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, bool write_request) 121 { 122 BUG_ON(which >= osd_req->r_num_ops); 123 124 /* return &osd_req->r_ops[which].extent.osd_data; */ 125 return write_request ? &osd_req->r_data_out : &osd_req->r_data_in; 126 } 127 EXPORT_SYMBOL(osd_req_op_extent_osd_data); 128 129 struct ceph_osd_data * 130 osd_req_op_cls_request_info(struct ceph_osd_request *osd_req, 131 unsigned int which) 132 { 133 BUG_ON(which >= osd_req->r_num_ops); 134 135 /* return &osd_req->r_ops[which].cls.request_info; */ 136 return &osd_req->r_data_out; /* Request data is outgoing */ 137 } 138 EXPORT_SYMBOL(osd_req_op_cls_request_info); /* ??? */ 139 140 struct ceph_osd_data * 141 osd_req_op_cls_response_data(struct ceph_osd_request *osd_req, 142 unsigned int which) 143 { 144 BUG_ON(which >= osd_req->r_num_ops); 145 146 /* return &osd_req->r_ops[which].cls.response_data; */ 147 return &osd_req->r_data_in; /* Response data is incoming */ 148 } 149 EXPORT_SYMBOL(osd_req_op_cls_response_data); /* ??? */ 150 151 void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *osd_req, 152 unsigned int which, bool write_request, 153 struct page **pages, u64 length, u32 alignment, 154 bool pages_from_pool, bool own_pages) 155 { 156 struct ceph_osd_data *osd_data; 157 158 osd_data = osd_req_op_extent_osd_data(osd_req, which, write_request); 159 ceph_osd_data_pages_init(osd_data, pages, length, alignment, 160 pages_from_pool, own_pages); 161 162 osd_req->r_ops[which].extent.osd_data = 163 osd_req_op_extent_osd_data(osd_req, which, write_request); 164 } 165 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pages); 166 167 void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *osd_req, 168 unsigned int which, bool write_request, 169 struct ceph_pagelist *pagelist) 170 { 171 struct ceph_osd_data *osd_data; 172 173 osd_data = osd_req_op_extent_osd_data(osd_req, which, write_request); 174 ceph_osd_data_pagelist_init(osd_data, pagelist); 175 176 osd_req->r_ops[which].extent.osd_data = 177 osd_req_op_extent_osd_data(osd_req, which, write_request); 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, bool write_request, 184 struct bio *bio, size_t bio_length) 185 { 186 struct ceph_osd_data *osd_data; 187 188 osd_data = osd_req_op_extent_osd_data(osd_req, which, write_request); 189 ceph_osd_data_bio_init(osd_data, bio, bio_length); 190 191 osd_req->r_ops[which].extent.osd_data = 192 osd_req_op_extent_osd_data(osd_req, which, write_request); 193 } 194 EXPORT_SYMBOL(osd_req_op_extent_osd_data_bio); 195 #endif /* CONFIG_BLOCK */ 196 197 static void osd_req_op_cls_request_info_pagelist( 198 struct ceph_osd_request *osd_req, 199 unsigned int which, struct ceph_pagelist *pagelist) 200 { 201 struct ceph_osd_data *osd_data; 202 203 osd_data = osd_req_op_cls_request_info(osd_req, which); 204 ceph_osd_data_pagelist_init(osd_data, pagelist); 205 206 osd_req->r_ops[which].cls.request_info = 207 osd_req_op_cls_request_info(osd_req, which); 208 } 209 210 void osd_req_op_cls_response_data_pages(struct ceph_osd_request *osd_req, 211 unsigned int which, struct page **pages, u64 length, 212 u32 alignment, bool pages_from_pool, bool own_pages) 213 { 214 struct ceph_osd_data *osd_data; 215 216 osd_data = osd_req_op_cls_response_data(osd_req, which); 217 ceph_osd_data_pages_init(osd_data, pages, length, alignment, 218 pages_from_pool, own_pages); 219 220 osd_req->r_ops[which].cls.response_data = 221 osd_req_op_cls_response_data(osd_req, which); 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 static void ceph_osd_data_release(struct ceph_osd_data *osd_data) 245 { 246 if (osd_data->type != CEPH_OSD_DATA_TYPE_PAGES) 247 return; 248 249 if (osd_data->own_pages) { 250 int num_pages; 251 252 num_pages = calc_pages_for((u64)osd_data->alignment, 253 (u64)osd_data->length); 254 ceph_release_page_vector(osd_data->pages, num_pages); 255 } 256 } 257 258 /* 259 * requests 260 */ 261 void ceph_osdc_release_request(struct kref *kref) 262 { 263 struct ceph_osd_request *req; 264 265 req = container_of(kref, struct ceph_osd_request, r_kref); 266 if (req->r_request) 267 ceph_msg_put(req->r_request); 268 if (req->r_reply) { 269 ceph_msg_revoke_incoming(req->r_reply); 270 ceph_msg_put(req->r_reply); 271 } 272 273 ceph_osd_data_release(&req->r_data_in); 274 ceph_osd_data_release(&req->r_data_out); 275 276 ceph_put_snap_context(req->r_snapc); 277 if (req->r_mempool) 278 mempool_free(req, req->r_osdc->req_mempool); 279 else 280 kfree(req); 281 } 282 EXPORT_SYMBOL(ceph_osdc_release_request); 283 284 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc, 285 struct ceph_snap_context *snapc, 286 unsigned int num_ops, 287 bool use_mempool, 288 gfp_t gfp_flags) 289 { 290 struct ceph_osd_request *req; 291 struct ceph_msg *msg; 292 size_t msg_size; 293 294 BUILD_BUG_ON(CEPH_OSD_MAX_OP > U16_MAX); 295 BUG_ON(num_ops > CEPH_OSD_MAX_OP); 296 297 msg_size = 4 + 4 + 8 + 8 + 4+8; 298 msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */ 299 msg_size += 1 + 8 + 4 + 4; /* pg_t */ 300 msg_size += 4 + MAX_OBJ_NAME_SIZE; 301 msg_size += 2 + num_ops*sizeof(struct ceph_osd_op); 302 msg_size += 8; /* snapid */ 303 msg_size += 8; /* snap_seq */ 304 msg_size += 8 * (snapc ? snapc->num_snaps : 0); /* snaps */ 305 msg_size += 4; 306 307 if (use_mempool) { 308 req = mempool_alloc(osdc->req_mempool, gfp_flags); 309 memset(req, 0, sizeof(*req)); 310 } else { 311 req = kzalloc(sizeof(*req), gfp_flags); 312 } 313 if (req == NULL) 314 return NULL; 315 316 req->r_osdc = osdc; 317 req->r_mempool = use_mempool; 318 req->r_num_ops = num_ops; 319 320 kref_init(&req->r_kref); 321 init_completion(&req->r_completion); 322 init_completion(&req->r_safe_completion); 323 RB_CLEAR_NODE(&req->r_node); 324 INIT_LIST_HEAD(&req->r_unsafe_item); 325 INIT_LIST_HEAD(&req->r_linger_item); 326 INIT_LIST_HEAD(&req->r_linger_osd); 327 INIT_LIST_HEAD(&req->r_req_lru_item); 328 INIT_LIST_HEAD(&req->r_osd_item); 329 330 /* create reply message */ 331 if (use_mempool) 332 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0); 333 else 334 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, 335 OSD_OPREPLY_FRONT_LEN, gfp_flags, true); 336 if (!msg) { 337 ceph_osdc_put_request(req); 338 return NULL; 339 } 340 req->r_reply = msg; 341 342 ceph_osd_data_init(&req->r_data_in); 343 ceph_osd_data_init(&req->r_data_out); 344 345 /* create request message; allow space for oid */ 346 if (use_mempool) 347 msg = ceph_msgpool_get(&osdc->msgpool_op, 0); 348 else 349 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true); 350 if (!msg) { 351 ceph_osdc_put_request(req); 352 return NULL; 353 } 354 355 memset(msg->front.iov_base, 0, msg->front.iov_len); 356 357 req->r_request = msg; 358 359 return req; 360 } 361 EXPORT_SYMBOL(ceph_osdc_alloc_request); 362 363 static bool osd_req_opcode_valid(u16 opcode) 364 { 365 switch (opcode) { 366 case CEPH_OSD_OP_READ: 367 case CEPH_OSD_OP_STAT: 368 case CEPH_OSD_OP_MAPEXT: 369 case CEPH_OSD_OP_MASKTRUNC: 370 case CEPH_OSD_OP_SPARSE_READ: 371 case CEPH_OSD_OP_NOTIFY: 372 case CEPH_OSD_OP_NOTIFY_ACK: 373 case CEPH_OSD_OP_ASSERT_VER: 374 case CEPH_OSD_OP_WRITE: 375 case CEPH_OSD_OP_WRITEFULL: 376 case CEPH_OSD_OP_TRUNCATE: 377 case CEPH_OSD_OP_ZERO: 378 case CEPH_OSD_OP_DELETE: 379 case CEPH_OSD_OP_APPEND: 380 case CEPH_OSD_OP_STARTSYNC: 381 case CEPH_OSD_OP_SETTRUNC: 382 case CEPH_OSD_OP_TRIMTRUNC: 383 case CEPH_OSD_OP_TMAPUP: 384 case CEPH_OSD_OP_TMAPPUT: 385 case CEPH_OSD_OP_TMAPGET: 386 case CEPH_OSD_OP_CREATE: 387 case CEPH_OSD_OP_ROLLBACK: 388 case CEPH_OSD_OP_WATCH: 389 case CEPH_OSD_OP_OMAPGETKEYS: 390 case CEPH_OSD_OP_OMAPGETVALS: 391 case CEPH_OSD_OP_OMAPGETHEADER: 392 case CEPH_OSD_OP_OMAPGETVALSBYKEYS: 393 case CEPH_OSD_OP_OMAPSETVALS: 394 case CEPH_OSD_OP_OMAPSETHEADER: 395 case CEPH_OSD_OP_OMAPCLEAR: 396 case CEPH_OSD_OP_OMAPRMKEYS: 397 case CEPH_OSD_OP_OMAP_CMP: 398 case CEPH_OSD_OP_CLONERANGE: 399 case CEPH_OSD_OP_ASSERT_SRC_VERSION: 400 case CEPH_OSD_OP_SRC_CMPXATTR: 401 case CEPH_OSD_OP_GETXATTR: 402 case CEPH_OSD_OP_GETXATTRS: 403 case CEPH_OSD_OP_CMPXATTR: 404 case CEPH_OSD_OP_SETXATTR: 405 case CEPH_OSD_OP_SETXATTRS: 406 case CEPH_OSD_OP_RESETXATTRS: 407 case CEPH_OSD_OP_RMXATTR: 408 case CEPH_OSD_OP_PULL: 409 case CEPH_OSD_OP_PUSH: 410 case CEPH_OSD_OP_BALANCEREADS: 411 case CEPH_OSD_OP_UNBALANCEREADS: 412 case CEPH_OSD_OP_SCRUB: 413 case CEPH_OSD_OP_SCRUB_RESERVE: 414 case CEPH_OSD_OP_SCRUB_UNRESERVE: 415 case CEPH_OSD_OP_SCRUB_STOP: 416 case CEPH_OSD_OP_SCRUB_MAP: 417 case CEPH_OSD_OP_WRLOCK: 418 case CEPH_OSD_OP_WRUNLOCK: 419 case CEPH_OSD_OP_RDLOCK: 420 case CEPH_OSD_OP_RDUNLOCK: 421 case CEPH_OSD_OP_UPLOCK: 422 case CEPH_OSD_OP_DNLOCK: 423 case CEPH_OSD_OP_CALL: 424 case CEPH_OSD_OP_PGLS: 425 case CEPH_OSD_OP_PGLS_FILTER: 426 return true; 427 default: 428 return false; 429 } 430 } 431 432 /* 433 * This is an osd op init function for opcodes that have no data or 434 * other information associated with them. It also serves as a 435 * common init routine for all the other init functions, below. 