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" 1499ccbd22SMilosz Tanski #include "cache.h" 153d14c5d2SYehuda Sadeh #include <linux/ceph/osd_client.h> 161d3576fdSSage Weil 171d3576fdSSage Weil /* 181d3576fdSSage Weil * Ceph address space ops. 191d3576fdSSage Weil * 201d3576fdSSage Weil * There are a few funny things going on here. 211d3576fdSSage Weil * 221d3576fdSSage Weil * The page->private field is used to reference a struct 231d3576fdSSage Weil * ceph_snap_context for _every_ dirty page. This indicates which 241d3576fdSSage Weil * snapshot the page was logically dirtied in, and thus which snap 251d3576fdSSage Weil * context needs to be associated with the osd write during writeback. 261d3576fdSSage Weil * 271d3576fdSSage Weil * Similarly, struct ceph_inode_info maintains a set of counters to 2825985edcSLucas De Marchi * count dirty pages on the inode. In the absence of snapshots, 291d3576fdSSage Weil * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count. 301d3576fdSSage Weil * 311d3576fdSSage Weil * When a snapshot is taken (that is, when the client receives 321d3576fdSSage Weil * notification that a snapshot was taken), each inode with caps and 331d3576fdSSage Weil * with dirty pages (dirty pages implies there is a cap) gets a new 341d3576fdSSage Weil * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending 351d3576fdSSage Weil * order, new snaps go to the tail). The i_wrbuffer_ref_head count is 361d3576fdSSage Weil * moved to capsnap->dirty. (Unless a sync write is currently in 371d3576fdSSage Weil * progress. In that case, the capsnap is said to be "pending", new 381d3576fdSSage Weil * writes cannot start, and the capsnap isn't "finalized" until the 391d3576fdSSage Weil * write completes (or fails) and a final size/mtime for the inode for 401d3576fdSSage Weil * that snap can be settled upon.) i_wrbuffer_ref_head is reset to 0. 411d3576fdSSage Weil * 421d3576fdSSage Weil * On writeback, we must submit writes to the osd IN SNAP ORDER. So, 431d3576fdSSage Weil * we look for the first capsnap in i_cap_snaps and write out pages in 441d3576fdSSage Weil * that snap context _only_. Then we move on to the next capsnap, 451d3576fdSSage Weil * eventually reaching the "live" or "head" context (i.e., pages that 461d3576fdSSage Weil * are not yet snapped) and are writing the most recently dirtied 471d3576fdSSage Weil * pages. 481d3576fdSSage Weil * 491d3576fdSSage Weil * Invalidate and so forth must take care to ensure the dirty page 501d3576fdSSage Weil * accounting is preserved. 511d3576fdSSage Weil */ 521d3576fdSSage Weil 532baba250SYehuda Sadeh #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10)) 542baba250SYehuda Sadeh #define CONGESTION_OFF_THRESH(congestion_kb) \ 552baba250SYehuda Sadeh (CONGESTION_ON_THRESH(congestion_kb) - \ 562baba250SYehuda Sadeh (CONGESTION_ON_THRESH(congestion_kb) >> 2)) 572baba250SYehuda Sadeh 5861600ef8SYan, Zheng static inline struct ceph_snap_context *page_snap_context(struct page *page) 5961600ef8SYan, Zheng { 6061600ef8SYan, Zheng if (PagePrivate(page)) 6161600ef8SYan, Zheng return (void *)page->private; 6261600ef8SYan, Zheng return NULL; 6361600ef8SYan, Zheng } 641d3576fdSSage Weil 651d3576fdSSage Weil /* 661d3576fdSSage Weil * Dirty a page. Optimistically adjust accounting, on the assumption 671d3576fdSSage Weil * that we won't race with invalidate. If we do, readjust. 681d3576fdSSage Weil */ 691d3576fdSSage Weil static int ceph_set_page_dirty(struct page *page) 701d3576fdSSage Weil { 711d3576fdSSage Weil struct address_space *mapping = page->mapping; 721d3576fdSSage Weil struct inode *inode; 731d3576fdSSage Weil struct ceph_inode_info *ci; 741d3576fdSSage Weil struct ceph_snap_context *snapc; 757d6e1f54SSha Zhengju int ret; 761d3576fdSSage Weil 771d3576fdSSage Weil if (unlikely(!mapping)) 781d3576fdSSage Weil return !TestSetPageDirty(page); 791d3576fdSSage Weil 807d6e1f54SSha Zhengju if (PageDirty(page)) { 811d3576fdSSage Weil dout("%p set_page_dirty %p idx %lu -- already dirty\n", 821d3576fdSSage Weil mapping->host, page, page->index); 837d6e1f54SSha Zhengju BUG_ON(!PagePrivate(page)); 841d3576fdSSage Weil return 0; 851d3576fdSSage Weil } 861d3576fdSSage Weil 871d3576fdSSage Weil inode = mapping->host; 881d3576fdSSage Weil ci = ceph_inode(inode); 891d3576fdSSage Weil 901d3576fdSSage Weil /* 911d3576fdSSage Weil * Note that we're grabbing a snapc ref here without holding 921d3576fdSSage Weil * any locks! 931d3576fdSSage Weil */ 941d3576fdSSage Weil snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context); 951d3576fdSSage Weil 961d3576fdSSage Weil /* dirty the head */ 97be655596SSage Weil spin_lock(&ci->i_ceph_lock); 987d8cb26dSSage Weil if (ci->i_head_snapc == NULL) 991d3576fdSSage Weil ci->i_head_snapc = ceph_get_snap_context(snapc); 1001d3576fdSSage Weil ++ci->i_wrbuffer_ref_head; 1011d3576fdSSage Weil if (ci->i_wrbuffer_ref == 0) 1020444d76aSDave Chinner ihold(inode); 1031d3576fdSSage Weil ++ci->i_wrbuffer_ref; 1041d3576fdSSage Weil dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d " 1051d3576fdSSage Weil "snapc %p seq %lld (%d snaps)\n", 1061d3576fdSSage Weil mapping->host, page, page->index, 1071d3576fdSSage Weil ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1, 1081d3576fdSSage Weil ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head, 1091d3576fdSSage Weil snapc, snapc->seq, snapc->num_snaps); 110be655596SSage Weil spin_unlock(&ci->i_ceph_lock); 1111d3576fdSSage Weil 1121d3576fdSSage Weil /* 1131d3576fdSSage Weil * Reference snap context in page->private. Also set 1141d3576fdSSage Weil * PagePrivate so that we get invalidatepage callback. 1151d3576fdSSage Weil */ 1167d6e1f54SSha Zhengju BUG_ON(PagePrivate(page)); 1171d3576fdSSage Weil page->private = (unsigned long)snapc; 1181d3576fdSSage Weil SetPagePrivate(page); 1191d3576fdSSage Weil 1207d6e1f54SSha Zhengju ret = __set_page_dirty_nobuffers(page); 1217d6e1f54SSha Zhengju WARN_ON(!PageLocked(page)); 1227d6e1f54SSha Zhengju WARN_ON(!page->mapping); 1231d3576fdSSage Weil 1247d6e1f54SSha Zhengju return ret; 1251d3576fdSSage Weil } 1261d3576fdSSage Weil 1271d3576fdSSage Weil /* 1281d3576fdSSage Weil * If we are truncating the full page (i.e. offset == 0), adjust the 1291d3576fdSSage Weil * dirty page counters appropriately. Only called if there is private 1301d3576fdSSage Weil * data on the page. 1311d3576fdSSage Weil */ 132d47992f8SLukas Czerner static void ceph_invalidatepage(struct page *page, unsigned int offset, 133d47992f8SLukas Czerner unsigned int length) 1341d3576fdSSage Weil { 1354ce1e9adSAlexander Beregalov struct inode *inode; 1361d3576fdSSage Weil struct ceph_inode_info *ci; 13761600ef8SYan, Zheng struct ceph_snap_context *snapc = page_snap_context(page); 1381d3576fdSSage Weil 1394ce1e9adSAlexander Beregalov inode = page->mapping->host; 140b150f5c1SMilosz Tanski ci = ceph_inode(inode); 141b150f5c1SMilosz Tanski 142b150f5c1SMilosz Tanski if (offset != 0 || length != PAGE_CACHE_SIZE) { 143b150f5c1SMilosz Tanski dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n", 144b150f5c1SMilosz Tanski inode, page, page->index, offset, length); 145b150f5c1SMilosz Tanski return; 146b150f5c1SMilosz Tanski } 1474ce1e9adSAlexander Beregalov 14899ccbd22SMilosz Tanski ceph_invalidate_fscache_page(inode, page); 14999ccbd22SMilosz Tanski 15099ccbd22SMilosz Tanski if (!PagePrivate(page)) 15199ccbd22SMilosz Tanski return; 15299ccbd22SMilosz Tanski 1531d3576fdSSage Weil /* 1541d3576fdSSage Weil * We can get non-dirty pages here due to races between 1551d3576fdSSage Weil * set_page_dirty and truncate_complete_page; just spit out a 1561d3576fdSSage Weil * warning, in case we end up with accounting problems later. 