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 *_)
  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 > 100) {
  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\n",
  99                                       vmpage, current->comm, current,
 100                                       vmpage->flags);
 101                         libcfs_debug_dumpstack(NULL);
 102                         LCONSOLE_WARN("Reproduced bug #18881,please contact:"
 103                                "jay <jinshan.xiong@sun.com>, thanks\n");
 104
 105                         lock_page(vmpage);
 106                         break;
 107                 }
 108         }
 109         /* DEBUG CODE END */
 110
 111         /* lock_page(vmpage); */
 112         wait_on_page_writeback(vmpage);
 113 }
 114
 115 static void vvp_page_assume(const struct lu_env *env,
 116                             const struct cl_page_slice *slice, struct cl_io *_)
 117 {
 118         cfs_page_t *vmpage = cl2vm_page(slice);
 119
 120         LASSERT(vmpage != NULL);
 121         LASSERT(PageLocked(vmpage));
 122         wait_on_page_writeback(vmpage);
 123 }
 124
 125 static void vvp_page_unassume(const struct lu_env *env,
 126                               const struct cl_page_slice *slice,
 127                               struct cl_io *_)
 128 {
 129         cfs_page_t *vmpage = cl2vm_page(slice);
 130
 131         LASSERT(vmpage != NULL);
 132         LASSERT(PageLocked(vmpage));
 133 }
 134
 135 static void vvp_page_disown(const struct lu_env *env,
 136                             const struct cl_page_slice *slice, struct cl_io *io)
 137 {
 138         cfs_page_t *vmpage = cl2vm_page(slice);
 139
 140         LASSERT(vmpage != NULL);
 141         LASSERT(PageLocked(vmpage));
 142
 143         unlock_page(cl2vm_page(slice));
 144 }
 145
 146 static void vvp_page_discard(const struct lu_env *env,
 147                              const struct cl_page_slice *slice, struct cl_io *_)
 148 {
 149         cfs_page_t           *vmpage  = cl2vm_page(slice);
 150         struct address_space *mapping = vmpage->mapping;
 151         struct ccc_page      *cpg     = cl2ccc_page(slice);
 152
 153         LASSERT(vmpage != NULL);
 154         LASSERT(PageLocked(vmpage));
 155
 156         if (cpg->cpg_defer_uptodate && !cpg->cpg_ra_used)
 157                 ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
 158
 159         /*
 160          * truncate_complete_page() calls
 161          * a_ops->invalidatepage()->cl_page_delete()->vvp_page_delete().
 162          */
 163         truncate_complete_page(mapping, vmpage);
 164 }
 165
 166 static int vvp_page_unmap(const struct lu_env *env,
 167                           const struct cl_page_slice *slice, struct cl_io *_)
 168 {
 169         cfs_page_t *vmpage = cl2vm_page(slice);
 170         __u64       offset = vmpage->index << CFS_PAGE_SHIFT;
 171
 172         LASSERT(vmpage != NULL);
 173         LASSERT(PageLocked(vmpage));
 174         /*
 175          * XXX is it safe to call this with the page lock held?
 176          */
 177         ll_teardown_mmaps(vmpage->mapping, offset, offset + CFS_PAGE_SIZE);
 178         return 0;
 179 }
 180
 181 static void vvp_page_delete(const struct lu_env *env,
 182                             const struct cl_page_slice *slice)
 183 {
 184         cfs_page_t       *vmpage = cl2vm_page(slice);
 185         struct inode     *inode  = vmpage->mapping->host;
 186         struct cl_object *obj    = slice->cpl_obj;
 187
 188         LASSERT(PageLocked(vmpage));
 189         LASSERT((struct cl_page *)vmpage->private == slice->cpl_page);
 190         LASSERT(inode == ccc_object_inode(obj));
 191
 192         vvp_write_complete(cl2ccc(obj), cl2ccc_page(slice));
 193         ClearPagePrivate(vmpage);
 194         vmpage->private = 0;
 195         /*
 196          * Reference from vmpage to cl_page is removed, but the reference back
 197          * is still here. It is removed later in vvp_page_fini().
