lustre/llite/vvp_page.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2014, Intel Corporation.
  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  *   Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_LLITE
  43
  44 #include <linux/atomic.h>
  45 #include <linux/bitops.h>
  46 #include <linux/mm.h>
  47 #include <linux/mutex.h>
  48 #include <linux/page-flags.h>
  49 #include <linux/pagemap.h>
  50
  51 #include <libcfs/libcfs.h>
  52 #include "llite_internal.h"
  53 #include "vvp_internal.h"
  54
  55 /*****************************************************************************
  56  *
  57  * Page operations.
  58  *
  59  */
  60
  61 static void vvp_page_fini_common(struct vvp_page *vpg)
  62 {
  63         struct page *vmpage = vpg->vpg_page;
  64
  65         LASSERT(vmpage != NULL);
  66         page_cache_release(vmpage);
  67 }
  68
  69 static void vvp_page_fini(const struct lu_env *env,
  70                           struct cl_page_slice *slice)
  71 {
  72         struct vvp_page *vpg     = cl2vvp_page(slice);
  73         struct page     *vmpage  = vpg->vpg_page;
  74
  75         /*
  76          * vmpage->private was already cleared when page was moved into
  77          * VPG_FREEING state.
  78          */
  79         LASSERT((struct cl_page *)vmpage->private != slice->cpl_page);
  80         vvp_page_fini_common(vpg);
  81 }
  82
  83 static int vvp_page_own(const struct lu_env *env,
  84                         const struct cl_page_slice *slice, struct cl_io *io,
  85                         int nonblock)
  86 {
  87         struct vvp_page *vpg    = cl2vvp_page(slice);
  88         struct page     *vmpage = vpg->vpg_page;
  89
  90         LASSERT(vmpage != NULL);
  91         if (nonblock) {
  92                 if (!trylock_page(vmpage))
  93                         return -EAGAIN;
  94
  95                 if (unlikely(PageWriteback(vmpage))) {
  96                         unlock_page(vmpage);
  97                         return -EAGAIN;
  98                 }
  99
 100                 return 0;
 101         }
 102
 103         lock_page(vmpage);
 104         wait_on_page_writeback(vmpage);
 105
 106         return 0;
 107 }
 108
 109 static void vvp_page_assume(const struct lu_env *env,
 110                             const struct cl_page_slice *slice,
 111                             struct cl_io *unused)
 112 {
 113         struct page *vmpage = cl2vm_page(slice);
 114
 115         LASSERT(vmpage != NULL);
 116         LASSERT(PageLocked(vmpage));
 117         wait_on_page_writeback(vmpage);
 118 }
 119
 120 static void vvp_page_unassume(const struct lu_env *env,
 121                               const struct cl_page_slice *slice,
 122                               struct cl_io *unused)
 123 {
 124         struct page *vmpage = cl2vm_page(slice);
 125
 126         LASSERT(vmpage != NULL);
 127         LASSERT(PageLocked(vmpage));
 128 }
 129
 130 static void vvp_page_disown(const struct lu_env *env,
 131                             const struct cl_page_slice *slice, struct cl_io *io)
 132 {
 133         struct page *vmpage = cl2vm_page(slice);
 134
 135         LASSERT(vmpage != NULL);
 136         LASSERT(PageLocked(vmpage));
 137
 138         unlock_page(cl2vm_page(slice));
 139 }
 140
 141 static void vvp_page_discard(const struct lu_env *env,
 142                              const struct cl_page_slice *slice,
 143                              struct cl_io *unused)
 144 {
 145         struct page     *vmpage = cl2vm_page(slice);
 146         struct vvp_page *vpg    = cl2vvp_page(slice);
 147
 148         LASSERT(vmpage != NULL);
 149         LASSERT(PageLocked(vmpage));
 150
 151         if (vpg->vpg_defer_uptodate && !vpg->vpg_ra_used)
 152                 ll_ra_stats_inc(vmpage->mapping->host, RA_STAT_DISCARDED);
 153
 154         ll_invalidate_page(vmpage);
 155 }
 156
 157 static void vvp_page_delete(const struct lu_env *env,
 158                             const struct cl_page_slice *slice)
 159 {
 160         struct page      *vmpage = cl2vm_page(slice);
 161         struct inode     *inode  = vmpage->mapping->host;
 162         struct cl_object *obj    = slice->cpl_obj;
 163         struct cl_page   *page   = slice->cpl_page;
 164         int refc;
 165
 166         LASSERT(PageLocked(vmpage));
 167         LASSERT((struct cl_page *)vmpage->private == page);
 168         LASSERT(inode == vvp_object_inode(obj));
 169
 170         vvp_write_complete(cl2vvp(obj), cl2vvp_page(slice));
 171
 172         /* Drop the reference count held in vvp_page_init */
 173         refc = atomic_dec_return(&page->cp_ref);
 174         LASSERTF(refc >= 1, "page = %p, refc = %d\n", page, refc);
 175
 176         ClearPageUptodate(vmpage);
 177         ClearPagePrivate(vmpage);
 178         vmpage->private = 0;
 179         /*
 180          * Reference from vmpage to cl_page is removed, but the reference back
 181          * is still here. It is removed later in vvp_page_fini().
