1 #include <linux/ceph/ceph_debug.h>
3 #include <linux/backing-dev.h>
6 #include <linux/pagemap.h>
7 #include <linux/writeback.h> /* generic_writepages */
8 #include <linux/slab.h>
9 #include <linux/pagevec.h>
10 #include <linux/task_io_accounting_ops.h>
13 #include "mds_client.h"
15 #include <linux/ceph/osd_client.h>
18 * Ceph address space ops.
20 * There are a few funny things going on here.
22 * The page->private field is used to reference a struct
23 * ceph_snap_context for _every_ dirty page. This indicates which
24 * snapshot the page was logically dirtied in, and thus which snap
25 * context needs to be associated with the osd write during writeback.
27 * Similarly, struct ceph_inode_info maintains a set of counters to
28 * count dirty pages on the inode. In the absence of snapshots,
29 * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
31 * When a snapshot is taken (that is, when the client receives
32 * notification that a snapshot was taken), each inode with caps and
33 * with dirty pages (dirty pages implies there is a cap) gets a new
34 * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
35 * order, new snaps go to the tail). The i_wrbuffer_ref_head count is
36 * moved to capsnap->dirty. (Unless a sync write is currently in
37 * progress. In that case, the capsnap is said to be "pending", new
38 * writes cannot start, and the capsnap isn't "finalized" until the
39 * write completes (or fails) and a final size/mtime for the inode for
40 * that snap can be settled upon.) i_wrbuffer_ref_head is reset to 0.
42 * On writeback, we must submit writes to the osd IN SNAP ORDER. So,
43 * we look for the first capsnap in i_cap_snaps and write out pages in
44 * that snap context _only_. Then we move on to the next capsnap,
45 * eventually reaching the "live" or "head" context (i.e., pages that
46 * are not yet snapped) and are writing the most recently dirtied
49 * Invalidate and so forth must take care to ensure the dirty page
50 * accounting is preserved.
53 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
54 #define CONGESTION_OFF_THRESH(congestion_kb) \
55 (CONGESTION_ON_THRESH(congestion_kb) - \
56 (CONGESTION_ON_THRESH(congestion_kb) >> 2))
58 static inline struct ceph_snap_context *page_snap_context(struct page *page)
60 if (PagePrivate(page))
61 return (void *)page->private;
66 * Dirty a page. Optimistically adjust accounting, on the assumption
67 * that we won't race with invalidate. If we do, readjust.
69 static int ceph_set_page_dirty(struct page *page)
71 struct address_space *mapping = page->mapping;
73 struct ceph_inode_info *ci;
74 struct ceph_snap_context *snapc;
77 if (unlikely(!mapping))
78 return !TestSetPageDirty(page);
80 if (PageDirty(page)) {
81 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
82 mapping->host, page, page->index);
83 BUG_ON(!PagePrivate(page));
87 inode = mapping->host;
88 ci = ceph_inode(inode);
91 * Note that we're grabbing a snapc ref here without holding
94 snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context);
97 spin_lock(&ci->i_ceph_lock);
98 if (ci->i_head_snapc == NULL)
99 ci->i_head_snapc = ceph_get_snap_context(snapc);
100 ++ci->i_wrbuffer_ref_head;
101 if (ci->i_wrbuffer_ref == 0)
103 ++ci->i_wrbuffer_ref;
104 dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
105 "snapc %p seq %lld (%d snaps)\n",
106 mapping->host, page, page->index,
107 ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
108 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
109 snapc, snapc->seq, snapc->num_snaps);
110 spin_unlock(&ci->i_ceph_lock);
113 * Reference snap context in page->private. Also set
114 * PagePrivate so that we get invalidatepage callback.
116 BUG_ON(PagePrivate(page));
117 page->private = (unsigned long)snapc;
118 SetPagePrivate(page);
120 ret = __set_page_dirty_nobuffers(page);
121 WARN_ON(!PageLocked(page));
122 WARN_ON(!page->mapping);
128 * If we are truncating the full page (i.e. offset == 0), adjust the
129 * dirty page counters appropriately. Only called if there is private
132 static void ceph_invalidatepage(struct page *page, unsigned int offset,
136 struct ceph_inode_info *ci;
137 struct ceph_snap_context *snapc = page_snap_context(page);
139 inode = page->mapping->host;
140 ci = ceph_inode(inode);
142 if (offset != 0 || length != PAGE_CACHE_SIZE) {
143 dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n",
144 inode, page, page->index, offset, length);
148 ceph_invalidate_fscache_page(inode, page);
150 if (!PagePrivate(page))
154 * We can get non-dirty pages here due to races between
155 * set_page_dirty and truncate_complete_page; just spit out a
156 * warning, in case we end up with accounting problems later.
