Commit | Line | Data |
---|---|---|
3d14c5d2 | 1 | #include <linux/ceph/ceph_debug.h> |
1d3576fd SW |
2 | |
3 | #include <linux/backing-dev.h> | |
4 | #include <linux/fs.h> | |
5 | #include <linux/mm.h> | |
6 | #include <linux/pagemap.h> | |
7 | #include <linux/writeback.h> /* generic_writepages */ | |
5a0e3ad6 | 8 | #include <linux/slab.h> |
1d3576fd SW |
9 | #include <linux/pagevec.h> |
10 | #include <linux/task_io_accounting_ops.h> | |
f361bf4a | 11 | #include <linux/signal.h> |
1d3576fd SW |
12 | |
13 | #include "super.h" | |
3d14c5d2 | 14 | #include "mds_client.h" |
99ccbd22 | 15 | #include "cache.h" |
3d14c5d2 | 16 | #include <linux/ceph/osd_client.h> |
1d3576fd SW |
17 | |
18 | /* | |
19 | * Ceph address space ops. | |
20 | * | |
21 | * There are a few funny things going on here. | |
22 | * | |
23 | * The page->private field is used to reference a struct | |
24 | * ceph_snap_context for _every_ dirty page. This indicates which | |
25 | * snapshot the page was logically dirtied in, and thus which snap | |
26 | * context needs to be associated with the osd write during writeback. | |
27 | * | |
28 | * Similarly, struct ceph_inode_info maintains a set of counters to | |
25985edc | 29 | * count dirty pages on the inode. In the absence of snapshots, |
1d3576fd SW |
30 | * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count. |
31 | * | |
32 | * When a snapshot is taken (that is, when the client receives | |
33 | * notification that a snapshot was taken), each inode with caps and | |
34 | * with dirty pages (dirty pages implies there is a cap) gets a new | |
35 | * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending | |
36 | * order, new snaps go to the tail). The i_wrbuffer_ref_head count is | |
37 | * moved to capsnap->dirty. (Unless a sync write is currently in | |
38 | * progress. In that case, the capsnap is said to be "pending", new | |
39 | * writes cannot start, and the capsnap isn't "finalized" until the | |
40 | * write completes (or fails) and a final size/mtime for the inode for | |
41 | * that snap can be settled upon.) i_wrbuffer_ref_head is reset to 0. | |
42 | * | |
43 | * On writeback, we must submit writes to the osd IN SNAP ORDER. So, | |
44 | * we look for the first capsnap in i_cap_snaps and write out pages in | |
45 | * that snap context _only_. Then we move on to the next capsnap, | |
46 | * eventually reaching the "live" or "head" context (i.e., pages that | |
47 | * are not yet snapped) and are writing the most recently dirtied | |
48 | * pages. | |
49 | * | |
50 | * Invalidate and so forth must take care to ensure the dirty page | |
51 | * accounting is preserved. | |
52 | */ | |
53 | ||
2baba250 YS |
54 | #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10)) |
55 | #define CONGESTION_OFF_THRESH(congestion_kb) \ | |
56 | (CONGESTION_ON_THRESH(congestion_kb) - \ | |
57 | (CONGESTION_ON_THRESH(congestion_kb) >> 2)) | |
58 | ||
61600ef8 YZ |
59 | static inline struct ceph_snap_context *page_snap_context(struct page *page) |
60 | { | |
61 | if (PagePrivate(page)) | |
62 | return (void *)page->private; | |
63 | return NULL; | |
64 | } | |
1d3576fd SW |
65 | |
66 | /* | |
67 | * Dirty a page. Optimistically adjust accounting, on the assumption | |
68 | * that we won't race with invalidate. If we do, readjust. | |
69 | */ | |
70 | static int ceph_set_page_dirty(struct page *page) | |
71 | { | |
72 | struct address_space *mapping = page->mapping; | |
73 | struct inode *inode; | |
74 | struct ceph_inode_info *ci; | |
1d3576fd | 75 | struct ceph_snap_context *snapc; |
7d6e1f54 | 76 | int ret; |
1d3576fd SW |
77 | |
78 | if (unlikely(!mapping)) | |
79 | return !TestSetPageDirty(page); | |
80 | ||
7d6e1f54 | 81 | if (PageDirty(page)) { |
1d3576fd SW |
82 | dout("%p set_page_dirty %p idx %lu -- already dirty\n", |
83 | mapping->host, page, page->index); | |
7d6e1f54 | 84 | BUG_ON(!PagePrivate(page)); |
1d3576fd SW |
85 | return 0; |
86 | } | |
87 | ||
88 | inode = mapping->host; | |
89 | ci = ceph_inode(inode); | |
90 | ||
1d3576fd | 91 | /* dirty the head */ |
be655596 | 92 | spin_lock(&ci->i_ceph_lock); |
5dda377c YZ |
93 | BUG_ON(ci->i_wr_ref == 0); // caller should hold Fw reference |
94 | if (__ceph_have_pending_cap_snap(ci)) { | |
95 | struct ceph_cap_snap *capsnap = | |
96 | list_last_entry(&ci->i_cap_snaps, | |
97 | struct ceph_cap_snap, | |
98 | ci_item); | |
99 | snapc = ceph_get_snap_context(capsnap->context); | |
100 | capsnap->dirty_pages++; | |
101 | } else { | |
102 | BUG_ON(!ci->i_head_snapc); | |
103 | snapc = ceph_get_snap_context(ci->i_head_snapc); | |
104 | ++ci->i_wrbuffer_ref_head; | |
105 | } | |
1d3576fd | 106 | if (ci->i_wrbuffer_ref == 0) |
0444d76a | 107 | ihold(inode); |
1d3576fd SW |
108 | ++ci->i_wrbuffer_ref; |
109 | dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d " | |
110 | "snapc %p seq %lld (%d snaps)\n", | |
111 | mapping->host, page, page->index, | |
112 | ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1, | |
113 | ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head, | |
114 | snapc, snapc->seq, snapc->num_snaps); | |
be655596 | 115 | spin_unlock(&ci->i_ceph_lock); |
1d3576fd | 116 | |
7d6e1f54 SZ |
117 | /* |
118 | * Reference snap context in page->private. Also set | |
119 | * PagePrivate so that we get invalidatepage callback. | |
120 | */ | |
121 | BUG_ON(PagePrivate(page)); | |
122 | page->private = (unsigned long)snapc; | |
123 | SetPagePrivate(page); | |
1d3576fd | 124 | |
7d6e1f54 SZ |
125 | ret = __set_page_dirty_nobuffers(page); |
126 | WARN_ON(!PageLocked(page)); | |
127 | WARN_ON(!page->mapping); | |
1d3576fd | 128 | |
7d6e1f54 | 129 | return ret; |
1d3576fd SW |
130 | } |
131 | ||
132 | /* | |
133 | * If we are truncating the full page (i.e. offset == 0), adjust the | |
134 | * dirty page counters appropriately. Only called if there is private | |
135 | * data on the page. | |
136 | */ | |
d47992f8 LC |
137 | static void ceph_invalidatepage(struct page *page, unsigned int offset, |
138 | unsigned int length) | |
1d3576fd | 139 | { |
4ce1e9ad | 140 | struct inode *inode; |
1d3576fd | 141 | struct ceph_inode_info *ci; |
61600ef8 | 142 | struct ceph_snap_context *snapc = page_snap_context(page); |
1d3576fd | 143 | |
4ce1e9ad | 144 | inode = page->mapping->host; |
b150f5c1 MT |
145 | ci = ceph_inode(inode); |
146 | ||
09cbfeaf | 147 | if (offset != 0 || length != PAGE_SIZE) { |
b150f5c1 MT |
148 | dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n", |
149 | inode, page, page->index, offset, length); | |
150 | return; | |
151 | } | |
4ce1e9ad | 152 | |
99ccbd22 MT |
153 | ceph_invalidate_fscache_page(inode, page); |
154 | ||
b072d774 | 155 | WARN_ON(!PageLocked(page)); |
99ccbd22 MT |
156 | if (!PagePrivate(page)) |
157 | return; | |
158 | ||
b150f5c1 | 159 | ClearPageChecked(page); |
1d3576fd | 160 | |
b150f5c1 MT |
161 | dout("%p invalidatepage %p idx %lu full dirty page\n", |
162 | inode, page, page->index); | |
163 | ||
164 | ceph_put_wrbuffer_cap_refs(ci, 1, snapc); | |
165 | ceph_put_snap_context(snapc); | |
166 | page->private = 0; | |
167 | ClearPagePrivate(page); | |
1d3576fd SW |
168 | } |
169 | ||
1d3576fd SW |
170 | static int ceph_releasepage(struct page *page, gfp_t g) |
171 | { | |
e55f1a18 N |
172 | dout("%p releasepage %p idx %lu (%sdirty)\n", page->mapping->host, |
173 | page, page->index, PageDirty(page) ? "" : "not "); | |
99ccbd22 MT |
174 | |
175 | /* Can we release the page from the cache? */ | |
176 | if (!ceph_release_fscache_page(page, g)) | |
177 | return 0; | |
178 | ||
179 | return !PagePrivate(page); | |
1d3576fd SW |
180 | } |
181 | ||
182 | /* | |
183 | * read a single page, without unlocking it. | |
184 | */ | |
dd2bc473 | 185 | static int ceph_do_readpage(struct file *filp, struct page *page) |
1d3576fd | 186 | { |
496ad9aa | 187 | struct inode *inode = file_inode(filp); |
1d3576fd | 188 | struct ceph_inode_info *ci = ceph_inode(inode); |
99ccbd22 | 189 | struct ceph_osd_client *osdc = |
3d14c5d2 | 190 | &ceph_inode_to_client(inode)->client->osdc; |
1d3576fd | 191 | int err = 0; |
83701246 | 192 | u64 off = page_offset(page); |
09cbfeaf | 193 | u64 len = PAGE_SIZE; |
1d3576fd | 194 | |
83701246 | 195 | if (off >= i_size_read(inode)) { |
09cbfeaf | 196 | zero_user_segment(page, 0, PAGE_SIZE); |
83701246 YZ |
197 | SetPageUptodate(page); |
198 | return 0; | |
199 | } | |
99ccbd22 | 200 | |
fcc02d2a YZ |
201 | if (ci->i_inline_version != CEPH_INLINE_NONE) { |
202 | /* | |
203 | * Uptodate inline data should have been added | |
204 | * into page cache while getting Fcr caps. | |
205 | */ | |
206 | if (off == 0) | |
207 | return -EINVAL; | |
09cbfeaf | 208 | zero_user_segment(page, 0, PAGE_SIZE); |
fcc02d2a YZ |
209 | SetPageUptodate(page); |
210 | return 0; | |
211 | } | |
83701246 YZ |
212 | |
213 | err = ceph_readpage_from_fscache(inode, page); | |
99ccbd22 | 214 | if (err == 0) |
dd2bc473 | 215 | return -EINPROGRESS; |
99ccbd22 | 216 | |
1d3576fd SW |
217 | dout("readpage inode %p file %p page %p index %lu\n", |
218 | inode, filp, page, page->index); | |
219 | err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout, | |
83701246 | 220 | off, &len, |
1d3576fd | 221 | ci->i_truncate_seq, ci->i_truncate_size, |
b7495fc2 | 222 | &page, 1, 0); |
1d3576fd SW |
223 | if (err == -ENOENT) |
224 | err = 0; | |
225 | if (err < 0) { | |
226 | SetPageError(page); | |
18302805 | 227 | ceph_fscache_readpage_cancel(inode, page); |
1d3576fd | 228 | goto out; |
1d3576fd | 229 | } |
09cbfeaf | 230 | if (err < PAGE_SIZE) |
23cd573b | 231 | /* zero fill remainder of page */ |
09cbfeaf | 232 | zero_user_segment(page, err, PAGE_SIZE); |
23cd573b ZZ |
233 | else |
234 | flush_dcache_page(page); | |
1d3576fd | 235 | |
23cd573b ZZ |
236 | SetPageUptodate(page); |
237 | ceph_readpage_to_fscache(inode, page); | |
99ccbd22 | 238 | |