 182          */
 183 }
 184
 185 static void vvp_page_export(const struct lu_env *env,
 186                             const struct cl_page_slice *slice,
 187                             int uptodate)
 188 {
 189         struct page *vmpage = cl2vm_page(slice);
 190
 191         LASSERT(vmpage != NULL);
 192         LASSERT(PageLocked(vmpage));
 193         if (uptodate)
 194                 SetPageUptodate(vmpage);
 195         else
 196                 ClearPageUptodate(vmpage);
 197 }
 198
 199 static int vvp_page_is_vmlocked(const struct lu_env *env,
 200                                 const struct cl_page_slice *slice)
 201 {
 202         return PageLocked(cl2vm_page(slice)) ? -EBUSY : -ENODATA;
 203 }
 204
 205 static int vvp_page_prep_read(const struct lu_env *env,
 206                               const struct cl_page_slice *slice,
 207                               struct cl_io *unused)
 208 {
 209         ENTRY;
 210         /* Skip the page already marked as PG_uptodate. */
 211         RETURN(PageUptodate(cl2vm_page(slice)) ? -EALREADY : 0);
 212 }
 213
 214 static int vvp_page_prep_write(const struct lu_env *env,
 215                                const struct cl_page_slice *slice,
 216                                struct cl_io *unused)
 217 {
 218         struct page *vmpage = cl2vm_page(slice);
 219
 220         LASSERT(PageLocked(vmpage));
 221         LASSERT(!PageDirty(vmpage));
 222
 223         set_page_writeback(vmpage);
 224         vvp_write_pending(cl2vvp(slice->cpl_obj), cl2vvp_page(slice));
 225
 226         return 0;
 227 }
 228
 229 /**
 230  * Handles page transfer errors at VM level.
 231  *
 232  * This takes inode as a separate argument, because inode on which error is to
 233  * be set can be different from \a vmpage inode in case of direct-io.
 234  */
 235 static void vvp_vmpage_error(struct inode *inode, struct page *vmpage, int ioret)
 236 {
 237         struct vvp_object *obj = cl_inode2vvp(inode);
 238
 239         if (ioret == 0) {
 240                 ClearPageError(vmpage);
 241                 obj->vob_discard_page_warned = 0;
 242         } else {
 243                 SetPageError(vmpage);
 244                 if (ioret == -ENOSPC)
 245                         set_bit(AS_ENOSPC, &inode->i_mapping->flags);
 246                 else
 247                         set_bit(AS_EIO, &inode->i_mapping->flags);
 248
 249                 if ((ioret == -ESHUTDOWN || ioret == -EINTR) &&
 250                      obj->vob_discard_page_warned == 0) {
 251                         obj->vob_discard_page_warned = 1;
 252                         ll_dirty_page_discard_warn(vmpage, ioret);
 253                 }
 254         }
 255 }
 256
 257 static void vvp_page_completion_read(const struct lu_env *env,
 258                                      const struct cl_page_slice *slice,
 259                                      int ioret)
 260 {
 261         struct vvp_page *vpg    = cl2vvp_page(slice);
 262         struct page     *vmpage = vpg->vpg_page;
 263         struct cl_page  *page   = slice->cpl_page;
 264         struct inode    *inode  = vvp_object_inode(page->cp_obj);
 265         ENTRY;
 266
 267         LASSERT(PageLocked(vmpage));
 268         CL_PAGE_HEADER(D_PAGE, env, page, "completing READ with %d\n", ioret);
 269
 270         if (vpg->vpg_defer_uptodate)
 271                 ll_ra_count_put(ll_i2sbi(inode), 1);
 272
 273         if (ioret == 0)  {
 274                 if (!vpg->vpg_defer_uptodate)
 275                         cl_page_export(env, page, 1);
 276         } else {
 277                 vpg->vpg_defer_uptodate = 0;
 278         }
 279
 280         if (page->cp_sync_io == NULL)
 281                 unlock_page(vmpage);
 282
 283         EXIT;
 284 }
 285
 286 static void vvp_page_completion_write(const struct lu_env *env,
 287                                       const struct cl_page_slice *slice,
 288                                       int ioret)
 289 {
 290         struct vvp_page *vpg    = cl2vvp_page(slice);
 291         struct cl_page  *pg     = slice->cpl_page;
 292         struct page     *vmpage = vpg->vpg_page;
