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 int __send_request(struct ceph_osd_client *osdc, 33 struct ceph_osd_request *req); 34 35 static int op_needs_trail(int op) 36 { 37 switch (op) { 38 case CEPH_OSD_OP_GETXATTR: 39 case CEPH_OSD_OP_SETXATTR: 40 case CEPH_OSD_OP_CMPXATTR: 41 case CEPH_OSD_OP_CALL: 42 case CEPH_OSD_OP_NOTIFY: 43 return 1; 44 default: 45 return 0; 46 } 47 } 48 49 static int op_has_extent(int op) 50 { 51 return (op == CEPH_OSD_OP_READ || 52 op == CEPH_OSD_OP_WRITE); 53 } 54 55 void ceph_calc_raw_layout(struct ceph_osd_client *osdc, 56 struct ceph_file_layout *layout, 57 u64 snapid, 58 u64 off, u64 *plen, u64 *bno, 59 struct ceph_osd_request *req, 60 struct ceph_osd_req_op *op) 61 { 62 struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base; 63 u64 orig_len = *plen; 64 u64 objoff, objlen; /* extent in object */ 65 66 reqhead->snapid = cpu_to_le64(snapid); 67 68 /* object extent? */ 69 ceph_calc_file_object_mapping(layout, off, plen, bno, 70 &objoff, &objlen); 71 if (*plen < orig_len) 72 dout(" skipping last %llu, final file extent %llu~%llu\n", 73 orig_len - *plen, off, *plen); 74 75 if (op_has_extent(op->op)) { 76 op->extent.offset = objoff; 77 op->extent.length = objlen; 78 } 79 req->r_num_pages = calc_pages_for(off, *plen); 80 req->r_page_alignment = off & ~PAGE_MASK; 81 if (op->op == CEPH_OSD_OP_WRITE) 82 op->payload_len = *plen; 83 84 dout("calc_layout bno=%llx %llu~%llu (%d pages)\n", 85 *bno, objoff, objlen, req->r_num_pages); 86 87 } 88 EXPORT_SYMBOL(ceph_calc_raw_layout); 89 90 /* 91 * Implement client access to distributed object storage cluster. 92 * 93 * All data objects are stored within a cluster/cloud of OSDs, or 94 * "object storage devices." (Note that Ceph OSDs have _nothing_ to 95 * do with the T10 OSD extensions to SCSI.) Ceph OSDs are simply 96 * remote daemons serving up and coordinating consistent and safe 97 * access to storage. 98 * 99 * Cluster membership and the mapping of data objects onto storage devices 100 * are described by the osd map. 101 * 102 * We keep track of pending OSD requests (read, write), resubmit 103 * requests to different OSDs when the cluster topology/data layout 104 * change, or retry the affected requests when the communications 105 * channel with an OSD is reset. 106 */ 107 108 /* 109 * calculate the mapping of a file extent onto an object, and fill out the 110 * request accordingly. shorten extent as necessary if it crosses an 111 * object boundary. 112 * 113 * fill osd op in request message. 114 */ 115 static void calc_layout(struct ceph_osd_client *osdc, 116 struct ceph_vino vino, 117 struct ceph_file_layout *layout, 118 u64 off, u64 *plen, 119 struct ceph_osd_request *req, 120 struct ceph_osd_req_op *op) 121 { 122 u64 bno; 123 124 ceph_calc_raw_layout(osdc, layout, vino.snap, off, 125 plen, &bno, req, op); 126 127 snprintf(req->r_oid, sizeof(req->r_oid), "%llx.%08llx", vino.ino, bno); 128 req->r_oid_len = strlen(req->r_oid); 129 } 130 131 /* 132 * requests 133 */ 134 void ceph_osdc_release_request(struct kref *kref) 135 { 136 struct ceph_osd_request *req = container_of(kref, 137 struct ceph_osd_request, 138 r_kref); 139 140 if (req->r_request) 141 ceph_msg_put(req->r_request); 142 if (req->r_reply) 143 ceph_msg_put(req->r_reply); 144 if (req->r_con_filling_msg) { 145 dout("release_request revoking pages %p from con %p\n", 146 req->r_pages, req->r_con_filling_msg); 147 ceph_con_revoke_message(req->r_con_filling_msg, 148 req->r_reply); 149 ceph_con_put(req->r_con_filling_msg); 150 } 151 if (req->r_own_pages) 152 ceph_release_page_vector(req->r_pages, 153 req->r_num_pages); 154 #ifdef CONFIG_BLOCK 155 if (req->r_bio) 156 bio_put(req->r_bio); 157 #endif 158 ceph_put_snap_context(req->r_snapc); 159 if (req->r_trail) { 160 ceph_pagelist_release(req->r_trail); 161 kfree(req->r_trail); 162 } 163 if (req->r_mempool) 164 mempool_free(req, req->r_osdc->req_mempool); 165 else 166 kfree(req); 167 } 168 EXPORT_SYMBOL(ceph_osdc_release_request); 169 170 static int get_num_ops(struct ceph_osd_req_op *ops, int *needs_trail) 171 { 172 int i = 0; 173 174 if (needs_trail) 175 *needs_trail = 0; 176 while (ops[i].op) { 177 if (needs_trail && op_needs_trail(ops[i].op)) 178 *needs_trail = 1; 179 i++; 180 } 181 182 return i; 183 } 184 185 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc, 186 int flags, 187 struct ceph_snap_context *snapc, 188 struct ceph_osd_req_op *ops, 189 bool use_mempool, 190 gfp_t gfp_flags, 191 struct page **pages, 192 struct bio *bio) 193 { 194 struct ceph_osd_request *req; 195 struct ceph_msg *msg; 196 int needs_trail; 197 int num_op = get_num_ops(ops, &needs_trail); 198 size_t msg_size = sizeof(struct ceph_osd_request_head); 199 200 msg_size += num_op*sizeof(struct ceph_osd_op); 201 202 if (use_mempool) { 203 req = mempool_alloc(osdc->req_mempool, gfp_flags); 204 memset(req, 0, sizeof(*req)); 205 } else { 206 req = kzalloc(sizeof(*req), gfp_flags); 207 } 208 if (req == NULL) 209 return NULL; 210 211 req->r_osdc = osdc; 212 req->r_mempool = use_mempool; 213 214 kref_init(&req->r_kref); 215 init_completion(&req->r_completion); 216 init_completion(&req->r_safe_completion); 217 INIT_LIST_HEAD(&req->r_unsafe_item); 218 INIT_LIST_HEAD(&req->r_linger_item); 219 INIT_LIST_HEAD(&req->r_linger_osd); 220 req->r_flags = flags; 221 222 WARN_ON((flags & (CEPH_OSD_FLAG_READ|CEPH_OSD_FLAG_WRITE)) == 0); 223 224 /* create reply message */ 225 if (use_mempool) 226 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0); 227 else 228 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, 229 OSD_OPREPLY_FRONT_LEN, gfp_flags); 230 if (!msg) { 231 ceph_osdc_put_request(req); 232 return NULL; 233 } 234 req->r_reply = msg; 235 236 /* allocate space for the trailing data */ 237 if (needs_trail) { 238 req->r_trail = kmalloc(sizeof(struct ceph_pagelist), gfp_flags); 239 if (!req->r_trail) { 240 ceph_osdc_put_request(req); 241 return NULL; 242 } 243 ceph_pagelist_init(req->r_trail); 244 } 245 /* create request message; allow space for oid */ 246 msg_size += 40; 247 if (snapc) 248 msg_size += sizeof(u64) * snapc->num_snaps; 249 if (use_mempool) 250 msg = ceph_msgpool_get(&osdc->msgpool_op, 0); 251 else 252 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags); 253 if (!msg) { 254 ceph_osdc_put_request(req); 255 return NULL; 256 } 257 258 msg->hdr.type = cpu_to_le16(CEPH_MSG_OSD_OP); 259 memset(msg->front.iov_base, 0, msg->front.iov_len); 260 261 req->r_request = msg; 262 req->r_pages = pages; 263 #ifdef CONFIG_BLOCK 264 if (bio) { 265 req->r_bio = bio; 266 bio_get(req->r_bio); 267 } 268 #endif 269 270 return req; 271 } 272 EXPORT_SYMBOL(ceph_osdc_alloc_request); 273 274 static void osd_req_encode_op(struct ceph_osd_request *req, 275 struct ceph_osd_op *dst, 276 struct ceph_osd_req_op *src) 277 { 278 dst->op = cpu_to_le16(src->op); 279 280 switch (dst->op) { 281 case CEPH_OSD_OP_READ: 282 case CEPH_OSD_OP_WRITE: 283 dst->extent.offset = 284 cpu_to_le64(src->extent.offset); 285 dst->extent.length = 286 cpu_to_le64(src->extent.length); 287 dst->extent.truncate_size = 288 cpu_to_le64(src->extent.truncate_size); 289 dst->extent.truncate_seq = 290 cpu_to_le32(src->extent.truncate_seq); 291 break; 292 293 case CEPH_OSD_OP_GETXATTR: 294 case CEPH_OSD_OP_SETXATTR: 295 case CEPH_OSD_OP_CMPXATTR: 296 BUG_ON(!req->r_trail); 297 298 dst->xattr.name_len = cpu_to_le32(src->xattr.name_len); 299 dst->xattr.value_len = cpu_to_le32(src->xattr.value_len); 300 dst->xattr.cmp_op = src->xattr.cmp_op; 301 dst->xattr.cmp_mode = src->xattr.cmp_mode; 302 ceph_pagelist_append(req->r_trail, src->xattr.name, 303 src->xattr.name_len); 304 ceph_pagelist_append(req->r_trail, src->xattr.val, 