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 266f6c7006SSage Weil 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_needs_trail(int op) 363d14c5d2SYehuda Sadeh { 373d14c5d2SYehuda Sadeh switch (op) { 383d14c5d2SYehuda Sadeh case CEPH_OSD_OP_GETXATTR: 393d14c5d2SYehuda Sadeh case CEPH_OSD_OP_SETXATTR: 403d14c5d2SYehuda Sadeh case CEPH_OSD_OP_CMPXATTR: 413d14c5d2SYehuda Sadeh case CEPH_OSD_OP_CALL: 42a40c4f10SYehuda Sadeh case CEPH_OSD_OP_NOTIFY: 433d14c5d2SYehuda Sadeh return 1; 443d14c5d2SYehuda Sadeh default: 453d14c5d2SYehuda Sadeh return 0; 463d14c5d2SYehuda Sadeh } 473d14c5d2SYehuda Sadeh } 483d14c5d2SYehuda Sadeh 493d14c5d2SYehuda Sadeh static int op_has_extent(int op) 503d14c5d2SYehuda Sadeh { 513d14c5d2SYehuda Sadeh return (op == CEPH_OSD_OP_READ || 523d14c5d2SYehuda Sadeh op == CEPH_OSD_OP_WRITE); 533d14c5d2SYehuda Sadeh } 543d14c5d2SYehuda Sadeh 553d14c5d2SYehuda Sadeh void ceph_calc_raw_layout(struct ceph_osd_client *osdc, 563d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 573d14c5d2SYehuda Sadeh u64 snapid, 583d14c5d2SYehuda Sadeh u64 off, u64 *plen, u64 *bno, 593d14c5d2SYehuda Sadeh struct ceph_osd_request *req, 603d14c5d2SYehuda Sadeh struct ceph_osd_req_op *op) 613d14c5d2SYehuda Sadeh { 623d14c5d2SYehuda Sadeh struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base; 633d14c5d2SYehuda Sadeh u64 orig_len = *plen; 643d14c5d2SYehuda Sadeh u64 objoff, objlen; /* extent in object */ 653d14c5d2SYehuda Sadeh 663d14c5d2SYehuda Sadeh reqhead->snapid = cpu_to_le64(snapid); 673d14c5d2SYehuda Sadeh 683d14c5d2SYehuda Sadeh /* object extent? */ 693d14c5d2SYehuda Sadeh ceph_calc_file_object_mapping(layout, off, plen, bno, 703d14c5d2SYehuda Sadeh &objoff, &objlen); 713d14c5d2SYehuda Sadeh if (*plen < orig_len) 723d14c5d2SYehuda Sadeh dout(" skipping last %llu, final file extent %llu~%llu\n", 733d14c5d2SYehuda Sadeh orig_len - *plen, off, *plen); 743d14c5d2SYehuda Sadeh 753d14c5d2SYehuda Sadeh if (op_has_extent(op->op)) { 763d14c5d2SYehuda Sadeh op->extent.offset = objoff; 773d14c5d2SYehuda Sadeh op->extent.length = objlen; 783d14c5d2SYehuda Sadeh } 793d14c5d2SYehuda Sadeh req->r_num_pages = calc_pages_for(off, *plen); 80b7495fc2SSage Weil req->r_page_alignment = off & ~PAGE_MASK; 813d14c5d2SYehuda Sadeh if (op->op == CEPH_OSD_OP_WRITE) 823d14c5d2SYehuda Sadeh op->payload_len = *plen; 833d14c5d2SYehuda Sadeh 843d14c5d2SYehuda Sadeh dout("calc_layout bno=%llx %llu~%llu (%d pages)\n", 853d14c5d2SYehuda Sadeh *bno, objoff, objlen, req->r_num_pages); 863d14c5d2SYehuda Sadeh 873d14c5d2SYehuda Sadeh } 883d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_calc_raw_layout); 893d14c5d2SYehuda Sadeh 903d14c5d2SYehuda Sadeh /* 913d14c5d2SYehuda Sadeh * Implement client access to distributed object storage cluster. 923d14c5d2SYehuda Sadeh * 933d14c5d2SYehuda Sadeh * All data objects are stored within a cluster/cloud of OSDs, or 943d14c5d2SYehuda Sadeh * "object storage devices." (Note that Ceph OSDs have _nothing_ to 953d14c5d2SYehuda Sadeh * do with the T10 OSD extensions to SCSI.) Ceph OSDs are simply 963d14c5d2SYehuda Sadeh * remote daemons serving up and coordinating consistent and safe 973d14c5d2SYehuda Sadeh * access to storage. 983d14c5d2SYehuda Sadeh * 993d14c5d2SYehuda Sadeh * Cluster membership and the mapping of data objects onto storage devices 1003d14c5d2SYehuda Sadeh * are described by the osd map. 1013d14c5d2SYehuda Sadeh * 1023d14c5d2SYehuda Sadeh * We keep track of pending OSD requests (read, write), resubmit 1033d14c5d2SYehuda Sadeh * requests to different OSDs when the cluster topology/data layout 1043d14c5d2SYehuda Sadeh * change, or retry the affected requests when the communications 1053d14c5d2SYehuda Sadeh * channel with an OSD is reset. 1063d14c5d2SYehuda Sadeh */ 1073d14c5d2SYehuda Sadeh 1083d14c5d2SYehuda Sadeh /* 1093d14c5d2SYehuda Sadeh * calculate the mapping of a file extent onto an object, and fill out the 1103d14c5d2SYehuda Sadeh * request accordingly. shorten extent as necessary if it crosses an 1113d14c5d2SYehuda Sadeh * object boundary. 1123d14c5d2SYehuda Sadeh * 1133d14c5d2SYehuda Sadeh * fill osd op in request message. 1143d14c5d2SYehuda Sadeh */ 1153d14c5d2SYehuda Sadeh static void calc_layout(struct ceph_osd_client *osdc, 1163d14c5d2SYehuda Sadeh struct ceph_vino vino, 1173d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 1183d14c5d2SYehuda Sadeh u64 off, u64 *plen, 1193d14c5d2SYehuda Sadeh struct ceph_osd_request *req, 1203d14c5d2SYehuda Sadeh struct ceph_osd_req_op *op) 1213d14c5d2SYehuda Sadeh { 1223d14c5d2SYehuda Sadeh u64 bno; 1233d14c5d2SYehuda Sadeh 1243d14c5d2SYehuda Sadeh ceph_calc_raw_layout(osdc, layout, vino.snap, off, 1253d14c5d2SYehuda Sadeh plen, &bno, req, op); 1263d14c5d2SYehuda Sadeh 1272dab036bSSage Weil snprintf(req->r_oid, sizeof(req->r_oid), "%llx.%08llx", vino.ino, bno); 1283d14c5d2SYehuda Sadeh req->r_oid_len = strlen(req->r_oid); 1293d14c5d2SYehuda Sadeh } 1303d14c5d2SYehuda Sadeh 1313d14c5d2SYehuda Sadeh /* 1323d14c5d2SYehuda Sadeh * requests 1333d14c5d2SYehuda Sadeh */ 1343d14c5d2SYehuda Sadeh void ceph_osdc_release_request(struct kref *kref) 1353d14c5d2SYehuda Sadeh { 1363d14c5d2SYehuda Sadeh struct ceph_osd_request *req = container_of(kref, 1373d14c5d2SYehuda Sadeh struct ceph_osd_request, 1383d14c5d2SYehuda Sadeh r_kref); 1393d14c5d2SYehuda Sadeh 1403d14c5d2SYehuda Sadeh if (req->r_request) 1413d14c5d2SYehuda Sadeh ceph_msg_put(req->r_request); 1423d14c5d2SYehuda Sadeh if (req->r_con_filling_msg) { 1433d14c5d2SYehuda Sadeh dout("release_request revoking pages %p from con %p\n", 1443d14c5d2SYehuda Sadeh req->r_pages, req->r_con_filling_msg); 1453d14c5d2SYehuda Sadeh ceph_con_revoke_message(req->r_con_filling_msg, 1463d14c5d2SYehuda Sadeh req->r_reply); 1473d14c5d2SYehuda Sadeh ceph_con_put(req->r_con_filling_msg); 1483d14c5d2SYehuda Sadeh } 149ab8cb34aSAlex Elder if (req->r_reply) 150ab8cb34aSAlex Elder ceph_msg_put(req->r_reply); 1513d14c5d2SYehuda Sadeh if (req->r_own_pages) 1523d14c5d2SYehuda Sadeh ceph_release_page_vector(req->r_pages, 1533d14c5d2SYehuda Sadeh req->r_num_pages); 1543d14c5d2SYehuda Sadeh #ifdef CONFIG_BLOCK 1553d14c5d2SYehuda Sadeh if (req->r_bio) 1563d14c5d2SYehuda Sadeh bio_put(req->r_bio); 1573d14c5d2SYehuda Sadeh #endif 1583d14c5d2SYehuda Sadeh ceph_put_snap_context(req->r_snapc); 1593d14c5d2SYehuda Sadeh if (req->r_trail) { 1603d14c5d2SYehuda Sadeh ceph_pagelist_release(req->r_trail); 1613d14c5d2SYehuda Sadeh kfree(req->r_trail); 1623d14c5d2SYehuda Sadeh } 1633d14c5d2SYehuda Sadeh if (req->r_mempool) 1643d14c5d2SYehuda Sadeh mempool_free(req, req->r_osdc->req_mempool); 1653d14c5d2SYehuda Sadeh else 1663d14c5d2SYehuda Sadeh kfree(req); 1673d14c5d2SYehuda Sadeh } 1683d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_release_request); 1693d14c5d2SYehuda Sadeh 1703d14c5d2SYehuda Sadeh static int get_num_ops(struct ceph_osd_req_op *ops, int *needs_trail) 1713d14c5d2SYehuda Sadeh { 1723d14c5d2SYehuda Sadeh int i = 0; 1733d14c5d2SYehuda Sadeh 1743d14c5d2SYehuda Sadeh if (needs_trail) 1753d14c5d2SYehuda Sadeh *needs_trail = 0; 1763d14c5d2SYehuda Sadeh while (ops[i].op) { 1773d14c5d2SYehuda Sadeh if (needs_trail && op_needs_trail(ops[i].op)) 1783d14c5d2SYehuda Sadeh *needs_trail = 1; 1793d14c5d2SYehuda Sadeh i++; 1803d14c5d2SYehuda Sadeh } 1813d14c5d2SYehuda Sadeh 1823d14c5d2SYehuda Sadeh return i; 1833d14c5d2SYehuda Sadeh } 1843d14c5d2SYehuda Sadeh 1853d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc, 1863d14c5d2SYehuda Sadeh int flags, 1873d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 1883d14c5d2SYehuda Sadeh struct ceph_osd_req_op *ops, 1893d14c5d2SYehuda Sadeh bool use_mempool, 1903d14c5d2SYehuda Sadeh gfp_t gfp_flags, 1913d14c5d2SYehuda Sadeh struct page **pages, 1923d14c5d2SYehuda Sadeh struct bio *bio) 1933d14c5d2SYehuda Sadeh { 1943d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 1953d14c5d2SYehuda Sadeh struct ceph_msg *msg; 1963d14c5d2SYehuda Sadeh int needs_trail; 1973d14c5d2SYehuda Sadeh int num_op = get_num_ops(ops, &needs_trail); 1983d14c5d2SYehuda Sadeh size_t msg_size = sizeof(struct ceph_osd_request_head); 1993d14c5d2SYehuda Sadeh 2003d14c5d2SYehuda Sadeh msg_size += num_op*sizeof(struct ceph_osd_op); 2013d14c5d2SYehuda Sadeh 2023d14c5d2SYehuda Sadeh if (use_mempool) { 2033d14c5d2SYehuda Sadeh req = mempool_alloc(osdc->req_mempool, gfp_flags); 2043d14c5d2SYehuda Sadeh memset(req, 0, sizeof(*req)); 2053d14c5d2SYehuda Sadeh } else { 2063d14c5d2SYehuda Sadeh req = kzalloc(sizeof(*req), gfp_flags); 2073d14c5d2SYehuda Sadeh } 2083d14c5d2SYehuda Sadeh if (req == NULL) 2093d14c5d2SYehuda Sadeh return NULL; 2103d14c5d2SYehuda Sadeh 2113d14c5d2SYehuda Sadeh req->r_osdc = osdc; 2123d14c5d2SYehuda Sadeh req->r_mempool = use_mempool; 2133d14c5d2SYehuda Sadeh 2143d14c5d2SYehuda Sadeh kref_init(&req->r_kref); 2153d14c5d2SYehuda Sadeh init_completion(&req->r_completion); 2163d14c5d2SYehuda Sadeh init_completion(&req->r_safe_completion); 2173d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&req->r_unsafe_item); 218a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&req->r_linger_item); 219a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&req->r_linger_osd); 220935b639aSSage Weil INIT_LIST_HEAD(&req->r_req_lru_item); 2213d14c5d2SYehuda Sadeh req->r_flags = flags; 2223d14c5d2SYehuda Sadeh 2233d14c5d2SYehuda Sadeh WARN_ON((flags & (CEPH_OSD_FLAG_READ|CEPH_OSD_FLAG_WRITE)) == 0); 2243d14c5d2SYehuda Sadeh 2253d14c5d2SYehuda Sadeh /* create reply message */ 2263d14c5d2SYehuda Sadeh if (use_mempool) 2273d14c5d2SYehuda Sadeh msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0); 2283d14c5d2SYehuda Sadeh else 2293d14c5d2SYehuda Sadeh msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, 230b61c2763SSage Weil OSD_OPREPLY_FRONT_LEN, gfp_flags, true); 2313d14c5d2SYehuda Sadeh if (!msg) { 2323d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 2333d14c5d2SYehuda Sadeh return NULL; 2343d14c5d2SYehuda Sadeh } 2353d14c5d2SYehuda Sadeh req->r_reply = msg; 2363d14c5d2SYehuda Sadeh 2373d14c5d2SYehuda Sadeh /* allocate space for the trailing data */ 2383d14c5d2SYehuda Sadeh if (needs_trail) { 2393d14c5d2SYehuda Sadeh req->r_trail = kmalloc(sizeof(struct ceph_pagelist), gfp_flags); 2403d14c5d2SYehuda Sadeh if (!req->r_trail) { 2413d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 2423d14c5d2SYehuda Sadeh return NULL; 2433d14c5d2SYehuda Sadeh } 2443d14c5d2SYehuda Sadeh ceph_pagelist_init(req->r_trail); 2453d14c5d2SYehuda Sadeh } 2463d14c5d2SYehuda Sadeh /* create request message; allow space for oid */ 247224736d9SStratos Psomadakis msg_size += MAX_OBJ_NAME_SIZE; 2483d14c5d2SYehuda Sadeh if (snapc) 2493d14c5d2SYehuda Sadeh msg_size += sizeof(u64) * snapc->num_snaps; 2503d14c5d2SYehuda Sadeh if (use_mempool) 2513d14c5d2SYehuda Sadeh msg = ceph_msgpool_get(&osdc->msgpool_op, 0); 2523d14c5d2SYehuda Sadeh else 253b61c2763SSage Weil msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true); 2543d14c5d2SYehuda Sadeh if (!msg) { 2553d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 2563d14c5d2SYehuda Sadeh return NULL; 2573d14c5d2SYehuda Sadeh } 2583d14c5d2SYehuda Sadeh 2593d14c5d2SYehuda Sadeh msg->hdr.type = cpu_to_le16(CEPH_MSG_OSD_OP); 2603d14c5d2SYehuda Sadeh memset(msg->front.iov_base, 0, msg->front.iov_len); 2613d14c5d2SYehuda Sadeh 2623d14c5d2SYehuda Sadeh req->r_request = msg; 2633d14c5d2SYehuda Sadeh req->r_pages = pages; 2643d14c5d2SYehuda Sadeh #ifdef CONFIG_BLOCK 2653d14c5d2SYehuda Sadeh if (bio) { 2663d14c5d2SYehuda Sadeh req->r_bio = bio; 2673d14c5d2SYehuda Sadeh bio_get(req->r_bio); 2683d14c5d2SYehuda Sadeh } 2693d14c5d2SYehuda Sadeh #endif 2703d14c5d2SYehuda Sadeh 2713d14c5d2SYehuda Sadeh return req; 2723d14c5d2SYehuda Sadeh } 2733d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_alloc_request); 2743d14c5d2SYehuda Sadeh 2753d14c5d2SYehuda Sadeh static void osd_req_encode_op(struct ceph_osd_request *req, 2763d14c5d2SYehuda Sadeh struct ceph_osd_op *dst, 2773d14c5d2SYehuda Sadeh struct ceph_osd_req_op *src) 2783d14c5d2SYehuda Sadeh { 2793d14c5d2SYehuda Sadeh dst->op = cpu_to_le16(src->op); 2803d14c5d2SYehuda Sadeh 281065a68f9SAlex Elder switch (src->op) { 2823d14c5d2SYehuda Sadeh case CEPH_OSD_OP_READ: 2833d14c5d2SYehuda Sadeh case CEPH_OSD_OP_WRITE: 2843d14c5d2SYehuda Sadeh dst->extent.offset = 2853d14c5d2SYehuda Sadeh cpu_to_le64(src->extent.offset); 2863d14c5d2SYehuda Sadeh dst->extent.length = 2873d14c5d2SYehuda Sadeh cpu_to_le64(src->extent.length); 2883d14c5d2SYehuda Sadeh dst->extent.truncate_size = 2893d14c5d2SYehuda Sadeh cpu_to_le64(src->extent.truncate_size); 2903d14c5d2SYehuda Sadeh dst->extent.truncate_seq = 2913d14c5d2SYehuda Sadeh cpu_to_le32(src->extent.truncate_seq); 2923d14c5d2SYehuda Sadeh break; 2933d14c5d2SYehuda Sadeh 2943d14c5d2SYehuda Sadeh case CEPH_OSD_OP_GETXATTR: 2953d14c5d2SYehuda Sadeh case CEPH_OSD_OP_SETXATTR: 2963d14c5d2SYehuda Sadeh case CEPH_OSD_OP_CMPXATTR: 2973d14c5d2SYehuda Sadeh BUG_ON(!req->r_trail); 2983d14c5d2SYehuda Sadeh 2993d14c5d2SYehuda Sadeh dst->xattr.name_len = cpu_to_le32(src->xattr.name_len); 3003d14c5d2SYehuda Sadeh dst->xattr.value_len = cpu_to_le32(src->xattr.value_len); 3013d14c5d2SYehuda Sadeh dst->xattr.cmp_op = src->xattr.cmp_op; 3023d14c5d2SYehuda Sadeh dst->xattr.cmp_mode = src->xattr.cmp_mode; 3033d14c5d2SYehuda Sadeh ceph_pagelist_append(req->r_trail, src->xattr.name, 3043d14c5d2SYehuda Sadeh src->xattr.name_len); 3053d14c5d2SYehuda Sadeh