lustre/llite/vvp_page.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
  30  * Use is subject to license terms.
  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  */
  40
  41 #define DEBUG_SUBSYSTEM S_LLITE
  42
  43 #ifndef __KERNEL__
  44 # error This file is kernel only.
  45 #endif
  46
  47 #include <obd.h>
  48 #include <lustre_lite.h>
  49
  50 #include "vvp_internal.h"
  51
  52 /*****************************************************************************
  53  *
  54  * Page operations.
  55  *
  56  */
  57
  58 static void vvp_page_fini_common(struct ccc_page *cp)
  59 {
  60         cfs_page_t *vmpage = cp->cpg_page;
  61
  62         LASSERT(vmpage != NULL);
  63         page_cache_release(vmpage);
  64         OBD_SLAB_FREE_PTR(cp, vvp_page_kmem);
  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         cfs_page_t *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 void vvp_page_own(const struct lu_env *env,
  82                          const struct cl_page_slice *slice, struct cl_io *_)
  83 {
  84         struct ccc_page *vpg    = cl2ccc_page(slice);
  85         cfs_page_t      *vmpage = vpg->cpg_page;
  86
  87         LASSERT(vmpage != NULL);
  88         lock_page(vmpage);
  89         wait_on_page_writeback(vmpage);
  90 }
  91
  92 static void vvp_page_assume(const struct lu_env *env,
  93                             const struct cl_page_slice *slice, struct cl_io *_)
  94 {
  95         cfs_page_t *vmpage = cl2vm_page(slice);
  96
  97         LASSERT(vmpage != NULL);
  98         LASSERT(PageLocked(vmpage));
  99         wait_on_page_writeback(vmpage);
 100 }
 101
 102 static void vvp_page_unassume(const struct lu_env *env,
 103                               const struct cl_page_slice *slice,
 104                               struct cl_io *_)
 105 {
 106         cfs_page_t *vmpage = cl2vm_page(slice);
 107
 108         LASSERT(vmpage != NULL);
 109         LASSERT(PageLocked(vmpage));
 110 }
 111
 112 static void vvp_page_disown(const struct lu_env *env,
 113                             const struct cl_page_slice *slice, struct cl_io *io)
 114 {
 115         cfs_page_t *vmpage = cl2vm_page(slice);
 116
 117         LASSERT(vmpage != NULL);
 118         LASSERT(PageLocked(vmpage));
 119
 120         unlock_page(cl2vm_page(slice));
 121 }
 122
 123 static void vvp_page_discard(const struct lu_env *env,
 124                              const struct cl_page_slice *slice, struct cl_io *_)
 125 {
 126         cfs_page_t           *vmpage  = cl2vm_page(slice);
 127         struct address_space *mapping = vmpage->mapping;
 128         struct ccc_page      *cpg     = cl2ccc_page(slice);
 129
 130         LASSERT(vmpage != NULL);
 131         LASSERT(PageLocked(vmpage));
 132
 133         if (cpg->cpg_defer_uptodate && !cpg->cpg_ra_used)
 134                 ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
 135
 136         /*
 137          * truncate_complete_page() calls
 138          * a_ops->invalidatepage()->cl_page_delete()->vvp_page_delete().
 139          */
 140         truncate_complete_page(mapping, vmpage);
 141 }
 142
 143 static int vvp_page_unmap(const struct lu_env *env,
 144                           const struct cl_page_slice *slice, struct cl_io *_)
 145 {
 146         cfs_page_t *vmpage = cl2vm_page(slice);
 147         __u64       offset = vmpage->index << CFS_PAGE_SHIFT;
 148
 149         LASSERT(vmpage != NULL);
 150         LASSERT(PageLocked(vmpage));
 151         /*
 152          * XXX is it safe to call this with the page lock held?
