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, 2015, 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);
  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);
  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);
 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);
 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);
 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 ccc_page      *cpg     = cl2ccc_page(slice);
 140
 141         LASSERT(vmpage);
 142         LASSERT(PageLocked(vmpage));
 143
 144         if (cpg->cpg_defer_uptodate && !cpg->cpg_ra_used)
 145                 ll_ra_stats_inc(vmpage->mapping, RA_STAT_DISCARDED);
 146
 147         ll_invalidate_page(vmpage);
 148 }
 149
 150 static void vvp_page_delete(const struct lu_env *env,
 151                             const struct cl_page_slice *slice)
 152 {
 153         struct page       *vmpage = cl2vm_page(slice);
 154         struct inode     *inode  = vmpage->mapping->host;
 155         struct cl_object *obj    = slice->cpl_obj;
 156         struct cl_page   *page   = slice->cpl_page;
 157         int refc;
 158
 159         LASSERT(PageLocked(vmpage));
 160         LASSERT((struct cl_page *)vmpage->private == page);
 161         LASSERT(inode == ccc_object_inode(obj));
 162
 163         vvp_write_complete(cl2ccc(obj), cl2ccc_page(slice));
 164
 165         /* Drop the reference count held in vvp_page_init */
 166         refc = atomic_dec_return(&page->cp_ref);
 167         LASSERTF(refc >= 1, "page = %p, refc = %d\n", page, refc);
 168
 169         ClearPageUptodate(vmpage);
 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                             int uptodate)
 181 {
 182         struct page *vmpage = cl2vm_page(slice);
 183
 184         LASSERT(vmpage);
 185         LASSERT(PageLocked(vmpage));
 186         if (uptodate)
 187                 SetPageUptodate(vmpage);
 188         else
 189                 ClearPageUptodate(vmpage);
 190 }
 191
 192 static int vvp_page_is_vmlocked(const struct lu_env *env,
 193                                 const struct cl_page_slice *slice)
 194 {
 195         return PageLocked(cl2vm_page(slice)) ? -EBUSY : -ENODATA;
 196 }
 197
 198 static int vvp_page_prep_read(const struct lu_env *env,
 199                               const struct cl_page_slice *slice,
 200                               struct cl_io *unused)
 201 {
 202         /* Skip the page already marked as PG_uptodate. */
 203         return PageUptodate(cl2vm_page(slice)) ? -EALREADY : 0;
 204 }
 205
 206 static int vvp_page_prep_write(const struct lu_env *env,
 207                                const struct cl_page_slice *slice,
 208                                struct cl_io *unused)
 209 {
 210         struct page *vmpage = cl2vm_page(slice);
 211         struct cl_page *pg = slice->cpl_page;
 212
 213         LASSERT(PageLocked(vmpage));
 214         LASSERT(!PageDirty(vmpage));
 215
 216         /* ll_writepage path is not a sync write, so need to set page writeback
 217          * flag
 218          */
 219         if (!pg->cp_sync_io)
 220                 set_page_writeback(vmpage);
 221
 222         vvp_write_pending(cl2ccc(slice->cpl_obj), cl2ccc_page(slice));
 223
 224         return 0;
 225 }
 226
 227 /**
 228  * Handles page transfer errors at VM level.
 229  *
 230  * This takes inode as a separate argument, because inode on which error is to
 231  * be set can be different from \a vmpage inode in case of direct-io.
