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         /* Drop the reference count held in vvp_page_init */
 171         refc = atomic_dec_return(&page->cp_ref);
 172         LASSERTF(refc >= 1, "page = %p, refc = %d\n", page, refc);
 173
 174         ClearPageUptodate(vmpage);
 175         ClearPagePrivate(vmpage);
 176         vmpage->private = 0;
 177         /*
 178          * Reference from vmpage to cl_page is removed, but the reference back
 179          * is still here. It is removed later in vvp_page_fini().
 180          */
 181 }
 182
 183 static void vvp_page_export(const struct lu_env *env,
 184                             const struct cl_page_slice *slice,
 185                             int uptodate)
 186 {
 187         struct page *vmpage = cl2vm_page(slice);
 188
 189         LASSERT(vmpage != NULL);
 190         LASSERT(PageLocked(vmpage));
 191         if (uptodate)
 192                 SetPageUptodate(vmpage);
 193         else
 194                 ClearPageUptodate(vmpage);
 195 }
 196
 197 static int vvp_page_is_vmlocked(const struct lu_env *env,
 198                                 const struct cl_page_slice *slice)
 199 {
 200         return PageLocked(cl2vm_page(slice)) ? -EBUSY : -ENODATA;
 201 }
 202
 203 static int vvp_page_prep_read(const struct lu_env *env,
 204                               const struct cl_page_slice *slice,
 205                               struct cl_io *unused)
 206 {
 207         ENTRY;
 208         /* Skip the page already marked as PG_uptodate. */
 209         RETURN(PageUptodate(cl2vm_page(slice)) ? -EALREADY : 0);
 210 }
 211
 212 static int vvp_page_prep_write(const struct lu_env *env,
 213                                const struct cl_page_slice *slice,
 214                                struct cl_io *unused)
 215 {
 216         struct page *vmpage = cl2vm_page(slice);
 217
 218         LASSERT(PageLocked(vmpage));
 219         LASSERT(!PageDirty(vmpage));
 220
 221         set_page_writeback(vmpage);
 222
 223         return 0;
 224 }
 225
 226 /**
 227  * Handles page transfer errors at VM level.
 228  *
 229  * This takes inode as a separate argument, because inode on which error is to
 230  * be set can be different from \a vmpage inode in case of direct-io.
 231  */
 232 static void vvp_vmpage_error(struct inode *inode, struct page *vmpage, int ioret)
 233 {
 234         struct vvp_object *obj = cl_inode2vvp(inode);
 235
 236         if (ioret == 0) {
 237                 ClearPageError(vmpage);
 238                 obj->vob_discard_page_warned = 0;
 239         } else {
 240                 SetPageError(vmpage);
 241                 if (ioret == -ENOSPC)
 242                         set_bit(AS_ENOSPC, &inode->i_mapping->flags);
 243                 else
 244                         set_bit(AS_EIO, &inode->i_mapping->flags);
 245
 246                 if ((ioret == -ESHUTDOWN || ioret == -EINTR) &&
 247                      obj->vob_discard_page_warned == 0) {
 248                         obj->vob_discard_page_warned = 1;
 249                         ll_dirty_page_discard_warn(vmpage, ioret);
 250                 }
 251         }
 252 }
 253
 254 static void vvp_page_completion_read(const struct lu_env *env,
 255                                      const struct cl_page_slice *slice,
 256                                      int ioret)
 257 {
 258         struct vvp_page *vpg    = cl2vvp_page(slice);
 259         struct page     *vmpage = vpg->vpg_page;
 260         struct cl_page  *page   = slice->cpl_page;
 261         struct inode    *inode  = vvp_object_inode(page->cp_obj);
 262         ENTRY;
 263
 264         LASSERT(PageLocked(vmpage));
 265         CL_PAGE_HEADER(D_PAGE, env, page, "completing READ with %d\n", ioret);
 266
 267         if (vpg->vpg_defer_uptodate)
 268                 ll_ra_count_put(ll_i2sbi(inode), 1);
 269
 270         if (ioret == 0)  {
 271                 if (!vpg->vpg_defer_uptodate)
 272                         cl_page_export(env, page, 1);
 273         } else {
 274                 vpg->vpg_defer_uptodate = 0;
 275         }
 276
 277         if (page->cp_sync_io == NULL)
 278                 unlock_page(vmpage);
 279
 280         EXIT;
 281 }
 282
 283 static void vvp_page_completion_write(const struct lu_env *env,
 284                                       const struct cl_page_slice *slice,
 285                                       int ioret)
 286 {
 287         struct vvp_page *vpg    = cl2vvp_page(slice);
 288         struct cl_page  *pg     = slice->cpl_page;
 289         struct page     *vmpage = vpg->vpg_page;
 290         ENTRY;
 291
 292         LASSERT(ergo(pg->cp_sync_io != NULL, PageLocked(vmpage)));
 293         LASSERT(PageWriteback(vmpage));
 294
 295         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 296
 297         /*
 298          * Only mark the page error only when it's an async write because
 299          * applications won't wait for IO to finish.
