| 1 | #include <linux/ceph/ceph_debug.h> |
| 2 | |
| 3 | #include <linux/module.h> |
| 4 | #include <linux/sched.h> |
| 5 | #include <linux/slab.h> |
| 6 | #include <linux/file.h> |
| 7 | #include <linux/mount.h> |
| 8 | #include <linux/namei.h> |
| 9 | #include <linux/writeback.h> |
| 10 | #include <linux/aio.h> |
| 11 | |
| 12 | #include "super.h" |
| 13 | #include "mds_client.h" |
| 14 | |
| 15 | /* |
| 16 | * Ceph file operations |
| 17 | * |
| 18 | * Implement basic open/close functionality, and implement |
| 19 | * read/write. |
| 20 | * |
| 21 | * We implement three modes of file I/O: |
| 22 | * - buffered uses the generic_file_aio_{read,write} helpers |
| 23 | * |
| 24 | * - synchronous is used when there is multi-client read/write |
| 25 | * sharing, avoids the page cache, and synchronously waits for an |
| 26 | * ack from the OSD. |
| 27 | * |
| 28 | * - direct io takes the variant of the sync path that references |
| 29 | * user pages directly. |
| 30 | * |
| 31 | * fsync() flushes and waits on dirty pages, but just queues metadata |
| 32 | * for writeback: since the MDS can recover size and mtime there is no |
| 33 | * need to wait for MDS acknowledgement. |
| 34 | */ |
| 35 | |
| 36 | |
| 37 | /* |
| 38 | * Prepare an open request. Preallocate ceph_cap to avoid an |
| 39 | * inopportune ENOMEM later. |
| 40 | */ |
| 41 | static struct ceph_mds_request * |
| 42 | prepare_open_request(struct super_block *sb, int flags, int create_mode) |
| 43 | { |
| 44 | struct ceph_fs_client *fsc = ceph_sb_to_client(sb); |
| 45 | struct ceph_mds_client *mdsc = fsc->mdsc; |
| 46 | struct ceph_mds_request *req; |
| 47 | int want_auth = USE_ANY_MDS; |
| 48 | int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN; |
| 49 | |
| 50 | if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC)) |
| 51 | want_auth = USE_AUTH_MDS; |
| 52 | |
| 53 | req = ceph_mdsc_create_request(mdsc, op, want_auth); |
| 54 | if (IS_ERR(req)) |
| 55 | goto out; |
| 56 | req->r_fmode = ceph_flags_to_mode(flags); |
| 57 | req->r_args.open.flags = cpu_to_le32(flags); |
| 58 | req->r_args.open.mode = cpu_to_le32(create_mode); |
| 59 | out: |
| 60 | return req; |
| 61 | } |
| 62 | |
| 63 | /* |
| 64 | * initialize private struct file data. |
| 65 | * if we fail, clean up by dropping fmode reference on the ceph_inode |
| 66 | */ |
| 67 | static int ceph_init_file(struct inode *inode, struct file *file, int fmode) |
| 68 | { |
| 69 | struct ceph_file_info *cf; |
| 70 | int ret = 0; |
| 71 | |
| 72 | switch (inode->i_mode & S_IFMT) { |
| 73 | case S_IFREG: |
| 74 | case S_IFDIR: |
| 75 | dout("init_file %p %p 0%o (regular)\n", inode, file, |
| 76 | inode->i_mode); |
| 77 | cf = kmem_cache_alloc(ceph_file_cachep, GFP_NOFS | __GFP_ZERO); |
| 78 | if (cf == NULL) { |
| 79 | ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */ |
| 80 | return -ENOMEM; |
| 81 | } |
| 82 | cf->fmode = fmode; |
| 83 | cf->next_offset = 2; |
| 84 | file->private_data = cf; |
| 85 | BUG_ON(inode->i_fop->release != ceph_release); |
| 86 | break; |
| 87 | |
| 88 | case S_IFLNK: |
| 89 | dout("init_file %p %p 0%o (symlink)\n", inode, file, |
| 90 | inode->i_mode); |
| 91 | ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */ |
| 92 | break; |
| 93 | |
| 94 | default: |
| 95 | dout("init_file %p %p 0%o (special)\n", inode, file, |
| 96 | inode->i_mode); |
| 97 | /* |
| 98 | * we need to drop the open ref now, since we don't |
| 99 | * have .release set to ceph_release. |
| 100 | */ |
| 101 | ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */ |
| 102 | BUG_ON(inode->i_fop->release == ceph_release); |
| 103 | |
| 104 | /* call the proper open fop */ |
| 105 | ret = inode->i_fop->open(inode, file); |
| 106 | } |
| 107 | return ret; |
| 108 | } |
| 109 | |
| 110 | /* |
| 111 | * If we already have the requisite capabilities, we can satisfy |
| 112 | * the open request locally (no need to request new caps from the |
| 113 | * MDS). We do, however, need to inform the MDS (asynchronously) |