 198          */
 199 }
 200
 201 static void vvp_page_export(const struct lu_env *env,
 202                             const struct cl_page_slice *slice)
 203 {
 204         cfs_page_t *vmpage = cl2vm_page(slice);
 205
 206         LASSERT(vmpage != NULL);
 207         LASSERT(PageLocked(vmpage));
 208         SetPageUptodate(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 *_)
 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 *_)
 229 {
 230         cfs_page_t *vmpage = cl2vm_page(slice);
 231         int result;
 232
 233         if (clear_page_dirty_for_io(vmpage)) {
 234                 set_page_writeback(vmpage);
 235                 vvp_write_pending(cl2ccc(slice->cpl_obj), cl2ccc_page(slice));
 236                 result = 0;
 237         } else
 238                 result = -EALREADY;
 239         return result;
 240 }
 241
 242 /**
 243  * Handles page transfer errors at VM level.
 244  *
 245  * This takes inode as a separate argument, because inode on which error is to
 246  * be set can be different from \a vmpage inode in case of direct-io.
 247  */
 248 static void vvp_vmpage_error(struct inode *inode, cfs_page_t *vmpage, int ioret)
 249 {
 250         if (ioret == 0)
 251                 ClearPageError(vmpage);
 252         else if (ioret != -EINTR) {
 253                 SetPageError(vmpage);
 254                 if (ioret == -ENOSPC)
 255                         set_bit(AS_ENOSPC, &inode->i_mapping->flags);
 256                 else
 257                         set_bit(AS_EIO, &inode->i_mapping->flags);
 258         }
 259 }
 260
 261 static void vvp_page_completion_common(const struct lu_env *env,
 262                                        struct ccc_page *cp, int ioret)
 263 {
 264         struct cl_page    *clp    = cp->cpg_cl.cpl_page;
 265         cfs_page_t        *vmpage = cp->cpg_page;
 266         struct inode      *inode  = ccc_object_inode(clp->cp_obj);
 267         struct cl_sync_io *anchor = cp->cpg_sync_io;
 268
 269         LINVRNT(cl_page_is_vmlocked(env, clp));
 270
 271         if (anchor != NULL) {
 272                 cp->cpg_sync_io  = NULL;
 273                 cl_sync_io_note(anchor, ioret);
 274         } else if (clp->cp_type == CPT_CACHEABLE) {
 275                 /*
 276                  * Only mark the page error only when it's a cacheable page
 277                  * and NOT a sync io.
 278                  *
 279                  * For sync IO and direct IO(CPT_TRANSIENT), the error is able
 280                  * to be seen by application, so we don't need to mark a page
 281                  * as error at all.
 282                  */
 283                 vvp_vmpage_error(inode, vmpage, ioret);
 284                 unlock_page(vmpage);
 285         }
 286 }
 287
 288 static void vvp_page_completion_read(const struct lu_env *env,
 289                                      const struct cl_page_slice *slice,
 290                                      int ioret)
 291 {
 292         struct ccc_page *cp    = cl2ccc_page(slice);
 293         struct cl_page  *page  = cl_page_top(slice->cpl_page);
 294         struct inode    *inode = ccc_object_inode(page->cp_obj);
 295         ENTRY;
 296
 297         CL_PAGE_HEADER(D_PAGE, env, page, "completing READ with %d\n", ioret);
 298
 299         if (cp->cpg_defer_uptodate)
 300                 ll_ra_count_put(ll_i2sbi(inode), 1);
 301
 302         if (ioret == 0)  {
 303                 /* XXX: do we need this for transient pages? */
 304                 if (!cp->cpg_defer_uptodate)
 305                         cl_page_export(env, page);
 306         } else
 307                 cp->cpg_defer_uptodate = 0;
 308         vvp_page_completion_common(env, cp, ioret);
 309
 310         EXIT;
 311 }
 312
 313 static void vvp_page_completion_write_common(const struct lu_env *env,
 314                                              const struct cl_page_slice *slice,
 315                                              int ioret)
 316 {
 317         struct ccc_page *cp = cl2ccc_page(slice);
 318
 319         /*
 320          * TODO: Actually it makes sense to add the page into oap pending
 321          * list again and so that we don't need to take the page out from
 322          * SoM write pending list, if we just meet a recoverable error,
 323          * -ENOMEM, etc.