1d3576fd SW |
239 | out: |
240 | return err < 0 ? err : 0; | |
241 | } | |
242 | ||
243 | static int ceph_readpage(struct file *filp, struct page *page) | |
244 | { | |
dd2bc473 YZ |
245 | int r = ceph_do_readpage(filp, page); |
246 | if (r != -EINPROGRESS) | |
247 | unlock_page(page); | |
248 | else | |
249 | r = 0; | |
1d3576fd SW |
250 | return r; |
251 | } | |
252 | ||
253 | /* | |
7c272194 | 254 | * Finish an async read(ahead) op. |
1d3576fd | 255 | */ |
85e084fe | 256 | static void finish_read(struct ceph_osd_request *req) |
1d3576fd | 257 | { |
7c272194 | 258 | struct inode *inode = req->r_inode; |
87060c10 | 259 | struct ceph_osd_data *osd_data; |
85e084fe ID |
260 | int rc = req->r_result <= 0 ? req->r_result : 0; |
261 | int bytes = req->r_result >= 0 ? req->r_result : 0; | |
e0c59487 | 262 | int num_pages; |
7c272194 | 263 | int i; |
1d3576fd | 264 | |
7c272194 SW |
265 | dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes); |
266 | ||
267 | /* unlock all pages, zeroing any data we didn't read */ | |
406e2c9f | 268 | osd_data = osd_req_op_extent_osd_data(req, 0); |
87060c10 AE |
269 | BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES); |
270 | num_pages = calc_pages_for((u64)osd_data->alignment, | |
271 | (u64)osd_data->length); | |
e0c59487 | 272 | for (i = 0; i < num_pages; i++) { |
87060c10 | 273 | struct page *page = osd_data->pages[i]; |
7c272194 | 274 | |
368e3585 YZ |
275 | if (rc < 0 && rc != -ENOENT) { |
276 | ceph_fscache_readpage_cancel(inode, page); | |
f36132a7 | 277 | goto unlock; |
368e3585 | 278 | } |
09cbfeaf | 279 | if (bytes < (int)PAGE_SIZE) { |
7c272194 SW |
280 | /* zero (remainder of) page */ |
281 | int s = bytes < 0 ? 0 : bytes; | |
09cbfeaf | 282 | zero_user_segment(page, s, PAGE_SIZE); |
1d3576fd | 283 | } |
7c272194 SW |
284 | dout("finish_read %p uptodate %p idx %lu\n", inode, page, |
285 | page->index); | |
286 | flush_dcache_page(page); | |
287 | SetPageUptodate(page); | |
99ccbd22 | 288 | ceph_readpage_to_fscache(inode, page); |
f36132a7 | 289 | unlock: |
7c272194 | 290 | unlock_page(page); |
09cbfeaf KS |
291 | put_page(page); |
292 | bytes -= PAGE_SIZE; | |
1d3576fd | 293 | } |
87060c10 | 294 | kfree(osd_data->pages); |
1d3576fd SW |
295 | } |
296 | ||
297 | /* | |
7c272194 SW |
298 | * start an async read(ahead) operation. return nr_pages we submitted |
299 | * a read for on success, or negative error code. | |
1d3576fd | 300 | */ |
0d66a487 | 301 | static int start_read(struct inode *inode, struct list_head *page_list, int max) |
1d3576fd | 302 | { |
3d14c5d2 YS |
303 | struct ceph_osd_client *osdc = |
304 | &ceph_inode_to_client(inode)->client->osdc; | |
7c272194 SW |
305 | struct ceph_inode_info *ci = ceph_inode(inode); |
306 | struct page *page = list_entry(page_list->prev, struct page, lru); | |
acead002 | 307 | struct ceph_vino vino; |
7c272194 SW |
308 | struct ceph_osd_request *req; |
309 | u64 off; | |
1d3576fd | 310 | u64 len; |
7c272194 SW |
311 | int i; |
312 | struct page **pages; | |
313 | pgoff_t next_index; | |
314 | int nr_pages = 0; | |
2b1ac852 YZ |
315 | int got = 0; |
316 | int ret = 0; | |
317 | ||
318 | if (!current->journal_info) { | |
319 | /* caller of readpages does not hold buffer and read caps | |
320 | * (fadvise, madvise and readahead cases) */ | |
321 | int want = CEPH_CAP_FILE_CACHE; | |
322 | ret = ceph_try_get_caps(ci, CEPH_CAP_FILE_RD, want, &got); | |
323 | if (ret < 0) { | |
324 | dout("start_read %p, error getting cap\n", inode); | |
325 | } else if (!(got & want)) { | |
326 | dout("start_read %p, no cache cap\n", inode); | |
327 | ret = 0; | |
328 | } | |
329 | if (ret <= 0) { | |
330 | if (got) | |
331 | ceph_put_cap_refs(ci, got); | |
332 | while (!list_empty(page_list)) { | |
333 | page = list_entry(page_list->prev, | |
334 | struct page, lru); | |
335 | list_del(&page->lru); | |
336 | put_page(page); | |
337 | } | |
338 | return ret; | |
339 | } | |
340 | } | |
1d3576fd | 341 | |
6285bc23 | 342 | off = (u64) page_offset(page); |
1d3576fd | 343 | |
7c272194 SW |
344 | /* count pages */ |
345 | next_index = page->index; | |
346 | list_for_each_entry_reverse(page, page_list, lru) { | |
347 | if (page->index != next_index) | |
348 | break; | |
349 | nr_pages++; | |
350 | next_index++; | |
0d66a487 SW |
351 | if (max && nr_pages == max) |
352 | break; | |
7c272194 | 353 | } |
09cbfeaf | 354 | len = nr_pages << PAGE_SHIFT; |
7c272194 SW |
355 | dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages, |
356 | off, len); | |
acead002 AE |
357 | vino = ceph_vino(inode); |
358 | req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len, | |
715e4cd4 | 359 | 0, 1, CEPH_OSD_OP_READ, |
acead002 | 360 | CEPH_OSD_FLAG_READ, NULL, |
7c272194 | 361 | ci->i_truncate_seq, ci->i_truncate_size, |
acead002 | 362 | false); |
2b1ac852 YZ |
363 | if (IS_ERR(req)) { |
364 | ret = PTR_ERR(req); | |
365 | goto out; | |
366 | } | |
1d3576fd | 367 | |
7c272194 | 368 | /* build page vector */ |
cf7b7e14 | 369 | nr_pages = calc_pages_for(0, len); |
687265e5 | 370 | pages = kmalloc(sizeof(*pages) * nr_pages, GFP_KERNEL); |
2b1ac852 YZ |
371 | if (!pages) { |
372 | ret = -ENOMEM; | |
373 | goto out_put; | |
374 | } | |
7c272194 SW |
375 | for (i = 0; i < nr_pages; ++i) { |
376 | page = list_entry(page_list->prev, struct page, lru); | |
377 | BUG_ON(PageLocked(page)); | |
1d3576fd | 378 | list_del(&page->lru); |
99ccbd22 | 379 | |
7c272194 SW |
380 | dout("start_read %p adding %p idx %lu\n", inode, page, |
381 | page->index); | |
382 | if (add_to_page_cache_lru(page, &inode->i_data, page->index, | |
687265e5 | 383 | GFP_KERNEL)) { |
d4d3aa38 | 384 | ceph_fscache_uncache_page(inode, page); |
09cbfeaf | 385 | put_page(page); |
7c272194 | 386 | dout("start_read %p add_to_page_cache failed %p\n", |
1d3576fd | 387 | inode, page); |
7c272194 | 388 | nr_pages = i; |
1afe4785 YZ |
389 | if (nr_pages > 0) { |
390 | len = nr_pages << PAGE_SHIFT; | |
d641df81 | 391 | osd_req_op_extent_update(req, 0, len); |
1afe4785 YZ |
392 | break; |
393 | } | |
7c272194 | 394 | goto out_pages; |
1d3576fd | 395 | } |
7c272194 | 396 | pages[i] = page; |
1d3576fd | 397 | } |
406e2c9f | 398 | osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false); |
7c272194 SW |
399 | req->r_callback = finish_read; |
400 | req->r_inode = inode; | |
401 | ||
402 | dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len); | |
403 | ret = ceph_osdc_start_request(osdc, req, false); | |
404 | if (ret < 0) | |
405 | goto out_pages; | |
406 | ceph_osdc_put_request(req); | |
2b1ac852 YZ |
407 | |
408 | /* After adding locked pages to page cache, the inode holds cache cap. | |
409 | * So we can drop our cap refs. */ | |
410 | if (got) | |
411 | ceph_put_cap_refs(ci, got); | |
412 | ||
7c272194 SW |
413 | return nr_pages; |
414 | ||
415 | out_pages: | |
1afe4785 YZ |
416 | for (i = 0; i < nr_pages; ++i) { |
417 | ceph_fscache_readpage_cancel(inode, pages[i]); | |
418 | unlock_page(pages[i]); | |
419 | } | |
420 | ceph_put_page_vector(pages, nr_pages, false); | |
2b1ac852 | 421 | out_put: |
7c272194 | 422 | ceph_osdc_put_request(req); |
2b1ac852 YZ |
423 | out: |
424 | if (got) | |
425 | ceph_put_cap_refs(ci, got); | |
7c272194 SW |
426 | return ret; |
427 | } | |
1d3576fd | 428 | |
7c272194 SW |
429 | |
430 | /* | |
431 | * Read multiple pages. Leave pages we don't read + unlock in page_list; | |
432 | * the caller (VM) cleans them up. | |
433 | */ | |
434 | static int ceph_readpages(struct file *file, struct address_space *mapping, | |
435 | struct list_head *page_list, unsigned nr_pages) | |
436 | { | |
496ad9aa | 437 | struct inode *inode = file_inode(file); |
0d66a487 | 438 | struct ceph_fs_client *fsc = ceph_inode_to_client(inode); |
7c272194 | 439 | int rc = 0; |
0d66a487 SW |
440 | int max = 0; |
441 | ||
83701246 YZ |
442 | if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE) |
443 | return -EINVAL; | |
444 | ||
99ccbd22 MT |
445 | rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list, |
446 | &nr_pages); | |
447 | ||
448 | if (rc == 0) | |
449 | goto out; | |
450 | ||
aa187926 YZ |
451 | max = fsc->mount_options->rsize >> PAGE_SHIFT; |
452 | dout("readpages %p file %p nr_pages %d max %d\n", | |
453 | inode, file, nr_pages, max); | |
7c272194 | 454 | while (!list_empty(page_list)) { |
0d66a487 | 455 | rc = start_read(inode, page_list, max); |
7c272194 SW |
456 | if (rc < 0) |
457 | goto out; | |
7c272194 | 458 | } |
1d3576fd | 459 | out: |
76be778b MT |
460 | ceph_fscache_readpages_cancel(inode, page_list); |
461 | ||
7c272194 | 462 | dout("readpages %p file %p ret %d\n", inode, file, rc); |
1d3576fd SW |
463 | return rc; |
464 | } | |
465 | ||
1f934b00 YZ |
466 | struct ceph_writeback_ctl |
467 | { | |
468 | loff_t i_size; | |
469 | u64 truncate_size; | |
470 | u32 truncate_seq; | |
471 | bool size_stable; | |
2a2d927e | 472 | bool head_snapc; |
1f934b00 YZ |
473 | }; |
474 | ||
1d3576fd SW |
475 | /* |
476 | * Get ref for the oldest snapc for an inode with dirty data... that is, the | |
477 | * only snap context we are allowed to write back. | |
1d3576fd | 478 | */ |
1f934b00 | 479 | static struct ceph_snap_context * |
05455e11 YZ |
480 | get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl, |
481 | struct ceph_snap_context *page_snapc) | |
1d3576fd SW |
482 | { |
483 | struct ceph_inode_info *ci = ceph_inode(inode); | |
484 | struct ceph_snap_context *snapc = NULL; | |
485 | struct ceph_cap_snap *capsnap = NULL; | |
486 | ||
be655596 | 487 | spin_lock(&ci->i_ceph_lock); |
1d3576fd SW |
488 | list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) { |
489 | dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap, | |
490 | capsnap->context, capsnap->dirty_pages); | |
05455e11 YZ |
491 | if (!capsnap->dirty_pages) |
492 | continue; | |
493 | ||
494 | /* get i_size, truncate_{seq,size} for page_snapc? */ | |
