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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * Implementation of cl_page for VVP layer.
  33  *
  34  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  35  *   Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_LLITE
  39
  40 #include <linux/atomic.h>
  41 #include <linux/bitops.h>
  42 #include <linux/mm.h>
  43 #include <linux/mutex.h>
  44 #include <linux/page-flags.h>
  45 #include <linux/pagemap.h>
  46
  47 #include <libcfs/libcfs.h>
  48 #include "llite_internal.h"
  49 #include "vvp_internal.h"
  50
  51 /*****************************************************************************
  52  *
  53  * Page operations.
  54  *
  55  */
  56
  57 static void vvp_page_fini_common(struct vvp_page *vpg, struct pagevec *pvec)
  58 {
  59         struct page *vmpage = vpg->vpg_page;
  60
  61         LASSERT(vmpage != NULL);
  62         if (pvec) {
  63                 if (!pagevec_add(pvec, vmpage))
  64                         pagevec_release(pvec);
  65         } else {
  66                 put_page(vmpage);
  67         }
  68 }
  69
  70 static void vvp_page_fini(const struct lu_env *env,
  71                           struct cl_page_slice *slice,
  72                           struct pagevec *pvec)
  73 {
  74         struct vvp_page *vpg     = cl2vvp_page(slice);
  75         struct page     *vmpage  = vpg->vpg_page;
  76
  77         /*
  78          * vmpage->private was already cleared when page was moved into
  79          * VPG_FREEING state.
  80          */
  81         LASSERT((struct cl_page *)vmpage->private != slice->cpl_page);
  82         vvp_page_fini_common(vpg, pvec);
  83 }
  84
  85 static int vvp_page_own(const struct lu_env *env,
  86                         const struct cl_page_slice *slice, struct cl_io *io,
  87                         int nonblock)
  88 {
  89         struct vvp_page *vpg    = cl2vvp_page(slice);
  90         struct page     *vmpage = vpg->vpg_page;
  91
  92         ENTRY;
  93
  94         LASSERT(vmpage != NULL);
  95         if (nonblock) {
  96                 if (!trylock_page(vmpage))
  97                         return -EAGAIN;
  98
  99                 if (unlikely(PageWriteback(vmpage))) {
 100                         unlock_page(vmpage);
 101                         return -EAGAIN;
 102                 }
 103
 104                 return 0;
 105         }
 106
 107         lock_page(vmpage);
 108         wait_on_page_writeback(vmpage);
 109
 110         RETURN(0);
 111 }
 112
 113 static void vvp_page_assume(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 != NULL);
 120         LASSERT(PageLocked(vmpage));
 121         wait_on_page_writeback(vmpage);
 122 }
 123
 124 static void vvp_page_unassume(const struct lu_env *env,
 125                               const struct cl_page_slice *slice,
 126                               struct cl_io *unused)
 127 {
 128         struct page *vmpage = cl2vm_page(slice);
 129
 130         LASSERT(vmpage != NULL);
 131         LASSERT(PageLocked(vmpage));
 132 }
 133
 134 static void vvp_page_disown(const struct lu_env *env,
 135                             const struct cl_page_slice *slice, struct cl_io *io)
 136 {
 137         struct page *vmpage = cl2vm_page(slice);
 138
 139         ENTRY;
 140
 141         LASSERT(vmpage != NULL);
 142         LASSERT(PageLocked(vmpage));
 143
 144         unlock_page(cl2vm_page(slice));
 145
 146         EXIT;
 147 }
 148
 149 static void vvp_page_discard(const struct lu_env *env,
 150                              const struct cl_page_slice *slice,
 151                              struct cl_io *unused)
 152 {
 153         struct page     *vmpage = cl2vm_page(slice);
 154         struct vvp_page *vpg    = cl2vvp_page(slice);
 155
 156         LASSERT(vmpage != NULL);
 157         LASSERT(PageLocked(vmpage));
 158
 159         if (vpg->vpg_defer_uptodate && !vpg->vpg_ra_used)
 160                 ll_ra_stats_inc(vmpage->mapping->host, RA_STAT_DISCARDED);
 161
 162         generic_error_remove_page(vmpage->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 == vvp_object_inode(obj));
 177
 178         /* Drop the reference count held in vvp_page_init */
 179         refc = atomic_dec_return(&page->cp_ref);
 180         LASSERTF(refc >= 1, "page = %p, refc = %d\n", page, refc);
 181
 182         ClearPagePrivate(vmpage);
 183         vmpage->private = 0;
 184         /*
 185          * Reference from vmpage to cl_page is removed, but the reference back
 186          * is still here. It is removed later in vvp_page_fini().
