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 #define DEBUG_SUBSYSTEM S_FILTER
29 #include <linux/config.h>
30 #include <linux/module.h>
31 #include <linux/pagemap.h> // XXX kill me soon
32 #include <linux/version.h>
34 #include <linux/obd_class.h>
35 #include <linux/lustre_fsfilt.h>
36 #include "filter_internal.h"
38 static int filter_start_page_read(struct inode *inode, struct niobuf_local *lnb)
40 struct address_space *mapping = inode->i_mapping;
42 unsigned long index = lnb->offset >> PAGE_SHIFT;
45 page = grab_cache_page(mapping, index); /* locked page */
47 return lnb->rc = PTR_ERR(page);
51 if (inode->i_size < lnb->offset + lnb->len - 1)
52 lnb->rc = inode->i_size - lnb->offset;
56 if (PageUptodate(page)) {
61 rc = mapping->a_ops->readpage(NULL, page);
63 CERROR("page index %lu, rc = %d\n", index, rc);
65 page_cache_release(page);
72 static int filter_finish_page_read(struct niobuf_local *lnb)
74 if (lnb->page == NULL)
77 if (PageUptodate(lnb->page))
80 wait_on_page(lnb->page);
81 if (!PageUptodate(lnb->page)) {
82 CERROR("page index %lu/offset "LPX64" not uptodate\n",
83 lnb->page->index, lnb->offset);
84 GOTO(err_page, lnb->rc = -EIO);
86 if (PageError(lnb->page)) {
87 CERROR("page index %lu/offset "LPX64" has error\n",
88 lnb->page->index, lnb->offset);
89 GOTO(err_page, lnb->rc = -EIO);
95 page_cache_release(lnb->page);
100 static struct page *lustre_get_page_write(struct inode *inode,
103 struct address_space *mapping = inode->i_mapping;
107 page = grab_cache_page(mapping, index); /* locked page */
110 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
111 * a no-op for most filesystems, because we write the whole
112 * page. For partial-page I/O this will read in the page.
114 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
116 CERROR("page index %lu, rc = %d\n", index, rc);
119 GOTO(err_unlock, rc);
121 /* XXX not sure if we need this if we are overwriting page */
122 if (PageError(page)) {
123 CERROR("error on page index %lu, rc = %d\n", index, rc);
125 GOTO(err_unlock, rc = -EIO);
132 page_cache_release(page);
136 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
137 int wait_on_page_locked(struct page *page)
139 waitfor_one_page(page);
143 /* We should only change the file mtime (and not the ctime, like
144 * update_inode_times() in generic_file_write()) when we only change data. */
145 static inline void inode_update_time(struct inode *inode, int ctime_too)
147 time_t now = CURRENT_TIME;
148 if (inode->i_mtime == now && (!ctime_too || inode->i_ctime == now))
150 inode->i_mtime = now;
152 inode->i_ctime = now;
153 mark_inode_dirty_sync(inode);
157 static int lustre_commit_write(struct niobuf_local *lnb)
159 struct page *page = lnb->page;
160 unsigned from = lnb->offset & ~PAGE_MASK;
161 unsigned to = from + lnb->len;
162 struct inode *inode = page->mapping->host;
165 LASSERT(to <= PAGE_SIZE);
166 err = page->mapping->a_ops->commit_write(NULL, page, from, to);
167 #warning 2.4 folks: wait_on_page_locked does NOT return its error here.