305 src->xattr.value_len); 306 break; 307 case CEPH_OSD_OP_CALL: 308 BUG_ON(!req->r_trail); 309 310 dst->cls.class_len = src->cls.class_len; 311 dst->cls.method_len = src->cls.method_len; 312 dst->cls.indata_len = cpu_to_le32(src->cls.indata_len); 313 314 ceph_pagelist_append(req->r_trail, src->cls.class_name, 315 src->cls.class_len); 316 ceph_pagelist_append(req->r_trail, src->cls.method_name, 317 src->cls.method_len); 318 ceph_pagelist_append(req->r_trail, src->cls.indata, 319 src->cls.indata_len); 320 break; 321 case CEPH_OSD_OP_ROLLBACK: 322 dst->snap.snapid = cpu_to_le64(src->snap.snapid); 323 break; 324 case CEPH_OSD_OP_STARTSYNC: 325 break; 326 case CEPH_OSD_OP_NOTIFY: 327 { 328 __le32 prot_ver = cpu_to_le32(src->watch.prot_ver); 329 __le32 timeout = cpu_to_le32(src->watch.timeout); 330 331 BUG_ON(!req->r_trail); 332 333 ceph_pagelist_append(req->r_trail, 334 &prot_ver, sizeof(prot_ver)); 335 ceph_pagelist_append(req->r_trail, 336 &timeout, sizeof(timeout)); 337 } 338 case CEPH_OSD_OP_NOTIFY_ACK: 339 case CEPH_OSD_OP_WATCH: 340 dst->watch.cookie = cpu_to_le64(src->watch.cookie); 341 dst->watch.ver = cpu_to_le64(src->watch.ver); 342 dst->watch.flag = src->watch.flag; 343 break; 344 default: 345 pr_err("unrecognized osd opcode %d\n", dst->op); 346 WARN_ON(1); 347 break; 348 } 349 dst->payload_len = cpu_to_le32(src->payload_len); 350 } 351 352 /* 353 * build new request AND message 354 * 355 */ 356 void ceph_osdc_build_request(struct ceph_osd_request *req, 357 u64 off, u64 *plen, 358 struct ceph_osd_req_op *src_ops, 359 struct ceph_snap_context *snapc, 360 struct timespec *mtime, 361 const char *oid, 362 int oid_len) 363 { 364 struct ceph_msg *msg = req->r_request; 365 struct ceph_osd_request_head *head; 366 struct ceph_osd_req_op *src_op; 367 struct ceph_osd_op *op; 368 void *p; 369 int num_op = get_num_ops(src_ops, NULL); 370 size_t msg_size = sizeof(*head) + num_op*sizeof(*op); 371 int flags = req->r_flags; 372 u64 data_len = 0; 373 int i; 374 375 head = msg->front.iov_base; 376 op = (void *)(head + 1); 377 p = (void *)(op + num_op); 378 379 req->r_snapc = ceph_get_snap_context(snapc); 380 381 head->client_inc = cpu_to_le32(1); /* always, for now. */ 382 head->flags = cpu_to_le32(flags); 383 if (flags & CEPH_OSD_FLAG_WRITE) 384 ceph_encode_timespec(&head->mtime, mtime); 385 head->num_ops = cpu_to_le16(num_op); 386 387 388 /* fill in oid */ 389 head->object_len = cpu_to_le32(oid_len); 390 memcpy(p, oid, oid_len); 391 p += oid_len; 392 393 src_op = src_ops; 394 while (src_op->op) { 395 osd_req_encode_op(req, op, src_op); 396 src_op++; 397 op++; 398 } 399 400 if (req->r_trail) 401 data_len += req->r_trail->length; 402 403 if (snapc) { 404 head->snap_seq = cpu_to_le64(snapc->seq); 405 head->num_snaps = cpu_to_le32(snapc->num_snaps); 406 for (i = 0; i < snapc->num_snaps; i++) { 407 put_unaligned_le64(snapc->snaps[i], p); 408 p += sizeof(u64); 409 } 410 } 411 412 if (flags & CEPH_OSD_FLAG_WRITE) { 413 req->r_request->hdr.data_off = cpu_to_le16(off); 414 req->r_request->hdr.data_len = cpu_to_le32(*plen + data_len); 415 } else if (data_len) { 416 req->r_request->hdr.data_off = 0; 417 req->r_request->hdr.data_len = cpu_to_le32(data_len); 418 } 419 420 req->r_request->page_alignment = req->r_page_alignment; 421 422 BUG_ON(p > msg->front.iov_base + msg->front.iov_len); 423 msg_size = p - msg->front.iov_base; 424 msg->front.iov_len = msg_size; 425 msg->hdr.front_len = cpu_to_le32(msg_size); 426 return; 427 } 428 EXPORT_SYMBOL(ceph_osdc_build_request); 429 430 /* 431 * build new request AND message, calculate layout, and adjust file 432 * extent as needed. 433 * 434 * if the file was recently truncated, we include information about its 435 * old and new size so that the object can be updated appropriately. (we 436 * avoid synchronously deleting truncated objects because it's slow.) 437 * 438 * if @do_sync, include a 'startsync' command so that the osd will flush 439 * data quickly. 440 */ 441 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc, 442 struct ceph_file_layout *layout, 443 struct ceph_vino vino, 444 u64 off, u64 *plen, 445 int opcode, int flags, 446 struct ceph_snap_context *snapc, 447 int do_sync, 448 u32 truncate_seq, 449 u64 truncate_size, 450 struct timespec *mtime, 451 bool use_mempool, int num_reply, 452 int page_align) 453 { 454 struct ceph_osd_req_op ops[3]; 455 struct ceph_osd_request *req; 456 457 ops[0].op = opcode; 458 ops[0].extent.truncate_seq = truncate_seq; 459 ops[0].extent.truncate_size = truncate_size; 460 ops[0].payload_len = 0; 461 462 if (do_sync) { 463 ops[1].op = CEPH_OSD_OP_STARTSYNC; 464 ops[1].payload_len = 0; 465 ops[2].op = 0; 466 } else 467 ops[1].op = 0; 468 469 req = ceph_osdc_alloc_request(osdc, flags, 470 snapc, ops, 471 use_mempool, 472 GFP_NOFS, NULL, NULL); 473 if (!req) 474 return NULL; 475 476 /* calculate max write size */ 477 calc_layout(osdc, vino, layout, off, plen, req, ops); 478 req->r_file_layout = *layout; /* keep a copy */ 479 480 /* in case it differs from natural alignment that calc_layout 481 filled in for us */ 482 req->r_page_alignment = page_align; 483 484 ceph_osdc_build_request(req, off, plen, ops, 485 snapc, 486 mtime, 487 req->r_oid, req->r_oid_len); 488 489 return req; 490 } 491 EXPORT_SYMBOL(ceph_osdc_new_request); 492 493 /* 494 * We keep osd requests in an rbtree, sorted by ->r_tid. 495 */ 496 static void __insert_request(struct ceph_osd_client *osdc, 497 struct ceph_osd_request *new) 498 { 499 struct rb_node **p = &osdc->requests.rb_node; 500 struct rb_node *parent = NULL; 501 struct ceph_osd_request *req = NULL; 502 503 while (*p) { 504 parent = *p; 505 req = rb_entry(parent, struct ceph_osd_request, r_node); 506 if (new->r_tid < req->r_tid) 507 p = &(*p)->rb_left; 508 else if (new->r_tid > req->r_tid) 509 p = &(*p)->rb_right; 510 else 511 BUG(); 512 } 513 514 rb_link_node(&new->r_node, parent, p); 515 rb_insert_color(&new->r_node, &osdc->requests); 516 } 517 518 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc, 519 u64 tid) 520 { 521 struct ceph_osd_request *req; 522 struct rb_node *n = osdc->requests.rb_node; 523 524 while (n) { 525 req = rb_entry(n, struct ceph_osd_request, r_node); 526 if (tid < req->r_tid) 527 n = n->rb_left; 528 else if (tid > req->r_tid) 529 n = n->rb_right; 530 else 531 return req; 532 } 533 return NULL; 534 } 535 536 static struct ceph_osd_request * 537 __lookup_request_ge(struct ceph_osd_client *osdc, 538 u64 tid) 539 { 540 struct ceph_osd_request *req; 541 struct rb_node *n = osdc->requests.rb_node; 542 543 while (n) { 544 req = rb_entry(n, struct ceph_osd_request, r_node); 545 if (tid < req->r_tid) { 546 if (!n->rb_left) 547 return req; 548 n = n->rb_left; 549 } else if (tid > req->r_tid) { 550 n = n->rb_right; 551 } else { 552 return req; 553 } 554 } 555 return NULL; 556 } 557 558 /* 559 * Resubmit requests pending on the given osd. 560 */ 561 static void __kick_osd_requests(struct ceph_osd_client *osdc, 562 struct ceph_osd *osd) 563 { 564 struct ceph_osd_request *req, *nreq; 565 int err; 566 567 dout("__kick_osd_requests osd%d\n", osd->o_osd); 568 err = __reset_osd(osdc, osd); 569 if (err == -EAGAIN) 570 return; 571 572 list_for_each_entry(req, &osd->o_requests, r_osd_item) { 573 list_move(&req->r_req_lru_item, &osdc->req_unsent); 574 dout("requeued %p tid %llu osd%d\n", req, req->r_tid, 575 osd->o_osd); 576 if (!req->r_linger) 577 req->r_flags |= CEPH_OSD_FLAG_RETRY; 578 } 579 580 list_for_each_entry_safe(req, nreq, &osd->o_linger_requests, 581 r_linger_osd) { 582 /* 583 * reregister request prior to unregistering linger so 584 * that r_osd is preserved. 