436 */ 437 static struct ceph_osd_req_op * 438 osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which, 439 u16 opcode) 440 { 441 struct ceph_osd_req_op *op; 442 443 BUG_ON(which >= osd_req->r_num_ops); 444 BUG_ON(!osd_req_opcode_valid(opcode)); 445 446 op = &osd_req->r_ops[which]; 447 memset(op, 0, sizeof (*op)); 448 op->op = opcode; 449 450 return op; 451 } 452 453 void osd_req_op_extent_init(struct ceph_osd_request *osd_req, 454 unsigned int which, u16 opcode, 455 u64 offset, u64 length, 456 u64 truncate_size, u32 truncate_seq) 457 { 458 struct ceph_osd_req_op *op = osd_req_op_init(osd_req, which, opcode); 459 size_t payload_len = 0; 460 461 BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE); 462 463 op->extent.offset = offset; 464 op->extent.length = length; 465 op->extent.truncate_size = truncate_size; 466 op->extent.truncate_seq = truncate_seq; 467 if (opcode == CEPH_OSD_OP_WRITE) 468 payload_len += length; 469 470 op->payload_len = payload_len; 471 } 472 EXPORT_SYMBOL(osd_req_op_extent_init); 473 474 void osd_req_op_extent_update(struct ceph_osd_request *osd_req, 475 unsigned int which, u64 length) 476 { 477 struct ceph_osd_req_op *op; 478 u64 previous; 479 480 BUG_ON(which >= osd_req->r_num_ops); 481 op = &osd_req->r_ops[which]; 482 previous = op->extent.length; 483 484 if (length == previous) 485 return; /* Nothing to do */ 486 BUG_ON(length > previous); 487 488 op->extent.length = length; 489 op->payload_len -= previous - length; 490 } 491 EXPORT_SYMBOL(osd_req_op_extent_update); 492 493 void osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which, 494 u16 opcode, const char *class, const char *method, 495 const void *request_data, size_t request_data_size) 496 { 497 struct ceph_osd_req_op *op = osd_req_op_init(osd_req, which, opcode); 498 struct ceph_pagelist *pagelist; 499 size_t payload_len = 0; 500 size_t size; 501 502 BUG_ON(opcode != CEPH_OSD_OP_CALL); 503 504 pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS); 505 BUG_ON(!pagelist); 506 ceph_pagelist_init(pagelist); 507 508 op->cls.class_name = class; 509 size = strlen(class); 510 BUG_ON(size > (size_t) U8_MAX); 511 op->cls.class_len = size; 512 ceph_pagelist_append(pagelist, class, size); 513 payload_len += size; 514 515 op->cls.method_name = method; 516 size = strlen(method); 517 BUG_ON(size > (size_t) U8_MAX); 518 op->cls.method_len = size; 519 ceph_pagelist_append(pagelist, method, size); 520 payload_len += size; 521 522 op->cls.request_data = request_data; 523 BUG_ON(request_data_size > (size_t) U32_MAX); 524 op->cls.request_data_len = (u32) request_data_size; 525 ceph_pagelist_append(pagelist, request_data, request_data_size); 526 payload_len += request_data_size; 527 528 osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist); 529 530 op->cls.argc = 0; /* currently unused */ 531 532 op->payload_len = payload_len; 533 } 534 EXPORT_SYMBOL(osd_req_op_cls_init); 535 536 void osd_req_op_watch_init(struct ceph_osd_request *osd_req, 537 unsigned int which, u16 opcode, 538 u64 cookie, u64 version, int flag) 539 { 540 struct ceph_osd_req_op *op = osd_req_op_init(osd_req, which, opcode); 541 542 BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH); 543 544 op->watch.cookie = cookie; 545 /* op->watch.ver = version; */ /* XXX 3847 */ 546 op->watch.ver = cpu_to_le64(version); 547 if (opcode == CEPH_OSD_OP_WATCH && flag) 548 op->watch.flag = (u8)1; 549 } 550 EXPORT_SYMBOL(osd_req_op_watch_init); 551 552 static u64 osd_req_encode_op(struct ceph_osd_request *req, 553 struct ceph_osd_op *dst, unsigned int which) 554 { 555 struct ceph_osd_req_op *src; 556 u64 request_data_len = 0; 557 558 BUG_ON(which >= req->r_num_ops); 559 src = &req->r_ops[which]; 560 if (WARN_ON(!osd_req_opcode_valid(src->op))) { 561 pr_err("unrecognized osd opcode %d\n", src->op); 562 563 return 0; 564 } 565 566 switch (src->op) { 567 case CEPH_OSD_OP_STAT: 568 break; 569 case CEPH_OSD_OP_READ: 570 case CEPH_OSD_OP_WRITE: 571 if (src->op == CEPH_OSD_OP_WRITE) 572 request_data_len = src->extent.length; 573 dst->extent.offset = cpu_to_le64(src->extent.offset); 574 dst->extent.length = cpu_to_le64(src->extent.length); 575 dst->extent.truncate_size = 576 cpu_to_le64(src->extent.truncate_size); 577 dst->extent.truncate_seq = 578 cpu_to_le32(src->extent.truncate_seq); 579 if (src->op == CEPH_OSD_OP_WRITE) 580 WARN_ON(src->extent.osd_data != &req->r_data_out); 581 else 582 WARN_ON(src->extent.osd_data != &req->r_data_in); 583 break; 584 case CEPH_OSD_OP_CALL: 585 dst->cls.class_len = src->cls.class_len; 586 dst->cls.method_len = src->cls.method_len; 587 dst->cls.indata_len = cpu_to_le32(src->cls.request_data_len); 588 WARN_ON(src->cls.response_data != &req->r_data_in); 589 WARN_ON(src->cls.request_info != &req->r_data_out); 590 BUG_ON(src->cls.request_info->type != 591 CEPH_OSD_DATA_TYPE_PAGELIST); 592 request_data_len = src->cls.request_info->pagelist->length; 593 break; 594 case CEPH_OSD_OP_STARTSYNC: 595 break; 596 case CEPH_OSD_OP_NOTIFY_ACK: 597 case CEPH_OSD_OP_WATCH: 598 dst->watch.cookie = cpu_to_le64(src->watch.cookie); 599 dst->watch.ver = cpu_to_le64(src->watch.ver); 600 dst->watch.flag = src->watch.flag; 601 break; 602 default: 603 pr_err("unsupported osd opcode %s\n", 604 ceph_osd_op_name(src->op)); 605 WARN_ON(1); 606 607 return 0; 608 } 609 dst->op = cpu_to_le16(src->op); 610 dst->payload_len = cpu_to_le32(src->payload_len); 611 612 return request_data_len; 613 } 614 615 /* 616 * build new request AND message, calculate layout, and adjust file 617 * extent as needed. 618 * 619 * if the file was recently truncated, we include information about its 620 * old and new size so that the object can be updated appropriately. (we 621 * avoid synchronously deleting truncated objects because it's slow.) 622 * 623 * if @do_sync, include a 'startsync' command so that the osd will flush 624 * data quickly. 625 */ 626 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc, 627 struct ceph_file_layout *layout, 628 struct ceph_vino vino, 629 u64 off, u64 *plen, int num_ops, 630 int opcode, int flags, 631 struct ceph_snap_context *snapc, 632 u32 truncate_seq, 633 u64 truncate_size, 634 bool use_mempool) 635 { 636 struct ceph_osd_request *req; 637 u64 objnum = 0; 638 u64 objoff = 0; 639 u64 objlen = 0; 640 u32 object_size; 641 u64 object_base; 642 int r; 643 644 BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE); 645 646 req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool, 647 GFP_NOFS); 648 if (!req) 649 return ERR_PTR(-ENOMEM); 650 651 req->r_flags = flags; 652 653 /* calculate max write size */ 654 r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen); 655 if (r < 0) { 656 ceph_osdc_put_request(req); 657 return ERR_PTR(r); 658 } 659 660 object_size = le32_to_cpu(layout->fl_object_size); 661 object_base = off - objoff; 662 if (truncate_size <= object_base) { 663 truncate_size = 0; 664 } else { 665 truncate_size -= object_base; 666 if (truncate_size > object_size) 667 truncate_size = object_size; 668 } 669 670 osd_req_op_extent_init(req, 0, opcode, objoff, objlen, 671 truncate_size, truncate_seq); 672 673 /* 674 * A second op in the ops array means the caller wants to 675 * also issue a include a 'startsync' command so that the 676 * osd will flush data quickly. 677 */ 678 if (num_ops > 1) 679 osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC); 680 681 req->r_file_layout = *layout; /* keep a copy */ 682 683 snprintf(req->r_oid, sizeof(req->r_oid), "%llx.