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 2008 Sun Microsystems, Inc.  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 void vvp_page_own(const struct lu_env *env,
  82                          const struct cl_page_slice *slice, struct cl_io *io)
  83 {
  84         struct ccc_page *vpg    = cl2ccc_page(slice);
  85         cfs_page_t      *vmpage = vpg->cpg_page;
  86         int count = 0;
  87
  88         LASSERT(vmpage != NULL);
  89
  90         /* DEBUG CODE FOR #18881 */
  91         while (TestSetPageLocked(vmpage)) {
  92                 cfs_schedule_timeout(CFS_TASK_INTERRUPTIBLE,
  93                                      cfs_time_seconds(1)/10);
  94                 if (++count > 600) {
  95                         CL_PAGE_DEBUG(D_ERROR, env,
  96                                       cl_page_top(slice->cpl_page),
  97                                       "XXX page %p blocked on acquiring the"
  98                                       " lock. process %s/%p, flags %lx,io %p\n",
  99                                       vmpage, current->comm, current,
 100                                       vmpage->flags, io);
 101                         libcfs_debug_dumpstack(NULL);
 102                         if (slice->cpl_page->cp_task) {
 103                                 cfs_task_t *tsk = slice->cpl_page->cp_task;
 104                                 LCONSOLE_WARN("The page was owned by %s\n",
 105                                               tsk->comm);
 106                                 libcfs_debug_dumpstack(tsk);
 107                         }
 108                         LCONSOLE_WARN("Reproduced bug #18881,please contact:"
 109                                "jay <jinshan.xiong@sun.com>, thanks\n");
 110
 111                         lock_page(vmpage);
 112                         break;
 113                 }
 114         }
 115         /* DEBUG CODE END */
 116
 117         /* lock_page(vmpage); */
 118         wait_on_page_writeback(vmpage);
 119 }
 120
 121 static void vvp_page_assume(const struct lu_env *env,
 122                             const struct cl_page_slice *slice,
 123                             struct cl_io *unused)
 124 {
 125         cfs_page_t *vmpage = cl2vm_page(slice);
 126
 127         LASSERT(vmpage != NULL);
 128         LASSERT(PageLocked(vmpage));
 129         wait_on_page_writeback(vmpage);
 130 }
 131
 132 static void vvp_page_unassume(const struct lu_env *env,
 133                               const struct cl_page_slice *slice,
 134                               struct cl_io *unused)
 135 {
 136         cfs_page_t *vmpage = cl2vm_page(slice);
 137
 138         LASSERT(vmpage != NULL);
 139         LASSERT(PageLocked(vmpage));
 140 }
 141
 142 static void vvp_page_disown(const struct lu_env *env,
 143                             const struct cl_page_slice *slice, struct cl_io *io)
 144 {
 145         cfs_page_t *vmpage = cl2vm_page(slice);
 146
 147         LASSERT(vmpage != NULL);
 148         LASSERT(PageLocked(vmpage));
 149
 150         unlock_page(cl2vm_page(slice));
 151 }
 152
 153 static void vvp_page_discard(const struct lu_env *env,
 154                              const struct cl_page_slice *slice,
 155                              struct cl_io *unused)
 156 {
 157         cfs_page_t           *vmpage  = cl2vm_page(slice);
 158         struct address_space *mapping = vmpage->mapping;
 159         struct ccc_page      *cpg     = cl2ccc_page(slice);
 160
 161         LASSERT(vmpage != NULL);
 162         LASSERT(PageLocked(vmpage));
 163
 164         if (cpg->cpg_defer_uptodate && !cpg->cpg_ra_used)
 165                 ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
 166
 167         /*
 168          * truncate_complete_page() calls
 169          * a_ops->invalidatepage()->cl_page_delete()->vvp_page_delete().
 170          */
 171         truncate_complete_page(mapping, vmpage);
 172 }
 173
 174 static int vvp_page_unmap(const struct lu_env *env,
 175                           const struct cl_page_slice *slice,
 176                           struct cl_io *unused)
 177 {
 178         cfs_page_t *vmpage = cl2vm_page(slice);
 179         __u64       offset = vmpage->index << CFS_PAGE_SHIFT;
 180
 181         LASSERT(vmpage != NULL);
 182         LASSERT(PageLocked(vmpage));
 183         /*
 184          * XXX is it safe to call this with the page lock held?