ceph_pagelist_append(req->r_trail, src->xattr.val, 3063d14c5d2SYehuda Sadeh src->xattr.value_len); 3073d14c5d2SYehuda Sadeh break; 3083d14c5d2SYehuda Sadeh case CEPH_OSD_OP_CALL: 3093d14c5d2SYehuda Sadeh BUG_ON(!req->r_trail); 3103d14c5d2SYehuda Sadeh 3113d14c5d2SYehuda Sadeh dst->cls.class_len = src->cls.class_len; 3123d14c5d2SYehuda Sadeh dst->cls.method_len = src->cls.method_len; 3133d14c5d2SYehuda Sadeh dst->cls.indata_len = cpu_to_le32(src->cls.indata_len); 3143d14c5d2SYehuda Sadeh 3153d14c5d2SYehuda Sadeh ceph_pagelist_append(req->r_trail, src->cls.class_name, 3163d14c5d2SYehuda Sadeh src->cls.class_len); 3173d14c5d2SYehuda Sadeh ceph_pagelist_append(req->r_trail, src->cls.method_name, 3183d14c5d2SYehuda Sadeh src->cls.method_len); 3193d14c5d2SYehuda Sadeh ceph_pagelist_append(req->r_trail, src->cls.indata, 3203d14c5d2SYehuda Sadeh src->cls.indata_len); 3213d14c5d2SYehuda Sadeh break; 3223d14c5d2SYehuda Sadeh case CEPH_OSD_OP_ROLLBACK: 3233d14c5d2SYehuda Sadeh dst->snap.snapid = cpu_to_le64(src->snap.snapid); 3243d14c5d2SYehuda Sadeh break; 3253d14c5d2SYehuda Sadeh case CEPH_OSD_OP_STARTSYNC: 3263d14c5d2SYehuda Sadeh break; 327a40c4f10SYehuda Sadeh case CEPH_OSD_OP_NOTIFY: 328a40c4f10SYehuda Sadeh { 329a40c4f10SYehuda Sadeh __le32 prot_ver = cpu_to_le32(src->watch.prot_ver); 330a40c4f10SYehuda Sadeh __le32 timeout = cpu_to_le32(src->watch.timeout); 331a40c4f10SYehuda Sadeh 332a40c4f10SYehuda Sadeh BUG_ON(!req->r_trail); 333a40c4f10SYehuda Sadeh 334a40c4f10SYehuda Sadeh ceph_pagelist_append(req->r_trail, 335a40c4f10SYehuda Sadeh &prot_ver, sizeof(prot_ver)); 336a40c4f10SYehuda Sadeh ceph_pagelist_append(req->r_trail, 337a40c4f10SYehuda Sadeh &timeout, sizeof(timeout)); 338a40c4f10SYehuda Sadeh } 339a40c4f10SYehuda Sadeh case CEPH_OSD_OP_NOTIFY_ACK: 340a40c4f10SYehuda Sadeh case CEPH_OSD_OP_WATCH: 341a40c4f10SYehuda Sadeh dst->watch.cookie = cpu_to_le64(src->watch.cookie); 342a40c4f10SYehuda Sadeh dst->watch.ver = cpu_to_le64(src->watch.ver); 343a40c4f10SYehuda Sadeh dst->watch.flag = src->watch.flag; 344a40c4f10SYehuda Sadeh break; 3453d14c5d2SYehuda Sadeh default: 3463d14c5d2SYehuda Sadeh pr_err("unrecognized osd opcode %d\n", dst->op); 3473d14c5d2SYehuda Sadeh WARN_ON(1); 3483d14c5d2SYehuda Sadeh break; 3493d14c5d2SYehuda Sadeh } 3503d14c5d2SYehuda Sadeh dst->payload_len = cpu_to_le32(src->payload_len); 3513d14c5d2SYehuda Sadeh } 3523d14c5d2SYehuda Sadeh 3533d14c5d2SYehuda Sadeh /* 3543d14c5d2SYehuda Sadeh * build new request AND message 3553d14c5d2SYehuda Sadeh * 3563d14c5d2SYehuda Sadeh */ 3573d14c5d2SYehuda Sadeh void ceph_osdc_build_request(struct ceph_osd_request *req, 3583d14c5d2SYehuda Sadeh u64 off, u64 *plen, 3593d14c5d2SYehuda Sadeh struct ceph_osd_req_op *src_ops, 3603d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 3613d14c5d2SYehuda Sadeh struct timespec *mtime, 3623d14c5d2SYehuda Sadeh const char *oid, 3633d14c5d2SYehuda Sadeh int oid_len) 3643d14c5d2SYehuda Sadeh { 3653d14c5d2SYehuda Sadeh struct ceph_msg *msg = req->r_request; 3663d14c5d2SYehuda Sadeh struct ceph_osd_request_head *head; 3673d14c5d2SYehuda Sadeh struct ceph_osd_req_op *src_op; 3683d14c5d2SYehuda Sadeh struct ceph_osd_op *op; 3693d14c5d2SYehuda Sadeh void *p; 3703d14c5d2SYehuda Sadeh int num_op = get_num_ops(src_ops, NULL); 3713d14c5d2SYehuda Sadeh size_t msg_size = sizeof(*head) + num_op*sizeof(*op); 3723d14c5d2SYehuda Sadeh int flags = req->r_flags; 3733d14c5d2SYehuda Sadeh u64 data_len = 0; 3743d14c5d2SYehuda Sadeh int i; 3753d14c5d2SYehuda Sadeh 3763d14c5d2SYehuda Sadeh head = msg->front.iov_base; 3773d14c5d2SYehuda Sadeh op = (void *)(head + 1); 3783d14c5d2SYehuda Sadeh p = (void *)(op + num_op); 3793d14c5d2SYehuda Sadeh 3803d14c5d2SYehuda Sadeh req->r_snapc = ceph_get_snap_context(snapc); 3813d14c5d2SYehuda Sadeh 3823d14c5d2SYehuda Sadeh head->client_inc = cpu_to_le32(1); /* always, for now. */ 3833d14c5d2SYehuda Sadeh head->flags = cpu_to_le32(flags); 3843d14c5d2SYehuda Sadeh if (flags & CEPH_OSD_FLAG_WRITE) 3853d14c5d2SYehuda Sadeh ceph_encode_timespec(&head->mtime, mtime); 3863d14c5d2SYehuda Sadeh head->num_ops = cpu_to_le16(num_op); 3873d14c5d2SYehuda Sadeh 3883d14c5d2SYehuda Sadeh 3893d14c5d2SYehuda Sadeh /* fill in oid */ 3903d14c5d2SYehuda Sadeh head->object_len = cpu_to_le32(oid_len); 3913d14c5d2SYehuda Sadeh memcpy(p, oid, oid_len); 3923d14c5d2SYehuda Sadeh p += oid_len; 3933d14c5d2SYehuda Sadeh 3943d14c5d2SYehuda Sadeh src_op = src_ops; 3953d14c5d2SYehuda Sadeh while (src_op->op) { 3963d14c5d2SYehuda Sadeh osd_req_encode_op(req, op, src_op); 3973d14c5d2SYehuda Sadeh src_op++; 3983d14c5d2SYehuda Sadeh op++; 3993d14c5d2SYehuda Sadeh } 4003d14c5d2SYehuda Sadeh 4013d14c5d2SYehuda Sadeh if (req->r_trail) 4023d14c5d2SYehuda Sadeh data_len += req->r_trail->length; 4033d14c5d2SYehuda Sadeh 4043d14c5d2SYehuda Sadeh if (snapc) { 4053d14c5d2SYehuda Sadeh head->snap_seq = cpu_to_le64(snapc->seq); 4063d14c5d2SYehuda Sadeh head->num_snaps = cpu_to_le32(snapc->num_snaps); 4073d14c5d2SYehuda Sadeh for (i = 0; i < snapc->num_snaps; i++) { 4083d14c5d2SYehuda Sadeh put_unaligned_le64(snapc->snaps[i], p); 4093d14c5d2SYehuda Sadeh p += sizeof(u64); 4103d14c5d2SYehuda Sadeh } 4113d14c5d2SYehuda Sadeh } 4123d14c5d2SYehuda Sadeh 4133d14c5d2SYehuda Sadeh if (flags & CEPH_OSD_FLAG_WRITE) { 4143d14c5d2SYehuda Sadeh req->r_request->hdr.data_off = cpu_to_le16(off); 4153d14c5d2SYehuda Sadeh req->r_request->hdr.data_len = cpu_to_le32(*plen + data_len); 4163d14c5d2SYehuda Sadeh } else if (data_len) { 4173d14c5d2SYehuda Sadeh req->r_request->hdr.data_off = 0; 4183d14c5d2SYehuda Sadeh req->r_request->hdr.data_len = cpu_to_le32(data_len); 4193d14c5d2SYehuda Sadeh } 4203d14c5d2SYehuda Sadeh 421c5c6b19dSSage Weil req->r_request->page_alignment = req->r_page_alignment; 422c5c6b19dSSage Weil 4233d14c5d2SYehuda Sadeh BUG_ON(p > msg->front.iov_base + msg->front.iov_len); 4243d14c5d2SYehuda Sadeh msg_size = p - msg->front.iov_base; 4253d14c5d2SYehuda Sadeh msg->front.iov_len = msg_size; 4263d14c5d2SYehuda Sadeh msg->hdr.front_len = cpu_to_le32(msg_size); 4273d14c5d2SYehuda Sadeh return; 4283d14c5d2SYehuda Sadeh } 4293d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_build_request); 4303d14c5d2SYehuda Sadeh 4313d14c5d2SYehuda Sadeh /* 4323d14c5d2SYehuda Sadeh * build new request AND message, calculate layout, and adjust file 4333d14c5d2SYehuda Sadeh * extent as needed. 4343d14c5d2SYehuda Sadeh * 4353d14c5d2SYehuda Sadeh * if the file was recently truncated, we include information about its 4363d14c5d2SYehuda Sadeh * old and new size so that the object can be updated appropriately. (we 4373d14c5d2SYehuda Sadeh * avoid synchronously deleting truncated objects because it's slow.) 4383d14c5d2SYehuda Sadeh * 4393d14c5d2SYehuda Sadeh * if @do_sync, include a 'startsync' command so that the osd will flush 4403d14c5d2SYehuda Sadeh * data quickly. 4413d14c5d2SYehuda Sadeh */ 4423d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc, 4433d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 4443d14c5d2SYehuda Sadeh struct ceph_vino vino, 4453d14c5d2SYehuda Sadeh u64 off, u64 *plen, 4463d14c5d2SYehuda Sadeh int opcode, int flags, 4473d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 4483d14c5d2SYehuda Sadeh int do_sync, 4493d14c5d2SYehuda Sadeh u32 truncate_seq, 4503d14c5d2SYehuda Sadeh u64 truncate_size, 4513d14c5d2SYehuda Sadeh struct timespec *mtime, 452b7495fc2SSage Weil bool use_mempool, int num_reply, 453b7495fc2SSage Weil int page_align) 4543d14c5d2SYehuda Sadeh { 4553d14c5d2SYehuda Sadeh struct ceph_osd_req_op ops[3]; 4563d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 4573d14c5d2SYehuda Sadeh 4583d14c5d2SYehuda Sadeh ops[0].op = opcode; 4593d14c5d2SYehuda Sadeh ops[0].extent.truncate_seq = truncate_seq; 4603d14c5d2SYehuda Sadeh ops[0].extent.truncate_size = truncate_size; 4613d14c5d2SYehuda Sadeh ops[0].payload_len = 0; 4623d14c5d2SYehuda Sadeh 4633d14c5d2SYehuda Sadeh if (do_sync) { 4643d14c5d2SYehuda Sadeh ops[1].op = CEPH_OSD_OP_STARTSYNC; 4653d14c5d2SYehuda Sadeh ops[1].payload_len = 0; 4663d14c5d2SYehuda Sadeh ops[2].op = 0; 4673d14c5d2SYehuda Sadeh } else 4683d14c5d2SYehuda Sadeh ops[1].op = 0; 4693d14c5d2SYehuda Sadeh 4703d14c5d2SYehuda Sadeh req = ceph_osdc_alloc_request(osdc, flags, 4713d14c5d2SYehuda Sadeh snapc, ops, 4723d14c5d2SYehuda Sadeh use_mempool, 4733d14c5d2SYehuda Sadeh GFP_NOFS, NULL, NULL); 4744ad12621SSage Weil if (!req) 4754ad12621SSage Weil return NULL; 4763d14c5d2SYehuda Sadeh 4773d14c5d2SYehuda Sadeh /* calculate max write size */ 4783d14c5d2SYehuda Sadeh calc_layout(osdc, vino, layout, off, plen, req, ops); 4793d14c5d2SYehuda Sadeh req->r_file_layout = *layout; /* keep a copy */ 4803d14c5d2SYehuda Sadeh 4819bb0ce2bSSage Weil /* in case it differs from natural (file) alignment that 4829bb0ce2bSSage Weil calc_layout filled in for us */ 4839bb0ce2bSSage Weil req->r_num_pages = calc_pages_for(page_align, *plen); 484b7495fc2SSage Weil req->r_page_alignment = page_align; 485b7495fc2SSage Weil 4863d14c5d2SYehuda Sadeh ceph_osdc_build_request(req, off, plen, ops, 4873d14c5d2SYehuda Sadeh snapc, 4883d14c5d2SYehuda Sadeh mtime, 4893d14c5d2SYehuda Sadeh req->r_oid, req->r_oid_len); 4903d14c5d2SYehuda Sadeh 4913d14c5d2SYehuda Sadeh return req; 4923d14c5d2SYehuda Sadeh } 4933d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_new_request); 4943d14c5d2SYehuda Sadeh 4953d14c5d2SYehuda Sadeh /* 4963d14c5d2SYehuda Sadeh * We keep osd requests in an rbtree, sorted by ->r_tid. 4973d14c5d2SYehuda Sadeh */ 4983d14c5d2SYehuda Sadeh static void __insert_request(struct ceph_osd_client *osdc, 4993d14c5d2SYehuda Sadeh struct ceph_osd_request *new) 5003d14c5d2SYehuda Sadeh { 5013d14c5d2SYehuda Sadeh struct rb_node **p = &osdc->requests.rb_node; 5023d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 5033d14c5d2SYehuda Sadeh struct ceph_osd_request *req = NULL; 5043d14c5d2SYehuda Sadeh 5053d14c5d2SYehuda Sadeh while (*p) { 5063d14c5d2SYehuda Sadeh parent = *p; 5073d14c5d2SYehuda Sadeh req = rb_entry(parent, struct ceph_osd_request, r_node); 5083d14c5d2SYehuda Sadeh if (new->r_tid < req->r_tid) 5093d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 5103d14c5d2SYehuda Sadeh else if (new->r_tid > req->r_tid) 5113d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 5123d14c5d2SYehuda Sadeh else 5133d14c5d2SYehuda Sadeh BUG(); 5143d14c5d2SYehuda Sadeh } 5153d14c5d2SYehuda Sadeh 5163d14c5d2SYehuda Sadeh rb_link_node(&new->r_node, parent, p); 5173d14c5d2SYehuda Sadeh rb_insert_color(&new->r_node, &osdc->requests); 5183d14c5d2SYehuda Sadeh } 5193d14c5d2SYehuda Sadeh 5203d14c5d2SYehuda Sadeh static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc, 5213d14c5d2SYehuda Sadeh u64 tid) 5223d14c5d2SYehuda Sadeh { 5233d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 5243d14c5d2SYehuda Sadeh struct rb_node *n = osdc->requests.rb_node; 5253d14c5d2SYehuda Sadeh 5263d14c5d2SYehuda Sadeh while (n) { 5273d14c5d2SYehuda Sadeh req = rb_entry(n, struct ceph_osd_request, r_node); 5283d14c5d2SYehuda Sadeh if (tid < req->r_tid) 5293d14c5d2SYehuda Sadeh n = n->rb_left; 5303d14c5d2SYehuda Sadeh else if (tid > req->r_tid) 5313d14c5d2SYehuda Sadeh n = n->rb_right; 5323d14c5d2SYehuda Sadeh else 5333d14c5d2SYehuda Sadeh return req; 5343d14c5d2SYehuda Sadeh } 5353d14c5d2SYehuda Sadeh return NULL; 5363d14c5d2SYehuda Sadeh } 5373d14c5d2SYehuda Sadeh 5383d14c5d2SYehuda Sadeh static struct ceph_osd_request * 5393d14c5d2SYehuda Sadeh __lookup_request_ge(struct ceph_osd_client *osdc, 5403d14c5d2SYehuda Sadeh u64 tid) 5413d14c5d2SYehuda Sadeh { 5423d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 5433d14c5d2SYehuda Sadeh struct rb_node *n = osdc->requests.rb_node; 5443d14c5d2SYehuda Sadeh 5453d14c5d2SYehuda Sadeh while (n) { 5463d14c5d2SYehuda Sadeh req = rb_entry(n, struct ceph_osd_request, r_node); 5473d14c5d2SYehuda Sadeh if (tid < req->r_tid) { 5483d14c5d2SYehuda Sadeh if (!n->rb_left) 5493d14c5d2SYehuda Sadeh return req; 5503d14c5d2SYehuda Sadeh n = n->rb_left; 5513d14c5d2SYehuda Sadeh } else if (tid > req->r_tid) { 5523d14c5d2SYehuda Sadeh n = n->rb_right; 5533d14c5d2SYehuda Sadeh } else { 5543d14c5d2SYehuda Sadeh return req; 5553d14c5d2SYehuda Sadeh } 5563d14c5d2SYehuda Sadeh } 5573d14c5d2SYehuda Sadeh return NULL; 5583d14c5d2SYehuda Sadeh } 5593d14c5d2SYehuda Sadeh 5606f6c7006SSage Weil /* 5616f6c7006SSage Weil * Resubmit requests pending on the given osd. 5626f6c7006SSage Weil */ 5636f6c7006SSage Weil static void __kick_osd_requests(struct ceph_osd_client *osdc, 5646f6c7006SSage Weil struct ceph_osd *osd) 5656f6c7006SSage Weil { 566a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 5676f6c7006SSage Weil int err; 5686f6c7006SSage Weil 5696f6c7006SSage Weil dout("__kick_osd_requests osd%d\n", osd->o_osd); 5706f6c7006SSage Weil err = __reset_osd(osdc, osd); 5716f6c7006SSage Weil if (err == -EAGAIN) 5726f6c7006SSage Weil return; 5736f6c7006SSage Weil 5746f6c7006SSage Weil list_for_each_entry(req, &osd->o_requests, r_osd_item) { 5756f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_unsent); 5766f6c7006SSage Weil dout("requeued %p tid %llu osd%d\n", req, req->r_tid, 5776f6c7006SSage Weil osd->o_osd); 578a40c4f10SYehuda Sadeh if (!req->r_linger) 5796f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 5806f6c7006SSage Weil } 581a40c4f10SYehuda Sadeh 582a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osd->o_linger_requests, 583a40c4f10SYehuda Sadeh r_linger_osd) { 58477f38e0eSSage Weil /* 58577f38e0eSSage Weil * reregister request prior to unregistering linger so 58677f38e0eSSage Weil * that r_osd is preserved. 