13d14c5d2SYehuda Sadeh #include <linux/ceph/ceph_debug.h> 23d14c5d2SYehuda Sadeh 33d14c5d2SYehuda Sadeh #include <linux/module.h> 43d14c5d2SYehuda Sadeh #include <linux/err.h> 53d14c5d2SYehuda Sadeh #include <linux/highmem.h> 63d14c5d2SYehuda Sadeh #include <linux/mm.h> 73d14c5d2SYehuda Sadeh #include <linux/pagemap.h> 83d14c5d2SYehuda Sadeh #include <linux/slab.h> 93d14c5d2SYehuda Sadeh #include <linux/uaccess.h> 103d14c5d2SYehuda Sadeh #ifdef CONFIG_BLOCK 113d14c5d2SYehuda Sadeh #include <linux/bio.h> 123d14c5d2SYehuda Sadeh #endif 133d14c5d2SYehuda Sadeh 143d14c5d2SYehuda Sadeh #include <linux/ceph/libceph.h> 153d14c5d2SYehuda Sadeh #include <linux/ceph/osd_client.h> 163d14c5d2SYehuda Sadeh #include <linux/ceph/messenger.h> 173d14c5d2SYehuda Sadeh #include <linux/ceph/decode.h> 183d14c5d2SYehuda Sadeh #include <linux/ceph/auth.h> 193d14c5d2SYehuda Sadeh #include <linux/ceph/pagelist.h> 203d14c5d2SYehuda Sadeh 213d14c5d2SYehuda Sadeh #define OSD_OP_FRONT_LEN 4096 223d14c5d2SYehuda Sadeh #define OSD_OPREPLY_FRONT_LEN 512 233d14c5d2SYehuda Sadeh 243d14c5d2SYehuda Sadeh static const struct ceph_connection_operations osd_con_ops; 253d14c5d2SYehuda Sadeh 26f9d25199SAlex Elder static void __send_queued(struct ceph_osd_client *osdc); 276f6c7006SSage Weil static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd); 28a40c4f10SYehuda Sadeh static void __register_request(struct ceph_osd_client *osdc, 29a40c4f10SYehuda Sadeh struct ceph_osd_request *req); 30a40c4f10SYehuda Sadeh static void __unregister_linger_request(struct ceph_osd_client *osdc, 31a40c4f10SYehuda Sadeh struct ceph_osd_request *req); 3256e925b6SSage Weil static void __send_request(struct ceph_osd_client *osdc, 33a40c4f10SYehuda Sadeh struct ceph_osd_request *req); 343d14c5d2SYehuda Sadeh 353d14c5d2SYehuda Sadeh static int op_has_extent(int op) 363d14c5d2SYehuda Sadeh { 373d14c5d2SYehuda Sadeh return (op == CEPH_OSD_OP_READ || 383d14c5d2SYehuda Sadeh op == CEPH_OSD_OP_WRITE); 393d14c5d2SYehuda Sadeh } 403d14c5d2SYehuda Sadeh 413d14c5d2SYehuda Sadeh /* 423d14c5d2SYehuda Sadeh * Implement client access to distributed object storage cluster. 433d14c5d2SYehuda Sadeh * 443d14c5d2SYehuda Sadeh * All data objects are stored within a cluster/cloud of OSDs, or 453d14c5d2SYehuda Sadeh * "object storage devices." (Note that Ceph OSDs have _nothing_ to 463d14c5d2SYehuda Sadeh * do with the T10 OSD extensions to SCSI.) Ceph OSDs are simply 473d14c5d2SYehuda Sadeh * remote daemons serving up and coordinating consistent and safe 483d14c5d2SYehuda Sadeh * access to storage. 493d14c5d2SYehuda Sadeh * 503d14c5d2SYehuda Sadeh * Cluster membership and the mapping of data objects onto storage devices 513d14c5d2SYehuda Sadeh * are described by the osd map. 523d14c5d2SYehuda Sadeh * 533d14c5d2SYehuda Sadeh * We keep track of pending OSD requests (read, write), resubmit 543d14c5d2SYehuda Sadeh * requests to different OSDs when the cluster topology/data layout 553d14c5d2SYehuda Sadeh * change, or retry the affected requests when the communications 563d14c5d2SYehuda Sadeh * channel with an OSD is reset. 573d14c5d2SYehuda Sadeh */ 583d14c5d2SYehuda Sadeh 593d14c5d2SYehuda Sadeh /* 603d14c5d2SYehuda Sadeh * calculate the mapping of a file extent onto an object, and fill out the 613d14c5d2SYehuda Sadeh * request accordingly. shorten extent as necessary if it crosses an 623d14c5d2SYehuda Sadeh * object boundary. 633d14c5d2SYehuda Sadeh * 643d14c5d2SYehuda Sadeh * fill osd op in request message. 653d14c5d2SYehuda Sadeh */ 66dbe0fc41SAlex Elder static int calc_layout(struct ceph_file_layout *layout, u64 off, u64 *plen, 673d14c5d2SYehuda Sadeh struct ceph_osd_request *req, 6847a05811SAlex Elder struct ceph_osd_req_op *op, u64 *bno) 693d14c5d2SYehuda Sadeh { 7060e56f13SAlex Elder u64 orig_len = *plen; 7160e56f13SAlex Elder u64 objoff = 0; 7260e56f13SAlex Elder u64 objlen = 0; 73d63b77f4SSage Weil int r; 743d14c5d2SYehuda Sadeh 7560e56f13SAlex Elder /* object extent? */ 7647a05811SAlex Elder r = ceph_calc_file_object_mapping(layout, off, orig_len, bno, 7760e56f13SAlex Elder &objoff, &objlen); 78d63b77f4SSage Weil if (r < 0) 79d63b77f4SSage Weil return r; 8060e56f13SAlex Elder if (objlen < orig_len) { 8160e56f13SAlex Elder *plen = objlen; 8260e56f13SAlex Elder dout(" skipping last %llu, final file extent %llu~%llu\n", 8360e56f13SAlex Elder orig_len - *plen, off, *plen); 8460e56f13SAlex Elder } 8560e56f13SAlex Elder 8660e56f13SAlex Elder if (op_has_extent(op->op)) { 8760e56f13SAlex Elder u32 osize = le32_to_cpu(layout->fl_object_size); 8860e56f13SAlex Elder op->extent.offset = objoff; 8960e56f13SAlex Elder op->extent.length = objlen; 9060e56f13SAlex Elder if (op->extent.truncate_size <= off - objoff) { 9160e56f13SAlex Elder op->extent.truncate_size = 0; 9260e56f13SAlex Elder } else { 9360e56f13SAlex Elder op->extent.truncate_size -= off - objoff; 9460e56f13SAlex Elder if (op->extent.truncate_size > osize) 9560e56f13SAlex Elder op->extent.truncate_size = osize; 9660e56f13SAlex Elder } 9760e56f13SAlex Elder } 9860e56f13SAlex Elder req->r_num_pages = calc_pages_for(off, *plen); 9960e56f13SAlex Elder req->r_page_alignment = off & ~PAGE_MASK; 10060e56f13SAlex Elder if (op->op == CEPH_OSD_OP_WRITE) 10160e56f13SAlex Elder op->payload_len = *plen; 10260e56f13SAlex Elder 10360e56f13SAlex Elder dout("calc_layout bno=%llx %llu~%llu (%d pages)\n", 10447a05811SAlex Elder *bno, objoff, objlen, req->r_num_pages); 1053d14c5d2SYehuda Sadeh 1063ff5f385SAlex Elder return 0; 1073d14c5d2SYehuda Sadeh } 1083d14c5d2SYehuda Sadeh 1093d14c5d2SYehuda Sadeh /* 1103d14c5d2SYehuda Sadeh * requests 1113d14c5d2SYehuda Sadeh */ 1123d14c5d2SYehuda Sadeh void ceph_osdc_release_request(struct kref *kref) 1133d14c5d2SYehuda Sadeh { 1143d14c5d2SYehuda Sadeh struct ceph_osd_request *req = container_of(kref, 1153d14c5d2SYehuda Sadeh struct ceph_osd_request, 1163d14c5d2SYehuda Sadeh r_kref); 1173d14c5d2SYehuda Sadeh 1183d14c5d2SYehuda Sadeh if (req->r_request) 1193d14c5d2SYehuda Sadeh ceph_msg_put(req->r_request); 1203d14c5d2SYehuda Sadeh if (req->r_con_filling_msg) { 1219cbb1d72SAlex Elder dout("%s revoking msg %p from con %p\n", __func__, 1229cbb1d72SAlex Elder req->r_reply, req->r_con_filling_msg); 1238921d114SAlex Elder ceph_msg_revoke_incoming(req->r_reply); 1240d47766fSSage Weil req->r_con_filling_msg->ops->put(req->r_con_filling_msg); 1259cbb1d72SAlex Elder req->r_con_filling_msg = NULL; 1263d14c5d2SYehuda Sadeh } 127ab8cb34aSAlex Elder if (req->r_reply) 128ab8cb34aSAlex Elder ceph_msg_put(req->r_reply); 1293d14c5d2SYehuda Sadeh if (req->r_own_pages) 1303d14c5d2SYehuda Sadeh ceph_release_page_vector(req->r_pages, 1313d14c5d2SYehuda Sadeh req->r_num_pages); 1323d14c5d2SYehuda Sadeh ceph_put_snap_context(req->r_snapc); 133c885837fSAlex Elder ceph_pagelist_release(&req->r_trail); 1343d14c5d2SYehuda Sadeh if (req->r_mempool) 1353d14c5d2SYehuda Sadeh mempool_free(req, req->r_osdc->req_mempool); 1363d14c5d2SYehuda Sadeh else 1373d14c5d2SYehuda Sadeh kfree(req); 1383d14c5d2SYehuda Sadeh } 1393d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_release_request); 1403d14c5d2SYehuda Sadeh 1413d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc, 1423d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 1431b83bef2SSage Weil unsigned int num_ops, 1443d14c5d2SYehuda Sadeh bool use_mempool, 14554a54007SAlex Elder gfp_t gfp_flags) 1463d14c5d2SYehuda Sadeh { 1473d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 1483d14c5d2SYehuda Sadeh struct ceph_msg *msg; 1491b83bef2SSage Weil size_t msg_size; 1503d14c5d2SYehuda Sadeh 1511b83bef2SSage Weil msg_size = 4 + 4 + 8 + 8 + 4+8; 1521b83bef2SSage Weil msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */ 1531b83bef2SSage Weil msg_size += 1 + 8 + 4 + 4; /* pg_t */ 1541b83bef2SSage Weil msg_size += 4 + MAX_OBJ_NAME_SIZE; 1551b83bef2SSage Weil msg_size += 2 + num_ops*sizeof(struct ceph_osd_op); 1561b83bef2SSage Weil msg_size += 8; /* snapid */ 1571b83bef2SSage Weil msg_size += 8; /* snap_seq */ 1581b83bef2SSage Weil msg_size += 8 * (snapc ? snapc->num_snaps : 0); /* snaps */ 1591b83bef2SSage Weil msg_size += 4; 1603d14c5d2SYehuda Sadeh 1613d14c5d2SYehuda Sadeh if (use_mempool) { 1623d14c5d2SYehuda Sadeh req = mempool_alloc(osdc->req_mempool, gfp_flags); 1633d14c5d2SYehuda Sadeh memset(req, 0, sizeof(*req)); 1643d14c5d2SYehuda Sadeh } else { 1653d14c5d2SYehuda Sadeh req = kzalloc(sizeof(*req), gfp_flags); 1663d14c5d2SYehuda Sadeh } 1673d14c5d2SYehuda Sadeh if (req == NULL) 1683d14c5d2SYehuda Sadeh return NULL; 1693d14c5d2SYehuda Sadeh 1703d14c5d2SYehuda Sadeh req->r_osdc = osdc; 1713d14c5d2SYehuda Sadeh req->r_mempool = use_mempool; 1723d14c5d2SYehuda Sadeh 1733d14c5d2SYehuda Sadeh kref_init(&req->r_kref); 1743d14c5d2SYehuda Sadeh init_completion(&req->r_completion); 1753d14c5d2SYehuda Sadeh init_completion(&req->r_safe_completion); 176a978fa20SAlex Elder RB_CLEAR_NODE(&req->r_node); 1773d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&req->r_unsafe_item); 178a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&req->r_linger_item); 179a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&req->r_linger_osd); 180935b639aSSage Weil INIT_LIST_HEAD(&req->r_req_lru_item); 181cd43045cSSage Weil INIT_LIST_HEAD(&req->r_osd_item); 182cd43045cSSage Weil 1833d14c5d2SYehuda Sadeh /* create reply message */ 1843d14c5d2SYehuda Sadeh if (use_mempool) 1853d14c5d2SYehuda Sadeh msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0); 1863d14c5d2SYehuda Sadeh else 1873d14c5d2SYehuda Sadeh msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, 