kernel/drivers/staging/lustre/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, 2012, 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 "../include/obd.h"
  45 #include "../include/lustre_lite.h"
  46
  47 #include "vvp_internal.h"
  48
  49 /*****************************************************************************
  50  *
  51  * Page operations.
  52  *
  53  */
  54
  55 static void vvp_page_fini_common(struct ccc_page *cp)
  56 {
  57         struct page *vmpage = cp->cpg_page;
  58
  59         LASSERT(vmpage != NULL);
  60         page_cache_release(vmpage);
  61 }
  62
  63 static void vvp_page_fini(const struct lu_env *env,
  64                           struct cl_page_slice *slice)
  65 {
  66         struct ccc_page *cp = cl2ccc_page(slice);
  67         struct page *vmpage  = cp->cpg_page;
  68
  69         /*
  70          * vmpage->private was already cleared when page was moved into
  71          * VPG_FREEING state.
  72          */
  73         LASSERT((struct cl_page *)vmpage->private != slice->cpl_page);
  74         vvp_page_fini_common(cp);
  75 }
  76
  77 static int vvp_page_own(const struct lu_env *env,
  78                         const struct cl_page_slice *slice, struct cl_io *io,
  79                         int nonblock)
  80 {
  81         struct ccc_page *vpg    = cl2ccc_page(slice);
  82         struct page      *vmpage = vpg->cpg_page;
  83
  84         LASSERT(vmpage != NULL);
  85         if (nonblock) {
  86                 if (!trylock_page(vmpage))
  87                         return -EAGAIN;
  88
  89                 if (unlikely(PageWriteback(vmpage))) {
  90                         unlock_page(vmpage);
  91                         return -EAGAIN;
  92                 }
  93
  94                 return 0;
  95         }
  96
  97         lock_page(vmpage);
  98         wait_on_page_writeback(vmpage);
  99         return 0;
 100 }
 101
 102 static void vvp_page_assume(const struct lu_env *env,
 103                             const struct cl_page_slice *slice,
 104                             struct cl_io *unused)
 105 {
 106         struct page *vmpage = cl2vm_page(slice);
 107
 108         LASSERT(vmpage != NULL);
 109         LASSERT(PageLocked(vmpage));
 110         wait_on_page_writeback(vmpage);
 111 }
 112
 113 static void vvp_page_unassume(const struct lu_env *env,
 114                               const struct cl_page_slice *slice,
 115                               struct cl_io *unused)
 116 {
 117         struct page *vmpage = cl2vm_page(slice);
 118
 119         LASSERT(vmpage != NULL);
 120         LASSERT(PageLocked(vmpage));
 121 }
 122
 123 static void vvp_page_disown(const struct lu_env *env,
 124                             const struct cl_page_slice *slice, struct cl_io *io)
 125 {
 126         struct page *vmpage = cl2vm_page(slice);
 127
 128         LASSERT(vmpage != NULL);
 129         LASSERT(PageLocked(vmpage));
 130
 131         unlock_page(cl2vm_page(slice));
 132 }
 133
 134 static void vvp_page_discard(const struct lu_env *env,
 135                              const struct cl_page_slice *slice,
 136                              struct cl_io *unused)
 137 {
 138         struct page        *vmpage  = cl2vm_page(slice);
 139         struct address_space *mapping;
 140         struct ccc_page      *cpg     = cl2ccc_page(slice);
 141
 142         LASSERT(vmpage != NULL);
 143         LASSERT(PageLocked(vmpage));
 144
 145         mapping = vmpage->mapping;
 146
 147         if (cpg->cpg_defer_uptodate && !cpg->cpg_ra_used)
 148                 ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
 149
 150         /*
 151          * truncate_complete_page() calls
 152          * a_ops->invalidatepage()->cl_page_delete()->vvp_page_delete().