158 if (!PageDirty(page))
159 pr_err("%p invalidatepage %p page not dirty\n", inode, page);
161 ClearPageChecked(page);
163 dout("%p invalidatepage %p idx %lu full dirty page\n",
164 inode, page, page->index);
166 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
167 ceph_put_snap_context(snapc);
169 ClearPagePrivate(page);
172 static int ceph_releasepage(struct page *page, gfp_t g)
174 struct inode *inode = page->mapping ? page->mapping->host : NULL;
175 dout("%p releasepage %p idx %lu\n", inode, page, page->index);
176 WARN_ON(PageDirty(page));
178 /* Can we release the page from the cache? */
179 if (!ceph_release_fscache_page(page, g))
182 return !PagePrivate(page);
186 * read a single page, without unlocking it.
188 static int readpage_nounlock(struct file *filp, struct page *page)
190 struct inode *inode = file_inode(filp);
191 struct ceph_inode_info *ci = ceph_inode(inode);
192 struct ceph_osd_client *osdc =
193 &ceph_inode_to_client(inode)->client->osdc;
195 u64 off = page_offset(page);
196 u64 len = PAGE_CACHE_SIZE;
198 if (off >= i_size_read(inode)) {
199 zero_user_segment(page, 0, PAGE_CACHE_SIZE);
200 SetPageUptodate(page);
204 if (ci->i_inline_version != CEPH_INLINE_NONE) {
206 * Uptodate inline data should have been added
207 * into page cache while getting Fcr caps.
211 zero_user_segment(page, 0, PAGE_CACHE_SIZE);
212 SetPageUptodate(page);
216 err = ceph_readpage_from_fscache(inode, page);
220 dout("readpage inode %p file %p page %p index %lu\n",
221 inode, filp, page, page->index);
222 err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
224 ci->i_truncate_seq, ci->i_truncate_size,
230 ceph_fscache_readpage_cancel(inode, page);
233 if (err < PAGE_CACHE_SIZE)
234 /* zero fill remainder of page */
235 zero_user_segment(page, err, PAGE_CACHE_SIZE);
237 flush_dcache_page(page);
239 SetPageUptodate(page);
240 ceph_readpage_to_fscache(inode, page);
243 return err < 0 ? err : 0;
246 static int ceph_readpage(struct file *filp, struct page *page)
248 int r = readpage_nounlock(filp, page);
254 * Finish an async read(ahead) op.
256 static void finish_read(struct ceph_osd_request *req, struct ceph_msg *msg)
258 struct inode *inode = req->r_inode;
259 struct ceph_osd_data *osd_data;
260 int rc = req->r_result;
261 int bytes = le32_to_cpu(msg->hdr.data_len);
265 dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
267 /* unlock all pages, zeroing any data we didn't read */
268 osd_data = osd_req_op_extent_osd_data(req, 0);
269 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
270 num_pages = calc_pages_for((u64)osd_data->alignment,
271 (u64)osd_data->length);
272 for (i = 0; i < num_pages; i++) {
273 struct page *page = osd_data->pages[i];
277 if (bytes < (int)PAGE_CACHE_SIZE) {
278 /* zero (remainder of) page */
279 int s = bytes < 0 ? 0 : bytes;
280 zero_user_segment(page, s, PAGE_CACHE_SIZE);
282 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
284 flush_dcache_page(page);
285 SetPageUptodate(page);
286 ceph_readpage_to_fscache(inode, page);
289 page_cache_release(page);
290 bytes -= PAGE_CACHE_SIZE;
292 kfree(osd_data->pages);
295 static void ceph_unlock_page_vector(struct page **pages, int num_pages)
299 for (i = 0; i < num_pages; i++)
300 unlock_page(pages[i]);
304 * start an async read(ahead) operation. return nr_pages we submitted
305 * a read for on success, or negative error code.
307 static int start_read(struct inode *inode, struct list_head *page_list, int max)
309 struct ceph_osd_client *osdc =
310 &ceph_inode_to_client(inode)->client->osdc;
311 struct ceph_inode_info *ci = ceph_inode(inode);
312 struct page *page = list_entry(page_list->prev, struct page, lru);
313 struct ceph_vino vino;
314 struct ceph_osd_request *req;
323 off = (u64) page_offset(page);
326 next_index = page->index;
327 list_for_each_entry_reverse(page, page_list, lru) {
328 if (page->index != next_index)
332 if (max && nr_pages == max)
335 len = nr_pages << PAGE_CACHE_SHIFT;
336 dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
338 vino = ceph_vino(inode);
339 req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
340 0, 1, CEPH_OSD_OP_READ,
341 CEPH_OSD_FLAG_READ, NULL,
342 ci->i_truncate_seq, ci->i_truncate_size,
347 /* build page vector */
348 nr_pages = calc_pages_for(0, len);
349 pages = kmalloc(sizeof(*pages) * nr_pages, GFP_NOFS);
353 for (i = 0; i < nr_pages; ++i) {
354 page = list_entry(page_list->prev, struct page, lru);
355 BUG_ON(PageLocked(page));
356 list_del(&page->lru);
358 dout("start_read %p adding %p idx %lu\n", inode, page,
360 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
362 ceph_fscache_uncache_page(inode, page);
363 page_cache_release(page);
364 dout("start_read %p add_to_page_cache failed %p\n",
371 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
372 req->r_callback = finish_read;
373 req->r_inode = inode;
375 ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
377 dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
378 ret = ceph_osdc_start_request(osdc, req, false);
381 ceph_osdc_put_request(req);
385 ceph_unlock_page_vector(pages, nr_pages);
386 ceph_release_page_vector(pages, nr_pages);
388 ceph_osdc_put_request(req);
394 * Read multiple pages. Leave pages we don't read + unlock in page_list;
395 * the caller (VM) cleans them up.
