lustre/llite/vvp_page.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  *
  32  * Copyright (c) 2011, Whamcloud, Inc.
  33  */
  34 /*
  35  * This file is part of Lustre, http://www.lustre.org/
  36  * Lustre is a trademark of Sun Microsystems, Inc.
  37  *
  38  * Implementation of cl_page for VVP layer.
  39  *
  40  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  41  */
  42
  43 #define DEBUG_SUBSYSTEM S_LLITE
  44
  45 #ifndef __KERNEL__
  46 # error This file is kernel only.
  47 #endif
  48
  49 #include <obd.h>
  50 #include <lustre_lite.h>
  51
  52 #include "vvp_internal.h"
  53
  54 /*****************************************************************************
  55  *
  56  * Page operations.
  57  *
  58  */
  59
  60 static void vvp_page_fini_common(struct ccc_page *cp)
  61 {
  62         cfs_page_t *vmpage = cp->cpg_page;
  63
  64         LASSERT(vmpage != NULL);
  65         page_cache_release(vmpage);
  66         OBD_SLAB_FREE_PTR(cp, vvp_page_kmem);
  67 }
  68
  69 static void vvp_page_fini(const struct lu_env *env,
  70                           struct cl_page_slice *slice)
  71 {
  72         struct ccc_page *cp = cl2ccc_page(slice);
  73         cfs_page_t *vmpage  = cp->cpg_page;
  74
  75         /*
  76          * vmpage->private was already cleared when page was moved into
  77          * VPG_FREEING state.
  78          */
  79         LASSERT((struct cl_page *)vmpage->private != slice->cpl_page);
  80         vvp_page_fini_common(cp);
  81 }
  82
  83 static int vvp_page_own(const struct lu_env *env,
  84                         const struct cl_page_slice *slice, struct cl_io *io,
  85                         int nonblock)
  86 {
  87         struct ccc_page *vpg    = cl2ccc_page(slice);
  88         cfs_page_t      *vmpage = vpg->cpg_page;
  89
  90         LASSERT(vmpage != NULL);
  91         if (nonblock) {
  92                 if (TestSetPageLocked(vmpage))
  93                         return -EAGAIN;
  94
  95                 if (unlikely(PageWriteback(vmpage))) {
  96                         unlock_page(vmpage);
  97                         return -EAGAIN;
  98                 }
  99
 100                 return 0;
 101         }
 102
 103         lock_page(vmpage);
 104         wait_on_page_writeback(vmpage);
 105         return 0;
 106 }
 107
 108 static void vvp_page_assume(const struct lu_env *env,
 109                             const struct cl_page_slice *slice,
 110                             struct cl_io *unused)
 111 {
 112         cfs_page_t *vmpage = cl2vm_page(slice);
 113
 114         LASSERT(vmpage != NULL);
 115         LASSERT(PageLocked(vmpage));
 116         wait_on_page_writeback(vmpage);
 117 }
 118
 119 static void vvp_page_unassume(const struct lu_env *env,
 120                               const struct cl_page_slice *slice,
 121                               struct cl_io *unused)
 122 {
 123         cfs_page_t *vmpage = cl2vm_page(slice);
 124
 125         LASSERT(vmpage != NULL);
 126         LASSERT(PageLocked(vmpage));
 127 }
 128
 129 static void vvp_page_disown(const struct lu_env *env,
 130                             const struct cl_page_slice *slice, struct cl_io *io)
 131 {
 132         cfs_page_t *vmpage = cl2vm_page(slice);
 133
 134         LASSERT(vmpage != NULL);
 135         LASSERT(PageLocked(vmpage));
 136
 137         unlock_page(cl2vm_page(slice));
 138 }
 139
 140 static void vvp_page_discard(const struct lu_env *env,
 141                              const struct cl_page_slice *slice,
 142                              struct cl_io *unused)
 143 {
 144         cfs_page_t           *vmpage  = cl2vm_page(slice);
 145         struct address_space *mapping = vmpage->mapping;
 146         struct ccc_page      *cpg     = cl2ccc_page(slice);
 147
 148         LASSERT(vmpage != NULL);
 149         LASSERT(PageLocked(vmpage));
 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         cfs_page_t *vmpage = cl2vm_page(slice);
 166         __u64       offset = vmpage->index << CFS_PAGE_SHIFT;
 167
 168         LASSERT(vmpage != NULL);
 169         LASSERT(PageLocked(vmpage));
 170         /*
 171          * XXX is it safe to call this with the page lock held?
