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                         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,
 117                             struct cl_io *unused)
 118 {
 119         cfs_page_t *vmpage = cl2vm_page(slice);
 120
 121         LASSERT(vmpage != NULL);
 122         LASSERT(PageLocked(vmpage));
 123         wait_on_page_writeback(vmpage);
 124 }
 125
 126 static void vvp_page_unassume(const struct lu_env *env,
 127                               const struct cl_page_slice *slice,
 128                               struct cl_io *unused)
 129 {
 130         cfs_page_t *vmpage = cl2vm_page(slice);
 131
 132         LASSERT(vmpage != NULL);
 133         LASSERT(PageLocked(vmpage));
 134 }
 135
 136 static void vvp_page_disown(const struct lu_env *env,
 137                             const struct cl_page_slice *slice, struct cl_io *io)
 138 {
 139         cfs_page_t *vmpage = cl2vm_page(slice);
 140
 141         LASSERT(vmpage != NULL);
 142         LASSERT(PageLocked(vmpage));
 143
 144         unlock_page(cl2vm_page(slice));
 145 }
 146
 147 static void vvp_page_discard(const struct lu_env *env,
 148                              const struct cl_page_slice *slice,
 149                              struct cl_io *unused)
 150 {
 151         cfs_page_t           *vmpage  = cl2vm_page(slice);
 152         struct address_space *mapping = vmpage->mapping;
 153         struct ccc_page      *cpg     = cl2ccc_page(slice);
 154
 155         LASSERT(vmpage != NULL);
 156         LASSERT(PageLocked(vmpage));
 157
 158         if (cpg->cpg_defer_uptodate && !cpg->cpg_ra_used)
 159                 ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
 160
 161         /*
 162          * truncate_complete_page() calls
 163          * a_ops->invalidatepage()->cl_page_delete()->vvp_page_delete().
 164          */
 165         truncate_complete_page(mapping, vmpage);
 166 }
 167
 168 static int vvp_page_unmap(const struct lu_env *env,
 169                           const struct cl_page_slice *slice,
 170                           struct cl_io *unused)
 171 {
 172         cfs_page_t *vmpage = cl2vm_page(slice);
 173         __u64       offset = vmpage->index << CFS_PAGE_SHIFT;
 174
 175         LASSERT(vmpage != NULL);
 176         LASSERT(PageLocked(vmpage));
 177         /*
 178          * XXX is it safe to call this with the page lock held?
 179          */
 180         ll_teardown_mmaps(vmpage->mapping, offset, offset + CFS_PAGE_SIZE);
 181         return 0;
 182 }
 183
 184 static void vvp_page_delete(const struct lu_env *env,
 185                             const struct cl_page_slice *slice)
 186 {
 187         cfs_page_t       *vmpage = cl2vm_page(slice);
 188         struct inode     *inode  = vmpage->mapping->host;
 189         struct cl_object *obj    = slice->cpl_obj;
 190
 191         LASSERT(PageLocked(vmpage));
 192         LASSERT((struct cl_page *)vmpage->private == slice->cpl_page);
 193         LASSERT(inode == ccc_object_inode(obj));
 194
 195         vvp_write_complete(cl2ccc(obj), cl2ccc_page(slice));
 196         ClearPagePrivate(vmpage);
 197         vmpage->private = 0;
 198         /*
 199          * Reference from vmpage to cl_page is removed, but the reference back
 200          * is still here. It is removed later in vvp_page_fini().
 201          */
 202 }
 203
 204 static void vvp_page_export(const struct lu_env *env,
 205                             const struct cl_page_slice *slice)
 206 {
 207         cfs_page_t *vmpage = cl2vm_page(slice);
 208
 209         LASSERT(vmpage != NULL);
 210         LASSERT(PageLocked(vmpage));
 211         SetPageUptodate(vmpage);
 212 }
 213
 214 static int vvp_page_is_vmlocked(const struct lu_env *env,
 215                                 const struct cl_page_slice *slice)
 216 {
 217         return PageLocked(cl2vm_page(slice)) ? -EBUSY : -ENODATA;
 218 }
 219
 220 static int vvp_page_prep_read(const struct lu_env *env,
 221                               const struct cl_page_slice *slice,
 222                               struct cl_io *unused)
 223 {
 224         ENTRY;
 225         /* Skip the page already marked as PG_uptodate. */
 226         RETURN(PageUptodate(cl2vm_page(slice)) ? -EALREADY : 0);
 227 }
 228
 229 static int vvp_page_prep_write(const struct lu_env *env,
 230                                const struct cl_page_slice *slice,
 231                                struct cl_io *unused)
 232 {
 233         cfs_page_t *vmpage = cl2vm_page(slice);
 234         int result;
 235
 236         if (clear_page_dirty_for_io(vmpage)) {
 237                 set_page_writeback(vmpage);
 238                 vvp_write_pending(cl2ccc(slice->cpl_obj), cl2ccc_page(slice));
 239                 result = 0;
 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_common(const struct lu_env *env,
 265                                        struct ccc_page *cp, int ioret)
 266 {
 267         struct cl_page    *clp    = cp->cpg_cl.cpl_page;
 268         cfs_page_t        *vmpage = cp->cpg_page;
 269         struct inode      *inode  = ccc_object_inode(clp->cp_obj);
 270         struct cl_sync_io *anchor = cp->cpg_sync_io;
 271
 272         LINVRNT(cl_page_is_vmlocked(env, clp));
 273
 274         if (anchor != NULL) {
 275                 cp->cpg_sync_io  = NULL;
 276                 cl_sync_io_note(anchor, ioret);
 277         } else if (clp->cp_type == CPT_CACHEABLE) {
 278                 /*
 279                  * Only mark the page error only when it's a cacheable page
 280                  * and NOT a sync io.
