1 2 #include <linux/ceph/ceph_debug.h> 3 4 #include <linux/module.h> 5 #include <linux/err.h> 6 #include <linux/highmem.h> 7 #include <linux/mm.h> 8 #include <linux/pagemap.h> 9 #include <linux/slab.h> 10 #include <linux/uaccess.h> 11 #ifdef CONFIG_BLOCK 12 #include <linux/bio.h> 13 #endif 14 15 #include <linux/ceph/libceph.h> 16 #include <linux/ceph/osd_client.h> 17 #include <linux/ceph/messenger.h> 18 #include <linux/ceph/decode.h> 19 #include <linux/ceph/auth.h> 20 #include <linux/ceph/pagelist.h> 21 22 #define OSD_OP_FRONT_LEN 4096 23 #define OSD_OPREPLY_FRONT_LEN 512 24 25 static struct kmem_cache *ceph_osd_request_cache; 26 27 static const struct ceph_connection_operations osd_con_ops; 28 29 static void __send_queued(struct ceph_osd_client *osdc); 30 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd); 31 static void __register_request(struct ceph_osd_client *osdc, 32 struct ceph_osd_request *req); 33 static void __unregister_request(struct ceph_osd_client *osdc, 34 struct ceph_osd_request *req); 35 static void __unregister_linger_request(struct ceph_osd_client *osdc, 36 struct ceph_osd_request *req); 37 static void __enqueue_request(struct ceph_osd_request *req); 38 static void __send_request(struct ceph_osd_client *osdc, 39 struct ceph_osd_request *req); 40 41 /* 42 * Implement client access to distributed object storage cluster. 43 * 44 * All data objects are stored within a cluster/cloud of OSDs, or 45 * "object storage devices." (Note that Ceph OSDs have _nothing_ to 46 * do with the T10 OSD extensions to SCSI.) Ceph OSDs are simply 47 * remote daemons serving up and coordinating consistent and safe 48 * access to storage. 49 * 50 * Cluster membership and the mapping of data objects onto storage devices 51 * are described by the osd map. 52 * 53 * We keep track of pending OSD requests (read, write), resubmit 54 * requests to different OSDs when the cluster topology/data layout 55 * change, or retry the affected requests when the communications 56 * channel with an OSD is reset. 57 */ 58 59 /* 60 * calculate the mapping of a file extent onto an object, and fill out the 61 * request accordingly. shorten extent as necessary if it crosses an 62 * object boundary. 63 * 64 * fill osd op in request message. 65 */ 66 static int calc_layout(struct ceph_file_layout *layout, u64 off, u64 *plen, 67 u64 *objnum, u64 *objoff, u64 *objlen) 68 { 69 u64 orig_len = *plen; 70 int r; 71 72 /* object extent? */ 73 r = ceph_calc_file_object_mapping(layout, off, orig_len, objnum, 74 objoff, objlen); 75 if (r < 0) 76 return r; 77 if (*objlen < orig_len) { 78 *plen = *objlen; 79 dout(" skipping last %llu, final file extent %llu~%llu\n", 80 orig_len - *plen, off, *plen); 81 } 82 83 dout("calc_layout objnum=%llx %llu~%llu\n", *objnum, *objoff, *objlen); 84 85 return 0; 86 } 87 88 static void ceph_osd_data_init(struct ceph_osd_data *osd_data) 89 { 90 memset(osd_data, 0, sizeof (*osd_data)); 91 osd_data->type = CEPH_OSD_DATA_TYPE_NONE; 92 } 93 94 static void ceph_osd_data_pages_init(struct ceph_osd_data *osd_data, 95 struct page **pages, u64 length, u32 alignment, 96 bool pages_from_pool, bool own_pages) 97 { 98 osd_data->type = CEPH_OSD_DATA_TYPE_PAGES; 99 osd_data->pages = pages; 100 osd_data->length = length; 101 osd_data->alignment = alignment; 102 osd_data->pages_from_pool = pages_from_pool; 103 osd_data->own_pages = own_pages; 104 } 105 106 static void ceph_osd_data_pagelist_init(struct ceph_osd_data *osd_data, 107 struct ceph_pagelist *pagelist) 108 { 109 osd_data->type = CEPH_OSD_DATA_TYPE_PAGELIST; 110 osd_data->pagelist = pagelist; 111 } 112 113 #ifdef CONFIG_BLOCK 114 static void ceph_osd_data_bio_init(struct ceph_osd_data *osd_data, 115 struct bio *bio, size_t bio_length) 116 { 117 osd_data->type = CEPH_OSD_DATA_TYPE_BIO; 118 osd_data->bio = bio; 119 osd_data->bio_length = bio_length; 120 } 121 #endif /* CONFIG_BLOCK */ 122 123 #define osd_req_op_data(oreq, whch, typ, fld) \ 124 ({ \ 125 struct ceph_osd_request *__oreq = (oreq); \ 126 unsigned int __whch = (whch); \ 127 BUG_ON(__whch >= __oreq->r_num_ops); \ 128 &__oreq->r_ops[__whch].typ.fld; \ 129 }) 130 131 static struct ceph_osd_data * 132 osd_req_op_raw_data_in(struct ceph_osd_request *osd_req, unsigned int which) 133 { 134 BUG_ON(which >= osd_req->r_num_ops); 135 136 return &osd_req->r_ops[which].raw_data_in; 137 } 138 139 struct ceph_osd_data * 140 osd_req_op_extent_osd_data(struct ceph_osd_request *osd_req, 141 unsigned int which) 142 { 143 return osd_req_op_data(osd_req, which, extent, osd_data); 144 } 145 EXPORT_SYMBOL(osd_req_op_extent_osd_data); 146 147 struct ceph_osd_data * 148 osd_req_op_cls_response_data(struct ceph_osd_request *osd_req, 149 unsigned int which) 150 { 151 return osd_req_op_data(osd_req, which, cls, response_data); 152 } 153 EXPORT_SYMBOL(osd_req_op_cls_response_data); /* ??? */ 154 155 void osd_req_op_raw_data_in_pages(struct ceph_osd_request *osd_req, 156 unsigned int which, struct page **pages, 157 u64 length, u32 alignment, 158 bool pages_from_pool, bool own_pages) 159 { 160 struct ceph_osd_data *osd_data; 161 162 osd_data = osd_req_op_raw_data_in(osd_req, which); 163 ceph_osd_data_pages_init(osd_data, pages, length, alignment, 164 pages_from_pool, own_pages); 165 } 166 EXPORT_SYMBOL(osd_req_op_raw_data_in_pages); 167 168 void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *osd_req, 169 unsigned int which, struct page **pages, 170 u64 length, u32 alignment, 171 bool pages_from_pool, bool own_pages) 172 { 173 struct ceph_osd_data *osd_data; 174 175 osd_data = osd_req_op_data(osd_req, which, extent, osd_data); 176 ceph_osd_data_pages_init(osd_data, pages, length, alignment, 177 pages_from_pool, own_pages); 178 } 179 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pages); 180 181 void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *osd_req, 182 unsigned int which, struct ceph_pagelist *pagelist) 183 { 184 struct ceph_osd_data *osd_data; 185 186 osd_data = osd_req_op_data(osd_req, which, extent, osd_data); 187 ceph_osd_data_pagelist_init(osd_data, pagelist); 188 } 189 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pagelist); 190 191 #ifdef CONFIG_BLOCK 192 void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *osd_req, 193 unsigned int which, struct bio *bio, size_t bio_length) 194 { 195 struct ceph_osd_data *osd_data; 196 197 osd_data = osd_req_op_data(osd_req, which, extent, osd_data); 198 ceph_osd_data_bio_init(osd_data, bio, bio_length); 199 } 200 EXPORT_SYMBOL(osd_req_op_extent_osd_data_bio); 201 #endif /* CONFIG_BLOCK */ 202 203 static void osd_req_op_cls_request_info_pagelist( 204 struct ceph_osd_request *osd_req, 205 unsigned int which, struct ceph_pagelist *pagelist) 206 { 207 struct ceph_osd_data *osd_data; 208 209 osd_data = osd_req_op_data(osd_req, which, cls, request_info); 210 ceph_osd_data_pagelist_init(osd_data, pagelist); 211 } 212 213 void osd_req_op_cls_request_data_pagelist( 214 struct ceph_osd_request *osd_req, 215 unsigned int which, struct ceph_pagelist *pagelist) 216 { 217 struct ceph_osd_data *osd_data; 218 219 osd_data = osd_req_op_data(osd_req, which, cls, request_data); 220 ceph_osd_data_pagelist_init(osd_data, pagelist); 221 } 222 EXPORT_SYMBOL(osd_req_op_cls_request_data_pagelist); 223 224 void osd_req_op_cls_request_data_pages(struct ceph_osd_request *osd_req, 225 unsigned int which, struct page **pages, u64 length, 226 u32 alignment, bool pages_from_pool, bool own_pages) 227 { 228 struct ceph_osd_data *osd_data; 229 230 osd_data = osd_req_op_data(osd_req, which, cls, request_data); 231 ceph_osd_data_pages_init(osd_data, pages, length, alignment, 232 pages_from_pool, own_pages); 233 } 234 EXPORT_SYMBOL(osd_req_op_cls_request_data_pages); 235 236 void osd_req_op_cls_response_data_pages(struct ceph_osd_request *osd_req, 237 unsigned int which, struct page **pages, u64 length, 238 u32 alignment, bool pages_from_pool, bool own_pages) 239 { 240 struct ceph_osd_data *osd_data; 241 242 osd_data = osd_req_op_data(osd_req, which, cls, response_data); 243 ceph_osd_data_pages_init(osd_data, pages, length, alignment, 244 pages_from_pool, own_pages); 245 } 246 EXPORT_SYMBOL(osd_req_op_cls_response_data_pages); 247 248 static u64 ceph_osd_data_length(struct ceph_osd_data *osd_data) 249 { 250 switch (osd_data->type) { 251 case CEPH_OSD_DATA_TYPE_NONE: 252 return 0; 253 case CEPH_OSD_DATA_TYPE_PAGES: 254 return osd_data->length; 255 case CEPH_OSD_DATA_TYPE_PAGELIST: 256 return (u64)osd_data->pagelist->length; 257 #ifdef CONFIG_BLOCK 258 case CEPH_OSD_DATA_TYPE_BIO: 259 return (u64)osd_data->bio_length; 260 #endif /* CONFIG_BLOCK */ 261 default: 262 WARN(true, "unrecognized data type %d\n", (int)osd_data->type); 263 return 0; 264 } 265 } 266 267 static void ceph_osd_data_release(struct ceph_osd_data *osd_data) 268 { 269 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES && osd_data->own_pages) { 270 int num_pages; 271 272 num_pages = calc_pages_for((u64)osd_data->alignment, 273 (u64)osd_data->length); 274 ceph_release_page_vector(osd_data->pages, num_pages); 275 } 276 ceph_osd_data_init(osd_data); 277 } 278 279 static void osd_req_op_data_release(struct ceph_osd_request *osd_req, 280 unsigned int which) 281 { 282 struct ceph_osd_req_op *op; 283 284 BUG_ON(which >= osd_req->r_num_ops); 285 op = &osd_req->r_ops[which]; 286 287 switch (op->op) { 288 case CEPH_OSD_OP_READ: 289 case CEPH_OSD_OP_WRITE: 290 case CEPH_OSD_OP_WRITEFULL: 291 ceph_osd_data_release(&op->extent.osd_data); 292 break; 293 case CEPH_OSD_OP_CALL: 294 ceph_osd_data_release(&op->cls.request_info); 295 ceph_osd_data_release(&op->cls.request_data); 296 ceph_osd_data_release(&op->cls.response_data); 297 break; 298 case CEPH_OSD_OP_SETXATTR: 299 case CEPH_OSD_OP_CMPXATTR: 300 ceph_osd_data_release(&op->xattr.osd_data); 301 break; 302 case CEPH_OSD_OP_STAT: 303 ceph_osd_data_release(&op->raw_data_in); 304 break; 305 default: 306 break; 307 } 308 } 309 310 /* 311 * requests 312 */ 313 static void ceph_osdc_release_request(struct kref *kref) 314 { 315 struct ceph_osd_request *req = container_of(kref, 316 struct ceph_osd_request, r_kref); 317 unsigned int which; 318 319 dout("%s %p (r_request %p r_reply %p)\n", __func__, req, 320 req->r_request, req->r_reply); 321 WARN_ON(!RB_EMPTY_NODE(&req->r_node)); 322 WARN_ON(!list_empty(&req->r_req_lru_item)); 323 WARN_ON(!list_empty(&req->r_osd_item)); 324 WARN_ON(!list_empty(&req->r_linger_item)); 325 WARN_ON(!list_empty(&req->r_linger_osd_item)); 326 WARN_ON(req->r_osd); 327 328 if (req->r_request) 329 ceph_msg_put(req->r_request); 330 if (req->r_reply) { 331 ceph_msg_revoke_incoming(req->r_reply); 332 ceph_msg_put(req->r_reply); 333 } 334 335 for (which = 0; which < req->r_num_ops; which++) 336 osd_req_op_data_release(req, which); 337 338 ceph_put_snap_context(req->r_snapc); 339 if (req->r_mempool) 340 mempool_free(req, req->r_osdc->req_mempool); 341 else 342 kmem_cache_free(ceph_osd_request_cache, req); 343 344 } 345 346 void ceph_osdc_get_request(struct ceph_osd_request *req) 347 { 348 dout("%s %p (was %d)\n", __func__, req, 349 atomic_read(&req->r_kref.refcount)); 350 kref_get(&req->r_kref); 351 } 352 EXPORT_SYMBOL(ceph_osdc_get_request); 