13d14c5d2SYehuda Sadeh #include <linux/ceph/ceph_debug.h> 23d14c5d2SYehuda Sadeh 33d14c5d2SYehuda Sadeh #include <linux/module.h> 43d14c5d2SYehuda Sadeh #include <linux/err.h> 53d14c5d2SYehuda Sadeh #include <linux/highmem.h> 63d14c5d2SYehuda Sadeh #include <linux/mm.h> 73d14c5d2SYehuda Sadeh #include <linux/pagemap.h> 83d14c5d2SYehuda Sadeh #include <linux/slab.h> 93d14c5d2SYehuda Sadeh #include <linux/uaccess.h> 103d14c5d2SYehuda Sadeh #ifdef CONFIG_BLOCK 113d14c5d2SYehuda Sadeh #include <linux/bio.h> 123d14c5d2SYehuda Sadeh #endif 133d14c5d2SYehuda Sadeh 143d14c5d2SYehuda Sadeh #include <linux/ceph/libceph.h> 153d14c5d2SYehuda Sadeh #include <linux/ceph/osd_client.h> 163d14c5d2SYehuda Sadeh #include <linux/ceph/messenger.h> 173d14c5d2SYehuda Sadeh #include <linux/ceph/decode.h> 183d14c5d2SYehuda Sadeh #include <linux/ceph/auth.h> 193d14c5d2SYehuda Sadeh #include <linux/ceph/pagelist.h> 203d14c5d2SYehuda Sadeh 213d14c5d2SYehuda Sadeh #define OSD_OP_FRONT_LEN 4096 223d14c5d2SYehuda Sadeh #define OSD_OPREPLY_FRONT_LEN 512 233d14c5d2SYehuda Sadeh 243d14c5d2SYehuda Sadeh static const struct ceph_connection_operations osd_con_ops; 253d14c5d2SYehuda Sadeh 26f9d25199SAlex Elder static void __send_queued(struct ceph_osd_client *osdc); 276f6c7006SSage Weil static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd); 28a40c4f10SYehuda Sadeh static void __register_request(struct ceph_osd_client *osdc, 29a40c4f10SYehuda Sadeh struct ceph_osd_request *req); 30a40c4f10SYehuda Sadeh static void __unregister_linger_request(struct ceph_osd_client *osdc, 31a40c4f10SYehuda Sadeh struct ceph_osd_request *req); 3256e925b6SSage Weil static void __send_request(struct ceph_osd_client *osdc, 33a40c4f10SYehuda Sadeh struct ceph_osd_request *req); 343d14c5d2SYehuda Sadeh 353d14c5d2SYehuda Sadeh static int op_has_extent(int op) 363d14c5d2SYehuda Sadeh { 373d14c5d2SYehuda Sadeh return (op == CEPH_OSD_OP_READ || 383d14c5d2SYehuda Sadeh op == CEPH_OSD_OP_WRITE); 393d14c5d2SYehuda Sadeh } 403d14c5d2SYehuda Sadeh 413d14c5d2SYehuda Sadeh /* 423d14c5d2SYehuda Sadeh * Implement client access to distributed object storage cluster. 433d14c5d2SYehuda Sadeh * 443d14c5d2SYehuda Sadeh * All data objects are stored within a cluster/cloud of OSDs, or 453d14c5d2SYehuda Sadeh * "object storage devices." (Note that Ceph OSDs have _nothing_ to 463d14c5d2SYehuda Sadeh * do with the T10 OSD extensions to SCSI.) Ceph OSDs are simply 473d14c5d2SYehuda Sadeh * remote daemons serving up and coordinating consistent and safe 483d14c5d2SYehuda Sadeh * access to storage. 493d14c5d2SYehuda Sadeh * 503d14c5d2SYehuda Sadeh * Cluster membership and the mapping of data objects onto storage devices 513d14c5d2SYehuda Sadeh * are described by the osd map. 523d14c5d2SYehuda Sadeh * 533d14c5d2SYehuda Sadeh * We keep track of pending OSD requests (read, write), resubmit 543d14c5d2SYehuda Sadeh * requests to different OSDs when the cluster topology/data layout 553d14c5d2SYehuda Sadeh * change, or retry the affected requests when the communications 563d14c5d2SYehuda Sadeh * channel with an OSD is reset. 573d14c5d2SYehuda Sadeh */ 583d14c5d2SYehuda Sadeh 593d14c5d2SYehuda Sadeh /* 603d14c5d2SYehuda Sadeh * calculate the mapping of a file extent onto an object, and fill out the 613d14c5d2SYehuda Sadeh * request accordingly. shorten extent as necessary if it crosses an 623d14c5d2SYehuda Sadeh * object boundary. 633d14c5d2SYehuda Sadeh * 643d14c5d2SYehuda Sadeh * fill osd op in request message. 653d14c5d2SYehuda Sadeh */ 66dbe0fc41SAlex Elder static int calc_layout(struct ceph_file_layout *layout, u64 off, u64 *plen, 6775d1c941SAlex Elder struct ceph_osd_req_op *op, u64 *objnum, 6875d1c941SAlex Elder u64 *objoff, u64 *objlen) 693d14c5d2SYehuda Sadeh { 7060e56f13SAlex Elder u64 orig_len = *plen; 71d63b77f4SSage Weil int r; 723d14c5d2SYehuda Sadeh 7360e56f13SAlex Elder /* object extent? */ 7475d1c941SAlex Elder r = ceph_calc_file_object_mapping(layout, off, orig_len, objnum, 7575d1c941SAlex Elder objoff, objlen); 76d63b77f4SSage Weil if (r < 0) 77d63b77f4SSage Weil return r; 7875d1c941SAlex Elder if (*objlen < orig_len) { 7975d1c941SAlex Elder *plen = *objlen; 8060e56f13SAlex Elder dout(" skipping last %llu, final file extent %llu~%llu\n", 8160e56f13SAlex Elder orig_len - *plen, off, *plen); 8260e56f13SAlex Elder } 8360e56f13SAlex Elder 8460e56f13SAlex Elder if (op_has_extent(op->op)) { 8560e56f13SAlex Elder u32 osize = le32_to_cpu(layout->fl_object_size); 8675d1c941SAlex Elder op->extent.offset = *objoff; 8775d1c941SAlex Elder op->extent.length = *objlen; 8875d1c941SAlex Elder if (op->extent.truncate_size <= off - *objoff) { 8960e56f13SAlex Elder op->extent.truncate_size = 0; 9060e56f13SAlex Elder } else { 9175d1c941SAlex Elder op->extent.truncate_size -= off - *objoff; 9260e56f13SAlex Elder if (op->extent.truncate_size > osize) 9360e56f13SAlex Elder op->extent.truncate_size = osize; 9460e56f13SAlex Elder } 9560e56f13SAlex Elder } 9660e56f13SAlex Elder if (op->op == CEPH_OSD_OP_WRITE) 9760e56f13SAlex Elder op->payload_len = *plen; 9860e56f13SAlex Elder 9975d1c941SAlex Elder dout("calc_layout objnum=%llx %llu~%llu\n", *objnum, *objoff, *objlen); 1003d14c5d2SYehuda Sadeh 1013ff5f385SAlex Elder return 0; 1023d14c5d2SYehuda Sadeh } 1033d14c5d2SYehuda Sadeh 1043d14c5d2SYehuda Sadeh /* 1053d14c5d2SYehuda Sadeh * requests 1063d14c5d2SYehuda Sadeh */ 1073d14c5d2SYehuda Sadeh void ceph_osdc_release_request(struct kref *kref) 1083d14c5d2SYehuda Sadeh { 109e0c59487SAlex Elder int num_pages; 1103d14c5d2SYehuda Sadeh struct ceph_osd_request *req = container_of(kref, 1113d14c5d2SYehuda Sadeh struct ceph_osd_request, 1123d14c5d2SYehuda Sadeh r_kref); 1133d14c5d2SYehuda Sadeh 1143d14c5d2SYehuda Sadeh if (req->r_request) 1153d14c5d2SYehuda Sadeh ceph_msg_put(req->r_request); 1163d14c5d2SYehuda Sadeh if (req->r_con_filling_msg) { 1179cbb1d72SAlex Elder dout("%s revoking msg %p from con %p\n", __func__, 1189cbb1d72SAlex Elder req->r_reply, req->r_con_filling_msg); 1198921d114SAlex Elder ceph_msg_revoke_incoming(req->r_reply); 1200d47766fSSage Weil req->r_con_filling_msg->ops->put(req->r_con_filling_msg); 1219cbb1d72SAlex Elder req->r_con_filling_msg = NULL; 1223d14c5d2SYehuda Sadeh } 123ab8cb34aSAlex Elder if (req->r_reply) 124ab8cb34aSAlex Elder ceph_msg_put(req->r_reply); 1250fff87ecSAlex Elder 1260fff87ecSAlex Elder if (req->r_data_in.type == CEPH_OSD_DATA_TYPE_PAGES && 127e0c59487SAlex Elder req->r_data_in.own_pages) { 128e0c59487SAlex Elder num_pages = calc_pages_for((u64)req->r_data_in.alignment, 129e0c59487SAlex Elder (u64)req->r_data_in.length); 130e0c59487SAlex Elder ceph_release_page_vector(req->r_data_in.pages, num_pages); 131e0c59487SAlex Elder } 1320fff87ecSAlex Elder if (req->r_data_out.type == CEPH_OSD_DATA_TYPE_PAGES && 133e0c59487SAlex Elder req->r_data_out.own_pages) { 134e0c59487SAlex Elder num_pages = calc_pages_for((u64)req->r_data_out.alignment, 135e0c59487SAlex Elder (u64)req->r_data_out.length); 136e0c59487SAlex Elder ceph_release_page_vector(req->r_data_out.pages, num_pages); 137e0c59487SAlex Elder } 1380fff87ecSAlex Elder 1393d14c5d2SYehuda Sadeh ceph_put_snap_context(req->r_snapc); 1403d14c5d2SYehuda Sadeh if (req->r_mempool) 1413d14c5d2SYehuda Sadeh mempool_free(req, req->r_osdc->req_mempool); 1423d14c5d2SYehuda Sadeh else 1433d14c5d2SYehuda Sadeh kfree(req); 1443d14c5d2SYehuda Sadeh } 1453d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_release_request); 1463d14c5d2SYehuda Sadeh 1473d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc, 1483d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 1491b83bef2SSage Weil unsigned int num_ops, 1503d14c5d2SYehuda Sadeh bool use_mempool, 15154a54007SAlex Elder gfp_t gfp_flags) 1523d14c5d2SYehuda Sadeh { 1533d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 1543d14c5d2SYehuda Sadeh struct ceph_msg *msg; 1551b83bef2SSage Weil size_t msg_size; 1563d14c5d2SYehuda Sadeh 1571b83bef2SSage Weil msg_size = 4 + 4 + 8 + 8 + 4+8; 1581b83bef2SSage Weil msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */ 1591b83bef2SSage Weil msg_size += 1 + 8 + 4 + 4; /* pg_t */ 1601b83bef2SSage Weil msg_size += 4 + MAX_OBJ_NAME_SIZE; 1611b83bef2SSage Weil msg_size += 2 + num_ops*sizeof(struct ceph_osd_op); 1621b83bef2SSage Weil msg_size += 8; /* snapid */ 1631b83bef2SSage Weil msg_size += 8; /* snap_seq */ 1641b83bef2SSage Weil msg_size += 8 * (snapc ? snapc->num_snaps : 0); /* snaps */ 1651b83bef2SSage Weil msg_size += 4; 1663d14c5d2SYehuda Sadeh 1673d14c5d2SYehuda Sadeh if (use_mempool) { 1683d14c5d2SYehuda Sadeh req = mempool_alloc(osdc->req_mempool, gfp_flags); 1693d14c5d2SYehuda Sadeh memset(req, 0, sizeof(*req)); 1703d14c5d2SYehuda Sadeh } else { 1713d14c5d2SYehuda Sadeh req = kzalloc(sizeof(*req), gfp_flags); 1723d14c5d2SYehuda Sadeh } 1733d14c5d2SYehuda Sadeh if (req == NULL) 1743d14c5d2SYehuda Sadeh return NULL; 1753d14c5d2SYehuda Sadeh 1763d14c5d2SYehuda Sadeh req->r_osdc = osdc; 1773d14c5d2SYehuda Sadeh req->r_mempool = use_mempool; 1783d14c5d2SYehuda Sadeh 1793d14c5d2SYehuda Sadeh kref_init(&req->r_kref); 1803d14c5d2SYehuda Sadeh init_completion(&req->r_completion); 1813d14c5d2SYehuda Sadeh init_completion(&req->r_safe_completion); 182a978fa20SAlex Elder RB_CLEAR_NODE(&req->r_node); 1833d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&req->r_unsafe_item); 184a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&req->r_linger_item); 185a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&req->r_linger_osd); 186935b639aSSage Weil INIT_LIST_HEAD(&req->r_req_lru_item); 187cd43045cSSage Weil INIT_LIST_HEAD(&req->r_osd_item); 188cd43045cSSage Weil 1893d14c5d2SYehuda Sadeh /* create reply message */ 1903d14c5d2SYehuda Sadeh if (use_mempool) 1913d14c5d2SYehuda Sadeh msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0); 1923d14c5d2SYehuda Sadeh else 1933d14c5d2SYehuda Sadeh msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, 194b61c2763SSage Weil OSD_OPREPLY_FRONT_LEN, gfp_flags, true); 1953d14c5d2SYehuda Sadeh if (!msg) { 1963d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 1973d14c5d2SYehuda Sadeh return NULL; 1983d14c5d2SYehuda Sadeh } 1993d14c5d2SYehuda Sadeh req->r_reply = msg; 2003d14c5d2SYehuda Sadeh 2010fff87ecSAlex Elder req->r_data_in.type = CEPH_OSD_DATA_TYPE_NONE; 2020fff87ecSAlex Elder req->r_data_out.type = CEPH_OSD_DATA_TYPE_NONE; 203d50b409fSSage Weil 2043d14c5d2SYehuda Sadeh /* create request message; allow space for oid */ 2053d14c5d2SYehuda Sadeh if (use_mempool) 2063d14c5d2SYehuda Sadeh msg = ceph_msgpool_get(&osdc->msgpool_op, 0); 2073d14c5d2SYehuda Sadeh else 208b61c2763SSage Weil msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true); 2093d14c5d2SYehuda Sadeh if (!msg) { 2103d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 2113d14c5d2SYehuda Sadeh return NULL; 2123d14c5d2SYehuda Sadeh } 2133d14c5d2SYehuda Sadeh 2143d14c5d2SYehuda Sadeh memset(msg->front.iov_base, 0, msg->front.iov_len); 2153d14c5d2SYehuda Sadeh 2163d14c5d2SYehuda Sadeh req->r_request = msg; 2173d14c5d2SYehuda Sadeh 2183d14c5d2SYehuda Sadeh return req; 2193d14c5d2SYehuda Sadeh } 2203d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_alloc_request); 2213d14c5d2SYehuda Sadeh 222a8dd0a37SAlex Elder static bool osd_req_opcode_valid(u16 opcode) 223a8dd0a37SAlex Elder { 224a8dd0a37SAlex Elder switch (opcode) { 225a8dd0a37SAlex Elder case CEPH_OSD_OP_READ: 226a8dd0a37SAlex Elder case CEPH_OSD_OP_STAT: 227a8dd0a37SAlex Elder case CEPH_OSD_OP_MAPEXT: 228a8dd0a37SAlex Elder case CEPH_OSD_OP_MASKTRUNC: 229a8dd0a37SAlex Elder case CEPH_OSD_OP_SPARSE_READ: 230a8dd0a37SAlex Elder case CEPH_OSD_OP_NOTIFY: 231a8dd0a37SAlex Elder case CEPH_OSD_OP_NOTIFY_ACK: 232a8dd0a37SAlex Elder case CEPH_OSD_OP_ASSERT_VER: 233a8dd0a37SAlex Elder case CEPH_OSD_OP_WRITE: 234a8dd0a37SAlex Elder case CEPH_OSD_OP_WRITEFULL: 235a8dd0a37SAlex Elder case CEPH_OSD_OP_TRUNCATE: 236a8dd0a37SAlex Elder case CEPH_OSD_OP_ZERO: 237a8dd0a37SAlex Elder case CEPH_OSD_OP_DELETE: 238a8dd0a37SAlex Elder case CEPH_OSD_OP_APPEND: 239a8dd0a37SAlex Elder case CEPH_OSD_OP_STARTSYNC: 240a8dd0a37SAlex Elder case CEPH_OSD_OP_SETTRUNC: 241a8dd0a37SAlex Elder case CEPH_OSD_OP_TRIMTRUNC: 242a8dd0a37SAlex Elder case CEPH_OSD_OP_TMAPUP: 243a8dd0a37SAlex Elder case CEPH_OSD_OP_TMAPPUT: 244a8dd0a37SAlex Elder case CEPH_OSD_OP_TMAPGET: 245a8dd0a37SAlex Elder case CEPH_OSD_OP_CREATE: 246a8dd0a37SAlex Elder case CEPH_OSD_OP_ROLLBACK: 247a8dd0a37SAlex Elder case CEPH_OSD_OP_WATCH: 248a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPGETKEYS: 249a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPGETVALS: 250a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPGETHEADER: 251a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPGETVALSBYKEYS: 252a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPSETVALS: 253a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPSETHEADER: 254a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPCLEAR: 255a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPRMKEYS: 256a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAP_CMP: 257a8dd0a37SAlex Elder case CEPH_OSD_OP_CLONERANGE: 258a8dd0a37SAlex Elder case CEPH_OSD_OP_ASSERT_SRC_VERSION: 259a8dd0a37SAlex Elder case CEPH_OSD_OP_SRC_CMPXATTR: 260a8dd0a37SAlex Elder case CEPH_OSD_OP_GETXATTR: 261a8dd0a37SAlex Elder case CEPH_OSD_OP_GETXATTRS: 262a8dd0a37SAlex Elder case CEPH_OSD_OP_CMPXATTR: 263a8dd0a37SAlex Elder case CEPH_OSD_OP_SETXATTR: 264a8dd0a37SAlex Elder case CEPH_OSD_OP_SETXATTRS: 265a8dd0a37SAlex Elder case CEPH_OSD_OP_RESETXATTRS: 266a8dd0a37SAlex Elder case CEPH_OSD_OP_RMXATTR: 267a8dd0a37SAlex Elder case CEPH_OSD_OP_PULL: 268a8dd0a37SAlex Elder case CEPH_OSD_OP_PUSH: 269a8dd0a37SAlex Elder case CEPH_OSD_OP_BALANCEREADS: 270a8dd0a37SAlex Elder case CEPH_OSD_OP_UNBALANCEREADS: 271a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB: 272a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB_RESERVE: 273a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB_UNRESERVE: 274a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB_STOP: 275a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB_MAP: 276a8dd0a37SAlex Elder case CEPH_OSD_OP_WRLOCK: 277a8dd0a37SAlex Elder case CEPH_OSD_OP_WRUNLOCK: 278a8dd0a37SAlex Elder case CEPH_OSD_OP_RDLOCK: 279a8dd0a37SAlex Elder case CEPH_OSD_OP_RDUNLOCK: 280a8dd0a37SAlex Elder case CEPH_OSD_OP_UPLOCK: 281a8dd0a37SAlex Elder case CEPH_OSD_OP_DNLOCK: 282a8dd0a37SAlex Elder case CEPH_OSD_OP_CALL: 283a8dd0a37SAlex Elder case CEPH_OSD_OP_PGLS: 284a8dd0a37SAlex Elder case CEPH_OSD_OP_PGLS_FILTER: 285a8dd0a37SAlex Elder return true; 286a8dd0a37SAlex Elder default: 287a8dd0a37SAlex Elder return false; 288a8dd0a37SAlex Elder } 289a8dd0a37SAlex Elder } 290a8dd0a37SAlex Elder 29133803f33SAlex Elder /* 29233803f33SAlex Elder * This is an osd op init function for opcodes that have no data or 29333803f33SAlex Elder * other information associated with them. It also serves as a 29433803f33SAlex Elder * common init routine for all the other init functions, below. 