1571d3576fdSSage Weil */ 1581d3576fdSSage Weil if (!PageDirty(page)) 1591d3576fdSSage Weil pr_err("%p invalidatepage %p page not dirty\n", inode, page); 1601d3576fdSSage Weil 1611d3576fdSSage Weil ClearPageChecked(page); 1621d3576fdSSage Weil 163569d39fcSLukas Czerner dout("%p invalidatepage %p idx %lu full dirty page\n", 164569d39fcSLukas Czerner inode, page, page->index); 165b150f5c1SMilosz Tanski 1661d3576fdSSage Weil ceph_put_wrbuffer_cap_refs(ci, 1, snapc); 1671d3576fdSSage Weil ceph_put_snap_context(snapc); 1681d3576fdSSage Weil page->private = 0; 1691d3576fdSSage Weil ClearPagePrivate(page); 1701d3576fdSSage Weil } 1711d3576fdSSage Weil 1721d3576fdSSage Weil static int ceph_releasepage(struct page *page, gfp_t g) 1731d3576fdSSage Weil { 1741d3576fdSSage Weil struct inode *inode = page->mapping ? page->mapping->host : NULL; 1751d3576fdSSage Weil dout("%p releasepage %p idx %lu\n", inode, page, page->index); 1761d3576fdSSage Weil WARN_ON(PageDirty(page)); 17799ccbd22SMilosz Tanski 17899ccbd22SMilosz Tanski /* Can we release the page from the cache? */ 17999ccbd22SMilosz Tanski if (!ceph_release_fscache_page(page, g)) 1801d3576fdSSage Weil return 0; 18199ccbd22SMilosz Tanski 18299ccbd22SMilosz Tanski return !PagePrivate(page); 1831d3576fdSSage Weil } 1841d3576fdSSage Weil 1851d3576fdSSage Weil /* 1861d3576fdSSage Weil * read a single page, without unlocking it. 1871d3576fdSSage Weil */ 1881d3576fdSSage Weil static int readpage_nounlock(struct file *filp, struct page *page) 1891d3576fdSSage Weil { 190496ad9aaSAl Viro struct inode *inode = file_inode(filp); 1911d3576fdSSage Weil struct ceph_inode_info *ci = ceph_inode(inode); 1923d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 1933d14c5d2SYehuda Sadeh &ceph_inode_to_client(inode)->client->osdc; 1941d3576fdSSage Weil int err = 0; 1951d3576fdSSage Weil u64 len = PAGE_CACHE_SIZE; 1961d3576fdSSage Weil 19799ccbd22SMilosz Tanski err = ceph_readpage_from_fscache(inode, page); 19899ccbd22SMilosz Tanski 19999ccbd22SMilosz Tanski if (err == 0) 20099ccbd22SMilosz Tanski goto out; 20199ccbd22SMilosz Tanski 2021d3576fdSSage Weil dout("readpage inode %p file %p page %p index %lu\n", 2031d3576fdSSage Weil inode, filp, page, page->index); 2041d3576fdSSage Weil err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout, 2056285bc23SAlex Elder (u64) page_offset(page), &len, 2061d3576fdSSage Weil ci->i_truncate_seq, ci->i_truncate_size, 207b7495fc2SSage Weil &page, 1, 0); 2081d3576fdSSage Weil if (err == -ENOENT) 2091d3576fdSSage Weil err = 0; 2101d3576fdSSage Weil if (err < 0) { 2111d3576fdSSage Weil SetPageError(page); 21218302805SLi Wang ceph_fscache_readpage_cancel(inode, page); 2131d3576fdSSage Weil goto out; 21456f91aadSLi Wang } else { 21556f91aadSLi Wang if (err < PAGE_CACHE_SIZE) { 2161d3576fdSSage Weil /* zero fill remainder of page */ 2171d3576fdSSage Weil zero_user_segment(page, err, PAGE_CACHE_SIZE); 21856f91aadSLi Wang } else { 21956f91aadSLi Wang flush_dcache_page(page); 22056f91aadSLi Wang } 2211d3576fdSSage Weil } 2221d3576fdSSage Weil SetPageUptodate(page); 2231d3576fdSSage Weil 224ff638b7dSLi Wang if (err >= 0) 22599ccbd22SMilosz Tanski ceph_readpage_to_fscache(inode, page); 22699ccbd22SMilosz Tanski 2271d3576fdSSage Weil out: 2281d3576fdSSage Weil return err < 0 ? err : 0; 2291d3576fdSSage Weil } 2301d3576fdSSage Weil 2311d3576fdSSage Weil static int ceph_readpage(struct file *filp, struct page *page) 2321d3576fdSSage Weil { 2331d3576fdSSage Weil int r = readpage_nounlock(filp, page); 2341d3576fdSSage Weil unlock_page(page); 2351d3576fdSSage Weil return r; 2361d3576fdSSage Weil } 2371d3576fdSSage Weil 2381d3576fdSSage Weil /* 2397c272194SSage Weil * Finish an async read(ahead) op. 2401d3576fdSSage Weil */ 2417c272194SSage Weil static void finish_read(struct ceph_osd_request *req, struct ceph_msg *msg) 2421d3576fdSSage Weil { 2437c272194SSage Weil struct inode *inode = req->r_inode; 24487060c10SAlex Elder struct ceph_osd_data *osd_data; 2451b83bef2SSage Weil int rc = req->r_result; 2461b83bef2SSage Weil int bytes = le32_to_cpu(msg->hdr.data_len); 247e0c59487SAlex Elder int num_pages; 2487c272194SSage Weil int i; 2497c272194SSage Weil 2507c272194SSage Weil dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes); 2517c272194SSage Weil 2527c272194SSage Weil /* unlock all pages, zeroing any data we didn't read */ 253406e2c9fSAlex Elder osd_data = osd_req_op_extent_osd_data(req, 0); 25487060c10SAlex Elder BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES); 25587060c10SAlex Elder num_pages = calc_pages_for((u64)osd_data->alignment, 25687060c10SAlex Elder (u64)osd_data->length); 257e0c59487SAlex Elder for (i = 0; i < num_pages; i++) { 25887060c10SAlex Elder struct page *page = osd_data->pages[i]; 2597c272194SSage Weil 260f36132a7SLi Wang if (rc < 0) 261f36132a7SLi Wang goto unlock; 2627c272194SSage Weil if (bytes < (int)PAGE_CACHE_SIZE) { 2637c272194SSage Weil /* zero (remainder of) page */ 2647c272194SSage Weil int s = bytes < 0 ? 0 : bytes; 2657c272194SSage Weil zero_user_segment(page, s, PAGE_CACHE_SIZE); 2667c272194SSage Weil } 2677c272194SSage Weil dout("finish_read %p uptodate %p idx %lu\n", inode, page, 2687c272194SSage Weil page->index); 2697c272194SSage Weil flush_dcache_page(page); 2707c272194SSage Weil SetPageUptodate(page); 27199ccbd22SMilosz Tanski ceph_readpage_to_fscache(inode, page); 272f36132a7SLi Wang unlock: 2737c272194SSage Weil unlock_page(page); 2747c272194SSage Weil page_cache_release(page); 2750fff87ecSAlex Elder bytes -= PAGE_CACHE_SIZE; 2767c272194SSage Weil } 27787060c10SAlex Elder kfree(osd_data->pages); 2787c272194SSage Weil } 2797c272194SSage Weil 2808884d53dSDavid Zafman static void ceph_unlock_page_vector(struct page **pages, int num_pages) 2818884d53dSDavid Zafman { 2828884d53dSDavid Zafman int i; 2838884d53dSDavid Zafman 2848884d53dSDavid Zafman for (i = 0; i < num_pages; i++) 2858884d53dSDavid Zafman unlock_page(pages[i]); 2868884d53dSDavid Zafman } 2878884d53dSDavid Zafman 2887c272194SSage Weil /* 2897c272194SSage Weil * start an async read(ahead) operation. return nr_pages we submitted 2907c272194SSage Weil * a read for on success, or negative error code. 2917c272194SSage Weil */ 2920d66a487SSage Weil static int start_read(struct inode *inode, struct list_head *page_list, int max) 2937c272194SSage Weil { 2947c272194SSage Weil struct ceph_osd_client *osdc = 2957c272194SSage Weil &ceph_inode_to_client(inode)->client->osdc; 2967c272194SSage Weil struct ceph_inode_info *ci = ceph_inode(inode); 2977c272194SSage Weil struct page *page = list_entry(page_list->prev, struct page, lru); 298acead002SAlex Elder struct ceph_vino vino; 2997c272194SSage Weil struct ceph_osd_request *req; 3007c272194SSage Weil u64 off; 3017c272194SSage Weil u64 len; 3027c272194SSage Weil int i; 3031d3576fdSSage Weil struct page **pages; 3047c272194SSage Weil pgoff_t next_index; 3057c272194SSage Weil int nr_pages = 0; 3067c272194SSage Weil int ret; 3077c272194SSage Weil 3086285bc23SAlex Elder off = (u64) page_offset(page); 3097c272194SSage Weil 3107c272194SSage Weil /* count pages */ 3117c272194SSage Weil next_index = page->index; 3127c272194SSage Weil list_for_each_entry_reverse(page, page_list, lru) { 3137c272194SSage Weil if (page->index != next_index) 3147c272194SSage Weil break; 3157c272194SSage Weil nr_pages++; 3167c272194SSage Weil next_index++; 3170d66a487SSage Weil if (max && nr_pages == max) 3180d66a487SSage Weil break; 3197c272194SSage Weil } 3207c272194SSage Weil len = nr_pages << PAGE_CACHE_SHIFT; 3217c272194SSage Weil dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages, 3227c272194SSage Weil off, len); 323acead002SAlex Elder vino = ceph_vino(inode); 324acead002SAlex Elder req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len, 32579528734SAlex Elder 1, CEPH_OSD_OP_READ, 326acead002SAlex Elder CEPH_OSD_FLAG_READ, NULL, 3277c272194SSage Weil ci->i_truncate_seq, ci->i_truncate_size, 328acead002SAlex Elder false); 3296816282dSSage Weil if (IS_ERR(req)) 3306816282dSSage Weil return PTR_ERR(req); 3311d3576fdSSage Weil 3321d3576fdSSage Weil /* build page vector */ 333cf7b7e14SAlex Elder nr_pages = calc_pages_for(0, len); 3347c272194SSage Weil pages = kmalloc(sizeof(*pages) * nr_pages, GFP_NOFS); 3357c272194SSage Weil ret = -ENOMEM; 3361d3576fdSSage Weil if (!pages) 3377c272194SSage Weil goto out; 3387c272194SSage Weil