13d14c5d2SYehuda Sadeh #include <linux/ceph/ceph_debug.h> 21d3576fdSSage Weil 31d3576fdSSage Weil #include <linux/backing-dev.h> 41d3576fdSSage Weil #include <linux/fs.h> 51d3576fdSSage Weil #include <linux/mm.h> 61d3576fdSSage Weil #include <linux/pagemap.h> 71d3576fdSSage Weil #include <linux/writeback.h> /* generic_writepages */ 85a0e3ad6STejun Heo #include <linux/slab.h> 91d3576fdSSage Weil #include <linux/pagevec.h> 101d3576fdSSage Weil #include <linux/task_io_accounting_ops.h> 111d3576fdSSage Weil 121d3576fdSSage Weil #include "super.h" 133d14c5d2SYehuda Sadeh #include "mds_client.h" 143d14c5d2SYehuda Sadeh #include <linux/ceph/osd_client.h> 151d3576fdSSage Weil 161d3576fdSSage Weil /* 171d3576fdSSage Weil * Ceph address space ops. 181d3576fdSSage Weil * 191d3576fdSSage Weil * There are a few funny things going on here. 201d3576fdSSage Weil * 211d3576fdSSage Weil * The page->private field is used to reference a struct 221d3576fdSSage Weil * ceph_snap_context for _every_ dirty page. This indicates which 231d3576fdSSage Weil * snapshot the page was logically dirtied in, and thus which snap 241d3576fdSSage Weil * context needs to be associated with the osd write during writeback. 251d3576fdSSage Weil * 261d3576fdSSage Weil * Similarly, struct ceph_inode_info maintains a set of counters to 2725985edcSLucas De Marchi * count dirty pages on the inode. In the absence of snapshots, 281d3576fdSSage Weil * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count. 291d3576fdSSage Weil * 301d3576fdSSage Weil * When a snapshot is taken (that is, when the client receives 311d3576fdSSage Weil * notification that a snapshot was taken), each inode with caps and 321d3576fdSSage Weil * with dirty pages (dirty pages implies there is a cap) gets a new 331d3576fdSSage Weil * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending 341d3576fdSSage Weil * order, new snaps go to the tail). The i_wrbuffer_ref_head count is 351d3576fdSSage Weil * moved to capsnap->dirty. (Unless a sync write is currently in 361d3576fdSSage Weil * progress. In that case, the capsnap is said to be "pending", new 371d3576fdSSage Weil * writes cannot start, and the capsnap isn't "finalized" until the 381d3576fdSSage Weil * write completes (or fails) and a final size/mtime for the inode for 391d3576fdSSage Weil * that snap can be settled upon.) i_wrbuffer_ref_head is reset to 0. 401d3576fdSSage Weil * 411d3576fdSSage Weil * On writeback, we must submit writes to the osd IN SNAP ORDER. So, 421d3576fdSSage Weil * we look for the first capsnap in i_cap_snaps and write out pages in 431d3576fdSSage Weil * that snap context _only_. Then we move on to the next capsnap, 441d3576fdSSage Weil * eventually reaching the "live" or "head" context (i.e., pages that 451d3576fdSSage Weil * are not yet snapped) and are writing the most recently dirtied 461d3576fdSSage Weil * pages. 471d3576fdSSage Weil * 481d3576fdSSage Weil * Invalidate and so forth must take care to ensure the dirty page 491d3576fdSSage Weil * accounting is preserved. 501d3576fdSSage Weil */ 511d3576fdSSage Weil 522baba250SYehuda Sadeh #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10)) 532baba250SYehuda Sadeh #define CONGESTION_OFF_THRESH(congestion_kb) \ 542baba250SYehuda Sadeh (CONGESTION_ON_THRESH(congestion_kb) - \ 552baba250SYehuda Sadeh (CONGESTION_ON_THRESH(congestion_kb) >> 2)) 562baba250SYehuda Sadeh 5761600ef8SYan, Zheng static inline struct ceph_snap_context *page_snap_context(struct page *page) 5861600ef8SYan, Zheng { 5961600ef8SYan, Zheng if (PagePrivate(page)) 6061600ef8SYan, Zheng return (void *)page->private; 6161600ef8SYan, Zheng return NULL; 6261600ef8SYan, Zheng } 631d3576fdSSage Weil 641d3576fdSSage Weil /* 651d3576fdSSage Weil * Dirty a page. Optimistically adjust accounting, on the assumption 661d3576fdSSage Weil * that we won't race with invalidate. If we do, readjust. 671d3576fdSSage Weil */ 681d3576fdSSage Weil static int ceph_set_page_dirty(struct page *page) 691d3576fdSSage Weil { 701d3576fdSSage Weil struct address_space *mapping = page->mapping; 711d3576fdSSage Weil struct inode *inode; 721d3576fdSSage Weil struct ceph_inode_info *ci; 731d3576fdSSage Weil int undo = 0; 741d3576fdSSage Weil struct ceph_snap_context *snapc; 751d3576fdSSage Weil 761d3576fdSSage Weil if (unlikely(!mapping)) 771d3576fdSSage Weil return !TestSetPageDirty(page); 781d3576fdSSage Weil 791d3576fdSSage Weil if (TestSetPageDirty(page)) { 801d3576fdSSage Weil dout("%p set_page_dirty %p idx %lu -- already dirty\n", 811d3576fdSSage Weil mapping->host, page, page->index); 821d3576fdSSage Weil return 0; 831d3576fdSSage Weil } 841d3576fdSSage Weil 851d3576fdSSage Weil inode = mapping->host; 861d3576fdSSage Weil ci = ceph_inode(inode); 871d3576fdSSage Weil 881d3576fdSSage Weil /* 891d3576fdSSage Weil * Note that we're grabbing a snapc ref here without holding 901d3576fdSSage Weil * any locks! 911d3576fdSSage Weil */ 921d3576fdSSage Weil snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context); 931d3576fdSSage Weil 941d3576fdSSage Weil /* dirty the head */ 95be655596SSage Weil spin_lock(&ci->i_ceph_lock); 967d8cb26dSSage Weil if (ci->i_head_snapc == NULL) 971d3576fdSSage Weil ci->i_head_snapc = ceph_get_snap_context(snapc); 981d3576fdSSage Weil ++ci->i_wrbuffer_ref_head; 991d3576fdSSage Weil if (ci->i_wrbuffer_ref == 0) 1000444d76aSDave Chinner ihold(inode); 1011d3576fdSSage Weil ++ci->i_wrbuffer_ref; 1021d3576fdSSage Weil dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d " 1031d3576fdSSage Weil "snapc %p seq %lld (%d snaps)\n", 1041d3576fdSSage Weil mapping->host, page, page->index, 1051d3576fdSSage Weil ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1, 1061d3576fdSSage Weil ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head, 1071d3576fdSSage Weil snapc, snapc->seq, snapc->num_snaps); 108be655596SSage Weil spin_unlock(&ci->i_ceph_lock); 1091d3576fdSSage Weil 1101d3576fdSSage Weil /* now adjust page */ 1111d3576fdSSage Weil spin_lock_irq(&mapping->tree_lock); 1121d3576fdSSage Weil if (page->mapping) { /* Race with truncate? */ 1131d3576fdSSage Weil WARN_ON_ONCE(!PageUptodate(page)); 114679ceaceSMichael Rubin account_page_dirtied(page, page->mapping); 1151d3576fdSSage Weil radix_tree_tag_set(&mapping->page_tree, 1161d3576fdSSage Weil page_index(page), PAGECACHE_TAG_DIRTY); 1171d3576fdSSage Weil 1181d3576fdSSage Weil /* 1191d3576fdSSage Weil * Reference snap context in page->private. Also set 1201d3576fdSSage Weil * PagePrivate so that we get invalidatepage callback. 1211d3576fdSSage Weil */ 1221d3576fdSSage Weil page->private = (unsigned long)snapc; 1231d3576fdSSage Weil SetPagePrivate(page); 1241d3576fdSSage Weil } else { 1251d3576fdSSage Weil dout("ANON set_page_dirty %p (raced truncate?)