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, 2013, 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 #ifndef __KERNEL__
  45 # error This file is kernel only.
  46 #endif
  47
  48 #include <obd.h>
  49 #include <lustre_lite.h>
  50
  51 #include "vvp_internal.h"
  52
  53 /*****************************************************************************
  54  *
  55  * Page operations.
  56  *
  57  */
  58
  59 static void vvp_page_fini_common(struct ccc_page *cp)
  60 {
  61         struct page *vmpage = cp->cpg_page;
  62
  63         LASSERT(vmpage != NULL);
  64         page_cache_release(vmpage);
  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         struct page *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 int vvp_page_own(const struct lu_env *env,
  82                         const struct cl_page_slice *slice, struct cl_io *io,
  83                         int nonblock)
  84 {
  85         struct ccc_page *vpg    = cl2ccc_page(slice);
  86         struct page      *vmpage = vpg->cpg_page;
  87
  88         LASSERT(vmpage != NULL);
  89         if (nonblock) {
  90                 if (!trylock_page(vmpage))
  91                         return -EAGAIN;
  92
  93                 if (unlikely(PageWriteback(vmpage))) {
  94                         unlock_page(vmpage);
  95                         return -EAGAIN;
  96                 }
  97
  98                 return 0;
  99         }
 100
 101         lock_page(vmpage);
 102         wait_on_page_writeback(vmpage);
 103         return 0;
 104 }
 105
 106 static void vvp_page_assume(const struct lu_env *env,
 107                             const struct cl_page_slice *slice,
 108                             struct cl_io *unused)
 109 {
 110         struct page *vmpage = cl2vm_page(slice);
 111
 112         LASSERT(vmpage != NULL);
 113         LASSERT(PageLocked(vmpage));
 114         wait_on_page_writeback(vmpage);
 115 }
 116
 117 static void vvp_page_unassume(const struct lu_env *env,
 118                               const struct cl_page_slice *slice,
 119                               struct cl_io *unused)
 120 {
 121         struct page *vmpage = cl2vm_page(slice);
 122
 123         LASSERT(vmpage != NULL);
 124         LASSERT(PageLocked(vmpage));
 125 }
 126
 127 static void vvp_page_disown(const struct lu_env *env,
 128                             const struct cl_page_slice *slice, struct cl_io *io)
 129 {
 130         struct page *vmpage = cl2vm_page(slice);
 131
 132         LASSERT(vmpage != NULL);
 133         LASSERT(PageLocked(vmpage));
 134
 135         unlock_page(cl2vm_page(slice));
 136 }
 137
 138 static void vvp_page_discard(const struct lu_env *env,
 139                              const struct cl_page_slice *slice,
 140                              struct cl_io *unused)
 141 {
 142         struct page          *vmpage  = cl2vm_page(slice);
 143         struct address_space *mapping;
 144         struct ccc_page      *cpg     = cl2ccc_page(slice);
 145         __u64 offset;
 146
 147         LASSERT(vmpage != NULL);
 148         LASSERT(PageLocked(vmpage));
 149
 150         mapping = vmpage->mapping;
 151
 152         if (cpg->cpg_defer_uptodate && !cpg->cpg_ra_used)
 153                 ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
 154
 155         offset = vmpage->index << PAGE_SHIFT;
 156         ll_teardown_mmaps(vmpage->mapping, offset, offset + PAGE_SIZE);
 157
 158         /*
 159          * truncate_complete_page() calls
 160          * a_ops->invalidatepage()->cl_page_delete()->vvp_page_delete().
