1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * linux/fs/obdfilter/filter_io.c
6 * Copyright (c) 2001-2003 Cluster File Systems, Inc.
7 * Author: Peter Braam <braam@clusterfs.com>
8 * Author: Andreas Dilger <adilger@clusterfs.com>
9 * Author: Phil Schwan <phil@clusterfs.com>
11 * This file is part of Lustre, http://www.lustre.org.
13 * Lustre is free software; you can redistribute it and/or
14 * modify it under the terms of version 2 of the GNU General Public
15 * License as published by the Free Software Foundation.
17 * Lustre is distributed in the hope that it will be useful,
18 * but WITHOUT ANY WARRANTY; without even the implied warranty of
19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
20 * GNU General Public License for more details.
22 * You should have received a copy of the GNU General Public License
23 * along with Lustre; if not, write to the Free Software
24 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
27 #include <linux/config.h>
28 #include <linux/module.h>
29 #include <linux/pagemap.h> // XXX kill me soon
30 #include <linux/version.h>
32 #define DEBUG_SUBSYSTEM S_FILTER
34 #include <linux/iobuf.h>
35 #include <linux/locks.h>
37 #include <linux/obd_class.h>
38 #include <linux/lustre_fsfilt.h>
39 #include "filter_internal.h"
42 /* We should only change the file mtime (and not the ctime, like
43 * update_inode_times() in generic_file_write()) when we only change data. */
44 void inode_update_time(struct inode *inode, int ctime_too)
46 time_t now = CURRENT_TIME;
47 if (inode->i_mtime == now && (!ctime_too || inode->i_ctime == now))
52 mark_inode_dirty_sync(inode);
55 /* Bug 2254 -- this is better done in ext3_map_inode_page, but this
56 * workaround will suffice until everyone has upgraded their kernels */
57 static void check_pending_bhs(unsigned long *blocks, int nr_pages, dev_t dev,
60 #if (LUSTRE_KERNEL_VERSION < 32)
61 struct buffer_head *bh;
64 for (i = 0; i < nr_pages; i++) {
65 bh = get_hash_table(dev, blocks[i], size);
68 if (!buffer_dirty(bh)) {
72 mark_buffer_clean(bh);
74 clear_bit(BH_Req, &bh->b_state);
80 /* when brw_kiovec() is asked to read from block -1UL it just zeros
81 * the page. this gives us a chance to verify the write mappings
83 static int filter_cleanup_mappings(int rw, struct kiobuf *iobuf,
86 int i, blocks_per_page_bits = PAGE_SHIFT - inode->i_blkbits;
89 for (i = 0 ; i < iobuf->nr_pages << blocks_per_page_bits; i++) {
90 if (iobuf->blocks[i] > 0)
93 if (rw == OBD_BRW_WRITE)
96 iobuf->blocks[i] = -1UL;
102 static void dump_page(int rw, unsigned long block, struct page *page)
104 char *blah = kmap(page);
105 CDEBUG(D_PAGE, "rw %d block %lu: %02x %02x %02x %02x\n", rw, block,
106 blah[0], blah[1], blah[2], blah[3]);
111 static void filter_clear_page_cache(struct inode *inode, struct kiobuf *iobuf)
116 for (i = 0; i < iobuf->nr_pages ; i++) {
117 page = find_lock_page(inode->i_mapping,
118 iobuf->maplist[i]->index);
121 if (page->mapping != NULL) {
122 block_flushpage(page, 0);
123 truncate_complete_page(page);
126 page_cache_release(page);
130 /* Must be called with i_sem taken for writes; this will drop it */
131 int filter_direct_io(int rw, struct dentry *dchild, void *buf,
132 struct obd_export *exp, struct iattr *attr,
133 struct obd_trans_info *oti, void **wait_handle)
135 struct obd_device *obd = exp->exp_obd;
136 struct inode *inode = dchild->d_inode;