495 | if (snapc && capsnap->context != page_snapc) | |
496 | continue; | |
497 | ||
498 | if (ctl) { | |
499 | if (capsnap->writing) { | |
500 | ctl->i_size = i_size_read(inode); | |
501 | ctl->size_stable = false; | |
502 | } else { | |
503 | ctl->i_size = capsnap->size; | |
504 | ctl->size_stable = true; | |
1f934b00 | 505 | } |
05455e11 YZ |
506 | ctl->truncate_size = capsnap->truncate_size; |
507 | ctl->truncate_seq = capsnap->truncate_seq; | |
2a2d927e | 508 | ctl->head_snapc = false; |
1d3576fd | 509 | } |
05455e11 YZ |
510 | |
511 | if (snapc) | |
512 | break; | |
513 | ||
514 | snapc = ceph_get_snap_context(capsnap->context); | |
515 | if (!page_snapc || | |
516 | page_snapc == snapc || | |
517 | page_snapc->seq > snapc->seq) | |
518 | break; | |
1d3576fd | 519 | } |
7d8cb26d | 520 | if (!snapc && ci->i_wrbuffer_ref_head) { |
80e755fe | 521 | snapc = ceph_get_snap_context(ci->i_head_snapc); |
1d3576fd SW |
522 | dout(" head snapc %p has %d dirty pages\n", |
523 | snapc, ci->i_wrbuffer_ref_head); | |
1f934b00 YZ |
524 | if (ctl) { |
525 | ctl->i_size = i_size_read(inode); | |
526 | ctl->truncate_size = ci->i_truncate_size; | |
527 | ctl->truncate_seq = ci->i_truncate_seq; | |
528 | ctl->size_stable = false; | |
2a2d927e | 529 | ctl->head_snapc = true; |
1f934b00 | 530 | } |
1d3576fd | 531 | } |
be655596 | 532 | spin_unlock(&ci->i_ceph_lock); |
1d3576fd SW |
533 | return snapc; |
534 | } | |
535 | ||
1f934b00 YZ |
536 | static u64 get_writepages_data_length(struct inode *inode, |
537 | struct page *page, u64 start) | |
538 | { | |
539 | struct ceph_inode_info *ci = ceph_inode(inode); | |
540 | struct ceph_snap_context *snapc = page_snap_context(page); | |
541 | struct ceph_cap_snap *capsnap = NULL; | |
542 | u64 end = i_size_read(inode); | |
543 | ||
544 | if (snapc != ci->i_head_snapc) { | |
545 | bool found = false; | |
546 | spin_lock(&ci->i_ceph_lock); | |
547 | list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) { | |
548 | if (capsnap->context == snapc) { | |
549 | if (!capsnap->writing) | |
550 | end = capsnap->size; | |
551 | found = true; | |
552 | break; | |
553 | } | |
554 | } | |
555 | spin_unlock(&ci->i_ceph_lock); | |
556 | WARN_ON(!found); | |
557 | } | |
558 | if (end > page_offset(page) + PAGE_SIZE) | |
559 | end = page_offset(page) + PAGE_SIZE; | |
560 | return end > start ? end - start : 0; | |
561 | } | |
562 | ||
1d3576fd SW |
563 | /* |
564 | * Write a single page, but leave the page locked. | |
565 | * | |
566 | * If we get a write error, set the page error bit, but still adjust the | |
567 | * dirty page accounting (i.e., page is no longer dirty). | |
568 | */ | |
569 | static int writepage_nounlock(struct page *page, struct writeback_control *wbc) | |
570 | { | |
571 | struct inode *inode; | |
572 | struct ceph_inode_info *ci; | |
3d14c5d2 | 573 | struct ceph_fs_client *fsc; |
6298a337 | 574 | struct ceph_snap_context *snapc, *oldest; |
fc2744aa | 575 | loff_t page_off = page_offset(page); |
2baba250 | 576 | long writeback_stat; |
43986881 | 577 | int err, len = PAGE_SIZE; |
1f934b00 | 578 | struct ceph_writeback_ctl ceph_wbc; |
1d3576fd SW |
579 | |
580 | dout("writepage %p idx %lu\n", page, page->index); | |
581 | ||
1d3576fd SW |
582 | inode = page->mapping->host; |
583 | ci = ceph_inode(inode); | |
3d14c5d2 | 584 | fsc = ceph_inode_to_client(inode); |
1d3576fd SW |
585 | |
586 | /* verify this is a writeable snap context */ | |
61600ef8 | 587 | snapc = page_snap_context(page); |
d37b1d99 | 588 | if (!snapc) { |
1d3576fd | 589 | dout("writepage %p page %p not dirty?\n", inode, page); |
43986881 | 590 | return 0; |
1d3576fd | 591 | } |
05455e11 | 592 | oldest = get_oldest_context(inode, &ceph_wbc, snapc); |
6298a337 | 593 | if (snapc->seq > oldest->seq) { |
1d3576fd | 594 | dout("writepage %p page %p snapc %p not writeable - noop\n", |
61600ef8 | 595 | inode, page, snapc); |
1d3576fd | 596 | /* we should only noop if called by kswapd */ |
fa71fefb | 597 | WARN_ON(!(current->flags & PF_MEMALLOC)); |
6298a337 | 598 | ceph_put_snap_context(oldest); |
fa71fefb | 599 | redirty_page_for_writepage(wbc, page); |
43986881 | 600 | return 0; |
1d3576fd | 601 | } |
6298a337 | 602 | ceph_put_snap_context(oldest); |
1d3576fd SW |
603 | |
604 | /* is this a partial page at end of file? */ | |
1f934b00 YZ |
605 | if (page_off >= ceph_wbc.i_size) { |
606 | dout("%p page eof %llu\n", page, ceph_wbc.i_size); | |
05455e11 | 607 | page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE); |
43986881 | 608 | return 0; |
fc2744aa | 609 | } |
43986881 | 610 | |
1f934b00 YZ |
611 | if (ceph_wbc.i_size < page_off + len) |
612 | len = ceph_wbc.i_size - page_off; | |
1d3576fd | 613 | |
1c0a9c2d YZ |
614 | dout("writepage %p page %p index %lu on %llu~%u snapc %p seq %lld\n", |
615 | inode, page, page->index, page_off, len, snapc, snapc->seq); | |
1d3576fd | 616 | |
3d14c5d2 | 617 | writeback_stat = atomic_long_inc_return(&fsc->writeback_count); |
2baba250 | 618 | if (writeback_stat > |
3d14c5d2 | 619 | CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb)) |
09dc9fc2 | 620 | set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC); |
2baba250 | 621 | |
1d3576fd | 622 | set_page_writeback(page); |
1f934b00 YZ |
623 | err = ceph_osdc_writepages(&fsc->client->osdc, ceph_vino(inode), |
624 | &ci->i_layout, snapc, page_off, len, | |
625 | ceph_wbc.truncate_seq, | |
626 | ceph_wbc.truncate_size, | |
24808826 | 627 | &inode->i_mtime, &page, 1); |
1d3576fd | 628 | if (err < 0) { |
ad15ec06 YZ |
629 | struct writeback_control tmp_wbc; |
630 | if (!wbc) | |
631 | wbc = &tmp_wbc; | |
632 | if (err == -ERESTARTSYS) { | |
633 | /* killed by SIGKILL */ | |
634 | dout("writepage interrupted page %p\n", page); | |
635 | redirty_page_for_writepage(wbc, page); | |
636 | end_page_writeback(page); | |
43986881 | 637 | return err; |
ad15ec06 YZ |
638 | } |
639 | dout("writepage setting page/mapping error %d %p\n", | |
640 | err, page); | |
1d3576fd SW |
641 | SetPageError(page); |
642 | mapping_set_error(&inode->i_data, err); | |
ad15ec06 | 643 | wbc->pages_skipped++; |
1d3576fd SW |
644 | } else { |
645 | dout("writepage cleaned page %p\n", page); | |
646 | err = 0; /* vfs expects us to return 0 */ | |
647 | } | |
648 | page->private = 0; | |
649 | ClearPagePrivate(page); | |
650 | end_page_writeback(page); | |
651 | ceph_put_wrbuffer_cap_refs(ci, 1, snapc); | |
6298a337 | 652 | ceph_put_snap_context(snapc); /* page's reference */ |
1d3576fd SW |
653 | return err; |
654 | } | |
655 | ||
656 | static int ceph_writepage(struct page *page, struct writeback_control *wbc) | |
657 | { | |
dbd646a8 YS |
658 | int err; |
659 | struct inode *inode = page->mapping->host; | |
660 | BUG_ON(!inode); | |
70b666c3 | 661 | ihold(inode); |
dbd646a8 | 662 | err = writepage_nounlock(page, wbc); |
ad15ec06 YZ |
663 | if (err == -ERESTARTSYS) { |
664 | /* direct memory reclaimer was killed by SIGKILL. return 0 | |
665 | * to prevent caller from setting mapping/page error */ | |
666 | err = 0; | |
667 | } | |
1d3576fd | 668 | unlock_page(page); |
dbd646a8 | 669 | iput(inode); |
1d3576fd SW |
670 | return err; |
671 | } | |
672 | ||
1d3576fd SW |
673 | /* |
674 | * lame release_pages helper. release_pages() isn't exported to | |
675 | * modules. | |
676 | */ | |
677 | static void ceph_release_pages(struct page **pages, int num) | |
678 | { | |
679 | struct pagevec pvec; | |
680 | int i; | |
681 | ||
682 | pagevec_init(&pvec, 0); | |
683 | for (i = 0; i < num; i++) { | |
684 | if (pagevec_add(&pvec, pages[i]) == 0) | |
685 | pagevec_release(&pvec); | |
686 | } | |
687 | pagevec_release(&pvec); | |
688 | } | |
689 | ||
1d3576fd SW |
690 | /* |
691 | * async writeback completion handler. | |
692 | * | |
693 | * If we get an error, set the mapping error bit, but not the individual | |
694 | * page error bits. | |
695 | */ | |
85e084fe | 696 | static void writepages_finish(struct ceph_osd_request *req) |
1d3576fd SW |
697 | { |
698 | struct inode *inode = req->r_inode; | |
1d3576fd | 699 | struct ceph_inode_info *ci = ceph_inode(inode); |
87060c10 | 700 | struct ceph_osd_data *osd_data; |
1d3576fd | 701 | struct page *page; |
5b64640c YZ |
702 | int num_pages, total_pages = 0; |
703 | int i, j; | |
704 | int rc = req->r_result; | |
1d3576fd SW |
705 | struct ceph_snap_context *snapc = req->r_snapc; |
706 | struct address_space *mapping = inode->i_mapping; | |
3d14c5d2 | 707 | struct ceph_fs_client *fsc = ceph_inode_to_client(inode); |
5b64640c | 708 | bool remove_page; |
1d3576fd | 709 | |
5b64640c | 710 | dout("writepages_finish %p rc %d\n", inode, rc); |
26544c62 | 711 | if (rc < 0) { |
1d3576fd | 712 | mapping_set_error(mapping, rc); |
26544c62 JL |
713 | ceph_set_error_write(ci); |
714 | } else { | |
715 | ceph_clear_error_write(ci); | |
716 | } | |
5b64640c YZ |
717 | |
718 | /* | |
719 | * We lost the cache cap, need to truncate the page before | |
720 | * it is unlocked, otherwise we'd truncate it later in the | |
721 | * page truncation thread, possibly losing some data that | |
722 | * raced its way in | |
723 | */ | |
724 | remove_page = !(ceph_caps_issued(ci) & | |
725 | (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)); | |
1d3576fd SW |
726 | |
727 | /* clean all pages */ | |
5b64640c YZ |
728 | for (i = 0; i < req->r_num_ops; i++) { |
729 | if (req->r_ops[i].op != CEPH_OSD_OP_WRITE) | |
730 | break; | |
e63dc5c7 | 731 | |
5b64640c YZ |
732 | osd_data = osd_req_op_extent_osd_data(req, i); |
733 | BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES); | |
734 | num_pages = calc_pages_for((u64)osd_data->alignment, | |
735 | (u64)osd_data->length); | |
736 | total_pages += num_pages; | |
737 | for (j = 0; j < num_pages; j++) { | |
738 | page = osd_data->pages[j]; | |
739 | BUG_ON(!page); | |
740 | WARN_ON(!PageUptodate(page)); | |