 161          */
 162         truncate_complete_page(mapping, vmpage);
 163 }
 164
 165 static void vvp_page_delete(const struct lu_env *env,
 166                             const struct cl_page_slice *slice)
 167 {
 168         struct page      *vmpage = cl2vm_page(slice);
 169         struct inode     *inode  = vmpage->mapping->host;
 170         struct cl_object *obj    = slice->cpl_obj;
 171         struct cl_page   *page   = slice->cpl_page;
 172         int refc;
 173
 174         LASSERT(PageLocked(vmpage));
 175         LASSERT((struct cl_page *)vmpage->private == page);
 176         LASSERT(inode == ccc_object_inode(obj));
 177
 178         vvp_write_complete(cl2ccc(obj), cl2ccc_page(slice));
 179
 180         /* Drop the reference count held in vvp_page_init */
 181         refc = atomic_dec_return(&page->cp_ref);
 182         LASSERTF(refc >= 1, "page = %p, refc = %d\n", page, refc);
 183
 184         ClearPageUptodate(vmpage);
 185         ClearPagePrivate(vmpage);
 186         vmpage->private = 0;
 187         /*
 188          * Reference from vmpage to cl_page is removed, but the reference back
 189          * is still here. It is removed later in vvp_page_fini().
 190          */
 191 }
 192
 193 static void vvp_page_export(const struct lu_env *env,
 194                             const struct cl_page_slice *slice,
 195                             int uptodate)
 196 {
 197         struct page *vmpage = cl2vm_page(slice);
 198
 199         LASSERT(vmpage != NULL);
 200         LASSERT(PageLocked(vmpage));
 201         if (uptodate)
 202                 SetPageUptodate(vmpage);
 203         else
 204                 ClearPageUptodate(vmpage);
 205 }
 206
 207 static int vvp_page_is_vmlocked(const struct lu_env *env,
 208                                 const struct cl_page_slice *slice)
 209 {
 210         return PageLocked(cl2vm_page(slice)) ? -EBUSY : -ENODATA;
 211 }
 212
 213 static int vvp_page_prep_read(const struct lu_env *env,
 214                               const struct cl_page_slice *slice,
 215                               struct cl_io *unused)
 216 {
 217         ENTRY;
 218         /* Skip the page already marked as PG_uptodate. */
 219         RETURN(PageUptodate(cl2vm_page(slice)) ? -EALREADY : 0);
 220 }
 221
 222 static int vvp_page_prep_write(const struct lu_env *env,
 223                                const struct cl_page_slice *slice,
 224                                struct cl_io *unused)
 225 {
 226         struct page *vmpage = cl2vm_page(slice);
 227
 228         LASSERT(PageLocked(vmpage));
 229         LASSERT(!PageDirty(vmpage));
 230
 231         set_page_writeback(vmpage);
 232         vvp_write_pending(cl2ccc(slice->cpl_obj), cl2ccc_page(slice));
 233
 234         return 0;
 235 }
 236
 237 /**
 238  * Handles page transfer errors at VM level.
 239  *
 240  * This takes inode as a separate argument, because inode on which error is to
 241  * be set can be different from \a vmpage inode in case of direct-io.
