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 address_space *mapping;
 144         struct ccc_page      *cpg     = cl2ccc_page(slice);
 145
 146         LASSERT(vmpage != NULL);
 147         LASSERT(PageLocked(vmpage));
 148
 149         mapping = vmpage->mapping;
 150
 151         if (cpg->cpg_defer_uptodate && !cpg->cpg_ra_used)
 152                 ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
 153
 154         /*
 155          * truncate_complete_page() calls
 156          * a_ops->invalidatepage()->cl_page_delete()->vvp_page_delete().
 157          */
 158         truncate_complete_page(mapping, vmpage);
 159 }
 160
 161 static int vvp_page_unmap(const struct lu_env *env,
 162                           const struct cl_page_slice *slice,
 163                           struct cl_io *unused)
 164 {
 165         struct page *vmpage = cl2vm_page(slice);
 166         __u64       offset;
 167
 168         LASSERT(vmpage != NULL);
 169         LASSERT(PageLocked(vmpage));
 170
 171         offset = vmpage->index << PAGE_CACHE_SHIFT;
 172
 173         /*
 174          * XXX is it safe to call this with the page lock held?
 175          */
 176         ll_teardown_mmaps(vmpage->mapping, offset, offset + PAGE_CACHE_SIZE);
 177         return 0;
 178 }
 179
 180 static void vvp_page_delete(const struct lu_env *env,
 181                             const struct cl_page_slice *slice)
 182 {
 183         struct page       *vmpage = cl2vm_page(slice);
 184         struct inode     *inode  = vmpage->mapping->host;
 185         struct cl_object *obj    = slice->cpl_obj;
 186
 187         LASSERT(PageLocked(vmpage));
 188         LASSERT((struct cl_page *)vmpage->private == slice->cpl_page);
 189         LASSERT(inode == ccc_object_inode(obj));
 190
 191         vvp_write_complete(cl2ccc(obj), cl2ccc_page(slice));
 192         ClearPagePrivate(vmpage);
 193         vmpage->private = 0;
 194         /*
 195          * Reference from vmpage to cl_page is removed, but the reference back
 196          * is still here. It is removed later in vvp_page_fini().
 197          */
 198 }
 199
 200 static void vvp_page_export(const struct lu_env *env,
 201                             const struct cl_page_slice *slice,
 202                             int uptodate)
 203 {
 204         struct page *vmpage = cl2vm_page(slice);
 205
 206         LASSERT(vmpage != NULL);
 207         LASSERT(PageLocked(vmpage));
 208         if (uptodate)
 209                 SetPageUptodate(vmpage);
 210         else
 211                 ClearPageUptodate(vmpage);
 212 }
 213
 214 static int vvp_page_is_vmlocked(const struct lu_env *env,
 215                                 const struct cl_page_slice *slice)
 216 {
 217         return PageLocked(cl2vm_page(slice)) ? -EBUSY : -ENODATA;
 218 }
 219
 220 static int vvp_page_prep_read(const struct lu_env *env,
 221                               const struct cl_page_slice *slice,
 222                               struct cl_io *unused)
 223 {
 224         ENTRY;
 225         /* Skip the page already marked as PG_uptodate. */
 226         RETURN(PageUptodate(cl2vm_page(slice)) ? -EALREADY : 0);
 227 }
 228
 229 static int vvp_page_prep_write(const struct lu_env *env,
 230                                const struct cl_page_slice *slice,
 231                                struct cl_io *unused)
 232 {
 233         struct page *vmpage = cl2vm_page(slice);
 234
 235         LASSERT(PageLocked(vmpage));
 236         LASSERT(!PageDirty(vmpage));
 237
 238         set_page_writeback(vmpage);
 239         vvp_write_pending(cl2ccc(slice->cpl_obj), cl2ccc_page(slice));
 240
 241         return 0;
 242 }
 243
 244 /**
 245  * Handles page transfer errors at VM level.
 246  *
 247  * This takes inode as a separate argument, because inode on which error is to
 248  * be set can be different from \a vmpage inode in case of direct-io.