 187          */
 188 }
 189
 190 static void vvp_page_export(const struct lu_env *env,
 191                             const struct cl_page_slice *slice,
 192                             int uptodate)
 193 {
 194         struct page *vmpage = cl2vm_page(slice);
 195
 196         LASSERT(vmpage != NULL);
 197         LASSERT(PageLocked(vmpage));
 198         if (uptodate)
 199                 SetPageUptodate(vmpage);
 200         else
 201                 ClearPageUptodate(vmpage);
 202 }
 203
 204 static int vvp_page_is_vmlocked(const struct lu_env *env,
 205                                 const struct cl_page_slice *slice)
 206 {
 207         return PageLocked(cl2vm_page(slice)) ? -EBUSY : -ENODATA;
 208 }
 209
 210 static int vvp_page_prep_read(const struct lu_env *env,
 211                               const struct cl_page_slice *slice,
 212                               struct cl_io *unused)
 213 {
 214         ENTRY;
 215         /* Skip the page already marked as PG_uptodate. */
 216         RETURN(PageUptodate(cl2vm_page(slice)) ? -EALREADY : 0);
 217 }
 218
 219 static int vvp_page_prep_write(const struct lu_env *env,
 220                                const struct cl_page_slice *slice,
 221                                struct cl_io *unused)
 222 {
 223         struct page *vmpage = cl2vm_page(slice);
 224         struct cl_page *pg = slice->cpl_page;
 225
 226         LASSERT(PageLocked(vmpage));
 227         LASSERT(!PageDirty(vmpage));
 228
 229         /* ll_writepage path is not a sync write, so need to set page writeback
 230          * flag */
 231         if (pg->cp_sync_io == NULL)
 232                 set_page_writeback(vmpage);
 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 vvp_object *obj = cl_inode2vvp(inode);
 246
 247         if (ioret == 0) {
 248                 ClearPageError(vmpage);
 249                 obj->vob_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                      ioret == -EIO) && obj->vob_discard_page_warned == 0) {
 259                         obj->vob_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 vvp_page *vpg    = cl2vvp_page(slice);
 270         struct page     *vmpage = vpg->vpg_page;
 271         struct cl_page  *page   = slice->cpl_page;
 272         struct inode    *inode  = vvp_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 (vpg->vpg_defer_uptodate)
 279                 ll_ra_count_put(ll_i2sbi(inode), 1);
 280
 281         if (ioret == 0)  {
 282                 if (!vpg->vpg_defer_uptodate)
 283                         cl_page_export(env, page, 1);
 284         } else if (vpg->vpg_defer_uptodate) {
 285                 vpg->vpg_defer_uptodate = 0;
 286                 if (ioret == -EWOULDBLOCK) {
 287                         /* mirror read failed, it needs to destroy the page
 288                          * because subpage would be from wrong osc when trying
 289                          * to read from a new mirror */
 290                         generic_error_remove_page(vmpage->mapping, vmpage);
 291                 }
 292         }
 293
 294         if (page->cp_sync_io == NULL)
 295                 unlock_page(vmpage);
 296
 297         EXIT;
 298 }
 299
 300 static void vvp_page_completion_write(const struct lu_env *env,
 301                                       const struct cl_page_slice *slice,
 302                                       int ioret)
 303 {
 304         struct vvp_page *vpg    = cl2vvp_page(slice);
 305         struct cl_page  *pg     = slice->cpl_page;
 306         struct page     *vmpage = vpg->vpg_page;
 307         ENTRY;
 308
 309         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 310
 311         if (pg->cp_sync_io != NULL) {
 312                 LASSERT(PageLocked(vmpage));
 313                 LASSERT(!PageWriteback(vmpage));
 314         } else {
 315                 LASSERT(PageWriteback(vmpage));
 316                 /*
 317                  * Only mark the page error only when it's an async write
 318                  * because applications won't wait for IO to finish.
 319                  */
 320                 vvp_vmpage_error(vvp_object_inode(pg->cp_obj), vmpage, ioret);
 321
 322                 end_page_writeback(vmpage);
 323         }
 324         EXIT;
 325 }
 326
 327 /**
 328  * Implements cl_page_operations::cpo_make_ready() method.
 329  *
 330  * This is called to yank a page from the transfer cache and to send it out as
 331  * a part of transfer. This function try-locks the page. If try-lock failed,
 332  * page is owned by some concurrent IO, and should be skipped (this is bad,
 333  * but hopefully rare situation, as it usually results in transfer being
 334  * shorter than possible).
 335  *
 336  * \retval 0      success, page can be placed into transfer
 337  *
 338  * \retval -EAGAIN page is either used by concurrent IO has been
 339  * truncated. Skip it.