585 */ 586 BUG_ON(!list_empty(&req->r_req_lru_item)); 587 __register_request(osdc, req); 588 list_add(&req->r_req_lru_item, &osdc->req_unsent); 589 list_add(&req->r_osd_item, &req->r_osd->o_requests); 590 __unregister_linger_request(osdc, req); 591 dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid, 592 osd->o_osd); 593 } 594 } 595 596 static void kick_osd_requests(struct ceph_osd_client *osdc, 597 struct ceph_osd *kickosd) 598 { 599 mutex_lock(&osdc->request_mutex); 600 __kick_osd_requests(osdc, kickosd); 601 mutex_unlock(&osdc->request_mutex); 602 } 603 604 /* 605 * If the osd connection drops, we need to resubmit all requests. 606 */ 607 static void osd_reset(struct ceph_connection *con) 608 { 609 struct ceph_osd *osd = con->private; 610 struct ceph_osd_client *osdc; 611 612 if (!osd) 613 return; 614 dout("osd_reset osd%d\n", osd->o_osd); 615 osdc = osd->o_osdc; 616 down_read(&osdc->map_sem); 617 kick_osd_requests(osdc, osd); 618 send_queued(osdc); 619 up_read(&osdc->map_sem); 620 } 621 622 /* 623 * Track open sessions with osds. 624 */ 625 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc) 626 { 627 struct ceph_osd *osd; 628 629 osd = kzalloc(sizeof(*osd), GFP_NOFS); 630 if (!osd) 631 return NULL; 632 633 atomic_set(&osd->o_ref, 1); 634 osd->o_osdc = osdc; 635 INIT_LIST_HEAD(&osd->o_requests); 636 INIT_LIST_HEAD(&osd->o_linger_requests); 637 INIT_LIST_HEAD(&osd->o_osd_lru); 638 osd->o_incarnation = 1; 639 640 ceph_con_init(osdc->client->msgr, &osd->o_con); 641 osd->o_con.private = osd; 642 osd->o_con.ops = &osd_con_ops; 643 osd->o_con.peer_name.type = CEPH_ENTITY_TYPE_OSD; 644 645 INIT_LIST_HEAD(&osd->o_keepalive_item); 646 return osd; 647 } 648 649 static struct ceph_osd *get_osd(struct ceph_osd *osd) 650 { 651 if (atomic_inc_not_zero(&osd->o_ref)) { 652 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1, 653 atomic_read(&osd->o_ref)); 654 return osd; 655 } else { 656 dout("get_osd %p FAIL\n", osd); 657 return NULL; 658 } 659 } 660 661 static void put_osd(struct ceph_osd *osd) 662 { 663 dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref), 664 atomic_read(&osd->o_ref) - 1); 665 if (atomic_dec_and_test(&osd->o_ref)) { 666 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth; 667 668 if (osd->o_authorizer) 669 ac->ops->destroy_authorizer(ac, osd->o_authorizer); 670 kfree(osd); 671 } 672 } 673 674 /* 675 * remove an osd from our map 676 */ 677 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 678 { 679 dout("__remove_osd %p\n", osd); 680 BUG_ON(!list_empty(&osd->o_requests)); 681 rb_erase(&osd->o_node, &osdc->osds); 682 list_del_init(&osd->o_osd_lru); 683 ceph_con_close(&osd->o_con); 684 put_osd(osd); 685 } 686 687 static void __move_osd_to_lru(struct ceph_osd_client *osdc, 688 struct ceph_osd *osd) 689 { 690 dout("__move_osd_to_lru %p\n", osd); 691 BUG_ON(!list_empty(&osd->o_osd_lru)); 692 list_add_tail(&osd->o_osd_lru, &osdc->osd_lru); 693 osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ; 694 } 695 696 static void __remove_osd_from_lru(struct ceph_osd *osd) 697 { 698 dout("__remove_osd_from_lru %p\n", osd); 699 if (!list_empty(&osd->o_osd_lru)) 700 list_del_init(&osd->o_osd_lru); 701 } 702 703 static void remove_old_osds(struct ceph_osd_client *osdc, int remove_all) 704 { 705 struct ceph_osd *osd, *nosd; 706 707 dout("__remove_old_osds %p\n", osdc); 708 mutex_lock(&osdc->request_mutex); 709 list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) { 710 if (!remove_all && time_before(jiffies, osd->lru_ttl)) 711 break; 712 __remove_osd(osdc, osd); 713 } 714 mutex_unlock(&osdc->request_mutex); 715 } 716 717 /* 718 * reset osd connect 719 */ 720 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 721 { 722 struct ceph_osd_request *req; 723 int ret = 0; 724 725 dout("__reset_osd %p osd%d\n", osd, osd->o_osd); 726 if (list_empty(&osd->o_requests) && 727 list_empty(&osd->o_linger_requests)) { 728 __remove_osd(osdc, osd); 729 } else if (memcmp(&osdc->osdmap->osd_addr[osd->o_osd], 730 &osd->o_con.peer_addr, 731 sizeof(osd->o_con.peer_addr)) == 0 && 732 !ceph_con_opened(&osd->o_con)) { 733 dout(" osd addr hasn't changed and connection never opened," 734 " letting msgr retry"); 735 /* touch each r_stamp for handle_timeout()'s benfit */ 736 list_for_each_entry(req, &osd->o_requests, r_osd_item) 737 req->r_stamp = jiffies; 738 ret = -EAGAIN; 739 } else { 740 ceph_con_close(&osd->o_con); 741 ceph_con_open(&osd->o_con, &osdc->osdmap->osd_addr[osd->o_osd]); 742 osd->o_incarnation++; 743 } 744 return ret; 745 } 746 747 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new) 748 { 749 struct rb_node **p = &osdc->osds.rb_node; 750 struct rb_node *parent = NULL; 751 struct ceph_osd *osd = NULL; 752 753 while (*p) { 754 parent = *p; 755 osd = rb_entry(parent, struct ceph_osd, o_node); 756 if (new->o_osd < osd->o_osd) 757 p = &(*p)->rb_left; 758 else if (new->o_osd > osd->o_osd) 759 p = &(*p)->rb_right; 760 else 761 BUG(); 762 } 763 764 rb_link_node(&new->o_node, parent, p); 765 rb_insert_color(&new->o_node, &osdc->osds); 766 } 767 768 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o) 769 { 770 struct ceph_osd *osd; 771 struct rb_node *n = osdc->osds.rb_node; 772 773 while (n) { 774 osd = rb_entry(n, struct ceph_osd, o_node); 775 if (o < osd->o_osd) 776 n = n->rb_left; 777 else if (o > osd->o_osd) 778 n = n->rb_right; 779 else 780 return osd; 781 } 782 return NULL; 783 } 784 785 static void __schedule_osd_timeout(struct ceph_osd_client *osdc) 786 { 787 schedule_delayed_work(&osdc->timeout_work, 788 osdc->client->options->osd_keepalive_timeout * HZ); 789 } 790 791 static void __cancel_osd_timeout(struct ceph_osd_client *osdc) 792 { 793 cancel_delayed_work(&osdc->timeout_work); 794 } 795 796 /* 797 * Register request, assign tid. If this is the first request, set up 798 * the timeout event. 799 */ 800 static void __register_request(struct ceph_osd_client *osdc, 801 struct ceph_osd_request *req) 802 { 803 req->r_tid = ++osdc->last_tid; 804 req->r_request->hdr.tid = cpu_to_le64(req->r_tid); 805 INIT_LIST_HEAD(&req->r_req_lru_item); 806 807 dout("__register_request %p tid %lld\n", req, req->r_tid); 808 __insert_request(osdc, req); 809 ceph_osdc_get_request(req); 810 osdc->num_requests++; 811 812 if (osdc->num_requests == 1) { 813 dout(" first request, scheduling timeout\n"); 814 __schedule_osd_timeout(osdc); 815 } 816 } 817 818 static void register_request(struct ceph_osd_client *osdc, 819 struct ceph_osd_request *req) 820 { 821 mutex_lock(&osdc->request_mutex); 822 __register_request(osdc, req); 823 mutex_unlock(&osdc->request_mutex); 824 } 825 826 /* 827 * called under osdc->request_mutex 828 */ 829 static void __unregister_request(struct ceph_osd_client *osdc, 830 struct ceph_osd_request *req) 831 { 832 dout("__unregister_request %p tid %lld\n", req, req->r_tid); 833 rb_erase(&req->r_node, &osdc->requests); 834 osdc->num_requests--; 835 836 if (req->r_osd) { 837 /* make sure the original request isn't in flight. */ 838 ceph_con_revoke(&req->r_osd->o_con, req->r_request); 839 840 list_del_init(&req->r_osd_item); 841 if (list_empty(&req->r_osd->o_requests) && 842 list_empty(&req->r_osd->o_linger_requests)) { 843 dout("moving osd to %p lru\n", req->r_osd); 844 __move_osd_to_lru(osdc, req->r_osd); 845 } 846 if (list_empty(&req->r_linger_item)) 847 req->r_osd = NULL; 848 } 849 850 ceph_osdc_put_request(req); 851 852 list_del_init(&req->r_req_lru_item); 853 if (osdc->num_requests == 0) { 854 dout(" no requests, canceling timeout\n"); 855 __cancel_osd_timeout(osdc); 856 } 857 } 858 859 /* 860 * Cancel a previously queued request message 861 */ 862 static void __cancel_request(struct ceph_osd_request *req) 863 { 864 if (req->r_sent && req->r_osd) { 