 153          */
 154         ll_teardown_mmaps(vmpage->mapping, offset, offset + CFS_PAGE_SIZE);
 155         return 0;
 156 }
 157
 158 static void vvp_page_delete(const struct lu_env *env,
 159                             const struct cl_page_slice *slice)
 160 {
 161         cfs_page_t       *vmpage = cl2vm_page(slice);
 162         struct inode     *inode  = vmpage->mapping->host;
 163         struct cl_object *obj    = slice->cpl_obj;
 164
 165         LASSERT(PageLocked(vmpage));
 166         LASSERT((struct cl_page *)vmpage->private == slice->cpl_page);
 167         LASSERT(inode == ccc_object_inode(obj));
 168
 169         vvp_write_complete(cl2ccc(obj), cl2ccc_page(slice));
 170         ClearPagePrivate(vmpage);
 171         vmpage->private = 0;
 172         /*
 173          * Reference from vmpage to cl_page is removed, but the reference back
 174          * is still here. It is removed later in vvp_page_fini().
 175          */
 176 }
 177
 178 static void vvp_page_export(const struct lu_env *env,
 179                             const struct cl_page_slice *slice)
 180 {
 181         cfs_page_t *vmpage = cl2vm_page(slice);
 182
 183         LASSERT(vmpage != NULL);
 184         LASSERT(PageLocked(vmpage));
 185         SetPageUptodate(vmpage);
 186 }
 187
 188 static int vvp_page_is_vmlocked(const struct lu_env *env,
 189                                 const struct cl_page_slice *slice)
 190 {
 191         return PageLocked(cl2vm_page(slice)) ? -EBUSY : -ENODATA;
 192 }
 193
 194 static int vvp_page_prep_read(const struct lu_env *env,
 195                               const struct cl_page_slice *slice,
 196                               struct cl_io *_)
 197 {
 198         ENTRY;
 199         /* Skip the page already marked as PG_uptodate. */
 200         RETURN(PageUptodate(cl2vm_page(slice)) ? -EALREADY : 0);
 201 }
 202
 203 static int vvp_page_prep_write(const struct lu_env *env,
 204                                const struct cl_page_slice *slice,
 205                                struct cl_io *_)
 206 {
 207         cfs_page_t *vmpage = cl2vm_page(slice);
 208         int result;
 209
 210         if (clear_page_dirty_for_io(vmpage)) {
 211                 set_page_writeback(vmpage);
 212                 vvp_write_pending(cl2ccc(slice->cpl_obj), cl2ccc_page(slice));
 213                 result = 0;
 214         } else
 215                 result = -EALREADY;
 216         return result;
 217 }
 218
 219 /**
 220  * Handles page transfer errors at VM level.
 221  *
 222  * This takes inode as a separate argument, because inode on which error is to
 223  * be set can be different from \a vmpage inode in case of direct-io.
 224  */
 225 static void vvp_vmpage_error(struct inode *inode, cfs_page_t *vmpage, int ioret)
 226 {
 227         if (ioret == 0)
 228                 ClearPageError(vmpage);
 229         else if (ioret != -EINTR) {
 230                 SetPageError(vmpage);
 231                 if (ioret == -ENOSPC)
 232                         set_bit(AS_ENOSPC, &inode->i_mapping->flags);
 233                 else
 234                         set_bit(AS_EIO, &inode->i_mapping->flags);
 235         }
 236 }
 237
 238 static void vvp_page_completion_common(const struct lu_env *env,
 239                                        struct ccc_page *cp, int ioret)
 240 {
 241         struct cl_page    *clp    = cp->cpg_cl.cpl_page;
 242         cfs_page_t        *vmpage = cp->cpg_page;
 243         struct inode      *inode  = ccc_object_inode(clp->cp_obj);
 244         struct cl_sync_io *anchor = cp->cpg_sync_io;
 245
 246         LINVRNT(cl_page_is_vmlocked(env, clp));
 247
 248         if (anchor != NULL) {
 249                 cp->cpg_sync_io  = NULL;
 250                 cl_sync_io_note(anchor, ioret);
 251         } else if (clp->cp_type == CPT_CACHEABLE) {
 252                 /*
 253                  * Only mark the page error only when it's a cacheable page
 254                  * and NOT a sync io.