29533803f33SAlex Elder */ 29633803f33SAlex Elder void osd_req_op_init(struct ceph_osd_req_op *op, u16 opcode) 29733803f33SAlex Elder { 29833803f33SAlex Elder BUG_ON(!osd_req_opcode_valid(opcode)); 29933803f33SAlex Elder 30033803f33SAlex Elder memset(op, 0, sizeof (*op)); 30133803f33SAlex Elder 30233803f33SAlex Elder op->op = opcode; 30333803f33SAlex Elder } 30433803f33SAlex Elder 30533803f33SAlex Elder void osd_req_op_extent_init(struct ceph_osd_req_op *op, u16 opcode, 30633803f33SAlex Elder u64 offset, u64 length, 30733803f33SAlex Elder u64 truncate_size, u32 truncate_seq) 30833803f33SAlex Elder { 30933803f33SAlex Elder size_t payload_len = 0; 31033803f33SAlex Elder 31133803f33SAlex Elder BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE); 31233803f33SAlex Elder 31333803f33SAlex Elder osd_req_op_init(op, opcode); 31433803f33SAlex Elder 31533803f33SAlex Elder op->extent.offset = offset; 31633803f33SAlex Elder op->extent.length = length; 31733803f33SAlex Elder op->extent.truncate_size = truncate_size; 31833803f33SAlex Elder op->extent.truncate_seq = truncate_seq; 31933803f33SAlex Elder if (opcode == CEPH_OSD_OP_WRITE) 32033803f33SAlex Elder payload_len += length; 32133803f33SAlex Elder 32233803f33SAlex Elder op->payload_len = payload_len; 32333803f33SAlex Elder } 32433803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_init); 32533803f33SAlex Elder 32633803f33SAlex Elder void osd_req_op_cls_init(struct ceph_osd_req_op *op, u16 opcode, 32733803f33SAlex Elder const char *class, const char *method, 32833803f33SAlex Elder const void *request_data, size_t request_data_size) 32933803f33SAlex Elder { 33033803f33SAlex Elder size_t payload_len = 0; 33133803f33SAlex Elder size_t size; 33233803f33SAlex Elder 33333803f33SAlex Elder BUG_ON(opcode != CEPH_OSD_OP_CALL); 33433803f33SAlex Elder 33533803f33SAlex Elder osd_req_op_init(op, opcode); 33633803f33SAlex Elder 33733803f33SAlex Elder op->cls.class_name = class; 33833803f33SAlex Elder size = strlen(class); 33933803f33SAlex Elder BUG_ON(size > (size_t) U8_MAX); 34033803f33SAlex Elder op->cls.class_len = size; 34133803f33SAlex Elder payload_len += size; 34233803f33SAlex Elder 34333803f33SAlex Elder op->cls.method_name = method; 34433803f33SAlex Elder size = strlen(method); 34533803f33SAlex Elder BUG_ON(size > (size_t) U8_MAX); 34633803f33SAlex Elder op->cls.method_len = size; 34733803f33SAlex Elder payload_len += size; 34833803f33SAlex Elder 34933803f33SAlex Elder op->cls.indata = request_data; 35033803f33SAlex Elder BUG_ON(request_data_size > (size_t) U32_MAX); 35133803f33SAlex Elder op->cls.indata_len = (u32) request_data_size; 35233803f33SAlex Elder payload_len += request_data_size; 35333803f33SAlex Elder 35433803f33SAlex Elder op->cls.argc = 0; /* currently unused */ 35533803f33SAlex Elder 35633803f33SAlex Elder op->payload_len = payload_len; 35733803f33SAlex Elder } 35833803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_init); 35933803f33SAlex Elder 36033803f33SAlex Elder void osd_req_op_watch_init(struct ceph_osd_req_op *op, u16 opcode, 36133803f33SAlex Elder u64 cookie, u64 version, int flag) 36233803f33SAlex Elder { 36333803f33SAlex Elder BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH); 36433803f33SAlex Elder 36533803f33SAlex Elder osd_req_op_init(op, opcode); 36633803f33SAlex Elder 36733803f33SAlex Elder op->watch.cookie = cookie; 36833803f33SAlex Elder /* op->watch.ver = version; */ /* XXX 3847 */ 36933803f33SAlex Elder op->watch.ver = cpu_to_le64(version); 37033803f33SAlex Elder if (opcode == CEPH_OSD_OP_WATCH && flag) 37133803f33SAlex Elder op->watch.flag = (u8) 1; 37233803f33SAlex Elder } 37333803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_watch_init); 37433803f33SAlex Elder 375175face2SAlex Elder static u64 osd_req_encode_op(struct ceph_osd_request *req, 3763d14c5d2SYehuda Sadeh struct ceph_osd_op *dst, 3773d14c5d2SYehuda Sadeh struct ceph_osd_req_op *src) 3783d14c5d2SYehuda Sadeh { 379175face2SAlex Elder u64 out_data_len = 0; 38095e072ebSAlex Elder struct ceph_pagelist *pagelist; 381175face2SAlex Elder 382a8dd0a37SAlex Elder if (WARN_ON(!osd_req_opcode_valid(src->op))) { 383a8dd0a37SAlex Elder pr_err("unrecognized osd opcode %d\n", src->op); 384a8dd0a37SAlex Elder 385a8dd0a37SAlex Elder return 0; 386a8dd0a37SAlex Elder } 3873d14c5d2SYehuda Sadeh 388065a68f9SAlex Elder switch (src->op) { 389fbfab539SAlex Elder case CEPH_OSD_OP_STAT: 390fbfab539SAlex Elder break; 3913d14c5d2SYehuda Sadeh case CEPH_OSD_OP_READ: 3923d14c5d2SYehuda Sadeh case CEPH_OSD_OP_WRITE: 393175face2SAlex Elder if (src->op == CEPH_OSD_OP_WRITE) 394175face2SAlex Elder out_data_len = src->extent.length; 395175face2SAlex Elder dst->extent.offset = cpu_to_le64(src->extent.offset); 396175face2SAlex Elder dst->extent.length = cpu_to_le64(src->extent.length); 3973d14c5d2SYehuda Sadeh dst->extent.truncate_size = 3983d14c5d2SYehuda Sadeh cpu_to_le64(src->extent.truncate_size); 3993d14c5d2SYehuda Sadeh dst->extent.truncate_seq = 4003d14c5d2SYehuda Sadeh cpu_to_le32(src->extent.truncate_seq); 4013d14c5d2SYehuda Sadeh break; 4023d14c5d2SYehuda Sadeh case CEPH_OSD_OP_CALL: 40395e072ebSAlex Elder pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS); 40495e072ebSAlex Elder BUG_ON(!pagelist); 40595e072ebSAlex Elder ceph_pagelist_init(pagelist); 40695e072ebSAlex Elder 4073d14c5d2SYehuda Sadeh dst->cls.class_len = src->cls.class_len; 4083d14c5d2SYehuda Sadeh dst->cls.method_len = src->cls.method_len; 4093d14c5d2SYehuda Sadeh dst->cls.indata_len = cpu_to_le32(src->cls.indata_len); 41095e072ebSAlex Elder ceph_pagelist_append(pagelist, src->cls.class_name, 4113d14c5d2SYehuda Sadeh src->cls.class_len); 41295e072ebSAlex Elder ceph_pagelist_append(pagelist, src->cls.method_name, 4133d14c5d2SYehuda Sadeh src->cls.method_len); 41495e072ebSAlex Elder ceph_pagelist_append(pagelist, src->cls.indata, 4153d14c5d2SYehuda Sadeh src->cls.indata_len); 41695e072ebSAlex Elder 41795e072ebSAlex Elder req->r_data_out.type = CEPH_OSD_DATA_TYPE_PAGELIST; 41895e072ebSAlex Elder req->r_data_out.pagelist = pagelist; 41995e072ebSAlex Elder out_data_len = pagelist->length; 4203d14c5d2SYehuda Sadeh break; 4213d14c5d2SYehuda Sadeh case CEPH_OSD_OP_STARTSYNC: 4223d14c5d2SYehuda Sadeh break; 423a40c4f10SYehuda Sadeh case CEPH_OSD_OP_NOTIFY_ACK: 424a40c4f10SYehuda Sadeh case CEPH_OSD_OP_WATCH: 425a40c4f10SYehuda Sadeh dst->watch.cookie = cpu_to_le64(src->watch.cookie); 426a40c4f10SYehuda Sadeh dst->watch.ver = cpu_to_le64(src->watch.ver); 427a40c4f10SYehuda Sadeh dst->watch.flag = src->watch.flag; 428a40c4f10SYehuda Sadeh break; 4293d14c5d2SYehuda Sadeh default: 4304c46459cSAlex Elder pr_err("unsupported osd opcode %s\n", 4318f63ca2dSAlex Elder ceph_osd_op_name(src->op)); 4324c46459cSAlex Elder WARN_ON(1); 433a8dd0a37SAlex Elder 434a8dd0a37SAlex Elder return 0; 4353d14c5d2SYehuda Sadeh } 436a8dd0a37SAlex Elder dst->op = cpu_to_le16(src->op); 4373d14c5d2SYehuda Sadeh dst->payload_len = cpu_to_le32(src->payload_len); 438175face2SAlex Elder 439175face2SAlex Elder return out_data_len; 4403d14c5d2SYehuda Sadeh } 4413d14c5d2SYehuda Sadeh 4423d14c5d2SYehuda Sadeh /* 4433d14c5d2SYehuda Sadeh * build new request AND message 4443d14c5d2SYehuda Sadeh * 4453d14c5d2SYehuda Sadeh */ 4463d14c5d2SYehuda Sadeh void ceph_osdc_build_request(struct ceph_osd_request *req, 447175face2SAlex Elder u64 off, unsigned int num_ops, 4483d14c5d2SYehuda Sadeh struct ceph_osd_req_op *src_ops, 4494d6b250bSAlex Elder struct ceph_snap_context *snapc, u64 snap_id, 450af77f26cSAlex Elder struct timespec *mtime) 4513d14c5d2SYehuda Sadeh { 4523d14c5d2SYehuda Sadeh struct ceph_msg *msg = req->r_request; 4533d14c5d2SYehuda Sadeh struct ceph_osd_req_op *src_op; 4543d14c5d2SYehuda Sadeh void *p; 4551b83bef2SSage Weil size_t msg_size; 4563d14c5d2SYehuda Sadeh int flags = req->r_flags; 457f44246e3SAlex Elder u64 data_len; 4583d14c5d2SYehuda Sadeh int i; 4593d14c5d2SYehuda Sadeh 4601b83bef2SSage Weil req->r_num_ops = num_ops; 4611b83bef2SSage Weil req->r_snapid = snap_id; 4623d14c5d2SYehuda Sadeh req->r_snapc = ceph_get_snap_context(snapc); 4633d14c5d2SYehuda Sadeh 4641b83bef2SSage Weil /* encode request */ 4651b83bef2SSage Weil msg->hdr.version = cpu_to_le16(4); 4663d14c5d2SYehuda Sadeh 4671b83bef2SSage Weil p = msg->front.iov_base; 4681b83bef2SSage Weil ceph_encode_32(&p, 1); /* client_inc is always 1 */ 4691b83bef2SSage Weil req->r_request_osdmap_epoch = p; 4701b83bef2SSage Weil p += 4; 4711b83bef2SSage Weil req->r_request_flags = p; 4721b83bef2SSage Weil p += 4; 4731b83bef2SSage Weil if (req->r_flags & CEPH_OSD_FLAG_WRITE) 4741b83bef2SSage Weil ceph_encode_timespec(p, mtime); 4751b83bef2SSage Weil p += sizeof(struct ceph_timespec); 4761b83bef2SSage Weil req->r_request_reassert_version = p; 4771b83bef2SSage Weil p += sizeof(struct ceph_eversion); /* will get filled in */ 4781b83bef2SSage Weil 4791b83bef2SSage Weil /* oloc */ 4801b83bef2SSage Weil ceph_encode_8(&p, 4); 4811b83bef2SSage Weil ceph_encode_8(&p, 4); 4821b83bef2SSage Weil ceph_encode_32(&p, 8 + 4 + 4); 4831b83bef2SSage Weil req->r_request_pool = p; 4841b83bef2SSage Weil p += 8; 4851b83bef2SSage Weil ceph_encode_32(&p, -1); /* preferred */ 4861b83bef2SSage Weil ceph_encode_32(&p, 0); /* key len */ 4871b83bef2SSage Weil 4881b83bef2SSage Weil ceph_encode_8(&p, 1); 4891b83bef2SSage Weil req->r_request_pgid = p; 4901b83bef2SSage Weil p += 8 + 4; 4911b83bef2SSage Weil ceph_encode_32(&p, -1); /* preferred */ 4921b83bef2SSage Weil 4931b83bef2SSage Weil /* oid */ 4941b83bef2SSage Weil ceph_encode_32(&p, req->r_oid_len); 495af77f26cSAlex Elder memcpy(p, req->r_oid, req->r_oid_len); 4961b83bef2SSage Weil dout("oid '%.*s' len %d\n", req->r_oid_len, req->r_oid, req->r_oid_len); 497af77f26cSAlex Elder p += req->r_oid_len; 4983d14c5d2SYehuda Sadeh 499175face2SAlex Elder /* ops--can imply data */ 5001b83bef2SSage Weil ceph_encode_16(&p, num_ops); 5013d14c5d2SYehuda Sadeh src_op = src_ops; 5021b83bef2SSage Weil req->r_request_ops = p; 503175face2SAlex Elder data_len = 0; 5041b83bef2SSage Weil for (i = 0; i < num_ops; i++, src_op++) { 505175face2SAlex Elder data_len += osd_req_encode_op(req, p, src_op); 5061b83bef2SSage Weil p += sizeof(struct ceph_osd_op); 5071b83bef2SSage Weil } 5083d14c5d2SYehuda Sadeh 5091b83bef2SSage Weil /* snaps */ 5101b83bef2SSage Weil ceph_encode_64(&p, req->r_snapid); 5111b83bef2SSage Weil ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0); 5121b83bef2SSage Weil ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0); 5131b83bef2SSage Weil if (req->r_snapc) { 5143d14c5d2SYehuda Sadeh for (i = 0; i < snapc->num_snaps; i++) { 5151b83bef2SSage Weil ceph_encode_64(&p, req->r_snapc->snaps[i]); 5163d14c5d2SYehuda Sadeh } 5173d14c5d2SYehuda Sadeh } 5183d14c5d2SYehuda Sadeh 5191b83bef2SSage Weil req->r_request_attempts = p; 5201b83bef2SSage Weil p += 4; 5211b83bef2SSage Weil 522175face2SAlex Elder /* data */ 5230baa1bd9SAlex Elder if (flags & CEPH_OSD_FLAG_WRITE) { 5240baa1bd9SAlex Elder u16 data_off; 5250baa1bd9SAlex Elder 5260baa1bd9SAlex Elder /* 5270baa1bd9SAlex Elder * The header "data_off" is a hint to the receiver 5280baa1bd9SAlex Elder * allowing it to align received data into its 5290baa1bd9SAlex Elder * buffers such that there's no need to re-copy 5300baa1bd9SAlex Elder * it before writing it to disk (direct I/O). 5310baa1bd9SAlex Elder */ 5320baa1bd9SAlex Elder data_off = (u16) (off & 0xffff); 5330baa1bd9SAlex Elder req->r_request->hdr.data_off = cpu_to_le16(data_off); 5340baa1bd9SAlex Elder } 535f44246e3SAlex Elder req->r_request->hdr.data_len = cpu_to_le32(data_len); 536c5c6b19dSSage Weil 5373d14c5d2SYehuda Sadeh BUG_ON(p > msg->front.iov_base + msg->front.iov_len); 5383d14c5d2SYehuda Sadeh msg_size = p - msg->front.iov_base; 5393d14c5d2SYehuda Sadeh msg->front.iov_len = msg_size; 5403d14c5d2SYehuda Sadeh msg->hdr.front_len = cpu_to_le32(msg_size); 5411b83bef2SSage Weil 5421b83bef2SSage Weil dout("build_request msg_size was %d num_ops %d\n", (int)msg_size, 5431b83bef2SSage Weil num_ops); 5443d14c5d2SYehuda Sadeh return; 5453d14c5d2SYehuda Sadeh } 5463d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_build_request); 5473d14c5d2SYehuda Sadeh 5483d14c5d2SYehuda Sadeh /* 5493d14c5d2SYehuda Sadeh * build new request AND message, calculate layout, and adjust file 5503d14c5d2SYehuda Sadeh * extent as needed. 5513d14c5d2SYehuda Sadeh * 5523d14c5d2SYehuda Sadeh * if the file was recently truncated, we include information about its 5533d14c5d2SYehuda Sadeh * old and new size so that the object can be updated appropriately. (we 5543d14c5d2SYehuda Sadeh * avoid synchronously deleting truncated objects because it's slow.) 5553d14c5d2SYehuda Sadeh * 5563d14c5d2SYehuda Sadeh * if @do_sync, include a 'startsync' command so that the osd will flush 5573d14c5d2SYehuda Sadeh * data quickly. 5583d14c5d2SYehuda Sadeh */ 5593d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc, 5603d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 5613d14c5d2SYehuda Sadeh struct ceph_vino vino, 5623d14c5d2SYehuda Sadeh u64 off, u64 *plen, 5633d14c5d2SYehuda Sadeh int opcode, int flags, 5643d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 5653d14c5d2SYehuda Sadeh int do_sync, 5663d14c5d2SYehuda Sadeh u32 truncate_seq, 5673d14c5d2SYehuda Sadeh u64 truncate_size, 5683d14c5d2SYehuda Sadeh struct timespec *mtime, 569153e5167SAlex Elder bool use_mempool) 5703d14c5d2SYehuda Sadeh { 571ae7ca4a3SAlex Elder struct ceph_osd_req_op ops[2]; 5723d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 573ae7ca4a3SAlex Elder unsigned int num_op = 1; 57475d1c941SAlex Elder u64 objnum = 0; 57575d1c941SAlex Elder u64 objoff = 0; 57675d1c941SAlex Elder u64 objlen = 0; 5776816282dSSage Weil int r; 5783d14c5d2SYehuda Sadeh 579ae7ca4a3SAlex Elder memset(&ops, 0, sizeof ops); 580ae7ca4a3SAlex Elder 5813d14c5d2SYehuda Sadeh ops[0].op = opcode; 5823d14c5d2SYehuda Sadeh ops[0].extent.truncate_seq = truncate_seq; 5833d14c5d2SYehuda Sadeh ops[0].extent.truncate_size = truncate_size; 5843d14c5d2SYehuda Sadeh 5853d14c5d2SYehuda Sadeh if (do_sync) { 5863d14c5d2SYehuda Sadeh ops[1].op = CEPH_OSD_OP_STARTSYNC; 587ae7ca4a3SAlex Elder num_op++; 588ae7ca4a3SAlex Elder } 5893d14c5d2SYehuda Sadeh 590ae7ca4a3SAlex Elder req = ceph_osdc_alloc_request(osdc, snapc, num_op, use_mempool, 591ae7ca4a3SAlex Elder GFP_NOFS); 5924ad12621SSage Weil if (!req) 5936816282dSSage Weil return ERR_PTR(-ENOMEM); 594d178a9e7SAlex Elder req->r_flags = flags; 5953d14c5d2SYehuda Sadeh 5963d14c5d2SYehuda Sadeh /* calculate max write size */ 59775d1c941SAlex Elder r = calc_layout(layout, off, plen, ops, &objnum, &objoff, &objlen); 5983ff5f385SAlex Elder if (r < 0) { 5993ff5f385SAlex Elder ceph_osdc_put_request(req); 6006816282dSSage Weil return ERR_PTR(r); 6013ff5f385SAlex Elder } 6023d14c5d2SYehuda Sadeh req->r_file_layout = *layout; /* keep a copy */ 6033d14c5d2SYehuda Sadeh 60475d1c941SAlex Elder snprintf(req->r_oid, sizeof(req->r_oid), "%llx.