741 | ||
742 | if (atomic_long_dec_return(&fsc->writeback_count) < | |
743 | CONGESTION_OFF_THRESH( | |
744 | fsc->mount_options->congestion_kb)) | |
09dc9fc2 | 745 | clear_bdi_congested(inode_to_bdi(inode), |
5b64640c YZ |
746 | BLK_RW_ASYNC); |
747 | ||
748 | ceph_put_snap_context(page_snap_context(page)); | |
749 | page->private = 0; | |
750 | ClearPagePrivate(page); | |
751 | dout("unlocking %p\n", page); | |
752 | end_page_writeback(page); | |
753 | ||
754 | if (remove_page) | |
755 | generic_error_remove_page(inode->i_mapping, | |
756 | page); | |
757 | ||
758 | unlock_page(page); | |
759 | } | |
760 | dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n", | |
761 | inode, osd_data->length, rc >= 0 ? num_pages : 0); | |
e63dc5c7 | 762 | |
5b64640c | 763 | ceph_release_pages(osd_data->pages, num_pages); |
1d3576fd | 764 | } |
1d3576fd | 765 | |
5b64640c YZ |
766 | ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc); |
767 | ||
768 | osd_data = osd_req_op_extent_osd_data(req, 0); | |
87060c10 AE |
769 | if (osd_data->pages_from_pool) |
770 | mempool_free(osd_data->pages, | |
640ef79d | 771 | ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool); |
1d3576fd | 772 | else |
87060c10 | 773 | kfree(osd_data->pages); |
1d3576fd SW |
774 | ceph_osdc_put_request(req); |
775 | } | |
776 | ||
1d3576fd SW |
777 | /* |
778 | * initiate async writeback | |
779 | */ | |
780 | static int ceph_writepages_start(struct address_space *mapping, | |
781 | struct writeback_control *wbc) | |
782 | { | |
783 | struct inode *inode = mapping->host; | |
1d3576fd | 784 | struct ceph_inode_info *ci = ceph_inode(inode); |
fc2744aa YZ |
785 | struct ceph_fs_client *fsc = ceph_inode_to_client(inode); |
786 | struct ceph_vino vino = ceph_vino(inode); | |
2a2d927e | 787 | pgoff_t index, start_index, end = -1; |
80e755fe | 788 | struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc; |
1d3576fd | 789 | struct pagevec pvec; |
1d3576fd | 790 | int rc = 0; |
93407472 | 791 | unsigned int wsize = i_blocksize(inode); |
1d3576fd | 792 | struct ceph_osd_request *req = NULL; |
1f934b00 | 793 | struct ceph_writeback_ctl ceph_wbc; |
590e9d98 YZ |
794 | bool should_loop, range_whole = false; |
795 | bool stop, done = false; | |
1d3576fd | 796 | |
3fb99d48 | 797 | dout("writepages_start %p (mode=%s)\n", inode, |
1d3576fd SW |
798 | wbc->sync_mode == WB_SYNC_NONE ? "NONE" : |
799 | (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD")); | |
800 | ||
52953d55 | 801 | if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) { |
6c93df5d YZ |
802 | if (ci->i_wrbuffer_ref > 0) { |
803 | pr_warn_ratelimited( | |
804 | "writepage_start %p %lld forced umount\n", | |
805 | inode, ceph_ino(inode)); | |
806 | } | |
a341d4df | 807 | mapping_set_error(mapping, -EIO); |
1d3576fd SW |
808 | return -EIO; /* we're in a forced umount, don't write! */ |
809 | } | |
95cca2b4 | 810 | if (fsc->mount_options->wsize < wsize) |
3d14c5d2 | 811 | wsize = fsc->mount_options->wsize; |
1d3576fd SW |
812 | |
813 | pagevec_init(&pvec, 0); | |
814 | ||
590e9d98 | 815 | start_index = wbc->range_cyclic ? mapping->writeback_index : 0; |
2a2d927e | 816 | index = start_index; |
1d3576fd SW |
817 | |
818 | retry: | |
819 | /* find oldest snap context with dirty data */ | |
05455e11 | 820 | snapc = get_oldest_context(inode, &ceph_wbc, NULL); |
1d3576fd SW |
821 | if (!snapc) { |
822 | /* hmm, why does writepages get called when there | |
823 | is no dirty data? */ | |
824 | dout(" no snap context with dirty data?\n"); | |
825 | goto out; | |
826 | } | |
827 | dout(" oldest snapc is %p seq %lld (%d snaps)\n", | |
828 | snapc, snapc->seq, snapc->num_snaps); | |
fc2744aa | 829 | |
2a2d927e YZ |
830 | should_loop = false; |
831 | if (ceph_wbc.head_snapc && snapc != last_snapc) { | |
832 | /* where to start/end? */ | |
833 | if (wbc->range_cyclic) { | |
834 | index = start_index; | |
835 | end = -1; | |
836 | if (index > 0) | |
837 | should_loop = true; | |
838 | dout(" cyclic, start at %lu\n", index); | |
839 | } else { | |
840 | index = wbc->range_start >> PAGE_SHIFT; | |
841 | end = wbc->range_end >> PAGE_SHIFT; | |
842 | if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX) | |
843 | range_whole = true; | |
844 | dout(" not cyclic, %lu to %lu\n", index, end); | |
845 | } | |
846 | } else if (!ceph_wbc.head_snapc) { | |
847 | /* Do not respect wbc->range_{start,end}. Dirty pages | |
848 | * in that range can be associated with newer snapc. | |
849 | * They are not writeable until we write all dirty pages | |
850 | * associated with 'snapc' get written */ | |
851 | if (index > 0 || wbc->sync_mode != WB_SYNC_NONE) | |
852 | should_loop = true; | |
853 | dout(" non-head snapc, range whole\n"); | |
1d3576fd | 854 | } |
2a2d927e YZ |
855 | |
856 | ceph_put_snap_context(last_snapc); | |
1d3576fd SW |
857 | last_snapc = snapc; |
858 | ||
590e9d98 YZ |
859 | stop = false; |
860 | while (!stop && index <= end) { | |
5b64640c | 861 | int num_ops = 0, op_idx; |
0e5ecac7 | 862 | unsigned i, pvec_pages, max_pages, locked_pages = 0; |
5b64640c | 863 | struct page **pages = NULL, **data_pages; |
e5975c7c | 864 | mempool_t *pool = NULL; /* Becomes non-null if mempool used */ |
1d3576fd | 865 | struct page *page; |
0e5ecac7 | 866 | pgoff_t strip_unit_end = 0; |
5b64640c | 867 | u64 offset = 0, len = 0; |
1d3576fd | 868 | |
0e5ecac7 | 869 | max_pages = wsize >> PAGE_SHIFT; |
1d3576fd SW |
870 | |
871 | get_more_pages: | |
0e5ecac7 YZ |
872 | pvec_pages = min_t(unsigned, PAGEVEC_SIZE, |
873 | max_pages - locked_pages); | |
874 | if (end - index < (u64)(pvec_pages - 1)) | |
875 | pvec_pages = (unsigned)(end - index) + 1; | |
876 | ||
1d3576fd SW |
877 | pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index, |
878 | PAGECACHE_TAG_DIRTY, | |
0e5ecac7 | 879 | pvec_pages); |
1d3576fd SW |
880 | dout("pagevec_lookup_tag got %d\n", pvec_pages); |
881 | if (!pvec_pages && !locked_pages) | |
882 | break; | |
883 | for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) { | |
884 | page = pvec.pages[i]; | |
885 | dout("? %p idx %lu\n", page, page->index); | |
886 | if (locked_pages == 0) | |
887 | lock_page(page); /* first page */ | |
888 | else if (!trylock_page(page)) | |
889 | break; | |
890 | ||
891 | /* only dirty pages, or our accounting breaks */ | |
892 | if (unlikely(!PageDirty(page)) || | |
893 | unlikely(page->mapping != mapping)) { | |
894 | dout("!dirty or !mapping %p\n", page); | |
895 | unlock_page(page); | |
0713e5f2 | 896 | continue; |
1d3576fd | 897 | } |
590e9d98 | 898 | if (page->index > end) { |
1d3576fd | 899 | dout("end of range %p\n", page); |
590e9d98 YZ |
900 | /* can't be range_cyclic (1st pass) because |
901 | * end == -1 in that case. */ | |
2a2d927e YZ |
902 | stop = true; |
903 | if (ceph_wbc.head_snapc) | |
904 | done = true; | |
1d3576fd SW |
905 | unlock_page(page); |
906 | break; | |
907 | } | |
5b64640c YZ |
908 | if (strip_unit_end && (page->index > strip_unit_end)) { |
909 | dout("end of strip unit %p\n", page); | |
1d3576fd SW |
910 | unlock_page(page); |
911 | break; | |
912 | } | |
1f934b00 YZ |
913 | if (page_offset(page) >= ceph_wbc.i_size) { |
914 | dout("%p page eof %llu\n", | |
915 | page, ceph_wbc.i_size); | |
590e9d98 YZ |
916 | /* not done if range_cyclic */ |
917 | stop = true; | |
1d3576fd SW |
918 | unlock_page(page); |
919 | break; | |
920 | } | |
921 | if (PageWriteback(page)) { | |
0713e5f2 YZ |
922 | if (wbc->sync_mode == WB_SYNC_NONE) { |
923 | dout("%p under writeback\n", page); | |
924 | unlock_page(page); | |
925 | continue; | |
926 | } | |
927 | dout("waiting on writeback %p\n", page); | |
928 | wait_on_page_writeback(page); | |
1d3576fd SW |
929 | } |
930 | ||
931 | /* only if matching snap context */ | |
61600ef8 | 932 | pgsnapc = page_snap_context(page); |
80e755fe SW |
933 | if (pgsnapc->seq > snapc->seq) { |
934 | dout("page snapc %p %lld > oldest %p %lld\n", | |
935 | pgsnapc, pgsnapc->seq, snapc, snapc->seq); | |
1d3576fd | 936 | unlock_page(page); |
0713e5f2 | 937 | continue; |
1d3576fd SW |
938 | } |
939 | ||
940 | if (!clear_page_dirty_for_io(page)) { | |
941 | dout("%p !clear_page_dirty_for_io\n", page); | |
942 | unlock_page(page); | |
0713e5f2 | 943 | continue; |
1d3576fd SW |
944 | } |
945 | ||
e5975c7c AE |
946 | /* |
947 | * We have something to write. If this is | |
948 | * the first locked page this time through, | |
5b64640c YZ |
949 | * calculate max possinle write size and |
950 | * allocate a page array | |
e5975c7c | 951 | */ |
1d3576fd | 952 | if (locked_pages == 0) { |
5b64640c YZ |
953 | u64 objnum; |
954 | u64 objoff; | |
955 | ||
1d3576fd | 956 | /* prepare async write request */ |
e5975c7c | 957 | offset = (u64)page_offset(page); |
1d3576fd | 958 | len = wsize; |
5b64640c YZ |
959 | |
960 | rc = ceph_calc_file_object_mapping(&ci->i_layout, | |
961 | offset, len, | |
962 | &objnum, &objoff, | |
963 | &len); | |
964 | if (rc < 0) { | |
8c71897b HC |
965 | unlock_page(page); |
966 | break; | |
967 | } | |
968 | ||
3fb99d48 | 969 | num_ops = 1; |
5b64640c | 970 | strip_unit_end = page->index + |
09cbfeaf | 971 | ((len - 1) >> PAGE_SHIFT); |
88486957 | 972 | |
5b64640c | 973 | BUG_ON(pages); |
88486957 | 974 | max_pages = calc_pages_for(0, (u64)len); |
fc2744aa YZ |
975 | pages = kmalloc(max_pages * sizeof (*pages), |
976 | GFP_NOFS); | |
88486957 AE |
977 | if (!pages) { |
978 | pool = fsc->wb_pagevec_pool; | |
88486957 | 979 | pages = mempool_alloc(pool, GFP_NOFS); |
e5975c7c | 980 | BUG_ON(!pages); |
88486957 | 981 | } |
5b64640c YZ |
982 | |
983 | len = 0; | |
984 | } else if (page->index != | |
09cbfeaf | 985 | (offset + len) >> PAGE_SHIFT) { |
5b64640c YZ |
986 | if (num_ops >= (pool ? CEPH_OSD_SLAB_OPS : |
987 | CEPH_OSD_MAX_OPS)) { | |
988 | redirty_page_for_writepage(wbc, page); | |
989 | unlock_page(page); | |
990 | break; | |