353 354 void ceph_osdc_put_request(struct ceph_osd_request *req) 355 { 356 dout("%s %p (was %d)\n", __func__, req, 357 atomic_read(&req->r_kref.refcount)); 358 kref_put(&req->r_kref, ceph_osdc_release_request); 359 } 360 EXPORT_SYMBOL(ceph_osdc_put_request); 361 362 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc, 363 struct ceph_snap_context *snapc, 364 unsigned int num_ops, 365 bool use_mempool, 366 gfp_t gfp_flags) 367 { 368 struct ceph_osd_request *req; 369 struct ceph_msg *msg; 370 size_t msg_size; 371 372 BUILD_BUG_ON(CEPH_OSD_MAX_OP > U16_MAX); 373 BUG_ON(num_ops > CEPH_OSD_MAX_OP); 374 375 msg_size = 4 + 4 + 8 + 8 + 4+8; 376 msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */ 377 msg_size += 1 + 8 + 4 + 4; /* pg_t */ 378 msg_size += 4 + CEPH_MAX_OID_NAME_LEN; /* oid */ 379 msg_size += 2 + num_ops*sizeof(struct ceph_osd_op); 380 msg_size += 8; /* snapid */ 381 msg_size += 8; /* snap_seq */ 382 msg_size += 8 * (snapc ? snapc->num_snaps : 0); /* snaps */ 383 msg_size += 4; 384 385 if (use_mempool) { 386 req = mempool_alloc(osdc->req_mempool, gfp_flags); 387 memset(req, 0, sizeof(*req)); 388 } else { 389 req = kmem_cache_zalloc(ceph_osd_request_cache, gfp_flags); 390 } 391 if (req == NULL) 392 return NULL; 393 394 req->r_osdc = osdc; 395 req->r_mempool = use_mempool; 396 req->r_num_ops = num_ops; 397 398 kref_init(&req->r_kref); 399 init_completion(&req->r_completion); 400 init_completion(&req->r_safe_completion); 401 RB_CLEAR_NODE(&req->r_node); 402 INIT_LIST_HEAD(&req->r_unsafe_item); 403 INIT_LIST_HEAD(&req->r_linger_item); 404 INIT_LIST_HEAD(&req->r_linger_osd_item); 405 INIT_LIST_HEAD(&req->r_req_lru_item); 406 INIT_LIST_HEAD(&req->r_osd_item); 407 408 req->r_base_oloc.pool = -1; 409 req->r_target_oloc.pool = -1; 410 411 /* create reply message */ 412 if (use_mempool) 413 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0); 414 else 415 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, 416 OSD_OPREPLY_FRONT_LEN, gfp_flags, true); 417 if (!msg) { 418 ceph_osdc_put_request(req); 419 return NULL; 420 } 421 req->r_reply = msg; 422 423 /* create request message; allow space for oid */ 424 if (use_mempool) 425 msg = ceph_msgpool_get(&osdc->msgpool_op, 0); 426 else 427 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true); 428 if (!msg) { 429 ceph_osdc_put_request(req); 430 return NULL; 431 } 432 433 memset(msg->front.iov_base, 0, msg->front.iov_len); 434 435 req->r_request = msg; 436 437 return req; 438 } 439 EXPORT_SYMBOL(ceph_osdc_alloc_request); 440 441 static bool osd_req_opcode_valid(u16 opcode) 442 { 443 switch (opcode) { 444 #define GENERATE_CASE(op, opcode, str) case CEPH_OSD_OP_##op: return true; 445 __CEPH_FORALL_OSD_OPS(GENERATE_CASE) 446 #undef GENERATE_CASE 447 default: 448 return false; 449 } 450 } 451 452 /* 453 * This is an osd op init function for opcodes that have no data or 454 * other information associated with them. It also serves as a 455 * common init routine for all the other init functions, below. 456 */ 457 static struct ceph_osd_req_op * 458 _osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which, 459 u16 opcode, u32 flags) 460 { 461 struct ceph_osd_req_op *op; 462 463 BUG_ON(which >= osd_req->r_num_ops); 464 BUG_ON(!osd_req_opcode_valid(opcode)); 465 466 op = &osd_req->r_ops[which]; 467 memset(op, 0, sizeof (*op)); 468 op->op = opcode; 469 op->flags = flags; 470 471 return op; 472 } 473 474 void osd_req_op_init(struct ceph_osd_request *osd_req, 475 unsigned int which, u16 opcode, u32 flags) 476 { 477 (void)_osd_req_op_init(osd_req, which, opcode, flags); 478 } 479 EXPORT_SYMBOL(osd_req_op_init); 480 481 void osd_req_op_extent_init(struct ceph_osd_request *osd_req, 482 unsigned int which, u16 opcode, 483 u64 offset, u64 length, 484 u64 truncate_size, u32 truncate_seq) 485 { 486 struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, 487 opcode, 0); 488 size_t payload_len = 0; 489 490 BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE && 491 opcode != CEPH_OSD_OP_WRITEFULL && opcode != CEPH_OSD_OP_ZERO && 492 opcode != CEPH_OSD_OP_TRUNCATE); 493 494 op->extent.offset = offset; 495 op->extent.length = length; 496 op->extent.truncate_size = truncate_size; 497 op->extent.truncate_seq = truncate_seq; 498 if (opcode == CEPH_OSD_OP_WRITE || opcode == CEPH_OSD_OP_WRITEFULL) 499 payload_len += length; 500 501 op->payload_len = payload_len; 502 } 503 EXPORT_SYMBOL(osd_req_op_extent_init); 504 505 void osd_req_op_extent_update(struct ceph_osd_request *osd_req, 506 unsigned int which, u64 length) 507 { 508 struct ceph_osd_req_op *op; 509 u64 previous; 510 511 BUG_ON(which >= osd_req->r_num_ops); 512 op = &osd_req->r_ops[which]; 513 previous = op->extent.length; 514 515 if (length == previous) 516 return; /* Nothing to do */ 517 BUG_ON(length > previous); 518 519 op->extent.length = length; 520 op->payload_len -= previous - length; 521 } 522 EXPORT_SYMBOL(osd_req_op_extent_update); 523 524 void osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which, 525 u16 opcode, const char *class, const char *method) 526 { 527 struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, 528 opcode, 0); 529 struct ceph_pagelist *pagelist; 530 size_t payload_len = 0; 531 size_t size; 532 533 BUG_ON(opcode != CEPH_OSD_OP_CALL); 534 535 pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS); 536 BUG_ON(!pagelist); 537 ceph_pagelist_init(pagelist); 538 539 op->cls.class_name = class; 540 size = strlen(class); 541 BUG_ON(size > (size_t) U8_MAX); 542 op->cls.class_len = size; 543 ceph_pagelist_append(pagelist, class, size); 544 payload_len += size; 545 546 op->cls.method_name = method; 547 size = strlen(method); 548 BUG_ON(size > (size_t) U8_MAX); 549 op->cls.method_len = size; 550 ceph_pagelist_append(pagelist, method, size); 551 payload_len += size; 552 553 osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist); 554 555 op->cls.argc = 0; /* currently unused */ 556 557 op->payload_len = payload_len; 558 } 559 EXPORT_SYMBOL(osd_req_op_cls_init); 560 561 int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which, 562 u16 opcode, const char *name, const void *value, 563 size_t size, u8 cmp_op, u8 cmp_mode) 564 { 565 struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, 566 opcode, 0); 567 struct ceph_pagelist *pagelist; 568 size_t payload_len; 569 570 BUG_ON(opcode != CEPH_OSD_OP_SETXATTR && opcode != CEPH_OSD_OP_CMPXATTR); 571 572 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS); 573 if (!pagelist) 574 return -ENOMEM; 575 576 ceph_pagelist_init(pagelist); 577 578 payload_len = strlen(name); 579 op->xattr.name_len = payload_len; 580 ceph_pagelist_append(pagelist, name, payload_len); 581 582 op->xattr.value_len = size; 583 ceph_pagelist_append(pagelist, value, size); 584 payload_len += size; 585 586 op->xattr.cmp_op = cmp_op; 587 op->xattr.cmp_mode = cmp_mode; 588 589 ceph_osd_data_pagelist_init(&op->xattr.osd_data, pagelist); 590 op->payload_len = payload_len; 591 return 0; 592 } 593 EXPORT_SYMBOL(osd_req_op_xattr_init); 594 595 void osd_req_op_watch_init(struct ceph_osd_request *osd_req, 596 unsigned int which, u16 opcode, 597 u64 cookie, u64 version, int flag) 598 { 599 struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, 600 opcode, 0); 601 602 BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH); 603 604 op->watch.cookie = cookie; 605 op->watch.ver = version; 606 if (opcode == CEPH_OSD_OP_WATCH && flag) 607 op->watch.flag = (u8)1; 608 } 609 EXPORT_SYMBOL(osd_req_op_watch_init); 610 611 void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req, 612 unsigned int which, 613 u64 expected_object_size, 614 u64 expected_write_size) 615 { 616 struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, 617 CEPH_OSD_OP_SETALLOCHINT, 618 0); 619 620 op->alloc_hint.expected_object_size = expected_object_size; 621 op->alloc_hint.expected_write_size = expected_write_size; 622 623 /* 624 * CEPH_OSD_OP_SETALLOCHINT op is advisory and therefore deemed 625 * not worth a feature bit. Set FAILOK per-op flag to make 626 * sure older osds don't trip over an unsupported opcode. 627 */ 628 op->flags |= CEPH_OSD_OP_FLAG_FAILOK; 629 } 630 EXPORT_SYMBOL(osd_req_op_alloc_hint_init); 631 632 static void ceph_osdc_msg_data_add(struct ceph_msg *msg, 633 struct ceph_osd_data *osd_data) 634 { 635 u64 length = ceph_osd_data_length(osd_data); 636 637 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 638 BUG_ON(length > (u64) SIZE_MAX); 639 if (length) 640 ceph_msg_data_add_pages(msg, osd_data->pages, 641 length, osd_data->alignment); 642 } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) { 643 BUG_ON(!length); 644 ceph_msg_data_add_pagelist(msg, osd_data->pagelist); 645 #ifdef CONFIG_BLOCK 646 } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) { 647 ceph_msg_data_add_bio(msg, osd_data->bio, length); 648 #endif 649 } else { 650 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE); 651 } 652 } 653 654 static u64 osd_req_encode_op(struct ceph_osd_request *req, 655 struct ceph_osd_op *dst, unsigned int which) 656 { 657 struct ceph_osd_req_op *src; 658 struct ceph_osd_data *osd_data; 659 u64 request_data_len = 0; 660 u64 data_length; 661 662 BUG_ON(which >= req->r_num_ops); 663 src = &req->r_ops[which]; 664 if (WARN_ON(!osd_req_opcode_valid(src->op))) { 665 pr_err("unrecognized osd opcode %d\n", src->op); 666 667 return 0; 668 } 669 670 switch (src->op) { 671 case CEPH_OSD_OP_STAT: 672 osd_data = &src->raw_data_in; 673 ceph_osdc_msg_data_add(req->r_reply, osd_data); 674 break; 675 case CEPH_OSD_OP_READ: 676 case CEPH_OSD_OP_WRITE: 677 case CEPH_OSD_OP_WRITEFULL: 678 case CEPH_OSD_OP_ZERO: 679 case CEPH_OSD_OP_TRUNCATE: 680 if (src->op == CEPH_OSD_OP_WRITE || 681 src->op == CEPH_OSD_OP_WRITEFULL) 682 request_data_len = src->extent.length; 683 dst->extent.offset = cpu_to_le64(src->extent.offset); 684 dst->extent.length = cpu_to_le64(src->extent.length); 685 dst->extent.truncate_size = 686 cpu_to_le64(src->extent.truncate_size); 687 dst->extent.truncate_seq = 688 cpu_to_le32(src->extent.truncate_seq); 689 osd_data = &src->extent.osd_data; 690 if (src->op == CEPH_OSD_OP_WRITE || 691 src->op == CEPH_OSD_OP_WRITEFULL) 692 ceph_osdc_msg_data_add(req->r_request, osd_data); 693 else 694 ceph_osdc_msg_data_add(req->r_reply, osd_data); 695 break; 696 case CEPH_OSD_OP_CALL: 697 dst->cls.class_len = src->cls.class_len; 698 dst->cls.method_len = src->cls.method_len; 699 osd_data = &src->cls.request_info; 700 ceph_osdc_msg_data_add(req->r_request, osd_data); 701 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGELIST); 702 request_data_len = osd_data->pagelist->length; 703 704 osd_data = &src->cls.request_data; 705 data_length = ceph_osd_data_length(osd_data); 706 if (data_length) { 707 BUG_ON(osd_data->type == CEPH_OSD_DATA_TYPE_NONE); 708 dst->cls.indata_len = cpu_to_le32(data_length); 709 ceph_osdc_msg_data_add(req->r_request, osd_data); 710 src->payload_len += data_length; 711 request_data_len += data_length; 712 } 713 osd_data = &src->cls.response_data; 714 ceph_osdc_msg_data_add(req->r_reply, osd_data); 715 break; 716 case CEPH_OSD_OP_STARTSYNC: 717 break; 718 case CEPH_OSD_OP_NOTIFY_ACK: 719 case CEPH_OSD_OP_WATCH: 720 dst->watch.cookie = cpu_to_le64(src->watch.cookie); 721 dst->watch.ver = cpu_to_le64(src->watch.ver); 722 dst->watch.flag = src->watch.flag; 723 break; 724 case CEPH_OSD_OP_SETALLOCHINT: 725 dst->alloc_hint.expected_object_size = 726 cpu_to_le64(src->alloc_hint.expected_object_size); 727 dst->alloc_hint.expected_write_size = 728 cpu_to_le64(src->alloc_hint.expected_write_size); 729 break; 730 case CEPH_OSD_OP_SETXATTR: 731 case CEPH_OSD_OP_CMPXATTR: 732 dst->xattr.name_len = cpu_to_le32(src->xattr.name_len); 733 dst->xattr.value_len = cpu_to_le32(src->xattr.value_len); 734 dst->xattr.cmp_op = src->xattr.cmp_op; 735 dst->xattr.cmp_mode = src->xattr.cmp_mode; 736 osd_data = &src->xattr.osd_data; 737 ceph_osdc_msg_data_add(req->r_request, osd_data); 738 request_data_len = osd_data->pagelist->length; 739 break; 740 case CEPH_OSD_OP_CREATE: 741 case CEPH_OSD_OP_DELETE: 742 break; 743 default: 744 pr_err("unsupported osd opcode %s\n", 745 ceph_osd_op_name(src->op)); 746 WARN_ON(1); 747 748 return 0; 749 } 750 751 dst->op = cpu_to_le16(src->op); 752 dst->flags = cpu_to_le32(src->flags); 753 dst->payload_len = cpu_to_le32(src->payload_len); 754 755 return request_data_len; 756 } 757 758 /* 759 * build new request AND message, calculate layout, and adjust file 760 * extent as needed. 