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, Whamcloud, Inc.
  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         cfs_page_t *vmpage = cp->cpg_page;
  62
  63         LASSERT(vmpage != NULL);
  64         page_cache_release(vmpage);
  65         OBD_SLAB_FREE_PTR(cp, vvp_page_kmem);
  66 }
  67
  68 static void vvp_page_fini(const struct lu_env *env,
  69                           struct cl_page_slice *slice)
  70 {
  71         struct ccc_page *cp = cl2ccc_page(slice);
  72         cfs_page_t *vmpage  = cp->cpg_page;
  73
  74         /*
  75          * vmpage->private was already cleared when page was moved into
  76          * VPG_FREEING state.
  77          */
  78         LASSERT((struct cl_page *)vmpage->private != slice->cpl_page);
  79         vvp_page_fini_common(cp);
  80 }
  81
  82 static int vvp_page_own(const struct lu_env *env,
  83                         const struct cl_page_slice *slice, struct cl_io *io,
  84                         int nonblock)
  85 {
  86         struct ccc_page *vpg    = cl2ccc_page(slice);
  87         cfs_page_t      *vmpage = vpg->cpg_page;
  88
  89         LASSERT(vmpage != NULL);
  90         if (nonblock) {
  91                 if (!trylock_page(vmpage))
  92                         return -EAGAIN;
  93
  94                 if (unlikely(PageWriteback(vmpage))) {
  95                         unlock_page(vmpage);
  96                         return -EAGAIN;
  97                 }
  98
  99                 return 0;
 100         }
 101
 102         lock_page(vmpage);
 103         wait_on_page_writeback(vmpage);
 104         return 0;
 105 }
 106
 107 static void vvp_page_assume(const struct lu_env *env,
 108                             const struct cl_page_slice *slice,
 109                             struct cl_io *unused)
 110 {
 111         cfs_page_t *vmpage = cl2vm_page(slice);
 112
 113         LASSERT(vmpage != NULL);
 114         LASSERT(PageLocked(vmpage));
 115         wait_on_page_writeback(vmpage);
 116 }
 117
 118 static void vvp_page_unassume(const struct lu_env *env,
 119                               const struct cl_page_slice *slice,
 120                               struct cl_io *unused)
 121 {
 122         cfs_page_t *vmpage = cl2vm_page(slice);
 123
 124         LASSERT(vmpage != NULL);
 125         LASSERT(PageLocked(vmpage));
 126 }
 127
 128 static void vvp_page_disown(const struct lu_env *env,
 129                             const struct cl_page_slice *slice, struct cl_io *io)
 130 {
 131         cfs_page_t *vmpage = cl2vm_page(slice);
 132
 133         LASSERT(vmpage != NULL);
 134         LASSERT(PageLocked(vmpage));
 135
 136         unlock_page(cl2vm_page(slice));
 137 }
 138
 139 static void vvp_page_discard(const struct lu_env *env,
 140                              const struct cl_page_slice *slice,
 141                              struct cl_io *unused)
 142 {
 143         cfs_page_t           *vmpage  = cl2vm_page(slice);
 144         struct address_space *mapping = vmpage->mapping;
 145         struct ccc_page      *cpg     = cl2ccc_page(slice);
 146
 147         LASSERT(vmpage != NULL);
 148         LASSERT(PageLocked(vmpage));
 149
 150         if (cpg->cpg_defer_uptodate && !cpg->cpg_ra_used)
 151                 ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
 152
 153         /*
 154          * truncate_complete_page() calls
 155          * a_ops->invalidatepage()->cl_page_delete()->vvp_page_delete().
 156          */
 157         truncate_complete_page(mapping, vmpage);
 158 }
 159
 160 static int vvp_page_unmap(const struct lu_env *env,
 161                           const struct cl_page_slice *slice,
 162                           struct cl_io *unused)
 163 {
 164         cfs_page_t *vmpage = cl2vm_page(slice);
 165         __u64       offset = vmpage->index << CFS_PAGE_SHIFT;
 166
 167         LASSERT(vmpage != NULL);
 168         LASSERT(PageLocked(vmpage));
 169         /*
 170          * XXX is it safe to call this with the page lock held?