 185          */
 186         ll_teardown_mmaps(vmpage->mapping, offset, offset + CFS_PAGE_SIZE);
 187         return 0;
 188 }
 189
 190 static void vvp_page_delete(const struct lu_env *env,
 191                             const struct cl_page_slice *slice)
 192 {
 193         cfs_page_t       *vmpage = cl2vm_page(slice);
 194         struct inode     *inode  = vmpage->mapping->host;
 195         struct cl_object *obj    = slice->cpl_obj;
 196
 197         LASSERT(PageLocked(vmpage));
 198         LASSERT((struct cl_page *)vmpage->private == slice->cpl_page);
 199         LASSERT(inode == ccc_object_inode(obj));
 200
 201         vvp_write_complete(cl2ccc(obj), cl2ccc_page(slice));
 202         ClearPagePrivate(vmpage);
 203         vmpage->private = 0;
 204         /*
 205          * Reference from vmpage to cl_page is removed, but the reference back
 206          * is still here. It is removed later in vvp_page_fini().
 207          */
 208 }
 209
 210 static void vvp_page_export(const struct lu_env *env,
 211                             const struct cl_page_slice *slice,
 212                             int uptodate)
 213 {
 214         cfs_page_t *vmpage = cl2vm_page(slice);
 215
 216         LASSERT(vmpage != NULL);
 217         LASSERT(PageLocked(vmpage));
 218         if (uptodate)
 219                 SetPageUptodate(vmpage);
 220         else
 221                 ClearPageUptodate(vmpage);
 222 }
 223
 224 static int vvp_page_is_vmlocked(const struct lu_env *env,
 225                                 const struct cl_page_slice *slice)
 226 {
 227         return PageLocked(cl2vm_page(slice)) ? -EBUSY : -ENODATA;
 228 }
 229
 230 static int vvp_page_prep_read(const struct lu_env *env,
 231                               const struct cl_page_slice *slice,
 232                               struct cl_io *unused)
 233 {
 234         ENTRY;
 235         /* Skip the page already marked as PG_uptodate. */
 236         RETURN(PageUptodate(cl2vm_page(slice)) ? -EALREADY : 0);
 237 }
 238
 239 static int vvp_page_prep_write(const struct lu_env *env,
 240                                const struct cl_page_slice *slice,
 241                                struct cl_io *unused)
 242 {
 243         cfs_page_t *vmpage = cl2vm_page(slice);
 244         int result;
 245
 246         if (clear_page_dirty_for_io(vmpage)) {
 247                 set_page_writeback(vmpage);
 248                 vvp_write_pending(cl2ccc(slice->cpl_obj), cl2ccc_page(slice));
 249                 result = 0;
 250         } else
 251                 result = -EALREADY;
 252         return result;
 253 }
 254
 255 /**
 256  * Handles page transfer errors at VM level.
 257  *
 258  * This takes inode as a separate argument, because inode on which error is to
 259  * be set can be different from \a vmpage inode in case of direct-io.
 260  */
 261 static void vvp_vmpage_error(struct inode *inode, cfs_page_t *vmpage, int ioret)
 262 {
 263         if (ioret == 0)
 264                 ClearPageError(vmpage);
 265         else if (ioret != -EINTR) {
 266                 SetPageError(vmpage);
 267                 if (ioret == -ENOSPC)
 268                         set_bit(AS_ENOSPC, &inode->i_mapping->flags);
 269                 else
 270                         set_bit(AS_EIO, &inode->i_mapping->flags);
 271         }
 272 }
 273
 274 static void vvp_page_completion_common(const struct lu_env *env,
 275                                        struct ccc_page *cp, int ioret)
 276 {
 277         struct cl_page    *clp    = cp->cpg_cl.cpl_page;
 278         cfs_page_t        *vmpage = cp->cpg_page;
 279         struct inode      *inode  = ccc_object_inode(clp->cp_obj);
 280
 281         LINVRNT(cl_page_is_vmlocked(env, clp));
 282
 283         if (!clp->cp_sync_io && clp->cp_type == CPT_CACHEABLE) {
 284                 /*
 285                  * Only mark the page error only when it's a cacheable page
 286                  * and NOT a sync io.