397 static int ceph_readpages(struct file *file, struct address_space *mapping,
398 struct list_head *page_list, unsigned nr_pages)
400 struct inode *inode = file_inode(file);
401 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
405 if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE)
408 rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list,
414 if (fsc->mount_options->rsize >= PAGE_CACHE_SIZE)
415 max = (fsc->mount_options->rsize + PAGE_CACHE_SIZE - 1)
418 dout("readpages %p file %p nr_pages %d max %d\n", inode,
421 while (!list_empty(page_list)) {
422 rc = start_read(inode, page_list, max);
428 ceph_fscache_readpages_cancel(inode, page_list);
430 dout("readpages %p file %p ret %d\n", inode, file, rc);
435 * Get ref for the oldest snapc for an inode with dirty data... that is, the
436 * only snap context we are allowed to write back.
438 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
441 struct ceph_inode_info *ci = ceph_inode(inode);
442 struct ceph_snap_context *snapc = NULL;
443 struct ceph_cap_snap *capsnap = NULL;
445 spin_lock(&ci->i_ceph_lock);
446 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
447 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
448 capsnap->context, capsnap->dirty_pages);
449 if (capsnap->dirty_pages) {
450 snapc = ceph_get_snap_context(capsnap->context);
452 *snap_size = capsnap->size;
456 if (!snapc && ci->i_wrbuffer_ref_head) {
457 snapc = ceph_get_snap_context(ci->i_head_snapc);
458 dout(" head snapc %p has %d dirty pages\n",
459 snapc, ci->i_wrbuffer_ref_head);
461 spin_unlock(&ci->i_ceph_lock);
466 * Write a single page, but leave the page locked.
468 * If we get a write error, set the page error bit, but still adjust the
469 * dirty page accounting (i.e., page is no longer dirty).
471 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
474 struct ceph_inode_info *ci;
475 struct ceph_fs_client *fsc;
476 struct ceph_osd_client *osdc;
477 struct ceph_snap_context *snapc, *oldest;
478 loff_t page_off = page_offset(page);
480 u64 truncate_size, snap_size = 0;
482 int err = 0, len = PAGE_CACHE_SIZE;
484 dout("writepage %p idx %lu\n", page, page->index);
486 if (!page->mapping || !page->mapping->host) {
487 dout("writepage %p - no mapping\n", page);
490 inode = page->mapping->host;
491 ci = ceph_inode(inode);
492 fsc = ceph_inode_to_client(inode);
493 osdc = &fsc->client->osdc;
495 /* verify this is a writeable snap context */
496 snapc = page_snap_context(page);
498 dout("writepage %p page %p not dirty?\n", inode, page);
501 oldest = get_oldest_context(inode, &snap_size);
502 if (snapc->seq > oldest->seq) {
503 dout("writepage %p page %p snapc %p not writeable - noop\n",
505 /* we should only noop if called by kswapd */
506 WARN_ON((current->flags & PF_MEMALLOC) == 0);
507 ceph_put_snap_context(oldest);
510 ceph_put_snap_context(oldest);
512 spin_lock(&ci->i_ceph_lock);
513 truncate_seq = ci->i_truncate_seq;
514 truncate_size = ci->i_truncate_size;
516 snap_size = i_size_read(inode);
517 spin_unlock(&ci->i_ceph_lock);
519 /* is this a partial page at end of file? */
520 if (page_off >= snap_size) {
521 dout("%p page eof %llu\n", page, snap_size);
524 if (snap_size < page_off + len)
525 len = snap_size - page_off;
527 dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
528 inode, page, page->index, page_off, len, snapc);
530 writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
532 CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
533 set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
535 ceph_readpage_to_fscache(inode, page);
537 set_page_writeback(page);
538 err = ceph_osdc_writepages(osdc, ceph_vino(inode),
539 &ci->i_layout, snapc,
541 truncate_seq, truncate_size,
542 &inode->i_mtime, &page, 1);
544 dout("writepage setting page/mapping error %d %p\n", err, page);
546 mapping_set_error(&inode->i_data, err);
548 wbc->pages_skipped++;
550 dout("writepage cleaned page %p\n", page);
551 err = 0; /* vfs expects us to return 0 */
554 ClearPagePrivate(page);
555 end_page_writeback(page);
556 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
557 ceph_put_snap_context(snapc); /* page's reference */
562 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
565 struct inode *inode = page->mapping->host;
568 err = writepage_nounlock(page, wbc);
576 * lame release_pages helper. release_pages() isn't exported to
579 static void ceph_release_pages(struct page **pages, int num)
584 pagevec_init(&pvec, 0);
585 for (i = 0; i < num; i++) {
586 if (pagevec_add(&pvec, pages[i]) == 0)
587 pagevec_release(&pvec);
589 pagevec_release(&pvec);
593 * async writeback completion handler.