865 ceph_con_revoke(&req->r_osd->o_con, req->r_request); 866 req->r_sent = 0; 867 } 868 } 869 870 static void __register_linger_request(struct ceph_osd_client *osdc, 871 struct ceph_osd_request *req) 872 { 873 dout("__register_linger_request %p\n", req); 874 list_add_tail(&req->r_linger_item, &osdc->req_linger); 875 list_add_tail(&req->r_linger_osd, &req->r_osd->o_linger_requests); 876 } 877 878 static void __unregister_linger_request(struct ceph_osd_client *osdc, 879 struct ceph_osd_request *req) 880 { 881 dout("__unregister_linger_request %p\n", req); 882 if (req->r_osd) { 883 list_del_init(&req->r_linger_item); 884 list_del_init(&req->r_linger_osd); 885 886 if (list_empty(&req->r_osd->o_requests) && 887 list_empty(&req->r_osd->o_linger_requests)) { 888 dout("moving osd to %p lru\n", req->r_osd); 889 __move_osd_to_lru(osdc, req->r_osd); 890 } 891 if (list_empty(&req->r_osd_item)) 892 req->r_osd = NULL; 893 } 894 } 895 896 void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc, 897 struct ceph_osd_request *req) 898 { 899 mutex_lock(&osdc->request_mutex); 900 if (req->r_linger) { 901 __unregister_linger_request(osdc, req); 902 ceph_osdc_put_request(req); 903 } 904 mutex_unlock(&osdc->request_mutex); 905 } 906 EXPORT_SYMBOL(ceph_osdc_unregister_linger_request); 907 908 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc, 909 struct ceph_osd_request *req) 910 { 911 if (!req->r_linger) { 912 dout("set_request_linger %p\n", req); 913 req->r_linger = 1; 914 /* 915 * caller is now responsible for calling 916 * unregister_linger_request 917 */ 918 ceph_osdc_get_request(req); 919 } 920 } 921 EXPORT_SYMBOL(ceph_osdc_set_request_linger); 922 923 /* 924 * Pick an osd (the first 'up' osd in the pg), allocate the osd struct 925 * (as needed), and set the request r_osd appropriately. If there is 926 * no up osd, set r_osd to NULL. Move the request to the appropriate list 927 * (unsent, homeless) or leave on in-flight lru. 928 * 929 * Return 0 if unchanged, 1 if changed, or negative on error. 930 * 931 * Caller should hold map_sem for read and request_mutex. 932 */ 933 static int __map_request(struct ceph_osd_client *osdc, 934 struct ceph_osd_request *req) 935 { 936 struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base; 937 struct ceph_pg pgid; 938 int acting[CEPH_PG_MAX_SIZE]; 939 int o = -1, num = 0; 940 int err; 941 942 dout("map_request %p tid %lld\n", req, req->r_tid); 943 err = ceph_calc_object_layout(&reqhead->layout, req->r_oid, 944 &req->r_file_layout, osdc->osdmap); 945 if (err) { 946 list_move(&req->r_req_lru_item, &osdc->req_notarget); 947 return err; 948 } 949 pgid = reqhead->layout.ol_pgid; 950 req->r_pgid = pgid; 951 952 err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting); 953 if (err > 0) { 954 o = acting[0]; 955 num = err; 956 } 957 958 if ((req->r_osd && req->r_osd->o_osd == o && 959 req->r_sent >= req->r_osd->o_incarnation && 960 req->r_num_pg_osds == num && 961 memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) || 962 (req->r_osd == NULL && o == -1)) 963 return 0; /* no change */ 964 965 dout("map_request tid %llu pgid %d.%x osd%d (was osd%d)\n", 966 req->r_tid, le32_to_cpu(pgid.pool), le16_to_cpu(pgid.ps), o, 967 req->r_osd ? req->r_osd->o_osd : -1); 968 969 /* record full pg acting set */ 970 memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num); 971 req->r_num_pg_osds = num; 972 973 if (req->r_osd) { 974 __cancel_request(req); 975 list_del_init(&req->r_osd_item); 976 req->r_osd = NULL; 977 } 978 979 req->r_osd = __lookup_osd(osdc, o); 980 if (!req->r_osd && o >= 0) { 981 err = -ENOMEM; 982 req->r_osd = create_osd(osdc); 983 if (!req->r_osd) { 984 list_move(&req->r_req_lru_item, &osdc->req_notarget); 985 goto out; 986 } 987 988 dout("map_request osd %p is osd%d\n", req->r_osd, o); 989 req->r_osd->o_osd = o; 990 req->r_osd->o_con.peer_name.num = cpu_to_le64(o); 991 __insert_osd(osdc, req->r_osd); 992 993 ceph_con_open(&req->r_osd->o_con, &osdc->osdmap->osd_addr[o]); 994 } 995 996 if (req->r_osd) { 997 __remove_osd_from_lru(req->r_osd); 998 list_add(&req->r_osd_item, &req->r_osd->o_requests); 999 list_move(&req->r_req_lru_item, &osdc->req_unsent); 1000 } else { 1001 list_move(&req->r_req_lru_item, &osdc->req_notarget); 1002 } 1003 err = 1; /* osd or pg changed */ 1004 1005 out: 1006 return err; 1007 } 1008 1009 /* 1010 * caller should hold map_sem (for read) and request_mutex 1011 */ 1012 static int __send_request(struct ceph_osd_client *osdc, 1013 struct ceph_osd_request *req) 1014 { 1015 struct ceph_osd_request_head *reqhead; 1016 1017 dout("send_request %p tid %llu to osd%d flags %d\n", 1018 req, req->r_tid, req->r_osd->o_osd, req->r_flags); 1019 1020 reqhead = req->r_request->front.iov_base; 1021 reqhead->osdmap_epoch = cpu_to_le32(osdc->osdmap->epoch); 1022 reqhead->flags |= cpu_to_le32(req->r_flags); /* e.g., RETRY */ 1023 reqhead->reassert_version = req->r_reassert_version; 1024 1025 req->r_stamp = jiffies; 1026 list_move_tail(&req->r_req_lru_item, &osdc->req_lru); 1027 1028 ceph_msg_get(req->r_request); /* send consumes a ref */ 1029 ceph_con_send(&req->r_osd->o_con, req->r_request); 1030 req->r_sent = req->r_osd->o_incarnation; 1031 return 0; 1032 } 1033 1034 /* 1035 * Send any requests in the queue (req_unsent). 1036 */ 1037 static void send_queued(struct ceph_osd_client *osdc) 1038 { 1039 struct ceph_osd_request *req, *tmp; 1040 1041 dout("send_queued\n"); 1042 mutex_lock(&osdc->request_mutex); 1043 list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) { 1044 __send_request(osdc, req); 1045 } 1046 mutex_unlock(&osdc->request_mutex); 1047 } 1048 1049 /* 1050 * Timeout callback, called every N seconds when 1 or more osd 1051 * requests has been active for more than N seconds. When this 1052 * happens, we ping all OSDs with requests who have timed out to 1053 * ensure any communications channel reset is detected. Reset the 1054 * request timeouts another N seconds in the future as we go. 1055 * Reschedule the timeout event another N seconds in future (unless 1056 * there are no open requests). 1057 */ 1058 static void handle_timeout(struct work_struct *work) 1059 { 1060 struct ceph_osd_client *osdc = 1061 container_of(work, struct ceph_osd_client, timeout_work.work); 1062 struct ceph_osd_request *req, *last_req = NULL; 1063 struct ceph_osd *osd; 1064 unsigned long timeout = osdc->client->options->osd_timeout * HZ; 1065 unsigned long keepalive = 1066 osdc->client->options->osd_keepalive_timeout * HZ; 1067 unsigned long last_stamp = 0; 1068 struct list_head slow_osds; 1069 dout("timeout\n"); 1070 down_read(&osdc->map_sem); 1071 1072 ceph_monc_request_next_osdmap(&osdc->client->monc); 1073 1074 mutex_lock(&osdc->request_mutex); 1075 1076 /* 1077 * reset osds that appear to be _really_ unresponsive. this 1078 * is a failsafe measure.. we really shouldn't be getting to 1079 * this point if the system is working properly. the monitors 1080 * should mark the osd as failed and we should find out about 1081 * it from an updated osd map. 1082 */ 1083 while (timeout && !list_empty(&osdc->req_lru)) { 1084 req = list_entry(osdc->req_lru.next, struct ceph_osd_request, 1085 r_req_lru_item); 1086 1087 if (time_before(jiffies, req->r_stamp + timeout)) 1088 break; 1089 1090 BUG_ON(req == last_req && req->r_stamp == last_stamp); 1091 last_req = req; 1092 last_stamp = req->r_stamp; 1093 1094 osd = req->r_osd; 1095 BUG_ON(!osd); 1096 pr_warning(" tid %llu timed out on osd%d, will reset osd\n", 1097 req->r_tid, osd->o_osd); 1098 __kick_osd_requests(osdc, osd); 1099 } 1100 1101 /* 1102 * ping osds that are a bit slow. this ensures that if there 1103 * is a break in the TCP connection we will notice, and reopen 1104 * a connection with that osd (from the fault callback). 