761 * 762 * if the file was recently truncated, we include information about its 763 * old and new size so that the object can be updated appropriately. (we 764 * avoid synchronously deleting truncated objects because it's slow.) 765 * 766 * if @do_sync, include a 'startsync' command so that the osd will flush 767 * data quickly. 768 */ 769 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc, 770 struct ceph_file_layout *layout, 771 struct ceph_vino vino, 772 u64 off, u64 *plen, 773 unsigned int which, int num_ops, 774 int opcode, int flags, 775 struct ceph_snap_context *snapc, 776 u32 truncate_seq, 777 u64 truncate_size, 778 bool use_mempool) 779 { 780 struct ceph_osd_request *req; 781 u64 objnum = 0; 782 u64 objoff = 0; 783 u64 objlen = 0; 784 int r; 785 786 BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE && 787 opcode != CEPH_OSD_OP_ZERO && opcode != CEPH_OSD_OP_TRUNCATE && 788 opcode != CEPH_OSD_OP_CREATE && opcode != CEPH_OSD_OP_DELETE); 789 790 req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool, 791 GFP_NOFS); 792 if (!req) 793 return ERR_PTR(-ENOMEM); 794 795 req->r_flags = flags; 796 797 /* calculate max write size */ 798 r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen); 799 if (r < 0) { 800 ceph_osdc_put_request(req); 801 return ERR_PTR(r); 802 } 803 804 if (opcode == CEPH_OSD_OP_CREATE || opcode == CEPH_OSD_OP_DELETE) { 805 osd_req_op_init(req, which, opcode, 0); 806 } else { 807 u32 object_size = le32_to_cpu(layout->fl_object_size); 808 u32 object_base = off - objoff; 809 if (!(truncate_seq == 1 && truncate_size == -1ULL)) { 810 if (truncate_size <= object_base) { 811 truncate_size = 0; 812 } else { 813 truncate_size -= object_base; 814 if (truncate_size > object_size) 815 truncate_size = object_size; 816 } 817 } 818 osd_req_op_extent_init(req, which, opcode, objoff, objlen, 819 truncate_size, truncate_seq); 820 } 821 822 req->r_base_oloc.pool = ceph_file_layout_pg_pool(*layout); 823 824 snprintf(req->r_base_oid.name, sizeof(req->r_base_oid.name), 825 "%llx.%08llx", vino.ino, objnum); 826 req->r_base_oid.name_len = strlen(req->r_base_oid.name); 827 828 return req; 829 } 830 EXPORT_SYMBOL(ceph_osdc_new_request); 831 832 /* 833 * We keep osd requests in an rbtree, sorted by ->r_tid. 834 */ 835 static void __insert_request(struct ceph_osd_client *osdc, 836 struct ceph_osd_request *new) 837 { 838 struct rb_node **p = &osdc->requests.rb_node; 839 struct rb_node *parent = NULL; 840 struct ceph_osd_request *req = NULL; 841 842 while (*p) { 843 parent = *p; 844 req = rb_entry(parent, struct ceph_osd_request, r_node); 845 if (new->r_tid < req->r_tid) 846 p = &(*p)->rb_left; 847 else if (new->r_tid > req->r_tid) 848 p = &(*p)->rb_right; 849 else 850 BUG(); 851 } 852 853 rb_link_node(&new->r_node, parent, p); 854 rb_insert_color(&new->r_node, &osdc->requests); 855 } 856 857 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc, 858 u64 tid) 859 { 860 struct ceph_osd_request *req; 861 struct rb_node *n = osdc->requests.rb_node; 862 863 while (n) { 864 req = rb_entry(n, struct ceph_osd_request, r_node); 865 if (tid < req->r_tid) 866 n = n->rb_left; 867 else if (tid > req->r_tid) 868 n = n->rb_right; 869 else 870 return req; 871 } 872 return NULL; 873 } 874 875 static struct ceph_osd_request * 876 __lookup_request_ge(struct ceph_osd_client *osdc, 877 u64 tid) 878 { 879 struct ceph_osd_request *req; 880 struct rb_node *n = osdc->requests.rb_node; 881 882 while (n) { 883 req = rb_entry(n, struct ceph_osd_request, r_node); 884 if (tid < req->r_tid) { 885 if (!n->rb_left) 886 return req; 887 n = n->rb_left; 888 } else if (tid > req->r_tid) { 889 n = n->rb_right; 890 } else { 891 return req; 892 } 893 } 894 return NULL; 895 } 896 897 static void __kick_linger_request(struct ceph_osd_request *req) 898 { 899 struct ceph_osd_client *osdc = req->r_osdc; 900 struct ceph_osd *osd = req->r_osd; 901 902 /* 903 * Linger requests need to be resent with a new tid to avoid 904 * the dup op detection logic on the OSDs. Achieve this with 905 * a re-register dance instead of open-coding. 906 */ 907 ceph_osdc_get_request(req); 908 if (!list_empty(&req->r_linger_item)) 909 __unregister_linger_request(osdc, req); 910 else 911 __unregister_request(osdc, req); 912 __register_request(osdc, req); 913 ceph_osdc_put_request(req); 914 915 /* 916 * Unless request has been registered as both normal and 917 * lingering, __unregister{,_linger}_request clears r_osd. 918 * However, here we need to preserve r_osd to make sure we 919 * requeue on the same OSD. 920 */ 921 WARN_ON(req->r_osd || !osd); 922 req->r_osd = osd; 923 924 dout("%s requeueing %p tid %llu\n", __func__, req, req->r_tid); 925 __enqueue_request(req); 926 } 927 928 /* 929 * Resubmit requests pending on the given osd. 930 */ 931 static void __kick_osd_requests(struct ceph_osd_client *osdc, 932 struct ceph_osd *osd) 933 { 934 struct ceph_osd_request *req, *nreq; 935 LIST_HEAD(resend); 936 LIST_HEAD(resend_linger); 937 int err; 938 939 dout("%s osd%d\n", __func__, osd->o_osd); 940 err = __reset_osd(osdc, osd); 941 if (err) 942 return; 943 944 /* 945 * Build up a list of requests to resend by traversing the 946 * osd's list of requests. Requests for a given object are 947 * sent in tid order, and that is also the order they're 948 * kept on this list. Therefore all requests that are in 949 * flight will be found first, followed by all requests that 950 * have not yet been sent. And to resend requests while 951 * preserving this order we will want to put any sent 952 * requests back on the front of the osd client's unsent 953 * list. 954 * 955 * So we build a separate ordered list of already-sent 956 * requests for the affected osd and splice it onto the 957 * front of the osd client's unsent list. Once we've seen a 958 * request that has not yet been sent we're done. Those 959 * requests are already sitting right where they belong. 960 */ 961 list_for_each_entry(req, &osd->o_requests, r_osd_item) { 962 if (!req->r_sent) 963 break; 964 965 if (!req->r_linger) { 966 dout("%s requeueing %p tid %llu\n", __func__, req, 967 req->r_tid); 968 list_move_tail(&req->r_req_lru_item, &resend); 969 req->r_flags |= CEPH_OSD_FLAG_RETRY; 970 } else { 971 list_move_tail(&req->r_req_lru_item, &resend_linger); 972 } 973 } 974 list_splice(&resend, &osdc->req_unsent); 975 976 /* 977 * Both registered and not yet registered linger requests are 978 * enqueued with a new tid on the same OSD. We add/move them 979 * to req_unsent/o_requests at the end to keep things in tid 980 * order. 981 */ 982 list_for_each_entry_safe(req, nreq, &osd->o_linger_requests, 983 r_linger_osd_item) { 984 WARN_ON(!list_empty(&req->r_req_lru_item)); 985 __kick_linger_request(req); 986 } 987 988 list_for_each_entry_safe(req, nreq, &resend_linger, r_req_lru_item) 989 __kick_linger_request(req); 990 } 991 992 /* 993 * If the osd connection drops, we need to resubmit all requests. 994 */ 995 static void osd_reset(struct ceph_connection *con) 996 { 997 struct ceph_osd *osd = con->private; 998 struct ceph_osd_client *osdc; 999 1000 if (!osd) 1001 return; 1002 dout("osd_reset osd%d\n", osd->o_osd); 1003 osdc = osd->o_osdc; 1004 down_read(&osdc->map_sem); 1005 mutex_lock(&osdc->request_mutex); 1006 __kick_osd_requests(osdc, osd); 1007 __send_queued(osdc); 1008 mutex_unlock(&osdc->request_mutex); 1009 up_read(&osdc->map_sem); 1010 } 1011 1012 /* 1013 * Track open sessions with osds. 1014 */ 1015 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum) 1016 { 1017 struct ceph_osd *osd; 1018 1019 osd = kzalloc(sizeof(*osd), GFP_NOFS); 1020 if (!osd) 1021 return NULL; 1022 1023 atomic_set(&osd->o_ref, 1); 1024 osd->o_osdc = osdc; 1025 osd->o_osd = onum; 1026 RB_CLEAR_NODE(&osd->o_node); 1027 INIT_LIST_HEAD(&osd->o_requests); 1028 INIT_LIST_HEAD(&osd->o_linger_requests); 1029 INIT_LIST_HEAD(&osd->o_osd_lru); 1030 osd->o_incarnation = 1; 1031 1032 ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr); 1033 1034 INIT_LIST_HEAD(&osd->o_keepalive_item); 1035 return osd; 1036 } 1037 1038 static struct ceph_osd *get_osd(struct ceph_osd *osd) 1039 { 1040 if (atomic_inc_not_zero(&osd->o_ref)) { 1041 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1, 1042 atomic_read(&osd->o_ref)); 1043 return osd; 1044 } else { 1045 dout("get_osd %p FAIL\n", osd); 1046 return NULL; 1047 } 1048 } 1049 1050 static void put_osd(struct ceph_osd *osd) 1051 { 1052 dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref), 1053 atomic_read(&osd->o_ref) - 1); 1054 if (atomic_dec_and_test(&osd->o_ref)) { 1055 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth; 1056 1057 if (osd->o_auth.authorizer) 1058 ceph_auth_destroy_authorizer(ac, osd->o_auth.authorizer); 1059 kfree(osd); 1060 } 1061 } 1062 1063 /* 1064 * remove an osd from our map 1065 */ 1066 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 1067 { 1068 dout("%s %p osd%d\n", __func__, osd, osd->o_osd); 1069 WARN_ON(!list_empty(&osd->o_requests)); 1070 WARN_ON(!list_empty(&osd->o_linger_requests)); 1071 1072 list_del_init(&osd->o_osd_lru); 1073 rb_erase(&osd->o_node, &osdc->osds); 1074 RB_CLEAR_NODE(&osd->o_node); 1075 } 1076 1077 static void remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 1078 { 1079 dout("%s %p osd%d\n", __func__, osd, osd->o_osd); 1080 1081 if (!RB_EMPTY_NODE(&osd->o_node)) { 1082 ceph_con_close(&osd->o_con); 1083 __remove_osd(osdc, osd); 1084 put_osd(osd); 1085 } 1086 } 1087 1088 static void remove_all_osds(struct ceph_osd_client *osdc) 1089 { 1090 dout("%s %p\n", __func__, osdc); 1091 mutex_lock(&osdc->request_mutex); 1092 while (!RB_EMPTY_ROOT(&osdc->osds)) { 1093 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds), 1094 struct ceph_osd, o_node); 1095 remove_osd(osdc, osd); 1096 } 1097 mutex_unlock(&osdc->request_mutex); 1098 } 1099 1100 static void __move_osd_to_lru(struct ceph_osd_client *osdc, 1101 struct ceph_osd *osd) 1102 { 1103 dout("%s %p\n", __func__, osd); 1104 BUG_ON(!list_empty(&osd->o_osd_lru)); 1105 1106 list_add_tail(&osd->o_osd_lru, &osdc->osd_lru); 1107 osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl; 1108 } 1109 1110 static void maybe_move_osd_to_lru(struct ceph_osd_client *osdc, 1111 struct ceph_osd *osd) 1112 { 1113 dout("%s %p\n", __func__, osd); 1114 1115 if (list_empty(&osd->o_requests) && 1116 list_empty(&osd->o_linger_requests)) 1117 __move_osd_to_lru(osdc, osd); 1118 } 1119 1120 static void __remove_osd_from_lru(struct ceph_osd *osd) 1121 { 1122 dout("__remove_osd_from_lru %p\n", osd); 1123 if (!list_empty(&osd->o_osd_lru)) 1124 list_del_init(&osd->o_osd_lru); 1125 } 1126 1127 static void remove_old_osds(struct ceph_osd_client *osdc) 1128 { 1129 struct ceph_osd *osd, *nosd; 1130 1131 dout("__remove_old_osds %p\n", osdc); 1132 mutex_lock(&osdc->request_mutex); 1133 list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) { 1134 if (time_before(jiffies, osd->lru_ttl)) 1135 break; 1136 remove_osd(osdc, osd); 1137 } 1138 mutex_unlock(&osdc->request_mutex); 1139 } 1140 1141 /* 1142 * reset osd connect 1143 */ 1144 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 1145 { 1146 struct ceph_entity_addr *peer_addr; 1147 1148 dout("__reset_osd %p osd%d\n", osd, osd->o_osd); 1149 if (list_empty(&osd->o_requests) && 1150 list_empty(&osd->o_linger_requests)) { 1151 remove_osd(osdc, osd); 1152 return -ENODEV; 1153 } 1154 1155 peer_addr = &osdc->osdmap->osd_addr[osd->o_osd]; 1156 if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) && 1157 !ceph_con_opened(&osd->o_con)) { 1158 struct ceph_osd_request *req; 1159 1160 dout("osd addr hasn't changed and connection never opened, " 1161 "letting msgr retry\n"); 1162 /* touch each r_stamp for handle_timeout()'s benfit */ 1163 list_for_each_entry(req, &osd->o_requests, r_osd_item) 1164 req->r_stamp = jiffies; 1165 1166 return -EAGAIN; 1167 } 1168 1169 ceph_con_close(&osd->o_con); 1170 ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr); 1171 osd->o_incarnation++; 1172 1173 return 0; 1174 } 1175 1176 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new) 1177 { 1178 struct rb_node **p = &osdc->osds.rb_node; 1179 struct rb_node *parent = NULL; 1180 struct ceph_osd *osd = NULL; 1181 1182 dout("__insert_osd %p osd%d\n", new, new->o_osd); 1183 while (*p) { 1184 parent = *p; 1185 osd = rb_entry(parent, struct ceph_osd, o_node); 1186 if (new->o_osd < osd->o_osd) 1187 p = &(*p)->rb_left; 1188 else if (new->o_osd > osd->o_osd) 1189 p = &(*p)->rb_right; 1190 else 1191 BUG(); 1192 } 1193 1194 rb_link_node(&new->o_node, parent, p); 1195 rb_insert_color(&new->o_node, &osdc->osds); 1196 } 1197 1198 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o) 1199 { 1200 struct ceph_osd *osd; 1201 struct rb_node *n = osdc->osds.rb_node; 1202 1203 while (n) { 1204 osd = rb_entry(n, struct ceph_osd, o_node); 1205 if (o < osd->o_osd) 1206 n = n->rb_left; 1207 else if (o > osd->o_osd) 1208 n = n->rb_right; 1209 else 1210 return osd; 1211 } 1212 return NULL; 1213 } 1214 1215 static void __schedule_osd_timeout(struct ceph_osd_client *osdc) 1216 { 1217 schedule_delayed_work(&osdc->timeout_work, 1218 osdc->client->options->osd_keepalive_timeout); 1219 } 1220 1221 static void __cancel_osd_timeout(struct ceph_osd_client *osdc) 1222 { 1223 cancel_delayed_work(&osdc->timeout_work); 1224 } 1225 1226 /* 1227 * Register request, assign tid. If this is the first request, set up 1228 * the timeout event. 1229 */ 1230 static void __register_request(struct ceph_osd_client *osdc, 1231 struct ceph_osd_request *req) 1232 { 1233 req->r_tid = ++osdc->last_tid; 1234 req->r_request->hdr.tid = cpu_to_le64(req->r_tid); 1235 dout("__register_request %p tid %lld\n", req, req->r_tid); 1236 __insert_request(osdc, req); 1237 ceph_osdc_get_request(req); 1238 osdc->num_requests++; 1239 if (osdc->num_requests == 1) { 1240 dout(" first request, scheduling timeout\n"); 1241 __schedule_osd_timeout(osdc); 1242 } 1243 } 1244 1245 /* 1246 * called under osdc->request_mutex 1247 */ 1248 static void __unregister_request(struct ceph_osd_client *osdc, 1249 struct ceph_osd_request *req) 1250 { 1251 if (RB_EMPTY_NODE(&req->r_node)) { 1252 dout("__unregister_request %p tid %lld not registered\n", 1253 req, req->r_tid); 1254 return; 1255 } 1256 1257 dout("__unregister_request %p tid %lld\n", req, req->r_tid); 1258 rb_erase(&req->r_node, &osdc->requests); 1259 RB_CLEAR_NODE(&req->r_node); 1260 osdc->num_requests--; 1261 1262 if (req->r_osd) { 1263 /* make sure the original request isn't in flight. */ 1264 ceph_msg_revoke(req->r_request); 1265 1266 list_del_init(&req->r_osd_item); 1267 maybe_move_osd_to_lru(osdc, req->r_osd); 1268 if (list_empty(&req->r_linger_osd_item)) 1269 req->r_osd = NULL; 1270 } 1271 1272 list_del_init(&req->r_req_lru_item); 1273 ceph_osdc_put_request(req); 1274 1275 if (osdc->num_requests == 0) { 1276 dout(" no requests, canceling timeout\n"); 1277 __cancel_osd_timeout(osdc); 1278 } 1279 } 1280 1281 /* 1282 * Cancel a previously queued request message 1283 */ 1284 static void __cancel_request(struct ceph_osd_request *req) 1285 { 1286 if (req->r_sent && req->r_osd) { 1287 ceph_msg_revoke(req->r_request); 1288 req->r_sent = 0; 1289 } 1290 } 1291 1292 static void __register_linger_request(struct ceph_osd_client *osdc, 1293 struct ceph_osd_request *req) 1294 { 1295 dout("%s %p tid %llu\n", __func__, req, req->r_tid); 1296 WARN_ON(!req->r_linger); 1297 1298 ceph_osdc_get_request(req); 1299 list_add_tail(&req->r_linger_item, &osdc->req_linger); 1300 if (req->r_osd) 1301 list_add_tail(&req->r_linger_osd_item, 1302 &req->r_osd->o_linger_requests); 1303 } 1304 1305 static void __unregister_linger_request(struct ceph_osd_client *osdc, 1306 struct ceph_osd_request *req) 1307 { 1308 WARN_ON(!req->r_linger); 1309 1310 if (list_empty(&req->r_linger_item)) { 1311 dout("%s %p tid %llu not registered\n", __func__, req, 1312 req->r_tid); 1313 return; 1314 } 1315 1316 dout("%s %p tid %llu\n", __func__, req, req->r_tid); 1317 list_del_init(&req->r_linger_item); 1318 1319 if (req->r_osd) { 1320 list_del_init(&req->r_linger_osd_item); 1321 maybe_move_osd_to_lru(osdc, req->r_osd); 1322 if (list_empty(&req->r_osd_item)) 1323 req->r_osd = NULL; 1324 } 1325 ceph_osdc_put_request(req); 1326 } 1327 1328 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc, 1329 struct ceph_osd_request *req) 1330 { 1331 if (!req->r_linger) { 1332 dout("set_request_linger %p\n", req); 1333 req->r_linger = 1; 1334 } 1335 } 1336 EXPORT_SYMBOL(ceph_osdc_set_request_linger); 1337 1338 /* 1339 * Returns whether a request should be blocked from being sent 1340 * based on the current osdmap and osd_client settings. 1341 * 1342 * Caller should hold map_sem for read. 1343 */ 1344 static bool __req_should_be_paused(struct ceph_osd_client *osdc, 1345 struct ceph_osd_request *req) 1346 { 1347 bool pauserd = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD); 1348 bool pausewr = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR) || 1349 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL); 1350 return (req->r_flags & CEPH_OSD_FLAG_READ && pauserd) || 1351 (req->r_flags & CEPH_OSD_FLAG_WRITE && pausewr); 1352 } 1353 1354 /* 1355 * Calculate mapping of a request to a PG. Takes tiering into account. 1356 */ 1357 static int __calc_request_pg(struct ceph_osdmap *osdmap, 1358 struct ceph_osd_request *req, 1359 struct ceph_pg *pg_out) 1360 { 1361 bool need_check_tiering; 1362 1363 need_check_tiering = false; 1364 if (req->r_target_oloc.pool == -1) { 1365 req->r_target_oloc = req->r_base_oloc; /* struct */ 1366 need_check_tiering = true; 1367 } 1368 if (req->r_target_oid.name_len == 0) { 1369 ceph_oid_copy(&req->r_target_oid, &req->r_base_oid); 1370 need_check_tiering = true; 1371 } 1372 1373 if (need_check_tiering && 1374 (req->r_flags & CEPH_OSD_FLAG_IGNORE_OVERLAY) == 0) { 1375 struct ceph_pg_pool_info *pi; 1376 1377 pi = ceph_pg_pool_by_id(osdmap, req->r_target_oloc.pool); 1378 if (pi) { 1379 if ((req->r_flags & CEPH_OSD_FLAG_READ) && 1380 pi->read_tier >= 0) 1381 req->r_target_oloc.pool = pi->read_tier; 1382 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) && 1383 pi->write_tier >= 0) 1384 req->r_target_oloc.pool = pi->write_tier; 1385 } 1386 /* !pi is caught in ceph_oloc_oid_to_pg() */ 1387 } 1388 1389 return ceph_oloc_oid_to_pg(osdmap, &req->r_target_oloc, 1390 &req->r_target_oid, pg_out); 1391 } 1392 1393 static void __enqueue_request(struct ceph_osd_request *req) 1394 { 1395 struct ceph_osd_client *osdc = req->r_osdc; 1396 1397 dout("%s %p tid %llu to osd%d\n", __func__, req, req->r_tid, 1398 req->r_osd ? req->r_osd->o_osd : -1); 1399 1400 if (req->r_osd) { 1401 __remove_osd_from_lru(req->r_osd); 1402 list_add_tail(&req->r_osd_item, &req->r_osd->o_requests); 1403 list_move_tail(&req->r_req_lru_item, &osdc->req_unsent); 1404 } else { 1405 list_move_tail(&req->r_req_lru_item, &osdc->req_notarget); 1406 } 1407 } 1408 1409 /* 1410 * Pick an osd (the first 'up' osd in the pg), allocate the osd struct 1411 * (as needed), and set the request r_osd appropriately. If there is 1412 * no up osd, set r_osd to NULL. Move the request to the appropriate list 1413 * (unsent, homeless) or leave on in-flight lru. 1414 * 1415 * Return 0 if unchanged, 1 if changed, or negative on error. 1416 * 1417 * Caller should hold map_sem for read and request_mutex. 1418 */ 1419 static int __map_request(struct ceph_osd_client *osdc, 1420 struct ceph_osd_request *req, int force_resend) 1421 { 1422 struct ceph_pg pgid; 1423 int acting[CEPH_PG_MAX_SIZE]; 1424 int num, o; 1425 int err; 1426 bool was_paused; 1427 1428 dout("map_request %p tid %lld\n", req, req->r_tid); 1429 1430 err = __calc_request_pg(osdc->osdmap, req, &pgid); 1431 if (err) { 1432 list_move(&req->r_req_lru_item, &osdc->req_notarget); 1433 return err; 1434 } 1435 req->r_pgid = pgid; 1436 1437 num = ceph_calc_pg_acting(osdc->osdmap, pgid, acting, &o); 1438 if (num < 0) 1439 num = 0; 1440 1441 was_paused = req->r_paused; 1442 req->r_paused = __req_should_be_paused(osdc, req); 1443 if (was_paused && !req->r_paused) 1444 force_resend = 1; 1445 1446 if ((!force_resend && 1447 req->r_osd && req->r_osd->o_osd == o && 1448 req->r_sent >= req->r_osd->o_incarnation && 1449 req->r_num_pg_osds == num && 1450 memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) || 1451 (req->r_osd == NULL && o == -1) || 1452 req->r_paused) 1453 return 0; /* no change */ 1454 1455 dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n", 1456 req->r_tid, pgid.pool, pgid.seed, o, 1457 req->r_osd ? req->r_osd->o_osd : -1); 1458 1459 /* record full pg acting set */ 1460 memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num); 1461 req->r_num_pg_osds = num; 1462 1463 if (req->r_osd) { 1464 __cancel_request(req); 1465 list_del_init(&req->r_osd_item); 1466 list_del_init(&req->r_linger_osd_item); 1467 req->r_osd = NULL; 1468 } 1469 1470 req->r_osd = __lookup_osd(osdc, o); 1471 if (!req->r_osd && o >= 0) { 1472 err = -ENOMEM; 1473 req->r_osd = create_osd(osdc, o); 1474 if (!req->r_osd) { 1475 list_move(&req->r_req_lru_item, &osdc->req_notarget); 1476 goto out; 1477 } 1478 1479 dout("map_request osd %p is osd%d\n", req->r_osd, o); 1480 __insert_osd(osdc, req->r_osd); 1481 1482 ceph_con_open(&req->r_osd->o_con, 1483 CEPH_ENTITY_TYPE_OSD, o, 1484 &osdc->osdmap->osd_addr[o]); 1485 } 1486 1487 __enqueue_request(req); 1488 err = 1; /* osd or pg changed */ 1489 1490 out: 1491 return err; 1492 } 1493 1494 /* 1495 * caller should hold map_sem (for read) and request_mutex 1496 */ 1497 static void __send_request(struct ceph_osd_client *osdc, 1498 struct ceph_osd_request *req) 1499 { 1500 void *p; 1501 1502 dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n", 1503 req, req->r_tid, req->r_osd->o_osd, req->r_flags, 1504 (unsigned long long)req->r_pgid.pool, req->r_pgid.seed); 1505 1506 /* fill in message content that changes each time we send it */ 1507 put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch); 1508 put_unaligned_le32(req->r_flags, req->r_request_flags); 1509 put_unaligned_le64(req->r_target_oloc.pool, req->r_request_pool); 1510 p = req->r_request_pgid; 1511 ceph_encode_64(&p, req->r_pgid.pool); 1512 ceph_encode_32(&p, req->r_pgid.seed); 1513 put_unaligned_le64(1, req->r_request_attempts); /* FIXME */ 1514 memcpy(req->r_request_reassert_version, &req->r_reassert_version, 1515 sizeof(req->r_reassert_version)); 1516 1517 req->r_stamp = jiffies; 1518 list_move_tail(&req->r_req_lru_item, &osdc->req_lru); 1519 1520 ceph_msg_get(req->r_request); /* send consumes a ref */ 1521 1522 req->r_sent = req->r_osd->o_incarnation; 1523 1524 ceph_con_send(&req->r_osd->o_con, req->r_request); 1525 } 1526 1527 /* 1528 * Send any requests in the queue (req_unsent). 