 249  */
 250 static void vvp_vmpage_error(struct inode *inode, struct page *vmpage, int ioret)
 251 {
 252         struct ccc_object *obj = cl_inode2ccc(inode);
 253
 254         if (ioret == 0) {
 255                 ClearPageError(vmpage);
 256                 obj->cob_discard_page_warned = 0;
 257         } else {
 258                 SetPageError(vmpage);
 259                 if (ioret == -ENOSPC)
 260                         set_bit(AS_ENOSPC, &inode->i_mapping->flags);
 261                 else
 262                         set_bit(AS_EIO, &inode->i_mapping->flags);
 263
 264                 if ((ioret == -ESHUTDOWN || ioret == -EINTR) &&
 265                      obj->cob_discard_page_warned == 0) {
 266                         obj->cob_discard_page_warned = 1;
 267                         ll_dirty_page_discard_warn(vmpage, ioret);
 268                 }
 269         }
 270 }
 271
 272 static void vvp_page_completion_read(const struct lu_env *env,
 273                                      const struct cl_page_slice *slice,
 274                                      int ioret)
 275 {
 276         struct ccc_page *cp     = cl2ccc_page(slice);
 277         struct page      *vmpage = cp->cpg_page;
 278         struct cl_page  *page   = cl_page_top(slice->cpl_page);
 279         struct inode    *inode  = ccc_object_inode(page->cp_obj);
 280         ENTRY;
 281
 282         LASSERT(PageLocked(vmpage));
 283         CL_PAGE_HEADER(D_PAGE, env, page, "completing READ with %d\n", ioret);
 284
 285         if (cp->cpg_defer_uptodate)
 286                 ll_ra_count_put(ll_i2sbi(inode), 1);
 287
 288         if (ioret == 0)  {
 289                 if (!cp->cpg_defer_uptodate)
 290                         cl_page_export(env, page, 1);
 291         } else
 292                 cp->cpg_defer_uptodate = 0;
 293
 294         if (page->cp_sync_io == NULL)
 295                 unlock_page(vmpage);
 296
 297         EXIT;
 298 }
 299
 300 static void vvp_page_completion_write(const struct lu_env *env,
 301                                       const struct cl_page_slice *slice,
 302                                       int ioret)
 303 {
 304         struct ccc_page *cp     = cl2ccc_page(slice);
 305         struct cl_page  *pg     = slice->cpl_page;
 306         struct page      *vmpage = cp->cpg_page;
 307         ENTRY;
 308
 309         LASSERT(ergo(pg->cp_sync_io != NULL, PageLocked(vmpage)));
 310         LASSERT(PageWriteback(vmpage));
 311
 312         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 313
 314         /*
 315          * TODO: Actually it makes sense to add the page into oap pending
 316          * list again and so that we don't need to take the page out from
 317          * SoM write pending list, if we just meet a recoverable error,
 318          * -ENOMEM, etc.
 319          * To implement this, we just need to return a non zero value in
 320          * ->cpo_completion method. The underlying transfer should be notified
 321          * and then re-add the page into pending transfer queue.  -jay
 322          */
 323
 324         cp->cpg_write_queued = 0;
 325         vvp_write_complete(cl2ccc(slice->cpl_obj), cp);
 326
 327         /*
 328          * Only mark the page error only when it's an async write because
 329          * applications won't wait for IO to finish.
 330          */
 331         if (pg->cp_sync_io == NULL)
 332                 vvp_vmpage_error(ccc_object_inode(pg->cp_obj), vmpage, ioret);
 333
 334         end_page_writeback(vmpage);
 335         EXIT;
 336 }
 337
 338 /**
 339  * Implements cl_page_operations::cpo_make_ready() method.
 340  *
 341  * This is called to yank a page from the transfer cache and to send it out as
 342  * a part of transfer. This function try-locks the page. If try-lock failed,
 343  * page is owned by some concurrent IO, and should be skipped (this is bad,
 344  * but hopefully rare situation, as it usually results in transfer being
 345  * shorter than possible).
 346  *
 347  * \retval 0      success, page can be placed into transfer
 348  *
 349  * \retval -EAGAIN page is either used by concurrent IO has been
 350  * truncated. Skip it.