1105 */ 1106 INIT_LIST_HEAD(&slow_osds); 1107 list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) { 1108 if (time_before(jiffies, req->r_stamp + keepalive)) 1109 break; 1110 1111 osd = req->r_osd; 1112 BUG_ON(!osd); 1113 dout(" tid %llu is slow, will send keepalive on osd%d\n", 1114 req->r_tid, osd->o_osd); 1115 list_move_tail(&osd->o_keepalive_item, &slow_osds); 1116 } 1117 while (!list_empty(&slow_osds)) { 1118 osd = list_entry(slow_osds.next, struct ceph_osd, 1119 o_keepalive_item); 1120 list_del_init(&osd->o_keepalive_item); 1121 ceph_con_keepalive(&osd->o_con); 1122 } 1123 1124 __schedule_osd_timeout(osdc); 1125 mutex_unlock(&osdc->request_mutex); 1126 send_queued(osdc); 1127 up_read(&osdc->map_sem); 1128 } 1129 1130 static void handle_osds_timeout(struct work_struct *work) 1131 { 1132 struct ceph_osd_client *osdc = 1133 container_of(work, struct ceph_osd_client, 1134 osds_timeout_work.work); 1135 unsigned long delay = 1136 osdc->client->options->osd_idle_ttl * HZ >> 2; 1137 1138 dout("osds timeout\n"); 1139 down_read(&osdc->map_sem); 1140 remove_old_osds(osdc, 0); 1141 up_read(&osdc->map_sem); 1142 1143 schedule_delayed_work(&osdc->osds_timeout_work, 1144 round_jiffies_relative(delay)); 1145 } 1146 1147 static void complete_request(struct ceph_osd_request *req) 1148 { 1149 if (req->r_safe_callback) 1150 req->r_safe_callback(req, NULL); 1151 complete_all(&req->r_safe_completion); /* fsync waiter */ 1152 } 1153 1154 /* 1155 * handle osd op reply. either call the callback if it is specified, 1156 * or do the completion to wake up the waiting thread. 1157 */ 1158 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg, 1159 struct ceph_connection *con) 1160 { 1161 struct ceph_osd_reply_head *rhead = msg->front.iov_base; 1162 struct ceph_osd_request *req; 1163 u64 tid; 1164 int numops, object_len, flags; 1165 s32 result; 1166 1167 tid = le64_to_cpu(msg->hdr.tid); 1168 if (msg->front.iov_len < sizeof(*rhead)) 1169 goto bad; 1170 numops = le32_to_cpu(rhead->num_ops); 1171 object_len = le32_to_cpu(rhead->object_len); 1172 result = le32_to_cpu(rhead->result); 1173 if (msg->front.iov_len != sizeof(*rhead) + object_len + 1174 numops * sizeof(struct ceph_osd_op)) 1175 goto bad; 1176 dout("handle_reply %p tid %llu result %d\n", msg, tid, (int)result); 1177 /* lookup */ 1178 mutex_lock(&osdc->request_mutex); 1179 req = __lookup_request(osdc, tid); 1180 if (req == NULL) { 1181 dout("handle_reply tid %llu dne\n", tid); 1182 mutex_unlock(&osdc->request_mutex); 1183 return; 1184 } 1185 ceph_osdc_get_request(req); 1186 flags = le32_to_cpu(rhead->flags); 1187 1188 /* 1189 * if this connection filled our message, drop our reference now, to 1190 * avoid a (safe but slower) revoke later. 1191 */ 1192 if (req->r_con_filling_msg == con && req->r_reply == msg) { 1193 dout(" dropping con_filling_msg ref %p\n", con); 1194 req->r_con_filling_msg = NULL; 1195 ceph_con_put(con); 1196 } 1197 1198 if (!req->r_got_reply) { 1199 unsigned bytes; 1200 1201 req->r_result = le32_to_cpu(rhead->result); 1202 bytes = le32_to_cpu(msg->hdr.data_len); 1203 dout("handle_reply result %d bytes %d\n", req->r_result, 1204 bytes); 1205 if (req->r_result == 0) 1206 req->r_result = bytes; 1207 1208 /* in case this is a write and we need to replay, */ 1209 req->r_reassert_version = rhead->reassert_version; 1210 1211 req->r_got_reply = 1; 1212 } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) { 1213 dout("handle_reply tid %llu dup ack\n", tid); 1214 mutex_unlock(&osdc->request_mutex); 1215 goto done; 1216 } 1217 1218 dout("handle_reply tid %llu flags %d\n", tid, flags); 1219 1220 if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK)) 1221 __register_linger_request(osdc, req); 1222 1223 /* either this is a read, or we got the safe response */ 1224 if (result < 0 || 1225 (flags & CEPH_OSD_FLAG_ONDISK) || 1226 ((flags & CEPH_OSD_FLAG_WRITE) == 0)) 1227 __unregister_request(osdc, req); 1228 1229 mutex_unlock(&osdc->request_mutex); 1230 1231 if (req->r_callback) 1232 req->r_callback(req, msg); 1233 else 1234 complete_all(&req->r_completion); 1235 1236 if (flags & CEPH_OSD_FLAG_ONDISK) 1237 complete_request(req); 1238 1239 done: 1240 dout("req=%p req->r_linger=%d\n", req, req->r_linger); 1241 ceph_osdc_put_request(req); 1242 return; 1243 1244 bad: 1245 pr_err("corrupt osd_op_reply got %d %d expected %d\n", 1246 (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len), 1247 (int)sizeof(*rhead)); 1248 ceph_msg_dump(msg); 1249 } 1250 1251 static void reset_changed_osds(struct ceph_osd_client *osdc) 1252 { 1253 struct rb_node *p, *n; 1254 1255 for (p = rb_first(&osdc->osds); p; p = n) { 1256 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node); 1257 1258 n = rb_next(p); 1259 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) || 1260 memcmp(&osd->o_con.peer_addr, 1261 ceph_osd_addr(osdc->osdmap, 1262 osd->o_osd), 1263 sizeof(struct ceph_entity_addr)) != 0) 1264 __reset_osd(osdc, osd); 1265 } 1266 } 1267 1268 /* 1269 * Requeue requests whose mapping to an OSD has changed. If requests map to 1270 * no osd, request a new map. 1271 * 1272 * Caller should hold map_sem for read and request_mutex. 1273 */ 1274 static void kick_requests(struct ceph_osd_client *osdc) 1275 { 1276 struct ceph_osd_request *req, *nreq; 1277 struct rb_node *p; 1278 int needmap = 0; 1279 int err; 1280 1281 dout("kick_requests\n"); 1282 mutex_lock(&osdc->request_mutex); 1283 for (p = rb_first(&osdc->requests); p; p = rb_next(p)) { 1284 req = rb_entry(p, struct ceph_osd_request, r_node); 1285 err = __map_request(osdc, req); 1286 if (err < 0) 1287 continue; /* error */ 1288 if (req->r_osd == NULL) { 1289 dout("%p tid %llu maps to no osd\n", req, req->r_tid); 1290 needmap++; /* request a newer map */ 1291 } else if (err > 0) { 1292 dout("%p tid %llu requeued on osd%d\n", req, req->r_tid, 1293 req->r_osd ? req->r_osd->o_osd : -1); 1294 if (!req->r_linger) 1295 req->r_flags |= CEPH_OSD_FLAG_RETRY; 1296 } 1297 } 1298 1299 list_for_each_entry_safe(req, nreq, &osdc->req_linger, 1300 r_linger_item) { 1301 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd); 1302 1303 err = __map_request(osdc, req); 1304 if (err == 0) 1305 continue; /* no change and no osd was specified */ 1306 if (err < 0) 1307 continue; /* hrm! */ 1308 if (req->r_osd == NULL) { 1309 dout("tid %llu maps to no valid osd\n", req->r_tid); 1310 needmap++; /* request a newer map */ 1311 continue; 1312 } 1313 1314 dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid, 1315 req->r_osd ? req->r_osd->o_osd : -1); 1316 __unregister_linger_request(osdc, req); 1317 __register_request(osdc, req); 1318 } 1319 mutex_unlock(&osdc->request_mutex); 1320 1321 if (needmap) { 1322 dout("%d requests for down osds, need new map\n", needmap); 1323 ceph_monc_request_next_osdmap(&osdc->client->monc); 1324 } 1325 } 1326 1327 1328 /* 1329 * Process updated osd map. 1330 * 1331 * The message contains any number of incremental and full maps, normally 1332 * indicating some sort of topology change in the cluster. Kick requests 1333 * off to different OSDs as needed. 