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  *
  31  * Implementation of cl_page for VVP layer.
  32  *
  33  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  34  *   Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_LLITE
  38
  39 #include <linux/atomic.h>
  40 #include <linux/bitops.h>
  41 #include <linux/mm.h>
  42 #include <linux/mutex.h>
  43 #include <linux/page-flags.h>
  44 #include <linux/pagemap.h>
  45
  46 #include <libcfs/libcfs.h>
  47 #include "llite_internal.h"
  48 #include "vvp_internal.h"
  49
  50 /*****************************************************************************
  51  *
  52  * Page operations.
  53  *
  54  */
  55 static void vvp_page_fini(const struct lu_env *env,
  56                           struct cl_page_slice *slice,
  57                           struct pagevec *pvec)
  58 {
  59         struct vvp_page *vpg     = cl2vvp_page(slice);
  60         struct page     *vmpage  = vpg->vpg_page;
  61
  62         /*
  63          * vmpage->private was already cleared when page was moved into
  64          * VPG_FREEING state.
  65          */
  66         LASSERT((struct cl_page *)vmpage->private != slice->cpl_page);
  67         LASSERT(vmpage != NULL);
  68         if (pvec) {
  69                 if (!pagevec_add(pvec, vmpage))
  70                         pagevec_release(pvec);
  71         } else {
  72                 put_page(vmpage);
  73         }
  74 }
  75
  76 static int vvp_page_own(const struct lu_env *env,
  77                         const struct cl_page_slice *slice, struct cl_io *io,
  78                         int nonblock)
  79 {
  80         struct vvp_page *vpg    = cl2vvp_page(slice);
  81         struct page     *vmpage = vpg->vpg_page;
  82
  83         ENTRY;
  84
  85         LASSERT(vmpage != NULL);
  86         if (nonblock) {
  87                 if (!trylock_page(vmpage))
  88                         return -EAGAIN;
  89
  90                 if (unlikely(PageWriteback(vmpage))) {
  91                         unlock_page(vmpage);
  92                         return -EAGAIN;
  93                 }
  94
  95                 return 0;
  96         }
  97
  98         lock_page(vmpage);
  99         wait_on_page_writeback(vmpage);
 100
 101         RETURN(0);
 102 }
 103
 104 static void vvp_page_assume(const struct lu_env *env,
 105                             const struct cl_page_slice *slice,
 106                             struct cl_io *unused)
 107 {
 108         struct page *vmpage = cl2vm_page(slice);
 109
 110         LASSERT(vmpage != NULL);
 111         LASSERT(PageLocked(vmpage));
 112         wait_on_page_writeback(vmpage);
 113 }
 114
 115 static void vvp_page_unassume(const struct lu_env *env,
 116                               const struct cl_page_slice *slice,
 117                               struct cl_io *unused)
 118 {
 119         struct page *vmpage = cl2vm_page(slice);
 120
 121         LASSERT(vmpage != NULL);
 122         LASSERT(PageLocked(vmpage));
 123 }
 124
 125 static void vvp_page_disown(const struct lu_env *env,
 126                             const struct cl_page_slice *slice, struct cl_io *io)
 127 {
 128         struct page *vmpage = cl2vm_page(slice);
 129
 130         ENTRY;
 131
 132         LASSERT(vmpage != NULL);
 133         LASSERT(PageLocked(vmpage));
 134
 135         unlock_page(cl2vm_page(slice));
 136
 137         EXIT;
 138 }
 139
 140 static void vvp_page_discard(const struct lu_env *env,
 141                              const struct cl_page_slice *slice,
 142                              struct cl_io *unused)
 143 {
 144         struct page     *vmpage = cl2vm_page(slice);
 145         struct vvp_page *vpg    = cl2vvp_page(slice);
 146
 147         LASSERT(vmpage != NULL);
 148         LASSERT(PageLocked(vmpage));
 149
 150         if (vpg->vpg_defer_uptodate && !vpg->vpg_ra_used && vmpage->mapping)
 151                 ll_ra_stats_inc(vmpage->mapping->host, RA_STAT_DISCARDED);
 152
 153         generic_error_remove_page(vmpage->mapping, vmpage);
 154 }
 155
 156 static void vvp_page_delete(const struct lu_env *env,
 157                             const struct cl_page_slice *slice)
 158 {
 159         struct page      *vmpage = cl2vm_page(slice);
 160         struct cl_page   *page   = slice->cpl_page;
 161         int refc;
 162
 163         LASSERT(PageLocked(vmpage));
 164         LASSERT((struct cl_page *)vmpage->private == page);
 165
 166
 167         /* Drop the reference count held in vvp_page_init */
 168         refc = atomic_dec_return(&page->cp_ref);
 169         LASSERTF(refc >= 1, "page = %p, refc = %d\n", page, refc);
 170
 171         ClearPagePrivate(vmpage);
 172         vmpage->private = 0;
 173
 174         /**
 175          * Vmpage might not be released due page refcount != 2,
 176          * clear Page uptodate here to avoid stale data.