595 * If we get an error, set the mapping error bit, but not the individual
598 static void writepages_finish(struct ceph_osd_request *req,
599 struct ceph_msg *msg)
601 struct inode *inode = req->r_inode;
602 struct ceph_inode_info *ci = ceph_inode(inode);
603 struct ceph_osd_data *osd_data;
608 struct ceph_snap_context *snapc = req->r_snapc;
609 struct address_space *mapping = inode->i_mapping;
610 int rc = req->r_result;
611 u64 bytes = req->r_ops[0].extent.length;
612 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
614 unsigned issued = ceph_caps_issued(ci);
616 osd_data = osd_req_op_extent_osd_data(req, 0);
617 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
618 num_pages = calc_pages_for((u64)osd_data->alignment,
619 (u64)osd_data->length);
622 * Assume we wrote the pages we originally sent. The
623 * osd might reply with fewer pages if our writeback
624 * raced with a truncation and was adjusted at the osd,
625 * so don't believe the reply.
630 mapping_set_error(mapping, rc);
632 dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
633 inode, rc, bytes, wrote);
635 /* clean all pages */
636 for (i = 0; i < num_pages; i++) {
637 page = osd_data->pages[i];
639 WARN_ON(!PageUptodate(page));
642 atomic_long_dec_return(&fsc->writeback_count);
644 CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
645 clear_bdi_congested(&fsc->backing_dev_info,
648 ceph_put_snap_context(page_snap_context(page));
650 ClearPagePrivate(page);
651 dout("unlocking %d %p\n", i, page);
652 end_page_writeback(page);
655 * We lost the cache cap, need to truncate the page before
656 * it is unlocked, otherwise we'd truncate it later in the
657 * page truncation thread, possibly losing some data that
660 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0)
661 generic_error_remove_page(inode->i_mapping, page);
665 dout("%p wrote+cleaned %d pages\n", inode, wrote);
666 ceph_put_wrbuffer_cap_refs(ci, num_pages, snapc);
668 ceph_release_pages(osd_data->pages, num_pages);
669 if (osd_data->pages_from_pool)
670 mempool_free(osd_data->pages,
671 ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
673 kfree(osd_data->pages);
674 ceph_osdc_put_request(req);
678 * initiate async writeback
680 static int ceph_writepages_start(struct address_space *mapping,
681 struct writeback_control *wbc)
683 struct inode *inode = mapping->host;
684 struct ceph_inode_info *ci = ceph_inode(inode);
685 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
686 struct ceph_vino vino = ceph_vino(inode);
687 pgoff_t index, start, end;
690 pgoff_t max_pages = 0, max_pages_ever = 0;
691 struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
695 unsigned wsize = 1 << inode->i_blkbits;
696 struct ceph_osd_request *req = NULL;
698 u64 truncate_size, snap_size;
702 * Include a 'sync' in the OSD request if this is a data
703 * integrity write (e.g., O_SYNC write or fsync()), or if our
704 * cap is being revoked.