 324          * To implement this, we just need to return a non zero value in
 325          * ->cpo_completion method. The underlying transfer should be notified
 326          * and then re-add the page into pending transfer queue.  -jay
 327          */
 328         cp->cpg_write_queued = 0;
 329         vvp_write_complete(cl2ccc(slice->cpl_obj), cp);
 330
 331         vvp_page_completion_common(env, cp, ioret);
 332 }
 333
 334 static void vvp_page_completion_write(const struct lu_env *env,
 335                                       const struct cl_page_slice *slice,
 336                                       int ioret)
 337 {
 338         struct ccc_page *cp     = cl2ccc_page(slice);
 339         struct cl_page  *pg     = slice->cpl_page;
 340         cfs_page_t      *vmpage = cp->cpg_page;
 341
 342         ENTRY;
 343
 344         LINVRNT(cl_page_is_vmlocked(env, pg));
 345         LASSERT(PageWriteback(vmpage));
 346
 347         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 348
 349         vvp_page_completion_write_common(env, slice, ioret);
 350         end_page_writeback(vmpage);
 351         EXIT;
 352 }
 353
 354 /**
 355  * Implements cl_page_operations::cpo_make_ready() method.
 356  *
 357  * This is called to yank a page from the transfer cache and to send it out as
 358  * a part of transfer. This function try-locks the page. If try-lock failed,
 359  * page is owned by some concurrent IO, and should be skipped (this is bad,
 360  * but hopefully rare situation, as it usually results in transfer being
 361  * shorter than possible).
 362  *
 363  * \retval 0      success, page can be placed into transfer
 364  *
 365  * \retval -EAGAIN page is either used by concurrent IO has been
 366  * truncated. Skip it.
 367  */
 368 static int vvp_page_make_ready(const struct lu_env *env,
 369                                const struct cl_page_slice *slice)
 370 {
 371         cfs_page_t *vmpage = cl2vm_page(slice);
 372         struct cl_page *pg = slice->cpl_page;
 373         int result;
 374
 375         result = -EAGAIN;
 376         /* we're trying to write, but the page is locked.. come back later */
 377         if (!TestSetPageLocked(vmpage)) {
 378                 if (pg->cp_state == CPS_CACHED) {
 379                         /*
 380                          * We can cancel IO if page wasn't dirty after all.
 381                          */
 382                         clear_page_dirty_for_io(vmpage);
 383                         /*
 384                          * This actually clears the dirty bit in the radix
 385                          * tree.
 386                          */
 387                         set_page_writeback(vmpage);
 388                         vvp_write_pending(cl2ccc(slice->cpl_obj),
 389                                           cl2ccc_page(slice));
 390                         CL_PAGE_HEADER(D_PAGE, env, pg, "readied\n");
 391                         result = 0;
 392                 } else
 393                         /*
 394                          * Page was concurrently truncated.