for (i = 0; i < nr_pages; ++i) { 3397c272194SSage Weil page = list_entry(page_list->prev, struct page, lru); 3407c272194SSage Weil BUG_ON(PageLocked(page)); 3417c272194SSage Weil list_del(&page->lru); 3421d3576fdSSage Weil 3437c272194SSage Weil dout("start_read %p adding %p idx %lu\n", inode, page, 3447c272194SSage Weil page->index); 3457c272194SSage Weil if (add_to_page_cache_lru(page, &inode->i_data, page->index, 3467c272194SSage Weil GFP_NOFS)) { 347d4d3aa38SMilosz Tanski ceph_fscache_uncache_page(inode, page); 3487c272194SSage Weil page_cache_release(page); 3497c272194SSage Weil dout("start_read %p add_to_page_cache failed %p\n", 3507c272194SSage Weil inode, page); 3517c272194SSage Weil nr_pages = i; 3527c272194SSage Weil goto out_pages; 3531d3576fdSSage Weil } 3547c272194SSage Weil pages[i] = page; 3551d3576fdSSage Weil } 356406e2c9fSAlex Elder osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false); 3577c272194SSage Weil req->r_callback = finish_read; 3587c272194SSage Weil req->r_inode = inode; 3597c272194SSage Weil 36079528734SAlex Elder ceph_osdc_build_request(req, off, NULL, vino.snap, NULL); 36102ee07d3SAlex Elder 3627c272194SSage Weil dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len); 3637c272194SSage Weil ret = ceph_osdc_start_request(osdc, req, false); 3647c272194SSage Weil if (ret < 0) 3657c272194SSage Weil goto out_pages; 3667c272194SSage Weil ceph_osdc_put_request(req); 3677c272194SSage Weil return nr_pages; 3687c272194SSage Weil 3697c272194SSage Weil out_pages: 3708884d53dSDavid Zafman ceph_unlock_page_vector(pages, nr_pages); 3717c272194SSage Weil ceph_release_page_vector(pages, nr_pages); 3727c272194SSage Weil out: 3737c272194SSage Weil ceph_osdc_put_request(req); 3747c272194SSage Weil return ret; 3751d3576fdSSage Weil } 3761d3576fdSSage Weil 3777c272194SSage Weil 3781d3576fdSSage Weil /* 3791d3576fdSSage Weil * Read multiple pages. Leave pages we don't read + unlock in page_list; 3801d3576fdSSage Weil * the caller (VM) cleans them up. 3811d3576fdSSage Weil */ 3821d3576fdSSage Weil static int ceph_readpages(struct file *file, struct address_space *mapping, 3831d3576fdSSage Weil struct list_head *page_list, unsigned nr_pages) 3841d3576fdSSage Weil { 385496ad9aaSAl Viro struct inode *inode = file_inode(file); 3860d66a487SSage Weil struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 3871d3576fdSSage Weil int rc = 0; 3880d66a487SSage Weil int max = 0; 3891d3576fdSSage Weil 39099ccbd22SMilosz Tanski rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list, 39199ccbd22SMilosz Tanski &nr_pages); 39299ccbd22SMilosz Tanski 39399ccbd22SMilosz Tanski if (rc == 0) 39499ccbd22SMilosz Tanski goto out; 39599ccbd22SMilosz Tanski 3960d66a487SSage Weil if (fsc->mount_options->rsize >= PAGE_CACHE_SIZE) 3970d66a487SSage Weil max = (fsc->mount_options->rsize + PAGE_CACHE_SIZE - 1) 3980d66a487SSage Weil >> PAGE_SHIFT; 3990d66a487SSage Weil 4002794a82aSAlex Elder dout("readpages %p file %p nr_pages %d max %d\n", inode, 4012794a82aSAlex Elder file, nr_pages, 4020d66a487SSage Weil max); 4037c272194SSage Weil while (!list_empty(page_list)) { 4040d66a487SSage Weil rc = start_read(inode, page_list, max); 4051d3576fdSSage Weil if (rc < 0) 4061d3576fdSSage Weil goto out; 4077c272194SSage Weil BUG_ON(rc == 0); 4081d3576fdSSage Weil } 4091d3576fdSSage Weil out: 41076be778bSMilosz Tanski ceph_fscache_readpages_cancel(inode, page_list); 41176be778bSMilosz Tanski 4127c272194SSage Weil dout("readpages %p file %p ret %d\n", inode, file, rc); 4131d3576fdSSage Weil return rc; 4141d3576fdSSage Weil } 4151d3576fdSSage Weil 4161d3576fdSSage Weil /* 4171d3576fdSSage Weil * Get ref for the oldest snapc for an inode with dirty data... that is, the 4181d3576fdSSage Weil * only snap context we are allowed to write back. 4191d3576fdSSage Weil */ 4206298a337SSage Weil static struct ceph_snap_context *get_oldest_context(struct inode *inode, 4211d3576fdSSage Weil u64 *snap_size) 4221d3576fdSSage Weil { 4231d3576fdSSage Weil struct ceph_inode_info *ci = ceph_inode(inode); 4241d3576fdSSage Weil struct ceph_snap_context *snapc = NULL; 4251d3576fdSSage Weil struct ceph_cap_snap *capsnap = NULL; 4261d3576fdSSage Weil 427be655596SSage Weil spin_lock(&ci->i_ceph_lock); 4281d3576fdSSage Weil list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) { 4291d3576fdSSage Weil dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap, 4301d3576fdSSage Weil capsnap->context, capsnap->dirty_pages); 4311d3576fdSSage Weil if (capsnap->dirty_pages) { 4321d3576fdSSage Weil snapc = ceph_get_snap_context(capsnap->context); 4331d3576fdSSage Weil if (snap_size) 4341d3576fdSSage Weil *snap_size = capsnap->size; 4351d3576fdSSage Weil break; 4361d3576fdSSage Weil } 4371d3576fdSSage Weil } 4387d8cb26dSSage Weil if (!snapc && ci->i_wrbuffer_ref_head) { 43980e755feSSage Weil snapc = ceph_get_snap_context(ci->i_head_snapc); 4401d3576fdSSage Weil dout(" head snapc %p has %d dirty pages\n", 4411d3576fdSSage Weil snapc, ci->i_wrbuffer_ref_head); 4421d3576fdSSage Weil } 443be655596SSage Weil spin_unlock(&ci->i_ceph_lock); 4441d3576fdSSage Weil return snapc; 4451d3576fdSSage Weil } 4461d3576fdSSage Weil 4471d3576fdSSage Weil /* 4481d3576fdSSage Weil * Write a single page, but leave the page locked. 4491d3576fdSSage Weil * 4501d3576fdSSage Weil * If we get a write error, set the page error bit, but still adjust the 4511d3576fdSSage Weil * dirty page accounting (i.e., page is no longer dirty). 4521d3576fdSSage Weil */ 4531d3576fdSSage Weil static int writepage_nounlock(struct page *page, struct writeback_control *wbc) 4541d3576fdSSage Weil { 4551d3576fdSSage Weil struct inode *inode; 4561d3576fdSSage Weil struct ceph_inode_info *ci; 4573d14c5d2SYehuda Sadeh struct ceph_fs_client *fsc; 4581d3576fdSSage Weil struct ceph_osd_client *osdc; 4596298a337SSage Weil struct ceph_snap_context *snapc, *oldest; 460fc2744aaSYan, Zheng loff_t page_off = page_offset(page); 4612baba250SYehuda Sadeh long writeback_stat; 462fc2744aaSYan, Zheng u64 truncate_size, snap_size = 0; 463fc2744aaSYan, Zheng u32 truncate_seq; 464fc2744aaSYan, Zheng int err = 0, len = PAGE_CACHE_SIZE; 4651d3576fdSSage Weil 4661d3576fdSSage Weil dout("writepage %p idx %lu\n", page, page->index); 4671d3576fdSSage Weil 4681d3576fdSSage Weil if (!page->mapping || !page->mapping->host) { 4691d3576fdSSage Weil dout("writepage %p - no mapping\n", page); 4701d3576fdSSage Weil return -EFAULT; 4711d3576fdSSage Weil } 4721d3576fdSSage Weil inode = page->mapping->host; 4731d3576fdSSage Weil ci = ceph_inode(inode); 4743d14c5d2SYehuda Sadeh fsc = ceph_inode_to_client(inode); 4753d14c5d2SYehuda Sadeh osdc = &fsc->client->osdc; 4761d3576fdSSage Weil 4771d3576fdSSage Weil /* verify this is a writeable snap context */ 47861600ef8SYan, Zheng snapc = page_snap_context(page); 4791d3576fdSSage Weil if (snapc == NULL) { 4801d3576fdSSage Weil dout("writepage %p page %p not dirty?