%08llx", 60575d1c941SAlex Elder vino.ino, objnum); 606dbe0fc41SAlex Elder req->r_oid_len = strlen(req->r_oid); 607dbe0fc41SAlex Elder 608175face2SAlex Elder ceph_osdc_build_request(req, off, num_op, ops, 609ae7ca4a3SAlex Elder snapc, vino.snap, mtime); 6103d14c5d2SYehuda Sadeh 6113d14c5d2SYehuda Sadeh return req; 6123d14c5d2SYehuda Sadeh } 6133d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_new_request); 6143d14c5d2SYehuda Sadeh 6153d14c5d2SYehuda Sadeh /* 6163d14c5d2SYehuda Sadeh * We keep osd requests in an rbtree, sorted by ->r_tid. 6173d14c5d2SYehuda Sadeh */ 6183d14c5d2SYehuda Sadeh static void __insert_request(struct ceph_osd_client *osdc, 6193d14c5d2SYehuda Sadeh struct ceph_osd_request *new) 6203d14c5d2SYehuda Sadeh { 6213d14c5d2SYehuda Sadeh struct rb_node **p = &osdc->requests.rb_node; 6223d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 6233d14c5d2SYehuda Sadeh struct ceph_osd_request *req = NULL; 6243d14c5d2SYehuda Sadeh 6253d14c5d2SYehuda Sadeh while (*p) { 6263d14c5d2SYehuda Sadeh parent = *p; 6273d14c5d2SYehuda Sadeh req = rb_entry(parent, struct ceph_osd_request, r_node); 6283d14c5d2SYehuda Sadeh if (new->r_tid < req->r_tid) 6293d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 6303d14c5d2SYehuda Sadeh else if (new->r_tid > req->r_tid) 6313d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 6323d14c5d2SYehuda Sadeh else 6333d14c5d2SYehuda Sadeh BUG(); 6343d14c5d2SYehuda Sadeh } 6353d14c5d2SYehuda Sadeh 6363d14c5d2SYehuda Sadeh rb_link_node(&new->r_node, parent, p); 6373d14c5d2SYehuda Sadeh rb_insert_color(&new->r_node, &osdc->requests); 6383d14c5d2SYehuda Sadeh } 6393d14c5d2SYehuda Sadeh 6403d14c5d2SYehuda Sadeh static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc, 6413d14c5d2SYehuda Sadeh u64 tid) 6423d14c5d2SYehuda Sadeh { 6433d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 6443d14c5d2SYehuda Sadeh struct rb_node *n = osdc->requests.rb_node; 6453d14c5d2SYehuda Sadeh 6463d14c5d2SYehuda Sadeh while (n) { 6473d14c5d2SYehuda Sadeh req = rb_entry(n, struct ceph_osd_request, r_node); 6483d14c5d2SYehuda Sadeh if (tid < req->r_tid) 6493d14c5d2SYehuda Sadeh n = n->rb_left; 6503d14c5d2SYehuda Sadeh else if (tid > req->r_tid) 6513d14c5d2SYehuda Sadeh n = n->rb_right; 6523d14c5d2SYehuda Sadeh else 6533d14c5d2SYehuda Sadeh return req; 6543d14c5d2SYehuda Sadeh } 6553d14c5d2SYehuda Sadeh return NULL; 6563d14c5d2SYehuda Sadeh } 6573d14c5d2SYehuda Sadeh 6583d14c5d2SYehuda Sadeh static struct ceph_osd_request * 6593d14c5d2SYehuda Sadeh __lookup_request_ge(struct ceph_osd_client *osdc, 6603d14c5d2SYehuda Sadeh u64 tid) 6613d14c5d2SYehuda Sadeh { 6623d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 6633d14c5d2SYehuda Sadeh struct rb_node *n = osdc->requests.rb_node; 6643d14c5d2SYehuda Sadeh 6653d14c5d2SYehuda Sadeh while (n) { 6663d14c5d2SYehuda Sadeh req = rb_entry(n, struct ceph_osd_request, r_node); 6673d14c5d2SYehuda Sadeh if (tid < req->r_tid) { 6683d14c5d2SYehuda Sadeh if (!n->rb_left) 6693d14c5d2SYehuda Sadeh return req; 6703d14c5d2SYehuda Sadeh n = n->rb_left; 6713d14c5d2SYehuda Sadeh } else if (tid > req->r_tid) { 6723d14c5d2SYehuda Sadeh n = n->rb_right; 6733d14c5d2SYehuda Sadeh } else { 6743d14c5d2SYehuda Sadeh return req; 6753d14c5d2SYehuda Sadeh } 6763d14c5d2SYehuda Sadeh } 6773d14c5d2SYehuda Sadeh return NULL; 6783d14c5d2SYehuda Sadeh } 6793d14c5d2SYehuda Sadeh 6806f6c7006SSage Weil /* 6816f6c7006SSage Weil * Resubmit requests pending on the given osd. 6826f6c7006SSage Weil */ 6836f6c7006SSage Weil static void __kick_osd_requests(struct ceph_osd_client *osdc, 6846f6c7006SSage Weil struct ceph_osd *osd) 6856f6c7006SSage Weil { 686a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 687e02493c0SAlex Elder LIST_HEAD(resend); 6886f6c7006SSage Weil int err; 6896f6c7006SSage Weil 6906f6c7006SSage Weil dout("__kick_osd_requests osd%d\n", osd->o_osd); 6916f6c7006SSage Weil err = __reset_osd(osdc, osd); 692685a7555SAlex Elder if (err) 6936f6c7006SSage Weil return; 694e02493c0SAlex Elder /* 695e02493c0SAlex Elder * Build up a list of requests to resend by traversing the 696e02493c0SAlex Elder * osd's list of requests. Requests for a given object are 697e02493c0SAlex Elder * sent in tid order, and that is also the order they're 698e02493c0SAlex Elder * kept on this list. Therefore all requests that are in 699e02493c0SAlex Elder * flight will be found first, followed by all requests that 700e02493c0SAlex Elder * have not yet been sent. And to resend requests while 701e02493c0SAlex Elder * preserving this order we will want to put any sent 702e02493c0SAlex Elder * requests back on the front of the osd client's unsent 703e02493c0SAlex Elder * list. 704e02493c0SAlex Elder * 705e02493c0SAlex Elder * So we build a separate ordered list of already-sent 706e02493c0SAlex Elder * requests for the affected osd and splice it onto the 707e02493c0SAlex Elder * front of the osd client's unsent list. Once we've seen a 708e02493c0SAlex Elder * request that has not yet been sent we're done. Those 709e02493c0SAlex Elder * requests are already sitting right where they belong. 710e02493c0SAlex Elder */ 7116f6c7006SSage Weil list_for_each_entry(req, &osd->o_requests, r_osd_item) { 712e02493c0SAlex Elder if (!req->r_sent) 713e02493c0SAlex Elder break; 714e02493c0SAlex Elder list_move_tail(&req->r_req_lru_item, &resend); 715e02493c0SAlex Elder dout("requeueing %p tid %llu osd%d\n", req, req->r_tid, 7166f6c7006SSage Weil osd->o_osd); 717a40c4f10SYehuda Sadeh if (!req->r_linger) 7186f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 7196f6c7006SSage Weil } 720e02493c0SAlex Elder list_splice(&resend, &osdc->req_unsent); 721a40c4f10SYehuda Sadeh 722e02493c0SAlex Elder /* 723e02493c0SAlex Elder * Linger requests are re-registered before sending, which 724e02493c0SAlex Elder * sets up a new tid for each. We add them to the unsent 725e02493c0SAlex Elder * list at the end to keep things in tid order. 726e02493c0SAlex Elder */ 727a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osd->o_linger_requests, 728a40c4f10SYehuda Sadeh r_linger_osd) { 72977f38e0eSSage Weil /* 73077f38e0eSSage Weil * reregister request prior to unregistering linger so 73177f38e0eSSage Weil * that r_osd is preserved. 73277f38e0eSSage Weil */ 73377f38e0eSSage Weil BUG_ON(!list_empty(&req->r_req_lru_item)); 734a40c4f10SYehuda Sadeh __register_request(osdc, req); 735e02493c0SAlex Elder list_add_tail(&req->r_req_lru_item, &osdc->req_unsent); 736ad885927SAlex Elder list_add_tail(&req->r_osd_item, &req->r_osd->o_requests); 73777f38e0eSSage Weil __unregister_linger_request(osdc, req); 738a40c4f10SYehuda Sadeh dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid, 739a40c4f10SYehuda Sadeh osd->o_osd); 740a40c4f10SYehuda Sadeh } 7416f6c7006SSage Weil } 7426f6c7006SSage Weil 7433d14c5d2SYehuda Sadeh /* 7443d14c5d2SYehuda Sadeh * If the osd connection drops, we need to resubmit all requests. 7453d14c5d2SYehuda Sadeh */ 7463d14c5d2SYehuda Sadeh static void osd_reset(struct ceph_connection *con) 7473d14c5d2SYehuda Sadeh { 7483d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 7493d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 7503d14c5d2SYehuda Sadeh 7513d14c5d2SYehuda Sadeh if (!osd) 7523d14c5d2SYehuda Sadeh return; 7533d14c5d2SYehuda Sadeh dout("osd_reset osd%d\n", osd->o_osd); 7543d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 7553d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 75683aff95eSSage Weil mutex_lock(&osdc->request_mutex); 75783aff95eSSage Weil __kick_osd_requests(osdc, osd); 758f9d25199SAlex Elder __send_queued(osdc); 75983aff95eSSage Weil mutex_unlock(&osdc->request_mutex); 7603d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 7613d14c5d2SYehuda Sadeh } 7623d14c5d2SYehuda Sadeh 7633d14c5d2SYehuda Sadeh /* 7643d14c5d2SYehuda Sadeh * Track open sessions with osds. 7653d14c5d2SYehuda Sadeh */ 766e10006f8SAlex Elder static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum) 7673d14c5d2SYehuda Sadeh { 7683d14c5d2SYehuda Sadeh struct ceph_osd *osd; 7693d14c5d2SYehuda Sadeh 7703d14c5d2SYehuda Sadeh osd = kzalloc(sizeof(*osd), GFP_NOFS); 7713d14c5d2SYehuda Sadeh if (!osd) 7723d14c5d2SYehuda Sadeh return NULL; 7733d14c5d2SYehuda Sadeh 7743d14c5d2SYehuda Sadeh atomic_set(&osd->o_ref, 1); 7753d14c5d2SYehuda Sadeh osd->o_osdc = osdc; 776e10006f8SAlex Elder osd->o_osd = onum; 777f407731dSAlex Elder RB_CLEAR_NODE(&osd->o_node); 7783d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_requests); 779a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osd->o_linger_requests); 7803d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_osd_lru); 7813d14c5d2SYehuda Sadeh osd->o_incarnation = 1; 7823d14c5d2SYehuda Sadeh 783b7a9e5ddSSage Weil ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr); 7843d14c5d2SYehuda Sadeh 7853d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_keepalive_item); 7863d14c5d2SYehuda Sadeh return osd; 7873d14c5d2SYehuda Sadeh } 7883d14c5d2SYehuda Sadeh 7893d14c5d2SYehuda Sadeh static struct ceph_osd *get_osd(struct ceph_osd *osd) 7903d14c5d2SYehuda Sadeh { 7913d14c5d2SYehuda Sadeh if (atomic_inc_not_zero(&osd->o_ref)) { 7923d14c5d2SYehuda Sadeh dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1, 7933d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref)); 7943d14c5d2SYehuda Sadeh return osd; 7953d14c5d2SYehuda Sadeh } else { 7963d14c5d2SYehuda Sadeh dout("get_osd %p FAIL\n", osd); 7973d14c5d2SYehuda Sadeh return NULL; 7983d14c5d2SYehuda Sadeh } 7993d14c5d2SYehuda Sadeh } 8003d14c5d2SYehuda Sadeh 8013d14c5d2SYehuda Sadeh static void put_osd(struct ceph_osd *osd) 8023d14c5d2SYehuda Sadeh { 8033d14c5d2SYehuda Sadeh dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref), 8043d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref) - 1); 805a255651dSAlex Elder if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) { 8063d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth; 8073d14c5d2SYehuda Sadeh 80827859f97SSage Weil ceph_auth_destroy_authorizer(ac, osd->o_auth.authorizer); 8093d14c5d2SYehuda Sadeh kfree(osd); 8103d14c5d2SYehuda Sadeh } 8113d14c5d2SYehuda Sadeh } 8123d14c5d2SYehuda Sadeh 8133d14c5d2SYehuda Sadeh /* 8143d14c5d2SYehuda Sadeh * remove an osd from our map 8153d14c5d2SYehuda Sadeh */ 8163d14c5d2SYehuda Sadeh static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 8173d14c5d2SYehuda Sadeh { 8183d14c5d2SYehuda Sadeh dout("__remove_osd %p\n", osd); 8193d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_requests)); 8203d14c5d2SYehuda Sadeh rb_erase(&osd->o_node, &osdc->osds); 8213d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 8223d14c5d2SYehuda Sadeh ceph_con_close(&osd->o_con); 8233d14c5d2SYehuda Sadeh put_osd(osd); 8243d14c5d2SYehuda Sadeh } 8253d14c5d2SYehuda Sadeh 826aca420bcSSage Weil static void remove_all_osds(struct ceph_osd_client *osdc) 827aca420bcSSage Weil { 828048a9d2dSJiaju Zhang dout("%s %p\n", __func__, osdc); 829aca420bcSSage Weil mutex_lock(&osdc->request_mutex); 830aca420bcSSage Weil while (!RB_EMPTY_ROOT(&osdc->osds)) { 831aca420bcSSage Weil struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds), 832aca420bcSSage Weil struct ceph_osd, o_node); 833aca420bcSSage Weil __remove_osd(osdc, osd); 834aca420bcSSage Weil } 835aca420bcSSage Weil mutex_unlock(&osdc->request_mutex); 836aca420bcSSage Weil } 837aca420bcSSage Weil 8383d14c5d2SYehuda Sadeh static void __move_osd_to_lru(struct ceph_osd_client *osdc, 8393d14c5d2SYehuda Sadeh struct ceph_osd *osd) 8403d14c5d2SYehuda Sadeh { 8413d14c5d2SYehuda Sadeh dout("__move_osd_to_lru %p\n", osd); 8423d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_osd_lru)); 8433d14c5d2SYehuda Sadeh list_add_tail(&osd->o_osd_lru, &osdc->osd_lru); 8443d14c5d2SYehuda Sadeh osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ; 8453d14c5d2SYehuda Sadeh } 8463d14c5d2SYehuda Sadeh 8473d14c5d2SYehuda Sadeh static void __remove_osd_from_lru(struct ceph_osd *osd) 8483d14c5d2SYehuda Sadeh { 8493d14c5d2SYehuda Sadeh dout("__remove_osd_from_lru %p\n", osd); 8503d14c5d2SYehuda Sadeh if (!list_empty(&osd->o_osd_lru)) 8513d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 8523d14c5d2SYehuda Sadeh } 8533d14c5d2SYehuda Sadeh 854aca420bcSSage Weil static void remove_old_osds(struct ceph_osd_client *osdc) 8553d14c5d2SYehuda Sadeh { 8563d14c5d2SYehuda Sadeh struct ceph_osd *osd, *nosd; 8573d14c5d2SYehuda Sadeh 8583d14c5d2SYehuda Sadeh dout("__remove_old_osds %p\n", osdc); 8593d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 8603d14c5d2SYehuda Sadeh list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) { 861aca420bcSSage Weil if (time_before(jiffies, osd->lru_ttl)) 8623d14c5d2SYehuda Sadeh break; 8633d14c5d2SYehuda Sadeh __remove_osd(osdc, osd); 8643d14c5d2SYehuda Sadeh } 8653d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 8663d14c5d2SYehuda Sadeh } 8673d14c5d2SYehuda Sadeh 8683d14c5d2SYehuda Sadeh /* 8693d14c5d2SYehuda Sadeh * reset osd connect 8703d14c5d2SYehuda Sadeh */ 8713d14c5d2SYehuda Sadeh static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 8723d14c5d2SYehuda Sadeh { 873c3acb181SAlex Elder struct ceph_entity_addr *peer_addr; 8743d14c5d2SYehuda Sadeh 8753d14c5d2SYehuda Sadeh dout("__reset_osd %p osd%d\n", osd, osd->o_osd); 876a40c4f10SYehuda Sadeh if (list_empty(&osd->o_requests) && 877a40c4f10SYehuda Sadeh list_empty(&osd->o_linger_requests)) { 8783d14c5d2SYehuda Sadeh __remove_osd(osdc, osd); 879c3acb181SAlex Elder 880c3acb181SAlex Elder return -ENODEV; 881c3acb181SAlex Elder } 882c3acb181SAlex Elder 883c3acb181SAlex Elder peer_addr = &osdc->osdmap->osd_addr[osd->o_osd]; 884c3acb181SAlex Elder if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) && 8853d14c5d2SYehuda Sadeh !ceph_con_opened(&osd->o_con)) { 886c3acb181SAlex Elder struct ceph_osd_request *req; 887c3acb181SAlex Elder 8883d14c5d2SYehuda Sadeh dout(" osd addr hasn't changed and connection never opened," 8893d14c5d2SYehuda Sadeh " letting msgr retry"); 8903d14c5d2SYehuda Sadeh /* touch each r_stamp for handle_timeout()'s benfit */ 8913d14c5d2SYehuda Sadeh list_for_each_entry(req, &osd->o_requests, r_osd_item) 8923d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 893c3acb181SAlex Elder 894c3acb181SAlex Elder return -EAGAIN; 8953d14c5d2SYehuda Sadeh } 896c3acb181SAlex Elder 897c3acb181SAlex Elder ceph_con_close(&osd->o_con); 898c3acb181SAlex Elder ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr); 899c3acb181SAlex Elder osd->o_incarnation++; 900c3acb181SAlex Elder 901c3acb181SAlex Elder return 0; 