 232  */
 233 static void vvp_vmpage_error(struct inode *inode, struct page *vmpage, int ioret)
 234 {
 235         struct ccc_object *obj = cl_inode2ccc(inode);
 236
 237         if (ioret == 0) {
 238                 ClearPageError(vmpage);
 239                 obj->cob_discard_page_warned = 0;
 240         } else {
 241                 SetPageError(vmpage);
 242                 if (ioret == -ENOSPC)
 243                         set_bit(AS_ENOSPC, &inode->i_mapping->flags);
 244                 else
 245                         set_bit(AS_EIO, &inode->i_mapping->flags);
 246
 247                 if ((ioret == -ESHUTDOWN || ioret == -EINTR) &&
 248                     obj->cob_discard_page_warned == 0) {
 249                         obj->cob_discard_page_warned = 1;
 250                         ll_dirty_page_discard_warn(vmpage, ioret);
 251                 }
 252         }
 253 }
 254
 255 static void vvp_page_completion_read(const struct lu_env *env,
 256                                      const struct cl_page_slice *slice,
 257                                      int ioret)
 258 {
 259         struct ccc_page *cp     = cl2ccc_page(slice);
 260         struct page      *vmpage = cp->cpg_page;
 261         struct cl_page  *page   = slice->cpl_page;
 262         struct inode    *inode  = ccc_object_inode(page->cp_obj);
 263
 264         LASSERT(PageLocked(vmpage));
 265         CL_PAGE_HEADER(D_PAGE, env, page, "completing READ with %d\n", ioret);
 266
 267         if (cp->cpg_defer_uptodate)
 268                 ll_ra_count_put(ll_i2sbi(inode), 1);
 269
 270         if (ioret == 0)  {
 271                 if (!cp->cpg_defer_uptodate)
 272                         cl_page_export(env, page, 1);
 273         } else
 274                 cp->cpg_defer_uptodate = 0;
 275
 276         if (!page->cp_sync_io)
 277                 unlock_page(vmpage);
 278 }
 279
 280 static void vvp_page_completion_write(const struct lu_env *env,
 281                                       const struct cl_page_slice *slice,
 282                                       int ioret)
 283 {
 284         struct ccc_page *cp     = cl2ccc_page(slice);
 285         struct cl_page  *pg     = slice->cpl_page;
 286         struct page      *vmpage = cp->cpg_page;
 287
 288         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 289
 290         /*
 291          * TODO: Actually it makes sense to add the page into oap pending
 292          * list again and so that we don't need to take the page out from
 293          * SoM write pending list, if we just meet a recoverable error,
 294          * -ENOMEM, etc.
 295          * To implement this, we just need to return a non zero value in
 296          * ->cpo_completion method. The underlying transfer should be notified
 297          * and then re-add the page into pending transfer queue.  -jay
 298          */
 299
 300         cp->cpg_write_queued = 0;
 301         vvp_write_complete(cl2ccc(slice->cpl_obj), cp);
 302
 303         if (pg->cp_sync_io) {
 304                 LASSERT(PageLocked(vmpage));
 305                 LASSERT(!PageWriteback(vmpage));
 306         } else {
 307                 LASSERT(PageWriteback(vmpage));
 308                 /*
 309                  * Only mark the page error only when it's an async write
 310                  * because applications won't wait for IO to finish.
 311                  */
 312                 vvp_vmpage_error(ccc_object_inode(pg->cp_obj), vmpage, ioret);
 313
 314                 end_page_writeback(vmpage);
 315         }
 316 }
 317
 318 /**
 319  * Implements cl_page_operations::cpo_make_ready() method.
 320  *
 321  * This is called to yank a page from the transfer cache and to send it out as
 322  * a part of transfer. This function try-locks the page. If try-lock failed,
 323  * page is owned by some concurrent IO, and should be skipped (this is bad,
 324  * but hopefully rare situation, as it usually results in transfer being
 325  * shorter than possible).
 326  *
 327  * \retval 0      success, page can be placed into transfer
 328  *
 329  * \retval -EAGAIN page is either used by concurrent IO has been
 330  * truncated. Skip it.
 331  */
 332 static int vvp_page_make_ready(const struct lu_env *env,
 333                                const struct cl_page_slice *slice)
 334 {
 335         struct page *vmpage = cl2vm_page(slice);
 336         struct cl_page *pg = slice->cpl_page;
 337         int result = 0;
 338
 339         lock_page(vmpage);
 340         if (clear_page_dirty_for_io(vmpage)) {
 341                 LASSERT(pg->cp_state == CPS_CACHED);
 342                 /* This actually clears the dirty bit in the radix tree. */
 343                 set_page_writeback(vmpage);
 344                 vvp_write_pending(cl2ccc(slice->cpl_obj), cl2ccc_page(slice));
 345                 CL_PAGE_HEADER(D_PAGE, env, pg, "readied\n");
 346         } else if (pg->cp_state == CPS_PAGEOUT) {
 347                 /* is it possible for osc_flush_async_page() to already
 348                  * make it ready?