 255                  *
 256                  * For sync IO and direct IO(CPT_TRANSIENT), the error is able
 257                  * to be seen by application, so we don't need to mark a page
 258                  * as error at all.
 259                  */
 260                 vvp_vmpage_error(inode, vmpage, ioret);
 261                 unlock_page(vmpage);
 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 cl_page  *page  = cl_page_top(slice->cpl_page);
 271         struct inode    *inode = ccc_object_inode(page->cp_obj);
 272         ENTRY;
 273
 274         CL_PAGE_HEADER(D_PAGE, env, page, "completing READ with %d\n", ioret);
 275
 276         if (cp->cpg_defer_uptodate)
 277                 ll_ra_count_put(ll_i2sbi(inode), 1);
 278
 279         if (ioret == 0)  {
 280                 /* XXX: do we need this for transient pages? */
 281                 if (!cp->cpg_defer_uptodate)
 282                         cl_page_export(env, page);
 283         } else
 284                 cp->cpg_defer_uptodate = 0;
 285         vvp_page_completion_common(env, cp, ioret);
 286
 287         EXIT;
 288 }
 289
 290 static void vvp_page_completion_write_common(const struct lu_env *env,
 291                                              const struct cl_page_slice *slice,
 292                                              int ioret)
 293 {
 294         struct ccc_page *cp = cl2ccc_page(slice);
 295
 296         /*
 297          * TODO: Actually it makes sense to add the page into oap pending
 298          * list again and so that we don't need to take the page out from
 299          * SoM write pending list, if we just meet a recoverable error,
 300          * -ENOMEM, etc.
 301          * To implement this, we just need to return a non zero value in
 302          * ->cpo_completion method. The underlying transfer should be notified
 303          * and then re-add the page into pending transfer queue.  -jay
 304          */
 305         cp->cpg_write_queued = 0;
 306         vvp_write_complete(cl2ccc(slice->cpl_obj), cp);
 307
 308         vvp_page_completion_common(env, cp, ioret);
 309 }
 310
 311 static void vvp_page_completion_write(const struct lu_env *env,
 312                                       const struct cl_page_slice *slice,
 313                                       int ioret)
 314 {
 315         struct ccc_page *cp     = cl2ccc_page(slice);
 316         struct cl_page  *pg     = slice->cpl_page;
 317         cfs_page_t      *vmpage = cp->cpg_page;
 318
 319         ENTRY;
 320
 321         LINVRNT(cl_page_is_vmlocked(env, pg));
 322         LASSERT(PageWriteback(vmpage));
 323
 324         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 325
 326         vvp_page_completion_write_common(env, slice, ioret);
 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         cfs_page_t *vmpage = cl2vm_page(slice);
 349         struct cl_page *pg = slice->cpl_page;
 350         int result;
 351
 352         result = -EAGAIN;
 353         /* we're trying to write, but the page is locked.. come back later */
 354         if (!TestSetPageLocked(vmpage)) {
 355                 if (pg->cp_state == CPS_CACHED) {
 356                         /*
 357                          * We can cancel IO if page wasn't dirty after all.
 358                          */
 359                         clear_page_dirty_for_io(vmpage);
 360                         /*
 361                          * This actually clears the dirty bit in the radix
 362                          * tree.
 363                          */
 364                         set_page_writeback(vmpage);
 365                         vvp_write_pending(cl2ccc(slice->cpl_obj),
 366                                           cl2ccc_page(slice));
 367                         CL_PAGE_HEADER(D_PAGE, env, pg, "readied\n");
 368                         result = 0;
 369                 } else
 370                         /*
 371                          * Page was concurrently truncated.