 153          */
 154         truncate_complete_page(mapping, vmpage);
 155 }
 156
 157 static int vvp_page_unmap(const struct lu_env *env,
 158                           const struct cl_page_slice *slice,
 159                           struct cl_io *unused)
 160 {
 161         struct page *vmpage = cl2vm_page(slice);
 162         __u64       offset;
 163
 164         LASSERT(vmpage != NULL);
 165         LASSERT(PageLocked(vmpage));
 166
 167         offset = vmpage->index << PAGE_CACHE_SHIFT;
 168
 169         /*
 170          * XXX is it safe to call this with the page lock held?
 171          */
 172         ll_teardown_mmaps(vmpage->mapping, offset, offset + PAGE_CACHE_SIZE);
 173         return 0;
 174 }
 175
 176 static void vvp_page_delete(const struct lu_env *env,
 177                             const struct cl_page_slice *slice)
 178 {
 179         struct page       *vmpage = cl2vm_page(slice);
 180         struct inode     *inode  = vmpage->mapping->host;
 181         struct cl_object *obj    = slice->cpl_obj;
 182
 183         LASSERT(PageLocked(vmpage));
 184         LASSERT((struct cl_page *)vmpage->private == slice->cpl_page);
 185         LASSERT(inode == ccc_object_inode(obj));
 186
 187         vvp_write_complete(cl2ccc(obj), cl2ccc_page(slice));
 188         ClearPagePrivate(vmpage);
 189         vmpage->private = 0;
 190         /*
 191          * Reference from vmpage to cl_page is removed, but the reference back
 192          * is still here. It is removed later in vvp_page_fini().
 193          */
 194 }
 195
 196 static void vvp_page_export(const struct lu_env *env,
 197                             const struct cl_page_slice *slice,
 198                             int uptodate)
 199 {
 200         struct page *vmpage = cl2vm_page(slice);
 201
 202         LASSERT(vmpage != NULL);
 203         LASSERT(PageLocked(vmpage));
 204         if (uptodate)
 205                 SetPageUptodate(vmpage);
 206         else
 207                 ClearPageUptodate(vmpage);
 208 }
 209
 210 static int vvp_page_is_vmlocked(const struct lu_env *env,
 211                                 const struct cl_page_slice *slice)
 212 {
 213         return PageLocked(cl2vm_page(slice)) ? -EBUSY : -ENODATA;
 214 }
 215
 216 static int vvp_page_prep_read(const struct lu_env *env,
 217                               const struct cl_page_slice *slice,
 218                               struct cl_io *unused)
 219 {
 220         /* Skip the page already marked as PG_uptodate. */
 221         return PageUptodate(cl2vm_page(slice)) ? -EALREADY : 0;
 222 }
 223
 224 static int vvp_page_prep_write(const struct lu_env *env,
 225                                const struct cl_page_slice *slice,
 226                                struct cl_io *unused)
 227 {
 228         struct page *vmpage = cl2vm_page(slice);
 229         struct cl_page *pg = slice->cpl_page;
 230
 231         LASSERT(PageLocked(vmpage));
 232         LASSERT(!PageDirty(vmpage));
 233
 234         /* ll_writepage path is not a sync write, so need to set page writeback
 235          * flag */
 236         if (!pg->cp_sync_io)
 237                 set_page_writeback(vmpage);
 238
 239         vvp_write_pending(cl2ccc(slice->cpl_obj), cl2ccc_page(slice));
 240
 241         return 0;
 242 }
 243
 244 /**
 245  * Handles page transfer errors at VM level.
 246  *
 247  * This takes inode as a separate argument, because inode on which error is to
 248  * be set can be different from \a vmpage inode in case of direct-io.
