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                 result = 0;
 213         } else
 214                 result = -EALREADY;
 215         return result;
 216 }
 217
 218 /**
 219  * Handles page transfer errors at VM level.
 220  *
 221  * This takes inode as a separate argument, because inode on which error is to
 222  * be set can be different from \a vmpage inode in case of direct-io.
 223  */
 224 static void vvp_vmpage_error(struct inode *inode, cfs_page_t *vmpage, int ioret)
 225 {
 226         if (ioret == 0)
 227                 ClearPageError(vmpage);
 228         else if (ioret != -EINTR) {
 229                 SetPageError(vmpage);
 230                 if (ioret == -ENOSPC)
 231                         set_bit(AS_ENOSPC, &inode->i_mapping->flags);
 232                 else
 233                         set_bit(AS_EIO, &inode->i_mapping->flags);
 234         }
 235 }
 236
 237 static void vvp_page_completion_common(const struct lu_env *env,
 238                                        struct ccc_page *cp, int ioret)
 239 {
 240         struct cl_page    *clp    = cp->cpg_cl.cpl_page;
 241         cfs_page_t        *vmpage = cp->cpg_page;
 242         struct inode      *inode  = ccc_object_inode(clp->cp_obj);
 243         struct cl_sync_io *anchor = cp->cpg_sync_io;
 244
 245         LINVRNT(cl_page_is_vmlocked(env, clp));
 246
 247         if (anchor != NULL) {
 248                 cp->cpg_sync_io  = NULL;
 249                 cl_sync_io_note(anchor, ioret);
 250         } else if (clp->cp_type == CPT_CACHEABLE) {
 251                 /*
 252                  * Don't assert the page writeback bit here because the lustre
 253                  * file may be as a backend of swap space. in this case, the
 254                  * page writeback is set by VM, and obvious we shouldn't clear
 255                  * it at all. Fortunately this type of pages are all TRANSIENT
 256                  * pages.
 257                  */
 258                 LASSERT(!PageWriteback(vmpage));
 259
 260                 /*
 261                  * Only mark the page error only when it's a cacheable page
 262                  * and NOT a sync io.
 263                  *
 264                  * For sync IO and direct IO(CPT_TRANSIENT), the error is able
 265                  * to be seen by application, so we don't need to mark a page
 266                  * as error at all.
 267                  */
 268                 vvp_vmpage_error(inode, vmpage, ioret);
 269                 unlock_page(vmpage);
 270         }
 271 }
 272
 273 static void vvp_page_completion_read(const struct lu_env *env,
 274                                      const struct cl_page_slice *slice,
 275                                      int ioret)
 276 {
 277         struct ccc_page *cp    = cl2ccc_page(slice);
 278         struct cl_page  *page  = cl_page_top(slice->cpl_page);
 279         struct inode    *inode = ccc_object_inode(page->cp_obj);
 280         ENTRY;
 281
 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                 /* XXX: do we need this for transient pages? */
 289                 if (!cp->cpg_defer_uptodate)
 290                         cl_page_export(env, page);
 291         } else
 292                 cp->cpg_defer_uptodate = 0;
 293         vvp_page_completion_common(env, cp, ioret);
 294
 295         EXIT;
 296 }
 297
 298 static void vvp_page_completion_write_common(const struct lu_env *env,
 299                                              const struct cl_page_slice *slice,
 300                                              int ioret)
 301 {
 302         struct ccc_page *cp = cl2ccc_page(slice);
 303
 304         /*
 305          * TODO: Actually it makes sense to add the page into oap pending
 306          * list again and so that we don't need to take the page out from
 307          * SoM write pending list, if we just meet a recoverable error,
 308          * -ENOMEM, etc.
 309          * To implement this, we just need to return a non zero value in
 310          * ->cpo_completion method. The underlying transfer should be notified
 311          * and then re-add the page into pending transfer queue.  -jay
 312          */
 313         cp->cpg_write_queued = 0;
 314         vvp_write_complete(cl2ccc(slice->cpl_obj), cp);
 315
 316         vvp_page_completion_common(env, cp, ioret);
 317 }
 318
 319 static void vvp_page_completion_write(const struct lu_env *env,
 320                                       const struct cl_page_slice *slice,
 321                                       int ioret)
 322 {
 323         struct ccc_page *cp     = cl2ccc_page(slice);
 324         struct cl_page  *pg     = slice->cpl_page;
 325         cfs_page_t      *vmpage = cp->cpg_page;
 326
 327         ENTRY;
 328
 329         LINVRNT(cl_page_is_vmlocked(env, pg));
 330         LASSERT(PageWriteback(vmpage));
 331
 332         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 333
 334         end_page_writeback(vmpage);
 335         LASSERT(!PageWriteback(vmpage));
 336
 337         vvp_page_completion_write_common(env, slice, ioret);
 338         EXIT;
 339 }
 340
 341 /**
 342  * Implements cl_page_operations::cpo_make_ready() method.