 242  */
 243 static void vvp_vmpage_error(struct inode *inode, struct page *vmpage, int ioret)
 244 {
 245         struct ccc_object *obj = cl_inode2ccc(inode);
 246
 247         if (ioret == 0) {
 248                 ClearPageError(vmpage);
 249                 obj->cob_discard_page_warned = 0;
 250         } else {
 251                 SetPageError(vmpage);
 252                 if (ioret == -ENOSPC)
 253                         set_bit(AS_ENOSPC, &inode->i_mapping->flags);
 254                 else
 255                         set_bit(AS_EIO, &inode->i_mapping->flags);
 256
 257                 if ((ioret == -ESHUTDOWN || ioret == -EINTR) &&
 258                      obj->cob_discard_page_warned == 0) {
 259                         obj->cob_discard_page_warned = 1;
 260                         ll_dirty_page_discard_warn(vmpage, ioret);
 261                 }
 262         }
 263 }
 264
 265 static void vvp_page_completion_read(const struct lu_env *env,
 266                                      const struct cl_page_slice *slice,
 267                                      int ioret)
 268 {
 269         struct ccc_page *cp     = cl2ccc_page(slice);
 270         struct page      *vmpage = cp->cpg_page;
 271         struct cl_page  *page   = cl_page_top(slice->cpl_page);
 272         struct inode    *inode  = ccc_object_inode(page->cp_obj);
 273         ENTRY;
 274
 275         LASSERT(PageLocked(vmpage));
 276         CL_PAGE_HEADER(D_PAGE, env, page, "completing READ with %d\n", ioret);
 277
 278         if (cp->cpg_defer_uptodate)
 279                 ll_ra_count_put(ll_i2sbi(inode), 1);
 280
 281         if (ioret == 0)  {
 282                 if (!cp->cpg_defer_uptodate)
 283                         cl_page_export(env, page, 1);
 284         } else
 285                 cp->cpg_defer_uptodate = 0;
 286
 287         if (page->cp_sync_io == NULL)
 288                 unlock_page(vmpage);
 289
 290         EXIT;
 291 }
 292
 293 static void vvp_page_completion_write(const struct lu_env *env,
 294                                       const struct cl_page_slice *slice,
 295                                       int ioret)
 296 {
 297         struct ccc_page *cp     = cl2ccc_page(slice);
 298         struct cl_page  *pg     = slice->cpl_page;
 299         struct page      *vmpage = cp->cpg_page;
 300         ENTRY;
 301
 302         LASSERT(ergo(pg->cp_sync_io != NULL, PageLocked(vmpage)));
 303         LASSERT(PageWriteback(vmpage));
 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         /*
 321          * Only mark the page error only when it's an async write because
 322          * applications won't wait for IO to finish.
 323          */
 324         if (pg->cp_sync_io == NULL)
 325                 vvp_vmpage_error(ccc_object_inode(pg->cp_obj), vmpage, ioret);
 326
 327         end_page_writeback(vmpage);
 328         EXIT;
 329 }
 330
 331 /**
 332  * Implements cl_page_operations::cpo_make_ready() method.
 333  *
 334  * This is called to yank a page from the transfer cache and to send it out as
 335  * a part of transfer. This function try-locks the page. If try-lock failed,
 336  * page is owned by some concurrent IO, and should be skipped (this is bad,
 337  * but hopefully rare situation, as it usually results in transfer being
 338  * shorter than possible).
 339  *
 340  * \retval 0      success, page can be placed into transfer
 341  *
 342  * \retval -EAGAIN page is either used by concurrent IO has been
 343  * truncated. Skip it.
 344  */
 345 static int vvp_page_make_ready(const struct lu_env *env,
 346                                const struct cl_page_slice *slice)
 347 {
 348         struct page *vmpage = cl2vm_page(slice);
 349         struct cl_page *pg = slice->cpl_page;
 350         int result = 0;
 351
 352         lock_page(vmpage);
 353         if (clear_page_dirty_for_io(vmpage)) {
 354                 LASSERT(pg->cp_state == CPS_CACHED);
 355                 /* This actually clears the dirty bit in the radix
 356                  * tree. */
 357                 set_page_writeback(vmpage);
 358                 vvp_write_pending(cl2ccc(slice->cpl_obj),
 359                                 cl2ccc_page(slice));
 360                 CL_PAGE_HEADER(D_PAGE, env, pg, "readied\n");
 361         } else if (pg->cp_state == CPS_PAGEOUT) {
 362                 /* is it possible for osc_flush_async_page() to already
 363                  * make it ready? */