%08llx", 684 vino.ino, objnum); 685 req->r_oid_len = strlen(req->r_oid); 686 687 return req; 688 } 689 EXPORT_SYMBOL(ceph_osdc_new_request); 690 691 /* 692 * We keep osd requests in an rbtree, sorted by ->r_tid. 693 */ 694 static void __insert_request(struct ceph_osd_client *osdc, 695 struct ceph_osd_request *new) 696 { 697 struct rb_node **p = &osdc->requests.rb_node; 698 struct rb_node *parent = NULL; 699 struct ceph_osd_request *req = NULL; 700 701 while (*p) { 702 parent = *p; 703 req = rb_entry(parent, struct ceph_osd_request, r_node); 704 if (new->r_tid < req->r_tid) 705 p = &(*p)->rb_left; 706 else if (new->r_tid > req->r_tid) 707 p = &(*p)->rb_right; 708 else 709 BUG(); 710 } 711 712 rb_link_node(&new->r_node, parent, p); 713 rb_insert_color(&new->r_node, &osdc->requests); 714 } 715 716 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc, 717 u64 tid) 718 { 719 struct ceph_osd_request *req; 720 struct rb_node *n = osdc->requests.rb_node; 721 722 while (n) { 723 req = rb_entry(n, struct ceph_osd_request, r_node); 724 if (tid < req->r_tid) 725 n = n->rb_left; 726 else if (tid > req->r_tid) 727 n = n->rb_right; 728 else 729 return req; 730 } 731 return NULL; 732 } 733 734 static struct ceph_osd_request * 735 __lookup_request_ge(struct ceph_osd_client *osdc, 736 u64 tid) 737 { 738 struct ceph_osd_request *req; 739 struct rb_node *n = osdc->requests.rb_node; 740 741 while (n) { 742 req = rb_entry(n, struct ceph_osd_request, r_node); 743 if (tid < req->r_tid) { 744 if (!n->rb_left) 745 return req; 746 n = n->rb_left; 747 } else if (tid > req->r_tid) { 748 n = n->rb_right; 749 } else { 750 return req; 751 } 752 } 753 return NULL; 754 } 755 756 /* 757 * Resubmit requests pending on the given osd. 758 */ 759 static void __kick_osd_requests(struct ceph_osd_client *osdc, 760 struct ceph_osd *osd) 761 { 762 struct ceph_osd_request *req, *nreq; 763 LIST_HEAD(resend); 764 int err; 765 766 dout("__kick_osd_requests osd%d\n", osd->o_osd); 767 err = __reset_osd(osdc, osd); 768 if (err) 769 return; 770 /* 771 * Build up a list of requests to resend by traversing the 772 * osd's list of requests. Requests for a given object are 773 * sent in tid order, and that is also the order they're 774 * kept on this list. Therefore all requests that are in 775 * flight will be found first, followed by all requests that 776 * have not yet been sent. And to resend requests while 777 * preserving this order we will want to put any sent 778 * requests back on the front of the osd client's unsent 779 * list. 780 * 781 * So we build a separate ordered list of already-sent 782 * requests for the affected osd and splice it onto the 783 * front of the osd client's unsent list. Once we've seen a 784 * request that has not yet been sent we're done. Those 785 * requests are already sitting right where they belong. 786 */ 787 list_for_each_entry(req, &osd->o_requests, r_osd_item) { 788 if (!req->r_sent) 789 break; 790 list_move_tail(&req->r_req_lru_item, &resend); 791 dout("requeueing %p tid %llu osd%d\n", req, req->r_tid, 792 osd->o_osd); 793 if (!req->r_linger) 794 req->r_flags |= CEPH_OSD_FLAG_RETRY; 795 } 796 list_splice(&resend, &osdc->req_unsent); 797 798 /* 799 * Linger requests are re-registered before sending, which 800 * sets up a new tid for each. We add them to the unsent 801 * list at the end to keep things in tid order. 802 */ 803 list_for_each_entry_safe(req, nreq, &osd->o_linger_requests, 804 r_linger_osd) { 805 /* 806 * reregister request prior to unregistering linger so 807 * that r_osd is preserved. 808 */ 809 BUG_ON(!list_empty(&req->r_req_lru_item)); 810 __register_request(osdc, req); 811 list_add_tail(&req->r_req_lru_item, &osdc->req_unsent); 812 list_add_tail(&req->r_osd_item, &req->r_osd->o_requests); 813 __unregister_linger_request(osdc, req); 814 dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid, 815 osd->o_osd); 816 } 817 } 818 819 /* 820 * If the osd connection drops, we need to resubmit all requests. 821 */ 822 static void osd_reset(struct ceph_connection *con) 823 { 824 struct ceph_osd *osd = con->private; 825 struct ceph_osd_client *osdc; 826 827 if (!osd) 828 return; 829 dout("osd_reset osd%d\n", osd->o_osd); 830 osdc = osd->o_osdc; 831 down_read(&osdc->map_sem); 832 mutex_lock(&osdc->request_mutex); 833 __kick_osd_requests(osdc, osd); 834 __send_queued(osdc); 835 mutex_unlock(&osdc->request_mutex); 836 up_read(&osdc->map_sem); 837 } 838 839 /* 840 * Track open sessions with osds. 841 */ 842 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum) 843 { 844 struct ceph_osd *osd; 845 846 osd = kzalloc(sizeof(*osd), GFP_NOFS); 847 if (!osd) 848 return NULL; 849 850 atomic_set(&osd->o_ref, 1); 851 osd->o_osdc = osdc; 852 osd->o_osd = onum; 853 RB_CLEAR_NODE(&osd->o_node); 854 INIT_LIST_HEAD(&osd->o_requests); 855 INIT_LIST_HEAD(&osd->o_linger_requests); 856 INIT_LIST_HEAD(&osd->o_osd_lru); 857 osd->o_incarnation = 1; 858 859 ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr); 860 861 INIT_LIST_HEAD(&osd->o_keepalive_item); 862 return osd; 863 } 864 865 static struct ceph_osd *get_osd(struct ceph_osd *osd) 866 { 867 if (atomic_inc_not_zero(&osd->o_ref)) { 868 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1, 869 atomic_read(&osd->o_ref)); 870 return osd; 871 } else { 872 dout("get_osd %p FAIL\n", osd); 873 return NULL; 874 } 875 } 876 877 static void put_osd(struct ceph_osd *osd) 878 { 879 dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref), 880 atomic_read(&osd->o_ref) - 1); 881 if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) { 882 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth; 883 884 ceph_auth_destroy_authorizer(ac, osd->o_auth.authorizer); 885 kfree(osd); 886 } 887 } 888 889 /* 890 * remove an osd from our map 891 */ 892 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 893 { 894 dout("__remove_osd %p\n", osd); 895 BUG_ON(!list_empty(&osd->o_requests)); 896 rb_erase(&osd->o_node, &osdc->osds); 897 list_del_init(&osd->o_osd_lru); 898 ceph_con_close(&osd->o_con); 899 put_osd(osd); 900 } 901 902 static void remove_all_osds(struct ceph_osd_client *osdc) 903 { 904 dout("%s %p\n", __func__, osdc); 905 mutex_lock(&osdc->request_mutex); 906 while (!RB_EMPTY_ROOT(&osdc->osds)) { 907 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds), 908 struct ceph_osd, o_node); 909 __remove_osd(osdc, osd); 910 } 911 mutex_unlock(&osdc->request_mutex); 912 } 913 914 static void __move_osd_to_lru(struct ceph_osd_client *osdc, 915 struct ceph_osd *osd) 916 { 917 dout("__move_osd_to_lru %p\n", osd); 918 BUG_ON(!list_empty(&osd->o_osd_lru)); 919 list_add_tail(&osd->o_osd_lru, &osdc->osd_lru); 920 osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ; 921 } 922 923 static void __remove_osd_from_lru(struct ceph_osd *osd) 924 { 925 dout("__remove_osd_from_lru %p\n", osd); 926 if (!list_empty(&osd->o_osd_lru)) 927 list_del_init(&osd->o_osd_lru); 928 } 929 930 static void remove_old_osds(struct ceph_osd_client *osdc) 931 { 932 struct ceph_osd *osd, *nosd; 933 934 dout("__remove_old_osds %p\n", osdc); 935 mutex_lock(&osdc->request_mutex); 936 list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) { 937 if (time_before(jiffies, osd->lru_ttl)) 938 break; 939 __remove_osd(osdc, osd); 940 } 941 mutex_unlock(&osdc->request_mutex); 942 } 943 944 /* 945 * reset osd connect 946 */ 947 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 948 { 949 struct ceph_entity_addr *peer_addr; 950 951 dout("__reset_osd %p osd%d\n", osd, osd->o_osd); 952 if (list_empty(&osd->o_requests) && 953 list_empty(&osd->o_linger_requests)) { 954 __remove_osd(osdc, osd); 955 956 return -ENODEV; 957 } 958 959 peer_addr = &osdc->osdmap->osd_addr[osd->o_osd]; 960 if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) && 961 !ceph_con_opened(&osd->o_con)) { 962 struct ceph_osd_request *req; 963 964 dout(" osd addr hasn't changed and connection never opened," 965 " letting msgr retry"); 966 /* touch each r_stamp for handle_timeout()'s benfit */ 967 list_for_each_entry(req, &osd->o_requests, r_osd_item) 968 req->r_stamp = jiffies; 969 970 return -EAGAIN; 971 } 972 973 ceph_con_close(&osd->o_con); 974 ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr); 975 osd->o_incarnation++; 976 977 return 0; 978 } 979 980 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new) 981 { 982 struct rb_node **p = &osdc->osds.rb_node; 983 struct rb_node *parent = NULL; 984 struct ceph_osd *osd = NULL; 985 986 dout("__insert_osd %p osd%d\n", new, new->o_osd); 987 while (*p) { 988 parent = *p; 989 osd = rb_entry(parent, struct ceph_osd, o_node); 990 if (new->o_osd < osd->o_osd) 991 p = &(*p)->rb_left; 992 else if (new->o_osd > osd->o_osd) 993 p = &(*p)->rb_right; 994 else 995 BUG(); 996 } 997 998 rb_link_node(&new->o_node, parent, p); 999 rb_insert_color(&new->o_node, &osdc->osds); 1000 } 1001 1002 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o) 1003 { 1004 struct ceph_osd *osd; 1005 struct rb_node *n = osdc->osds.rb_node; 1006 1007 while (n) { 1008 osd = rb_entry(n, struct ceph_osd, o_node); 1009 if (o < osd->o_osd) 1010 n = n->rb_left; 1011 else if (o > osd->o_osd) 1012 n = n->rb_right; 1013 else 1014 return osd; 1015 } 1016 return NULL; 1017 } 1018 1019 static void __schedule_osd_timeout(struct ceph_osd_client *osdc) 1020 { 1021 schedule_delayed_work(&osdc->timeout_work, 1022 osdc->client->options->osd_keepalive_timeout * HZ); 1023 } 1024 1025 static void __cancel_osd_timeout(struct ceph_osd_client *osdc) 1026 { 1027 cancel_delayed_work(&osdc->timeout_work); 1028 } 1029 1030 /* 1031 * Register request, assign tid. If this is the first request, set up 1032 * the timeout event. 