 281                  *
 282                  * For sync IO and direct IO(CPT_TRANSIENT), the error is able
 283                  * to be seen by application, so we don't need to mark a page
 284                  * as error at all.
 285                  */
 286                 vvp_vmpage_error(inode, vmpage, ioret);
 287                 unlock_page(vmpage);
 288         }
 289 }
 290
 291 static void vvp_page_completion_read(const struct lu_env *env,
 292                                      const struct cl_page_slice *slice,
 293                                      int ioret)
 294 {
 295         struct ccc_page *cp    = cl2ccc_page(slice);
 296         struct cl_page  *page  = cl_page_top(slice->cpl_page);
 297         struct inode    *inode = ccc_object_inode(page->cp_obj);
 298         ENTRY;
 299
 300         CL_PAGE_HEADER(D_PAGE, env, page, "completing READ with %d\n", ioret);
 301
 302         if (cp->cpg_defer_uptodate)
 303                 ll_ra_count_put(ll_i2sbi(inode), 1);
 304
 305         if (ioret == 0)  {
 306                 /* XXX: do we need this for transient pages? */
 307                 if (!cp->cpg_defer_uptodate)
 308                         cl_page_export(env, page);
 309         } else
 310                 cp->cpg_defer_uptodate = 0;
 311         vvp_page_completion_common(env, cp, ioret);
 312
 313         EXIT;
 314 }
 315
 316 static void vvp_page_completion_write_common(const struct lu_env *env,
 317                                              const struct cl_page_slice *slice,
 318                                              int ioret)
 319 {
 320         struct ccc_page *cp = cl2ccc_page(slice);
 321
 322         /*
 323          * TODO: Actually it makes sense to add the page into oap pending
 324          * list again and so that we don't need to take the page out from
 325          * SoM write pending list, if we just meet a recoverable error,
 326          * -ENOMEM, etc.
 327          * To implement this, we just need to return a non zero value in
 328          * ->cpo_completion method. The underlying transfer should be notified
 329          * and then re-add the page into pending transfer queue.  -jay
 330          */
 331         cp->cpg_write_queued = 0;
 332         vvp_write_complete(cl2ccc(slice->cpl_obj), cp);
 333
 334         vvp_page_completion_common(env, cp, ioret);
 335 }
 336
 337 static void vvp_page_completion_write(const struct lu_env *env,
 338                                       const struct cl_page_slice *slice,
 339                                       int ioret)
 340 {
 341         struct ccc_page *cp     = cl2ccc_page(slice);
 342         struct cl_page  *pg     = slice->cpl_page;
 343         cfs_page_t      *vmpage = cp->cpg_page;
 344
 345         ENTRY;
 346
 347         LINVRNT(cl_page_is_vmlocked(env, pg));
 348         LASSERT(PageWriteback(vmpage));
 349
 350         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 351
 352         vvp_page_completion_write_common(env, slice, ioret);
 353         end_page_writeback(vmpage);
 354         EXIT;
 355 }
 356
 357 /**
 358  * Implements cl_page_operations::cpo_make_ready() method.
 359  *
 360  * This is called to yank a page from the transfer cache and to send it out as
 361  * a part of transfer. This function try-locks the page. If try-lock failed,
 362  * page is owned by some concurrent IO, and should be skipped (this is bad,
 363  * but hopefully rare situation, as it usually results in transfer being
 364  * shorter than possible).
 365  *
 366  * \retval 0      success, page can be placed into transfer
 367  *
 368  * \retval -EAGAIN page is either used by concurrent IO has been
 369  * truncated. Skip it.