\n", inode, page); 4811d3576fdSSage Weil goto out; 4821d3576fdSSage Weil } 4836298a337SSage Weil oldest = get_oldest_context(inode, &snap_size); 4846298a337SSage Weil if (snapc->seq > oldest->seq) { 4851d3576fdSSage Weil dout("writepage %p page %p snapc %p not writeable - noop\n", 48661600ef8SYan, Zheng inode, page, snapc); 4871d3576fdSSage Weil /* we should only noop if called by kswapd */ 4881d3576fdSSage Weil WARN_ON((current->flags & PF_MEMALLOC) == 0); 4896298a337SSage Weil ceph_put_snap_context(oldest); 4901d3576fdSSage Weil goto out; 4911d3576fdSSage Weil } 4926298a337SSage Weil ceph_put_snap_context(oldest); 4931d3576fdSSage Weil 494fc2744aaSYan, Zheng spin_lock(&ci->i_ceph_lock); 495fc2744aaSYan, Zheng truncate_seq = ci->i_truncate_seq; 496fc2744aaSYan, Zheng truncate_size = ci->i_truncate_size; 497fc2744aaSYan, Zheng if (!snap_size) 498fc2744aaSYan, Zheng snap_size = i_size_read(inode); 499fc2744aaSYan, Zheng spin_unlock(&ci->i_ceph_lock); 500fc2744aaSYan, Zheng 5011d3576fdSSage Weil /* is this a partial page at end of file? */ 502fc2744aaSYan, Zheng if (page_off >= snap_size) { 503fc2744aaSYan, Zheng dout("%p page eof %llu\n", page, snap_size); 504fc2744aaSYan, Zheng goto out; 505fc2744aaSYan, Zheng } 506fc2744aaSYan, Zheng if (snap_size < page_off + len) 507fc2744aaSYan, Zheng len = snap_size - page_off; 5081d3576fdSSage Weil 509ae00d4f3SSage Weil dout("writepage %p page %p index %lu on %llu~%u snapc %p\n", 510ae00d4f3SSage Weil inode, page, page->index, page_off, len, snapc); 5111d3576fdSSage Weil 5123d14c5d2SYehuda Sadeh writeback_stat = atomic_long_inc_return(&fsc->writeback_count); 5132baba250SYehuda Sadeh if (writeback_stat > 5143d14c5d2SYehuda Sadeh CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb)) 5153d14c5d2SYehuda Sadeh set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC); 5162baba250SYehuda Sadeh 51799ccbd22SMilosz Tanski ceph_readpage_to_fscache(inode, page); 51899ccbd22SMilosz Tanski 5191d3576fdSSage Weil set_page_writeback(page); 5201d3576fdSSage Weil err = ceph_osdc_writepages(osdc, ceph_vino(inode), 5211d3576fdSSage Weil &ci->i_layout, snapc, 5221d3576fdSSage Weil page_off, len, 523fc2744aaSYan, Zheng truncate_seq, truncate_size, 52424808826SAlex Elder &inode->i_mtime, &page, 1); 5251d3576fdSSage Weil if (err < 0) { 5261d3576fdSSage Weil dout("writepage setting page/mapping error %d %p\n", err, page); 5271d3576fdSSage Weil SetPageError(page); 5281d3576fdSSage Weil mapping_set_error(&inode->i_data, err); 5291d3576fdSSage Weil if (wbc) 5301d3576fdSSage Weil wbc->pages_skipped++; 5311d3576fdSSage Weil } else { 5321d3576fdSSage Weil dout("writepage cleaned page %p\n", page); 5331d3576fdSSage Weil err = 0; /* vfs expects us to return 0 */ 5341d3576fdSSage Weil } 5351d3576fdSSage Weil page->private = 0; 5361d3576fdSSage Weil ClearPagePrivate(page); 5371d3576fdSSage Weil end_page_writeback(page); 5381d3576fdSSage Weil ceph_put_wrbuffer_cap_refs(ci, 1, snapc); 5396298a337SSage Weil ceph_put_snap_context(snapc); /* page's reference */ 5401d3576fdSSage Weil out: 5411d3576fdSSage Weil return err; 5421d3576fdSSage Weil } 5431d3576fdSSage Weil 5441d3576fdSSage Weil static int ceph_writepage(struct page *page, struct writeback_control *wbc) 5451d3576fdSSage Weil { 546dbd646a8SYehuda Sadeh int err; 547dbd646a8SYehuda Sadeh struct inode *inode = page->mapping->host; 548dbd646a8SYehuda Sadeh BUG_ON(!inode); 54970b666c3SSage Weil ihold(inode); 550dbd646a8SYehuda Sadeh err = writepage_nounlock(page, wbc); 5511d3576fdSSage Weil unlock_page(page); 552dbd646a8SYehuda Sadeh iput(inode); 5531d3576fdSSage Weil return err; 5541d3576fdSSage Weil } 5551d3576fdSSage Weil 5561d3576fdSSage Weil 5571d3576fdSSage Weil /* 5581d3576fdSSage Weil * lame release_pages helper. release_pages() isn't exported to 5591d3576fdSSage Weil * modules. 5601d3576fdSSage Weil */ 5611d3576fdSSage Weil static void ceph_release_pages(struct page **pages, int num) 5621d3576fdSSage Weil { 5631d3576fdSSage Weil struct pagevec pvec; 5641d3576fdSSage Weil int i; 5651d3576fdSSage Weil 5661d3576fdSSage Weil pagevec_init(&pvec, 0); 5671d3576fdSSage Weil for (i = 0; i < num; i++) { 5681d3576fdSSage Weil if (pagevec_add(&pvec, pages[i]) == 0) 5691d3576fdSSage Weil pagevec_release(&pvec); 5701d3576fdSSage Weil } 5711d3576fdSSage Weil pagevec_release(&pvec); 5721d3576fdSSage Weil } 5731d3576fdSSage Weil 5741d3576fdSSage Weil /* 5751d3576fdSSage Weil * async writeback completion handler. 5761d3576fdSSage Weil * 5771d3576fdSSage Weil * If we get an error, set the mapping error bit, but not the individual 5781d3576fdSSage Weil * page error bits. 5791d3576fdSSage Weil */ 5801d3576fdSSage Weil static void writepages_finish(struct ceph_osd_request *req, 5811d3576fdSSage Weil struct ceph_msg *msg) 5821d3576fdSSage Weil { 5831d3576fdSSage Weil struct inode *inode = req->r_inode; 5841d3576fdSSage Weil struct ceph_inode_info *ci = ceph_inode(inode); 58587060c10SAlex Elder struct ceph_osd_data *osd_data; 5861d3576fdSSage Weil unsigned wrote; 5871d3576fdSSage Weil struct page *page; 588e0c59487SAlex Elder int num_pages; 5891d3576fdSSage Weil int i; 5901d3576fdSSage Weil struct ceph_snap_context *snapc = req->r_snapc; 5911d3576fdSSage Weil struct address_space *mapping = inode->i_mapping; 5921b83bef2SSage Weil int rc = req->r_result; 59379528734SAlex Elder u64 bytes = req->r_ops[0].extent.length; 5943d14c5d2SYehuda Sadeh struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 5952baba250SYehuda Sadeh long writeback_stat; 5967ff899daSSage Weil unsigned issued = ceph_caps_issued(ci); 5971d3576fdSSage Weil 598406e2c9fSAlex Elder osd_data = osd_req_op_extent_osd_data(req, 0); 59987060c10SAlex Elder BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES); 60087060c10SAlex Elder num_pages = calc_pages_for((u64)osd_data->alignment, 60187060c10SAlex Elder (u64)osd_data->length); 6021d3576fdSSage Weil if (rc >= 0) { 60379788c69SSage Weil /* 60479788c69SSage Weil * Assume we wrote the pages we originally sent. The 60579788c69SSage Weil * osd might reply with fewer pages if our writeback 60679788c69SSage Weil * raced with a truncation and was adjusted at the osd, 60779788c69SSage Weil * so don't believe the reply. 60879788c69SSage Weil */ 609e0c59487SAlex Elder wrote = num_pages; 6101d3576fdSSage Weil } else { 6111d3576fdSSage Weil wrote = 0; 6121d3576fdSSage Weil mapping_set_error(mapping, rc); 6131d3576fdSSage Weil } 6141d3576fdSSage Weil dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n", 6151d3576fdSSage Weil inode, rc, bytes, wrote); 6161d3576fdSSage Weil 6171d3576fdSSage Weil /* clean all pages */ 618e0c59487SAlex Elder for (i = 0; i < num_pages; i++) { 61987060c10SAlex Elder page = osd_data->pages[i]; 6201d3576fdSSage Weil BUG_ON(!page); 6211d3576fdSSage Weil WARN_ON(!PageUptodate(page)); 6221d3576fdSSage Weil 6232baba250SYehuda Sadeh writeback_stat = 6243d14c5d2SYehuda Sadeh atomic_long_dec_return(&fsc->writeback_count); 6252baba250SYehuda Sadeh if (writeback_stat < 6263d14c5d2SYehuda Sadeh CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb)) 6273d14c5d2SYehuda Sadeh clear_bdi_congested(&fsc->backing_dev_info, 6282baba250SYehuda Sadeh BLK_RW_ASYNC); 6292baba250SYehuda Sadeh 63061600ef8SYan, Zheng ceph_put_snap_context(page_snap_context(page)); 6311d3576fdSSage Weil page->private = 0; 6321d3576fdSSage Weil ClearPagePrivate(page); 6331d3576fdSSage Weil dout("unlocking %d %p\n", i, page); 6341d3576fdSSage Weil end_page_writeback(page); 635e63dc5c7SYehuda Sadeh 636e63dc5c7SYehuda Sadeh /* 637e63dc5c7SYehuda Sadeh * We lost the cache cap, need to truncate the page before 638e63dc5c7SYehuda Sadeh * it is unlocked, otherwise we'd truncate it later in the 639e63dc5c7SYehuda Sadeh * page truncation thread, possibly losing some data that 640e63dc5c7SYehuda Sadeh * raced its way in 641e63dc5c7SYehuda Sadeh */ 6422962507cSSage Weil if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) 643e63dc5c7SYehuda Sadeh generic_error_remove_page(inode->i_mapping, page); 644e63dc5c7SYehuda Sadeh 6451d3576fdSSage Weil unlock_page(page); 6461d3576fdSSage Weil } 6471d3576fdSSage Weil dout("%p wrote+cleaned %d pages\n", inode, wrote); 648e0c59487SAlex Elder ceph_put_wrbuffer_cap_refs(ci, num_pages, snapc); 6491d3576fdSSage Weil 65087060c10SAlex Elder ceph_release_pages(osd_data->pages, num_pages); 65187060c10SAlex Elder if (osd_data->pages_from_pool) 65287060c10SAlex Elder mempool_free(osd_data->pages, 653640ef79dSCheng Renquan ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool); 6541d3576fdSSage Weil else 65587060c10SAlex Elder kfree(osd_data->pages); 6561d3576fdSSage Weil ceph_osdc_put_request(req); 6571d3576fdSSage Weil } 6581d3576fdSSage Weil 6591d3576fdSSage Weil /* 6601d3576fdSSage Weil * initiate async writeback 6611d3576fdSSage Weil */ 6621d3576fdSSage Weil static int ceph_writepages_start(struct address_space *mapping, 6631d3576fdSSage Weil struct writeback_control *wbc) 6641d3576fdSSage Weil { 6651d3576fdSSage Weil struct inode *inode = mapping->host; 6661d3576fdSSage Weil struct ceph_inode_info *ci = ceph_inode(inode); 667fc2744aaSYan, Zheng struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 668fc2744aaSYan, Zheng struct ceph_vino vino = ceph_vino(inode); 6691d3576fdSSage Weil pgoff_t index, start, end; 6701d3576fdSSage Weil int range_whole = 0; 6711d3576fdSSage Weil int should_loop = 1; 6721d3576fdSSage Weil pgoff_t max_pages = 0, max_pages_ever = 0; 67380e755feSSage Weil struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc; 6741d3576fdSSage Weil struct pagevec pvec; 6751d3576fdSSage Weil int done = 0; 6761d3576fdSSage Weil int rc = 0; 6771d3576fdSSage Weil unsigned wsize = 1 << inode->i_blkbits; 6781d3576fdSSage Weil struct ceph_osd_request *req = NULL; 6791d3576fdSSage Weil int do_sync; 680fc2744aaSYan, Zheng u64 truncate_size, snap_size; 681fc2744aaSYan, Zheng u32 truncate_seq; 6821d3576fdSSage Weil 6831d3576fdSSage Weil /* 6841d3576fdSSage Weil * Include a 'sync' in the OSD request if this is a data 6851d3576fdSSage Weil * integrity write (e.g., O_SYNC write or fsync()), or if our 6861d3576fdSSage Weil * cap is being revoked. 