 395                          */
 396                         LASSERT(pg->cp_state == CPS_FREEING);
 397         }
 398         RETURN(result);
 399 }
 400
 401 static int vvp_page_print(const struct lu_env *env,
 402                           const struct cl_page_slice *slice,
 403                           void *cookie, lu_printer_t printer)
 404 {
 405         struct ccc_page *vp = cl2ccc_page(slice);
 406         cfs_page_t      *vmpage = vp->cpg_page;
 407
 408         (*printer)(env, cookie, LUSTRE_VVP_NAME"-page@%p(%d:%d:%d) "
 409                    "vm@%p ",
 410                    vp, vp->cpg_defer_uptodate, vp->cpg_ra_used,
 411                    vp->cpg_write_queued, vmpage);
 412         if (vmpage != NULL) {
 413                 (*printer)(env, cookie, "%lx %d:%d %lx %lu %slru",
 414                            (long)vmpage->flags, page_count(vmpage),
 415                            page_mapcount(vmpage), vmpage->private,
 416                            page_index(vmpage),
 417                            list_empty(&vmpage->lru) ? "not-" : "");
 418         }
 419         (*printer)(env, cookie, "\n");
 420         return 0;
 421 }
 422
 423 static const struct cl_page_operations vvp_page_ops = {
 424         .cpo_own           = vvp_page_own,
 425         .cpo_assume        = vvp_page_assume,
 426         .cpo_unassume      = vvp_page_unassume,
 427         .cpo_disown        = vvp_page_disown,
 428         .cpo_vmpage        = ccc_page_vmpage,
 429         .cpo_discard       = vvp_page_discard,
 430         .cpo_delete        = vvp_page_delete,
 431         .cpo_unmap         = vvp_page_unmap,
 432         .cpo_export        = vvp_page_export,
 433         .cpo_is_vmlocked   = vvp_page_is_vmlocked,
 434         .cpo_fini          = vvp_page_fini,
 435         .cpo_print         = vvp_page_print,
 436         .cpo_is_under_lock = ccc_page_is_under_lock,
 437         .io = {
 438                 [CRT_READ] = {
 439                         .cpo_prep        = vvp_page_prep_read,
 440                         .cpo_completion  = vvp_page_completion_read,
 441                         .cpo_make_ready  = ccc_fail,
 442                 },
 443                 [CRT_WRITE] = {
 444                         .cpo_prep        = vvp_page_prep_write,
 445                         .cpo_completion  = vvp_page_completion_write,
 446                         .cpo_make_ready  = vvp_page_make_ready,
 447                 }
 448         }
 449 };
 450
 451 static void vvp_transient_page_verify(const struct cl_page *page)
 452 {
 453         struct inode *inode = ccc_object_inode(page->cp_obj);
 454
 455         LASSERT(!TRYLOCK_INODE_MUTEX(inode));
 456         /* LASSERT_SEM_LOCKED(&inode->i_alloc_sem); */
 457 }
 458
 459 static void vvp_transient_page_own(const struct lu_env *env,
 460                                    const struct cl_page_slice *slice,
 461                                    struct cl_io *_)
 462 {
 463         vvp_transient_page_verify(slice->cpl_page);
 464 }
 465
 466 static void vvp_transient_page_assume(const struct lu_env *env,
 467                                       const struct cl_page_slice *slice,
 468                                       struct cl_io *_)
 469 {
 470         vvp_transient_page_verify(slice->cpl_page);
 471 }
 472
 473 static void vvp_transient_page_unassume(const struct lu_env *env,
 474                                         const struct cl_page_slice *slice,
 475                                         struct cl_io *_)
 476 {
 477         vvp_transient_page_verify(slice->cpl_page);
 478 }
 479
 480 static void vvp_transient_page_disown(const struct lu_env *env,
 481                                       const struct cl_page_slice *slice,
 482                                       struct cl_io *_)
 483 {
 484         vvp_transient_page_verify(slice->cpl_page);
 485 }
 486
 487 static void vvp_transient_page_discard(const struct lu_env *env,
 488                                        const struct cl_page_slice *slice,
 489                                        struct cl_io *_)
 490 {
 491         struct cl_page *page = slice->cpl_page;
 492
 493         vvp_transient_page_verify(slice->cpl_page);
 494
 495         /*
 496          * For transient pages, remove it from the radix tree.