137 struct kiobuf *iobuf = buf;
138 int rc, create = (rw == OBD_BRW_WRITE), *created = NULL, committed = 0;
139 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits, cleanup_phase = 0;
140 struct semaphore *sem = NULL;
143 LASSERTF(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ, "%x\n", rw);
145 if (iobuf->nr_pages == 0)
146 GOTO(cleanup, rc = 0);
148 if (iobuf->nr_pages * blocks_per_page > KIO_MAX_SECTORS)
149 GOTO(cleanup, rc = -EINVAL);
151 OBD_ALLOC(created, sizeof(*created) * iobuf->nr_pages*blocks_per_page);
153 GOTO(cleanup, rc = -ENOMEM);
156 rc = lock_kiovec(1, &iobuf, 1);
161 if (rw == OBD_BRW_WRITE) {
163 sem = &obd->u.filter.fo_alloc_lock;
166 rc = fsfilt_map_inode_pages(obd, inode, iobuf->maplist,
167 iobuf->nr_pages, iobuf->blocks, created,
172 rc = filter_cleanup_mappings(rw, iobuf, inode);
176 if (rw == OBD_BRW_WRITE) {
177 filter_tally_write(&obd->u.filter, iobuf->maplist,
178 iobuf->nr_pages, iobuf->blocks,
181 if (attr->ia_size > inode->i_size)
182 attr->ia_valid |= ATTR_SIZE;
183 rc = fsfilt_setattr(obd, dchild, oti->oti_handle, attr, 0);
188 rc = filter_finish_transno(exp, oti, 0);
192 rc = fsfilt_commit_async(obd,inode,oti->oti_handle,wait_handle);
198 /* these are our hacks to keep our directio/bh IO coherent with ext3's
199 * page cache use. Most notably ext3 reads file data into the page
200 * cache when it is zeroing the tail of partial-block truncates and
201 * leaves it there, sometimes generating io from it at later truncates.
202 * Someday very soon we'll be performing our brw_kiovec() IO to and
203 * from the page cache. */
205 check_pending_bhs(iobuf->blocks, iobuf->nr_pages, inode->i_dev,
206 1 << inode->i_blkbits);
208 rc = filemap_fdatasync(inode->i_mapping);
210 rc = fsync_inode_data_buffers(inode);
212 rc = filemap_fdatawait(inode->i_mapping);
216 /* be careful to call this after fsync_inode_data_buffers has waited
217 * for IO to complete before we evict it from the cache */
218 filter_clear_page_cache(inode, iobuf);
220 rc = fsfilt_send_bio(rw, obd, inode, iobuf);
222 CDEBUG(D_INFO, "tried to write %d pages, rc = %d\n",
223 iobuf->nr_pages, rc);
230 if (!committed && (rw == OBD_BRW_WRITE)) {
231 int err = fsfilt_commit_async(obd, inode,
232 oti->oti_handle, wait_handle);
233 oti->oti_handle = NULL;
235 CERROR("can't close transaction: %d\n", err);
237 * this is error path, so we prefer to return
238 * original error, not this one
242 switch(cleanup_phase) {
245 unlock_kiovec(1, &iobuf);
247 OBD_FREE(created, sizeof(*created) *
248 iobuf->nr_pages*blocks_per_page);
250 if (cleanup_phase != 3 && rw == OBD_BRW_WRITE)
254 CERROR("corrupt cleanup_phase (%d)?\n", cleanup_phase);
261 /* See if there are unallocated parts in given file region */
262 int filter_range_is_mapped(struct inode *inode, obd_size offset, int len)
264 int (*fs_bmap)(struct address_space *, long) =
265 inode->i_mapping->a_ops->bmap;
268 /* We can't know if the range is mapped already or not */
272 offset >>= inode->i_blkbits;
273 len >>= inode->i_blkbits;
275 for (j = 0; j < len; j++)
276 if (fs_bmap(inode->i_mapping, offset + j) == 0)
283 /* some kernels require alloc_kiovec callers to zero members through the use of
284 * map_user_kiobuf and unmap_.. we don't use those, so we have a little helper
285 * that makes sure we don't break the rules. */