1334 */ 1335 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg) 1336 { 1337 void *p, *end, *next; 1338 u32 nr_maps, maplen; 1339 u32 epoch; 1340 struct ceph_osdmap *newmap = NULL, *oldmap; 1341 int err; 1342 struct ceph_fsid fsid; 1343 1344 dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0); 1345 p = msg->front.iov_base; 1346 end = p + msg->front.iov_len; 1347 1348 /* verify fsid */ 1349 ceph_decode_need(&p, end, sizeof(fsid), bad); 1350 ceph_decode_copy(&p, &fsid, sizeof(fsid)); 1351 if (ceph_check_fsid(osdc->client, &fsid) < 0) 1352 return; 1353 1354 down_write(&osdc->map_sem); 1355 1356 /* incremental maps */ 1357 ceph_decode_32_safe(&p, end, nr_maps, bad); 1358 dout(" %d inc maps\n", nr_maps); 1359 while (nr_maps > 0) { 1360 ceph_decode_need(&p, end, 2*sizeof(u32), bad); 1361 epoch = ceph_decode_32(&p); 1362 maplen = ceph_decode_32(&p); 1363 ceph_decode_need(&p, end, maplen, bad); 1364 next = p + maplen; 1365 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) { 1366 dout("applying incremental map %u len %d\n", 1367 epoch, maplen); 1368 newmap = osdmap_apply_incremental(&p, next, 1369 osdc->osdmap, 1370 osdc->client->msgr); 1371 if (IS_ERR(newmap)) { 1372 err = PTR_ERR(newmap); 1373 goto bad; 1374 } 1375 BUG_ON(!newmap); 1376 if (newmap != osdc->osdmap) { 1377 ceph_osdmap_destroy(osdc->osdmap); 1378 osdc->osdmap = newmap; 1379 } 1380 kick_requests(osdc); 1381 reset_changed_osds(osdc); 1382 } else { 1383 dout("ignoring incremental map %u len %d\n", 1384 epoch, maplen); 1385 } 1386 p = next; 1387 nr_maps--; 1388 } 1389 if (newmap) 1390 goto done; 1391 1392 /* full maps */ 1393 ceph_decode_32_safe(&p, end, nr_maps, bad); 1394 dout(" %d full maps\n", nr_maps); 1395 while (nr_maps) { 1396 ceph_decode_need(&p, end, 2*sizeof(u32), bad); 1397 epoch = ceph_decode_32(&p); 1398 maplen = ceph_decode_32(&p); 1399 ceph_decode_need(&p, end, maplen, bad); 1400 if (nr_maps > 1) { 1401 dout("skipping non-latest full map %u len %d\n", 1402 epoch, maplen); 1403 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) { 1404 dout("skipping full map %u len %d, " 1405 "older than our %u\n", epoch, maplen, 1406 osdc->osdmap->epoch); 1407 } else { 1408 dout("taking full map %u len %d\n", epoch, maplen); 1409 newmap = osdmap_decode(&p, p+maplen); 1410 if (IS_ERR(newmap)) { 1411 err = PTR_ERR(newmap); 1412 goto bad; 1413 } 1414 BUG_ON(!newmap); 1415 oldmap = osdc->osdmap; 1416 osdc->osdmap = newmap; 1417 if (oldmap) 1418 ceph_osdmap_destroy(oldmap); 1419 kick_requests(osdc); 1420 } 1421 p += maplen; 1422 nr_maps--; 1423 } 1424 1425 done: 1426 downgrade_write(&osdc->map_sem); 1427 ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch); 1428 1429 /* 1430 * subscribe to subsequent osdmap updates if full to ensure 1431 * we find out when we are no longer full and stop returning 1432 * ENOSPC. 1433 */ 1434 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL)) 1435 ceph_monc_request_next_osdmap(&osdc->client->monc); 1436 1437 send_queued(osdc); 1438 up_read(&osdc->map_sem); 1439 wake_up_all(&osdc->client->auth_wq); 1440 return; 1441 1442 bad: 1443 pr_err("osdc handle_map corrupt msg\n"); 1444 ceph_msg_dump(msg); 1445 up_write(&osdc->map_sem); 1446 return; 1447 } 1448 1449 /* 1450 * watch/notify callback event infrastructure 1451 * 1452 * These callbacks are used both for watch and notify operations. 1453 */ 1454 static void __release_event(struct kref *kref) 1455 { 1456 struct ceph_osd_event *event = 1457 container_of(kref, struct ceph_osd_event, kref); 1458 1459 dout("__release_event %p\n", event); 1460 kfree(event); 1461 } 1462 1463 static void get_event(struct ceph_osd_event *event) 1464 { 1465 kref_get(&event->kref); 1466 } 1467 1468 void ceph_osdc_put_event(struct ceph_osd_event *event) 1469 { 1470 kref_put(&event->kref, __release_event); 1471 } 1472 EXPORT_SYMBOL(ceph_osdc_put_event); 1473 1474 static void __insert_event(struct ceph_osd_client *osdc, 1475 struct ceph_osd_event *new) 1476 { 1477 struct rb_node **p = &osdc->event_tree.rb_node; 1478 struct rb_node *parent = NULL; 1479 struct ceph_osd_event *event = NULL; 1480 1481 while (*p) { 1482 parent = *p; 1483 event = rb_entry(parent, struct ceph_osd_event, node); 1484 if (new->cookie < event->cookie) 1485 p = &(*p)->rb_left; 1486 else if (new->cookie > event->cookie) 1487 p = &(*p)->rb_right; 1488 else 1489 BUG(); 1490 } 1491 1492 rb_link_node(&new->node, parent, p); 1493 rb_insert_color(&new->node, &osdc->event_tree); 1494 } 1495 1496 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc, 1497 u64 cookie) 1498 { 1499 struct rb_node **p = &osdc->event_tree.rb_node; 1500 struct rb_node *parent = NULL; 1501 struct ceph_osd_event *event = NULL; 1502 1503 while (*p) { 1504 parent = *p; 1505 event = rb_entry(parent, struct ceph_osd_event, node); 1506 if (cookie < event->cookie) 1507 p = &(*p)->rb_left; 1508 else if (cookie > event->cookie) 1509 p = &(*p)->rb_right; 1510 else 1511 return event; 1512 } 1513 return NULL; 1514 } 1515 1516 static void __remove_event(struct ceph_osd_event *event) 1517 { 1518 struct ceph_osd_client *osdc = event->osdc; 1519 1520 if (!RB_EMPTY_NODE(&event->node)) { 1521 dout("__remove_event removed %p\n", event); 1522 rb_erase(&event->node, &osdc->event_tree); 1523 ceph_osdc_put_event(event); 1524 } else { 1525 dout("__remove_event didn't remove %p\n", event); 1526 } 1527 } 1528 1529 int ceph_osdc_create_event(struct ceph_osd_client *osdc, 1530 void (*event_cb)(u64, u64, u8, void *), 1531 int one_shot, void *data, 1532 struct ceph_osd_event **pevent) 1533 { 1534 struct ceph_osd_event *event; 1535 1536 event = kmalloc(sizeof(*event), GFP_NOIO); 1537 if (!event) 1538 return -ENOMEM; 1539 1540 dout("create_event %p\n", event); 1541 event->cb = event_cb; 1542 event->one_shot = one_shot; 1543 event->data = data; 1544 event->osdc = osdc; 1545 INIT_LIST_HEAD(&event->osd_node); 1546 kref_init(&event->kref); /* one ref for us */ 1547 kref_get(&event->kref); /* one ref for the caller */ 1548 init_completion(&event->completion); 1549 1550 spin_lock(&osdc->event_lock); 1551 event->cookie = ++osdc->event_count; 1552 __insert_event(osdc, event); 1553 spin_unlock(&osdc->event_lock); 1554 1555 *pevent = event; 1556 return 0; 1557 } 1558 EXPORT_SYMBOL(ceph_osdc_create_event); 1559 1560 void ceph_osdc_cancel_event(struct ceph_osd_event *event) 1561 { 1562 struct ceph_osd_client *osdc = event->osdc; 1563 1564 dout("cancel_event %p\n", event); 1565 spin_lock(&osdc->event_lock); 1566 __remove_event(event); 1567 spin_unlock(&osdc->event_lock); 1568 ceph_osdc_put_event(event); /* caller's */ 1569 } 1570 EXPORT_SYMBOL(ceph_osdc_cancel_event); 1571 1572 1573 static void do_event_work(struct work_struct *work) 1574 { 1575 struct ceph_osd_event_work *event_work = 1576 container_of(work, struct ceph_osd_event_work, work); 1577 struct ceph_osd_event *event = event_work->event; 1578 u64 ver = event_work->ver; 1579 u64 notify_id = event_work->notify_id; 1580 u8 opcode = event_work->opcode; 1581 1582 dout("do_event_work completing %p\n", event); 1583 event->cb(ver, notify_id, opcode, event->data); 1584 complete(&event->completion); 1585 dout("do_event_work completed %p\n", event); 1586 ceph_osdc_put_event(event); 1587 kfree(event_work); 1588 } 1589 1590 1591 /* 1592 * Process osd watch notifications 1593 */ 1594 void handle_watch_notify(struct ceph_osd_client *osdc, struct ceph_msg *msg) 1595 { 1596 void *p, *end; 1597 u8 proto_ver; 1598 u64 cookie, ver, notify_id; 1599 u8 opcode; 1600 struct ceph_osd_event *event; 1601 struct ceph_osd_event_work *event_work; 1602 1603 p = msg->front.iov_base; 1604 end = p + msg->front.iov_len; 1605 1606 ceph_decode_8_safe(&p, end, proto_ver, bad); 1607 ceph_decode_8_safe(&p, end, opcode, bad); 1608 ceph_decode_64_safe(&p, end, cookie, bad); 1609 ceph_decode_64_safe(&p, end, ver, bad); 1610 ceph_decode_64_safe(&p, end, notify_id, bad); 1611 1612 spin_lock(&osdc->event_lock); 1613 event = __find_event(osdc, cookie); 1614 if (event) { 1615 get_event(event); 1616 if (event->one_shot) 1617 __remove_event(event); 1618 } 1619 spin_unlock(&osdc->event_lock); 1620 dout("handle_watch_notify cookie %lld ver %lld event %p\n", 1621 cookie, ver, event); 