6871d3576fdSSage Weil */ 688c62988ecSmajianpeng if ((wbc->sync_mode == WB_SYNC_ALL) || 689c62988ecSmajianpeng ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER)) 6901d3576fdSSage Weil do_sync = 1; 6911d3576fdSSage Weil dout("writepages_start %p dosync=%d (mode=%s)\n", 6921d3576fdSSage Weil inode, do_sync, 6931d3576fdSSage Weil wbc->sync_mode == WB_SYNC_NONE ? "NONE" : 6941d3576fdSSage Weil (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD")); 6951d3576fdSSage Weil 6963d14c5d2SYehuda Sadeh if (fsc->mount_state == CEPH_MOUNT_SHUTDOWN) { 697f3ae1b97SFabian Frederick pr_warn("writepage_start %p on forced umount\n", inode); 6981d3576fdSSage Weil return -EIO; /* we're in a forced umount, don't write! */ 6991d3576fdSSage Weil } 7003d14c5d2SYehuda Sadeh if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize) 7013d14c5d2SYehuda Sadeh wsize = fsc->mount_options->wsize; 7021d3576fdSSage Weil if (wsize < PAGE_CACHE_SIZE) 7031d3576fdSSage Weil wsize = PAGE_CACHE_SIZE; 7041d3576fdSSage Weil max_pages_ever = wsize >> PAGE_CACHE_SHIFT; 7051d3576fdSSage Weil 7061d3576fdSSage Weil pagevec_init(&pvec, 0); 7071d3576fdSSage Weil 7081d3576fdSSage Weil /* where to start/end? */ 7091d3576fdSSage Weil if (wbc->range_cyclic) { 7101d3576fdSSage Weil start = mapping->writeback_index; /* Start from prev offset */ 7111d3576fdSSage Weil end = -1; 7121d3576fdSSage Weil dout(" cyclic, start at %lu\n", start); 7131d3576fdSSage Weil } else { 7141d3576fdSSage Weil start = wbc->range_start >> PAGE_CACHE_SHIFT; 7151d3576fdSSage Weil end = wbc->range_end >> PAGE_CACHE_SHIFT; 7161d3576fdSSage Weil if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX) 7171d3576fdSSage Weil range_whole = 1; 7181d3576fdSSage Weil should_loop = 0; 7191d3576fdSSage Weil dout(" not cyclic, %lu to %lu\n", start, end); 7201d3576fdSSage Weil } 7211d3576fdSSage Weil index = start; 7221d3576fdSSage Weil 7231d3576fdSSage Weil retry: 7241d3576fdSSage Weil /* find oldest snap context with dirty data */ 7251d3576fdSSage Weil ceph_put_snap_context(snapc); 7261ac0fc8aSYan, Zheng snap_size = 0; 7271d3576fdSSage Weil snapc = get_oldest_context(inode, &snap_size); 7281d3576fdSSage Weil if (!snapc) { 7291d3576fdSSage Weil /* hmm, why does writepages get called when there 7301d3576fdSSage Weil is no dirty data? */ 7311d3576fdSSage Weil dout(" no snap context with dirty data?\n"); 7321d3576fdSSage Weil goto out; 7331d3576fdSSage Weil } 7341ac0fc8aSYan, Zheng if (snap_size == 0) 7351ac0fc8aSYan, Zheng snap_size = i_size_read(inode); 7361d3576fdSSage Weil dout(" oldest snapc is %p seq %lld (%d snaps)\n", 7371d3576fdSSage Weil snapc, snapc->seq, snapc->num_snaps); 738fc2744aaSYan, Zheng 739fc2744aaSYan, Zheng spin_lock(&ci->i_ceph_lock); 740fc2744aaSYan, Zheng truncate_seq = ci->i_truncate_seq; 741fc2744aaSYan, Zheng truncate_size = ci->i_truncate_size; 742fc2744aaSYan, Zheng if (!snap_size) 743fc2744aaSYan, Zheng snap_size = i_size_read(inode); 744fc2744aaSYan, Zheng spin_unlock(&ci->i_ceph_lock); 745fc2744aaSYan, Zheng 7461d3576fdSSage Weil if (last_snapc && snapc != last_snapc) { 7471d3576fdSSage Weil /* if we switched to a newer snapc, restart our scan at the 7481d3576fdSSage Weil * start of the original file range. */ 7491d3576fdSSage Weil dout(" snapc differs from last pass, restarting at %lu\n", 7501d3576fdSSage Weil index); 7511d3576fdSSage Weil index = start; 7521d3576fdSSage Weil } 7531d3576fdSSage Weil last_snapc = snapc; 7541d3576fdSSage Weil 7551d3576fdSSage Weil while (!done && index <= end) { 756e5975c7cSAlex Elder int num_ops = do_sync ? 2 : 1; 7571d3576fdSSage Weil unsigned i; 7581d3576fdSSage Weil int first; 7591d3576fdSSage Weil pgoff_t next; 7601d3576fdSSage Weil int pvec_pages, locked_pages; 761e5975c7cSAlex Elder struct page **pages = NULL; 762e5975c7cSAlex Elder mempool_t *pool = NULL; /* Becomes non-null if mempool used */ 7631d3576fdSSage Weil struct page *page; 7641d3576fdSSage Weil int want; 7651d3576fdSSage Weil u64 offset, len; 7662baba250SYehuda Sadeh long writeback_stat; 7671d3576fdSSage Weil 7681d3576fdSSage Weil next = 0; 7691d3576fdSSage Weil locked_pages = 0; 7701d3576fdSSage Weil max_pages = max_pages_ever; 7711d3576fdSSage Weil 7721d3576fdSSage Weil get_more_pages: 7731d3576fdSSage Weil first = -1; 7741d3576fdSSage Weil want = min(end - index, 7751d3576fdSSage Weil min((pgoff_t)PAGEVEC_SIZE, 7761d3576fdSSage Weil max_pages - (pgoff_t)locked_pages) - 1) 7771d3576fdSSage Weil + 1; 7781d3576fdSSage Weil pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index, 7791d3576fdSSage Weil PAGECACHE_TAG_DIRTY, 7801d3576fdSSage Weil want); 7811d3576fdSSage Weil dout("pagevec_lookup_tag got %d\n", pvec_pages); 7821d3576fdSSage Weil if (!pvec_pages && !locked_pages) 7831d3576fdSSage Weil break; 7841d3576fdSSage Weil for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) { 7851d3576fdSSage Weil page = pvec.pages[i]; 7861d3576fdSSage Weil dout("? %p idx %lu\n", page, page->index); 7871d3576fdSSage Weil if (locked_pages == 0) 7881d3576fdSSage Weil lock_page(page); /* first page */ 7891d3576fdSSage Weil else if (!trylock_page(page)) 7901d3576fdSSage Weil break; 7911d3576fdSSage Weil 7921d3576fdSSage Weil /* only dirty pages, or our accounting breaks */ 7931d3576fdSSage Weil if (unlikely(!PageDirty(page)) || 7941d3576fdSSage Weil unlikely(page->mapping != mapping)) { 7951d3576fdSSage Weil dout("!dirty or !mapping %p\n", page); 7961d3576fdSSage Weil unlock_page(page); 7971d3576fdSSage Weil break; 7981d3576fdSSage Weil } 7991d3576fdSSage Weil if (!wbc->range_cyclic && page->index > end) { 8001d3576fdSSage Weil dout("end of range %p\n", page); 8011d3576fdSSage Weil done = 1; 8021d3576fdSSage Weil unlock_page(page); 8031d3576fdSSage Weil break; 8041d3576fdSSage Weil } 8051d3576fdSSage Weil if (next && (page->index != next)) { 8061d3576fdSSage Weil dout("not consecutive %p\n", page); 8071d3576fdSSage Weil unlock_page(page); 8081d3576fdSSage Weil break; 8091d3576fdSSage Weil } 8101d3576fdSSage Weil if (wbc->sync_mode != WB_SYNC_NONE) { 8111d3576fdSSage Weil dout("waiting on writeback %p\n", page); 8121d3576fdSSage Weil wait_on_page_writeback(page); 8131d3576fdSSage Weil } 8141ac0fc8aSYan, Zheng if (page_offset(page) >= snap_size) { 8151ac0fc8aSYan, Zheng dout("%p page eof %llu\n", page, snap_size); 8161d3576fdSSage Weil done = 1; 8171d3576fdSSage Weil unlock_page(page); 8181d3576fdSSage Weil break; 8191d3576fdSSage Weil } 8201d3576fdSSage Weil if (PageWriteback(page)) { 8211d3576fdSSage Weil dout("%p under writeback\n", page); 8221d3576fdSSage Weil unlock_page(page); 8231d3576fdSSage Weil break; 8241d3576fdSSage Weil } 8251d3576fdSSage Weil 8261d3576fdSSage Weil /* only if matching snap context */ 82761600ef8SYan, Zheng pgsnapc = page_snap_context(page); 82880e755feSSage Weil if (pgsnapc->seq > snapc->seq) { 82980e755feSSage Weil dout("page snapc %p %lld > oldest %p %lld\n", 83080e755feSSage Weil pgsnapc, pgsnapc->seq, snapc, snapc->seq); 8311d3576fdSSage Weil unlock_page(page); 8321d3576fdSSage Weil if (!locked_pages) 8331d3576fdSSage Weil continue; /* keep looking for snap */ 8341d3576fdSSage Weil break; 8351d3576fdSSage Weil } 8361d3576fdSSage Weil 8371d3576fdSSage Weil if (!clear_page_dirty_for_io(page)) { 8381d3576fdSSage Weil dout("%p !clear_page_dirty_for_io\n", page); 8391d3576fdSSage Weil unlock_page(page); 8401d3576fdSSage Weil break; 8411d3576fdSSage Weil } 8421d3576fdSSage Weil 843e5975c7cSAlex Elder /* 844e5975c7cSAlex Elder * We have something to write. If this is 845e5975c7cSAlex Elder * the first locked page this time through, 846e5975c7cSAlex Elder * allocate an osd request and a page array 847e5975c7cSAlex Elder * that it will use. 848e5975c7cSAlex Elder */ 8491d3576fdSSage Weil if (locked_pages == 0) { 850e5975c7cSAlex Elder BUG_ON(pages); 8511d3576fdSSage Weil /* prepare async write request */ 8526285bc23SAlex Elder offset = (u64)page_offset(page); 8531d3576fdSSage Weil len = wsize; 854fc2744aaSYan, Zheng req = ceph_osdc_new_request(&fsc->client->osdc, 855fc2744aaSYan, Zheng &ci->i_layout, vino, 856fc2744aaSYan, Zheng offset, &len, num_ops, 857fc2744aaSYan, Zheng CEPH_OSD_OP_WRITE, 858fc2744aaSYan, Zheng CEPH_OSD_FLAG_WRITE | 859fc2744aaSYan, Zheng CEPH_OSD_FLAG_ONDISK, 860fc2744aaSYan, Zheng snapc, truncate_seq, 861fc2744aaSYan, Zheng truncate_size, true); 8626816282dSSage Weil if (IS_ERR(req)) { 8636816282dSSage Weil rc = PTR_ERR(req); 8648c71897bSHenry C Chang unlock_page(page); 8658c71897bSHenry C Chang break; 8668c71897bSHenry C Chang } 8678c71897bSHenry C Chang 86888486957SAlex Elder req->r_callback = writepages_finish; 86988486957SAlex Elder req->r_inode = inode; 87088486957SAlex Elder 87188486957SAlex Elder max_pages = calc_pages_for(0, (u64)len); 872fc2744aaSYan, Zheng pages = kmalloc(max_pages * sizeof (*pages), 873fc2744aaSYan, Zheng GFP_NOFS); 87488486957SAlex Elder if (!pages) { 87588486957SAlex Elder pool = fsc->wb_pagevec_pool; 87688486957SAlex Elder pages = mempool_alloc(pool, GFP_NOFS); 877e5975c7cSAlex Elder BUG_ON(!pages); 87888486957SAlex Elder } 8791d3576fdSSage Weil } 8801d3576fdSSage Weil 8811d3576fdSSage Weil /* note position of first page in pvec */ 8821d3576fdSSage Weil if (first < 0) 8831d3576fdSSage Weil first = i; 8841d3576fdSSage Weil dout("%p will write page %p idx %lu\n", 8851d3576fdSSage Weil inode, page, page->index); 8862baba250SYehuda Sadeh 887213c99eeSSage Weil writeback_stat = 8883d14c5d2SYehuda Sadeh atomic_long_inc_return(&fsc->writeback_count); 889213c99eeSSage Weil if (writeback_stat > CONGESTION_ON_THRESH( 8903d14c5d2SYehuda Sadeh fsc->mount_options->congestion_kb)) { 8913d14c5d2SYehuda Sadeh set_bdi_congested(&fsc->backing_dev_info, 892213c99eeSSage Weil BLK_RW_ASYNC); 8932baba250SYehuda Sadeh } 8942baba250SYehuda Sadeh 8951d3576fdSSage Weil set_page_writeback(page); 896e5975c7cSAlex Elder pages[locked_pages] = page; 8971d3576fdSSage Weil locked_pages++; 8981d3576fdSSage Weil next = page->index + 1; 8991d3576fdSSage Weil } 9001d3576fdSSage Weil 9011d3576fdSSage Weil /* did we get anything? */ 9021d3576fdSSage Weil if (!locked_pages) 9031d3576fdSSage Weil goto release_pvec_pages; 9041d3576fdSSage Weil if (i) { 9051d3576fdSSage Weil int j; 9061d3576fdSSage Weil BUG_ON(!locked_pages || first < 0); 9071d3576fdSSage Weil 9081d3576fdSSage Weil if (pvec_pages && i == pvec_pages && 9091d3576fdSSage Weil locked_pages < max_pages) { 9101d3576fdSSage Weil dout("reached end pvec, trying for more\n"); 9111d3576fdSSage Weil pagevec_reinit(&pvec); 9121d3576fdSSage Weil goto get_more_pages; 9131d3576fdSSage Weil } 9141d3576fdSSage Weil 9151d3576fdSSage Weil /* shift unused pages over in the pvec... we 9161d3576fdSSage Weil * will need to release them below. */ 9171d3576fdSSage Weil for (j = i; j < pvec_pages; j++) { 9181d3576fdSSage Weil dout(" pvec leftover page %p\n", 9191d3576fdSSage Weil pvec.pages[j]); 9201d3576fdSSage Weil pvec.pages[j-i+first] = pvec.pages[j]; 9211d3576fdSSage Weil } 9221d3576fdSSage Weil pvec.nr -= i-first; 9231d3576fdSSage Weil } 9241d3576fdSSage Weil 925e5975c7cSAlex Elder /* Format the osd request message and submit the write */ 926e5975c7cSAlex Elder 927e5975c7cSAlex Elder offset = page_offset(pages[0]); 9281ac0fc8aSYan, Zheng len = min(snap_size - offset, 9291d3576fdSSage Weil (u64)locked_pages << PAGE_CACHE_SHIFT); 9301d3576fdSSage Weil dout("writepages got %d pages at %llu~%llu\n", 9311d3576fdSSage Weil locked_pages, offset, len); 9321d3576fdSSage Weil 933406e2c9fSAlex Elder osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, 934a4ce40a9SAlex Elder !!pool, false); 935e5975c7cSAlex Elder 936e5975c7cSAlex Elder pages = NULL; /* request message now owns the pages array */ 937e5975c7cSAlex Elder pool = NULL; 938e5975c7cSAlex Elder 939e5975c7cSAlex Elder /* Update the write op length in case we changed it */ 940e5975c7cSAlex Elder 941c99d2d4aSAlex Elder osd_req_op_extent_update(req, 0, len); 942e5975c7cSAlex Elder 943e5975c7cSAlex Elder vino = ceph_vino(inode); 94479528734SAlex Elder ceph_osdc_build_request(req, offset, snapc, vino.snap, 94579528734SAlex Elder &inode->i_mtime); 9461d3576fdSSage Weil 9479d6fcb08SSage Weil rc = ceph_osdc_start_request(&fsc->client->osdc, req, true); 9489d6fcb08SSage Weil BUG_ON(rc); 9491d3576fdSSage Weil req = NULL; 9501d3576fdSSage Weil 9511d3576fdSSage Weil /* continue? */ 9521d3576fdSSage Weil index = next; 9531d3576fdSSage Weil wbc->nr_to_write -= locked_pages; 9541d3576fdSSage Weil if (wbc->nr_to_write <= 0) 9551d3576fdSSage Weil done = 1; 9561d3576fdSSage Weil 9571d3576fdSSage Weil release_pvec_pages: 9581d3576fdSSage Weil dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr, 9591d3576fdSSage Weil pvec.nr ? pvec.pages[0] : NULL); 9601d3576fdSSage Weil pagevec_release(&pvec); 9611d3576fdSSage Weil 9621d3576fdSSage Weil if (locked_pages && !done) 9631d3576fdSSage Weil goto retry; 9641d3576fdSSage Weil } 9651d3576fdSSage Weil 9661d3576fdSSage Weil if (should_loop && !done) { 9671d3576fdSSage Weil /* more to do; loop back to beginning of file */ 9681d3576fdSSage Weil dout("writepages looping back to beginning of file\n"); 9691d3576fdSSage Weil should_loop = 0; 9701d3576fdSSage Weil index = 0; 9711d3576fdSSage Weil goto retry; 9721d3576fdSSage Weil } 9731d3576fdSSage Weil 9741d3576fdSSage Weil if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) 9751d3576fdSSage Weil mapping->writeback_index = index; 9761d3576fdSSage Weil 9771d3576fdSSage Weil out: 9781d3576fdSSage Weil if (req) 9791d3576fdSSage Weil ceph_osdc_put_request(req); 9801d3576fdSSage Weil ceph_put_snap_context(snapc); 9811d3576fdSSage Weil dout("writepages done, rc = %d\n", rc); 9821d3576fdSSage Weil return rc; 9831d3576fdSSage Weil } 9841d3576fdSSage Weil 9851d3576fdSSage Weil 9861d3576fdSSage Weil 9871d3576fdSSage Weil /* 9881d3576fdSSage Weil * See if a given @snapc is either writeable, or already written. 9891d3576fdSSage Weil */ 9901d3576fdSSage Weil static int context_is_writeable_or_written(struct inode *inode, 9911d3576fdSSage Weil struct ceph_snap_context *snapc) 9921d3576fdSSage Weil { 9931d3576fdSSage Weil struct ceph_snap_context *oldest = get_oldest_context(inode, NULL); 9946298a337SSage Weil int ret = !oldest || snapc->seq <= oldest->seq; 9956298a337SSage Weil 9966298a337SSage Weil ceph_put_snap_context(oldest); 9976298a337SSage Weil return ret; 9981d3576fdSSage Weil } 9991d3576fdSSage Weil 10001d3576fdSSage Weil /* 10011d3576fdSSage Weil * We are only allowed to write into/dirty the page if the page is 10021d3576fdSSage Weil * clean, or already dirty within the same snap context. 10038f883c24SSage Weil * 10048f883c24SSage Weil * called with page locked. 10058f883c24SSage Weil * return success with page locked, 10068f883c24SSage Weil * or any failure (incl -EAGAIN) with page unlocked. 