286 static void clear_kiobuf(struct kiobuf *iobuf)
290 for (i = 0; i < iobuf->array_len; i++)
291 iobuf->maplist[i] = NULL;
298 int filter_alloc_iobuf(int rw, int num_pages, void **ret)
301 struct kiobuf *iobuf;
304 LASSERTF(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ, "%x\n", rw);
306 rc = alloc_kiovec(1, &iobuf);
310 rc = expand_kiobuf(iobuf, num_pages);
312 free_kiovec(1, &iobuf);
316 #ifdef HAVE_KIOBUF_DOVARY
317 iobuf->dovary = 0; /* this prevents corruption, not present in 2.4.20 */
324 void filter_free_iobuf(void *buf)
326 struct kiobuf *iobuf = buf;
329 free_kiovec(1, &iobuf);
332 int filter_iobuf_add_page(struct obd_device *obd, void *buf,
333 struct inode *inode, struct page *page)
335 struct kiobuf *iobuf = buf;
337 iobuf->maplist[iobuf->nr_pages++] = page;
338 iobuf->length += PAGE_SIZE;
343 int filter_commitrw_write(struct obd_export *exp, struct obdo *oa, int objcount,
344 struct obd_ioobj *obj, int niocount,
345 struct niobuf_local *res, struct obd_trans_info *oti,
348 struct obd_device *obd = exp->exp_obd;
349 struct lvfs_run_ctxt saved;
350 struct niobuf_local *lnb;
351 struct fsfilt_objinfo fso;
352 struct iattr iattr = { 0 };
354 struct inode *inode = NULL;
355 int i, n, cleanup_phase = 0, err;
356 unsigned long now = jiffies; /* DEBUGGING OST TIMEOUTS */
359 LASSERT(oti != NULL);
360 LASSERT(objcount == 1);
361 LASSERT(current->journal_info == NULL);
366 rc = filter_alloc_iobuf(OBD_BRW_WRITE, obj->ioo_bufcnt, &iobuf);
371 fso.fso_dentry = res->dentry;
372 fso.fso_bufcnt = obj->ioo_bufcnt;
373 inode = res->dentry->d_inode;
375 for (i = 0, lnb = res, n = 0; i < obj->ioo_bufcnt; i++, lnb++) {
378 /* If overwriting an existing block, we don't need a grant */
379 if (!(lnb->flags & OBD_BRW_GRANTED) && lnb->rc == -ENOSPC &&
380 filter_range_is_mapped(inode, lnb->offset, lnb->len))
383 if (lnb->rc) /* ENOSPC, network RPC error */
386 filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
387 /* We expect these pages to be in offset order, but we'll
389 this_size = lnb->offset + lnb->len;
390 if (this_size > iattr.ia_size)
391 iattr.ia_size = this_size;
394 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
398 oti->oti_handle = fsfilt_brw_start(obd, objcount, &fso, niocount, res,
400 if (IS_ERR(oti->oti_handle)) {
402 rc = PTR_ERR(oti->oti_handle);
403 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
404 "error starting transaction: rc = %d\n", rc);
405 oti->oti_handle = NULL;
409 if (time_after(jiffies, now + 15 * HZ))
410 CERROR("slow brw_start %lus\n", (jiffies - now) / HZ);
412 iattr_from_obdo(&iattr,oa,OBD_MD_FLATIME|OBD_MD_FLMTIME|OBD_MD_FLCTIME);
413 /* filter_direct_io drops i_sem */
414 rc = filter_direct_io(OBD_BRW_WRITE, res->dentry, iobuf, exp, &iattr,
417 obdo_from_inode(oa, inode, FILTER_VALID_FLAGS);
419 if (time_after(jiffies, now + 15 * HZ))
420 CERROR("slow direct_io %lus\n", (jiffies - now) / HZ);
422 err = fsfilt_commit_wait(obd, inode, wait_handle);
426 LASSERT(oti->oti_transno <= obd->obd_last_committed);
427 if (time_after(jiffies, now + 15 * HZ))
428 CERROR("slow commitrw commit %lus\n", (jiffies - now) / HZ);
430 filter_grant_commit(exp, niocount, res);
432 switch (cleanup_phase) {
434 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
435 LASSERT(current->journal_info == NULL);
437 filter_free_iobuf(iobuf);
439 filter_free_dio_pages(objcount, obj, niocount, res);