| 114 | * if our wanted caps set expands. |
| 115 | */ |
| 116 | int ceph_open(struct inode *inode, struct file *file) |
| 117 | { |
| 118 | struct ceph_inode_info *ci = ceph_inode(inode); |
| 119 | struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb); |
| 120 | struct ceph_mds_client *mdsc = fsc->mdsc; |
| 121 | struct ceph_mds_request *req; |
| 122 | struct ceph_file_info *cf = file->private_data; |
| 123 | struct inode *parent_inode = NULL; |
| 124 | int err; |
| 125 | int flags, fmode, wanted; |
| 126 | |
| 127 | if (cf) { |
| 128 | dout("open file %p is already opened\n", file); |
| 129 | return 0; |
| 130 | } |
| 131 | |
| 132 | /* filter out O_CREAT|O_EXCL; vfs did that already. yuck. */ |
| 133 | flags = file->f_flags & ~(O_CREAT|O_EXCL); |
| 134 | if (S_ISDIR(inode->i_mode)) |
| 135 | flags = O_DIRECTORY; /* mds likes to know */ |
| 136 | |
| 137 | dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode, |
| 138 | ceph_vinop(inode), file, flags, file->f_flags); |
| 139 | fmode = ceph_flags_to_mode(flags); |
| 140 | wanted = ceph_caps_for_mode(fmode); |
| 141 | |
| 142 | /* snapped files are read-only */ |
| 143 | if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE)) |
| 144 | return -EROFS; |
| 145 | |
| 146 | /* trivially open snapdir */ |
| 147 | if (ceph_snap(inode) == CEPH_SNAPDIR) { |
| 148 | spin_lock(&ci->i_ceph_lock); |
| 149 | __ceph_get_fmode(ci, fmode); |
| 150 | spin_unlock(&ci->i_ceph_lock); |
| 151 | return ceph_init_file(inode, file, fmode); |
| 152 | } |
| 153 | |
| 154 | /* |
| 155 | * No need to block if we have caps on the auth MDS (for |
| 156 | * write) or any MDS (for read). Update wanted set |
| 157 | * asynchronously. |
| 158 | */ |
| 159 | spin_lock(&ci->i_ceph_lock); |
| 160 | if (__ceph_is_any_real_caps(ci) && |
| 161 | (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) { |
| 162 | int mds_wanted = __ceph_caps_mds_wanted(ci); |
| 163 | int issued = __ceph_caps_issued(ci, NULL); |
| 164 | |
| 165 | dout("open %p fmode %d want %s issued %s using existing\n", |
| 166 | inode, fmode, ceph_cap_string(wanted), |
| 167 | ceph_cap_string(issued)); |
| 168 | __ceph_get_fmode(ci, fmode); |
| 169 | spin_unlock(&ci->i_ceph_lock); |
| 170 | |
| 171 | /* adjust wanted? */ |
| 172 | if ((issued & wanted) != wanted && |
| 173 | (mds_wanted & wanted) != wanted && |
| 174 | ceph_snap(inode) != CEPH_SNAPDIR) |
| 175 | ceph_check_caps(ci, 0, NULL); |
| 176 | |
| 177 | return ceph_init_file(inode, file, fmode); |
| 178 | } else if (ceph_snap(inode) != CEPH_NOSNAP && |
| 179 | (ci->i_snap_caps & wanted) == wanted) { |
| 180 | __ceph_get_fmode(ci, fmode); |
| 181 | spin_unlock(&ci->i_ceph_lock); |
| 182 | return ceph_init_file(inode, file, fmode); |
| 183 | } |
| 184 | spin_unlock(&ci->i_ceph_lock); |
| 185 | |
| 186 | dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted)); |
| 187 | req = prepare_open_request(inode->i_sb, flags, 0); |
| 188 | if (IS_ERR(req)) { |
| 189 | err = PTR_ERR(req); |
| 190 | goto out; |
| 191 | } |
| 192 | req->r_inode = inode; |
| 193 | ihold(inode); |
| 194 | req->r_num_caps = 1; |
| 195 | if (flags & (O_CREAT|O_TRUNC)) |
| 196 | parent_inode = ceph_get_dentry_parent_inode(file->f_dentry); |
| 197 | err = ceph_mdsc_do_request(mdsc, parent_inode, req); |
| 198 | iput(parent_inode); |
| 199 | if (!err) |
| 200 | err = ceph_init_file(inode, file, req->r_fmode); |
| 201 | ceph_mdsc_put_request(req); |
| 202 | dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode)); |
| 203 | out: |
| 204 | return err; |
| 205 | } |
| 206 | |
| 207 | |
| 208 | /* |
| 209 | * Do a lookup + open with a single request. If we get a non-existent |
| 210 | * file or symlink, return 1 so the VFS can retry. |
| 211 | */ |
| 212 | int ceph_atomic_open(struct inode *dir, struct dentry *dentry, |
| 213 | struct file *file, unsigned flags, umode_t mode, |
| 214 | int *opened) |
| 215 | { |
| 216 | struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb); |
| 217 | struct ceph_mds_client *mdsc = fsc->mdsc; |
| 218 | struct ceph_mds_request *req; |