1622 if (event) { 1623 event_work = kmalloc(sizeof(*event_work), GFP_NOIO); 1624 if (!event_work) { 1625 dout("ERROR: could not allocate event_work\n"); 1626 goto done_err; 1627 } 1628 INIT_WORK(&event_work->work, do_event_work); 1629 event_work->event = event; 1630 event_work->ver = ver; 1631 event_work->notify_id = notify_id; 1632 event_work->opcode = opcode; 1633 if (!queue_work(osdc->notify_wq, &event_work->work)) { 1634 dout("WARNING: failed to queue notify event work\n"); 1635 goto done_err; 1636 } 1637 } 1638 1639 return; 1640 1641 done_err: 1642 complete(&event->completion); 1643 ceph_osdc_put_event(event); 1644 return; 1645 1646 bad: 1647 pr_err("osdc handle_watch_notify corrupt msg\n"); 1648 return; 1649 } 1650 1651 int ceph_osdc_wait_event(struct ceph_osd_event *event, unsigned long timeout) 1652 { 1653 int err; 1654 1655 dout("wait_event %p\n", event); 1656 err = wait_for_completion_interruptible_timeout(&event->completion, 1657 timeout * HZ); 1658 ceph_osdc_put_event(event); 1659 if (err > 0) 1660 err = 0; 1661 dout("wait_event %p returns %d\n", event, err); 1662 return err; 1663 } 1664 EXPORT_SYMBOL(ceph_osdc_wait_event); 1665 1666 /* 1667 * Register request, send initial attempt. 1668 */ 1669 int ceph_osdc_start_request(struct ceph_osd_client *osdc, 1670 struct ceph_osd_request *req, 1671 bool nofail) 1672 { 1673 int rc = 0; 1674 1675 req->r_request->pages = req->r_pages; 1676 req->r_request->nr_pages = req->r_num_pages; 1677 #ifdef CONFIG_BLOCK 1678 req->r_request->bio = req->r_bio; 1679 #endif 1680 req->r_request->trail = req->r_trail; 1681 1682 register_request(osdc, req); 1683 1684 down_read(&osdc->map_sem); 1685 mutex_lock(&osdc->request_mutex); 1686 /* 1687 * a racing kick_requests() may have sent the message for us 1688 * while we dropped request_mutex above, so only send now if 1689 * the request still han't been touched yet. 1690 */ 1691 if (req->r_sent == 0) { 1692 rc = __map_request(osdc, req); 1693 if (rc < 0) { 1694 if (nofail) { 1695 dout("osdc_start_request failed map, " 1696 " will retry %lld\n", req->r_tid); 1697 rc = 0; 1698 } 1699 goto out_unlock; 1700 } 1701 if (req->r_osd == NULL) { 1702 dout("send_request %p no up osds in pg\n", req); 1703 ceph_monc_request_next_osdmap(&osdc->client->monc); 1704 } else { 1705 rc = __send_request(osdc, req); 1706 if (rc) { 1707 if (nofail) { 1708 dout("osdc_start_request failed send, " 1709 " will retry %lld\n", req->r_tid); 1710 rc = 0; 1711 } else { 1712 __unregister_request(osdc, req); 1713 } 1714 } 1715 } 1716 } 1717 1718 out_unlock: 1719 mutex_unlock(&osdc->request_mutex); 1720 up_read(&osdc->map_sem); 1721 return rc; 1722 } 1723 EXPORT_SYMBOL(ceph_osdc_start_request); 1724 1725 /* 1726 * wait for a request to complete 1727 */ 1728 int ceph_osdc_wait_request(struct ceph_osd_client *osdc, 1729 struct ceph_osd_request *req) 1730 { 1731 int rc; 1732 1733 rc = wait_for_completion_interruptible(&req->r_completion); 1734 if (rc < 0) { 1735 mutex_lock(&osdc->request_mutex); 1736 __cancel_request(req); 1737 __unregister_request(osdc, req); 1738 mutex_unlock(&osdc->request_mutex); 1739 complete_request(req); 1740 dout("wait_request tid %llu canceled/timed out\n", req->r_tid); 1741 return rc; 1742 } 1743 1744 dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result); 1745 return req->r_result; 1746 } 1747 EXPORT_SYMBOL(ceph_osdc_wait_request); 1748 1749 /* 1750 * sync - wait for all in-flight requests to flush. avoid starvation. 1751 */ 1752 void ceph_osdc_sync(struct ceph_osd_client *osdc) 1753 { 1754 struct ceph_osd_request *req; 1755 u64 last_tid, next_tid = 0; 1756 1757 mutex_lock(&osdc->request_mutex); 1758 last_tid = osdc->last_tid; 1759 while (1) { 1760 req = __lookup_request_ge(osdc, next_tid); 1761 if (!req) 1762 break; 1763 if (req->r_tid > last_tid) 1764 break; 1765 1766 next_tid = req->r_tid + 1; 1767 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0) 1768 continue; 1769 1770 ceph_osdc_get_request(req); 1771 mutex_unlock(&osdc->request_mutex); 1772 dout("sync waiting on tid %llu (last is %llu)\n", 1773 req->r_tid, last_tid); 1774 wait_for_completion(&req->r_safe_completion); 1775 mutex_lock(&osdc->request_mutex); 1776 ceph_osdc_put_request(req); 1777 } 1778 mutex_unlock(&osdc->request_mutex); 1779 dout("sync done (thru tid %llu)\n", last_tid); 1780 } 1781 EXPORT_SYMBOL(ceph_osdc_sync); 1782 1783 /* 1784 * init, shutdown 1785 */ 1786 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client) 1787 { 1788 int err; 1789 1790 dout("init\n"); 1791 osdc->client = client; 1792 osdc->osdmap = NULL; 1793 init_rwsem(&osdc->map_sem); 1794 init_completion(&osdc->map_waiters); 1795 osdc->last_requested_map = 0; 1796 mutex_init(&osdc->request_mutex); 1797 osdc->last_tid = 0; 1798 osdc->osds = RB_ROOT; 1799 INIT_LIST_HEAD(&osdc->osd_lru); 1800 osdc->requests = RB_ROOT; 1801 INIT_LIST_HEAD(&osdc->req_lru); 1802 INIT_LIST_HEAD(&osdc->req_unsent); 1803 INIT_LIST_HEAD(&osdc->req_notarget); 1804 INIT_LIST_HEAD(&osdc->req_linger); 1805 osdc->num_requests = 0; 1806 INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout); 1807 INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout); 1808 spin_lock_init(&osdc->event_lock); 1809 osdc->event_tree = RB_ROOT; 1810 osdc->event_count = 0; 1811 1812 schedule_delayed_work(&osdc->osds_timeout_work, 1813 round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ)); 1814 1815 err = -ENOMEM; 1816 osdc->req_mempool = mempool_create_kmalloc_pool(10, 1817 sizeof(struct ceph_osd_request)); 1818 if (!osdc->req_mempool) 1819 goto out; 1820 1821 err = ceph_msgpool_init(&osdc->msgpool_op, OSD_OP_FRONT_LEN, 10, true, 1822 "osd_op"); 1823 if (err < 0) 1824 goto out_mempool; 1825 err = ceph_msgpool_init(&osdc->msgpool_op_reply, 1826 OSD_OPREPLY_FRONT_LEN, 10, true, 1827 "osd_op_reply"); 1828 if (err < 0) 1829 goto out_msgpool; 1830 1831 osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify"); 1832 if (IS_ERR(osdc->notify_wq)) { 1833 err = PTR_ERR(osdc->notify_wq); 1834 osdc->notify_wq = NULL; 1835 goto out_msgpool; 1836 } 1837 return 0; 1838 1839 out_msgpool: 1840 ceph_msgpool_destroy(&osdc->msgpool_op); 1841 out_mempool: 1842 mempool_destroy(osdc->req_mempool); 1843 out: 1844 return err; 1845 } 1846 EXPORT_SYMBOL(ceph_osdc_init); 1847 1848 void ceph_osdc_stop(struct ceph_osd_client *osdc) 1849 { 1850 flush_workqueue(osdc->notify_wq); 1851 destroy_workqueue(osdc->notify_wq); 1852 cancel_delayed_work_sync(&osdc->timeout_work); 1853 cancel_delayed_work_sync(&osdc->osds_timeout_work); 1854 if (osdc->osdmap) { 1855 ceph_osdmap_destroy(osdc->osdmap); 1856 osdc->osdmap = NULL; 1857 } 1858 remove_old_osds(osdc, 1); 1859 WARN_ON(!RB_EMPTY_ROOT(&osdc->osds)); 1860 mempool_destroy(osdc->req_mempool); 1861 ceph_msgpool_destroy(&osdc->msgpool_op); 1862 ceph_msgpool_destroy(&osdc->msgpool_op_reply); 1863 } 1864 EXPORT_SYMBOL(ceph_osdc_stop); 1865 1866 /* 1867 * Read some contiguous pages. If we cross a stripe boundary, shorten 1868 * *plen. Return number of bytes read, or error. 