 171          */
 172         ll_teardown_mmaps(vmpage->mapping, offset, offset + CFS_PAGE_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         cfs_page_t       *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         cfs_page_t *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         ENTRY;
 221         /* Skip the page already marked as PG_uptodate. */
 222         RETURN(PageUptodate(cl2vm_page(slice)) ? -EALREADY : 0);
 223 }
 224
 225 static int vvp_page_prep_write(const struct lu_env *env,
 226                                const struct cl_page_slice *slice,
 227                                struct cl_io *unused)
 228 {
 229         cfs_page_t *vmpage = cl2vm_page(slice);
 230
 231         LASSERT(PageLocked(vmpage));
 232         LASSERT(!PageDirty(vmpage));
 233
 234         set_page_writeback(vmpage);
 235         vvp_write_pending(cl2ccc(slice->cpl_obj), cl2ccc_page(slice));
 236
 237         return 0;
 238 }
 239
 240 /**
 241  * Handles page transfer errors at VM level.
 242  *
 243  * This takes inode as a separate argument, because inode on which error is to
 244  * be set can be different from \a vmpage inode in case of direct-io.
 245  */
 246 static void vvp_vmpage_error(struct inode *inode, cfs_page_t *vmpage, int ioret)
 247 {
 248         if (ioret == 0)
 249                 ClearPageError(vmpage);
 250         else if (ioret != -EINTR) {
 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 }
 258
 259 static void vvp_page_completion_read(const struct lu_env *env,
 260                                      const struct cl_page_slice *slice,
 261                                      int ioret)
 262 {
 263         struct ccc_page *cp     = cl2ccc_page(slice);
 264         cfs_page_t      *vmpage = cp->cpg_page;
 265         struct cl_page  *page   = cl_page_top(slice->cpl_page);
 266         struct inode    *inode  = ccc_object_inode(page->cp_obj);
 267         ENTRY;
 268
 269         LASSERT(PageLocked(vmpage));
 270         CL_PAGE_HEADER(D_PAGE, env, page, "completing READ with %d\n", ioret);
 271
 272         if (cp->cpg_defer_uptodate)
 273                 ll_ra_count_put(ll_i2sbi(inode), 1);
 274
 275         if (ioret == 0)  {
 276                 if (!cp->cpg_defer_uptodate)
 277                         cl_page_export(env, page, 1);
 278         } else
 279                 cp->cpg_defer_uptodate = 0;
 280
 281         if (page->cp_sync_io == NULL)
 282                 unlock_page(vmpage);
 283
 284         EXIT;
 285 }
 286
 287 static void vvp_page_completion_write(const struct lu_env *env,
 288                                       const struct cl_page_slice *slice,
 289                                       int ioret)
 290 {
 291         struct ccc_page *cp     = cl2ccc_page(slice);
 292         struct cl_page  *pg     = slice->cpl_page;
 293         cfs_page_t      *vmpage = cp->cpg_page;
 294         ENTRY;
 295
 296         LASSERT(ergo(pg->cp_sync_io != NULL, PageLocked(vmpage)));
 297         LASSERT(PageWriteback(vmpage));
 298
 299         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 300
 301         /*
 302          * TODO: Actually it makes sense to add the page into oap pending
 303          * list again and so that we don't need to take the page out from
 304          * SoM write pending list, if we just meet a recoverable error,
 305          * -ENOMEM, etc.
 306          * To implement this, we just need to return a non zero value in
 307          * ->cpo_completion method. The underlying transfer should be notified
 308          * and then re-add the page into pending transfer queue.  -jay
 309          */
 310
 311         cp->cpg_write_queued = 0;
 312         vvp_write_complete(cl2ccc(slice->cpl_obj), cp);
 313
 314         /*
 315          * Only mark the page error only when it's an async write because
 316          * applications won't wait for IO to finish.
 317          */
 318         if (pg->cp_sync_io == NULL)
 319                 vvp_vmpage_error(ccc_object_inode(pg->cp_obj), vmpage, ioret);
 320
 321         end_page_writeback(vmpage);
 322         EXIT;
 323 }
 324
 325 /**
 326  * Implements cl_page_operations::cpo_make_ready() method.
 327  *
 328  * This is called to yank a page from the transfer cache and to send it out as
 329  * a part of transfer. This function try-locks the page. If try-lock failed,
 330  * page is owned by some concurrent IO, and should be skipped (this is bad,
 331  * but hopefully rare situation, as it usually results in transfer being
 332  * shorter than possible).
 333  *
 334  * \retval 0      success, page can be placed into transfer
 335  *
 336  * \retval -EAGAIN page is either used by concurrent IO has been
 337  * truncated. Skip it.