1033 */ 1034 static void __register_request(struct ceph_osd_client *osdc, 1035 struct ceph_osd_request *req) 1036 { 1037 req->r_tid = ++osdc->last_tid; 1038 req->r_request->hdr.tid = cpu_to_le64(req->r_tid); 1039 dout("__register_request %p tid %lld\n", req, req->r_tid); 1040 __insert_request(osdc, req); 1041 ceph_osdc_get_request(req); 1042 osdc->num_requests++; 1043 if (osdc->num_requests == 1) { 1044 dout(" first request, scheduling timeout\n"); 1045 __schedule_osd_timeout(osdc); 1046 } 1047 } 1048 1049 /* 1050 * called under osdc->request_mutex 1051 */ 1052 static void __unregister_request(struct ceph_osd_client *osdc, 1053 struct ceph_osd_request *req) 1054 { 1055 if (RB_EMPTY_NODE(&req->r_node)) { 1056 dout("__unregister_request %p tid %lld not registered\n", 1057 req, req->r_tid); 1058 return; 1059 } 1060 1061 dout("__unregister_request %p tid %lld\n", req, req->r_tid); 1062 rb_erase(&req->r_node, &osdc->requests); 1063 osdc->num_requests--; 1064 1065 if (req->r_osd) { 1066 /* make sure the original request isn't in flight. */ 1067 ceph_msg_revoke(req->r_request); 1068 1069 list_del_init(&req->r_osd_item); 1070 if (list_empty(&req->r_osd->o_requests) && 1071 list_empty(&req->r_osd->o_linger_requests)) { 1072 dout("moving osd to %p lru\n", req->r_osd); 1073 __move_osd_to_lru(osdc, req->r_osd); 1074 } 1075 if (list_empty(&req->r_linger_item)) 1076 req->r_osd = NULL; 1077 } 1078 1079 list_del_init(&req->r_req_lru_item); 1080 ceph_osdc_put_request(req); 1081 1082 if (osdc->num_requests == 0) { 1083 dout(" no requests, canceling timeout\n"); 1084 __cancel_osd_timeout(osdc); 1085 } 1086 } 1087 1088 /* 1089 * Cancel a previously queued request message 1090 */ 1091 static void __cancel_request(struct ceph_osd_request *req) 1092 { 1093 if (req->r_sent && req->r_osd) { 1094 ceph_msg_revoke(req->r_request); 1095 req->r_sent = 0; 1096 } 1097 } 1098 1099 static void __register_linger_request(struct ceph_osd_client *osdc, 1100 struct ceph_osd_request *req) 1101 { 1102 dout("__register_linger_request %p\n", req); 1103 list_add_tail(&req->r_linger_item, &osdc->req_linger); 1104 if (req->r_osd) 1105 list_add_tail(&req->r_linger_osd, 1106 &req->r_osd->o_linger_requests); 1107 } 1108 1109 static void __unregister_linger_request(struct ceph_osd_client *osdc, 1110 struct ceph_osd_request *req) 1111 { 1112 dout("__unregister_linger_request %p\n", req); 1113 list_del_init(&req->r_linger_item); 1114 if (req->r_osd) { 1115 list_del_init(&req->r_linger_osd); 1116 1117 if (list_empty(&req->r_osd->o_requests) && 1118 list_empty(&req->r_osd->o_linger_requests)) { 1119 dout("moving osd to %p lru\n", req->r_osd); 1120 __move_osd_to_lru(osdc, req->r_osd); 1121 } 1122 if (list_empty(&req->r_osd_item)) 1123 req->r_osd = NULL; 1124 } 1125 } 1126 1127 void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc, 1128 struct ceph_osd_request *req) 1129 { 1130 mutex_lock(&osdc->request_mutex); 1131 if (req->r_linger) { 1132 __unregister_linger_request(osdc, req); 1133 ceph_osdc_put_request(req); 1134 } 1135 mutex_unlock(&osdc->request_mutex); 1136 } 1137 EXPORT_SYMBOL(ceph_osdc_unregister_linger_request); 1138 1139 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc, 1140 struct ceph_osd_request *req) 1141 { 1142 if (!req->r_linger) { 1143 dout("set_request_linger %p\n", req); 1144 req->r_linger = 1; 1145 /* 1146 * caller is now responsible for calling 1147 * unregister_linger_request 1148 */ 1149 ceph_osdc_get_request(req); 1150 } 1151 } 1152 EXPORT_SYMBOL(ceph_osdc_set_request_linger); 1153 1154 /* 1155 * Pick an osd (the first 'up' osd in the pg), allocate the osd struct 1156 * (as needed), and set the request r_osd appropriately. If there is 1157 * no up osd, set r_osd to NULL. Move the request to the appropriate list 1158 * (unsent, homeless) or leave on in-flight lru. 1159 * 1160 * Return 0 if unchanged, 1 if changed, or negative on error. 1161 * 1162 * Caller should hold map_sem for read and request_mutex. 1163 */ 1164 static int __map_request(struct ceph_osd_client *osdc, 1165 struct ceph_osd_request *req, int force_resend) 1166 { 1167 struct ceph_pg pgid; 1168 int acting[CEPH_PG_MAX_SIZE]; 1169 int o = -1, num = 0; 1170 int err; 1171 1172 dout("map_request %p tid %lld\n", req, req->r_tid); 1173 err = ceph_calc_ceph_pg(&pgid, req->r_oid, osdc->osdmap, 1174 ceph_file_layout_pg_pool(req->r_file_layout)); 1175 if (err) { 1176 list_move(&req->r_req_lru_item, &osdc->req_notarget); 1177 return err; 1178 } 1179 req->r_pgid = pgid; 1180 1181 err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting); 1182 if (err > 0) { 1183 o = acting[0]; 1184 num = err; 1185 } 1186 1187 if ((!force_resend && 1188 req->r_osd && req->r_osd->o_osd == o && 1189 req->r_sent >= req->r_osd->o_incarnation && 1190 req->r_num_pg_osds == num && 1191 memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) || 1192 (req->r_osd == NULL && o == -1)) 1193 return 0; /* no change */ 1194 1195 dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n", 1196 req->r_tid, pgid.pool, pgid.seed, o, 1197 req->r_osd ? req->r_osd->o_osd : -1); 1198 1199 /* record full pg acting set */ 1200 memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num); 1201 req->r_num_pg_osds = num; 1202 1203 if (req->r_osd) { 1204 __cancel_request(req); 1205 list_del_init(&req->r_osd_item); 1206 req->r_osd = NULL; 1207 } 1208 1209 req->r_osd = __lookup_osd(osdc, o); 1210 if (!req->r_osd && o >= 0) { 1211 err = -ENOMEM; 1212 req->r_osd = create_osd(osdc, o); 1213 if (!req->r_osd) { 1214 list_move(&req->r_req_lru_item, &osdc->req_notarget); 1215 goto out; 1216 } 1217 1218 dout("map_request osd %p is osd%d\n", req->r_osd, o); 1219 __insert_osd(osdc, req->r_osd); 1220 1221 ceph_con_open(&req->r_osd->o_con, 1222 CEPH_ENTITY_TYPE_OSD, o, 1223 &osdc->osdmap->osd_addr[o]); 1224 } 1225 1226 if (req->r_osd) { 1227 __remove_osd_from_lru(req->r_osd); 1228 list_add_tail(&req->r_osd_item, &req->r_osd->o_requests); 1229 list_move_tail(&req->r_req_lru_item, &osdc->req_unsent); 1230 } else { 1231 list_move_tail(&req->r_req_lru_item, &osdc->req_notarget); 1232 } 1233 err = 1; /* osd or pg changed */ 1234 1235 out: 1236 return err; 1237 } 1238 1239 /* 1240 * caller should hold map_sem (for read) and request_mutex 1241 */ 1242 static void __send_request(struct ceph_osd_client *osdc, 1243 struct ceph_osd_request *req) 1244 { 1245 void *p; 1246 1247 dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n", 1248 req, req->r_tid, req->r_osd->o_osd, req->r_flags, 1249 (unsigned long long)req->r_pgid.pool, req->r_pgid.seed); 1250 1251 /* fill in message content that changes each time we send it */ 1252 put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch); 1253 put_unaligned_le32(req->r_flags, req->r_request_flags); 1254 put_unaligned_le64(req->r_pgid.pool, req->r_request_pool); 1255 p = req->r_request_pgid; 1256 ceph_encode_64(&p, req->r_pgid.pool); 1257 ceph_encode_32(&p, req->r_pgid.seed); 1258 put_unaligned_le64(1, req->r_request_attempts); /* FIXME */ 1259 memcpy(req->r_request_reassert_version, &req->r_reassert_version, 1260 sizeof(req->r_reassert_version)); 1261 1262 req->r_stamp = jiffies; 1263 list_move_tail(&req->r_req_lru_item, &osdc->req_lru); 1264 1265 ceph_msg_get(req->r_request); /* send consumes a ref */ 1266 ceph_con_send(&req->r_osd->o_con, req->r_request); 1267 req->r_sent = req->r_osd->o_incarnation; 1268 } 1269 1270 /* 1271 * Send any requests in the queue (req_unsent). 1272 */ 1273 static void __send_queued(struct ceph_osd_client *osdc) 1274 { 1275 struct ceph_osd_request *req, *tmp; 1276 1277 dout("__send_queued\n"); 1278 list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) 1279 __send_request(osdc, req); 1280 } 1281 1282 /* 1283 * Timeout callback, called every N seconds when 1 or more osd 1284 * requests has been active for more than N seconds. When this 1285 * happens, we ping all OSDs with requests who have timed out to 1286 * ensure any communications channel reset is detected. Reset the 1287 * request timeouts another N seconds in the future as we go. 1288 * Reschedule the timeout event another N seconds in future (unless 1289 * there are no open requests). 1290 */ 1291 static void handle_timeout(struct work_struct *work) 1292 { 1293 struct ceph_osd_client *osdc = 1294 container_of(work, struct ceph_osd_client, timeout_work.work); 1295 struct ceph_osd_request *req; 1296 struct ceph_osd *osd; 1297 unsigned long keepalive = 1298 osdc->client->options->osd_keepalive_timeout * HZ; 1299 struct list_head slow_osds; 1300 dout("timeout\n"); 1301 down_read(&osdc->map_sem); 1302 1303 ceph_monc_request_next_osdmap(&osdc->client->monc); 1304 1305 mutex_lock(&osdc->request_mutex); 1306 1307 /* 1308 * ping osds that are a bit slow. this ensures that if there 1309 * is a break in the TCP connection we will notice, and reopen 1310 * a connection with that osd (from the fault callback). 