1 #include <linux/ceph/ceph_debug.h> 2 3 #include <linux/module.h> 4 #include <linux/err.h> 5 #include <linux/highmem.h> 6 #include <linux/mm.h> 7 #include <linux/pagemap.h> 8 #include <linux/slab.h> 9 #include <linux/uaccess.h> 10 #ifdef CONFIG_BLOCK 11 #include <linux/bio.h> 12 #endif 13 14 #include <linux/ceph/libceph.h> 15 #include <linux/ceph/osd_client.h> 16 #include <linux/ceph/messenger.h> 17 #include <linux/ceph/decode.h> 18 #include <linux/ceph/auth.h> 19 #include <linux/ceph/pagelist.h> 20 21 #define OSD_OP_FRONT_LEN 4096 22 #define OSD_OPREPLY_FRONT_LEN 512 23 24 static const struct ceph_connection_operations osd_con_ops; 25 26 static void __send_queued(struct ceph_osd_client *osdc); 27 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd); 28 static void __register_request(struct ceph_osd_client *osdc, 29 struct ceph_osd_request *req); 30 static void __unregister_linger_request(struct ceph_osd_client *osdc, 31 struct ceph_osd_request *req); 32 static void __send_request(struct ceph_osd_client *osdc, 33 struct ceph_osd_request *req); 34 35 /* 36 * Implement client access to distributed object storage cluster. 37 * 38 * All data objects are stored within a cluster/cloud of OSDs, or 39 * "object storage devices." (Note that Ceph OSDs have _nothing_ to 40 * do with the T10 OSD extensions to SCSI.) Ceph OSDs are simply 41 * remote daemons serving up and coordinating consistent and safe 42 * access to storage. 43 * 44 * Cluster membership and the mapping of data objects onto storage devices 45 * are described by the osd map. 46 * 47 * We keep track of pending OSD requests (read, write), resubmit 48 * requests to different OSDs when the cluster topology/data layout 49 * change, or retry the affected requests when the communications 50 * channel with an OSD is reset. 51 */ 52 53 /* 54 * calculate the mapping of a file extent onto an object, and fill out the 55 * request accordingly. shorten extent as necessary if it crosses an 56 * object boundary. 57 * 58 * fill osd op in request message. 59 */ 60 static int calc_layout(struct ceph_file_layout *layout, u64 off, u64 *plen, 61 u64 *objnum, u64 *objoff, u64 *objlen) 62 { 63 u64 orig_len = *plen; 64 int r; 65 66 /* object extent? */ 67 r = ceph_calc_file_object_mapping(layout, off, orig_len, objnum, 68 objoff, objlen); 69 if (r < 0) 70 return r; 71 if (*objlen < orig_len) { 72 *plen = *objlen; 73 dout(" skipping last %llu, final file extent %llu~%llu\n", 74 orig_len - *plen, off, *plen); 75 } 76 77 dout("calc_layout objnum=%llx %llu~%llu\n", *objnum, *objoff, *objlen); 78 79 return 0; 80 } 81 82 /* 83 * requests 84 */ 85 void ceph_osdc_release_request(struct kref *kref) 86 { 87 int num_pages; 88 struct ceph_osd_request *req = container_of(kref, 89 struct ceph_osd_request, 90 r_kref); 91 92 if (req->r_request) 93 ceph_msg_put(req->r_request); 94 if (req->r_reply) { 95 ceph_msg_revoke_incoming(req->r_reply); 96 ceph_msg_put(req->r_reply); 97 } 98 99 if (req->r_data_in.type == CEPH_OSD_DATA_TYPE_PAGES && 100 req->r_data_in.own_pages) { 101 num_pages = calc_pages_for((u64)req->r_data_in.alignment, 102 (u64)req->r_data_in.length); 103 ceph_release_page_vector(req->r_data_in.pages, num_pages); 104 } 105 if (req->r_data_out.type == CEPH_OSD_DATA_TYPE_PAGES && 106 req->r_data_out.own_pages) { 107 num_pages = calc_pages_for((u64)req->r_data_out.alignment, 108 (u64)req->r_data_out.length); 109 ceph_release_page_vector(req->r_data_out.pages, num_pages); 110 } 111 112 ceph_put_snap_context(req->r_snapc); 113 if (req->r_mempool) 114 mempool_free(req, req->r_osdc->req_mempool); 115 else 116 kfree(req); 117 } 118 EXPORT_SYMBOL(ceph_osdc_release_request); 119 120 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc, 121 struct ceph_snap_context *snapc, 122 unsigned int num_ops, 123 bool use_mempool, 124 gfp_t gfp_flags) 125 { 126 struct ceph_osd_request *req; 127 struct ceph_msg *msg; 128 size_t msg_size; 129 130 msg_size = 4 + 4 + 8 + 8 + 4+8; 131 msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */ 132 msg_size += 1 + 8 + 4 + 4; /* pg_t */ 133 msg_size += 4 + MAX_OBJ_NAME_SIZE; 134 msg_size += 2 + num_ops*sizeof(struct ceph_osd_op); 135 msg_size += 8; /* snapid */ 136 msg_size += 8; /* snap_seq */ 137 msg_size += 8 * (snapc ? snapc->num_snaps : 0); /* snaps */ 138 msg_size += 4; 139 140 if (use_mempool) { 141 req = mempool_alloc(osdc->req_mempool, gfp_flags); 142 memset(req, 0, sizeof(*req)); 143 } else { 144 req = kzalloc(sizeof(*req), gfp_flags); 145 } 146 if (req == NULL) 147 return NULL; 148 149 req->r_osdc = osdc; 150 req->r_mempool = use_mempool; 151 152 kref_init(&req->r_kref); 153 init_completion(&req->r_completion); 154 init_completion(&req->r_safe_completion); 155 RB_CLEAR_NODE(&req->r_node); 156 INIT_LIST_HEAD(&req->r_unsafe_item); 157 INIT_LIST_HEAD(&req->r_linger_item); 158 INIT_LIST_HEAD(&req->r_linger_osd); 159 INIT_LIST_HEAD(&req->r_req_lru_item); 160 INIT_LIST_HEAD(&req->r_osd_item); 161 162 /* create reply message */ 163 if (use_mempool) 164 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0); 165 else 166 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, 167 OSD_OPREPLY_FRONT_LEN, gfp_flags, true); 168 if (!msg) { 169 ceph_osdc_put_request(req); 170 return NULL; 171 } 172 req->r_reply = msg; 173 174 req->r_data_in.type = CEPH_OSD_DATA_TYPE_NONE; 175 req->r_data_out.type = CEPH_OSD_DATA_TYPE_NONE; 176 177 /* create request message; allow space for oid */ 178 if (use_mempool) 179 msg = ceph_msgpool_get(&osdc->msgpool_op, 0); 180 else 181 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true); 182 if (!msg) { 183 ceph_osdc_put_request(req); 184 return NULL; 185 } 186 187 memset(msg->front.iov_base, 0, msg->front.iov_len); 188 189 req->r_request = msg; 190 191 return req; 192 } 193 EXPORT_SYMBOL(ceph_osdc_alloc_request); 194 195 static bool osd_req_opcode_valid(u16 opcode) 196 { 197 switch (opcode) { 198 case CEPH_OSD_OP_READ: 199 case CEPH_OSD_OP_STAT: 200 case CEPH_OSD_OP_MAPEXT: 201 case CEPH_OSD_OP_MASKTRUNC: 202 case CEPH_OSD_OP_SPARSE_READ: 203 case CEPH_OSD_OP_NOTIFY: 204 case CEPH_OSD_OP_NOTIFY_ACK: 205 case CEPH_OSD_OP_ASSERT_VER: 206 case CEPH_OSD_OP_WRITE: 207 case CEPH_OSD_OP_WRITEFULL: 208 case CEPH_OSD_OP_TRUNCATE: 209 case CEPH_OSD_OP_ZERO: 210 case CEPH_OSD_OP_DELETE: 211 case CEPH_OSD_OP_APPEND: 212 case CEPH_OSD_OP_STARTSYNC: 213 case CEPH_OSD_OP_SETTRUNC: 214 case CEPH_OSD_OP_TRIMTRUNC: 215 case CEPH_OSD_OP_TMAPUP: 216 case CEPH_OSD_OP_TMAPPUT: 217 case CEPH_OSD_OP_TMAPGET: 218 case CEPH_OSD_OP_CREATE: 219 case CEPH_OSD_OP_ROLLBACK: 220 case CEPH_OSD_OP_WATCH: 221 case CEPH_OSD_OP_OMAPGETKEYS: 222 case CEPH_OSD_OP_OMAPGETVALS: 223 case CEPH_OSD_OP_OMAPGETHEADER: 224 case CEPH_OSD_OP_OMAPGETVALSBYKEYS: 225 case CEPH_OSD_OP_OMAPSETVALS: 226 case CEPH_OSD_OP_OMAPSETHEADER: 227 case CEPH_OSD_OP_OMAPCLEAR: 228 case CEPH_OSD_OP_OMAPRMKEYS: 229 case CEPH_OSD_OP_OMAP_CMP: 230 case CEPH_OSD_OP_CLONERANGE: 231 case CEPH_OSD_OP_ASSERT_SRC_VERSION: 232 case CEPH_OSD_OP_SRC_CMPXATTR: 233 case CEPH_OSD_OP_GETXATTR: 234 case CEPH_OSD_OP_GETXATTRS: 235 case CEPH_OSD_OP_CMPXATTR: 236 case CEPH_OSD_OP_SETXATTR: 237 case CEPH_OSD_OP_SETXATTRS: 238 case CEPH_OSD_OP_RESETXATTRS: 239 case CEPH_OSD_OP_RMXATTR: 240 case CEPH_OSD_OP_PULL: 241 case CEPH_OSD_OP_PUSH: 242 case CEPH_OSD_OP_BALANCEREADS: 243 case CEPH_OSD_OP_UNBALANCEREADS: 244 case CEPH_OSD_OP_SCRUB: 245 case CEPH_OSD_OP_SCRUB_RESERVE: 246 case CEPH_OSD_OP_SCRUB_UNRESERVE: 247 case CEPH_OSD_OP_SCRUB_STOP: 248 case CEPH_OSD_OP_SCRUB_MAP: 249 case CEPH_OSD_OP_WRLOCK: 250 case CEPH_OSD_OP_WRUNLOCK: 251 case CEPH_OSD_OP_RDLOCK: 252 case CEPH_OSD_OP_RDUNLOCK: 253 case CEPH_OSD_OP_UPLOCK: 254 case CEPH_OSD_OP_DNLOCK: 255 case CEPH_OSD_OP_CALL: 256 case CEPH_OSD_OP_PGLS: 257 case CEPH_OSD_OP_PGLS_FILTER: 258 return true; 259 default: 260 return false; 261 } 262 } 263 264 /* 265 * This is an osd op init function for opcodes that have no data or 266 * other information associated with them. It also serves as a 267 * common init routine for all the other init functions, below. 