706 if ((wbc->sync_mode == WB_SYNC_ALL) ||
707 ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
709 dout("writepages_start %p dosync=%d (mode=%s)\n",
711 wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
712 (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
714 if (fsc->mount_state == CEPH_MOUNT_SHUTDOWN) {
715 pr_warn("writepage_start %p on forced umount\n", inode);
716 return -EIO; /* we're in a forced umount, don't write! */
718 if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
719 wsize = fsc->mount_options->wsize;
720 if (wsize < PAGE_CACHE_SIZE)
721 wsize = PAGE_CACHE_SIZE;
722 max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
724 pagevec_init(&pvec, 0);
726 /* where to start/end? */
727 if (wbc->range_cyclic) {
728 start = mapping->writeback_index; /* Start from prev offset */
730 dout(" cyclic, start at %lu\n", start);
732 start = wbc->range_start >> PAGE_CACHE_SHIFT;
733 end = wbc->range_end >> PAGE_CACHE_SHIFT;
734 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
737 dout(" not cyclic, %lu to %lu\n", start, end);
742 /* find oldest snap context with dirty data */
743 ceph_put_snap_context(snapc);
745 snapc = get_oldest_context(inode, &snap_size);
747 /* hmm, why does writepages get called when there
749 dout(" no snap context with dirty data?\n");
753 snap_size = i_size_read(inode);
754 dout(" oldest snapc is %p seq %lld (%d snaps)\n",
755 snapc, snapc->seq, snapc->num_snaps);
757 spin_lock(&ci->i_ceph_lock);
758 truncate_seq = ci->i_truncate_seq;
759 truncate_size = ci->i_truncate_size;
761 snap_size = i_size_read(inode);
762 spin_unlock(&ci->i_ceph_lock);
764 if (last_snapc && snapc != last_snapc) {
765 /* if we switched to a newer snapc, restart our scan at the
766 * start of the original file range. */
767 dout(" snapc differs from last pass, restarting at %lu\n",
773 while (!done && index <= end) {
777 int pvec_pages, locked_pages;
778 struct page **pages = NULL;
779 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
787 max_pages = max_pages_ever;
791 want = min(end - index,
792 min((pgoff_t)PAGEVEC_SIZE,
793 max_pages - (pgoff_t)locked_pages) - 1)
795 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
798 dout("pagevec_lookup_tag got %d\n", pvec_pages);
799 if (!pvec_pages && !locked_pages)
801 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
802 page = pvec.pages[i];
803 dout("? %p idx %lu\n", page, page->index);
804 if (locked_pages == 0)
805 lock_page(page); /* first page */
806 else if (!trylock_page(page))
809 /* only dirty pages, or our accounting breaks */
810 if (unlikely(!PageDirty(page)) ||
811 unlikely(page->mapping != mapping)) {
812 dout("!dirty or !mapping %p\n", page);
816 if (!wbc->range_cyclic && page->index > end) {
817 dout("end of range %p\n", page);
822 if (next && (page->index != next)) {
823 dout("not consecutive %p\n", page);
827 if (wbc->sync_mode != WB_SYNC_NONE) {
828 dout("waiting on writeback %p\n", page);
829 wait_on_page_writeback(page);
831 if (page_offset(page) >= snap_size) {
832 dout("%p page eof %llu\n", page, snap_size);
837 if (PageWriteback(page)) {
838 dout("%p under writeback\n", page);
843 /* only if matching snap context */
844 pgsnapc = page_snap_context(page);
845 if (pgsnapc->seq > snapc->seq) {
846 dout("page snapc %p %lld > oldest %p %lld\n",
847 pgsnapc, pgsnapc->seq, snapc, snapc->seq);
850 continue; /* keep looking for snap */
854 if (!clear_page_dirty_for_io(page)) {
855 dout("%p !clear_page_dirty_for_io\n", page);
861 * We have something to write. If this is
862 * the first locked page this time through,
863 * allocate an osd request and a page array
866 if (locked_pages == 0) {
868 /* prepare async write request */
869 offset = (u64)page_offset(page);
871 req = ceph_osdc_new_request(&fsc->client->osdc,
876 CEPH_OSD_FLAG_WRITE |
877 CEPH_OSD_FLAG_ONDISK,
879 truncate_size, true);
887 osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC);
889 req->r_callback = writepages_finish;
890 req->r_inode = inode;
892 max_pages = calc_pages_for(0, (u64)len);
893 pages = kmalloc(max_pages * sizeof (*pages),
896 pool = fsc->wb_pagevec_pool;
897 pages = mempool_alloc(pool, GFP_NOFS);
902 /* note position of first page in pvec */
905 dout("%p will write page %p idx %lu\n",
906 inode, page, page->index);
909 atomic_long_inc_return(&fsc->writeback_count);
910 if (writeback_stat > CONGESTION_ON_THRESH(
911 fsc->mount_options->congestion_kb)) {
912 set_bdi_congested(&fsc->backing_dev_info,
916 set_page_writeback(page);
917 pages[locked_pages] = page;
919 next = page->index + 1;
922 /* did we get anything? */
924 goto release_pvec_pages;
927 BUG_ON(!locked_pages || first < 0);
929 if (pvec_pages && i == pvec_pages &&
930 locked_pages < max_pages) {
931 dout("reached end pvec, trying for more\n");
932 pagevec_reinit(&pvec);
936 /* shift unused pages over in the pvec... we
937 * will need to release them below. */
938 for (j = i; j < pvec_pages; j++) {
939 dout(" pvec leftover page %p\n",
941 pvec.pages[j-i+first] = pvec.pages[j];
946 /* Format the osd request message and submit the write */
948 offset = page_offset(pages[0]);
949 len = min(snap_size - offset,
950 (u64)locked_pages << PAGE_CACHE_SHIFT);
951 dout("writepages got %d pages at %llu~%llu\n",
952 locked_pages, offset, len);
954 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
957 pages = NULL; /* request message now owns the pages array */
960 /* Update the write op length in case we changed it */
962 osd_req_op_extent_update(req, 0, len);
964 vino = ceph_vino(inode);
965 ceph_osdc_build_request(req, offset, snapc, vino.snap,
968 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
974 wbc->nr_to_write -= locked_pages;
975 if (wbc->nr_to_write <= 0)
979 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
980 pvec.nr ? pvec.pages[0] : NULL);
981 pagevec_release(&pvec);
983 if (locked_pages && !done)
987 if (should_loop && !done) {
988 /* more to do; loop back to beginning of file */
989 dout("writepages looping back to beginning of file\n");
995 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
996 mapping->writeback_index = index;
1000 ceph_osdc_put_request(req);
1001 ceph_put_snap_context(snapc);
1002 dout("writepages done, rc = %d\n", rc);
1009 * See if a given @snapc is either writeable, or already written.