 177          */
 178         ClearPageUptodate(vmpage);
 179         /*
 180          * Reference from vmpage to cl_page is removed, but the reference back
 181          * is still here. It is removed later in vvp_page_fini().
 182          */
 183 }
 184
 185 static void vvp_page_export(const struct lu_env *env,
 186                             const struct cl_page_slice *slice,
 187                             int uptodate)
 188 {
 189         struct page *vmpage = cl2vm_page(slice);
 190
 191         LASSERT(vmpage != NULL);
 192         LASSERT(PageLocked(vmpage));
 193         if (uptodate)
 194                 SetPageUptodate(vmpage);
 195         else
 196                 ClearPageUptodate(vmpage);
 197 }
 198
 199 static int vvp_page_is_vmlocked(const struct lu_env *env,
 200                                 const struct cl_page_slice *slice)
 201 {
 202         return PageLocked(cl2vm_page(slice)) ? -EBUSY : -ENODATA;
 203 }
 204
 205 static int vvp_page_prep_read(const struct lu_env *env,
 206                               const struct cl_page_slice *slice,
 207                               struct cl_io *unused)
 208 {
 209         ENTRY;
 210         /* Skip the page already marked as PG_uptodate. */
 211         RETURN(PageUptodate(cl2vm_page(slice)) ? -EALREADY : 0);
 212 }
 213
 214 static int vvp_page_prep_write(const struct lu_env *env,
 215                                const struct cl_page_slice *slice,
 216                                struct cl_io *unused)
 217 {
 218         struct page *vmpage = cl2vm_page(slice);
 219         struct cl_page *pg = slice->cpl_page;
 220
 221         LASSERT(PageLocked(vmpage));
 222         LASSERT(!PageDirty(vmpage));
 223
 224         /* ll_writepage path is not a sync write, so need to set page writeback
 225          * flag
 226          */
 227         if (pg->cp_sync_io == NULL)
 228                 set_page_writeback(vmpage);
 229
 230         return 0;
 231 }
 232
 233 /**
 234  * Handles page transfer errors at VM level.
 235  *
 236  * This takes inode as a separate argument, because inode on which error is to
 237  * be set can be different from \a vmpage inode in case of direct-io.