 287                  *
 288                  * For sync IO and direct IO(CPT_TRANSIENT), the error is able
 289                  * to be seen by application, so we don't need to mark a page
 290                  * as error at all.
 291                  */
 292                 vvp_vmpage_error(inode, vmpage, ioret);
 293                 unlock_page(vmpage);
 294         }
 295 }
 296
 297 static void vvp_page_completion_read(const struct lu_env *env,
 298                                      const struct cl_page_slice *slice,
 299                                      int ioret)
 300 {
 301         struct ccc_page *cp    = cl2ccc_page(slice);
 302         struct cl_page  *page  = cl_page_top(slice->cpl_page);
 303         struct inode    *inode = ccc_object_inode(page->cp_obj);
 304         ENTRY;
 305
 306         CL_PAGE_HEADER(D_PAGE, env, page, "completing READ with %d\n", ioret);
 307
 308         if (cp->cpg_defer_uptodate)
 309                 ll_ra_count_put(ll_i2sbi(inode), 1);
 310
 311         if (ioret == 0)  {
 312                 /* XXX: do we need this for transient pages? */
 313                 if (!cp->cpg_defer_uptodate)
 314                         cl_page_export(env, page, 1);
 315         } else
 316                 cp->cpg_defer_uptodate = 0;
 317         vvp_page_completion_common(env, cp, ioret);
 318
 319         EXIT;
 320 }
 321
 322 static void vvp_page_completion_write_common(const struct lu_env *env,
 323                                              const struct cl_page_slice *slice,
 324                                              int ioret)
 325 {
 326         struct ccc_page *cp = cl2ccc_page(slice);
 327
 328         /*
 329          * TODO: Actually it makes sense to add the page into oap pending
 330          * list again and so that we don't need to take the page out from
 331          * SoM write pending list, if we just meet a recoverable error,
 332          * -ENOMEM, etc.
 333          * To implement this, we just need to return a non zero value in
 334          * ->cpo_completion method. The underlying transfer should be notified
 335          * and then re-add the page into pending transfer queue.  -jay
 336          */
 337         cp->cpg_write_queued = 0;
 338         vvp_write_complete(cl2ccc(slice->cpl_obj), cp);
 339
 340         vvp_page_completion_common(env, cp, ioret);
 341 }
 342
 343 static void vvp_page_completion_write(const struct lu_env *env,
 344                                       const struct cl_page_slice *slice,
 345                                       int ioret)
 346 {
 347         struct ccc_page *cp     = cl2ccc_page(slice);
 348         struct cl_page  *pg     = slice->cpl_page;
 349         cfs_page_t      *vmpage = cp->cpg_page;
 350
 351         ENTRY;
 352
 353         LINVRNT(cl_page_is_vmlocked(env, pg));
 354         LASSERT(PageWriteback(vmpage));
 355
 356         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 357
 358         vvp_page_completion_write_common(env, slice, ioret);
 359         end_page_writeback(vmpage);
 360         EXIT;
 361 }
 362
 363 /**
 364  * Implements cl_page_operations::cpo_make_ready() method.
 365  *
 366  * This is called to yank a page from the transfer cache and to send it out as
 367  * a part of transfer. This function try-locks the page. If try-lock failed,
 368  * page is owned by some concurrent IO, and should be skipped (this is bad,
 369  * but hopefully rare situation, as it usually results in transfer being
 370  * shorter than possible).
 371  *
 372  * \retval 0      success, page can be placed into transfer
 373  *
 374  * \retval -EAGAIN page is either used by concurrent IO has been
 375  * truncated. Skip it.
 376  */
 377 static int vvp_page_make_ready(const struct lu_env *env,
 378                                const struct cl_page_slice *slice)
 379 {
 380         cfs_page_t *vmpage = cl2vm_page(slice);
 381         struct cl_page *pg = slice->cpl_page;
 382         int result;
 383
 384         result = -EAGAIN;
 385         /* we're trying to write, but the page is locked.. come back later */
 386         if (!TestSetPageLocked(vmpage)) {
 387                 if (pg->cp_state == CPS_CACHED) {
 388                         /*
 389                          * We can cancel IO if page wasn't dirty after all.