 338  */
 339 static int vvp_page_make_ready(const struct lu_env *env,
 340                                const struct cl_page_slice *slice)
 341 {
 342         cfs_page_t *vmpage = cl2vm_page(slice);
 343         struct cl_page *pg = slice->cpl_page;
 344         int result = 0;
 345
 346         lock_page(vmpage);
 347         if (clear_page_dirty_for_io(vmpage)) {
 348                 LASSERT(pg->cp_state == CPS_CACHED);
 349                 /* This actually clears the dirty bit in the radix
 350                  * tree. */
 351                 set_page_writeback(vmpage);
 352                 vvp_write_pending(cl2ccc(slice->cpl_obj),
 353                                 cl2ccc_page(slice));
 354                 CL_PAGE_HEADER(D_PAGE, env, pg, "readied\n");
 355         } else if (pg->cp_state == CPS_PAGEOUT) {
 356                 /* is it possible for osc_flush_async_page() to already
 357                  * make it ready? */
 358                 result = -EALREADY;
 359         } else {
 360                 CL_PAGE_DEBUG(D_ERROR, env, pg, "Unexpecting page state %d.\n",
 361                               pg->cp_state);
 362                 LBUG();
 363         }
 364         unlock_page(vmpage);
 365         RETURN(result);
 366 }
 367
 368 static int vvp_page_print(const struct lu_env *env,
 369                           const struct cl_page_slice *slice,
 370                           void *cookie, lu_printer_t printer)
 371 {
 372         struct ccc_page *vp = cl2ccc_page(slice);
 373         cfs_page_t      *vmpage = vp->cpg_page;
 374
 375         (*printer)(env, cookie, LUSTRE_VVP_NAME"-page@%p(%d:%d:%d) "
 376                    "vm@%p ",
 377                    vp, vp->cpg_defer_uptodate, vp->cpg_ra_used,
 378                    vp->cpg_write_queued, vmpage);
 379         if (vmpage != NULL) {
 380                 (*printer)(env, cookie, "%lx %d:%d %lx %lu %slru",
 381                            (long)vmpage->flags, page_count(vmpage),
 382                            page_mapcount(vmpage), vmpage->private,
 383                            page_index(vmpage),
 384                            list_empty(&vmpage->lru) ? "not-" : "");
 385         }
 386         (*printer)(env, cookie, "\n");
 387         return 0;
 388 }
 389
 390 static const struct cl_page_operations vvp_page_ops = {
 391         .cpo_own           = vvp_page_own,
 392         .cpo_assume        = vvp_page_assume,
 393         .cpo_unassume      = vvp_page_unassume,
 394         .cpo_disown        = vvp_page_disown,
 395         .cpo_vmpage        = ccc_page_vmpage,
 396         .cpo_discard       = vvp_page_discard,
 397         .cpo_delete        = vvp_page_delete,
 398         .cpo_unmap         = vvp_page_unmap,
 399         .cpo_export        = vvp_page_export,
 400         .cpo_is_vmlocked   = vvp_page_is_vmlocked,
 401         .cpo_fini          = vvp_page_fini,
 402         .cpo_print         = vvp_page_print,
 403         .cpo_is_under_lock = ccc_page_is_under_lock,
 404         .io = {
 405                 [CRT_READ] = {
 406                         .cpo_prep        = vvp_page_prep_read,
 407                         .cpo_completion  = vvp_page_completion_read,
 408                         .cpo_make_ready  = ccc_fail,
 409                 },
 410                 [CRT_WRITE] = {
 411                         .cpo_prep        = vvp_page_prep_write,
 412                         .cpo_completion  = vvp_page_completion_write,
 413                         .cpo_make_ready  = vvp_page_make_ready,
 414                 }
 415         }
 416 };
 417
 418 static void vvp_transient_page_verify(const struct cl_page *page)
 419 {
 420         struct inode *inode = ccc_object_inode(page->cp_obj);
 421
 422         LASSERT(!mutex_trylock(&inode->i_mutex));
 423         /* LASSERT_SEM_LOCKED(&inode->i_alloc_sem); */
 424 }
 425
 426 static int vvp_transient_page_own(const struct lu_env *env,
 427                                   const struct cl_page_slice *slice,
 428                                   struct cl_io *unused, int nonblock)
 429 {
 430         vvp_transient_page_verify(slice->cpl_page);
 431         return 0;
 432 }
 433
 434 static void vvp_transient_page_assume(const struct lu_env *env,
 435                                       const struct cl_page_slice *slice,
 436                                       struct cl_io *unused)
 437 {
 438         vvp_transient_page_verify(slice->cpl_page);
 439 }
 440
 441 static void vvp_transient_page_unassume(const struct lu_env *env,
 442                                         const struct cl_page_slice *slice,
 443                                         struct cl_io *unused)
 444 {
 445         vvp_transient_page_verify(slice->cpl_page);
 446 }
 447
 448 static void vvp_transient_page_disown(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_discard(const struct lu_env *env,
 456                                        const struct cl_page_slice *slice,
 457                                        struct cl_io *unused)
 458 {
 459         struct cl_page *page = slice->cpl_page;
 460
 461         vvp_transient_page_verify(slice->cpl_page);
 462
 463         /*
 464          * For transient pages, remove it from the radix tree.
