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 /*
  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         cfs_page_t *vmpage = cl2vm_page(slice);
 229         int result;
 230
 231         if (clear_page_dirty_for_io(vmpage)) {
 232                 set_page_writeback(vmpage);
 233                 vvp_write_pending(cl2ccc(slice->cpl_obj), cl2ccc_page(slice));
 234                 result = 0;
 235         } else
 236                 result = -EALREADY;
 237         return result;
 238 }
 239
 240 /**
 241  * Handles page transfer errors at VM level.
 242  *
 243  * This takes inode as a separate argument, because inode on which error is to
 244  * be set can be different from \a vmpage inode in case of direct-io.
 245  */
 246 static void vvp_vmpage_error(struct inode *inode, cfs_page_t *vmpage, int ioret)
 247 {
 248         if (ioret == 0)
 249                 ClearPageError(vmpage);
 250         else if (ioret != -EINTR) {
 251                 SetPageError(vmpage);
 252                 if (ioret == -ENOSPC)
 253                         set_bit(AS_ENOSPC, &inode->i_mapping->flags);
 254                 else
 255                         set_bit(AS_EIO, &inode->i_mapping->flags);
 256         }
 257 }
 258
 259 static void vvp_page_completion_common(const struct lu_env *env,
 260                                        struct ccc_page *cp, int ioret)
 261 {
 262         struct cl_page    *clp    = cp->cpg_cl.cpl_page;
 263         cfs_page_t        *vmpage = cp->cpg_page;
 264         struct inode      *inode  = ccc_object_inode(clp->cp_obj);
 265
 266         LINVRNT(cl_page_is_vmlocked(env, clp));
 267
 268         if (!clp->cp_sync_io && clp->cp_type == CPT_CACHEABLE) {
 269                 /*
 270                  * Only mark the page error only when it's a cacheable page
 271                  * and NOT a sync io.
 272                  *
 273                  * For sync IO and direct IO(CPT_TRANSIENT), the error is able
 274                  * to be seen by application, so we don't need to mark a page
 275                  * as error at all.
 276                  */
 277                 vvp_vmpage_error(inode, vmpage, ioret);
 278                 unlock_page(vmpage);
 279         }
 280 }
 281
 282 static void vvp_page_completion_read(const struct lu_env *env,
 283                                      const struct cl_page_slice *slice,
 284                                      int ioret)
 285 {
 286         struct ccc_page *cp    = cl2ccc_page(slice);
 287         struct cl_page  *page  = cl_page_top(slice->cpl_page);
 288         struct inode    *inode = ccc_object_inode(page->cp_obj);
 289         ENTRY;
 290
 291         CL_PAGE_HEADER(D_PAGE, env, page, "completing READ with %d\n", ioret);
 292
 293         if (cp->cpg_defer_uptodate)
 294                 ll_ra_count_put(ll_i2sbi(inode), 1);
 295
 296         if (ioret == 0)  {
 297                 /* XXX: do we need this for transient pages? */
 298                 if (!cp->cpg_defer_uptodate)
 299                         cl_page_export(env, page, 1);
 300         } else
 301                 cp->cpg_defer_uptodate = 0;
 302         vvp_page_completion_common(env, cp, ioret);
 303
 304         EXIT;
 305 }
 306
 307 static void vvp_page_completion_write_common(const struct lu_env *env,
 308                                              const struct cl_page_slice *slice,
 309                                              int ioret)
 310 {
 311         struct ccc_page *cp = cl2ccc_page(slice);
 312
 313         /*
 314          * TODO: Actually it makes sense to add the page into oap pending
 315          * list again and so that we don't need to take the page out from
 316          * SoM write pending list, if we just meet a recoverable error,
 317          * -ENOMEM, etc.
 318          * To implement this, we just need to return a non zero value in
 319          * ->cpo_completion method. The underlying transfer should be notified
 320          * and then re-add the page into pending transfer queue.  -jay
 321          */
 322         cp->cpg_write_queued = 0;
 323         vvp_write_complete(cl2ccc(slice->cpl_obj), cp);
 324
 325         vvp_page_completion_common(env, cp, ioret);
 326 }
 327
 328 static void vvp_page_completion_write(const struct lu_env *env,
 329                                       const struct cl_page_slice *slice,
 330                                       int ioret)
 331 {
 332         struct ccc_page *cp     = cl2ccc_page(slice);
 333         struct cl_page  *pg     = slice->cpl_page;
 334         cfs_page_t      *vmpage = cp->cpg_page;
 335
 336         ENTRY;
 337
 338         LINVRNT(cl_page_is_vmlocked(env, pg));
 339         LASSERT(PageWriteback(vmpage));
 340
 341         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 342
 343         vvp_page_completion_write_common(env, slice, ioret);
 344         end_page_writeback(vmpage);
 345         EXIT;
 346 }
 347
 348 /**
 349  * Implements cl_page_operations::cpo_make_ready() method.