188b61c2763SSage Weil OSD_OPREPLY_FRONT_LEN, gfp_flags, true); 1893d14c5d2SYehuda Sadeh if (!msg) { 1903d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 1913d14c5d2SYehuda Sadeh return NULL; 1923d14c5d2SYehuda Sadeh } 1933d14c5d2SYehuda Sadeh req->r_reply = msg; 1943d14c5d2SYehuda Sadeh 195c885837fSAlex Elder ceph_pagelist_init(&req->r_trail); 196d50b409fSSage Weil 1973d14c5d2SYehuda Sadeh /* create request message; allow space for oid */ 1983d14c5d2SYehuda Sadeh if (use_mempool) 1993d14c5d2SYehuda Sadeh msg = ceph_msgpool_get(&osdc->msgpool_op, 0); 2003d14c5d2SYehuda Sadeh else 201b61c2763SSage Weil msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true); 2023d14c5d2SYehuda Sadeh if (!msg) { 2033d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 2043d14c5d2SYehuda Sadeh return NULL; 2053d14c5d2SYehuda Sadeh } 2063d14c5d2SYehuda Sadeh 2073d14c5d2SYehuda Sadeh memset(msg->front.iov_base, 0, msg->front.iov_len); 2083d14c5d2SYehuda Sadeh 2093d14c5d2SYehuda Sadeh req->r_request = msg; 2103d14c5d2SYehuda Sadeh 2113d14c5d2SYehuda Sadeh return req; 2123d14c5d2SYehuda Sadeh } 2133d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_alloc_request); 2143d14c5d2SYehuda Sadeh 2153d14c5d2SYehuda Sadeh static void osd_req_encode_op(struct ceph_osd_request *req, 2163d14c5d2SYehuda Sadeh struct ceph_osd_op *dst, 2173d14c5d2SYehuda Sadeh struct ceph_osd_req_op *src) 2183d14c5d2SYehuda Sadeh { 2193d14c5d2SYehuda Sadeh dst->op = cpu_to_le16(src->op); 2203d14c5d2SYehuda Sadeh 221065a68f9SAlex Elder switch (src->op) { 222fbfab539SAlex Elder case CEPH_OSD_OP_STAT: 223fbfab539SAlex Elder break; 2243d14c5d2SYehuda Sadeh case CEPH_OSD_OP_READ: 2253d14c5d2SYehuda Sadeh case CEPH_OSD_OP_WRITE: 2263d14c5d2SYehuda Sadeh dst->extent.offset = 2273d14c5d2SYehuda Sadeh cpu_to_le64(src->extent.offset); 2283d14c5d2SYehuda Sadeh dst->extent.length = 2293d14c5d2SYehuda Sadeh cpu_to_le64(src->extent.length); 2303d14c5d2SYehuda Sadeh dst->extent.truncate_size = 2313d14c5d2SYehuda Sadeh cpu_to_le64(src->extent.truncate_size); 2323d14c5d2SYehuda Sadeh dst->extent.truncate_seq = 2333d14c5d2SYehuda Sadeh cpu_to_le32(src->extent.truncate_seq); 2343d14c5d2SYehuda Sadeh break; 2353d14c5d2SYehuda Sadeh case CEPH_OSD_OP_CALL: 2363d14c5d2SYehuda Sadeh dst->cls.class_len = src->cls.class_len; 2373d14c5d2SYehuda Sadeh dst->cls.method_len = src->cls.method_len; 2383d14c5d2SYehuda Sadeh dst->cls.indata_len = cpu_to_le32(src->cls.indata_len); 2393d14c5d2SYehuda Sadeh 240c885837fSAlex Elder ceph_pagelist_append(&req->r_trail, src->cls.class_name, 2413d14c5d2SYehuda Sadeh src->cls.class_len); 242c885837fSAlex Elder ceph_pagelist_append(&req->r_trail, src->cls.method_name, 2433d14c5d2SYehuda Sadeh src->cls.method_len); 244c885837fSAlex Elder ceph_pagelist_append(&req->r_trail, src->cls.indata, 2453d14c5d2SYehuda Sadeh src->cls.indata_len); 2463d14c5d2SYehuda Sadeh break; 2473d14c5d2SYehuda Sadeh case CEPH_OSD_OP_STARTSYNC: 2483d14c5d2SYehuda Sadeh break; 249a40c4f10SYehuda Sadeh case CEPH_OSD_OP_NOTIFY_ACK: 250a40c4f10SYehuda Sadeh case CEPH_OSD_OP_WATCH: 251a40c4f10SYehuda Sadeh dst->watch.cookie = cpu_to_le64(src->watch.cookie); 252a40c4f10SYehuda Sadeh dst->watch.ver = cpu_to_le64(src->watch.ver); 253a40c4f10SYehuda Sadeh dst->watch.flag = src->watch.flag; 254a40c4f10SYehuda Sadeh break; 2553d14c5d2SYehuda Sadeh default: 2563d14c5d2SYehuda Sadeh pr_err("unrecognized osd opcode %d\n", dst->op); 2573d14c5d2SYehuda Sadeh WARN_ON(1); 2583d14c5d2SYehuda Sadeh break; 2594c46459cSAlex Elder case CEPH_OSD_OP_MAPEXT: 2604c46459cSAlex Elder case CEPH_OSD_OP_MASKTRUNC: 2614c46459cSAlex Elder case CEPH_OSD_OP_SPARSE_READ: 262a9f36c3eSAlex Elder case CEPH_OSD_OP_NOTIFY: 2634c46459cSAlex Elder case CEPH_OSD_OP_ASSERT_VER: 2644c46459cSAlex Elder case CEPH_OSD_OP_WRITEFULL: 2654c46459cSAlex Elder case CEPH_OSD_OP_TRUNCATE: 2664c46459cSAlex Elder case CEPH_OSD_OP_ZERO: 2674c46459cSAlex Elder case CEPH_OSD_OP_DELETE: 2684c46459cSAlex Elder case CEPH_OSD_OP_APPEND: 2694c46459cSAlex Elder case CEPH_OSD_OP_SETTRUNC: 2704c46459cSAlex Elder case CEPH_OSD_OP_TRIMTRUNC: 2714c46459cSAlex Elder case CEPH_OSD_OP_TMAPUP: 2724c46459cSAlex Elder case CEPH_OSD_OP_TMAPPUT: 2734c46459cSAlex Elder case CEPH_OSD_OP_TMAPGET: 2744c46459cSAlex Elder case CEPH_OSD_OP_CREATE: 275a9f36c3eSAlex Elder case CEPH_OSD_OP_ROLLBACK: 2764c46459cSAlex Elder case CEPH_OSD_OP_OMAPGETKEYS: 2774c46459cSAlex Elder case CEPH_OSD_OP_OMAPGETVALS: 2784c46459cSAlex Elder case CEPH_OSD_OP_OMAPGETHEADER: 2794c46459cSAlex Elder case CEPH_OSD_OP_OMAPGETVALSBYKEYS: 2804c46459cSAlex Elder case CEPH_OSD_OP_MODE_RD: 2814c46459cSAlex Elder case CEPH_OSD_OP_OMAPSETVALS: 2824c46459cSAlex Elder case CEPH_OSD_OP_OMAPSETHEADER: 2834c46459cSAlex Elder case CEPH_OSD_OP_OMAPCLEAR: 2844c46459cSAlex Elder case CEPH_OSD_OP_OMAPRMKEYS: 2854c46459cSAlex Elder case CEPH_OSD_OP_OMAP_CMP: 2864c46459cSAlex Elder case CEPH_OSD_OP_CLONERANGE: 2874c46459cSAlex Elder case CEPH_OSD_OP_ASSERT_SRC_VERSION: 2884c46459cSAlex Elder case CEPH_OSD_OP_SRC_CMPXATTR: 289a9f36c3eSAlex Elder case CEPH_OSD_OP_GETXATTR: 2904c46459cSAlex Elder case CEPH_OSD_OP_GETXATTRS: 291a9f36c3eSAlex Elder case CEPH_OSD_OP_CMPXATTR: 292a9f36c3eSAlex Elder case CEPH_OSD_OP_SETXATTR: 2934c46459cSAlex Elder case CEPH_OSD_OP_SETXATTRS: 2944c46459cSAlex Elder case CEPH_OSD_OP_RESETXATTRS: 2954c46459cSAlex Elder case CEPH_OSD_OP_RMXATTR: 2964c46459cSAlex Elder case CEPH_OSD_OP_PULL: 2974c46459cSAlex Elder case CEPH_OSD_OP_PUSH: 2984c46459cSAlex Elder case CEPH_OSD_OP_BALANCEREADS: 2994c46459cSAlex Elder case CEPH_OSD_OP_UNBALANCEREADS: 3004c46459cSAlex Elder case CEPH_OSD_OP_SCRUB: 3014c46459cSAlex Elder case CEPH_OSD_OP_SCRUB_RESERVE: 3024c46459cSAlex Elder case CEPH_OSD_OP_SCRUB_UNRESERVE: 3034c46459cSAlex Elder case CEPH_OSD_OP_SCRUB_STOP: 3044c46459cSAlex Elder case CEPH_OSD_OP_SCRUB_MAP: 3054c46459cSAlex Elder case CEPH_OSD_OP_WRLOCK: 3064c46459cSAlex Elder case CEPH_OSD_OP_WRUNLOCK: 3074c46459cSAlex Elder case CEPH_OSD_OP_RDLOCK: 3084c46459cSAlex Elder case CEPH_OSD_OP_RDUNLOCK: 3094c46459cSAlex Elder case CEPH_OSD_OP_UPLOCK: 3104c46459cSAlex Elder case CEPH_OSD_OP_DNLOCK: 3114c46459cSAlex Elder case CEPH_OSD_OP_PGLS: 3124c46459cSAlex Elder case CEPH_OSD_OP_PGLS_FILTER: 3134c46459cSAlex Elder pr_err("unsupported osd opcode %s\n", 3144c46459cSAlex Elder ceph_osd_op_name(dst->op)); 3154c46459cSAlex Elder WARN_ON(1); 3164c46459cSAlex Elder break; 3173d14c5d2SYehuda Sadeh } 3183d14c5d2SYehuda Sadeh dst->payload_len = cpu_to_le32(src->payload_len); 3193d14c5d2SYehuda Sadeh } 3203d14c5d2SYehuda Sadeh 3213d14c5d2SYehuda Sadeh /* 3223d14c5d2SYehuda Sadeh * build new request AND message 3233d14c5d2SYehuda Sadeh * 3243d14c5d2SYehuda Sadeh */ 3253d14c5d2SYehuda Sadeh void ceph_osdc_build_request(struct ceph_osd_request *req, 3261b83bef2SSage Weil u64 off, u64 len, unsigned int num_ops, 3273d14c5d2SYehuda Sadeh struct ceph_osd_req_op *src_ops, 3284d6b250bSAlex Elder struct ceph_snap_context *snapc, u64 snap_id, 329af77f26cSAlex Elder struct timespec *mtime) 3303d14c5d2SYehuda Sadeh { 3313d14c5d2SYehuda Sadeh struct ceph_msg *msg = req->r_request; 3323d14c5d2SYehuda Sadeh struct ceph_osd_req_op *src_op; 3333d14c5d2SYehuda Sadeh void *p; 3341b83bef2SSage Weil size_t msg_size; 3353d14c5d2SYehuda Sadeh int flags = req->r_flags; 336f44246e3SAlex Elder u64 data_len; 3373d14c5d2SYehuda Sadeh int i; 3383d14c5d2SYehuda Sadeh 3391b83bef2SSage Weil req->r_num_ops = num_ops; 3401b83bef2SSage Weil req->r_snapid = snap_id; 3413d14c5d2SYehuda Sadeh req->r_snapc = ceph_get_snap_context(snapc); 3423d14c5d2SYehuda Sadeh 3431b83bef2SSage Weil /* encode request */ 3441b83bef2SSage Weil msg->hdr.version = cpu_to_le16(4); 3453d14c5d2SYehuda Sadeh 3461b83bef2SSage Weil p = msg->front.iov_base; 3471b83bef2SSage Weil ceph_encode_32(&p, 1); /* client_inc is always 1 */ 3481b83bef2SSage Weil req->r_request_osdmap_epoch = p; 3491b83bef2SSage Weil p += 4; 3501b83bef2SSage Weil req->r_request_flags = p; 3511b83bef2SSage Weil p += 4; 3521b83bef2SSage Weil if (req->r_flags & CEPH_OSD_FLAG_WRITE) 3531b83bef2SSage Weil ceph_encode_timespec(p, mtime); 3541b83bef2SSage Weil p += sizeof(struct ceph_timespec); 3551b83bef2SSage Weil req->r_request_reassert_version = p; 3561b83bef2SSage Weil p += sizeof(struct ceph_eversion); /* will get filled in */ 3571b83bef2SSage Weil 3581b83bef2SSage Weil /* oloc */ 3591b83bef2SSage Weil ceph_encode_8(&p, 4); 3601b83bef2SSage Weil ceph_encode_8(&p, 4); 3611b83bef2SSage Weil ceph_encode_32(&p, 8 + 4 + 4); 3621b83bef2SSage Weil req->r_request_pool = p; 3631b83bef2SSage Weil p += 8; 3641b83bef2SSage Weil ceph_encode_32(&p, -1); /* preferred */ 3651b83bef2SSage Weil ceph_encode_32(&p, 0); /* key len */ 3661b83bef2SSage Weil 3671b83bef2SSage Weil ceph_encode_8(&p, 1); 3681b83bef2SSage Weil req->r_request_pgid = p; 3691b83bef2SSage Weil p += 8 + 4; 3701b83bef2SSage Weil ceph_encode_32(&p, -1); /* preferred */ 3711b83bef2SSage Weil 3721b83bef2SSage Weil /* oid */ 3731b83bef2SSage Weil ceph_encode_32(&p, req->r_oid_len); 374af77f26cSAlex Elder memcpy(p, req->r_oid, req->r_oid_len); 3751b83bef2SSage Weil dout("oid '%.