmpage = vp->cpg_page;
 388
 389         (*printer)(env, cookie, LUSTRE_VVP_NAME"-page@%p(%d:%d:%d) "
 390                    "vm@%p ",
 391                    vp, vp->cpg_defer_uptodate, vp->cpg_ra_used,
 392                    vp->cpg_write_queued, vmpage);
 393         if (vmpage != NULL) {
 394                 (*printer)(env, cookie, "%lx %d:%d %lx %lu %slru",
 395                            (long)vmpage->flags, page_count(vmpage),
 396                            page_mapcount(vmpage), vmpage->private,
 397                            page_index(vmpage),
 398                            list_empty(&vmpage->lru) ? "not-" : "");
 399         }
 400         (*printer)(env, cookie, "\n");
 401         return 0;
 402 }
 403
 404 static const struct cl_page_operations vvp_page_ops = {
 405         .cpo_own           = vvp_page_own,
 406         .cpo_assume        = vvp_page_assume,
 407         .cpo_unassume      = vvp_page_unassume,
 408         .cpo_disown        = vvp_page_disown,
 409         .cpo_vmpage        = ccc_page_vmpage,
 410         .cpo_discard       = vvp_page_discard,
 411         .cpo_delete        = vvp_page_delete,
 412         .cpo_unmap         = vvp_page_unmap,
 413         .cpo_export        = vvp_page_export,
 414         .cpo_is_vmlocked   = vvp_page_is_vmlocked,
 415         .cpo_fini          = vvp_page_fini,
 416         .cpo_print         = vvp_page_print,
 417         .cpo_is_under_lock = ccc_page_is_under_lock,
 418         .io = {
 419                 [CRT_READ] = {
 420                         .cpo_prep        = vvp_page_prep_read,
 421                         .cpo_completion  = vvp_page_completion_read,
 422                         .cpo_make_ready  = ccc_fail,
 423                 },
 424                 [CRT_WRITE] = {
 425                         .cpo_prep        = vvp_page_prep_write,
 426                         .cpo_completion  = vvp_page_completion_write,
 427                         .cpo_make_ready  = vvp_page_make_ready,
 428                 }
 429         }
 430 };
 431
 432 static void vvp_transient_page_verify(const struct cl_page *page)
 433 {
 434         struct inode *inode = ccc_object_inode(page->cp_obj);
 435
 436         LASSERT(!mutex_trylock(&inode->i_mutex));
 437 }
 438
 439 static int vvp_transient_page_own(const struct lu_env *env,
 440                                   const struct cl_page_slice *slice,
 441                                   struct cl_io *unused, int nonblock)
 442 {
 443         vvp_transient_page_verify(slice->cpl_page);
 444         return 0;
 445 }
 446
 447 static void vvp_transient_page_assume(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_unassume(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_disown(const struct lu_env *env,
 462                                       const struct cl_page_slice *slice,
 463                                       struct cl_io *unused)
 464 {
 465         vvp_transient_page_verify(slice->cpl_page);
 466 }
 467
 468 static void vvp_transient_page_discard(const struct lu_env *env,
 469                                        const struct cl_page_slice *slice,
 470                                        struct cl_io *unused)
 471 {
 472         struct cl_page *page = slice->cpl_page;
 473
 474         vvp_transient_page_verify(slice->cpl_page);
 475
 476         /*
 477          * For transient pages, remove it from the radix tree.
 478          */
 479         cl_page_delete(env, page);
 480 }
 481
 482 static int vvp_transient_page_is_vmlocked(const struct lu_env *env,
 483                                           const struct cl_page_slice *slice)
 484 {
 485         struct inode    *inode = ccc_object_inode(slice->cpl_obj);
 486         int     locked;
 487
 488         locked = !mutex_trylock(&inode->i_mutex);
 489         if (!locked)
 490                 mutex_unlock(&inode->i_mutex);
 491         return locked ? -EBUSY : -ENODATA;
 492 }
 493
 494 static void
 495 vvp_transient_page_completion(const struct lu_env *env,
 496                               const struct cl_page_slice *slice,
 497                               int ioret)
 498 {
 499         vvp_transient_page_verify(slice->cpl_page);
 500 }
 501
 502 static void vvp_transient_page_fini(const struct lu_env *env,
 503                                     struct cl_page_slice *slice)
 504 {
 505         struct ccc_page *cp = cl2ccc_page(slice);
 506         struct cl_page *clp = slice->cpl_page;
 507         struct ccc_object *clobj = cl2ccc(clp->cp_obj);
 508
 509         vvp_page_fini_common(cp);
 510         LASSERT(!mutex_trylock(&clobj->cob_inode->i_mutex));
 511         clobj->cob_transient_pages--;
 512 }
 513
 514 static const struct cl_page_operations vvp_transient_page_ops = {
 515         .cpo_own           = vvp_transient_page_own,
 516         .cpo_assume        = vvp_transient_page_assume,
 517         .cpo_unassume      = vvp_transient_page_unassume,
 518         .cpo_disown        = vvp_transient_page_disown,
 519         .cpo_discard       = vvp_transient_page_discard,
 520         .cpo_vmpage        = ccc_page_vmpage,
 521         .cpo_fini          = vvp_transient_page_fini,
 522         .cpo_is_vmlocked   = vvp_transient_page_is_vmlocked,
 523         .cpo_print         = vvp_page_print,
 524         .cpo_is_under_lock = ccc_page_is_under_lock,
 525         .io = {
 526                 [CRT_READ] = {
 527                         .cpo_prep        = ccc_transient_page_prep,
 528                         .cpo_completion  = vvp_transient_page_completion,
 529                 },
 530                 [CRT_WRITE] = {
 531                         .cpo_prep        = ccc_transient_page_prep,
 532                         .cpo_completion  = vvp_transient_page_completion,
 533                 }
 534         }
 535 };
 536
 537 struct cl_page *vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 538                               struct cl_page *page, cfs_page_t *vmpage)
 539 {
 540         struct ccc_page *cpg;
 541         int result;
 542
 543         CLOBINVRNT(env, obj, ccc_object_invariant(obj));
 544
 545         OBD_SLAB_ALLOC_PTR_GFP(cpg, vvp_page_kmem, CFS_ALLOC_IO);
 546         if (cpg != NULL) {
 547                 cpg->cpg_page = vmpage;
 548                 page_cache_get(vmpage);
 549
 550                 CFS_INIT_LIST_HEAD(&cpg->cpg_pending_linkage);
 551                 if (page->cp_type == CPT_CACHEABLE) {
 552                         SetPagePrivate(vmpage);
 553                         vmpage->private = (unsigned long)page;
 554                         cl_page_slice_add(page, &cpg->cpg_cl, obj,
 555                                           &vvp_page_ops);
 556                 } else {
 557                         struct ccc_object *clobj = cl2ccc(obj);
 558
 559                         LASSERT(!mutex_trylock(&clobj->cob_inode->i_mutex));
 560                         cl_page_slice_add(page, &cpg->cpg_cl, obj,
 561                                           &vvp_transient_page_ops);
 562                         clobj->cob_transient_pages++;
 563                 }
 564                 result = 0;
 565         } else
 566                 result = -ENOMEM;
 567         return ERR_PTR(result);
 568 }
 569