| 219 | struct dentry *dn; |
| 220 | int err; |
| 221 | |
| 222 | dout("atomic_open %p dentry %p '%.*s' %s flags %d mode 0%o\n", |
| 223 | dir, dentry, dentry->d_name.len, dentry->d_name.name, |
| 224 | d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode); |
| 225 | |
| 226 | if (dentry->d_name.len > NAME_MAX) |
| 227 | return -ENAMETOOLONG; |
| 228 | |
| 229 | err = ceph_init_dentry(dentry); |
| 230 | if (err < 0) |
| 231 | return err; |
| 232 | |
| 233 | /* do the open */ |
| 234 | req = prepare_open_request(dir->i_sb, flags, mode); |
| 235 | if (IS_ERR(req)) |
| 236 | return PTR_ERR(req); |
| 237 | req->r_dentry = dget(dentry); |
| 238 | req->r_num_caps = 2; |
| 239 | if (flags & O_CREAT) { |
| 240 | req->r_dentry_drop = CEPH_CAP_FILE_SHARED; |
| 241 | req->r_dentry_unless = CEPH_CAP_FILE_EXCL; |
| 242 | } |
| 243 | req->r_locked_dir = dir; /* caller holds dir->i_mutex */ |
| 244 | err = ceph_mdsc_do_request(mdsc, |
| 245 | (flags & (O_CREAT|O_TRUNC)) ? dir : NULL, |
| 246 | req); |
| 247 | if (err) |
| 248 | goto out_err; |
| 249 | |
| 250 | err = ceph_handle_snapdir(req, dentry, err); |
| 251 | if (err == 0 && (flags & O_CREAT) && !req->r_reply_info.head->is_dentry) |
| 252 | err = ceph_handle_notrace_create(dir, dentry); |
| 253 | |
| 254 | if (d_unhashed(dentry)) { |
| 255 | dn = ceph_finish_lookup(req, dentry, err); |
| 256 | if (IS_ERR(dn)) |
| 257 | err = PTR_ERR(dn); |
| 258 | } else { |
| 259 | /* we were given a hashed negative dentry */ |
| 260 | dn = NULL; |
| 261 | } |
| 262 | if (err) |
| 263 | goto out_err; |
| 264 | if (dn || dentry->d_inode == NULL || S_ISLNK(dentry->d_inode->i_mode)) { |
| 265 | /* make vfs retry on splice, ENOENT, or symlink */ |
| 266 | dout("atomic_open finish_no_open on dn %p\n", dn); |
| 267 | err = finish_no_open(file, dn); |
| 268 | } else { |
| 269 | dout("atomic_open finish_open on dn %p\n", dn); |
| 270 | if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) { |
| 271 | *opened |= FILE_CREATED; |
| 272 | } |
| 273 | err = finish_open(file, dentry, ceph_open, opened); |
| 274 | } |
| 275 | |
| 276 | out_err: |
| 277 | ceph_mdsc_put_request(req); |
| 278 | dout("atomic_open result=%d\n", err); |
| 279 | return err; |
| 280 | } |
| 281 | |
| 282 | int ceph_release(struct inode *inode, struct file *file) |
| 283 | { |
| 284 | struct ceph_inode_info *ci = ceph_inode(inode); |
| 285 | struct ceph_file_info *cf = file->private_data; |
| 286 | |
| 287 | dout("release inode %p file %p\n", inode, file); |
| 288 | ceph_put_fmode(ci, cf->fmode); |
| 289 | if (cf->last_readdir) |
| 290 | ceph_mdsc_put_request(cf->last_readdir); |
| 291 | kfree(cf->last_name); |
| 292 | kfree(cf->dir_info); |
| 293 | dput(cf->dentry); |
| 294 | kmem_cache_free(ceph_file_cachep, cf); |
| 295 | |
| 296 | /* wake up anyone waiting for caps on this inode */ |
| 297 | wake_up_all(&ci->i_cap_wq); |
| 298 | return 0; |
| 299 | } |
| 300 | |
| 301 | /* |
| 302 | * Read a range of bytes striped over one or more objects. Iterate over |
| 303 | * objects we stripe over. (That's not atomic, but good enough for now.) |
| 304 | * |
| 305 | * If we get a short result from the OSD, check against i_size; we need to |
| 306 | * only return a short read to the caller if we hit EOF. |
| 307 | */ |
| 308 | static int striped_read(struct inode *inode, |
| 309 | u64 off, u64 len, |
| 310 | struct page **pages, int num_pages, |
| 311 | int *checkeof, bool o_direct, |
| 312 | unsigned long buf_align) |
| 313 | { |
| 314 | struct ceph_fs_client *fsc = ceph_inode_to_client(inode); |
| 315 | struct ceph_inode_info *ci = ceph_inode(inode); |
| 316 | u64 pos, this_len, left; |
| 317 | int io_align, page_align; |
| 318 | int pages_left; |
| 319 | int read; |
| 320 | struct page **page_pos; |
| 321 | int ret; |
| 322 | bool hit_stripe, was_short; |
| 323 | |
| 324 | /* |
| 325 | * we may need to do multiple reads. not atomic, unfortunately. |
| 326 | */ |
| 327 | pos = off; |
| 328 | left = len; |
| 329 | page_pos = pages; |