1011 static int context_is_writeable_or_written(struct inode *inode,
1012 struct ceph_snap_context *snapc)
1014 struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
1015 int ret = !oldest || snapc->seq <= oldest->seq;
1017 ceph_put_snap_context(oldest);
1022 * We are only allowed to write into/dirty the page if the page is
1023 * clean, or already dirty within the same snap context.
1025 * called with page locked.
1026 * return success with page locked,
1027 * or any failure (incl -EAGAIN) with page unlocked.
1029 static int ceph_update_writeable_page(struct file *file,
1030 loff_t pos, unsigned len,
1033 struct inode *inode = file_inode(file);
1034 struct ceph_inode_info *ci = ceph_inode(inode);
1035 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1036 loff_t page_off = pos & PAGE_CACHE_MASK;
1037 int pos_in_page = pos & ~PAGE_CACHE_MASK;
1038 int end_in_page = pos_in_page + len;
1041 struct ceph_snap_context *snapc, *oldest;
1044 /* writepages currently holds page lock, but if we change that later, */
1045 wait_on_page_writeback(page);
1047 /* check snap context */
1048 BUG_ON(!ci->i_snap_realm);
1049 down_read(&mdsc->snap_rwsem);
1050 BUG_ON(!ci->i_snap_realm->cached_context);
1051 snapc = page_snap_context(page);
1052 if (snapc && snapc != ci->i_head_snapc) {
1054 * this page is already dirty in another (older) snap
1055 * context! is it writeable now?
1057 oldest = get_oldest_context(inode, NULL);
1058 up_read(&mdsc->snap_rwsem);
1060 if (snapc->seq > oldest->seq) {
1061 ceph_put_snap_context(oldest);
1062 dout(" page %p snapc %p not current or oldest\n",
1065 * queue for writeback, and wait for snapc to
1066 * be writeable or written
1068 snapc = ceph_get_snap_context(snapc);
1070 ceph_queue_writeback(inode);
1071 r = wait_event_interruptible(ci->i_cap_wq,
1072 context_is_writeable_or_written(inode, snapc));
1073 ceph_put_snap_context(snapc);
1074 if (r == -ERESTARTSYS)
1078 ceph_put_snap_context(oldest);
1080 /* yay, writeable, do it now (without dropping page lock) */
1081 dout(" page %p snapc %p not current, but oldest\n",
1083 if (!clear_page_dirty_for_io(page))
1085 r = writepage_nounlock(page, NULL);
1091 if (PageUptodate(page)) {
1092 dout(" page %p already uptodate\n", page);
1097 if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
1100 /* past end of file? */
1101 i_size = inode->i_size; /* caller holds i_mutex */
1103 if (page_off >= i_size ||
1104 (pos_in_page == 0 && (pos+len) >= i_size &&
1105 end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
1106 dout(" zeroing %p 0 - %d and %d - %d\n",
1107 page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
1108 zero_user_segments(page,
1110 end_in_page, PAGE_CACHE_SIZE);
1114 /* we need to read it. */
1115 up_read(&mdsc->snap_rwsem);
1116 r = readpage_nounlock(file, page);
1126 * We are only allowed to write into/dirty the page if the page is
1127 * clean, or already dirty within the same snap context.