1869 */ 1870 int ceph_osdc_readpages(struct ceph_osd_client *osdc, 1871 struct ceph_vino vino, struct ceph_file_layout *layout, 1872 u64 off, u64 *plen, 1873 u32 truncate_seq, u64 truncate_size, 1874 struct page **pages, int num_pages, int page_align) 1875 { 1876 struct ceph_osd_request *req; 1877 int rc = 0; 1878 1879 dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino, 1880 vino.snap, off, *plen); 1881 req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 1882 CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ, 1883 NULL, 0, truncate_seq, truncate_size, NULL, 1884 false, 1, page_align); 1885 if (!req) 1886 return -ENOMEM; 1887 1888 /* it may be a short read due to an object boundary */ 1889 req->r_pages = pages; 1890 1891 dout("readpages final extent is %llu~%llu (%d pages align %d)\n", 1892 off, *plen, req->r_num_pages, page_align); 1893 1894 rc = ceph_osdc_start_request(osdc, req, false); 1895 if (!rc) 1896 rc = ceph_osdc_wait_request(osdc, req); 1897 1898 ceph_osdc_put_request(req); 1899 dout("readpages result %d\n", rc); 1900 return rc; 1901 } 1902 EXPORT_SYMBOL(ceph_osdc_readpages); 1903 1904 /* 1905 * do a synchronous write on N pages 1906 */ 1907 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino, 1908 struct ceph_file_layout *layout, 1909 struct ceph_snap_context *snapc, 1910 u64 off, u64 len, 1911 u32 truncate_seq, u64 truncate_size, 1912 struct timespec *mtime, 1913 struct page **pages, int num_pages, 1914 int flags, int do_sync, bool nofail) 1915 { 1916 struct ceph_osd_request *req; 1917 int rc = 0; 1918 int page_align = off & ~PAGE_MASK; 1919 1920 BUG_ON(vino.snap != CEPH_NOSNAP); 1921 req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 1922 CEPH_OSD_OP_WRITE, 1923 flags | CEPH_OSD_FLAG_ONDISK | 1924 CEPH_OSD_FLAG_WRITE, 1925 snapc, do_sync, 1926 truncate_seq, truncate_size, mtime, 1927 nofail, 1, page_align); 1928 if (!req) 1929 return -ENOMEM; 1930 1931 /* it may be a short write due to an object boundary */ 1932 req->r_pages = pages; 1933 dout("writepages %llu~%llu (%d pages)\n", off, len, 1934 req->r_num_pages); 1935 1936 rc = ceph_osdc_start_request(osdc, req, nofail); 1937 if (!rc) 1938 rc = ceph_osdc_wait_request(osdc, req); 1939 1940 ceph_osdc_put_request(req); 1941 if (rc == 0) 1942 rc = len; 1943 dout("writepages result %d\n", rc); 1944 return rc; 1945 } 1946 EXPORT_SYMBOL(ceph_osdc_writepages); 1947 1948 /* 1949 * handle incoming message 1950 */ 1951 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg) 1952 { 1953 struct ceph_osd *osd = con->private; 1954 struct ceph_osd_client *osdc; 1955 int type = le16_to_cpu(msg->hdr.type); 1956 1957 if (!osd) 1958 goto out; 1959 osdc = osd->o_osdc; 1960 1961 switch (type) { 1962 case CEPH_MSG_OSD_MAP: 1963 ceph_osdc_handle_map(osdc, msg); 1964 break; 1965 case CEPH_MSG_OSD_OPREPLY: 1966 handle_reply(osdc, msg, con); 1967 break; 1968 case CEPH_MSG_WATCH_NOTIFY: 1969 handle_watch_notify(osdc, msg); 1970 break; 1971 1972 default: 1973 pr_err("received unknown message type %d %s\n", type, 1974 ceph_msg_type_name(type)); 1975 } 1976 out: 1977 ceph_msg_put(msg); 1978 } 1979 1980 /* 1981 * lookup and return message for incoming reply. set up reply message 1982 * pages. 1983 */ 1984 static struct ceph_msg *get_reply(struct ceph_connection *con, 1985 struct ceph_msg_header *hdr, 1986 int *skip) 1987 { 1988 struct ceph_osd *osd = con->private; 1989 struct ceph_osd_client *osdc = osd->o_osdc; 1990 struct ceph_msg *m; 1991 struct ceph_osd_request *req; 1992 int front = le32_to_cpu(hdr->front_len); 1993 int data_len = le32_to_cpu(hdr->data_len); 1994 u64 tid; 1995 1996 tid = le64_to_cpu(hdr->tid); 1997 mutex_lock(&osdc->request_mutex); 1998 req = __lookup_request(osdc, tid); 1999 if (!req) { 2000 *skip = 1; 2001 m = NULL; 2002 pr_info("get_reply unknown tid %llu from osd%d\n", tid, 2003 osd->o_osd); 2004 goto out; 2005 } 2006 2007 if (req->r_con_filling_msg) { 2008 dout("get_reply revoking msg %p from old con %p\n", 2009 req->r_reply, req->r_con_filling_msg); 2010 ceph_con_revoke_message(req->r_con_filling_msg, req->r_reply); 2011 ceph_con_put(req->r_con_filling_msg); 2012 req->r_con_filling_msg = NULL; 2013 } 2014 2015 if (front > req->r_reply->front.iov_len) { 2016 pr_warning("get_reply front %d > preallocated %d\n", 2017 front, (int)req->r_reply->front.iov_len); 2018 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS); 2019 if (!m) 2020 goto out; 2021 ceph_msg_put(req->r_reply); 2022 req->r_reply = m; 2023 } 2024 m = ceph_msg_get(req->r_reply); 2025 2026 if (data_len > 0) { 2027 int want = calc_pages_for(req->r_page_alignment, data_len); 2028 2029 if (unlikely(req->r_num_pages < want)) { 2030 pr_warning("tid %lld reply %d > expected %d pages\n", 2031 tid, want, m->nr_pages); 2032 *skip = 1; 2033 ceph_msg_put(m); 2034 m = NULL; 2035 goto out; 2036 } 2037 m->pages = req->r_pages; 2038 m->nr_pages = req->r_num_pages; 2039 m->page_alignment = req->r_page_alignment; 2040 #ifdef CONFIG_BLOCK 2041 m->bio = req->r_bio; 2042 #endif 2043 } 2044 *skip = 0; 2045 req->r_con_filling_msg = ceph_con_get(con); 2046 dout("get_reply tid %lld %p\n", tid, m); 2047 2048 out: 2049 mutex_unlock(&osdc->request_mutex); 2050 return m; 2051 2052 } 2053 2054 static struct ceph_msg *alloc_msg(struct ceph_connection *con, 2055 struct ceph_msg_header *hdr, 2056 int *skip) 2057 { 2058 struct ceph_osd *osd = con->private; 2059 int type = le16_to_cpu(hdr->type); 2060 int front = le32_to_cpu(hdr->front_len); 2061 2062 switch (type) { 2063 case CEPH_MSG_OSD_MAP: 2064 case CEPH_MSG_WATCH_NOTIFY: 2065 return ceph_msg_new(type, front, GFP_NOFS); 2066 case CEPH_MSG_OSD_OPREPLY: 2067 return get_reply(con, hdr, skip); 2068 default: 2069 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type, 2070 osd->o_osd); 2071 *skip = 1; 2072 return NULL; 2073 } 2074 } 2075 2076 /* 2077 * Wrappers to refcount containing ceph_osd struct 2078 */ 2079 static struct ceph_connection *get_osd_con(struct ceph_connection *con) 2080 { 2081 struct ceph_osd *osd = con->private; 2082 if (get_osd(osd)) 2083 return con; 2084 return NULL; 2085 } 2086 2087 static void put_osd_con(struct ceph_connection *con) 2088 { 2089 struct ceph_osd *osd = con->private; 2090 put_osd(osd); 2091 } 2092 2093 /* 2094 * authentication 2095 */ 2096 static int get_authorizer(struct ceph_connection *con, 2097 void **buf, int *len, int *proto, 2098 void **reply_buf, int *reply_len, int force_new) 2099 { 2100 struct ceph_osd *o = con->private; 2101 struct ceph_osd_client *osdc = o->o_osdc; 2102 struct ceph_auth_client *ac = osdc->client->monc.auth; 2103 int ret = 0; 2104 2105 if (force_new && o->o_authorizer) { 2106 ac->ops->destroy_authorizer(ac, o->o_authorizer); 2107 o->o_authorizer = NULL; 2108 } 2109 if (o->o_authorizer == NULL) { 2110 ret = ac->ops->create_authorizer( 2111 ac, CEPH_ENTITY_TYPE_OSD, 2112 &o->o_authorizer, 2113 &o->o_authorizer_buf, 2114 &o->o_authorizer_buf_len, 2115 &o->o_authorizer_reply_buf, 2116 &o->o_authorizer_reply_buf_len); 2117 if (ret) 2118 return ret; 2119 } 2120 2121 *proto = ac->protocol; 2122 *buf = o->o_authorizer_buf; 2123 *len = o->o_authorizer_buf_len; 2124 *reply_buf = o->o_authorizer_reply_buf; 2125 *reply_len = o->o_authorizer_reply_buf_len; 2126 return 0; 2127 } 2128 2129 2130 static int verify_authorizer_reply(struct ceph_connection *con, int len) 2131 { 2132 struct ceph_osd *o = con->private; 2133 struct ceph_osd_client *osdc = o->o_osdc; 2134 struct ceph_auth_client *ac = osdc->client->monc.auth; 2135 2136 return ac->ops->verify_authorizer_reply(ac, o->o_authorizer, len); 2137 } 2138 2139 static int invalidate_authorizer(struct ceph_connection *con) 2140 { 2141 struct ceph_osd *o = con->private; 2142 struct ceph_osd_client *osdc = o->o_osdc; 2143 struct ceph_auth_client *ac = osdc->client->monc.auth; 2144 2145 if (ac->ops->invalidate_authorizer) 2146 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD); 2147 2148 return ceph_monc_validate_auth(&osdc->client->monc); 2149 } 2150 2151 static const struct ceph_connection_operations osd_con_ops = { 2152 .get = get_osd_con, 2153 .put = put_osd_con, 2154 .dispatch = dispatch, 2155 .get_authorizer = get_authorizer, 2156 .verify_authorizer_reply = verify_authorizer_reply, 2157 .invalidate_authorizer = invalidate_authorizer, 2158 .alloc_msg = alloc_msg, 2159 .fault = osd_reset, 2160 }; 2161