58777f38e0eSSage Weil */ 58877f38e0eSSage Weil BUG_ON(!list_empty(&req->r_req_lru_item)); 589a40c4f10SYehuda Sadeh __register_request(osdc, req); 59077f38e0eSSage Weil list_add(&req->r_req_lru_item, &osdc->req_unsent); 59177f38e0eSSage Weil list_add(&req->r_osd_item, &req->r_osd->o_requests); 59277f38e0eSSage Weil __unregister_linger_request(osdc, req); 593a40c4f10SYehuda Sadeh dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid, 594a40c4f10SYehuda Sadeh osd->o_osd); 595a40c4f10SYehuda Sadeh } 5966f6c7006SSage Weil } 5976f6c7006SSage Weil 5986f6c7006SSage Weil static void kick_osd_requests(struct ceph_osd_client *osdc, 5996f6c7006SSage Weil struct ceph_osd *kickosd) 6006f6c7006SSage Weil { 6016f6c7006SSage Weil mutex_lock(&osdc->request_mutex); 6026f6c7006SSage Weil __kick_osd_requests(osdc, kickosd); 6036f6c7006SSage Weil mutex_unlock(&osdc->request_mutex); 6046f6c7006SSage Weil } 6053d14c5d2SYehuda Sadeh 6063d14c5d2SYehuda Sadeh /* 6073d14c5d2SYehuda Sadeh * If the osd connection drops, we need to resubmit all requests. 6083d14c5d2SYehuda Sadeh */ 6093d14c5d2SYehuda Sadeh static void osd_reset(struct ceph_connection *con) 6103d14c5d2SYehuda Sadeh { 6113d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 6123d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 6133d14c5d2SYehuda Sadeh 6143d14c5d2SYehuda Sadeh if (!osd) 6153d14c5d2SYehuda Sadeh return; 6163d14c5d2SYehuda Sadeh dout("osd_reset osd%d\n", osd->o_osd); 6173d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 6183d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 6196f6c7006SSage Weil kick_osd_requests(osdc, osd); 6206f6c7006SSage Weil send_queued(osdc); 6213d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 6223d14c5d2SYehuda Sadeh } 6233d14c5d2SYehuda Sadeh 6243d14c5d2SYehuda Sadeh /* 6253d14c5d2SYehuda Sadeh * Track open sessions with osds. 6263d14c5d2SYehuda Sadeh */ 627e10006f8SAlex Elder static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum) 6283d14c5d2SYehuda Sadeh { 6293d14c5d2SYehuda Sadeh struct ceph_osd *osd; 6303d14c5d2SYehuda Sadeh 6313d14c5d2SYehuda Sadeh osd = kzalloc(sizeof(*osd), GFP_NOFS); 6323d14c5d2SYehuda Sadeh if (!osd) 6333d14c5d2SYehuda Sadeh return NULL; 6343d14c5d2SYehuda Sadeh 6353d14c5d2SYehuda Sadeh atomic_set(&osd->o_ref, 1); 6363d14c5d2SYehuda Sadeh osd->o_osdc = osdc; 637e10006f8SAlex Elder osd->o_osd = onum; 6383d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_requests); 639a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osd->o_linger_requests); 6403d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_osd_lru); 6413d14c5d2SYehuda Sadeh osd->o_incarnation = 1; 6423d14c5d2SYehuda Sadeh 64315d9882cSAlex Elder ceph_con_init(&osdc->client->msgr, &osd->o_con); 6443d14c5d2SYehuda Sadeh osd->o_con.private = osd; 6453d14c5d2SYehuda Sadeh osd->o_con.ops = &osd_con_ops; 6463d14c5d2SYehuda Sadeh osd->o_con.peer_name.type = CEPH_ENTITY_TYPE_OSD; 647e10006f8SAlex Elder osd->o_con.peer_name.num = cpu_to_le64(onum); 6483d14c5d2SYehuda Sadeh 6493d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_keepalive_item); 6503d14c5d2SYehuda Sadeh return osd; 6513d14c5d2SYehuda Sadeh } 6523d14c5d2SYehuda Sadeh 6533d14c5d2SYehuda Sadeh static struct ceph_osd *get_osd(struct ceph_osd *osd) 6543d14c5d2SYehuda Sadeh { 6553d14c5d2SYehuda Sadeh if (atomic_inc_not_zero(&osd->o_ref)) { 6563d14c5d2SYehuda Sadeh dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1, 6573d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref)); 6583d14c5d2SYehuda Sadeh return osd; 6593d14c5d2SYehuda Sadeh } else { 6603d14c5d2SYehuda Sadeh dout("get_osd %p FAIL\n", osd); 6613d14c5d2SYehuda Sadeh return NULL; 6623d14c5d2SYehuda Sadeh } 6633d14c5d2SYehuda Sadeh } 6643d14c5d2SYehuda Sadeh 6653d14c5d2SYehuda Sadeh static void put_osd(struct ceph_osd *osd) 6663d14c5d2SYehuda Sadeh { 6673d14c5d2SYehuda Sadeh dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref), 6683d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref) - 1); 669a255651dSAlex Elder if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) { 6703d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth; 6713d14c5d2SYehuda Sadeh 672a255651dSAlex Elder if (ac->ops && ac->ops->destroy_authorizer) 6736c4a1915SAlex Elder ac->ops->destroy_authorizer(ac, osd->o_auth.authorizer); 6743d14c5d2SYehuda Sadeh kfree(osd); 6753d14c5d2SYehuda Sadeh } 6763d14c5d2SYehuda Sadeh } 6773d14c5d2SYehuda Sadeh 6783d14c5d2SYehuda Sadeh /* 6793d14c5d2SYehuda Sadeh * remove an osd from our map 6803d14c5d2SYehuda Sadeh */ 6813d14c5d2SYehuda Sadeh static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 6823d14c5d2SYehuda Sadeh { 6833d14c5d2SYehuda Sadeh dout("__remove_osd %p\n", osd); 6843d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_requests)); 6853d14c5d2SYehuda Sadeh rb_erase(&osd->o_node, &osdc->osds); 6863d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 6873d14c5d2SYehuda Sadeh ceph_con_close(&osd->o_con); 6883d14c5d2SYehuda Sadeh put_osd(osd); 6893d14c5d2SYehuda Sadeh } 6903d14c5d2SYehuda Sadeh 691aca420bcSSage Weil static void remove_all_osds(struct ceph_osd_client *osdc) 692aca420bcSSage Weil { 693aca420bcSSage Weil dout("__remove_old_osds %p\n", osdc); 694aca420bcSSage Weil mutex_lock(&osdc->request_mutex); 695aca420bcSSage Weil while (!RB_EMPTY_ROOT(&osdc->osds)) { 696aca420bcSSage Weil struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds), 697aca420bcSSage Weil struct ceph_osd, o_node); 698aca420bcSSage Weil __remove_osd(osdc, osd); 699aca420bcSSage Weil } 700aca420bcSSage Weil mutex_unlock(&osdc->request_mutex); 701aca420bcSSage Weil } 702aca420bcSSage Weil 7033d14c5d2SYehuda Sadeh static void __move_osd_to_lru(struct ceph_osd_client *osdc, 7043d14c5d2SYehuda Sadeh struct ceph_osd *osd) 7053d14c5d2SYehuda Sadeh { 7063d14c5d2SYehuda Sadeh dout("__move_osd_to_lru %p\n", osd); 7073d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_osd_lru)); 7083d14c5d2SYehuda Sadeh list_add_tail(&osd->o_osd_lru, &osdc->osd_lru); 7093d14c5d2SYehuda Sadeh osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ; 7103d14c5d2SYehuda Sadeh } 7113d14c5d2SYehuda Sadeh 7123d14c5d2SYehuda Sadeh static void __remove_osd_from_lru(struct ceph_osd *osd) 7133d14c5d2SYehuda Sadeh { 7143d14c5d2SYehuda Sadeh dout("__remove_osd_from_lru %p\n", osd); 7153d14c5d2SYehuda Sadeh if (!list_empty(&osd->o_osd_lru)) 7163d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 7173d14c5d2SYehuda Sadeh } 7183d14c5d2SYehuda Sadeh 719aca420bcSSage Weil static void remove_old_osds(struct ceph_osd_client *osdc) 7203d14c5d2SYehuda Sadeh { 7213d14c5d2SYehuda Sadeh struct ceph_osd *osd, *nosd; 7223d14c5d2SYehuda Sadeh 7233d14c5d2SYehuda Sadeh dout("__remove_old_osds %p\n", osdc); 7243d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 7253d14c5d2SYehuda Sadeh list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) { 726aca420bcSSage Weil if (time_before(jiffies, osd->lru_ttl)) 7273d14c5d2SYehuda Sadeh break; 7283d14c5d2SYehuda Sadeh __remove_osd(osdc, osd); 7293d14c5d2SYehuda Sadeh } 7303d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 7313d14c5d2SYehuda Sadeh } 7323d14c5d2SYehuda Sadeh 7333d14c5d2SYehuda Sadeh /* 7343d14c5d2SYehuda Sadeh * reset osd connect 7353d14c5d2SYehuda Sadeh */ 7363d14c5d2SYehuda Sadeh static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 7373d14c5d2SYehuda Sadeh { 7383d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 7393d14c5d2SYehuda Sadeh int ret = 0; 7403d14c5d2SYehuda Sadeh 7413d14c5d2SYehuda Sadeh dout("__reset_osd %p osd%d\n", osd, osd->o_osd); 742a40c4f10SYehuda Sadeh if (list_empty(&osd->o_requests) && 743a40c4f10SYehuda Sadeh list_empty(&osd->o_linger_requests)) { 7443d14c5d2SYehuda Sadeh __remove_osd(osdc, osd); 7453d14c5d2SYehuda Sadeh } else if (memcmp(&osdc->osdmap->osd_addr[osd->o_osd], 7463d14c5d2SYehuda Sadeh &osd->o_con.peer_addr, 7473d14c5d2SYehuda Sadeh sizeof(osd->o_con.peer_addr)) == 0 && 7483d14c5d2SYehuda Sadeh !ceph_con_opened(&osd->o_con)) { 7493d14c5d2SYehuda Sadeh dout(" osd addr hasn't changed and connection never opened," 7503d14c5d2SYehuda Sadeh " letting msgr retry"); 7513d14c5d2SYehuda Sadeh /* touch each r_stamp for handle_timeout()'s benfit */ 7523d14c5d2SYehuda Sadeh list_for_each_entry(req, &osd->o_requests, r_osd_item) 7533d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 7543d14c5d2SYehuda Sadeh ret = -EAGAIN; 7553d14c5d2SYehuda Sadeh } else { 7563d14c5d2SYehuda Sadeh ceph_con_close(&osd->o_con); 7573d14c5d2SYehuda Sadeh ceph_con_open(&osd->o_con, &osdc->osdmap->osd_addr[osd->o_osd]); 7583d14c5d2SYehuda Sadeh osd->o_incarnation++; 7593d14c5d2SYehuda Sadeh } 7603d14c5d2SYehuda Sadeh return ret; 7613d14c5d2SYehuda Sadeh } 7623d14c5d2SYehuda Sadeh 7633d14c5d2SYehuda Sadeh static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new) 7643d14c5d2SYehuda Sadeh { 7653d14c5d2SYehuda Sadeh struct rb_node **p = &osdc->osds.rb_node; 7663d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 7673d14c5d2SYehuda Sadeh struct ceph_osd *osd = NULL; 7683d14c5d2SYehuda Sadeh 769aca420bcSSage Weil dout("__insert_osd %p osd%d\n", new, new->o_osd); 7703d14c5d2SYehuda Sadeh while (*p) { 7713d14c5d2SYehuda Sadeh parent = *p; 7723d14c5d2SYehuda Sadeh osd = rb_entry(parent, struct ceph_osd, o_node); 7733d14c5d2SYehuda Sadeh if (new->o_osd < osd->o_osd) 7743d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 7753d14c5d2SYehuda Sadeh else if (new->o_osd > osd->o_osd) 7763d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 7773d14c5d2SYehuda Sadeh else 7783d14c5d2SYehuda Sadeh BUG(); 7793d14c5d2SYehuda Sadeh } 7803d14c5d2SYehuda Sadeh 7813d14c5d2SYehuda Sadeh rb_link_node(&new->o_node, parent, p); 7823d14c5d2SYehuda Sadeh rb_insert_color(&new->o_node, &osdc->osds); 7833d14c5d2SYehuda Sadeh } 7843d14c5d2SYehuda Sadeh 7853d14c5d2SYehuda Sadeh static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o) 7863d14c5d2SYehuda Sadeh { 7873d14c5d2SYehuda Sadeh struct ceph_osd *osd; 7883d14c5d2SYehuda Sadeh struct rb_node *n = osdc->osds.rb_node; 7893d14c5d2SYehuda Sadeh 7903d14c5d2SYehuda Sadeh while (n) { 7913d14c5d2SYehuda Sadeh osd = rb_entry(n, struct ceph_osd, o_node); 7923d14c5d2SYehuda Sadeh if (o < osd->o_osd) 7933d14c5d2SYehuda Sadeh n = n->rb_left; 7943d14c5d2SYehuda Sadeh else if (o > osd->o_osd) 7953d14c5d2SYehuda Sadeh n = n->rb_right; 7963d14c5d2SYehuda Sadeh else 7973d14c5d2SYehuda Sadeh return osd; 7983d14c5d2SYehuda Sadeh } 7993d14c5d2SYehuda Sadeh return NULL; 8003d14c5d2SYehuda Sadeh } 8013d14c5d2SYehuda Sadeh 8023d14c5d2SYehuda Sadeh static void __schedule_osd_timeout(struct ceph_osd_client *osdc) 8033d14c5d2SYehuda Sadeh { 8043d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->timeout_work, 8053d14c5d2SYehuda Sadeh osdc->client->options->osd_keepalive_timeout * HZ); 8063d14c5d2SYehuda Sadeh } 8073d14c5d2SYehuda Sadeh 8083d14c5d2SYehuda Sadeh static void __cancel_osd_timeout(struct ceph_osd_client *osdc) 8093d14c5d2SYehuda Sadeh { 8103d14c5d2SYehuda Sadeh cancel_delayed_work(&osdc->timeout_work); 8113d14c5d2SYehuda Sadeh } 8123d14c5d2SYehuda Sadeh 8133d14c5d2SYehuda Sadeh /* 8143d14c5d2SYehuda Sadeh * Register request, assign tid. If this is the first request, set up 8153d14c5d2SYehuda Sadeh * the timeout event. 8163d14c5d2SYehuda Sadeh */ 817a40c4f10SYehuda Sadeh static void __register_request(struct ceph_osd_client *osdc, 8183d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 8193d14c5d2SYehuda Sadeh { 8203d14c5d2SYehuda Sadeh req->r_tid = ++osdc->last_tid; 8213d14c5d2SYehuda Sadeh req->r_request->hdr.tid = cpu_to_le64(req->r_tid); 82277f38e0eSSage Weil dout("__register_request %p tid %lld\n", req, req->r_tid); 8233d14c5d2SYehuda Sadeh __insert_request(osdc, req); 8243d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 8253d14c5d2SYehuda Sadeh osdc->num_requests++; 8263d14c5d2SYehuda Sadeh if (osdc->num_requests == 1) { 8273d14c5d2SYehuda Sadeh dout(" first request, scheduling timeout\n"); 8283d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 8293d14c5d2SYehuda Sadeh } 830a40c4f10SYehuda Sadeh } 831a40c4f10SYehuda Sadeh 832a40c4f10SYehuda Sadeh static void register_request(struct ceph_osd_client *osdc, 833a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 834a40c4f10SYehuda Sadeh { 835a40c4f10SYehuda Sadeh mutex_lock(&osdc->request_mutex); 836a40c4f10SYehuda Sadeh __register_request(osdc, req); 8373d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 8383d14c5d2SYehuda Sadeh } 8393d14c5d2SYehuda Sadeh 8403d14c5d2SYehuda Sadeh /* 8413d14c5d2SYehuda Sadeh * called under osdc->request_mutex 8423d14c5d2SYehuda Sadeh */ 8433d14c5d2SYehuda Sadeh static void __unregister_request(struct ceph_osd_client *osdc, 8443d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 8453d14c5d2SYehuda Sadeh { 84635f9f8a0SSage Weil if (RB_EMPTY_NODE(&req->r_node)) { 84735f9f8a0SSage Weil dout("__unregister_request %p tid %lld not registered\n", 84835f9f8a0SSage Weil req, req->r_tid); 84935f9f8a0SSage Weil return; 85035f9f8a0SSage Weil } 85135f9f8a0SSage Weil 8523d14c5d2SYehuda Sadeh dout("__unregister_request %p tid %lld\n", req, req->r_tid); 8533d14c5d2SYehuda Sadeh rb_erase(&req->r_node, &osdc->requests); 8543d14c5d2SYehuda Sadeh osdc->num_requests--; 8553d14c5d2SYehuda Sadeh 8563d14c5d2SYehuda Sadeh if (req->r_osd) { 