10071d3576fdSSage Weil */ 10084af6b225SYehuda Sadeh static int ceph_update_writeable_page(struct file *file, 10094af6b225SYehuda Sadeh loff_t pos, unsigned len, 10104af6b225SYehuda Sadeh struct page *page) 10111d3576fdSSage Weil { 1012496ad9aaSAl Viro struct inode *inode = file_inode(file); 10131d3576fdSSage Weil struct ceph_inode_info *ci = ceph_inode(inode); 10143d14c5d2SYehuda Sadeh struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc; 10151d3576fdSSage Weil loff_t page_off = pos & PAGE_CACHE_MASK; 10161d3576fdSSage Weil int pos_in_page = pos & ~PAGE_CACHE_MASK; 10171d3576fdSSage Weil int end_in_page = pos_in_page + len; 10181d3576fdSSage Weil loff_t i_size; 10191d3576fdSSage Weil int r; 102080e755feSSage Weil struct ceph_snap_context *snapc, *oldest; 10211d3576fdSSage Weil 10221d3576fdSSage Weil retry_locked: 10231d3576fdSSage Weil /* writepages currently holds page lock, but if we change that later, */ 10241d3576fdSSage Weil wait_on_page_writeback(page); 10251d3576fdSSage Weil 10261d3576fdSSage Weil /* check snap context */ 10271d3576fdSSage Weil BUG_ON(!ci->i_snap_realm); 10281d3576fdSSage Weil down_read(&mdsc->snap_rwsem); 10291d3576fdSSage Weil BUG_ON(!ci->i_snap_realm->cached_context); 103061600ef8SYan, Zheng snapc = page_snap_context(page); 103180e755feSSage Weil if (snapc && snapc != ci->i_head_snapc) { 10321d3576fdSSage Weil /* 10331d3576fdSSage Weil * this page is already dirty in another (older) snap 10341d3576fdSSage Weil * context! is it writeable now? 10351d3576fdSSage Weil */ 103680e755feSSage Weil oldest = get_oldest_context(inode, NULL); 10371d3576fdSSage Weil up_read(&mdsc->snap_rwsem); 10381d3576fdSSage Weil 103980e755feSSage Weil if (snapc->seq > oldest->seq) { 10406298a337SSage Weil ceph_put_snap_context(oldest); 10411d3576fdSSage Weil dout(" page %p snapc %p not current or oldest\n", 10426298a337SSage Weil page, snapc); 10431d3576fdSSage Weil /* 10441d3576fdSSage Weil * queue for writeback, and wait for snapc to 10451d3576fdSSage Weil * be writeable or written 10461d3576fdSSage Weil */ 10476298a337SSage Weil snapc = ceph_get_snap_context(snapc); 10481d3576fdSSage Weil unlock_page(page); 10493c6f6b79SSage Weil ceph_queue_writeback(inode); 10508f883c24SSage Weil r = wait_event_interruptible(ci->i_cap_wq, 10511d3576fdSSage Weil context_is_writeable_or_written(inode, snapc)); 10521d3576fdSSage Weil ceph_put_snap_context(snapc); 10538f883c24SSage Weil if (r == -ERESTARTSYS) 10548f883c24SSage Weil return r; 10554af6b225SYehuda Sadeh return -EAGAIN; 10561d3576fdSSage Weil } 10576298a337SSage Weil ceph_put_snap_context(oldest); 10581d3576fdSSage Weil 10591d3576fdSSage Weil /* yay, writeable, do it now (without dropping page lock) */ 10601d3576fdSSage Weil dout(" page %p snapc %p not current, but oldest\n", 10611d3576fdSSage Weil page, snapc); 10621d3576fdSSage Weil if (!clear_page_dirty_for_io(page)) 10631d3576fdSSage Weil goto retry_locked; 10641d3576fdSSage Weil r = writepage_nounlock(page, NULL); 10651d3576fdSSage Weil if (r < 0) 10661d3576fdSSage Weil goto fail_nosnap; 10671d3576fdSSage Weil goto retry_locked; 10681d3576fdSSage Weil } 10691d3576fdSSage Weil 10701d3576fdSSage Weil if (PageUptodate(page)) { 10711d3576fdSSage Weil dout(" page %p already uptodate\n", page); 10721d3576fdSSage Weil return 0; 10731d3576fdSSage Weil } 10741d3576fdSSage Weil 10751d3576fdSSage Weil /* full page? */ 10761d3576fdSSage Weil if (pos_in_page == 0 && len == PAGE_CACHE_SIZE) 10771d3576fdSSage Weil return 0; 10781d3576fdSSage Weil 10791d3576fdSSage Weil /* past end of file? */ 10801d3576fdSSage Weil i_size = inode->i_size; /* caller holds i_mutex */ 10811d3576fdSSage Weil 10821d3576fdSSage Weil if (i_size + len > inode->i_sb->s_maxbytes) { 10831d3576fdSSage Weil /* file is too big */ 10841d3576fdSSage Weil r = -EINVAL; 10851d3576fdSSage Weil goto fail; 10861d3576fdSSage Weil } 10871d3576fdSSage Weil 10881d3576fdSSage Weil if (page_off >= i_size || 10891d3576fdSSage Weil (pos_in_page == 0 && (pos+len) >= i_size && 10901d3576fdSSage Weil end_in_page - pos_in_page != PAGE_CACHE_SIZE)) { 10911d3576fdSSage Weil dout(" zeroing %p 0 - %d and %d - %d\n", 10921d3576fdSSage Weil page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE); 10931d3576fdSSage Weil zero_user_segments(page, 10941d3576fdSSage Weil 0, pos_in_page, 10951d3576fdSSage Weil end_in_page, PAGE_CACHE_SIZE); 10961d3576fdSSage Weil return 0; 10971d3576fdSSage Weil } 10981d3576fdSSage Weil 10991d3576fdSSage Weil /* we need to read it. */ 11001d3576fdSSage Weil up_read(&mdsc->snap_rwsem); 11011d3576fdSSage Weil r = readpage_nounlock(file, page); 11021d3576fdSSage Weil if (r < 0) 11031d3576fdSSage Weil goto fail_nosnap; 11041d3576fdSSage Weil goto retry_locked; 11051d3576fdSSage Weil 11061d3576fdSSage Weil fail: 11071d3576fdSSage Weil up_read(&mdsc->snap_rwsem); 11081d3576fdSSage Weil fail_nosnap: 11091d3576fdSSage Weil unlock_page(page); 11101d3576fdSSage Weil return r; 11111d3576fdSSage Weil } 11121d3576fdSSage Weil 11131d3576fdSSage Weil /* 11144af6b225SYehuda Sadeh * We are only allowed to write into/dirty the page if the page is 11154af6b225SYehuda Sadeh * clean, or already dirty within the same snap context. 11164af6b225SYehuda Sadeh */ 11174af6b225SYehuda Sadeh static int ceph_write_begin(struct file *file, struct address_space *mapping, 11184af6b225SYehuda Sadeh loff_t pos, unsigned len, unsigned flags, 11194af6b225SYehuda Sadeh struct page **pagep, void **fsdata) 11204af6b225SYehuda Sadeh { 1121496ad9aaSAl Viro struct inode *inode = file_inode(file); 11224af6b225SYehuda Sadeh struct page *page; 11234af6b225SYehuda Sadeh pgoff_t index = pos >> PAGE_CACHE_SHIFT; 11247971bd92SSage Weil int r; 11254af6b225SYehuda Sadeh 11264af6b225SYehuda Sadeh do { 11274af6b225SYehuda Sadeh /* get a page */ 11284af6b225SYehuda Sadeh page = grab_cache_page_write_begin(mapping, index, 0); 11297971bd92SSage Weil if (!page) 11307971bd92SSage Weil return -ENOMEM; 11317971bd92SSage Weil *pagep = page; 11324af6b225SYehuda Sadeh 11334af6b225SYehuda Sadeh dout("write_begin file %p inode %p page %p %d~%d\n", file, 11344af6b225SYehuda Sadeh inode, page, (int)pos, (int)len); 11354af6b225SYehuda Sadeh 11364af6b225SYehuda Sadeh r = ceph_update_writeable_page(file, pos, len, page); 11374af6b225SYehuda Sadeh } while (r == -EAGAIN); 11384af6b225SYehuda Sadeh 11394af6b225SYehuda Sadeh return r; 11404af6b225SYehuda Sadeh } 11414af6b225SYehuda Sadeh 11424af6b225SYehuda Sadeh /* 11431d3576fdSSage Weil * we don't do anything in here that simple_write_end doesn't do 11441d3576fdSSage Weil * except adjust dirty page accounting and drop read lock on 11451d3576fdSSage Weil * mdsc->snap_rwsem. 11461d3576fdSSage Weil */ 11471d3576fdSSage Weil static int ceph_write_end(struct file *file, struct address_space *mapping, 11481d3576fdSSage Weil loff_t pos, unsigned len, unsigned copied, 11491d3576fdSSage Weil struct page *page, void *fsdata) 11501d3576fdSSage Weil { 1151496ad9aaSAl Viro struct inode *inode = file_inode(file); 11523d14c5d2SYehuda Sadeh struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 11533d14c5d2SYehuda Sadeh struct ceph_mds_client *mdsc = fsc->mdsc; 11541d3576fdSSage Weil unsigned from = pos & (PAGE_CACHE_SIZE - 1); 11551d3576fdSSage Weil int check_cap = 0; 11561d3576fdSSage Weil 11571d3576fdSSage Weil dout("write_end file %p inode %p page %p %d~%d (%d)\n", file, 11581d3576fdSSage Weil inode, page, (int)pos, (int)copied, (int)len); 11591d3576fdSSage Weil 11601d3576fdSSage Weil /* zero the stale part of the page if we did a short copy */ 11611d3576fdSSage Weil if (copied < len) 11621d3576fdSSage Weil zero_user_segment(page, from+copied, len); 11631d3576fdSSage Weil 11641d3576fdSSage Weil /* did file size increase? */ 11651d3576fdSSage Weil /* (no need for i_size_read(); we caller holds i_mutex */ 11661d3576fdSSage Weil if (pos+copied > inode->i_size) 11671d3576fdSSage Weil check_cap = ceph_inode_set_size(inode, pos+copied); 11681d3576fdSSage Weil 11691d3576fdSSage Weil if (!PageUptodate(page)) 11701d3576fdSSage Weil SetPageUptodate(page); 11711d3576fdSSage Weil 11721d3576fdSSage Weil set_page_dirty(page); 11731d3576fdSSage Weil 11741d3576fdSSage Weil unlock_page(page); 11751d3576fdSSage Weil up_read(&mdsc->snap_rwsem); 11761d3576fdSSage Weil page_cache_release(page); 11771d3576fdSSage Weil 11781d3576fdSSage Weil if (check_cap) 11791d3576fdSSage Weil ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL); 11801d3576fdSSage Weil 11811d3576fdSSage Weil return copied; 11821d3576fdSSage Weil } 11831d3576fdSSage Weil 11841d3576fdSSage Weil /* 11851d3576fdSSage Weil * we set .direct_IO to indicate direct io is supported, but since we 11861d3576fdSSage Weil * intercept O_DIRECT reads and writes early, this function should 11871d3576fdSSage Weil * never get called. 