\n", page); 1261d3576fdSSage Weil undo = 1; 1271d3576fdSSage Weil } 1281d3576fdSSage Weil 1291d3576fdSSage Weil spin_unlock_irq(&mapping->tree_lock); 1301d3576fdSSage Weil 1311d3576fdSSage Weil if (undo) 1321d3576fdSSage Weil /* whoops, we failed to dirty the page */ 1331d3576fdSSage Weil ceph_put_wrbuffer_cap_refs(ci, 1, snapc); 1341d3576fdSSage Weil 1351d3576fdSSage Weil __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); 1361d3576fdSSage Weil 1371d3576fdSSage Weil BUG_ON(!PageDirty(page)); 1381d3576fdSSage Weil return 1; 1391d3576fdSSage Weil } 1401d3576fdSSage Weil 1411d3576fdSSage Weil /* 1421d3576fdSSage Weil * If we are truncating the full page (i.e. offset == 0), adjust the 1431d3576fdSSage Weil * dirty page counters appropriately. Only called if there is private 1441d3576fdSSage Weil * data on the page. 1451d3576fdSSage Weil */ 1461d3576fdSSage Weil static void ceph_invalidatepage(struct page *page, unsigned long offset) 1471d3576fdSSage Weil { 1484ce1e9adSAlexander Beregalov struct inode *inode; 1491d3576fdSSage Weil struct ceph_inode_info *ci; 15061600ef8SYan, Zheng struct ceph_snap_context *snapc = page_snap_context(page); 1511d3576fdSSage Weil 1521d3576fdSSage Weil BUG_ON(!PageLocked(page)); 1531d3576fdSSage Weil BUG_ON(!PagePrivate(page)); 1541d3576fdSSage Weil BUG_ON(!page->mapping); 1551d3576fdSSage Weil 1564ce1e9adSAlexander Beregalov inode = page->mapping->host; 1574ce1e9adSAlexander Beregalov 1581d3576fdSSage Weil /* 1591d3576fdSSage Weil * We can get non-dirty pages here due to races between 1601d3576fdSSage Weil * set_page_dirty and truncate_complete_page; just spit out a 1611d3576fdSSage Weil * warning, in case we end up with accounting problems later. 1621d3576fdSSage Weil */ 1631d3576fdSSage Weil if (!PageDirty(page)) 1641d3576fdSSage Weil pr_err("%p invalidatepage %p page not dirty\n", inode, page); 1651d3576fdSSage Weil 1661d3576fdSSage Weil if (offset == 0) 1671d3576fdSSage Weil ClearPageChecked(page); 1681d3576fdSSage Weil 1691d3576fdSSage Weil ci = ceph_inode(inode); 1701d3576fdSSage Weil if (offset == 0) { 1711d3576fdSSage Weil dout("%p invalidatepage %p idx %lu full dirty page %lu\n", 1721d3576fdSSage Weil inode, page, page->index, offset); 1731d3576fdSSage Weil ceph_put_wrbuffer_cap_refs(ci, 1, snapc); 1741d3576fdSSage Weil ceph_put_snap_context(snapc); 1751d3576fdSSage Weil page->private = 0; 1761d3576fdSSage Weil ClearPagePrivate(page); 1771d3576fdSSage Weil } else { 1781d3576fdSSage Weil dout("%p invalidatepage %p idx %lu partial dirty page\n", 1791d3576fdSSage Weil inode, page, page->index); 1801d3576fdSSage Weil } 1811d3576fdSSage Weil } 1821d3576fdSSage Weil 1831d3576fdSSage Weil /* just a sanity check */ 1841d3576fdSSage Weil static int ceph_releasepage(struct page *page, gfp_t g) 1851d3576fdSSage Weil { 1861d3576fdSSage Weil struct inode *inode = page->mapping ? page->mapping->host : NULL; 1871d3576fdSSage Weil dout("%p releasepage %p idx %lu\n", inode, page, page->index); 1881d3576fdSSage Weil WARN_ON(PageDirty(page)); 1891d3576fdSSage Weil WARN_ON(PagePrivate(page)); 1901d3576fdSSage Weil return 0; 1911d3576fdSSage Weil } 1921d3576fdSSage Weil 1931d3576fdSSage Weil /* 1941d3576fdSSage Weil * read a single page, without unlocking it. 1951d3576fdSSage Weil */ 1961d3576fdSSage Weil static int readpage_nounlock(struct file *filp, struct page *page) 1971d3576fdSSage Weil { 198496ad9aaSAl Viro struct inode *inode = file_inode(filp); 1991d3576fdSSage Weil struct ceph_inode_info *ci = ceph_inode(inode); 2003d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 2013d14c5d2SYehuda Sadeh &ceph_inode_to_client(inode)->client->osdc; 2021d3576fdSSage Weil int err = 0; 2031d3576fdSSage Weil u64 len = PAGE_CACHE_SIZE; 2041d3576fdSSage Weil 2051d3576fdSSage Weil dout("readpage inode %p file %p page %p index %lu\n", 2061d3576fdSSage Weil inode, filp, page, page->index); 2071d3576fdSSage Weil err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout, 2086285bc23SAlex Elder (u64) page_offset(page), &len, 2091d3576fdSSage Weil ci->i_truncate_seq, ci->i_truncate_size, 210b7495fc2SSage Weil &page, 1, 0); 2111d3576fdSSage Weil if (err == -ENOENT) 2121d3576fdSSage Weil err = 0; 2131d3576fdSSage Weil if (err < 0) { 2141d3576fdSSage Weil SetPageError(page); 2151d3576fdSSage Weil goto out; 2161d3576fdSSage Weil } else if (err < PAGE_CACHE_SIZE) { 2171d3576fdSSage Weil /* zero fill remainder of page */ 2181d3576fdSSage Weil zero_user_segment(page, err, PAGE_CACHE_SIZE); 2191d3576fdSSage Weil } 2201d3576fdSSage Weil SetPageUptodate(page); 2211d3576fdSSage Weil 2221d3576fdSSage Weil out: 2231d3576fdSSage Weil return err < 0 ? err : 0; 2241d3576fdSSage Weil } 2251d3576fdSSage Weil 2261d3576fdSSage Weil static int ceph_readpage(struct file *filp, struct page *page) 2271d3576fdSSage Weil { 2281d3576fdSSage Weil int r = readpage_nounlock(filp, page); 2291d3576fdSSage Weil unlock_page(page); 2301d3576fdSSage Weil return r; 2311d3576fdSSage Weil } 2321d3576fdSSage Weil 2331d3576fdSSage Weil /* 2347c272194SSage Weil * Finish an async read(ahead) op. 2351d3576fdSSage Weil */ 2367c272194SSage Weil static void finish_read(struct ceph_osd_request *req, struct ceph_msg *msg) 2371d3576fdSSage Weil { 2387c272194SSage Weil struct inode *inode = req->r_inode; 23987060c10SAlex Elder struct ceph_osd_data *osd_data; 2401b83bef2SSage Weil int rc = req->r_result; 2411b83bef2SSage Weil int bytes = le32_to_cpu(msg->hdr.data_len); 242e0c59487SAlex Elder int num_pages; 2437c272194SSage Weil int i; 2447c272194SSage Weil 2457c272194SSage Weil dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes); 2467c272194SSage Weil 2477c272194SSage Weil /* unlock all pages, zeroing any data we didn't read */ 248a4ce40a9SAlex Elder osd_data = osd_req_op_extent_osd_data(req, 0, false); 24987060c10SAlex Elder BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES); 25087060c10SAlex Elder num_pages = calc_pages_for((u64)osd_data->alignment, 25187060c10SAlex Elder (u64)osd_data->length); 252e0c59487SAlex Elder for (i = 0; i < num_pages; i++) { 25387060c10SAlex Elder struct page *page = osd_data->pages[i]; 2547c272194SSage Weil 2557c272194SSage Weil if (bytes < (int)PAGE_CACHE_SIZE) { 2567c272194SSage Weil /* zero (remainder of) page */ 2577c272194SSage Weil int s = bytes < 0 ? 0 : bytes; 2587c272194SSage Weil zero_user_segment(page, s, PAGE_CACHE_SIZE); 2597c272194SSage Weil } 2607c272194SSage Weil dout("finish_read %p uptodate %p idx %lu\n", inode, page, 2617c272194SSage Weil page->index); 2627c272194SSage Weil flush_dcache_page(page); 2637c272194SSage Weil SetPageUptodate(page); 2647c272194SSage Weil unlock_page(page); 2657c272194SSage Weil page_cache_release(page); 2660fff87ecSAlex Elder bytes -= PAGE_CACHE_SIZE; 2677c272194SSage Weil } 26887060c10SAlex Elder kfree(osd_data->pages); 2697c272194SSage Weil } 2707c272194SSage Weil 2718884d53dSDavid Zafman static void ceph_unlock_page_vector(struct page **pages, int num_pages) 2728884d53dSDavid Zafman { 2738884d53dSDavid Zafman int i; 2748884d53dSDavid Zafman 2758884d53dSDavid Zafman for (i = 0; i < num_pages; i++) 2768884d53dSDavid Zafman unlock_page(pages[i]); 2778884d53dSDavid Zafman } 2788884d53dSDavid Zafman 2797c272194SSage Weil /* 2807c272194SSage Weil * start an async read(ahead) operation. return nr_pages we submitted 2817c272194SSage Weil * a read for on success, or negative error code. 2827c272194SSage Weil */ 2830d66a487SSage Weil static int start_read(struct inode *inode, struct list_head *page_list, int max) 2847c272194SSage Weil { 2857c272194SSage Weil struct ceph_osd_client *osdc = 2867c272194SSage Weil &ceph_inode_to_client(inode)->client->osdc; 2877c272194SSage Weil struct ceph_inode_info *ci = ceph_inode(inode); 2887c272194SSage Weil struct page *page = list_entry(page_list->prev, struct page, lru); 289acead002SAlex Elder struct ceph_vino vino; 2907c272194SSage Weil struct ceph_osd_request *req; 2917c272194SSage Weil u64 off; 2927c272194SSage Weil u64 len; 2937c272194SSage Weil int i; 2941d3576fdSSage Weil struct page **pages; 2957c272194SSage Weil pgoff_t next_index; 2967c272194SSage Weil int nr_pages = 0; 2977c272194SSage Weil int ret; 2987c272194SSage Weil 2996285bc23SAlex Elder off = (u64) page_offset(page); 3007c272194SSage Weil 3017c272194SSage Weil /* count pages */ 3027c272194SSage Weil next_index = page->index; 3037c272194SSage Weil list_for_each_entry_reverse(page, page_list, lru) { 3047c272194SSage Weil if (page->index != next_index) 3057c272194SSage Weil break; 3067c272194SSage Weil nr_pages++; 3077c272194SSage Weil next_index++; 3080d66a487SSage Weil if (max && nr_pages == max) 3090d66a487SSage Weil break; 3107c272194SSage Weil } 3117c272194SSage Weil len = nr_pages << PAGE_CACHE_SHIFT; 3127c272194SSage Weil dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages, 3137c272194SSage Weil off, len); 314acead002SAlex Elder vino = ceph_vino(inode); 315acead002SAlex Elder req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len, 31679528734SAlex Elder 1, CEPH_OSD_OP_READ, 317acead002SAlex Elder CEPH_OSD_FLAG_READ, NULL, 3187c272194SSage Weil ci->i_truncate_seq, ci->i_truncate_size, 319acead002SAlex Elder false); 3206816282dSSage Weil if (IS_ERR(req)) 3216816282dSSage Weil return PTR_ERR(req); 3221d3576fdSSage Weil 3231d3576fdSSage Weil /* build page vector */ 324cf7b7e14SAlex Elder nr_pages = calc_pages_for(0, len); 3257c272194SSage Weil pages = kmalloc(sizeof(*pages) * nr_pages, GFP_NOFS); 3267c272194SSage Weil ret = -ENOMEM; 3271d3576fdSSage Weil if (!pages) 3287c272194SSage Weil goto out; 3297c272194SSage Weil for (i = 0; i < nr_pages; ++i) { 3307c272194SSage Weil page = list_entry(page_list->prev, struct page, lru); 3317c272194SSage Weil BUG_ON(PageLocked(page)); 3327c272194SSage Weil list_del(&page->lru); 3331d3576fdSSage Weil 3347c272194SSage Weil dout("start_read %p adding %p idx %lu\n", inode, page, 3357c272194SSage Weil page->index); 3367c272194SSage Weil if (add_to_page_cache_lru(page, &inode->i_data, page->index, 3377c272194SSage Weil GFP_NOFS)) { 3387c272194SSage Weil page_cache_release(page); 3397c272194SSage Weil dout("start_read %p add_to_page_cache failed %p\n", 3407c272194SSage Weil inode, page); 3417c272194SSage Weil nr_pages = i; 3427c272194SSage Weil goto out_pages; 3431d3576fdSSage Weil } 3447c272194SSage Weil pages[i] = page; 3451d3576fdSSage Weil } 346a4ce40a9SAlex Elder osd_req_op_extent_osd_data_pages(req, 0, false, pages, len, 0, 34743bfe5deSAlex Elder false, false); 3487c272194SSage Weil req->r_callback = finish_read; 3497c272194SSage Weil req->r_inode = inode; 3507c272194SSage Weil 35179528734SAlex Elder ceph_osdc_build_request(req, off, NULL, vino.snap, NULL); 35202ee07d3SAlex Elder 3537c272194SSage Weil dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len); 3547c272194SSage Weil ret = ceph_osdc_start_request(osdc, req, false); 3557c272194SSage Weil if (ret < 0) 3567c272194SSage Weil goto out_pages; 3577c272194SSage Weil ceph_osdc_put_request(req); 3587c272194SSage Weil return nr_pages; 3597c272194SSage Weil 3607c272194SSage Weil out_pages: 3618884d53dSDavid Zafman ceph_unlock_page_vector(pages, nr_pages); 3627c272194SSage Weil ceph_release_page_vector(pages, nr_pages); 3637c272194SSage Weil out: 3647c272194SSage Weil ceph_osdc_put_request(req); 3657c272194SSage Weil return ret; 3661d3576fdSSage Weil } 3671d3576fdSSage Weil 3687c272194SSage Weil 3691d3576fdSSage Weil /* 3701d3576fdSSage Weil * Read multiple pages. Leave pages we don't read + unlock in page_list; 3711d3576fdSSage Weil * the caller (VM) cleans them up. 3721d3576fdSSage Weil */ 3731d3576fdSSage Weil static int ceph_readpages(struct file *file, struct address_space *mapping, 3741d3576fdSSage Weil struct list_head *page_list, unsigned nr_pages) 3751d3576fdSSage Weil { 376496ad9aaSAl Viro struct inode *inode = file_inode(file); 3770d66a487SSage Weil struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 3781d3576fdSSage Weil int rc = 0; 3790d66a487SSage Weil int max = 0; 3801d3576fdSSage Weil 3810d66a487SSage Weil if (fsc->mount_options->rsize >= PAGE_CACHE_SIZE) 3820d66a487SSage Weil max = (fsc->mount_options->rsize + PAGE_CACHE_SIZE - 1) 3830d66a487SSage Weil >> PAGE_SHIFT; 3840d66a487SSage Weil 3852794a82aSAlex Elder dout("readpages %p file %p nr_pages %d max %d\n", inode, 3862794a82aSAlex Elder file, nr_pages, 3870d66a487SSage Weil max); 3887c272194SSage Weil while (!list_empty(page_list)) { 3890d66a487SSage Weil rc = start_read(inode, page_list, max); 3901d3576fdSSage Weil if (rc < 0) 3911d3576fdSSage Weil goto out; 3927c272194SSage Weil BUG_ON(rc == 0); 3931d3576fdSSage Weil } 3941d3576fdSSage Weil out: 3957c272194SSage Weil dout("readpages %p file %p ret %d\n", inode, file, rc); 3961d3576fdSSage Weil return rc; 3971d3576fdSSage Weil } 3981d3576fdSSage Weil 3991d3576fdSSage Weil /* 4001d3576fdSSage Weil * Get ref for the oldest snapc for an inode with dirty data... that is, the 4011d3576fdSSage Weil * only snap context we are allowed to write back. 4021d3576fdSSage Weil */ 4036298a337SSage Weil static struct ceph_snap_context *get_oldest_context(struct inode *inode, 4041d3576fdSSage Weil u64 *snap_size) 4051d3576fdSSage Weil { 4061d3576fdSSage Weil struct ceph_inode_info *ci = ceph_inode(inode); 4071d3576fdSSage Weil struct ceph_snap_context *snapc = NULL; 4081d3576fdSSage Weil struct ceph_cap_snap *capsnap = NULL; 4091d3576fdSSage Weil 410be655596SSage Weil spin_lock(&ci->i_ceph_lock); 4111d3576fdSSage Weil list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) { 4121d3576fdSSage Weil dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap, 4131d3576fdSSage Weil capsnap->context, capsnap->dirty_pages); 4141d3576fdSSage Weil if (capsnap->dirty_pages) { 4151d3576fdSSage Weil snapc = ceph_get_snap_context(capsnap->context); 4161d3576fdSSage Weil if (snap_size) 4171d3576fdSSage Weil *snap_size = capsnap->size; 4181d3576fdSSage Weil break; 4191d3576fdSSage Weil } 4201d3576fdSSage Weil } 4217d8cb26dSSage Weil if (!snapc && ci->i_wrbuffer_ref_head) { 42280e755feSSage Weil snapc = ceph_get_snap_context(ci->i_head_snapc); 4231d3576fdSSage Weil dout(" head snapc %p has %d dirty pages\n", 4241d3576fdSSage Weil snapc, ci->i_wrbuffer_ref_head); 4251d3576fdSSage Weil } 426be655596SSage Weil spin_unlock(&ci->i_ceph_lock); 4271d3576fdSSage Weil return snapc; 4281d3576fdSSage Weil } 4291d3576fdSSage Weil 4301d3576fdSSage Weil /* 4311d3576fdSSage Weil * Write a single page, but leave the page locked. 4321d3576fdSSage Weil * 4331d3576fdSSage Weil * If we get a write error, set the page error bit, but still adjust the 4341d3576fdSSage Weil * dirty page accounting (i.e., page is no longer dirty). 4351d3576fdSSage Weil */ 4361d3576fdSSage Weil static int writepage_nounlock(struct page *page, struct writeback_control *wbc) 4371d3576fdSSage Weil { 4381d3576fdSSage Weil struct inode *inode; 4391d3576fdSSage Weil struct ceph_inode_info *ci; 4403d14c5d2SYehuda Sadeh struct ceph_fs_client *fsc; 4411d3576fdSSage Weil struct ceph_osd_client *osdc; 4426285bc23SAlex Elder loff_t page_off = page_offset(page); 4431d3576fdSSage Weil int len = PAGE_CACHE_SIZE; 4441d3576fdSSage Weil loff_t i_size; 4451d3576fdSSage Weil int err = 0; 4466298a337SSage Weil struct ceph_snap_context *snapc, *oldest; 4471d3576fdSSage Weil u64 snap_size = 0; 4482baba250SYehuda Sadeh long writeback_stat; 4491d3576fdSSage Weil 4501d3576fdSSage Weil dout("writepage %p idx %lu\n", page, page->index); 4511d3576fdSSage Weil 4521d3576fdSSage Weil if (!page->mapping || !page->mapping->host) { 4531d3576fdSSage Weil dout("writepage %p - no mapping\n", page); 4541d3576fdSSage Weil return -EFAULT; 4551d3576fdSSage Weil } 4561d3576fdSSage Weil inode = page->mapping->host; 4571d3576fdSSage Weil ci = ceph_inode(inode); 4583d14c5d2SYehuda Sadeh fsc = ceph_inode_to_client(inode); 4593d14c5d2SYehuda Sadeh osdc = &fsc->client->osdc; 4601d3576fdSSage Weil 4611d3576fdSSage Weil /* verify this is a writeable snap context */ 46261600ef8SYan, Zheng snapc = page_snap_context(page); 4631d3576fdSSage Weil if (snapc == NULL) { 4641d3576fdSSage Weil dout("writepage %p page %p not dirty?