8573d14c5d2SYehuda Sadeh /* make sure the original request isn't in flight. */ 8583d14c5d2SYehuda Sadeh ceph_con_revoke(&req->r_osd->o_con, req->r_request); 8593d14c5d2SYehuda Sadeh 8603d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 861a40c4f10SYehuda Sadeh if (list_empty(&req->r_osd->o_requests) && 862a40c4f10SYehuda Sadeh list_empty(&req->r_osd->o_linger_requests)) { 863a40c4f10SYehuda Sadeh dout("moving osd to %p lru\n", req->r_osd); 8643d14c5d2SYehuda Sadeh __move_osd_to_lru(osdc, req->r_osd); 865a40c4f10SYehuda Sadeh } 866fbdb9190SSage Weil if (list_empty(&req->r_linger_item)) 8673d14c5d2SYehuda Sadeh req->r_osd = NULL; 8683d14c5d2SYehuda Sadeh } 8693d14c5d2SYehuda Sadeh 8703d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 8713d14c5d2SYehuda Sadeh 8723d14c5d2SYehuda Sadeh list_del_init(&req->r_req_lru_item); 8733d14c5d2SYehuda Sadeh if (osdc->num_requests == 0) { 8743d14c5d2SYehuda Sadeh dout(" no requests, canceling timeout\n"); 8753d14c5d2SYehuda Sadeh __cancel_osd_timeout(osdc); 8763d14c5d2SYehuda Sadeh } 8773d14c5d2SYehuda Sadeh } 8783d14c5d2SYehuda Sadeh 8793d14c5d2SYehuda Sadeh /* 8803d14c5d2SYehuda Sadeh * Cancel a previously queued request message 8813d14c5d2SYehuda Sadeh */ 8823d14c5d2SYehuda Sadeh static void __cancel_request(struct ceph_osd_request *req) 8833d14c5d2SYehuda Sadeh { 8843d14c5d2SYehuda Sadeh if (req->r_sent && req->r_osd) { 8853d14c5d2SYehuda Sadeh ceph_con_revoke(&req->r_osd->o_con, req->r_request); 8863d14c5d2SYehuda Sadeh req->r_sent = 0; 8873d14c5d2SYehuda Sadeh } 8883d14c5d2SYehuda Sadeh } 8893d14c5d2SYehuda Sadeh 890a40c4f10SYehuda Sadeh static void __register_linger_request(struct ceph_osd_client *osdc, 891a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 892a40c4f10SYehuda Sadeh { 893a40c4f10SYehuda Sadeh dout("__register_linger_request %p\n", req); 894a40c4f10SYehuda Sadeh list_add_tail(&req->r_linger_item, &osdc->req_linger); 895a40c4f10SYehuda Sadeh list_add_tail(&req->r_linger_osd, &req->r_osd->o_linger_requests); 896a40c4f10SYehuda Sadeh } 897a40c4f10SYehuda Sadeh 898a40c4f10SYehuda Sadeh static void __unregister_linger_request(struct ceph_osd_client *osdc, 899a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 900a40c4f10SYehuda Sadeh { 901a40c4f10SYehuda Sadeh dout("__unregister_linger_request %p\n", req); 902a40c4f10SYehuda Sadeh if (req->r_osd) { 903a40c4f10SYehuda Sadeh list_del_init(&req->r_linger_item); 904a40c4f10SYehuda Sadeh list_del_init(&req->r_linger_osd); 905a40c4f10SYehuda Sadeh 906a40c4f10SYehuda Sadeh if (list_empty(&req->r_osd->o_requests) && 907a40c4f10SYehuda Sadeh list_empty(&req->r_osd->o_linger_requests)) { 908a40c4f10SYehuda Sadeh dout("moving osd to %p lru\n", req->r_osd); 909a40c4f10SYehuda Sadeh __move_osd_to_lru(osdc, req->r_osd); 910a40c4f10SYehuda Sadeh } 911fbdb9190SSage Weil if (list_empty(&req->r_osd_item)) 912a40c4f10SYehuda Sadeh req->r_osd = NULL; 913a40c4f10SYehuda Sadeh } 914a40c4f10SYehuda Sadeh } 915a40c4f10SYehuda Sadeh 916a40c4f10SYehuda Sadeh void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc, 917a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 918a40c4f10SYehuda Sadeh { 919a40c4f10SYehuda Sadeh mutex_lock(&osdc->request_mutex); 920a40c4f10SYehuda Sadeh if (req->r_linger) { 921a40c4f10SYehuda Sadeh __unregister_linger_request(osdc, req); 922a40c4f10SYehuda Sadeh ceph_osdc_put_request(req); 923a40c4f10SYehuda Sadeh } 924a40c4f10SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 925a40c4f10SYehuda Sadeh } 926a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_unregister_linger_request); 927a40c4f10SYehuda Sadeh 928a40c4f10SYehuda Sadeh void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc, 929a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 930a40c4f10SYehuda Sadeh { 931a40c4f10SYehuda Sadeh if (!req->r_linger) { 932a40c4f10SYehuda Sadeh dout("set_request_linger %p\n", req); 933a40c4f10SYehuda Sadeh req->r_linger = 1; 934a40c4f10SYehuda Sadeh /* 935a40c4f10SYehuda Sadeh * caller is now responsible for calling 936a40c4f10SYehuda Sadeh * unregister_linger_request 937a40c4f10SYehuda Sadeh */ 938a40c4f10SYehuda Sadeh ceph_osdc_get_request(req); 939a40c4f10SYehuda Sadeh } 940a40c4f10SYehuda Sadeh } 941a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_set_request_linger); 942a40c4f10SYehuda Sadeh 9433d14c5d2SYehuda Sadeh /* 9443d14c5d2SYehuda Sadeh * Pick an osd (the first 'up' osd in the pg), allocate the osd struct 9453d14c5d2SYehuda Sadeh * (as needed), and set the request r_osd appropriately. If there is 94625985edcSLucas De Marchi * no up osd, set r_osd to NULL. Move the request to the appropriate list 9476f6c7006SSage Weil * (unsent, homeless) or leave on in-flight lru. 9483d14c5d2SYehuda Sadeh * 9493d14c5d2SYehuda Sadeh * Return 0 if unchanged, 1 if changed, or negative on error. 9503d14c5d2SYehuda Sadeh * 9513d14c5d2SYehuda Sadeh * Caller should hold map_sem for read and request_mutex. 9523d14c5d2SYehuda Sadeh */ 9536f6c7006SSage Weil static int __map_request(struct ceph_osd_client *osdc, 95438d6453cSSage Weil struct ceph_osd_request *req, int force_resend) 9553d14c5d2SYehuda Sadeh { 9563d14c5d2SYehuda Sadeh struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base; 9573d14c5d2SYehuda Sadeh struct ceph_pg pgid; 9583d14c5d2SYehuda Sadeh int acting[CEPH_PG_MAX_SIZE]; 9593d14c5d2SYehuda Sadeh int o = -1, num = 0; 9603d14c5d2SYehuda Sadeh int err; 9613d14c5d2SYehuda Sadeh 9626f6c7006SSage Weil dout("map_request %p tid %lld\n", req, req->r_tid); 9633d14c5d2SYehuda Sadeh err = ceph_calc_object_layout(&reqhead->layout, req->r_oid, 9643d14c5d2SYehuda Sadeh &req->r_file_layout, osdc->osdmap); 9656f6c7006SSage Weil if (err) { 9666f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 9673d14c5d2SYehuda Sadeh return err; 9686f6c7006SSage Weil } 9693d14c5d2SYehuda Sadeh pgid = reqhead->layout.ol_pgid; 9703d14c5d2SYehuda Sadeh req->r_pgid = pgid; 9713d14c5d2SYehuda Sadeh 9723d14c5d2SYehuda Sadeh err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting); 9733d14c5d2SYehuda Sadeh if (err > 0) { 9743d14c5d2SYehuda Sadeh o = acting[0]; 9753d14c5d2SYehuda Sadeh num = err; 9763d14c5d2SYehuda Sadeh } 9773d14c5d2SYehuda Sadeh 97838d6453cSSage Weil if ((!force_resend && 97938d6453cSSage Weil req->r_osd && req->r_osd->o_osd == o && 9803d14c5d2SYehuda Sadeh req->r_sent >= req->r_osd->o_incarnation && 9813d14c5d2SYehuda Sadeh req->r_num_pg_osds == num && 9823d14c5d2SYehuda Sadeh memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) || 9833d14c5d2SYehuda Sadeh (req->r_osd == NULL && o == -1)) 9843d14c5d2SYehuda Sadeh return 0; /* no change */ 9853d14c5d2SYehuda Sadeh 9866f6c7006SSage Weil dout("map_request tid %llu pgid %d.%x osd%d (was osd%d)\n", 9873d14c5d2SYehuda Sadeh req->r_tid, le32_to_cpu(pgid.pool), le16_to_cpu(pgid.ps), o, 9883d14c5d2SYehuda Sadeh req->r_osd ? req->r_osd->o_osd : -1); 9893d14c5d2SYehuda Sadeh 9903d14c5d2SYehuda Sadeh /* record full pg acting set */ 9913d14c5d2SYehuda Sadeh memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num); 9923d14c5d2SYehuda Sadeh req->r_num_pg_osds = num; 9933d14c5d2SYehuda Sadeh 9943d14c5d2SYehuda Sadeh if (req->r_osd) { 9953d14c5d2SYehuda Sadeh __cancel_request(req); 9963d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 9973d14c5d2SYehuda Sadeh req->r_osd = NULL; 9983d14c5d2SYehuda Sadeh } 9993d14c5d2SYehuda Sadeh 10003d14c5d2SYehuda Sadeh req->r_osd = __lookup_osd(osdc, o); 10013d14c5d2SYehuda Sadeh if (!req->r_osd && o >= 0) { 10023d14c5d2SYehuda Sadeh err = -ENOMEM; 1003e10006f8SAlex Elder req->r_osd = create_osd(osdc, o); 10046f6c7006SSage Weil if (!req->r_osd) { 10056f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 10063d14c5d2SYehuda Sadeh goto out; 10076f6c7006SSage Weil } 10083d14c5d2SYehuda Sadeh 10096f6c7006SSage Weil dout("map_request osd %p is osd%d\n", req->r_osd, o); 10103d14c5d2SYehuda Sadeh __insert_osd(osdc, req->r_osd); 10113d14c5d2SYehuda Sadeh 10123d14c5d2SYehuda Sadeh ceph_con_open(&req->r_osd->o_con, &osdc->osdmap->osd_addr[o]); 10133d14c5d2SYehuda Sadeh } 10143d14c5d2SYehuda Sadeh 10153d14c5d2SYehuda Sadeh if (req->r_osd) { 10163d14c5d2SYehuda Sadeh __remove_osd_from_lru(req->r_osd); 10173d14c5d2SYehuda Sadeh list_add(&req->r_osd_item, &req->r_osd->o_requests); 10186f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_unsent); 10196f6c7006SSage Weil } else { 10206f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 10213d14c5d2SYehuda Sadeh } 10223d14c5d2SYehuda Sadeh err = 1; /* osd or pg changed */ 10233d14c5d2SYehuda Sadeh 10243d14c5d2SYehuda Sadeh out: 10253d14c5d2SYehuda Sadeh return err; 10263d14c5d2SYehuda Sadeh } 10273d14c5d2SYehuda Sadeh 10283d14c5d2SYehuda Sadeh /* 10293d14c5d2SYehuda Sadeh * caller should hold map_sem (for read) and request_mutex 10303d14c5d2SYehuda Sadeh */ 103156e925b6SSage Weil static void __send_request(struct ceph_osd_client *osdc, 10323d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 10333d14c5d2SYehuda Sadeh { 10343d14c5d2SYehuda Sadeh struct ceph_osd_request_head *reqhead; 10353d14c5d2SYehuda Sadeh 10363d14c5d2SYehuda Sadeh dout("send_request %p tid %llu to osd%d flags %d\n", 10373d14c5d2SYehuda Sadeh req, req->r_tid, req->r_osd->o_osd, req->r_flags); 10383d14c5d2SYehuda Sadeh 10393d14c5d2SYehuda Sadeh reqhead = req->r_request->front.iov_base; 10403d14c5d2SYehuda Sadeh reqhead->osdmap_epoch = cpu_to_le32(osdc->osdmap->epoch); 10413d14c5d2SYehuda Sadeh reqhead->flags |= cpu_to_le32(req->r_flags); /* e.g., RETRY */ 10423d14c5d2SYehuda Sadeh reqhead->reassert_version = req->r_reassert_version; 10433d14c5d2SYehuda Sadeh 10443d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 10453d14c5d2SYehuda Sadeh list_move_tail(&req->r_req_lru_item, &osdc->req_lru); 10463d14c5d2SYehuda Sadeh 10473d14c5d2SYehuda Sadeh ceph_msg_get(req->r_request); /* send consumes a ref */ 10483d14c5d2SYehuda Sadeh ceph_con_send(&req->r_osd->o_con, req->r_request); 10493d14c5d2SYehuda Sadeh req->r_sent = req->r_osd->o_incarnation; 10503d14c5d2SYehuda Sadeh } 10513d14c5d2SYehuda Sadeh 10523d14c5d2SYehuda Sadeh /* 10536f6c7006SSage Weil * Send any requests in the queue (req_unsent). 10546f6c7006SSage Weil */ 10556f6c7006SSage Weil static void send_queued(struct ceph_osd_client *osdc) 10566f6c7006SSage Weil { 10576f6c7006SSage Weil struct ceph_osd_request *req, *tmp; 10586f6c7006SSage Weil 10596f6c7006SSage Weil dout("send_queued\n"); 10606f6c7006SSage Weil mutex_lock(&osdc->request_mutex); 10616f6c7006SSage Weil list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) { 10626f6c7006SSage Weil __send_request(osdc, req); 10636f6c7006SSage Weil } 10646f6c7006SSage Weil mutex_unlock(&osdc->request_mutex); 10656f6c7006SSage Weil } 10666f6c7006SSage Weil 10676f6c7006SSage Weil /* 10683d14c5d2SYehuda Sadeh * Timeout callback, called every N seconds when 1 or more osd 10693d14c5d2SYehuda Sadeh * requests has been active for more than N seconds. When this 10703d14c5d2SYehuda Sadeh * happens, we ping all OSDs with requests who have timed out to 10713d14c5d2SYehuda Sadeh * ensure any communications channel reset is detected. Reset the 10723d14c5d2SYehuda Sadeh * request timeouts another N seconds in the future as we go. 10733d14c5d2SYehuda Sadeh * Reschedule the timeout event another N seconds in future (unless 10743d14c5d2SYehuda Sadeh * there are no open requests). 10753d14c5d2SYehuda Sadeh */ 10763d14c5d2SYehuda Sadeh static void handle_timeout(struct work_struct *work) 10773d14c5d2SYehuda Sadeh { 10783d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 10793d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, timeout_work.work); 10803d14c5d2SYehuda Sadeh struct ceph_osd_request *req, *last_req = NULL; 10813d14c5d2SYehuda Sadeh struct ceph_osd *osd; 10823d14c5d2SYehuda Sadeh unsigned long timeout = osdc->client->options->osd_timeout * HZ; 10833d14c5d2SYehuda Sadeh unsigned long keepalive = 10843d14c5d2SYehuda Sadeh osdc->client->options->osd_keepalive_timeout * HZ; 10853d14c5d2SYehuda Sadeh unsigned long last_stamp = 0; 10863d14c5d2SYehuda Sadeh struct list_head slow_osds; 10873d14c5d2SYehuda Sadeh dout("timeout\n"); 10883d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 10893d14c5d2SYehuda Sadeh 10903d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 10913d14c5d2SYehuda Sadeh 10923d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 10933d14c5d2SYehuda Sadeh 10943d14c5d2SYehuda Sadeh /* 10953d14c5d2SYehuda Sadeh * reset osds that appear to be _really_ unresponsive. this 10963d14c5d2SYehuda Sadeh * is a failsafe measure.. we really shouldn't be getting to 10973d14c5d2SYehuda Sadeh * this point if the system is working properly. the monitors 10983d14c5d2SYehuda Sadeh * should mark the osd as failed and we should find out about 10993d14c5d2SYehuda Sadeh * it from an updated osd map. 11003d14c5d2SYehuda Sadeh */ 11013d14c5d2SYehuda Sadeh while (timeout && !list_empty(&osdc->req_lru)) { 11023d14c5d2SYehuda Sadeh req = list_entry(osdc->req_lru.next, struct ceph_osd_request, 11033d14c5d2SYehuda Sadeh r_req_lru_item); 11043d14c5d2SYehuda Sadeh 11054cf9d544SSage Weil /* hasn't been long enough since we sent it? */ 11063d14c5d2SYehuda Sadeh if (time_before(jiffies, req->r_stamp + timeout)) 11073d14c5d2SYehuda Sadeh break; 11083d14c5d2SYehuda Sadeh 11094cf9d544SSage Weil /* hasn't been long enough since it was acked? */ 11104cf9d544SSage Weil if (req->r_request->ack_stamp == 0 || 11114cf9d544SSage Weil time_before(jiffies, req->r_request->ack_stamp + timeout)) 11124cf9d544SSage Weil break; 11134cf9d544SSage Weil 11143d14c5d2SYehuda Sadeh BUG_ON(req == last_req && req->r_stamp == last_stamp); 11153d14c5d2SYehuda Sadeh last_req = req; 11163d14c5d2SYehuda Sadeh last_stamp = req->r_stamp; 11173d14c5d2SYehuda Sadeh 11183d14c5d2SYehuda Sadeh osd = req->r_osd; 11193d14c5d2SYehuda Sadeh BUG_ON(!osd); 11203d14c5d2SYehuda Sadeh pr_warning(" tid %llu timed out on osd%d, will reset osd\n", 11213d14c5d2SYehuda Sadeh req->r_tid, osd->o_osd); 11226f6c7006SSage Weil __kick_osd_requests(osdc, osd); 11233d14c5d2SYehuda Sadeh } 11243d14c5d2SYehuda Sadeh 11253d14c5d2SYehuda Sadeh /* 11263d14c5d2SYehuda Sadeh * ping osds that are a bit slow. this ensures that if there 11273d14c5d2SYehuda Sadeh * is a break in the TCP connection we will notice, and reopen 11283d14c5d2SYehuda Sadeh * a connection with that osd (from the fault callback). 