9023d14c5d2SYehuda Sadeh } 9033d14c5d2SYehuda Sadeh 9043d14c5d2SYehuda Sadeh static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new) 9053d14c5d2SYehuda Sadeh { 9063d14c5d2SYehuda Sadeh struct rb_node **p = &osdc->osds.rb_node; 9073d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 9083d14c5d2SYehuda Sadeh struct ceph_osd *osd = NULL; 9093d14c5d2SYehuda Sadeh 910aca420bcSSage Weil dout("__insert_osd %p osd%d\n", new, new->o_osd); 9113d14c5d2SYehuda Sadeh while (*p) { 9123d14c5d2SYehuda Sadeh parent = *p; 9133d14c5d2SYehuda Sadeh osd = rb_entry(parent, struct ceph_osd, o_node); 9143d14c5d2SYehuda Sadeh if (new->o_osd < osd->o_osd) 9153d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 9163d14c5d2SYehuda Sadeh else if (new->o_osd > osd->o_osd) 9173d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 9183d14c5d2SYehuda Sadeh else 9193d14c5d2SYehuda Sadeh BUG(); 9203d14c5d2SYehuda Sadeh } 9213d14c5d2SYehuda Sadeh 9223d14c5d2SYehuda Sadeh rb_link_node(&new->o_node, parent, p); 9233d14c5d2SYehuda Sadeh rb_insert_color(&new->o_node, &osdc->osds); 9243d14c5d2SYehuda Sadeh } 9253d14c5d2SYehuda Sadeh 9263d14c5d2SYehuda Sadeh static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o) 9273d14c5d2SYehuda Sadeh { 9283d14c5d2SYehuda Sadeh struct ceph_osd *osd; 9293d14c5d2SYehuda Sadeh struct rb_node *n = osdc->osds.rb_node; 9303d14c5d2SYehuda Sadeh 9313d14c5d2SYehuda Sadeh while (n) { 9323d14c5d2SYehuda Sadeh osd = rb_entry(n, struct ceph_osd, o_node); 9333d14c5d2SYehuda Sadeh if (o < osd->o_osd) 9343d14c5d2SYehuda Sadeh n = n->rb_left; 9353d14c5d2SYehuda Sadeh else if (o > osd->o_osd) 9363d14c5d2SYehuda Sadeh n = n->rb_right; 9373d14c5d2SYehuda Sadeh else 9383d14c5d2SYehuda Sadeh return osd; 9393d14c5d2SYehuda Sadeh } 9403d14c5d2SYehuda Sadeh return NULL; 9413d14c5d2SYehuda Sadeh } 9423d14c5d2SYehuda Sadeh 9433d14c5d2SYehuda Sadeh static void __schedule_osd_timeout(struct ceph_osd_client *osdc) 9443d14c5d2SYehuda Sadeh { 9453d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->timeout_work, 9463d14c5d2SYehuda Sadeh osdc->client->options->osd_keepalive_timeout * HZ); 9473d14c5d2SYehuda Sadeh } 9483d14c5d2SYehuda Sadeh 9493d14c5d2SYehuda Sadeh static void __cancel_osd_timeout(struct ceph_osd_client *osdc) 9503d14c5d2SYehuda Sadeh { 9513d14c5d2SYehuda Sadeh cancel_delayed_work(&osdc->timeout_work); 9523d14c5d2SYehuda Sadeh } 9533d14c5d2SYehuda Sadeh 9543d14c5d2SYehuda Sadeh /* 9553d14c5d2SYehuda Sadeh * Register request, assign tid. If this is the first request, set up 9563d14c5d2SYehuda Sadeh * the timeout event. 9573d14c5d2SYehuda Sadeh */ 958a40c4f10SYehuda Sadeh static void __register_request(struct ceph_osd_client *osdc, 9593d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 9603d14c5d2SYehuda Sadeh { 9613d14c5d2SYehuda Sadeh req->r_tid = ++osdc->last_tid; 9623d14c5d2SYehuda Sadeh req->r_request->hdr.tid = cpu_to_le64(req->r_tid); 96377f38e0eSSage Weil dout("__register_request %p tid %lld\n", req, req->r_tid); 9643d14c5d2SYehuda Sadeh __insert_request(osdc, req); 9653d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 9663d14c5d2SYehuda Sadeh osdc->num_requests++; 9673d14c5d2SYehuda Sadeh if (osdc->num_requests == 1) { 9683d14c5d2SYehuda Sadeh dout(" first request, scheduling timeout\n"); 9693d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 9703d14c5d2SYehuda Sadeh } 971a40c4f10SYehuda Sadeh } 972a40c4f10SYehuda Sadeh 9733d14c5d2SYehuda Sadeh /* 9743d14c5d2SYehuda Sadeh * called under osdc->request_mutex 9753d14c5d2SYehuda Sadeh */ 9763d14c5d2SYehuda Sadeh static void __unregister_request(struct ceph_osd_client *osdc, 9773d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 9783d14c5d2SYehuda Sadeh { 97935f9f8a0SSage Weil if (RB_EMPTY_NODE(&req->r_node)) { 98035f9f8a0SSage Weil dout("__unregister_request %p tid %lld not registered\n", 98135f9f8a0SSage Weil req, req->r_tid); 98235f9f8a0SSage Weil return; 98335f9f8a0SSage Weil } 98435f9f8a0SSage Weil 9853d14c5d2SYehuda Sadeh dout("__unregister_request %p tid %lld\n", req, req->r_tid); 9863d14c5d2SYehuda Sadeh rb_erase(&req->r_node, &osdc->requests); 9873d14c5d2SYehuda Sadeh osdc->num_requests--; 9883d14c5d2SYehuda Sadeh 9893d14c5d2SYehuda Sadeh if (req->r_osd) { 9903d14c5d2SYehuda Sadeh /* make sure the original request isn't in flight. */ 9916740a845SAlex Elder ceph_msg_revoke(req->r_request); 9923d14c5d2SYehuda Sadeh 9933d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 994a40c4f10SYehuda Sadeh if (list_empty(&req->r_osd->o_requests) && 995a40c4f10SYehuda Sadeh list_empty(&req->r_osd->o_linger_requests)) { 996a40c4f10SYehuda Sadeh dout("moving osd to %p lru\n", req->r_osd); 9973d14c5d2SYehuda Sadeh __move_osd_to_lru(osdc, req->r_osd); 998a40c4f10SYehuda Sadeh } 999fbdb9190SSage Weil if (list_empty(&req->r_linger_item)) 10003d14c5d2SYehuda Sadeh req->r_osd = NULL; 10013d14c5d2SYehuda Sadeh } 10023d14c5d2SYehuda Sadeh 10037d5f2481SAlex Elder list_del_init(&req->r_req_lru_item); 10043d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 10053d14c5d2SYehuda Sadeh 10063d14c5d2SYehuda Sadeh if (osdc->num_requests == 0) { 10073d14c5d2SYehuda Sadeh dout(" no requests, canceling timeout\n"); 10083d14c5d2SYehuda Sadeh __cancel_osd_timeout(osdc); 10093d14c5d2SYehuda Sadeh } 10103d14c5d2SYehuda Sadeh } 10113d14c5d2SYehuda Sadeh 10123d14c5d2SYehuda Sadeh /* 10133d14c5d2SYehuda Sadeh * Cancel a previously queued request message 10143d14c5d2SYehuda Sadeh */ 10153d14c5d2SYehuda Sadeh static void __cancel_request(struct ceph_osd_request *req) 10163d14c5d2SYehuda Sadeh { 10173d14c5d2SYehuda Sadeh if (req->r_sent && req->r_osd) { 10186740a845SAlex Elder ceph_msg_revoke(req->r_request); 10193d14c5d2SYehuda Sadeh req->r_sent = 0; 10203d14c5d2SYehuda Sadeh } 10213d14c5d2SYehuda Sadeh } 10223d14c5d2SYehuda Sadeh 1023a40c4f10SYehuda Sadeh static void __register_linger_request(struct ceph_osd_client *osdc, 1024a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1025a40c4f10SYehuda Sadeh { 1026a40c4f10SYehuda Sadeh dout("__register_linger_request %p\n", req); 1027a40c4f10SYehuda Sadeh list_add_tail(&req->r_linger_item, &osdc->req_linger); 10286194ea89SSage Weil if (req->r_osd) 10296194ea89SSage Weil list_add_tail(&req->r_linger_osd, 10306194ea89SSage Weil &req->r_osd->o_linger_requests); 1031a40c4f10SYehuda Sadeh } 1032a40c4f10SYehuda Sadeh 1033a40c4f10SYehuda Sadeh static void __unregister_linger_request(struct ceph_osd_client *osdc, 1034a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1035a40c4f10SYehuda Sadeh { 1036a40c4f10SYehuda Sadeh dout("__unregister_linger_request %p\n", req); 1037a40c4f10SYehuda Sadeh list_del_init(&req->r_linger_item); 103861c74035SAlex Elder if (req->r_osd) { 1039a40c4f10SYehuda Sadeh list_del_init(&req->r_linger_osd); 1040a40c4f10SYehuda Sadeh 1041a40c4f10SYehuda Sadeh if (list_empty(&req->r_osd->o_requests) && 1042a40c4f10SYehuda Sadeh list_empty(&req->r_osd->o_linger_requests)) { 1043a40c4f10SYehuda Sadeh dout("moving osd to %p lru\n", req->r_osd); 1044a40c4f10SYehuda Sadeh __move_osd_to_lru(osdc, req->r_osd); 1045a40c4f10SYehuda Sadeh } 1046fbdb9190SSage Weil if (list_empty(&req->r_osd_item)) 1047a40c4f10SYehuda Sadeh req->r_osd = NULL; 1048a40c4f10SYehuda Sadeh } 1049a40c4f10SYehuda Sadeh } 1050a40c4f10SYehuda Sadeh 1051a40c4f10SYehuda Sadeh void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc, 1052a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1053a40c4f10SYehuda Sadeh { 1054a40c4f10SYehuda Sadeh mutex_lock(&osdc->request_mutex); 1055a40c4f10SYehuda Sadeh if (req->r_linger) { 1056a40c4f10SYehuda Sadeh __unregister_linger_request(osdc, req); 1057a40c4f10SYehuda Sadeh ceph_osdc_put_request(req); 1058a40c4f10SYehuda Sadeh } 1059a40c4f10SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 1060a40c4f10SYehuda Sadeh } 1061a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_unregister_linger_request); 1062a40c4f10SYehuda Sadeh 1063a40c4f10SYehuda Sadeh void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc, 1064a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1065a40c4f10SYehuda Sadeh { 1066a40c4f10SYehuda Sadeh if (!req->r_linger) { 1067a40c4f10SYehuda Sadeh dout("set_request_linger %p\n", req); 1068a40c4f10SYehuda Sadeh req->r_linger = 1; 1069a40c4f10SYehuda Sadeh /* 1070a40c4f10SYehuda Sadeh * caller is now responsible for calling 1071a40c4f10SYehuda Sadeh * unregister_linger_request 1072a40c4f10SYehuda Sadeh */ 1073a40c4f10SYehuda Sadeh ceph_osdc_get_request(req); 1074a40c4f10SYehuda Sadeh } 1075a40c4f10SYehuda Sadeh } 1076a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_set_request_linger); 1077a40c4f10SYehuda Sadeh 10783d14c5d2SYehuda Sadeh /* 10793d14c5d2SYehuda Sadeh * Pick an osd (the first 'up' osd in the pg), allocate the osd struct 10803d14c5d2SYehuda Sadeh * (as needed), and set the request r_osd appropriately. If there is 108125985edcSLucas De Marchi * no up osd, set r_osd to NULL. Move the request to the appropriate list 10826f6c7006SSage Weil * (unsent, homeless) or leave on in-flight lru. 10833d14c5d2SYehuda Sadeh * 10843d14c5d2SYehuda Sadeh * Return 0 if unchanged, 1 if changed, or negative on error. 10853d14c5d2SYehuda Sadeh * 10863d14c5d2SYehuda Sadeh * Caller should hold map_sem for read and request_mutex. 10873d14c5d2SYehuda Sadeh */ 10886f6c7006SSage Weil static int __map_request(struct ceph_osd_client *osdc, 108938d6453cSSage Weil struct ceph_osd_request *req, int force_resend) 10903d14c5d2SYehuda Sadeh { 10915b191d99SSage Weil struct ceph_pg pgid; 10923d14c5d2SYehuda Sadeh int acting[CEPH_PG_MAX_SIZE]; 10933d14c5d2SYehuda Sadeh int o = -1, num = 0; 10943d14c5d2SYehuda Sadeh int err; 10953d14c5d2SYehuda Sadeh 10966f6c7006SSage Weil dout("map_request %p tid %lld\n", req, req->r_tid); 109741766f87SAlex Elder err = ceph_calc_ceph_pg(&pgid, req->r_oid, osdc->osdmap, 109841766f87SAlex Elder ceph_file_layout_pg_pool(req->r_file_layout)); 10996f6c7006SSage Weil if (err) { 11006f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 11013d14c5d2SYehuda Sadeh return err; 11026f6c7006SSage Weil } 11033d14c5d2SYehuda Sadeh req->r_pgid = pgid; 11043d14c5d2SYehuda Sadeh 11053d14c5d2SYehuda Sadeh err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting); 11063d14c5d2SYehuda Sadeh if (err > 0) { 11073d14c5d2SYehuda Sadeh o = acting[0]; 11083d14c5d2SYehuda Sadeh num = err; 11093d14c5d2SYehuda Sadeh } 11103d14c5d2SYehuda Sadeh 111138d6453cSSage Weil if ((!force_resend && 111238d6453cSSage Weil req->r_osd && req->r_osd->o_osd == o && 11133d14c5d2SYehuda Sadeh req->r_sent >= req->r_osd->o_incarnation && 11143d14c5d2SYehuda Sadeh req->r_num_pg_osds == num && 11153d14c5d2SYehuda Sadeh memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) || 11163d14c5d2SYehuda Sadeh (req->r_osd == NULL && o == -1)) 11173d14c5d2SYehuda Sadeh return 0; /* no change */ 11183d14c5d2SYehuda Sadeh 11195b191d99SSage Weil dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n", 11205b191d99SSage Weil req->r_tid, pgid.pool, pgid.seed, o, 11213d14c5d2SYehuda Sadeh req->r_osd ? req->r_osd->o_osd : -1); 11223d14c5d2SYehuda Sadeh 11233d14c5d2SYehuda Sadeh /* record full pg acting set */ 11243d14c5d2SYehuda Sadeh memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num); 11253d14c5d2SYehuda Sadeh req->r_num_pg_osds = num; 11263d14c5d2SYehuda Sadeh 11273d14c5d2SYehuda Sadeh if (req->r_osd) { 11283d14c5d2SYehuda Sadeh __cancel_request(req); 11293d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 11303d14c5d2SYehuda Sadeh req->r_osd = NULL; 11313d14c5d2SYehuda Sadeh } 11323d14c5d2SYehuda Sadeh 11333d14c5d2SYehuda Sadeh req->r_osd = __lookup_osd(osdc, o); 11343d14c5d2SYehuda Sadeh if (!req->r_osd && o >= 0) { 11353d14c5d2SYehuda Sadeh err = -ENOMEM; 1136e10006f8SAlex Elder req->r_osd = create_osd(osdc, o); 11376f6c7006SSage Weil if (!req->r_osd) { 11386f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 11393d14c5d2SYehuda Sadeh goto out; 11406f6c7006SSage Weil } 11413d14c5d2SYehuda Sadeh 11426f6c7006SSage Weil dout("map_request osd %p is osd%d\n", req->r_osd, o); 11433d14c5d2SYehuda Sadeh __insert_osd(osdc, req->r_osd); 11443d14c5d2SYehuda Sadeh 1145b7a9e5ddSSage Weil ceph_con_open(&req->r_osd->o_con, 1146b7a9e5ddSSage Weil CEPH_ENTITY_TYPE_OSD, o, 1147b7a9e5ddSSage Weil &osdc->osdmap->osd_addr[o]); 11483d14c5d2SYehuda Sadeh } 11493d14c5d2SYehuda Sadeh 11503d14c5d2SYehuda Sadeh if (req->r_osd) { 11513d14c5d2SYehuda Sadeh __remove_osd_from_lru(req->r_osd); 1152ad885927SAlex Elder list_add_tail(&req->r_osd_item, &req->r_osd->o_requests); 1153ad885927SAlex Elder list_move_tail(&req->r_req_lru_item, &osdc->req_unsent); 11546f6c7006SSage Weil } else { 1155ad885927SAlex Elder list_move_tail(&req->r_req_lru_item, &osdc->req_notarget); 11563d14c5d2SYehuda Sadeh } 11573d14c5d2SYehuda Sadeh err = 1; /* osd or pg changed */ 11583d14c5d2SYehuda Sadeh 11593d14c5d2SYehuda Sadeh out: 11603d14c5d2SYehuda Sadeh return err; 11613d14c5d2SYehuda Sadeh } 11623d14c5d2SYehuda Sadeh 11633d14c5d2SYehuda Sadeh /* 11643d14c5d2SYehuda Sadeh * caller should hold map_sem (for read) and request_mutex 11653d14c5d2SYehuda Sadeh */ 116656e925b6SSage Weil static void __send_request(struct ceph_osd_client *osdc, 11673d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 11683d14c5d2SYehuda Sadeh { 11691b83bef2SSage Weil void *p; 11703d14c5d2SYehuda Sadeh 11711b83bef2SSage Weil dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n", 11721b83bef2SSage Weil req, req->r_tid, req->r_osd->o_osd, req->r_flags, 11731b83bef2SSage Weil (unsigned long long)req->r_pgid.pool, req->r_pgid.seed); 11743d14c5d2SYehuda Sadeh 11751b83bef2SSage Weil /* fill in message content that changes each time we send it */ 11761b83bef2SSage Weil put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch); 11771b83bef2SSage Weil put_unaligned_le32(req->r_flags, req->r_request_flags); 11781b83bef2SSage Weil put_unaligned_le64(req->r_pgid.pool, req->r_request_pool); 11791b83bef2SSage Weil p = req->r_request_pgid; 11801b83bef2SSage Weil ceph_encode_64(&p, req->r_pgid.pool); 11811b83bef2SSage Weil ceph_encode_32(&p, req->r_pgid.seed); 11821b83bef2SSage Weil put_unaligned_le64(1, req->r_request_attempts); /* FIXME */ 11831b83bef2SSage Weil memcpy(req->r_request_reassert_version, &req->r_reassert_version, 11841b83bef2SSage Weil sizeof(req->r_reassert_version)); 11852169aea6SSage Weil 11863d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 11873d14c5d2SYehuda Sadeh list_move_tail(&req->r_req_lru_item, &osdc->req_lru); 11883d14c5d2SYehuda Sadeh 11893d14c5d2SYehuda Sadeh ceph_msg_get(req->r_request); /* send consumes a ref */ 11903d14c5d2SYehuda Sadeh ceph_con_send(&req->r_osd->o_con, req->r_request); 11913d14c5d2SYehuda Sadeh req->r_sent = req->r_osd->o_incarnation; 11923d14c5d2SYehuda Sadeh } 11933d14c5d2SYehuda Sadeh 11943d14c5d2SYehuda Sadeh /* 11956f6c7006SSage Weil * Send any requests in the queue (req_unsent). 