1129 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1130 loff_t pos, unsigned len, unsigned flags,
1131 struct page **pagep, void **fsdata)
1133 struct inode *inode = file_inode(file);
1135 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1140 page = grab_cache_page_write_begin(mapping, index, 0);
1145 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1146 inode, page, (int)pos, (int)len);
1148 r = ceph_update_writeable_page(file, pos, len, page);
1149 } while (r == -EAGAIN);
1155 * we don't do anything in here that simple_write_end doesn't do
1156 * except adjust dirty page accounting and drop read lock on
1159 static int ceph_write_end(struct file *file, struct address_space *mapping,
1160 loff_t pos, unsigned len, unsigned copied,
1161 struct page *page, void *fsdata)
1163 struct inode *inode = file_inode(file);
1164 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1165 struct ceph_mds_client *mdsc = fsc->mdsc;
1166 unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1169 dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1170 inode, page, (int)pos, (int)copied, (int)len);
1172 /* zero the stale part of the page if we did a short copy */
1174 zero_user_segment(page, from+copied, len);
1176 /* did file size increase? */
1177 /* (no need for i_size_read(); we caller holds i_mutex */
1178 if (pos+copied > inode->i_size)
1179 check_cap = ceph_inode_set_size(inode, pos+copied);
1181 if (!PageUptodate(page))
1182 SetPageUptodate(page);
1184 set_page_dirty(page);
1187 up_read(&mdsc->snap_rwsem);
1188 page_cache_release(page);
1191 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1197 * we set .direct_IO to indicate direct io is supported, but since we
1198 * intercept O_DIRECT reads and writes early, this function should
1201 static ssize_t ceph_direct_io(int rw, struct kiocb *iocb,
1202 struct iov_iter *iter,
1209 const struct address_space_operations ceph_aops = {
1210 .readpage = ceph_readpage,
1211 .readpages = ceph_readpages,
1212 .writepage = ceph_writepage,
1213 .writepages = ceph_writepages_start,
1214 .write_begin = ceph_write_begin,
1215 .write_end = ceph_write_end,
1216 .set_page_dirty = ceph_set_page_dirty,
1217 .invalidatepage = ceph_invalidatepage,
1218 .releasepage = ceph_releasepage,
1219 .direct_IO = ceph_direct_io,
1226 static int ceph_filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1228 struct inode *inode = file_inode(vma->vm_file);
1229 struct ceph_inode_info *ci = ceph_inode(inode);
1230 struct ceph_file_info *fi = vma->vm_file->private_data;
1231 struct page *pinned_page = NULL;
1232 loff_t off = vmf->pgoff << PAGE_CACHE_SHIFT;
1235 dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1236 inode, ceph_vinop(inode), off, (size_t)PAGE_CACHE_SIZE);
1237 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1238 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1240 want = CEPH_CAP_FILE_CACHE;
1243 ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want,
1244 -1, &got, &pinned_page);
1247 if (ret != -ERESTARTSYS) {
1249 return VM_FAULT_SIGBUS;
1252 dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1253 inode, off, (size_t)PAGE_CACHE_SIZE, ceph_cap_string(got));
1255 if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1256 ci->i_inline_version == CEPH_INLINE_NONE)
1257 ret = filemap_fault(vma, vmf);
1261 dout("filemap_fault %p %llu~%zd dropping cap refs on %s ret %d\n",
1262 inode, off, (size_t)PAGE_CACHE_SIZE, ceph_cap_string(got), ret);
1264 page_cache_release(pinned_page);
1265 ceph_put_cap_refs(ci, got);
1270 /* read inline data */
1271 if (off >= PAGE_CACHE_SIZE) {
1272 /* does not support inline data > PAGE_SIZE */
1273 ret = VM_FAULT_SIGBUS;
1276 struct address_space *mapping = inode->i_mapping;
1277 struct page *page = find_or_create_page(mapping, 0,
1278 mapping_gfp_mask(mapping) &
1284 ret1 = __ceph_do_getattr(inode, page,
1285 CEPH_STAT_CAP_INLINE_DATA, true);
1286 if (ret1 < 0 || off >= i_size_read(inode)) {
1288 page_cache_release(page);
1289 ret = VM_FAULT_SIGBUS;
1292 if (ret1 < PAGE_CACHE_SIZE)
1293 zero_user_segment(page, ret1, PAGE_CACHE_SIZE);
1295 flush_dcache_page(page);
1296 SetPageUptodate(page);
1298 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1301 dout("filemap_fault %p %llu~%zd read inline data ret %d\n",
1302 inode, off, (size_t)PAGE_CACHE_SIZE, ret);
1307 * Reuse write_begin here for simplicity.