 351  */
 352 static int vvp_page_make_ready(const struct lu_env *env,
 353                                const struct cl_page_slice *slice)
 354 {
 355         struct page *vmpage = cl2vm_page(slice);
 356         struct cl_page *pg = slice->cpl_page;
 357         int result = 0;
 358
 359         lock_page(vmpage);
 360         if (clear_page_dirty_for_io(vmpage)) {
 361                 LASSERT(pg->cp_state == CPS_CACHED);
 362                 /* This actually clears the dirty bit in the radix
 363                  * tree. */
 364                 set_page_writeback(vmpage);
 365                 vvp_write_pending(cl2ccc(slice->cpl_obj),
 366                                 cl2ccc_page(slice));
 367                 CL_PAGE_HEADER(D_PAGE, env, pg, "readied\n");
 368         } else if (pg->cp_state == CPS_PAGEOUT) {
 369                 /* is it possible for osc_flush_async_page() to already
 370                  * make it ready? */
 371                 result = -EALREADY;
 372         } else {
 373                 CL_PAGE_DEBUG(D_ERROR, env, pg, "Unexpecting page state %d.\n",
 374                               pg->cp_state);
 375                 LBUG();
 376         }
 377         unlock_page(vmpage);
 378         RETURN(result);
 379 }
 380
 381 static int vvp_page_print(const struct lu_env *env,
 382                           const struct cl_page_slice *slice,
 383                           void *cookie, lu_printer_t printer)
 384 {
 385         struct ccc_page *vp = cl2ccc_page(slice);
 386         struct page      *vmpage = vp->cpg_page;
 387
 388         (*printer)(env, cookie, LUSTRE_VVP_NAME"-page@%p(%d:%d:%d) "
 389                    "vm@%p ",
 390                    vp, vp->cpg_defer_uptodate, vp->cpg_ra_used,
 391                    vp->cpg_write_queued, vmpage);
 392         if (vmpage != NULL) {
 393                 (*printer)(env, cookie, "%lx %d:%d %lx %lu %slru",
 394                            (long)vmpage->flags, page_count(vmpage),
 395                            page_mapcount(vmpage), vmpage->private,
 396                            page_index(vmpage),
 397                            list_empty(&vmpage->lru) ? "not-" : "");
 398         }
 399         (*printer)(env, cookie, "\n");
 400         return 0;
 401 }
 402
 403 static const struct cl_page_operations vvp_page_ops = {
 404         .cpo_own           = vvp_page_own,
 405         .cpo_assume        = vvp_page_assume,
 406         .cpo_unassume      = vvp_page_unassume,
 407         .cpo_disown        = vvp_page_disown,
 408         .cpo_vmpage        = ccc_page_vmpage,
 409         .cpo_discard       = vvp_page_discard,
 410         .cpo_delete        = vvp_page_delete,
 411         .cpo_unmap         = vvp_page_unmap,
 412         .cpo_export        = vvp_page_export,
 413         .cpo_is_vmlocked   = vvp_page_is_vmlocked,
 414         .cpo_fini          = vvp_page_fini,
 415         .cpo_print         = vvp_page_print,
 416         .cpo_is_under_lock = ccc_page_is_under_lock,
 417         .io = {
 418                 [CRT_READ] = {
 419                         .cpo_prep        = vvp_page_prep_read,
 420                         .cpo_completion  = vvp_page_completion_read,
 421                         .cpo_make_ready  = ccc_fail,
 422                 },
 423                 [CRT_WRITE] = {
 424                         .cpo_prep        = vvp_page_prep_write,
 425                         .cpo_completion  = vvp_page_completion_write,
 426                         .cpo_make_ready  = vvp_page_make_ready,
 427                 }
 428         }
 429 };
 430
 431 static void vvp_transient_page_verify(const struct cl_page *page)
 432 {
 433         struct inode *inode = ccc_object_inode(page->cp_obj);
 434
 435         LASSERT(!mutex_trylock(&inode->i_mutex));
 436 }
 437
 438 static int vvp_transient_page_own(const struct lu_env *env,
 439                                   const struct cl_page_slice *slice,
 440                                   struct cl_io *unused, int nonblock)
 441 {
 442         vvp_transient_page_verify(slice->cpl_page);
 443         return 0;
 444 }
 445
 446 static void vvp_transient_page_assume(const struct lu_env *env,
 447                                       const struct cl_page_slice *slice,
 448                                       struct cl_io *unused)
 449 {
 450         vvp_transient_page_verify(slice->cpl_page);
 451 }
 452
 453 static void vvp_transient_page_unassume(const struct lu_env *env,
 454                                         const struct cl_page_slice *slice,
 455                                         struct cl_io *unused)
 456 {
 457         vvp_transient_page_verify(slice->cpl_page);
 458 }
 459
 460 static void vvp_transient_page_disown(const struct lu_env *env,
 461                                       const struct cl_page_slice *slice,
 462                                       struct cl_io *unused)
 463 {
 464         vvp_transient_page_verify(slice->cpl_page);
 465 }
 466
 467 static void vvp_transient_page_discard(const struct lu_env *env,
 468                                        const struct cl_page_slice *slice,
 469                                        struct cl_io *unused)
 470 {
 471         struct cl_page *page = slice->cpl_page;
 472
 473         vvp_transient_page_verify(slice->cpl_page);
 474
 475         /*
 476          * For transient pages, remove it from the radix tree.