11881d3576fdSSage Weil */ 11891d3576fdSSage Weil static ssize_t ceph_direct_io(int rw, struct kiocb *iocb, 11901d3576fdSSage Weil const struct iovec *iov, 11911d3576fdSSage Weil loff_t pos, unsigned long nr_segs) 11921d3576fdSSage Weil { 11931d3576fdSSage Weil WARN_ON(1); 11941d3576fdSSage Weil return -EINVAL; 11951d3576fdSSage Weil } 11961d3576fdSSage Weil 11971d3576fdSSage Weil const struct address_space_operations ceph_aops = { 11981d3576fdSSage Weil .readpage = ceph_readpage, 11991d3576fdSSage Weil .readpages = ceph_readpages, 12001d3576fdSSage Weil .writepage = ceph_writepage, 12011d3576fdSSage Weil .writepages = ceph_writepages_start, 12021d3576fdSSage Weil .write_begin = ceph_write_begin, 12031d3576fdSSage Weil .write_end = ceph_write_end, 12041d3576fdSSage Weil .set_page_dirty = ceph_set_page_dirty, 12051d3576fdSSage Weil .invalidatepage = ceph_invalidatepage, 12061d3576fdSSage Weil .releasepage = ceph_releasepage, 12071d3576fdSSage Weil .direct_IO = ceph_direct_io, 12081d3576fdSSage Weil }; 12091d3576fdSSage Weil 12101d3576fdSSage Weil 12111d3576fdSSage Weil /* 12121d3576fdSSage Weil * vm ops 12131d3576fdSSage Weil */ 121461f68816SYan, Zheng static int ceph_filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf) 121561f68816SYan, Zheng { 121661f68816SYan, Zheng struct inode *inode = file_inode(vma->vm_file); 121761f68816SYan, Zheng struct ceph_inode_info *ci = ceph_inode(inode); 121861f68816SYan, Zheng struct ceph_file_info *fi = vma->vm_file->private_data; 121961f68816SYan, Zheng loff_t off = vmf->pgoff << PAGE_CACHE_SHIFT; 122061f68816SYan, Zheng int want, got, ret; 122161f68816SYan, Zheng 122261f68816SYan, Zheng dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n", 122337b52fe6SIlya Dryomov inode, ceph_vinop(inode), off, (size_t)PAGE_CACHE_SIZE); 122461f68816SYan, Zheng if (fi->fmode & CEPH_FILE_MODE_LAZY) 122561f68816SYan, Zheng want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO; 122661f68816SYan, Zheng else 122761f68816SYan, Zheng want = CEPH_CAP_FILE_CACHE; 122861f68816SYan, Zheng while (1) { 122961f68816SYan, Zheng got = 0; 123061f68816SYan, Zheng ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, &got, -1); 123161f68816SYan, Zheng if (ret == 0) 123261f68816SYan, Zheng break; 123361f68816SYan, Zheng if (ret != -ERESTARTSYS) { 123461f68816SYan, Zheng WARN_ON(1); 123561f68816SYan, Zheng return VM_FAULT_SIGBUS; 123661f68816SYan, Zheng } 123761f68816SYan, Zheng } 123861f68816SYan, Zheng dout("filemap_fault %p %llu~%zd got cap refs on %s\n", 123937b52fe6SIlya Dryomov inode, off, (size_t)PAGE_CACHE_SIZE, ceph_cap_string(got)); 124061f68816SYan, Zheng 124161f68816SYan, Zheng ret = filemap_fault(vma, vmf); 124261f68816SYan, Zheng 124361f68816SYan, Zheng dout("filemap_fault %p %llu~%zd dropping cap refs on %s ret %d\n", 124437b52fe6SIlya Dryomov inode, off, (size_t)PAGE_CACHE_SIZE, ceph_cap_string(got), ret); 124561f68816SYan, Zheng ceph_put_cap_refs(ci, got); 124661f68816SYan, Zheng 124761f68816SYan, Zheng return ret; 124861f68816SYan, Zheng } 12491d3576fdSSage Weil 12501d3576fdSSage Weil /* 12511d3576fdSSage Weil * Reuse write_begin here for simplicity. 12521d3576fdSSage Weil */ 12531d3576fdSSage Weil static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf) 12541d3576fdSSage Weil { 1255496ad9aaSAl Viro struct inode *inode = file_inode(vma->vm_file); 125661f68816SYan, Zheng struct ceph_inode_info *ci = ceph_inode(inode); 125761f68816SYan, Zheng struct ceph_file_info *fi = vma->vm_file->private_data; 12583d14c5d2SYehuda Sadeh struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc; 125961f68816SYan, Zheng struct page *page = vmf->page; 12606285bc23SAlex Elder loff_t off = page_offset(page); 126161f68816SYan, Zheng loff_t size = i_size_read(inode); 126261f68816SYan, Zheng size_t len; 126361f68816SYan, Zheng int want, got, ret; 12641d3576fdSSage Weil 12651d3576fdSSage Weil if (off + PAGE_CACHE_SIZE <= size) 12661d3576fdSSage Weil len = PAGE_CACHE_SIZE; 12671d3576fdSSage Weil else 12681d3576fdSSage Weil len = size & ~PAGE_CACHE_MASK; 12691d3576fdSSage Weil 127061f68816SYan, Zheng dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n", 127161f68816SYan, Zheng inode, ceph_vinop(inode), off, len, size); 127261f68816SYan, Zheng if (fi->fmode & CEPH_FILE_MODE_LAZY) 127361f68816SYan, Zheng want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO; 127461f68816SYan, Zheng else 127561f68816SYan, Zheng want = CEPH_CAP_FILE_BUFFER; 127661f68816SYan, Zheng while (1) { 127761f68816SYan, Zheng got = 0; 127861f68816SYan, Zheng ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, &got, off + len); 127961f68816SYan, Zheng if (ret == 0) 128061f68816SYan, Zheng break; 128161f68816SYan, Zheng if (ret != -ERESTARTSYS) { 128261f68816SYan, Zheng WARN_ON(1); 128361f68816SYan, Zheng return VM_FAULT_SIGBUS; 128461f68816SYan, Zheng } 128561f68816SYan, Zheng } 128661f68816SYan, Zheng dout("page_mkwrite %p %llu~%zd got cap refs on %s\n", 128761f68816SYan, Zheng inode, off, len, ceph_cap_string(got)); 128861f68816SYan, Zheng 128961f68816SYan, Zheng /* Update time before taking page lock */ 129061f68816SYan, Zheng file_update_time(vma->vm_file); 12914af6b225SYehuda Sadeh 12924af6b225SYehuda Sadeh lock_page(page); 12934af6b225SYehuda Sadeh 12944af6b225SYehuda Sadeh ret = VM_FAULT_NOPAGE; 12954af6b225SYehuda Sadeh if ((off > size) || 12964af6b225SYehuda Sadeh (page->mapping != inode->i_mapping)) 12974af6b225SYehuda Sadeh goto out; 12984af6b225SYehuda Sadeh 12994af6b225SYehuda Sadeh ret = ceph_update_writeable_page(vma->vm_file, off, len, page); 13004af6b225SYehuda Sadeh if (ret == 0) { 13014af6b225SYehuda Sadeh /* success. we'll keep the page locked. */ 13021d3576fdSSage Weil set_page_dirty(page); 13031d3576fdSSage Weil up_read(&mdsc->snap_rwsem); 13041d3576fdSSage Weil ret = VM_FAULT_LOCKED; 13051d3576fdSSage Weil } else { 13064af6b225SYehuda Sadeh if (ret == -ENOMEM) 13074af6b225SYehuda Sadeh ret = VM_FAULT_OOM; 13084af6b225SYehuda Sadeh else 13091d3576fdSSage Weil ret = VM_FAULT_SIGBUS; 13101d3576fdSSage Weil } 13114af6b225SYehuda Sadeh out: 131261f68816SYan, Zheng if (ret != VM_FAULT_LOCKED) { 13134af6b225SYehuda Sadeh unlock_page(page); 131461f68816SYan, Zheng } else { 131561f68816SYan, Zheng int dirty; 131661f68816SYan, Zheng spin_lock(&ci->i_ceph_lock); 131761f68816SYan, Zheng dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR); 131861f68816SYan, Zheng spin_unlock(&ci->i_ceph_lock); 131961f68816SYan, Zheng if (dirty) 132061f68816SYan, Zheng __mark_inode_dirty(inode, dirty); 132161f68816SYan, Zheng } 132261f68816SYan, Zheng 132361f68816SYan, Zheng dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %d\n", 132461f68816SYan, Zheng inode, off, len, ceph_cap_string(got), ret); 132561f68816SYan, Zheng ceph_put_cap_refs(ci, got); 132661f68816SYan, Zheng 13271d3576fdSSage Weil return ret; 13281d3576fdSSage Weil } 13291d3576fdSSage Weil 13301d3576fdSSage Weil static struct vm_operations_struct ceph_vmops = { 133161f68816SYan, Zheng .fault = ceph_filemap_fault, 13321d3576fdSSage Weil .page_mkwrite = ceph_page_mkwrite, 13330b173bc4SKonstantin Khlebnikov .remap_pages = generic_file_remap_pages, 13341d3576fdSSage Weil }; 13351d3576fdSSage Weil 13361d3576fdSSage Weil int ceph_mmap(struct file *file, struct vm_area_struct *vma) 13371d3576fdSSage Weil { 13381d3576fdSSage Weil struct address_space *mapping = file->f_mapping; 13391d3576fdSSage Weil 13401d3576fdSSage Weil if (!mapping->a_ops->readpage) 13411d3576fdSSage Weil return -ENOEXEC; 13421d3576fdSSage Weil file_accessed(file); 13431d3576fdSSage Weil vma->vm_ops = &ceph_vmops; 13441d3576fdSSage Weil return 0; 13451d3576fdSSage Weil } 1346