 238  */
 239 static void vvp_vmpage_error(struct inode *inode, struct page *vmpage,
 240                              int ioret)
 241 {
 242         struct vvp_object *obj = cl_inode2vvp(inode);
 243
 244         if (ioret == 0) {
 245                 ClearPageError(vmpage);
 246                 obj->vob_discard_page_warned = 0;
 247         } else {
 248                 SetPageError(vmpage);
 249                 if (ioret == -ENOSPC)
 250                         set_bit(AS_ENOSPC, &inode->i_mapping->flags);
 251                 else
 252                         set_bit(AS_EIO, &inode->i_mapping->flags);
 253
 254                 if ((ioret == -ESHUTDOWN || ioret == -EINTR ||
 255                      ioret == -EIO) && obj->vob_discard_page_warned == 0) {
 256                         obj->vob_discard_page_warned = 1;
 257                         ll_dirty_page_discard_warn(inode, ioret);
 258                 }
 259         }
 260 }
 261
 262 static void vvp_page_completion_read(const struct lu_env *env,
 263                                      const struct cl_page_slice *slice,
 264                                      int ioret)
 265 {
 266         struct vvp_page *vpg    = cl2vvp_page(slice);
 267         struct page     *vmpage = vpg->vpg_page;
 268         struct cl_page  *page   = slice->cpl_page;
 269         struct inode    *inode  = vvp_object_inode(page->cp_obj);
 270
 271         ENTRY;
 272         LASSERT(PageLocked(vmpage));
 273         CL_PAGE_HEADER(D_PAGE, env, page, "completing READ with %d\n", ioret);
 274
 275         if (vpg->vpg_defer_uptodate)
 276                 ll_ra_count_put(ll_i2sbi(inode), 1);
 277
 278         if (ioret == 0)  {
 279                 if (!vpg->vpg_defer_uptodate)
 280                         cl_page_export(env, page, 1);
 281         } else if (vpg->vpg_defer_uptodate) {
 282                 vpg->vpg_defer_uptodate = 0;
 283                 if (ioret == -EAGAIN) {
 284                         /* mirror read failed, it needs to destroy the page
 285                          * because subpage would be from wrong osc when trying
 286                          * to read from a new mirror
 287                          */
 288                         generic_error_remove_page(vmpage->mapping, vmpage);
 289                 }
 290         }
 291
 292         if (page->cp_sync_io == NULL)
 293                 unlock_page(vmpage);
 294
 295         EXIT;
 296 }
 297
 298 static void vvp_page_completion_write(const struct lu_env *env,
 299                                       const struct cl_page_slice *slice,
 300                                       int ioret)
 301 {
 302         struct vvp_page *vpg    = cl2vvp_page(slice);
 303         struct cl_page  *pg     = slice->cpl_page;
 304         struct page     *vmpage = vpg->vpg_page;
 305
 306         ENTRY;
 307         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 308
 309         if (pg->cp_sync_io != NULL) {
 310                 LASSERT(PageLocked(vmpage));
 311                 LASSERT(!PageWriteback(vmpage));
 312         } else {
 313                 LASSERT(PageWriteback(vmpage));
 314                 /*
 315                  * Only mark the page error only when it's an async write
 316                  * because applications won't wait for IO to finish.
 317                  */
 318                 vvp_vmpage_error(vvp_object_inode(pg->cp_obj), vmpage, ioret);
 319
 320                 end_page_writeback(vmpage);
 321         }
 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         struct page *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                  */
 352                 set_page_writeback(vmpage);
 353                 CL_PAGE_HEADER(D_PAGE, env, pg, "readied\n");
 354         } else if (pg->cp_state == CPS_PAGEOUT) {
 355                 /* is it possible for osc_flush_async_page() to already
 356                  * make it ready?
 357                  */
 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,
 376                    LUSTRE_VVP_NAME"-page@%p(%d:%d) 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 const struct cl_page_operations vvp_transient_page_ops = {
 447         .cpo_discard            = vvp_transient_page_discard,
 448         .cpo_is_vmlocked        = vvp_transient_page_is_vmlocked,
 449         .cpo_print              = vvp_page_print,
 450 };
 451
 452 int vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 453                 struct cl_page *page, pgoff_t index)
 454 {
 455         struct vvp_page *vpg = cl_object_page_slice(obj, page);
 456         struct page     *vmpage = page->cp_vmpage;
 457
 458         CLOBINVRNT(env, obj, vvp_object_invariant(obj));
 459
 460         vpg->vpg_page = vmpage;
 461
 462         if (page->cp_type == CPT_TRANSIENT) {
 463                 /* DIO pages are referenced by userspace, we don't need to take
 464                  * a reference on them. (contrast with get_page() call above)
 465                  */
 466                 cl_page_slice_add(page, &vpg->vpg_cl, obj,
 467                                   &vvp_transient_page_ops);
 468         } else {
 469                 get_page(vmpage);
 470                 /* in cache, decref in vvp_page_delete */
 471                 atomic_inc(&page->cp_ref);
 472                 SetPagePrivate(vmpage);
 473                 vmpage->private = (unsigned long)page;
 474                 cl_page_slice_add(page, &vpg->vpg_cl, obj,
 475                                 &vvp_page_ops);
 476         }
 477
 478         return 0;
 479 }