*s' len %d\n", req->r_oid_len, req->r_oid, req->r_oid_len); 376af77f26cSAlex Elder p += req->r_oid_len; 3773d14c5d2SYehuda Sadeh 3781b83bef2SSage Weil /* ops */ 3791b83bef2SSage Weil ceph_encode_16(&p, num_ops); 3803d14c5d2SYehuda Sadeh src_op = src_ops; 3811b83bef2SSage Weil req->r_request_ops = p; 3821b83bef2SSage Weil for (i = 0; i < num_ops; i++, src_op++) { 3831b83bef2SSage Weil osd_req_encode_op(req, p, src_op); 3841b83bef2SSage Weil p += sizeof(struct ceph_osd_op); 3851b83bef2SSage Weil } 3863d14c5d2SYehuda Sadeh 3871b83bef2SSage Weil /* snaps */ 3881b83bef2SSage Weil ceph_encode_64(&p, req->r_snapid); 3891b83bef2SSage Weil ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0); 3901b83bef2SSage Weil ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0); 3911b83bef2SSage Weil if (req->r_snapc) { 3923d14c5d2SYehuda Sadeh for (i = 0; i < snapc->num_snaps; i++) { 3931b83bef2SSage Weil ceph_encode_64(&p, req->r_snapc->snaps[i]); 3943d14c5d2SYehuda Sadeh } 3953d14c5d2SYehuda Sadeh } 3963d14c5d2SYehuda Sadeh 3971b83bef2SSage Weil req->r_request_attempts = p; 3981b83bef2SSage Weil p += 4; 3991b83bef2SSage Weil 400f44246e3SAlex Elder data_len = req->r_trail.length; 4013d14c5d2SYehuda Sadeh if (flags & CEPH_OSD_FLAG_WRITE) { 4023d14c5d2SYehuda Sadeh req->r_request->hdr.data_off = cpu_to_le16(off); 403f44246e3SAlex Elder data_len += len; 4043d14c5d2SYehuda Sadeh } 405f44246e3SAlex Elder req->r_request->hdr.data_len = cpu_to_le32(data_len); 406c5c6b19dSSage Weil req->r_request->page_alignment = req->r_page_alignment; 407c5c6b19dSSage Weil 4083d14c5d2SYehuda Sadeh BUG_ON(p > msg->front.iov_base + msg->front.iov_len); 4093d14c5d2SYehuda Sadeh msg_size = p - msg->front.iov_base; 4103d14c5d2SYehuda Sadeh msg->front.iov_len = msg_size; 4113d14c5d2SYehuda Sadeh msg->hdr.front_len = cpu_to_le32(msg_size); 4121b83bef2SSage Weil 4131b83bef2SSage Weil dout("build_request msg_size was %d num_ops %d\n", (int)msg_size, 4141b83bef2SSage Weil num_ops); 4153d14c5d2SYehuda Sadeh return; 4163d14c5d2SYehuda Sadeh } 4173d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_build_request); 4183d14c5d2SYehuda Sadeh 4193d14c5d2SYehuda Sadeh /* 4203d14c5d2SYehuda Sadeh * build new request AND message, calculate layout, and adjust file 4213d14c5d2SYehuda Sadeh * extent as needed. 4223d14c5d2SYehuda Sadeh * 4233d14c5d2SYehuda Sadeh * if the file was recently truncated, we include information about its 4243d14c5d2SYehuda Sadeh * old and new size so that the object can be updated appropriately. (we 4253d14c5d2SYehuda Sadeh * avoid synchronously deleting truncated objects because it's slow.) 4263d14c5d2SYehuda Sadeh * 4273d14c5d2SYehuda Sadeh * if @do_sync, include a 'startsync' command so that the osd will flush 4283d14c5d2SYehuda Sadeh * data quickly. 4293d14c5d2SYehuda Sadeh */ 4303d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc, 4313d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 4323d14c5d2SYehuda Sadeh struct ceph_vino vino, 4333d14c5d2SYehuda Sadeh u64 off, u64 *plen, 4343d14c5d2SYehuda Sadeh int opcode, int flags, 4353d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 4363d14c5d2SYehuda Sadeh int do_sync, 4373d14c5d2SYehuda Sadeh u32 truncate_seq, 4383d14c5d2SYehuda Sadeh u64 truncate_size, 4393d14c5d2SYehuda Sadeh struct timespec *mtime, 440a3bea47eSAlex Elder bool use_mempool, 441b7495fc2SSage Weil int page_align) 4423d14c5d2SYehuda Sadeh { 443ae7ca4a3SAlex Elder struct ceph_osd_req_op ops[2]; 4443d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 445ae7ca4a3SAlex Elder unsigned int num_op = 1; 44647a05811SAlex Elder u64 bno = 0; 4476816282dSSage Weil int r; 4483d14c5d2SYehuda Sadeh 449ae7ca4a3SAlex Elder memset(&ops, 0, sizeof ops); 450ae7ca4a3SAlex Elder 4513d14c5d2SYehuda Sadeh ops[0].op = opcode; 4523d14c5d2SYehuda Sadeh ops[0].extent.truncate_seq = truncate_seq; 4533d14c5d2SYehuda Sadeh ops[0].extent.truncate_size = truncate_size; 4543d14c5d2SYehuda Sadeh 4553d14c5d2SYehuda Sadeh if (do_sync) { 4563d14c5d2SYehuda Sadeh ops[1].op = CEPH_OSD_OP_STARTSYNC; 457ae7ca4a3SAlex Elder num_op++; 458ae7ca4a3SAlex Elder } 4593d14c5d2SYehuda Sadeh 460ae7ca4a3SAlex Elder req = ceph_osdc_alloc_request(osdc, snapc, num_op, use_mempool, 461ae7ca4a3SAlex Elder GFP_NOFS); 4624ad12621SSage Weil if (!req) 4636816282dSSage Weil return ERR_PTR(-ENOMEM); 464d178a9e7SAlex Elder req->r_flags = flags; 4653d14c5d2SYehuda Sadeh 4663d14c5d2SYehuda Sadeh /* calculate max write size */ 467dbe0fc41SAlex Elder r = calc_layout(layout, off, plen, req, ops, &bno); 4683ff5f385SAlex Elder if (r < 0) { 4693ff5f385SAlex Elder ceph_osdc_put_request(req); 4706816282dSSage Weil return ERR_PTR(r); 4713ff5f385SAlex Elder } 47247a05811SAlex Elder 4733d14c5d2SYehuda Sadeh req->r_file_layout = *layout; /* keep a copy */ 4743d14c5d2SYehuda Sadeh 475dbe0fc41SAlex Elder snprintf(req->r_oid, sizeof(req->r_oid), "%llx.%08llx", vino.ino, bno); 476dbe0fc41SAlex Elder req->r_oid_len = strlen(req->r_oid); 477dbe0fc41SAlex Elder 4789bb0ce2bSSage Weil /* in case it differs from natural (file) alignment that 4799bb0ce2bSSage Weil calc_layout filled in for us */ 4809bb0ce2bSSage Weil req->r_num_pages = calc_pages_for(page_align, *plen); 481b7495fc2SSage Weil req->r_page_alignment = page_align; 482b7495fc2SSage Weil 483ae7ca4a3SAlex Elder ceph_osdc_build_request(req, off, *plen, num_op, ops, 484ae7ca4a3SAlex Elder snapc, vino.snap, mtime); 4853d14c5d2SYehuda Sadeh 4863d14c5d2SYehuda Sadeh return req; 4873d14c5d2SYehuda Sadeh } 4883d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_new_request); 4893d14c5d2SYehuda Sadeh 4903d14c5d2SYehuda Sadeh /* 4913d14c5d2SYehuda Sadeh * We keep osd requests in an rbtree, sorted by ->r_tid. 4923d14c5d2SYehuda Sadeh */ 4933d14c5d2SYehuda Sadeh static void __insert_request(struct ceph_osd_client *osdc, 4943d14c5d2SYehuda Sadeh struct ceph_osd_request *new) 4953d14c5d2SYehuda Sadeh { 4963d14c5d2SYehuda Sadeh struct rb_node **p = &osdc->requests.rb_node; 4973d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 4983d14c5d2SYehuda Sadeh struct ceph_osd_request *req = NULL; 4993d14c5d2SYehuda Sadeh 5003d14c5d2SYehuda Sadeh while (*p) { 5013d14c5d2SYehuda Sadeh parent = *p; 5023d14c5d2SYehuda Sadeh req = rb_entry(parent, struct ceph_osd_request, r_node); 5033d14c5d2SYehuda Sadeh if (new->r_tid < req->r_tid) 5043d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 5053d14c5d2SYehuda Sadeh else if (new->r_tid > req->r_tid) 5063d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 5073d14c5d2SYehuda Sadeh else 5083d14c5d2SYehuda Sadeh BUG(); 5093d14c5d2SYehuda Sadeh } 5103d14c5d2SYehuda Sadeh 5113d14c5d2SYehuda Sadeh rb_link_node(&new->r_node, parent, p); 5123d14c5d2SYehuda Sadeh rb_insert_color(&new->r_node, &osdc->requests); 5133d14c5d2SYehuda Sadeh } 5143d14c5d2SYehuda Sadeh 5153d14c5d2SYehuda Sadeh static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc, 5163d14c5d2SYehuda Sadeh u64 tid) 5173d14c5d2SYehuda Sadeh { 5183d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 5193d14c5d2SYehuda Sadeh struct rb_node *n = osdc->requests.rb_node; 5203d14c5d2SYehuda Sadeh 5213d14c5d2SYehuda Sadeh while (n) { 5223d14c5d2SYehuda Sadeh req = rb_entry(n, struct ceph_osd_request, r_node); 5233d14c5d2SYehuda Sadeh if (tid < req->r_tid) 5243d14c5d2SYehuda Sadeh n = n->rb_left; 5253d14c5d2SYehuda Sadeh else if (tid > req->r_tid) 5263d14c5d2SYehuda Sadeh n = n->rb_right; 5273d14c5d2SYehuda Sadeh else 5283d14c5d2SYehuda Sadeh return req; 5293d14c5d2SYehuda Sadeh } 5303d14c5d2SYehuda Sadeh return NULL; 5313d14c5d2SYehuda Sadeh } 5323d14c5d2SYehuda Sadeh 5333d14c5d2SYehuda Sadeh static struct ceph_osd_request * 5343d14c5d2SYehuda Sadeh __lookup_request_ge(struct ceph_osd_client *osdc, 5353d14c5d2SYehuda Sadeh u64 tid) 5363d14c5d2SYehuda Sadeh { 5373d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 5383d14c5d2SYehuda Sadeh struct rb_node *n = osdc->requests.rb_node; 5393d14c5d2SYehuda Sadeh 5403d14c5d2SYehuda Sadeh while (n) { 5413d14c5d2SYehuda Sadeh req = rb_entry(n, struct ceph_osd_request, r_node); 5423d14c5d2SYehuda Sadeh if (tid < req->r_tid) { 5433d14c5d2SYehuda Sadeh if (!n->rb_left) 5443d14c5d2SYehuda Sadeh return req; 5453d14c5d2SYehuda Sadeh n = n->rb_left; 5463d14c5d2SYehuda Sadeh } else if (tid > req->r_tid) { 5473d14c5d2SYehuda Sadeh n = n->rb_right; 5483d14c5d2SYehuda Sadeh } else { 5493d14c5d2SYehuda Sadeh return req; 5503d14c5d2SYehuda Sadeh } 5513d14c5d2SYehuda Sadeh } 5523d14c5d2SYehuda Sadeh return NULL; 5533d14c5d2SYehuda Sadeh } 5543d14c5d2SYehuda Sadeh 5556f6c7006SSage Weil /* 5566f6c7006SSage Weil * Resubmit requests pending on the given osd. 5576f6c7006SSage Weil */ 5586f6c7006SSage Weil static void __kick_osd_requests(struct ceph_osd_client *osdc, 5596f6c7006SSage Weil struct ceph_osd *osd) 5606f6c7006SSage Weil { 561a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 5626f6c7006SSage Weil int err; 5636f6c7006SSage Weil 5646f6c7006SSage Weil dout("__kick_osd_requests osd%d\n", osd->o_osd); 5656f6c7006SSage Weil err = __reset_osd(osdc, osd); 566685a7555SAlex Elder if (err) 5676f6c7006SSage Weil return; 5686f6c7006SSage Weil 5696f6c7006SSage Weil list_for_each_entry(req, &osd->o_requests, r_osd_item) { 5706f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_unsent); 5716f6c7006SSage Weil dout("requeued %p tid %llu osd%d\n", req, req->r_tid, 5726f6c7006SSage Weil osd->o_osd); 573a40c4f10SYehuda Sadeh if (!req->r_linger) 5746f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 5756f6c7006SSage Weil } 576a40c4f10SYehuda Sadeh 577a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osd->o_linger_requests, 578a40c4f10SYehuda Sadeh r_linger_osd) { 57977f38e0eSSage Weil /* 58077f38e0eSSage Weil * reregister request prior to unregistering linger so 58177f38e0eSSage Weil * that r_osd is preserved. 