 343  *
 344  * This is called to yank a page from the transfer cache and to send it out as
 345  * a part of transfer. This function try-locks the page. If try-lock failed,
 346  * page is owned by some concurrent IO, and should be skipped (this is bad,
 347  * but hopefully rare situation, as it usually results in transfer being
 348  * shorter than possible).
 349  *
 350  * \retval 0      success, page can be placed into transfer
 351  *
 352  * \retval -EAGAIN page is either used by concurrent IO has been
 353  * truncated. Skip it.
 354  */
 355 static int vvp_page_make_ready(const struct lu_env *env,
 356                                const struct cl_page_slice *slice)
 357 {
 358         cfs_page_t *vmpage = cl2vm_page(slice);
 359         struct cl_page *pg = slice->cpl_page;
 360         int result;
 361
 362         result = -EAGAIN;
 363         /* we're trying to write, but the page is locked.. come back later */
 364         if (!TestSetPageLocked(vmpage)) {
 365                 if (pg->cp_state == CPS_CACHED) {
 366                         /*
 367                          * We can cancel IO if page wasn't dirty after all.
 368                          */
 369                         clear_page_dirty_for_io(vmpage);
 370                         /*
 371                          * This actually clears the dirty bit in the radix
 372                          * tree.
 373                          */
 374                         set_page_writeback(vmpage);
 375
 376                         CL_PAGE_HEADER(D_PAGE, env, pg, "readied\n");
 377                         result = 0;
 378                 } else
 379                         /*
 380                          * Page was concurrently truncated.
 381                          */
 382                         LASSERT(pg->cp_state == CPS_FREEING);
 383         }
 384         RETURN(result);
 385 }
 386
 387 static int vvp_page_print(const struct lu_env *env,
 388                           const struct cl_page_slice *slice,
 389                           void *cookie, lu_printer_t printer)
 390 {
 391         struct ccc_page *vp = cl2ccc_page(slice);
 392         cfs_page_t      *vmpage = vp->cpg_page;
 393
 394         (*printer)(env, cookie, LUSTRE_VVP_NAME"-page@%p(%d:%d:%d) "
 395                    "vm@%p ",
 396                    vp, vp->cpg_defer_uptodate, vp->cpg_ra_used,
 397                    vp->cpg_write_queued, vmpage);
 398         if (vmpage != NULL) {
 399                 (*printer)(env, cookie, "%lx %d:%d %lx %lu %slru",
 400                            (long)vmpage->flags, page_count(vmpage),
 401                            page_mapcount(vmpage), vmpage->private,
 402                            page_index(vmpage),
 403                            list_empty(&vmpage->lru) ? "not-" : "");
 404         }
 405         (*printer)(env, cookie, "\n");
 406         return 0;
 407 }
 408
 409 static const struct cl_page_operations vvp_page_ops = {
 410         .cpo_own           = vvp_page_own,
 411         .cpo_assume        = vvp_page_assume,
 412         .cpo_unassume      = vvp_page_unassume,
 413         .cpo_disown        = vvp_page_disown,
 414         .cpo_vmpage        = ccc_page_vmpage,
 415         .cpo_discard       = vvp_page_discard,
 416         .cpo_delete        = vvp_page_delete,
 417         .cpo_unmap         = vvp_page_unmap,
 418         .cpo_export        = vvp_page_export,
 419         .cpo_is_vmlocked   = vvp_page_is_vmlocked,
 420         .cpo_fini          = vvp_page_fini,
 421         .cpo_print         = vvp_page_print,
 422         .cpo_is_under_lock = ccc_page_is_under_lock,
 423         .io = {
 424                 [CRT_READ] = {
 425                         .cpo_prep        = vvp_page_prep_read,
 426                         .cpo_completion  = vvp_page_completion_read,
 427                         .cpo_make_ready  = ccc_fail,
 428                 },
 429                 [CRT_WRITE] = {
 430                         .cpo_prep        = vvp_page_prep_write,
 431                         .cpo_completion  = vvp_page_completion_write,
 432                         .cpo_make_ready  = vvp_page_make_ready,
 433                 }
 434         }
 435 };
 436
 437 static void vvp_transient_page_verify(const struct cl_page *page)
 438 {
 439         struct inode *inode = ccc_object_inode(page->cp_obj);
 440
 441         LASSERT(!TRYLOCK_INODE_MUTEX(inode));
 442         /* LASSERT_SEM_LOCKED(&inode->i_alloc_sem); */
 443 }
 444
 445 static void vvp_transient_page_own(const struct lu_env *env,
 446                                    const struct cl_page_slice *slice,
 447                                    struct cl_io *_)
 448 {
 449         vvp_transient_page_verify(slice->cpl_page);
 450 }
 451
 452 static void vvp_transient_page_assume(const struct lu_env *env,
 453                                       const struct cl_page_slice *slice,
 454                                       struct cl_io *_)
 455 {
 456         vvp_transient_page_verify(slice->cpl_page);
 457 }
 458
 459 static void vvp_transient_page_unassume(const struct lu_env *env,
 460                                         const struct cl_page_slice *slice,
 461                                         struct cl_io *_)
 462 {
 463         vvp_transient_page_verify(slice->cpl_page);
 464 }
 465
 466 static void vvp_transient_page_disown(const struct lu_env *env,
 467                                       const struct cl_page_slice *slice,
 468                                       struct cl_io *_)
 469 {
 470         vvp_transient_page_verify(slice->cpl_page);
 471 }
 472
 473 static void vvp_transient_page_discard(const struct lu_env *env,
 474                                        const struct cl_page_slice *slice,
 475                                        struct cl_io *_)
 476 {
 477         struct cl_page *page = slice->cpl_page;
 478
 479         vvp_transient_page_verify(slice->cpl_page);
 480
 481         /*
 482          * For transient pages, remove it from the radix tree.