168 if (!err && IS_SYNC(inode))
169 wait_on_page_locked(page);
170 //SetPageUptodate(page); // the client commit_write will do this
172 SetPageReferenced(page);
174 page_cache_release(page);
178 int filter_get_page_write(struct inode *inode, struct niobuf_local *lnb,
181 unsigned long index = lnb->offset >> PAGE_SHIFT;
182 struct address_space *mapping = inode->i_mapping;
186 //ASSERT_PAGE_INDEX(index, GOTO(err, rc = -EINVAL));
188 page = grab_cache_page_nowait(mapping, index); /* locked page */
190 page = grab_cache_page(mapping, index); /* locked page */
193 /* This page is currently locked, so get a temporary page instead. */
195 CDEBUG(D_INFO, "ino %lu page %ld locked\n", inode->i_ino,index);
196 page = alloc_pages(GFP_KERNEL, 0); /* locked page */
198 CERROR("no memory for a temp page\n");
199 GOTO(err, rc = -ENOMEM);
203 lnb->flags |= N_LOCAL_TEMP_PAGE;
204 } else if (!IS_ERR(page)) {
207 rc = mapping->a_ops->prepare_write(NULL, page,
208 lnb->offset & ~PAGE_MASK,
212 CERROR("page index %lu, rc = %d\n", index, rc);
213 GOTO(err_unlock, rc);
215 /* XXX not sure if we need this if we are overwriting page */
216 if (PageError(page)) {
217 CERROR("error on page index %lu, rc = %d\n", index, rc);
219 GOTO(err_unlock, rc = -EIO);
228 page_cache_release(page);
233 static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa,
234 int objcount, struct obd_ioobj *obj,
235 int niocount, struct niobuf_remote *nb,
236 struct niobuf_local *res,
237 struct obd_trans_info *oti)
239 struct obd_run_ctxt saved;
241 struct niobuf_remote *rnb;
242 struct niobuf_local *lnb;
243 struct fsfilt_objinfo *fso;
244 struct dentry *dentry;
246 int rc = 0, i, j, tot_bytes = 0;
247 unsigned long now = jiffies;
250 /* We are currently not supporting multi-obj BRW_READ RPCS at all */
251 LASSERT(objcount == 1);
253 OBD_ALLOC(fso, objcount * sizeof(*fso));
257 memset(res, 0, niocount * sizeof(*res));
259 push_ctxt(&saved, &exp->exp_obd->u.filter.fo_ctxt, NULL);
260 for (i = 0, o = obj; i < objcount; i++, o++) {
261 struct filter_dentry_data *fdd;
262 LASSERT(o->ioo_bufcnt);
264 dentry = filter_oa2dentry(exp->exp_obd, oa);
266 GOTO(out_objinfo, rc = PTR_ERR(dentry));
268 if (dentry->d_inode == NULL) {
269 CERROR("trying to BRW to non-existent file "LPU64"\n",
272 GOTO(out_objinfo, rc = -ENOENT);
275 fso[i].fso_dentry = dentry;
276 fso[i].fso_bufcnt = o->ioo_bufcnt;
278 fdd = dentry->d_fsdata;
279 if (fdd == NULL || !atomic_read(&fdd->fdd_open_count))
280 CDEBUG(D_PAGE, "I/O to unopened object "LPU64"\n",
284 if (time_after(jiffies, now + 15 * HZ))
285 CERROR("slow prep setup %lus\n", (jiffies - now) / HZ);
287 for (i = 0, o = obj, rnb = nb, lnb = res; i < objcount; i++, o++) {
288 dentry = fso[i].fso_dentry;
289 inode = dentry->d_inode;
291 for (j = 0; j < o->ioo_bufcnt; j++, rnb++, lnb++) {
293 lnb->dentry = dentry;
295 lnb->dentry = dget(dentry);
297 lnb->offset = rnb->offset;
299 lnb->flags = rnb->flags;
300 lnb->start = jiffies;
302 if (inode->i_size <= rnb->offset) {
303 /* If there's no more data, abort early.
