/* when brw_kiovec() is asked to read from block -1UL it just zeros
* the page. this gives us a chance to verify the write mappings
* as well */
-static int filter_cleanup_mappings(int rw, struct kiobuf *iobuf,
+static int filter_cleanup_mappings(int rw, struct kiobuf *iobuf,
struct inode *inode)
{
int i, blocks_per_page_bits = PAGE_SHIFT - inode->i_blkbits;
}
#endif
-static void filter_clear_page_cache(struct inode *inode, struct kiobuf *iobuf)
+/* These are our hacks to keep our directio/bh IO coherent with ext3's
+ * page cache use. Most notably ext3 reads file data into the page
+ * cache when it is zeroing the tail of partial-block truncates and
+ * leaves it there, sometimes generating io from it at later truncates.
+ * This removes the partial page and its buffers from the page cache,
+ * so it should only ever cause a wait in rare cases, as otherwise we
+ * always do full-page IO to the OST.
+ *
+ * The call to truncate_complete_page() will call journal_flushpage() to
+ * free the buffers and drop the page from cache. The buffers should not
+ * be dirty, because we already called fdatasync/fdatawait on them.
+ */
+static int filter_clear_page_cache(struct inode *inode, struct kiobuf *iobuf)
{
struct page *page;
- int i;
+ int i, rc, rc2;
+
+ check_pending_bhs(KIOBUF_GET_BLOCKS(iobuf), iobuf->nr_pages,
+ inode->i_dev, 1 << inode->i_blkbits);
+
+ /* This is nearly generic_osync_inode, without the waiting on the inode
+ rc = generic_osync_inode(inode, inode->i_mapping,
+ OSYNC_DATA|OSYNC_METADATA);
+ */
+ rc = filemap_fdatasync(inode->i_mapping);
+ rc2 = fsync_inode_data_buffers(inode);
+ if (rc == 0)
+ rc = rc2;
+ rc2 = filemap_fdatawait(inode->i_mapping);
+ if (rc == 0)
+ rc = rc2;
+ if (rc != 0)
+ RETURN(rc);
+ /* be careful to call this after fsync_inode_data_buffers has waited
+ * for IO to complete before we evict it from the cache */
for (i = 0; i < iobuf->nr_pages ; i++) {
page = find_lock_page(inode->i_mapping,
iobuf->maplist[i]->index);
if (page == NULL)
continue;
- if (page->mapping != NULL) {
- block_flushpage(page, 0);
- truncate_complete_page(page);
- }
+ if (page->mapping != NULL)
+ ll_truncate_complete_page(page);
+
unlock_page(page);
page_cache_release(page);
}
+
+ return 0;
}
/* Must be called with i_sem taken for writes; this will drop it */
if (iobuf->nr_pages * blocks_per_page > KIO_MAX_SECTORS)
GOTO(cleanup, rc = -EINVAL);
- if (iobuf->nr_pages * blocks_per_page >
+ if (iobuf->nr_pages * blocks_per_page >
OBDFILTER_CREATED_SCRATCHPAD_ENTRIES)
GOTO(cleanup, rc = -EINVAL);
GOTO(cleanup, rc);
}
- /* these are our hacks to keep our directio/bh IO coherent with ext3's
- * page cache use. Most notably ext3 reads file data into the page
- * cache when it is zeroing the tail of partial-block truncates and
- * leaves it there, sometimes generating io from it at later truncates.
- * Someday very soon we'll be performing our brw_kiovec() IO to and
- * from the page cache. */
- check_pending_bhs(KIOBUF_GET_BLOCKS(iobuf), iobuf->nr_pages,
- inode->i_dev, 1 << inode->i_blkbits);
-
- rc = filemap_fdatasync(inode->i_mapping);
- if (rc == 0)
- rc = fsync_inode_data_buffers(inode);
- if (rc == 0)
- rc = filemap_fdatawait(inode->i_mapping);
+ rc = filter_clear_page_cache(inode, iobuf);
if (rc < 0)
GOTO(cleanup, rc);
- /* be careful to call this after fsync_inode_data_buffers has waited
- * for IO to complete before we evict it from the cache */
- filter_clear_page_cache(inode, iobuf);
-
rc = fsfilt_send_bio(rw, obd, inode, iobuf);
CDEBUG(D_INFO, "tried to write %d pages, rc = %d\n",
int rc)
{
struct obd_device *obd = exp->exp_obd;
- struct obd_run_ctxt saved;
+ struct lvfs_run_ctxt saved;
struct niobuf_local *lnb;
struct fsfilt_objinfo fso;
struct iattr iattr = { 0 };
iattr.ia_size = this_size;
}
- push_ctxt(&saved, &obd->obd_ctxt, NULL);
+ push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
cleanup_phase = 2;
down(&inode->i_sem);
fsfilt_check_slow(now, obd_timeout, "direct_io");
err = fsfilt_commit_wait(obd, inode, wait_handle);
- if (err)
+ if (err) {
+ CERROR("Failure to commit OST transaction (%d)?\n", err);
rc = err;
+ }
if (obd_sync_filter && !err)
- LASSERT(oti->oti_transno <= obd->obd_last_committed);
+ LASSERTF(oti->oti_transno <= obd->obd_last_committed,
+ "oti_transno "LPU64" last_committed "LPU64"\n",
+ oti->oti_transno, obd->obd_last_committed);
fsfilt_check_slow(now, obd_timeout, "commitrw commit");
cleanup:
switch (cleanup_phase) {
case 2:
- pop_ctxt(&saved, &obd->obd_ctxt, NULL);
+ pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
LASSERT(current->journal_info == NULL);
case 1:
filter_free_iobuf(iobuf);
git://git.whamcloud.com / fs / lustre-release.git / blobdiff