 477          */
 478         cl_page_delete(env, page);
 479 }
 480
 481 static int vvp_transient_page_is_vmlocked(const struct lu_env *env,
 482                                           const struct cl_page_slice *slice)
 483 {
 484         struct inode    *inode = ccc_object_inode(slice->cpl_obj);
 485         int     locked;
 486
 487         locked = !mutex_trylock(&inode->i_mutex);
 488         if (!locked)
 489                 mutex_unlock(&inode->i_mutex);
 490         return locked ? -EBUSY : -ENODATA;
 491 }
 492
 493 static void
 494 vvp_transient_page_completion(const struct lu_env *env,
 495                               const struct cl_page_slice *slice,
 496                               int ioret)
 497 {
 498         vvp_transient_page_verify(slice->cpl_page);
 499 }
 500
 501 static void vvp_transient_page_fini(const struct lu_env *env,
 502                                     struct cl_page_slice *slice)
 503 {
 504         struct ccc_page *cp = cl2ccc_page(slice);
 505         struct cl_page *clp = slice->cpl_page;
 506         struct ccc_object *clobj = cl2ccc(clp->cp_obj);
 507
 508         vvp_page_fini_common(cp);
 509         LASSERT(!mutex_trylock(&clobj->cob_inode->i_mutex));
 510         clobj->cob_transient_pages--;
 511 }
 512
 513 static const struct cl_page_operations vvp_transient_page_ops = {
 514         .cpo_own           = vvp_transient_page_own,
 515         .cpo_assume        = vvp_transient_page_assume,
 516         .cpo_unassume      = vvp_transient_page_unassume,
 517         .cpo_disown        = vvp_transient_page_disown,
 518         .cpo_discard       = vvp_transient_page_discard,
 519         .cpo_vmpage        = ccc_page_vmpage,
 520         .cpo_fini          = vvp_transient_page_fini,
 521         .cpo_is_vmlocked   = vvp_transient_page_is_vmlocked,
 522         .cpo_print         = vvp_page_print,
 523         .cpo_is_under_lock = ccc_page_is_under_lock,
 524         .io = {
 525                 [CRT_READ] = {
 526                         .cpo_prep        = ccc_transient_page_prep,
 527                         .cpo_completion  = vvp_transient_page_completion,
 528                 },
 529                 [CRT_WRITE] = {
 530                         .cpo_prep        = ccc_transient_page_prep,
 531                         .cpo_completion  = vvp_transient_page_completion,
 532                 }
 533         }
 534 };
 535
 536 int vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 537                 struct cl_page *page, struct page *vmpage)
 538 {
 539         struct ccc_page *cpg = cl_object_page_slice(obj, page);
 540
 541         CLOBINVRNT(env, obj, ccc_object_invariant(obj));
 542
 543         cpg->cpg_page = vmpage;
 544         page_cache_get(vmpage);
 545
 546         CFS_INIT_LIST_HEAD(&cpg->cpg_pending_linkage);
 547         if (page->cp_type == CPT_CACHEABLE) {
 548                 SetPagePrivate(vmpage);
 549                 vmpage->private = (unsigned long)page;
 550                 cl_page_slice_add(page, &cpg->cpg_cl, obj,
 551                                 &vvp_page_ops);
 552         } else {
 553                 struct ccc_object *clobj = cl2ccc(obj);
 554
 555                 LASSERT(!mutex_trylock(&clobj->cob_inode->i_mutex));
 556                 cl_page_slice_add(page, &cpg->cpg_cl, obj,
 557                                 &vvp_transient_page_ops);
 558                 clobj->cob_transient_pages++;
 559         }
 560         return 0;
 561 }
 562