304 * lnb->page == NULL and lnb->rc == 0, so it's
305 * easy to detect later. */
310 rc = filter_start_page_read(inode, lnb);
314 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
315 "page err %u@"LPU64" %u/%u %p: rc %d\n",
316 lnb->len, lnb->offset, j, o->ioo_bufcnt,
322 tot_bytes += lnb->rc;
323 if (lnb->rc < lnb->len)
324 break; /* short read */
328 if (time_after(jiffies, now + 15 * HZ))
329 CERROR("slow prep get page %lus\n", (jiffies - now) / HZ);
331 lprocfs_counter_add(exp->exp_obd->obd_stats, LPROC_FILTER_READ_BYTES,
333 while (lnb-- > res) {
334 rc = filter_finish_page_read(lnb);
336 CERROR("error page %u@"LPU64" %u %p: rc %d\n", lnb->len,
337 lnb->offset, (int)(lnb - res), lnb->dentry, rc);
343 if (time_after(jiffies, now + 15 * HZ))
344 CERROR("slow prep finish page %lus\n", (jiffies - now) / HZ);
348 OBD_FREE(fso, objcount * sizeof(*fso));
349 /* we saved the journal handle into oti->oti_handle instead */
350 current->journal_info = NULL;
351 pop_ctxt(&saved, &exp->exp_obd->u.filter.fo_ctxt, NULL);
355 while (lnb-- > res) {
356 page_cache_release(lnb->page);
359 goto out; /* dropped the dentry refs already (one per page) */
362 for (i = 0; i < objcount && fso[i].fso_dentry; i++)
363 f_dput(fso[i].fso_dentry);
367 /* We need to balance prepare_write() calls with commit_write() calls.
368 * If the page has been prepared, but we have no data for it, we don't
369 * want to overwrite valid data on disk, but we still need to zero out
370 * data for space which was newly allocated. Like part of what happens
371 * in __block_prepare_write() for newly allocated blocks.
373 * XXX currently __block_prepare_write() creates buffers for all the
374 * pages, and the filesystems mark these buffers as BH_New if they
375 * were newly allocated from disk. We use the BH_New flag similarly. */
376 static int filter_commit_write(struct niobuf_local *lnb, int err)
378 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
380 unsigned block_start, block_end;
381 struct buffer_head *bh, *head = lnb->page->buffers;
382 unsigned blocksize = head->b_size;
384 /* debugging: just seeing if this ever happens */
385 CDEBUG(err == -ENOSPC ? D_INODE : D_ERROR,
386 "called for ino %lu:%lu on err %d\n",
387 lnb->page->mapping->host->i_ino, lnb->page->index, err);
389 /* Currently one buffer per page, but in the future... */
390 for (bh = head, block_start = 0; bh != head || !block_start;
391 block_start = block_end, bh = bh->b_this_page) {
392 block_end = block_start + blocksize;
393 if (buffer_new(bh)) {
394 memset(kmap(lnb->page) + block_start, 0,
401 return lustre_commit_write(lnb);
404 /* If we ever start to support multi-object BRW RPCs, we will need to get locks
405 * on mulitple inodes. That isn't all, because there still exists the
406 * possibility of a truncate starting a new transaction while holding the ext3
407 * rwsem = write while some writes (which have started their transactions here)
408 * blocking on the ext3 rwsem = read => lock inversion.