268 */ 269 void osd_req_op_init(struct ceph_osd_req_op *op, u16 opcode) 270 { 271 BUG_ON(!osd_req_opcode_valid(opcode)); 272 273 memset(op, 0, sizeof (*op)); 274 275 op->op = opcode; 276 } 277 278 void osd_req_op_extent_init(struct ceph_osd_req_op *op, u16 opcode, 279 u64 offset, u64 length, 280 u64 truncate_size, u32 truncate_seq) 281 { 282 size_t payload_len = 0; 283 284 BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE); 285 286 osd_req_op_init(op, opcode); 287 288 op->extent.offset = offset; 289 op->extent.length = length; 290 op->extent.truncate_size = truncate_size; 291 op->extent.truncate_seq = truncate_seq; 292 if (opcode == CEPH_OSD_OP_WRITE) 293 payload_len += length; 294 295 op->payload_len = payload_len; 296 } 297 EXPORT_SYMBOL(osd_req_op_extent_init); 298 299 void osd_req_op_extent_update(struct ceph_osd_req_op *op, u64 length) 300 { 301 u64 previous = op->extent.length; 302 303 if (length == previous) 304 return; /* Nothing to do */ 305 BUG_ON(length > previous); 306 307 op->extent.length = length; 308 op->payload_len -= previous - length; 309 } 310 EXPORT_SYMBOL(osd_req_op_extent_update); 311 312 void osd_req_op_cls_init(struct ceph_osd_req_op *op, u16 opcode, 313 const char *class, const char *method, 314 const void *request_data, size_t request_data_size) 315 { 316 size_t payload_len = 0; 317 size_t size; 318 319 BUG_ON(opcode != CEPH_OSD_OP_CALL); 320 321 osd_req_op_init(op, opcode); 322 323 op->cls.class_name = class; 324 size = strlen(class); 325 BUG_ON(size > (size_t) U8_MAX); 326 op->cls.class_len = size; 327 payload_len += size; 328 329 op->cls.method_name = method; 330 size = strlen(method); 331 BUG_ON(size > (size_t) U8_MAX); 332 op->cls.method_len = size; 333 payload_len += size; 334 335 op->cls.indata = request_data; 336 BUG_ON(request_data_size > (size_t) U32_MAX); 337 op->cls.indata_len = (u32) request_data_size; 338 payload_len += request_data_size; 339 340 op->cls.argc = 0; /* currently unused */ 341 342 op->payload_len = payload_len; 343 } 344 EXPORT_SYMBOL(osd_req_op_cls_init); 345 346 void osd_req_op_watch_init(struct ceph_osd_req_op *op, u16 opcode, 347 u64 cookie, u64 version, int flag) 348 { 349 BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH); 350 351 osd_req_op_init(op, opcode); 352 353 op->watch.cookie = cookie; 354 /* op->watch.ver = version; */ /* XXX 3847 */ 355 op->watch.ver = cpu_to_le64(version); 356 if (opcode == CEPH_OSD_OP_WATCH && flag) 357 op->watch.flag = (u8) 1; 358 } 359 EXPORT_SYMBOL(osd_req_op_watch_init); 360 361 static u64 osd_req_encode_op(struct ceph_osd_request *req, 362 struct ceph_osd_op *dst, 363 struct ceph_osd_req_op *src) 364 { 365 u64 out_data_len = 0; 366 struct ceph_pagelist *pagelist; 367 368 if (WARN_ON(!osd_req_opcode_valid(src->op))) { 369 pr_err("unrecognized osd opcode %d\n", src->op); 370 371 return 0; 372 } 373 374 switch (src->op) { 375 case CEPH_OSD_OP_STAT: 376 break; 377 case CEPH_OSD_OP_READ: 378 case CEPH_OSD_OP_WRITE: 379 if (src->op == CEPH_OSD_OP_WRITE) 380 out_data_len = src->extent.length; 381 dst->extent.offset = cpu_to_le64(src->extent.offset); 382 dst->extent.length = cpu_to_le64(src->extent.length); 383 dst->extent.truncate_size = 384 cpu_to_le64(src->extent.truncate_size); 385 dst->extent.truncate_seq = 386 cpu_to_le32(src->extent.truncate_seq); 387 break; 388 case CEPH_OSD_OP_CALL: 389 pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS); 390 BUG_ON(!pagelist); 391 ceph_pagelist_init(pagelist); 392 393 dst->cls.class_len = src->cls.class_len; 394 dst->cls.method_len = src->cls.method_len; 395 dst->cls.indata_len = cpu_to_le32(src->cls.indata_len); 396 ceph_pagelist_append(pagelist, src->cls.class_name, 397 src->cls.class_len); 398 ceph_pagelist_append(pagelist, src->cls.method_name, 399 src->cls.method_len); 400 ceph_pagelist_append(pagelist, src->cls.indata, 401 src->cls.indata_len); 402 403 req->r_data_out.type = CEPH_OSD_DATA_TYPE_PAGELIST; 404 req->r_data_out.pagelist = pagelist; 405 out_data_len = pagelist->length; 406 break; 407 case CEPH_OSD_OP_STARTSYNC: 408 break; 409 case CEPH_OSD_OP_NOTIFY_ACK: 410 case CEPH_OSD_OP_WATCH: 411 dst->watch.cookie = cpu_to_le64(src->watch.cookie); 412 dst->watch.ver = cpu_to_le64(src->watch.ver); 413 dst->watch.flag = src->watch.flag; 414 break; 415 default: 416 pr_err("unsupported osd opcode %s\n", 417 ceph_osd_op_name(src->op)); 418 WARN_ON(1); 419 420 return 0; 421 } 422 dst->op = cpu_to_le16(src->op); 423 dst->payload_len = cpu_to_le32(src->payload_len); 424 425 return out_data_len; 426 } 427 428 /* 429 * build new request AND message 430 * 431 */ 432 void ceph_osdc_build_request(struct ceph_osd_request *req, 433 u64 off, unsigned int num_ops, 434 struct ceph_osd_req_op *src_ops, 435 struct ceph_snap_context *snapc, u64 snap_id, 436 struct timespec *mtime) 437 { 438 struct ceph_msg *msg = req->r_request; 439 struct ceph_osd_req_op *src_op; 440 void *p; 441 size_t msg_size; 442 int flags = req->r_flags; 443 u64 data_len; 444 int i; 445 446 req->r_num_ops = num_ops; 447 req->r_snapid = snap_id; 448 req->r_snapc = ceph_get_snap_context(snapc); 449 450 /* encode request */ 451 msg->hdr.version = cpu_to_le16(4); 452 453 p = msg->front.iov_base; 454 ceph_encode_32(&p, 1); /* client_inc is always 1 */ 455 req->r_request_osdmap_epoch = p; 456 p += 4; 457 req->r_request_flags = p; 458 p += 4; 459 if (req->r_flags & CEPH_OSD_FLAG_WRITE) 460 ceph_encode_timespec(p, mtime); 461 p += sizeof(struct ceph_timespec); 462 req->r_request_reassert_version = p; 463 p += sizeof(struct ceph_eversion); /* will get filled in */ 464 465 /* oloc */ 466 ceph_encode_8(&p, 4); 467 ceph_encode_8(&p, 4); 468 ceph_encode_32(&p, 8 + 4 + 4); 469 req->r_request_pool = p; 470 p += 8; 471 ceph_encode_32(&p, -1); /* preferred */ 472 ceph_encode_32(&p, 0); /* key len */ 473 474 ceph_encode_8(&p, 1); 475 req->r_request_pgid = p; 476 p += 8 + 4; 477 ceph_encode_32(&p, -1); /* preferred */ 478 479 /* oid */ 480 ceph_encode_32(&p, req->r_oid_len); 481 memcpy(p, req->r_oid, req->r_oid_len); 482 dout("oid '%.*s' len %d\n", req->r_oid_len, req->r_oid, req->r_oid_len); 483 p += req->r_oid_len; 484 485 /* ops--can imply data */ 486 ceph_encode_16(&p, num_ops); 487 src_op = src_ops; 488 req->r_request_ops = p; 489 data_len = 0; 490 for (i = 0; i < num_ops; i++, src_op++) { 491 data_len += osd_req_encode_op(req, p, src_op); 492 p += sizeof(struct ceph_osd_op); 493 } 494 495 /* snaps */ 496 ceph_encode_64(&p, req->r_snapid); 497 ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0); 498 ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0); 499 if (req->r_snapc) { 500 for (i = 0; i < snapc->num_snaps; i++) { 501 ceph_encode_64(&p, req->r_snapc->snaps[i]); 502 } 503 } 504 505 req->r_request_attempts = p; 506 p += 4; 507 508 /* data */ 509 if (flags & CEPH_OSD_FLAG_WRITE) { 510 u16 data_off; 511 512 /* 513 * The header "data_off" is a hint to the receiver 514 * allowing it to align received data into its 515 * buffers such that there's no need to re-copy 516 * it before writing it to disk (direct I/O). 517 */ 518 data_off = (u16) (off & 0xffff); 519 req->r_request->hdr.data_off = cpu_to_le16(data_off); 520 } 521 req->r_request->hdr.data_len = cpu_to_le32(data_len); 522 523 BUG_ON(p > msg->front.iov_base + msg->front.iov_len); 524 msg_size = p - msg->front.iov_base; 525 msg->front.iov_len = msg_size; 526 msg->hdr.front_len = cpu_to_le32(msg_size); 527 528 dout("build_request msg_size was %d\n", (int)msg_size); 529 } 530 EXPORT_SYMBOL(ceph_osdc_build_request); 531 532 /* 533 * build new request AND message, calculate layout, and adjust file 534 * extent as needed. 535 * 536 * if the file was recently truncated, we include information about its 537 * old and new size so that the object can be updated appropriately. (we 538 * avoid synchronously deleting truncated objects because it's slow.) 539 * 540 * if @do_sync, include a 'startsync' command so that the osd will flush 541 * data quickly. 542 */ 543 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc, 544 struct ceph_file_layout *layout, 545 struct ceph_vino vino, 546 u64 off, u64 *plen, int num_ops, 547 struct ceph_osd_req_op *ops, 548 int opcode, int flags, 549 struct ceph_snap_context *snapc, 550 u32 truncate_seq, 551 u64 truncate_size, 552 bool use_mempool) 553 { 554 struct ceph_osd_request *req; 555 u64 objnum = 0; 556 u64 objoff = 0; 557 u64 objlen = 0; 558 u32 object_size; 559 u64 object_base; 560 int r; 561 562 BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE); 563 564 req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool, 565 GFP_NOFS); 566 if (!req) 567 return ERR_PTR(-ENOMEM); 568 req->r_flags = flags; 569 570 /* calculate max write size */ 571 r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen); 572 if (r < 0) { 573 ceph_osdc_put_request(req); 574 return ERR_PTR(r); 575 } 576 577 object_size = le32_to_cpu(layout->fl_object_size); 578 object_base = off - objoff; 579 if (truncate_size <= object_base) { 580 truncate_size = 0; 581 } else { 582 truncate_size -= object_base; 583 if (truncate_size > object_size) 584 truncate_size = object_size; 585 } 586 587 osd_req_op_extent_init(&ops[0], opcode, objoff, objlen, 588 truncate_size, truncate_seq); 589 /* 590 * A second op in the ops array means the caller wants to 591 * also issue a include a 'startsync' command so that the 592 * osd will flush data quickly. 