\n", inode, page); 4651d3576fdSSage Weil goto out; 4661d3576fdSSage Weil } 4676298a337SSage Weil oldest = get_oldest_context(inode, &snap_size); 4686298a337SSage Weil if (snapc->seq > oldest->seq) { 4691d3576fdSSage Weil dout("writepage %p page %p snapc %p not writeable - noop\n", 47061600ef8SYan, Zheng inode, page, snapc); 4711d3576fdSSage Weil /* we should only noop if called by kswapd */ 4721d3576fdSSage Weil WARN_ON((current->flags & PF_MEMALLOC) == 0); 4736298a337SSage Weil ceph_put_snap_context(oldest); 4741d3576fdSSage Weil goto out; 4751d3576fdSSage Weil } 4766298a337SSage Weil ceph_put_snap_context(oldest); 4771d3576fdSSage Weil 4781d3576fdSSage Weil /* is this a partial page at end of file? */ 4791d3576fdSSage Weil if (snap_size) 4801d3576fdSSage Weil i_size = snap_size; 4811d3576fdSSage Weil else 4821d3576fdSSage Weil i_size = i_size_read(inode); 4831d3576fdSSage Weil if (i_size < page_off + len) 4841d3576fdSSage Weil len = i_size - page_off; 4851d3576fdSSage Weil 486ae00d4f3SSage Weil dout("writepage %p page %p index %lu on %llu~%u snapc %p\n", 487ae00d4f3SSage Weil inode, page, page->index, page_off, len, snapc); 4881d3576fdSSage Weil 4893d14c5d2SYehuda Sadeh writeback_stat = atomic_long_inc_return(&fsc->writeback_count); 4902baba250SYehuda Sadeh if (writeback_stat > 4913d14c5d2SYehuda Sadeh CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb)) 4923d14c5d2SYehuda Sadeh set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC); 4932baba250SYehuda Sadeh 4941d3576fdSSage Weil set_page_writeback(page); 4951d3576fdSSage Weil err = ceph_osdc_writepages(osdc, ceph_vino(inode), 4961d3576fdSSage Weil &ci->i_layout, snapc, 4971d3576fdSSage Weil page_off, len, 4981d3576fdSSage Weil ci->i_truncate_seq, ci->i_truncate_size, 49924808826SAlex Elder &inode->i_mtime, &page, 1); 5001d3576fdSSage Weil if (err < 0) { 5011d3576fdSSage Weil dout("writepage setting page/mapping error %d %p\n", err, page); 5021d3576fdSSage Weil SetPageError(page); 5031d3576fdSSage Weil mapping_set_error(&inode->i_data, err); 5041d3576fdSSage Weil if (wbc) 5051d3576fdSSage Weil wbc->pages_skipped++; 5061d3576fdSSage Weil } else { 5071d3576fdSSage Weil dout("writepage cleaned page %p\n", page); 5081d3576fdSSage Weil err = 0; /* vfs expects us to return 0 */ 5091d3576fdSSage Weil } 5101d3576fdSSage Weil page->private = 0; 5111d3576fdSSage Weil ClearPagePrivate(page); 5121d3576fdSSage Weil end_page_writeback(page); 5131d3576fdSSage Weil ceph_put_wrbuffer_cap_refs(ci, 1, snapc); 5146298a337SSage Weil ceph_put_snap_context(snapc); /* page's reference */ 5151d3576fdSSage Weil out: 5161d3576fdSSage Weil return err; 5171d3576fdSSage Weil } 5181d3576fdSSage Weil 5191d3576fdSSage Weil static int ceph_writepage(struct page *page, struct writeback_control *wbc) 5201d3576fdSSage Weil { 521dbd646a8SYehuda Sadeh int err; 522dbd646a8SYehuda Sadeh struct inode *inode = page->mapping->host; 523dbd646a8SYehuda Sadeh BUG_ON(!inode); 52470b666c3SSage Weil ihold(inode); 525dbd646a8SYehuda Sadeh err = writepage_nounlock(page, wbc); 5261d3576fdSSage Weil unlock_page(page); 527dbd646a8SYehuda Sadeh iput(inode); 5281d3576fdSSage Weil return err; 5291d3576fdSSage Weil } 5301d3576fdSSage Weil 5311d3576fdSSage Weil 5321d3576fdSSage Weil /* 5331d3576fdSSage Weil * lame release_pages helper. release_pages() isn't exported to 5341d3576fdSSage Weil * modules. 5351d3576fdSSage Weil */ 5361d3576fdSSage Weil static void ceph_release_pages(struct page **pages, int num) 5371d3576fdSSage Weil { 5381d3576fdSSage Weil struct pagevec pvec; 5391d3576fdSSage Weil int i; 5401d3576fdSSage Weil 5411d3576fdSSage Weil pagevec_init(&pvec, 0); 5421d3576fdSSage Weil for (i = 0; i < num; i++) { 5431d3576fdSSage Weil if (pagevec_add(&pvec, pages[i]) == 0) 5441d3576fdSSage Weil pagevec_release(&pvec); 5451d3576fdSSage Weil } 5461d3576fdSSage Weil pagevec_release(&pvec); 5471d3576fdSSage Weil } 5481d3576fdSSage Weil 5491d3576fdSSage Weil 5501d3576fdSSage Weil /* 5511d3576fdSSage Weil * async writeback completion handler. 5521d3576fdSSage Weil * 5531d3576fdSSage Weil * If we get an error, set the mapping error bit, but not the individual 5541d3576fdSSage Weil * page error bits. 5551d3576fdSSage Weil */ 5561d3576fdSSage Weil static void writepages_finish(struct ceph_osd_request *req, 5571d3576fdSSage Weil struct ceph_msg *msg) 5581d3576fdSSage Weil { 5591d3576fdSSage Weil struct inode *inode = req->r_inode; 5601d3576fdSSage Weil struct ceph_inode_info *ci = ceph_inode(inode); 56187060c10SAlex Elder struct ceph_osd_data *osd_data; 5621d3576fdSSage Weil unsigned wrote; 5631d3576fdSSage Weil struct page *page; 564e0c59487SAlex Elder int num_pages; 5651d3576fdSSage Weil int i; 5661d3576fdSSage Weil struct ceph_snap_context *snapc = req->r_snapc; 5671d3576fdSSage Weil struct address_space *mapping = inode->i_mapping; 5681b83bef2SSage Weil int rc = req->r_result; 56979528734SAlex Elder u64 bytes = req->r_ops[0].extent.length; 5703d14c5d2SYehuda Sadeh struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 5712baba250SYehuda Sadeh long writeback_stat; 5727ff899daSSage Weil unsigned issued = ceph_caps_issued(ci); 5731d3576fdSSage Weil 574a4ce40a9SAlex Elder osd_data = osd_req_op_extent_osd_data(req, 0, true); 57587060c10SAlex Elder BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES); 57687060c10SAlex Elder num_pages = calc_pages_for((u64)osd_data->alignment, 57787060c10SAlex Elder (u64)osd_data->length); 5781d3576fdSSage Weil if (rc >= 0) { 57979788c69SSage Weil /* 58079788c69SSage Weil * Assume we wrote the pages we originally sent. The 58179788c69SSage Weil * osd might reply with fewer pages if our writeback 58279788c69SSage Weil * raced with a truncation and was adjusted at the osd, 58379788c69SSage Weil * so don't believe the reply. 58479788c69SSage Weil */ 585e0c59487SAlex Elder wrote = num_pages; 5861d3576fdSSage Weil } else { 5871d3576fdSSage Weil wrote = 0; 5881d3576fdSSage Weil mapping_set_error(mapping, rc); 5891d3576fdSSage Weil } 5901d3576fdSSage Weil dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n", 5911d3576fdSSage Weil inode, rc, bytes, wrote); 5921d3576fdSSage Weil 5931d3576fdSSage Weil /* clean all pages */ 594e0c59487SAlex Elder for (i = 0; i < num_pages; i++) { 59587060c10SAlex Elder page = osd_data->pages[i]; 5961d3576fdSSage Weil BUG_ON(!page); 5971d3576fdSSage Weil WARN_ON(!PageUptodate(page)); 5981d3576fdSSage Weil 5992baba250SYehuda Sadeh writeback_stat = 6003d14c5d2SYehuda Sadeh atomic_long_dec_return(&fsc->writeback_count); 6012baba250SYehuda Sadeh if (writeback_stat < 6023d14c5d2SYehuda Sadeh CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb)) 6033d14c5d2SYehuda Sadeh clear_bdi_congested(&fsc->backing_dev_info, 6042baba250SYehuda Sadeh BLK_RW_ASYNC); 6052baba250SYehuda Sadeh 60661600ef8SYan, Zheng ceph_put_snap_context(page_snap_context(page)); 6071d3576fdSSage Weil page->private = 0; 6081d3576fdSSage Weil ClearPagePrivate(page); 6091d3576fdSSage Weil dout("unlocking %d %p\n", i, page); 6101d3576fdSSage Weil end_page_writeback(page); 611e63dc5c7SYehuda Sadeh 612e63dc5c7SYehuda Sadeh /* 613e63dc5c7SYehuda Sadeh * We lost the cache cap, need to truncate the page before 614e63dc5c7SYehuda Sadeh * it is unlocked, otherwise we'd truncate it later in the 615e63dc5c7SYehuda Sadeh * page truncation thread, possibly losing some data that 616e63dc5c7SYehuda Sadeh * raced its way in 617e63dc5c7SYehuda Sadeh */ 6182962507cSSage Weil if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) 619e63dc5c7SYehuda Sadeh generic_error_remove_page(inode->i_mapping, page); 