58277f38e0eSSage Weil */ 58377f38e0eSSage Weil BUG_ON(!list_empty(&req->r_req_lru_item)); 584a40c4f10SYehuda Sadeh __register_request(osdc, req); 58577f38e0eSSage Weil list_add(&req->r_req_lru_item, &osdc->req_unsent); 58677f38e0eSSage Weil list_add(&req->r_osd_item, &req->r_osd->o_requests); 58777f38e0eSSage Weil __unregister_linger_request(osdc, req); 588a40c4f10SYehuda Sadeh dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid, 589a40c4f10SYehuda Sadeh osd->o_osd); 590a40c4f10SYehuda Sadeh } 5916f6c7006SSage Weil } 5926f6c7006SSage Weil 5933d14c5d2SYehuda Sadeh /* 5943d14c5d2SYehuda Sadeh * If the osd connection drops, we need to resubmit all requests. 5953d14c5d2SYehuda Sadeh */ 5963d14c5d2SYehuda Sadeh static void osd_reset(struct ceph_connection *con) 5973d14c5d2SYehuda Sadeh { 5983d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 5993d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 6003d14c5d2SYehuda Sadeh 6013d14c5d2SYehuda Sadeh if (!osd) 6023d14c5d2SYehuda Sadeh return; 6033d14c5d2SYehuda Sadeh dout("osd_reset osd%d\n", osd->o_osd); 6043d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 6053d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 60683aff95eSSage Weil mutex_lock(&osdc->request_mutex); 60783aff95eSSage Weil __kick_osd_requests(osdc, osd); 608f9d25199SAlex Elder __send_queued(osdc); 60983aff95eSSage Weil mutex_unlock(&osdc->request_mutex); 6103d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 6113d14c5d2SYehuda Sadeh } 6123d14c5d2SYehuda Sadeh 6133d14c5d2SYehuda Sadeh /* 6143d14c5d2SYehuda Sadeh * Track open sessions with osds. 6153d14c5d2SYehuda Sadeh */ 616e10006f8SAlex Elder static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum) 6173d14c5d2SYehuda Sadeh { 6183d14c5d2SYehuda Sadeh struct ceph_osd *osd; 6193d14c5d2SYehuda Sadeh 6203d14c5d2SYehuda Sadeh osd = kzalloc(sizeof(*osd), GFP_NOFS); 6213d14c5d2SYehuda Sadeh if (!osd) 6223d14c5d2SYehuda Sadeh return NULL; 6233d14c5d2SYehuda Sadeh 6243d14c5d2SYehuda Sadeh atomic_set(&osd->o_ref, 1); 6253d14c5d2SYehuda Sadeh osd->o_osdc = osdc; 626e10006f8SAlex Elder osd->o_osd = onum; 627f407731dSAlex Elder RB_CLEAR_NODE(&osd->o_node); 6283d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_requests); 629a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osd->o_linger_requests); 6303d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_osd_lru); 6313d14c5d2SYehuda Sadeh osd->o_incarnation = 1; 6323d14c5d2SYehuda Sadeh 633b7a9e5ddSSage Weil ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr); 6343d14c5d2SYehuda Sadeh 6353d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_keepalive_item); 6363d14c5d2SYehuda Sadeh return osd; 6373d14c5d2SYehuda Sadeh } 6383d14c5d2SYehuda Sadeh 6393d14c5d2SYehuda Sadeh static struct ceph_osd *get_osd(struct ceph_osd *osd) 6403d14c5d2SYehuda Sadeh { 6413d14c5d2SYehuda Sadeh if (atomic_inc_not_zero(&osd->o_ref)) { 6423d14c5d2SYehuda Sadeh dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1, 6433d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref)); 6443d14c5d2SYehuda Sadeh return osd; 6453d14c5d2SYehuda Sadeh } else { 6463d14c5d2SYehuda Sadeh dout("get_osd %p FAIL\n", osd); 6473d14c5d2SYehuda Sadeh return NULL; 6483d14c5d2SYehuda Sadeh } 6493d14c5d2SYehuda Sadeh } 6503d14c5d2SYehuda Sadeh 6513d14c5d2SYehuda Sadeh static void put_osd(struct ceph_osd *osd) 6523d14c5d2SYehuda Sadeh { 6533d14c5d2SYehuda Sadeh dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref), 6543d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref) - 1); 655a255651dSAlex Elder if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) { 6563d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth; 6573d14c5d2SYehuda Sadeh 658a255651dSAlex Elder if (ac->ops && ac->ops->destroy_authorizer) 6596c4a1915SAlex Elder ac->ops->destroy_authorizer(ac, osd->o_auth.authorizer); 6603d14c5d2SYehuda Sadeh kfree(osd); 6613d14c5d2SYehuda Sadeh } 6623d14c5d2SYehuda Sadeh } 6633d14c5d2SYehuda Sadeh 6643d14c5d2SYehuda Sadeh /* 6653d14c5d2SYehuda Sadeh * remove an osd from our map 6663d14c5d2SYehuda Sadeh */ 6673d14c5d2SYehuda Sadeh static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 6683d14c5d2SYehuda Sadeh { 6693d14c5d2SYehuda Sadeh dout("__remove_osd %p\n", osd); 6703d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_requests)); 6713d14c5d2SYehuda Sadeh rb_erase(&osd->o_node, &osdc->osds); 6723d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 6733d14c5d2SYehuda Sadeh ceph_con_close(&osd->o_con); 6743d14c5d2SYehuda Sadeh put_osd(osd); 6753d14c5d2SYehuda Sadeh } 6763d14c5d2SYehuda Sadeh 677aca420bcSSage Weil static void remove_all_osds(struct ceph_osd_client *osdc) 678aca420bcSSage Weil { 679048a9d2dSJiaju Zhang dout("%s %p\n", __func__, osdc); 680aca420bcSSage Weil mutex_lock(&osdc->request_mutex); 681aca420bcSSage Weil while (!RB_EMPTY_ROOT(&osdc->osds)) { 682aca420bcSSage Weil struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds), 683aca420bcSSage Weil struct ceph_osd, o_node); 684aca420bcSSage Weil __remove_osd(osdc, osd); 685aca420bcSSage Weil } 686aca420bcSSage Weil mutex_unlock(&osdc->request_mutex); 687aca420bcSSage Weil } 688aca420bcSSage Weil 6893d14c5d2SYehuda Sadeh static void __move_osd_to_lru(struct ceph_osd_client *osdc, 6903d14c5d2SYehuda Sadeh struct ceph_osd *osd) 6913d14c5d2SYehuda Sadeh { 6923d14c5d2SYehuda Sadeh dout("__move_osd_to_lru %p\n", osd); 6933d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_osd_lru)); 6943d14c5d2SYehuda Sadeh list_add_tail(&osd->o_osd_lru, &osdc->osd_lru); 6953d14c5d2SYehuda Sadeh osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ; 6963d14c5d2SYehuda Sadeh } 6973d14c5d2SYehuda Sadeh 6983d14c5d2SYehuda Sadeh static void __remove_osd_from_lru(struct ceph_osd *osd) 6993d14c5d2SYehuda Sadeh { 7003d14c5d2SYehuda Sadeh dout("__remove_osd_from_lru %p\n", osd); 7013d14c5d2SYehuda Sadeh if (!list_empty(&osd->o_osd_lru)) 7023d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 7033d14c5d2SYehuda Sadeh } 7043d14c5d2SYehuda Sadeh 705aca420bcSSage Weil static void remove_old_osds(struct ceph_osd_client *osdc) 7063d14c5d2SYehuda Sadeh { 7073d14c5d2SYehuda Sadeh struct ceph_osd *osd, *nosd; 7083d14c5d2SYehuda Sadeh 7093d14c5d2SYehuda Sadeh dout("__remove_old_osds %p\n", osdc); 7103d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 7113d14c5d2SYehuda Sadeh list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) { 712aca420bcSSage Weil if (time_before(jiffies, osd->lru_ttl)) 7133d14c5d2SYehuda Sadeh break; 7143d14c5d2SYehuda Sadeh __remove_osd(osdc, osd); 7153d14c5d2SYehuda Sadeh } 7163d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 7173d14c5d2SYehuda Sadeh } 7183d14c5d2SYehuda Sadeh 7193d14c5d2SYehuda Sadeh /* 7203d14c5d2SYehuda Sadeh * reset osd connect 7213d14c5d2SYehuda Sadeh */ 7223d14c5d2SYehuda Sadeh static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 7233d14c5d2SYehuda Sadeh { 724c3acb181SAlex Elder struct ceph_entity_addr *peer_addr; 7253d14c5d2SYehuda Sadeh 7263d14c5d2SYehuda Sadeh dout("__reset_osd %p osd%d\n", osd, osd->o_osd); 727a40c4f10SYehuda Sadeh if (list_empty(&osd->o_requests) && 728a40c4f10SYehuda Sadeh list_empty(&osd->o_linger_requests)) { 7293d14c5d2SYehuda Sadeh __remove_osd(osdc, osd); 730c3acb181SAlex Elder 731c3acb181SAlex Elder return -ENODEV; 732c3acb181SAlex Elder } 733c3acb181SAlex Elder 734c3acb181SAlex Elder peer_addr = &osdc->osdmap->osd_addr[osd->o_osd]; 735c3acb181SAlex Elder if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) && 7363d14c5d2SYehuda Sadeh !ceph_con_opened(&osd->o_con)) { 737c3acb181SAlex Elder struct ceph_osd_request *req; 738c3acb181SAlex Elder 7393d14c5d2SYehuda Sadeh dout(" osd addr hasn't changed and connection never opened," 7403d14c5d2SYehuda Sadeh " letting msgr retry"); 7413d14c5d2SYehuda Sadeh /* touch each r_stamp for handle_timeout()'s benfit */ 7423d14c5d2SYehuda Sadeh list_for_each_entry(req, &osd->o_requests, r_osd_item) 7433d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 744c3acb181SAlex Elder 745c3acb181SAlex Elder return -EAGAIN; 7463d14c5d2SYehuda Sadeh } 747c3acb181SAlex Elder 748c3acb181SAlex Elder ceph_con_close(&osd->o_con); 749c3acb181SAlex Elder ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr); 750c3acb181SAlex Elder osd->o_incarnation++; 751c3acb181SAlex Elder 752c3acb181SAlex Elder return 0; 7533d14c5d2SYehuda Sadeh } 7543d14c5d2SYehuda Sadeh 7553d14c5d2SYehuda Sadeh static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new) 7563d14c5d2SYehuda Sadeh { 7573d14c5d2SYehuda Sadeh struct rb_node **p = &osdc->osds.rb_node; 7583d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 7593d14c5d2SYehuda Sadeh struct ceph_osd *osd = NULL; 7603d14c5d2SYehuda Sadeh 761aca420bcSSage Weil dout("__insert_osd %p osd%d\n", new, new->o_osd); 7623d14c5d2SYehuda Sadeh while (*p) { 7633d14c5d2SYehuda Sadeh parent = *p; 7643d14c5d2SYehuda Sadeh osd = rb_entry(parent, struct ceph_osd, o_node); 7653d14c5d2SYehuda Sadeh if (new->o_osd < osd->o_osd) 7663d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 7673d14c5d2SYehuda Sadeh else if (new->o_osd > osd->o_osd) 7683d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 7693d14c5d2SYehuda Sadeh else 7703d14c5d2SYehuda Sadeh BUG(); 7713d14c5d2SYehuda Sadeh } 7723d14c5d2SYehuda Sadeh 7733d14c5d2SYehuda Sadeh rb_link_node(&new->o_node, parent, p); 7743d14c5d2SYehuda Sadeh rb_insert_color(&new->o_node, &osdc->osds); 7753d14c5d2SYehuda Sadeh } 7763d14c5d2SYehuda Sadeh 7773d14c5d2SYehuda Sadeh static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o) 7783d14c5d2SYehuda Sadeh { 7793d14c5d2SYehuda Sadeh struct ceph_osd *osd; 7803d14c5d2SYehuda Sadeh struct rb_node *n = osdc->osds.rb_node; 7813d14c5d2SYehuda Sadeh 7823d14c5d2SYehuda Sadeh while (n) { 7833d14c5d2SYehuda Sadeh osd = rb_entry(n, struct ceph_osd, o_node); 7843d14c5d2SYehuda Sadeh if (o < osd->o_osd) 7853d14c5d2SYehuda Sadeh n = n->rb_left; 7863d14c5d2SYehuda Sadeh else if (o > osd->o_osd) 7873d14c5d2SYehuda Sadeh n = n->rb_right; 7883d14c5d2SYehuda Sadeh else 7893d14c5d2SYehuda Sadeh return osd; 7903d14c5d2SYehuda Sadeh } 7913d14c5d2SYehuda Sadeh return NULL; 7923d14c5d2SYehuda Sadeh } 7933d14c5d2SYehuda Sadeh 7943d14c5d2SYehuda Sadeh static void __schedule_osd_timeout(struct ceph_osd_client *osdc) 7953d14c5d2SYehuda Sadeh { 7963d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->timeout_work, 7973d14c5d2SYehuda Sadeh osdc->client->options->osd_keepalive_timeout * HZ); 7983d14c5d2SYehuda Sadeh } 7993d14c5d2SYehuda Sadeh 8003d14c5d2SYehuda Sadeh static void __cancel_osd_timeout(struct ceph_osd_client *osdc) 8013d14c5d2SYehuda Sadeh { 8023d14c5d2SYehuda Sadeh cancel_delayed_work(&osdc->timeout_work); 8033d14c5d2SYehuda Sadeh } 8043d14c5d2SYehuda Sadeh 8053d14c5d2SYehuda Sadeh /* 8063d14c5d2SYehuda Sadeh * Register request, assign tid. If this is the first request, set up 8073d14c5d2SYehuda Sadeh * the timeout event. 