410 * The handling gets very ugly when dealing with locked pages. It may be easier
411 * to just get rid of the locked page code (which has problems of its own) and
412 * either discover we do not need it anymore (i.e. it was a symptom of another
413 * bug) or ensure we get the page locks in an appropriate order. */
414 static int filter_preprw_write(int cmd, struct obd_export *exp, struct obdo *oa,
415 int objcount, struct obd_ioobj *obj,
416 int niocount, struct niobuf_remote *nb,
417 struct niobuf_local *res,
418 struct obd_trans_info *oti)
420 struct obd_run_ctxt saved;
422 struct niobuf_remote *rnb;
423 struct niobuf_local *lnb;
424 struct fsfilt_objinfo *fso;
425 struct dentry *dentry;
426 int pglocked = 0, rc = 0, i, j, tot_bytes = 0;
427 unsigned long now = jiffies;
429 LASSERT(objcount == 1);
431 OBD_ALLOC(fso, objcount * sizeof(*fso));
435 memset(res, 0, niocount * sizeof(*res));
437 push_ctxt(&saved, &exp->exp_obd->u.filter.fo_ctxt, NULL);
438 for (i = 0, o = obj; i < objcount; i++, o++) {
439 struct filter_dentry_data *fdd;
440 LASSERT(o->ioo_bufcnt);
442 dentry = filter_oa2dentry(exp->exp_obd, oa);
444 GOTO(out_objinfo, rc = PTR_ERR(dentry));
446 if (dentry->d_inode == NULL) {
447 CERROR("trying to BRW to non-existent file "LPU64"\n",
450 GOTO(out_objinfo, rc = -ENOENT);
453 fso[i].fso_dentry = dentry;
454 fso[i].fso_bufcnt = o->ioo_bufcnt;
456 down(&dentry->d_inode->i_sem);
457 fdd = dentry->d_fsdata;
458 if (fdd == NULL || !atomic_read(&fdd->fdd_open_count))
459 CDEBUG(D_PAGE, "I/O to unopened object "LPU64"\n",
463 if (time_after(jiffies, now + 15 * HZ))
464 CERROR("slow prep setup %lus\n", (jiffies - now) / HZ);
466 LASSERT(oti != NULL);
467 oti->oti_handle = fsfilt_brw_start(exp->exp_obd, objcount, fso,
469 if (IS_ERR(oti->oti_handle)) {
470 rc = PTR_ERR(oti->oti_handle);
471 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
472 "error starting transaction: rc = %d\n", rc);
473 oti->oti_handle = NULL;
474 GOTO(out_objinfo, rc);
477 for (i = 0, o = obj, rnb = nb, lnb = res; i < objcount; i++, o++) {
478 dentry = fso[i].fso_dentry;
479 for (j = 0; j < o->ioo_bufcnt; j++, rnb++, lnb++) {
481 lnb->dentry = dentry;
483 lnb->dentry = dget(dentry);
485 lnb->offset = rnb->offset;
487 lnb->flags = rnb->flags;
488 lnb->start = jiffies;
490 rc = filter_get_page_write(dentry->d_inode, lnb,
493 up(&dentry->d_inode->i_sem);
496 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
497 "page err %u@"LPU64" %u/%u %p: rc %d\n",
498 lnb->len, lnb->offset, j, o->ioo_bufcnt,
503 tot_bytes += lnb->len;
507 if (time_after(jiffies, now + 15 * HZ))
508 CERROR("slow prep get page %lus\n", (jiffies - now) / HZ);
510 lprocfs_counter_add(exp->exp_obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
515 OBD_FREE(fso, objcount * sizeof(*fso));
516 /* we saved the journal handle into oti->oti_handle instead */
517 current->journal_info = NULL;
518 pop_ctxt(&saved, &exp->exp_obd->u.filter.fo_ctxt, NULL);
522 while (lnb-- > res) {
523 filter_commit_write(lnb, rc);
524 up(&lnb->dentry->d_inode->i_sem);
527 filter_finish_transno(exp, oti, rc);
528 fsfilt_commit(exp->exp_obd,
529 filter_parent(exp->exp_obd,S_IFREG,obj->ioo_id)->d_inode,
531 goto out; /* dropped the dentry refs already (one per page) */
534 for (i = 0; i < objcount && fso[i].fso_dentry; i++) {
535 up(&fso[i].fso_dentry->d_inode->i_sem);
536 f_dput(fso[i].fso_dentry);
541 int filter_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
542 int objcount, struct obd_ioobj *obj, int niocount,
543 struct niobuf_remote *nb, struct niobuf_local *res,
544 struct obd_trans_info *oti)
546 if (cmd == OBD_BRW_WRITE)
547 return filter_preprw_write(cmd, exp, oa, objcount, obj,
548 niocount, nb, res, oti);
550 if (cmd == OBD_BRW_READ)
551 return filter_preprw_read(cmd, exp, oa, objcount, obj,
552 niocount, nb, res, oti);
559 /* It is highly unlikely that we would ever get an error here. The page we want
560 * to get was previously locked, so it had to have already allocated the space,
561 * and we were just writing over the same data, so there would be no hole in the
564 * XXX: possibility of a race with truncate could exist, need to check that.