620e63dc5c7SYehuda Sadeh 6211d3576fdSSage Weil unlock_page(page); 6221d3576fdSSage Weil } 6231d3576fdSSage Weil dout("%p wrote+cleaned %d pages\n", inode, wrote); 624e0c59487SAlex Elder ceph_put_wrbuffer_cap_refs(ci, num_pages, snapc); 6251d3576fdSSage Weil 62687060c10SAlex Elder ceph_release_pages(osd_data->pages, num_pages); 62787060c10SAlex Elder if (osd_data->pages_from_pool) 62887060c10SAlex Elder mempool_free(osd_data->pages, 629640ef79dSCheng Renquan ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool); 6301d3576fdSSage Weil else 63187060c10SAlex Elder kfree(osd_data->pages); 6321d3576fdSSage Weil ceph_osdc_put_request(req); 6331d3576fdSSage Weil } 6341d3576fdSSage Weil 63594fe8420SAlex Elder static struct ceph_osd_request * 63694fe8420SAlex Elder ceph_writepages_osd_request(struct inode *inode, u64 offset, u64 *len, 63779528734SAlex Elder struct ceph_snap_context *snapc, int num_ops) 63894fe8420SAlex Elder { 63994fe8420SAlex Elder struct ceph_fs_client *fsc; 64094fe8420SAlex Elder struct ceph_inode_info *ci; 64194fe8420SAlex Elder struct ceph_vino vino; 64294fe8420SAlex Elder 64394fe8420SAlex Elder fsc = ceph_inode_to_client(inode); 64494fe8420SAlex Elder ci = ceph_inode(inode); 64594fe8420SAlex Elder vino = ceph_vino(inode); 64694fe8420SAlex Elder /* BUG_ON(vino.snap != CEPH_NOSNAP); */ 64794fe8420SAlex Elder 64894fe8420SAlex Elder return ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, 64979528734SAlex Elder vino, offset, len, num_ops, CEPH_OSD_OP_WRITE, 65094fe8420SAlex Elder CEPH_OSD_FLAG_WRITE|CEPH_OSD_FLAG_ONDISK, 65194fe8420SAlex Elder snapc, ci->i_truncate_seq, ci->i_truncate_size, true); 65294fe8420SAlex Elder } 65394fe8420SAlex Elder 6541d3576fdSSage Weil /* 6551d3576fdSSage Weil * initiate async writeback 6561d3576fdSSage Weil */ 6571d3576fdSSage Weil static int ceph_writepages_start(struct address_space *mapping, 6581d3576fdSSage Weil struct writeback_control *wbc) 6591d3576fdSSage Weil { 6601d3576fdSSage Weil struct inode *inode = mapping->host; 6611d3576fdSSage Weil struct ceph_inode_info *ci = ceph_inode(inode); 6623d14c5d2SYehuda Sadeh struct ceph_fs_client *fsc; 6631d3576fdSSage Weil pgoff_t index, start, end; 6641d3576fdSSage Weil int range_whole = 0; 6651d3576fdSSage Weil int should_loop = 1; 6661d3576fdSSage Weil pgoff_t max_pages = 0, max_pages_ever = 0; 66780e755feSSage Weil struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc; 6681d3576fdSSage Weil struct pagevec pvec; 6691d3576fdSSage Weil int done = 0; 6701d3576fdSSage Weil int rc = 0; 6711d3576fdSSage Weil unsigned wsize = 1 << inode->i_blkbits; 6721d3576fdSSage Weil struct ceph_osd_request *req = NULL; 6731d3576fdSSage Weil int do_sync; 6741d3576fdSSage Weil u64 snap_size = 0; 6751d3576fdSSage Weil 6761d3576fdSSage Weil /* 6771d3576fdSSage Weil * Include a 'sync' in the OSD request if this is a data 6781d3576fdSSage Weil * integrity write (e.g., O_SYNC write or fsync()), or if our 6791d3576fdSSage Weil * cap is being revoked. 6801d3576fdSSage Weil */ 6811d3576fdSSage Weil do_sync = wbc->sync_mode == WB_SYNC_ALL; 6821d3576fdSSage Weil if (ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER)) 6831d3576fdSSage Weil do_sync = 1; 6841d3576fdSSage Weil dout("writepages_start %p dosync=%d (mode=%s)\n", 6851d3576fdSSage Weil inode, do_sync, 6861d3576fdSSage Weil wbc->sync_mode == WB_SYNC_NONE ? "NONE" : 6871d3576fdSSage Weil (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD")); 6881d3576fdSSage Weil 6893d14c5d2SYehuda Sadeh fsc = ceph_inode_to_client(inode); 6903d14c5d2SYehuda Sadeh if (fsc->mount_state == CEPH_MOUNT_SHUTDOWN) { 6911d3576fdSSage Weil pr_warning("writepage_start %p on forced umount\n", inode); 6921d3576fdSSage Weil return -EIO; /* we're in a forced umount, don't write! */ 6931d3576fdSSage Weil } 6943d14c5d2SYehuda Sadeh if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize) 6953d14c5d2SYehuda Sadeh wsize = fsc->mount_options->wsize; 6961d3576fdSSage Weil if (wsize < PAGE_CACHE_SIZE) 6971d3576fdSSage Weil wsize = PAGE_CACHE_SIZE; 6981d3576fdSSage Weil max_pages_ever = wsize >> PAGE_CACHE_SHIFT; 6991d3576fdSSage Weil 7001d3576fdSSage Weil pagevec_init(&pvec, 0); 7011d3576fdSSage Weil 7021d3576fdSSage Weil /* where to start/end? */ 7031d3576fdSSage Weil if (wbc->range_cyclic) { 7041d3576fdSSage Weil start = mapping->writeback_index; /* Start from prev offset */ 7051d3576fdSSage Weil end = -1; 7061d3576fdSSage Weil dout(" cyclic, start at %lu\n", start); 7071d3576fdSSage Weil } else { 7081d3576fdSSage Weil start = wbc->range_start >> PAGE_CACHE_SHIFT; 7091d3576fdSSage Weil end = wbc->range_end >> PAGE_CACHE_SHIFT; 7101d3576fdSSage Weil if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX) 7111d3576fdSSage Weil range_whole = 1; 7121d3576fdSSage Weil should_loop = 0; 7131d3576fdSSage Weil dout(" not cyclic, %lu to %lu\n", start, end); 7141d3576fdSSage Weil } 7151d3576fdSSage Weil index = start; 7161d3576fdSSage Weil 7171d3576fdSSage Weil retry: 7181d3576fdSSage Weil /* find oldest snap context with dirty data */ 7191d3576fdSSage Weil ceph_put_snap_context(snapc); 7201d3576fdSSage Weil snapc = get_oldest_context(inode, &snap_size); 7211d3576fdSSage Weil if (!snapc) { 7221d3576fdSSage Weil /* hmm, why does writepages get called when there 7231d3576fdSSage Weil is no dirty data? */ 7241d3576fdSSage Weil dout(" no snap context with dirty data?\n"); 7251d3576fdSSage Weil goto out; 7261d3576fdSSage Weil } 7271d3576fdSSage Weil dout(" oldest snapc is %p seq %lld (%d snaps)\n", 7281d3576fdSSage Weil snapc, snapc->seq, snapc->num_snaps); 7291d3576fdSSage Weil if (last_snapc && snapc != last_snapc) { 7301d3576fdSSage Weil /* if we switched to a newer snapc, restart our scan at the 7311d3576fdSSage Weil * start of the original file range. */ 7321d3576fdSSage Weil dout(" snapc differs from last pass, restarting at %lu\n", 7331d3576fdSSage Weil index); 7341d3576fdSSage Weil index = start; 7351d3576fdSSage Weil } 7361d3576fdSSage Weil last_snapc = snapc; 7371d3576fdSSage Weil 7381d3576fdSSage Weil while (!done && index <= end) { 739e5975c7cSAlex Elder int num_ops = do_sync ? 2 : 1; 740e5975c7cSAlex Elder struct ceph_vino vino; 7411d3576fdSSage Weil unsigned i; 7421d3576fdSSage Weil int first; 7431d3576fdSSage Weil pgoff_t next; 7441d3576fdSSage Weil int pvec_pages, locked_pages; 745e5975c7cSAlex Elder struct page **pages = NULL; 746e5975c7cSAlex Elder mempool_t *pool = NULL; /* Becomes non-null if mempool used */ 7471d3576fdSSage Weil struct page *page; 7481d3576fdSSage Weil int want; 7491d3576fdSSage Weil u64 offset, len; 7502baba250SYehuda Sadeh long writeback_stat; 7511d3576fdSSage Weil 7521d3576fdSSage Weil next = 0; 7531d3576fdSSage Weil locked_pages = 0; 7541d3576fdSSage Weil max_pages = max_pages_ever; 7551d3576fdSSage Weil 7561d3576fdSSage Weil get_more_pages: 7571d3576fdSSage Weil first = -1; 7581d3576fdSSage Weil want = min(end - index, 7591d3576fdSSage Weil min((pgoff_t)PAGEVEC_SIZE, 7601d3576fdSSage Weil max_pages - (pgoff_t)locked_pages) - 1) 7611d3576fdSSage Weil + 1; 7621d3576fdSSage Weil pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index, 7631d3576fdSSage Weil PAGECACHE_TAG_DIRTY, 7641d3576fdSSage Weil want); 7651d3576fdSSage Weil dout("pagevec_lookup_tag got %d\n", pvec_pages); 7661d3576fdSSage Weil if (!pvec_pages && !locked_pages) 7671d3576fdSSage Weil break; 7681d3576fdSSage Weil for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) { 7691d3576fdSSage Weil page = pvec.pages[i]; 7701d3576fdSSage Weil dout("? %p idx %lu\n", page, page->index); 7711d3576fdSSage Weil if (locked_pages == 0) 7721d3576fdSSage Weil