8083d14c5d2SYehuda Sadeh */ 809a40c4f10SYehuda Sadeh static void __register_request(struct ceph_osd_client *osdc, 8103d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 8113d14c5d2SYehuda Sadeh { 8123d14c5d2SYehuda Sadeh req->r_tid = ++osdc->last_tid; 8133d14c5d2SYehuda Sadeh req->r_request->hdr.tid = cpu_to_le64(req->r_tid); 81477f38e0eSSage Weil dout("__register_request %p tid %lld\n", req, req->r_tid); 8153d14c5d2SYehuda Sadeh __insert_request(osdc, req); 8163d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 8173d14c5d2SYehuda Sadeh osdc->num_requests++; 8183d14c5d2SYehuda Sadeh if (osdc->num_requests == 1) { 8193d14c5d2SYehuda Sadeh dout(" first request, scheduling timeout\n"); 8203d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 8213d14c5d2SYehuda Sadeh } 822a40c4f10SYehuda Sadeh } 823a40c4f10SYehuda Sadeh 824a40c4f10SYehuda Sadeh static void register_request(struct ceph_osd_client *osdc, 825a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 826a40c4f10SYehuda Sadeh { 827a40c4f10SYehuda Sadeh mutex_lock(&osdc->request_mutex); 828a40c4f10SYehuda Sadeh __register_request(osdc, req); 8293d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 8303d14c5d2SYehuda Sadeh } 8313d14c5d2SYehuda Sadeh 8323d14c5d2SYehuda Sadeh /* 8333d14c5d2SYehuda Sadeh * called under osdc->request_mutex 8343d14c5d2SYehuda Sadeh */ 8353d14c5d2SYehuda Sadeh static void __unregister_request(struct ceph_osd_client *osdc, 8363d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 8373d14c5d2SYehuda Sadeh { 83835f9f8a0SSage Weil if (RB_EMPTY_NODE(&req->r_node)) { 83935f9f8a0SSage Weil dout("__unregister_request %p tid %lld not registered\n", 84035f9f8a0SSage Weil req, req->r_tid); 84135f9f8a0SSage Weil return; 84235f9f8a0SSage Weil } 84335f9f8a0SSage Weil 8443d14c5d2SYehuda Sadeh dout("__unregister_request %p tid %lld\n", req, req->r_tid); 8453d14c5d2SYehuda Sadeh rb_erase(&req->r_node, &osdc->requests); 8463d14c5d2SYehuda Sadeh osdc->num_requests--; 8473d14c5d2SYehuda Sadeh 8483d14c5d2SYehuda Sadeh if (req->r_osd) { 8493d14c5d2SYehuda Sadeh /* make sure the original request isn't in flight. */ 8506740a845SAlex Elder ceph_msg_revoke(req->r_request); 8513d14c5d2SYehuda Sadeh 8523d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 853a40c4f10SYehuda Sadeh if (list_empty(&req->r_osd->o_requests) && 854a40c4f10SYehuda Sadeh list_empty(&req->r_osd->o_linger_requests)) { 855a40c4f10SYehuda Sadeh dout("moving osd to %p lru\n", req->r_osd); 8563d14c5d2SYehuda Sadeh __move_osd_to_lru(osdc, req->r_osd); 857a40c4f10SYehuda Sadeh } 858fbdb9190SSage Weil if (list_empty(&req->r_linger_item)) 8593d14c5d2SYehuda Sadeh req->r_osd = NULL; 8603d14c5d2SYehuda Sadeh } 8613d14c5d2SYehuda Sadeh 8627d5f2481SAlex Elder list_del_init(&req->r_req_lru_item); 8633d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 8643d14c5d2SYehuda Sadeh 8653d14c5d2SYehuda Sadeh if (osdc->num_requests == 0) { 8663d14c5d2SYehuda Sadeh dout(" no requests, canceling timeout\n"); 8673d14c5d2SYehuda Sadeh __cancel_osd_timeout(osdc); 8683d14c5d2SYehuda Sadeh } 8693d14c5d2SYehuda Sadeh } 8703d14c5d2SYehuda Sadeh 8713d14c5d2SYehuda Sadeh /* 8723d14c5d2SYehuda Sadeh * Cancel a previously queued request message 8733d14c5d2SYehuda Sadeh */ 8743d14c5d2SYehuda Sadeh static void __cancel_request(struct ceph_osd_request *req) 8753d14c5d2SYehuda Sadeh { 8763d14c5d2SYehuda Sadeh if (req->r_sent && req->r_osd) { 8776740a845SAlex Elder ceph_msg_revoke(req->r_request); 8783d14c5d2SYehuda Sadeh req->r_sent = 0; 8793d14c5d2SYehuda Sadeh } 8803d14c5d2SYehuda Sadeh } 8813d14c5d2SYehuda Sadeh 882a40c4f10SYehuda Sadeh static void __register_linger_request(struct ceph_osd_client *osdc, 883a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 884a40c4f10SYehuda Sadeh { 885a40c4f10SYehuda Sadeh dout("__register_linger_request %p\n", req); 886a40c4f10SYehuda Sadeh list_add_tail(&req->r_linger_item, &osdc->req_linger); 8876194ea89SSage Weil if (req->r_osd) 8886194ea89SSage Weil list_add_tail(&req->r_linger_osd, 8896194ea89SSage Weil &req->r_osd->o_linger_requests); 890a40c4f10SYehuda Sadeh } 891a40c4f10SYehuda Sadeh 892a40c4f10SYehuda Sadeh static void __unregister_linger_request(struct ceph_osd_client *osdc, 893a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 894a40c4f10SYehuda Sadeh { 895a40c4f10SYehuda Sadeh dout("__unregister_linger_request %p\n", req); 896a40c4f10SYehuda Sadeh list_del_init(&req->r_linger_item); 89761c74035SAlex Elder if (req->r_osd) { 898a40c4f10SYehuda Sadeh list_del_init(&req->r_linger_osd); 899a40c4f10SYehuda Sadeh 900a40c4f10SYehuda Sadeh if (list_empty(&req->r_osd->o_requests) && 901a40c4f10SYehuda Sadeh list_empty(&req->r_osd->o_linger_requests)) { 902a40c4f10SYehuda Sadeh dout("moving osd to %p lru\n", req->r_osd); 903a40c4f10SYehuda Sadeh __move_osd_to_lru(osdc, req->r_osd); 904a40c4f10SYehuda Sadeh } 905fbdb9190SSage Weil if (list_empty(&req->r_osd_item)) 906a40c4f10SYehuda Sadeh req->r_osd = NULL; 907a40c4f10SYehuda Sadeh } 908a40c4f10SYehuda Sadeh } 909a40c4f10SYehuda Sadeh 910a40c4f10SYehuda Sadeh void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc, 911a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 912a40c4f10SYehuda Sadeh { 913a40c4f10SYehuda Sadeh mutex_lock(&osdc->request_mutex); 914a40c4f10SYehuda Sadeh if (req->r_linger) { 915a40c4f10SYehuda Sadeh __unregister_linger_request(osdc, req); 916a40c4f10SYehuda Sadeh ceph_osdc_put_request(req); 917a40c4f10SYehuda Sadeh } 918a40c4f10SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 919a40c4f10SYehuda Sadeh } 920a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_unregister_linger_request); 921a40c4f10SYehuda Sadeh 922a40c4f10SYehuda Sadeh void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc, 923a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 924a40c4f10SYehuda Sadeh { 925a40c4f10SYehuda Sadeh if (!req->r_linger) { 926a40c4f10SYehuda Sadeh dout("set_request_linger %p\n", req); 927a40c4f10SYehuda Sadeh req->r_linger = 1; 928a40c4f10SYehuda Sadeh /* 929a40c4f10SYehuda Sadeh * caller is now responsible for calling 930a40c4f10SYehuda Sadeh * unregister_linger_request 931a40c4f10SYehuda Sadeh */ 932a40c4f10SYehuda Sadeh ceph_osdc_get_request(req); 933a40c4f10SYehuda Sadeh } 934a40c4f10SYehuda Sadeh } 935a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_set_request_linger); 936a40c4f10SYehuda Sadeh 9373d14c5d2SYehuda Sadeh /* 9383d14c5d2SYehuda Sadeh * Pick an osd (the first 'up' osd in the pg), allocate the osd struct 9393d14c5d2SYehuda Sadeh * (as needed), and set the request r_osd appropriately. If there is 94025985edcSLucas De Marchi * no up osd, set r_osd to NULL. Move the request to the appropriate list 9416f6c7006SSage Weil * (unsent, homeless) or leave on in-flight lru. 9423d14c5d2SYehuda Sadeh * 9433d14c5d2SYehuda Sadeh * Return 0 if unchanged, 1 if changed, or negative on error. 9443d14c5d2SYehuda Sadeh * 9453d14c5d2SYehuda Sadeh * Caller should hold map_sem for read and request_mutex. 