 372                          */
 373                         LASSERT(pg->cp_state == CPS_FREEING);
 374         }
 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         cfs_page_t      *vmpage = vp->cpg_page;
 384
 385         (*printer)(env, cookie, LUSTRE_VVP_NAME"-page@%p(%d:%d:%d) "
 386                    "vm@%p ",
 387                    vp, vp->cpg_defer_uptodate, vp->cpg_ra_used,
 388                    vp->cpg_write_queued, vmpage);
 389         if (vmpage != NULL) {
 390                 (*printer)(env, cookie, "%lx %d:%d %lx %lu %slru",
 391                            (long)vmpage->flags, page_count(vmpage),
 392                            page_mapcount(vmpage), vmpage->private,
 393                            page_index(vmpage),
 394                            list_empty(&vmpage->lru) ? "not-" : "");
 395         }
 396         (*printer)(env, cookie, "\n");
 397         return 0;
 398 }
 399
 400 static const struct cl_page_operations vvp_page_ops = {
 401         .cpo_own           = vvp_page_own,
 402         .cpo_assume        = vvp_page_assume,
 403         .cpo_unassume      = vvp_page_unassume,
 404         .cpo_disown        = vvp_page_disown,
 405         .cpo_vmpage        = ccc_page_vmpage,
 406         .cpo_discard       = vvp_page_discard,
 407         .cpo_delete        = vvp_page_delete,
 408         .cpo_unmap         = vvp_page_unmap,
 409         .cpo_export        = vvp_page_export,
 410         .cpo_is_vmlocked   = vvp_page_is_vmlocked,
 411         .cpo_fini          = vvp_page_fini,
 412         .cpo_print         = vvp_page_print,
 413         .cpo_is_under_lock = ccc_page_is_under_lock,
 414         .io = {
 415                 [CRT_READ] = {
 416                         .cpo_prep        = vvp_page_prep_read,
 417                         .cpo_completion  = vvp_page_completion_read,
 418                         .cpo_make_ready  = ccc_fail,
 419                 },
 420                 [CRT_WRITE] = {
 421                         .cpo_prep        = vvp_page_prep_write,
 422                         .cpo_completion  = vvp_page_completion_write,
 423                         .cpo_make_ready  = vvp_page_make_ready,
 424                 }
 425         }
 426 };
 427
 428 static void vvp_transient_page_verify(const struct cl_page *page)
 429 {
 430         struct inode *inode = ccc_object_inode(page->cp_obj);
 431
 432         LASSERT(!TRYLOCK_INODE_MUTEX(inode));
 433         /* LASSERT_SEM_LOCKED(&inode->i_alloc_sem); */
 434 }
 435
 436 static void vvp_transient_page_own(const struct lu_env *env,
 437                                    const struct cl_page_slice *slice,
 438                                    struct cl_io *_)
 439 {
 440         vvp_transient_page_verify(slice->cpl_page);
 441 }
 442
 443 static void vvp_transient_page_assume(const struct lu_env *env,
 444                                       const struct cl_page_slice *slice,
 445                                       struct cl_io *_)
 446 {
 447         vvp_transient_page_verify(slice->cpl_page);
 448 }
 449
 450 static void vvp_transient_page_unassume(const struct lu_env *env,
 451                                         const struct cl_page_slice *slice,
 452                                         struct cl_io *_)
 453 {
 454         vvp_transient_page_verify(slice->cpl_page);
 455 }
 456
 457 static void vvp_transient_page_disown(const struct lu_env *env,
 458                                       const struct cl_page_slice *slice,
 459                                       struct cl_io *_)
 460 {
 461         vvp_transient_page_verify(slice->cpl_page);
 462 }
 463
 464 static void vvp_transient_page_discard(const struct lu_env *env,
 465                                        const struct cl_page_slice *slice,
 466                                        struct cl_io *_)
 467 {
 468         struct cl_page *page = slice->cpl_page;
 469
 470         vvp_transient_page_verify(slice->cpl_page);
 471
 472         /*
 473          * For transient pages, remove it from the radix tree.