| 330 | pages_left = num_pages; |
| 331 | read = 0; |
| 332 | io_align = off & ~PAGE_MASK; |
| 333 | |
| 334 | more: |
| 335 | if (o_direct) |
| 336 | page_align = (pos - io_align + buf_align) & ~PAGE_MASK; |
| 337 | else |
| 338 | page_align = pos & ~PAGE_MASK; |
| 339 | this_len = left; |
| 340 | ret = ceph_osdc_readpages(&fsc->client->osdc, ceph_vino(inode), |
| 341 | &ci->i_layout, pos, &this_len, |
| 342 | ci->i_truncate_seq, |
| 343 | ci->i_truncate_size, |
| 344 | page_pos, pages_left, page_align); |
| 345 | if (ret == -ENOENT) |
| 346 | ret = 0; |
| 347 | hit_stripe = this_len < left; |
| 348 | was_short = ret >= 0 && ret < this_len; |
| 349 | dout("striped_read %llu~%llu (read %u) got %d%s%s\n", pos, left, read, |
| 350 | ret, hit_stripe ? " HITSTRIPE" : "", was_short ? " SHORT" : ""); |
| 351 | |
| 352 | if (ret >= 0) { |
| 353 | int didpages; |
| 354 | if (was_short && (pos + ret < inode->i_size)) { |
| 355 | u64 tmp = min(this_len - ret, |
| 356 | inode->i_size - pos - ret); |
| 357 | dout(" zero gap %llu to %llu\n", |
| 358 | pos + ret, pos + ret + tmp); |
| 359 | ceph_zero_page_vector_range(page_align + read + ret, |
| 360 | tmp, pages); |
| 361 | ret += tmp; |
| 362 | } |
| 363 | |
| 364 | didpages = (page_align + ret) >> PAGE_CACHE_SHIFT; |
| 365 | pos += ret; |
| 366 | read = pos - off; |
| 367 | left -= ret; |
| 368 | page_pos += didpages; |
| 369 | pages_left -= didpages; |
| 370 | |
| 371 | /* hit stripe and need continue*/ |
| 372 | if (left && hit_stripe && pos < inode->i_size) |
| 373 | goto more; |
| 374 | } |
| 375 | |
| 376 | if (read > 0) { |
| 377 | ret = read; |
| 378 | /* did we bounce off eof? */ |
| 379 | if (pos + left > inode->i_size) |
| 380 | *checkeof = 1; |
| 381 | } |
| 382 | |
| 383 | dout("striped_read returns %d\n", ret); |
| 384 | return ret; |
| 385 | } |
| 386 | |
| 387 | /* |
| 388 | * Completely synchronous read and write methods. Direct from __user |
| 389 | * buffer to osd, or directly to user pages (if O_DIRECT). |
| 390 | * |
| 391 | * If the read spans object boundary, just do multiple reads. |
| 392 | */ |
| 393 | static ssize_t ceph_sync_read(struct file *file, char __user *data, |
| 394 | unsigned len, loff_t *poff, int *checkeof) |
| 395 | { |
| 396 | struct inode *inode = file_inode(file); |
| 397 | struct page **pages; |
| 398 | u64 off = *poff; |
| 399 | int num_pages, ret; |
| 400 | |
| 401 | dout("sync_read on file %p %llu~%u %s\n", file, off, len, |
| 402 | (file->f_flags & O_DIRECT) ? "O_DIRECT" : ""); |
| 403 | |
| 404 | if (file->f_flags & O_DIRECT) { |
| 405 | num_pages = calc_pages_for((unsigned long)data, len); |
| 406 | pages = ceph_get_direct_page_vector(data, num_pages, true); |
| 407 | } else { |
| 408 | num_pages = calc_pages_for(off, len); |
| 409 | pages = ceph_alloc_page_vector(num_pages, GFP_NOFS); |
| 410 | } |
| 411 | if (IS_ERR(pages)) |
| 412 | return PTR_ERR(pages); |
| 413 | |
| 414 | /* |
| 415 | * flush any page cache pages in this range. this |
| 416 | * will make concurrent normal and sync io slow, |
| 417 | * but it will at least behave sensibly when they are |
| 418 | * in sequence. |
| 419 | */ |
| 420 | ret = filemap_write_and_wait(inode->i_mapping); |
| 421 | if (ret < 0) |
| 422 | goto done; |
| 423 | |
| 424 | ret = striped_read(inode, off, len, pages, num_pages, checkeof, |
| 425 | file->f_flags & O_DIRECT, |
| 426 | (unsigned long)data & ~PAGE_MASK); |
| 427 | |
| 428 | if (ret >= 0 && (file->f_flags & O_DIRECT) == 0) |
| 429 | ret = ceph_copy_page_vector_to_user(pages, data, off, ret); |
| 430 | if (ret >= 0) |
| 431 | *poff = off + ret; |
| 432 | |
| 433 | done: |
| 434 | if (file->f_flags & O_DIRECT) |
| 435 | ceph_put_page_vector(pages, num_pages, true); |
| 436 | else |
| 437 | ceph_release_page_vector(pages, num_pages); |
| 438 | dout("sync_read result %d\n", ret); |
| 439 | return ret; |
| 440 | } |
| 441 | |
| 442 | /* |
| 443 | * Write commit request unsafe callback, called to tell us when a |
| 444 | * request is unsafe (that is, in flight--has been handed to the |