 497          */
 498         cl_page_delete(env, page);
 499 }
 500
 501 static int vvp_transient_page_is_vmlocked(const struct lu_env *env,
 502                                           const struct cl_page_slice *slice)
 503 {
 504         struct inode    *inode = ccc_object_inode(slice->cpl_obj);
 505         int              locked;
 506
 507         locked = !TRYLOCK_INODE_MUTEX(inode);
 508         if (!locked)
 509                 UNLOCK_INODE_MUTEX(inode);
 510         return locked ? -EBUSY : -ENODATA;
 511 }
 512
 513 static void
 514 vvp_transient_page_completion_write(const struct lu_env *env,
 515                                     const struct cl_page_slice *slice,
 516                                     int ioret)
 517 {
 518         vvp_transient_page_verify(slice->cpl_page);
 519         vvp_page_completion_write_common(env, slice, ioret);
 520 }
 521
 522
 523 static void vvp_transient_page_fini(const struct lu_env *env,
 524                                     struct cl_page_slice *slice)
 525 {
 526         struct ccc_page *cp = cl2ccc_page(slice);
 527         struct cl_page *clp = slice->cpl_page;
 528         struct ccc_object *clobj = cl2ccc(clp->cp_obj);
 529
 530         vvp_page_fini_common(cp);
 531         LASSERT(!TRYLOCK_INODE_MUTEX(clobj->cob_inode));
 532         clobj->cob_transient_pages--;
 533 }
 534
 535 static const struct cl_page_operations vvp_transient_page_ops = {
 536         .cpo_own           = vvp_transient_page_own,
 537         .cpo_assume        = vvp_transient_page_assume,
 538         .cpo_unassume      = vvp_transient_page_unassume,
 539         .cpo_disown        = vvp_transient_page_disown,
 540         .cpo_discard       = vvp_transient_page_discard,
 541         .cpo_vmpage        = ccc_page_vmpage,
 542         .cpo_fini          = vvp_transient_page_fini,
 543         .cpo_is_vmlocked   = vvp_transient_page_is_vmlocked,
 544         .cpo_print         = vvp_page_print,
 545         .cpo_is_under_lock = ccc_page_is_under_lock,
 546         .io = {
 547                 [CRT_READ] = {
 548                         .cpo_prep        = ccc_transient_page_prep,
 549                         .cpo_completion  = vvp_page_completion_read,
 550                 },
 551                 [CRT_WRITE] = {
 552                         .cpo_prep        = ccc_transient_page_prep,
 553                         .cpo_completion  = vvp_transient_page_completion_write,
 554                 }
 555         }
 556 };
 557
 558 struct cl_page *vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 559                               struct cl_page *page, cfs_page_t *vmpage)
 560 {
 561         struct ccc_page *cpg;
 562         int result;
 563
 564         CLOBINVRNT(env, obj, ccc_object_invariant(obj));
 565
 566         OBD_SLAB_ALLOC_PTR_GFP(cpg, vvp_page_kmem, CFS_ALLOC_IO);
 567         if (cpg != NULL) {
 568                 cpg->cpg_page = vmpage;
 569                 page_cache_get(vmpage);
 570
 571                 CFS_INIT_LIST_HEAD(&cpg->cpg_pending_linkage);
 572                 if (page->cp_type == CPT_CACHEABLE) {
 573                         SetPagePrivate(vmpage);
 574                         vmpage->private = (unsigned long)page;
 575                         cl_page_slice_add(page, &cpg->cpg_cl, obj,
 576                                           &vvp_page_ops);
 577                 } else {
 578                         struct ccc_object *clobj = cl2ccc(obj);
 579
 580                         LASSERT(!TRYLOCK_INODE_MUTEX(clobj->cob_inode));
 581                         cl_page_slice_add(page, &cpg->cpg_cl, obj,
 582                                           &vvp_transient_page_ops);
 583                         clobj->cob_transient_pages++;
 584                 }
 585                 result = 0;
 586         } else
 587                 result = -ENOMEM;
 588         return ERR_PTR(result);
 589 }
 590