 293         ENTRY;
 294
 295         LASSERT(ergo(pg->cp_sync_io != NULL, PageLocked(vmpage)));
 296         LASSERT(PageWriteback(vmpage));
 297
 298         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 299
 300         /*
 301          * TODO: Actually it makes sense to add the page into oap pending
 302          * list again and so that we don't need to take the page out from
 303          * SoM write pending list, if we just meet a recoverable error,
 304          * -ENOMEM, etc.
 305          * To implement this, we just need to return a non zero value in
 306          * ->cpo_completion method. The underlying transfer should be notified
 307          * and then re-add the page into pending transfer queue.  -jay
 308          */
 309
 310         vpg->vpg_write_queued = 0;
 311         vvp_write_complete(cl2vvp(slice->cpl_obj), vpg);
 312
 313         /*
 314          * Only mark the page error only when it's an async write because
 315          * applications won't wait for IO to finish.
 316          */
 317         if (pg->cp_sync_io == NULL)
 318                 vvp_vmpage_error(vvp_object_inode(pg->cp_obj), vmpage, ioret);
 319
 320         end_page_writeback(vmpage);
 321         EXIT;
 322 }
 323
 324 /**
 325  * Implements cl_page_operations::cpo_make_ready() method.
 326  *
 327  * This is called to yank a page from the transfer cache and to send it out as
 328  * a part of transfer. This function try-locks the page. If try-lock failed,
 329  * page is owned by some concurrent IO, and should be skipped (this is bad,
 330  * but hopefully rare situation, as it usually results in transfer being
 331  * shorter than possible).
 332  *
 333  * \retval 0      success, page can be placed into transfer
 334  *
 335  * \retval -EAGAIN page is either used by concurrent IO has been
 336  * truncated. Skip it.
 337  */
 338 static int vvp_page_make_ready(const struct lu_env *env,
 339                                const struct cl_page_slice *slice)
 340 {
 341         struct page *vmpage = cl2vm_page(slice);
 342         struct cl_page *pg = slice->cpl_page;
 343         int result = 0;
 344
 345         lock_page(vmpage);
 346         if (clear_page_dirty_for_io(vmpage)) {
 347                 LASSERT(pg->cp_state == CPS_CACHED);
 348                 /* This actually clears the dirty bit in the radix
 349                  * tree. */
 350                 set_page_writeback(vmpage);
 351                 vvp_write_pending(cl2vvp(slice->cpl_obj),
 352                                 cl2vvp_page(slice));
 353                 CL_PAGE_HEADER(D_PAGE, env, pg, "readied\n");
 354         } else if (pg->cp_state == CPS_PAGEOUT) {
 355                 /* is it possible for osc_flush_async_page() to already
 356                  * make it ready? */
 357                 result = -EALREADY;
 358         } else {
 359                 CL_PAGE_DEBUG(D_ERROR, env, pg, "Unexpecting page state %d.\n",
 360                               pg->cp_state);
 361                 LBUG();
 362         }
 363         unlock_page(vmpage);
 364         RETURN(result);
 365 }
 366
 367 static int vvp_page_is_under_lock(const struct lu_env *env,
 368                                   const struct cl_page_slice *slice,
 369                                   struct cl_io *io, pgoff_t *max_index)
 370 {
 371         ENTRY;
 372
 373         if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE ||
 374             io->ci_type == CIT_FAULT) {
 375                 struct ccc_io *cio = ccc_env_io(env);
 376
 377                 if (unlikely(cio->cui_fd->fd_flags & LL_FILE_GROUP_LOCKED))
 378                         *max_index = CL_PAGE_EOF;
 379         }
 380         RETURN(0);
 381 }
 382
 383
 384 static int vvp_page_print(const struct lu_env *env,
 385                           const struct cl_page_slice *slice,
 386                           void *cookie, lu_printer_t printer)
 387 {
 388         struct vvp_page *vpg    = cl2vvp_page(slice);
 389         struct page     *vmpage = vpg->vpg_page;
 390
 391         (*printer)(env, cookie, LUSTRE_VVP_NAME"-page@%p(%d:%d:%d) "