 350  *
 351  * This is called to yank a page from the transfer cache and to send it out as
 352  * a part of transfer. This function try-locks the page. If try-lock failed,
 353  * page is owned by some concurrent IO, and should be skipped (this is bad,
 354  * but hopefully rare situation, as it usually results in transfer being
 355  * shorter than possible).
 356  *
 357  * \retval 0      success, page can be placed into transfer
 358  *
 359  * \retval -EAGAIN page is either used by concurrent IO has been
 360  * truncated. Skip it.
 361  */
 362 static int vvp_page_make_ready(const struct lu_env *env,
 363                                const struct cl_page_slice *slice)
 364 {
 365         cfs_page_t *vmpage = cl2vm_page(slice);
 366         struct cl_page *pg = slice->cpl_page;
 367         int result;
 368
 369         result = -EAGAIN;
 370         /* we're trying to write, but the page is locked.. come back later */
 371         if (!TestSetPageLocked(vmpage)) {
 372                 if (pg->cp_state == CPS_CACHED) {
 373                         /*
 374                          * We can cancel IO if page wasn't dirty after all.
 375                          */
 376                         clear_page_dirty_for_io(vmpage);
 377                         /*
 378                          * This actually clears the dirty bit in the radix
 379                          * tree.
 380                          */
 381                         set_page_writeback(vmpage);
 382                         vvp_write_pending(cl2ccc(slice->cpl_obj),
 383                                           cl2ccc_page(slice));
 384                         CL_PAGE_HEADER(D_PAGE, env, pg, "readied\n");
 385                         result = 0;
 386                 } else
 387                         /*
 388                          * Page was concurrently truncated.
 389                          */
 390                         LASSERT(pg->cp_state == CPS_FREEING);
 391         }
 392         RETURN(result);
 393 }
 394
 395 static int vvp_page_print(const struct lu_env *env,
 396                           const struct cl_page_slice *slice,
 397                           void *cookie, lu_printer_t printer)
 398 {
 399         struct ccc_page *vp = cl2ccc_page(slice);
 400         cfs_page_t      *vmpage = vp->cpg_page;
 401
 402         (*printer)(env, cookie, LUSTRE_VVP_NAME"-page@%p(%d:%d:%d) "
 403                    "vm@%p ",
 404                    vp, vp->cpg_defer_uptodate, vp->cpg_ra_used,
 405                    vp->cpg_write_queued, vmpage);
 406         if (vmpage != NULL) {
 407                 (*printer)(env, cookie, "%lx %d:%d %lx %lu %slru",
 408                            (long)vmpage->flags, page_count(vmpage),
 409                            page_mapcount(vmpage), vmpage->private,
 410                            page_index(vmpage),
 411                            list_empty(&vmpage->lru) ? "not-" : "");
 412         }
 413         (*printer)(env, cookie, "\n");
 414         return 0;
 415 }
 416
 417 static const struct cl_page_operations vvp_page_ops = {
 418         .cpo_own           = vvp_page_own,
 419         .cpo_assume        = vvp_page_assume,
 420         .cpo_unassume      = vvp_page_unassume,
 421         .cpo_disown        = vvp_page_disown,
 422         .cpo_vmpage        = ccc_page_vmpage,
 423         .cpo_discard       = vvp_page_discard,
 424         .cpo_delete        = vvp_page_delete,
 425         .cpo_unmap         = vvp_page_unmap,
 426         .cpo_export        = vvp_page_export,
 427         .cpo_is_vmlocked   = vvp_page_is_vmlocked,
 428         .cpo_fini          = vvp_page_fini,
 429         .cpo_print         = vvp_page_print,
 430         .cpo_is_under_lock = ccc_page_is_under_lock,
 431         .io = {
 432                 [CRT_READ] = {
 433                         .cpo_prep        = vvp_page_prep_read,
 434                         .cpo_completion  = vvp_page_completion_read,
 435                         .cpo_make_ready  = ccc_fail,
 436                 },
 437                 [CRT_WRITE] = {
 438                         .cpo_prep        = vvp_page_prep_write,
 439                         .cpo_completion  = vvp_page_completion_write,
 440                         .cpo_make_ready  = vvp_page_make_ready,
 441                 }
 442         }
 443 };
 444
 445 static void vvp_transient_page_verify(const struct cl_page *page)
 446 {
 447         struct inode *inode = ccc_object_inode(page->cp_obj);
 448
 449         LASSERT(!TRYLOCK_INODE_MUTEX(inode));
 450         /* LASSERT_SEM_LOCKED(&inode->i_alloc_sem); */
 451 }
 452
 453 static int vvp_transient_page_own(const struct lu_env *env,
 454                                   const struct cl_page_slice *slice,
 455                                   struct cl_io *unused, int nonblock)
 456 {
 457         vvp_transient_page_verify(slice->cpl_page);
 458         return 0;
 459 }
 460
 461 static void vvp_transient_page_assume(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_unassume(const struct lu_env *env,
 469                                         const struct cl_page_slice *slice,
 470                                         struct cl_io *unused)
 471 {
 472         vvp_transient_page_verify(slice->cpl_page);
 473 }
 474
 475 static void vvp_transient_page_disown(const struct lu_env *env,
 476                                       const struct cl_page_slice *slice,
 477                                       struct cl_io *unused)
 478 {
 479         vvp_transient_page_verify(slice->cpl_page);
 480 }
 481
 482 static void vvp_transient_page_discard(const struct lu_env *env,
 483                                        const struct cl_page_slice *slice,
 484                                        struct cl_io *unused)
 485 {
 486         struct cl_page *page = slice->cpl_page;
 487
 488         vvp_transient_page_verify(slice->cpl_page);
 489
 490         /*
 491          * For transient pages, remove it from the radix tree.