565 * There are no guarantees w.r.t. write order even on a local filesystem,
566 * although the normal response would be to return the number of bytes
567 * successfully written and leave the rest to the app. */
568 static int filter_write_locked_page(struct niobuf_local *lnb)
571 void *lpage_addr, *lnb_addr;
575 lpage = lustre_get_page_write(lnb->dentry->d_inode, lnb->page->index);
578 CERROR("error getting locked page index %ld: rc = %d\n",
579 lnb->page->index, rc);
581 lustre_commit_write(lnb);
585 /* 2 kmaps == vanishingly small deadlock opportunity */
586 lpage_addr = kmap(lpage);
587 lnb_addr = kmap(lnb->page);
589 memcpy(lpage_addr, lnb_addr, PAGE_SIZE);
594 page_cache_release(lnb->page);
597 rc = lustre_commit_write(lnb);
599 CERROR("error committing locked page %ld: rc = %d\n",
600 lnb->page->index, rc);
604 int filter_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
605 int objcount, struct obd_ioobj *obj, int niocount,
606 struct niobuf_local *res, struct obd_trans_info *oti)
608 struct obd_run_ctxt saved;
610 struct niobuf_local *lnb;
611 struct obd_device *obd = exp->exp_obd;
612 int found_locked = 0, rc = 0, i;
613 int nested_trans = current->journal_info != NULL;
614 unsigned long now = jiffies; /* DEBUGGING OST TIMEOUTS */
617 push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
619 if (cmd & OBD_BRW_WRITE) {
621 LASSERT(current->journal_info == NULL ||
622 current->journal_info == oti->oti_handle);
623 current->journal_info = oti->oti_handle;
626 for (i = 0, o = obj, lnb = res; i < objcount; i++, o++) {
630 /* If all of the page reads were beyond EOF, let's pretend
631 * this read didn't really happen at all. */
632 if (lnb->dentry == NULL) {
633 oa->o_valid = OBD_MD_FLID|(oa->o_valid&OBD_MD_FLCKSUM);
637 inode = igrab(lnb->dentry->d_inode);
639 if (cmd & OBD_BRW_WRITE) {
640 /* FIXME: MULTI OBJECT BRW */
641 if (oa && oa->o_valid & (OBD_MD_FLMTIME|OBD_MD_FLCTIME))
642 obdo_refresh_inode(inode, oa, OBD_MD_FLATIME |
646 inode_update_time(lnb->dentry->d_inode, 1);
647 } else if (oa && oa->o_valid & OBD_MD_FLATIME) {
648 /* Note that we don't necessarily write this to disk */
649 obdo_refresh_inode(inode, oa, OBD_MD_FLATIME);
652 for (j = 0 ; j < o->ioo_bufcnt ; j++, lnb++) {
653 if (lnb->page == NULL) {
657 if (lnb->flags & N_LOCAL_TEMP_PAGE) {
662 if (time_after(jiffies, lnb->start + 15 * HZ))
663 CERROR("slow commitrw %lusi (%lus)\n",
664 (jiffies - lnb->start) / HZ,
665 (jiffies - now) / HZ);
667 if (cmd & OBD_BRW_WRITE) {
668 int err = filter_commit_write(lnb, 0);
673 page_cache_release(lnb->page);
677 if (time_after(jiffies, lnb->start + 15 * HZ))
678 CERROR("slow commit_write %lus (%lus)\n",
679 (jiffies - lnb->start) / HZ,