 300          */
 301         if (pg->cp_sync_io == NULL)
 302                 vvp_vmpage_error(vvp_object_inode(pg->cp_obj), vmpage, ioret);
 303
 304         end_page_writeback(vmpage);
 305         EXIT;
 306 }
 307
 308 /**
 309  * Implements cl_page_operations::cpo_make_ready() method.
 310  *
 311  * This is called to yank a page from the transfer cache and to send it out as
 312  * a part of transfer. This function try-locks the page. If try-lock failed,
 313  * page is owned by some concurrent IO, and should be skipped (this is bad,
 314  * but hopefully rare situation, as it usually results in transfer being
 315  * shorter than possible).
 316  *
 317  * \retval 0      success, page can be placed into transfer
 318  *
 319  * \retval -EAGAIN page is either used by concurrent IO has been
 320  * truncated. Skip it.
 321  */
 322 static int vvp_page_make_ready(const struct lu_env *env,
 323                                const struct cl_page_slice *slice)
 324 {
 325         struct page *vmpage = cl2vm_page(slice);
 326         struct cl_page *pg = slice->cpl_page;
 327         int result = 0;
 328
 329         lock_page(vmpage);
 330         if (clear_page_dirty_for_io(vmpage)) {
 331                 LASSERT(pg->cp_state == CPS_CACHED);
 332                 /* This actually clears the dirty bit in the radix
 333                  * tree. */
 334                 set_page_writeback(vmpage);
 335                 CL_PAGE_HEADER(D_PAGE, env, pg, "readied\n");
 336         } else if (pg->cp_state == CPS_PAGEOUT) {
 337                 /* is it possible for osc_flush_async_page() to already
 338                  * make it ready? */
 339                 result = -EALREADY;
 340         } else {
 341                 CL_PAGE_DEBUG(D_ERROR, env, pg, "Unexpecting page state %d.\n",
 342                               pg->cp_state);
 343                 LBUG();
 344         }
 345         unlock_page(vmpage);
 346         RETURN(result);
 347 }
 348
 349 static int vvp_page_is_under_lock(const struct lu_env *env,
 350                                   const struct cl_page_slice *slice,
 351                                   struct cl_io *io, pgoff_t *max_index)
 352 {
 353         ENTRY;
 354
 355         if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE ||
 356             io->ci_type == CIT_FAULT) {
 357                 struct vvp_io *vio = vvp_env_io(env);
 358
 359                 if (unlikely(vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED))
 360                         *max_index = CL_PAGE_EOF;
 361         }
 362         RETURN(0);
 363 }
 364
 365
 366 static int vvp_page_print(const struct lu_env *env,
 367                           const struct cl_page_slice *slice,
 368                           void *cookie, lu_printer_t printer)
 369 {
 370         struct vvp_page *vpg    = cl2vvp_page(slice);
 371         struct page     *vmpage = vpg->vpg_page;
 372
 373         (*printer)(env, cookie, LUSTRE_VVP_NAME"-page@%p(%d:%d) "
 374                    "vm@%p ",
 375                    vpg, vpg->vpg_defer_uptodate, vpg->vpg_ra_used, vmpage);
 376
 377         if (vmpage != NULL) {
 378                 (*printer)(env, cookie, "%lx %d:%d %lx %lu %slru",
 379                            (long)vmpage->flags, page_count(vmpage),
 380                            page_mapcount(vmpage), vmpage->private,
 381                            page_index(vmpage),
 382                            list_empty(&vmpage->lru) ? "not-" : "");
 383         }
 384
 385         (*printer)(env, cookie, "\n");
 386
 387         return 0;
 388 }
 389
 390 static int vvp_page_fail(const struct lu_env *env,
 391                          const struct cl_page_slice *slice)
 392 {
 393         /*
 394          * Cached read?