 492          */
 493         cl_page_delete(env, page);
 494 }
 495
 496 static int vvp_transient_page_is_vmlocked(const struct lu_env *env,
 497                                           const struct cl_page_slice *slice)
 498 {
 499         struct inode    *inode = ccc_object_inode(slice->cpl_obj);
 500         int              locked;
 501
 502         locked = !TRYLOCK_INODE_MUTEX(inode);
 503         if (!locked)
 504                 UNLOCK_INODE_MUTEX(inode);
 505         return locked ? -EBUSY : -ENODATA;
 506 }
 507
 508 static void
 509 vvp_transient_page_completion_write(const struct lu_env *env,
 510                                     const struct cl_page_slice *slice,
 511                                     int ioret)
 512 {
 513         vvp_transient_page_verify(slice->cpl_page);
 514         vvp_page_completion_write_common(env, slice, ioret);
 515 }
 516
 517
 518 static void vvp_transient_page_fini(const struct lu_env *env,
 519                                     struct cl_page_slice *slice)
 520 {
 521         struct ccc_page *cp = cl2ccc_page(slice);
 522         struct cl_page *clp = slice->cpl_page;
 523         struct ccc_object *clobj = cl2ccc(clp->cp_obj);
 524
 525         vvp_page_fini_common(cp);
 526         LASSERT(!TRYLOCK_INODE_MUTEX(clobj->cob_inode));
 527         clobj->cob_transient_pages--;
 528 }
 529
 530 static const struct cl_page_operations vvp_transient_page_ops = {
 531         .cpo_own           = vvp_transient_page_own,
 532         .cpo_assume        = vvp_transient_page_assume,
 533         .cpo_unassume      = vvp_transient_page_unassume,
 534         .cpo_disown        = vvp_transient_page_disown,
 535         .cpo_discard       = vvp_transient_page_discard,
 536         .cpo_vmpage        = ccc_page_vmpage,
 537         .cpo_fini          = vvp_transient_page_fini,
 538         .cpo_is_vmlocked   = vvp_transient_page_is_vmlocked,
 539         .cpo_print         = vvp_page_print,
 540         .cpo_is_under_lock = ccc_page_is_under_lock,
 541         .io = {
 542                 [CRT_READ] = {
 543                         .cpo_prep        = ccc_transient_page_prep,
 544                         .cpo_completion  = vvp_page_completion_read,
 545                 },
 546                 [CRT_WRITE] = {
 547                         .cpo_prep        = ccc_transient_page_prep,
 548                         .cpo_completion  = vvp_transient_page_completion_write,
 549                 }
 550         }
 551 };
 552
 553 struct cl_page *vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 554                               struct cl_page *page, cfs_page_t *vmpage)
 555 {
 556         struct ccc_page *cpg;
 557         int result;
 558
 559         CLOBINVRNT(env, obj, ccc_object_invariant(obj));
 560
 561         OBD_SLAB_ALLOC_PTR_GFP(cpg, vvp_page_kmem, CFS_ALLOC_IO);
 562         if (cpg != NULL) {
 563                 cpg->cpg_page = vmpage;
 564                 page_cache_get(vmpage);
 565
 566                 CFS_INIT_LIST_HEAD(&cpg->cpg_pending_linkage);
 567                 if (page->cp_type == CPT_CACHEABLE) {
 568                         SetPagePrivate(vmpage);
 569                         vmpage->private = (unsigned long)page;
 570                         cl_page_slice_add(page, &cpg->cpg_cl, obj,
 571                                           &vvp_page_ops);
 572                 } else {
 573                         struct ccc_object *clobj = cl2ccc(obj);
 574
 575                         LASSERT(!TRYLOCK_INODE_MUTEX(clobj->cob_inode));
 576                         cl_page_slice_add(page, &cpg->cpg_cl, obj,
 577                                           &vvp_transient_page_ops);
 578                         clobj->cob_transient_pages++;
 579                 }
 580                 result = 0;
 581         } else
 582                 result = -ENOMEM;
 583         return ERR_PTR(result);
 584 }
 585