| 445 | * messenger to send to its target osd). It is called again when |
| 446 | * we've received a response message indicating the request is |
| 447 | * "safe" (its CEPH_OSD_FLAG_ONDISK flag is set), or when a request |
| 448 | * is completed early (and unsuccessfully) due to a timeout or |
| 449 | * interrupt. |
| 450 | * |
| 451 | * This is used if we requested both an ACK and ONDISK commit reply |
| 452 | * from the OSD. |
| 453 | */ |
| 454 | static void ceph_sync_write_unsafe(struct ceph_osd_request *req, bool unsafe) |
| 455 | { |
| 456 | struct ceph_inode_info *ci = ceph_inode(req->r_inode); |
| 457 | |
| 458 | dout("%s %p tid %llu %ssafe\n", __func__, req, req->r_tid, |
| 459 | unsafe ? "un" : ""); |
| 460 | if (unsafe) { |
| 461 | ceph_get_cap_refs(ci, CEPH_CAP_FILE_WR); |
| 462 | spin_lock(&ci->i_unsafe_lock); |
| 463 | list_add_tail(&req->r_unsafe_item, |
| 464 | &ci->i_unsafe_writes); |
| 465 | spin_unlock(&ci->i_unsafe_lock); |
| 466 | } else { |
| 467 | spin_lock(&ci->i_unsafe_lock); |
| 468 | list_del_init(&req->r_unsafe_item); |
| 469 | spin_unlock(&ci->i_unsafe_lock); |
| 470 | ceph_put_cap_refs(ci, CEPH_CAP_FILE_WR); |
| 471 | } |
| 472 | } |
| 473 | |
| 474 | /* |
| 475 | * Synchronous write, straight from __user pointer or user pages (if |
| 476 | * O_DIRECT). |
| 477 | * |
| 478 | * If write spans object boundary, just do multiple writes. (For a |
| 479 | * correct atomic write, we should e.g. take write locks on all |
| 480 | * objects, rollback on failure, etc.) |
| 481 | */ |
| 482 | static ssize_t ceph_sync_write(struct file *file, const char __user *data, |
| 483 | size_t left, loff_t pos, loff_t *ppos) |
| 484 | { |
| 485 | struct inode *inode = file_inode(file); |
| 486 | struct ceph_inode_info *ci = ceph_inode(inode); |
| 487 | struct ceph_fs_client *fsc = ceph_inode_to_client(inode); |
| 488 | struct ceph_snap_context *snapc; |
| 489 | struct ceph_vino vino; |
| 490 | struct ceph_osd_request *req; |
| 491 | int num_ops = 1; |
| 492 | struct page **pages; |
| 493 | int num_pages; |
| 494 | u64 len; |
| 495 | int written = 0; |
| 496 | int flags; |
| 497 | int check_caps = 0; |
| 498 | int page_align, io_align; |
| 499 | unsigned long buf_align; |
| 500 | int ret; |
| 501 | struct timespec mtime = CURRENT_TIME; |
| 502 | bool own_pages = false; |
| 503 | |
| 504 | if (ceph_snap(file_inode(file)) != CEPH_NOSNAP) |
| 505 | return -EROFS; |
| 506 | |
| 507 | dout("sync_write on file %p %lld~%u %s\n", file, pos, |
| 508 | (unsigned)left, (file->f_flags & O_DIRECT) ? "O_DIRECT" : ""); |
| 509 | |
| 510 | ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + left); |
| 511 | if (ret < 0) |
| 512 | return ret; |
| 513 | |
| 514 | ret = invalidate_inode_pages2_range(inode->i_mapping, |
| 515 | pos >> PAGE_CACHE_SHIFT, |
| 516 | (pos + left) >> PAGE_CACHE_SHIFT); |
| 517 | if (ret < 0) |
| 518 | dout("invalidate_inode_pages2_range returned %d\n", ret); |
| 519 | |
| 520 | flags = CEPH_OSD_FLAG_ORDERSNAP | |
| 521 | CEPH_OSD_FLAG_ONDISK | |
| 522 | CEPH_OSD_FLAG_WRITE; |
| 523 | if ((file->f_flags & (O_SYNC|O_DIRECT)) == 0) |
| 524 | flags |= CEPH_OSD_FLAG_ACK; |
| 525 | else |
| 526 | num_ops++; /* Also include a 'startsync' command. */ |
| 527 | |
| 528 | /* |
| 529 | * we may need to do multiple writes here if we span an object |
| 530 | * boundary. this isn't atomic, unfortunately. :( |
| 531 | */ |
| 532 | more: |
| 533 | io_align = pos & ~PAGE_MASK; |
| 534 | buf_align = (unsigned long)data & ~PAGE_MASK; |
| 535 | len = left; |
| 536 | |
| 537 | snapc = ci->i_snap_realm->cached_context; |
| 538 | vino = ceph_vino(inode); |
| 539 | req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, |
| 540 | vino, pos, &len, num_ops, |
| 541 | CEPH_OSD_OP_WRITE, flags, snapc, |
| 542 | ci->i_truncate_seq, ci->i_truncate_size, |
| 543 | false); |
| 544 | if (IS_ERR(req)) |
| 545 | return PTR_ERR(req); |
| 546 | |
| 547 | /* write from beginning of first page, regardless of io alignment */ |
| 548 | page_align = file->f_flags & O_DIRECT ? buf_align : io_align; |