 390                          */
 391                         clear_page_dirty_for_io(vmpage);
 392                         /*
 393                          * This actually clears the dirty bit in the radix
 394                          * tree.
 395                          */
 396                         set_page_writeback(vmpage);
 397                         vvp_write_pending(cl2ccc(slice->cpl_obj),
 398                                           cl2ccc_page(slice));
 399                         CL_PAGE_HEADER(D_PAGE, env, pg, "readied\n");
 400                         result = 0;
 401                 } else
 402                         /*
 403                          * Page was concurrently truncated.
 404                          */
 405                         LASSERT(pg->cp_state == CPS_FREEING);
 406         }
 407         RETURN(result);
 408 }
 409
 410 static int vvp_page_print(const struct lu_env *env,
 411                           const struct cl_page_slice *slice,
 412                           void *cookie, lu_printer_t printer)
 413 {
 414         struct ccc_page *vp = cl2ccc_page(slice);
 415         cfs_page_t      *vmpage = vp->cpg_page;
 416
 417         (*printer)(env, cookie, LUSTRE_VVP_NAME"-page@%p(%d:%d:%d) "
 418                    "vm@%p ",
 419                    vp, vp->cpg_defer_uptodate, vp->cpg_ra_used,
 420                    vp->cpg_write_queued, vmpage);
 421         if (vmpage != NULL) {
 422                 (*printer)(env, cookie, "%lx %d:%d %lx %lu %slru",
 423                            (long)vmpage->flags, page_count(vmpage),
 424                            page_mapcount(vmpage), vmpage->private,
 425                            page_index(vmpage),
 426                            list_empty(&vmpage->lru) ? "not-" : "");
 427         }
 428         (*printer)(env, cookie, "\n");
 429         return 0;
 430 }
 431
 432 static const struct cl_page_operations vvp_page_ops = {
 433         .cpo_own           = vvp_page_own,
 434         .cpo_assume        = vvp_page_assume,
 435         .cpo_unassume      = vvp_page_unassume,
 436         .cpo_disown        = vvp_page_disown,
 437         .cpo_vmpage        = ccc_page_vmpage,
 438         .cpo_discard       = vvp_page_discard,
 439         .cpo_delete        = vvp_page_delete,
 440         .cpo_unmap         = vvp_page_unmap,
 441         .cpo_export        = vvp_page_export,
 442         .cpo_is_vmlocked   = vvp_page_is_vmlocked,
 443         .cpo_fini          = vvp_page_fini,
 444         .cpo_print         = vvp_page_print,
 445         .cpo_is_under_lock = ccc_page_is_under_lock,
 446         .io = {
 447                 [CRT_READ] = {
 448                         .cpo_prep        = vvp_page_prep_read,
 449                         .cpo_completion  = vvp_page_completion_read,
 450                         .cpo_make_ready  = ccc_fail,
 451                 },
 452                 [CRT_WRITE] = {
 453                         .cpo_prep        = vvp_page_prep_write,
 454                         .cpo_completion  = vvp_page_completion_write,
 455                         .cpo_make_ready  = vvp_page_make_ready,
 456                 }
 457         }
 458 };
 459
 460 static void vvp_transient_page_verify(const struct cl_page *page)
 461 {
 462         struct inode *inode = ccc_object_inode(page->cp_obj);
 463
 464         LASSERT(!TRYLOCK_INODE_MUTEX(inode));
 465         /* LASSERT_SEM_LOCKED(&inode->i_alloc_sem); */
 466 }
 467
 468 static void vvp_transient_page_own(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_assume(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_unassume(const struct lu_env *env,
 483                                         const struct cl_page_slice *slice,
 484                                         struct cl_io *unused)
 485 {
 486         vvp_transient_page_verify(slice->cpl_page);
 487 }
 488
 489 static void vvp_transient_page_disown(const struct lu_env *env,
 490                                       const struct cl_page_slice *slice,
 491                                       struct cl_io *unused)
 492 {
 493         vvp_transient_page_verify(slice->cpl_page);
 494 }
 495
 496 static void vvp_transient_page_discard(const struct lu_env *env,
 497                                        const struct cl_page_slice *slice,
 498                                        struct cl_io *unused)
 499 {
 500         struct cl_page *page = slice->cpl_page;
 501
 502         vvp_transient_page_verify(slice->cpl_page);
 503
 504         /*
 505          * For transient pages, remove it from the radix tree.
