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_discard(const struct lu_env *env,
  56                              const struct cl_page_slice *slice,
  57                              struct cl_io *unused)
  58 {
  59         struct cl_page *cp = slice->cpl_page;
  60         struct page *vmpage = cp->cp_vmpage;
  61
  62         if (cp->cp_defer_uptodate && !cp->cp_ra_used && vmpage->mapping != NULL)
  63                 ll_ra_stats_inc(vmpage->mapping->host, RA_STAT_DISCARDED);
  64 }
  65
  66 static int vvp_page_prep_read(const struct lu_env *env,
  67                               const struct cl_page_slice *slice,
  68                               struct cl_io *unused)
  69 {
  70         ENTRY;
  71         /* Skip the page already marked as PG_uptodate. */
  72         RETURN(PageUptodate(cl2vm_page(slice)) ? -EALREADY : 0);
  73 }
  74
  75 static int vvp_page_prep_write(const struct lu_env *env,
  76                                const struct cl_page_slice *slice,
  77                                struct cl_io *unused)
  78 {
  79         struct page *vmpage = cl2vm_page(slice);
  80         struct cl_page *pg = slice->cpl_page;
  81
  82         LASSERT(PageLocked(vmpage));
  83         LASSERT(!PageDirty(vmpage));
  84
  85         /* ll_writepage path is not a sync write, so need to set page writeback
  86          * flag
  87          */
  88         if (pg->cp_sync_io == NULL)
  89                 set_page_writeback(vmpage);
  90
  91         return 0;
  92 }
  93
  94 /**
  95  * Handles page transfer errors at VM level.
  96  *
  97  * This takes inode as a separate argument, because inode on which error is to
  98  * be set can be different from \a vmpage inode in case of direct-io.
  99  */
 100 static void vvp_vmpage_error(struct inode *inode, struct page *vmpage,
 101                              int ioret)
 102 {
 103         struct vvp_object *obj = cl_inode2vvp(inode);
 104
 105         if (ioret == 0) {
 106                 ClearPageError(vmpage);
 107                 obj->vob_discard_page_warned = 0;
 108         } else {
 109                 SetPageError(vmpage);
 110                 if (ioret == -ENOSPC)
 111                         set_bit(AS_ENOSPC, &inode->i_mapping->flags);
 112                 else
 113                         set_bit(AS_EIO, &inode->i_mapping->flags);
 114
 115                 if ((ioret == -ESHUTDOWN || ioret == -EINTR ||
 116                      ioret == -EIO) && obj->vob_discard_page_warned == 0) {
 117                         obj->vob_discard_page_warned = 1;
 118                         ll_dirty_page_discard_warn(inode, ioret);
 119                 }
 120         }
 121 }
 122
 123 static void vvp_page_completion_read(const struct lu_env *env,
 124                                      const struct cl_page_slice *slice,
 125                                      int ioret)
 126 {
 127         struct cl_page *cp = slice->cpl_page;
 128         struct page *vmpage = cp->cp_vmpage;
 129         struct inode *inode = vvp_object_inode(cp->cp_obj);
 130
 131         ENTRY;
 132         LASSERT(PageLocked(vmpage));
 133         CL_PAGE_HEADER(D_PAGE, env, cp, "completing READ with %d\n", ioret);
 134
 135         if (cp->cp_defer_uptodate)
 136                 ll_ra_count_put(ll_i2sbi(inode), 1);
 137
 138         if (ioret == 0)  {
 139                 if (!cp->cp_defer_uptodate)
 140                         SetPageUptodate(vmpage);
 141         } else if (cp->cp_defer_uptodate) {
 142                 cp->cp_defer_uptodate = 0;
 143                 if (ioret == -EAGAIN) {
 144                         /* mirror read failed, it needs to destroy the page
 145                          * because subpage would be from wrong osc when trying
 146                          * to read from a new mirror
 147                          */
 148                         generic_error_remove_page(vmpage->mapping, vmpage);
 149                 }
 150         }
 151
 152         if (cp->cp_sync_io == NULL)
 153                 unlock_page(vmpage);
 154
 155         EXIT;
 156 }
 157
 158 static void vvp_page_completion_write(const struct lu_env *env,
 159                                       const struct cl_page_slice *slice,
 160                                       int ioret)
 161 {
 162         struct vvp_page *vpg    = cl2vvp_page(slice);
 163         struct cl_page  *pg     = slice->cpl_page;
 164         struct page     *vmpage = vpg->vpg_page;
 165
 166         ENTRY;
 167         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 168
 169         if (pg->cp_sync_io != NULL) {
 170                 LASSERT(PageLocked(vmpage));
 171                 LASSERT(!PageWriteback(vmpage));
 172         } else {
 173                 LASSERT(PageWriteback(vmpage));
 174                 /*
 175                  * Only mark the page error only when it's an async write
 176                  * because applications won't wait for IO to finish.