 364                 result = -EALREADY;
 365         } else {
 366                 CL_PAGE_DEBUG(D_ERROR, env, pg, "Unexpecting page state %d.\n",
 367                               pg->cp_state);
 368                 LBUG();
 369         }
 370         unlock_page(vmpage);
 371         RETURN(result);
 372 }
 373
 374 static int vvp_page_print(const struct lu_env *env,
 375                           const struct cl_page_slice *slice,
 376                           void *cookie, lu_printer_t printer)
 377 {
 378         struct ccc_page *vp = cl2ccc_page(slice);
 379         struct page      *vmpage = vp->cpg_page;
 380
 381         (*printer)(env, cookie, LUSTRE_VVP_NAME"-page@%p(%d:%d:%d) "
 382                    "vm@%p ",
 383                    vp, vp->cpg_defer_uptodate, vp->cpg_ra_used,
 384                    vp->cpg_write_queued, vmpage);
 385         if (vmpage != NULL) {
 386                 (*printer)(env, cookie, "%lx %d:%d %lx %lu %slru",
 387                            (long)vmpage->flags, page_count(vmpage),
 388                            page_mapcount(vmpage), vmpage->private,
 389                            page_index(vmpage),
 390                            list_empty(&vmpage->lru) ? "not-" : "");
 391         }
 392         (*printer)(env, cookie, "\n");
 393         return 0;
 394 }
 395
 396 static const struct cl_page_operations vvp_page_ops = {
 397         .cpo_own           = vvp_page_own,
 398         .cpo_assume        = vvp_page_assume,
 399         .cpo_unassume      = vvp_page_unassume,
 400         .cpo_disown        = vvp_page_disown,
 401         .cpo_vmpage        = ccc_page_vmpage,
 402         .cpo_discard       = vvp_page_discard,
 403         .cpo_delete        = vvp_page_delete,
 404         .cpo_export        = vvp_page_export,
 405         .cpo_is_vmlocked   = vvp_page_is_vmlocked,
 406         .cpo_fini          = vvp_page_fini,
 407         .cpo_print         = vvp_page_print,
 408         .cpo_is_under_lock = ccc_page_is_under_lock,
 409         .io = {
 410                 [CRT_READ] = {
 411                         .cpo_prep        = vvp_page_prep_read,
 412                         .cpo_completion  = vvp_page_completion_read,
 413                         .cpo_make_ready  = ccc_fail,
 414                 },
 415                 [CRT_WRITE] = {
 416                         .cpo_prep        = vvp_page_prep_write,
 417                         .cpo_completion  = vvp_page_completion_write,
 418                         .cpo_make_ready  = vvp_page_make_ready,
 419                 }
 420         }
 421 };
 422
 423 static void vvp_transient_page_verify(const struct cl_page *page)
 424 {
 425         struct inode *inode = ccc_object_inode(page->cp_obj);
 426
 427         LASSERT(!mutex_trylock(&inode->i_mutex));
 428 }
 429
 430 static int vvp_transient_page_own(const struct lu_env *env,
 431                                   const struct cl_page_slice *slice,
 432                                   struct cl_io *unused, int nonblock)
 433 {
 434         vvp_transient_page_verify(slice->cpl_page);
 435         return 0;
 436 }
 437
 438 static void vvp_transient_page_assume(const struct lu_env *env,
 439                                       const struct cl_page_slice *slice,
 440                                       struct cl_io *unused)
 441 {
 442         vvp_transient_page_verify(slice->cpl_page);
 443 }
 444
 445 static void vvp_transient_page_unassume(const struct lu_env *env,
 446                                         const struct cl_page_slice *slice,
 447                                         struct cl_io *unused)
 448 {
 449         vvp_transient_page_verify(slice->cpl_page);
 450 }
 451
 452 static void vvp_transient_page_disown(const struct lu_env *env,
 453                                       const struct cl_page_slice *slice,
 454                                       struct cl_io *unused)
 455 {
 456         vvp_transient_page_verify(slice->cpl_page);
 457 }
 458
 459 static void vvp_transient_page_discard(const struct lu_env *env,
 460                                        const struct cl_page_slice *slice,
 461                                        struct cl_io *unused)
 462 {
 463         struct cl_page *page = slice->cpl_page;
 464
 465         vvp_transient_page_verify(slice->cpl_page);
 466
 467         /*
 468          * For transient pages, remove it from the radix tree.