 392                    "vm@%p ",
 393                    vpg, vpg->vpg_defer_uptodate, vpg->vpg_ra_used,
 394                    vpg->vpg_write_queued, vmpage);
 395
 396         if (vmpage != NULL) {
 397                 (*printer)(env, cookie, "%lx %d:%d %lx %lu %slru",
 398                            (long)vmpage->flags, page_count(vmpage),
 399                            page_mapcount(vmpage), vmpage->private,
 400                            page_index(vmpage),
 401                            list_empty(&vmpage->lru) ? "not-" : "");
 402         }
 403
 404         (*printer)(env, cookie, "\n");
 405
 406         return 0;
 407 }
 408
 409 static int vvp_page_fail(const struct lu_env *env,
 410                          const struct cl_page_slice *slice)
 411 {
 412         /*
 413          * Cached read?
 414          */
 415         LBUG();
 416
 417         return 0;
 418 }
 419
 420 static const struct cl_page_operations vvp_page_ops = {
 421         .cpo_own           = vvp_page_own,
 422         .cpo_assume        = vvp_page_assume,
 423         .cpo_unassume      = vvp_page_unassume,
 424         .cpo_disown        = vvp_page_disown,
 425         .cpo_discard       = vvp_page_discard,
 426         .cpo_delete        = vvp_page_delete,
 427         .cpo_export        = vvp_page_export,
 428         .cpo_is_vmlocked   = vvp_page_is_vmlocked,
 429         .cpo_fini          = vvp_page_fini,
 430         .cpo_print         = vvp_page_print,
 431         .cpo_is_under_lock = vvp_page_is_under_lock,
 432         .io = {
 433                 [CRT_READ] = {
 434                         .cpo_prep       = vvp_page_prep_read,
 435                         .cpo_completion = vvp_page_completion_read,
 436                         .cpo_make_ready = vvp_page_fail,
 437                 },
 438                 [CRT_WRITE] = {
 439                         .cpo_prep       = vvp_page_prep_write,
 440                         .cpo_completion = vvp_page_completion_write,
 441                         .cpo_make_ready = vvp_page_make_ready,
 442                 },
 443         },
 444 };
 445
 446 static int vvp_transient_page_prep(const struct lu_env *env,
 447                                    const struct cl_page_slice *slice,
 448                                    struct cl_io *unused)
 449 {
 450         ENTRY;
 451         /* transient page should always be sent. */
 452         RETURN(0);
 453 }
 454
 455 static void vvp_transient_page_verify(const struct cl_page *page)
 456 {
 457 }
 458
 459 static int vvp_transient_page_own(const struct lu_env *env,
 460                                   const struct cl_page_slice *slice,
 461                                   struct cl_io *unused, int nonblock)
 462 {
 463         vvp_transient_page_verify(slice->cpl_page);
 464         return 0;
 465 }
 466
 467 static void vvp_transient_page_assume(const struct lu_env *env,
 468                                       const struct cl_page_slice *slice,
 469                                       struct cl_io *unused)
 470 {
 471         vvp_transient_page_verify(slice->cpl_page);
 472 }
 473
 474 static void vvp_transient_page_unassume(const struct lu_env *env,
 475                                         const struct cl_page_slice *slice,
 476                                         struct cl_io *unused)
 477 {
 478         vvp_transient_page_verify(slice->cpl_page);
 479 }
 480
 481 static void vvp_transient_page_disown(const struct lu_env *env,
 482                                       const struct cl_page_slice *slice,
 483                                       struct cl_io *unused)
 484 {
 485         vvp_transient_page_verify(slice->cpl_page);
 486 }
 487
 488 static void vvp_transient_page_discard(const struct lu_env *env,
 489                                        const struct cl_page_slice *slice,
 490                                        struct cl_io *unused)
 491 {
 492         struct cl_page *page = slice->cpl_page;
 493
 494         vvp_transient_page_verify(slice->cpl_page);
 495
 496         /*
 497          * For transient pages, remove it from the radix tree.