9463d14c5d2SYehuda Sadeh */ 9476f6c7006SSage Weil static int __map_request(struct ceph_osd_client *osdc, 94838d6453cSSage Weil struct ceph_osd_request *req, int force_resend) 9493d14c5d2SYehuda Sadeh { 9505b191d99SSage Weil struct ceph_pg pgid; 9513d14c5d2SYehuda Sadeh int acting[CEPH_PG_MAX_SIZE]; 9523d14c5d2SYehuda Sadeh int o = -1, num = 0; 9533d14c5d2SYehuda Sadeh int err; 9543d14c5d2SYehuda Sadeh 9556f6c7006SSage Weil dout("map_request %p tid %lld\n", req, req->r_tid); 9562169aea6SSage Weil err = ceph_calc_object_layout(&pgid, req->r_oid, 9573d14c5d2SYehuda Sadeh &req->r_file_layout, osdc->osdmap); 9586f6c7006SSage Weil if (err) { 9596f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 9603d14c5d2SYehuda Sadeh return err; 9616f6c7006SSage Weil } 9623d14c5d2SYehuda Sadeh req->r_pgid = pgid; 9633d14c5d2SYehuda Sadeh 9643d14c5d2SYehuda Sadeh err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting); 9653d14c5d2SYehuda Sadeh if (err > 0) { 9663d14c5d2SYehuda Sadeh o = acting[0]; 9673d14c5d2SYehuda Sadeh num = err; 9683d14c5d2SYehuda Sadeh } 9693d14c5d2SYehuda Sadeh 97038d6453cSSage Weil if ((!force_resend && 97138d6453cSSage Weil req->r_osd && req->r_osd->o_osd == o && 9723d14c5d2SYehuda Sadeh req->r_sent >= req->r_osd->o_incarnation && 9733d14c5d2SYehuda Sadeh req->r_num_pg_osds == num && 9743d14c5d2SYehuda Sadeh memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) || 9753d14c5d2SYehuda Sadeh (req->r_osd == NULL && o == -1)) 9763d14c5d2SYehuda Sadeh return 0; /* no change */ 9773d14c5d2SYehuda Sadeh 9785b191d99SSage Weil dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n", 9795b191d99SSage Weil req->r_tid, pgid.pool, pgid.seed, o, 9803d14c5d2SYehuda Sadeh req->r_osd ? req->r_osd->o_osd : -1); 9813d14c5d2SYehuda Sadeh 9823d14c5d2SYehuda Sadeh /* record full pg acting set */ 9833d14c5d2SYehuda Sadeh memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num); 9843d14c5d2SYehuda Sadeh req->r_num_pg_osds = num; 9853d14c5d2SYehuda Sadeh 9863d14c5d2SYehuda Sadeh if (req->r_osd) { 9873d14c5d2SYehuda Sadeh __cancel_request(req); 9883d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 9893d14c5d2SYehuda Sadeh req->r_osd = NULL; 9903d14c5d2SYehuda Sadeh } 9913d14c5d2SYehuda Sadeh 9923d14c5d2SYehuda Sadeh req->r_osd = __lookup_osd(osdc, o); 9933d14c5d2SYehuda Sadeh if (!req->r_osd && o >= 0) { 9943d14c5d2SYehuda Sadeh err = -ENOMEM; 995e10006f8SAlex Elder req->r_osd = create_osd(osdc, o); 9966f6c7006SSage Weil if (!req->r_osd) { 9976f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 9983d14c5d2SYehuda Sadeh goto out; 9996f6c7006SSage Weil } 10003d14c5d2SYehuda Sadeh 10016f6c7006SSage Weil dout("map_request osd %p is osd%d\n", req->r_osd, o); 10023d14c5d2SYehuda Sadeh __insert_osd(osdc, req->r_osd); 10033d14c5d2SYehuda Sadeh 1004b7a9e5ddSSage Weil ceph_con_open(&req->r_osd->o_con, 1005b7a9e5ddSSage Weil CEPH_ENTITY_TYPE_OSD, o, 1006b7a9e5ddSSage Weil &osdc->osdmap->osd_addr[o]); 10073d14c5d2SYehuda Sadeh } 10083d14c5d2SYehuda Sadeh 10093d14c5d2SYehuda Sadeh if (req->r_osd) { 10103d14c5d2SYehuda Sadeh __remove_osd_from_lru(req->r_osd); 10113d14c5d2SYehuda Sadeh list_add(&req->r_osd_item, &req->r_osd->o_requests); 10126f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_unsent); 10136f6c7006SSage Weil } else { 10146f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 10153d14c5d2SYehuda Sadeh } 10163d14c5d2SYehuda Sadeh err = 1; /* osd or pg changed */ 10173d14c5d2SYehuda Sadeh 10183d14c5d2SYehuda Sadeh out: 10193d14c5d2SYehuda Sadeh return err; 10203d14c5d2SYehuda Sadeh } 10213d14c5d2SYehuda Sadeh 10223d14c5d2SYehuda Sadeh /* 10233d14c5d2SYehuda Sadeh * caller should hold map_sem (for read) and request_mutex 10243d14c5d2SYehuda Sadeh */ 102556e925b6SSage Weil static void __send_request(struct ceph_osd_client *osdc, 10263d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 10273d14c5d2SYehuda Sadeh { 10281b83bef2SSage Weil void *p; 10293d14c5d2SYehuda Sadeh 10301b83bef2SSage Weil dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n", 10311b83bef2SSage Weil req, req->r_tid, req->r_osd->o_osd, req->r_flags, 10321b83bef2SSage Weil (unsigned long long)req->r_pgid.pool, req->r_pgid.seed); 10333d14c5d2SYehuda Sadeh 10341b83bef2SSage Weil /* fill in message content that changes each time we send it */ 10351b83bef2SSage Weil put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch); 10361b83bef2SSage Weil put_unaligned_le32(req->r_flags, req->r_request_flags); 10371b83bef2SSage Weil put_unaligned_le64(req->r_pgid.pool, req->r_request_pool); 10381b83bef2SSage Weil p = req->r_request_pgid; 10391b83bef2SSage Weil ceph_encode_64(&p, req->r_pgid.pool); 10401b83bef2SSage Weil ceph_encode_32(&p, req->r_pgid.seed); 10411b83bef2SSage Weil put_unaligned_le64(1, req->r_request_attempts); /* FIXME */ 10421b83bef2SSage Weil memcpy(req->r_request_reassert_version, &req->r_reassert_version, 10431b83bef2SSage Weil sizeof(req->r_reassert_version)); 10442169aea6SSage Weil 10453d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 10463d14c5d2SYehuda Sadeh list_move_tail(&req->r_req_lru_item, &osdc->req_lru); 10473d14c5d2SYehuda Sadeh 10483d14c5d2SYehuda Sadeh ceph_msg_get(req->r_request); /* send consumes a ref */ 10493d14c5d2SYehuda Sadeh ceph_con_send(&req->r_osd->o_con, req->r_request); 10503d14c5d2SYehuda Sadeh req->r_sent = req->r_osd->o_incarnation; 10513d14c5d2SYehuda Sadeh } 10523d14c5d2SYehuda Sadeh 10533d14c5d2SYehuda Sadeh /* 10546f6c7006SSage Weil * Send any requests in the queue (req_unsent). 10556f6c7006SSage Weil */ 1056f9d25199SAlex Elder static void __send_queued(struct ceph_osd_client *osdc) 10576f6c7006SSage Weil { 10586f6c7006SSage Weil struct ceph_osd_request *req, *tmp; 10596f6c7006SSage Weil 1060f9d25199SAlex Elder dout("__send_queued\n"); 1061f9d25199SAlex Elder list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) 10626f6c7006SSage Weil __send_request(osdc, req); 10636f6c7006SSage Weil } 10646f6c7006SSage Weil 10656f6c7006SSage Weil /* 10663d14c5d2SYehuda Sadeh * Timeout callback, called every N seconds when 1 or more osd 10673d14c5d2SYehuda Sadeh * requests has been active for more than N seconds. When this 10683d14c5d2SYehuda Sadeh * happens, we ping all OSDs with requests who have timed out to 10693d14c5d2SYehuda Sadeh * ensure any communications channel reset is detected. Reset the 10703d14c5d2SYehuda Sadeh * request timeouts another N seconds in the future as we go. 10713d14c5d2SYehuda Sadeh * Reschedule the timeout event another N seconds in future (unless 10723d14c5d2SYehuda Sadeh * there are no open requests). 10733d14c5d2SYehuda Sadeh */ 10743d14c5d2SYehuda Sadeh static void handle_timeout(struct work_struct *work) 10753d14c5d2SYehuda Sadeh { 10763d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 10773d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, timeout_work.work); 107883aff95eSSage Weil struct ceph_osd_request *req; 10793d14c5d2SYehuda Sadeh struct ceph_osd *osd; 10803d14c5d2SYehuda Sadeh unsigned long keepalive = 10813d14c5d2SYehuda Sadeh osdc->client->options->osd_keepalive_timeout * HZ; 10823d14c5d2SYehuda Sadeh struct list_head slow_osds; 10833d14c5d2SYehuda Sadeh dout("timeout\n"); 10843d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 10853d14c5d2SYehuda Sadeh 10863d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 10873d14c5d2SYehuda Sadeh 10883d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 10893d14c5d2SYehuda Sadeh 10903d14c5d2SYehuda Sadeh /* 10913d14c5d2SYehuda Sadeh * ping osds that are a bit slow. this ensures that if there 10923d14c5d2SYehuda Sadeh * is a break in the TCP connection we will notice, and reopen 10933d14c5d2SYehuda Sadeh * a connection with that osd (from the fault callback). 10943d14c5d2SYehuda Sadeh */ 10953d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&slow_osds); 10963d14c5d2SYehuda Sadeh list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) { 10973d14c5d2SYehuda Sadeh if (time_before(jiffies, req->r_stamp + keepalive)) 10983d14c5d2SYehuda Sadeh break; 10993d14c5d2SYehuda Sadeh 11003d14c5d2SYehuda Sadeh osd = req->r_osd; 11013d14c5d2SYehuda Sadeh BUG_ON(!osd); 11023d14c5d2SYehuda Sadeh dout(" tid %llu is slow, will send keepalive on osd%d\n", 11033d14c5d2SYehuda Sadeh req->r_tid, osd->o_osd); 11043d14c5d2SYehuda Sadeh list_move_tail(&osd->o_keepalive_item, &slow_osds); 11053d14c5d2SYehuda Sadeh } 11063d14c5d2SYehuda Sadeh while (!list_empty(&slow_osds)) { 11073d14c5d2SYehuda Sadeh osd = list_entry(slow_osds.next, struct ceph_osd, 11083d14c5d2SYehuda Sadeh o_keepalive_item); 11093d14c5d2SYehuda Sadeh list_del_init(&osd->o_keepalive_item); 11103d14c5d2SYehuda Sadeh ceph_con_keepalive(&osd->o_con); 11113d14c5d2SYehuda Sadeh } 11123d14c5d2SYehuda Sadeh 11133d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 1114f9d25199SAlex Elder __send_queued(osdc); 11153d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 11163d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 11173d14c5d2SYehuda Sadeh } 11183d14c5d2SYehuda Sadeh 11193d14c5d2SYehuda Sadeh static void handle_osds_timeout(struct work_struct *work) 11203d14c5d2SYehuda Sadeh { 11213d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 11223d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, 11233d14c5d2SYehuda Sadeh osds_timeout_work.work); 11243d14c5d2SYehuda Sadeh unsigned long delay = 11253d14c5d2SYehuda Sadeh osdc->client->options->osd_idle_ttl * HZ >> 2; 11263d14c5d2SYehuda Sadeh 11273d14c5d2SYehuda Sadeh dout("osds timeout\n"); 11283d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 1129aca420bcSSage Weil remove_old_osds(osdc); 11303d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 11313d14c5d2SYehuda Sadeh 11323d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 11333d14c5d2SYehuda Sadeh round_jiffies_relative(delay)); 11343d14c5d2SYehuda Sadeh } 11353d14c5d2SYehuda Sadeh 113625845472SSage Weil static void complete_request(struct ceph_osd_request *req) 113725845472SSage Weil { 113825845472SSage Weil if (req->r_safe_callback) 113925845472SSage Weil req->r_safe_callback(req, NULL); 114025845472SSage Weil complete_all(&req->r_safe_completion); /* fsync waiter */ 114125845472SSage Weil } 114225845472SSage Weil 11431b83bef2SSage Weil static int __decode_pgid(void **p, void *end, struct ceph_pg *pgid) 11441b83bef2SSage Weil { 11451b83bef2SSage Weil __u8 v; 11461b83bef2SSage Weil 11471b83bef2SSage Weil ceph_decode_need(p, end, 1 + 8 + 4 + 4, bad); 11481b83bef2SSage Weil v = ceph_decode_8(p); 11491b83bef2SSage Weil if (v > 1) { 11501b83bef2SSage Weil pr_warning("do not understand pg encoding %d > 1", v); 11511b83bef2SSage Weil return -EINVAL; 11521b83bef2SSage Weil } 11531b83bef2SSage Weil pgid->pool = ceph_decode_64(p); 11541b83bef2SSage Weil pgid->seed = ceph_decode_32(p); 11551b83bef2SSage Weil *p += 4; 11561b83bef2SSage Weil return 0; 11571b83bef2SSage Weil 11581b83bef2SSage Weil bad: 11591b83bef2SSage Weil pr_warning("incomplete pg encoding"); 11601b83bef2SSage Weil return -EINVAL; 11611b83bef2SSage Weil } 11621b83bef2SSage Weil 11633d14c5d2SYehuda Sadeh /* 11643d14c5d2SYehuda Sadeh * handle osd op reply. either call the callback if it is specified, 11653d14c5d2SYehuda Sadeh * or do the completion to wake up the waiting thread. 