680 (jiffies - now) / HZ);
683 /* FIXME: MULTI OBJECT BRW */
685 oa->o_valid = OBD_MD_FLID|(oa->o_valid&OBD_MD_FLCKSUM);
686 obdo_from_inode(oa, inode, FILTER_VALID_FLAGS);
689 if (cmd & OBD_BRW_WRITE)
695 for (i = 0, o = obj, lnb = res; found_locked > 0 && i < objcount;
699 for (j = 0 ; j < o->ioo_bufcnt ; j++, lnb++) {
701 if (!(lnb->flags & N_LOCAL_TEMP_PAGE))
704 if (time_after(jiffies, lnb->start + 15 * HZ))
705 CERROR("slow commitrw locked %lus (%lus)\n",
706 (jiffies - lnb->start) / HZ,
707 (jiffies - now) / HZ);
709 err = filter_write_locked_page(lnb);
715 if (time_after(jiffies, lnb->start + 15 * HZ))
716 CERROR("slow commit_write locked %lus (%lus)\n",
717 (jiffies - lnb->start) / HZ,
718 (jiffies - now) / HZ);
722 if (cmd & OBD_BRW_WRITE) {
723 /* We just want any dentry for the commit, for now */
724 struct dentry *dparent = filter_parent(obd, S_IFREG, 0);
727 rc = filter_finish_transno(exp, oti, rc);
728 err = fsfilt_commit(obd, dparent->d_inode, oti->oti_handle,
733 LASSERT(oti->oti_transno <= obd->obd_last_committed);
734 if (time_after(jiffies, now + 15 * HZ))
735 CERROR("slow commitrw commit %lus\n", (jiffies-now)/HZ);
738 LASSERT(nested_trans || current->journal_info == NULL);
739 pop_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
743 int filter_brw(int cmd, struct lustre_handle *conn, struct obdo *oa,
744 struct lov_stripe_md *lsm, obd_count oa_bufs,
745 struct brw_page *pga, struct obd_trans_info *oti)
747 struct obd_export *exp;
748 struct obd_ioobj ioo;
749 struct niobuf_local *lnb;
750 struct niobuf_remote *rnb;
755 exp = class_conn2export(conn);
757 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",conn->cookie);
761 OBD_ALLOC(lnb, oa_bufs * sizeof(struct niobuf_local));
762 OBD_ALLOC(rnb, oa_bufs * sizeof(struct niobuf_remote));
764 if (lnb == NULL || rnb == NULL)
765 GOTO(out, ret = -ENOMEM);
767 for (i = 0; i < oa_bufs; i++) {
768 rnb[i].offset = pga[i].off;
769 rnb[i].len = pga[i].count;
772 ioo.ioo_id = oa->o_id;
774 ioo.ioo_type = oa->o_mode & S_IFMT;
775 ioo.ioo_bufcnt = oa_bufs;
777 ret = filter_preprw(cmd, exp, oa, 1, &ioo, oa_bufs, rnb, lnb, oti);
781 for (i = 0; i < oa_bufs; i++) {
782 void *virt = kmap(pga[i].pg);
783 obd_off off = pga[i].off & ~PAGE_MASK;
784 void *addr = kmap(lnb[i].page);
786 /* 2 kmaps == vanishingly small deadlock opportunity */
788 if (cmd & OBD_BRW_WRITE)
789 memcpy(addr + off, virt + off, pga[i].count);
791 memcpy(virt + off, addr + off, pga[i].count);
797 ret = filter_commitrw(cmd, exp, oa, 1, &ioo, oa_bufs, lnb, oti);
801 OBD_FREE(lnb, oa_bufs * sizeof(struct niobuf_local));
803 OBD_FREE(rnb, oa_bufs * sizeof(struct niobuf_remote));
804 class_export_put(exp);