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 const struct ceph_connection_operations osd_con_ops; 26 27 static void __send_queued(struct ceph_osd_client *osdc); 28 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd); 29 static void __register_request(struct ceph_osd_client *osdc, 30 struct ceph_osd_request *req); 31 static void __unregister_linger_request(struct ceph_osd_client *osdc, 32 struct ceph_osd_request *req); 33 static void __send_request(struct ceph_osd_client *osdc, 34 struct ceph_osd_request *req); 35 36 /* 37 * Implement client access to distributed object storage cluster. 38 * 39 * All data objects are stored within a cluster/cloud of OSDs, or 40 * "object storage devices." (Note that Ceph OSDs have _nothing_ to 41 * do with the T10 OSD extensions to SCSI.) Ceph OSDs are simply 42 * remote daemons serving up and coordinating consistent and safe 43 * access to storage. 44 * 45 * Cluster membership and the mapping of data objects onto storage devices 46 * are described by the osd map. 47 * 48 * We keep track of pending OSD requests (read, write), resubmit 49 * requests to different OSDs when the cluster topology/data layout 50 * change, or retry the affected requests when the communications 51 * channel with an OSD is reset. 52 */ 53 54 /* 55 * calculate the mapping of a file extent onto an object, and fill out the 56 * request accordingly. shorten extent as necessary if it crosses an 57 * object boundary. 58 * 59 * fill osd op in request message. 60 */ 61 static int calc_layout(struct ceph_file_layout *layout, u64 off, u64 *plen, 62 u64 *objnum, u64 *objoff, u64 *objlen) 63 { 64 u64 orig_len = *plen; 65 int r; 66 67 /* object extent? */ 68 r = ceph_calc_file_object_mapping(layout, off, orig_len, objnum, 69 objoff, objlen); 70 if (r < 0) 71 return r; 72 if (*objlen < orig_len) { 73 *plen = *objlen; 74 dout(" skipping last %llu, final file extent %llu~%llu\n", 75 orig_len - *plen, off, *plen); 76 } 77 78 dout("calc_layout objnum=%llx %llu~%llu\n", *objnum, *objoff, *objlen); 79 80 return 0; 81 } 82 83 static void ceph_osd_data_init(struct ceph_osd_data *osd_data) 84 { 85 memset(osd_data, 0, sizeof (*osd_data)); 86 osd_data->type = CEPH_OSD_DATA_TYPE_NONE; 87 } 88 89 static void ceph_osd_data_pages_init(struct ceph_osd_data *osd_data, 90 struct page **pages, u64 length, u32 alignment, 91 bool pages_from_pool, bool own_pages) 92 { 93 osd_data->type = CEPH_OSD_DATA_TYPE_PAGES; 94 osd_data->pages = pages; 95 osd_data->length = length; 96 osd_data->alignment = alignment; 97 osd_data->pages_from_pool = pages_from_pool; 98 osd_data->own_pages = own_pages; 99 } 100 101 static void ceph_osd_data_pagelist_init(struct ceph_osd_data *osd_data, 102 struct ceph_pagelist *pagelist) 103 { 104 osd_data->type = CEPH_OSD_DATA_TYPE_PAGELIST; 105 osd_data->pagelist = pagelist; 106 } 107 108 #ifdef CONFIG_BLOCK 109 static void ceph_osd_data_bio_init(struct ceph_osd_data *osd_data, 110 struct bio *bio, size_t bio_length) 111 { 112 osd_data->type = CEPH_OSD_DATA_TYPE_BIO; 113 osd_data->bio = bio; 114 osd_data->bio_length = bio_length; 115 } 116 #endif /* CONFIG_BLOCK */ 117 118 struct ceph_osd_data * 119 osd_req_op_extent_osd_data(struct ceph_osd_request *osd_req, 120 unsigned int which) 121 { 122 BUG_ON(which >= osd_req->r_num_ops); 123 124 return &osd_req->r_ops[which].extent.osd_data; 125 } 126 EXPORT_SYMBOL(osd_req_op_extent_osd_data); 127 128 struct ceph_osd_data * 129 osd_req_op_cls_request_info(struct ceph_osd_request *osd_req, 130 unsigned int which) 131 { 132 BUG_ON(which >= osd_req->r_num_ops); 133 134 return &osd_req->r_ops[which].cls.request_info; 135 } 136 EXPORT_SYMBOL(osd_req_op_cls_request_info); /* ??? */ 137 138 struct ceph_osd_data * 139 osd_req_op_cls_request_data(struct ceph_osd_request *osd_req, 140 unsigned int which) 141 { 142 BUG_ON(which >= osd_req->r_num_ops); 143 144 return &osd_req->r_ops[which].cls.request_data; 145 } 146 EXPORT_SYMBOL(osd_req_op_cls_request_data); /* ??? */ 147 148 struct ceph_osd_data * 149 osd_req_op_cls_response_data(struct ceph_osd_request *osd_req, 150 unsigned int which) 151 { 152 BUG_ON(which >= osd_req->r_num_ops); 153 154 return &osd_req->r_ops[which].cls.response_data; 155 } 156 EXPORT_SYMBOL(osd_req_op_cls_response_data); /* ??? */ 157 158 void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *osd_req, 159 unsigned int which, struct page **pages, 160 u64 length, u32 alignment, 161 bool pages_from_pool, bool own_pages) 162 { 163 struct ceph_osd_data *osd_data; 164 165 osd_data = osd_req_op_extent_osd_data(osd_req, which); 166 ceph_osd_data_pages_init(osd_data, pages, length, alignment, 167 pages_from_pool, own_pages); 168 } 169 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pages); 170 171 void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *osd_req, 172 unsigned int which, struct ceph_pagelist *pagelist) 173 { 174 struct ceph_osd_data *osd_data; 175 176 osd_data = osd_req_op_extent_osd_data(osd_req, which); 177 ceph_osd_data_pagelist_init(osd_data, pagelist); 178 } 179 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pagelist); 180 181 #ifdef CONFIG_BLOCK 182 void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *osd_req, 183 unsigned int which, struct bio *bio, size_t bio_length) 184 { 185 struct ceph_osd_data *osd_data; 186 osd_data = osd_req_op_extent_osd_data(osd_req, which); 187 ceph_osd_data_bio_init(osd_data, bio, bio_length); 188 } 189 EXPORT_SYMBOL(osd_req_op_extent_osd_data_bio); 190 #endif /* CONFIG_BLOCK */ 191 192 static void osd_req_op_cls_request_info_pagelist( 193 struct ceph_osd_request *osd_req, 194 unsigned int which, struct ceph_pagelist *pagelist) 195 { 196 struct ceph_osd_data *osd_data; 197 198 osd_data = osd_req_op_cls_request_info(osd_req, which); 199 ceph_osd_data_pagelist_init(osd_data, pagelist); 200 } 201 202 void osd_req_op_cls_request_data_pagelist( 203 struct ceph_osd_request *osd_req, 204 unsigned int which, struct ceph_pagelist *pagelist) 205 { 206 struct ceph_osd_data *osd_data; 207 208 osd_data = osd_req_op_cls_request_data(osd_req, which); 209 ceph_osd_data_pagelist_init(osd_data, pagelist); 210 } 211 EXPORT_SYMBOL(osd_req_op_cls_request_data_pagelist); 212 213 void osd_req_op_cls_response_data_pages(struct ceph_osd_request *osd_req, 214 unsigned int which, struct page **pages, u64 length, 215 u32 alignment, bool pages_from_pool, bool own_pages) 216 { 217 struct ceph_osd_data *osd_data; 218 219 osd_data = osd_req_op_cls_response_data(osd_req, which); 220 ceph_osd_data_pages_init(osd_data, pages, length, alignment, 221 pages_from_pool, own_pages); 222 } 223 EXPORT_SYMBOL(osd_req_op_cls_response_data_pages); 224 225 static u64 ceph_osd_data_length(struct ceph_osd_data *osd_data) 226 { 227 switch (osd_data->type) { 228 case CEPH_OSD_DATA_TYPE_NONE: 229 return 0; 230 case CEPH_OSD_DATA_TYPE_PAGES: 231 return osd_data->length; 232 case CEPH_OSD_DATA_TYPE_PAGELIST: 233 return (u64)osd_data->pagelist->length; 234 #ifdef CONFIG_BLOCK 235 case CEPH_OSD_DATA_TYPE_BIO: 236 return (u64)osd_data->bio_length; 237 #endif /* CONFIG_BLOCK */ 238 default: 239 WARN(true, "unrecognized data type %d\n", (int)osd_data->type); 240 return 0; 241 } 242 } 243 244 245 static void ceph_osd_data_release(struct ceph_osd_data *osd_data) 246 { 247 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES && osd_data->own_pages) { 248 int num_pages; 249 250 num_pages = calc_pages_for((u64)osd_data->alignment, 251 (u64)osd_data->length); 252 ceph_release_page_vector(osd_data->pages, num_pages); 253 } 254 ceph_osd_data_init(osd_data); 255 } 256 257 static void osd_req_op_data_release(struct ceph_osd_request *osd_req, 258 unsigned int which) 259 { 260 struct ceph_osd_req_op *op; 261 262 BUG_ON(which >= osd_req->r_num_ops); 263 op = &osd_req->r_ops[which]; 264 265 switch (op->op) { 266 case CEPH_OSD_OP_READ: 267 case CEPH_OSD_OP_WRITE: 268 ceph_osd_data_release(&op->extent.osd_data); 269 break; 270 case CEPH_OSD_OP_CALL: 271 ceph_osd_data_release(&op->cls.request_info); 272 ceph_osd_data_release(&op->cls.request_data); 273 ceph_osd_data_release(&op->cls.response_data); 274 break; 275 default: 276 break; 277 } 278 } 279 280 /* 281 * requests 282 */ 283 void ceph_osdc_release_request(struct kref *kref) 284 { 285 struct ceph_osd_request *req; 286 unsigned int which; 287 288 req = container_of(kref, struct ceph_osd_request, r_kref); 289 if (req->r_request) 290 ceph_msg_put(req->r_request); 291 if (req->r_reply) { 292 ceph_msg_revoke_incoming(req->r_reply); 293 ceph_msg_put(req->r_reply); 294 } 295 296 for (which = 0; which < req->r_num_ops; which++) 297 osd_req_op_data_release(req, which); 298 299 ceph_put_snap_context(req->r_snapc); 300 if (req->r_mempool) 301 mempool_free(req, req->r_osdc->req_mempool); 302 else 303 kfree(req); 304 } 305 EXPORT_SYMBOL(ceph_osdc_release_request); 306 307 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc, 308 struct ceph_snap_context *snapc, 309 unsigned int num_ops, 310 bool use_mempool, 311 gfp_t gfp_flags) 312 { 313 struct ceph_osd_request *req; 314 struct ceph_msg *msg; 315 size_t msg_size; 316 317 BUILD_BUG_ON(CEPH_OSD_MAX_OP > U16_MAX); 318 BUG_ON(num_ops > CEPH_OSD_MAX_OP); 319 320 msg_size = 4 + 4 + 8 + 8 + 4+8; 321 msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */ 322 msg_size += 1 + 8 + 4 + 4; /* pg_t */ 323 msg_size += 4 + MAX_OBJ_NAME_SIZE; 324 msg_size += 2 + num_ops*sizeof(struct ceph_osd_op); 325 msg_size += 8; /* snapid */ 326 msg_size += 8; /* snap_seq */ 327 msg_size += 8 * (snapc ? snapc->num_snaps : 0); /* snaps */ 328 msg_size += 4; 329 330 if (use_mempool) { 331 req = mempool_alloc(osdc->req_mempool, gfp_flags); 332 memset(req, 0, sizeof(*req)); 333 } else { 334 req = kzalloc(sizeof(*req), gfp_flags); 335 } 336 if (req == NULL) 337 return NULL; 338 339 req->r_osdc = osdc; 340 req->r_mempool = use_mempool; 341 req->r_num_ops = num_ops; 342 343 kref_init(&req->r_kref); 344 init_completion(&req->r_completion); 345 init_completion(&req->r_safe_completion); 346 RB_CLEAR_NODE(&req->r_node); 347 INIT_LIST_HEAD(&req->r_unsafe_item); 348 INIT_LIST_HEAD(&req->r_linger_item); 349 INIT_LIST_HEAD(&req->r_linger_osd); 350 INIT_LIST_HEAD(&req->r_req_lru_item); 351 INIT_LIST_HEAD(&req->r_osd_item); 352 353 /* create reply message */ 354 if (use_mempool) 355 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0); 356 else 357 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, 358 OSD_OPREPLY_FRONT_LEN, gfp_flags, true); 359 if (!msg) { 360 ceph_osdc_put_request(req); 361 return NULL; 362 } 363 req->r_reply = msg; 364 365 /* create request message; allow space for oid */ 366 if (use_mempool) 367 msg = ceph_msgpool_get(&osdc->msgpool_op, 0); 368 else 369 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true); 370 if (!msg) { 371 ceph_osdc_put_request(req); 372 return NULL; 373 } 374 375 memset(msg->front.iov_base, 0, msg->front.iov_len); 376 377 req->r_request = msg; 378 379 return req; 380 } 381 EXPORT_SYMBOL(ceph_osdc_alloc_request); 382 383 static bool osd_req_opcode_valid(u16 opcode) 384 { 385 switch (opcode) { 386 case CEPH_OSD_OP_READ: 387 case CEPH_OSD_OP_STAT: 388 case CEPH_OSD_OP_MAPEXT: 389 case CEPH_OSD_OP_MASKTRUNC: 390 case CEPH_OSD_OP_SPARSE_READ: 391 case CEPH_OSD_OP_NOTIFY: 392 case CEPH_OSD_OP_NOTIFY_ACK: 393 case CEPH_OSD_OP_ASSERT_VER: 394 case CEPH_OSD_OP_WRITE: 395 case CEPH_OSD_OP_WRITEFULL: 396 case CEPH_OSD_OP_TRUNCATE: 397 case CEPH_OSD_OP_ZERO: 398 case CEPH_OSD_OP_DELETE: 399 case CEPH_OSD_OP_APPEND: 400 case CEPH_OSD_OP_STARTSYNC: 401 case CEPH_OSD_OP_SETTRUNC: 402 case CEPH_OSD_OP_TRIMTRUNC: 403 case CEPH_OSD_OP_TMAPUP: 404 case CEPH_OSD_OP_TMAPPUT: 405 case CEPH_OSD_OP_TMAPGET: 406 case CEPH_OSD_OP_CREATE: 407 case CEPH_OSD_OP_ROLLBACK: 408 case CEPH_OSD_OP_WATCH: 409 case CEPH_OSD_OP_OMAPGETKEYS: 410 case CEPH_OSD_OP_OMAPGETVALS: 411 case CEPH_OSD_OP_OMAPGETHEADER: 412 case CEPH_OSD_OP_OMAPGETVALSBYKEYS: 413 case CEPH_OSD_OP_OMAPSETVALS: 414 case CEPH_OSD_OP_OMAPSETHEADER: 415 case CEPH_OSD_OP_OMAPCLEAR: 416 case CEPH_OSD_OP_OMAPRMKEYS: 417 case CEPH_OSD_OP_OMAP_CMP: 418 case CEPH_OSD_OP_CLONERANGE: 419 case CEPH_OSD_OP_ASSERT_SRC_VERSION: 420 case CEPH_OSD_OP_SRC_CMPXATTR: 421 case CEPH_OSD_OP_GETXATTR: 422 case CEPH_OSD_OP_GETXATTRS: 423 case CEPH_OSD_OP_CMPXATTR: 424 case CEPH_OSD_OP_SETXATTR: 425 case CEPH_OSD_OP_SETXATTRS: 426 case CEPH_OSD_OP_RESETXATTRS: 427 case CEPH_OSD_OP_RMXATTR: 428 case CEPH_OSD_OP_PULL: 429 case CEPH_OSD_OP_PUSH: 430 case CEPH_OSD_OP_BALANCEREADS: 431 case CEPH_OSD_OP_UNBALANCEREADS: 432 case CEPH_OSD_OP_SCRUB: 433 case CEPH_OSD_OP_SCRUB_RESERVE: 434 case CEPH_OSD_OP_SCRUB_UNRESERVE: 435 case CEPH_OSD_OP_SCRUB_STOP: 436 case CEPH_OSD_OP_SCRUB_MAP: 437 case CEPH_OSD_OP_WRLOCK: 438 case CEPH_OSD_OP_WRUNLOCK: 439 case CEPH_OSD_OP_RDLOCK: 440 case CEPH_OSD_OP_RDUNLOCK: 441 case CEPH_OSD_OP_UPLOCK: 442 case CEPH_OSD_OP_DNLOCK: 443 case CEPH_OSD_OP_CALL: 444 case CEPH_OSD_OP_PGLS: 445 case CEPH_OSD_OP_PGLS_FILTER: 446 return true; 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) 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; 469 470 return op; 471 } 472 473 void osd_req_op_extent_init(struct ceph_osd_request *osd_req, 474 unsigned int which, u16 opcode, 475 u64 offset, u64 length, 476 u64 truncate_size, u32 truncate_seq) 477 { 478 struct ceph_osd_req_op *op = osd_req_op_init(osd_req, which, opcode); 479 size_t payload_len = 0; 480 481 BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE); 482 483 op->extent.offset = offset; 484 op->extent.length = length; 485 op->extent.truncate_size = truncate_size; 486 op->extent.truncate_seq = truncate_seq; 487 if (opcode == CEPH_OSD_OP_WRITE) 488 payload_len += length; 489 490 op->payload_len = payload_len; 491 } 492 EXPORT_SYMBOL(osd_req_op_extent_init); 493 494 void osd_req_op_extent_update(struct ceph_osd_request *osd_req, 495 unsigned int which, u64 length) 496 { 497 struct ceph_osd_req_op *op; 498 u64 previous; 499 500 BUG_ON(which >= osd_req->r_num_ops); 501 op = &osd_req->r_ops[which]; 502 previous = op->extent.length; 503 504 if (length == previous) 505 return; /* Nothing to do */ 506 BUG_ON(length > previous); 507 508 op->extent.length = length; 509 op->payload_len -= previous - length; 510 } 511 EXPORT_SYMBOL(osd_req_op_extent_update); 512 513 void osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which, 514 u16 opcode, const char *class, const char *method) 515 { 516 struct ceph_osd_req_op *op = osd_req_op_init(osd_req, which, opcode); 517 struct ceph_pagelist *pagelist; 518 size_t payload_len = 0; 519 size_t size; 520 521 BUG_ON(opcode != CEPH_OSD_OP_CALL); 522 523 pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS); 524 BUG_ON(!pagelist); 525 ceph_pagelist_init(pagelist); 526 527 op->cls.class_name = class; 528 size = strlen(class); 529 BUG_ON(size > (size_t) U8_MAX); 530 op->cls.class_len = size; 531 ceph_pagelist_append(pagelist, class, size); 532 payload_len += size; 533 534 op->cls.method_name = method; 535 size = strlen(method); 536 BUG_ON(size > (size_t) U8_MAX); 537 op->cls.method_len = size; 538 ceph_pagelist_append(pagelist, method, size); 539 payload_len += size; 540 541 osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist); 542 543 op->cls.argc = 0; /* currently unused */ 544 545 op->payload_len = payload_len; 546 } 547 EXPORT_SYMBOL(osd_req_op_cls_init); 548 549 void osd_req_op_watch_init(struct ceph_osd_request *osd_req, 550 unsigned int which, u16 opcode, 551 u64 cookie, u64 version, int flag) 552 { 553 struct ceph_osd_req_op *op = osd_req_op_init(osd_req, which, opcode); 554 555 BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH); 556 557 op->watch.cookie = cookie; 558 /* op->watch.ver = version; */ /* XXX 3847 */ 559 op->watch.ver = cpu_to_le64(version); 560 if (opcode == CEPH_OSD_OP_WATCH && flag) 561 op->watch.flag = (u8)1; 562 } 563 EXPORT_SYMBOL(osd_req_op_watch_init); 564 565 static void ceph_osdc_msg_data_add(struct ceph_msg *msg, 566 struct ceph_osd_data *osd_data) 567 { 568 u64 length = ceph_osd_data_length(osd_data); 569 570 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 571 BUG_ON(length > (u64) SIZE_MAX); 572 if (length) 573 ceph_msg_data_add_pages(msg, osd_data->pages, 574 length, osd_data->alignment); 575 } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) { 576 BUG_ON(!length); 577 ceph_msg_data_add_pagelist(msg, osd_data->pagelist); 578 #ifdef CONFIG_BLOCK 579 } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) { 580 ceph_msg_data_add_bio(msg, osd_data->bio, length); 581 #endif 582 } else { 583 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE); 584 } 585 } 586 587 static u64 osd_req_encode_op(struct ceph_osd_request *req, 588 struct ceph_osd_op *dst, unsigned int which) 589 { 590 struct ceph_osd_req_op *src; 591 struct ceph_osd_data *osd_data; 592 u64 request_data_len = 0; 593 u64 data_length; 594 595 BUG_ON(which >= req->r_num_ops); 596 src = &req->r_ops[which]; 597 if (WARN_ON(!osd_req_opcode_valid(src->op))) { 598 pr_err("unrecognized osd opcode %d\n", src->op); 599 600 return 0; 601 } 602 603 switch (src->op) { 604 case CEPH_OSD_OP_STAT: 605 break; 606 case CEPH_OSD_OP_READ: 607 case CEPH_OSD_OP_WRITE: 608 if (src->op == CEPH_OSD_OP_WRITE) 609 request_data_len = src->extent.length; 610 dst->extent.offset = cpu_to_le64(src->extent.offset); 611 dst->extent.length = cpu_to_le64(src->extent.length); 612 dst->extent.truncate_size = 613 cpu_to_le64(src->extent.truncate_size); 614 dst->extent.truncate_seq = 615 cpu_to_le32(src->extent.truncate_seq); 616 osd_data = &src->extent.osd_data; 617 if (src->op == CEPH_OSD_OP_WRITE) 618 ceph_osdc_msg_data_add(req->r_request, osd_data); 619 else 620 ceph_osdc_msg_data_add(req->r_reply, osd_data); 621 break; 622 case CEPH_OSD_OP_CALL: 623 dst->cls.class_len = src->cls.class_len; 624 dst->cls.method_len = src->cls.method_len; 625 osd_data = &src->cls.request_info; 626 ceph_osdc_msg_data_add(req->r_request, osd_data); 627 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGELIST); 628 request_data_len = osd_data->pagelist->length; 629 630 osd_data = &src->cls.request_data; 631 data_length = ceph_osd_data_length(osd_data); 632 if (data_length) { 633 BUG_ON(osd_data->type == CEPH_OSD_DATA_TYPE_NONE); 634 dst->cls.indata_len = cpu_to_le32(data_length); 635 ceph_osdc_msg_data_add(req->r_request, osd_data); 636 src->payload_len += data_length; 637 request_data_len += data_length; 638 } 639 osd_data = &src->cls.response_data; 640 ceph_osdc_msg_data_add(req->r_reply, osd_data); 641 break; 642 case CEPH_OSD_OP_STARTSYNC: 643 break; 644 case CEPH_OSD_OP_NOTIFY_ACK: 645 case CEPH_OSD_OP_WATCH: 646 dst->watch.cookie = cpu_to_le64(src->watch.cookie); 647 dst->watch.ver = cpu_to_le64(src->watch.ver); 648 dst->watch.flag = src->watch.flag; 649 break; 650 default: 651 pr_err("unsupported osd opcode %s\n", 652 ceph_osd_op_name(src->op)); 653 WARN_ON(1); 654 655 return 0; 656 } 657 dst->op = cpu_to_le16(src->op); 658 dst->payload_len = cpu_to_le32(src->payload_len); 659 660 return request_data_len; 661 } 662 663 /* 664 * build new request AND message, calculate layout, and adjust file 665 * extent as needed. 