lock_page(page); /* first page */ 7731d3576fdSSage Weil else if (!trylock_page(page)) 7741d3576fdSSage Weil break; 7751d3576fdSSage Weil 7761d3576fdSSage Weil /* only dirty pages, or our accounting breaks */ 7771d3576fdSSage Weil if (unlikely(!PageDirty(page)) || 7781d3576fdSSage Weil unlikely(page->mapping != mapping)) { 7791d3576fdSSage Weil dout("!dirty or !mapping %p\n", page); 7801d3576fdSSage Weil unlock_page(page); 7811d3576fdSSage Weil break; 7821d3576fdSSage Weil } 7831d3576fdSSage Weil if (!wbc->range_cyclic && page->index > end) { 7841d3576fdSSage Weil dout("end of range %p\n", page); 7851d3576fdSSage Weil done = 1; 7861d3576fdSSage Weil unlock_page(page); 7871d3576fdSSage Weil break; 7881d3576fdSSage Weil } 7891d3576fdSSage Weil if (next && (page->index != next)) { 7901d3576fdSSage Weil dout("not consecutive %p\n", page); 7911d3576fdSSage Weil unlock_page(page); 7921d3576fdSSage Weil break; 7931d3576fdSSage Weil } 7941d3576fdSSage Weil if (wbc->sync_mode != WB_SYNC_NONE) { 7951d3576fdSSage Weil dout("waiting on writeback %p\n", page); 7961d3576fdSSage Weil wait_on_page_writeback(page); 7971d3576fdSSage Weil } 7981d3576fdSSage Weil if ((snap_size && page_offset(page) > snap_size) || 7991d3576fdSSage Weil (!snap_size && 8001d3576fdSSage Weil page_offset(page) > i_size_read(inode))) { 8011d3576fdSSage Weil dout("%p page eof %llu\n", page, snap_size ? 8021d3576fdSSage Weil snap_size : i_size_read(inode)); 8031d3576fdSSage Weil done = 1; 8041d3576fdSSage Weil unlock_page(page); 8051d3576fdSSage Weil break; 8061d3576fdSSage Weil } 8071d3576fdSSage Weil if (PageWriteback(page)) { 8081d3576fdSSage Weil dout("%p under writeback\n", page); 8091d3576fdSSage Weil unlock_page(page); 8101d3576fdSSage Weil break; 8111d3576fdSSage Weil } 8121d3576fdSSage Weil 8131d3576fdSSage Weil /* only if matching snap context */ 81461600ef8SYan, Zheng pgsnapc = page_snap_context(page); 81580e755feSSage Weil if (pgsnapc->seq > snapc->seq) { 81680e755feSSage Weil dout("page snapc %p %lld > oldest %p %lld\n", 81780e755feSSage Weil pgsnapc, pgsnapc->seq, snapc, snapc->seq); 8181d3576fdSSage Weil unlock_page(page); 8191d3576fdSSage Weil if (!locked_pages) 8201d3576fdSSage Weil continue; /* keep looking for snap */ 8211d3576fdSSage Weil break; 8221d3576fdSSage Weil } 8231d3576fdSSage Weil 8241d3576fdSSage Weil if (!clear_page_dirty_for_io(page)) { 8251d3576fdSSage Weil dout("%p !clear_page_dirty_for_io\n", page); 8261d3576fdSSage Weil unlock_page(page); 8271d3576fdSSage Weil break; 8281d3576fdSSage Weil } 8291d3576fdSSage Weil 830e5975c7cSAlex Elder /* 831e5975c7cSAlex Elder * We have something to write. If this is 832e5975c7cSAlex Elder * the first locked page this time through, 833e5975c7cSAlex Elder * allocate an osd request and a page array 834e5975c7cSAlex Elder * that it will use. 835e5975c7cSAlex Elder */ 8361d3576fdSSage Weil if (locked_pages == 0) { 83788486957SAlex Elder size_t size; 838e5975c7cSAlex Elder 839e5975c7cSAlex Elder BUG_ON(pages); 840acead002SAlex Elder 8411d3576fdSSage Weil /* prepare async write request */ 8426285bc23SAlex Elder offset = (u64)page_offset(page); 8431d3576fdSSage Weil len = wsize; 84494fe8420SAlex Elder req = ceph_writepages_osd_request(inode, 84594fe8420SAlex Elder offset, &len, snapc, 84679528734SAlex Elder num_ops); 8478c71897bSHenry C Chang 8486816282dSSage Weil if (IS_ERR(req)) { 8496816282dSSage Weil rc = PTR_ERR(req); 8508c71897bSHenry C Chang unlock_page(page); 8518c71897bSHenry C Chang break; 8528c71897bSHenry C Chang } 8538c71897bSHenry C Chang 85488486957SAlex Elder req->r_callback = writepages_finish; 85588486957SAlex Elder req->r_inode = inode; 85688486957SAlex Elder 85788486957SAlex Elder max_pages = calc_pages_for(0, (u64)len); 85888486957SAlex Elder size = max_pages * sizeof (*pages); 85988486957SAlex Elder pages = kmalloc(size, GFP_NOFS); 86088486957SAlex Elder if (!pages) { 86188486957SAlex Elder pool = fsc->wb_pagevec_pool; 86288486957SAlex Elder pages = mempool_alloc(pool, GFP_NOFS); 863e5975c7cSAlex Elder BUG_ON(!pages); 86488486957SAlex Elder } 8651d3576fdSSage Weil } 8661d3576fdSSage Weil 8671d3576fdSSage Weil /* note position of first page in pvec */ 8681d3576fdSSage Weil if (first < 0) 8691d3576fdSSage Weil first = i; 8701d3576fdSSage Weil dout("%p will write page %p idx %lu\n", 8711d3576fdSSage Weil inode, page, page->index); 8722baba250SYehuda Sadeh 873213c99eeSSage Weil writeback_stat = 8743d14c5d2SYehuda Sadeh atomic_long_inc_return(&fsc->writeback_count); 875213c99eeSSage Weil if (writeback_stat > CONGESTION_ON_THRESH( 8763d14c5d2SYehuda Sadeh fsc->mount_options->congestion_kb)) { 8773d14c5d2SYehuda Sadeh set_bdi_congested(&fsc->backing_dev_info, 878213c99eeSSage Weil BLK_RW_ASYNC); 8792baba250SYehuda Sadeh } 8802baba250SYehuda Sadeh 8811d3576fdSSage Weil set_page_writeback(page); 882e5975c7cSAlex Elder pages[locked_pages] = page; 8831d3576fdSSage Weil locked_pages++; 8841d3576fdSSage Weil next = page->index + 1; 8851d3576fdSSage Weil } 8861d3576fdSSage Weil 8871d3576fdSSage Weil /* did we get anything? */ 8881d3576fdSSage Weil if (!locked_pages) 8891d3576fdSSage Weil goto release_pvec_pages; 8901d3576fdSSage Weil if (i) { 8911d3576fdSSage Weil int j; 8921d3576fdSSage Weil BUG_ON(!locked_pages || first < 0); 8931d3576fdSSage Weil 8941d3576fdSSage Weil if (pvec_pages && i == pvec_pages && 8951d3576fdSSage Weil locked_pages < max_pages) { 8961d3576fdSSage Weil dout("reached end pvec, trying for more\n"); 8971d3576fdSSage Weil pagevec_reinit(&pvec); 8981d3576fdSSage Weil goto get_more_pages; 8991d3576fdSSage Weil } 9001d3576fdSSage Weil 9011d3576fdSSage Weil /* shift unused pages over in the pvec... we 9021d3576fdSSage Weil * will need to release them below. */ 9031d3576fdSSage Weil for (j = i; j < pvec_pages; j++) { 9041d3576fdSSage Weil dout(" pvec leftover page %p\n", 9051d3576fdSSage Weil pvec.pages[j]); 9061d3576fdSSage Weil pvec.pages[j-i+first] = pvec.pages[j]; 9071d3576fdSSage Weil } 9081d3576fdSSage Weil pvec.nr -= i-first; 9091d3576fdSSage Weil } 9101d3576fdSSage Weil 911e5975c7cSAlex Elder /* Format the osd request message and submit the write */ 912e5975c7cSAlex Elder 913e5975c7cSAlex Elder offset = page_offset(pages[0]); 9141d3576fdSSage Weil len = min((snap_size ? snap_size : i_size_read(inode)) - offset, 9151d3576fdSSage Weil (u64)locked_pages << PAGE_CACHE_SHIFT); 9161d3576fdSSage Weil dout("writepages got %d pages at %llu~%llu\n", 9171d3576fdSSage Weil locked_pages, offset, len); 9181d3576fdSSage Weil 919a4ce40a9SAlex Elder osd_req_op_extent_osd_data_pages(req, 0, true, pages, len, 0, 920a4ce40a9SAlex Elder !!pool, false); 921e5975c7cSAlex Elder 922e5975c7cSAlex Elder pages = NULL; /* request message now owns the pages array */ 923e5975c7cSAlex Elder pool = NULL; 924e5975c7cSAlex Elder 925e5975c7cSAlex Elder /* Update the write op length in case we changed it */ 926e5975c7cSAlex Elder 927c99d2d4aSAlex Elder osd_req_op_extent_update(req, 0, len); 928e5975c7cSAlex Elder 929e5975c7cSAlex Elder vino = ceph_vino(inode); 93079528734SAlex Elder ceph_osdc_build_request(req, offset, snapc, vino.snap, 93179528734SAlex Elder &inode->i_mtime); 9321d3576fdSSage Weil 9339d6fcb08SSage Weil rc = ceph_osdc_start_request(&fsc->client->osdc, req, true); 9349d6fcb08SSage Weil BUG_ON(rc); 9351d3576fdSSage Weil req = NULL; 9361d3576fdSSage Weil 9371d3576fdSSage Weil /* continue? */ 9381d3576fdSSage Weil index = next; 9391d3576fdSSage Weil wbc->nr_to_write -= locked_pages; 9401d3576fdSSage Weil if (wbc->nr_to_write <= 0) 9411d3576fdSSage Weil done = 1; 9421d3576fdSSage Weil 9431d3576fdSSage Weil release_pvec_pages: 