 349                  */
 350                 result = -EALREADY;
 351         } else {
 352                 CL_PAGE_DEBUG(D_ERROR, env, pg, "Unexpecting page state %d.\n",
 353                               pg->cp_state);
 354                 LBUG();
 355         }
 356         unlock_page(vmpage);
 357         return result;
 358 }
 359
 360 static int vvp_page_print(const struct lu_env *env,
 361                           const struct cl_page_slice *slice,
 362                           void *cookie, lu_printer_t printer)
 363 {
 364         struct ccc_page *vp = cl2ccc_page(slice);
 365         struct page      *vmpage = vp->cpg_page;
 366
 367         (*printer)(env, cookie, LUSTRE_VVP_NAME "-page@%p(%d:%d:%d) vm@%p ",
 368                    vp, vp->cpg_defer_uptodate, vp->cpg_ra_used,
 369                    vp->cpg_write_queued, vmpage);
 370         if (vmpage) {
 371                 (*printer)(env, cookie, "%lx %d:%d %lx %lu %slru",
 372                            (long)vmpage->flags, page_count(vmpage),
 373                            page_mapcount(vmpage), vmpage->private,
 374                            page_index(vmpage),
 375                            list_empty(&vmpage->lru) ? "not-" : "");
 376         }
 377         (*printer)(env, cookie, "\n");
 378         return 0;
 379 }
 380
 381 static const struct cl_page_operations vvp_page_ops = {
 382         .cpo_own           = vvp_page_own,
 383         .cpo_assume     = vvp_page_assume,
 384         .cpo_unassume      = vvp_page_unassume,
 385         .cpo_disown     = vvp_page_disown,
 386         .cpo_discard       = vvp_page_discard,
 387         .cpo_delete     = vvp_page_delete,
 388         .cpo_export     = vvp_page_export,
 389         .cpo_is_vmlocked   = vvp_page_is_vmlocked,
 390         .cpo_fini         = vvp_page_fini,
 391         .cpo_print       = vvp_page_print,
 392         .cpo_is_under_lock = ccc_page_is_under_lock,
 393         .io = {
 394                 [CRT_READ] = {
 395                         .cpo_prep       = vvp_page_prep_read,
 396                         .cpo_completion  = vvp_page_completion_read,
 397                         .cpo_make_ready  = ccc_fail,
 398                 },
 399                 [CRT_WRITE] = {
 400                         .cpo_prep       = vvp_page_prep_write,
 401                         .cpo_completion  = vvp_page_completion_write,
 402                         .cpo_make_ready  = vvp_page_make_ready,
 403                 }
 404         }
 405 };
 406
 407 static void vvp_transient_page_verify(const struct cl_page *page)
 408 {
 409         struct inode *inode = ccc_object_inode(page->cp_obj);
 410
 411         LASSERT(!inode_trylock(inode));
 412 }
 413
 414 static int vvp_transient_page_own(const struct lu_env *env,
 415                                   const struct cl_page_slice *slice,
 416                                   struct cl_io *unused, int nonblock)
 417 {
 418         vvp_transient_page_verify(slice->cpl_page);
 419         return 0;
 420 }
 421
 422 static void vvp_transient_page_assume(const struct lu_env *env,
 423                                       const struct cl_page_slice *slice,
 424                                       struct cl_io *unused)
 425 {
 426         vvp_transient_page_verify(slice->cpl_page);
 427 }
 428
 429 static void vvp_transient_page_unassume(const struct lu_env *env,
 430                                         const struct cl_page_slice *slice,
 431                                         struct cl_io *unused)
 432 {
 433         vvp_transient_page_verify(slice->cpl_page);
 434 }
 435
 436 static void vvp_transient_page_disown(const struct lu_env *env,
 437                                       const struct cl_page_slice *slice,
 438                                       struct cl_io *unused)
 439 {
 440         vvp_transient_page_verify(slice->cpl_page);
 441 }
 442
 443 static void vvp_transient_page_discard(const struct lu_env *env,
 444                                        const struct cl_page_slice *slice,
 445                                        struct cl_io *unused)
 446 {
 447         struct cl_page *page = slice->cpl_page;
 448
 449         vvp_transient_page_verify(slice->cpl_page);
 450
 451         /*
 452          * For transient pages, remove it from the radix tree.