 395          */
 396         LBUG();
 397
 398         return 0;
 399 }
 400
 401 static const struct cl_page_operations vvp_page_ops = {
 402         .cpo_own           = vvp_page_own,
 403         .cpo_assume        = vvp_page_assume,
 404         .cpo_unassume      = vvp_page_unassume,
 405         .cpo_disown        = vvp_page_disown,
 406         .cpo_discard       = vvp_page_discard,
 407         .cpo_delete        = vvp_page_delete,
 408         .cpo_export        = vvp_page_export,
 409         .cpo_is_vmlocked   = vvp_page_is_vmlocked,
 410         .cpo_fini          = vvp_page_fini,
 411         .cpo_print         = vvp_page_print,
 412         .cpo_is_under_lock = vvp_page_is_under_lock,
 413         .io = {
 414                 [CRT_READ] = {
 415                         .cpo_prep       = vvp_page_prep_read,
 416                         .cpo_completion = vvp_page_completion_read,
 417                         .cpo_make_ready = vvp_page_fail,
 418                 },
 419                 [CRT_WRITE] = {
 420                         .cpo_prep       = vvp_page_prep_write,
 421                         .cpo_completion = vvp_page_completion_write,
 422                         .cpo_make_ready = vvp_page_make_ready,
 423                 },
 424         },
 425 };
 426
 427 static int vvp_transient_page_prep(const struct lu_env *env,
 428                                    const struct cl_page_slice *slice,
 429                                    struct cl_io *unused)
 430 {
 431         ENTRY;
 432         /* transient page should always be sent. */
 433         RETURN(0);
 434 }
 435
 436 static void vvp_transient_page_verify(const struct cl_page *page)
 437 {
 438 }
 439
 440 static int vvp_transient_page_own(const struct lu_env *env,
 441                                   const struct cl_page_slice *slice,
 442                                   struct cl_io *unused, int nonblock)
 443 {
 444         vvp_transient_page_verify(slice->cpl_page);
 445         return 0;
 446 }
 447
 448 static void vvp_transient_page_assume(const struct lu_env *env,
 449                                       const struct cl_page_slice *slice,
 450                                       struct cl_io *unused)
 451 {
 452         vvp_transient_page_verify(slice->cpl_page);
 453 }
 454
 455 static void vvp_transient_page_unassume(const struct lu_env *env,
 456                                         const struct cl_page_slice *slice,
 457                                         struct cl_io *unused)
 458 {
 459         vvp_transient_page_verify(slice->cpl_page);
 460 }
 461
 462 static void vvp_transient_page_disown(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_discard(const struct lu_env *env,
 470                                        const struct cl_page_slice *slice,
 471                                        struct cl_io *unused)
 472 {
 473         struct cl_page *page = slice->cpl_page;
 474
 475         vvp_transient_page_verify(slice->cpl_page);
 476
 477         /*
 478          * For transient pages, remove it from the radix tree.
 479          */
 480         cl_page_delete(env, page);
 481 }
 482
 483 static int vvp_transient_page_is_vmlocked(const struct lu_env *env,
 484                                           const struct cl_page_slice *slice)
 485 {
 486         struct inode    *inode = vvp_object_inode(slice->cpl_obj);
 487         int     locked;
 488
 489         locked = !mutex_trylock(&inode->i_mutex);
 490         if (!locked)
 491                 mutex_unlock(&inode->i_mutex);
 492         return locked ? -EBUSY : -ENODATA;
 493 }
 494
 495 static void
 496 vvp_transient_page_completion(const struct lu_env *env,
 497                               const struct cl_page_slice *slice,
 498                               int ioret)
 499 {
 500         vvp_transient_page_verify(slice->cpl_page);
 501 }
 502
 503 static void vvp_transient_page_fini(const struct lu_env *env,
 504                                     struct cl_page_slice *slice)
 505 {
 506         struct vvp_page *vpg = cl2vvp_page(slice);
 507         struct cl_page *clp = slice->cpl_page;
 508         struct vvp_object *clobj = cl2vvp(clp->cp_obj);
 509
 510         vvp_page_fini_common(vpg);
 511         atomic_dec(&clobj->vob_transient_pages);
 512 }
 513
 514 static const struct cl_page_operations vvp_transient_page_ops = {
 515         .cpo_own                = vvp_transient_page_own,
 516         .cpo_assume             = vvp_transient_page_assume,
 517         .cpo_unassume           = vvp_transient_page_unassume,
 518         .cpo_disown             = vvp_transient_page_disown,
 519         .cpo_discard            = vvp_transient_page_discard,
 520         .cpo_fini               = vvp_transient_page_fini,
 521         .cpo_is_vmlocked        = vvp_transient_page_is_vmlocked,
 522         .cpo_print              = vvp_page_print,
 523         .cpo_is_under_lock      = vvp_page_is_under_lock,
 524         .io = {
 525                 [CRT_READ] = {
 526                         .cpo_prep       = vvp_transient_page_prep,
 527                         .cpo_completion = vvp_transient_page_completion,
 528                 },
 529                 [CRT_WRITE] = {
 530                         .cpo_prep       = vvp_transient_page_prep,
 531                         .cpo_completion = vvp_transient_page_completion,
 532                 }
 533         }
 534 };
 535
 536 int vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 537                 struct cl_page *page, pgoff_t index)
 538 {
 539         struct vvp_page *vpg = cl_object_page_slice(obj, page);
 540         struct page     *vmpage = page->cp_vmpage;
 541
 542         CLOBINVRNT(env, obj, vvp_object_invariant(obj));
 543
 544         vpg->vpg_page = vmpage;
 545         page_cache_get(vmpage);
 546
 547         if (page->cp_type == CPT_CACHEABLE) {
 548                 /* in cache, decref in vvp_page_delete */
 549                 atomic_inc(&page->cp_ref);
 550                 SetPagePrivate(vmpage);
 551                 vmpage->private = (unsigned long)page;
 552                 cl_page_slice_add(page, &vpg->vpg_cl, obj, index,
 553                                 &vvp_page_ops);
 554         } else {
 555                 struct vvp_object *clobj = cl2vvp(obj);
 556
 557                 cl_page_slice_add(page, &vpg->vpg_cl, obj, index,
 558                                 &vvp_transient_page_ops);
 559                 atomic_inc(&clobj->vob_transient_pages);
 560         }
 561         return 0;
 562 }