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, Whamcloud, Inc.
  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  */
  40
  41 #define DEBUG_SUBSYSTEM S_LLITE
  42
  43 #ifndef __KERNEL__
  44 # error This file is kernel only.
  45 #endif
  46
  47 #include <obd.h>
  48 #include <lustre_lite.h>
  49
  50 #include "vvp_internal.h"
  51
  52 /*****************************************************************************
  53  *
  54  * Page operations.
  55  *
  56  */
  57
  58 static void vvp_page_fini_common(struct ccc_page *cp)
  59 {
  60         cfs_page_t *vmpage = cp->cpg_page;
  61
  62         LASSERT(vmpage != NULL);
  63         page_cache_release(vmpage);
  64         OBD_SLAB_FREE_PTR(cp, vvp_page_kmem);
  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         cfs_page_t *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         cfs_page_t      *vmpage = vpg->cpg_page;
  87
  88         LASSERT(vmpage != NULL);
  89         if (nonblock) {
  90                 if (TestSetPageLocked(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         cfs_page_t *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         cfs_page_t *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         cfs_page_t *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         cfs_page_t           *vmpage  = cl2vm_page(slice);
 143         struct address_space *mapping = vmpage->mapping;
 144         struct ccc_page      *cpg     = cl2ccc_page(slice);
 145
 146         LASSERT(vmpage != NULL);
 147         LASSERT(PageLocked(vmpage));
 148
 149         if (cpg->cpg_defer_uptodate && !cpg->cpg_ra_used)
 150                 ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
 151
 152         /*
 153          * truncate_complete_page() calls
 154          * a_ops->invalidatepage()->cl_page_delete()->vvp_page_delete().
 155          */
 156         truncate_complete_page(mapping, vmpage);
 157 }
 158
 159 static int vvp_page_unmap(const struct lu_env *env,
 160                           const struct cl_page_slice *slice,
 161                           struct cl_io *unused)
 162 {
 163         cfs_page_t *vmpage = cl2vm_page(slice);
 164         __u64       offset = vmpage->index << CFS_PAGE_SHIFT;
 165
 166         LASSERT(vmpage != NULL);
 167         LASSERT(PageLocked(vmpage));
 168         /*
 169          * XXX is it safe to call this with the page lock held?
 170          */
 171         ll_teardown_mmaps(vmpage->mapping, offset, offset + CFS_PAGE_SIZE);
 172         return 0;
 173 }
 174
 175 static void vvp_page_delete(const struct lu_env *env,
 176                             const struct cl_page_slice *slice)
 177 {
 178         cfs_page_t       *vmpage = cl2vm_page(slice);
 179         struct inode     *inode  = vmpage->mapping->host;
 180         struct cl_object *obj    = slice->cpl_obj;
 181
 182         LASSERT(PageLocked(vmpage));
 183         LASSERT((struct cl_page *)vmpage->private == slice->cpl_page);
 184         LASSERT(inode == ccc_object_inode(obj));
 185
 186         vvp_write_complete(cl2ccc(obj), cl2ccc_page(slice));
 187         ClearPagePrivate(vmpage);
 188         vmpage->private = 0;
 189         /*
 190          * Reference from vmpage to cl_page is removed, but the reference back
 191          * is still here. It is removed later in vvp_page_fini().