11293d14c5d2SYehuda Sadeh */ 11303d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&slow_osds); 11313d14c5d2SYehuda Sadeh list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) { 11323d14c5d2SYehuda Sadeh if (time_before(jiffies, req->r_stamp + keepalive)) 11333d14c5d2SYehuda Sadeh break; 11343d14c5d2SYehuda Sadeh 11353d14c5d2SYehuda Sadeh osd = req->r_osd; 11363d14c5d2SYehuda Sadeh BUG_ON(!osd); 11373d14c5d2SYehuda Sadeh dout(" tid %llu is slow, will send keepalive on osd%d\n", 11383d14c5d2SYehuda Sadeh req->r_tid, osd->o_osd); 11393d14c5d2SYehuda Sadeh list_move_tail(&osd->o_keepalive_item, &slow_osds); 11403d14c5d2SYehuda Sadeh } 11413d14c5d2SYehuda Sadeh while (!list_empty(&slow_osds)) { 11423d14c5d2SYehuda Sadeh osd = list_entry(slow_osds.next, struct ceph_osd, 11433d14c5d2SYehuda Sadeh o_keepalive_item); 11443d14c5d2SYehuda Sadeh list_del_init(&osd->o_keepalive_item); 11453d14c5d2SYehuda Sadeh ceph_con_keepalive(&osd->o_con); 11463d14c5d2SYehuda Sadeh } 11473d14c5d2SYehuda Sadeh 11483d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 11493d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 11506f6c7006SSage Weil send_queued(osdc); 11513d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 11523d14c5d2SYehuda Sadeh } 11533d14c5d2SYehuda Sadeh 11543d14c5d2SYehuda Sadeh static void handle_osds_timeout(struct work_struct *work) 11553d14c5d2SYehuda Sadeh { 11563d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 11573d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, 11583d14c5d2SYehuda Sadeh osds_timeout_work.work); 11593d14c5d2SYehuda Sadeh unsigned long delay = 11603d14c5d2SYehuda Sadeh osdc->client->options->osd_idle_ttl * HZ >> 2; 11613d14c5d2SYehuda Sadeh 11623d14c5d2SYehuda Sadeh dout("osds timeout\n"); 11633d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 1164aca420bcSSage Weil remove_old_osds(osdc); 11653d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 11663d14c5d2SYehuda Sadeh 11673d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 11683d14c5d2SYehuda Sadeh round_jiffies_relative(delay)); 11693d14c5d2SYehuda Sadeh } 11703d14c5d2SYehuda Sadeh 117125845472SSage Weil static void complete_request(struct ceph_osd_request *req) 117225845472SSage Weil { 117325845472SSage Weil if (req->r_safe_callback) 117425845472SSage Weil req->r_safe_callback(req, NULL); 117525845472SSage Weil complete_all(&req->r_safe_completion); /* fsync waiter */ 117625845472SSage Weil } 117725845472SSage Weil 11783d14c5d2SYehuda Sadeh /* 11793d14c5d2SYehuda Sadeh * handle osd op reply. either call the callback if it is specified, 11803d14c5d2SYehuda Sadeh * or do the completion to wake up the waiting thread. 11813d14c5d2SYehuda Sadeh */ 11823d14c5d2SYehuda Sadeh static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg, 11833d14c5d2SYehuda Sadeh struct ceph_connection *con) 11843d14c5d2SYehuda Sadeh { 11853d14c5d2SYehuda Sadeh struct ceph_osd_reply_head *rhead = msg->front.iov_base; 11863d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 11873d14c5d2SYehuda Sadeh u64 tid; 11883d14c5d2SYehuda Sadeh int numops, object_len, flags; 11893d14c5d2SYehuda Sadeh s32 result; 11903d14c5d2SYehuda Sadeh 11913d14c5d2SYehuda Sadeh tid = le64_to_cpu(msg->hdr.tid); 11923d14c5d2SYehuda Sadeh if (msg->front.iov_len < sizeof(*rhead)) 11933d14c5d2SYehuda Sadeh goto bad; 11943d14c5d2SYehuda Sadeh numops = le32_to_cpu(rhead->num_ops); 11953d14c5d2SYehuda Sadeh object_len = le32_to_cpu(rhead->object_len); 11963d14c5d2SYehuda Sadeh result = le32_to_cpu(rhead->result); 11973d14c5d2SYehuda Sadeh if (msg->front.iov_len != sizeof(*rhead) + object_len + 11983d14c5d2SYehuda Sadeh numops * sizeof(struct ceph_osd_op)) 11993d14c5d2SYehuda Sadeh goto bad; 12003d14c5d2SYehuda Sadeh dout("handle_reply %p tid %llu result %d\n", msg, tid, (int)result); 12013d14c5d2SYehuda Sadeh /* lookup */ 12023d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 12033d14c5d2SYehuda Sadeh req = __lookup_request(osdc, tid); 12043d14c5d2SYehuda Sadeh if (req == NULL) { 12053d14c5d2SYehuda Sadeh dout("handle_reply tid %llu dne\n", tid); 12063d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 12073d14c5d2SYehuda Sadeh return; 12083d14c5d2SYehuda Sadeh } 12093d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 12103d14c5d2SYehuda Sadeh flags = le32_to_cpu(rhead->flags); 12113d14c5d2SYehuda Sadeh 12123d14c5d2SYehuda Sadeh /* 12133d14c5d2SYehuda Sadeh * if this connection filled our message, drop our reference now, to 12143d14c5d2SYehuda Sadeh * avoid a (safe but slower) revoke later. 12153d14c5d2SYehuda Sadeh */ 12163d14c5d2SYehuda Sadeh if (req->r_con_filling_msg == con && req->r_reply == msg) { 12173d14c5d2SYehuda Sadeh dout(" dropping con_filling_msg ref %p\n", con); 12183d14c5d2SYehuda Sadeh req->r_con_filling_msg = NULL; 12193d14c5d2SYehuda Sadeh ceph_con_put(con); 12203d14c5d2SYehuda Sadeh } 12213d14c5d2SYehuda Sadeh 12223d14c5d2SYehuda Sadeh if (!req->r_got_reply) { 12233d14c5d2SYehuda Sadeh unsigned bytes; 12243d14c5d2SYehuda Sadeh 12253d14c5d2SYehuda Sadeh req->r_result = le32_to_cpu(rhead->result); 12263d14c5d2SYehuda Sadeh bytes = le32_to_cpu(msg->hdr.data_len); 12273d14c5d2SYehuda Sadeh dout("handle_reply result %d bytes %d\n", req->r_result, 12283d14c5d2SYehuda Sadeh bytes); 12293d14c5d2SYehuda Sadeh if (req->r_result == 0) 12303d14c5d2SYehuda Sadeh req->r_result = bytes; 12313d14c5d2SYehuda Sadeh 12323d14c5d2SYehuda Sadeh /* in case this is a write and we need to replay, */ 12333d14c5d2SYehuda Sadeh req->r_reassert_version = rhead->reassert_version; 12343d14c5d2SYehuda Sadeh 12353d14c5d2SYehuda Sadeh req->r_got_reply = 1; 12363d14c5d2SYehuda Sadeh } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) { 12373d14c5d2SYehuda Sadeh dout("handle_reply tid %llu dup ack\n", tid); 12383d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 12393d14c5d2SYehuda Sadeh goto done; 12403d14c5d2SYehuda Sadeh } 12413d14c5d2SYehuda Sadeh 12423d14c5d2SYehuda Sadeh dout("handle_reply tid %llu flags %d\n", tid, flags); 12433d14c5d2SYehuda Sadeh 1244a40c4f10SYehuda Sadeh if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK)) 1245a40c4f10SYehuda Sadeh __register_linger_request(osdc, req); 1246a40c4f10SYehuda Sadeh 12473d14c5d2SYehuda Sadeh /* either this is a read, or we got the safe response */ 12483d14c5d2SYehuda Sadeh if (result < 0 || 12493d14c5d2SYehuda Sadeh (flags & CEPH_OSD_FLAG_ONDISK) || 12503d14c5d2SYehuda Sadeh ((flags & CEPH_OSD_FLAG_WRITE) == 0)) 12513d14c5d2SYehuda Sadeh __unregister_request(osdc, req); 12523d14c5d2SYehuda Sadeh 12533d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 12543d14c5d2SYehuda Sadeh 12553d14c5d2SYehuda Sadeh if (req->r_callback) 12563d14c5d2SYehuda Sadeh req->r_callback(req, msg); 12573d14c5d2SYehuda Sadeh else 12583d14c5d2SYehuda Sadeh complete_all(&req->r_completion); 12593d14c5d2SYehuda Sadeh 126025845472SSage Weil if (flags & CEPH_OSD_FLAG_ONDISK) 126125845472SSage Weil complete_request(req); 12623d14c5d2SYehuda Sadeh 12633d14c5d2SYehuda Sadeh done: 1264a40c4f10SYehuda Sadeh dout("req=%p req->r_linger=%d\n", req, req->r_linger); 12653d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 12663d14c5d2SYehuda Sadeh return; 12673d14c5d2SYehuda Sadeh 12683d14c5d2SYehuda Sadeh bad: 12693d14c5d2SYehuda Sadeh pr_err("corrupt osd_op_reply got %d %d expected %d\n", 12703d14c5d2SYehuda Sadeh (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len), 12713d14c5d2SYehuda Sadeh (int)sizeof(*rhead)); 12723d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 12733d14c5d2SYehuda Sadeh } 12743d14c5d2SYehuda Sadeh 12756f6c7006SSage Weil static void reset_changed_osds(struct ceph_osd_client *osdc) 12763d14c5d2SYehuda Sadeh { 12773d14c5d2SYehuda Sadeh struct rb_node *p, *n; 12783d14c5d2SYehuda Sadeh 12793d14c5d2SYehuda Sadeh for (p = rb_first(&osdc->osds); p; p = n) { 12806f6c7006SSage Weil struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node); 12813d14c5d2SYehuda Sadeh 12823d14c5d2SYehuda Sadeh n = rb_next(p); 12833d14c5d2SYehuda Sadeh if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) || 12843d14c5d2SYehuda Sadeh memcmp(&osd->o_con.peer_addr, 12853d14c5d2SYehuda Sadeh ceph_osd_addr(osdc->osdmap, 12863d14c5d2SYehuda Sadeh osd->o_osd), 12873d14c5d2SYehuda Sadeh sizeof(struct ceph_entity_addr)) != 0) 12883d14c5d2SYehuda Sadeh __reset_osd(osdc, osd); 12893d14c5d2SYehuda Sadeh } 12903d14c5d2SYehuda Sadeh } 12913d14c5d2SYehuda Sadeh 12923d14c5d2SYehuda Sadeh /* 12936f6c7006SSage Weil * Requeue requests whose mapping to an OSD has changed. If requests map to 12946f6c7006SSage Weil * no osd, request a new map. 12953d14c5d2SYehuda Sadeh * 12963d14c5d2SYehuda Sadeh * Caller should hold map_sem for read and request_mutex. 12973d14c5d2SYehuda Sadeh */ 129838d6453cSSage Weil static void kick_requests(struct ceph_osd_client *osdc, int force_resend) 12993d14c5d2SYehuda Sadeh { 1300a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 13016f6c7006SSage Weil struct rb_node *p; 13026f6c7006SSage Weil int needmap = 0; 13036f6c7006SSage Weil int err; 13043d14c5d2SYehuda Sadeh 130538d6453cSSage Weil dout("kick_requests %s\n", force_resend ? " (force resend)" : ""); 13063d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 13076f6c7006SSage Weil for (p = rb_first(&osdc->requests); p; p = rb_next(p)) { 13086f6c7006SSage Weil req = rb_entry(p, struct ceph_osd_request, r_node); 130938d6453cSSage Weil err = __map_request(osdc, req, force_resend); 13106f6c7006SSage Weil if (err < 0) 13116f6c7006SSage Weil continue; /* error */ 13126f6c7006SSage Weil if (req->r_osd == NULL) { 13136f6c7006SSage Weil dout("%p tid %llu maps to no osd\n", req, req->r_tid); 13146f6c7006SSage Weil needmap++; /* request a newer map */ 13156f6c7006SSage Weil } else if (err > 0) { 13166f6c7006SSage Weil dout("%p tid %llu requeued on osd%d\n", req, req->r_tid, 13176f6c7006SSage Weil req->r_osd ? req->r_osd->o_osd : -1); 1318a40c4f10SYehuda Sadeh if (!req->r_linger) 13196f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 13206f6c7006SSage Weil } 13216f6c7006SSage Weil } 1322a40c4f10SYehuda Sadeh 1323a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osdc->req_linger, 1324a40c4f10SYehuda Sadeh r_linger_item) { 1325a40c4f10SYehuda Sadeh dout("linger req=%p req->r_osd=%p\n", req, req->r_osd); 1326a40c4f10SYehuda Sadeh 132738d6453cSSage Weil err = __map_request(osdc, req, force_resend); 1328a40c4f10SYehuda Sadeh if (err == 0) 1329a40c4f10SYehuda Sadeh continue; /* no change and no osd was specified */ 1330a40c4f10SYehuda Sadeh if (err < 0) 1331a40c4f10SYehuda Sadeh continue; /* hrm! */ 1332a40c4f10SYehuda Sadeh if (req->r_osd == NULL) { 1333a40c4f10SYehuda Sadeh dout("tid %llu maps to no valid osd\n", req->r_tid); 1334a40c4f10SYehuda Sadeh needmap++; /* request a newer map */ 1335a40c4f10SYehuda Sadeh continue; 1336a40c4f10SYehuda Sadeh } 1337a40c4f10SYehuda Sadeh 1338a40c4f10SYehuda Sadeh dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid, 1339a40c4f10SYehuda Sadeh req->r_osd ? req->r_osd->o_osd : -1); 1340a40c4f10SYehuda Sadeh __unregister_linger_request(osdc, req); 1341a40c4f10SYehuda Sadeh __register_request(osdc, req); 1342a40c4f10SYehuda Sadeh } 13433d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 13443d14c5d2SYehuda Sadeh 13453d14c5d2SYehuda Sadeh if (needmap) { 13463d14c5d2SYehuda Sadeh dout("%d requests for down osds, need new map\n", needmap); 13473d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 13483d14c5d2SYehuda Sadeh } 13493d14c5d2SYehuda Sadeh } 13506f6c7006SSage Weil 13516f6c7006SSage Weil 13523d14c5d2SYehuda Sadeh /* 13533d14c5d2SYehuda Sadeh * Process updated osd map. 13543d14c5d2SYehuda Sadeh * 13553d14c5d2SYehuda Sadeh * The message contains any number of incremental and full maps, normally 13563d14c5d2SYehuda Sadeh * indicating some sort of topology change in the cluster. Kick requests 13573d14c5d2SYehuda Sadeh * off to different OSDs as needed. 