 465          */
 466         cl_page_delete(env, page);
 467 }
 468
 469 static int vvp_transient_page_is_vmlocked(const struct lu_env *env,
 470                                           const struct cl_page_slice *slice)
 471 {
 472         struct inode    *inode = ccc_object_inode(slice->cpl_obj);
 473         int     locked;
 474
 475         locked = !mutex_trylock(&inode->i_mutex);
 476         if (!locked)
 477                 mutex_unlock(&inode->i_mutex);
 478         return locked ? -EBUSY : -ENODATA;
 479 }
 480
 481 static void
 482 vvp_transient_page_completion(const struct lu_env *env,
 483                               const struct cl_page_slice *slice,
 484                               int ioret)
 485 {
 486         vvp_transient_page_verify(slice->cpl_page);
 487 }
 488
 489 static void vvp_transient_page_fini(const struct lu_env *env,
 490                                     struct cl_page_slice *slice)
 491 {
 492         struct ccc_page *cp = cl2ccc_page(slice);
 493         struct cl_page *clp = slice->cpl_page;
 494         struct ccc_object *clobj = cl2ccc(clp->cp_obj);
 495
 496         vvp_page_fini_common(cp);
 497         LASSERT(!mutex_trylock(&clobj->cob_inode->i_mutex));
 498         clobj->cob_transient_pages--;
 499 }
 500
 501 static const struct cl_page_operations vvp_transient_page_ops = {
 502         .cpo_own           = vvp_transient_page_own,
 503         .cpo_assume        = vvp_transient_page_assume,
 504         .cpo_unassume      = vvp_transient_page_unassume,
 505         .cpo_disown        = vvp_transient_page_disown,
 506         .cpo_discard       = vvp_transient_page_discard,
 507         .cpo_vmpage        = ccc_page_vmpage,
 508         .cpo_fini          = vvp_transient_page_fini,
 509         .cpo_is_vmlocked   = vvp_transient_page_is_vmlocked,
 510         .cpo_print         = vvp_page_print,
 511         .cpo_is_under_lock = ccc_page_is_under_lock,
 512         .io = {
 513                 [CRT_READ] = {
 514                         .cpo_prep        = ccc_transient_page_prep,
 515                         .cpo_completion  = vvp_transient_page_completion,
 516                 },
 517                 [CRT_WRITE] = {
 518                         .cpo_prep        = ccc_transient_page_prep,
 519                         .cpo_completion  = vvp_transient_page_completion,
 520                 }
 521         }
 522 };
 523
 524 struct cl_page *vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 525                               struct cl_page *page, cfs_page_t *vmpage)
 526 {
 527         struct ccc_page *cpg;
 528         int result;
 529
 530         CLOBINVRNT(env, obj, ccc_object_invariant(obj));
 531
 532         OBD_SLAB_ALLOC_PTR_GFP(cpg, vvp_page_kmem, CFS_ALLOC_IO);
 533         if (cpg != NULL) {
 534                 cpg->cpg_page = vmpage;
 535                 page_cache_get(vmpage);
 536
 537                 CFS_INIT_LIST_HEAD(&cpg->cpg_pending_linkage);
 538                 if (page->cp_type == CPT_CACHEABLE) {
 539                         SetPagePrivate(vmpage);
 540                         vmpage->private = (unsigned long)page;
 541                         cl_page_slice_add(page, &cpg->cpg_cl, obj,
 542                                           &vvp_page_ops);
 543                 } else {
 544                         struct ccc_object *clobj = cl2ccc(obj);
 545
 546                         LASSERT(!mutex_trylock(&clobj->cob_inode->i_mutex));
 547                         cl_page_slice_add(page, &cpg->cpg_cl, obj,
 548                                           &vvp_transient_page_ops);
 549                         clobj->cob_transient_pages++;
 550                 }
 551                 result = 0;
 552         } else
 553                 result = -ENOMEM;
 554         return ERR_PTR(result);
 555 }
 556