991 | } | |
992 | ||
993 | num_ops++; | |
994 | offset = (u64)page_offset(page); | |
995 | len = 0; | |
1d3576fd SW |
996 | } |
997 | ||
998 | /* note position of first page in pvec */ | |
1d3576fd SW |
999 | dout("%p will write page %p idx %lu\n", |
1000 | inode, page, page->index); | |
2baba250 | 1001 | |
5b64640c YZ |
1002 | if (atomic_long_inc_return(&fsc->writeback_count) > |
1003 | CONGESTION_ON_THRESH( | |
3d14c5d2 | 1004 | fsc->mount_options->congestion_kb)) { |
09dc9fc2 | 1005 | set_bdi_congested(inode_to_bdi(inode), |
213c99ee | 1006 | BLK_RW_ASYNC); |
2baba250 YS |
1007 | } |
1008 | ||
0713e5f2 YZ |
1009 | |
1010 | pages[locked_pages++] = page; | |
1011 | pvec.pages[i] = NULL; | |
1012 | ||
09cbfeaf | 1013 | len += PAGE_SIZE; |
1d3576fd SW |
1014 | } |
1015 | ||
1016 | /* did we get anything? */ | |
1017 | if (!locked_pages) | |
1018 | goto release_pvec_pages; | |
1019 | if (i) { | |
0713e5f2 YZ |
1020 | unsigned j, n = 0; |
1021 | /* shift unused page to beginning of pvec */ | |
1022 | for (j = 0; j < pvec_pages; j++) { | |
1023 | if (!pvec.pages[j]) | |
1024 | continue; | |
1025 | if (n < j) | |
1026 | pvec.pages[n] = pvec.pages[j]; | |
1027 | n++; | |
1028 | } | |
1029 | pvec.nr = n; | |
1d3576fd SW |
1030 | |
1031 | if (pvec_pages && i == pvec_pages && | |
1032 | locked_pages < max_pages) { | |
1033 | dout("reached end pvec, trying for more\n"); | |
0713e5f2 | 1034 | pagevec_release(&pvec); |
1d3576fd SW |
1035 | goto get_more_pages; |
1036 | } | |
1d3576fd SW |
1037 | } |
1038 | ||
5b64640c | 1039 | new_request: |
e5975c7c | 1040 | offset = page_offset(pages[0]); |
5b64640c YZ |
1041 | len = wsize; |
1042 | ||
1043 | req = ceph_osdc_new_request(&fsc->client->osdc, | |
1044 | &ci->i_layout, vino, | |
1045 | offset, &len, 0, num_ops, | |
1f934b00 YZ |
1046 | CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE, |
1047 | snapc, ceph_wbc.truncate_seq, | |
1048 | ceph_wbc.truncate_size, false); | |
5b64640c YZ |
1049 | if (IS_ERR(req)) { |
1050 | req = ceph_osdc_new_request(&fsc->client->osdc, | |
1051 | &ci->i_layout, vino, | |
1052 | offset, &len, 0, | |
1053 | min(num_ops, | |
1054 | CEPH_OSD_SLAB_OPS), | |
1055 | CEPH_OSD_OP_WRITE, | |
54ea0046 | 1056 | CEPH_OSD_FLAG_WRITE, |
1f934b00 YZ |
1057 | snapc, ceph_wbc.truncate_seq, |
1058 | ceph_wbc.truncate_size, true); | |
5b64640c | 1059 | BUG_ON(IS_ERR(req)); |
e1966b49 | 1060 | } |
5b64640c | 1061 | BUG_ON(len < page_offset(pages[locked_pages - 1]) + |
09cbfeaf | 1062 | PAGE_SIZE - offset); |
5b64640c YZ |
1063 | |
1064 | req->r_callback = writepages_finish; | |
1065 | req->r_inode = inode; | |
1d3576fd | 1066 | |
5b64640c YZ |
1067 | /* Format the osd request message and submit the write */ |
1068 | len = 0; | |
1069 | data_pages = pages; | |
1070 | op_idx = 0; | |
1071 | for (i = 0; i < locked_pages; i++) { | |
1072 | u64 cur_offset = page_offset(pages[i]); | |
1073 | if (offset + len != cur_offset) { | |
3fb99d48 | 1074 | if (op_idx + 1 == req->r_num_ops) |
5b64640c YZ |
1075 | break; |
1076 | osd_req_op_extent_dup_last(req, op_idx, | |
1077 | cur_offset - offset); | |
1078 | dout("writepages got pages at %llu~%llu\n", | |
1079 | offset, len); | |
1080 | osd_req_op_extent_osd_data_pages(req, op_idx, | |
1081 | data_pages, len, 0, | |
a4ce40a9 | 1082 | !!pool, false); |
5b64640c | 1083 | osd_req_op_extent_update(req, op_idx, len); |
e5975c7c | 1084 | |
5b64640c YZ |
1085 | len = 0; |
1086 | offset = cur_offset; | |
1087 | data_pages = pages + i; | |
1088 | op_idx++; | |
1089 | } | |
1090 | ||
1091 | set_page_writeback(pages[i]); | |
09cbfeaf | 1092 | len += PAGE_SIZE; |
5b64640c YZ |
1093 | } |
1094 | ||
1f934b00 YZ |
1095 | if (ceph_wbc.size_stable) { |
1096 | len = min(len, ceph_wbc.i_size - offset); | |
5b64640c YZ |
1097 | } else if (i == locked_pages) { |
1098 | /* writepages_finish() clears writeback pages | |
1099 | * according to the data length, so make sure | |
1100 | * data length covers all locked pages */ | |
09cbfeaf | 1101 | u64 min_len = len + 1 - PAGE_SIZE; |
1f934b00 YZ |
1102 | len = get_writepages_data_length(inode, pages[i - 1], |
1103 | offset); | |
5b64640c YZ |
1104 | len = max(len, min_len); |
1105 | } | |
1106 | dout("writepages got pages at %llu~%llu\n", offset, len); | |
e5975c7c | 1107 | |
5b64640c YZ |
1108 | osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len, |
1109 | 0, !!pool, false); | |
1110 | osd_req_op_extent_update(req, op_idx, len); | |
e5975c7c | 1111 | |
5b64640c YZ |
1112 | BUG_ON(op_idx + 1 != req->r_num_ops); |
1113 | ||
1114 | pool = NULL; | |
1115 | if (i < locked_pages) { | |
1116 | BUG_ON(num_ops <= req->r_num_ops); | |
1117 | num_ops -= req->r_num_ops; | |
5b64640c YZ |
1118 | locked_pages -= i; |
1119 | ||
1120 | /* allocate new pages array for next request */ | |
1121 | data_pages = pages; | |
1122 | pages = kmalloc(locked_pages * sizeof (*pages), | |
1123 | GFP_NOFS); | |
1124 | if (!pages) { | |
1125 | pool = fsc->wb_pagevec_pool; | |
1126 | pages = mempool_alloc(pool, GFP_NOFS); | |
1127 | BUG_ON(!pages); | |
1128 | } | |
1129 | memcpy(pages, data_pages + i, | |
1130 | locked_pages * sizeof(*pages)); | |
1131 | memset(data_pages + i, 0, | |
1132 | locked_pages * sizeof(*pages)); | |
1133 | } else { | |
1134 | BUG_ON(num_ops != req->r_num_ops); | |
1135 | index = pages[i - 1]->index + 1; | |
1136 | /* request message now owns the pages array */ | |
1137 | pages = NULL; | |
1138 | } | |
e5975c7c | 1139 | |
bb873b53 | 1140 | req->r_mtime = inode->i_mtime; |
9d6fcb08 SW |
1141 | rc = ceph_osdc_start_request(&fsc->client->osdc, req, true); |
1142 | BUG_ON(rc); | |
1d3576fd SW |
1143 | req = NULL; |
1144 | ||
5b64640c YZ |
1145 | wbc->nr_to_write -= i; |
1146 | if (pages) | |
1147 | goto new_request; | |
1148 | ||
2a2d927e YZ |
1149 | /* |
1150 | * We stop writing back only if we are not doing | |
1151 | * integrity sync. In case of integrity sync we have to | |
1152 | * keep going until we have written all the pages | |
1153 | * we tagged for writeback prior to entering this loop. | |
1154 | */ | |
1155 | if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE) | |
1156 | done = stop = true; | |
1d3576fd SW |
1157 | |
1158 | release_pvec_pages: | |
1159 | dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr, | |
1160 | pvec.nr ? pvec.pages[0] : NULL); | |
1161 | pagevec_release(&pvec); | |
1d3576fd SW |
1162 | } |
1163 | ||
1164 | if (should_loop && !done) { | |
1165 | /* more to do; loop back to beginning of file */ | |
1166 | dout("writepages looping back to beginning of file\n"); | |
2a2d927e YZ |
1167 | end = start_index - 1; /* OK even when start_index == 0 */ |
1168 | start_index = 0; | |
1d3576fd SW |
1169 | index = 0; |
1170 | goto retry; | |
1171 | } | |
1172 | ||
1173 | if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) | |
1174 | mapping->writeback_index = index; | |
1175 | ||
1176 | out: | |
3ed97d63 | 1177 | ceph_osdc_put_request(req); |
2a2d927e YZ |
1178 | ceph_put_snap_context(last_snapc); |
1179 | dout("writepages dend - startone, rc = %d\n", rc); | |
1d3576fd SW |
1180 | return rc; |
1181 | } | |
1182 | ||
1183 | ||
1184 | ||
1185 | /* | |
1186 | * See if a given @snapc is either writeable, or already written. | |
1187 | */ | |
1188 | static int context_is_writeable_or_written(struct inode *inode, | |
1189 | struct ceph_snap_context *snapc) | |
1190 | { | |
05455e11 | 1191 | struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL); |
6298a337 SW |
1192 | int ret = !oldest || snapc->seq <= oldest->seq; |
1193 | ||
1194 | ceph_put_snap_context(oldest); | |
1195 | return ret; | |
1d3576fd SW |
1196 | } |
1197 | ||
1198 | /* | |
1199 | * We are only allowed to write into/dirty the page if the page is | |
1200 | * clean, or already dirty within the same snap context. | |
8f883c24 SW |
1201 | * |
1202 | * called with page locked. | |
1203 | * return success with page locked, | |
1204 | * or any failure (incl -EAGAIN) with page unlocked. | |
1d3576fd | 1205 | */ |
4af6b225 YS |
1206 | static int ceph_update_writeable_page(struct file *file, |
1207 | loff_t pos, unsigned len, | |
1208 | struct page *page) | |
1d3576fd | 1209 | { |
496ad9aa | 1210 | struct inode *inode = file_inode(file); |
6c93df5d | 1211 | struct ceph_fs_client *fsc = ceph_inode_to_client(inode); |
1d3576fd | 1212 | struct ceph_inode_info *ci = ceph_inode(inode); |
09cbfeaf KS |
1213 | loff_t page_off = pos & PAGE_MASK; |
1214 | int pos_in_page = pos & ~PAGE_MASK; | |
1d3576fd SW |
1215 | int end_in_page = pos_in_page + len; |
1216 | loff_t i_size; | |
1d3576fd | 1217 | int r; |
80e755fe | 1218 | struct ceph_snap_context *snapc, *oldest; |
1d3576fd | 1219 | |
52953d55 | 1220 | if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) { |
6c93df5d YZ |
1221 | dout(" page %p forced umount\n", page); |
1222 | unlock_page(page); | |
1223 | return -EIO; | |
1224 | } | |
1225 | ||
1d3576fd SW |
1226 | retry_locked: |
1227 | /* writepages currently holds page lock, but if we change that later, */ | |
1228 | wait_on_page_writeback(page); | |
1229 | ||
61600ef8 | 1230 | snapc = page_snap_context(page); |
80e755fe | 1231 | if (snapc && snapc != ci->i_head_snapc) { |
1d3576fd SW |
1232 | /* |
1233 | * this page is already dirty in another (older) snap | |
1234 | * context! is it writeable now? | |
1235 | */ | |
05455e11 | 1236 | oldest = get_oldest_context(inode, NULL, NULL); |
80e755fe | 1237 | if (snapc->seq > oldest->seq) { |
6298a337 | 1238 | ceph_put_snap_context(oldest); |
1d3576fd | 1239 | dout(" page %p snapc %p not current or oldest\n", |
6298a337 | 1240 | page, snapc); |
1d3576fd SW |
1241 | /* |
1242 | * queue for writeback, and wait for snapc to | |
1243 | * be writeable or written | |
1244 | */ | |
6298a337 | 1245 | snapc = ceph_get_snap_context(snapc); |
1d3576fd | 1246 | unlock_page(page); |