11966f6c7006SSage Weil */ 1197f9d25199SAlex Elder static void __send_queued(struct ceph_osd_client *osdc) 11986f6c7006SSage Weil { 11996f6c7006SSage Weil struct ceph_osd_request *req, *tmp; 12006f6c7006SSage Weil 1201f9d25199SAlex Elder dout("__send_queued\n"); 1202f9d25199SAlex Elder list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) 12036f6c7006SSage Weil __send_request(osdc, req); 12046f6c7006SSage Weil } 12056f6c7006SSage Weil 12066f6c7006SSage Weil /* 12073d14c5d2SYehuda Sadeh * Timeout callback, called every N seconds when 1 or more osd 12083d14c5d2SYehuda Sadeh * requests has been active for more than N seconds. When this 12093d14c5d2SYehuda Sadeh * happens, we ping all OSDs with requests who have timed out to 12103d14c5d2SYehuda Sadeh * ensure any communications channel reset is detected. Reset the 12113d14c5d2SYehuda Sadeh * request timeouts another N seconds in the future as we go. 12123d14c5d2SYehuda Sadeh * Reschedule the timeout event another N seconds in future (unless 12133d14c5d2SYehuda Sadeh * there are no open requests). 12143d14c5d2SYehuda Sadeh */ 12153d14c5d2SYehuda Sadeh static void handle_timeout(struct work_struct *work) 12163d14c5d2SYehuda Sadeh { 12173d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 12183d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, timeout_work.work); 121983aff95eSSage Weil struct ceph_osd_request *req; 12203d14c5d2SYehuda Sadeh struct ceph_osd *osd; 12213d14c5d2SYehuda Sadeh unsigned long keepalive = 12223d14c5d2SYehuda Sadeh osdc->client->options->osd_keepalive_timeout * HZ; 12233d14c5d2SYehuda Sadeh struct list_head slow_osds; 12243d14c5d2SYehuda Sadeh dout("timeout\n"); 12253d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 12263d14c5d2SYehuda Sadeh 12273d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 12283d14c5d2SYehuda Sadeh 12293d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 12303d14c5d2SYehuda Sadeh 12313d14c5d2SYehuda Sadeh /* 12323d14c5d2SYehuda Sadeh * ping osds that are a bit slow. this ensures that if there 12333d14c5d2SYehuda Sadeh * is a break in the TCP connection we will notice, and reopen 12343d14c5d2SYehuda Sadeh * a connection with that osd (from the fault callback). 12353d14c5d2SYehuda Sadeh */ 12363d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&slow_osds); 12373d14c5d2SYehuda Sadeh list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) { 12383d14c5d2SYehuda Sadeh if (time_before(jiffies, req->r_stamp + keepalive)) 12393d14c5d2SYehuda Sadeh break; 12403d14c5d2SYehuda Sadeh 12413d14c5d2SYehuda Sadeh osd = req->r_osd; 12423d14c5d2SYehuda Sadeh BUG_ON(!osd); 12433d14c5d2SYehuda Sadeh dout(" tid %llu is slow, will send keepalive on osd%d\n", 12443d14c5d2SYehuda Sadeh req->r_tid, osd->o_osd); 12453d14c5d2SYehuda Sadeh list_move_tail(&osd->o_keepalive_item, &slow_osds); 12463d14c5d2SYehuda Sadeh } 12473d14c5d2SYehuda Sadeh while (!list_empty(&slow_osds)) { 12483d14c5d2SYehuda Sadeh osd = list_entry(slow_osds.next, struct ceph_osd, 12493d14c5d2SYehuda Sadeh o_keepalive_item); 12503d14c5d2SYehuda Sadeh list_del_init(&osd->o_keepalive_item); 12513d14c5d2SYehuda Sadeh ceph_con_keepalive(&osd->o_con); 12523d14c5d2SYehuda Sadeh } 12533d14c5d2SYehuda Sadeh 12543d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 1255f9d25199SAlex Elder __send_queued(osdc); 12563d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 12573d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 12583d14c5d2SYehuda Sadeh } 12593d14c5d2SYehuda Sadeh 12603d14c5d2SYehuda Sadeh static void handle_osds_timeout(struct work_struct *work) 12613d14c5d2SYehuda Sadeh { 12623d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 12633d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, 12643d14c5d2SYehuda Sadeh osds_timeout_work.work); 12653d14c5d2SYehuda Sadeh unsigned long delay = 12663d14c5d2SYehuda Sadeh osdc->client->options->osd_idle_ttl * HZ >> 2; 12673d14c5d2SYehuda Sadeh 12683d14c5d2SYehuda Sadeh dout("osds timeout\n"); 12693d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 1270aca420bcSSage Weil remove_old_osds(osdc); 12713d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 12723d14c5d2SYehuda Sadeh 12733d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 12743d14c5d2SYehuda Sadeh round_jiffies_relative(delay)); 12753d14c5d2SYehuda Sadeh } 12763d14c5d2SYehuda Sadeh 127725845472SSage Weil static void complete_request(struct ceph_osd_request *req) 127825845472SSage Weil { 127925845472SSage Weil if (req->r_safe_callback) 128025845472SSage Weil req->r_safe_callback(req, NULL); 128125845472SSage Weil complete_all(&req->r_safe_completion); /* fsync waiter */ 128225845472SSage Weil } 128325845472SSage Weil 12841b83bef2SSage Weil static int __decode_pgid(void **p, void *end, struct ceph_pg *pgid) 12851b83bef2SSage Weil { 12861b83bef2SSage Weil __u8 v; 12871b83bef2SSage Weil 12881b83bef2SSage Weil ceph_decode_need(p, end, 1 + 8 + 4 + 4, bad); 12891b83bef2SSage Weil v = ceph_decode_8(p); 12901b83bef2SSage Weil if (v > 1) { 12911b83bef2SSage Weil pr_warning("do not understand pg encoding %d > 1", v); 12921b83bef2SSage Weil return -EINVAL; 12931b83bef2SSage Weil } 12941b83bef2SSage Weil pgid->pool = ceph_decode_64(p); 12951b83bef2SSage Weil pgid->seed = ceph_decode_32(p); 12961b83bef2SSage Weil *p += 4; 12971b83bef2SSage Weil return 0; 12981b83bef2SSage Weil 12991b83bef2SSage Weil bad: 13001b83bef2SSage Weil pr_warning("incomplete pg encoding"); 13011b83bef2SSage Weil return -EINVAL; 13021b83bef2SSage Weil } 13031b83bef2SSage Weil 13043d14c5d2SYehuda Sadeh /* 13053d14c5d2SYehuda Sadeh * handle osd op reply. either call the callback if it is specified, 13063d14c5d2SYehuda Sadeh * or do the completion to wake up the waiting thread. 13073d14c5d2SYehuda Sadeh */ 13083d14c5d2SYehuda Sadeh static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg, 13093d14c5d2SYehuda Sadeh struct ceph_connection *con) 13103d14c5d2SYehuda Sadeh { 13111b83bef2SSage Weil void *p, *end; 13123d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 13133d14c5d2SYehuda Sadeh u64 tid; 13141b83bef2SSage Weil int object_len; 13151b83bef2SSage Weil int numops, payload_len, flags; 13163d14c5d2SYehuda Sadeh s32 result; 13171b83bef2SSage Weil s32 retry_attempt; 13181b83bef2SSage Weil struct ceph_pg pg; 13191b83bef2SSage Weil int err; 13201b83bef2SSage Weil u32 reassert_epoch; 13211b83bef2SSage Weil u64 reassert_version; 13221b83bef2SSage Weil u32 osdmap_epoch; 13230d5af164SAlex Elder int already_completed; 13241b83bef2SSage Weil int i; 13253d14c5d2SYehuda Sadeh 13263d14c5d2SYehuda Sadeh tid = le64_to_cpu(msg->hdr.tid); 13271b83bef2SSage Weil dout("handle_reply %p tid %llu\n", msg, tid); 13281b83bef2SSage Weil 13291b83bef2SSage Weil p = msg->front.iov_base; 13301b83bef2SSage Weil end = p + msg->front.iov_len; 13311b83bef2SSage Weil 13321b83bef2SSage Weil ceph_decode_need(&p, end, 4, bad); 13331b83bef2SSage Weil object_len = ceph_decode_32(&p); 13341b83bef2SSage Weil ceph_decode_need(&p, end, object_len, bad); 13351b83bef2SSage Weil p += object_len; 13361b83bef2SSage Weil 13371b83bef2SSage Weil err = __decode_pgid(&p, end, &pg); 13381b83bef2SSage Weil if (err) 13393d14c5d2SYehuda Sadeh goto bad; 13401b83bef2SSage Weil 13411b83bef2SSage Weil ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad); 13421b83bef2SSage Weil flags = ceph_decode_64(&p); 13431b83bef2SSage Weil result = ceph_decode_32(&p); 13441b83bef2SSage Weil reassert_epoch = ceph_decode_32(&p); 13451b83bef2SSage Weil reassert_version = ceph_decode_64(&p); 13461b83bef2SSage Weil osdmap_epoch = ceph_decode_32(&p); 13471b83bef2SSage Weil 13483d14c5d2SYehuda Sadeh /* lookup */ 13493d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 13503d14c5d2SYehuda Sadeh req = __lookup_request(osdc, tid); 13513d14c5d2SYehuda Sadeh if (req == NULL) { 13523d14c5d2SYehuda Sadeh dout("handle_reply tid %llu dne\n", tid); 13533d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 13543d14c5d2SYehuda Sadeh return; 13553d14c5d2SYehuda Sadeh } 13563d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 13571b83bef2SSage Weil 13581b83bef2SSage Weil dout("handle_reply %p tid %llu req %p result %d\n", msg, tid, 13591b83bef2SSage Weil req, result); 13601b83bef2SSage Weil 13611b83bef2SSage Weil ceph_decode_need(&p, end, 4, bad); 13621b83bef2SSage Weil numops = ceph_decode_32(&p); 13631b83bef2SSage Weil if (numops > CEPH_OSD_MAX_OP) 13641b83bef2SSage Weil goto bad_put; 13651b83bef2SSage Weil if (numops != req->r_num_ops) 13661b83bef2SSage Weil goto bad_put; 13671b83bef2SSage Weil payload_len = 0; 13681b83bef2SSage Weil ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad); 13691b83bef2SSage Weil for (i = 0; i < numops; i++) { 13701b83bef2SSage Weil struct ceph_osd_op *op = p; 13711b83bef2SSage Weil int len; 13721b83bef2SSage Weil 13731b83bef2SSage Weil len = le32_to_cpu(op->payload_len); 13741b83bef2SSage Weil req->r_reply_op_len[i] = len; 13751b83bef2SSage Weil dout(" op %d has %d bytes\n", i, len); 13761b83bef2SSage Weil payload_len += len; 13771b83bef2SSage Weil p += sizeof(*op); 13781b83bef2SSage Weil } 13791b83bef2SSage Weil if (payload_len != le32_to_cpu(msg->hdr.data_len)) { 13801b83bef2SSage Weil pr_warning("sum of op payload lens %d != data_len %d", 13811b83bef2SSage Weil payload_len, le32_to_cpu(msg->hdr.data_len)); 13821b83bef2SSage Weil goto bad_put; 13831b83bef2SSage Weil } 13841b83bef2SSage Weil 13851b83bef2SSage Weil ceph_decode_need(&p, end, 4 + numops * 4, bad); 13861b83bef2SSage Weil retry_attempt = ceph_decode_32(&p); 13871b83bef2SSage Weil for (i = 0; i < numops; i++) 13881b83bef2SSage Weil req->r_reply_op_result[i] = ceph_decode_32(&p); 13893d14c5d2SYehuda Sadeh 13903d14c5d2SYehuda Sadeh /* 13913d14c5d2SYehuda Sadeh * if this connection filled our message, drop our reference now, to 13923d14c5d2SYehuda Sadeh * avoid a (safe but slower) revoke later. 13933d14c5d2SYehuda Sadeh */ 13943d14c5d2SYehuda Sadeh if (req->r_con_filling_msg == con && req->r_reply == msg) { 13953d14c5d2SYehuda Sadeh dout(" dropping con_filling_msg ref %p\n", con); 13963d14c5d2SYehuda Sadeh req->r_con_filling_msg = NULL; 13970d47766fSSage Weil con->ops->put(con); 13983d14c5d2SYehuda Sadeh } 13993d14c5d2SYehuda Sadeh 14003d14c5d2SYehuda Sadeh if (!req->r_got_reply) { 140195c96174SEric Dumazet unsigned int bytes; 14023d14c5d2SYehuda Sadeh 14031b83bef2SSage Weil req->r_result = result; 14043d14c5d2SYehuda Sadeh bytes = le32_to_cpu(msg->hdr.data_len); 14053d14c5d2SYehuda Sadeh dout("handle_reply result %d bytes %d\n", req->r_result, 14063d14c5d2SYehuda Sadeh bytes); 14073d14c5d2SYehuda Sadeh if (req->r_result == 0) 14083d14c5d2SYehuda Sadeh req->r_result = bytes; 14093d14c5d2SYehuda Sadeh 14103d14c5d2SYehuda Sadeh /* in case this is a write and we need to replay, */ 14111b83bef2SSage Weil req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch); 14121b83bef2SSage Weil req->r_reassert_version.version = cpu_to_le64(reassert_version); 14133d14c5d2SYehuda Sadeh 14143d14c5d2SYehuda Sadeh req->r_got_reply = 1; 14153d14c5d2SYehuda Sadeh } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) { 14163d14c5d2SYehuda Sadeh dout("handle_reply tid %llu dup ack\n", tid); 14173d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 14183d14c5d2SYehuda Sadeh goto done; 14193d14c5d2SYehuda Sadeh } 14203d14c5d2SYehuda Sadeh 14213d14c5d2SYehuda Sadeh dout("handle_reply tid %llu flags %d\n", tid, flags); 14223d14c5d2SYehuda Sadeh 1423a40c4f10SYehuda Sadeh if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK)) 1424a40c4f10SYehuda Sadeh __register_linger_request(osdc, req); 1425a40c4f10SYehuda Sadeh 14263d14c5d2SYehuda Sadeh /* either this is a read, or we got the safe response */ 14273d14c5d2SYehuda Sadeh if (result < 0 || 14283d14c5d2SYehuda Sadeh (flags & CEPH_OSD_FLAG_ONDISK) || 14293d14c5d2SYehuda Sadeh ((flags & CEPH_OSD_FLAG_WRITE) == 0)) 14303d14c5d2SYehuda Sadeh __unregister_request(osdc, req); 14313d14c5d2SYehuda Sadeh 14320d5af164SAlex Elder already_completed = req->r_completed; 14330d5af164SAlex Elder req->r_completed = 1; 14343d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 14350d5af164SAlex Elder if (already_completed) 14360d5af164SAlex Elder goto done; 14373d14c5d2SYehuda Sadeh 14383d14c5d2SYehuda Sadeh if (req->r_callback) 14393d14c5d2SYehuda Sadeh req->r_callback(req, msg); 14403d14c5d2SYehuda Sadeh else 14413d14c5d2SYehuda Sadeh complete_all(&req->r_completion); 14423d14c5d2SYehuda Sadeh 144325845472SSage Weil if (flags & CEPH_OSD_FLAG_ONDISK) 144425845472SSage Weil complete_request(req); 14453d14c5d2SYehuda Sadeh 14463d14c5d2SYehuda Sadeh done: 1447a40c4f10SYehuda Sadeh dout("req=%p req->r_linger=%d\n", req, req->r_linger); 14483d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 14493d14c5d2SYehuda Sadeh return; 14503d14c5d2SYehuda Sadeh 14511b83bef2SSage Weil bad_put: 14521b83bef2SSage Weil ceph_osdc_put_request(req); 14533d14c5d2SYehuda Sadeh bad: 14541b83bef2SSage Weil pr_err("corrupt osd_op_reply got %d %d\n", 14551b83bef2SSage Weil (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len)); 14563d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 14573d14c5d2SYehuda Sadeh } 14583d14c5d2SYehuda Sadeh 14596f6c7006SSage Weil static void reset_changed_osds(struct ceph_osd_client *osdc) 14603d14c5d2SYehuda Sadeh { 14613d14c5d2SYehuda Sadeh struct rb_node *p, *n; 14623d14c5d2SYehuda Sadeh 14633d14c5d2SYehuda Sadeh for (p = rb_first(&osdc->osds); p; p = n) { 14646f6c7006SSage Weil struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node); 14653d14c5d2SYehuda Sadeh 14663d14c5d2SYehuda Sadeh n = rb_next(p); 14673d14c5d2SYehuda Sadeh if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) || 14683d14c5d2SYehuda Sadeh memcmp(&osd->o_con.peer_addr, 14693d14c5d2SYehuda Sadeh ceph_osd_addr(osdc->osdmap, 14703d14c5d2SYehuda Sadeh osd->o_osd), 14713d14c5d2SYehuda Sadeh sizeof(struct ceph_entity_addr)) != 0) 14723d14c5d2SYehuda Sadeh __reset_osd(osdc, osd); 14733d14c5d2SYehuda Sadeh } 14743d14c5d2SYehuda Sadeh } 14753d14c5d2SYehuda Sadeh 14763d14c5d2SYehuda Sadeh /* 14776f6c7006SSage Weil * Requeue requests whose mapping to an OSD has changed. If requests map to 14786f6c7006SSage Weil * no osd, request a new map. 14793d14c5d2SYehuda Sadeh * 1480e6d50f67SAlex Elder * Caller should hold map_sem for read. 14813d14c5d2SYehuda Sadeh */ 148238d6453cSSage Weil static void kick_requests(struct ceph_osd_client *osdc, int force_resend) 14833d14c5d2SYehuda Sadeh { 1484a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 14856f6c7006SSage Weil struct rb_node *p; 14866f6c7006SSage Weil int needmap = 0; 14876f6c7006SSage Weil int err; 14883d14c5d2SYehuda Sadeh 148938d6453cSSage Weil dout("kick_requests %s\n", force_resend ? " (force resend)" : ""); 14903d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 14916194ea89SSage