1311 */ 1312 INIT_LIST_HEAD(&slow_osds); 1313 list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) { 1314 if (time_before(jiffies, req->r_stamp + keepalive)) 1315 break; 1316 1317 osd = req->r_osd; 1318 BUG_ON(!osd); 1319 dout(" tid %llu is slow, will send keepalive on osd%d\n", 1320 req->r_tid, osd->o_osd); 1321 list_move_tail(&osd->o_keepalive_item, &slow_osds); 1322 } 1323 while (!list_empty(&slow_osds)) { 1324 osd = list_entry(slow_osds.next, struct ceph_osd, 1325 o_keepalive_item); 1326 list_del_init(&osd->o_keepalive_item); 1327 ceph_con_keepalive(&osd->o_con); 1328 } 1329 1330 __schedule_osd_timeout(osdc); 1331 __send_queued(osdc); 1332 mutex_unlock(&osdc->request_mutex); 1333 up_read(&osdc->map_sem); 1334 } 1335 1336 static void handle_osds_timeout(struct work_struct *work) 1337 { 1338 struct ceph_osd_client *osdc = 1339 container_of(work, struct ceph_osd_client, 1340 osds_timeout_work.work); 1341 unsigned long delay = 1342 osdc->client->options->osd_idle_ttl * HZ >> 2; 1343 1344 dout("osds timeout\n"); 1345 down_read(&osdc->map_sem); 1346 remove_old_osds(osdc); 1347 up_read(&osdc->map_sem); 1348 1349 schedule_delayed_work(&osdc->osds_timeout_work, 1350 round_jiffies_relative(delay)); 1351 } 1352 1353 static void complete_request(struct ceph_osd_request *req) 1354 { 1355 if (req->r_safe_callback) 1356 req->r_safe_callback(req, NULL); 1357 complete_all(&req->r_safe_completion); /* fsync waiter */ 1358 } 1359 1360 /* 1361 * handle osd op reply. either call the callback if it is specified, 1362 * or do the completion to wake up the waiting thread. 1363 */ 1364 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg, 1365 struct ceph_connection *con) 1366 { 1367 void *p, *end; 1368 struct ceph_osd_request *req; 1369 u64 tid; 1370 int object_len; 1371 unsigned int numops; 1372 int payload_len, flags; 1373 s32 result; 1374 s32 retry_attempt; 1375 struct ceph_pg pg; 1376 int err; 1377 u32 reassert_epoch; 1378 u64 reassert_version; 1379 u32 osdmap_epoch; 1380 int already_completed; 1381 u32 bytes; 1382 unsigned int i; 1383 1384 tid = le64_to_cpu(msg->hdr.tid); 1385 dout("handle_reply %p tid %llu\n", msg, tid); 1386 1387 p = msg->front.iov_base; 1388 end = p + msg->front.iov_len; 1389 1390 ceph_decode_need(&p, end, 4, bad); 1391 object_len = ceph_decode_32(&p); 1392 ceph_decode_need(&p, end, object_len, bad); 1393 p += object_len; 1394 1395 err = ceph_decode_pgid(&p, end, &pg); 1396 if (err) 1397 goto bad; 1398 1399 ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad); 1400 flags = ceph_decode_64(&p); 1401 result = ceph_decode_32(&p); 1402 reassert_epoch = ceph_decode_32(&p); 1403 reassert_version = ceph_decode_64(&p); 1404 osdmap_epoch = ceph_decode_32(&p); 1405 1406 /* lookup */ 1407 mutex_lock(&osdc->request_mutex); 1408 req = __lookup_request(osdc, tid); 1409 if (req == NULL) { 1410 dout("handle_reply tid %llu dne\n", tid); 1411 goto bad_mutex; 1412 } 1413 ceph_osdc_get_request(req); 1414 1415 dout("handle_reply %p tid %llu req %p result %d\n", msg, tid, 1416 req, result); 1417 1418 ceph_decode_need(&p, end, 4, bad); 1419 numops = ceph_decode_32(&p); 1420 if (numops > CEPH_OSD_MAX_OP) 1421 goto bad_put; 1422 if (numops != req->r_num_ops) 1423 goto bad_put; 1424 payload_len = 0; 1425 ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad); 1426 for (i = 0; i < numops; i++) { 1427 struct ceph_osd_op *op = p; 1428 int len; 1429 1430 len = le32_to_cpu(op->payload_len); 1431 req->r_reply_op_len[i] = len; 1432 dout(" op %d has %d bytes\n", i, len); 1433 payload_len += len; 1434 p += sizeof(*op); 1435 } 1436 bytes = le32_to_cpu(msg->hdr.data_len); 1437 if (payload_len != bytes) { 1438 pr_warning("sum of op payload lens %d != data_len %d", 1439 payload_len, bytes); 1440 goto bad_put; 1441 } 1442 1443 ceph_decode_need(&p, end, 4 + numops * 4, bad); 1444 retry_attempt = ceph_decode_32(&p); 1445 for (i = 0; i < numops; i++) 1446 req->r_reply_op_result[i] = ceph_decode_32(&p); 1447 1448 if (!req->r_got_reply) { 1449 1450 req->r_result = result; 1451 dout("handle_reply result %d bytes %d\n", req->r_result, 1452 bytes); 1453 if (req->r_result == 0) 1454 req->r_result = bytes; 1455 1456 /* in case this is a write and we need to replay, */ 1457 req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch); 1458 req->r_reassert_version.version = cpu_to_le64(reassert_version); 1459 1460 req->r_got_reply = 1; 1461 } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) { 1462 dout("handle_reply tid %llu dup ack\n", tid); 1463 mutex_unlock(&osdc->request_mutex); 1464 goto done; 1465 } 1466 1467 dout("handle_reply tid %llu flags %d\n", tid, flags); 1468 1469 if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK)) 1470 __register_linger_request(osdc, req); 1471 1472 /* either this is a read, or we got the safe response */ 1473 if (result < 0 || 1474 (flags & CEPH_OSD_FLAG_ONDISK) || 1475 ((flags & CEPH_OSD_FLAG_WRITE) == 0)) 1476 __unregister_request(osdc, req); 1477 1478 already_completed = req->r_completed; 1479 req->r_completed = 1; 1480 mutex_unlock(&osdc->request_mutex); 1481 if (already_completed) 1482 goto done; 1483 1484 if (req->r_callback) 1485 req->r_callback(req, msg); 1486 else 1487 complete_all(&req->r_completion); 1488 1489 if (flags & CEPH_OSD_FLAG_ONDISK) 1490 complete_request(req); 1491 1492 done: 1493 dout("req=%p req->r_linger=%d\n", req, req->r_linger); 1494 ceph_osdc_put_request(req); 1495 return; 1496 1497 bad_put: 1498 ceph_osdc_put_request(req); 1499 bad_mutex: 1500 mutex_unlock(&osdc->request_mutex); 1501 bad: 1502 pr_err("corrupt osd_op_reply got %d %d\n", 1503 (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len)); 1504 ceph_msg_dump(msg); 1505 } 1506 1507 static void reset_changed_osds(struct ceph_osd_client *osdc) 1508 { 1509 struct rb_node *p, *n; 1510 1511 for (p = rb_first(&osdc->osds); p; p = n) { 1512 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node); 1513 1514 n = rb_next(p); 1515 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) || 1516 memcmp(&osd->o_con.peer_addr, 1517 ceph_osd_addr(osdc->osdmap, 1518 osd->o_osd), 1519 sizeof(struct ceph_entity_addr)) != 0) 1520 __reset_osd(osdc, osd); 1521 } 1522 } 1523 1524 /* 1525 * Requeue requests whose mapping to an OSD has changed. If requests map to 1526 * no osd, request a new map. 1527 * 1528 * Caller should hold map_sem for read. 1529 */ 1530 static void kick_requests(struct ceph_osd_client *osdc, int force_resend) 1531 { 1532 struct ceph_osd_request *req, *nreq; 1533 struct rb_node *p; 1534 int needmap = 0; 1535 int err; 1536 1537 dout("kick_requests %s\n", force_resend ? " (force resend)" : ""); 1538 mutex_lock(&osdc->request_mutex); 1539 for (p = rb_first(&osdc->requests); p; ) { 1540 req = rb_entry(p, struct ceph_osd_request, r_node); 1541 p = rb_next(p); 1542 1543 /* 1544 * For linger requests that have not yet been 1545 * registered, move them to the linger list; they'll 1546 * be sent to the osd in the loop below. Unregister 1547 * the request before re-registering it as a linger 1548 * request to ensure the __map_request() below 1549 * will decide it needs to be sent. 1550 */ 1551 if (req->r_linger && list_empty(&req->r_linger_item)) { 1552 dout("%p tid %llu restart on osd%d\n", 1553 req, req->r_tid, 1554 req->r_osd ? req->r_osd->o_osd : -1); 1555 __unregister_request(osdc, req); 1556 __register_linger_request(osdc, req); 1557 continue; 1558 } 1559 1560 err = __map_request(osdc, req, force_resend); 1561 if (err < 0) 1562 continue; /* error */ 1563 if (req->r_osd == NULL) { 1564 dout("%p tid %llu maps to no osd\n", req, req->r_tid); 1565 needmap++; /* request a newer map */ 1566 } else if (err > 0) { 1567 if (!req->r_linger) { 1568 dout("%p tid %llu requeued on osd%d\n", req, 1569 req->r_tid, 1570 req->r_osd ? req->r_osd->o_osd : -1); 1571 req->r_flags |= CEPH_OSD_FLAG_RETRY; 1572 } 1573 } 1574 } 1575 1576 list_for_each_entry_safe(req, nreq, &osdc->req_linger, 1577 r_linger_item) { 1578 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd); 1579 1580 err = __map_request(osdc, req, force_resend); 1581 dout("__map_request returned %d\n", err); 1582 if (err == 0) 1583 continue; /* no change and no osd was specified */ 1584 if (err < 0) 1585 continue; /* hrm! */ 1586 if (req->r_osd == NULL) { 1587 dout("tid %llu maps to no valid osd\n", req->r_tid); 1588 needmap++; /* request a newer map */ 1589 continue; 1590 } 1591 1592 dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid, 1593 req->r_osd ? req->r_osd->o_osd : -1); 1594 __register_request(osdc, req); 1595 __unregister_linger_request(osdc, req); 1596 } 1597 mutex_unlock(&osdc->request_mutex); 1598 1599 if (needmap) { 1600 dout("%d requests for down osds, need new map\n", needmap); 1601 ceph_monc_request_next_osdmap(&osdc->client->monc); 1602 } 1603 reset_changed_osds(osdc); 1604 } 1605 1606 1607 /* 1608 * Process updated osd map. 1609 * 1610 * The message contains any number of incremental and full maps, normally 1611 * indicating some sort of topology change in the cluster. Kick requests 1612 * off to different OSDs as needed. 