3c6f6b79 | 1247 | ceph_queue_writeback(inode); |
a78bbd4b | 1248 | r = wait_event_killable(ci->i_cap_wq, |
1d3576fd SW |
1249 | context_is_writeable_or_written(inode, snapc)); |
1250 | ceph_put_snap_context(snapc); | |
8f883c24 SW |
1251 | if (r == -ERESTARTSYS) |
1252 | return r; | |
4af6b225 | 1253 | return -EAGAIN; |
1d3576fd | 1254 | } |
6298a337 | 1255 | ceph_put_snap_context(oldest); |
1d3576fd SW |
1256 | |
1257 | /* yay, writeable, do it now (without dropping page lock) */ | |
1258 | dout(" page %p snapc %p not current, but oldest\n", | |
1259 | page, snapc); | |
1260 | if (!clear_page_dirty_for_io(page)) | |
1261 | goto retry_locked; | |
1262 | r = writepage_nounlock(page, NULL); | |
1263 | if (r < 0) | |
dd2bc473 | 1264 | goto fail_unlock; |
1d3576fd SW |
1265 | goto retry_locked; |
1266 | } | |
1267 | ||
1268 | if (PageUptodate(page)) { | |
1269 | dout(" page %p already uptodate\n", page); | |
1270 | return 0; | |
1271 | } | |
1272 | ||
1273 | /* full page? */ | |
09cbfeaf | 1274 | if (pos_in_page == 0 && len == PAGE_SIZE) |
1d3576fd SW |
1275 | return 0; |
1276 | ||
1277 | /* past end of file? */ | |
99c88e69 | 1278 | i_size = i_size_read(inode); |
1d3576fd | 1279 | |
1d3576fd SW |
1280 | if (page_off >= i_size || |
1281 | (pos_in_page == 0 && (pos+len) >= i_size && | |
09cbfeaf | 1282 | end_in_page - pos_in_page != PAGE_SIZE)) { |
1d3576fd | 1283 | dout(" zeroing %p 0 - %d and %d - %d\n", |
09cbfeaf | 1284 | page, pos_in_page, end_in_page, (int)PAGE_SIZE); |
1d3576fd SW |
1285 | zero_user_segments(page, |
1286 | 0, pos_in_page, | |
09cbfeaf | 1287 | end_in_page, PAGE_SIZE); |
1d3576fd SW |
1288 | return 0; |
1289 | } | |
1290 | ||
1291 | /* we need to read it. */ | |
dd2bc473 YZ |
1292 | r = ceph_do_readpage(file, page); |
1293 | if (r < 0) { | |
1294 | if (r == -EINPROGRESS) | |
1295 | return -EAGAIN; | |
1296 | goto fail_unlock; | |
1297 | } | |
1d3576fd | 1298 | goto retry_locked; |
dd2bc473 | 1299 | fail_unlock: |
1d3576fd SW |
1300 | unlock_page(page); |
1301 | return r; | |
1302 | } | |
1303 | ||
4af6b225 YS |
1304 | /* |
1305 | * We are only allowed to write into/dirty the page if the page is | |
1306 | * clean, or already dirty within the same snap context. | |
1307 | */ | |
1308 | static int ceph_write_begin(struct file *file, struct address_space *mapping, | |
1309 | loff_t pos, unsigned len, unsigned flags, | |
1310 | struct page **pagep, void **fsdata) | |
1311 | { | |
496ad9aa | 1312 | struct inode *inode = file_inode(file); |
4af6b225 | 1313 | struct page *page; |
09cbfeaf | 1314 | pgoff_t index = pos >> PAGE_SHIFT; |
7971bd92 | 1315 | int r; |
4af6b225 YS |
1316 | |
1317 | do { | |
8f883c24 | 1318 | /* get a page */ |
4af6b225 | 1319 | page = grab_cache_page_write_begin(mapping, index, 0); |
7971bd92 SW |
1320 | if (!page) |
1321 | return -ENOMEM; | |
4af6b225 YS |
1322 | |
1323 | dout("write_begin file %p inode %p page %p %d~%d\n", file, | |
213c99ee | 1324 | inode, page, (int)pos, (int)len); |
4af6b225 YS |
1325 | |
1326 | r = ceph_update_writeable_page(file, pos, len, page); | |
c1d00b2d | 1327 | if (r < 0) |
09cbfeaf | 1328 | put_page(page); |
c1d00b2d TK |
1329 | else |
1330 | *pagep = page; | |
4af6b225 YS |
1331 | } while (r == -EAGAIN); |
1332 | ||
1333 | return r; | |
1334 | } | |
1335 | ||
1d3576fd SW |
1336 | /* |
1337 | * we don't do anything in here that simple_write_end doesn't do | |
5dda377c | 1338 | * except adjust dirty page accounting |
1d3576fd SW |
1339 | */ |
1340 | static int ceph_write_end(struct file *file, struct address_space *mapping, | |
1341 | loff_t pos, unsigned len, unsigned copied, | |
1342 | struct page *page, void *fsdata) | |
1343 | { | |
496ad9aa | 1344 | struct inode *inode = file_inode(file); |
efb0ca76 | 1345 | bool check_cap = false; |
1d3576fd SW |
1346 | |
1347 | dout("write_end file %p inode %p page %p %d~%d (%d)\n", file, | |
1348 | inode, page, (int)pos, (int)copied, (int)len); | |
1349 | ||
1350 | /* zero the stale part of the page if we did a short copy */ | |
b9de313c AV |
1351 | if (!PageUptodate(page)) { |
1352 | if (copied < len) { | |
1353 | copied = 0; | |
1354 | goto out; | |
1355 | } | |
1356 | SetPageUptodate(page); | |
1357 | } | |
1d3576fd SW |
1358 | |
1359 | /* did file size increase? */ | |
99c88e69 | 1360 | if (pos+copied > i_size_read(inode)) |
1d3576fd SW |
1361 | check_cap = ceph_inode_set_size(inode, pos+copied); |
1362 | ||
1d3576fd SW |
1363 | set_page_dirty(page); |
1364 | ||
b9de313c | 1365 | out: |
1d3576fd | 1366 | unlock_page(page); |
09cbfeaf | 1367 | put_page(page); |
1d3576fd SW |
1368 | |
1369 | if (check_cap) | |
1370 | ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL); | |
1371 | ||
1372 | return copied; | |
1373 | } | |
1374 | ||
1375 | /* | |
1376 | * we set .direct_IO to indicate direct io is supported, but since we | |
1377 | * intercept O_DIRECT reads and writes early, this function should | |
1378 | * never get called. | |
1379 | */ | |
c8b8e32d | 1380 | static ssize_t ceph_direct_io(struct kiocb *iocb, struct iov_iter *iter) |
1d3576fd SW |
1381 | { |
1382 | WARN_ON(1); | |
1383 | return -EINVAL; | |
1384 | } | |
1385 | ||
1386 | const struct address_space_operations ceph_aops = { | |
1387 | .readpage = ceph_readpage, | |
1388 | .readpages = ceph_readpages, | |
1389 | .writepage = ceph_writepage, | |
1390 | .writepages = ceph_writepages_start, | |
1391 | .write_begin = ceph_write_begin, | |
1392 | .write_end = ceph_write_end, | |
1393 | .set_page_dirty = ceph_set_page_dirty, | |
1394 | .invalidatepage = ceph_invalidatepage, | |
1395 | .releasepage = ceph_releasepage, | |
1396 | .direct_IO = ceph_direct_io, | |
1397 | }; | |
1398 | ||
4f7e89f6 YZ |
1399 | static void ceph_block_sigs(sigset_t *oldset) |
1400 | { | |
1401 | sigset_t mask; | |
1402 | siginitsetinv(&mask, sigmask(SIGKILL)); | |
1403 | sigprocmask(SIG_BLOCK, &mask, oldset); | |
1404 | } | |
1405 | ||
1406 | static void ceph_restore_sigs(sigset_t *oldset) | |
1407 | { | |
1408 | sigprocmask(SIG_SETMASK, oldset, NULL); | |
1409 | } | |
1d3576fd SW |
1410 | |
1411 | /* | |
1412 | * vm ops | |
1413 | */ | |
11bac800 | 1414 | static int ceph_filemap_fault(struct vm_fault *vmf) |
61f68816 | 1415 | { |
11bac800 | 1416 | struct vm_area_struct *vma = vmf->vma; |
61f68816 YZ |
1417 | struct inode *inode = file_inode(vma->vm_file); |
1418 | struct ceph_inode_info *ci = ceph_inode(inode); | |
1419 | struct ceph_file_info *fi = vma->vm_file->private_data; | |
3738daa6 | 1420 | struct page *pinned_page = NULL; |
09cbfeaf | 1421 | loff_t off = vmf->pgoff << PAGE_SHIFT; |
61f68816 | 1422 | int want, got, ret; |
4f7e89f6 YZ |
1423 | sigset_t oldset; |
1424 | ||
1425 | ceph_block_sigs(&oldset); | |
61f68816 YZ |
1426 | |
1427 | dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n", | |
09cbfeaf | 1428 | inode, ceph_vinop(inode), off, (size_t)PAGE_SIZE); |
61f68816 YZ |
1429 | if (fi->fmode & CEPH_FILE_MODE_LAZY) |
1430 | want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO; | |
1431 | else | |
1432 | want = CEPH_CAP_FILE_CACHE; | |
4f7e89f6 YZ |
1433 | |
1434 | got = 0; | |
1435 | ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page); | |
6ce026e4 | 1436 | if (ret < 0) |
4f7e89f6 | 1437 | goto out_restore; |
6ce026e4 | 1438 | |
61f68816 | 1439 | dout("filemap_fault %p %llu~%zd got cap refs on %s\n", |
09cbfeaf | 1440 | inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got)); |
61f68816 | 1441 | |
83701246 | 1442 | if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) || |
2b1ac852 YZ |
1443 | ci->i_inline_version == CEPH_INLINE_NONE) { |
1444 | current->journal_info = vma->vm_file; | |
11bac800 | 1445 | ret = filemap_fault(vmf); |
2b1ac852 YZ |
1446 | current->journal_info = NULL; |
1447 | } else | |
83701246 | 1448 | ret = -EAGAIN; |
61f68816 YZ |
1449 | |
1450 | dout("filemap_fault %p %llu~%zd dropping cap refs on %s ret %d\n", | |
09cbfeaf | 1451 | inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got), ret); |
3738daa6 | 1452 | if (pinned_page) |
09cbfeaf | 1453 | put_page(pinned_page); |
61f68816 YZ |
1454 | ceph_put_cap_refs(ci, got); |
1455 | ||
83701246 | 1456 | if (ret != -EAGAIN) |
4f7e89f6 | 1457 | goto out_restore; |
83701246 YZ |
1458 | |
1459 | /* read inline data */ | |
09cbfeaf | 1460 | if (off >= PAGE_SIZE) { |
83701246 YZ |
1461 | /* does not support inline data > PAGE_SIZE */ |
1462 | ret = VM_FAULT_SIGBUS; | |
1463 | } else { | |
1464 | int ret1; | |
1465 | struct address_space *mapping = inode->i_mapping; | |
1466 | struct page *page = find_or_create_page(mapping, 0, | |
c62d2555 MH |
1467 | mapping_gfp_constraint(mapping, |
1468 | ~__GFP_FS)); | |
83701246 YZ |
1469 | if (!page) { |
1470 | ret = VM_FAULT_OOM; | |
4f7e89f6 | 1471 | goto out_inline; |
83701246 YZ |
1472 | } |
1473 | ret1 = __ceph_do_getattr(inode, page, | |
1474 | CEPH_STAT_CAP_INLINE_DATA, true); | |
1475 | if (ret1 < 0 || off >= i_size_read(inode)) { | |
1476 | unlock_page(page); | |
09cbfeaf | 1477 | put_page(page); |
6ce026e4 YZ |
1478 | if (ret1 < 0) |
1479 | ret = ret1; | |
1480 | else | |
1481 | ret = VM_FAULT_SIGBUS; | |
4f7e89f6 | 1482 | goto out_inline; |
83701246 | 1483 | } |
09cbfeaf KS |
1484 | if (ret1 < PAGE_SIZE) |
1485 | zero_user_segment(page, ret1, PAGE_SIZE); | |
83701246 YZ |
1486 | else |
1487 | flush_dcache_page(page); | |
1488 | SetPageUptodate(page); | |
1489 | vmf->page = page; | |
1490 | ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED; | |
4f7e89f6 YZ |
1491 | out_inline: |
1492 | dout("filemap_fault %p %llu~%zd read inline data ret %d\n", | |
1493 | inode, off, (size_t)PAGE_SIZE, ret); | |
83701246 | 1494 | } |
4f7e89f6 YZ |
1495 | out_restore: |
1496 | ceph_restore_sigs(&oldset); | |
6ce026e4 YZ |
1497 | if (ret < 0) |
1498 | ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS; | |
1499 | ||
61f68816 YZ |
1500 | return ret; |
1501 | } | |
1d3576fd SW |
