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 void vvp_page_delete(const struct lu_env *env,
  67                             const struct cl_page_slice *slice)
  68 {
  69         struct cl_page *cp = slice->cpl_page;
  70
  71         if (cp->cp_type == CPT_CACHEABLE) {
  72                 struct page *vmpage = cp->cp_vmpage;
  73                 struct inode *inode = vmpage->mapping->host;
  74
  75                 LASSERT(PageLocked(vmpage));
  76                 LASSERT((struct cl_page *)vmpage->private == cp);
  77
  78                 /* Drop the reference count held in vvp_page_init */
  79                 refcount_dec(&cp->cp_ref);
  80
  81                 ClearPagePrivate(vmpage);
  82                 vmpage->private = 0;
  83
  84                 /* clearpageuptodate prevents the page being read by the
  85                  * kernel after it has been deleted from Lustre, which avoids
  86                  * potential stale data reads.  The seqlock allows us to see
  87                  * that a page was potentially deleted and catch the resulting
  88                  * SIGBUS - see ll_filemap_fault() (LU-16160)
  89                  */
  90                 if (PageUptodate(vmpage)) {
  91                         write_seqlock(&ll_i2info(inode)->lli_page_inv_lock);
  92                         ClearPageUptodate(vmpage);
  93                         write_sequnlock(&ll_i2info(inode)->lli_page_inv_lock);
  94                 }
  95                 /* The reference from vmpage to cl_page is removed,
  96                  * but the reference back is still here. It is removed
  97                  * later in cl_page_free().
  98                  */
  99         }
 100 }
 101
 102 /**
 103  * Handles page transfer errors at VM level.
 104  *
 105  * This takes inode as a separate argument, because inode on which error is to
 106  * be set can be different from \a vmpage inode in case of direct-io.
 107  */
 108 static void vvp_vmpage_error(struct inode *inode, struct page *vmpage,
 109                              int ioret)
 110 {
 111         struct vvp_object *obj = cl_inode2vvp(inode);
 112
 113         if (ioret == 0) {
 114                 ClearPageError(vmpage);
 115                 obj->vob_discard_page_warned = 0;
 116         } else {
 117                 SetPageError(vmpage);
 118                 if (ioret != -ENOSPC &&
 119                     CFS_FAIL_CHECK(OBD_FAIL_LLITE_PANIC_ON_ESTALE))
 120                         LBUG();
 121
 122                 mapping_set_error(inode->i_mapping, ioret);
 123
 124                 if ((ioret == -ESHUTDOWN || ioret == -EINTR ||
 125                      ioret == -EIO) && obj->vob_discard_page_warned == 0) {
 126                         obj->vob_discard_page_warned = 1;
 127                         ll_dirty_page_discard_warn(inode, ioret);
 128                 }
 129         }
 130 }
 131
 132 static void vvp_page_completion_read(const struct lu_env *env,
 133                                      const struct cl_page_slice *slice,
 134                                      int ioret)
 135 {
 136         struct cl_page *cp = slice->cpl_page;
 137         struct page *vmpage = cp->cp_vmpage;
 138         struct inode *inode = vvp_object_inode(cp->cp_obj);
 139
 140         ENTRY;
 141         LASSERT(PageLocked(vmpage));
 142         CL_PAGE_HEADER(D_PAGE, env, cp, "completing READ with %d\n", ioret);
 143
 144         if (cp->cp_defer_uptodate)
 145                 ll_ra_count_put(ll_i2sbi(inode), 1);
 146
 147         if (ioret == 0)  {
 148                 /**
 149                  * cp_defer_uptodate is used for readahead page, and the
 150                  * vmpage Uptodate bit is deferred to set in ll_readpage/
 151                  * ll_io_read_page.
 152                  */
 153                 if (!cp->cp_defer_uptodate)
 154                         SetPageUptodate(vmpage);
 155         } else if (cp->cp_defer_uptodate) {
 156                 cp->cp_defer_uptodate = 0;
 157                 if (ioret == -EAGAIN) {
 158                         /* mirror read failed, it needs to destroy the page
 159                          * because subpage would be from wrong osc when trying
 160                          * to read from a new mirror
 161                          */
 162                         generic_error_remove_folio(vmpage->mapping,
 163                                                    page_folio(vmpage));
 164                 }
 165         }
 166
 167         if (cp->cp_sync_io == NULL)
 168                 unlock_page(vmpage);
 169
 170         EXIT;
 171 }
 172
 173 static void vvp_page_completion_write(const struct lu_env *env,
 174                                       const struct cl_page_slice *slice,
 175                                       int ioret)
 176 {
 177         struct cl_page *cp = slice->cpl_page;
 178         struct page *vmpage = cp->cp_vmpage;
 179
 180         ENTRY;
 181         CL_PAGE_HEADER(D_PAGE, env, cp, "completing WRITE with %d\n", ioret);
 182
 183         if (cp->cp_sync_io != NULL) {
 184                 LASSERT(PageLocked(vmpage));
 185                 LASSERT(!PageWriteback(vmpage));
 186         } else {
 187                 LASSERT(PageWriteback(vmpage));
 188                 /*
 189                  * Only mark the page error only when it's an async write
 190                  * because applications won't wait for IO to finish.
 191                  */
 192                 vvp_vmpage_error(vvp_object_inode(cp->cp_obj), vmpage, ioret);
 193
 194                 end_page_writeback(vmpage);
 195         }
 196         EXIT;
 197 }
 198
 199 static const struct cl_page_operations vvp_page_ops = {
 200         .cpo_delete        = vvp_page_delete,
 201         .cpo_discard       = vvp_page_discard,
 202         .io = {
 203                 [CRT_READ] = {
 204                         .cpo_completion = vvp_page_completion_read,
 205                 },
 206                 [CRT_WRITE] = {
 207                         .cpo_completion = vvp_page_completion_write,
 208                 },
 209         },
 210 };
 211
 212 static const struct cl_page_operations vvp_transient_page_ops = {
 213 };
 214
 215 int vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 216                 struct cl_page *page, pgoff_t index)
 217 {
 218         struct cl_page_slice *cpl = cl_object_page_slice(obj, page);
 219         struct page *vmpage = page->cp_vmpage;
 220
 221         CLOBINVRNT(env, obj, vvp_object_invariant(obj));
 222
 223         if (page->cp_type == CPT_TRANSIENT) {
 224                 /* DIO pages are referenced by userspace, we don't need to take
 225                  * a reference on them. (contrast with get_page() call above)
 226                  */
 227                 cl_page_slice_add(page, cpl, obj,
 228                                   &vvp_transient_page_ops);
 229         } else {
 230                 get_page(vmpage);
 231                 /* in cache, decref in cl_page_delete() */
 232                 refcount_inc(&page->cp_ref);
 233                 SetPagePrivate(vmpage);
 234                 vmpage->private = (unsigned long)page;
 235                 cl_page_slice_add(page, cpl, obj, &vvp_page_ops);
 236         }
 237
 238         return 0;
 239 }