 370  */
 371 static int vvp_page_make_ready(const struct lu_env *env,
 372                                const struct cl_page_slice *slice)
 373 {
 374         cfs_page_t *vmpage = cl2vm_page(slice);
 375         struct cl_page *pg = slice->cpl_page;
 376         int result;
 377
 378         result = -EAGAIN;
 379         /* we're trying to write, but the page is locked.. come back later */
 380         if (!TestSetPageLocked(vmpage)) {
 381                 if (pg->cp_state == CPS_CACHED) {
 382                         /*
 383                          * We can cancel IO if page wasn't dirty after all.
 384                          */
 385                         clear_page_dirty_for_io(vmpage);
 386                         /*
 387                          * This actually clears the dirty bit in the radix
 388                          * tree.
 389                          */
 390                         set_page_writeback(vmpage);
 391                         vvp_write_pending(cl2ccc(slice->cpl_obj),
 392                                           cl2ccc_page(slice));
 393                         CL_PAGE_HEADER(D_PAGE, env, pg, "readied\n");
 394                         result = 0;
 395                 } else
 396                         /*
 397                          * Page was concurrently truncated.
 398                          */
 399                         LASSERT(pg->cp_state == CPS_FREEING);
 400         }
 401         RETURN(result);
 402 }
 403
 404 static int vvp_page_print(const struct lu_env *env,
 405                           const struct cl_page_slice *slice,
 406                           void *cookie, lu_printer_t printer)
 407 {
 408         struct ccc_page *vp = cl2ccc_page(slice);
 409         cfs_page_t      *vmpage = vp->cpg_page;
 410
 411         (*printer)(env, cookie, LUSTRE_VVP_NAME"-page@%p(%d:%d:%d) "
 412                    "vm@%p ",
 413                    vp, vp->cpg_defer_uptodate, vp->cpg_ra_used,
 414                    vp->cpg_write_queued, vmpage);
 415         if (vmpage != NULL) {
 416                 (*printer)(env, cookie, "%lx %d:%d %lx %lu %slru",
 417                            (long)vmpage->flags, page_count(vmpage),
 418                            page_mapcount(vmpage), vmpage->private,
 419                            page_index(vmpage),
 420                            list_empty(&vmpage->lru) ? "not-" : "");
 421         }
 422         (*printer)(env, cookie, "\n");
 423         return 0;
 424 }
 425
 426 static const struct cl_page_operations vvp_page_ops = {
 427         .cpo_own           = vvp_page_own,
 428         .cpo_assume        = vvp_page_assume,
 429         .cpo_unassume      = vvp_page_unassume,
 430         .cpo_disown        = vvp_page_disown,
 431         .cpo_vmpage        = ccc_page_vmpage,
 432         .cpo_discard       = vvp_page_discard,
 433         .cpo_delete        = vvp_page_delete,
 434         .cpo_unmap         = vvp_page_unmap,
 435         .cpo_export        = vvp_page_export,
 436         .cpo_is_vmlocked   = vvp_page_is_vmlocked,
 437         .cpo_fini          = vvp_page_fini,
 438         .cpo_print         = vvp_page_print,
 439         .cpo_is_under_lock = ccc_page_is_under_lock,
 440         .io = {
 441                 [CRT_READ] = {
 442                         .cpo_prep        = vvp_page_prep_read,
 443                         .cpo_completion  = vvp_page_completion_read,
 444                         .cpo_make_ready  = ccc_fail,
 445                 },
 446                 [CRT_WRITE] = {
 447                         .cpo_prep        = vvp_page_prep_write,
 448                         .cpo_completion  = vvp_page_completion_write,
 449                         .cpo_make_ready  = vvp_page_make_ready,
 450                 }
 451         }
 452 };
 453
 454 static void vvp_transient_page_verify(const struct cl_page *page)
 455 {
 456         struct inode *inode = ccc_object_inode(page->cp_obj);
 457
 458         LASSERT(!TRYLOCK_INODE_MUTEX(inode));
 459         /* LASSERT_SEM_LOCKED(&inode->i_alloc_sem); */
 460 }
 461
 462 static void vvp_transient_page_own(const struct lu_env *env,
 463                                    const struct cl_page_slice *slice,
 464                                    struct cl_io *unused)
 465 {
 466         vvp_transient_page_verify(slice->cpl_page);
 467 }
 468
 469 static void vvp_transient_page_assume(const struct lu_env *env,
 470                                       const struct cl_page_slice *slice,
 471                                       struct cl_io *unused)
 472 {
 473         vvp_transient_page_verify(slice->cpl_page);
 474 }
 475
 476 static void vvp_transient_page_unassume(const struct lu_env *env,
 477                                         const struct cl_page_slice *slice,
 478                                         struct cl_io *unused)
 479 {
 480         vvp_transient_page_verify(slice->cpl_page);
 481 }
 482
 483 static void vvp_transient_page_disown(const struct lu_env *env,
 484                                       const struct cl_page_slice *slice,
 485                                       struct cl_io *unused)
 486 {
 487         vvp_transient_page_verify(slice->cpl_page);
 488 }
 489
 490 static void vvp_transient_page_discard(const struct lu_env *env,
 491                                        const struct cl_page_slice *slice,
 492                                        struct cl_io *unused)
 493 {
 494         struct cl_page *page = slice->cpl_page;
 495
 496         vvp_transient_page_verify(slice->cpl_page);
 497
 498         /*
 499          * For transient pages, remove it from the radix tree.