| 549 | num_pages = calc_pages_for(page_align, len); |
| 550 | if (file->f_flags & O_DIRECT) { |
| 551 | pages = ceph_get_direct_page_vector(data, num_pages, false); |
| 552 | if (IS_ERR(pages)) { |
| 553 | ret = PTR_ERR(pages); |
| 554 | goto out; |
| 555 | } |
| 556 | |
| 557 | /* |
| 558 | * throw out any page cache pages in this range. this |
| 559 | * may block. |
| 560 | */ |
| 561 | truncate_inode_pages_range(inode->i_mapping, pos, |
| 562 | (pos+len) | (PAGE_CACHE_SIZE-1)); |
| 563 | } else { |
| 564 | pages = ceph_alloc_page_vector(num_pages, GFP_NOFS); |
| 565 | if (IS_ERR(pages)) { |
| 566 | ret = PTR_ERR(pages); |
| 567 | goto out; |
| 568 | } |
| 569 | ret = ceph_copy_user_to_page_vector(pages, data, pos, len); |
| 570 | if (ret < 0) { |
| 571 | ceph_release_page_vector(pages, num_pages); |
| 572 | goto out; |
| 573 | } |
| 574 | |
| 575 | if ((file->f_flags & O_SYNC) == 0) { |
| 576 | /* get a second commit callback */ |
| 577 | req->r_unsafe_callback = ceph_sync_write_unsafe; |
| 578 | req->r_inode = inode; |
| 579 | own_pages = true; |
| 580 | } |
| 581 | } |
| 582 | osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align, |
| 583 | false, own_pages); |
| 584 | |
| 585 | /* BUG_ON(vino.snap != CEPH_NOSNAP); */ |
| 586 | ceph_osdc_build_request(req, pos, snapc, vino.snap, &mtime); |
| 587 | |
| 588 | ret = ceph_osdc_start_request(&fsc->client->osdc, req, false); |
| 589 | if (!ret) |
| 590 | ret = ceph_osdc_wait_request(&fsc->client->osdc, req); |
| 591 | |
| 592 | if (file->f_flags & O_DIRECT) |
| 593 | ceph_put_page_vector(pages, num_pages, false); |
| 594 | else if (file->f_flags & O_SYNC) |
| 595 | ceph_release_page_vector(pages, num_pages); |
| 596 | |
| 597 | out: |
| 598 | ceph_osdc_put_request(req); |
| 599 | if (ret == 0) { |
| 600 | pos += len; |
| 601 | written += len; |
| 602 | left -= len; |
| 603 | data += len; |
| 604 | if (left) |
| 605 | goto more; |
| 606 | |
| 607 | ret = written; |
| 608 | *ppos = pos; |
| 609 | if (pos > i_size_read(inode)) |
| 610 | check_caps = ceph_inode_set_size(inode, pos); |
| 611 | if (check_caps) |
| 612 | ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, |
| 613 | NULL); |
| 614 | } else if (ret != -EOLDSNAPC && written > 0) { |
| 615 | ret = written; |
| 616 | } |
| 617 | return ret; |
| 618 | } |
| 619 | |
| 620 | /* |
| 621 | * Wrap generic_file_aio_read with checks for cap bits on the inode. |
| 622 | * Atomically grab references, so that those bits are not released |
| 623 | * back to the MDS mid-read. |
| 624 | * |
| 625 | * Hmm, the sync read case isn't actually async... should it be? |
| 626 | */ |
| 627 | static ssize_t ceph_aio_read(struct kiocb *iocb, const struct iovec *iov, |
| 628 | unsigned long nr_segs, loff_t pos) |
| 629 | { |
| 630 | struct file *filp = iocb->ki_filp; |
| 631 | struct ceph_file_info *fi = filp->private_data; |
| 632 | loff_t *ppos = &iocb->ki_pos; |
| 633 | size_t len = iov->iov_len; |
| 634 | struct inode *inode = file_inode(filp); |
| 635 | struct ceph_inode_info *ci = ceph_inode(inode); |
| 636 | void __user *base = iov->iov_base; |
| 637 | ssize_t ret; |
| 638 | int want, got = 0; |
| 639 | int checkeof = 0, read = 0; |
| 640 | |
| 641 | dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n", |
| 642 | inode, ceph_vinop(inode), pos, (unsigned)len, inode); |
| 643 | again: |
| 644 | if (fi->fmode & CEPH_FILE_MODE_LAZY) |
| 645 | want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO; |
| 646 | else |
| 647 | want = CEPH_CAP_FILE_CACHE; |
| 648 | ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, &got, -1); |
| 649 | if (ret < 0) |
| 650 | goto out; |
| 651 | dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n", |
| 652 | inode, ceph_vinop(inode), pos, (unsigned)len, |
| 653 | ceph_cap_string(got)); |
| 654 | |
| 655 | if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 || |
| 656 | (iocb->ki_filp->f_flags & O_DIRECT) || |
| 657 | (inode->i_sb->s_flags & MS_SYNCHRONOUS) || |
| 658 | (fi->flags & CEPH_F_SYNC)) |