 192          */
 193 }
 194
 195 static void vvp_page_export(const struct lu_env *env,
 196                             const struct cl_page_slice *slice,
 197                             int uptodate)
 198 {
 199         cfs_page_t *vmpage = cl2vm_page(slice);
 200
 201         LASSERT(vmpage != NULL);
 202         LASSERT(PageLocked(vmpage));
 203         if (uptodate)
 204                 SetPageUptodate(vmpage);
 205         else
 206                 ClearPageUptodate(vmpage);
 207 }
 208
 209 static int vvp_page_is_vmlocked(const struct lu_env *env,
 210                                 const struct cl_page_slice *slice)
 211 {
 212         return PageLocked(cl2vm_page(slice)) ? -EBUSY : -ENODATA;
 213 }
 214
 215 static int vvp_page_prep_read(const struct lu_env *env,
 216                               const struct cl_page_slice *slice,
 217                               struct cl_io *unused)
 218 {
 219         ENTRY;
 220         /* Skip the page already marked as PG_uptodate. */
 221         RETURN(PageUptodate(cl2vm_page(slice)) ? -EALREADY : 0);
 222 }
 223
 224 static int vvp_page_prep_write(const struct lu_env *env,
 225                                const struct cl_page_slice *slice,
 226                                struct cl_io *unused)
 227 {
 228         struct cl_page *cp     = slice->cpl_page;
 229         cfs_page_t     *vmpage = cl2vm_page(slice);
 230         int result;
 231
 232         if (clear_page_dirty_for_io(vmpage)) {
 233                 set_page_writeback(vmpage);
 234                 vvp_write_pending(cl2ccc(slice->cpl_obj), cl2ccc_page(slice));
 235                 result = 0;
 236
 237                 /* only turn on writeback for async write. */
 238                 if (cp->cp_sync_io == NULL)
 239                         unlock_page(vmpage);
 240         } else
 241                 result = -EALREADY;
 242         return result;
 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         if (ioret == 0)
 254                 ClearPageError(vmpage);
 255         else if (ioret != -EINTR) {
 256                 SetPageError(vmpage);
 257                 if (ioret == -ENOSPC)
 258                         set_bit(AS_ENOSPC, &inode->i_mapping->flags);
 259                 else
 260                         set_bit(AS_EIO, &inode->i_mapping->flags);
 261         }
 262 }
 263
 264 static void vvp_page_completion_read(const struct lu_env *env,
 265                                      const struct cl_page_slice *slice,
 266                                      int ioret)
 267 {
 268         struct ccc_page *cp     = cl2ccc_page(slice);
 269         cfs_page_t      *vmpage = cp->cpg_page;
 270         struct cl_page  *page   = cl_page_top(slice->cpl_page);
 271         struct inode    *inode  = ccc_object_inode(page->cp_obj);
 272         ENTRY;
 273
 274         LASSERT(PageLocked(vmpage));
 275         CL_PAGE_HEADER(D_PAGE, env, page, "completing READ with %d\n", ioret);
 276
 277         if (cp->cpg_defer_uptodate)
 278                 ll_ra_count_put(ll_i2sbi(inode), 1);
 279
 280         if (ioret == 0)  {
 281                 if (!cp->cpg_defer_uptodate)
 282                         cl_page_export(env, page, 1);
 283         } else
 284                 cp->cpg_defer_uptodate = 0;
 285
 286         if (page->cp_sync_io == NULL)
 287                 unlock_page(vmpage);
 288
 289         EXIT;
 290 }
 291
 292 static void vvp_page_completion_write(const struct lu_env *env,
 293                                       const struct cl_page_slice *slice,
 294                                       int ioret)
 295 {
 296         struct ccc_page *cp     = cl2ccc_page(slice);
 297         struct cl_page  *pg     = slice->cpl_page;
 298         cfs_page_t      *vmpage = cp->cpg_page;
 299         ENTRY;
 300
 301         LASSERT(ergo(pg->cp_sync_io != NULL, PageLocked(vmpage)));
 302         LASSERT(PageWriteback(vmpage));
 303
 304         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 305
 306         /*
 307          * TODO: Actually it makes sense to add the page into oap pending
 308          * list again and so that we don't need to take the page out from
 309          * SoM write pending list, if we just meet a recoverable error,
 310          * -ENOMEM, etc.
 311          * To implement this, we just need to return a non zero value in
 312          * ->cpo_completion method. The underlying transfer should be notified
 313          * and then re-add the page into pending transfer queue.  -jay
 314          */
 315
 316         cp->cpg_write_queued = 0;
 317         vvp_write_complete(cl2ccc(slice->cpl_obj), cp);
 318
 319         /*
 320          * Only mark the page error only when it's an async write because
 321          * applications won't wait for IO to finish.
 322          */
 323         if (pg->cp_sync_io == NULL)
 324                 vvp_vmpage_error(ccc_object_inode(pg->cp_obj), vmpage, ioret);
 325
 326         end_page_writeback(vmpage);
 327         EXIT;
 328 }
 329
 330 /**
 331  * Implements cl_page_operations::cpo_make_ready() method.
 332  *
 333  * This is called to yank a page from the transfer cache and to send it out as
 334  * a part of transfer. This function try-locks the page. If try-lock failed,
 335  * page is owned by some concurrent IO, and should be skipped (this is bad,
 336  * but hopefully rare situation, as it usually results in transfer being
 337  * shorter than possible).
 338  *
 339  * \retval 0      success, page can be placed into transfer
 340  *
 341  * \retval -EAGAIN page is either used by concurrent IO has been
 342  * truncated. Skip it.
 343  */
 344 static int vvp_page_make_ready(const struct lu_env *env,
 345                                const struct cl_page_slice *slice)
 346 {
 347         cfs_page_t *vmpage = cl2vm_page(slice);
 348         struct cl_page *pg = slice->cpl_page;
 349         int result;
 350
 351         result = -EAGAIN;
 352         /* we're trying to write, but the page is locked.. come back later */
 353         if (!TestSetPageLocked(vmpage)) {
 354                 if (pg->cp_state == CPS_CACHED) {
 355                         /*
 356                          * We can cancel IO if page wasn't dirty after all.