593 */ 594 if (num_ops > 1) 595 osd_req_op_init(&ops[1], CEPH_OSD_OP_STARTSYNC); 596 597 req->r_file_layout = *layout; /* keep a copy */ 598 599 snprintf(req->r_oid, sizeof(req->r_oid), "%llx.%08llx", 600 vino.ino, objnum); 601 req->r_oid_len = strlen(req->r_oid); 602 603 return req; 604 } 605 EXPORT_SYMBOL(ceph_osdc_new_request); 606 607 /* 608 * We keep osd requests in an rbtree, sorted by ->r_tid. 609 */ 610 static void __insert_request(struct ceph_osd_client *osdc, 611 struct ceph_osd_request *new) 612 { 613 struct rb_node **p = &osdc->requests.rb_node; 614 struct rb_node *parent = NULL; 615 struct ceph_osd_request *req = NULL; 616 617 while (*p) { 618 parent = *p; 619 req = rb_entry(parent, struct ceph_osd_request, r_node); 620 if (new->r_tid < req->r_tid) 621 p = &(*p)->rb_left; 622 else if (new->r_tid > req->r_tid) 623 p = &(*p)->rb_right; 624 else 625 BUG(); 626 } 627 628 rb_link_node(&new->r_node, parent, p); 629 rb_insert_color(&new->r_node, &osdc->requests); 630 } 631 632 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc, 633 u64 tid) 634 { 635 struct ceph_osd_request *req; 636 struct rb_node *n = osdc->requests.rb_node; 637 638 while (n) { 639 req = rb_entry(n, struct ceph_osd_request, r_node); 640 if (tid < req->r_tid) 641 n = n->rb_left; 642 else if (tid > req->r_tid) 643 n = n->rb_right; 644 else 645 return req; 646 } 647 return NULL; 648 } 649 650 static struct ceph_osd_request * 651 __lookup_request_ge(struct ceph_osd_client *osdc, 652 u64 tid) 653 { 654 struct ceph_osd_request *req; 655 struct rb_node *n = osdc->requests.rb_node; 656 657 while (n) { 658 req = rb_entry(n, struct ceph_osd_request, r_node); 659 if (tid < req->r_tid) { 660 if (!n->rb_left) 661 return req; 662 n = n->rb_left; 663 } else if (tid > req->r_tid) { 664 n = n->rb_right; 665 } else { 666 return req; 667 } 668 } 669 return NULL; 670 } 671 672 /* 673 * Resubmit requests pending on the given osd. 674 */ 675 static void __kick_osd_requests(struct ceph_osd_client *osdc, 676 struct ceph_osd *osd) 677 { 678 struct ceph_osd_request *req, *nreq; 679 LIST_HEAD(resend); 680 int err; 681 682 dout("__kick_osd_requests osd%d\n", osd->o_osd); 683 err = __reset_osd(osdc, osd); 684 if (err) 685 return; 686 /* 687 * Build up a list of requests to resend by traversing the 688 * osd's list of requests. Requests for a given object are 689 * sent in tid order, and that is also the order they're 690 * kept on this list. Therefore all requests that are in 691 * flight will be found first, followed by all requests that 692 * have not yet been sent. And to resend requests while 693 * preserving this order we will want to put any sent 694 * requests back on the front of the osd client's unsent 695 * list. 696 * 697 * So we build a separate ordered list of already-sent 698 * requests for the affected osd and splice it onto the 699 * front of the osd client's unsent list. Once we've seen a 700 * request that has not yet been sent we're done. Those 701 * requests are already sitting right where they belong. 702 */ 703 list_for_each_entry(req, &osd->o_requests, r_osd_item) { 704 if (!req->r_sent) 705 break; 706 list_move_tail(&req->r_req_lru_item, &resend); 707 dout("requeueing %p tid %llu osd%d\n", req, req->r_tid, 708 osd->o_osd); 709 if (!req->r_linger) 710 req->r_flags |= CEPH_OSD_FLAG_RETRY; 711 } 712 list_splice(&resend, &osdc->req_unsent); 713 714 /* 715 * Linger requests are re-registered before sending, which 716 * sets up a new tid for each. We add them to the unsent 717 * list at the end to keep things in tid order. 718 */ 719 list_for_each_entry_safe(req, nreq, &osd->o_linger_requests, 720 r_linger_osd) { 721 /* 722 * reregister request prior to unregistering linger so 723 * that r_osd is preserved. 724 */ 725 BUG_ON(!list_empty(&req->r_req_lru_item)); 726 __register_request(osdc, req); 727 list_add_tail(&req->r_req_lru_item, &osdc->req_unsent); 728 list_add_tail(&req->r_osd_item, &req->r_osd->o_requests); 729 __unregister_linger_request(osdc, req); 730 dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid, 731 osd->o_osd); 732 } 733 } 734 735 /* 736 * If the osd connection drops, we need to resubmit all requests. 737 */ 738 static void osd_reset(struct ceph_connection *con) 739 { 740 struct ceph_osd *osd = con->private; 741 struct ceph_osd_client *osdc; 742 743 if (!osd) 744 return; 745 dout("osd_reset osd%d\n", osd->o_osd); 746 osdc = osd->o_osdc; 747 down_read(&osdc->map_sem); 748 mutex_lock(&osdc->request_mutex); 749 __kick_osd_requests(osdc, osd); 750 __send_queued(osdc); 751 mutex_unlock(&osdc->request_mutex); 752 up_read(&osdc->map_sem); 753 } 754 755 /* 756 * Track open sessions with osds. 757 */ 758 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum) 759 { 760 struct ceph_osd *osd; 761 762 osd = kzalloc(sizeof(*osd), GFP_NOFS); 763 if (!osd) 764 return NULL; 765 766 atomic_set(&osd->o_ref, 1); 767 osd->o_osdc = osdc; 768 osd->o_osd = onum; 769 RB_CLEAR_NODE(&osd->o_node); 770 INIT_LIST_HEAD(&osd->o_requests); 771 INIT_LIST_HEAD(&osd->o_linger_requests); 772 INIT_LIST_HEAD(&osd->o_osd_lru); 773 osd->o_incarnation = 1; 774 775 ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr); 776 777 INIT_LIST_HEAD(&osd->o_keepalive_item); 778 return osd; 779 } 780 781 static struct ceph_osd *get_osd(struct ceph_osd *osd) 782 { 783 if (atomic_inc_not_zero(&osd->o_ref)) { 784 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1, 785 atomic_read(&osd->o_ref)); 786 return osd; 787 } else { 788 dout("get_osd %p FAIL\n", osd); 789 return NULL; 790 } 791 } 792 793 static void put_osd(struct ceph_osd *osd) 794 { 795 dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref), 796 atomic_read(&osd->o_ref) - 1); 797 if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) { 798 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth; 799 800 ceph_auth_destroy_authorizer(ac, osd->o_auth.authorizer); 801 kfree(osd); 802 } 803 } 804 805 /* 806 * remove an osd from our map 807 */ 808 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 809 { 810 dout("__remove_osd %p\n", osd); 811 BUG_ON(!list_empty(&osd->o_requests)); 812 rb_erase(&osd->o_node, &osdc->osds); 813 list_del_init(&osd->o_osd_lru); 814 ceph_con_close(&osd->o_con); 815 put_osd(osd); 816 } 817 818 static void remove_all_osds(struct ceph_osd_client *osdc) 819 { 820 dout("%s %p\n", __func__, osdc); 821 mutex_lock(&osdc->request_mutex); 822 while (!RB_EMPTY_ROOT(&osdc->osds)) { 823 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds), 824 struct ceph_osd, o_node); 825 __remove_osd(osdc, osd); 826 } 827 mutex_unlock(&osdc->request_mutex); 828 } 829 830 static void __move_osd_to_lru(struct ceph_osd_client *osdc, 831 struct ceph_osd *osd) 832 { 833 dout("__move_osd_to_lru %p\n", osd); 834 BUG_ON(!list_empty(&osd->o_osd_lru)); 835 list_add_tail(&osd->o_osd_lru, &osdc->osd_lru); 836 osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ; 837 } 838 839 static void __remove_osd_from_lru(struct ceph_osd *osd) 840 { 841 dout("__remove_osd_from_lru %p\n", osd); 842 if (!list_empty(&osd->o_osd_lru)) 843 list_del_init(&osd->o_osd_lru); 844 } 845 846 static void remove_old_osds(struct ceph_osd_client *osdc) 847 { 848 struct ceph_osd *osd, *nosd; 849 850 dout("__remove_old_osds %p\n", osdc); 851 mutex_lock(&osdc->request_mutex); 852 list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) { 853 if (time_before(jiffies, osd->lru_ttl)) 854 break; 855 __remove_osd(osdc, osd); 856 } 857 mutex_unlock(&osdc->request_mutex); 858 } 859 860 /* 861 * reset osd connect 862 */ 863 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 864 { 865 struct ceph_entity_addr *peer_addr; 866 867 dout("__reset_osd %p osd%d\n", osd, osd->o_osd); 868 if (list_empty(&osd->o_requests) && 869 list_empty(&osd->o_linger_requests)) { 870 __remove_osd(osdc, osd); 871 872 return -ENODEV; 873 } 874 875 peer_addr = &osdc->osdmap->osd_addr[osd->o_osd]; 876 if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) && 877 !ceph_con_opened(&osd->o_con)) { 878 struct ceph_osd_request *req; 879 880 dout(" osd addr hasn't changed and connection never opened," 881 " letting msgr retry"); 882 /* touch each r_stamp for handle_timeout()'s benfit */ 883 list_for_each_entry(req, &osd->o_requests, r_osd_item) 884 req->r_stamp = jiffies; 885 886 return -EAGAIN; 887 } 888 889 ceph_con_close(&osd->o_con); 890 ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr); 891 osd->o_incarnation++; 892 893 return 0; 894 } 895 896 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new) 897 { 898 struct rb_node **p = &osdc->osds.rb_node; 899 struct rb_node *parent = NULL; 900 struct ceph_osd *osd = NULL; 901 902 dout("__insert_osd %p osd%d\n", new, new->o_osd); 903 while (*p) { 904 parent = *p; 905 osd = rb_entry(parent, struct ceph_osd, o_node); 906 if (new->o_osd < osd->o_osd) 907 p = &(*p)->rb_left; 908 else if (new->o_osd > osd->o_osd) 909 p = &(*p)->rb_right; 910 else 911 BUG(); 912 } 913 914 rb_link_node(&new->o_node, parent, p); 915 rb_insert_color(&new->o_node, &osdc->osds); 916 } 917 918 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o) 919 { 920 struct ceph_osd *osd; 921 struct rb_node *n = osdc->osds.rb_node; 922 923 while (n) { 924 osd = rb_entry(n, struct ceph_osd, o_node); 925 if (o < osd->o_osd) 926 n = n->rb_left; 927 else if (o > osd->o_osd) 928 n = n->rb_right; 929 else 930 return osd; 931 } 932 return NULL; 933 } 934 935 static void __schedule_osd_timeout(struct ceph_osd_client *osdc) 936 { 937 schedule_delayed_work(&osdc->timeout_work, 938 osdc->client->options->osd_keepalive_timeout * HZ); 939 } 940 941 static void __cancel_osd_timeout(struct ceph_osd_client *osdc) 942 { 943 cancel_delayed_work(&osdc->timeout_work); 944 } 945 946 /* 947 * Register request, assign tid. If this is the first request, set up 948 * the timeout event. 949 */ 950 static void __register_request(struct ceph_osd_client *osdc, 951 struct ceph_osd_request *req) 952 { 953 req->r_tid = ++osdc->last_tid; 954 req->r_request->hdr.tid = cpu_to_le64(req->r_tid); 955 dout("__register_request %p tid %lld\n", req, req->r_tid); 956 __insert_request(osdc, req); 957 ceph_osdc_get_request(req); 958 osdc->num_requests++; 959 if (osdc->num_requests == 1) { 960 dout(" first request, scheduling timeout\n"); 961 __schedule_osd_timeout(osdc); 962 } 963 } 964 965 /* 966 * called under osdc->request_mutex 967 */ 968 static void __unregister_request(struct ceph_osd_client *osdc, 969 struct ceph_osd_request *req) 970 { 971 if (RB_EMPTY_NODE(&req->r_node)) { 972 dout("__unregister_request %p tid %lld not registered\n", 973 req, req->r_tid); 974 return; 975 } 976 977 dout("__unregister_request %p tid %lld\n", req, req->r_tid); 978 rb_erase(&req->r_node, &osdc->requests); 979 osdc->num_requests--; 980 981 if (req->r_osd) { 982 /* make sure the original request isn't in flight. */ 983 ceph_msg_revoke(req->r_request); 984 985 list_del_init(&req->r_osd_item); 986 if (list_empty(&req->r_osd->o_requests) && 987 list_empty(&req->r_osd->o_linger_requests)) { 988 dout("moving osd to %p lru\n", req->r_osd); 989 __move_osd_to_lru(osdc, req->r_osd); 990 } 991 if (list_empty(&req->r_linger_item)) 992 req->r_osd = NULL; 993 } 994 995 list_del_init(&req->r_req_lru_item); 996 ceph_osdc_put_request(req); 997 998 if (osdc->num_requests == 0) { 999 dout(" no requests, canceling timeout\n"); 1000 __cancel_osd_timeout(osdc); 1001 } 1002 } 1003 1004 /* 1005 * Cancel a previously queued request message 1006 */ 1007 static void __cancel_request(struct ceph_osd_request *req) 1008 { 1009 if (req->r_sent && req->r_osd) { 1010 ceph_msg_revoke(req->r_request); 1011 req->r_sent = 0; 1012 } 1013 } 1014 1015 static void __register_linger_request(struct ceph_osd_client *osdc, 1016 struct ceph_osd_request *req) 1017 { 1018 dout("__register_linger_request %p\n", req); 1019 list_add_tail(&req->r_linger_item, &osdc->req_linger); 1020 if (req->r_osd) 1021 list_add_tail(&req->r_linger_osd, 1022 &req->r_osd->o_linger_requests); 1023 } 1024 1025 static void __unregister_linger_request(struct ceph_osd_client *osdc, 1026 struct ceph_osd_request *req) 1027 { 1028 dout("__unregister_linger_request %p\n", req); 1029 list_del_init(&req->r_linger_item); 1030 if (req->r_osd) { 1031 list_del_init(&req->r_linger_osd); 1032 1033 if (list_empty(&req->r_osd->o_requests) && 1034 list_empty(&req->r_osd->o_linger_requests)) { 1035 dout("moving osd to %p lru\n", req->r_osd); 1036 __move_osd_to_lru(osdc, req->r_osd); 1037 } 1038 if (list_empty(&req->r_osd_item)) 1039 req->r_osd = NULL; 1040 } 1041 } 1042 1043 void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc, 1044 struct ceph_osd_request *req) 1045 { 1046 mutex_lock(&osdc->request_mutex); 1047 if (req->r_linger) { 1048 __unregister_linger_request(osdc, req); 1049 ceph_osdc_put_request(req); 1050 } 1051 mutex_unlock(&osdc->request_mutex); 1052 } 1053 EXPORT_SYMBOL(ceph_osdc_unregister_linger_request); 1054 1055 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc, 1056 struct ceph_osd_request *req) 1057 { 1058 if (!req->r_linger) { 1059 dout("set_request_linger %p\n", req); 1060 req->r_linger = 1; 1061 /* 1062 * caller is now responsible for calling 1063 * unregister_linger_request 1064 */ 1065 ceph_osdc_get_request(req); 1066 } 1067 } 1068 EXPORT_SYMBOL(ceph_osdc_set_request_linger); 1069 1070 /* 1071 * Pick an osd (the first 'up' osd in the pg), allocate the osd struct 1072 * (as needed), and set the request r_osd appropriately. If there is 1073 * no up osd, set r_osd to NULL. Move the request to the appropriate list 1074 * (unsent, homeless) or leave on in-flight lru. 1075 * 1076 * Return 0 if unchanged, 1 if changed, or negative on error. 1077 * 1078 * Caller should hold map_sem for read and request_mutex. 1079 */ 1080 static int __map_request(struct ceph_osd_client *osdc, 1081 struct ceph_osd_request *req, int force_resend) 1082 { 1083 struct ceph_pg pgid; 1084 int acting[CEPH_PG_MAX_SIZE]; 1085 int o = -1, num = 0; 1086 int err; 1087 1088 dout("map_request %p tid %lld\n", req, req->r_tid); 1089 err = ceph_calc_ceph_pg(&pgid, req->r_oid, osdc->osdmap, 1090 ceph_file_layout_pg_pool(req->r_file_layout)); 1091 if (err) { 1092 list_move(&req->r_req_lru_item, &osdc->req_notarget); 1093 return err; 1094 } 1095 req->r_pgid = pgid; 1096 1097 err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting); 1098 if (err > 0) { 1099 o = acting[0]; 1100 num = err; 1101 } 1102 1103 if ((!force_resend && 1104 req->r_osd && req->r_osd->o_osd == o && 1105 req->r_sent >= req->r_osd->o_incarnation && 1106 req->r_num_pg_osds == num && 1107 memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) || 1108 (req->r_osd == NULL && o == -1)) 1109 return 0; /* no change */ 1110 1111 dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n", 1112 req->r_tid, pgid.pool, pgid.seed, o, 1113 req->r_osd ? req->r_osd->o_osd : -1); 1114 1115 /* record full pg acting set */ 1116 memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num); 1117 req->r_num_pg_osds = num; 1118 1119 if (req->r_osd) { 1120 __cancel_request(req); 1121 list_del_init(&req->r_osd_item); 1122 req->r_osd = NULL; 1123 } 1124 1125 req->r_osd = __lookup_osd(osdc, o); 1126 if (!req->r_osd && o >= 0) { 1127 err = -ENOMEM; 1128 req->r_osd = create_osd(osdc, o); 1129 if (!req->r_osd) { 1130 list_move(&req->r_req_lru_item, &osdc->req_notarget); 1131 goto out; 1132 } 1133 1134 dout("map_request osd %p is osd%d\n", req->r_osd, o); 1135 __insert_osd(osdc, req->r_osd); 1136 1137 ceph_con_open(&req->r_osd->o_con, 1138 CEPH_ENTITY_TYPE_OSD, o, 1139 &osdc->osdmap->osd_addr[o]); 1140 } 1141 1142 if (req->r_osd) { 1143 __remove_osd_from_lru(req->r_osd); 1144 list_add_tail(&req->r_osd_item, &req->r_osd->o_requests); 1145 list_move_tail(&req->r_req_lru_item, &osdc->req_unsent); 1146 } else { 1147 list_move_tail(&req->r_req_lru_item, &osdc->req_notarget); 1148 } 1149 err = 1; /* osd or pg changed */ 1150 1151 out: 1152 return err; 1153 } 1154 1155 /* 1156 * caller should hold map_sem (for read) and request_mutex 1157 */ 1158 static void __send_request(struct ceph_osd_client *osdc, 1159 struct ceph_osd_request *req) 1160 { 1161 void *p; 1162 1163 dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n", 1164 req, req->r_tid, req->r_osd->o_osd, req->r_flags, 1165 (unsigned long long)req->r_pgid.pool, req->r_pgid.seed); 1166 1167 /* fill in message content that changes each time we send it */ 1168 put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch); 1169 put_unaligned_le32(req->r_flags, req->r_request_flags); 1170 put_unaligned_le64(req->r_pgid.pool, req->r_request_pool); 1171 p = req->r_request_pgid; 1172 ceph_encode_64(&p, req->r_pgid.pool); 1173 ceph_encode_32(&p, req->r_pgid.seed); 1174 put_unaligned_le64(1, req->r_request_attempts); /* FIXME */ 1175 memcpy(req->r_request_reassert_version, &req->r_reassert_version, 1176 sizeof(req->r_reassert_version)); 1177 1178 req->r_stamp = jiffies; 1179 list_move_tail(&req->r_req_lru_item, &osdc->req_lru); 1180 1181 ceph_msg_get(req->r_request); /* send consumes a ref */ 1182 ceph_con_send(&req->r_osd->o_con, req->r_request); 1183 req->r_sent = req->r_osd->o_incarnation; 1184 } 1185 1186 /* 1187 * Send any requests in the queue (req_unsent). 