| 659 | /* hmm, this isn't really async... */ |
| 660 | ret = ceph_sync_read(filp, base, len, ppos, &checkeof); |
| 661 | else |
| 662 | ret = generic_file_aio_read(iocb, iov, nr_segs, pos); |
| 663 | |
| 664 | out: |
| 665 | dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n", |
| 666 | inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret); |
| 667 | ceph_put_cap_refs(ci, got); |
| 668 | |
| 669 | if (checkeof && ret >= 0) { |
| 670 | int statret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE); |
| 671 | |
| 672 | /* hit EOF or hole? */ |
| 673 | if (statret == 0 && *ppos < inode->i_size) { |
| 674 | dout("aio_read sync_read hit hole, ppos %lld < size %lld, reading more\n", *ppos, inode->i_size); |
| 675 | read += ret; |
| 676 | base += ret; |
| 677 | len -= ret; |
| 678 | checkeof = 0; |
| 679 | goto again; |
| 680 | } |
| 681 | } |
| 682 | if (ret >= 0) |
| 683 | ret += read; |
| 684 | |
| 685 | return ret; |
| 686 | } |
| 687 | |
| 688 | /* |
| 689 | * Take cap references to avoid releasing caps to MDS mid-write. |
| 690 | * |
| 691 | * If we are synchronous, and write with an old snap context, the OSD |
| 692 | * may return EOLDSNAPC. In that case, retry the write.. _after_ |
| 693 | * dropping our cap refs and allowing the pending snap to logically |
| 694 | * complete _before_ this write occurs. |
| 695 | * |
| 696 | * If we are near ENOSPC, write synchronously. |
| 697 | */ |
| 698 | static ssize_t ceph_aio_write(struct kiocb *iocb, const struct iovec *iov, |
| 699 | unsigned long nr_segs, loff_t pos) |
| 700 | { |
| 701 | struct file *file = iocb->ki_filp; |
| 702 | struct ceph_file_info *fi = file->private_data; |
| 703 | struct inode *inode = file_inode(file); |
| 704 | struct ceph_inode_info *ci = ceph_inode(inode); |
| 705 | struct ceph_osd_client *osdc = |
| 706 | &ceph_sb_to_client(inode->i_sb)->client->osdc; |
| 707 | ssize_t count, written = 0; |
| 708 | int err, want, got; |
| 709 | bool hold_mutex; |
| 710 | |
| 711 | if (ceph_snap(inode) != CEPH_NOSNAP) |
| 712 | return -EROFS; |
| 713 | |
| 714 | sb_start_write(inode->i_sb); |
| 715 | mutex_lock(&inode->i_mutex); |
| 716 | hold_mutex = true; |
| 717 | |
| 718 | err = generic_segment_checks(iov, &nr_segs, &count, VERIFY_READ); |
| 719 | if (err) |
| 720 | goto out; |
| 721 | |
| 722 | /* We can write back this queue in page reclaim */ |
| 723 | current->backing_dev_info = file->f_mapping->backing_dev_info; |
| 724 | |
| 725 | err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode)); |
| 726 | if (err) |
| 727 | goto out; |
| 728 | |
| 729 | if (count == 0) |
| 730 | goto out; |
| 731 | |
| 732 | err = file_remove_suid(file); |
| 733 | if (err) |
| 734 | goto out; |
| 735 | |
| 736 | err = file_update_time(file); |
| 737 | if (err) |
| 738 | goto out; |
| 739 | |
| 740 | retry_snap: |
| 741 | if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL)) { |
| 742 | err = -ENOSPC; |
| 743 | goto out; |
| 744 | } |
| 745 | |
| 746 | dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n", |
| 747 | inode, ceph_vinop(inode), pos, count, inode->i_size); |
| 748 | if (fi->fmode & CEPH_FILE_MODE_LAZY) |
| 749 | want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO; |
| 750 | else |
| 751 | want = CEPH_CAP_FILE_BUFFER; |
| 752 | got = 0; |
| 753 | err = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, &got, pos + count); |
| 754 | if (err < 0) |
| 755 | goto out; |
| 756 | |
| 757 | dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n", |
| 758 | inode, ceph_vinop(inode), pos, count, ceph_cap_string(got)); |
| 759 | |
| 760 | if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 || |
| 761 | (iocb->ki_filp->f_flags & O_DIRECT) || |
| 762 | (inode->i_sb->s_flags & MS_SYNCHRONOUS) || |
| 763 | (fi->flags & CEPH_F_SYNC)) { |
| 764 | mutex_unlock(&inode->i_mutex); |
| 765 | written = ceph_sync_write(file, iov->iov_base, count, |
| 766 | pos, &iocb->ki_pos); |
| 767 | } else { |
| 768 | written = generic_file_buffered_write(iocb, iov, nr_segs, |