 249  */
 250 static void vvp_vmpage_error(struct inode *inode, struct page *vmpage, int ioret)
 251 {
 252         struct ccc_object *obj = cl_inode2ccc(inode);
 253
 254         if (ioret == 0) {
 255                 ClearPageError(vmpage);
 256                 obj->cob_discard_page_warned = 0;
 257         } else {
 258                 SetPageError(vmpage);
 259                 if (ioret == -ENOSPC)
 260                         set_bit(AS_ENOSPC, &inode->i_mapping->flags);
 261                 else
 262                         set_bit(AS_EIO, &inode->i_mapping->flags);
 263
 264                 if ((ioret == -ESHUTDOWN || ioret == -EINTR) &&
 265                      obj->cob_discard_page_warned == 0) {
 266                         obj->cob_discard_page_warned = 1;
 267                         ll_dirty_page_discard_warn(vmpage, ioret);
 268                 }
 269         }
 270 }
 271
 272 static void vvp_page_completion_read(const struct lu_env *env,
 273                                      const struct cl_page_slice *slice,
 274                                      int ioret)
 275 {
 276         struct ccc_page *cp     = cl2ccc_page(slice);
 277         struct page      *vmpage = cp->cpg_page;
 278         struct cl_page  *page   = cl_page_top(slice->cpl_page);
 279         struct inode    *inode  = ccc_object_inode(page->cp_obj);
 280
 281         LASSERT(PageLocked(vmpage));
 282         CL_PAGE_HEADER(D_PAGE, env, page, "completing READ with %d\n", ioret);
 283
 284         if (cp->cpg_defer_uptodate)
 285                 ll_ra_count_put(ll_i2sbi(inode), 1);
 286
 287         if (ioret == 0)  {
 288                 if (!cp->cpg_defer_uptodate)
 289                         cl_page_export(env, page, 1);
 290         } else
 291                 cp->cpg_defer_uptodate = 0;
 292
 293         if (page->cp_sync_io == NULL)
 294                 unlock_page(vmpage);
 295 }
 296
 297 static void vvp_page_completion_write(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  *pg     = slice->cpl_page;
 303         struct page      *vmpage = cp->cpg_page;
 304
 305         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 306
 307         /*
 308          * TODO: Actually it makes sense to add the page into oap pending
 309          * list again and so that we don't need to take the page out from
 310          * SoM write pending list, if we just meet a recoverable error,
 311          * -ENOMEM, etc.
 312          * To implement this, we just need to return a non zero value in
 313          * ->cpo_completion method. The underlying transfer should be notified
 314          * and then re-add the page into pending transfer queue.  -jay
 315          */
 316
 317         cp->cpg_write_queued = 0;
 318         vvp_write_complete(cl2ccc(slice->cpl_obj), cp);
 319
 320         if (pg->cp_sync_io != NULL) {
 321                 LASSERT(PageLocked(vmpage));
 322                 LASSERT(!PageWriteback(vmpage));
 323         } else {
 324                 LASSERT(PageWriteback(vmpage));
 325                 /*
 326                  * Only mark the page error only when it's an async write
 327                  * because applications won't wait for IO to finish.
 328                  */
 329                 vvp_vmpage_error(ccc_object_inode(pg->cp_obj), vmpage, ioret);
 330
 331                 end_page_writeback(vmpage);
 332         }
 333 }
 334
 335 /**
 336  * Implements cl_page_operations::cpo_make_ready() method.
 337  *
 338  * This is called to yank a page from the transfer cache and to send it out as
 339  * a part of transfer. This function try-locks the page. If try-lock failed,
 340  * page is owned by some concurrent IO, and should be skipped (this is bad,
 341  * but hopefully rare situation, as it usually results in transfer being
 342  * shorter than possible).
 343  *
 344  * \retval 0      success, page can be placed into transfer
 345  *
 346  * \retval -EAGAIN page is either used by concurrent IO has been
 347  * truncated. Skip it.