 172          */
 173         ll_teardown_mmaps(vmpage->mapping, offset, offset + CFS_PAGE_SIZE);
 174         return 0;
 175 }
 176
 177 static void vvp_page_delete(const struct lu_env *env,
 178                             const struct cl_page_slice *slice)
 179 {
 180         cfs_page_t       *vmpage = cl2vm_page(slice);
 181         struct inode     *inode  = vmpage->mapping->host;
 182         struct cl_object *obj    = slice->cpl_obj;
 183
 184         LASSERT(PageLocked(vmpage));
 185         LASSERT((struct cl_page *)vmpage->private == slice->cpl_page);
 186         LASSERT(inode == ccc_object_inode(obj));
 187
 188         vvp_write_complete(cl2ccc(obj), cl2ccc_page(slice));
 189         ClearPagePrivate(vmpage);
 190         vmpage->private = 0;
 191         /*
 192          * Reference from vmpage to cl_page is removed, but the reference back
 193          * is still here. It is removed later in vvp_page_fini().
 194          */
 195 }
 196
 197 static void vvp_page_export(const struct lu_env *env,
 198                             const struct cl_page_slice *slice,
 199                             int uptodate)
 200 {
 201         cfs_page_t *vmpage = cl2vm_page(slice);
 202
 203         LASSERT(vmpage != NULL);
 204         LASSERT(PageLocked(vmpage));
 205         if (uptodate)
 206                 SetPageUptodate(vmpage);
 207         else
 208                 ClearPageUptodate(vmpage);
 209 }
 210
 211 static int vvp_page_is_vmlocked(const struct lu_env *env,
 212                                 const struct cl_page_slice *slice)
 213 {
 214         return PageLocked(cl2vm_page(slice)) ? -EBUSY : -ENODATA;
 215 }
 216
 217 static int vvp_page_prep_read(const struct lu_env *env,
 218                               const struct cl_page_slice *slice,
 219                               struct cl_io *unused)
 220 {
 221         ENTRY;
 222         /* Skip the page already marked as PG_uptodate. */
 223         RETURN(PageUptodate(cl2vm_page(slice)) ? -EALREADY : 0);
 224 }
 225
 226 static int vvp_page_prep_write(const struct lu_env *env,
 227                                const struct cl_page_slice *slice,
 228                                struct cl_io *unused)
 229 {
 230         struct cl_page *cp     = slice->cpl_page;
 231         cfs_page_t     *vmpage = cl2vm_page(slice);
 232         int result;
 233
 234         if (clear_page_dirty_for_io(vmpage)) {
 235                 set_page_writeback(vmpage);
 236                 vvp_write_pending(cl2ccc(slice->cpl_obj), cl2ccc_page(slice));
 237                 result = 0;
 238
 239                 /* only turn on writeback for async write. */
 240                 if (cp->cp_sync_io == NULL)
 241                         unlock_page(vmpage);
 242         } else
 243                 result = -EALREADY;
 244         return result;
 245 }
 246
 247 /**
 248  * Handles page transfer errors at VM level.
 249  *
 250  * This takes inode as a separate argument, because inode on which error is to
 251  * be set can be different from \a vmpage inode in case of direct-io.