 483          */
 484         cl_page_delete(env, page);
 485 }
 486
 487 static int vvp_transient_page_is_vmlocked(const struct lu_env *env,
 488                                           const struct cl_page_slice *slice)
 489 {
 490         struct inode    *inode = ccc_object_inode(slice->cpl_obj);
 491         int              locked;
 492
 493         locked = !TRYLOCK_INODE_MUTEX(inode);
 494         if (!locked)
 495                 UNLOCK_INODE_MUTEX(inode);
 496         return locked ? -EBUSY : -ENODATA;
 497 }
 498
 499 static void
 500 vvp_transient_page_completion_write(const struct lu_env *env,
 501                                     const struct cl_page_slice *slice,
 502                                     int ioret)
 503 {
 504         vvp_transient_page_verify(slice->cpl_page);
 505         vvp_page_completion_write_common(env, slice, ioret);
 506 }
 507
 508
 509 static void vvp_transient_page_fini(const struct lu_env *env,
 510                                     struct cl_page_slice *slice)
 511 {
 512         struct ccc_page *cp = cl2ccc_page(slice);
 513         struct cl_page *clp = slice->cpl_page;
 514         struct ccc_object *clobj = cl2ccc(clp->cp_obj);
 515
 516         vvp_page_fini_common(cp);
 517         LASSERT(!TRYLOCK_INODE_MUTEX(clobj->cob_inode));
 518         clobj->cob_transient_pages--;
 519 }
 520
 521 static const struct cl_page_operations vvp_transient_page_ops = {
 522         .cpo_own           = vvp_transient_page_own,
 523         .cpo_assume        = vvp_transient_page_assume,
 524         .cpo_unassume      = vvp_transient_page_unassume,
 525         .cpo_disown        = vvp_transient_page_disown,
 526         .cpo_discard       = vvp_transient_page_discard,
 527         .cpo_vmpage        = ccc_page_vmpage,
 528         .cpo_fini          = vvp_transient_page_fini,
 529         .cpo_is_vmlocked   = vvp_transient_page_is_vmlocked,
 530         .cpo_print         = vvp_page_print,
 531         .cpo_is_under_lock = ccc_page_is_under_lock,
 532         .io = {
 533                 [CRT_READ] = {
 534                         .cpo_prep        = ccc_transient_page_prep,
 535                         .cpo_completion  = vvp_page_completion_read,
 536                 },
 537                 [CRT_WRITE] = {
 538                         .cpo_prep        = ccc_transient_page_prep,
 539                         .cpo_completion  = vvp_transient_page_completion_write,
 540                 }
 541         }
 542 };
 543
 544 struct cl_page *vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 545                               struct cl_page *page, cfs_page_t *vmpage)
 546 {
 547         struct ccc_page *cpg;
 548         int result;
 549
 550         CLOBINVRNT(env, obj, ccc_object_invariant(obj));
 551
 552         OBD_SLAB_ALLOC_PTR_GFP(cpg, vvp_page_kmem, CFS_ALLOC_IO);
 553         if (cpg != NULL) {
 554                 cpg->cpg_page = vmpage;
 555                 page_cache_get(vmpage);
 556
 557                 CFS_INIT_LIST_HEAD(&cpg->cpg_pending_linkage);
 558                 if (page->cp_type == CPT_CACHEABLE) {
 559                         SetPagePrivate(vmpage);
 560                         vmpage->private = (unsigned long)page;
 561                         cl_page_slice_add(page, &cpg->cpg_cl, obj,
 562                                           &vvp_page_ops);
 563                 } else {
 564                         struct ccc_object *clobj = cl2ccc(obj);
 565
 566                         LASSERT(!TRYLOCK_INODE_MUTEX(clobj->cob_inode));
 567                         cl_page_slice_add(page, &cpg->cpg_cl, obj,
 568                                           &vvp_transient_page_ops);
 569                         clobj->cob_transient_pages++;
 570                 }
 571                 result = 0;
 572         } else
 573                 result = -ENOMEM;
 574         return ERR_PTR(result);
 575 }
 576