 498          */
 499         cl_page_delete(env, page);
 500 }
 501
 502 static int vvp_transient_page_is_vmlocked(const struct lu_env *env,
 503                                           const struct cl_page_slice *slice)
 504 {
 505         struct inode    *inode = vvp_object_inode(slice->cpl_obj);
 506         int     locked;
 507
 508         locked = !mutex_trylock(&inode->i_mutex);
 509         if (!locked)
 510                 mutex_unlock(&inode->i_mutex);
 511         return locked ? -EBUSY : -ENODATA;
 512 }
 513
 514 static void
 515 vvp_transient_page_completion(const struct lu_env *env,
 516                               const struct cl_page_slice *slice,
 517                               int ioret)
 518 {
 519         vvp_transient_page_verify(slice->cpl_page);
 520 }
 521
 522 static void vvp_transient_page_fini(const struct lu_env *env,
 523                                     struct cl_page_slice *slice)
 524 {
 525         struct vvp_page *vpg = cl2vvp_page(slice);
 526         struct cl_page *clp = slice->cpl_page;
 527         struct vvp_object *clobj = cl2vvp(clp->cp_obj);
 528
 529         vvp_page_fini_common(vpg);
 530         atomic_dec(&clobj->vob_transient_pages);
 531 }
 532
 533 static const struct cl_page_operations vvp_transient_page_ops = {
 534         .cpo_own                = vvp_transient_page_own,
 535         .cpo_assume             = vvp_transient_page_assume,
 536         .cpo_unassume           = vvp_transient_page_unassume,
 537         .cpo_disown             = vvp_transient_page_disown,
 538         .cpo_discard            = vvp_transient_page_discard,
 539         .cpo_fini               = vvp_transient_page_fini,
 540         .cpo_is_vmlocked        = vvp_transient_page_is_vmlocked,
 541         .cpo_print              = vvp_page_print,
 542         .cpo_is_under_lock      = vvp_page_is_under_lock,
 543         .io = {
 544                 [CRT_READ] = {
 545                         .cpo_prep       = vvp_transient_page_prep,
 546                         .cpo_completion = vvp_transient_page_completion,
 547                 },
 548                 [CRT_WRITE] = {
 549                         .cpo_prep       = vvp_transient_page_prep,
 550                         .cpo_completion = vvp_transient_page_completion,
 551                 }
 552         }
 553 };
 554
 555 int vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 556                 struct cl_page *page, pgoff_t index)
 557 {
 558         struct vvp_page *vpg = cl_object_page_slice(obj, page);
 559         struct page     *vmpage = page->cp_vmpage;
 560
 561         CLOBINVRNT(env, obj, vvp_object_invariant(obj));
 562
 563         vpg->vpg_page = vmpage;
 564         page_cache_get(vmpage);
 565
 566         INIT_LIST_HEAD(&vpg->vpg_pending_linkage);
 567         if (page->cp_type == CPT_CACHEABLE) {
 568                 /* in cache, decref in vvp_page_delete */
 569                 atomic_inc(&page->cp_ref);
 570                 SetPagePrivate(vmpage);
 571                 vmpage->private = (unsigned long)page;
 572                 cl_page_slice_add(page, &vpg->vpg_cl, obj, index,
 573                                 &vvp_page_ops);
 574         } else {
 575                 struct vvp_object *clobj = cl2vvp(obj);
 576
 577                 cl_page_slice_add(page, &vpg->vpg_cl, obj, index,
 578                                 &vvp_transient_page_ops);
 579                 atomic_inc(&clobj->vob_transient_pages);
 580         }
 581         return 0;
 582 }