1529 */ 1530 static void __send_queued(struct ceph_osd_client *osdc) 1531 { 1532 struct ceph_osd_request *req, *tmp; 1533 1534 dout("__send_queued\n"); 1535 list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) 1536 __send_request(osdc, req); 1537 } 1538 1539 /* 1540 * Caller should hold map_sem for read and request_mutex. 1541 */ 1542 static int __ceph_osdc_start_request(struct ceph_osd_client *osdc, 1543 struct ceph_osd_request *req, 1544 bool nofail) 1545 { 1546 int rc; 1547 1548 __register_request(osdc, req); 1549 req->r_sent = 0; 1550 req->r_got_reply = 0; 1551 rc = __map_request(osdc, req, 0); 1552 if (rc < 0) { 1553 if (nofail) { 1554 dout("osdc_start_request failed map, " 1555 " will retry %lld\n", req->r_tid); 1556 rc = 0; 1557 } else { 1558 __unregister_request(osdc, req); 1559 } 1560 return rc; 1561 } 1562 1563 if (req->r_osd == NULL) { 1564 dout("send_request %p no up osds in pg\n", req); 1565 ceph_monc_request_next_osdmap(&osdc->client->monc); 1566 } else { 1567 __send_queued(osdc); 1568 } 1569 1570 return 0; 1571 } 1572 1573 /* 1574 * Timeout callback, called every N seconds when 1 or more osd 1575 * requests has been active for more than N seconds. When this 1576 * happens, we ping all OSDs with requests who have timed out to 1577 * ensure any communications channel reset is detected. Reset the 1578 * request timeouts another N seconds in the future as we go. 1579 * Reschedule the timeout event another N seconds in future (unless 1580 * there are no open requests). 1581 */ 1582 static void handle_timeout(struct work_struct *work) 1583 { 1584 struct ceph_osd_client *osdc = 1585 container_of(work, struct ceph_osd_client, timeout_work.work); 1586 struct ceph_options *opts = osdc->client->options; 1587 struct ceph_osd_request *req; 1588 struct ceph_osd *osd; 1589 struct list_head slow_osds; 1590 dout("timeout\n"); 1591 down_read(&osdc->map_sem); 1592 1593 ceph_monc_request_next_osdmap(&osdc->client->monc); 1594 1595 mutex_lock(&osdc->request_mutex); 1596 1597 /* 1598 * ping osds that are a bit slow. this ensures that if there 1599 * is a break in the TCP connection we will notice, and reopen 1600 * a connection with that osd (from the fault callback). 1601 */ 1602 INIT_LIST_HEAD(&slow_osds); 1603 list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) { 1604 if (time_before(jiffies, 1605 req->r_stamp + opts->osd_keepalive_timeout)) 1606 break; 1607 1608 osd = req->r_osd; 1609 BUG_ON(!osd); 1610 dout(" tid %llu is slow, will send keepalive on osd%d\n", 1611 req->r_tid, osd->o_osd); 1612 list_move_tail(&osd->o_keepalive_item, &slow_osds); 1613 } 1614 while (!list_empty(&slow_osds)) { 1615 osd = list_entry(slow_osds.next, struct ceph_osd, 1616 o_keepalive_item); 1617 list_del_init(&osd->o_keepalive_item); 1618 ceph_con_keepalive(&osd->o_con); 1619 } 1620 1621 __schedule_osd_timeout(osdc); 1622 __send_queued(osdc); 1623 mutex_unlock(&osdc->request_mutex); 1624 up_read(&osdc->map_sem); 1625 } 1626 1627 static void handle_osds_timeout(struct work_struct *work) 1628 { 1629 struct ceph_osd_client *osdc = 1630 container_of(work, struct ceph_osd_client, 1631 osds_timeout_work.work); 1632 unsigned long delay = osdc->client->options->osd_idle_ttl / 4; 1633 1634 dout("osds timeout\n"); 1635 down_read(&osdc->map_sem); 1636 remove_old_osds(osdc); 1637 up_read(&osdc->map_sem); 1638 1639 schedule_delayed_work(&osdc->osds_timeout_work, 1640 round_jiffies_relative(delay)); 1641 } 1642 1643 static int ceph_oloc_decode(void **p, void *end, 1644 struct ceph_object_locator *oloc) 1645 { 1646 u8 struct_v, struct_cv; 1647 u32 len; 1648 void *struct_end; 1649 int ret = 0; 1650 1651 ceph_decode_need(p, end, 1 + 1 + 4, e_inval); 1652 struct_v = ceph_decode_8(p); 1653 struct_cv = ceph_decode_8(p); 1654 if (struct_v < 3) { 1655 pr_warn("got v %d < 3 cv %d of ceph_object_locator\n", 1656 struct_v, struct_cv); 1657 goto e_inval; 1658 } 1659 if (struct_cv > 6) { 1660 pr_warn("got v %d cv %d > 6 of ceph_object_locator\n", 1661 struct_v, struct_cv); 1662 goto e_inval; 1663 } 1664 len = ceph_decode_32(p); 1665 ceph_decode_need(p, end, len, e_inval); 1666 struct_end = *p + len; 1667 1668 oloc->pool = ceph_decode_64(p); 1669 *p += 4; /* skip preferred */ 1670 1671 len = ceph_decode_32(p); 1672 if (len > 0) { 1673 pr_warn("ceph_object_locator::key is set\n"); 1674 goto e_inval; 1675 } 1676 1677 if (struct_v >= 5) { 1678 len = ceph_decode_32(p); 1679 if (len > 0) { 1680 pr_warn("ceph_object_locator::nspace is set\n"); 1681 goto e_inval; 1682 } 1683 } 1684 1685 if (struct_v >= 6) { 1686 s64 hash = ceph_decode_64(p); 1687 if (hash != -1) { 1688 pr_warn("ceph_object_locator::hash is set\n"); 1689 goto e_inval; 1690 } 1691 } 1692 1693 /* skip the rest */ 1694 *p = struct_end; 1695 out: 1696 return ret; 1697 1698 e_inval: 1699 ret = -EINVAL; 1700 goto out; 1701 } 1702 1703 static int ceph_redirect_decode(void **p, void *end, 1704 struct ceph_request_redirect *redir) 1705 { 1706 u8 struct_v, struct_cv; 1707 u32 len; 1708 void *struct_end; 1709 int ret; 1710 1711 ceph_decode_need(p, end, 1 + 1 + 4, e_inval); 1712 struct_v = ceph_decode_8(p); 1713 struct_cv = ceph_decode_8(p); 1714 if (struct_cv > 1) { 1715 pr_warn("got v %d cv %d > 1 of ceph_request_redirect\n", 1716 struct_v, struct_cv); 1717 goto e_inval; 1718 } 1719 len = ceph_decode_32(p); 1720 ceph_decode_need(p, end, len, e_inval); 1721 struct_end = *p + len; 1722 1723 ret = ceph_oloc_decode(p, end, &redir->oloc); 1724 if (ret) 1725 goto out; 1726 1727 len = ceph_decode_32(p); 1728 if (len > 0) { 1729 pr_warn("ceph_request_redirect::object_name is set\n"); 1730 goto e_inval; 1731 } 1732 1733 len = ceph_decode_32(p); 1734 *p += len; /* skip osd_instructions */ 1735 1736 /* skip the rest */ 1737 *p = struct_end; 1738 out: 1739 return ret; 1740 1741 e_inval: 1742 ret = -EINVAL; 1743 goto out; 1744 } 1745 1746 static void complete_request(struct ceph_osd_request *req) 1747 { 1748 complete_all(&req->r_safe_completion); /* fsync waiter */ 1749 } 1750 1751 /* 1752 * handle osd op reply. either call the callback if it is specified, 1753 * or do the completion to wake up the waiting thread. 1754 */ 1755 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg) 1756 { 1757 void *p, *end; 1758 struct ceph_osd_request *req; 1759 struct ceph_request_redirect redir; 1760 u64 tid; 1761 int object_len; 1762 unsigned int numops; 1763 int payload_len, flags; 1764 s32 result; 1765 s32 retry_attempt; 1766 struct ceph_pg pg; 1767 int err; 1768 u32 reassert_epoch; 1769 u64 reassert_version; 1770 u32 osdmap_epoch; 1771 int already_completed; 1772 u32 bytes; 1773 u8 decode_redir; 1774 unsigned int i; 1775 1776 tid = le64_to_cpu(msg->hdr.tid); 1777 dout("handle_reply %p tid %llu\n", msg, tid); 1778 1779 p = msg->front.iov_base; 1780 end = p + msg->front.iov_len; 1781 1782 ceph_decode_need(&p, end, 4, bad); 1783 object_len = ceph_decode_32(&p); 1784 ceph_decode_need(&p, end, object_len, bad); 1785 p += object_len; 1786 1787 err = ceph_decode_pgid(&p, end, &pg); 1788 if (err) 1789 goto bad; 1790 1791 ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad); 1792 flags = ceph_decode_64(&p); 1793 result = ceph_decode_32(&p); 1794 reassert_epoch = ceph_decode_32(&p); 1795 reassert_version = ceph_decode_64(&p); 1796 osdmap_epoch = ceph_decode_32(&p); 1797 1798 /* lookup */ 1799 down_read(&osdc->map_sem); 1800 mutex_lock(&osdc->request_mutex); 1801 req = __lookup_request(osdc, tid); 1802 if (req == NULL) { 1803 dout("handle_reply tid %llu dne\n", tid); 1804 goto bad_mutex; 1805 } 1806 ceph_osdc_get_request(req); 1807 1808 dout("handle_reply %p tid %llu req %p result %d\n", msg, tid, 1809 req, result); 1810 1811 ceph_decode_need(&p, end, 4, bad_put); 1812 numops = ceph_decode_32(&p); 1813 if (numops > CEPH_OSD_MAX_OP) 1814 goto bad_put; 1815 if (numops != req->r_num_ops) 1816 goto bad_put; 1817 payload_len = 0; 1818 ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad_put); 1819 for (i = 0; i < numops; i++) { 1820 struct ceph_osd_op *op = p; 1821 int len; 1822 1823 len = le32_to_cpu(op->payload_len); 1824 req->r_reply_op_len[i] = len; 1825 dout(" op %d has %d bytes\n", i, len); 1826 payload_len += len; 1827 p += sizeof(*op); 1828 } 1829 bytes = le32_to_cpu(msg->hdr.data_len); 1830 if (payload_len != bytes) { 1831 pr_warn("sum of op payload lens %d != data_len %d\n", 1832 payload_len, bytes); 1833 goto bad_put; 1834 } 1835 1836 ceph_decode_need(&p, end, 4 + numops * 4, bad_put); 1837 retry_attempt = ceph_decode_32(&p); 1838 for (i = 0; i < numops; i++) 1839 req->r_reply_op_result[i] = ceph_decode_32(&p); 1840 1841 if (le16_to_cpu(msg->hdr.version) >= 6) { 1842 p += 8 + 4; /* skip replay_version */ 1843 p += 8; /* skip user_version */ 1844 1845 if (le16_to_cpu(msg->hdr.version) >= 7) 1846 ceph_decode_8_safe(&p, end, decode_redir, bad_put); 1847 else 1848 decode_redir = 1; 1849 } else { 1850 decode_redir = 0; 1851 } 1852 1853 if (decode_redir) { 1854 err = ceph_redirect_decode(&p, end, &redir); 1855 if (err) 1856 goto bad_put; 1857 } else { 1858 redir.oloc.pool = -1; 1859 } 1860 1861 if (redir.oloc.pool != -1) { 1862 dout("redirect pool %lld\n", redir.oloc.pool); 1863 1864 __unregister_request(osdc, req); 1865 1866 req->r_target_oloc = redir.oloc; /* struct */ 1867 1868 /* 1869 * Start redirect requests with nofail=true. If 1870 * mapping fails, request will end up on the notarget 1871 * list, waiting for the new osdmap (which can take 1872 * a while), even though the original request mapped 1873 * successfully. In the future we might want to follow 1874 * original request's nofail setting here. 1875 */ 1876 err = __ceph_osdc_start_request(osdc, req, true); 1877 BUG_ON(err); 1878 1879 goto out_unlock; 1880 } 1881 1882 already_completed = req->r_got_reply; 1883 if (!req->r_got_reply) { 1884 req->r_result = result; 1885 dout("handle_reply result %d bytes %d\n", req->r_result, 1886 bytes); 1887 if (req->r_result == 0) 1888 req->r_result = bytes; 1889 1890 /* in case this is a write and we need to replay, */ 1891 req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch); 1892 req->r_reassert_version.version = cpu_to_le64(reassert_version); 1893 1894 req->r_got_reply = 1; 1895 } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) { 1896 dout("handle_reply tid %llu dup ack\n", tid); 1897 goto out_unlock; 1898 } 1899 1900 dout("handle_reply tid %llu flags %d\n", tid, flags); 1901 1902 if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK)) 1903 __register_linger_request(osdc, req); 1904 1905 /* either this is a read, or we got the safe response */ 1906 if (result < 0 || 1907 (flags & CEPH_OSD_FLAG_ONDISK) || 1908 ((flags & CEPH_OSD_FLAG_WRITE) == 0)) 1909 __unregister_request(osdc, req); 1910 1911 mutex_unlock(&osdc->request_mutex); 1912 up_read(&osdc->map_sem); 1913 1914 if (!already_completed) { 1915 if (req->r_unsafe_callback && 1916 result >= 0 && !