1613 */ 1614 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg) 1615 { 1616 void *p, *end, *next; 1617 u32 nr_maps, maplen; 1618 u32 epoch; 1619 struct ceph_osdmap *newmap = NULL, *oldmap; 1620 int err; 1621 struct ceph_fsid fsid; 1622 1623 dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0); 1624 p = msg->front.iov_base; 1625 end = p + msg->front.iov_len; 1626 1627 /* verify fsid */ 1628 ceph_decode_need(&p, end, sizeof(fsid), bad); 1629 ceph_decode_copy(&p, &fsid, sizeof(fsid)); 1630 if (ceph_check_fsid(osdc->client, &fsid) < 0) 1631 return; 1632 1633 down_write(&osdc->map_sem); 1634 1635 /* incremental maps */ 1636 ceph_decode_32_safe(&p, end, nr_maps, bad); 1637 dout(" %d inc maps\n", nr_maps); 1638 while (nr_maps > 0) { 1639 ceph_decode_need(&p, end, 2*sizeof(u32), bad); 1640 epoch = ceph_decode_32(&p); 1641 maplen = ceph_decode_32(&p); 1642 ceph_decode_need(&p, end, maplen, bad); 1643 next = p + maplen; 1644 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) { 1645 dout("applying incremental map %u len %d\n", 1646 epoch, maplen); 1647 newmap = osdmap_apply_incremental(&p, next, 1648 osdc->osdmap, 1649 &osdc->client->msgr); 1650 if (IS_ERR(newmap)) { 1651 err = PTR_ERR(newmap); 1652 goto bad; 1653 } 1654 BUG_ON(!newmap); 1655 if (newmap != osdc->osdmap) { 1656 ceph_osdmap_destroy(osdc->osdmap); 1657 osdc->osdmap = newmap; 1658 } 1659 kick_requests(osdc, 0); 1660 } else { 1661 dout("ignoring incremental map %u len %d\n", 1662 epoch, maplen); 1663 } 1664 p = next; 1665 nr_maps--; 1666 } 1667 if (newmap) 1668 goto done; 1669 1670 /* full maps */ 1671 ceph_decode_32_safe(&p, end, nr_maps, bad); 1672 dout(" %d full maps\n", nr_maps); 1673 while (nr_maps) { 1674 ceph_decode_need(&p, end, 2*sizeof(u32), bad); 1675 epoch = ceph_decode_32(&p); 1676 maplen = ceph_decode_32(&p); 1677 ceph_decode_need(&p, end, maplen, bad); 1678 if (nr_maps > 1) { 1679 dout("skipping non-latest full map %u len %d\n", 1680 epoch, maplen); 1681 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) { 1682 dout("skipping full map %u len %d, " 1683 "older than our %u\n", epoch, maplen, 1684 osdc->osdmap->epoch); 1685 } else { 1686 int skipped_map = 0; 1687 1688 dout("taking full map %u len %d\n", epoch, maplen); 1689 newmap = osdmap_decode(&p, p+maplen); 1690 if (IS_ERR(newmap)) { 1691 err = PTR_ERR(newmap); 1692 goto bad; 1693 } 1694 BUG_ON(!newmap); 1695 oldmap = osdc->osdmap; 1696 osdc->osdmap = newmap; 1697 if (oldmap) { 1698 if (oldmap->epoch + 1 < newmap->epoch) 1699 skipped_map = 1; 1700 ceph_osdmap_destroy(oldmap); 1701 } 1702 kick_requests(osdc, skipped_map); 1703 } 1704 p += maplen; 1705 nr_maps--; 1706 } 1707 1708 done: 1709 downgrade_write(&osdc->map_sem); 1710 ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch); 1711 1712 /* 1713 * subscribe to subsequent osdmap updates if full to ensure 1714 * we find out when we are no longer full and stop returning 1715 * ENOSPC. 1716 */ 1717 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL)) 1718 ceph_monc_request_next_osdmap(&osdc->client->monc); 1719 1720 mutex_lock(&osdc->request_mutex); 1721 __send_queued(osdc); 1722 mutex_unlock(&osdc->request_mutex); 1723 up_read(&osdc->map_sem); 1724 wake_up_all(&osdc->client->auth_wq); 1725 return; 1726 1727 bad: 1728 pr_err("osdc handle_map corrupt msg\n"); 1729 ceph_msg_dump(msg); 1730 up_write(&osdc->map_sem); 1731 return; 1732 } 1733 1734 /* 1735 * watch/notify callback event infrastructure 1736 * 1737 * These callbacks are used both for watch and notify operations. 1738 */ 1739 static void __release_event(struct kref *kref) 1740 { 1741 struct ceph_osd_event *event = 1742 container_of(kref, struct ceph_osd_event, kref); 1743 1744 dout("__release_event %p\n", event); 1745 kfree(event); 1746 } 1747 1748 static void get_event(struct ceph_osd_event *event) 1749 { 1750 kref_get(&event->kref); 1751 } 1752 1753 void ceph_osdc_put_event(struct ceph_osd_event *event) 1754 { 1755 kref_put(&event->kref, __release_event); 1756 } 1757 EXPORT_SYMBOL(ceph_osdc_put_event); 1758 1759 static void __insert_event(struct ceph_osd_client *osdc, 1760 struct ceph_osd_event *new) 1761 { 1762 struct rb_node **p = &osdc->event_tree.rb_node; 1763 struct rb_node *parent = NULL; 1764 struct ceph_osd_event *event = NULL; 1765 1766 while (*p) { 1767 parent = *p; 1768 event = rb_entry(parent, struct ceph_osd_event, node); 1769 if (new->cookie < event->cookie) 1770 p = &(*p)->rb_left; 1771 else if (new->cookie > event->cookie) 1772 p = &(*p)->rb_right; 1773 else 1774 BUG(); 1775 } 1776 1777 rb_link_node(&new->node, parent, p); 1778 rb_insert_color(&new->node, &osdc->event_tree); 1779 } 1780 1781 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc, 1782 u64 cookie) 1783 { 1784 struct rb_node **p = &osdc->event_tree.rb_node; 1785 struct rb_node *parent = NULL; 1786 struct ceph_osd_event *event = NULL; 1787 1788 while (*p) { 1789 parent = *p; 1790 event = rb_entry(parent, struct ceph_osd_event, node); 1791 if (cookie < event->cookie) 1792 p = &(*p)->rb_left; 1793 else if (cookie > event->cookie) 1794 p = &(*p)->rb_right; 1795 else 1796 return event; 1797 } 1798 return NULL; 1799 } 1800 1801 static void __remove_event(struct ceph_osd_event *event) 1802 { 1803 struct ceph_osd_client *osdc = event->osdc; 1804 1805 if (!RB_EMPTY_NODE(&event->node)) { 1806 dout("__remove_event removed %p\n", event); 1807 rb_erase(&event->node, &osdc->event_tree); 1808 ceph_osdc_put_event(event); 1809 } else { 1810 dout("__remove_event didn't remove %p\n", event); 1811 } 1812 } 1813 1814 int ceph_osdc_create_event(struct ceph_osd_client *osdc, 1815 void (*event_cb)(u64, u64, u8, void *), 1816 void *data, struct ceph_osd_event **pevent) 1817 { 1818 struct ceph_osd_event *event; 1819 1820 event = kmalloc(sizeof(*event), GFP_NOIO); 1821 if (!event) 1822 return -ENOMEM; 1823 1824 dout("create_event %p\n", event); 1825 event->cb = event_cb; 1826 event->one_shot = 0; 1827 event->data = data; 1828 event->osdc = osdc; 1829 INIT_LIST_HEAD(&event->osd_node); 1830 RB_CLEAR_NODE(&event->node); 1831 kref_init(&event->kref); /* one ref for us */ 1832 kref_get(&event->kref); /* one ref for the caller */ 1833 1834 spin_lock(&osdc->event_lock); 1835 event->cookie = ++osdc->event_count; 1836 __insert_event(osdc, event); 1837 spin_unlock(&osdc->event_lock); 1838 1839 *pevent = event; 1840 return 0; 1841 } 1842 EXPORT_SYMBOL(ceph_osdc_create_event); 1843 1844 void ceph_osdc_cancel_event(struct ceph_osd_event *event) 1845 { 1846 struct ceph_osd_client *osdc = event->osdc; 1847 1848 dout("cancel_event %p\n", event); 1849 spin_lock(&osdc->event_lock); 1850 __remove_event(event); 1851 spin_unlock(&osdc->event_lock); 1852 ceph_osdc_put_event(event); /* caller's */ 1853 } 1854 EXPORT_SYMBOL(ceph_osdc_cancel_event); 1855 1856 1857 static void do_event_work(struct work_struct *work) 1858 { 1859 struct ceph_osd_event_work *event_work = 1860 container_of(work, struct ceph_osd_event_work, work); 1861 struct ceph_osd_event *event = event_work->event; 1862 u64 ver = event_work->ver; 1863 u64 notify_id = event_work->notify_id; 1864 u8 opcode = event_work->opcode; 1865 1866 dout("do_event_work completing %p\n", event); 1867 event->cb(ver, notify_id, opcode, event->data); 1868 dout("do_event_work completed %p\n", event); 1869 ceph_osdc_put_event(event); 1870 kfree(event_work); 1871 } 1872 1873 1874 /* 1875 * Process osd watch notifications 1876 */ 1877 static void handle_watch_notify(struct ceph_osd_client *osdc, 1878 struct ceph_msg *msg) 1879 { 1880 void *p, *end; 1881 u8 proto_ver; 1882 u64 cookie, ver, notify_id; 1883 u8 opcode; 1884 struct ceph_osd_event *event; 1885 struct ceph_osd_event_work *event_work; 1886 1887 p = msg->front.iov_base; 1888 end = p + msg->front.iov_len; 1889 1890 ceph_decode_8_safe(&p, end, proto_ver, bad); 1891 ceph_decode_8_safe(&p, end, opcode, bad); 1892 ceph_decode_64_safe(&p, end, cookie, bad); 1893 ceph_decode_64_safe(&p, end, ver, bad); 1894 ceph_decode_64_safe(&p, end, notify_id, bad); 1895 1896 spin_lock(&osdc->event_lock); 1897 event = __find_event(osdc, cookie); 1898 if (event) { 1899 BUG_ON(event->one_shot); 1900 get_event(event); 1901 } 1902 spin_unlock(&osdc->event_lock); 1903 dout("handle_watch_notify cookie %lld ver %lld event %p\n", 1904 cookie, ver, event); 1905 if (event) { 1906 event_work = kmalloc(sizeof(*event_work), GFP_NOIO); 1907 if (!event_work) { 1908 dout("ERROR: could not allocate event_work\n"); 1909 goto done_err; 1910 } 1911 INIT_WORK(&event_work->work, do_event_work); 1912 event_work->event = event; 1913 event_work->ver = ver; 1914 event_work->notify_id = notify_id; 1915 event_work->opcode = opcode; 1916 if (!queue_work(osdc->notify_wq, &event_work->work)) { 1917 dout("WARNING: failed to queue notify event work\n"); 1918 goto done_err; 1919 } 1920 } 1921 1922 return; 1923 