 348  */
 349 static int vvp_page_make_ready(const struct lu_env *env,
 350                                const struct cl_page_slice *slice)
 351 {
 352         struct page *vmpage = cl2vm_page(slice);
 353         struct cl_page *pg = slice->cpl_page;
 354         int result = 0;
 355
 356         lock_page(vmpage);
 357         if (clear_page_dirty_for_io(vmpage)) {
 358                 LASSERT(pg->cp_state == CPS_CACHED);
 359                 /* This actually clears the dirty bit in the radix
 360                  * tree. */
 361                 set_page_writeback(vmpage);
 362                 vvp_write_pending(cl2ccc(slice->cpl_obj),
 363                                 cl2ccc_page(slice));
 364                 CL_PAGE_HEADER(D_PAGE, env, pg, "readied\n");
 365         } else if (pg->cp_state == CPS_PAGEOUT) {
 366                 /* is it possible for osc_flush_async_page() to already
 367                  * make it ready? */
 368                 result = -EALREADY;
 369         } else {
 370                 CL_PAGE_DEBUG(D_ERROR, env, pg, "Unexpecting page state %d.\n",
 371                               pg->cp_state);
 372                 LBUG();
 373         }
 374         unlock_page(vmpage);
 375         return result;
 376 }
 377
 378 static int vvp_page_print(const struct lu_env *env,
 379                           const struct cl_page_slice *slice,
 380                           void *cookie, lu_printer_t printer)
 381 {
 382         struct ccc_page *vp = cl2ccc_page(slice);
 383         struct page      *vmpage = vp->cpg_page;
 384
 385         (*printer)(env, cookie, LUSTRE_VVP_NAME "-page@%p(%d:%d:%d) vm@%p ",
 386                    vp, vp->cpg_defer_uptodate, vp->cpg_ra_used,
 387                    vp->cpg_write_queued, vmpage);
 388         if (vmpage != NULL) {
 389                 (*printer)(env, cookie, "%lx %d:%d %lx %lu %slru",
 390                            (long)vmpage->flags, page_count(vmpage),
 391                            page_mapcount(vmpage), vmpage->private,
 392                            page_index(vmpage),
 393                            list_empty(&vmpage->lru) ? "not-" : "");
 394         }
 395         (*printer)(env, cookie, "\n");
 396         return 0;
 397 }
 398
 399 static const struct cl_page_operations vvp_page_ops = {
 400         .cpo_own           = vvp_page_own,
 401         .cpo_assume     = vvp_page_assume,
 402         .cpo_unassume      = vvp_page_unassume,
 403         .cpo_disown     = vvp_page_disown,
 404         .cpo_vmpage     = ccc_page_vmpage,
 405         .cpo_discard       = vvp_page_discard,
 406         .cpo_delete     = vvp_page_delete,
 407         .cpo_unmap       = vvp_page_unmap,
 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 = ccc_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  = ccc_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 void vvp_transient_page_verify(const struct cl_page *page)
 428 {
 429         struct inode *inode = ccc_object_inode(page->cp_obj);
 430
 431         LASSERT(!mutex_trylock(&inode->i_mutex));
 432 }
 433
 434 static int vvp_transient_page_own(const struct lu_env *env,
 435                                   const struct cl_page_slice *slice,
 436                                   struct cl_io *unused, int nonblock)
 437 {
 438         vvp_transient_page_verify(slice->cpl_page);
 439         return 0;
 440 }
 441
 442 static void vvp_transient_page_assume(const struct lu_env *env,
 443                                       const struct cl_page_slice *slice,
 444                                       struct cl_io *unused)
 445 {
 446         vvp_transient_page_verify(slice->cpl_page);
 447 }
 448
 449 static void vvp_transient_page_unassume(const struct lu_env *env,
 450                                         const struct cl_page_slice *slice,
 451                                         struct cl_io *unused)
 452 {
 453         vvp_transient_page_verify(slice->cpl_page);
 454 }
 455
 456 static void vvp_transient_page_disown(const struct lu_env *env,
 457                                       const struct cl_page_slice *slice,
 458                                       struct cl_io *unused)
 459 {
 460         vvp_transient_page_verify(slice->cpl_page);
 461 }
 462
 463 static void vvp_transient_page_discard(const struct lu_env *env,
 464                                        const struct cl_page_slice *slice,
 465                                        struct cl_io *unused)
 466 {
 467         struct cl_page *page = slice->cpl_page;
 468
 469         vvp_transient_page_verify(slice->cpl_page);
 470
 471         /*
 472          * For transient pages, remove it from the radix tree.