666 * 667 * if the file was recently truncated, we include information about its 668 * old and new size so that the object can be updated appropriately. (we 669 * avoid synchronously deleting truncated objects because it's slow.) 670 * 671 * if @do_sync, include a 'startsync' command so that the osd will flush 672 * data quickly. 673 */ 674 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc, 675 struct ceph_file_layout *layout, 676 struct ceph_vino vino, 677 u64 off, u64 *plen, int num_ops, 678 int opcode, int flags, 679 struct ceph_snap_context *snapc, 680 u32 truncate_seq, 681 u64 truncate_size, 682 bool use_mempool) 683 { 684 struct ceph_osd_request *req; 685 u64 objnum = 0; 686 u64 objoff = 0; 687 u64 objlen = 0; 688 u32 object_size; 689 u64 object_base; 690 int r; 691 692 BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE); 693 694 req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool, 695 GFP_NOFS); 696 if (!req) 697 return ERR_PTR(-ENOMEM); 698 699 req->r_flags = flags; 700 701 /* calculate max write size */ 702 r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen); 703 if (r < 0) { 704 ceph_osdc_put_request(req); 705 return ERR_PTR(r); 706 } 707 708 object_size = le32_to_cpu(layout->fl_object_size); 709 object_base = off - objoff; 710 if (truncate_size <= object_base) { 711 truncate_size = 0; 712 } else { 713 truncate_size -= object_base; 714 if (truncate_size > object_size) 715 truncate_size = object_size; 716 } 717 718 osd_req_op_extent_init(req, 0, opcode, objoff, objlen, 719 truncate_size, truncate_seq); 720 721 /* 722 * A second op in the ops array means the caller wants to 723 * also issue a include a 'startsync' command so that the 724 * osd will flush data quickly. 725 */ 726 if (num_ops > 1) 727 osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC); 728 729 req->r_file_layout = *layout; /* keep a copy */ 730 731 snprintf(req->r_oid, sizeof(req->r_oid), "%llx.%08llx", 732 vino.ino, objnum); 733 req->r_oid_len = strlen(req->r_oid); 734 735 return req; 736 } 737 EXPORT_SYMBOL(ceph_osdc_new_request); 738 739 /* 740 * We keep osd requests in an rbtree, sorted by ->r_tid. 741 */ 742 static void __insert_request(struct ceph_osd_client *osdc, 743 struct ceph_osd_request *new) 744 { 745 struct rb_node **p = &osdc->requests.rb_node; 746 struct rb_node *parent = NULL; 747 struct ceph_osd_request *req = NULL; 748 749 while (*p) { 750 parent = *p; 751 req = rb_entry(parent, struct ceph_osd_request, r_node); 752 if (new->r_tid < req->r_tid) 753 p = &(*p)->rb_left; 754 else if (new->r_tid > req->r_tid) 755 p = &(*p)->rb_right; 756 else 757 BUG(); 758 } 759 760 rb_link_node(&new->r_node, parent, p); 761 rb_insert_color(&new->r_node, &osdc->requests); 762 } 763 764 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc, 765 u64 tid) 766 { 767 struct ceph_osd_request *req; 768 struct rb_node *n = osdc->requests.rb_node; 769 770 while (n) { 771 req = rb_entry(n, struct ceph_osd_request, r_node); 772 if (tid < req->r_tid) 773 n = n->rb_left; 774 else if (tid > req->r_tid) 775 n = n->rb_right; 776 else 777 return req; 778 } 779 return NULL; 780 } 781 782 static struct ceph_osd_request * 783 __lookup_request_ge(struct ceph_osd_client *osdc, 784 u64 tid) 785 { 786 struct ceph_osd_request *req; 787 struct rb_node *n = osdc->requests.rb_node; 788 789 while (n) { 790 req = rb_entry(n, struct ceph_osd_request, r_node); 791 if (tid < req->r_tid) { 792 if (!n->rb_left) 793 return req; 794 n = n->rb_left; 795 } else if (tid > req->r_tid) { 796 n = n->rb_right; 797 } else { 798 return req; 799 } 800 } 801 return NULL; 802 } 803 804 /* 805 * Resubmit requests pending on the given osd. 806 */ 807 static void __kick_osd_requests(struct ceph_osd_client *osdc, 808 struct ceph_osd *osd) 809 { 810 struct ceph_osd_request *req, *nreq; 811 LIST_HEAD(resend); 812 int err; 813 814 dout("__kick_osd_requests osd%d\n", osd->o_osd); 815 err = __reset_osd(osdc, osd); 816 if (err) 817 return; 818 /* 819 * Build up a list of requests to resend by traversing the 820 * osd's list of requests. Requests for a given object are 821 * sent in tid order, and that is also the order they're 822 * kept on this list. Therefore all requests that are in 823 * flight will be found first, followed by all requests that 824 * have not yet been sent. And to resend requests while 825 * preserving this order we will want to put any sent 826 * requests back on the front of the osd client's unsent 827 * list. 828 * 829 * So we build a separate ordered list of already-sent 830 * requests for the affected osd and splice it onto the 831 * front of the osd client's unsent list. Once we've seen a 832 * request that has not yet been sent we're done. Those 833 * requests are already sitting right where they belong. 834 */ 835 list_for_each_entry(req, &osd->o_requests, r_osd_item) { 836 if (!req->r_sent) 837 break; 838 list_move_tail(&req->r_req_lru_item, &resend); 839 dout("requeueing %p tid %llu osd%d\n", req, req->r_tid, 840 osd->o_osd); 841 if (!req->r_linger) 842 req->r_flags |= CEPH_OSD_FLAG_RETRY; 843 } 844 list_splice(&resend, &osdc->req_unsent); 845 846 /* 847 * Linger requests are re-registered before sending, which 848 * sets up a new tid for each. We add them to the unsent 849 * list at the end to keep things in tid order. 850 */ 851 list_for_each_entry_safe(req, nreq, &osd->o_linger_requests, 852 r_linger_osd) { 853 /* 854 * reregister request prior to unregistering linger so 855 * that r_osd is preserved. 856 */ 857 BUG_ON(!list_empty(&req->r_req_lru_item)); 858 __register_request(osdc, req); 859 list_add_tail(&req->r_req_lru_item, &osdc->req_unsent); 860 list_add_tail(&req->r_osd_item, &req->r_osd->o_requests); 861 __unregister_linger_request(osdc, req); 862 dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid, 863 osd->o_osd); 864 } 865 } 866 867 /* 868 * If the osd connection drops, we need to resubmit all requests. 869 */ 870 static void osd_reset(struct ceph_connection *con) 871 { 872 struct ceph_osd *osd = con->private; 873 struct ceph_osd_client *osdc; 874 875 if (!osd) 876 return; 877 dout("osd_reset osd%d\n", osd->o_osd); 878 osdc = osd->o_osdc; 879 down_read(&osdc->map_sem); 880 mutex_lock(&osdc->request_mutex); 881 __kick_osd_requests(osdc, osd); 882 __send_queued(osdc); 883 mutex_unlock(&osdc->request_mutex); 884 up_read(&osdc->map_sem); 885 } 886 887 /* 888 * Track open sessions with osds. 889 */ 890 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum) 891 { 892 struct ceph_osd *osd; 893 894 osd = kzalloc(sizeof(*osd), GFP_NOFS); 895 if (!osd) 896 return NULL; 897 898 atomic_set(&osd->o_ref, 1); 899 osd->o_osdc = osdc; 900 osd->o_osd = onum; 901 RB_CLEAR_NODE(&osd->o_node); 902 INIT_LIST_HEAD(&osd->o_requests); 903 INIT_LIST_HEAD(&osd->o_linger_requests); 904 INIT_LIST_HEAD(&osd->o_osd_lru); 905 osd->o_incarnation = 1; 906 907 ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr); 908 909 INIT_LIST_HEAD(&osd->o_keepalive_item); 910 return osd; 911 } 912 913 static struct ceph_osd *get_osd(struct ceph_osd *osd) 914 { 915 if (atomic_inc_not_zero(&osd->o_ref)) { 916 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1, 917 atomic_read(&osd->o_ref)); 918 return osd; 919 } else { 920 dout("get_osd %p FAIL\n", osd); 921 return NULL; 922 } 923 } 924 925 static void put_osd(struct ceph_osd *osd) 926 { 927 dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref), 928 atomic_read(&osd->o_ref) - 1); 929 if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) { 930 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth; 931 932 ceph_auth_destroy_authorizer(ac, osd->o_auth.authorizer); 933 kfree(osd); 934 } 935 } 936 937 /* 938 * remove an osd from our map 939 */ 940 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 941 { 942 dout("__remove_osd %p\n", osd); 943 BUG_ON(!list_empty(&osd->o_requests)); 944 rb_erase(&osd->o_node, &osdc->osds); 945 list_del_init(&osd->o_osd_lru); 946 ceph_con_close(&osd->o_con); 947 put_osd(osd); 948 } 949 950 static void remove_all_osds(struct ceph_osd_client *osdc) 951 { 952 dout("%s %p\n", __func__, osdc); 953 mutex_lock(&osdc->request_mutex); 954 while (!RB_EMPTY_ROOT(&osdc->osds)) { 955 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds), 956 struct ceph_osd, o_node); 957 __remove_osd(osdc, osd); 958 } 959 mutex_unlock(&osdc->request_mutex); 960 } 961 962 static void __move_osd_to_lru(struct ceph_osd_client *osdc, 963 struct ceph_osd *osd) 964 { 965 dout("__move_osd_to_lru %p\n", osd); 966 BUG_ON(!list_empty(&osd->o_osd_lru)); 967 list_add_tail(&osd->o_osd_lru, &osdc->osd_lru); 968 osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ; 969 } 970 971 static void __remove_osd_from_lru(struct ceph_osd *osd) 972 { 973 dout("__remove_osd_from_lru %p\n", osd); 974 if (!list_empty(&osd->o_osd_lru)) 975 list_del_init(&osd->o_osd_lru); 976 } 977 978 static void remove_old_osds(struct ceph_osd_client *osdc) 979 { 980 struct ceph_osd *osd, *nosd; 981 982 dout("__remove_old_osds %p\n", osdc); 983 mutex_lock(&osdc->request_mutex); 984 list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) { 985 if (time_before(jiffies, osd->lru_ttl)) 986 break; 987 __remove_osd(osdc, osd); 988 } 989 mutex_unlock(&osdc->request_mutex); 990 } 991 992 /* 993 * reset osd connect 994 */ 995 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 996 { 997 struct ceph_entity_addr *peer_addr; 998 999 dout("__reset_osd %p osd%d\n", osd, osd->o_osd); 1000 if (list_empty(&osd->o_requests) && 1001 list_empty(&osd->o_linger_requests)) { 1002 __remove_osd(osdc, osd); 1003 1004 return -ENODEV; 1005 } 1006 1007 peer_addr = &osdc->osdmap->osd_addr[osd->o_osd]; 1008 if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) && 1009 !ceph_con_opened(&osd->o_con)) { 1010 struct ceph_osd_request *req; 1011 1012 dout(" osd addr hasn't changed and connection never opened," 1013 " letting msgr retry"); 1014 /* touch each r_stamp for handle_timeout()'s benfit */ 1015 list_for_each_entry(req, &osd->o_requests, r_osd_item) 1016 req->r_stamp = jiffies; 1017 1018 return -EAGAIN; 1019 } 1020 1021 ceph_con_close(&osd->o_con); 1022 ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr); 1023 osd->o_incarnation++; 1024 1025 return 0; 1026 } 1027 1028 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new) 1029 { 1030 struct rb_node **p = &osdc->osds.rb_node; 1031 struct rb_node *parent = NULL; 1032 struct ceph_osd *osd = NULL; 1033 1034 dout("__insert_osd %p osd%d\n", new, new->o_osd); 1035 while (*p) { 1036 parent = *p; 1037 osd = rb_entry(parent, struct ceph_osd, o_node); 1038 if (new->o_osd < osd->o_osd) 1039 p = &(*p)->rb_left; 1040 else if (new->o_osd > osd->o_osd) 1041 p = &(*p)->rb_right; 1042 else 1043 BUG(); 1044 } 1045 1046 rb_link_node(&new->o_node, parent, p); 1047 rb_insert_color(&new->o_node, &osdc->osds); 1048 } 1049 1050 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o) 1051 { 1052 struct ceph_osd *osd; 1053 struct rb_node *n = osdc->osds.rb_node; 1054 1055 while (n) { 1056 osd = rb_entry(n, struct ceph_osd, o_node); 1057 if (o < osd->o_osd) 1058 n = n->rb_left; 1059 else if (o > osd->o_osd) 1060 n = n->rb_right; 1061 else 1062 return osd; 1063 } 1064 return NULL; 1065 } 1066 1067 static void __schedule_osd_timeout(struct ceph_osd_client *osdc) 1068 { 1069 schedule_delayed_work(&osdc->timeout_work, 1070 osdc->client->options->osd_keepalive_timeout * HZ); 1071 } 1072 1073 static void __cancel_osd_timeout(struct ceph_osd_client *osdc) 1074 { 1075 cancel_delayed_work(&osdc->timeout_work); 1076 } 1077 1078 /* 1079 * Register request, assign tid. If this is the first request, set up 1080 * the timeout event. 1081 */ 1082 static void __register_request(struct ceph_osd_client *osdc, 1083 struct ceph_osd_request *req) 1084 { 1085 req->r_tid = ++osdc->last_tid; 1086 req->r_request->hdr.tid = cpu_to_le64(req->r_tid); 1087 dout("__register_request %p tid %lld\n", req, req->r_tid); 1088 __insert_request(osdc, req); 1089 ceph_osdc_get_request(req); 1090 osdc->num_requests++; 1091 if (osdc->num_requests == 1) { 1092 dout(" first request, scheduling timeout\n"); 1093 __schedule_osd_timeout(osdc); 1094 } 1095 } 1096 1097 /* 1098 * called under osdc->request_mutex 1099 */ 1100 static void __unregister_request(struct ceph_osd_client *osdc, 1101 struct ceph_osd_request *req) 1102 { 1103 if (RB_EMPTY_NODE(&req->r_node)) { 1104 dout("__unregister_request %p tid %lld not registered\n", 1105 req, req->r_tid); 1106 return; 1107 } 1108 1109 dout("__unregister_request %p tid %lld\n", req, req->r_tid); 1110 rb_erase(&req->r_node, &osdc->requests); 1111 osdc->num_requests--; 1112 1113 if (req->r_osd) { 1114 /* make sure the original request isn't in flight. */ 1115 ceph_msg_revoke(req->r_request); 1116 1117 list_del_init(&req->r_osd_item); 1118 if (list_empty(&req->r_osd->o_requests) && 1119 list_empty(&req->r_osd->o_linger_requests)) { 1120 dout("moving osd to %p lru\n", req->r_osd); 1121 __move_osd_to_lru(osdc, req->r_osd); 1122 } 1123 if (list_empty(&req->r_linger_item)) 1124 req->r_osd = NULL; 1125 } 1126 1127 list_del_init(&req->r_req_lru_item); 1128 ceph_osdc_put_request(req); 1129 1130 if (osdc->num_requests == 0) { 1131 dout(" no requests, canceling timeout\n"); 1132 __cancel_osd_timeout(osdc); 1133 } 1134 } 1135 1136 /* 1137 * Cancel a previously queued request message 1138 */ 1139 static void __cancel_request(struct ceph_osd_request *req) 1140 { 1141 if (req->r_sent && req->r_osd) { 1142 ceph_msg_revoke(req->r_request); 1143 req->r_sent = 0; 1144 } 1145 } 1146 1147 static void __register_linger_request(struct ceph_osd_client *osdc, 1148 struct ceph_osd_request *req) 1149 { 1150 dout("__register_linger_request %p\n", req); 1151 list_add_tail(&req->r_linger_item, &osdc->req_linger); 1152 if (req->r_osd) 1153 list_add_tail(&req->r_linger_osd, 1154 &req->r_osd->o_linger_requests); 1155 } 1156 1157 static void __unregister_linger_request(struct ceph_osd_client *osdc, 1158 struct ceph_osd_request *req) 1159 { 1160 dout("__unregister_linger_request %p\n", req); 1161 list_del_init(&req->r_linger_item); 1162 if (req->r_osd) { 1163 list_del_init(&req->r_linger_osd); 1164 1165 if (list_empty(&req->r_osd->o_requests) && 1166 list_empty(&req->r_osd->o_linger_requests)) { 1167 dout("moving osd to %p lru\n", req->r_osd); 1168 __move_osd_to_lru(osdc, req->r_osd); 1169 } 1170 if (list_empty(&req->r_osd_item)) 1171 req->r_osd = NULL; 1172 } 1173 } 1174 1175 void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc, 1176 struct ceph_osd_request *req) 1177 { 1178 mutex_lock(&osdc->request_mutex); 1179 if (req->r_linger) { 1180 __unregister_linger_request(osdc, req); 1181 ceph_osdc_put_request(req); 1182 } 1183 mutex_unlock(&osdc->request_mutex); 1184 } 1185 EXPORT_SYMBOL(ceph_osdc_unregister_linger_request); 1186 1187 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc, 1188 struct ceph_osd_request *req) 1189 { 1190 if (!req->r_linger) { 1191 dout("set_request_linger %p\n", req); 1192 req->r_linger = 1; 1193 /* 1194 * caller is now responsible for calling 1195 * unregister_linger_request 1196 */ 1197 ceph_osdc_get_request(req); 1198 } 1199 } 1200 EXPORT_SYMBOL(ceph_osdc_set_request_linger); 1201 1202 /* 1203 * Pick an osd (the first 'up' osd in the pg), allocate the osd struct 1204 * (as needed), and set the request r_osd appropriately. If there is 1205 * no up osd, set r_osd to NULL. Move the request to the appropriate list 1206 * (unsent, homeless) or leave on in-flight lru. 1207 * 1208 * Return 0 if unchanged, 1 if changed, or negative on error. 1209 * 1210 * Caller should hold map_sem for read and request_mutex. 1211 */ 1212 static int __map_request(struct ceph_osd_client *osdc, 1213 struct ceph_osd_request *req, int force_resend) 1214 { 1215 struct ceph_pg pgid; 1216 int acting[CEPH_PG_MAX_SIZE]; 1217 int o = -1, num = 0; 1218 int err; 1219 1220 dout("map_request %p tid %lld\n", req, req->r_tid); 1221 err = ceph_calc_ceph_pg(&pgid, req->r_oid, osdc->osdmap, 1222 ceph_file_layout_pg_pool(req->r_file_layout)); 1223 if (err) { 1224 list_move(&req->r_req_lru_item, &osdc->req_notarget); 1225 return err; 1226 } 1227 req->r_pgid = pgid; 1228 1229 err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting); 1230 if (err > 0) { 1231 o = acting[0]; 1232 num = err; 1233 } 1234 1235 if ((!force_resend && 1236 req->r_osd && req->r_osd->o_osd == o && 1237 req->r_sent >= req->r_osd->o_incarnation && 1238 req->r_num_pg_osds == num && 1239 memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) || 1240 (req->r_osd == NULL && o == -1)) 1241 return 0; /* no change */ 1242 1243 dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n", 1244 req->r_tid, pgid.pool, pgid.seed, o, 1245 req->r_osd ? req->r_osd->o_osd : -1); 1246 1247 /* record full pg acting set */ 1248 memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num); 1249 req->r_num_pg_osds = num; 1250 1251 if (req->r_osd) { 1252 __cancel_request(req); 1253 list_del_init(&req->r_osd_item); 1254 req->r_osd = NULL; 1255 } 1256 1257 req->r_osd = __lookup_osd(osdc, o); 1258 if (!req->r_osd && o >= 0) { 1259 err = -ENOMEM; 1260 req->r_osd = create_osd(osdc, o); 1261 if (!req->r_osd) { 1262 list_move(&req->r_req_lru_item, &osdc->req_notarget); 1263 goto out; 1264 } 1265 1266 dout("map_request osd %p is osd%d\n", req->r_osd, o); 1267 __insert_osd(osdc, req->r_osd); 1268 1269 ceph_con_open(&req->r_osd->o_con, 1270 CEPH_ENTITY_TYPE_OSD, o, 1271 &osdc->osdmap->osd_addr[o]); 1272 } 1273 1274 if (req->r_osd) { 1275 __remove_osd_from_lru(req->r_osd); 1276 list_add_tail(&req->r_osd_item, &req->r_osd->o_requests); 1277 list_move_tail(&req->r_req_lru_item, &osdc->req_unsent); 1278 } else { 1279 list_move_tail(&req->r_req_lru_item, &osdc->req_notarget); 1280 } 1281 err = 1; /* osd or pg changed */ 1282 1283 out: 1284 return err; 1285 } 1286 1287 /* 1288 * caller should hold map_sem (for read) and request_mutex 1289 */ 1290 static void __send_request(struct ceph_osd_client *osdc, 1291 struct ceph_osd_request *req) 1292 { 1293 void *p; 1294 1295 dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n", 1296 req, req->r_tid, req->r_osd->o_osd, req->r_flags, 1297 (unsigned long long)req->r_pgid.pool, req->r_pgid.seed); 1298 1299 /* fill in message content that changes each time we send it */ 1300 put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch); 1301 put_unaligned_le32(req->r_flags, req->r_request_flags); 1302 put_unaligned_le64(req->r_pgid.pool, req->r_request_pool); 1303 p = req->r_request_pgid; 1304 ceph_encode_64(&p, req->r_pgid.pool); 1305 ceph_encode_32(&p, req->r_pgid.seed); 1306 put_unaligned_le64(1, req->r_request_attempts); /* FIXME */ 1307 memcpy(req->r_request_reassert_version, &req->r_reassert_version, 1308 sizeof(req->r_reassert_version)); 1309 1310 req->r_stamp = jiffies; 1311 list_move_tail(&req->r_req_lru_item, &osdc->req_lru); 1312 1313 ceph_msg_get(req->r_request); /* send consumes a ref */ 1314 1315 /* Mark the request unsafe if this is the first timet's being sent. */ 1316 1317 if (!req->r_sent && req->r_unsafe_callback) 1318 req->r_unsafe_callback(req, true); 1319 req->r_sent = req->r_osd->o_incarnation; 1320 1321 ceph_con_send(&req->r_osd->o_con, req->r_request); 1322 } 1323 1324 /* 1325 * Send any requests in the queue (req_unsent). 