1502 | |
1503 | /* | |
1504 | * Reuse write_begin here for simplicity. | |
1505 | */ | |
11bac800 | 1506 | static int ceph_page_mkwrite(struct vm_fault *vmf) |
1d3576fd | 1507 | { |
11bac800 | 1508 | struct vm_area_struct *vma = vmf->vma; |
496ad9aa | 1509 | struct inode *inode = file_inode(vma->vm_file); |
61f68816 YZ |
1510 | struct ceph_inode_info *ci = ceph_inode(inode); |
1511 | struct ceph_file_info *fi = vma->vm_file->private_data; | |
f66fd9f0 | 1512 | struct ceph_cap_flush *prealloc_cf; |
61f68816 | 1513 | struct page *page = vmf->page; |
6285bc23 | 1514 | loff_t off = page_offset(page); |
61f68816 YZ |
1515 | loff_t size = i_size_read(inode); |
1516 | size_t len; | |
1517 | int want, got, ret; | |
4f7e89f6 | 1518 | sigset_t oldset; |
3ca9c3bd | 1519 | |
f66fd9f0 YZ |
1520 | prealloc_cf = ceph_alloc_cap_flush(); |
1521 | if (!prealloc_cf) | |
6ce026e4 | 1522 | return VM_FAULT_OOM; |
f66fd9f0 | 1523 | |
4f7e89f6 | 1524 | ceph_block_sigs(&oldset); |
f66fd9f0 | 1525 | |
28127bdd YZ |
1526 | if (ci->i_inline_version != CEPH_INLINE_NONE) { |
1527 | struct page *locked_page = NULL; | |
1528 | if (off == 0) { | |
1529 | lock_page(page); | |
1530 | locked_page = page; | |
1531 | } | |
1532 | ret = ceph_uninline_data(vma->vm_file, locked_page); | |
1533 | if (locked_page) | |
1534 | unlock_page(locked_page); | |
6ce026e4 | 1535 | if (ret < 0) |
f66fd9f0 | 1536 | goto out_free; |
28127bdd YZ |
1537 | } |
1538 | ||
09cbfeaf KS |
1539 | if (off + PAGE_SIZE <= size) |
1540 | len = PAGE_SIZE; | |
1d3576fd | 1541 | else |
09cbfeaf | 1542 | len = size & ~PAGE_MASK; |
1d3576fd | 1543 | |
61f68816 YZ |
1544 | dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n", |
1545 | inode, ceph_vinop(inode), off, len, size); | |
1546 | if (fi->fmode & CEPH_FILE_MODE_LAZY) | |
1547 | want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO; | |
1548 | else | |
1549 | want = CEPH_CAP_FILE_BUFFER; | |
4f7e89f6 YZ |
1550 | |
1551 | got = 0; | |
1552 | ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, off + len, | |
1553 | &got, NULL); | |
6ce026e4 | 1554 | if (ret < 0) |
4f7e89f6 | 1555 | goto out_free; |
6ce026e4 | 1556 | |
61f68816 YZ |
1557 | dout("page_mkwrite %p %llu~%zd got cap refs on %s\n", |
1558 | inode, off, len, ceph_cap_string(got)); | |
1559 | ||
1560 | /* Update time before taking page lock */ | |
1561 | file_update_time(vma->vm_file); | |
4af6b225 | 1562 | |
f0b33df5 YZ |
1563 | do { |
1564 | lock_page(page); | |
4af6b225 | 1565 | |
f0b33df5 YZ |
1566 | if ((off > size) || (page->mapping != inode->i_mapping)) { |
1567 | unlock_page(page); | |
1568 | ret = VM_FAULT_NOPAGE; | |
1569 | break; | |
1570 | } | |
1571 | ||
1572 | ret = ceph_update_writeable_page(vma->vm_file, off, len, page); | |
1573 | if (ret >= 0) { | |
1574 | /* success. we'll keep the page locked. */ | |
1575 | set_page_dirty(page); | |
1576 | ret = VM_FAULT_LOCKED; | |
1577 | } | |
1578 | } while (ret == -EAGAIN); | |
4af6b225 | 1579 | |
28127bdd YZ |
1580 | if (ret == VM_FAULT_LOCKED || |
1581 | ci->i_inline_version != CEPH_INLINE_NONE) { | |
61f68816 YZ |
1582 | int dirty; |
1583 | spin_lock(&ci->i_ceph_lock); | |
28127bdd | 1584 | ci->i_inline_version = CEPH_INLINE_NONE; |
f66fd9f0 YZ |
1585 | dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR, |
1586 | &prealloc_cf); | |
61f68816 YZ |
1587 | spin_unlock(&ci->i_ceph_lock); |
1588 | if (dirty) | |
1589 | __mark_inode_dirty(inode, dirty); | |
1590 | } | |
1591 | ||
1592 | dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %d\n", | |
1593 | inode, off, len, ceph_cap_string(got), ret); | |
1594 | ceph_put_cap_refs(ci, got); | |
f66fd9f0 | 1595 | out_free: |
4f7e89f6 | 1596 | ceph_restore_sigs(&oldset); |
f66fd9f0 | 1597 | ceph_free_cap_flush(prealloc_cf); |
6ce026e4 YZ |
1598 | if (ret < 0) |
1599 | ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS; | |
1d3576fd SW |
1600 | return ret; |
1601 | } | |
1602 | ||
31c542a1 YZ |
1603 | void ceph_fill_inline_data(struct inode *inode, struct page *locked_page, |
1604 | char *data, size_t len) | |
1605 | { | |
1606 | struct address_space *mapping = inode->i_mapping; | |
1607 | struct page *page; | |
1608 | ||
1609 | if (locked_page) { | |
1610 | page = locked_page; | |
1611 | } else { | |
1612 | if (i_size_read(inode) == 0) | |
1613 | return; | |
1614 | page = find_or_create_page(mapping, 0, | |
c62d2555 MH |
1615 | mapping_gfp_constraint(mapping, |
1616 | ~__GFP_FS)); | |
31c542a1 YZ |
1617 | if (!page) |
1618 | return; | |
1619 | if (PageUptodate(page)) { | |
1620 | unlock_page(page); | |
09cbfeaf | 1621 | put_page(page); |
31c542a1 YZ |
1622 | return; |
1623 | } | |
1624 | } | |
1625 | ||
0668ff52 | 1626 | dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n", |
31c542a1 YZ |
1627 | inode, ceph_vinop(inode), len, locked_page); |
1628 | ||
1629 | if (len > 0) { | |
1630 | void *kaddr = kmap_atomic(page); | |
1631 | memcpy(kaddr, data, len); | |
1632 | kunmap_atomic(kaddr); | |
1633 | } | |
1634 | ||
1635 | if (page != locked_page) { | |
09cbfeaf KS |
1636 | if (len < PAGE_SIZE) |
1637 | zero_user_segment(page, len, PAGE_SIZE); | |
31c542a1 YZ |
1638 | else |
1639 | flush_dcache_page(page); | |
1640 | ||
1641 | SetPageUptodate(page); | |
1642 | unlock_page(page); | |
09cbfeaf | 1643 | put_page(page); |
31c542a1 YZ |
1644 | } |
1645 | } | |
1646 | ||
28127bdd YZ |
1647 | int ceph_uninline_data(struct file *filp, struct page *locked_page) |
1648 | { | |
1649 | struct inode *inode = file_inode(filp); | |
1650 | struct ceph_inode_info *ci = ceph_inode(inode); | |
1651 | struct ceph_fs_client *fsc = ceph_inode_to_client(inode); | |
1652 | struct ceph_osd_request *req; | |
1653 | struct page *page = NULL; | |
1654 | u64 len, inline_version; | |
1655 | int err = 0; | |
1656 | bool from_pagecache = false; | |
1657 | ||
1658 | spin_lock(&ci->i_ceph_lock); | |
1659 | inline_version = ci->i_inline_version; | |
1660 | spin_unlock(&ci->i_ceph_lock); | |
1661 | ||
1662 | dout("uninline_data %p %llx.%llx inline_version %llu\n", | |
1663 | inode, ceph_vinop(inode), inline_version); | |
1664 | ||
1665 | if (inline_version == 1 || /* initial version, no data */ | |
1666 | inline_version == CEPH_INLINE_NONE) | |
1667 | goto out; | |
1668 | ||
1669 | if (locked_page) { | |
1670 | page = locked_page; | |
1671 | WARN_ON(!PageUptodate(page)); | |
1672 | } else if (ceph_caps_issued(ci) & | |
1673 | (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) { | |
1674 | page = find_get_page(inode->i_mapping, 0); | |
1675 | if (page) { | |
1676 | if (PageUptodate(page)) { | |
1677 | from_pagecache = true; | |
1678 | lock_page(page); | |
1679 | } else { | |
09cbfeaf | 1680 | put_page(page); |
28127bdd YZ |
1681 | page = NULL; |
1682 | } | |
1683 | } | |
1684 | } | |
1685 | ||
1686 | if (page) { | |
1687 | len = i_size_read(inode); | |
09cbfeaf KS |
1688 | if (len > PAGE_SIZE) |
1689 | len = PAGE_SIZE; | |
28127bdd YZ |
1690 | } else { |
1691 | page = __page_cache_alloc(GFP_NOFS); | |
1692 | if (!page) { | |
1693 | err = -ENOMEM; | |
1694 | goto out; | |
1695 | } | |
1696 | err = __ceph_do_getattr(inode, page, | |
1697 | CEPH_STAT_CAP_INLINE_DATA, true); | |
1698 | if (err < 0) { | |
1699 | /* no inline data */ | |
1700 | if (err == -ENODATA) | |
1701 | err = 0; | |
1702 | goto out; | |
1703 | } | |
1704 | len = err; | |
1705 | } | |
1706 | ||
1707 | req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, | |
1708 | ceph_vino(inode), 0, &len, 0, 1, | |
54ea0046 | 1709 | CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE, |
34b759b4 | 1710 | NULL, 0, 0, false); |
28127bdd YZ |
1711 | if (IS_ERR(req)) { |
1712 | err = PTR_ERR(req); | |
1713 | goto out; | |
1714 | } | |
1715 | ||
bb873b53 | 1716 | req->r_mtime = inode->i_mtime; |
28127bdd YZ |
1717 | err = ceph_osdc_start_request(&fsc->client->osdc, req, false); |
1718 | if (!err) | |
1719 | err = ceph_osdc_wait_request(&fsc->client->osdc, req); | |
1720 | ceph_osdc_put_request(req); | |
1721 | if (err < 0) | |
1722 | goto out; | |
1723 | ||
1724 | req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, | |
1725 | ceph_vino(inode), 0, &len, 1, 3, | |
54ea0046 | 1726 | CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE, |
34b759b4 ID |
1727 | NULL, ci->i_truncate_seq, |
1728 | ci->i_truncate_size, false); | |
28127bdd YZ |
1729 | if (IS_ERR(req)) { |
1730 | err = PTR_ERR(req); | |
1731 | goto out; | |
1732 | } | |
1733 | ||
1734 | osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false); | |
1735 | ||
ec137c10 YZ |
1736 | { |
1737 | __le64 xattr_buf = cpu_to_le64(inline_version); | |
1738 | err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR, | |
1739 | "inline_version", &xattr_buf, | |
1740 | sizeof(xattr_buf), | |
1741 | CEPH_OSD_CMPXATTR_OP_GT, | |
1742 | CEPH_OSD_CMPXATTR_MODE_U64); | |
1743 | if (err) | |
1744 | goto out_put; | |
1745 | } | |
1746 | ||
1747 | { | |
1748 | char xattr_buf[32]; | |
1749 | int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf), | |
1750 | "%llu", inline_version); | |
1751 | err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR, | |
1752 | "inline_version", | |
1753 | xattr_buf, xattr_len, 0, 0); | |
1754 | if (err) | |
1755 | goto out_put; | |
1756 | } | |
28127bdd | 1757 | |
bb873b53 | 1758 | req->r_mtime = inode->i_mtime; |
28127bdd YZ |
1759 | err = ceph_osdc_start_request(&fsc->client->osdc, req, false); |
1760 | if (!err) | |
1761 | err = ceph_osdc_wait_request(&fsc->client->osdc, req); | |
1762 | out_put: | |
1763 | ceph_osdc_put_request(req); | |
1764 | if (err == -ECANCELED) | |
1765 | err = 0; | |
1766 | out: | |
1767 | if (page && page != locked_page) { | |
1768 | if (from_pagecache) { | |
1769 | unlock_page(page); | |
09cbfeaf | 1770 | put_page(page); |
28127bdd YZ |
1771 | } else |
1772 | __free_pages(page, 0); | |
1773 | } | |
1774 | ||