1188 */ 1189 static void __send_queued(struct ceph_osd_client *osdc) 1190 { 1191 struct ceph_osd_request *req, *tmp; 1192 1193 dout("__send_queued\n"); 1194 list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) 1195 __send_request(osdc, req); 1196 } 1197 1198 /* 1199 * Timeout callback, called every N seconds when 1 or more osd 1200 * requests has been active for more than N seconds. When this 1201 * happens, we ping all OSDs with requests who have timed out to 1202 * ensure any communications channel reset is detected. Reset the 1203 * request timeouts another N seconds in the future as we go. 1204 * Reschedule the timeout event another N seconds in future (unless 1205 * there are no open requests). 1206 */ 1207 static void handle_timeout(struct work_struct *work) 1208 { 1209 struct ceph_osd_client *osdc = 1210 container_of(work, struct ceph_osd_client, timeout_work.work); 1211 struct ceph_osd_request *req; 1212 struct ceph_osd *osd; 1213 unsigned long keepalive = 1214 osdc->client->options->osd_keepalive_timeout * HZ; 1215 struct list_head slow_osds; 1216 dout("timeout\n"); 1217 down_read(&osdc->map_sem); 1218 1219 ceph_monc_request_next_osdmap(&osdc->client->monc); 1220 1221 mutex_lock(&osdc->request_mutex); 1222 1223 /* 1224 * ping osds that are a bit slow. this ensures that if there 1225 * is a break in the TCP connection we will notice, and reopen 1226 * a connection with that osd (from the fault callback). 1227 */ 1228 INIT_LIST_HEAD(&slow_osds); 1229 list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) { 1230 if (time_before(jiffies, req->r_stamp + keepalive)) 1231 break; 1232 1233 osd = req->r_osd; 1234 BUG_ON(!osd); 1235 dout(" tid %llu is slow, will send keepalive on osd%d\n", 1236 req->r_tid, osd->o_osd); 1237 list_move_tail(&osd->o_keepalive_item, &slow_osds); 1238 } 1239 while (!list_empty(&slow_osds)) { 1240 osd = list_entry(slow_osds.next, struct ceph_osd, 1241 o_keepalive_item); 1242 list_del_init(&osd->o_keepalive_item); 1243 ceph_con_keepalive(&osd->o_con); 1244 } 1245 1246 __schedule_osd_timeout(osdc); 1247 __send_queued(osdc); 1248 mutex_unlock(&osdc->request_mutex); 1249 up_read(&osdc->map_sem); 1250 } 1251 1252 static void handle_osds_timeout(struct work_struct *work) 1253 { 1254 struct ceph_osd_client *osdc = 1255 container_of(work, struct ceph_osd_client, 1256 osds_timeout_work.work); 1257 unsigned long delay = 1258 osdc->client->options->osd_idle_ttl * HZ >> 2; 1259 1260 dout("osds timeout\n"); 1261 down_read(&osdc->map_sem); 1262 remove_old_osds(osdc); 1263 up_read(&osdc->map_sem); 1264 1265 schedule_delayed_work(&osdc->osds_timeout_work, 1266 round_jiffies_relative(delay)); 1267 } 1268 1269 static void complete_request(struct ceph_osd_request *req) 1270 { 1271 if (req->r_safe_callback) 1272 req->r_safe_callback(req, NULL); 1273 complete_all(&req->r_safe_completion); /* fsync waiter */ 1274 } 1275 1276 /* 1277 * handle osd op reply. either call the callback if it is specified, 1278 * or do the completion to wake up the waiting thread. 1279 */ 1280 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg, 1281 struct ceph_connection *con) 1282 { 1283 void *p, *end; 1284 struct ceph_osd_request *req; 1285 u64 tid; 1286 int object_len; 1287 int numops, payload_len, flags; 1288 s32 result; 1289 s32 retry_attempt; 1290 struct ceph_pg pg; 1291 int err; 1292 u32 reassert_epoch; 1293 u64 reassert_version; 1294 u32 osdmap_epoch; 1295 int already_completed; 1296 u32 bytes; 1297 int i; 1298 1299 tid = le64_to_cpu(msg->hdr.tid); 1300 dout("handle_reply %p tid %llu\n", msg, tid); 1301 1302 p = msg->front.iov_base; 1303 end = p + msg->front.iov_len; 1304 1305 ceph_decode_need(&p, end, 4, bad); 1306 object_len = ceph_decode_32(&p); 1307 ceph_decode_need(&p, end, object_len, bad); 1308 p += object_len; 1309 1310 err = ceph_decode_pgid(&p, end, &pg); 1311 if (err) 1312 goto bad; 1313 1314 ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad); 1315 flags = ceph_decode_64(&p); 1316 result = ceph_decode_32(&p); 1317 reassert_epoch = ceph_decode_32(&p); 1318 reassert_version = ceph_decode_64(&p); 1319 osdmap_epoch = ceph_decode_32(&p); 1320 1321 /* lookup */ 1322 mutex_lock(&osdc->request_mutex); 1323 req = __lookup_request(osdc, tid); 1324 if (req == NULL) { 1325 dout("handle_reply tid %llu dne\n", tid); 1326 goto bad_mutex; 1327 } 1328 ceph_osdc_get_request(req); 1329 1330 dout("handle_reply %p tid %llu req %p result %d\n", msg, tid, 1331 req, result); 1332 1333 ceph_decode_need(&p, end, 4, bad); 1334 numops = ceph_decode_32(&p); 1335 if (numops > CEPH_OSD_MAX_OP) 1336 goto bad_put; 1337 if (numops != req->r_num_ops) 1338 goto bad_put; 1339 payload_len = 0; 1340 ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad); 1341 for (i = 0; i < numops; i++) { 1342 struct ceph_osd_op *op = p; 1343 int len; 1344 1345 len = le32_to_cpu(op->payload_len); 1346 req->r_reply_op_len[i] = len; 1347 dout(" op %d has %d bytes\n", i, len); 1348 payload_len += len; 1349 p += sizeof(*op); 1350 } 1351 bytes = le32_to_cpu(msg->hdr.data_len); 1352 if (payload_len != bytes) { 1353 pr_warning("sum of op payload lens %d != data_len %d", 1354 payload_len, bytes); 1355 goto bad_put; 1356 } 1357 1358 ceph_decode_need(&p, end, 4 + numops * 4, bad); 1359 retry_attempt = ceph_decode_32(&p); 1360 for (i = 0; i < numops; i++) 1361 req->r_reply_op_result[i] = ceph_decode_32(&p); 1362 1363 if (!req->r_got_reply) { 1364 1365 req->r_result = result; 1366 dout("handle_reply result %d bytes %d\n", req->r_result, 1367 bytes); 1368 if (req->r_result == 0) 1369 req->r_result = bytes; 1370 1371 /* in case this is a write and we need to replay, */ 1372 req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch); 1373 req->r_reassert_version.version = cpu_to_le64(reassert_version); 1374 1375 req->r_got_reply = 1; 1376 } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) { 1377 dout("handle_reply tid %llu dup ack\n", tid); 1378 mutex_unlock(&osdc->request_mutex); 1379 goto done; 1380 } 1381 1382 dout("handle_reply tid %llu flags %d\n", tid, flags); 1383 1384 if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK)) 1385 __register_linger_request(osdc, req); 1386 1387 /* either this is a read, or we got the safe response */ 1388 if (result < 0 || 1389 (flags & CEPH_OSD_FLAG_ONDISK) || 1390 ((flags & CEPH_OSD_FLAG_WRITE) == 0)) 1391 __unregister_request(osdc, req); 1392 1393 already_completed = req->r_completed; 1394 req->r_completed = 1; 1395 mutex_unlock(&osdc->request_mutex); 1396 if (already_completed) 1397 goto done; 1398 1399 if (req->r_callback) 1400 req->r_callback(req, msg); 1401 else 1402 complete_all(&req->r_completion); 1403 1404 if (flags & CEPH_OSD_FLAG_ONDISK) 1405 complete_request(req); 1406 1407 done: 1408 dout("req=%p req->r_linger=%d\n", req, req->r_linger); 1409 ceph_osdc_put_request(req); 1410 return; 1411 1412 bad_put: 1413 ceph_osdc_put_request(req); 1414 bad_mutex: 1415 mutex_unlock(&osdc->request_mutex); 1416 bad: 1417 pr_err("corrupt osd_op_reply got %d %d\n", 1418 (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len)); 1419 ceph_msg_dump(msg); 1420 } 1421 1422 static void reset_changed_osds(struct ceph_osd_client *osdc) 1423 { 1424 struct rb_node *p, *n; 1425 1426 for (p = rb_first(&osdc->osds); p; p = n) { 1427 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node); 1428 1429 n = rb_next(p); 1430 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) || 1431 memcmp(&osd->o_con.peer_addr, 1432 ceph_osd_addr(osdc->osdmap, 1433 osd->o_osd), 1434 sizeof(struct ceph_entity_addr)) != 0) 1435 __reset_osd(osdc, osd); 1436 } 1437 } 1438 1439 /* 1440 * Requeue requests whose mapping to an OSD has changed. If requests map to 1441 * no osd, request a new map. 1442 * 1443 * Caller should hold map_sem for read. 1444 */ 1445 static void kick_requests(struct ceph_osd_client *osdc, int force_resend) 1446 { 1447 struct ceph_osd_request *req, *nreq; 1448 struct rb_node *p; 1449 int needmap = 0; 1450 int err; 1451 1452 dout("kick_requests %s\n", force_resend ? " (force resend)" : ""); 1453 mutex_lock(&osdc->request_mutex); 1454 for (p = rb_first(&osdc->requests); p; ) { 1455 req = rb_entry(p, struct ceph_osd_request, r_node); 1456 p = rb_next(p); 1457 1458 /* 1459 * For linger requests that have not yet been 1460 * registered, move them to the linger list; they'll 1461 * be sent to the osd in the loop