 340  */
 341 static int vvp_page_make_ready(const struct lu_env *env,
 342                                const struct cl_page_slice *slice)
 343 {
 344         struct page *vmpage = cl2vm_page(slice);
 345         struct cl_page *pg = slice->cpl_page;
 346         int result = 0;
 347
 348         lock_page(vmpage);
 349         if (clear_page_dirty_for_io(vmpage)) {
 350                 LASSERT(pg->cp_state == CPS_CACHED);
 351                 /* This actually clears the dirty bit in the radix
 352                  * tree. */
 353                 set_page_writeback(vmpage);
 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 vvp_page *vpg    = cl2vvp_page(slice);
 373         struct page     *vmpage = vpg->vpg_page;
 374
 375         (*printer)(env, cookie, LUSTRE_VVP_NAME"-page@%p(%d:%d) "
 376                    "vm@%p ",
 377                    vpg, vpg->vpg_defer_uptodate, vpg->vpg_ra_used, vmpage);
 378
 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
 387         (*printer)(env, cookie, "\n");
 388
 389         return 0;
 390 }
 391
 392 static int vvp_page_fail(const struct lu_env *env,
 393                          const struct cl_page_slice *slice)
 394 {
 395         /*
 396          * Cached read?
 397          */
 398         LBUG();
 399
 400         return 0;
 401 }
 402
 403 static const struct cl_page_operations vvp_page_ops = {
 404         .cpo_own           = vvp_page_own,
 405         .cpo_assume        = vvp_page_assume,
 406         .cpo_unassume      = vvp_page_unassume,
 407         .cpo_disown        = vvp_page_disown,
 408         .cpo_discard       = vvp_page_discard,
 409         .cpo_delete        = vvp_page_delete,
 410         .cpo_export        = vvp_page_export,
 411         .cpo_is_vmlocked   = vvp_page_is_vmlocked,
 412         .cpo_fini          = vvp_page_fini,
 413         .cpo_print         = vvp_page_print,
 414         .io = {
 415                 [CRT_READ] = {
 416                         .cpo_prep       = vvp_page_prep_read,
 417                         .cpo_completion = vvp_page_completion_read,
 418                         .cpo_make_ready = vvp_page_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_discard(const struct lu_env *env,
 429                                        const struct cl_page_slice *slice,
 430                                        struct cl_io *unused)
 431 {
 432         struct cl_page *page = slice->cpl_page;
 433
 434         /*
 435          * For transient pages, remove it from the radix tree.
 436          */
 437         cl_page_delete(env, page);
 438 }
 439
 440 static int vvp_transient_page_is_vmlocked(const struct lu_env *env,
 441                                           const struct cl_page_slice *slice)
 442 {
 443         return -EBUSY;
 444 }
 445
 446 static void vvp_transient_page_fini(const struct lu_env *env,
 447                                     struct cl_page_slice *slice,
 448                                     struct pagevec *pvec)
 449 {
 450         struct vvp_page *vpg = cl2vvp_page(slice);
 451         struct vvp_object *clobj = cl2vvp(slice->cpl_obj);
 452
 453         vvp_page_fini_common(vpg, pvec);
 454         atomic_dec(&clobj->vob_transient_pages);
 455 }
 456
 457 static const struct cl_page_operations vvp_transient_page_ops = {
 458         .cpo_discard            = vvp_transient_page_discard,
 459         .cpo_fini               = vvp_transient_page_fini,
 460         .cpo_is_vmlocked        = vvp_transient_page_is_vmlocked,
 461         .cpo_print              = vvp_page_print,
 462 };
 463
 464 int vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 465                 struct cl_page *page, pgoff_t index)
 466 {
 467         struct vvp_page *vpg = cl_object_page_slice(obj, page);
 468         struct page     *vmpage = page->cp_vmpage;
 469
 470         CLOBINVRNT(env, obj, vvp_object_invariant(obj));
 471
 472         vpg->vpg_page = vmpage;
 473         get_page(vmpage);
 474
 475         if (page->cp_type == CPT_CACHEABLE) {
 476                 /* in cache, decref in vvp_page_delete */
 477                 atomic_inc(&page->cp_ref);
 478                 SetPagePrivate(vmpage);
 479                 vmpage->private = (unsigned long)page;
 480                 cl_page_slice_add(page, &vpg->vpg_cl, obj, index,
 481                                 &vvp_page_ops);
 482         } else {
 483                 struct vvp_object *clobj = cl2vvp(obj);
 484
 485                 cl_page_slice_add(page, &vpg->vpg_cl, obj, index,
 486                                 &vvp_transient_page_ops);
 487                 atomic_inc(&clobj->vob_transient_pages);
 488         }
 489         return 0;
 490 }