9441d3576fdSSage Weil dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr, 9451d3576fdSSage Weil pvec.nr ? pvec.pages[0] : NULL); 9461d3576fdSSage Weil pagevec_release(&pvec); 9471d3576fdSSage Weil 9481d3576fdSSage Weil if (locked_pages && !done) 9491d3576fdSSage Weil goto retry; 9501d3576fdSSage Weil } 9511d3576fdSSage Weil 9521d3576fdSSage Weil if (should_loop && !done) { 9531d3576fdSSage Weil /* more to do; loop back to beginning of file */ 9541d3576fdSSage Weil dout("writepages looping back to beginning of file\n"); 9551d3576fdSSage Weil should_loop = 0; 9561d3576fdSSage Weil index = 0; 9571d3576fdSSage Weil goto retry; 9581d3576fdSSage Weil } 9591d3576fdSSage Weil 9601d3576fdSSage Weil if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) 9611d3576fdSSage Weil mapping->writeback_index = index; 9621d3576fdSSage Weil 9631d3576fdSSage Weil out: 9641d3576fdSSage Weil if (req) 9651d3576fdSSage Weil ceph_osdc_put_request(req); 9661d3576fdSSage Weil ceph_put_snap_context(snapc); 9671d3576fdSSage Weil dout("writepages done, rc = %d\n", rc); 9681d3576fdSSage Weil return rc; 9691d3576fdSSage Weil } 9701d3576fdSSage Weil 9711d3576fdSSage Weil 9721d3576fdSSage Weil 9731d3576fdSSage Weil /* 9741d3576fdSSage Weil * See if a given @snapc is either writeable, or already written. 9751d3576fdSSage Weil */ 9761d3576fdSSage Weil static int context_is_writeable_or_written(struct inode *inode, 9771d3576fdSSage Weil struct ceph_snap_context *snapc) 9781d3576fdSSage Weil { 9791d3576fdSSage Weil struct ceph_snap_context *oldest = get_oldest_context(inode, NULL); 9806298a337SSage Weil int ret = !oldest || snapc->seq <= oldest->seq; 9816298a337SSage Weil 9826298a337SSage Weil ceph_put_snap_context(oldest); 9836298a337SSage Weil return ret; 9841d3576fdSSage Weil } 9851d3576fdSSage Weil 9861d3576fdSSage Weil /* 9871d3576fdSSage Weil * We are only allowed to write into/dirty the page if the page is 9881d3576fdSSage Weil * clean, or already dirty within the same snap context. 9898f883c24SSage Weil * 9908f883c24SSage Weil * called with page locked. 9918f883c24SSage Weil * return success with page locked, 9928f883c24SSage Weil * or any failure (incl -EAGAIN) with page unlocked. 9931d3576fdSSage Weil */ 9944af6b225SYehuda Sadeh static int ceph_update_writeable_page(struct file *file, 9954af6b225SYehuda Sadeh loff_t pos, unsigned len, 9964af6b225SYehuda Sadeh struct page *page) 9971d3576fdSSage Weil { 998496ad9aaSAl Viro struct inode *inode = file_inode(file); 9991d3576fdSSage Weil struct ceph_inode_info *ci = ceph_inode(inode); 10003d14c5d2SYehuda Sadeh struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc; 10011d3576fdSSage Weil loff_t page_off = pos & PAGE_CACHE_MASK; 10021d3576fdSSage Weil int pos_in_page = pos & ~PAGE_CACHE_MASK; 10031d3576fdSSage Weil int end_in_page = pos_in_page + len; 10041d3576fdSSage Weil loff_t i_size; 10051d3576fdSSage Weil int r; 100680e755feSSage Weil struct ceph_snap_context *snapc, *oldest; 10071d3576fdSSage Weil 10081d3576fdSSage Weil retry_locked: 10091d3576fdSSage Weil /* writepages currently holds page lock, but if we change that later, */ 10101d3576fdSSage Weil wait_on_page_writeback(page); 10111d3576fdSSage Weil 10121d3576fdSSage Weil /* check snap context */ 10131d3576fdSSage Weil BUG_ON(!ci->i_snap_realm); 10141d3576fdSSage Weil down_read(&mdsc->snap_rwsem); 10151d3576fdSSage Weil BUG_ON(!ci->i_snap_realm->cached_context); 101661600ef8SYan, Zheng snapc = page_snap_context(page); 101780e755feSSage Weil if (snapc && snapc != ci->i_head_snapc) { 10181d3576fdSSage Weil /* 10191d3576fdSSage Weil * this page is already dirty in another (older) snap 10201d3576fdSSage Weil * context! is it writeable now? 10211d3576fdSSage Weil */ 102280e755feSSage Weil oldest = get_oldest_context(inode, NULL); 10231d3576fdSSage Weil up_read(&mdsc->snap_rwsem); 10241d3576fdSSage Weil 102580e755feSSage Weil if (snapc->seq > oldest->seq) { 10266298a337SSage Weil ceph_put_snap_context(oldest); 10271d3576fdSSage Weil dout(" page %p snapc %p not current or oldest\n", 10286298a337SSage Weil page, snapc); 10291d3576fdSSage Weil /* 10301d3576fdSSage Weil * queue for writeback, and wait for snapc to 10311d3576fdSSage Weil * be writeable or written 10321d3576fdSSage Weil */ 10336298a337SSage Weil snapc = ceph_get_snap_context(snapc); 10341d3576fdSSage Weil unlock_page(page); 10353c6f6b79SSage Weil ceph_queue_writeback(inode); 10368f883c24SSage Weil r = wait_event_interruptible(ci->i_cap_wq, 10371d3576fdSSage Weil context_is_writeable_or_written(inode, snapc)); 10381d3576fdSSage Weil ceph_put_snap_context(snapc); 10398f883c24SSage Weil if (r == -ERESTARTSYS) 10408f883c24SSage Weil return r; 10414af6b225SYehuda Sadeh return -EAGAIN; 10421d3576fdSSage Weil } 10436298a337SSage Weil ceph_put_snap_context(oldest); 10441d3576fdSSage Weil 10451d3576fdSSage Weil /* yay, writeable, do it now (without dropping page lock) */ 10461d3576fdSSage Weil dout(" page %p snapc %p not current, but oldest\n", 10471d3576fdSSage Weil page, snapc); 10481d3576fdSSage Weil if (!clear_page_dirty_for_io(page)) 10491d3576fdSSage Weil goto retry_locked; 10501d3576fdSSage Weil r = writepage_nounlock(page, NULL); 10511d3576fdSSage Weil if (r < 0) 10521d3576fdSSage Weil goto fail_nosnap; 10531d3576fdSSage Weil goto retry_locked; 10541d3576fdSSage Weil } 10551d3576fdSSage Weil 10561d3576fdSSage Weil if (PageUptodate(page)) { 10571d3576fdSSage Weil dout(" page %p already uptodate\n", page); 10581d3576fdSSage Weil return 0; 10591d3576fdSSage Weil } 10601d3576fdSSage Weil 10611d3576fdSSage Weil /* full page? */ 10621d3576fdSSage Weil if (pos_in_page == 0 && len == PAGE_CACHE_SIZE) 10631d3576fdSSage Weil return 0; 10641d3576fdSSage Weil 10651d3576fdSSage Weil /* past end of file? */ 10661d3576fdSSage Weil i_size = inode->i_size; /* caller holds i_mutex */ 10671d3576fdSSage Weil 10681d3576fdSSage Weil if (i_size + len > inode->i_sb->s_maxbytes) { 10691d3576fdSSage Weil /* file is too big */ 10701d3576fdSSage Weil r = -EINVAL; 10711d3576fdSSage Weil goto fail; 10721d3576fdSSage Weil } 10731d3576fdSSage Weil 10741d3576fdSSage Weil if (page_off >= i_size || 10751d3576fdSSage Weil (pos_in_page == 0 && (pos+len) >= i_size && 10761d3576fdSSage Weil end_in_page - pos_in_page != PAGE_CACHE_SIZE)) { 10771d3576fdSSage Weil dout(" zeroing %p 0 - %d and %d - %d\n", 10781d3576fdSSage Weil page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE); 10791d3576fdSSage Weil zero_user_segments(page, 10801d3576fdSSage Weil 0, pos_in_page, 10811d3576fdSSage Weil end_in_page, PAGE_CACHE_SIZE); 10821d3576fdSSage Weil return 0; 10831d3576fdSSage Weil } 10841d3576fdSSage Weil 10851d3576fdSSage Weil /* we need to read it. */ 10861d3576fdSSage Weil up_read(&mdsc->snap_rwsem); 10871d3576fdSSage Weil r = readpage_nounlock(file, page); 10881d3576fdSSage Weil if (r < 0) 10891d3576fdSSage Weil goto fail_nosnap; 10901d3576fdSSage Weil goto retry_locked; 10911d3576fdSSage Weil 10921d3576fdSSage Weil fail: 10931d3576fdSSage Weil up_read(&mdsc->snap_rwsem); 10941d3576fdSSage Weil fail_nosnap: 10951d3576fdSSage Weil unlock_page(page); 10961d3576fdSSage Weil return r; 10971d3576fdSSage Weil } 10981d3576fdSSage Weil 10991d3576fdSSage Weil /* 11004af6b225SYehuda Sadeh * We are only allowed to write into/dirty the page if the page is 11014af6b225SYehuda Sadeh * clean, or already dirty within the same snap context. 