below. Unregister 1462 * the request before re-registering it as a linger 1463 * request to ensure the __map_request() below 1464 * will decide it needs to be sent. 1465 */ 1466 if (req->r_linger && list_empty(&req->r_linger_item)) { 1467 dout("%p tid %llu restart on osd%d\n", 1468 req, req->r_tid, 1469 req->r_osd ? req->r_osd->o_osd : -1); 1470 __unregister_request(osdc, req); 1471 __register_linger_request(osdc, req); 1472 continue; 1473 } 1474 1475 err = __map_request(osdc, req, force_resend); 1476 if (err < 0) 1477 continue; /* error */ 1478 if (req->r_osd == NULL) { 1479 dout("%p tid %llu maps to no osd\n", req, req->r_tid); 1480 needmap++; /* request a newer map */ 1481 } else if (err > 0) { 1482 if (!req->r_linger) { 1483 dout("%p tid %llu requeued on osd%d\n", req, 1484 req->r_tid, 1485 req->r_osd ? req->r_osd->o_osd : -1); 1486 req->r_flags |= CEPH_OSD_FLAG_RETRY; 1487 } 1488 } 1489 } 1490 1491 list_for_each_entry_safe(req, nreq, &osdc->req_linger, 1492 r_linger_item) { 1493 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd); 1494 1495 err = __map_request(osdc, req, force_resend); 1496 dout("__map_request returned %d\n", err); 1497 if (err == 0) 1498 continue; /* no change and no osd was specified */ 1499 if (err < 0) 1500 continue; /* hrm! */ 1501 if (req->r_osd == NULL) { 1502 dout("tid %llu maps to no valid osd\n", req->r_tid); 1503 needmap++; /* request a newer map */ 1504 continue; 1505 } 1506 1507 dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid, 1508 req->r_osd ? req->r_osd->o_osd : -1); 1509 __register_request(osdc, req); 1510 __unregister_linger_request(osdc, req); 1511 } 1512 mutex_unlock(&osdc->request_mutex); 1513 1514 if (needmap) { 1515 dout("%d requests for down osds, need new map\n", needmap); 1516 ceph_monc_request_next_osdmap(&osdc->client->monc); 1517 } 1518 reset_changed_osds(osdc); 1519 } 1520 1521 1522 /* 1523 * Process updated osd map. 1524 * 1525 * The message contains any number of incremental and full maps, normally 1526 * indicating some sort of topology change in the cluster. Kick requests 1527 * off to different OSDs as needed. 1528 */ 1529 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg) 1530 { 1531 void *p, *end, *next; 1532 u32 nr_maps, maplen; 1533 u32 epoch; 1534 struct ceph_osdmap *newmap = NULL, *oldmap; 1535 int err; 1536 struct ceph_fsid fsid; 1537 1538 dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0); 1539 p = msg->front.iov_base; 1540 end = p + msg->front.iov_len; 1541 1542 /* verify fsid */ 1543 ceph_decode_need(&p, end, sizeof(fsid), bad); 1544 ceph_decode_copy(&p, &fsid, sizeof(fsid)); 1545 if (ceph_check_fsid(osdc->client, &fsid) < 0) 1546 return; 1547 1548 down_write(&osdc->map_sem); 1549 1550 /* incremental maps */ 1551 ceph_decode_32_safe(&p, end, nr_maps, bad); 1552 dout(" %d inc maps\n", nr_maps); 1553 while (nr_maps > 0) { 1554 ceph_decode_need(&p, end, 2*sizeof(u32), bad); 1555 epoch = ceph_decode_32(&p); 1556 maplen = ceph_decode_32(&p); 1557 ceph_decode_need(&p, end, maplen, bad); 1558 next = p + maplen; 1559 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) { 1560 dout("applying incremental map %u len %d\n", 1561 epoch, maplen); 1562 newmap = osdmap_apply_incremental(&p, next, 1563 osdc->osdmap, 1564 &osdc->client->msgr); 1565 if (IS_ERR(newmap)) { 1566 err = PTR_ERR(newmap); 1567 goto bad; 1568 } 1569 BUG_ON(!newmap); 1570 if (newmap != osdc->osdmap) { 1571 ceph_osdmap_destroy(osdc->osdmap); 1572 osdc->osdmap = newmap; 1573 } 1574 kick_requests(osdc, 0); 1575 } else { 1576 dout("ignoring incremental map %u len %d\n", 1577 epoch, maplen); 1578 } 1579 p = next; 1580 nr_maps--; 1581 } 1582 if (newmap) 1583 goto done; 1584 1585 /* full maps */ 1586 ceph_decode_32_safe(&p, end, nr_maps, bad); 1587 dout(" %d full maps\n", nr_maps); 1588 while (nr_maps) { 1589 ceph_decode_need(&p, end, 2*sizeof(u32), bad); 1590 epoch = ceph_decode_32(&p); 1591 maplen = ceph_decode_32(&p); 1592 ceph_decode_need(&p, end, maplen, bad); 1593 if (nr_maps > 1) { 1594 dout("skipping non-latest full map %u len %d\n", 1595 epoch, maplen); 1596 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) { 1597 dout("skipping full map %u len %d, " 1598 "older than our %u\n", epoch, maplen, 1599 osdc->osdmap->epoch); 1600 } else { 1601 int skipped_map = 0; 1602 1603 dout("taking full map %u len %d\n", epoch, maplen); 1604 newmap = osdmap_decode(&p, p+maplen); 1605 if (IS_ERR(newmap)) { 1606 err = PTR_ERR(newmap); 1607 goto bad; 1608 } 1609 BUG_ON(!newmap); 1610 oldmap = osdc->osdmap; 1611 osdc->osdmap = newmap; 1612 if (oldmap) { 1613 if (oldmap->epoch + 1 < newmap->epoch) 1614 skipped_map = 1; 1615 ceph_osdmap_destroy(oldmap); 1616 } 1617 kick_requests(osdc, skipped_map); 1618 } 1619 p += maplen; 1620 nr_maps--; 1621 } 1622 1623 done: 1624 downgrade_write(&osdc->map_sem); 1625 ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch); 1626 1627 /* 1628 * subscribe to subsequent osdmap updates if full to ensure 1629 * we find out when we are no longer full and stop returning 1630 * ENOSPC. 1631 */ 1632 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL)) 1633 ceph_monc_request_next_osdmap(&osdc->client->monc); 1634 1635 mutex_lock(&osdc->request_mutex); 1636 __send_queued(osdc); 1637 mutex_unlock(&osdc->request_mutex); 1638 up_read(&osdc->map_sem); 1639 wake_up_all(&osdc->client->auth_wq); 1640 return; 1641 1642 bad: 1643 pr_err("osdc handle_map corrupt msg\n"); 1644 ceph_msg_dump(msg); 1645 up_write(&osdc->map_sem); 1646 return; 1647 } 1648 1649 /* 1650 * watch/notify callback event infrastructure 1651 * 1652 * These callbacks are used both for watch and notify operations. 1653 */ 1654 static void __release_event(struct kref *kref) 1655 { 1656 struct ceph_osd_event *event = 1657 container_of(kref, struct ceph_osd_event, kref); 1658 1659 dout("__release_event %p\n", event); 1660 kfree(event); 1661 } 1662 1663 static void get_event(struct ceph_osd_event *event) 1664 { 1665 kref_get(&event->kref); 1666 } 1667 1668 void ceph_osdc_put_event(struct ceph_osd_event *event) 1669 { 1670 kref_put(&event->kref, __release_event); 1671 } 1672 EXPORT_SYMBOL(ceph_osdc_put_event); 1673 1674 static void __insert_event(struct ceph_osd_client *osdc, 1675 struct ceph_osd_event *new) 1676 { 1677 struct rb_node **p = &osdc->event_tree.rb_node; 1678 struct rb_node *parent = NULL; 1679 struct ceph_osd_event *event = NULL; 1680 1681 while (*p) { 1682 parent = *p; 1683 event = rb_entry(parent, struct ceph_osd_event, node); 1684 if (new->cookie < event->cookie) 1685 p = &(*p)->rb_left; 1686 else if (new->cookie > event->cookie) 1687 p = &(*p)->rb_right; 1688 else 1689 BUG(); 1690 } 1691 1692 rb_link_node(&new->node, parent, p); 1693 rb_insert_color(&new->node, &osdc->event_tree); 1694 } 1695 1696 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc, 1697 u64 cookie) 1698 { 1699 struct rb_node **p = &osdc->event_tree.rb_node; 1700 struct rb_node *parent = NULL; 1701 struct ceph_osd_event *event = NULL; 1702 1703 while (*p) { 1704 parent = *p; 1705 event = rb_entry(parent, struct ceph_osd_event, node); 1706 if (cookie < event->cookie) 1707 p = &(*p)->rb_left; 1708 else if (cookie > event->cookie) 1709 p = &(*p)->rb_right; 1710 else 1711 return event; 1712 } 1713 return NULL; 1714 } 1715 1716 static void __remove_event(struct ceph_osd_event *event) 1717 { 1718 struct ceph_osd_client *osdc = event->osdc; 1719 1720 if (!RB_EMPTY_NODE(&event->node)) { 1721 dout("__remove_event removed %p\n", event); 1722 rb_erase(&event->node, &osdc->event_tree); 1723 ceph_osdc_put_event(event); 1724 } else { 1725 dout("__remove_event didn't remove %p\n", event); 1726 } 1727 } 1728 1729 int ceph_osdc_create_event(struct ceph_osd_client *osdc, 1730 void (*event_cb)(u64, u64, u8, void *), 1731 void *data, struct ceph_osd_event **pevent) 1732 { 1733 struct ceph_osd_event *event; 1734 1735 event = kmalloc(sizeof(*event), GFP_NOIO); 1736 if (!event) 1737 return -ENOMEM; 1738 1739 dout("create_event %p\n", event); 1740 event->cb = event_cb; 1741 event->one_shot = 0; 1742 event->data = data; 1743 event->osdc = osdc; 1744 INIT_LIST_HEAD(&event->osd_node); 1745 RB_CLEAR_NODE(&event->node); 1746 kref_init(&event->kref); /* one ref for us */ 1747 kref_get(&event->kref); /* one ref for the caller */ 1748 1749 spin_lock(&osdc->event_lock); 1750 event->cookie = ++osdc->event_count; 1751 __insert_event(osdc, event); 1752 spin_unlock(&osdc->event_lock); 1753 1754 *pevent = event; 1755 return 0; 1756 } 1757 EXPORT_SYMBOL(ceph_osdc_create_event); 1758 1759 void ceph_osdc_cancel_event(struct ceph_osd_event *event) 1760 { 1761 struct ceph_osd_client *osdc = event->osdc; 1762 1763 dout("cancel_event %p\n", event); 1764 spin_lock(&osdc->event_lock); 1765 __remove_event(event); 1766 spin_unlock(&osdc->event_lock); 1767 ceph_osdc_put_event(event); /* caller's */ 1768 } 1769 EXPORT_SYMBOL(ceph_osdc_cancel_event); 1770 1771 1772 static void do_event_work(struct work_struct *work) 1773 { 1774 struct ceph_osd_event_work *event_work = 1775 container_of(work, struct ceph_osd_event_work, work); 1776 struct ceph_osd_event *event = event_work->event; 1777 u64 ver = event_work->ver; 1778 u64 notify_id = event_work->notify_id; 1779 u8 opcode = event_work->opcode; 1780 1781 dout("do_event_work completing %p\n", event); 1782 event->cb(ver, notify_id, opcode, event->data); 1783 dout("do_event_work completed %p\n", event); 1784 ceph_osdc_put_event(event); 1785 kfree(event_work); 1786 } 1787 1788 1789 /* 1790 * Process osd watch notifications 1791 */ 1792 static void handle_watch_notify(struct ceph_osd_client *osdc, 1793 struct ceph_msg *msg) 1794 { 1795 void *p, *end; 1796 u8 proto_ver; 1797 u64 cookie, ver, notify_id; 1798 u8 opcode; 1799 struct ceph_osd_event *event; 1800 struct ceph_osd_event_work *event_work; 1801 1802 p = msg->front.iov_base; 1803 end = p + msg->front.iov_len; 1804 1805 ceph_decode_8_safe(&p, end, proto_ver, bad); 1806 ceph_decode_8_safe(&p, end, opcode, bad); 1807 ceph_decode_64_safe(&p, end, cookie, bad); 1808 ceph_decode_64_safe(&p, end, ver, bad); 1809 ceph_decode_64_safe(&p, end, notify_id, bad); 1810 1811 spin_lock(&osdc->event_lock); 1812 event = __find_event(osdc, cookie); 1813 if (event) { 1814 BUG_ON(event->one_shot); 1815 get_event(event); 1816 } 1817 spin_unlock(&osdc->event_lock); 1818 dout("handle_watch_notify cookie %lld ver %lld event %p\n", 1819 cookie, ver, event); 1820 if (event) { 1821 event_work = kmalloc(sizeof(*event_work), GFP_NOIO); 1822 if (!event_work) { 1823 dout("ERROR: could not allocate event_work\n"); 1824 goto done_err; 1825 } 1826 INIT_WORK(&event_work->work, do_event_work); 1827 event_work->event = event; 1828 event_work->ver = ver; 1829 event_work->notify_id = notify_id; 1830 event_work->opcode = opcode; 1831 if (!queue_work(osdc->notify_wq, &event_work->work)) { 1832 dout("WARNING: failed to queue notify event work\n"); 1833 goto done_err; 1834 } 1835 } 1836 1837 return; 1838 1839 done_err: 1840 ceph_osdc_put_event(event); 1841 return; 1842 1843 bad: 1844 pr_err("osdc handle_watch_notify corrupt msg\n"); 1845 return; 1846 } 1847 1848 static void ceph_osdc_msg_data_set(struct ceph_msg *msg, 1849 struct ceph_osd_data *osd_data) 1850 { 1851 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 1852 BUG_ON(osd_data->length > (u64) SIZE_MAX); 1853 if (osd_data->length) 1854 ceph_msg_data_set_pages(msg, osd_data->pages, 1855 osd_data->length, osd_data->alignment); 1856 } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) { 1857 BUG_ON(!osd_data->pagelist->length); 1858 ceph_msg_data_set_pagelist(msg, osd_data->pagelist); 1859 #ifdef CONFIG_BLOCK 1860 } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) { 1861 ceph_msg_data_set_bio(msg, osd_data->bio, osd_data->bio_length); 1862 #endif 1863 } else { 1864 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE); 1865 } 1866 } 1867 1868 /* 1869 * Register request, send initial attempt. 1870 */ 1871 int ceph_osdc_start_request(struct ceph_osd_client *osdc, 1872 struct ceph_osd_request *req, 1873 bool nofail) 1874 { 1875 int rc = 0; 1876 1877 /* Set up response incoming data and request outgoing data fields */ 1878 1879 ceph_osdc_msg_data_set(req->r_reply, &req->r_data_in); 1880 ceph_osdc_msg_data_set(req->r_request, &req->r_data_out); 1881 1882 down_read(&osdc->map_sem); 1883 mutex_lock(&osdc->request_mutex); 1884 __register_request(osdc, req); 1885 WARN_ON(req->r_sent); 1886 rc = __map_request(osdc, req, 0); 1887 if (rc < 0) { 1888 if (nofail) { 1889 dout("osdc_start_request failed map, " 1890 " will retry %lld\n", req->r_tid); 1891 rc = 0; 1892 } 1893 goto out_unlock; 1894 } 1895 if (req->r_osd == NULL) { 1896 dout("send_request %p no up osds in pg\n", req); 1897 ceph_monc_request_next_osdmap(&osdc->client->monc); 1898 } else { 1899 __send_queued(osdc); 1900 } 1901 rc = 0; 1902 out_unlock: 1903 mutex_unlock(&osdc->request_mutex); 1904 up_read(&osdc->map_sem); 1905 return rc; 1906 } 1907 EXPORT_SYMBOL(ceph_osdc_start_request); 1908 1909 /* 1910 * wait for a request to complete 1911 */ 1912 int ceph_osdc_wait_request(struct ceph_osd_client *osdc, 1913 struct ceph_osd_request *req) 1914 { 1915 int rc; 1916 1917 rc = wait_for_completion_interruptible(&req->r_completion); 1918 if (rc < 0) { 1919 mutex_lock(&osdc->request_mutex); 1920 __cancel_request(req); 1921 __unregister_request(osdc, req); 1922 mutex_unlock(&osdc->request_mutex); 1923 complete_request(req); 1924 dout("wait_request tid %llu canceled/timed out\n", req->r_tid); 1925 return rc; 1926 } 1927 1928 dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result); 1929 return req->r_result; 1930 } 1931 EXPORT_SYMBOL(ceph_osdc_wait_request); 1932 1933 /* 1934 * sync - wait for all in-flight requests to flush. avoid starvation. 1935 */ 1936 void ceph_osdc_sync(struct ceph_osd_client *osdc) 1937 { 1938 struct ceph_osd_request *req; 1939 u64 last_tid, next_tid = 0; 1940 1941 mutex_lock(&osdc->request_mutex); 1942 last_tid = osdc->last_tid; 1943 while (1) { 1944 req = __lookup_request_ge(osdc, next_tid); 1945 if (!req) 1946 break; 1947 if (req->r_tid > last_tid) 1948 break; 1949 1950 next_tid = req->r_tid + 1; 1951 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0) 1952 continue; 1953 1954 ceph_osdc_get_request(req); 1955 mutex_unlock(&osdc->request_mutex); 1956 dout("sync waiting on tid %llu (last is %llu)\n", 1957 req->r_tid, last_tid); 1958 wait_for_completion(&req->r_safe_completion); 1959 mutex_lock(&osdc->request_mutex); 1960 ceph_osdc_put_request(req); 1961 } 1962 mutex_unlock(&osdc->request_mutex); 1963 dout("sync done (thru tid %llu)\n", last_tid); 1964 } 1965 EXPORT_SYMBOL(ceph_osdc_sync); 1966 1967 /* 1968 * init, shutdown 1969 */ 1970 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client) 1971 { 1972 int err; 1973 1974 dout("init\n"); 1975 osdc->client = client; 1976 osdc->osdmap = NULL; 1977 init_rwsem(&osdc->map_sem); 1978 init_completion(&osdc->map_waiters); 1979 osdc->last_requested_map = 0; 1980 mutex_init(&osdc->request_mutex); 1981 osdc->last_tid = 0; 1982 osdc->osds = RB_ROOT; 1983 INIT_LIST_HEAD(&osdc->osd_lru); 1984 osdc->requests = RB_ROOT; 1985 INIT_LIST_HEAD(&osdc->req_lru); 1986 INIT_LIST_HEAD(&osdc->req_unsent); 1987 INIT_LIST_HEAD(&osdc->req_notarget); 1988 INIT_LIST_HEAD(&osdc->req_linger); 1989 osdc->num_requests = 0; 1990 INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout); 1991 INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout); 1992 spin_lock_init(&osdc->event_lock); 1993 osdc->event_tree = RB_ROOT; 1994 osdc->event_count = 0; 1995 1996 schedule_delayed_work(&osdc->osds_timeout_work, 1997 round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ)); 1998 1999 err = -ENOMEM; 2000 osdc->req_mempool = mempool_create_kmalloc_pool(10, 2001 sizeof(struct ceph_osd_request)); 2002 if (!osdc->req_mempool) 2003 goto out; 2004 2005 err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP, 2006 OSD_OP_FRONT_LEN, 10, true, 2007 "osd_op"); 2008 if (err < 0) 2009 goto out_mempool; 2010 err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY, 2011 OSD_OPREPLY_FRONT_LEN, 10, true, 2012 "osd_op_reply"); 2013 if (err < 0) 2014 goto out_msgpool; 2015 2016 osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify"); 2017 if (IS_ERR(osdc->notify_wq)) { 2018 err = PTR_ERR(osdc->notify_wq); 2019 osdc->notify_wq = NULL; 2020 goto out_msgpool; 2021 } 2022 return 0; 2023 2024 out_msgpool: 2025 ceph_msgpool_destroy(&osdc->msgpool_op); 2026 out_mempool: 2027 mempool_destroy(osdc->req_mempool); 2028 out: 2029 return err; 2030 } 2031 2032 void ceph_osdc_stop(struct ceph_osd_client *osdc) 2033 { 2034 flush_workqueue(osdc->notify_wq); 2035 destroy_workqueue(osdc->notify_wq); 2036 cancel_delayed_work_sync(&osdc->timeout_work); 2037 cancel_delayed_work_sync(&osdc->osds_timeout_work); 2038 if (osdc->osdmap) { 2039 ceph_osdmap_destroy(osdc->osdmap); 2040 osdc->osdmap = NULL; 2041 } 2042 remove_all_osds(osdc); 2043 mempool_destroy(osdc->req_mempool); 2044 ceph_msgpool_destroy(&osdc->msgpool_op); 2045 ceph_msgpool_destroy(&osdc->msgpool_op_reply); 2046 } 2047 2048 /* 2049 * Read some contiguous pages. If we cross a stripe boundary, shorten 2050 * *plen. Return number of bytes read, or error. 