 500          */
 501         cl_page_delete(env, page);
 502 }
 503
 504 static int vvp_transient_page_is_vmlocked(const struct lu_env *env,
 505                                           const struct cl_page_slice *slice)
 506 {
 507         struct inode    *inode = ccc_object_inode(slice->cpl_obj);
 508         int              locked;
 509
 510         locked = !TRYLOCK_INODE_MUTEX(inode);
 511         if (!locked)
 512                 UNLOCK_INODE_MUTEX(inode);
 513         return locked ? -EBUSY : -ENODATA;
 514 }
 515
 516 static void
 517 vvp_transient_page_completion_write(const struct lu_env *env,
 518                                     const struct cl_page_slice *slice,
 519                                     int ioret)
 520 {
 521         vvp_transient_page_verify(slice->cpl_page);
 522         vvp_page_completion_write_common(env, slice, ioret);
 523 }
 524
 525
 526 static void vvp_transient_page_fini(const struct lu_env *env,
 527                                     struct cl_page_slice *slice)
 528 {
 529         struct ccc_page *cp = cl2ccc_page(slice);
 530         struct cl_page *clp = slice->cpl_page;
 531         struct ccc_object *clobj = cl2ccc(clp->cp_obj);
 532
 533         vvp_page_fini_common(cp);
 534         LASSERT(!TRYLOCK_INODE_MUTEX(clobj->cob_inode));
 535         clobj->cob_transient_pages--;
 536 }
 537
 538 static const struct cl_page_operations vvp_transient_page_ops = {
 539         .cpo_own           = vvp_transient_page_own,
 540         .cpo_assume        = vvp_transient_page_assume,
 541         .cpo_unassume      = vvp_transient_page_unassume,
 542         .cpo_disown        = vvp_transient_page_disown,
 543         .cpo_discard       = vvp_transient_page_discard,
 544         .cpo_vmpage        = ccc_page_vmpage,
 545         .cpo_fini          = vvp_transient_page_fini,
 546         .cpo_is_vmlocked   = vvp_transient_page_is_vmlocked,
 547         .cpo_print         = vvp_page_print,
 548         .cpo_is_under_lock = ccc_page_is_under_lock,
 549         .io = {
 550                 [CRT_READ] = {
 551                         .cpo_prep        = ccc_transient_page_prep,
 552                         .cpo_completion  = vvp_page_completion_read,
 553                 },
 554                 [CRT_WRITE] = {
 555                         .cpo_prep        = ccc_transient_page_prep,
 556                         .cpo_completion  = vvp_transient_page_completion_write,
 557                 }
 558         }
 559 };
 560
 561 struct cl_page *vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 562                               struct cl_page *page, cfs_page_t *vmpage)
 563 {
 564         struct ccc_page *cpg;
 565         int result;
 566
 567         CLOBINVRNT(env, obj, ccc_object_invariant(obj));
 568
 569         OBD_SLAB_ALLOC_PTR_GFP(cpg, vvp_page_kmem, CFS_ALLOC_IO);
 570         if (cpg != NULL) {
 571                 cpg->cpg_page = vmpage;
 572                 page_cache_get(vmpage);
 573
 574                 CFS_INIT_LIST_HEAD(&cpg->cpg_pending_linkage);
 575                 if (page->cp_type == CPT_CACHEABLE) {
 576                         SetPagePrivate(vmpage);
 577                         vmpage->private = (unsigned long)page;
 578                         cl_page_slice_add(page, &cpg->cpg_cl, obj,
 579                                           &vvp_page_ops);
 580                 } else {
 581                         struct ccc_object *clobj = cl2ccc(obj);
 582
 583                         LASSERT(!TRYLOCK_INODE_MUTEX(clobj->cob_inode));
 584                         cl_page_slice_add(page, &cpg->cpg_cl, obj,
 585                                           &vvp_transient_page_ops);
 586                         clobj->cob_transient_pages++;
 587                 }
 588                 result = 0;
 589         } else
 590                 result = -ENOMEM;
 591         return ERR_PTR(result);
 592 }
 593