| 769 | pos, &iocb->ki_pos, |
| 770 | count, 0); |
| 771 | mutex_unlock(&inode->i_mutex); |
| 772 | } |
| 773 | hold_mutex = false; |
| 774 | |
| 775 | if (written >= 0) { |
| 776 | int dirty; |
| 777 | spin_lock(&ci->i_ceph_lock); |
| 778 | dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR); |
| 779 | spin_unlock(&ci->i_ceph_lock); |
| 780 | if (dirty) |
| 781 | __mark_inode_dirty(inode, dirty); |
| 782 | } |
| 783 | |
| 784 | dout("aio_write %p %llx.%llx %llu~%u dropping cap refs on %s\n", |
| 785 | inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len, |
| 786 | ceph_cap_string(got)); |
| 787 | ceph_put_cap_refs(ci, got); |
| 788 | |
| 789 | if (written >= 0 && |
| 790 | ((file->f_flags & O_SYNC) || IS_SYNC(file->f_mapping->host) || |
| 791 | ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_NEARFULL))) { |
| 792 | err = vfs_fsync_range(file, pos, pos + written - 1, 1); |
| 793 | if (err < 0) |
| 794 | written = err; |
| 795 | } |
| 796 | |
| 797 | if (written == -EOLDSNAPC) { |
| 798 | dout("aio_write %p %llx.%llx %llu~%u got EOLDSNAPC, retrying\n", |
| 799 | inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len); |
| 800 | mutex_lock(&inode->i_mutex); |
| 801 | hold_mutex = true; |
| 802 | goto retry_snap; |
| 803 | } |
| 804 | out: |
| 805 | if (hold_mutex) |
| 806 | mutex_unlock(&inode->i_mutex); |
| 807 | sb_end_write(inode->i_sb); |
| 808 | current->backing_dev_info = NULL; |
| 809 | |
| 810 | return written ? written : err; |
| 811 | } |
| 812 | |
| 813 | /* |
| 814 | * llseek. be sure to verify file size on SEEK_END. |
| 815 | */ |
| 816 | static loff_t ceph_llseek(struct file *file, loff_t offset, int whence) |
| 817 | { |
| 818 | struct inode *inode = file->f_mapping->host; |
| 819 | int ret; |
| 820 | |
| 821 | mutex_lock(&inode->i_mutex); |
| 822 | __ceph_do_pending_vmtruncate(inode, false); |
| 823 | |
| 824 | if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) { |
| 825 | ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE); |
| 826 | if (ret < 0) { |
| 827 | offset = ret; |
| 828 | goto out; |
| 829 | } |
| 830 | } |
| 831 | |
| 832 | switch (whence) { |
| 833 | case SEEK_END: |
| 834 | offset += inode->i_size; |
| 835 | break; |
| 836 | case SEEK_CUR: |
| 837 | /* |
| 838 | * Here we special-case the lseek(fd, 0, SEEK_CUR) |
| 839 | * position-querying operation. Avoid rewriting the "same" |
| 840 | * f_pos value back to the file because a concurrent read(), |
| 841 | * write() or lseek() might have altered it |
| 842 | */ |
| 843 | if (offset == 0) { |
| 844 | offset = file->f_pos; |
| 845 | goto out; |
| 846 | } |
| 847 | offset += file->f_pos; |
| 848 | break; |
| 849 | case SEEK_DATA: |
| 850 | if (offset >= inode->i_size) { |
| 851 | ret = -ENXIO; |
| 852 | goto out; |
| 853 | } |
| 854 | break; |
| 855 | case SEEK_HOLE: |
| 856 | if (offset >= inode->i_size) { |
| 857 | ret = -ENXIO; |
| 858 | goto out; |
| 859 | } |
| 860 | offset = inode->i_size; |
| 861 | break; |
| 862 | } |
| 863 | |
| 864 | if (offset < 0 || offset > inode->i_sb->s_maxbytes) { |
| 865 | offset = -EINVAL; |
| 866 | goto out; |
| 867 | } |
| 868 | |
| 869 | /* Special lock needed here? */ |
| 870 | if (offset != file->f_pos) { |
| 871 | file->f_pos = offset; |
| 872 | file->f_version = 0; |
| 873 | } |
| 874 | |
| 875 | out: |
| 876 | mutex_unlock(&inode->i_mutex); |
| 877 | return offset; |
| 878 | } |
| 879 | |
| 880 | const struct file_operations ceph_file_fops = { |
| 881 | .open = ceph_open, |
| 882 | .release = ceph_release, |
| 883 | .llseek = ceph_llseek, |
| 884 | .read = do_sync_read, |
| 885 | .write = do_sync_write, |
| 886 | .aio_read = ceph_aio_read, |
| 887 | .aio_write = ceph_aio_write, |
| 888 | .mmap = ceph_mmap, |
| 889 | .fsync = ceph_fsync, |
| 890 | .lock = ceph_lock, |
| 891 | .flock = ceph_flock, |
| 892 | .splice_read = generic_file_splice_read, |
| 893 | .splice_write = generic_file_splice_write, |
| 894 | .unlocked_ioctl = ceph_ioctl, |
| 895 | .compat_ioctl = ceph_ioctl, |
| 896 | }; |
| 897 | |