 453          */
 454         cl_page_delete(env, page);
 455 }
 456
 457 static int vvp_transient_page_is_vmlocked(const struct lu_env *env,
 458                                           const struct cl_page_slice *slice)
 459 {
 460         struct inode    *inode = ccc_object_inode(slice->cpl_obj);
 461         int     locked;
 462
 463         locked = !inode_trylock(inode);
 464         if (!locked)
 465                 inode_unlock(inode);
 466         return locked ? -EBUSY : -ENODATA;
 467 }
 468
 469 static void
 470 vvp_transient_page_completion(const struct lu_env *env,
 471                               const struct cl_page_slice *slice,
 472                               int ioret)
 473 {
 474         vvp_transient_page_verify(slice->cpl_page);
 475 }
 476
 477 static void vvp_transient_page_fini(const struct lu_env *env,
 478                                     struct cl_page_slice *slice)
 479 {
 480         struct ccc_page *cp = cl2ccc_page(slice);
 481         struct cl_page *clp = slice->cpl_page;
 482         struct ccc_object *clobj = cl2ccc(clp->cp_obj);
 483
 484         vvp_page_fini_common(cp);
 485         LASSERT(!inode_trylock(clobj->cob_inode));
 486         clobj->cob_transient_pages--;
 487 }
 488
 489 static const struct cl_page_operations vvp_transient_page_ops = {
 490         .cpo_own           = vvp_transient_page_own,
 491         .cpo_assume     = vvp_transient_page_assume,
 492         .cpo_unassume      = vvp_transient_page_unassume,
 493         .cpo_disown     = vvp_transient_page_disown,
 494         .cpo_discard       = vvp_transient_page_discard,
 495         .cpo_fini         = vvp_transient_page_fini,
 496         .cpo_is_vmlocked   = vvp_transient_page_is_vmlocked,
 497         .cpo_print       = vvp_page_print,
 498         .cpo_is_under_lock = ccc_page_is_under_lock,
 499         .io = {
 500                 [CRT_READ] = {
 501                         .cpo_prep       = ccc_transient_page_prep,
 502                         .cpo_completion  = vvp_transient_page_completion,
 503                 },
 504                 [CRT_WRITE] = {
 505                         .cpo_prep       = ccc_transient_page_prep,
 506                         .cpo_completion  = vvp_transient_page_completion,
 507                 }
 508         }
 509 };
 510
 511 int vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 512                 struct cl_page *page, pgoff_t index)
 513 {
 514         struct ccc_page *cpg = cl_object_page_slice(obj, page);
 515         struct page     *vmpage = page->cp_vmpage;
 516
 517         CLOBINVRNT(env, obj, ccc_object_invariant(obj));
 518
 519         cpg->cpg_cl.cpl_index = index;
 520         cpg->cpg_page = vmpage;
 521         page_cache_get(vmpage);
 522
 523         INIT_LIST_HEAD(&cpg->cpg_pending_linkage);
 524         if (page->cp_type == CPT_CACHEABLE) {
 525                 /* in cache, decref in vvp_page_delete */
 526                 atomic_inc(&page->cp_ref);
 527                 SetPagePrivate(vmpage);
 528                 vmpage->private = (unsigned long)page;
 529                 cl_page_slice_add(page, &cpg->cpg_cl, obj, &vvp_page_ops);
 530         } else {
 531                 struct ccc_object *clobj = cl2ccc(obj);
 532
 533                 LASSERT(!inode_trylock(clobj->cob_inode));
 534                 cl_page_slice_add(page, &cpg->cpg_cl, obj,
 535                                   &vvp_transient_page_ops);
 536                 clobj->cob_transient_pages++;
 537         }
 538         return 0;
 539 }