1326 */ 1327 static void __send_queued(struct ceph_osd_client *osdc) 1328 { 1329 struct ceph_osd_request *req, *tmp; 1330 1331 dout("__send_queued\n"); 1332 list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) 1333 __send_request(osdc, req); 1334 } 1335 1336 /* 1337 * Timeout callback, called every N seconds when 1 or more osd 1338 * requests has been active for more than N seconds. When this 1339 * happens, we ping all OSDs with requests who have timed out to 1340 * ensure any communications channel reset is detected. Reset the 1341 * request timeouts another N seconds in the future as we go. 1342 * Reschedule the timeout event another N seconds in future (unless 1343 * there are no open requests). 1344 */ 1345 static void handle_timeout(struct work_struct *work) 1346 { 1347 struct ceph_osd_client *osdc = 1348 container_of(work, struct ceph_osd_client, timeout_work.work); 1349 struct ceph_osd_request *req; 1350 struct ceph_osd *osd; 1351 unsigned long keepalive = 1352 osdc->client->options->osd_keepalive_timeout * HZ; 1353 struct list_head slow_osds; 1354 dout("timeout\n"); 1355 down_read(&osdc->map_sem); 1356 1357 ceph_monc_request_next_osdmap(&osdc->client->monc); 1358 1359 mutex_lock(&osdc->request_mutex); 1360 1361 /* 1362 * ping osds that are a bit slow. this ensures that if there 1363 * is a break in the TCP connection we will notice, and reopen 1364 * a connection with that osd (from the fault callback). 1365 */ 1366 INIT_LIST_HEAD(&slow_osds); 1367 list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) { 1368 if (time_before(jiffies, req->r_stamp + keepalive)) 1369 break; 1370 1371 osd = req->r_osd; 1372 BUG_ON(!osd); 1373 dout(" tid %llu is slow, will send keepalive on osd%d\n", 1374 req->r_tid, osd->o_osd); 1375 list_move_tail(&osd->o_keepalive_item, &slow_osds); 1376 } 1377 while (!list_empty(&slow_osds)) { 1378 osd = list_entry(slow_osds.next, struct ceph_osd, 1379 o_keepalive_item); 1380 list_del_init(&osd->o_keepalive_item); 1381 ceph_con_keepalive(&osd->o_con); 1382 } 1383 1384 __schedule_osd_timeout(osdc); 1385 __send_queued(osdc); 1386 mutex_unlock(&osdc->request_mutex); 1387 up_read(&osdc->map_sem); 1388 } 1389 1390 static void handle_osds_timeout(struct work_struct *work) 1391 { 1392 struct ceph_osd_client *osdc = 1393 container_of(work, struct ceph_osd_client, 1394 osds_timeout_work.work); 1395 unsigned long delay = 1396 osdc->client->options->osd_idle_ttl * HZ >> 2; 1397 1398 dout("osds timeout\n"); 1399 down_read(&osdc->map_sem); 1400 remove_old_osds(osdc); 1401 up_read(&osdc->map_sem); 1402 1403 schedule_delayed_work(&osdc->osds_timeout_work, 1404 round_jiffies_relative(delay)); 1405 } 1406 1407 static void complete_request(struct ceph_osd_request *req) 1408 { 1409 if (req->r_unsafe_callback) 1410 req->r_unsafe_callback(req, false); 1411 complete_all(&req->r_safe_completion); /* fsync waiter */ 1412 } 1413 1414 /* 1415 * handle osd op reply. either call the callback if it is specified, 1416 * or do the completion to wake up the waiting thread. 1417 */ 1418 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg, 1419 struct ceph_connection *con) 1420 { 1421 void *p, *end; 1422 struct ceph_osd_request *req; 1423 u64 tid; 1424 int object_len; 1425 unsigned int numops; 1426 int payload_len, flags; 1427 s32 result; 1428 s32 retry_attempt; 1429 struct ceph_pg pg; 1430 int err; 1431 u32 reassert_epoch; 1432 u64 reassert_version; 1433 u32 osdmap_epoch; 1434 int already_completed; 1435 u32 bytes; 1436 unsigned int i; 1437 1438 tid = le64_to_cpu(msg->hdr.tid); 1439 dout("handle_reply %p tid %llu\n", msg, tid); 1440 1441 p = msg->front.iov_base; 1442 end = p + msg->front.iov_len; 1443 1444 ceph_decode_need(&p, end, 4, bad); 1445 object_len = ceph_decode_32(&p); 1446 ceph_decode_need(&p, end, object_len, bad); 1447 p += object_len; 1448 1449 err = ceph_decode_pgid(&p, end, &pg); 1450 if (err) 1451 goto bad; 1452 1453 ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad); 1454 flags = ceph_decode_64(&p); 1455 result = ceph_decode_32(&p); 1456 reassert_epoch = ceph_decode_32(&p); 1457 reassert_version = ceph_decode_64(&p); 1458 osdmap_epoch = ceph_decode_32(&p); 1459 1460 /* lookup */ 1461 mutex_lock(&osdc->request_mutex); 1462 req = __lookup_request(osdc, tid); 1463 if (req == NULL) { 1464 dout("handle_reply tid %llu dne\n", tid); 1465 goto bad_mutex; 1466 } 1467 ceph_osdc_get_request(req); 1468 1469 dout("handle_reply %p tid %llu req %p result %d\n", msg, tid, 1470 req, result); 1471 1472 ceph_decode_need(&p, end, 4, bad); 1473 numops = ceph_decode_32(&p); 1474 if (numops > CEPH_OSD_MAX_OP) 1475 goto bad_put; 1476 if (numops != req->r_num_ops) 1477 goto bad_put; 1478 payload_len = 0; 1479 ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad); 1480 for (i = 0; i < numops; i++) { 1481 struct ceph_osd_op *op = p; 1482 int len; 1483 1484 len = le32_to_cpu(op->payload_len); 1485 req->r_reply_op_len[i] = len; 1486 dout(" op %d has %d bytes\n", i, len); 1487 payload_len += len; 1488 p += sizeof(*op); 1489 } 1490 bytes = le32_to_cpu(msg->hdr.data_len); 1491 if (payload_len != bytes) { 1492 pr_warning("sum of op payload lens %d != data_len %d", 1493 payload_len, bytes); 1494 goto bad_put; 1495 } 1496 1497 ceph_decode_need(&p, end, 4 + numops * 4, bad); 1498 retry_attempt = ceph_decode_32(&p); 1499 for (i = 0; i < numops; i++) 1500 req->r_reply_op_result[i] = ceph_decode_32(&p); 1501 1502 if (!req->r_got_reply) { 1503 1504 req->r_result = result; 1505 dout("handle_reply result %d bytes %d\n", req->r_result, 1506 bytes); 1507 if (req->r_result == 0) 1508 req->r_result = bytes; 1509 1510 /* in case this is a write and we need to replay, */ 1511 req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch); 1512 req->r_reassert_version.version = cpu_to_le64(reassert_version); 1513 1514 req->r_got_reply = 1; 1515 } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) { 1516 dout("handle_reply tid %llu dup ack\n", tid); 1517 mutex_unlock(&osdc->request_mutex); 1518 goto done; 1519 } 1520 1521 dout("handle_reply tid %llu flags %d\n", tid, flags); 1522 1523 if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK)) 1524 __register_linger_request(osdc, req); 1525 1526 /* either this is a read, or we got the safe response */ 1527 if (result < 0 || 1528 (flags & CEPH_OSD_FLAG_ONDISK) || 1529 ((flags & CEPH_OSD_FLAG_WRITE) == 0)) 1530 __unregister_request(osdc, req); 1531 1532 already_completed = req->r_completed; 1533 req->r_completed = 1; 1534 mutex_unlock(&osdc->request_mutex); 1535 if (already_completed) 1536 goto done; 1537 1538 if (req->r_callback) 1539 req->r_callback(req, msg); 1540 else 1541 complete_all(&req->r_completion); 1542 1543 if (flags & CEPH_OSD_FLAG_ONDISK) 1544 complete_request(req); 1545 1546 done: 1547 dout("req=%p req->r_linger=%d\n", req, req->r_linger); 1548 ceph_osdc_put_request(req); 1549 return; 1550 1551 bad_put: 1552 ceph_osdc_put_request(req); 1553 bad_mutex: 1554 mutex_unlock(&osdc->request_mutex); 1555 bad: 1556 pr_err("corrupt osd_op_reply got %d %d\n", 1557 (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len)); 1558 ceph_msg_dump(msg); 1559 } 1560 1561 static void reset_changed_osds(struct ceph_osd_client *osdc) 1562 { 1563 struct rb_node *p, *n; 1564 1565 for (p = rb_first(&osdc->osds); p; p = n) { 1566 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node); 1567 1568 n = rb_next(p); 1569 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) || 1570 memcmp(&osd->o_con.peer_addr, 1571 ceph_osd_addr(osdc->osdmap, 1572 osd->o_osd), 1573 sizeof(struct ceph_entity_addr)) != 0) 1574 __reset_osd(osdc, osd); 1575 } 1576 } 1577 1578 /* 1579 * Requeue requests whose mapping to an OSD has changed. If requests map to 1580 * no osd, request a new map. 1581 * 1582 * Caller should hold map_sem for read. 1583 */ 1584 static void kick_requests(struct ceph_osd_client *osdc, int force_resend) 1585 { 1586 struct ceph_osd_request *req, *nreq; 1587 struct rb_node *p; 1588 int needmap = 0; 1589 int err; 1590 1591 dout("kick_requests %s\n", force_resend ? " (force resend)" : ""); 1592 