 506          */
 507         cl_page_delete(env, page);
 508 }
 509
 510 static int vvp_transient_page_is_vmlocked(const struct lu_env *env,
 511                                           const struct cl_page_slice *slice)
 512 {
 513         struct inode    *inode = ccc_object_inode(slice->cpl_obj);
 514         int              locked;
 515
 516         locked = !TRYLOCK_INODE_MUTEX(inode);
 517         if (!locked)
 518                 UNLOCK_INODE_MUTEX(inode);
 519         return locked ? -EBUSY : -ENODATA;
 520 }
 521
 522 static void
 523 vvp_transient_page_completion_write(const struct lu_env *env,
 524                                     const struct cl_page_slice *slice,
 525                                     int ioret)
 526 {
 527         vvp_transient_page_verify(slice->cpl_page);
 528         vvp_page_completion_write_common(env, slice, ioret);
 529 }
 530
 531
 532 static void vvp_transient_page_fini(const struct lu_env *env,
 533                                     struct cl_page_slice *slice)
 534 {
 535         struct ccc_page *cp = cl2ccc_page(slice);
 536         struct cl_page *clp = slice->cpl_page;
 537         struct ccc_object *clobj = cl2ccc(clp->cp_obj);
 538
 539         vvp_page_fini_common(cp);
 540         LASSERT(!TRYLOCK_INODE_MUTEX(clobj->cob_inode));
 541         clobj->cob_transient_pages--;
 542 }
 543
 544 static const struct cl_page_operations vvp_transient_page_ops = {
 545         .cpo_own           = vvp_transient_page_own,
 546         .cpo_assume        = vvp_transient_page_assume,
 547         .cpo_unassume      = vvp_transient_page_unassume,
 548         .cpo_disown        = vvp_transient_page_disown,
 549         .cpo_discard       = vvp_transient_page_discard,
 550         .cpo_vmpage        = ccc_page_vmpage,
 551         .cpo_fini          = vvp_transient_page_fini,
 552         .cpo_is_vmlocked   = vvp_transient_page_is_vmlocked,
 553         .cpo_print         = vvp_page_print,
 554         .cpo_is_under_lock = ccc_page_is_under_lock,
 555         .io = {
 556                 [CRT_READ] = {
 557                         .cpo_prep        = ccc_transient_page_prep,
 558                         .cpo_completion  = vvp_page_completion_read,
 559                 },
 560                 [CRT_WRITE] = {
 561                         .cpo_prep        = ccc_transient_page_prep,
 562                         .cpo_completion  = vvp_transient_page_completion_write,
 563                 }
 564         }
 565 };
 566
 567 struct cl_page *vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 568                               struct cl_page *page, cfs_page_t *vmpage)
 569 {
 570         struct ccc_page *cpg;
 571         int result;
 572
 573         CLOBINVRNT(env, obj, ccc_object_invariant(obj));
 574
 575         OBD_SLAB_ALLOC_PTR_GFP(cpg, vvp_page_kmem, CFS_ALLOC_IO);
 576         if (cpg != NULL) {
 577                 cpg->cpg_page = vmpage;
 578                 page_cache_get(vmpage);
 579
 580                 CFS_INIT_LIST_HEAD(&cpg->cpg_pending_linkage);
 581                 if (page->cp_type == CPT_CACHEABLE) {
 582                         SetPagePrivate(vmpage);
 583                         vmpage->private = (unsigned long)page;
 584                         cl_page_slice_add(page, &cpg->cpg_cl, obj,
 585                                           &vvp_page_ops);
 586                 } else {
 587                         struct ccc_object *clobj = cl2ccc(obj);
 588
 589                         LASSERT(!TRYLOCK_INODE_MUTEX(clobj->cob_inode));
 590                         cl_page_slice_add(page, &cpg->cpg_cl, obj,
 591                                           &vvp_transient_page_ops);
 592                         clobj->cob_transient_pages++;
 593                 }
 594                 result = 0;
 595         } else
 596                 result = -ENOMEM;
 597         return ERR_PTR(result);
 598 }
 599