(flags & CEPH_OSD_FLAG_ONDISK)) 1917 req->r_unsafe_callback(req, true); 1918 if (req->r_callback) 1919 req->r_callback(req, msg); 1920 else 1921 complete_all(&req->r_completion); 1922 } 1923 1924 if (flags & CEPH_OSD_FLAG_ONDISK) { 1925 if (req->r_unsafe_callback && already_completed) 1926 req->r_unsafe_callback(req, false); 1927 complete_request(req); 1928 } 1929 1930 out: 1931 dout("req=%p req->r_linger=%d\n", req, req->r_linger); 1932 ceph_osdc_put_request(req); 1933 return; 1934 out_unlock: 1935 mutex_unlock(&osdc->request_mutex); 1936 up_read(&osdc->map_sem); 1937 goto out; 1938 1939 bad_put: 1940 req->r_result = -EIO; 1941 __unregister_request(osdc, req); 1942 if (req->r_callback) 1943 req->r_callback(req, msg); 1944 else 1945 complete_all(&req->r_completion); 1946 complete_request(req); 1947 ceph_osdc_put_request(req); 1948 bad_mutex: 1949 mutex_unlock(&osdc->request_mutex); 1950 up_read(&osdc->map_sem); 1951 bad: 1952 pr_err("corrupt osd_op_reply got %d %d\n", 1953 (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len)); 1954 ceph_msg_dump(msg); 1955 } 1956 1957 static void reset_changed_osds(struct ceph_osd_client *osdc) 1958 { 1959 struct rb_node *p, *n; 1960 1961 dout("%s %p\n", __func__, osdc); 1962 for (p = rb_first(&osdc->osds); p; p = n) { 1963 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node); 1964 1965 n = rb_next(p); 1966 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) || 1967 memcmp(&osd->o_con.peer_addr, 1968 ceph_osd_addr(osdc->osdmap, 1969 osd->o_osd), 1970 sizeof(struct ceph_entity_addr)) != 0) 1971 __reset_osd(osdc, osd); 1972 } 1973 } 1974 1975 /* 1976 * Requeue requests whose mapping to an OSD has changed. If requests map to 1977 * no osd, request a new map. 1978 * 1979 * Caller should hold map_sem for read. 1980 */ 1981 static void kick_requests(struct ceph_osd_client *osdc, bool force_resend, 1982 bool force_resend_writes) 1983 { 1984 struct ceph_osd_request *req, *nreq; 1985 struct rb_node *p; 1986 int needmap = 0; 1987 int err; 1988 bool force_resend_req; 1989 1990 dout("kick_requests %s %s\n", force_resend ? " (force resend)" : "", 1991 force_resend_writes ? " (force resend writes)" : ""); 1992 mutex_lock(&osdc->request_mutex); 1993 for (p = rb_first(&osdc->requests); p; ) { 1994 req = rb_entry(p, struct ceph_osd_request, r_node); 1995 p = rb_next(p); 1996 1997 /* 1998 * For linger requests that have not yet been 1999 * registered, move them to the linger list; they'll 2000 * be sent to the osd in the loop below. Unregister 2001 * the request before re-registering it as a linger 2002 * request to ensure the __map_request() below 2003 * will decide it needs to be sent. 2004 */ 2005 if (req->r_linger && list_empty(&req->r_linger_item)) { 2006 dout("%p tid %llu restart on osd%d\n", 2007 req, req->r_tid, 2008 req->r_osd ? req->r_osd->o_osd : -1); 2009 ceph_osdc_get_request(req); 2010 __unregister_request(osdc, req); 2011 __register_linger_request(osdc, req); 2012 ceph_osdc_put_request(req); 2013 continue; 2014 } 2015 2016 force_resend_req = force_resend || 2017 (force_resend_writes && 2018 req->r_flags & CEPH_OSD_FLAG_WRITE); 2019 err = __map_request(osdc, req, force_resend_req); 2020 if (err < 0) 2021 continue; /* error */ 2022 if (req->r_osd == NULL) { 2023 dout("%p tid %llu maps to no osd\n", req, req->r_tid); 2024 needmap++; /* request a newer map */ 2025 } else if (err > 0) { 2026 if (!req->r_linger) { 2027 dout("%p tid %llu requeued on osd%d\n", req, 2028 req->r_tid, 2029 req->r_osd ? req->r_osd->o_osd : -1); 2030 req->r_flags |= CEPH_OSD_FLAG_RETRY; 2031 } 2032 } 2033 } 2034 2035 list_for_each_entry_safe(req, nreq, &osdc->req_linger, 2036 r_linger_item) { 2037 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd); 2038 2039 err = __map_request(osdc, req, 2040 force_resend || force_resend_writes); 2041 dout("__map_request returned %d\n", err); 2042 if (err < 0) 2043 continue; /* hrm! */ 2044 if (req->r_osd == NULL || err > 0) { 2045 if (req->r_osd == NULL) { 2046 dout("lingering %p tid %llu maps to no osd\n", 2047 req, req->r_tid); 2048 /* 2049 * A homeless lingering request makes 2050 * no sense, as it's job is to keep 2051 * a particular OSD connection open. 2052 * Request a newer map and kick the 2053 * request, knowing that it won't be 2054 * resent until we actually get a map 2055 * that can tell us where to send it. 2056 */ 2057 needmap++; 2058 } 2059 2060 dout("kicking lingering %p tid %llu osd%d\n", req, 2061 req->r_tid, req->r_osd ? req->r_osd->o_osd : -1); 2062 __register_request(osdc, req); 2063 __unregister_linger_request(osdc, req); 2064 } 2065 } 2066 reset_changed_osds(osdc); 2067 mutex_unlock(&osdc->request_mutex); 2068 2069 if (needmap) { 2070 dout("%d requests for down osds, need new map\n", needmap); 2071 ceph_monc_request_next_osdmap(&osdc->client->monc); 2072 } 2073 } 2074 2075 2076 /* 2077 * Process updated osd map. 2078 * 2079 * The message contains any number of incremental and full maps, normally 2080 * indicating some sort of topology change in the cluster. Kick requests 2081 * off to different OSDs as needed. 2082 */ 2083 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg) 2084 { 2085 void *p, *end, *next; 2086 u32 nr_maps, maplen; 2087 u32 epoch; 2088 struct ceph_osdmap *newmap = NULL, *oldmap; 2089 int err; 2090 struct ceph_fsid fsid; 2091 bool was_full; 2092 2093 dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0); 2094 p = msg->front.iov_base; 2095 end = p + msg->front.iov_len; 2096 2097 /* verify fsid */ 2098 ceph_decode_need(&p, end, sizeof(fsid), bad); 2099 ceph_decode_copy(&p, &fsid, sizeof(fsid)); 2100 if (ceph_check_fsid(osdc->client, &fsid) < 0) 2101 return; 2102 2103 down_write(&osdc->map_sem); 2104 2105 was_full = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL); 2106 2107 /* incremental maps */ 2108 ceph_decode_32_safe(&p, end, nr_maps, bad); 2109 dout(" %d inc maps\n", nr_maps); 2110 while (nr_maps > 0) { 2111 ceph_decode_need(&p, end, 2*sizeof(u32), bad); 2112 epoch = ceph_decode_32(&p); 2113 maplen = ceph_decode_32(&p); 2114 ceph_decode_need(&p, end, maplen, bad); 2115 next = p + maplen; 2116 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) { 2117 dout("applying incremental map %u len %d\n", 2118 epoch, maplen); 2119 newmap = osdmap_apply_incremental(&p, next, 2120 osdc->osdmap, 2121 &osdc->client->msgr); 2122 if (IS_ERR(newmap)) { 2123 err = PTR_ERR(newmap); 2124 goto bad; 2125 } 2126 BUG_ON(!newmap); 2127 if (newmap != osdc->osdmap) { 2128 ceph_osdmap_destroy(osdc->osdmap); 2129 osdc->osdmap = newmap; 2130 } 2131 was_full = was_full || 2132 ceph_osdmap_flag(osdc->osdmap, 2133 CEPH_OSDMAP_FULL); 2134 kick_requests(osdc, 0, was_full); 2135 } else { 2136 dout("ignoring incremental map %u len %d\n", 2137 epoch, maplen); 2138 } 2139 p = next; 2140 nr_maps--; 2141 } 2142 if (newmap) 2143 goto done; 2144 2145 /* full maps */ 2146 ceph_decode_32_safe(&p, end, nr_maps, bad); 2147 dout(" %d full maps\n", nr_maps); 2148 while (nr_maps) { 2149 ceph_decode_need(&p, end, 2*sizeof(u32), bad); 2150 epoch = ceph_decode_32(&p); 2151 maplen = ceph_decode_32(&p); 2152 ceph_decode_need(&p, end, maplen, bad); 2153 if (nr_maps > 1) { 2154 dout("skipping non-latest full map %u len %d\n", 2155 epoch, maplen); 2156 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) { 2157 dout("skipping full map %u len %d, " 2158 "older than our %u\n", epoch, maplen, 2159 osdc->osdmap->epoch); 2160 } else { 2161 int skipped_map = 0; 2162 2163 dout("taking full map %u len %d\n", epoch, maplen); 2164 newmap = ceph_osdmap_decode(&p, p+maplen); 2165 if (IS_ERR(newmap)) { 2166 err = PTR_ERR(newmap); 2167 goto bad; 2168 } 2169 BUG_ON(!newmap); 2170 oldmap = osdc->osdmap; 2171 osdc->osdmap = newmap; 2172 if (oldmap) { 2173 if (oldmap->epoch + 1 < newmap->epoch) 2174 skipped_map = 1; 2175 ceph_osdmap_destroy(oldmap); 2176 } 2177 was_full = was_full || 2178 ceph_osdmap_flag(osdc->osdmap, 2179 CEPH_OSDMAP_FULL); 2180 kick_requests(osdc, skipped_map, was_full); 2181 } 2182 p += maplen; 2183 nr_maps--; 2184 } 2185 2186 if (!osdc->osdmap) 2187 goto bad; 2188 done: 2189 downgrade_write(&osdc->map_sem); 2190 ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch); 2191 2192 /* 2193 * subscribe to subsequent osdmap updates if full to ensure 2194 * we find out when we are no longer full and stop returning 2195 * ENOSPC. 2196 */ 2197 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) || 2198 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD) || 2199 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR)) 2200 ceph_monc_request_next_osdmap(&osdc->client->monc); 2201 2202 mutex_lock(&osdc->request_mutex); 2203 __send_queued(osdc); 2204 mutex_unlock(&osdc->request_mutex); 2205 up_read(&osdc->map_sem); 2206 wake_up_all(&osdc->client->auth_wq); 2207 return; 2208 2209 bad: 2210 pr_err("osdc handle_map corrupt msg\n"); 2211 ceph_msg_dump(msg); 2212 up_write(&osdc->map_sem); 2213 } 2214 2215 /* 2216 * watch/notify callback event infrastructure 2217 * 2218 * These callbacks are used both for watch and notify operations. 