 252  */
 253 static void vvp_vmpage_error(struct inode *inode, cfs_page_t *vmpage, int ioret)
 254 {
 255         if (ioret == 0)
 256                 ClearPageError(vmpage);
 257         else if (ioret != -EINTR) {
 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 }
 265
 266 static void vvp_page_completion_read(const struct lu_env *env,
 267                                      const struct cl_page_slice *slice,
 268                                      int ioret)
 269 {
 270         struct ccc_page *cp     = cl2ccc_page(slice);
 271         cfs_page_t      *vmpage = cp->cpg_page;
 272         struct cl_page  *page   = cl_page_top(slice->cpl_page);
 273         struct inode    *inode  = ccc_object_inode(page->cp_obj);
 274         ENTRY;
 275
 276         LASSERT(PageLocked(vmpage));
 277         CL_PAGE_HEADER(D_PAGE, env, page, "completing READ with %d\n", ioret);
 278
 279         if (cp->cpg_defer_uptodate)
 280                 ll_ra_count_put(ll_i2sbi(inode), 1);
 281
 282         if (ioret == 0)  {
 283                 if (!cp->cpg_defer_uptodate)
 284                         cl_page_export(env, page, 1);
 285         } else
 286                 cp->cpg_defer_uptodate = 0;
 287
 288         if (page->cp_sync_io == NULL)
 289                 unlock_page(vmpage);
 290
 291         EXIT;
 292 }
 293
 294 static void vvp_page_completion_write(const struct lu_env *env,
 295                                       const struct cl_page_slice *slice,
 296                                       int ioret)
 297 {
 298         struct ccc_page *cp     = cl2ccc_page(slice);
 299         struct cl_page  *pg     = slice->cpl_page;
 300         cfs_page_t      *vmpage = cp->cpg_page;
 301         ENTRY;
 302
 303         LASSERT(ergo(pg->cp_sync_io != NULL, PageLocked(vmpage)));
 304         LASSERT(PageWriteback(vmpage));
 305
 306         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 307
 308         /*
 309          * TODO: Actually it makes sense to add the page into oap pending
 310          * list again and so that we don't need to take the page out from
 311          * SoM write pending list, if we just meet a recoverable error,
 312          * -ENOMEM, etc.
 313          * To implement this, we just need to return a non zero value in
 314          * ->cpo_completion method. The underlying transfer should be notified
 315          * and then re-add the page into pending transfer queue.  -jay
 316          */
 317
 318         cp->cpg_write_queued = 0;
 319         vvp_write_complete(cl2ccc(slice->cpl_obj), cp);
 320
 321         /*
 322          * Only mark the page error only when it's an async write because
 323          * applications won't wait for IO to finish.
 324          */
 325         if (pg->cp_sync_io == NULL)
 326                 vvp_vmpage_error(ccc_object_inode(pg->cp_obj), vmpage, ioret);
 327
 328         end_page_writeback(vmpage);
 329         EXIT;
 330 }
 331
 332 /**
 333  * Implements cl_page_operations::cpo_make_ready() method.
 334  *
 335  * This is called to yank a page from the transfer cache and to send it out as
 336  * a part of transfer. This function try-locks the page. If try-lock failed,
 337  * page is owned by some concurrent IO, and should be skipped (this is bad,
 338  * but hopefully rare situation, as it usually results in transfer being
 339  * shorter than possible).
 340  *
 341  * \retval 0      success, page can be placed into transfer
 342  *
 343  * \retval -EAGAIN page is either used by concurrent IO has been
 344  * truncated. Skip it.
 345  */
 346 static int vvp_page_make_ready(const struct lu_env *env,
 347                                const struct cl_page_slice *slice)
 348 {
 349         cfs_page_t *vmpage = cl2vm_page(slice);
 350         struct cl_page *pg = slice->cpl_page;
 351         int result;
 352
 353         result = -EAGAIN;
 354         /* we're trying to write, but the page is locked.. come back later */
 355         if (!TestSetPageLocked(vmpage)) {
 356                 if (pg->cp_state == CPS_CACHED) {
 357                         /*
 358                          * We can cancel IO if page wasn't dirty after all.
 359                          */
 360                         clear_page_dirty_for_io(vmpage);
 361                         /*
 362                          * This actually clears the dirty bit in the radix
 363                          * tree.
 364                          */
 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                         result = 0;
 370                 } else
 371                         /*
 372                          * Page was concurrently truncated.