 177                  */
 178                 vvp_vmpage_error(vvp_object_inode(pg->cp_obj), vmpage, ioret);
 179
 180                 end_page_writeback(vmpage);
 181         }
 182         EXIT;
 183 }
 184
 185 /**
 186  * Implements cl_page_operations::cpo_make_ready() method.
 187  *
 188  * This is called to yank a page from the transfer cache and to send it out as
 189  * a part of transfer. This function try-locks the page. If try-lock failed,
 190  * page is owned by some concurrent IO, and should be skipped (this is bad,
 191  * but hopefully rare situation, as it usually results in transfer being
 192  * shorter than possible).
 193  *
 194  * \retval 0      success, page can be placed into transfer
 195  *
 196  * \retval -EAGAIN page is either used by concurrent IO has been
 197  * truncated. Skip it.
 198  */
 199 static int vvp_page_make_ready(const struct lu_env *env,
 200                                const struct cl_page_slice *slice)
 201 {
 202         struct page *vmpage = cl2vm_page(slice);
 203         struct cl_page *pg = slice->cpl_page;
 204         int result = 0;
 205
 206         lock_page(vmpage);
 207         if (clear_page_dirty_for_io(vmpage)) {
 208                 LASSERT(pg->cp_state == CPS_CACHED);
 209                 /* This actually clears the dirty bit in the radix
 210                  * tree.
 211                  */
 212                 set_page_writeback(vmpage);
 213                 CL_PAGE_HEADER(D_PAGE, env, pg, "readied\n");
 214         } else if (pg->cp_state == CPS_PAGEOUT) {
 215                 /* is it possible for osc_flush_async_page() to already
 216                  * make it ready?
 217                  */
 218                 result = -EALREADY;
 219         } else {
 220                 CL_PAGE_DEBUG(D_ERROR, env, pg, "Unexpecting page state %d.\n",
 221                               pg->cp_state);
 222                 LBUG();
 223         }
 224         unlock_page(vmpage);
 225         RETURN(result);
 226 }
 227
 228 static int vvp_page_print(const struct lu_env *env,
 229                           const struct cl_page_slice *slice,
 230                           void *cookie, lu_printer_t printer)
 231 {
 232         struct vvp_page *vpg    = cl2vvp_page(slice);
 233         struct page     *vmpage = vpg->vpg_page;
 234
 235         (*printer)(env, cookie,
 236                    LUSTRE_VVP_NAME"-page@%p vm@%p ", vpg, vmpage);
 237
 238         if (vmpage != NULL) {
 239                 (*printer)(env, cookie, "%lx %d:%d %lx %lu %slru",
 240                            (long)vmpage->flags, page_count(vmpage),
 241                            page_mapcount(vmpage), vmpage->private,
 242                            page_index(vmpage),
 243                            list_empty(&vmpage->lru) ? "not-" : "");
 244         }
 245
 246         (*printer)(env, cookie, "\n");
 247
 248         return 0;
 249 }
 250
 251 static int vvp_page_fail(const struct lu_env *env,
 252                          const struct cl_page_slice *slice)
 253 {
 254         /*
 255          * Cached read?
 256          */
 257         LBUG();
 258
 259         return 0;
 260 }
 261
 262 static const struct cl_page_operations vvp_page_ops = {
 263         .cpo_discard       = vvp_page_discard,
 264         .cpo_print         = vvp_page_print,
 265         .io = {
 266                 [CRT_READ] = {
 267                         .cpo_prep       = vvp_page_prep_read,
 268                         .cpo_completion = vvp_page_completion_read,
 269                         .cpo_make_ready = vvp_page_fail,
 270                 },
 271                 [CRT_WRITE] = {
 272                         .cpo_prep       = vvp_page_prep_write,
 273                         .cpo_completion = vvp_page_completion_write,
 274                         .cpo_make_ready = vvp_page_make_ready,
 275                 },
 276         },
 277 };
 278
 279 static const struct cl_page_operations vvp_transient_page_ops = {
 280         .cpo_print              = vvp_page_print,
 281 };
 282
 283 int vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 284                 struct cl_page *page, pgoff_t index)
 285 {
 286         struct vvp_page *vpg = cl_object_page_slice(obj, page);
 287         struct page     *vmpage = page->cp_vmpage;
 288
 289         CLOBINVRNT(env, obj, vvp_object_invariant(obj));
 290
 291         vpg->vpg_page = vmpage;
 292
 293         if (page->cp_type == CPT_TRANSIENT) {
 294                 /* DIO pages are referenced by userspace, we don't need to take
 295                  * a reference on them. (contrast with get_page() call above)
 296                  */
 297                 cl_page_slice_add(page, &vpg->vpg_cl, obj,
 298                                   &vvp_transient_page_ops);
 299         } else {
 300                 get_page(vmpage);
 301                 /* in cache, decref in cl_page_delete() */
 302                 atomic_inc(&page->cp_ref);
 303                 SetPagePrivate(vmpage);
 304                 vmpage->private = (unsigned long)page;
 305                 cl_page_slice_add(page, &vpg->vpg_cl, obj,
 306                                 &vvp_page_ops);
 307         }
 308
 309         return 0;
 310 }