1775 | dout("uninline_data %p %llx.%llx inline_version %llu = %d\n", | |
1776 | inode, ceph_vinop(inode), inline_version, err); | |
1777 | return err; | |
1778 | } | |
1779 | ||
7cbea8dc | 1780 | static const struct vm_operations_struct ceph_vmops = { |
61f68816 | 1781 | .fault = ceph_filemap_fault, |
1d3576fd SW |
1782 | .page_mkwrite = ceph_page_mkwrite, |
1783 | }; | |
1784 | ||
1785 | int ceph_mmap(struct file *file, struct vm_area_struct *vma) | |
1786 | { | |
1787 | struct address_space *mapping = file->f_mapping; | |
1788 | ||
1789 | if (!mapping->a_ops->readpage) | |
1790 | return -ENOEXEC; | |
1791 | file_accessed(file); | |
1792 | vma->vm_ops = &ceph_vmops; | |
1d3576fd SW |
1793 | return 0; |
1794 | } | |
10183a69 YZ |
1795 | |
1796 | enum { | |
1797 | POOL_READ = 1, | |
1798 | POOL_WRITE = 2, | |
1799 | }; | |
1800 | ||
779fe0fb YZ |
1801 | static int __ceph_pool_perm_get(struct ceph_inode_info *ci, |
1802 | s64 pool, struct ceph_string *pool_ns) | |
10183a69 YZ |
1803 | { |
1804 | struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode); | |
1805 | struct ceph_mds_client *mdsc = fsc->mdsc; | |
1806 | struct ceph_osd_request *rd_req = NULL, *wr_req = NULL; | |
1807 | struct rb_node **p, *parent; | |
1808 | struct ceph_pool_perm *perm; | |
1809 | struct page **pages; | |
779fe0fb | 1810 | size_t pool_ns_len; |
10183a69 YZ |
1811 | int err = 0, err2 = 0, have = 0; |
1812 | ||
1813 | down_read(&mdsc->pool_perm_rwsem); | |
1814 | p = &mdsc->pool_perm_tree.rb_node; | |
1815 | while (*p) { | |
1816 | perm = rb_entry(*p, struct ceph_pool_perm, node); | |
1817 | if (pool < perm->pool) | |
1818 | p = &(*p)->rb_left; | |
1819 | else if (pool > perm->pool) | |
1820 | p = &(*p)->rb_right; | |
1821 | else { | |
779fe0fb YZ |
1822 | int ret = ceph_compare_string(pool_ns, |
1823 | perm->pool_ns, | |
1824 | perm->pool_ns_len); | |
1825 | if (ret < 0) | |
1826 | p = &(*p)->rb_left; | |
1827 | else if (ret > 0) | |
1828 | p = &(*p)->rb_right; | |
1829 | else { | |
1830 | have = perm->perm; | |
1831 | break; | |
1832 | } | |
10183a69 YZ |
1833 | } |
1834 | } | |
1835 | up_read(&mdsc->pool_perm_rwsem); | |
1836 | if (*p) | |
1837 | goto out; | |
1838 | ||
779fe0fb YZ |
1839 | if (pool_ns) |
1840 | dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n", | |
1841 | pool, (int)pool_ns->len, pool_ns->str); | |
1842 | else | |
1843 | dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool); | |
10183a69 YZ |
1844 | |
1845 | down_write(&mdsc->pool_perm_rwsem); | |
779fe0fb | 1846 | p = &mdsc->pool_perm_tree.rb_node; |
10183a69 YZ |
1847 | parent = NULL; |
1848 | while (*p) { | |
1849 | parent = *p; | |
1850 | perm = rb_entry(parent, struct ceph_pool_perm, node); | |
1851 | if (pool < perm->pool) | |
1852 | p = &(*p)->rb_left; | |
1853 | else if (pool > perm->pool) | |
1854 | p = &(*p)->rb_right; | |
1855 | else { | |
779fe0fb YZ |
1856 | int ret = ceph_compare_string(pool_ns, |
1857 | perm->pool_ns, | |
1858 | perm->pool_ns_len); | |
1859 | if (ret < 0) | |
1860 | p = &(*p)->rb_left; | |
1861 | else if (ret > 0) | |
1862 | p = &(*p)->rb_right; | |
1863 | else { | |
1864 | have = perm->perm; | |
1865 | break; | |
1866 | } | |
10183a69 YZ |
1867 | } |
1868 | } | |
1869 | if (*p) { | |
1870 | up_write(&mdsc->pool_perm_rwsem); | |
1871 | goto out; | |
1872 | } | |
1873 | ||
34b759b4 | 1874 | rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL, |
10183a69 YZ |
1875 | 1, false, GFP_NOFS); |
1876 | if (!rd_req) { | |
1877 | err = -ENOMEM; | |
1878 | goto out_unlock; | |
1879 | } | |
1880 | ||
1881 | rd_req->r_flags = CEPH_OSD_FLAG_READ; | |
1882 | osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0); | |
1883 | rd_req->r_base_oloc.pool = pool; | |
779fe0fb YZ |
1884 | if (pool_ns) |
1885 | rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns); | |
d30291b9 | 1886 | ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino); |
10183a69 | 1887 | |
13d1ad16 ID |
1888 | err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS); |
1889 | if (err) | |
1890 | goto out_unlock; | |
10183a69 | 1891 | |
34b759b4 | 1892 | wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL, |
10183a69 YZ |
1893 | 1, false, GFP_NOFS); |
1894 | if (!wr_req) { | |
1895 | err = -ENOMEM; | |
1896 | goto out_unlock; | |
1897 | } | |
1898 | ||
54ea0046 | 1899 | wr_req->r_flags = CEPH_OSD_FLAG_WRITE; |
10183a69 | 1900 | osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL); |
63244fa1 | 1901 | ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc); |
d30291b9 | 1902 | ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid); |
10183a69 | 1903 | |
13d1ad16 ID |
1904 | err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS); |
1905 | if (err) | |
1906 | goto out_unlock; | |
10183a69 YZ |
1907 | |
1908 | /* one page should be large enough for STAT data */ | |
1909 | pages = ceph_alloc_page_vector(1, GFP_KERNEL); | |
1910 | if (IS_ERR(pages)) { | |
1911 | err = PTR_ERR(pages); | |
1912 | goto out_unlock; | |
1913 | } | |
1914 | ||
1915 | osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE, | |
1916 | 0, false, true); | |
10183a69 YZ |
1917 | err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false); |
1918 | ||
bb873b53 | 1919 | wr_req->r_mtime = ci->vfs_inode.i_mtime; |
a1f4020a | 1920 | wr_req->r_abort_on_full = true; |
10183a69 YZ |
1921 | err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false); |
1922 | ||
1923 | if (!err) | |
1924 | err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req); | |
1925 | if (!err2) | |
1926 | err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req); | |
1927 | ||
1928 | if (err >= 0 || err == -ENOENT) | |
1929 | have |= POOL_READ; | |
1930 | else if (err != -EPERM) | |
1931 | goto out_unlock; | |
1932 | ||
1933 | if (err2 == 0 || err2 == -EEXIST) | |
1934 | have |= POOL_WRITE; | |
1935 | else if (err2 != -EPERM) { | |
1936 | err = err2; | |
1937 | goto out_unlock; | |
1938 | } | |
1939 | ||
779fe0fb YZ |
1940 | pool_ns_len = pool_ns ? pool_ns->len : 0; |
1941 | perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS); | |
10183a69 YZ |
1942 | if (!perm) { |
1943 | err = -ENOMEM; | |
1944 | goto out_unlock; | |
1945 | } | |
1946 | ||
1947 | perm->pool = pool; | |
1948 | perm->perm = have; | |
779fe0fb YZ |
1949 | perm->pool_ns_len = pool_ns_len; |
1950 | if (pool_ns_len > 0) | |
1951 | memcpy(perm->pool_ns, pool_ns->str, pool_ns_len); | |
1952 | perm->pool_ns[pool_ns_len] = 0; | |
1953 | ||
10183a69 YZ |
1954 | rb_link_node(&perm->node, parent, p); |
1955 | rb_insert_color(&perm->node, &mdsc->pool_perm_tree); | |
1956 | err = 0; | |
1957 | out_unlock: | |
1958 | up_write(&mdsc->pool_perm_rwsem); | |
1959 | ||
3ed97d63 ID |
1960 | ceph_osdc_put_request(rd_req); |
1961 | ceph_osdc_put_request(wr_req); | |
10183a69 YZ |
1962 | out: |
1963 | if (!err) | |
1964 | err = have; | |
779fe0fb YZ |
1965 | if (pool_ns) |
1966 | dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n", | |
1967 | pool, (int)pool_ns->len, pool_ns->str, err); | |
1968 | else | |
1969 | dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err); | |
10183a69 YZ |
1970 | return err; |
1971 | } | |
1972 | ||
1973 | int ceph_pool_perm_check(struct ceph_inode_info *ci, int need) | |
1974 | { | |
7627151e | 1975 | s64 pool; |
779fe0fb | 1976 | struct ceph_string *pool_ns; |
10183a69 YZ |
1977 | int ret, flags; |
1978 | ||
80e80fbb YZ |
1979 | if (ci->i_vino.snap != CEPH_NOSNAP) { |
1980 | /* | |
1981 | * Pool permission check needs to write to the first object. | |
1982 | * But for snapshot, head of the first object may have alread | |
1983 | * been deleted. Skip check to avoid creating orphan object. | |
1984 | */ | |
1985 | return 0; | |
1986 | } | |
1987 | ||
10183a69 YZ |
1988 | if (ceph_test_mount_opt(ceph_inode_to_client(&ci->vfs_inode), |
1989 | NOPOOLPERM)) | |
1990 | return 0; | |
1991 | ||
1992 | spin_lock(&ci->i_ceph_lock); | |
1993 | flags = ci->i_ceph_flags; | |
7627151e | 1994 | pool = ci->i_layout.pool_id; |
10183a69 YZ |
1995 | spin_unlock(&ci->i_ceph_lock); |
1996 | check: | |
1997 | if (flags & CEPH_I_POOL_PERM) { | |
1998 | if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) { | |
7627151e | 1999 | dout("ceph_pool_perm_check pool %lld no read perm\n", |
10183a69 YZ |
2000 | pool); |
2001 | return -EPERM; | |
2002 | } | |
2003 | if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) { | |
7627151e | 2004 | dout("ceph_pool_perm_check pool %lld no write perm\n", |
10183a69 YZ |
2005 | pool); |
2006 | return -EPERM; | |
2007 | } | |
2008 | return 0; | |
2009 | } | |
2010 | ||
779fe0fb YZ |
2011 | pool_ns = ceph_try_get_string(ci->i_layout.pool_ns); |
2012 | ret = __ceph_pool_perm_get(ci, pool, pool_ns); | |
2013 | ceph_put_string(pool_ns); | |
10183a69 YZ |
2014 | if (ret < 0) |
2015 | return ret; | |
2016 | ||
2017 | flags = CEPH_I_POOL_PERM; | |
2018 | if (ret & POOL_READ) | |
2019 | flags |= CEPH_I_POOL_RD; | |
2020 | if (ret & POOL_WRITE) | |
2021 | flags |= CEPH_I_POOL_WR; | |
2022 | ||
2023 | spin_lock(&ci->i_ceph_lock); | |
779fe0fb YZ |
2024 | if (pool == ci->i_layout.pool_id && |
2025 | pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) { | |
2026 | ci->i_ceph_flags |= flags; | |
10183a69 | 2027 | } else { |
7627151e | 2028 | pool = ci->i_layout.pool_id; |
10183a69 YZ |
2029 | flags = ci->i_ceph_flags; |
2030 | } | |
2031 | spin_unlock(&ci->i_ceph_lock); | |
2032 | goto check; | |
2033 | } | |
2034 | ||
2035 | void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc) | |
2036 | { | |
2037 | struct ceph_pool_perm *perm; | |
2038 | struct rb_node *n; | |
2039 | ||
2040 | while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) { | |
2041 | n = rb_first(&mdsc->pool_perm_tree); | |
2042 | perm = rb_entry(n, struct ceph_pool_perm, node); | |
2043 | rb_erase(n, &mdsc->pool_perm_tree); | |
2044 | kfree(perm); | |
2045 | } | |
2046 | } |