mutex_lock(&osdc->request_mutex); 1593 for (p = rb_first(&osdc->requests); p; ) { 1594 req = rb_entry(p, struct ceph_osd_request, r_node); 1595 p = rb_next(p); 1596 1597 /* 1598 * For linger requests that have not yet been 1599 * registered, move them to the linger list; they'll 1600 * be sent to the osd in the loop below. Unregister 1601 * the request before re-registering it as a linger 1602 * request to ensure the __map_request() below 1603 * will decide it needs to be sent. 1604 */ 1605 if (req->r_linger && list_empty(&req->r_linger_item)) { 1606 dout("%p tid %llu restart on osd%d\n", 1607 req, req->r_tid, 1608 req->r_osd ? req->r_osd->o_osd : -1); 1609 __unregister_request(osdc, req); 1610 __register_linger_request(osdc, req); 1611 continue; 1612 } 1613 1614 err = __map_request(osdc, req, force_resend); 1615 if (err < 0) 1616 continue; /* error */ 1617 if (req->r_osd == NULL) { 1618 dout("%p tid %llu maps to no osd\n", req, req->r_tid); 1619 needmap++; /* request a newer map */ 1620 } else if (err > 0) { 1621 if (!req->r_linger) { 1622 dout("%p tid %llu requeued on osd%d\n", req, 1623 req->r_tid, 1624 req->r_osd ? req->r_osd->o_osd : -1); 1625 req->r_flags |= CEPH_OSD_FLAG_RETRY; 1626 } 1627 } 1628 } 1629 1630 list_for_each_entry_safe(req, nreq, &osdc->req_linger, 1631 r_linger_item) { 1632 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd); 1633 1634 err = __map_request(osdc, req, force_resend); 1635 dout("__map_request returned %d\n", err); 1636 if (err == 0) 1637 continue; /* no change and no osd was specified */ 1638 if (err < 0) 1639 continue; /* hrm! */ 1640 if (req->r_osd == NULL) { 1641 dout("tid %llu maps to no valid osd\n", req->r_tid); 1642 needmap++; /* request a newer map */ 1643 continue; 1644 } 1645 1646 dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid, 1647 req->r_osd ? req->r_osd->o_osd : -1); 1648 __register_request(osdc, req); 1649 __unregister_linger_request(osdc, req); 1650 } 1651 mutex_unlock(&osdc->request_mutex); 1652 1653 if (needmap) { 1654 dout("%d requests for down osds, need new map\n", needmap); 1655 ceph_monc_request_next_osdmap(&osdc->client->monc); 1656 } 1657 reset_changed_osds(osdc); 1658 } 1659 1660 1661 /* 1662 * Process updated osd map. 1663 * 1664 * The message contains any number of incremental and full maps, normally 1665 * indicating some sort of topology change in the cluster. Kick requests 1666 * off to different OSDs as needed. 1667 */ 1668 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg) 1669 { 1670 void *p, *end, *next; 1671 u32 nr_maps, maplen; 1672 u32 epoch; 1673 struct ceph_osdmap *newmap = NULL, *oldmap; 1674 int err; 1675 struct ceph_fsid fsid; 1676 1677 dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0); 1678 p = msg->front.iov_base; 1679 end = p + msg->front.iov_len; 1680 1681 /* verify fsid */ 1682 ceph_decode_need(&p, end, sizeof(fsid), bad); 1683 ceph_decode_copy(&p, &fsid, sizeof(fsid)); 1684 if (ceph_check_fsid(osdc->client, &fsid) < 0) 1685 return; 1686 1687 down_write(&osdc->map_sem); 1688 1689 /* incremental maps */ 1690 ceph_decode_32_safe(&p, end, nr_maps, bad); 1691 dout(" %d inc maps\n", nr_maps); 1692 while (nr_maps > 0) { 1693 ceph_decode_need(&p, end, 2*sizeof(u32), bad); 1694 epoch = ceph_decode_32(&p); 1695 maplen = ceph_decode_32(&p); 1696 ceph_decode_need(&p, end, maplen, bad); 1697 next = p + maplen; 1698 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) { 1699 dout("applying incremental map %u len %d\n", 1700 epoch, maplen); 1701 newmap = osdmap_apply_incremental(&p, next, 1702 osdc->osdmap, 1703 &osdc->client->msgr); 1704 if (IS_ERR(newmap)) { 1705 err = PTR_ERR(newmap); 1706 goto bad; 1707 } 1708 BUG_ON(!newmap); 1709 if (newmap != osdc->osdmap) { 1710 ceph_osdmap_destroy(osdc->osdmap); 1711 osdc->osdmap = newmap; 1712 } 1713 kick_requests(osdc, 0); 1714 } else { 1715 dout("ignoring incremental map %u len %d\n", 1716 epoch, maplen); 1717 } 1718 p = next; 1719 nr_maps--; 1720 } 1721 if (newmap) 1722 goto done; 1723 1724 /* full maps */ 1725 ceph_decode_32_safe(&p, end, nr_maps, bad); 1726 dout(" %d full maps\n", nr_maps); 1727 while (nr_maps) { 1728 ceph_decode_need(&p, end, 2*sizeof(u32), bad); 1729 epoch = ceph_decode_32(&p); 1730 maplen = ceph_decode_32(&p); 1731 ceph_decode_need(&p, end, maplen, bad); 1732 if (nr_maps > 1) { 1733 dout("skipping non-latest full map %u len %d\n", 1734 epoch, maplen); 1735 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) { 1736 dout("skipping full map %u len %d, " 1737 "older than our %u\n", epoch, maplen, 1738 osdc->osdmap->epoch); 1739 } else { 1740 int skipped_map = 0; 1741 1742 dout("taking full map %u len %d\n", epoch, maplen); 1743 newmap = osdmap_decode(&p, p+maplen); 1744 if (IS_ERR(newmap)) { 1745 err = PTR_ERR(newmap); 1746 goto bad; 1747 } 1748 BUG_ON(!newmap); 1749 oldmap = osdc->osdmap; 1750 osdc->osdmap = newmap; 1751 if (oldmap) { 1752 if (oldmap->epoch + 1 < newmap->epoch) 1753 skipped_map = 1; 1754 ceph_osdmap_destroy(oldmap); 1755 } 1756 kick_requests(osdc, skipped_map); 1757 } 1758 p += maplen; 1759 nr_maps--; 1760 } 1761 1762 done: 1763 downgrade_write(&osdc->map_sem); 1764 ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch); 1765 1766 /* 1767 * subscribe to subsequent osdmap updates if full to ensure 1768 * we find out when we are no longer full and stop returning 1769 * ENOSPC. 1770 */ 1771 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL)) 1772 ceph_monc_request_next_osdmap(&osdc->client->monc); 1773 1774 mutex_lock(&osdc->request_mutex); 1775 __send_queued(osdc); 1776 mutex_unlock(&osdc->request_mutex); 1777 up_read(&osdc->map_sem); 1778 wake_up_all(&osdc->client->auth_wq); 1779 return; 1780 1781 bad: 1782 pr_err("osdc handle_map corrupt msg\n"); 1783 ceph_msg_dump(msg); 1784 up_write(&osdc->map_sem); 1785 return; 1786 } 1787 1788 /* 1789 * watch/notify callback event infrastructure 1790 * 1791 * These callbacks are used both for watch and notify operations. 1792 */ 1793 static void __release_event(struct kref *kref) 1794 { 1795 struct ceph_osd_event *event = 1796 container_of(kref, struct ceph_osd_event, kref); 1797 1798 dout("__release_event %p\n", event); 1799 kfree(event); 1800 } 1801 1802 static void get_event(struct ceph_osd_event *event) 1803 { 1804 kref_get(&event->kref); 1805 } 1806 1807 void ceph_osdc_put_event(struct ceph_osd_event *event) 1808 { 1809 kref_put(&event->kref, __release_event); 1810 } 1811 EXPORT_SYMBOL(ceph_osdc_put_event); 1812 1813 static void __insert_event(struct ceph_osd_client *osdc, 1814 struct ceph_osd_event *new) 1815 { 1816 struct rb_node **p = &osdc->event_tree.rb_node; 1817 struct rb_node *parent = NULL; 1818 struct ceph_osd_event *event = NULL; 1819 1820 while (*p) { 1821 parent = *p; 1822 event = rb_entry(parent, struct ceph_osd_event, node); 1823 if (new->cookie < event->cookie) 1824 p = &(*p)->rb_left; 1825 else if (new->cookie > event->cookie) 1826 p = &(*p)->rb_right; 1827 else 1828 BUG(); 1829 } 1830 1831 rb_link_node(&new->node, parent, p); 1832 rb_insert_color(&new->node, &osdc->event_tree); 1833 } 1834 1835 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc, 1836 u64 cookie) 1837 { 1838 struct rb_node **p = &osdc->event_tree.rb_node; 1839 struct rb_node *parent = NULL; 1840 struct ceph_osd_event *event = NULL; 1841 1842 while (*p) { 1843 parent = *p; 1844 event = rb_entry(parent, struct ceph_osd_event, node); 1845 if (cookie < event->cookie) 1846 p = &(*p)->rb_left; 1847 else if (cookie > event->cookie) 1848 p = &(*p)->rb_right; 1849 else 1850 return event; 1851 } 1852 return NULL; 1853 } 1854 1855 static void __remove_event(struct ceph_osd_event *event) 1856 { 1857 struct ceph_osd_client *osdc = event->osdc; 1858 1859 if (!RB_EMPTY_NODE(&event->node)) { 1860 dout("__remove_event removed %p\n", event); 1861 rb_erase(&event->node, &osdc->event_tree); 1862 ceph_osdc_put_event(event); 1863 } else { 1864 dout("__remove_event didn't remove %p\n", event); 1865 } 1866 } 1867 1868 int ceph_osdc_create_event(struct ceph_osd_client *osdc, 1869 void (*event_cb)(u64, u64, u8, void *), 1870 void *data, struct ceph_osd_event **pevent) 1871 { 1872 struct ceph_osd_event *event; 1873 1874 event = kmalloc(sizeof(*event), GFP_NOIO); 1875 if (!event) 1876 return -ENOMEM; 1877 1878 dout("create_event %p\n", event); 1879 event->cb = event_cb; 1880 event->one_shot = 0; 1881 event->data = data; 1882 event->osdc = osdc; 1883 INIT_LIST_HEAD(&event->osd_node); 1884 RB_CLEAR_NODE(&event->node); 1885 kref_init(&event->kref); /* one ref for us */ 1886 kref_get(&event->kref); /* one ref for the caller */ 1887 1888 spin_lock(&osdc->event_lock); 1889 event->cookie = ++osdc->event_count; 1890 __insert_event(osdc, event); 1891 spin_unlock(&osdc->event_lock); 1892 1893 *pevent = event; 1894 return 0; 1895 } 1896 EXPORT_SYMBOL(ceph_osdc_create_event); 1897 1898 void ceph_osdc_cancel_event(struct ceph_osd_event *event) 1899 { 1900 struct ceph_osd_client *osdc = event->osdc; 1901 1902 dout("cancel_event %p\n", event); 1903 spin_lock(&osdc->event_lock); 1904 __remove_event(event); 1905 spin_unlock(&osdc->event_lock); 1906 ceph_osdc_put_event(event); /* caller's */ 1907 } 1908 EXPORT_SYMBOL(ceph_osdc_cancel_event); 1909 1910 1911 static void do_event_work(struct work_struct *work) 1912 { 1913 struct ceph_osd_event_work *event_work = 1914 container_of(work, struct ceph_osd_event_work, work); 1915 struct ceph_osd_event *event = event_work->event; 1916 u64 ver = event_work->ver; 1917 u64 notify_id = event_work->notify_id; 1918 u8 opcode = event_work->opcode; 1919 1920 dout("do_event_work completing %p\n", event); 1921 event->cb(ver, notify_id, opcode, event->data); 1922 dout("do_event_work completed %p\n", event); 1923 ceph_osdc_put_event(event); 1924 kfree(event_work); 1925 } 1926 1927 1928 /* 1929 * Process osd watch notifications 1930 */ 1931 static void handle_watch_notify(struct ceph_osd_client *osdc, 1932 struct ceph_msg *msg) 1933 { 1934 void *p, *end; 1935 u8 proto_ver; 1936 u64 cookie, ver, notify_id; 1937 u8 opcode; 1938 struct ceph_osd_event *event; 1939 struct ceph_osd_event_work *event_work; 1940 1941 p = msg->front.iov_base; 1942 end = p + msg->front.iov_len; 1943 1944 ceph_decode_8_safe(&p, end, proto_ver, bad); 1945 ceph_decode_8_safe(&p, end, opcode, bad); 1946 ceph_decode_64_safe(&p, end, cookie, bad); 1947 ceph_decode_64_safe(&p, end, ver, bad); 1948 ceph_decode_64_safe(&p, end, notify_id, bad); 1949 1950 spin_lock(&osdc->event_lock); 1951 event = __find_event(osdc, cookie); 1952 if (event) { 1953 BUG_ON(event->one_shot); 1954 get_event(event); 1955 } 1956 spin_unlock(&osdc->event_lock); 1957 dout("handle_watch_notify cookie %lld ver %lld event %p\n", 1958 cookie, ver, event); 1959 if (event) { 1960 event_work = kmalloc(sizeof(*event_work), GFP_NOIO); 1961 if (!event_work) { 1962 dout("ERROR: could not allocate event_work\n"); 1963 goto done_err; 1964 } 1965 INIT_WORK(&event_work->work, do_event_work); 1966 event_work->event = event; 1967 event_work->ver = ver; 1968 event_work->notify_id = notify_id; 1969 event_work->opcode = opcode; 1970 if (!queue_work(osdc->notify_wq, &event_work->work)) { 1971 dout("WARNING: failed to queue notify event work\n"); 1972 goto done_err; 1973 } 1974 } 1975 1976 return; 1977 1978 done_err: 1979 ceph_osdc_put_event(event); 1980 return; 1981 1982 bad: 1983 pr_err("osdc handle_watch_notify corrupt msg\n"); 1984 return; 1985 } 1986 1987 /* 1988 * build new request AND message 1989 * 1990 */ 1991 void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off, 1992 struct ceph_snap_context *snapc, u64 snap_id, 1993 struct timespec *mtime) 1994 { 1995 struct ceph_msg *msg = req->r_request; 1996 void *p; 1997 size_t msg_size; 1998 int flags = req->r_flags; 1999 u64 data_len; 2000 unsigned int i; 2001 2002 req->r_snapid = snap_id; 2003 req->r_snapc = ceph_get_snap_context(snapc); 2004 2005 /* encode request */ 2006 msg->hdr.version = cpu_to_le16(4); 2007 2008 p = msg->front.iov_base; 2009 ceph_encode_32(&p, 1); /* client_inc is always 1 */ 2010 req->r_request_osdmap_epoch = p; 2011 p += 4; 2012 req->r_request_flags = p; 2013 p += 4; 2014 if (req->r_flags & CEPH_OSD_FLAG_WRITE) 2015 ceph_encode_timespec(p, mtime); 2016 p += sizeof(struct ceph_timespec); 2017 req->r_request_reassert_version = p; 2018 p += sizeof(struct ceph_eversion); /* will get filled in */ 2019 2020 /* oloc */ 2021 ceph_encode_8(&p, 4); 2022 ceph_encode_8(&p, 4); 2023 ceph_encode_32(&p, 8 + 4 + 4); 2024 req->r_request_pool = p; 2025 p += 8; 2026 ceph_encode_32(&p, -1); /* preferred */ 2027 ceph_encode_32(&p, 0); /* key len */ 2028 2029 ceph_encode_8(&p, 1); 2030 req->r_request_pgid = p; 2031 p += 8 + 4; 2032 ceph_encode_32(&p, -1); /* preferred */ 2033 2034 /* oid */ 2035 ceph_encode_32(&p, req->r_oid_len); 2036 memcpy(p, req->r_oid, req->r_oid_len); 2037 dout("oid '%.*s' len %d\n", req->r_oid_len, req->r_oid, req->r_oid_len); 2038 p += req->r_oid_len; 2039 2040 /* ops--can imply data */ 2041 ceph_encode_16(&p, (u16)req->r_num_ops); 2042 data_len = 0; 2043 for (i = 0; i < req->r_num_ops; i++) { 2044 data_len += osd_req_encode_op(req, p, i); 2045 p += sizeof(struct ceph_osd_op); 2046 } 2047 2048 /* snaps */ 2049 ceph_encode_64(&p, req->r_snapid); 2050 ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0); 2051 ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0); 2052 if (req->r_snapc) { 2053 for (i = 0; i < snapc->num_snaps; i++) { 2054 ceph_encode_64(&p, req->r_snapc->snaps[i]); 2055 } 2056 } 2057 2058 req->r_request_attempts = p; 2059 p += 4; 2060 2061 /* data */ 2062 if (flags & CEPH_OSD_FLAG_WRITE) { 2063 u16 data_off; 2064 2065 /* 2066 * The header "data_off" is a hint to the receiver 2067 * allowing it to align received data into its 2068 * buffers such that there's no need to re-copy 2069 * it before writing it to disk (direct I/O). 2070 */ 2071 data_off = (u16) (off & 0xffff); 2072 req->r_request->hdr.data_off = cpu_to_le16(data_off); 2073 } 2074 req->r_request->hdr.data_len = cpu_to_le32(data_len); 2075 2076 BUG_ON(p > msg->front.iov_base + msg->front.iov_len); 2077 msg_size = p - msg->front.iov_base; 2078 msg->front.iov_len = msg_size; 2079 msg->hdr.front_len = cpu_to_le32(msg_size); 2080 2081 dout("build_request msg_size was %d\n", (int)msg_size); 2082 } 2083 EXPORT_SYMBOL(ceph_osdc_build_request); 2084 2085 /* 2086 * Register request, send initial attempt. 2087 */ 2088 int ceph_osdc_start_request(struct ceph_osd_client *osdc, 2089 struct ceph_osd_request *req, 2090 bool nofail) 2091 { 2092 int rc = 0; 2093 2094 down_read(&osdc->map_sem); 2095 mutex_lock(&osdc->request_mutex); 2096 __register_request(osdc, req); 2097 WARN_ON(req->r_sent); 2098 rc = __map_request(osdc, req, 0); 2099 if (rc < 0) { 2100 if (nofail) { 2101 dout("osdc_start_request failed map, " 2102 " will retry %lld\n", req->r_tid); 2103 rc = 0; 2104 } 2105 goto out_unlock; 2106 } 2107 if (req->r_osd == NULL) { 2108 dout("send_request %p no up osds in pg\n", req); 2109 ceph_monc_request_next_osdmap(&osdc->client->monc); 2110 } else { 2111 __send_queued(osdc); 2112 } 2113 rc = 0; 2114 out_unlock: 2115 mutex_unlock(&osdc->request_mutex); 2116 up_read(&osdc->map_sem); 2117 return rc; 2118 } 2119 EXPORT_SYMBOL(ceph_osdc_start_request); 2120 2121 /* 2122 * wait for a request to complete 2123 */ 2124 int ceph_osdc_wait_request(struct ceph_osd_client *osdc, 2125 struct ceph_osd_request *req) 2126 { 2127 int rc; 2128 2129 rc = wait_for_completion_interruptible(&req->r_completion); 2130 if (rc < 0) { 2131 mutex_lock(&osdc->request_mutex); 2132 __cancel_request(req); 2133 __unregister_request(osdc, req); 2134 mutex_unlock(&osdc->request_mutex); 2135 complete_request(req); 2136 dout("wait_request tid %llu canceled/timed out\n", req->r_tid); 2137 return rc; 2138 } 2139 2140 dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result); 2141 return req->r_result; 2142 } 2143 EXPORT_SYMBOL(ceph_osdc_wait_request); 2144 2145 /* 2146 * sync - wait for all in-flight requests to flush. avoid starvation. 2147 */ 2148 void ceph_osdc_sync(struct ceph_osd_client *osdc) 2149 { 2150 struct ceph_osd_request *req; 2151 u64 last_tid, next_tid = 0; 2152 2153 mutex_lock(&osdc->request_mutex); 2154 last_tid = osdc->last_tid; 2155 while (1) { 2156 req = __lookup_request_ge(osdc, next_tid); 2157 if (!req) 2158 break; 2159 if (req->r_tid > last_tid) 2160 break; 2161 2162 next_tid = req->r_tid + 1; 2163 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0) 2164 continue; 2165 2166 ceph_osdc_get_request(req); 2167 mutex_unlock(&osdc->request_mutex); 2168 dout("sync waiting on tid %llu (last is %llu)\n", 2169 req->r_tid, last_tid); 2170 wait_for_completion(&req->r_safe_completion); 2171 mutex_lock(&osdc->request_mutex); 2172 ceph_osdc_put_request(req); 2173 } 2174 mutex_unlock(&osdc->request_mutex); 2175 dout("sync done (thru tid %llu)\n", last_tid); 2176 } 2177 EXPORT_SYMBOL(ceph_osdc_sync); 2178 2179 /* 2180 * init, shutdown 2181 */ 2182 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client) 2183 { 2184 int err; 2185 2186 dout("init\n"); 2187 osdc->client = client; 2188 osdc->osdmap = NULL; 2189 init_rwsem(&osdc->map_sem); 2190 init_completion(&osdc->map_waiters); 2191 osdc->last_requested_map = 0; 2192 mutex_init(&osdc->request_mutex); 2193 osdc->last_tid = 0; 2194 osdc->osds = RB_ROOT; 2195 INIT_LIST_HEAD(&osdc->osd_lru); 2196 osdc->requests = RB_ROOT; 2197 INIT_LIST_HEAD(&osdc->req_lru); 2198 INIT_LIST_HEAD(&osdc->req_unsent); 2199 INIT_LIST_HEAD(&osdc->req_notarget); 2200 INIT_LIST_HEAD(&osdc->req_linger); 2201 osdc->num_requests = 0; 2202 INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout); 2203 INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout); 2204 spin_lock_init(&osdc->event_lock); 2205 osdc->event_tree = RB_ROOT; 2206 osdc->event_count = 0; 2207 2208 schedule_delayed_work(&osdc->osds_timeout_work, 2209 round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ)); 2210 2211 err = -ENOMEM; 2212 osdc->req_mempool = mempool_create_kmalloc_pool(10, 2213 sizeof(struct ceph_osd_request)); 2214 if (!osdc->req_mempool) 2215 goto out; 2216 2217 err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP, 2218 OSD_OP_FRONT_LEN, 10, true, 2219 "osd_op"); 2220 if (err < 0) 2221 goto out_mempool; 2222 err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY, 2223 OSD_OPREPLY_FRONT_LEN, 10, true, 2224 "osd_op_reply"); 2225 if (err < 0) 2226 goto out_msgpool; 2227 2228 osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify"); 2229 if (IS_ERR(osdc->notify_wq)) { 2230 err = PTR_ERR(osdc->notify_wq); 2231 osdc->notify_wq = NULL; 2232 goto out_msgpool; 2233 } 2234 return 0; 2235 2236 out_msgpool: 2237 ceph_msgpool_destroy(&osdc->msgpool_op); 2238 out_mempool: 2239 mempool_destroy(osdc->req_mempool); 2240 out: 2241 return err; 2242 } 2243 2244 void ceph_osdc_stop(struct ceph_osd_client *osdc) 2245 { 2246 flush_workqueue(osdc->notify_wq); 2247 destroy_workqueue(osdc->notify_wq); 2248 cancel_delayed_work_sync(&osdc->timeout_work); 2249 cancel_delayed_work_sync(&osdc->osds_timeout_work); 2250 if (osdc->osdmap) { 2251 ceph_osdmap_destroy(osdc->osdmap); 2252 osdc->osdmap = NULL; 2253 } 2254 remove_all_osds(osdc); 2255 mempool_destroy(osdc->req_mempool); 2256 ceph_msgpool_destroy(&osdc->msgpool_op); 2257 ceph_msgpool_destroy(&osdc->msgpool_op_reply); 2258 } 2259 2260 /* 2261 * Read some contiguous pages. If we cross a stripe boundary, shorten 2262 * *plen. Return number of bytes read, or error. 