13583d14c5d2SYehuda Sadeh */ 13593d14c5d2SYehuda Sadeh void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg) 13603d14c5d2SYehuda Sadeh { 13613d14c5d2SYehuda Sadeh void *p, *end, *next; 13623d14c5d2SYehuda Sadeh u32 nr_maps, maplen; 13633d14c5d2SYehuda Sadeh u32 epoch; 13643d14c5d2SYehuda Sadeh struct ceph_osdmap *newmap = NULL, *oldmap; 13653d14c5d2SYehuda Sadeh int err; 13663d14c5d2SYehuda Sadeh struct ceph_fsid fsid; 13673d14c5d2SYehuda Sadeh 13683d14c5d2SYehuda Sadeh dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0); 13693d14c5d2SYehuda Sadeh p = msg->front.iov_base; 13703d14c5d2SYehuda Sadeh end = p + msg->front.iov_len; 13713d14c5d2SYehuda Sadeh 13723d14c5d2SYehuda Sadeh /* verify fsid */ 13733d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, sizeof(fsid), bad); 13743d14c5d2SYehuda Sadeh ceph_decode_copy(&p, &fsid, sizeof(fsid)); 13753d14c5d2SYehuda Sadeh if (ceph_check_fsid(osdc->client, &fsid) < 0) 13763d14c5d2SYehuda Sadeh return; 13773d14c5d2SYehuda Sadeh 13783d14c5d2SYehuda Sadeh down_write(&osdc->map_sem); 13793d14c5d2SYehuda Sadeh 13803d14c5d2SYehuda Sadeh /* incremental maps */ 13813d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 13823d14c5d2SYehuda Sadeh dout(" %d inc maps\n", nr_maps); 13833d14c5d2SYehuda Sadeh while (nr_maps > 0) { 13843d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 13853d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 13863d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 13873d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 13883d14c5d2SYehuda Sadeh next = p + maplen; 13893d14c5d2SYehuda Sadeh if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) { 13903d14c5d2SYehuda Sadeh dout("applying incremental map %u len %d\n", 13913d14c5d2SYehuda Sadeh epoch, maplen); 13923d14c5d2SYehuda Sadeh newmap = osdmap_apply_incremental(&p, next, 13933d14c5d2SYehuda Sadeh osdc->osdmap, 139415d9882cSAlex Elder &osdc->client->msgr); 13953d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 13963d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 13973d14c5d2SYehuda Sadeh goto bad; 13983d14c5d2SYehuda Sadeh } 13993d14c5d2SYehuda Sadeh BUG_ON(!newmap); 14003d14c5d2SYehuda Sadeh if (newmap != osdc->osdmap) { 14013d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 14023d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 14033d14c5d2SYehuda Sadeh } 140438d6453cSSage Weil kick_requests(osdc, 0); 14056f6c7006SSage Weil reset_changed_osds(osdc); 14063d14c5d2SYehuda Sadeh } else { 14073d14c5d2SYehuda Sadeh dout("ignoring incremental map %u len %d\n", 14083d14c5d2SYehuda Sadeh epoch, maplen); 14093d14c5d2SYehuda Sadeh } 14103d14c5d2SYehuda Sadeh p = next; 14113d14c5d2SYehuda Sadeh nr_maps--; 14123d14c5d2SYehuda Sadeh } 14133d14c5d2SYehuda Sadeh if (newmap) 14143d14c5d2SYehuda Sadeh goto done; 14153d14c5d2SYehuda Sadeh 14163d14c5d2SYehuda Sadeh /* full maps */ 14173d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 14183d14c5d2SYehuda Sadeh dout(" %d full maps\n", nr_maps); 14193d14c5d2SYehuda Sadeh while (nr_maps) { 14203d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 14213d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 14223d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 14233d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 14243d14c5d2SYehuda Sadeh if (nr_maps > 1) { 14253d14c5d2SYehuda Sadeh dout("skipping non-latest full map %u len %d\n", 14263d14c5d2SYehuda Sadeh epoch, maplen); 14273d14c5d2SYehuda Sadeh } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) { 14283d14c5d2SYehuda Sadeh dout("skipping full map %u len %d, " 14293d14c5d2SYehuda Sadeh "older than our %u\n", epoch, maplen, 14303d14c5d2SYehuda Sadeh osdc->osdmap->epoch); 14313d14c5d2SYehuda Sadeh } else { 143238d6453cSSage Weil int skipped_map = 0; 143338d6453cSSage Weil 14343d14c5d2SYehuda Sadeh dout("taking full map %u len %d\n", epoch, maplen); 14353d14c5d2SYehuda Sadeh newmap = osdmap_decode(&p, p+maplen); 14363d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 14373d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 14383d14c5d2SYehuda Sadeh goto bad; 14393d14c5d2SYehuda Sadeh } 14403d14c5d2SYehuda Sadeh BUG_ON(!newmap); 14413d14c5d2SYehuda Sadeh oldmap = osdc->osdmap; 14423d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 144338d6453cSSage Weil if (oldmap) { 144438d6453cSSage Weil if (oldmap->epoch + 1 < newmap->epoch) 144538d6453cSSage Weil skipped_map = 1; 14463d14c5d2SYehuda Sadeh ceph_osdmap_destroy(oldmap); 144738d6453cSSage Weil } 144838d6453cSSage Weil kick_requests(osdc, skipped_map); 14493d14c5d2SYehuda Sadeh } 14503d14c5d2SYehuda Sadeh p += maplen; 14513d14c5d2SYehuda Sadeh nr_maps--; 14523d14c5d2SYehuda Sadeh } 14533d14c5d2SYehuda Sadeh 14543d14c5d2SYehuda Sadeh done: 14553d14c5d2SYehuda Sadeh downgrade_write(&osdc->map_sem); 14563d14c5d2SYehuda Sadeh ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch); 1457cd634fb6SSage Weil 1458cd634fb6SSage Weil /* 1459cd634fb6SSage Weil * subscribe to subsequent osdmap updates if full to ensure 1460cd634fb6SSage Weil * we find out when we are no longer full and stop returning 1461cd634fb6SSage Weil * ENOSPC. 1462cd634fb6SSage Weil */ 1463cd634fb6SSage Weil if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL)) 1464cd634fb6SSage Weil ceph_monc_request_next_osdmap(&osdc->client->monc); 1465cd634fb6SSage Weil 14666f6c7006SSage Weil send_queued(osdc); 14673d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 14683d14c5d2SYehuda Sadeh wake_up_all(&osdc->client->auth_wq); 14693d14c5d2SYehuda Sadeh return; 14703d14c5d2SYehuda Sadeh 14713d14c5d2SYehuda Sadeh bad: 14723d14c5d2SYehuda Sadeh pr_err("osdc handle_map corrupt msg\n"); 14733d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 14743d14c5d2SYehuda Sadeh up_write(&osdc->map_sem); 14753d14c5d2SYehuda Sadeh return; 14763d14c5d2SYehuda Sadeh } 14773d14c5d2SYehuda Sadeh 14783d14c5d2SYehuda Sadeh /* 1479a40c4f10SYehuda Sadeh * watch/notify callback event infrastructure 1480a40c4f10SYehuda Sadeh * 1481a40c4f10SYehuda Sadeh * These callbacks are used both for watch and notify operations. 1482a40c4f10SYehuda Sadeh */ 1483a40c4f10SYehuda Sadeh static void __release_event(struct kref *kref) 1484a40c4f10SYehuda Sadeh { 1485a40c4f10SYehuda Sadeh struct ceph_osd_event *event = 1486a40c4f10SYehuda Sadeh container_of(kref, struct ceph_osd_event, kref); 1487a40c4f10SYehuda Sadeh 1488a40c4f10SYehuda Sadeh dout("__release_event %p\n", event); 1489a40c4f10SYehuda Sadeh kfree(event); 1490a40c4f10SYehuda Sadeh } 1491a40c4f10SYehuda Sadeh 1492a40c4f10SYehuda Sadeh static void get_event(struct ceph_osd_event *event) 1493a40c4f10SYehuda Sadeh { 1494a40c4f10SYehuda Sadeh kref_get(&event->kref); 1495a40c4f10SYehuda Sadeh } 1496a40c4f10SYehuda Sadeh 1497a40c4f10SYehuda Sadeh void ceph_osdc_put_event(struct ceph_osd_event *event) 1498a40c4f10SYehuda Sadeh { 1499a40c4f10SYehuda Sadeh kref_put(&event->kref, __release_event); 1500a40c4f10SYehuda Sadeh } 1501a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_put_event); 1502a40c4f10SYehuda Sadeh 1503a40c4f10SYehuda Sadeh static void __insert_event(struct ceph_osd_client *osdc, 1504a40c4f10SYehuda Sadeh struct ceph_osd_event *new) 1505a40c4f10SYehuda Sadeh { 1506a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 1507a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 1508a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 1509a40c4f10SYehuda Sadeh 1510a40c4f10SYehuda Sadeh while (*p) { 1511a40c4f10SYehuda Sadeh parent = *p; 1512a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 1513a40c4f10SYehuda Sadeh if (new->cookie < event->cookie) 1514a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 1515a40c4f10SYehuda Sadeh else if (new->cookie > event->cookie) 1516a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 1517a40c4f10SYehuda Sadeh else 1518a40c4f10SYehuda Sadeh BUG(); 1519a40c4f10SYehuda Sadeh } 1520a40c4f10SYehuda Sadeh 1521a40c4f10SYehuda Sadeh rb_link_node(&new->node, parent, p); 1522a40c4f10SYehuda Sadeh rb_insert_color(&new->node, &osdc->event_tree); 1523a40c4f10SYehuda Sadeh } 1524a40c4f10SYehuda Sadeh 1525a40c4f10SYehuda Sadeh static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc, 1526a40c4f10SYehuda Sadeh u64 cookie) 1527a40c4f10SYehuda Sadeh { 1528a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 1529a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 1530a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 1531a40c4f10SYehuda Sadeh 1532a40c4f10SYehuda Sadeh while (*p) { 1533a40c4f10SYehuda Sadeh parent = *p; 1534a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 1535a40c4f10SYehuda Sadeh if (cookie < event->cookie) 1536a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 1537a40c4f10SYehuda Sadeh else if (cookie > event->cookie) 1538a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 1539a40c4f10SYehuda Sadeh else 1540a40c4f10SYehuda Sadeh return event; 1541a40c4f10SYehuda Sadeh } 1542a40c4f10SYehuda Sadeh return NULL; 1543a40c4f10SYehuda Sadeh } 1544a40c4f10SYehuda Sadeh 1545a40c4f10SYehuda Sadeh static void __remove_event(struct ceph_osd_event *event) 1546a40c4f10SYehuda Sadeh { 1547a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 1548a40c4f10SYehuda Sadeh 1549a40c4f10SYehuda Sadeh if (!RB_EMPTY_NODE(&event->node)) { 1550a40c4f10SYehuda Sadeh dout("__remove_event removed %p\n", event); 1551a40c4f10SYehuda Sadeh rb_erase(&event->node, &osdc->event_tree); 1552a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 1553a40c4f10SYehuda Sadeh } else { 1554a40c4f10SYehuda Sadeh dout("__remove_event didn't remove %p\n", event); 1555a40c4f10SYehuda Sadeh } 1556a40c4f10SYehuda Sadeh } 1557a40c4f10SYehuda Sadeh 1558a40c4f10SYehuda Sadeh int ceph_osdc_create_event(struct ceph_osd_client *osdc, 1559a40c4f10SYehuda Sadeh void (*event_cb)(u64, u64, u8, void *), 1560a40c4f10SYehuda Sadeh int one_shot, void *data, 1561a40c4f10SYehuda Sadeh struct ceph_osd_event **pevent) 1562a40c4f10SYehuda Sadeh { 1563a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 1564a40c4f10SYehuda Sadeh 1565a40c4f10SYehuda Sadeh event = kmalloc(sizeof(*event), GFP_NOIO); 1566a40c4f10SYehuda Sadeh if (!event) 1567a40c4f10SYehuda Sadeh return -ENOMEM; 1568a40c4f10SYehuda Sadeh 1569a40c4f10SYehuda Sadeh dout("create_event %p\n", event); 1570a40c4f10SYehuda Sadeh event->cb = event_cb; 1571a40c4f10SYehuda Sadeh event->one_shot = one_shot; 1572a40c4f10SYehuda Sadeh event->data = data; 1573a40c4f10SYehuda Sadeh event->osdc = osdc; 1574a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&event->osd_node); 1575a40c4f10SYehuda Sadeh kref_init(&event->kref); /* one ref for us */ 1576a40c4f10SYehuda Sadeh kref_get(&event->kref); /* one ref for the caller */ 1577a40c4f10SYehuda Sadeh init_completion(&event->completion); 1578a40c4f10SYehuda Sadeh 1579a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 1580a40c4f10SYehuda Sadeh event->cookie = ++osdc->event_count; 1581a40c4f10SYehuda Sadeh __insert_event(osdc, event); 1582a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 1583a40c4f10SYehuda Sadeh 1584a40c4f10SYehuda Sadeh *pevent = event; 1585a40c4f10SYehuda Sadeh return 0; 1586a40c4f10SYehuda Sadeh } 1587a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_create_event); 1588a40c4f10SYehuda Sadeh 1589a40c4f10SYehuda Sadeh void ceph_osdc_cancel_event(struct ceph_osd_event *event) 1590a40c4f10SYehuda Sadeh { 1591a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 1592a40c4f10SYehuda Sadeh 1593a40c4f10SYehuda Sadeh dout("cancel_event %p\n", event); 1594a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 1595a40c4f10SYehuda Sadeh __remove_event(event); 1596a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 1597a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); /* caller's */ 1598a40c4f10SYehuda Sadeh } 1599a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_cancel_event); 1600a40c4f10SYehuda Sadeh 1601a40c4f10SYehuda Sadeh 1602a40c4f10SYehuda Sadeh static void do_event_work(struct work_struct *work) 1603a40c4f10SYehuda Sadeh { 1604a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work = 1605a40c4f10SYehuda Sadeh container_of(work, struct ceph_osd_event_work, work); 1606a40c4f10SYehuda Sadeh struct ceph_osd_event *event = event_work->event; 1607a40c4f10SYehuda Sadeh u64 ver = event_work->ver; 1608a40c4f10SYehuda Sadeh u64 notify_id = event_work->notify_id; 1609a40c4f10SYehuda Sadeh u8 opcode = event_work->opcode; 1610a40c4f10SYehuda Sadeh 1611a40c4f10SYehuda Sadeh dout("do_event_work completing %p\n", event); 1612a40c4f10SYehuda Sadeh event->cb(ver, notify_id, opcode, event->data); 1613a40c4f10SYehuda Sadeh complete(&event->completion); 1614a40c4f10SYehuda Sadeh dout("do_event_work completed %p\n", event); 1615a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 1616a40c4f10SYehuda Sadeh kfree(event_work); 1617a40c4f10SYehuda Sadeh } 1618a40c4f10SYehuda Sadeh 1619a40c4f10SYehuda Sadeh 1620a40c4f10SYehuda Sadeh /* 1621a40c4f10SYehuda Sadeh * Process osd watch notifications 1622a40c4f10SYehuda Sadeh */ 1623a40c4f10SYehuda Sadeh void handle_watch_notify(struct ceph_osd_client *osdc, struct ceph_msg *msg) 1624a40c4f10SYehuda Sadeh { 1625a40c4f10SYehuda Sadeh void *p, *end; 1626a40c4f10SYehuda Sadeh u8 proto_ver; 1627a40c4f10SYehuda Sadeh u64 cookie, ver, notify_id; 1628a40c4f10SYehuda Sadeh u8 opcode; 1629a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 1630a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work; 1631a40c4f10SYehuda Sadeh 1632a40c4f10SYehuda Sadeh p = msg->front.iov_base; 1633a40c4f10SYehuda Sadeh end = p + msg->front.iov_len; 1634a40c4f10SYehuda Sadeh 1635a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, proto_ver, bad); 1636a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, opcode, bad); 1637a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, cookie, bad); 1638a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, ver, bad); 1639a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, notify_id, bad); 1640a40c4f10SYehuda Sadeh 1641a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 