11663d14c5d2SYehuda Sadeh */ 11673d14c5d2SYehuda Sadeh static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg, 11683d14c5d2SYehuda Sadeh struct ceph_connection *con) 11693d14c5d2SYehuda Sadeh { 11701b83bef2SSage Weil void *p, *end; 11713d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 11723d14c5d2SYehuda Sadeh u64 tid; 11731b83bef2SSage Weil int object_len; 11741b83bef2SSage Weil int numops, payload_len, flags; 11753d14c5d2SYehuda Sadeh s32 result; 11761b83bef2SSage Weil s32 retry_attempt; 11771b83bef2SSage Weil struct ceph_pg pg; 11781b83bef2SSage Weil int err; 11791b83bef2SSage Weil u32 reassert_epoch; 11801b83bef2SSage Weil u64 reassert_version; 11811b83bef2SSage Weil u32 osdmap_epoch; 11821b83bef2SSage Weil int i; 11833d14c5d2SYehuda Sadeh 11843d14c5d2SYehuda Sadeh tid = le64_to_cpu(msg->hdr.tid); 11851b83bef2SSage Weil dout("handle_reply %p tid %llu\n", msg, tid); 11861b83bef2SSage Weil 11871b83bef2SSage Weil p = msg->front.iov_base; 11881b83bef2SSage Weil end = p + msg->front.iov_len; 11891b83bef2SSage Weil 11901b83bef2SSage Weil ceph_decode_need(&p, end, 4, bad); 11911b83bef2SSage Weil object_len = ceph_decode_32(&p); 11921b83bef2SSage Weil ceph_decode_need(&p, end, object_len, bad); 11931b83bef2SSage Weil p += object_len; 11941b83bef2SSage Weil 11951b83bef2SSage Weil err = __decode_pgid(&p, end, &pg); 11961b83bef2SSage Weil if (err) 11973d14c5d2SYehuda Sadeh goto bad; 11981b83bef2SSage Weil 11991b83bef2SSage Weil ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad); 12001b83bef2SSage Weil flags = ceph_decode_64(&p); 12011b83bef2SSage Weil result = ceph_decode_32(&p); 12021b83bef2SSage Weil reassert_epoch = ceph_decode_32(&p); 12031b83bef2SSage Weil reassert_version = ceph_decode_64(&p); 12041b83bef2SSage Weil osdmap_epoch = ceph_decode_32(&p); 12051b83bef2SSage Weil 12063d14c5d2SYehuda Sadeh /* lookup */ 12073d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 12083d14c5d2SYehuda Sadeh req = __lookup_request(osdc, tid); 12093d14c5d2SYehuda Sadeh if (req == NULL) { 12103d14c5d2SYehuda Sadeh dout("handle_reply tid %llu dne\n", tid); 12113d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 12123d14c5d2SYehuda Sadeh return; 12133d14c5d2SYehuda Sadeh } 12143d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 12151b83bef2SSage Weil 12161b83bef2SSage Weil dout("handle_reply %p tid %llu req %p result %d\n", msg, tid, 12171b83bef2SSage Weil req, result); 12181b83bef2SSage Weil 12191b83bef2SSage Weil ceph_decode_need(&p, end, 4, bad); 12201b83bef2SSage Weil numops = ceph_decode_32(&p); 12211b83bef2SSage Weil if (numops > CEPH_OSD_MAX_OP) 12221b83bef2SSage Weil goto bad_put; 12231b83bef2SSage Weil if (numops != req->r_num_ops) 12241b83bef2SSage Weil goto bad_put; 12251b83bef2SSage Weil payload_len = 0; 12261b83bef2SSage Weil ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad); 12271b83bef2SSage Weil for (i = 0; i < numops; i++) { 12281b83bef2SSage Weil struct ceph_osd_op *op = p; 12291b83bef2SSage Weil int len; 12301b83bef2SSage Weil 12311b83bef2SSage Weil len = le32_to_cpu(op->payload_len); 12321b83bef2SSage Weil req->r_reply_op_len[i] = len; 12331b83bef2SSage Weil dout(" op %d has %d bytes\n", i, len); 12341b83bef2SSage Weil payload_len += len; 12351b83bef2SSage Weil p += sizeof(*op); 12361b83bef2SSage Weil } 12371b83bef2SSage Weil if (payload_len != le32_to_cpu(msg->hdr.data_len)) { 12381b83bef2SSage Weil pr_warning("sum of op payload lens %d != data_len %d", 12391b83bef2SSage Weil payload_len, le32_to_cpu(msg->hdr.data_len)); 12401b83bef2SSage Weil goto bad_put; 12411b83bef2SSage Weil } 12421b83bef2SSage Weil 12431b83bef2SSage Weil ceph_decode_need(&p, end, 4 + numops * 4, bad); 12441b83bef2SSage Weil retry_attempt = ceph_decode_32(&p); 12451b83bef2SSage Weil for (i = 0; i < numops; i++) 12461b83bef2SSage Weil req->r_reply_op_result[i] = ceph_decode_32(&p); 12473d14c5d2SYehuda Sadeh 12483d14c5d2SYehuda Sadeh /* 12493d14c5d2SYehuda Sadeh * if this connection filled our message, drop our reference now, to 12503d14c5d2SYehuda Sadeh * avoid a (safe but slower) revoke later. 12513d14c5d2SYehuda Sadeh */ 12523d14c5d2SYehuda Sadeh if (req->r_con_filling_msg == con && req->r_reply == msg) { 12533d14c5d2SYehuda Sadeh dout(" dropping con_filling_msg ref %p\n", con); 12543d14c5d2SYehuda Sadeh req->r_con_filling_msg = NULL; 12550d47766fSSage Weil con->ops->put(con); 12563d14c5d2SYehuda Sadeh } 12573d14c5d2SYehuda Sadeh 12583d14c5d2SYehuda Sadeh if (!req->r_got_reply) { 125995c96174SEric Dumazet unsigned int bytes; 12603d14c5d2SYehuda Sadeh 12611b83bef2SSage Weil req->r_result = result; 12623d14c5d2SYehuda Sadeh bytes = le32_to_cpu(msg->hdr.data_len); 12633d14c5d2SYehuda Sadeh dout("handle_reply result %d bytes %d\n", req->r_result, 12643d14c5d2SYehuda Sadeh bytes); 12653d14c5d2SYehuda Sadeh if (req->r_result == 0) 12663d14c5d2SYehuda Sadeh req->r_result = bytes; 12673d14c5d2SYehuda Sadeh 12683d14c5d2SYehuda Sadeh /* in case this is a write and we need to replay, */ 12691b83bef2SSage Weil req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch); 12701b83bef2SSage Weil req->r_reassert_version.version = cpu_to_le64(reassert_version); 12713d14c5d2SYehuda Sadeh 12723d14c5d2SYehuda Sadeh req->r_got_reply = 1; 12733d14c5d2SYehuda Sadeh } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) { 12743d14c5d2SYehuda Sadeh dout("handle_reply tid %llu dup ack\n", tid); 12753d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 12763d14c5d2SYehuda Sadeh goto done; 12773d14c5d2SYehuda Sadeh } 12783d14c5d2SYehuda Sadeh 12793d14c5d2SYehuda Sadeh dout("handle_reply tid %llu flags %d\n", tid, flags); 12803d14c5d2SYehuda Sadeh 1281a40c4f10SYehuda Sadeh if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK)) 1282a40c4f10SYehuda Sadeh __register_linger_request(osdc, req); 1283a40c4f10SYehuda Sadeh 12843d14c5d2SYehuda Sadeh /* either this is a read, or we got the safe response */ 12853d14c5d2SYehuda Sadeh if (result < 0 || 12863d14c5d2SYehuda Sadeh (flags & CEPH_OSD_FLAG_ONDISK) || 12873d14c5d2SYehuda Sadeh ((flags & CEPH_OSD_FLAG_WRITE) == 0)) 12883d14c5d2SYehuda Sadeh __unregister_request(osdc, req); 12893d14c5d2SYehuda Sadeh 12903d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 12913d14c5d2SYehuda Sadeh 12923d14c5d2SYehuda Sadeh if (req->r_callback) 12933d14c5d2SYehuda Sadeh req->r_callback(req, msg); 12943d14c5d2SYehuda Sadeh else 12953d14c5d2SYehuda Sadeh complete_all(&req->r_completion); 12963d14c5d2SYehuda Sadeh 129725845472SSage Weil if (flags & CEPH_OSD_FLAG_ONDISK) 129825845472SSage Weil complete_request(req); 12993d14c5d2SYehuda Sadeh 13003d14c5d2SYehuda Sadeh done: 1301a40c4f10SYehuda Sadeh dout("req=%p req->r_linger=%d\n", req, req->r_linger); 13023d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 13033d14c5d2SYehuda Sadeh return; 13043d14c5d2SYehuda Sadeh 13051b83bef2SSage Weil bad_put: 13061b83bef2SSage Weil ceph_osdc_put_request(req); 13073d14c5d2SYehuda Sadeh bad: 13081b83bef2SSage Weil pr_err("corrupt osd_op_reply got %d %d\n", 13091b83bef2SSage Weil (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len)); 13103d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 13113d14c5d2SYehuda Sadeh } 13123d14c5d2SYehuda Sadeh 13136f6c7006SSage Weil static void reset_changed_osds(struct ceph_osd_client *osdc) 13143d14c5d2SYehuda Sadeh { 13153d14c5d2SYehuda Sadeh struct rb_node *p, *n; 13163d14c5d2SYehuda Sadeh 13173d14c5d2SYehuda Sadeh for (p = rb_first(&osdc->osds); p; p = n) { 13186f6c7006SSage Weil struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node); 13193d14c5d2SYehuda Sadeh 13203d14c5d2SYehuda Sadeh n = rb_next(p); 13213d14c5d2SYehuda Sadeh if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) || 13223d14c5d2SYehuda Sadeh memcmp(&osd->o_con.peer_addr, 13233d14c5d2SYehuda Sadeh ceph_osd_addr(osdc->osdmap, 13243d14c5d2SYehuda Sadeh osd->o_osd), 13253d14c5d2SYehuda Sadeh sizeof(struct ceph_entity_addr)) != 0) 13263d14c5d2SYehuda Sadeh __reset_osd(osdc, osd); 13273d14c5d2SYehuda Sadeh } 13283d14c5d2SYehuda Sadeh } 13293d14c5d2SYehuda Sadeh 13303d14c5d2SYehuda Sadeh /* 13316f6c7006SSage Weil * Requeue requests whose mapping to an OSD has changed. If requests map to 13326f6c7006SSage Weil * no osd, request a new map. 13333d14c5d2SYehuda Sadeh * 1334e6d50f67SAlex Elder * Caller should hold map_sem for read. 13353d14c5d2SYehuda Sadeh */ 133638d6453cSSage Weil static void kick_requests(struct ceph_osd_client *osdc, int force_resend) 13373d14c5d2SYehuda Sadeh { 1338a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 13396f6c7006SSage Weil struct rb_node *p; 13406f6c7006SSage Weil int needmap = 0; 13416f6c7006SSage Weil int err; 13423d14c5d2SYehuda Sadeh 134338d6453cSSage Weil dout("kick_requests %s\n", force_resend ? " (force resend)" : ""); 13443d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 13456194ea89SSage Weil for (p = rb_first(&osdc->requests); p; ) { 13466f6c7006SSage Weil req = rb_entry(p, struct ceph_osd_request, r_node); 13476194ea89SSage Weil p = rb_next(p); 1348ab60b16dSAlex Elder 1349ab60b16dSAlex Elder /* 1350ab60b16dSAlex Elder * For linger requests that have not yet been 