1924 done_err: 1925 ceph_osdc_put_event(event); 1926 return; 1927 1928 bad: 1929 pr_err("osdc handle_watch_notify corrupt msg\n"); 1930 return; 1931 } 1932 1933 static void ceph_osdc_msg_data_set(struct ceph_msg *msg, 1934 struct ceph_osd_data *osd_data) 1935 { 1936 u64 length = ceph_osd_data_length(osd_data); 1937 1938 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 1939 BUG_ON(length > (u64) SIZE_MAX); 1940 if (length) 1941 ceph_msg_data_set_pages(msg, osd_data->pages, 1942 length, osd_data->alignment); 1943 } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) { 1944 BUG_ON(!length); 1945 ceph_msg_data_set_pagelist(msg, osd_data->pagelist); 1946 #ifdef CONFIG_BLOCK 1947 } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) { 1948 ceph_msg_data_set_bio(msg, osd_data->bio, length); 1949 #endif 1950 } else { 1951 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE); 1952 } 1953 } 1954 1955 /* 1956 * build new request AND message 1957 * 1958 */ 1959 void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off, 1960 struct ceph_snap_context *snapc, u64 snap_id, 1961 struct timespec *mtime) 1962 { 1963 struct ceph_msg *msg = req->r_request; 1964 void *p; 1965 size_t msg_size; 1966 int flags = req->r_flags; 1967 u64 data_len; 1968 unsigned int i; 1969 1970 /* Set up response incoming data and request outgoing data fields */ 1971 1972 ceph_osdc_msg_data_set(req->r_reply, &req->r_data_in); 1973 ceph_osdc_msg_data_set(req->r_request, &req->r_data_out); 1974 1975 req->r_snapid = snap_id; 1976 req->r_snapc = ceph_get_snap_context(snapc); 1977 1978 /* encode request */ 1979 msg->hdr.version = cpu_to_le16(4); 1980 1981 p = msg->front.iov_base; 1982 ceph_encode_32(&p, 1); /* client_inc is always 1 */ 1983 req->r_request_osdmap_epoch = p; 1984 p += 4; 1985 req->r_request_flags = p; 1986 p += 4; 1987 if (req->r_flags & CEPH_OSD_FLAG_WRITE) 1988 ceph_encode_timespec(p, mtime); 1989 p += sizeof(struct ceph_timespec); 1990 req->r_request_reassert_version = p; 1991 p += sizeof(struct ceph_eversion); /* will get filled in */ 1992 1993 /* oloc */ 1994 ceph_encode_8(&p, 4); 1995 ceph_encode_8(&p, 4); 1996 ceph_encode_32(&p, 8 + 4 + 4); 1997 req->r_request_pool = p; 1998 p += 8; 1999 ceph_encode_32(&p, -1); /* preferred */ 2000 ceph_encode_32(&p, 0); /* key len */ 2001 2002 ceph_encode_8(&p, 1); 2003 req->r_request_pgid = p; 2004 p += 8 + 4; 2005 ceph_encode_32(&p, -1); /* preferred */ 2006 2007 /* oid */ 2008 ceph_encode_32(&p, req->r_oid_len); 2009 memcpy(p, req->r_oid, req->r_oid_len); 2010 dout("oid '%.*s' len %d\n", req->r_oid_len, req->r_oid, req->r_oid_len); 2011 p += req->r_oid_len; 2012 2013 /* ops--can imply data */ 2014 ceph_encode_16(&p, (u16)req->r_num_ops); 2015 data_len = 0; 2016 for (i = 0; i < req->r_num_ops; i++) { 2017 data_len += osd_req_encode_op(req, p, i); 2018 p += sizeof(struct ceph_osd_op); 2019 } 2020 2021 /* snaps */ 2022 ceph_encode_64(&p, req->r_snapid); 2023 ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0); 2024 ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0); 2025 if (req->r_snapc) { 2026 for (i = 0; i < snapc->num_snaps; i++) { 2027 ceph_encode_64(&p, req->r_snapc->snaps[i]); 2028 } 2029 } 2030 2031 req->r_request_attempts = p; 2032 p += 4; 2033 2034 /* data */ 2035 if (flags & CEPH_OSD_FLAG_WRITE) { 2036 u16 data_off; 2037 2038 /* 2039 * The header "data_off" is a hint to the receiver 2040 * allowing it to align received data into its 2041 * buffers such that there's no need to re-copy 2042 * it before writing it to disk (direct I/O). 2043 */ 2044 data_off = (u16) (off & 0xffff); 2045 req->r_request->hdr.data_off = cpu_to_le16(data_off); 2046 } 2047 req->r_request->hdr.data_len = cpu_to_le32(data_len); 2048 2049 BUG_ON(p > msg->front.iov_base + msg->front.iov_len); 2050 msg_size = p - msg->front.iov_base; 2051 msg->front.iov_len = msg_size; 2052 msg->hdr.front_len = cpu_to_le32(msg_size); 2053 2054 dout("build_request msg_size was %d\n", (int)msg_size); 2055 } 2056 EXPORT_SYMBOL(ceph_osdc_build_request); 2057 2058 /* 2059 * Register request, send initial attempt. 2060 */ 2061 int ceph_osdc_start_request(struct ceph_osd_client *osdc, 2062 struct ceph_osd_request *req, 2063 bool nofail) 2064 { 2065 int rc = 0; 2066 2067 down_read(&osdc->map_sem); 2068 mutex_lock(&osdc->request_mutex); 2069 __register_request(osdc, req); 2070 WARN_ON(req->r_sent); 2071 rc = __map_request(osdc, req, 0); 2072 if (rc < 0) { 2073 if (nofail) { 2074 dout("osdc_start_request failed map, " 2075 " will retry %lld\n", req->r_tid); 2076 rc = 0; 2077 } 2078 goto out_unlock; 2079 } 2080 if (req->r_osd == NULL) { 2081 dout("send_request %p no up osds in pg\n", req); 2082 ceph_monc_request_next_osdmap(&osdc->client->monc); 2083 } else { 2084 __send_queued(osdc); 2085 } 2086 rc = 0; 2087 out_unlock: 2088 mutex_unlock(&osdc->request_mutex); 2089 up_read(&osdc->map_sem); 2090 return rc; 2091 } 2092 EXPORT_SYMBOL(ceph_osdc_start_request); 2093 2094 /* 2095 * wait for a request to complete 2096 */ 2097 int ceph_osdc_wait_request(struct ceph_osd_client *osdc, 2098 struct ceph_osd_request *req) 2099 { 2100 int rc; 2101 2102 rc = wait_for_completion_interruptible(&req->r_completion); 2103 if (rc < 0) { 2104 mutex_lock(&osdc->request_mutex); 2105 __cancel_request(req); 2106 __unregister_request(osdc, req); 2107 mutex_unlock(&osdc->request_mutex); 2108 complete_request(req); 2109 dout("wait_request tid %llu canceled/timed out\n", req->r_tid); 2110 return rc; 2111 } 2112 2113 dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result); 2114 return req->r_result; 2115 } 2116 EXPORT_SYMBOL(ceph_osdc_wait_request); 2117 2118 /* 2119 * sync - wait for all in-flight requests to flush. avoid starvation. 2120 */ 2121 void ceph_osdc_sync(struct ceph_osd_client *osdc) 2122 { 2123 struct ceph_osd_request *req; 2124 u64 last_tid, next_tid = 0; 2125 2126 mutex_lock(&osdc->request_mutex); 2127 last_tid = osdc->last_tid; 2128 while (1) { 2129 req = __lookup_request_ge(osdc, next_tid); 2130 if (!req) 2131 break; 2132 if (req->r_tid > last_tid) 2133 break; 2134 2135 next_tid = req->r_tid + 1; 2136 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0) 2137 continue; 2138 2139 ceph_osdc_get_request(req); 2140 mutex_unlock(&osdc->request_mutex); 2141 dout("sync waiting on tid %llu (last is %llu)\n", 2142 req->r_tid, last_tid); 2143 wait_for_completion(&req->r_safe_completion); 2144 mutex_lock(&osdc->request_mutex); 2145 ceph_osdc_put_request(req); 2146 } 2147 mutex_unlock(&osdc->request_mutex); 2148 dout("sync done (thru tid %llu)\n", last_tid); 2149 } 2150 EXPORT_SYMBOL(ceph_osdc_sync); 2151 2152 /* 2153 * init, shutdown 2154 */ 2155 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client) 2156 { 2157 int err; 2158 2159 dout("init\n"); 2160 osdc->client = client; 2161 osdc->osdmap = NULL; 2162 init_rwsem(&osdc->map_sem); 2163 init_completion(&osdc->map_waiters); 2164 osdc->last_requested_map = 0; 2165 mutex_init(&osdc->request_mutex); 2166 osdc->last_tid = 0; 2167 osdc->osds = RB_ROOT; 2168 INIT_LIST_HEAD(&osdc->osd_lru); 2169 osdc->requests = RB_ROOT; 2170 INIT_LIST_HEAD(&osdc->req_lru); 2171 INIT_LIST_HEAD(&osdc->req_unsent); 2172 INIT_LIST_HEAD(&osdc->req_notarget); 2173 INIT_LIST_HEAD(&osdc->req_linger); 2174 osdc->num_requests = 0; 2175 INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout); 2176 INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout); 2177 spin_lock_init(&osdc->event_lock); 2178 osdc->event_tree = RB_ROOT; 2179 osdc->event_count = 0; 2180 2181 schedule_delayed_work(&osdc->osds_timeout_work, 2182 round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ)); 2183 2184 err = -ENOMEM; 2185 osdc->req_mempool = mempool_create_kmalloc_pool(10, 2186 sizeof(struct ceph_osd_request)); 2187 if (!osdc->req_mempool) 2188 goto out; 2189 2190 err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP, 2191 OSD_OP_FRONT_LEN, 10, true, 2192 "osd_op"); 2193 if (err < 0) 2194 goto out_mempool; 2195 err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY, 2196 OSD_OPREPLY_FRONT_LEN, 10, true, 2197 "osd_op_reply"); 2198 if (err < 0) 2199 goto out_msgpool; 2200 2201 osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify"); 2202 if (IS_ERR(osdc->notify_wq)) { 2203 err = PTR_ERR(osdc->notify_wq); 2204 osdc->notify_wq = NULL; 2205 goto out_msgpool; 2206 } 2207 return 0; 2208 2209 out_msgpool: 2210 ceph_msgpool_destroy(&osdc->msgpool_op); 2211 out_mempool: 2212 mempool_destroy(osdc->req_mempool); 2213 out: 2214 return err; 2215 } 2216 2217 void ceph_osdc_stop(struct ceph_osd_client *osdc) 2218 { 2219 flush_workqueue(osdc->notify_wq); 2220 destroy_workqueue(osdc->notify_wq); 2221 cancel_delayed_work_sync(&osdc->timeout_work); 2222 cancel_delayed_work_sync(&osdc->osds_timeout_work); 2223 if (osdc->osdmap) { 2224 ceph_osdmap_destroy(osdc->osdmap); 2225 osdc->osdmap = NULL; 2226 } 2227 remove_all_osds(osdc); 2228 mempool_destroy(osdc->req_mempool); 2229 ceph_msgpool_destroy(&osdc->msgpool_op); 2230 ceph_msgpool_destroy(&osdc->msgpool_op_reply); 2231 } 2232 2233 /* 2234 * Read some contiguous pages. If we cross a stripe boundary, shorten 2235 * *plen. Return number of bytes read, or error. 