1642a40c4f10SYehuda Sadeh event = __find_event(osdc, cookie); 1643a40c4f10SYehuda Sadeh if (event) { 1644a40c4f10SYehuda Sadeh get_event(event); 1645a40c4f10SYehuda Sadeh if (event->one_shot) 1646a40c4f10SYehuda Sadeh __remove_event(event); 1647a40c4f10SYehuda Sadeh } 1648a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 1649a40c4f10SYehuda Sadeh dout("handle_watch_notify cookie %lld ver %lld event %p\n", 1650a40c4f10SYehuda Sadeh cookie, ver, event); 1651a40c4f10SYehuda Sadeh if (event) { 1652a40c4f10SYehuda Sadeh event_work = kmalloc(sizeof(*event_work), GFP_NOIO); 1653a40c4f10SYehuda Sadeh if (!event_work) { 1654a40c4f10SYehuda Sadeh dout("ERROR: could not allocate event_work\n"); 1655a40c4f10SYehuda Sadeh goto done_err; 1656a40c4f10SYehuda Sadeh } 16576b0ae409SMariusz Kozlowski INIT_WORK(&event_work->work, do_event_work); 1658a40c4f10SYehuda Sadeh event_work->event = event; 1659a40c4f10SYehuda Sadeh event_work->ver = ver; 1660a40c4f10SYehuda Sadeh event_work->notify_id = notify_id; 1661a40c4f10SYehuda Sadeh event_work->opcode = opcode; 1662a40c4f10SYehuda Sadeh if (!queue_work(osdc->notify_wq, &event_work->work)) { 1663a40c4f10SYehuda Sadeh dout("WARNING: failed to queue notify event work\n"); 1664a40c4f10SYehuda Sadeh goto done_err; 1665a40c4f10SYehuda Sadeh } 1666a40c4f10SYehuda Sadeh } 1667a40c4f10SYehuda Sadeh 1668a40c4f10SYehuda Sadeh return; 1669a40c4f10SYehuda Sadeh 1670a40c4f10SYehuda Sadeh done_err: 1671a40c4f10SYehuda Sadeh complete(&event->completion); 1672a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 1673a40c4f10SYehuda Sadeh return; 1674a40c4f10SYehuda Sadeh 1675a40c4f10SYehuda Sadeh bad: 1676a40c4f10SYehuda Sadeh pr_err("osdc handle_watch_notify corrupt msg\n"); 1677a40c4f10SYehuda Sadeh return; 1678a40c4f10SYehuda Sadeh } 1679a40c4f10SYehuda Sadeh 1680a40c4f10SYehuda Sadeh int ceph_osdc_wait_event(struct ceph_osd_event *event, unsigned long timeout) 1681a40c4f10SYehuda Sadeh { 1682a40c4f10SYehuda Sadeh int err; 1683a40c4f10SYehuda Sadeh 1684a40c4f10SYehuda Sadeh dout("wait_event %p\n", event); 1685a40c4f10SYehuda Sadeh err = wait_for_completion_interruptible_timeout(&event->completion, 1686a40c4f10SYehuda Sadeh timeout * HZ); 1687a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 1688a40c4f10SYehuda Sadeh if (err > 0) 1689a40c4f10SYehuda Sadeh err = 0; 1690a40c4f10SYehuda Sadeh dout("wait_event %p returns %d\n", event, err); 1691a40c4f10SYehuda Sadeh return err; 1692a40c4f10SYehuda Sadeh } 1693a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_wait_event); 1694a40c4f10SYehuda Sadeh 1695a40c4f10SYehuda Sadeh /* 16963d14c5d2SYehuda Sadeh * Register request, send initial attempt. 16973d14c5d2SYehuda Sadeh */ 16983d14c5d2SYehuda Sadeh int ceph_osdc_start_request(struct ceph_osd_client *osdc, 16993d14c5d2SYehuda Sadeh struct ceph_osd_request *req, 17003d14c5d2SYehuda Sadeh bool nofail) 17013d14c5d2SYehuda Sadeh { 17023d14c5d2SYehuda Sadeh int rc = 0; 17033d14c5d2SYehuda Sadeh 17043d14c5d2SYehuda Sadeh req->r_request->pages = req->r_pages; 17053d14c5d2SYehuda Sadeh req->r_request->nr_pages = req->r_num_pages; 17063d14c5d2SYehuda Sadeh #ifdef CONFIG_BLOCK 17073d14c5d2SYehuda Sadeh req->r_request->bio = req->r_bio; 17083d14c5d2SYehuda Sadeh #endif 17093d14c5d2SYehuda Sadeh req->r_request->trail = req->r_trail; 17103d14c5d2SYehuda Sadeh 17113d14c5d2SYehuda Sadeh register_request(osdc, req); 17123d14c5d2SYehuda Sadeh 17133d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 17143d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 17153d14c5d2SYehuda Sadeh /* 17163d14c5d2SYehuda Sadeh * a racing kick_requests() may have sent the message for us 17173d14c5d2SYehuda Sadeh * while we dropped request_mutex above, so only send now if 17183d14c5d2SYehuda Sadeh * the request still han't been touched yet. 17193d14c5d2SYehuda Sadeh */ 17203d14c5d2SYehuda Sadeh if (req->r_sent == 0) { 172138d6453cSSage Weil rc = __map_request(osdc, req, 0); 17229d6fcb08SSage Weil if (rc < 0) { 17239d6fcb08SSage Weil if (nofail) { 17249d6fcb08SSage Weil dout("osdc_start_request failed map, " 17259d6fcb08SSage Weil " will retry %lld\n", req->r_tid); 17269d6fcb08SSage Weil rc = 0; 17279d6fcb08SSage Weil } 1728234af26fSDan Carpenter goto out_unlock; 17299d6fcb08SSage Weil } 17306f6c7006SSage Weil if (req->r_osd == NULL) { 17316f6c7006SSage Weil dout("send_request %p no up osds in pg\n", req); 17326f6c7006SSage Weil ceph_monc_request_next_osdmap(&osdc->client->monc); 17336f6c7006SSage Weil } else { 173456e925b6SSage Weil __send_request(osdc, req); 173556e925b6SSage Weil } 17363d14c5d2SYehuda Sadeh rc = 0; 17376f6c7006SSage Weil } 1738234af26fSDan Carpenter 1739234af26fSDan Carpenter out_unlock: 17403d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 17413d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 17423d14c5d2SYehuda Sadeh return rc; 17433d14c5d2SYehuda Sadeh } 17443d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_start_request); 17453d14c5d2SYehuda Sadeh 17463d14c5d2SYehuda Sadeh /* 17473d14c5d2SYehuda Sadeh * wait for a request to complete 17483d14c5d2SYehuda Sadeh */ 17493d14c5d2SYehuda Sadeh int ceph_osdc_wait_request(struct ceph_osd_client *osdc, 17503d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 17513d14c5d2SYehuda Sadeh { 17523d14c5d2SYehuda Sadeh int rc; 17533d14c5d2SYehuda Sadeh 17543d14c5d2SYehuda Sadeh rc = wait_for_completion_interruptible(&req->r_completion); 17553d14c5d2SYehuda Sadeh if (rc < 0) { 17563d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 17573d14c5d2SYehuda Sadeh __cancel_request(req); 17583d14c5d2SYehuda Sadeh __unregister_request(osdc, req); 17593d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 176025845472SSage Weil complete_request(req); 17613d14c5d2SYehuda Sadeh dout("wait_request tid %llu canceled/timed out\n", req->r_tid); 17623d14c5d2SYehuda Sadeh return rc; 17633d14c5d2SYehuda Sadeh } 17643d14c5d2SYehuda Sadeh 17653d14c5d2SYehuda Sadeh dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result); 17663d14c5d2SYehuda Sadeh return req->r_result; 17673d14c5d2SYehuda Sadeh } 17683d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_wait_request); 17693d14c5d2SYehuda Sadeh 17703d14c5d2SYehuda Sadeh /* 17713d14c5d2SYehuda Sadeh * sync - wait for all in-flight requests to flush. avoid starvation. 17723d14c5d2SYehuda Sadeh */ 17733d14c5d2SYehuda Sadeh void ceph_osdc_sync(struct ceph_osd_client *osdc) 17743d14c5d2SYehuda Sadeh { 17753d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 17763d14c5d2SYehuda Sadeh u64 last_tid, next_tid = 0; 17773d14c5d2SYehuda Sadeh 17783d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 17793d14c5d2SYehuda Sadeh last_tid = osdc->last_tid; 17803d14c5d2SYehuda Sadeh while (1) { 17813d14c5d2SYehuda Sadeh req = __lookup_request_ge(osdc, next_tid); 17823d14c5d2SYehuda Sadeh if (!req) 17833d14c5d2SYehuda Sadeh break; 17843d14c5d2SYehuda Sadeh if (req->r_tid > last_tid) 17853d14c5d2SYehuda Sadeh break; 17863d14c5d2SYehuda Sadeh 17873d14c5d2SYehuda Sadeh next_tid = req->r_tid + 1; 17883d14c5d2SYehuda Sadeh if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0) 17893d14c5d2SYehuda Sadeh continue; 17903d14c5d2SYehuda Sadeh 17913d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 17923d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 17933d14c5d2SYehuda Sadeh dout("sync waiting on tid %llu (last is %llu)\n", 17943d14c5d2SYehuda Sadeh req->r_tid, last_tid); 17953d14c5d2SYehuda Sadeh wait_for_completion(&req->r_safe_completion); 17963d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 17973d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 17983d14c5d2SYehuda Sadeh } 17993d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 18003d14c5d2SYehuda Sadeh dout("sync done (thru tid %llu)\n", last_tid); 18013d14c5d2SYehuda Sadeh } 18023d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_sync); 18033d14c5d2SYehuda Sadeh 18043d14c5d2SYehuda Sadeh /* 18053d14c5d2SYehuda Sadeh * init, shutdown 18063d14c5d2SYehuda Sadeh */ 18073d14c5d2SYehuda Sadeh int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client) 18083d14c5d2SYehuda Sadeh { 18093d14c5d2SYehuda Sadeh int err; 18103d14c5d2SYehuda Sadeh 18113d14c5d2SYehuda Sadeh dout("init\n"); 18123d14c5d2SYehuda Sadeh osdc->client = client; 18133d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 18143d14c5d2SYehuda Sadeh init_rwsem(&osdc->map_sem); 18153d14c5d2SYehuda Sadeh init_completion(&osdc->map_waiters); 18163d14c5d2SYehuda Sadeh osdc->last_requested_map = 0; 18173d14c5d2SYehuda Sadeh mutex_init(&osdc->request_mutex); 18183d14c5d2SYehuda Sadeh osdc->last_tid = 0; 18193d14c5d2SYehuda Sadeh osdc->osds = RB_ROOT; 18203d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->osd_lru); 18213d14c5d2SYehuda Sadeh osdc->requests = RB_ROOT; 18223d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_lru); 18236f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_unsent); 18246f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_notarget); 1825a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_linger); 18263d14c5d2SYehuda Sadeh osdc->num_requests = 0; 18273d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout); 18283d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout); 1829a40c4f10SYehuda Sadeh spin_lock_init(&osdc->event_lock); 1830a40c4f10SYehuda Sadeh osdc->event_tree = RB_ROOT; 1831a40c4f10SYehuda Sadeh osdc->event_count = 0; 18323d14c5d2SYehuda Sadeh 18333d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 18343d14c5d2SYehuda Sadeh round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ)); 18353d14c5d2SYehuda Sadeh 18363d14c5d2SYehuda Sadeh err = -ENOMEM; 18373d14c5d2SYehuda Sadeh osdc->req_mempool = mempool_create_kmalloc_pool(10, 18383d14c5d2SYehuda Sadeh sizeof(struct ceph_osd_request)); 18393d14c5d2SYehuda Sadeh if (!osdc->req_mempool) 18403d14c5d2SYehuda Sadeh goto out; 18413d14c5d2SYehuda Sadeh 18423d14c5d2SYehuda Sadeh err = ceph_msgpool_init(&osdc->msgpool_op, OSD_OP_FRONT_LEN, 10, true, 18433d14c5d2SYehuda Sadeh "osd_op"); 18443d14c5d2SYehuda Sadeh if (err < 0) 18453d14c5d2SYehuda Sadeh goto out_mempool; 18463d14c5d2SYehuda Sadeh err = ceph_msgpool_init(&osdc->msgpool_op_reply, 18473d14c5d2SYehuda Sadeh OSD_OPREPLY_FRONT_LEN, 10, true, 18483d14c5d2SYehuda Sadeh "osd_op_reply"); 18493d14c5d2SYehuda Sadeh if (err < 0) 18503d14c5d2SYehuda Sadeh goto out_msgpool; 1851a40c4f10SYehuda Sadeh 1852a40c4f10SYehuda Sadeh osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify"); 1853a40c4f10SYehuda Sadeh if (IS_ERR(osdc->notify_wq)) { 1854a40c4f10SYehuda Sadeh err = PTR_ERR(osdc->notify_wq); 1855a40c4f10SYehuda Sadeh osdc->notify_wq = NULL; 1856a40c4f10SYehuda Sadeh goto out_msgpool; 1857a40c4f10SYehuda Sadeh } 18583d14c5d2SYehuda Sadeh return 0; 18593d14c5d2SYehuda Sadeh 18603d14c5d2SYehuda Sadeh out_msgpool: 18613d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 18623d14c5d2SYehuda Sadeh out_mempool: 18633d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 18643d14c5d2SYehuda Sadeh out: 18653d14c5d2SYehuda Sadeh return err; 18663d14c5d2SYehuda Sadeh } 18673d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_init); 18683d14c5d2SYehuda Sadeh 18693d14c5d2SYehuda Sadeh void ceph_osdc_stop(struct ceph_osd_client *osdc) 18703d14c5d2SYehuda Sadeh { 1871a40c4f10SYehuda Sadeh flush_workqueue(osdc->notify_wq); 1872a40c4f10SYehuda Sadeh destroy_workqueue(osdc->notify_wq); 18733d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->timeout_work); 18743d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->osds_timeout_work); 18753d14c5d2SYehuda Sadeh if (osdc->osdmap) { 18763d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 18773d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 18783d14c5d2SYehuda Sadeh } 1879aca420bcSSage Weil remove_all_osds(osdc); 18803d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 18813d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 18823d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op_reply); 18833d14c5d2SYehuda Sadeh } 18843d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_stop); 18853d14c5d2SYehuda Sadeh 18863d14c5d2SYehuda Sadeh /* 18873d14c5d2SYehuda Sadeh * Read some contiguous pages. If we cross a stripe boundary, shorten 18883d14c5d2SYehuda Sadeh * *plen. Return number of bytes read, or error. 