 373                          */
 374                         LASSERT(pg->cp_state == CPS_FREEING);
 375                 unlock_page(vmpage);
 376         }
 377         RETURN(result);
 378 }
 379
 380 static int vvp_page_print(const struct lu_env *env,
 381                           const struct cl_page_slice *slice,
 382                           void *cookie, lu_printer_t printer)
 383 {
 384         struct ccc_page *vp = cl2ccc_page(slice);
 385         cfs_page_t      *vmpage = vp->cpg_page;
 386
 387         (*printer)(env, cookie, LUSTRE_VVP_NAME"-page@%p(%d:%d:%d) "
 388                    "vm@%p ",
 389                    vp, vp->cpg_defer_uptodate, vp->cpg_ra_used,
 390                    vp->cpg_write_queued, vmpage);
 391         if (vmpage != NULL) {
 392                 (*printer)(env, cookie, "%lx %d:%d %lx %lu %slru",
 393                            (long)vmpage->flags, page_count(vmpage),
 394                            page_mapcount(vmpage), vmpage->private,
 395                            page_index(vmpage),
 396                            list_empty(&vmpage->lru) ? "not-" : "");
 397         }
 398         (*printer)(env, cookie, "\n");
 399         return 0;
 400 }
 401
 402 static const struct cl_page_operations vvp_page_ops = {
 403         .cpo_own           = vvp_page_own,
 404         .cpo_assume        = vvp_page_assume,
 405         .cpo_unassume      = vvp_page_unassume,
 406         .cpo_disown        = vvp_page_disown,
 407         .cpo_vmpage        = ccc_page_vmpage,
 408         .cpo_discard       = vvp_page_discard,
 409         .cpo_delete        = vvp_page_delete,
 410         .cpo_unmap         = vvp_page_unmap,
 411         .cpo_export        = vvp_page_export,
 412         .cpo_is_vmlocked   = vvp_page_is_vmlocked,
 413         .cpo_fini          = vvp_page_fini,
 414         .cpo_print         = vvp_page_print,
 415         .cpo_is_under_lock = ccc_page_is_under_lock,
 416         .io = {
 417                 [CRT_READ] = {
 418                         .cpo_prep        = vvp_page_prep_read,
 419                         .cpo_completion  = vvp_page_completion_read,
 420                         .cpo_make_ready  = ccc_fail,
 421                 },
 422                 [CRT_WRITE] = {
 423                         .cpo_prep        = vvp_page_prep_write,
 424                         .cpo_completion  = vvp_page_completion_write,
 425                         .cpo_make_ready  = vvp_page_make_ready,
 426                 }
 427         }
 428 };
 429
 430 static void vvp_transient_page_verify(const struct cl_page *page)
 431 {
 432         struct inode *inode = ccc_object_inode(page->cp_obj);
 433
 434         LASSERT(!TRYLOCK_INODE_MUTEX(inode));
 435         /* LASSERT_SEM_LOCKED(&inode->i_alloc_sem); */
 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 = !TRYLOCK_INODE_MUTEX(inode);
 488         if (!locked)
 489                 UNLOCK_INODE_MUTEX(inode);
 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(!TRYLOCK_INODE_MUTEX(clobj->cob_inode));
 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 struct cl_page *vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 537                               struct cl_page *page, cfs_page_t *vmpage)
 538 {
 539         struct ccc_page *cpg;
 540         int result;
 541
 542         CLOBINVRNT(env, obj, ccc_object_invariant(obj));
 543
 544         OBD_SLAB_ALLOC_PTR_GFP(cpg, vvp_page_kmem, CFS_ALLOC_IO);
 545         if (cpg != NULL) {
 546                 cpg->cpg_page = vmpage;
 547                 page_cache_get(vmpage);
 548
 549                 CFS_INIT_LIST_HEAD(&cpg->cpg_pending_linkage);
 550                 if (page->cp_type == CPT_CACHEABLE) {
 551                         SetPagePrivate(vmpage);
 552                         vmpage->private = (unsigned long)page;
 553                         cl_page_slice_add(page, &cpg->cpg_cl, obj,
 554                                           &vvp_page_ops);
 555                 } else {
 556                         struct ccc_object *clobj = cl2ccc(obj);
 557
 558                         LASSERT(!TRYLOCK_INODE_MUTEX(clobj->cob_inode));
 559                         cl_page_slice_add(page, &cpg->cpg_cl, obj,
 560                                           &vvp_transient_page_ops);
 561                         clobj->cob_transient_pages++;
 562                 }
 563                 result = 0;
 564         } else
 565                 result = -ENOMEM;
 566         return ERR_PTR(result);
 567 }
 568