Weil for (p = rb_first(&osdc->requests); p; ) { 14926f6c7006SSage Weil req = rb_entry(p, struct ceph_osd_request, r_node); 14936194ea89SSage Weil p = rb_next(p); 1494ab60b16dSAlex Elder 1495ab60b16dSAlex Elder /* 1496ab60b16dSAlex Elder * For linger requests that have not yet been 1497ab60b16dSAlex Elder * registered, move them to the linger list; they'll 1498ab60b16dSAlex Elder * be sent to the osd in the loop below. Unregister 1499ab60b16dSAlex Elder * the request before re-registering it as a linger 1500ab60b16dSAlex Elder * request to ensure the __map_request() below 1501ab60b16dSAlex Elder * will decide it needs to be sent. 1502ab60b16dSAlex Elder */ 1503ab60b16dSAlex Elder if (req->r_linger && list_empty(&req->r_linger_item)) { 1504ab60b16dSAlex Elder dout("%p tid %llu restart on osd%d\n", 1505ab60b16dSAlex Elder req, req->r_tid, 1506ab60b16dSAlex Elder req->r_osd ? req->r_osd->o_osd : -1); 1507ab60b16dSAlex Elder __unregister_request(osdc, req); 1508ab60b16dSAlex Elder __register_linger_request(osdc, req); 1509ab60b16dSAlex Elder continue; 1510ab60b16dSAlex Elder } 1511ab60b16dSAlex Elder 151238d6453cSSage Weil err = __map_request(osdc, req, force_resend); 15136f6c7006SSage Weil if (err < 0) 15146f6c7006SSage Weil continue; /* error */ 15156f6c7006SSage Weil if (req->r_osd == NULL) { 15166f6c7006SSage Weil dout("%p tid %llu maps to no osd\n", req, req->r_tid); 15176f6c7006SSage Weil needmap++; /* request a newer map */ 15186f6c7006SSage Weil } else if (err > 0) { 15196194ea89SSage Weil if (!req->r_linger) { 15206194ea89SSage Weil dout("%p tid %llu requeued on osd%d\n", req, 15216194ea89SSage Weil req->r_tid, 15226f6c7006SSage Weil req->r_osd ? req->r_osd->o_osd : -1); 15236f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 15246f6c7006SSage Weil } 15256f6c7006SSage Weil } 15266194ea89SSage Weil } 1527a40c4f10SYehuda Sadeh 1528a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osdc->req_linger, 1529a40c4f10SYehuda Sadeh r_linger_item) { 1530a40c4f10SYehuda Sadeh dout("linger req=%p req->r_osd=%p\n", req, req->r_osd); 1531a40c4f10SYehuda Sadeh 153238d6453cSSage Weil err = __map_request(osdc, req, force_resend); 1533ab60b16dSAlex Elder dout("__map_request returned %d\n", err); 1534a40c4f10SYehuda Sadeh if (err == 0) 1535a40c4f10SYehuda Sadeh continue; /* no change and no osd was specified */ 1536a40c4f10SYehuda Sadeh if (err < 0) 1537a40c4f10SYehuda Sadeh continue; /* hrm! */ 1538a40c4f10SYehuda Sadeh if (req->r_osd == NULL) { 1539a40c4f10SYehuda Sadeh dout("tid %llu maps to no valid osd\n", req->r_tid); 1540a40c4f10SYehuda Sadeh needmap++; /* request a newer map */ 1541a40c4f10SYehuda Sadeh continue; 1542a40c4f10SYehuda Sadeh } 1543a40c4f10SYehuda Sadeh 1544a40c4f10SYehuda Sadeh dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid, 1545a40c4f10SYehuda Sadeh req->r_osd ? req->r_osd->o_osd : -1); 1546a40c4f10SYehuda Sadeh __register_request(osdc, req); 1547c89ce05eSAlex Elder __unregister_linger_request(osdc, req); 1548a40c4f10SYehuda Sadeh } 15493d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 15503d14c5d2SYehuda Sadeh 15513d14c5d2SYehuda Sadeh if (needmap) { 15523d14c5d2SYehuda Sadeh dout("%d requests for down osds, need new map\n", needmap); 15533d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 15543d14c5d2SYehuda Sadeh } 1555e6d50f67SAlex Elder reset_changed_osds(osdc); 15563d14c5d2SYehuda Sadeh } 15576f6c7006SSage Weil 15586f6c7006SSage Weil 15593d14c5d2SYehuda Sadeh /* 15603d14c5d2SYehuda Sadeh * Process updated osd map. 15613d14c5d2SYehuda Sadeh * 15623d14c5d2SYehuda Sadeh * The message contains any number of incremental and full maps, normally 15633d14c5d2SYehuda Sadeh * indicating some sort of topology change in the cluster. Kick requests 15643d14c5d2SYehuda Sadeh * off to different OSDs as needed. 15653d14c5d2SYehuda Sadeh */ 15663d14c5d2SYehuda Sadeh void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg) 15673d14c5d2SYehuda Sadeh { 15683d14c5d2SYehuda Sadeh void *p, *end, *next; 15693d14c5d2SYehuda Sadeh u32 nr_maps, maplen; 15703d14c5d2SYehuda Sadeh u32 epoch; 15713d14c5d2SYehuda Sadeh struct ceph_osdmap *newmap = NULL, *oldmap; 15723d14c5d2SYehuda Sadeh int err; 15733d14c5d2SYehuda Sadeh struct ceph_fsid fsid; 15743d14c5d2SYehuda Sadeh 15753d14c5d2SYehuda Sadeh dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0); 15763d14c5d2SYehuda Sadeh p = msg->front.iov_base; 15773d14c5d2SYehuda Sadeh end = p + msg->front.iov_len; 15783d14c5d2SYehuda Sadeh 15793d14c5d2SYehuda Sadeh /* verify fsid */ 15803d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, sizeof(fsid), bad); 15813d14c5d2SYehuda Sadeh ceph_decode_copy(&p, &fsid, sizeof(fsid)); 15823d14c5d2SYehuda Sadeh if (ceph_check_fsid(osdc->client, &fsid) < 0) 15833d14c5d2SYehuda Sadeh return; 15843d14c5d2SYehuda Sadeh 15853d14c5d2SYehuda Sadeh down_write(&osdc->map_sem); 15863d14c5d2SYehuda Sadeh 15873d14c5d2SYehuda Sadeh /* incremental maps */ 15883d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 15893d14c5d2SYehuda Sadeh dout(" %d inc maps\n", nr_maps); 15903d14c5d2SYehuda Sadeh while (nr_maps > 0) { 15913d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 15923d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 15933d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 15943d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 15953d14c5d2SYehuda Sadeh next = p + maplen; 15963d14c5d2SYehuda Sadeh if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) { 15973d14c5d2SYehuda Sadeh dout("applying incremental map %u len %d\n", 15983d14c5d2SYehuda Sadeh epoch, maplen); 15993d14c5d2SYehuda Sadeh newmap = osdmap_apply_incremental(&p, next, 16003d14c5d2SYehuda Sadeh osdc->osdmap, 160115d9882cSAlex Elder &osdc->client->msgr); 16023d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 16033d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 16043d14c5d2SYehuda Sadeh goto bad; 16053d14c5d2SYehuda Sadeh } 16063d14c5d2SYehuda Sadeh BUG_ON(!newmap); 16073d14c5d2SYehuda Sadeh if (newmap != osdc->osdmap) { 16083d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 16093d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 16103d14c5d2SYehuda Sadeh } 161138d6453cSSage Weil kick_requests(osdc, 0); 16123d14c5d2SYehuda Sadeh } else { 16133d14c5d2SYehuda Sadeh dout("ignoring incremental map %u len %d\n", 16143d14c5d2SYehuda Sadeh epoch, maplen); 16153d14c5d2SYehuda Sadeh } 16163d14c5d2SYehuda Sadeh p = next; 16173d14c5d2SYehuda Sadeh nr_maps--; 16183d14c5d2SYehuda Sadeh } 16193d14c5d2SYehuda Sadeh if (newmap) 16203d14c5d2SYehuda Sadeh goto done; 16213d14c5d2SYehuda Sadeh 16223d14c5d2SYehuda Sadeh /* full maps */ 16233d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 16243d14c5d2SYehuda Sadeh dout(" %d full maps\n", nr_maps); 16253d14c5d2SYehuda Sadeh while (nr_maps) { 16263d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 16273d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 16283d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 16293d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 16303d14c5d2SYehuda Sadeh if (nr_maps > 1) { 16313d14c5d2SYehuda Sadeh dout("skipping non-latest full map %u len %d\n", 16323d14c5d2SYehuda Sadeh epoch, maplen); 16333d14c5d2SYehuda Sadeh } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) { 16343d14c5d2SYehuda Sadeh dout("skipping full map %u len %d, " 16353d14c5d2SYehuda Sadeh "older than our %u\n", epoch, maplen, 16363d14c5d2SYehuda Sadeh osdc->osdmap->epoch); 16373d14c5d2SYehuda Sadeh } else { 163838d6453cSSage Weil int skipped_map = 0; 163938d6453cSSage Weil 16403d14c5d2SYehuda Sadeh dout("taking full map %u len %d\n", epoch, maplen); 16413d14c5d2SYehuda Sadeh newmap = osdmap_decode(&p, p+maplen); 16423d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 16433d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 16443d14c5d2SYehuda Sadeh goto bad; 16453d14c5d2SYehuda Sadeh } 16463d14c5d2SYehuda Sadeh BUG_ON(!newmap); 16473d14c5d2SYehuda Sadeh oldmap = osdc->osdmap; 16483d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 164938d6453cSSage Weil if (oldmap) { 165038d6453cSSage Weil if (oldmap->epoch + 1 < newmap->epoch) 165138d6453cSSage Weil skipped_map = 1; 16523d14c5d2SYehuda Sadeh ceph_osdmap_destroy(oldmap); 165338d6453cSSage Weil } 165438d6453cSSage Weil kick_requests(osdc, skipped_map); 16553d14c5d2SYehuda Sadeh } 16563d14c5d2SYehuda Sadeh p += maplen; 16573d14c5d2SYehuda Sadeh nr_maps--; 16583d14c5d2SYehuda Sadeh } 16593d14c5d2SYehuda Sadeh 16603d14c5d2SYehuda Sadeh done: 16613d14c5d2SYehuda Sadeh downgrade_write(&osdc->map_sem); 16623d14c5d2SYehuda Sadeh ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch); 1663cd634fb6SSage Weil 1664cd634fb6SSage Weil /* 1665cd634fb6SSage Weil * subscribe to subsequent osdmap updates if full to ensure 1666cd634fb6SSage Weil * we find out when we are no longer full and stop returning 1667cd634fb6SSage Weil * ENOSPC. 1668cd634fb6SSage Weil */ 1669cd634fb6SSage Weil if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL)) 1670cd634fb6SSage Weil ceph_monc_request_next_osdmap(&osdc->client->monc); 1671cd634fb6SSage Weil 1672f9d25199SAlex Elder mutex_lock(&osdc->request_mutex); 1673f9d25199SAlex Elder __send_queued(osdc); 1674f9d25199SAlex Elder mutex_unlock(&osdc->request_mutex); 16753d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 16763d14c5d2SYehuda Sadeh wake_up_all(&osdc->client->auth_wq); 16773d14c5d2SYehuda Sadeh return; 16783d14c5d2SYehuda Sadeh 16793d14c5d2SYehuda Sadeh bad: 16803d14c5d2SYehuda Sadeh pr_err("osdc handle_map corrupt msg\n"); 16813d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 16823d14c5d2SYehuda Sadeh up_write(&osdc->map_sem); 16833d14c5d2SYehuda Sadeh return; 16843d14c5d2SYehuda Sadeh } 16853d14c5d2SYehuda Sadeh 16863d14c5d2SYehuda Sadeh /* 1687a40c4f10SYehuda Sadeh * watch/notify callback event infrastructure 1688a40c4f10SYehuda Sadeh * 1689a40c4f10SYehuda Sadeh * These callbacks are used both for watch and notify operations. 1690a40c4f10SYehuda Sadeh */ 1691a40c4f10SYehuda Sadeh static void __release_event(struct kref *kref) 1692a40c4f10SYehuda Sadeh { 1693a40c4f10SYehuda Sadeh struct ceph_osd_event *event = 1694a40c4f10SYehuda Sadeh container_of(kref, struct ceph_osd_event, kref); 1695a40c4f10SYehuda Sadeh 1696a40c4f10SYehuda Sadeh dout("__release_event %p\n", event); 1697a40c4f10SYehuda Sadeh kfree(event); 1698a40c4f10SYehuda Sadeh } 1699a40c4f10SYehuda Sadeh 1700a40c4f10SYehuda Sadeh static void get_event(struct ceph_osd_event *event) 1701a40c4f10SYehuda Sadeh { 1702a40c4f10SYehuda Sadeh kref_get(&event->kref); 1703a40c4f10SYehuda Sadeh } 1704a40c4f10SYehuda Sadeh 1705a40c4f10SYehuda Sadeh void ceph_osdc_put_event(struct ceph_osd_event *event) 1706a40c4f10SYehuda Sadeh { 1707a40c4f10SYehuda Sadeh kref_put(&event->kref, __release_event); 1708a40c4f10SYehuda Sadeh } 1709a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_put_event); 1710a40c4f10SYehuda Sadeh 1711a40c4f10SYehuda Sadeh static void __insert_event(struct ceph_osd_client *osdc, 1712a40c4f10SYehuda Sadeh struct ceph_osd_event *new) 1713a40c4f10SYehuda Sadeh { 1714a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 1715a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 1716a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 1717a40c4f10SYehuda Sadeh 1718a40c4f10SYehuda Sadeh while (*p) { 1719a40c4f10SYehuda Sadeh parent = *p; 1720a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 1721a40c4f10SYehuda Sadeh if (new->cookie < event->cookie) 1722a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 1723a40c4f10SYehuda Sadeh else if (new->cookie > event->cookie) 1724a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 1725a40c4f10SYehuda Sadeh else 1726a40c4f10SYehuda Sadeh BUG(); 1727a40c4f10SYehuda Sadeh } 1728a40c4f10SYehuda Sadeh 1729a40c4f10SYehuda Sadeh rb_link_node(&new->node, parent, p); 1730a40c4f10SYehuda Sadeh rb_insert_color(&new->node, &osdc->event_tree); 1731a40c4f10SYehuda Sadeh } 1732a40c4f10SYehuda Sadeh 1733a40c4f10SYehuda Sadeh static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc, 1734a40c4f10SYehuda Sadeh u64 cookie) 1735a40c4f10SYehuda Sadeh { 1736a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 1737a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 1738a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 1739a40c4f10SYehuda Sadeh 1740a40c4f10SYehuda Sadeh while (*p) { 1741a40c4f10SYehuda Sadeh parent = *p; 1742a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 1743a40c4f10SYehuda Sadeh if (cookie < event->cookie) 1744a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 1745a40c4f10SYehuda Sadeh else if (cookie > event->cookie) 1746a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 1747a40c4f10SYehuda Sadeh else 1748a40c4f10SYehuda Sadeh return event; 1749a40c4f10SYehuda Sadeh } 1750a40c4f10SYehuda Sadeh return NULL; 1751a40c4f10SYehuda Sadeh } 1752a40c4f10SYehuda Sadeh 1753a40c4f10SYehuda Sadeh static void __remove_event(struct ceph_osd_event *event) 1754a40c4f10SYehuda Sadeh { 1755a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 1756a40c4f10SYehuda Sadeh 1757a40c4f10SYehuda Sadeh if (!RB_EMPTY_NODE(&event->node)) { 1758a40c4f10SYehuda Sadeh dout("__remove_event removed %p\n", event); 1759a40c4f10SYehuda Sadeh rb_erase(&event->node, &osdc->event_tree); 1760a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 1761a40c4f10SYehuda Sadeh } else { 1762a40c4f10SYehuda Sadeh dout("__remove_event didn't remove %p\n", event); 1763a40c4f10SYehuda Sadeh } 1764a40c4f10SYehuda Sadeh } 1765a40c4f10SYehuda Sadeh 1766a40c4f10SYehuda Sadeh int ceph_osdc_create_event(struct ceph_osd_client *osdc, 1767a40c4f10SYehuda Sadeh void (*event_cb)(u64, u64, u8, void *), 17683c663bbdSAlex Elder void *data, struct ceph_osd_event **pevent) 1769a40c4f10SYehuda Sadeh { 1770a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 1771a40c4f10SYehuda Sadeh 1772a40c4f10SYehuda Sadeh event = kmalloc(sizeof(*event), GFP_NOIO); 1773a40c4f10SYehuda Sadeh if (!event) 1774a40c4f10SYehuda Sadeh return -ENOMEM; 1775a40c4f10SYehuda Sadeh 1776a40c4f10SYehuda Sadeh dout("create_event %p\n", event); 1777a40c4f10SYehuda Sadeh event->cb = event_cb; 17783c663bbdSAlex Elder event->one_shot = 0; 1779a40c4f10SYehuda Sadeh event->data = data; 1780a40c4f10SYehuda Sadeh event->osdc = osdc; 1781a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&event->osd_node); 17823ee5234dSAlex Elder RB_CLEAR_NODE(&event->node); 1783a40c4f10SYehuda Sadeh kref_init(&event->kref); /* one ref for us */ 1784a40c4f10SYehuda Sadeh kref_get(&event->kref); /* one ref for the caller */ 1785a40c4f10SYehuda Sadeh 1786a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 1787a40c4f10SYehuda Sadeh event->cookie = ++osdc->event_count; 1788a40c4f10SYehuda Sadeh __insert_event(osdc, event); 