2219 */ 2220 static void __release_event(struct kref *kref) 2221 { 2222 struct ceph_osd_event *event = 2223 container_of(kref, struct ceph_osd_event, kref); 2224 2225 dout("__release_event %p\n", event); 2226 kfree(event); 2227 } 2228 2229 static void get_event(struct ceph_osd_event *event) 2230 { 2231 kref_get(&event->kref); 2232 } 2233 2234 void ceph_osdc_put_event(struct ceph_osd_event *event) 2235 { 2236 kref_put(&event->kref, __release_event); 2237 } 2238 EXPORT_SYMBOL(ceph_osdc_put_event); 2239 2240 static void __insert_event(struct ceph_osd_client *osdc, 2241 struct ceph_osd_event *new) 2242 { 2243 struct rb_node **p = &osdc->event_tree.rb_node; 2244 struct rb_node *parent = NULL; 2245 struct ceph_osd_event *event = NULL; 2246 2247 while (*p) { 2248 parent = *p; 2249 event = rb_entry(parent, struct ceph_osd_event, node); 2250 if (new->cookie < event->cookie) 2251 p = &(*p)->rb_left; 2252 else if (new->cookie > event->cookie) 2253 p = &(*p)->rb_right; 2254 else 2255 BUG(); 2256 } 2257 2258 rb_link_node(&new->node, parent, p); 2259 rb_insert_color(&new->node, &osdc->event_tree); 2260 } 2261 2262 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc, 2263 u64 cookie) 2264 { 2265 struct rb_node **p = &osdc->event_tree.rb_node; 2266 struct rb_node *parent = NULL; 2267 struct ceph_osd_event *event = NULL; 2268 2269 while (*p) { 2270 parent = *p; 2271 event = rb_entry(parent, struct ceph_osd_event, node); 2272 if (cookie < event->cookie) 2273 p = &(*p)->rb_left; 2274 else if (cookie > event->cookie) 2275 p = &(*p)->rb_right; 2276 else 2277 return event; 2278 } 2279 return NULL; 2280 } 2281 2282 static void __remove_event(struct ceph_osd_event *event) 2283 { 2284 struct ceph_osd_client *osdc = event->osdc; 2285 2286 if (!RB_EMPTY_NODE(&event->node)) { 2287 dout("__remove_event removed %p\n", event); 2288 rb_erase(&event->node, &osdc->event_tree); 2289 ceph_osdc_put_event(event); 2290 } else { 2291 dout("__remove_event didn't remove %p\n", event); 2292 } 2293 } 2294 2295 int ceph_osdc_create_event(struct ceph_osd_client *osdc, 2296 void (*event_cb)(u64, u64, u8, void *), 2297 void *data, struct ceph_osd_event **pevent) 2298 { 2299 struct ceph_osd_event *event; 2300 2301 event = kmalloc(sizeof(*event), GFP_NOIO); 2302 if (!event) 2303 return -ENOMEM; 2304 2305 dout("create_event %p\n", event); 2306 event->cb = event_cb; 2307 event->one_shot = 0; 2308 event->data = data; 2309 event->osdc = osdc; 2310 INIT_LIST_HEAD(&event->osd_node); 2311 RB_CLEAR_NODE(&event->node); 2312 kref_init(&event->kref); /* one ref for us */ 2313 kref_get(&event->kref); /* one ref for the caller */ 2314 2315 spin_lock(&osdc->event_lock); 2316 event->cookie = ++osdc->event_count; 2317 __insert_event(osdc, event); 2318 spin_unlock(&osdc->event_lock); 2319 2320 *pevent = event; 2321 return 0; 2322 } 2323 EXPORT_SYMBOL(ceph_osdc_create_event); 2324 2325 void ceph_osdc_cancel_event(struct ceph_osd_event *event) 2326 { 2327 struct ceph_osd_client *osdc = event->osdc; 2328 2329 dout("cancel_event %p\n", event); 2330 spin_lock(&osdc->event_lock); 2331 __remove_event(event); 2332 spin_unlock(&osdc->event_lock); 2333 ceph_osdc_put_event(event); /* caller's */ 2334 } 2335 EXPORT_SYMBOL(ceph_osdc_cancel_event); 2336 2337 2338 static void do_event_work(struct work_struct *work) 2339 { 2340 struct ceph_osd_event_work *event_work = 2341 container_of(work, struct ceph_osd_event_work, work); 2342 struct ceph_osd_event *event = event_work->event; 2343 u64 ver = event_work->ver; 2344 u64 notify_id = event_work->notify_id; 2345 u8 opcode = event_work->opcode; 2346 2347 dout("do_event_work completing %p\n", event); 2348 event->cb(ver, notify_id, opcode, event->data); 2349 dout("do_event_work completed %p\n", event); 2350 ceph_osdc_put_event(event); 2351 kfree(event_work); 2352 } 2353 2354 2355 /* 2356 * Process osd watch notifications 2357 */ 2358 static void handle_watch_notify(struct ceph_osd_client *osdc, 2359 struct ceph_msg *msg) 2360 { 2361 void *p, *end; 2362 u8 proto_ver; 2363 u64 cookie, ver, notify_id; 2364 u8 opcode; 2365 struct ceph_osd_event *event; 2366 struct ceph_osd_event_work *event_work; 2367 2368 p = msg->front.iov_base; 2369 end = p + msg->front.iov_len; 2370 2371 ceph_decode_8_safe(&p, end, proto_ver, bad); 2372 ceph_decode_8_safe(&p, end, opcode, bad); 2373 ceph_decode_64_safe(&p, end, cookie, bad); 2374 ceph_decode_64_safe(&p, end, ver, bad); 2375 ceph_decode_64_safe(&p, end, notify_id, bad); 2376 2377 spin_lock(&osdc->event_lock); 2378 event = __find_event(osdc, cookie); 2379 if (event) { 2380 BUG_ON(event->one_shot); 2381 get_event(event); 2382 } 2383 spin_unlock(&osdc->event_lock); 2384 dout("handle_watch_notify cookie %lld ver %lld event %p\n", 2385 cookie, ver, event); 2386 if (event) { 2387 event_work = kmalloc(sizeof(*event_work), GFP_NOIO); 2388 if (!event_work) { 2389 pr_err("couldn't allocate event_work\n"); 2390 ceph_osdc_put_event(event); 2391 return; 2392 } 2393 INIT_WORK(&event_work->work, do_event_work); 2394 event_work->event = event; 2395 event_work->ver = ver; 2396 event_work->notify_id = notify_id; 2397 event_work->opcode = opcode; 2398 2399 queue_work(osdc->notify_wq, &event_work->work); 2400 } 2401 2402 return; 2403 2404 bad: 2405 pr_err("osdc handle_watch_notify corrupt msg\n"); 2406 } 2407 2408 /* 2409 * build new request AND message 2410 * 2411 */ 2412 void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off, 2413 struct ceph_snap_context *snapc, u64 snap_id, 2414 struct timespec *mtime) 2415 { 2416 struct ceph_msg *msg = req->r_request; 2417 void *p; 2418 size_t msg_size; 2419 int flags = req->r_flags; 2420 u64 data_len; 2421 unsigned int i; 2422 2423 req->r_snapid = snap_id; 2424 req->r_snapc = ceph_get_snap_context(snapc); 2425 2426 /* encode request */ 2427 msg->hdr.version = cpu_to_le16(4); 2428 2429 p = msg->front.iov_base; 2430 ceph_encode_32(&p, 1); /* client_inc is always 1 */ 2431 req->r_request_osdmap_epoch = p; 2432 p += 4; 2433 req->r_request_flags = p; 2434 p += 4; 2435 if (req->r_flags & CEPH_OSD_FLAG_WRITE) 2436 ceph_encode_timespec(p, mtime); 2437 p += sizeof(struct ceph_timespec); 2438 req->r_request_reassert_version = p; 2439 p += sizeof(struct ceph_eversion); /* will get filled in */ 2440 2441 /* oloc */ 2442 ceph_encode_8(&p, 4); 2443 ceph_encode_8(&p, 4); 2444 ceph_encode_32(&p, 8 + 4 + 4); 2445 req->r_request_pool = p; 2446 p += 8; 2447 ceph_encode_32(&p, -1); /* preferred */ 2448 ceph_encode_32(&p, 0); /* key len */ 2449 2450 ceph_encode_8(&p, 1); 2451 req->r_request_pgid = p; 2452 p += 8 + 4; 2453 ceph_encode_32(&p, -1); /* preferred */ 2454 2455 /* oid */ 2456 ceph_encode_32(&p, req->r_base_oid.name_len); 2457 memcpy(p, req->r_base_oid.name, req->r_base_oid.name_len); 2458 dout("oid '%.*s' len %d\n", req->r_base_oid.name_len, 2459 req->r_base_oid.name, req->r_base_oid.name_len); 2460 p += req->r_base_oid.name_len; 2461 2462 /* ops--can imply data */ 2463 ceph_encode_16(&p, (u16)req->r_num_ops); 2464 data_len = 0; 2465 for (i = 0; i < req->r_num_ops; i++) { 2466 data_len += osd_req_encode_op(req, p, i); 2467 p += sizeof(struct ceph_osd_op); 2468 } 2469 2470 /* snaps */ 2471 ceph_encode_64(&p, req->r_snapid); 2472 ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0); 2473 ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0); 2474 if (req->r_snapc) { 2475 for (i = 0; i < snapc->num_snaps; i++) { 2476 ceph_encode_64(&p, req->r_snapc->snaps[i]); 2477 } 2478 } 2479 2480 req->r_request_attempts = p; 2481 p += 4; 2482 2483 /* data */ 2484 if (flags & CEPH_OSD_FLAG_WRITE) { 2485 u16 data_off; 2486 2487 /* 2488 * The header "data_off" is a hint to the receiver 2489 * allowing it to align received data into its 2490 * buffers such that there's no need to re-copy 2491 * it before writing it to disk (direct I/O). 2492 */ 2493 data_off = (u16) (off & 0xffff); 2494 req->r_request->hdr.data_off = cpu_to_le16(data_off); 2495 } 2496 req->r_request->hdr.data_len = cpu_to_le32(data_len); 2497 2498 BUG_ON(p > msg->front.iov_base + msg->front.iov_len); 2499 msg_size = p - msg->front.iov_base; 2500 msg->front.iov_len = msg_size; 2501 msg->hdr.front_len = cpu_to_le32(msg_size); 2502 2503 dout("build_request msg_size was %d\n", (int)msg_size); 2504 } 2505 EXPORT_SYMBOL(ceph_osdc_build_request); 2506 2507 /* 2508 * Register request, send initial attempt. 2509 */ 2510 int ceph_osdc_start_request(struct ceph_osd_client *osdc, 2511 struct ceph_osd_request *req, 2512 bool nofail) 2513 { 2514 int rc; 2515 2516 down_read(&osdc->map_sem); 2517 mutex_lock(&osdc->request_mutex); 2518 2519 rc = __ceph_osdc_start_request(osdc, req, nofail); 2520 2521 mutex_unlock(&osdc->request_mutex); 2522 up_read(&osdc->map_sem); 2523 2524 return rc; 2525 } 2526 EXPORT_SYMBOL(ceph_osdc_start_request); 2527 2528 /* 2529 * Unregister a registered request. The request is not completed (i.e. 2530 * no callbacks or wakeups) - higher layers are supposed to know what 2531 * they are canceling. 2532 */ 2533 void ceph_osdc_cancel_request(struct ceph_osd_request *req) 2534 { 2535 struct ceph_osd_client *osdc = req->r_osdc; 2536 2537 mutex_lock(&osdc->request_mutex); 2538 if (req->r_linger) 2539 __unregister_linger_request(osdc, req); 2540 __unregister_request(osdc, req); 2541 mutex_unlock(&osdc->request_mutex); 2542 2543 dout("%s %p tid %llu canceled\n", __func__, req, req->r_tid); 2544 } 2545 EXPORT_SYMBOL(ceph_osdc_cancel_request); 2546 2547 /* 2548 * wait for a request to complete 2549 */ 2550 int ceph_osdc_wait_request(struct ceph_osd_client *osdc, 2551 struct ceph_osd_request *req) 2552 { 2553 int rc; 2554 2555 dout("%s %p tid %llu\n", __func__, req, req->r_tid); 2556 2557 rc = wait_for_completion_interruptible(&req->r_completion); 2558 if (rc < 0) { 2559 dout("%s %p tid %llu interrupted\n", __func__, req, req->r_tid); 2560 ceph_osdc_cancel_request(req); 2561 complete_request(req); 2562 return rc; 2563 } 2564 2565 dout("%s %p tid %llu result %d\n", __func__, req, req->r_tid, 2566 req->r_result); 2567 return req->r_result; 2568 } 2569 EXPORT_SYMBOL(ceph_osdc_wait_request); 2570 2571 /* 2572 * sync - wait for all in-flight requests to flush. avoid starvation. 2573 */ 2574 void ceph_osdc_sync(struct ceph_osd_client *osdc) 2575 { 2576 struct ceph_osd_request *req; 2577 u64 last_tid, next_tid = 0; 2578 2579 mutex_lock(&osdc->request_mutex); 2580 last_tid = osdc->last_tid; 2581 while (1) { 2582 req = __lookup_request_ge(osdc, next_tid); 2583 if (!req) 2584 break; 2585 if (req->r_tid > last_tid) 2586 break; 2587 2588 next_tid = req->r_tid + 1; 2589 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0) 2590 continue; 2591 2592 ceph_osdc_get_request(req); 2593 mutex_unlock(&osdc->request_mutex); 2594 dout("sync waiting on tid %llu (last is %llu)\n", 2595 req->r_tid, last_tid); 2596 wait_for_completion(&req->r_safe_completion); 2597 mutex_lock(&osdc->request_mutex); 2598 ceph_osdc_put_request(req); 2599 } 2600 mutex_unlock(&osdc->request_mutex); 2601 dout("sync done (thru tid %llu)\n", last_tid); 2602 } 2603 EXPORT_SYMBOL(ceph_osdc_sync); 2604 2605 /* 2606 * Call all pending notify callbacks - for use after a watch is 2607 * unregistered, to make sure no more callbacks for it will be invoked 2608 */ 2609 void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc) 2610 { 2611 flush_workqueue(osdc->notify_wq); 2612 } 2613 EXPORT_SYMBOL(ceph_osdc_flush_notifies); 2614 2615 2616 /* 2617 * init, shutdown 2618 */ 2619 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client) 2620 { 2621 int err; 2622 2623 dout("init\n"); 2624 osdc->client = client; 2625 osdc->osdmap = NULL; 2626 init_rwsem(&osdc->map_sem); 2627 init_completion(&osdc->map_waiters); 2628 osdc->last_requested_map = 0; 2629 mutex_init(&osdc->request_mutex); 2630 osdc->last_tid = 0; 2631 osdc->osds = RB_ROOT; 2632 INIT_LIST_HEAD(&osdc->osd_lru); 2633 osdc->requests = RB_ROOT; 2634 INIT_LIST_HEAD(&osdc->req_lru); 2635 INIT_LIST_HEAD(&osdc->req_unsent); 2636 INIT_LIST_HEAD(&osdc->req_notarget); 2637 INIT_LIST_HEAD(&osdc->req_linger); 2638 osdc->num_requests = 0; 2639 INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout); 2640 INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout); 2641 spin_lock_init(&osdc->event_lock); 2642 osdc->event_tree = RB_ROOT; 2643 osdc->event_count = 0; 2644 2645 schedule_delayed_work(&osdc->osds_timeout_work, 2646 round_jiffies_relative(osdc->client->options->osd_idle_ttl)); 2647 2648 err = -ENOMEM; 2649 osdc->req_mempool = mempool_create_kmalloc_pool(10, 2650 sizeof(struct ceph_osd_request)); 2651 if (!osdc->req_mempool) 2652 goto out; 2653 2654 err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP, 2655 OSD_OP_FRONT_LEN, 10, true, 2656 "osd_op"); 2657 if (err < 0) 2658 goto out_mempool; 2659 err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY, 2660 OSD_OPREPLY_FRONT_LEN, 10, true, 2661 "osd_op_reply"); 2662 if (err < 0) 2663 goto out_msgpool; 2664 2665 err = -ENOMEM; 2666 osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify"); 2667 if (!osdc->notify_wq) 2668 goto out_msgpool_reply; 2669 2670 return 0; 2671 2672 out_msgpool_reply: 2673 ceph_msgpool_destroy(&osdc->msgpool_op_reply); 2674 out_msgpool: 2675 ceph_msgpool_destroy(&osdc->msgpool_op); 2676 out_mempool: 2677 mempool_destroy(osdc->req_mempool); 2678 out: 2679 return err; 2680 } 2681 2682 void ceph_osdc_stop(struct ceph_osd_client *osdc) 2683 { 2684 flush_workqueue(osdc->notify_wq); 2685 destroy_workqueue(osdc->notify_wq); 2686 cancel_delayed_work_sync(&osdc->timeout_work); 2687 cancel_delayed_work_sync(&osdc->osds_timeout_work); 2688 if (osdc->osdmap) { 2689 ceph_osdmap_destroy(osdc->osdmap); 2690 osdc->osdmap = NULL; 2691 } 2692 remove_all_osds(osdc); 2693 mempool_destroy(osdc->req_mempool); 2694 ceph_msgpool_destroy(&osdc->msgpool_op); 2695 ceph_msgpool_destroy(&osdc->msgpool_op_reply); 2696 } 2697 2698 /* 2699 * Read some contiguous pages. If we cross a stripe boundary, shorten 2700 * *plen. Return number of bytes read, or error. 2701 */ 2702 int ceph_osdc_readpages(struct ceph_osd_client *osdc, 2703 struct ceph_vino vino, struct ceph_file_layout *layout, 2704 u64 off, u64 *plen, 2705 u32 truncate_seq, u64 truncate_size, 2706 struct page **pages, int num_pages, int page_align) 2707 { 2708 struct ceph_osd_request *req; 2709 int rc = 0; 2710 2711 dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino, 2712 vino.snap, off, *plen); 2713 req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 0, 1, 2714 CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ, 2715 NULL, truncate_seq, truncate_size, 2716 false); 2717 if (IS_ERR(req)) 2718 return PTR_ERR(req); 2719 2720 /* it may be a short read due to an object boundary */ 2721 2722 osd_req_op_extent_osd_data_pages(req, 0, 2723 pages, *plen, page_align, false, false); 2724 2725 dout("readpages final extent is %llu~%llu (%llu bytes align %d)\n", 2726 off, *plen, *plen, page_align); 2727 2728 ceph_osdc_build_request(req, off, NULL, vino.snap, NULL); 2729 2730 rc = ceph_osdc_start_request(osdc, req, false); 2731 if (!rc) 2732 rc = ceph_osdc_wait_request(osdc, req); 2733 2734 ceph_osdc_put_request(req); 2735 dout("readpages result %d\n", rc); 2736 return rc; 2737 } 2738 EXPORT_SYMBOL(ceph_osdc_readpages); 2739 2740 /* 2741 * do a synchronous write on N pages 2742 */ 2743 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino, 2744 struct ceph_file_layout *layout, 2745 struct ceph_snap_context *snapc, 2746 u64 off, u64 len, 2747 u32 truncate_seq, u64 truncate_size, 2748 struct timespec *mtime, 2749 struct page **pages, int num_pages) 2750 { 2751 struct ceph_osd_request *req; 2752 int rc = 0; 2753 int page_align = off & ~PAGE_MASK; 2754 2755 BUG_ON(vino.snap != CEPH_NOSNAP); /* snapshots aren't writeable */ 2756 req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 0, 1, 2757 CEPH_OSD_OP_WRITE, 2758 CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE, 2759 snapc, truncate_seq, truncate_size, 2760 true); 2761 if (IS_ERR(req)) 2762 return PTR_ERR(req); 2763 2764 /* it may be a short write due to an object boundary */ 2765 osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align, 2766 false, false); 2767 dout("writepages %llu~%llu (%llu bytes)\n", off, len, len); 2768 2769 ceph_osdc_build_request(req, off, snapc, CEPH_NOSNAP, mtime); 2770 2771 rc = ceph_osdc_start_request(osdc, req, true); 2772 if (!rc) 2773 rc = ceph_osdc_wait_request(osdc, req); 2774 2775 ceph_osdc_put_request(req); 2776 if (rc == 0) 2777 rc = len; 2778 dout("writepages result %d\n", rc); 2779 return rc; 2780 } 2781 EXPORT_SYMBOL(ceph_osdc_writepages); 2782 2783 int ceph_osdc_setup(void) 2784 { 2785 BUG_ON(ceph_osd_request_cache); 2786 ceph_osd_request_cache = kmem_cache_create("ceph_osd_request", 2787 sizeof (struct ceph_osd_request), 2788 __alignof__(struct ceph_osd_request), 2789 0, NULL); 2790 2791 return ceph_osd_request_cache ? 0 : -ENOMEM; 2792 } 2793 EXPORT_SYMBOL(ceph_osdc_setup); 2794 2795 void ceph_osdc_cleanup(void) 2796 { 2797 BUG_ON(!ceph_osd_request_cache); 2798 kmem_cache_destroy(ceph_osd_request_cache); 2799 ceph_osd_request_cache = NULL; 2800 } 2801 EXPORT_SYMBOL(ceph_osdc_cleanup); 2802 2803 /* 2804 * handle incoming message 2805 */ 2806 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg) 2807 { 2808 struct ceph_osd *osd = con->private; 2809 struct ceph_osd_client *osdc; 2810 int type = le16_to_cpu(msg->hdr.type); 2811 2812 if (!osd) 2813 goto out; 2814 osdc = osd->o_osdc; 2815 2816 switch (type) { 2817 case CEPH_MSG_OSD_MAP: 2818 ceph_osdc_handle_map(osdc, msg); 2819 break; 2820 case CEPH_MSG_OSD_OPREPLY: 2821 handle_reply(osdc, msg); 2822 break; 2823 case CEPH_MSG_WATCH_NOTIFY: 2824 handle_watch_notify(osdc, msg); 2825 break; 2826 2827 default: 2828 pr_err("received unknown message type %d %s\n", type, 2829 ceph_msg_type_name(type)); 2830 } 2831 out: 2832 ceph_msg_put(msg); 2833 } 2834 2835 /* 2836 * Lookup and return message for incoming reply. Don't try to do 2837 * anything about a larger than preallocated data portion of the 2838 * message at the moment - for now, just skip the message. 2839 */ 2840 static struct ceph_msg *get_reply(struct ceph_connection *con, 2841 struct ceph_msg_header *hdr, 2842 int *skip) 2843 { 2844 struct ceph_osd *osd = con->private; 2845 struct ceph_osd_client *osdc = osd->o_osdc; 2846 struct ceph_msg *m; 2847 struct ceph_osd_request *req; 2848 int front_len = le32_to_cpu(hdr->front_len); 2849 int data_len = le32_to_cpu(hdr->data_len); 2850 u64 tid; 2851 2852 tid = le64_to_cpu(hdr->tid); 2853 mutex_lock(&osdc->request_mutex); 2854 req = __lookup_request(osdc, tid); 2855 if (!req) { 2856 pr_warn("%s osd%d tid %llu unknown, skipping\n", 2857 __func__, osd->o_osd, tid); 2858 m = NULL; 2859 *skip = 1; 2860 goto out; 2861 } 2862 2863 ceph_msg_revoke_incoming(req->r_reply); 2864 2865 if (front_len > req->r_reply->front_alloc_len) { 2866 pr_warn("%s osd%d tid %llu front %d > preallocated %d\n", 2867 __func__, osd->o_osd, req->r_tid, front_len, 2868 req->r_reply->front_alloc_len); 2869 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front_len, GFP_NOFS, 2870 false); 2871 if (!m) 2872 goto out; 2873 ceph_msg_put(req->r_reply); 2874 req->r_reply = m; 2875 } 2876 2877 if (data_len > req->r_reply->data_length) { 2878 pr_warn("%s osd%d tid %llu data %d > preallocated %zu, skipping\n", 2879 __func__, osd->o_osd, req->r_tid, data_len, 2880 req->r_reply->data_length); 2881 m = NULL; 2882 *skip = 1; 2883 goto out; 2884 } 2885 2886 m = ceph_msg_get(req->r_reply); 2887 dout("get_reply tid %lld %p\n", tid, m); 2888 2889 out: 2890 mutex_unlock(&osdc->request_mutex); 2891 return m; 2892 } 2893 2894 static struct ceph_msg *alloc_msg(struct ceph_connection *con, 2895 struct ceph_msg_header *hdr, 2896 int *skip) 2897 { 2898 struct ceph_osd *osd = con->private; 2899 int type = le16_to_cpu(hdr->type); 2900 int front = le32_to_cpu(hdr->front_len); 2901 2902 *skip = 0; 2903 switch (type) { 2904 case CEPH_MSG_OSD_MAP: 2905 case CEPH_MSG_WATCH_NOTIFY: 2906 return ceph_msg_new(type, front, GFP_NOFS, false); 2907 case CEPH_MSG_OSD_OPREPLY: 2908 return get_reply(con, hdr, skip); 2909 default: 2910 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type, 2911 osd->o_osd); 2912 *skip = 1; 2913 return NULL; 2914 } 2915 } 2916 2917 /* 2918 * Wrappers to refcount containing ceph_osd struct 2919 */ 2920 static struct ceph_connection *get_osd_con(struct ceph_connection *con) 2921 { 2922 struct ceph_osd *osd = con->private; 2923 if (get_osd(osd)) 2924 return con; 2925 return NULL; 2926 } 2927 2928 static void put_osd_con(struct ceph_connection *con) 2929 { 2930 struct ceph_osd *osd = con->private; 2931 put_osd(osd); 2932 } 2933 2934 /* 2935 * authentication 2936 */ 2937 /* 2938 * Note: returned pointer is the address of a structure that's 2939 * managed separately. Caller must *not* attempt to free it. 2940 */ 2941 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con, 2942 int *proto, int force_new) 2943 { 2944 struct ceph_osd *o = con->private; 2945 struct ceph_osd_client *osdc = o->o_osdc; 2946 struct ceph_auth_client *ac = osdc->client->monc.auth; 2947 struct ceph_auth_handshake *auth = &o->o_auth; 2948 2949 if (force_new && auth->authorizer) { 2950 ceph_auth_destroy_authorizer(ac, auth->authorizer); 2951 auth->authorizer = NULL; 2952 } 2953 if (!auth->authorizer) { 2954 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 2955 auth); 2956 if (ret) 2957 return ERR_PTR(ret); 2958 } else { 2959 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 2960 auth); 2961 if (ret) 2962 return ERR_PTR(ret); 2963 } 2964 *proto = ac->protocol; 2965 2966 return auth; 2967 } 2968 2969 2970 static int verify_authorizer_reply(struct ceph_connection *con, int len) 2971 { 2972 struct ceph_osd *o = con->private; 2973 struct ceph_osd_client *osdc = o->o_osdc; 2974 struct ceph_auth_client *ac = osdc->client->monc.auth; 2975 2976 return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len); 2977 } 2978 2979 static int invalidate_authorizer(struct ceph_connection *con) 2980 { 2981 struct ceph_osd *o = con->private; 2982 struct ceph_osd_client *osdc = o->o_osdc; 2983 struct ceph_auth_client *ac = osdc->client->monc.auth; 2984 2985 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD); 2986 return ceph_monc_validate_auth(&osdc->client->monc); 2987 } 2988 2989 static int osd_sign_message(struct ceph_msg *msg) 2990 { 2991 struct ceph_osd *o = msg->con->private; 2992 struct ceph_auth_handshake *auth = &o->o_auth; 2993 2994 return ceph_auth_sign_message(auth, msg); 2995 } 2996 2997 static int osd_check_message_signature(struct ceph_msg *msg) 2998 { 2999 struct ceph_osd *o = msg->con->private; 3000 struct ceph_auth_handshake *auth = &o->o_auth; 3001 3002 return ceph_auth_check_message_signature(auth, msg); 3003 } 3004 3005 static const struct ceph_connection_operations osd_con_ops = { 3006 .get = get_osd_con, 3007 .put = put_osd_con, 3008 .dispatch = dispatch, 3009 .get_authorizer = get_authorizer, 3010 .verify_authorizer_reply = verify_authorizer_reply, 3011 .invalidate_authorizer = invalidate_authorizer, 3012 .alloc_msg = alloc_msg, 3013 .sign_message = osd_sign_message, 3014 .check_message_signature = osd_check_message_signature, 3015 .fault = osd_reset, 3016 }; 3017