2263 */ 2264 int ceph_osdc_readpages(struct ceph_osd_client *osdc, 2265 struct ceph_vino vino, struct ceph_file_layout *layout, 2266 u64 off, u64 *plen, 2267 u32 truncate_seq, u64 truncate_size, 2268 struct page **pages, int num_pages, int page_align) 2269 { 2270 struct ceph_osd_request *req; 2271 int rc = 0; 2272 2273 dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino, 2274 vino.snap, off, *plen); 2275 req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 1, 2276 CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ, 2277 NULL, truncate_seq, truncate_size, 2278 false); 2279 if (IS_ERR(req)) 2280 return PTR_ERR(req); 2281 2282 /* it may be a short read due to an object boundary */ 2283 2284 osd_req_op_extent_osd_data_pages(req, 0, 2285 pages, *plen, page_align, false, false); 2286 2287 dout("readpages final extent is %llu~%llu (%llu bytes align %d)\n", 2288 off, *plen, *plen, page_align); 2289 2290 ceph_osdc_build_request(req, off, NULL, vino.snap, NULL); 2291 2292 rc = ceph_osdc_start_request(osdc, req, false); 2293 if (!rc) 2294 rc = ceph_osdc_wait_request(osdc, req); 2295 2296 ceph_osdc_put_request(req); 2297 dout("readpages result %d\n", rc); 2298 return rc; 2299 } 2300 EXPORT_SYMBOL(ceph_osdc_readpages); 2301 2302 /* 2303 * do a synchronous write on N pages 2304 */ 2305 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino, 2306 struct ceph_file_layout *layout, 2307 struct ceph_snap_context *snapc, 2308 u64 off, u64 len, 2309 u32 truncate_seq, u64 truncate_size, 2310 struct timespec *mtime, 2311 struct page **pages, int num_pages) 2312 { 2313 struct ceph_osd_request *req; 2314 int rc = 0; 2315 int page_align = off & ~PAGE_MASK; 2316 2317 BUG_ON(vino.snap != CEPH_NOSNAP); /* snapshots aren't writeable */ 2318 req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 1, 2319 CEPH_OSD_OP_WRITE, 2320 CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE, 2321 snapc, truncate_seq, truncate_size, 2322 true); 2323 if (IS_ERR(req)) 2324 return PTR_ERR(req); 2325 2326 /* it may be a short write due to an object boundary */ 2327 osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align, 2328 false, false); 2329 dout("writepages %llu~%llu (%llu bytes)\n", off, len, len); 2330 2331 ceph_osdc_build_request(req, off, snapc, CEPH_NOSNAP, mtime); 2332 2333 rc = ceph_osdc_start_request(osdc, req, true); 2334 if (!rc) 2335 rc = ceph_osdc_wait_request(osdc, req); 2336 2337 ceph_osdc_put_request(req); 2338 if (rc == 0) 2339 rc = len; 2340 dout("writepages result %d\n", rc); 2341 return rc; 2342 } 2343 EXPORT_SYMBOL(ceph_osdc_writepages); 2344 2345 /* 2346 * handle incoming message 2347 */ 2348 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg) 2349 { 2350 struct ceph_osd *osd = con->private; 2351 struct ceph_osd_client *osdc; 2352 int type = le16_to_cpu(msg->hdr.type); 2353 2354 if (!osd) 2355 goto out; 2356 osdc = osd->o_osdc; 2357 2358 switch (type) { 2359 case CEPH_MSG_OSD_MAP: 2360 ceph_osdc_handle_map(osdc, msg); 2361 break; 2362 case CEPH_MSG_OSD_OPREPLY: 2363 handle_reply(osdc, msg, con); 2364 break; 2365 case CEPH_MSG_WATCH_NOTIFY: 2366 handle_watch_notify(osdc, msg); 2367 break; 2368 2369 default: 2370 pr_err("received unknown message type %d %s\n", type, 2371 ceph_msg_type_name(type)); 2372 } 2373 out: 2374 ceph_msg_put(msg); 2375 } 2376 2377 /* 2378 * lookup and return message for incoming reply. set up reply message 2379 * pages. 2380 */ 2381 static struct ceph_msg *get_reply(struct ceph_connection *con, 2382 struct ceph_msg_header *hdr, 2383 int *skip) 2384 { 2385 struct ceph_osd *osd = con->private; 2386 struct ceph_osd_client *osdc = osd->o_osdc; 2387 struct ceph_msg *m; 2388 struct ceph_osd_request *req; 2389 int front = le32_to_cpu(hdr->front_len); 2390 int data_len = le32_to_cpu(hdr->data_len); 2391 u64 tid; 2392 2393 tid = le64_to_cpu(hdr->tid); 2394 mutex_lock(&osdc->request_mutex); 2395 req = __lookup_request(osdc, tid); 2396 if (!req) { 2397 *skip = 1; 2398 m = NULL; 2399 dout("get_reply unknown tid %llu from osd%d\n", tid, 2400 osd->o_osd); 2401 goto out; 2402 } 2403 2404 if (req->r_reply->con) 2405 dout("%s revoking msg %p from old con %p\n", __func__, 2406 req->r_reply, req->r_reply->con); 2407 ceph_msg_revoke_incoming(req->r_reply); 2408 2409 if (front > req->r_reply->front.iov_len) { 2410 pr_warning("get_reply front %d > preallocated %d\n", 2411 front, (int)req->r_reply->front.iov_len); 2412 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS, false); 2413 if (!m) 2414 goto out; 2415 ceph_msg_put(req->r_reply); 2416 req->r_reply = m; 2417 } 2418 m = ceph_msg_get(req->r_reply); 2419 2420 if (data_len > 0) { 2421 struct ceph_osd_data *osd_data; 2422 2423 /* 2424 * XXX This is assuming there is only one op containing 2425 * XXX page data. Probably OK for reads, but this 2426 * XXX ought to be done more generally. 2427 */ 2428 osd_data = osd_req_op_extent_osd_data(req, 0); 2429 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 2430 if (osd_data->pages && 2431 unlikely(osd_data->length < data_len)) { 2432 2433 pr_warning("tid %lld reply has %d bytes " 2434 "we had only %llu bytes ready\n", 2435 tid, data_len, osd_data->length); 2436 *skip = 1; 2437 ceph_msg_put(m); 2438 m = NULL; 2439 goto out; 2440 } 2441 } 2442 } 2443 *skip = 0; 2444 dout("get_reply tid %lld %p\n", tid, m); 2445 2446 out: 2447 mutex_unlock(&osdc->request_mutex); 2448 return m; 2449 2450 } 2451 2452 static struct ceph_msg *alloc_msg(struct ceph_connection *con, 2453 struct ceph_msg_header *hdr, 2454 int *skip) 2455 { 2456 struct ceph_osd *osd = con->private; 2457 int type = le16_to_cpu(hdr->type); 2458 int front = le32_to_cpu(hdr->front_len); 2459 2460 *skip = 0; 2461 switch (type) { 2462 case CEPH_MSG_OSD_MAP: 2463 case CEPH_MSG_WATCH_NOTIFY: 2464 return ceph_msg_new(type, front, GFP_NOFS, false); 2465 case CEPH_MSG_OSD_OPREPLY: 2466 return get_reply(con, hdr, skip); 2467 default: 2468 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type, 2469 osd->o_osd); 2470 *skip = 1; 2471 return NULL; 2472 } 2473 } 2474 2475 /* 2476 * Wrappers to refcount containing ceph_osd struct 2477 */ 2478 static struct ceph_connection *get_osd_con(struct ceph_connection *con) 2479 { 2480 struct ceph_osd *osd = con->private; 2481 if (get_osd(osd)) 2482 return con; 2483 return NULL; 2484 } 2485 2486 static void put_osd_con(struct ceph_connection *con) 2487 { 2488 struct ceph_osd *osd = con->private; 2489 put_osd(osd); 2490 } 2491 2492 /* 2493 * authentication 2494 */ 2495 /* 2496 * Note: returned pointer is the address of a structure that's 2497 * managed separately. Caller must *not* attempt to free it. 2498 */ 2499 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con, 2500 int *proto, int force_new) 2501 { 2502 struct ceph_osd *o = con->private; 2503 struct ceph_osd_client *osdc = o->o_osdc; 2504 struct ceph_auth_client *ac = osdc->client->monc.auth; 2505 struct ceph_auth_handshake *auth = &o->o_auth; 2506 2507 if (force_new && auth->authorizer) { 2508 ceph_auth_destroy_authorizer(ac, auth->authorizer); 2509 auth->authorizer = NULL; 2510 } 2511 if (!auth->authorizer) { 2512 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 2513 auth); 2514 if (ret) 2515 return ERR_PTR(ret); 2516 } else { 2517 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 2518 auth); 2519 if (ret) 2520 return ERR_PTR(ret); 2521 } 2522 *proto = ac->protocol; 2523 2524 return auth; 2525 } 2526 2527 2528 static int verify_authorizer_reply(struct ceph_connection *con, int len) 2529 { 2530 struct ceph_osd *o = con->private; 2531 struct ceph_osd_client *osdc = o->o_osdc; 2532 struct ceph_auth_client *ac = osdc->client->monc.auth; 2533 2534 return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len); 2535 } 2536 2537 static int invalidate_authorizer(struct ceph_connection *con) 2538 { 2539 struct ceph_osd *o = con->private; 2540 struct ceph_osd_client *osdc = o->o_osdc; 2541 struct ceph_auth_client *ac = osdc->client->monc.auth; 2542 2543 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD); 2544 return ceph_monc_validate_auth(&osdc->client->monc); 2545 } 2546 2547 static const struct ceph_connection_operations osd_con_ops = { 2548 .get = get_osd_con, 2549 .put = put_osd_con, 2550 .dispatch = dispatch, 2551 .get_authorizer = get_authorizer, 2552 .verify_authorizer_reply = verify_authorizer_reply, 2553 .invalidate_authorizer = invalidate_authorizer, 2554 .alloc_msg = alloc_msg, 2555 .fault = osd_reset, 2556 }; 2557