 473          */
 474         cl_page_delete(env, page);
 475 }
 476
 477 static int vvp_transient_page_is_vmlocked(const struct lu_env *env,
 478                                           const struct cl_page_slice *slice)
 479 {
 480         struct inode    *inode = ccc_object_inode(slice->cpl_obj);
 481         int     locked;
 482
 483         locked = !mutex_trylock(&inode->i_mutex);
 484         if (!locked)
 485                 mutex_unlock(&inode->i_mutex);
 486         return locked ? -EBUSY : -ENODATA;
 487 }
 488
 489 static void
 490 vvp_transient_page_completion(const struct lu_env *env,
 491                               const struct cl_page_slice *slice,
 492                               int ioret)
 493 {
 494         vvp_transient_page_verify(slice->cpl_page);
 495 }
 496
 497 static void vvp_transient_page_fini(const struct lu_env *env,
 498                                     struct cl_page_slice *slice)
 499 {
 500         struct ccc_page *cp = cl2ccc_page(slice);
 501         struct cl_page *clp = slice->cpl_page;
 502         struct ccc_object *clobj = cl2ccc(clp->cp_obj);
 503
 504         vvp_page_fini_common(cp);
 505         LASSERT(!mutex_trylock(&clobj->cob_inode->i_mutex));
 506         clobj->cob_transient_pages--;
 507 }
 508
 509 static const struct cl_page_operations vvp_transient_page_ops = {
 510         .cpo_own           = vvp_transient_page_own,
 511         .cpo_assume     = vvp_transient_page_assume,
 512         .cpo_unassume      = vvp_transient_page_unassume,
 513         .cpo_disown     = vvp_transient_page_disown,
 514         .cpo_discard       = vvp_transient_page_discard,
 515         .cpo_vmpage     = ccc_page_vmpage,
 516         .cpo_fini         = vvp_transient_page_fini,
 517         .cpo_is_vmlocked   = vvp_transient_page_is_vmlocked,
 518         .cpo_print       = vvp_page_print,
 519         .cpo_is_under_lock = ccc_page_is_under_lock,
 520         .io = {
 521                 [CRT_READ] = {
 522                         .cpo_prep       = ccc_transient_page_prep,
 523                         .cpo_completion  = vvp_transient_page_completion,
 524                 },
 525                 [CRT_WRITE] = {
 526                         .cpo_prep       = ccc_transient_page_prep,
 527                         .cpo_completion  = vvp_transient_page_completion,
 528                 }
 529         }
 530 };
 531
 532 int vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 533                 struct cl_page *page, struct page *vmpage)
 534 {
 535         struct ccc_page *cpg = cl_object_page_slice(obj, page);
 536
 537         CLOBINVRNT(env, obj, ccc_object_invariant(obj));
 538
 539         cpg->cpg_page = vmpage;
 540         page_cache_get(vmpage);
 541
 542         INIT_LIST_HEAD(&cpg->cpg_pending_linkage);
 543         if (page->cp_type == CPT_CACHEABLE) {
 544                 SetPagePrivate(vmpage);
 545                 vmpage->private = (unsigned long)page;
 546                 cl_page_slice_add(page, &cpg->cpg_cl, obj,
 547                                 &vvp_page_ops);
 548         } else {
 549                 struct ccc_object *clobj = cl2ccc(obj);
 550
 551                 LASSERT(!mutex_trylock(&clobj->cob_inode->i_mutex));
 552                 cl_page_slice_add(page, &cpg->cpg_cl, obj,
 553                                 &vvp_transient_page_ops);
 554                 clobj->cob_transient_pages++;
 555         }
 556         return 0;
 557 }