1309 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1311 struct inode *inode = file_inode(vma->vm_file);
1312 struct ceph_inode_info *ci = ceph_inode(inode);
1313 struct ceph_file_info *fi = vma->vm_file->private_data;
1314 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1315 struct page *page = vmf->page;
1316 loff_t off = page_offset(page);
1317 loff_t size = i_size_read(inode);
1321 if (ci->i_inline_version != CEPH_INLINE_NONE) {
1322 struct page *locked_page = NULL;
1327 ret = ceph_uninline_data(vma->vm_file, locked_page);
1329 unlock_page(locked_page);
1331 return VM_FAULT_SIGBUS;
1334 if (off + PAGE_CACHE_SIZE <= size)
1335 len = PAGE_CACHE_SIZE;
1337 len = size & ~PAGE_CACHE_MASK;
1339 dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1340 inode, ceph_vinop(inode), off, len, size);
1341 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1342 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1344 want = CEPH_CAP_FILE_BUFFER;
1347 ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, off + len,
1351 if (ret != -ERESTARTSYS) {
1353 return VM_FAULT_SIGBUS;
1356 dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1357 inode, off, len, ceph_cap_string(got));
1359 /* Update time before taking page lock */
1360 file_update_time(vma->vm_file);
1364 ret = VM_FAULT_NOPAGE;
1366 (page->mapping != inode->i_mapping))
1369 ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1371 /* success. we'll keep the page locked. */
1372 set_page_dirty(page);
1373 up_read(&mdsc->snap_rwsem);
1374 ret = VM_FAULT_LOCKED;
1379 ret = VM_FAULT_SIGBUS;
1382 if (ret != VM_FAULT_LOCKED)
1384 if (ret == VM_FAULT_LOCKED ||
1385 ci->i_inline_version != CEPH_INLINE_NONE) {
1387 spin_lock(&ci->i_ceph_lock);
1388 ci->i_inline_version = CEPH_INLINE_NONE;
1389 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR);
1390 spin_unlock(&ci->i_ceph_lock);
1392 __mark_inode_dirty(inode, dirty);
1395 dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %d\n",
1396 inode, off, len, ceph_cap_string(got), ret);
1397 ceph_put_cap_refs(ci, got);
1402 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1403 char *data, size_t len)
1405 struct address_space *mapping = inode->i_mapping;
1411 if (i_size_read(inode) == 0)
1413 page = find_or_create_page(mapping, 0,
1414 mapping_gfp_mask(mapping) & ~__GFP_FS);
1417 if (PageUptodate(page)) {
1419 page_cache_release(page);
1424 dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1425 inode, ceph_vinop(inode), len, locked_page);
1428 void *kaddr = kmap_atomic(page);
1429 memcpy(kaddr, data, len);
1430 kunmap_atomic(kaddr);
1433 if (page != locked_page) {
1434 if (len < PAGE_CACHE_SIZE)
1435 zero_user_segment(page, len, PAGE_CACHE_SIZE);
1437 flush_dcache_page(page);
1439 SetPageUptodate(page);
1441 page_cache_release(page);
1445 int ceph_uninline_data(struct file *filp, struct page *locked_page)
1447 struct inode *inode = file_inode(filp);
1448 struct ceph_inode_info *ci = ceph_inode(inode);
1449 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1450 struct ceph_osd_request *req;
1451 struct page *page = NULL;
1452 u64 len, inline_version;
1454 bool from_pagecache = false;
1456 spin_lock(&ci->i_ceph_lock);
1457 inline_version = ci->i_inline_version;
1458 spin_unlock(&ci->i_ceph_lock);
1460 dout("uninline_data %p %llx.%llx inline_version %llu\n",
1461 inode, ceph_vinop(inode), inline_version);
1463 if (inline_version == 1 || /* initial version, no data */
1464 inline_version == CEPH_INLINE_NONE)
1469 WARN_ON(!PageUptodate(page));
1470 } else if (ceph_caps_issued(ci) &
1471 (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1472 page = find_get_page(inode->i_mapping, 0);
1474 if (PageUptodate(page)) {
1475 from_pagecache = true;
1478 page_cache_release(page);
1485 len = i_size_read(inode);
1486 if (len > PAGE_CACHE_SIZE)
1487 len = PAGE_CACHE_SIZE;
1489 page = __page_cache_alloc(GFP_NOFS);
1494 err = __ceph_do_getattr(inode, page,
1495 CEPH_STAT_CAP_INLINE_DATA, true);
1497 /* no inline data */
1498 if (err == -ENODATA)
1505 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1506 ceph_vino(inode), 0, &len, 0, 1,
1508 CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
1509 ci->i_snap_realm->cached_context,
1516 ceph_osdc_build_request(req, 0, NULL, CEPH_NOSNAP, &inode->i_mtime);
1517 err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1519 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1520 ceph_osdc_put_request(req);
1524 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1525 ceph_vino(inode), 0, &len, 1, 3,
1527 CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
1528 ci->i_snap_realm->cached_context,
1529 ci->i_truncate_seq, ci->i_truncate_size,
1536 osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1538 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1539 "inline_version", &inline_version,
1540 sizeof(inline_version),
1541 CEPH_OSD_CMPXATTR_OP_GT,
1542 CEPH_OSD_CMPXATTR_MODE_U64);
1546 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1547 "inline_version", &inline_version,
1548 sizeof(inline_version), 0, 0);
1552 ceph_osdc_build_request(req, 0, NULL, CEPH_NOSNAP, &inode->i_mtime);
1553 err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1555 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1557 ceph_osdc_put_request(req);
1558 if (err == -ECANCELED)
1561 if (page && page != locked_page) {
1562 if (from_pagecache) {
1564 page_cache_release(page);
1566 __free_pages(page, 0);
1569 dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1570 inode, ceph_vinop(inode), inline_version, err);
1574 static struct vm_operations_struct ceph_vmops = {
1575 .fault = ceph_filemap_fault,
1576 .page_mkwrite = ceph_page_mkwrite,
1579 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1581 struct address_space *mapping = file->f_mapping;
1583 if (!mapping->a_ops->readpage)
1585 file_accessed(file);
1586 vma->vm_ops = &ceph_vmops;