11024af6b225SYehuda Sadeh */ 11034af6b225SYehuda Sadeh static int ceph_write_begin(struct file *file, struct address_space *mapping, 11044af6b225SYehuda Sadeh loff_t pos, unsigned len, unsigned flags, 11054af6b225SYehuda Sadeh struct page **pagep, void **fsdata) 11064af6b225SYehuda Sadeh { 1107496ad9aaSAl Viro struct inode *inode = file_inode(file); 11084af6b225SYehuda Sadeh struct page *page; 11094af6b225SYehuda Sadeh pgoff_t index = pos >> PAGE_CACHE_SHIFT; 11107971bd92SSage Weil int r; 11114af6b225SYehuda Sadeh 11124af6b225SYehuda Sadeh do { 11134af6b225SYehuda Sadeh /* get a page */ 11144af6b225SYehuda Sadeh page = grab_cache_page_write_begin(mapping, index, 0); 11157971bd92SSage Weil if (!page) 11167971bd92SSage Weil return -ENOMEM; 11177971bd92SSage Weil *pagep = page; 11184af6b225SYehuda Sadeh 11194af6b225SYehuda Sadeh dout("write_begin file %p inode %p page %p %d~%d\n", file, 11204af6b225SYehuda Sadeh inode, page, (int)pos, (int)len); 11214af6b225SYehuda Sadeh 11224af6b225SYehuda Sadeh r = ceph_update_writeable_page(file, pos, len, page); 11234af6b225SYehuda Sadeh } while (r == -EAGAIN); 11244af6b225SYehuda Sadeh 11254af6b225SYehuda Sadeh return r; 11264af6b225SYehuda Sadeh } 11274af6b225SYehuda Sadeh 11284af6b225SYehuda Sadeh /* 11291d3576fdSSage Weil * we don't do anything in here that simple_write_end doesn't do 11301d3576fdSSage Weil * except adjust dirty page accounting and drop read lock on 11311d3576fdSSage Weil * mdsc->snap_rwsem. 11321d3576fdSSage Weil */ 11331d3576fdSSage Weil static int ceph_write_end(struct file *file, struct address_space *mapping, 11341d3576fdSSage Weil loff_t pos, unsigned len, unsigned copied, 11351d3576fdSSage Weil struct page *page, void *fsdata) 11361d3576fdSSage Weil { 1137496ad9aaSAl Viro struct inode *inode = file_inode(file); 11383d14c5d2SYehuda Sadeh struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 11393d14c5d2SYehuda Sadeh struct ceph_mds_client *mdsc = fsc->mdsc; 11401d3576fdSSage Weil unsigned from = pos & (PAGE_CACHE_SIZE - 1); 11411d3576fdSSage Weil int check_cap = 0; 11421d3576fdSSage Weil 11431d3576fdSSage Weil dout("write_end file %p inode %p page %p %d~%d (%d)\n", file, 11441d3576fdSSage Weil inode, page, (int)pos, (int)copied, (int)len); 11451d3576fdSSage Weil 11461d3576fdSSage Weil /* zero the stale part of the page if we did a short copy */ 11471d3576fdSSage Weil if (copied < len) 11481d3576fdSSage Weil zero_user_segment(page, from+copied, len); 11491d3576fdSSage Weil 11501d3576fdSSage Weil /* did file size increase? */ 11511d3576fdSSage Weil /* (no need for i_size_read(); we caller holds i_mutex */ 11521d3576fdSSage Weil if (pos+copied > inode->i_size) 11531d3576fdSSage Weil check_cap = ceph_inode_set_size(inode, pos+copied); 11541d3576fdSSage Weil 11551d3576fdSSage Weil if (!PageUptodate(page)) 11561d3576fdSSage Weil SetPageUptodate(page); 11571d3576fdSSage Weil 11581d3576fdSSage Weil set_page_dirty(page); 11591d3576fdSSage Weil 11601d3576fdSSage Weil unlock_page(page); 11611d3576fdSSage Weil up_read(&mdsc->snap_rwsem); 11621d3576fdSSage Weil page_cache_release(page); 11631d3576fdSSage Weil 11641d3576fdSSage Weil if (check_cap) 11651d3576fdSSage Weil ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL); 11661d3576fdSSage Weil 11671d3576fdSSage Weil return copied; 11681d3576fdSSage Weil } 11691d3576fdSSage Weil 11701d3576fdSSage Weil /* 11711d3576fdSSage Weil * we set .direct_IO to indicate direct io is supported, but since we 11721d3576fdSSage Weil * intercept O_DIRECT reads and writes early, this function should 11731d3576fdSSage Weil * never get called. 11741d3576fdSSage Weil */ 11751d3576fdSSage Weil static ssize_t ceph_direct_io(int rw, struct kiocb *iocb, 11761d3576fdSSage Weil const struct iovec *iov, 11771d3576fdSSage Weil loff_t pos, unsigned long nr_segs) 11781d3576fdSSage Weil { 11791d3576fdSSage Weil WARN_ON(1); 11801d3576fdSSage Weil return -EINVAL; 11811d3576fdSSage Weil } 11821d3576fdSSage Weil 11831d3576fdSSage Weil const struct address_space_operations ceph_aops = { 11841d3576fdSSage Weil .readpage = ceph_readpage, 11851d3576fdSSage Weil .readpages = ceph_readpages, 11861d3576fdSSage Weil .writepage = ceph_writepage, 11871d3576fdSSage Weil .writepages = ceph_writepages_start, 11881d3576fdSSage Weil .write_begin = ceph_write_begin, 11891d3576fdSSage Weil .write_end = ceph_write_end, 11901d3576fdSSage Weil .set_page_dirty = ceph_set_page_dirty, 11911d3576fdSSage Weil .invalidatepage = ceph_invalidatepage, 11921d3576fdSSage Weil .releasepage = ceph_releasepage, 11931d3576fdSSage Weil .direct_IO = ceph_direct_io, 11941d3576fdSSage Weil }; 11951d3576fdSSage Weil 11961d3576fdSSage Weil 11971d3576fdSSage Weil /* 11981d3576fdSSage Weil * vm ops 11991d3576fdSSage Weil */ 12001d3576fdSSage Weil 12011d3576fdSSage Weil /* 12021d3576fdSSage Weil * Reuse write_begin here for simplicity. 12031d3576fdSSage Weil */ 12041d3576fdSSage Weil static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf) 12051d3576fdSSage Weil { 1206496ad9aaSAl Viro struct inode *inode = file_inode(vma->vm_file); 12071d3576fdSSage Weil struct page *page = vmf->page; 12083d14c5d2SYehuda Sadeh struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc; 12096285bc23SAlex Elder loff_t off = page_offset(page); 12101d3576fdSSage Weil loff_t size, len; 12111d3576fdSSage Weil int ret; 12121d3576fdSSage Weil 12133ca9c3bdSJan Kara /* Update time before taking page lock */ 12143ca9c3bdSJan Kara file_update_time(vma->vm_file); 12153ca9c3bdSJan Kara 12161d3576fdSSage Weil size = i_size_read(inode); 12171d3576fdSSage Weil if (off + PAGE_CACHE_SIZE <= size) 12181d3576fdSSage Weil len = PAGE_CACHE_SIZE; 12191d3576fdSSage Weil else 12201d3576fdSSage Weil len = size & ~PAGE_CACHE_MASK; 12211d3576fdSSage Weil 12221d3576fdSSage Weil dout("page_mkwrite %p %llu~%llu page %p idx %lu\n", inode, 12231d3576fdSSage Weil off, len, page, page->index); 12244af6b225SYehuda Sadeh 12254af6b225SYehuda Sadeh lock_page(page); 12264af6b225SYehuda Sadeh 12274af6b225SYehuda Sadeh ret = VM_FAULT_NOPAGE; 12284af6b225SYehuda Sadeh if ((off > size) || 12294af6b225SYehuda Sadeh (page->mapping != inode->i_mapping)) 12304af6b225SYehuda Sadeh goto out; 12314af6b225SYehuda Sadeh 12324af6b225SYehuda Sadeh ret = ceph_update_writeable_page(vma->vm_file, off, len, page); 12334af6b225SYehuda Sadeh if (ret == 0) { 12344af6b225SYehuda Sadeh /* success. we'll keep the page locked. */ 12351d3576fdSSage Weil set_page_dirty(page); 12361d3576fdSSage Weil up_read(&mdsc->snap_rwsem); 12371d3576fdSSage Weil ret = VM_FAULT_LOCKED; 12381d3576fdSSage Weil } else { 12394af6b225SYehuda Sadeh if (ret == -ENOMEM) 12404af6b225SYehuda Sadeh ret = VM_FAULT_OOM; 12414af6b225SYehuda Sadeh else 12421d3576fdSSage Weil ret = VM_FAULT_SIGBUS; 12431d3576fdSSage Weil } 12444af6b225SYehuda Sadeh out: 12451d3576fdSSage Weil dout("page_mkwrite %p %llu~%llu = %d\n", inode, off, len, ret); 12464af6b225SYehuda Sadeh if (ret != VM_FAULT_LOCKED) 12474af6b225SYehuda Sadeh unlock_page(page); 12481d3576fdSSage Weil return ret; 12491d3576fdSSage Weil } 12501d3576fdSSage Weil 12511d3576fdSSage Weil static struct vm_operations_struct ceph_vmops = { 12521d3576fdSSage Weil .fault = filemap_fault, 12531d3576fdSSage Weil .page_mkwrite = ceph_page_mkwrite, 12540b173bc4SKonstantin Khlebnikov .remap_pages = generic_file_remap_pages, 12551d3576fdSSage Weil }; 12561d3576fdSSage Weil 12571d3576fdSSage Weil int ceph_mmap(struct file *file, struct vm_area_struct *vma) 12581d3576fdSSage Weil { 12591d3576fdSSage Weil struct address_space *mapping = file->f_mapping; 12601d3576fdSSage Weil 12611d3576fdSSage Weil if (!mapping->a_ops->readpage) 12621d3576fdSSage Weil return -ENOEXEC; 12631d3576fdSSage Weil file_accessed(file); 12641d3576fdSSage Weil vma->vm_ops = &ceph_vmops; 12651d3576fdSSage Weil return 0; 12661d3576fdSSage Weil } 1267