 357                          */
 358                         clear_page_dirty_for_io(vmpage);
 359                         /*
 360                          * This actually clears the dirty bit in the radix
 361                          * tree.
 362                          */
 363                         set_page_writeback(vmpage);
 364                         vvp_write_pending(cl2ccc(slice->cpl_obj),
 365                                           cl2ccc_page(slice));
 366                         CL_PAGE_HEADER(D_PAGE, env, pg, "readied\n");
 367                         result = 0;
 368                 } else
 369                         /*
 370                          * Page was concurrently truncated.
 371                          */
 372                         LASSERT(pg->cp_state == CPS_FREEING);
 373                 unlock_page(vmpage);
 374         }
 375         RETURN(result);
 376 }
 377
 378 static int vvp_page_print(const struct lu_env *env,
 379                           const struct cl_page_slice *slice,
 380                           void *cookie, lu_printer_t printer)
 381 {
 382         struct ccc_page *vp = cl2ccc_page(slice);
 383         cfs_page_t      *vmpage = vp->cpg_page;
 384
 385         (*printer)(env, cookie, LUSTRE_VVP_NAME"-page@%p(%d:%d:%d) "
 386                    "vm@%p ",
 387                    vp, vp->cpg_defer_uptodate, vp->cpg_ra_used,
 388                    vp->cpg_write_queued, vmpage);
 389         if (vmpage != NULL) {
 390                 (*printer)(env, cookie, "%lx %d:%d %lx %lu %slru",
 391                            (long)vmpage->flags, page_count(vmpage),
 392                            page_mapcount(vmpage), vmpage->private,
 393                            page_index(vmpage),
 394                            list_empty(&vmpage->lru) ? "not-" : "");
 395         }
 396         (*printer)(env, cookie, "\n");
 397         return 0;
 398 }
 399
 400 static const struct cl_page_operations vvp_page_ops = {
 401         .cpo_own           = vvp_page_own,
 402         .cpo_assume        = vvp_page_assume,
 403         .cpo_unassume      = vvp_page_unassume,
 404         .cpo_disown        = vvp_page_disown,
 405         .cpo_vmpage        = ccc_page_vmpage,
 406         .cpo_discard       = vvp_page_discard,
 407         .cpo_delete        = vvp_page_delete,
 408         .cpo_unmap         = vvp_page_unmap,
 409         .cpo_export        = vvp_page_export,
 410         .cpo_is_vmlocked   = vvp_page_is_vmlocked,
 411         .cpo_fini          = vvp_page_fini,
 412         .cpo_print         = vvp_page_print,
 413         .cpo_is_under_lock = ccc_page_is_under_lock,
 414         .io = {
 415                 [CRT_READ] = {
 416                         .cpo_prep        = vvp_page_prep_read,
 417                         .cpo_completion  = vvp_page_completion_read,
 418                         .cpo_make_ready  = ccc_fail,
 419                 },
 420                 [CRT_WRITE] = {
 421                         .cpo_prep        = vvp_page_prep_write,
 422                         .cpo_completion  = vvp_page_completion_write,
 423                         .cpo_make_ready  = vvp_page_make_ready,
 424                 }
 425         }
 426 };
 427
 428 static void vvp_transient_page_verify(const struct cl_page *page)
 429 {
 430         struct inode *inode = ccc_object_inode(page->cp_obj);
 431
 432         LASSERT(!TRYLOCK_INODE_MUTEX(inode));
 433         /* LASSERT_SEM_LOCKED(&inode->i_alloc_sem); */
 434 }
 435
 436 static int vvp_transient_page_own(const struct lu_env *env,
 437                                   const struct cl_page_slice *slice,
 438                                   struct cl_io *unused, int nonblock)
 439 {
 440         vvp_transient_page_verify(slice->cpl_page);
 441         return 0;
 442 }
 443
 444 static void vvp_transient_page_assume(const struct lu_env *env,
 445                                       const struct cl_page_slice *slice,
 446                                       struct cl_io *unused)
 447 {
 448         vvp_transient_page_verify(slice->cpl_page);
 449 }
 450
 451 static void vvp_transient_page_unassume(const struct lu_env *env,
 452                                         const struct cl_page_slice *slice,
 453                                         struct cl_io *unused)
 454 {
 455         vvp_transient_page_verify(slice->cpl_page);
 456 }
 457
 458 static void vvp_transient_page_disown(const struct lu_env *env,
 459                                       const struct cl_page_slice *slice,
 460                                       struct cl_io *unused)
 461 {
 462         vvp_transient_page_verify(slice->cpl_page);
 463 }
 464
 465 static void vvp_transient_page_discard(const struct lu_env *env,
 466                                        const struct cl_page_slice *slice,
 467                                        struct cl_io *unused)
 468 {
 469         struct cl_page *page = slice->cpl_page;
 470
 471         vvp_transient_page_verify(slice->cpl_page);
 472
 473         /*
 474          * For transient pages, remove it from the radix tree.