2236 */ 2237 int ceph_osdc_readpages(struct ceph_osd_client *osdc, 2238 struct ceph_vino vino, struct ceph_file_layout *layout, 2239 u64 off, u64 *plen, 2240 u32 truncate_seq, u64 truncate_size, 2241 struct page **pages, int num_pages, int page_align) 2242 { 2243 struct ceph_osd_request *req; 2244 int rc = 0; 2245 2246 dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino, 2247 vino.snap, off, *plen); 2248 req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 1, 2249 CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ, 2250 NULL, truncate_seq, truncate_size, 2251 false); 2252 if (IS_ERR(req)) 2253 return PTR_ERR(req); 2254 2255 /* it may be a short read due to an object boundary */ 2256 2257 osd_req_op_extent_osd_data_pages(req, 0, false, 2258 pages, *plen, page_align, false, false); 2259 2260 dout("readpages final extent is %llu~%llu (%llu bytes align %d)\n", 2261 off, *plen, *plen, page_align); 2262 2263 ceph_osdc_build_request(req, off, NULL, vino.snap, NULL); 2264 2265 rc = ceph_osdc_start_request(osdc, req, false); 2266 if (!rc) 2267 rc = ceph_osdc_wait_request(osdc, req); 2268 2269 ceph_osdc_put_request(req); 2270 dout("readpages result %d\n", rc); 2271 return rc; 2272 } 2273 EXPORT_SYMBOL(ceph_osdc_readpages); 2274 2275 /* 2276 * do a synchronous write on N pages 2277 */ 2278 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino, 2279 struct ceph_file_layout *layout, 2280 struct ceph_snap_context *snapc, 2281 u64 off, u64 len, 2282 u32 truncate_seq, u64 truncate_size, 2283 struct timespec *mtime, 2284 struct page **pages, int num_pages) 2285 { 2286 struct ceph_osd_request *req; 2287 int rc = 0; 2288 int page_align = off & ~PAGE_MASK; 2289 2290 BUG_ON(vino.snap != CEPH_NOSNAP); /* snapshots aren't writeable */ 2291 req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 1, 2292 CEPH_OSD_OP_WRITE, 2293 CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE, 2294 snapc, truncate_seq, truncate_size, 2295 true); 2296 if (IS_ERR(req)) 2297 return PTR_ERR(req); 2298 2299 /* it may be a short write due to an object boundary */ 2300 osd_req_op_extent_osd_data_pages(req, 0, true, pages, len, page_align, 2301 false, false); 2302 dout("writepages %llu~%llu (%llu bytes)\n", off, len, len); 2303 2304 ceph_osdc_build_request(req, off, snapc, CEPH_NOSNAP, mtime); 2305 2306 rc = ceph_osdc_start_request(osdc, req, true); 2307 if (!rc) 2308 rc = ceph_osdc_wait_request(osdc, req); 2309 2310 ceph_osdc_put_request(req); 2311 if (rc == 0) 2312 rc = len; 2313 dout("writepages result %d\n", rc); 2314 return rc; 2315 } 2316 EXPORT_SYMBOL(ceph_osdc_writepages); 2317 2318 /* 2319 * handle incoming message 2320 */ 2321 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg) 2322 { 2323 struct ceph_osd *osd = con->private; 2324 struct ceph_osd_client *osdc; 2325 int type = le16_to_cpu(msg->hdr.type); 2326 2327 if (!osd) 2328 goto out; 2329 osdc = osd->o_osdc; 2330 2331 switch (type) { 2332 case CEPH_MSG_OSD_MAP: 2333 ceph_osdc_handle_map(osdc, msg); 2334 break; 2335 case CEPH_MSG_OSD_OPREPLY: 2336 handle_reply(osdc, msg, con); 2337 break; 2338 case CEPH_MSG_WATCH_NOTIFY: 2339 handle_watch_notify(osdc, msg); 2340 break; 2341 2342 default: 2343 pr_err("received unknown message type %d %s\n", type, 2344 ceph_msg_type_name(type)); 2345 } 2346 out: 2347 ceph_msg_put(msg); 2348 } 2349 2350 /* 2351 * lookup and return message for incoming reply. set up reply message 2352 * pages. 2353 */ 2354 static struct ceph_msg *get_reply(struct ceph_connection *con, 2355 struct ceph_msg_header *hdr, 2356 int *skip) 2357 { 2358 struct ceph_osd *osd = con->private; 2359 struct ceph_osd_client *osdc = osd->o_osdc; 2360 struct ceph_msg *m; 2361 struct ceph_osd_request *req; 2362 int front = le32_to_cpu(hdr->front_len); 2363 int data_len = le32_to_cpu(hdr->data_len); 2364 u64 tid; 2365 2366 tid = le64_to_cpu(hdr->tid); 2367 mutex_lock(&osdc->request_mutex); 2368 req = __lookup_request(osdc, tid); 2369 if (!req) { 2370 *skip = 1; 2371 m = NULL; 2372 dout("get_reply unknown tid %llu from osd%d\n", tid, 2373 osd->o_osd); 2374 goto out; 2375 } 2376 2377 if (req->r_reply->con) 2378 dout("%s revoking msg %p from old con %p\n", __func__, 2379 req->r_reply, req->r_reply->con); 2380 ceph_msg_revoke_incoming(req->r_reply); 2381 2382 if (front > req->r_reply->front.iov_len) { 2383 pr_warning("get_reply front %d > preallocated %d\n", 2384 front, (int)req->r_reply->front.iov_len); 2385 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS, false); 2386 if (!m) 2387 goto out; 2388 ceph_msg_put(req->r_reply); 2389 req->r_reply = m; 2390 } 2391 m = ceph_msg_get(req->r_reply); 2392 2393 if (data_len > 0) { 2394 struct ceph_osd_data *osd_data; 2395 2396 /* 2397 * XXX This is assuming there is only one op containing 2398 * XXX page data. Probably OK for reads, but this 2399 * XXX ought to be done more generally. 2400 */ 2401 osd_data = osd_req_op_extent_osd_data(req, 0, false); 2402 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 2403 if (osd_data->pages && 2404 unlikely(osd_data->length < data_len)) { 2405 2406 pr_warning("tid %lld reply has %d bytes " 2407 "we had only %llu bytes ready\n", 2408 tid, data_len, osd_data->length); 2409 *skip = 1; 2410 ceph_msg_put(m); 2411 m = NULL; 2412 goto out; 2413 } 2414 } 2415 } 2416 *skip = 0; 2417 dout("get_reply tid %lld %p\n", tid, m); 2418 2419 out: 2420 mutex_unlock(&osdc->request_mutex); 2421 return m; 2422 2423 } 2424 2425 static struct ceph_msg *alloc_msg(struct ceph_connection *con, 2426 struct ceph_msg_header *hdr, 2427 int *skip) 2428 { 2429 struct ceph_osd *osd = con->private; 2430 int type = le16_to_cpu(hdr->type); 2431 int front = le32_to_cpu(hdr->front_len); 2432 2433 *skip = 0; 2434 switch (type) { 2435 case CEPH_MSG_OSD_MAP: 2436 case CEPH_MSG_WATCH_NOTIFY: 2437 return ceph_msg_new(type, front, GFP_NOFS, false); 2438 case CEPH_MSG_OSD_OPREPLY: 2439 return get_reply(con, hdr, skip); 2440 default: 2441 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type, 2442 osd->o_osd); 2443 *skip = 1; 2444 return NULL; 2445 } 2446 } 2447 2448 /* 2449 * Wrappers to refcount containing ceph_osd struct 2450 */ 2451 static struct ceph_connection *get_osd_con(struct ceph_connection *con) 2452 { 2453 struct ceph_osd *osd = con->private; 2454 if (get_osd(osd)) 2455 return con; 2456 return NULL; 2457 } 2458 2459 static void put_osd_con(struct ceph_connection *con) 2460 { 2461 struct ceph_osd *osd = con->private; 2462 put_osd(osd); 2463 } 2464 2465 /* 2466 * authentication 2467 */ 2468 /* 2469 * Note: returned pointer is the address of a structure that's 2470 * managed separately. Caller must *not* attempt to free it. 2471 */ 2472 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con, 2473 int *proto, int force_new) 2474 { 2475 struct ceph_osd *o = con->private; 2476 struct ceph_osd_client *osdc = o->o_osdc; 2477 struct ceph_auth_client *ac = osdc->client->monc.auth; 2478 struct ceph_auth_handshake *auth = &o->o_auth; 2479 2480 if (force_new && auth->authorizer) { 2481 ceph_auth_destroy_authorizer(ac, auth->authorizer); 2482 auth->authorizer = NULL; 2483 } 2484 if (!auth->authorizer) { 2485 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 2486 auth); 2487 if (ret) 2488 return ERR_PTR(ret); 2489 } else { 2490 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 2491 auth); 2492 if (ret) 2493 return ERR_PTR(ret); 2494 } 2495 *proto = ac->protocol; 2496 2497 return auth; 2498 } 2499 2500 2501 static int verify_authorizer_reply(struct ceph_connection *con, int len) 2502 { 2503 struct ceph_osd *o = con->private; 2504 struct ceph_osd_client *osdc = o->o_osdc; 2505 struct ceph_auth_client *ac = osdc->client->monc.auth; 2506 2507 return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len); 2508 } 2509 2510 static int invalidate_authorizer(struct ceph_connection *con) 2511 { 2512 struct ceph_osd *o = con->private; 2513 struct ceph_osd_client *osdc = o->o_osdc; 2514 struct ceph_auth_client *ac = osdc->client->monc.auth; 2515 2516 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD); 2517 return ceph_monc_validate_auth(&osdc->client->monc); 2518 } 2519 2520 static const struct ceph_connection_operations osd_con_ops = { 2521 .get = get_osd_con, 2522 .put = put_osd_con, 2523 .dispatch = dispatch, 2524 .get_authorizer = get_authorizer, 2525 .verify_authorizer_reply = verify_authorizer_reply, 2526 .invalidate_authorizer = invalidate_authorizer, 2527 .alloc_msg = alloc_msg, 2528 .fault = osd_reset, 2529 }; 2530