 474          */
 475         cl_page_delete(env, page);
 476 }
 477
 478 static int vvp_transient_page_is_vmlocked(const struct lu_env *env,
 479                                           const struct cl_page_slice *slice)
 480 {
 481         struct inode    *inode = ccc_object_inode(slice->cpl_obj);
 482         int              locked;
 483
 484         locked = !TRYLOCK_INODE_MUTEX(inode);
 485         if (!locked)
 486                 UNLOCK_INODE_MUTEX(inode);
 487         return locked ? -EBUSY : -ENODATA;
 488 }
 489
 490 static void
 491 vvp_transient_page_completion_write(const struct lu_env *env,
 492                                     const struct cl_page_slice *slice,
 493                                     int ioret)
 494 {
 495         vvp_transient_page_verify(slice->cpl_page);
 496         vvp_page_completion_write_common(env, slice, ioret);
 497 }
 498
 499
 500 static void vvp_transient_page_fini(const struct lu_env *env,
 501                                     struct cl_page_slice *slice)
 502 {
 503         struct ccc_page *cp = cl2ccc_page(slice);
 504         struct cl_page *clp = slice->cpl_page;
 505         struct ccc_object *clobj = cl2ccc(clp->cp_obj);
 506
 507         vvp_page_fini_common(cp);
 508         LASSERT(!TRYLOCK_INODE_MUTEX(clobj->cob_inode));
 509         clobj->cob_transient_pages--;
 510 }
 511
 512 static const struct cl_page_operations vvp_transient_page_ops = {
 513         .cpo_own           = vvp_transient_page_own,
 514         .cpo_assume        = vvp_transient_page_assume,
 515         .cpo_unassume      = vvp_transient_page_unassume,
 516         .cpo_disown        = vvp_transient_page_disown,
 517         .cpo_discard       = vvp_transient_page_discard,
 518         .cpo_vmpage        = ccc_page_vmpage,
 519         .cpo_fini          = vvp_transient_page_fini,
 520         .cpo_is_vmlocked   = vvp_transient_page_is_vmlocked,
 521         .cpo_print         = vvp_page_print,
 522         .cpo_is_under_lock = ccc_page_is_under_lock,
 523         .io = {
 524                 [CRT_READ] = {
 525                         .cpo_prep        = ccc_transient_page_prep,
 526                         .cpo_completion  = vvp_page_completion_read,
 527                 },
 528                 [CRT_WRITE] = {
 529                         .cpo_prep        = ccc_transient_page_prep,
 530                         .cpo_completion  = vvp_transient_page_completion_write,
 531                 }
 532         }
 533 };
 534
 535 struct cl_page *vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 536                               struct cl_page *page, cfs_page_t *vmpage)
 537 {
 538         struct ccc_page *cpg;
 539         int result;
 540
 541         CLOBINVRNT(env, obj, ccc_object_invariant(obj));
 542
 543         OBD_SLAB_ALLOC_PTR_GFP(cpg, vvp_page_kmem, CFS_ALLOC_IO);
 544         if (cpg != NULL) {
 545                 cpg->cpg_page = vmpage;
 546                 page_cache_get(vmpage);
 547
 548                 CFS_INIT_LIST_HEAD(&cpg->cpg_pending_linkage);
 549                 if (page->cp_type == CPT_CACHEABLE) {
 550                         SetPagePrivate(vmpage);
 551                         vmpage->private = (unsigned long)page;
 552                         cl_page_slice_add(page, &cpg->cpg_cl, obj,
 553                                           &vvp_page_ops);
 554                 } else {
 555                         struct ccc_object *clobj = cl2ccc(obj);
 556
 557                         LASSERT(!TRYLOCK_INODE_MUTEX(clobj->cob_inode));
 558                         cl_page_slice_add(page, &cpg->cpg_cl, obj,
 559                                           &vvp_transient_page_ops);
 560                         clobj->cob_transient_pages++;
 561                 }
 562                 result = 0;
 563         } else
 564                 result = -ENOMEM;
 565         return ERR_PTR(result);
 566 }
 567