2051 */ 2052 int ceph_osdc_readpages(struct ceph_osd_client *osdc, 2053 struct ceph_vino vino, struct ceph_file_layout *layout, 2054 u64 off, u64 *plen, 2055 u32 truncate_seq, u64 truncate_size, 2056 struct page **pages, int num_pages, int page_align) 2057 { 2058 struct ceph_osd_request *req; 2059 struct ceph_osd_data *osd_data; 2060 struct ceph_osd_req_op op; 2061 int rc = 0; 2062 2063 dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino, 2064 vino.snap, off, *plen); 2065 req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 1, &op, 2066 CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ, 2067 NULL, truncate_seq, truncate_size, 2068 false); 2069 if (IS_ERR(req)) 2070 return PTR_ERR(req); 2071 2072 /* it may be a short read due to an object boundary */ 2073 2074 osd_data = &req->r_data_in; 2075 osd_data->type = CEPH_OSD_DATA_TYPE_PAGES; 2076 osd_data->pages = pages; 2077 osd_data->length = *plen; 2078 osd_data->alignment = page_align; 2079 2080 dout("readpages final extent is %llu~%llu (%llu bytes align %d)\n", 2081 off, *plen, osd_data->length, page_align); 2082 2083 ceph_osdc_build_request(req, off, 1, &op, NULL, vino.snap, NULL); 2084 2085 rc = ceph_osdc_start_request(osdc, req, false); 2086 if (!rc) 2087 rc = ceph_osdc_wait_request(osdc, req); 2088 2089 ceph_osdc_put_request(req); 2090 dout("readpages result %d\n", rc); 2091 return rc; 2092 } 2093 EXPORT_SYMBOL(ceph_osdc_readpages); 2094 2095 /* 2096 * do a synchronous write on N pages 2097 */ 2098 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino, 2099 struct ceph_file_layout *layout, 2100 struct ceph_snap_context *snapc, 2101 u64 off, u64 len, 2102 u32 truncate_seq, u64 truncate_size, 2103 struct timespec *mtime, 2104 struct page **pages, int num_pages) 2105 { 2106 struct ceph_osd_request *req; 2107 struct ceph_osd_data *osd_data; 2108 struct ceph_osd_req_op op; 2109 int rc = 0; 2110 int page_align = off & ~PAGE_MASK; 2111 2112 BUG_ON(vino.snap != CEPH_NOSNAP); /* snapshots aren't writeable */ 2113 req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 1, &op, 2114 CEPH_OSD_OP_WRITE, 2115 CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE, 2116 snapc, truncate_seq, truncate_size, 2117 true); 2118 if (IS_ERR(req)) 2119 return PTR_ERR(req); 2120 2121 /* it may be a short write due to an object boundary */ 2122 osd_data = &req->r_data_out; 2123 osd_data->type = CEPH_OSD_DATA_TYPE_PAGES; 2124 osd_data->pages = pages; 2125 osd_data->length = len; 2126 osd_data->alignment = page_align; 2127 dout("writepages %llu~%llu (%llu bytes)\n", off, len, osd_data->length); 2128 2129 ceph_osdc_build_request(req, off, 1, &op, snapc, CEPH_NOSNAP, mtime); 2130 2131 rc = ceph_osdc_start_request(osdc, req, true); 2132 if (!rc) 2133 rc = ceph_osdc_wait_request(osdc, req); 2134 2135 ceph_osdc_put_request(req); 2136 if (rc == 0) 2137 rc = len; 2138 dout("writepages result %d\n", rc); 2139 return rc; 2140 } 2141 EXPORT_SYMBOL(ceph_osdc_writepages); 2142 2143 /* 2144 * handle incoming message 2145 */ 2146 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg) 2147 { 2148 struct ceph_osd *osd = con->private; 2149 struct ceph_osd_client *osdc; 2150 int type = le16_to_cpu(msg->hdr.type); 2151 2152 if (!osd) 2153 goto out; 2154 osdc = osd->o_osdc; 2155 2156 switch (type) { 2157 case CEPH_MSG_OSD_MAP: 2158 ceph_osdc_handle_map(osdc, msg); 2159 break; 2160 case CEPH_MSG_OSD_OPREPLY: 2161 handle_reply(osdc, msg, con); 2162 break; 2163 case CEPH_MSG_WATCH_NOTIFY: 2164 handle_watch_notify(osdc, msg); 2165 break; 2166 2167 default: 2168 pr_err("received unknown message type %d %s\n", type, 2169 ceph_msg_type_name(type)); 2170 } 2171 out: 2172 ceph_msg_put(msg); 2173 } 2174 2175 /* 2176 * lookup and return message for incoming reply. set up reply message 2177 * pages. 2178 */ 2179 static struct ceph_msg *get_reply(struct ceph_connection *con, 2180 struct ceph_msg_header *hdr, 2181 int *skip) 2182 { 2183 struct ceph_osd *osd = con->private; 2184 struct ceph_osd_client *osdc = osd->o_osdc; 2185 struct ceph_msg *m; 2186 struct ceph_osd_request *req; 2187 int front = le32_to_cpu(hdr->front_len); 2188 int data_len = le32_to_cpu(hdr->data_len); 2189 u64 tid; 2190 2191 tid = le64_to_cpu(hdr->tid); 2192 mutex_lock(&osdc->request_mutex); 2193 req = __lookup_request(osdc, tid); 2194 if (!req) { 2195 *skip = 1; 2196 m = NULL; 2197 dout("get_reply unknown tid %llu from osd%d\n", tid, 2198 osd->o_osd); 2199 goto out; 2200 } 2201 2202 if (req->r_reply->con) 2203 dout("%s revoking msg %p from old con %p\n", __func__, 2204 req->r_reply, req->r_reply->con); 2205 ceph_msg_revoke_incoming(req->r_reply); 2206 2207 if (front > req->r_reply->front.iov_len) { 2208 pr_warning("get_reply front %d > preallocated %d\n", 2209 front, (int)req->r_reply->front.iov_len); 2210 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS, false); 2211 if (!m) 2212 goto out; 2213 ceph_msg_put(req->r_reply); 2214 req->r_reply = m; 2215 } 2216 m = ceph_msg_get(req->r_reply); 2217 2218 if (data_len > 0) { 2219 struct ceph_osd_data *osd_data = &req->r_data_in; 2220 2221 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 2222 if (osd_data->pages && 2223 unlikely(osd_data->length < data_len)) { 2224 2225 pr_warning("tid %lld reply has %d bytes " 2226 "we had only %llu bytes ready\n", 2227 tid, data_len, osd_data->length); 2228 *skip = 1; 2229 ceph_msg_put(m); 2230 m = NULL; 2231 goto out; 2232 } 2233 } 2234 } 2235 *skip = 0; 2236 dout("get_reply tid %lld %p\n", tid, m); 2237 2238 out: 2239 mutex_unlock(&osdc->request_mutex); 2240 return m; 2241 2242 } 2243 2244 static struct ceph_msg *alloc_msg(struct ceph_connection *con, 2245 struct ceph_msg_header *hdr, 2246 int *skip) 2247 { 2248 struct ceph_osd *osd = con->private; 2249 int type = le16_to_cpu(hdr->type); 2250 int front = le32_to_cpu(hdr->front_len); 2251 2252 *skip = 0; 2253 switch (type) { 2254 case CEPH_MSG_OSD_MAP: 2255 case CEPH_MSG_WATCH_NOTIFY: 2256 return ceph_msg_new(type, front, GFP_NOFS, false); 2257 case CEPH_MSG_OSD_OPREPLY: 2258 return get_reply(con, hdr, skip); 2259 default: 2260 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type, 2261 osd->o_osd); 2262 *skip = 1; 2263 return NULL; 2264 } 2265 } 2266 2267 /* 2268 * Wrappers to refcount containing ceph_osd struct 2269 */ 2270 static struct ceph_connection *get_osd_con(struct ceph_connection *con) 2271 { 2272 struct ceph_osd *osd = con->private; 2273 if (get_osd(osd)) 2274 return con; 2275 return NULL; 2276 } 2277 2278 static void put_osd_con(struct ceph_connection *con) 2279 { 2280 struct ceph_osd *osd = con->private; 2281 put_osd(osd); 2282 } 2283 2284 /* 2285 * authentication 2286 */ 2287 /* 2288 * Note: returned pointer is the address of a structure that's 2289 * managed separately. Caller must *not* attempt to free it. 2290 */ 2291 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con, 2292 int *proto, int force_new) 2293 { 2294 struct ceph_osd *o = con->private; 2295 struct ceph_osd_client *osdc = o->o_osdc; 2296 struct ceph_auth_client *ac = osdc->client->monc.auth; 2297 struct ceph_auth_handshake *auth = &o->o_auth; 2298 2299 if (force_new && auth->authorizer) { 2300 ceph_auth_destroy_authorizer(ac, auth->authorizer); 2301 auth->authorizer = NULL; 2302 } 2303 if (!auth->authorizer) { 2304 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 2305 auth); 2306 if (ret) 2307 return ERR_PTR(ret); 2308 } else { 2309 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 2310 auth); 2311 if (ret) 2312 return ERR_PTR(ret); 2313 } 2314 *proto = ac->protocol; 2315 2316 return auth; 2317 } 2318 2319 2320 static int verify_authorizer_reply(struct ceph_connection *con, int len) 2321 { 2322 struct ceph_osd *o = con->private; 2323 struct ceph_osd_client *osdc = o->o_osdc; 2324 struct ceph_auth_client *ac = osdc->client->monc.auth; 2325 2326 return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len); 2327 } 2328 2329 static int invalidate_authorizer(struct ceph_connection *con) 2330 { 2331 struct ceph_osd *o = con->private; 2332 struct ceph_osd_client *osdc = o->o_osdc; 2333 struct ceph_auth_client *ac = osdc->client->monc.auth; 2334 2335 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD); 2336 return ceph_monc_validate_auth(&osdc->client->monc); 2337 } 2338 2339 static const struct ceph_connection_operations osd_con_ops = { 2340 .get = get_osd_con, 2341 .put = put_osd_con, 2342 .dispatch = dispatch, 2343 .get_authorizer = get_authorizer, 2344 .verify_authorizer_reply = verify_authorizer_reply, 2345 .invalidate_authorizer = invalidate_authorizer, 2346 .alloc_msg = alloc_msg, 2347 .fault = osd_reset, 2348 }; 2349