 475          */
 476         cl_page_delete(env, page);
 477 }
 478
 479 static int vvp_transient_page_is_vmlocked(const struct lu_env *env,
 480                                           const struct cl_page_slice *slice)
 481 {
 482         struct inode    *inode = ccc_object_inode(slice->cpl_obj);
 483         int              locked;
 484
 485         locked = !TRYLOCK_INODE_MUTEX(inode);
 486         if (!locked)
 487                 UNLOCK_INODE_MUTEX(inode);
 488         return locked ? -EBUSY : -ENODATA;
 489 }
 490
 491 static void
 492 vvp_transient_page_completion(const struct lu_env *env,
 493                               const struct cl_page_slice *slice,
 494                               int ioret)
 495 {
 496         vvp_transient_page_verify(slice->cpl_page);
 497 }
 498
 499 static void vvp_transient_page_fini(const struct lu_env *env,
 500                                     struct cl_page_slice *slice)
 501 {
 502         struct ccc_page *cp = cl2ccc_page(slice);
 503         struct cl_page *clp = slice->cpl_page;
 504         struct ccc_object *clobj = cl2ccc(clp->cp_obj);
 505
 506         vvp_page_fini_common(cp);
 507         LASSERT(!TRYLOCK_INODE_MUTEX(clobj->cob_inode));
 508         clobj->cob_transient_pages--;
 509 }
 510
 511 static const struct cl_page_operations vvp_transient_page_ops = {
 512         .cpo_own           = vvp_transient_page_own,
 513         .cpo_assume        = vvp_transient_page_assume,
 514         .cpo_unassume      = vvp_transient_page_unassume,
 515         .cpo_disown        = vvp_transient_page_disown,
 516         .cpo_discard       = vvp_transient_page_discard,
 517         .cpo_vmpage        = ccc_page_vmpage,
 518         .cpo_fini          = vvp_transient_page_fini,
 519         .cpo_is_vmlocked   = vvp_transient_page_is_vmlocked,
 520         .cpo_print         = vvp_page_print,
 521         .cpo_is_under_lock = ccc_page_is_under_lock,
 522         .io = {
 523                 [CRT_READ] = {
 524                         .cpo_prep        = ccc_transient_page_prep,
 525                         .cpo_completion  = vvp_transient_page_completion,
 526                 },
 527                 [CRT_WRITE] = {
 528                         .cpo_prep        = ccc_transient_page_prep,
 529                         .cpo_completion  = vvp_transient_page_completion,
 530                 }
 531         }
 532 };
 533
 534 struct cl_page *vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 535                               struct cl_page *page, cfs_page_t *vmpage)
 536 {
 537         struct ccc_page *cpg;
 538         int result;
 539
 540         CLOBINVRNT(env, obj, ccc_object_invariant(obj));
 541
 542         OBD_SLAB_ALLOC_PTR_GFP(cpg, vvp_page_kmem, CFS_ALLOC_IO);
 543         if (cpg != NULL) {
 544                 cpg->cpg_page = vmpage;
 545                 page_cache_get(vmpage);
 546
 547                 CFS_INIT_LIST_HEAD(&cpg->cpg_pending_linkage);
 548                 if (page->cp_type == CPT_CACHEABLE) {
 549                         SetPagePrivate(vmpage);
 550                         vmpage->private = (unsigned long)page;
 551                         cl_page_slice_add(page, &cpg->cpg_cl, obj,
 552                                           &vvp_page_ops);
 553                 } else {
 554                         struct ccc_object *clobj = cl2ccc(obj);
 555
 556                         LASSERT(!TRYLOCK_INODE_MUTEX(clobj->cob_inode));
 557                         cl_page_slice_add(page, &cpg->cpg_cl, obj,
 558                                           &vvp_transient_page_ops);
 559                         clobj->cob_transient_pages++;
 560                 }
 561                 result = 0;
 562         } else
 563                 result = -ENOMEM;
 564         return ERR_PTR(result);
 565 }
 566