 469          */
 470         cl_page_delete(env, page);
 471 }
 472
 473 static int vvp_transient_page_is_vmlocked(const struct lu_env *env,
 474                                           const struct cl_page_slice *slice)
 475 {
 476         struct inode    *inode = ccc_object_inode(slice->cpl_obj);
 477         int     locked;
 478
 479         locked = !mutex_trylock(&inode->i_mutex);
 480         if (!locked)
 481                 mutex_unlock(&inode->i_mutex);
 482         return locked ? -EBUSY : -ENODATA;
 483 }
 484
 485 static void
 486 vvp_transient_page_completion(const struct lu_env *env,
 487                               const struct cl_page_slice *slice,
 488                               int ioret)
 489 {
 490         vvp_transient_page_verify(slice->cpl_page);
 491 }
 492
 493 static void vvp_transient_page_fini(const struct lu_env *env,
 494                                     struct cl_page_slice *slice)
 495 {
 496         struct ccc_page *cp = cl2ccc_page(slice);
 497         struct cl_page *clp = slice->cpl_page;
 498         struct ccc_object *clobj = cl2ccc(clp->cp_obj);
 499
 500         vvp_page_fini_common(cp);
 501         LASSERT(!mutex_trylock(&clobj->cob_inode->i_mutex));
 502         clobj->cob_transient_pages--;
 503 }
 504
 505 static const struct cl_page_operations vvp_transient_page_ops = {
 506         .cpo_own           = vvp_transient_page_own,
 507         .cpo_assume        = vvp_transient_page_assume,
 508         .cpo_unassume      = vvp_transient_page_unassume,
 509         .cpo_disown        = vvp_transient_page_disown,
 510         .cpo_discard       = vvp_transient_page_discard,
 511         .cpo_vmpage        = ccc_page_vmpage,
 512         .cpo_fini          = vvp_transient_page_fini,
 513         .cpo_is_vmlocked   = vvp_transient_page_is_vmlocked,
 514         .cpo_print         = vvp_page_print,
 515         .cpo_is_under_lock = ccc_page_is_under_lock,
 516         .io = {
 517                 [CRT_READ] = {
 518                         .cpo_prep        = ccc_transient_page_prep,
 519                         .cpo_completion  = vvp_transient_page_completion,
 520                 },
 521                 [CRT_WRITE] = {
 522                         .cpo_prep        = ccc_transient_page_prep,
 523                         .cpo_completion  = vvp_transient_page_completion,
 524                 }
 525         }
 526 };
 527
 528 int vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 529                 struct cl_page *page, struct page *vmpage)
 530 {
 531         struct ccc_page *cpg = cl_object_page_slice(obj, page);
 532
 533         CLOBINVRNT(env, obj, ccc_object_invariant(obj));
 534
 535         cpg->cpg_page = vmpage;
 536         page_cache_get(vmpage);
 537
 538         CFS_INIT_LIST_HEAD(&cpg->cpg_pending_linkage);
 539         if (page->cp_type == CPT_CACHEABLE) {
 540                 /* in cache, decref in vvp_page_delete */
 541                 atomic_inc(&page->cp_ref);
 542                 SetPagePrivate(vmpage);
 543                 vmpage->private = (unsigned long)page;
 544                 cl_page_slice_add(page, &cpg->cpg_cl, obj,
 545                                 &vvp_page_ops);
 546         } else {
 547                 struct ccc_object *clobj = cl2ccc(obj);
 548
 549                 LASSERT(!mutex_trylock(&clobj->cob_inode->i_mutex));
 550                 cl_page_slice_add(page, &cpg->cpg_cl, obj,
 551                                 &vvp_transient_page_ops);
 552                 clobj->cob_transient_pages++;
 553         }
 554         return 0;
 555 }
 556