1789a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 1790a40c4f10SYehuda Sadeh 1791a40c4f10SYehuda Sadeh *pevent = event; 1792a40c4f10SYehuda Sadeh return 0; 1793a40c4f10SYehuda Sadeh } 1794a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_create_event); 1795a40c4f10SYehuda Sadeh 1796a40c4f10SYehuda Sadeh void ceph_osdc_cancel_event(struct ceph_osd_event *event) 1797a40c4f10SYehuda Sadeh { 1798a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 1799a40c4f10SYehuda Sadeh 1800a40c4f10SYehuda Sadeh dout("cancel_event %p\n", event); 1801a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 1802a40c4f10SYehuda Sadeh __remove_event(event); 1803a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 1804a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); /* caller's */ 1805a40c4f10SYehuda Sadeh } 1806a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_cancel_event); 1807a40c4f10SYehuda Sadeh 1808a40c4f10SYehuda Sadeh 1809a40c4f10SYehuda Sadeh static void do_event_work(struct work_struct *work) 1810a40c4f10SYehuda Sadeh { 1811a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work = 1812a40c4f10SYehuda Sadeh container_of(work, struct ceph_osd_event_work, work); 1813a40c4f10SYehuda Sadeh struct ceph_osd_event *event = event_work->event; 1814a40c4f10SYehuda Sadeh u64 ver = event_work->ver; 1815a40c4f10SYehuda Sadeh u64 notify_id = event_work->notify_id; 1816a40c4f10SYehuda Sadeh u8 opcode = event_work->opcode; 1817a40c4f10SYehuda Sadeh 1818a40c4f10SYehuda Sadeh dout("do_event_work completing %p\n", event); 1819a40c4f10SYehuda Sadeh event->cb(ver, notify_id, opcode, event->data); 1820a40c4f10SYehuda Sadeh dout("do_event_work completed %p\n", event); 1821a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 1822a40c4f10SYehuda Sadeh kfree(event_work); 1823a40c4f10SYehuda Sadeh } 1824a40c4f10SYehuda Sadeh 1825a40c4f10SYehuda Sadeh 1826a40c4f10SYehuda Sadeh /* 1827a40c4f10SYehuda Sadeh * Process osd watch notifications 1828a40c4f10SYehuda Sadeh */ 18293c663bbdSAlex Elder static void handle_watch_notify(struct ceph_osd_client *osdc, 18303c663bbdSAlex Elder struct ceph_msg *msg) 1831a40c4f10SYehuda Sadeh { 1832a40c4f10SYehuda Sadeh void *p, *end; 1833a40c4f10SYehuda Sadeh u8 proto_ver; 1834a40c4f10SYehuda Sadeh u64 cookie, ver, notify_id; 1835a40c4f10SYehuda Sadeh u8 opcode; 1836a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 1837a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work; 1838a40c4f10SYehuda Sadeh 1839a40c4f10SYehuda Sadeh p = msg->front.iov_base; 1840a40c4f10SYehuda Sadeh end = p + msg->front.iov_len; 1841a40c4f10SYehuda Sadeh 1842a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, proto_ver, bad); 1843a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, opcode, bad); 1844a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, cookie, bad); 1845a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, ver, bad); 1846a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, notify_id, bad); 1847a40c4f10SYehuda Sadeh 1848a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 1849a40c4f10SYehuda Sadeh event = __find_event(osdc, cookie); 1850a40c4f10SYehuda Sadeh if (event) { 18513c663bbdSAlex Elder BUG_ON(event->one_shot); 1852a40c4f10SYehuda Sadeh get_event(event); 1853a40c4f10SYehuda Sadeh } 1854a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 1855a40c4f10SYehuda Sadeh dout("handle_watch_notify cookie %lld ver %lld event %p\n", 1856a40c4f10SYehuda Sadeh cookie, ver, event); 1857a40c4f10SYehuda Sadeh if (event) { 1858a40c4f10SYehuda Sadeh event_work = kmalloc(sizeof(*event_work), GFP_NOIO); 1859a40c4f10SYehuda Sadeh if (!event_work) { 1860a40c4f10SYehuda Sadeh dout("ERROR: could not allocate event_work\n"); 1861a40c4f10SYehuda Sadeh goto done_err; 1862a40c4f10SYehuda Sadeh } 18636b0ae409SMariusz Kozlowski INIT_WORK(&event_work->work, do_event_work); 1864a40c4f10SYehuda Sadeh event_work->event = event; 1865a40c4f10SYehuda Sadeh event_work->ver = ver; 1866a40c4f10SYehuda Sadeh event_work->notify_id = notify_id; 1867a40c4f10SYehuda Sadeh event_work->opcode = opcode; 1868a40c4f10SYehuda Sadeh if (!queue_work(osdc->notify_wq, &event_work->work)) { 1869a40c4f10SYehuda Sadeh dout("WARNING: failed to queue notify event work\n"); 1870a40c4f10SYehuda Sadeh goto done_err; 1871a40c4f10SYehuda Sadeh } 1872a40c4f10SYehuda Sadeh } 1873a40c4f10SYehuda Sadeh 1874a40c4f10SYehuda Sadeh return; 1875a40c4f10SYehuda Sadeh 1876a40c4f10SYehuda Sadeh done_err: 1877a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 1878a40c4f10SYehuda Sadeh return; 1879a40c4f10SYehuda Sadeh 1880a40c4f10SYehuda Sadeh bad: 1881a40c4f10SYehuda Sadeh pr_err("osdc handle_watch_notify corrupt msg\n"); 1882a40c4f10SYehuda Sadeh return; 1883a40c4f10SYehuda Sadeh } 1884a40c4f10SYehuda Sadeh 188570636773SAlex Elder static void ceph_osdc_msg_data_set(struct ceph_msg *msg, 188670636773SAlex Elder struct ceph_osd_data *osd_data) 188770636773SAlex Elder { 188870636773SAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 188970636773SAlex Elder BUG_ON(osd_data->length > (u64) SIZE_MAX); 189070636773SAlex Elder if (osd_data->length) 189170636773SAlex Elder ceph_msg_data_set_pages(msg, osd_data->pages, 189270636773SAlex Elder osd_data->length, osd_data->alignment); 18939a5e6d09SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) { 18949a5e6d09SAlex Elder BUG_ON(!osd_data->pagelist->length); 18959a5e6d09SAlex Elder ceph_msg_data_set_pagelist(msg, osd_data->pagelist); 189670636773SAlex Elder #ifdef CONFIG_BLOCK 189770636773SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) { 189870636773SAlex Elder ceph_msg_data_set_bio(msg, osd_data->bio); 189970636773SAlex Elder #endif 190070636773SAlex Elder } else { 190170636773SAlex Elder BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE); 190270636773SAlex Elder } 190370636773SAlex Elder } 190470636773SAlex Elder 1905a40c4f10SYehuda Sadeh /* 19063d14c5d2SYehuda Sadeh * Register request, send initial attempt. 19073d14c5d2SYehuda Sadeh */ 19083d14c5d2SYehuda Sadeh int ceph_osdc_start_request(struct ceph_osd_client *osdc, 19093d14c5d2SYehuda Sadeh struct ceph_osd_request *req, 19103d14c5d2SYehuda Sadeh bool nofail) 19113d14c5d2SYehuda Sadeh { 19123d14c5d2SYehuda Sadeh int rc = 0; 19133d14c5d2SYehuda Sadeh 191470636773SAlex Elder /* Set up response incoming data and request outgoing data fields */ 19150fff87ecSAlex Elder 191670636773SAlex Elder ceph_osdc_msg_data_set(req->r_reply, &req->r_data_in); 191770636773SAlex Elder ceph_osdc_msg_data_set(req->r_request, &req->r_data_out); 19183d14c5d2SYehuda Sadeh 19193d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 19203d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 1921dc4b870cSAlex Elder __register_request(osdc, req); 192292451b49SAlex Elder WARN_ON(req->r_sent); 192338d6453cSSage Weil rc = __map_request(osdc, req, 0); 19249d6fcb08SSage Weil if (rc < 0) { 19259d6fcb08SSage Weil if (nofail) { 19269d6fcb08SSage Weil dout("osdc_start_request failed map, " 19279d6fcb08SSage Weil " will retry %lld\n", req->r_tid); 19289d6fcb08SSage Weil rc = 0; 19299d6fcb08SSage Weil } 1930234af26fSDan Carpenter goto out_unlock; 19319d6fcb08SSage Weil } 19326f6c7006SSage Weil if (req->r_osd == NULL) { 19336f6c7006SSage Weil dout("send_request %p no up osds in pg\n", req); 19346f6c7006SSage Weil ceph_monc_request_next_osdmap(&osdc->client->monc); 19356f6c7006SSage Weil } else { 19367e2766a1SAlex Elder __send_queued(osdc); 193756e925b6SSage Weil } 19383d14c5d2SYehuda Sadeh rc = 0; 1939234af26fSDan Carpenter out_unlock: 19403d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 19413d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 19423d14c5d2SYehuda Sadeh return rc; 19433d14c5d2SYehuda Sadeh } 19443d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_start_request); 19453d14c5d2SYehuda Sadeh 19463d14c5d2SYehuda Sadeh /* 19473d14c5d2SYehuda Sadeh * wait for a request to complete 19483d14c5d2SYehuda Sadeh */ 19493d14c5d2SYehuda Sadeh int ceph_osdc_wait_request(struct ceph_osd_client *osdc, 19503d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 19513d14c5d2SYehuda Sadeh { 19523d14c5d2SYehuda Sadeh int rc; 19533d14c5d2SYehuda Sadeh 19543d14c5d2SYehuda Sadeh rc = wait_for_completion_interruptible(&req->r_completion); 19553d14c5d2SYehuda Sadeh if (rc < 0) { 19563d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 19573d14c5d2SYehuda Sadeh __cancel_request(req); 19583d14c5d2SYehuda Sadeh __unregister_request(osdc, req); 19593d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 196025845472SSage Weil complete_request(req); 19613d14c5d2SYehuda Sadeh dout("wait_request tid %llu canceled/timed out\n", req->r_tid); 19623d14c5d2SYehuda Sadeh return rc; 19633d14c5d2SYehuda Sadeh } 19643d14c5d2SYehuda Sadeh 19653d14c5d2SYehuda Sadeh dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result); 19663d14c5d2SYehuda Sadeh return req->r_result; 19673d14c5d2SYehuda Sadeh } 19683d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_wait_request); 19693d14c5d2SYehuda Sadeh 19703d14c5d2SYehuda Sadeh /* 19713d14c5d2SYehuda Sadeh * sync - wait for all in-flight requests to flush. avoid starvation. 19723d14c5d2SYehuda Sadeh */ 19733d14c5d2SYehuda Sadeh void ceph_osdc_sync(struct ceph_osd_client *osdc) 19743d14c5d2SYehuda Sadeh { 19753d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 19763d14c5d2SYehuda Sadeh u64 last_tid, next_tid = 0; 19773d14c5d2SYehuda Sadeh 19783d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 19793d14c5d2SYehuda Sadeh last_tid = osdc->last_tid; 19803d14c5d2SYehuda Sadeh while (1) { 19813d14c5d2SYehuda Sadeh req = __lookup_request_ge(osdc, next_tid); 19823d14c5d2SYehuda Sadeh if (!req) 19833d14c5d2SYehuda Sadeh break; 19843d14c5d2SYehuda Sadeh if (req->r_tid > last_tid) 19853d14c5d2SYehuda Sadeh break; 19863d14c5d2SYehuda Sadeh 19873d14c5d2SYehuda Sadeh next_tid = req->r_tid + 1; 19883d14c5d2SYehuda Sadeh if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0) 19893d14c5d2SYehuda Sadeh continue; 19903d14c5d2SYehuda Sadeh 19913d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 19923d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 19933d14c5d2SYehuda Sadeh dout("sync waiting on tid %llu (last is %llu)\n", 19943d14c5d2SYehuda Sadeh req->r_tid, last_tid); 19953d14c5d2SYehuda Sadeh wait_for_completion(&req->r_safe_completion); 19963d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 19973d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 19983d14c5d2SYehuda Sadeh } 19993d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 20003d14c5d2SYehuda Sadeh dout("sync done (thru tid %llu)\n", last_tid); 20013d14c5d2SYehuda Sadeh } 20023d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_sync); 20033d14c5d2SYehuda Sadeh 20043d14c5d2SYehuda Sadeh /* 20053d14c5d2SYehuda Sadeh * init, shutdown 20063d14c5d2SYehuda Sadeh */ 20073d14c5d2SYehuda Sadeh int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client) 20083d14c5d2SYehuda Sadeh { 20093d14c5d2SYehuda Sadeh int err; 20103d14c5d2SYehuda Sadeh 20113d14c5d2SYehuda Sadeh dout("init\n"); 20123d14c5d2SYehuda Sadeh osdc->client = client; 20133d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 20143d14c5d2SYehuda Sadeh init_rwsem(&osdc->map_sem); 20153d14c5d2SYehuda Sadeh init_completion(&osdc->map_waiters); 20163d14c5d2SYehuda Sadeh osdc->last_requested_map = 0; 20173d14c5d2SYehuda Sadeh mutex_init(&osdc->request_mutex); 20183d14c5d2SYehuda Sadeh osdc->last_tid = 0; 20193d14c5d2SYehuda Sadeh osdc->osds = RB_ROOT; 20203d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->osd_lru); 20213d14c5d2SYehuda Sadeh osdc->requests = RB_ROOT; 20223d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_lru); 20236f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_unsent); 20246f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_notarget); 2025a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_linger); 20263d14c5d2SYehuda Sadeh osdc->num_requests = 0; 20273d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout); 20283d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout); 2029a40c4f10SYehuda Sadeh spin_lock_init(&osdc->event_lock); 2030a40c4f10SYehuda Sadeh osdc->event_tree = RB_ROOT; 2031a40c4f10SYehuda Sadeh osdc->event_count = 0; 20323d14c5d2SYehuda Sadeh 20333d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 20343d14c5d2SYehuda Sadeh round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ)); 20353d14c5d2SYehuda Sadeh 20363d14c5d2SYehuda Sadeh err = -ENOMEM; 20373d14c5d2SYehuda Sadeh osdc->req_mempool = mempool_create_kmalloc_pool(10, 20383d14c5d2SYehuda Sadeh sizeof(struct ceph_osd_request)); 20393d14c5d2SYehuda Sadeh if (!osdc->req_mempool) 20403d14c5d2SYehuda Sadeh goto out; 20413d14c5d2SYehuda Sadeh 2042d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP, 2043d50b409fSSage Weil OSD_OP_FRONT_LEN, 10, true, 20443d14c5d2SYehuda Sadeh "osd_op"); 20453d14c5d2SYehuda Sadeh if (err < 0) 20463d14c5d2SYehuda Sadeh goto out_mempool; 2047d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY, 20483d14c5d2SYehuda Sadeh OSD_OPREPLY_FRONT_LEN, 10, true, 20493d14c5d2SYehuda Sadeh "osd_op_reply"); 20503d14c5d2SYehuda Sadeh if (err < 0) 20513d14c5d2SYehuda Sadeh goto out_msgpool; 2052a40c4f10SYehuda Sadeh 2053a40c4f10SYehuda Sadeh osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify"); 2054a40c4f10SYehuda Sadeh if (IS_ERR(osdc->notify_wq)) { 2055a40c4f10SYehuda Sadeh err = PTR_ERR(osdc->notify_wq); 2056a40c4f10SYehuda Sadeh osdc->notify_wq = NULL; 2057a40c4f10SYehuda Sadeh goto out_msgpool; 2058a40c4f10SYehuda Sadeh } 20593d14c5d2SYehuda Sadeh return 0; 20603d14c5d2SYehuda Sadeh 20613d14c5d2SYehuda Sadeh out_msgpool: 20623d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 20633d14c5d2SYehuda Sadeh out_mempool: 20643d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 20653d14c5d2SYehuda Sadeh out: 20663d14c5d2SYehuda Sadeh return err; 20673d14c5d2SYehuda Sadeh } 20683d14c5d2SYehuda Sadeh 20693d14c5d2SYehuda Sadeh void ceph_osdc_stop(struct ceph_osd_client *osdc) 20703d14c5d2SYehuda Sadeh { 2071a40c4f10SYehuda Sadeh flush_workqueue(osdc->notify_wq); 2072a40c4f10SYehuda Sadeh destroy_workqueue(osdc->notify_wq); 20733d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->timeout_work); 20743d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->osds_timeout_work); 20753d14c5d2SYehuda Sadeh if (osdc->osdmap) { 20763d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 20773d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 20783d14c5d2SYehuda Sadeh } 2079aca420bcSSage Weil remove_all_osds(osdc); 20803d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 20813d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 20823d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op_reply); 20833d14c5d2SYehuda Sadeh } 20843d14c5d2SYehuda Sadeh 20853d14c5d2SYehuda Sadeh /* 20863d14c5d2SYehuda Sadeh * Read some contiguous pages. If we cross a stripe boundary, shorten 20873d14c5d2SYehuda Sadeh * *plen. Return number of bytes read, or error. 