18893d14c5d2SYehuda Sadeh */ 18903d14c5d2SYehuda Sadeh int ceph_osdc_readpages(struct ceph_osd_client *osdc, 18913d14c5d2SYehuda Sadeh struct ceph_vino vino, struct ceph_file_layout *layout, 18923d14c5d2SYehuda Sadeh u64 off, u64 *plen, 18933d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 1894b7495fc2SSage Weil struct page **pages, int num_pages, int page_align) 18953d14c5d2SYehuda Sadeh { 18963d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 18973d14c5d2SYehuda Sadeh int rc = 0; 18983d14c5d2SYehuda Sadeh 18993d14c5d2SYehuda Sadeh dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino, 19003d14c5d2SYehuda Sadeh vino.snap, off, *plen); 19013d14c5d2SYehuda Sadeh req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 19023d14c5d2SYehuda Sadeh CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ, 19033d14c5d2SYehuda Sadeh NULL, 0, truncate_seq, truncate_size, NULL, 1904b7495fc2SSage Weil false, 1, page_align); 19053d14c5d2SYehuda Sadeh if (!req) 19063d14c5d2SYehuda Sadeh return -ENOMEM; 19073d14c5d2SYehuda Sadeh 19083d14c5d2SYehuda Sadeh /* it may be a short read due to an object boundary */ 19093d14c5d2SYehuda Sadeh req->r_pages = pages; 19103d14c5d2SYehuda Sadeh 1911b7495fc2SSage Weil dout("readpages final extent is %llu~%llu (%d pages align %d)\n", 1912b7495fc2SSage Weil off, *plen, req->r_num_pages, page_align); 19133d14c5d2SYehuda Sadeh 19143d14c5d2SYehuda Sadeh rc = ceph_osdc_start_request(osdc, req, false); 19153d14c5d2SYehuda Sadeh if (!rc) 19163d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 19173d14c5d2SYehuda Sadeh 19183d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 19193d14c5d2SYehuda Sadeh dout("readpages result %d\n", rc); 19203d14c5d2SYehuda Sadeh return rc; 19213d14c5d2SYehuda Sadeh } 19223d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_readpages); 19233d14c5d2SYehuda Sadeh 19243d14c5d2SYehuda Sadeh /* 19253d14c5d2SYehuda Sadeh * do a synchronous write on N pages 19263d14c5d2SYehuda Sadeh */ 19273d14c5d2SYehuda Sadeh int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino, 19283d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 19293d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 19303d14c5d2SYehuda Sadeh u64 off, u64 len, 19313d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 19323d14c5d2SYehuda Sadeh struct timespec *mtime, 19333d14c5d2SYehuda Sadeh struct page **pages, int num_pages, 19343d14c5d2SYehuda Sadeh int flags, int do_sync, bool nofail) 19353d14c5d2SYehuda Sadeh { 19363d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 19373d14c5d2SYehuda Sadeh int rc = 0; 1938b7495fc2SSage Weil int page_align = off & ~PAGE_MASK; 19393d14c5d2SYehuda Sadeh 19403d14c5d2SYehuda Sadeh BUG_ON(vino.snap != CEPH_NOSNAP); 19413d14c5d2SYehuda Sadeh req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 19423d14c5d2SYehuda Sadeh CEPH_OSD_OP_WRITE, 19433d14c5d2SYehuda Sadeh flags | CEPH_OSD_FLAG_ONDISK | 19443d14c5d2SYehuda Sadeh CEPH_OSD_FLAG_WRITE, 19453d14c5d2SYehuda Sadeh snapc, do_sync, 19463d14c5d2SYehuda Sadeh truncate_seq, truncate_size, mtime, 1947b7495fc2SSage Weil nofail, 1, page_align); 19483d14c5d2SYehuda Sadeh if (!req) 19493d14c5d2SYehuda Sadeh return -ENOMEM; 19503d14c5d2SYehuda Sadeh 19513d14c5d2SYehuda Sadeh /* it may be a short write due to an object boundary */ 19523d14c5d2SYehuda Sadeh req->r_pages = pages; 19533d14c5d2SYehuda Sadeh dout("writepages %llu~%llu (%d pages)\n", off, len, 19543d14c5d2SYehuda Sadeh req->r_num_pages); 19553d14c5d2SYehuda Sadeh 19563d14c5d2SYehuda Sadeh rc = ceph_osdc_start_request(osdc, req, nofail); 19573d14c5d2SYehuda Sadeh if (!rc) 19583d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 19593d14c5d2SYehuda Sadeh 19603d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 19613d14c5d2SYehuda Sadeh if (rc == 0) 19623d14c5d2SYehuda Sadeh rc = len; 19633d14c5d2SYehuda Sadeh dout("writepages result %d\n", rc); 19643d14c5d2SYehuda Sadeh return rc; 19653d14c5d2SYehuda Sadeh } 19663d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_writepages); 19673d14c5d2SYehuda Sadeh 19683d14c5d2SYehuda Sadeh /* 19693d14c5d2SYehuda Sadeh * handle incoming message 19703d14c5d2SYehuda Sadeh */ 19713d14c5d2SYehuda Sadeh static void dispatch(struct ceph_connection *con, struct ceph_msg *msg) 19723d14c5d2SYehuda Sadeh { 19733d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 19743d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 19753d14c5d2SYehuda Sadeh int type = le16_to_cpu(msg->hdr.type); 19763d14c5d2SYehuda Sadeh 19773d14c5d2SYehuda Sadeh if (!osd) 19783d14c5d2SYehuda Sadeh goto out; 19793d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 19803d14c5d2SYehuda Sadeh 19813d14c5d2SYehuda Sadeh switch (type) { 19823d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 19833d14c5d2SYehuda Sadeh ceph_osdc_handle_map(osdc, msg); 19843d14c5d2SYehuda Sadeh break; 19853d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 19863d14c5d2SYehuda Sadeh handle_reply(osdc, msg, con); 19873d14c5d2SYehuda Sadeh break; 1988a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 1989a40c4f10SYehuda Sadeh handle_watch_notify(osdc, msg); 1990a40c4f10SYehuda Sadeh break; 19913d14c5d2SYehuda Sadeh 19923d14c5d2SYehuda Sadeh default: 19933d14c5d2SYehuda Sadeh pr_err("received unknown message type %d %s\n", type, 19943d14c5d2SYehuda Sadeh ceph_msg_type_name(type)); 19953d14c5d2SYehuda Sadeh } 19963d14c5d2SYehuda Sadeh out: 19973d14c5d2SYehuda Sadeh ceph_msg_put(msg); 19983d14c5d2SYehuda Sadeh } 19993d14c5d2SYehuda Sadeh 20003d14c5d2SYehuda Sadeh /* 20013d14c5d2SYehuda Sadeh * lookup and return message for incoming reply. set up reply message 20023d14c5d2SYehuda Sadeh * pages. 20033d14c5d2SYehuda Sadeh */ 20043d14c5d2SYehuda Sadeh static struct ceph_msg *get_reply(struct ceph_connection *con, 20053d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 20063d14c5d2SYehuda Sadeh int *skip) 20073d14c5d2SYehuda Sadeh { 20083d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 20093d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = osd->o_osdc; 20103d14c5d2SYehuda Sadeh struct ceph_msg *m; 20113d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 20123d14c5d2SYehuda Sadeh int front = le32_to_cpu(hdr->front_len); 20133d14c5d2SYehuda Sadeh int data_len = le32_to_cpu(hdr->data_len); 20143d14c5d2SYehuda Sadeh u64 tid; 20153d14c5d2SYehuda Sadeh 20163d14c5d2SYehuda Sadeh tid = le64_to_cpu(hdr->tid); 20173d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 20183d14c5d2SYehuda Sadeh req = __lookup_request(osdc, tid); 20193d14c5d2SYehuda Sadeh if (!req) { 20203d14c5d2SYehuda Sadeh *skip = 1; 20213d14c5d2SYehuda Sadeh m = NULL; 20223d14c5d2SYehuda Sadeh pr_info("get_reply unknown tid %llu from osd%d\n", tid, 20233d14c5d2SYehuda Sadeh osd->o_osd); 20243d14c5d2SYehuda Sadeh goto out; 20253d14c5d2SYehuda Sadeh } 20263d14c5d2SYehuda Sadeh 20273d14c5d2SYehuda Sadeh if (req->r_con_filling_msg) { 20283d14c5d2SYehuda Sadeh dout("get_reply revoking msg %p from old con %p\n", 20293d14c5d2SYehuda Sadeh req->r_reply, req->r_con_filling_msg); 20303d14c5d2SYehuda Sadeh ceph_con_revoke_message(req->r_con_filling_msg, req->r_reply); 20313d14c5d2SYehuda Sadeh ceph_con_put(req->r_con_filling_msg); 20323d14c5d2SYehuda Sadeh req->r_con_filling_msg = NULL; 20333d14c5d2SYehuda Sadeh } 20343d14c5d2SYehuda Sadeh 20353d14c5d2SYehuda Sadeh if (front > req->r_reply->front.iov_len) { 20363d14c5d2SYehuda Sadeh pr_warning("get_reply front %d > preallocated %d\n", 20373d14c5d2SYehuda Sadeh front, (int)req->r_reply->front.iov_len); 2038b61c2763SSage Weil m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS, false); 20393d14c5d2SYehuda Sadeh if (!m) 20403d14c5d2SYehuda Sadeh goto out; 20413d14c5d2SYehuda Sadeh ceph_msg_put(req->r_reply); 20423d14c5d2SYehuda Sadeh req->r_reply = m; 20433d14c5d2SYehuda Sadeh } 20443d14c5d2SYehuda Sadeh m = ceph_msg_get(req->r_reply); 20453d14c5d2SYehuda Sadeh 20463d14c5d2SYehuda Sadeh if (data_len > 0) { 2047b7495fc2SSage Weil int want = calc_pages_for(req->r_page_alignment, data_len); 20483d14c5d2SYehuda Sadeh 20493d14c5d2SYehuda Sadeh if (unlikely(req->r_num_pages < want)) { 20509bb0ce2bSSage Weil pr_warning("tid %lld reply has %d bytes %d pages, we" 20519bb0ce2bSSage Weil " had only %d pages ready\n", tid, data_len, 20529bb0ce2bSSage Weil want, req->r_num_pages); 20533d14c5d2SYehuda Sadeh *skip = 1; 20543d14c5d2SYehuda Sadeh ceph_msg_put(m); 20553d14c5d2SYehuda Sadeh m = NULL; 20563d14c5d2SYehuda Sadeh goto out; 20573d14c5d2SYehuda Sadeh } 20583d14c5d2SYehuda Sadeh m->pages = req->r_pages; 20593d14c5d2SYehuda Sadeh m->nr_pages = req->r_num_pages; 2060c5c6b19dSSage Weil m->page_alignment = req->r_page_alignment; 20613d14c5d2SYehuda Sadeh #ifdef CONFIG_BLOCK 20623d14c5d2SYehuda Sadeh m->bio = req->r_bio; 20633d14c5d2SYehuda Sadeh #endif 20643d14c5d2SYehuda Sadeh } 20653d14c5d2SYehuda Sadeh *skip = 0; 20663d14c5d2SYehuda Sadeh req->r_con_filling_msg = ceph_con_get(con); 20673d14c5d2SYehuda Sadeh dout("get_reply tid %lld %p\n", tid, m); 20683d14c5d2SYehuda Sadeh 20693d14c5d2SYehuda Sadeh out: 20703d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 20713d14c5d2SYehuda Sadeh return m; 20723d14c5d2SYehuda Sadeh 20733d14c5d2SYehuda Sadeh } 20743d14c5d2SYehuda Sadeh 20753d14c5d2SYehuda Sadeh static struct ceph_msg *alloc_msg(struct ceph_connection *con, 20763d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 20773d14c5d2SYehuda Sadeh int *skip) 20783d14c5d2SYehuda Sadeh { 20793d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 20803d14c5d2SYehuda Sadeh int type = le16_to_cpu(hdr->type); 20813d14c5d2SYehuda Sadeh int front = le32_to_cpu(hdr->front_len); 20823d14c5d2SYehuda Sadeh 20833d14c5d2SYehuda Sadeh switch (type) { 20843d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 2085a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 2086b61c2763SSage Weil return ceph_msg_new(type, front, GFP_NOFS, false); 20873d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 20883d14c5d2SYehuda Sadeh return get_reply(con, hdr, skip); 20893d14c5d2SYehuda Sadeh default: 20903d14c5d2SYehuda Sadeh pr_info("alloc_msg unexpected msg type %d from osd%d\n", type, 20913d14c5d2SYehuda Sadeh osd->o_osd); 20923d14c5d2SYehuda Sadeh *skip = 1; 20933d14c5d2SYehuda Sadeh return NULL; 20943d14c5d2SYehuda Sadeh } 20953d14c5d2SYehuda Sadeh } 20963d14c5d2SYehuda Sadeh 20973d14c5d2SYehuda Sadeh /* 20983d14c5d2SYehuda Sadeh * Wrappers to refcount containing ceph_osd struct 20993d14c5d2SYehuda Sadeh */ 21003d14c5d2SYehuda Sadeh static struct ceph_connection *get_osd_con(struct ceph_connection *con) 21013d14c5d2SYehuda Sadeh { 21023d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 21033d14c5d2SYehuda Sadeh if (get_osd(osd)) 21043d14c5d2SYehuda Sadeh return con; 21053d14c5d2SYehuda Sadeh return NULL; 21063d14c5d2SYehuda Sadeh } 21073d14c5d2SYehuda Sadeh 21083d14c5d2SYehuda Sadeh static void put_osd_con(struct ceph_connection *con) 21093d14c5d2SYehuda Sadeh { 21103d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 21113d14c5d2SYehuda Sadeh put_osd(osd); 21123d14c5d2SYehuda Sadeh } 21133d14c5d2SYehuda Sadeh 21143d14c5d2SYehuda Sadeh /* 21153d14c5d2SYehuda Sadeh * authentication 21163d14c5d2SYehuda Sadeh */ 2117a3530df3SAlex Elder /* 2118a3530df3SAlex Elder * Note: returned pointer is the address of a structure that's 2119a3530df3SAlex Elder * managed separately. Caller must *not* attempt to free it. 2120a3530df3SAlex Elder */ 2121a3530df3SAlex Elder static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con, 21228f43fb53SAlex Elder int *proto, int force_new) 21233d14c5d2SYehuda Sadeh { 21243d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 21253d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 21263d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 212774f1869fSAlex Elder struct ceph_auth_handshake *auth = &o->o_auth; 21283d14c5d2SYehuda Sadeh 212974f1869fSAlex Elder if (force_new && auth->authorizer) { 2130a255651dSAlex Elder if (ac->ops && ac->ops->destroy_authorizer) 213174f1869fSAlex Elder ac->ops->destroy_authorizer(ac, auth->authorizer); 213274f1869fSAlex Elder auth->authorizer = NULL; 21333d14c5d2SYehuda Sadeh } 2134a255651dSAlex Elder if (!auth->authorizer && ac->ops && ac->ops->create_authorizer) { 2135a3530df3SAlex Elder int ret = ac->ops->create_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 2136a3530df3SAlex Elder auth); 21373d14c5d2SYehuda Sadeh if (ret) 2138a3530df3SAlex Elder return ERR_PTR(ret); 21393d14c5d2SYehuda Sadeh } 21403d14c5d2SYehuda Sadeh *proto = ac->protocol; 214174f1869fSAlex Elder 2142a3530df3SAlex Elder return auth; 21433d14c5d2SYehuda Sadeh } 21443d14c5d2SYehuda Sadeh 21453d14c5d2SYehuda Sadeh 21463d14c5d2SYehuda Sadeh static int verify_authorizer_reply(struct ceph_connection *con, int len) 21473d14c5d2SYehuda Sadeh { 21483d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 21493d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 21503d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 21513d14c5d2SYehuda Sadeh 2152a255651dSAlex Elder /* 2153a255651dSAlex Elder * XXX If ac->ops or ac->ops->verify_authorizer_reply is null, 2154a255651dSAlex Elder * XXX which do we do: succeed or fail? 2155a255651dSAlex Elder */ 21566c4a1915SAlex Elder return ac->ops->verify_authorizer_reply(ac, o->o_auth.authorizer, len); 21573d14c5d2SYehuda Sadeh } 21583d14c5d2SYehuda Sadeh 21593d14c5d2SYehuda Sadeh static int invalidate_authorizer(struct ceph_connection *con) 21603d14c5d2SYehuda Sadeh { 21613d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 21623d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 21633d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 21643d14c5d2SYehuda Sadeh 2165a255651dSAlex Elder if (ac->ops && ac->ops->invalidate_authorizer) 21663d14c5d2SYehuda Sadeh ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD); 21673d14c5d2SYehuda Sadeh 21683d14c5d2SYehuda Sadeh return ceph_monc_validate_auth(&osdc->client->monc); 21693d14c5d2SYehuda Sadeh } 21703d14c5d2SYehuda Sadeh 21713d14c5d2SYehuda Sadeh static const struct ceph_connection_operations osd_con_ops = { 21723d14c5d2SYehuda Sadeh .get = get_osd_con, 21733d14c5d2SYehuda Sadeh .put = put_osd_con, 21743d14c5d2SYehuda Sadeh .dispatch = dispatch, 21753d14c5d2SYehuda Sadeh .get_authorizer = get_authorizer, 21763d14c5d2SYehuda Sadeh .verify_authorizer_reply = verify_authorizer_reply, 21773d14c5d2SYehuda Sadeh .invalidate_authorizer = invalidate_authorizer, 21783d14c5d2SYehuda Sadeh .alloc_msg = alloc_msg, 21793d14c5d2SYehuda Sadeh .fault = osd_reset, 21803d14c5d2SYehuda Sadeh }; 2181