20883d14c5d2SYehuda Sadeh */ 20893d14c5d2SYehuda Sadeh int ceph_osdc_readpages(struct ceph_osd_client *osdc, 20903d14c5d2SYehuda Sadeh struct ceph_vino vino, struct ceph_file_layout *layout, 20913d14c5d2SYehuda Sadeh u64 off, u64 *plen, 20923d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 2093b7495fc2SSage Weil struct page **pages, int num_pages, int page_align) 20943d14c5d2SYehuda Sadeh { 20953d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 20960fff87ecSAlex Elder struct ceph_osd_data *osd_data; 20973d14c5d2SYehuda Sadeh int rc = 0; 20983d14c5d2SYehuda Sadeh 20993d14c5d2SYehuda Sadeh dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino, 21003d14c5d2SYehuda Sadeh vino.snap, off, *plen); 21013d14c5d2SYehuda Sadeh req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 21023d14c5d2SYehuda Sadeh CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ, 21033d14c5d2SYehuda Sadeh NULL, 0, truncate_seq, truncate_size, NULL, 2104153e5167SAlex Elder false); 21056816282dSSage Weil if (IS_ERR(req)) 21066816282dSSage Weil return PTR_ERR(req); 21073d14c5d2SYehuda Sadeh 21083d14c5d2SYehuda Sadeh /* it may be a short read due to an object boundary */ 21090fff87ecSAlex Elder 21100fff87ecSAlex Elder osd_data = &req->r_data_in; 21110fff87ecSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_PAGES; 21120fff87ecSAlex Elder osd_data->pages = pages; 2113e0c59487SAlex Elder osd_data->length = *plen; 21140fff87ecSAlex Elder osd_data->alignment = page_align; 21153d14c5d2SYehuda Sadeh 2116e0c59487SAlex Elder dout("readpages final extent is %llu~%llu (%llu bytes align %d)\n", 2117e0c59487SAlex Elder off, *plen, osd_data->length, page_align); 21183d14c5d2SYehuda Sadeh 21193d14c5d2SYehuda Sadeh rc = ceph_osdc_start_request(osdc, req, false); 21203d14c5d2SYehuda Sadeh if (!rc) 21213d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 21223d14c5d2SYehuda Sadeh 21233d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 21243d14c5d2SYehuda Sadeh dout("readpages result %d\n", rc); 21253d14c5d2SYehuda Sadeh return rc; 21263d14c5d2SYehuda Sadeh } 21273d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_readpages); 21283d14c5d2SYehuda Sadeh 21293d14c5d2SYehuda Sadeh /* 21303d14c5d2SYehuda Sadeh * do a synchronous write on N pages 21313d14c5d2SYehuda Sadeh */ 21323d14c5d2SYehuda Sadeh int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino, 21333d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 21343d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 21353d14c5d2SYehuda Sadeh u64 off, u64 len, 21363d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 21373d14c5d2SYehuda Sadeh struct timespec *mtime, 213824808826SAlex Elder struct page **pages, int num_pages) 21393d14c5d2SYehuda Sadeh { 21403d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 21410fff87ecSAlex Elder struct ceph_osd_data *osd_data; 21423d14c5d2SYehuda Sadeh int rc = 0; 2143b7495fc2SSage Weil int page_align = off & ~PAGE_MASK; 21443d14c5d2SYehuda Sadeh 21453d14c5d2SYehuda Sadeh BUG_ON(vino.snap != CEPH_NOSNAP); 21463d14c5d2SYehuda Sadeh req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 21473d14c5d2SYehuda Sadeh CEPH_OSD_OP_WRITE, 214824808826SAlex Elder CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE, 2149fbf8685fSAlex Elder snapc, 0, 21503d14c5d2SYehuda Sadeh truncate_seq, truncate_size, mtime, 2151153e5167SAlex Elder true); 21526816282dSSage Weil if (IS_ERR(req)) 21536816282dSSage Weil return PTR_ERR(req); 21543d14c5d2SYehuda Sadeh 21553d14c5d2SYehuda Sadeh /* it may be a short write due to an object boundary */ 21560fff87ecSAlex Elder osd_data = &req->r_data_out; 21570fff87ecSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_PAGES; 21580fff87ecSAlex Elder osd_data->pages = pages; 2159e0c59487SAlex Elder osd_data->length = len; 21600fff87ecSAlex Elder osd_data->alignment = page_align; 2161e0c59487SAlex Elder dout("writepages %llu~%llu (%llu bytes)\n", off, len, osd_data->length); 21623d14c5d2SYehuda Sadeh 216387f979d3SAlex Elder rc = ceph_osdc_start_request(osdc, req, true); 21643d14c5d2SYehuda Sadeh if (!rc) 21653d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 21663d14c5d2SYehuda Sadeh 21673d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 21683d14c5d2SYehuda Sadeh if (rc == 0) 21693d14c5d2SYehuda Sadeh rc = len; 21703d14c5d2SYehuda Sadeh dout("writepages result %d\n", rc); 21713d14c5d2SYehuda Sadeh return rc; 21723d14c5d2SYehuda Sadeh } 21733d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_writepages); 21743d14c5d2SYehuda Sadeh 21753d14c5d2SYehuda Sadeh /* 21763d14c5d2SYehuda Sadeh * handle incoming message 21773d14c5d2SYehuda Sadeh */ 21783d14c5d2SYehuda Sadeh static void dispatch(struct ceph_connection *con, struct ceph_msg *msg) 21793d14c5d2SYehuda Sadeh { 21803d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 21813d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 21823d14c5d2SYehuda Sadeh int type = le16_to_cpu(msg->hdr.type); 21833d14c5d2SYehuda Sadeh 21843d14c5d2SYehuda Sadeh if (!osd) 21853d14c5d2SYehuda Sadeh goto out; 21863d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 21873d14c5d2SYehuda Sadeh 21883d14c5d2SYehuda Sadeh switch (type) { 21893d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 21903d14c5d2SYehuda Sadeh ceph_osdc_handle_map(osdc, msg); 21913d14c5d2SYehuda Sadeh break; 21923d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 21933d14c5d2SYehuda Sadeh handle_reply(osdc, msg, con); 21943d14c5d2SYehuda Sadeh break; 2195a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 2196a40c4f10SYehuda Sadeh handle_watch_notify(osdc, msg); 2197a40c4f10SYehuda Sadeh break; 21983d14c5d2SYehuda Sadeh 21993d14c5d2SYehuda Sadeh default: 22003d14c5d2SYehuda Sadeh pr_err("received unknown message type %d %s\n", type, 22013d14c5d2SYehuda Sadeh ceph_msg_type_name(type)); 22023d14c5d2SYehuda Sadeh } 22033d14c5d2SYehuda Sadeh out: 22043d14c5d2SYehuda Sadeh ceph_msg_put(msg); 22053d14c5d2SYehuda Sadeh } 22063d14c5d2SYehuda Sadeh 22073d14c5d2SYehuda Sadeh /* 22083d14c5d2SYehuda Sadeh * lookup and return message for incoming reply. set up reply message 22093d14c5d2SYehuda Sadeh * pages. 22103d14c5d2SYehuda Sadeh */ 22113d14c5d2SYehuda Sadeh static struct ceph_msg *get_reply(struct ceph_connection *con, 22123d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 22133d14c5d2SYehuda Sadeh int *skip) 22143d14c5d2SYehuda Sadeh { 22153d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 22163d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = osd->o_osdc; 22173d14c5d2SYehuda Sadeh struct ceph_msg *m; 22183d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 22193d14c5d2SYehuda Sadeh int front = le32_to_cpu(hdr->front_len); 22203d14c5d2SYehuda Sadeh int data_len = le32_to_cpu(hdr->data_len); 22213d14c5d2SYehuda Sadeh u64 tid; 22223d14c5d2SYehuda Sadeh 22233d14c5d2SYehuda Sadeh tid = le64_to_cpu(hdr->tid); 22243d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 22253d14c5d2SYehuda Sadeh req = __lookup_request(osdc, tid); 22263d14c5d2SYehuda Sadeh if (!req) { 22273d14c5d2SYehuda Sadeh *skip = 1; 22283d14c5d2SYehuda Sadeh m = NULL; 2229756a16a5SSage Weil dout("get_reply unknown tid %llu from osd%d\n", tid, 22303d14c5d2SYehuda Sadeh osd->o_osd); 22313d14c5d2SYehuda Sadeh goto out; 22323d14c5d2SYehuda Sadeh } 22333d14c5d2SYehuda Sadeh 22343d14c5d2SYehuda Sadeh if (req->r_con_filling_msg) { 22358921d114SAlex Elder dout("%s revoking msg %p from old con %p\n", __func__, 22363d14c5d2SYehuda Sadeh req->r_reply, req->r_con_filling_msg); 22378921d114SAlex Elder ceph_msg_revoke_incoming(req->r_reply); 22380d47766fSSage Weil req->r_con_filling_msg->ops->put(req->r_con_filling_msg); 22393d14c5d2SYehuda Sadeh req->r_con_filling_msg = NULL; 22403d14c5d2SYehuda Sadeh } 22413d14c5d2SYehuda Sadeh 22423d14c5d2SYehuda Sadeh if (front > req->r_reply->front.iov_len) { 22433d14c5d2SYehuda Sadeh pr_warning("get_reply front %d > preallocated %d\n", 22443d14c5d2SYehuda Sadeh front, (int)req->r_reply->front.iov_len); 2245b61c2763SSage Weil m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS, false); 22463d14c5d2SYehuda Sadeh if (!m) 22473d14c5d2SYehuda Sadeh goto out; 22483d14c5d2SYehuda Sadeh ceph_msg_put(req->r_reply); 22493d14c5d2SYehuda Sadeh req->r_reply = m; 22503d14c5d2SYehuda Sadeh } 22513d14c5d2SYehuda Sadeh m = ceph_msg_get(req->r_reply); 22523d14c5d2SYehuda Sadeh 22533d14c5d2SYehuda Sadeh if (data_len > 0) { 22540fff87ecSAlex Elder struct ceph_osd_data *osd_data = &req->r_data_in; 22550fff87ecSAlex Elder 22560fff87ecSAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 22570fff87ecSAlex Elder if (osd_data->pages && 2258e0c59487SAlex Elder unlikely(osd_data->length < data_len)) { 22592ac2b7a6SAlex Elder 2260e0c59487SAlex Elder pr_warning("tid %lld reply has %d bytes " 2261e0c59487SAlex Elder "we had only %llu bytes ready\n", 2262e0c59487SAlex Elder tid, data_len, osd_data->length); 22633d14c5d2SYehuda Sadeh *skip = 1; 22643d14c5d2SYehuda Sadeh ceph_msg_put(m); 22653d14c5d2SYehuda Sadeh m = NULL; 22663d14c5d2SYehuda Sadeh goto out; 22673d14c5d2SYehuda Sadeh } 22683d14c5d2SYehuda Sadeh } 22692ac2b7a6SAlex Elder } 22703d14c5d2SYehuda Sadeh *skip = 0; 22710d47766fSSage Weil req->r_con_filling_msg = con->ops->get(con); 22723d14c5d2SYehuda Sadeh dout("get_reply tid %lld %p\n", tid, m); 22733d14c5d2SYehuda Sadeh 22743d14c5d2SYehuda Sadeh out: 22753d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 22763d14c5d2SYehuda Sadeh return m; 22773d14c5d2SYehuda Sadeh 22783d14c5d2SYehuda Sadeh } 22793d14c5d2SYehuda Sadeh 22803d14c5d2SYehuda Sadeh static struct ceph_msg *alloc_msg(struct ceph_connection *con, 22813d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 22823d14c5d2SYehuda Sadeh int *skip) 22833d14c5d2SYehuda Sadeh { 22843d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 22853d14c5d2SYehuda Sadeh int type = le16_to_cpu(hdr->type); 22863d14c5d2SYehuda Sadeh int front = le32_to_cpu(hdr->front_len); 22873d14c5d2SYehuda Sadeh 22881c20f2d2SAlex Elder *skip = 0; 22893d14c5d2SYehuda Sadeh switch (type) { 22903d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 2291a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 2292b61c2763SSage Weil return ceph_msg_new(type, front, GFP_NOFS, false); 22933d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 22943d14c5d2SYehuda Sadeh return get_reply(con, hdr, skip); 22953d14c5d2SYehuda Sadeh default: 22963d14c5d2SYehuda Sadeh pr_info("alloc_msg unexpected msg type %d from osd%d\n", type, 22973d14c5d2SYehuda Sadeh osd->o_osd); 22983d14c5d2SYehuda Sadeh *skip = 1; 22993d14c5d2SYehuda Sadeh return NULL; 23003d14c5d2SYehuda Sadeh } 23013d14c5d2SYehuda Sadeh } 23023d14c5d2SYehuda Sadeh 23033d14c5d2SYehuda Sadeh /* 23043d14c5d2SYehuda Sadeh * Wrappers to refcount containing ceph_osd struct 23053d14c5d2SYehuda Sadeh */ 23063d14c5d2SYehuda Sadeh static struct ceph_connection *get_osd_con(struct ceph_connection *con) 23073d14c5d2SYehuda Sadeh { 23083d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 23093d14c5d2SYehuda Sadeh if (get_osd(osd)) 23103d14c5d2SYehuda Sadeh return con; 23113d14c5d2SYehuda Sadeh return NULL; 23123d14c5d2SYehuda Sadeh } 23133d14c5d2SYehuda Sadeh 23143d14c5d2SYehuda Sadeh static void put_osd_con(struct ceph_connection *con) 23153d14c5d2SYehuda Sadeh { 23163d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 23173d14c5d2SYehuda Sadeh put_osd(osd); 23183d14c5d2SYehuda Sadeh } 23193d14c5d2SYehuda Sadeh 23203d14c5d2SYehuda Sadeh /* 23213d14c5d2SYehuda Sadeh * authentication 23223d14c5d2SYehuda Sadeh */ 2323a3530df3SAlex Elder /* 2324a3530df3SAlex Elder * Note: returned pointer is the address of a structure that's 2325a3530df3SAlex Elder * managed separately. Caller must *not* attempt to free it. 2326a3530df3SAlex Elder */ 2327a3530df3SAlex Elder static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con, 23288f43fb53SAlex Elder int *proto, int force_new) 23293d14c5d2SYehuda Sadeh { 23303d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 23313d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 23323d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 233374f1869fSAlex Elder struct ceph_auth_handshake *auth = &o->o_auth; 23343d14c5d2SYehuda Sadeh 233574f1869fSAlex Elder if (force_new && auth->authorizer) { 233627859f97SSage Weil ceph_auth_destroy_authorizer(ac, auth->authorizer); 233774f1869fSAlex Elder auth->authorizer = NULL; 23383d14c5d2SYehuda Sadeh } 233927859f97SSage Weil if (!auth->authorizer) { 234027859f97SSage Weil int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 2341a3530df3SAlex Elder auth); 23423d14c5d2SYehuda Sadeh if (ret) 2343a3530df3SAlex Elder return ERR_PTR(ret); 234427859f97SSage Weil } else { 234527859f97SSage Weil int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 23460bed9b5cSSage Weil auth); 23470bed9b5cSSage Weil if (ret) 23480bed9b5cSSage Weil return ERR_PTR(ret); 23493d14c5d2SYehuda Sadeh } 23503d14c5d2SYehuda Sadeh *proto = ac->protocol; 235174f1869fSAlex Elder 2352a3530df3SAlex Elder return auth; 23533d14c5d2SYehuda Sadeh } 23543d14c5d2SYehuda Sadeh 23553d14c5d2SYehuda Sadeh 23563d14c5d2SYehuda Sadeh static int verify_authorizer_reply(struct ceph_connection *con, int len) 23573d14c5d2SYehuda Sadeh { 23583d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 23593d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 23603d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 23613d14c5d2SYehuda Sadeh 236227859f97SSage Weil return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len); 23633d14c5d2SYehuda Sadeh } 23643d14c5d2SYehuda Sadeh 23653d14c5d2SYehuda Sadeh static int invalidate_authorizer(struct ceph_connection *con) 23663d14c5d2SYehuda Sadeh { 23673d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 23683d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 23693d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 23703d14c5d2SYehuda Sadeh 237127859f97SSage Weil ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD); 23723d14c5d2SYehuda Sadeh return ceph_monc_validate_auth(&osdc->client->monc); 23733d14c5d2SYehuda Sadeh } 23743d14c5d2SYehuda Sadeh 23753d14c5d2SYehuda Sadeh static const struct ceph_connection_operations osd_con_ops = { 23763d14c5d2SYehuda Sadeh .get = get_osd_con, 23773d14c5d2SYehuda Sadeh .put = put_osd_con, 23783d14c5d2SYehuda Sadeh .dispatch = dispatch, 23793d14c5d2SYehuda Sadeh .get_authorizer = get_authorizer, 23803d14c5d2SYehuda Sadeh .verify_authorizer_reply = verify_authorizer_reply, 23813d14c5d2SYehuda Sadeh .invalidate_authorizer = invalidate_authorizer, 23823d14c5d2SYehuda Sadeh .alloc_msg = alloc_msg, 23833d14c5d2SYehuda Sadeh .fault = osd_reset, 23843d14c5d2SYehuda Sadeh }; 2385