1351ab60b16dSAlex Elder * registered, move them to the linger list; they'll 1352ab60b16dSAlex Elder * be sent to the osd in the loop below. Unregister 1353ab60b16dSAlex Elder * the request before re-registering it as a linger 1354ab60b16dSAlex Elder * request to ensure the __map_request() below 1355ab60b16dSAlex Elder * will decide it needs to be sent. 1356ab60b16dSAlex Elder */ 1357ab60b16dSAlex Elder if (req->r_linger && list_empty(&req->r_linger_item)) { 1358ab60b16dSAlex Elder dout("%p tid %llu restart on osd%d\n", 1359ab60b16dSAlex Elder req, req->r_tid, 1360ab60b16dSAlex Elder req->r_osd ? req->r_osd->o_osd : -1); 1361ab60b16dSAlex Elder __unregister_request(osdc, req); 1362ab60b16dSAlex Elder __register_linger_request(osdc, req); 1363ab60b16dSAlex Elder continue; 1364ab60b16dSAlex Elder } 1365ab60b16dSAlex Elder 136638d6453cSSage Weil err = __map_request(osdc, req, force_resend); 13676f6c7006SSage Weil if (err < 0) 13686f6c7006SSage Weil continue; /* error */ 13696f6c7006SSage Weil if (req->r_osd == NULL) { 13706f6c7006SSage Weil dout("%p tid %llu maps to no osd\n", req, req->r_tid); 13716f6c7006SSage Weil needmap++; /* request a newer map */ 13726f6c7006SSage Weil } else if (err > 0) { 13736194ea89SSage Weil if (!req->r_linger) { 13746194ea89SSage Weil dout("%p tid %llu requeued on osd%d\n", req, 13756194ea89SSage Weil req->r_tid, 13766f6c7006SSage Weil req->r_osd ? req->r_osd->o_osd : -1); 13776f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 13786f6c7006SSage Weil } 13796f6c7006SSage Weil } 13806194ea89SSage Weil } 1381a40c4f10SYehuda Sadeh 1382a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osdc->req_linger, 1383a40c4f10SYehuda Sadeh r_linger_item) { 1384a40c4f10SYehuda Sadeh dout("linger req=%p req->r_osd=%p\n", req, req->r_osd); 1385a40c4f10SYehuda Sadeh 138638d6453cSSage Weil err = __map_request(osdc, req, force_resend); 1387ab60b16dSAlex Elder dout("__map_request returned %d\n", err); 1388a40c4f10SYehuda Sadeh if (err == 0) 1389a40c4f10SYehuda Sadeh continue; /* no change and no osd was specified */ 1390a40c4f10SYehuda Sadeh if (err < 0) 1391a40c4f10SYehuda Sadeh continue; /* hrm! */ 1392a40c4f10SYehuda Sadeh if (req->r_osd == NULL) { 1393a40c4f10SYehuda Sadeh dout("tid %llu maps to no valid osd\n", req->r_tid); 1394a40c4f10SYehuda Sadeh needmap++; /* request a newer map */ 1395a40c4f10SYehuda Sadeh continue; 1396a40c4f10SYehuda Sadeh } 1397a40c4f10SYehuda Sadeh 1398a40c4f10SYehuda Sadeh dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid, 1399a40c4f10SYehuda Sadeh req->r_osd ? req->r_osd->o_osd : -1); 1400a40c4f10SYehuda Sadeh __register_request(osdc, req); 1401c89ce05eSAlex Elder __unregister_linger_request(osdc, req); 1402a40c4f10SYehuda Sadeh } 14033d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 14043d14c5d2SYehuda Sadeh 14053d14c5d2SYehuda Sadeh if (needmap) { 14063d14c5d2SYehuda Sadeh dout("%d requests for down osds, need new map\n", needmap); 14073d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 14083d14c5d2SYehuda Sadeh } 1409e6d50f67SAlex Elder reset_changed_osds(osdc); 14103d14c5d2SYehuda Sadeh } 14116f6c7006SSage Weil 14126f6c7006SSage Weil 14133d14c5d2SYehuda Sadeh /* 14143d14c5d2SYehuda Sadeh * Process updated osd map. 14153d14c5d2SYehuda Sadeh * 14163d14c5d2SYehuda Sadeh * The message contains any number of incremental and full maps, normally 14173d14c5d2SYehuda Sadeh * indicating some sort of topology change in the cluster. Kick requests 14183d14c5d2SYehuda Sadeh * off to different OSDs as needed. 14193d14c5d2SYehuda Sadeh */ 14203d14c5d2SYehuda Sadeh void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg) 14213d14c5d2SYehuda Sadeh { 14223d14c5d2SYehuda Sadeh void *p, *end, *next; 14233d14c5d2SYehuda Sadeh u32 nr_maps, maplen; 14243d14c5d2SYehuda Sadeh u32 epoch; 14253d14c5d2SYehuda Sadeh struct ceph_osdmap *newmap = NULL, *oldmap; 14263d14c5d2SYehuda Sadeh int err; 14273d14c5d2SYehuda Sadeh struct ceph_fsid fsid; 14283d14c5d2SYehuda Sadeh 14293d14c5d2SYehuda Sadeh dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0); 14303d14c5d2SYehuda Sadeh p = msg->front.iov_base; 14313d14c5d2SYehuda Sadeh end = p + msg->front.iov_len; 14323d14c5d2SYehuda Sadeh 14333d14c5d2SYehuda Sadeh /* verify fsid */ 14343d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, sizeof(fsid), bad); 14353d14c5d2SYehuda Sadeh ceph_decode_copy(&p, &fsid, sizeof(fsid)); 14363d14c5d2SYehuda Sadeh if (ceph_check_fsid(osdc->client, &fsid) < 0) 14373d14c5d2SYehuda Sadeh return; 14383d14c5d2SYehuda Sadeh 14393d14c5d2SYehuda Sadeh down_write(&osdc->map_sem); 14403d14c5d2SYehuda Sadeh 14413d14c5d2SYehuda Sadeh /* incremental maps */ 14423d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 14433d14c5d2SYehuda Sadeh dout(" %d inc maps\n", nr_maps); 14443d14c5d2SYehuda Sadeh while (nr_maps > 0) { 14453d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 14463d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 14473d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 14483d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 14493d14c5d2SYehuda Sadeh next = p + maplen; 14503d14c5d2SYehuda Sadeh if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) { 14513d14c5d2SYehuda Sadeh dout("applying incremental map %u len %d\n", 14523d14c5d2SYehuda Sadeh epoch, maplen); 14533d14c5d2SYehuda Sadeh newmap = osdmap_apply_incremental(&p, next, 14543d14c5d2SYehuda Sadeh osdc->osdmap, 145515d9882cSAlex Elder &osdc->client->msgr); 14563d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 14573d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 14583d14c5d2SYehuda Sadeh goto bad; 14593d14c5d2SYehuda Sadeh } 14603d14c5d2SYehuda Sadeh BUG_ON(!newmap); 14613d14c5d2SYehuda Sadeh if (newmap != osdc->osdmap) { 14623d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 14633d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 14643d14c5d2SYehuda Sadeh } 146538d6453cSSage Weil kick_requests(osdc, 0); 14663d14c5d2SYehuda Sadeh } else { 14673d14c5d2SYehuda Sadeh dout("ignoring incremental map %u len %d\n", 14683d14c5d2SYehuda Sadeh epoch, maplen); 14693d14c5d2SYehuda Sadeh } 14703d14c5d2SYehuda Sadeh p = next; 14713d14c5d2SYehuda Sadeh nr_maps--; 14723d14c5d2SYehuda Sadeh } 14733d14c5d2SYehuda Sadeh if (newmap) 14743d14c5d2SYehuda Sadeh goto done; 14753d14c5d2SYehuda Sadeh 14763d14c5d2SYehuda Sadeh /* full maps */ 14773d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 14783d14c5d2SYehuda Sadeh dout(" %d full maps\n", nr_maps); 14793d14c5d2SYehuda Sadeh while (nr_maps) { 14803d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 14813d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 14823d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 14833d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 14843d14c5d2SYehuda Sadeh if (nr_maps > 1) { 14853d14c5d2SYehuda Sadeh dout("skipping non-latest full map %u len %d\n", 14863d14c5d2SYehuda Sadeh epoch, maplen); 14873d14c5d2SYehuda Sadeh } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) { 14883d14c5d2SYehuda Sadeh dout("skipping full map %u len %d, " 14893d14c5d2SYehuda Sadeh "older than our %u\n", epoch, maplen, 14903d14c5d2SYehuda Sadeh osdc->osdmap->epoch); 14913d14c5d2SYehuda Sadeh } else { 149238d6453cSSage Weil int skipped_map = 0; 149338d6453cSSage Weil 14943d14c5d2SYehuda Sadeh dout("taking full map %u len %d\n", epoch, maplen); 14953d14c5d2SYehuda Sadeh newmap = osdmap_decode(&p, p+maplen); 14963d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 14973d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 14983d14c5d2SYehuda Sadeh goto bad; 14993d14c5d2SYehuda Sadeh } 15003d14c5d2SYehuda Sadeh BUG_ON(!newmap); 15013d14c5d2SYehuda Sadeh oldmap = osdc->osdmap; 15023d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 150338d6453cSSage Weil if (oldmap) { 150438d6453cSSage Weil if (oldmap->epoch + 1 < newmap->epoch) 150538d6453cSSage Weil skipped_map = 1; 15063d14c5d2SYehuda Sadeh ceph_osdmap_destroy(oldmap); 150738d6453cSSage Weil } 150838d6453cSSage Weil kick_requests(osdc, skipped_map); 15093d14c5d2SYehuda Sadeh } 15103d14c5d2SYehuda Sadeh p += maplen; 15113d14c5d2SYehuda Sadeh nr_maps--; 15123d14c5d2SYehuda Sadeh } 15133d14c5d2SYehuda Sadeh 15143d14c5d2SYehuda Sadeh done: 15153d14c5d2SYehuda Sadeh downgrade_write(&osdc->map_sem); 15163d14c5d2SYehuda Sadeh ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch); 1517cd634fb6SSage Weil 1518cd634fb6SSage Weil /* 1519cd634fb6SSage Weil * subscribe to subsequent osdmap updates if full to ensure 1520cd634fb6SSage Weil * we find out when we are no longer full and stop returning 1521cd634fb6SSage Weil * ENOSPC. 1522cd634fb6SSage Weil */ 1523cd634fb6SSage Weil if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL)) 1524cd634fb6SSage Weil ceph_monc_request_next_osdmap(&osdc->client->monc); 1525cd634fb6SSage Weil 1526f9d25199SAlex Elder mutex_lock(&osdc->request_mutex); 1527f9d25199SAlex Elder __send_queued(osdc); 1528f9d25199SAlex Elder mutex_unlock(&osdc->request_mutex); 15293d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 15303d14c5d2SYehuda Sadeh wake_up_all(&osdc->client->auth_wq); 15313d14c5d2SYehuda Sadeh return; 15323d14c5d2SYehuda Sadeh 15333d14c5d2SYehuda Sadeh bad: 15343d14c5d2SYehuda Sadeh pr_err("osdc handle_map corrupt msg\n"); 15353d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 15363d14c5d2SYehuda Sadeh up_write(&osdc->map_sem); 15373d14c5d2SYehuda Sadeh return; 15383d14c5d2SYehuda Sadeh } 15393d14c5d2SYehuda Sadeh 15403d14c5d2SYehuda Sadeh /* 1541a40c4f10SYehuda Sadeh * watch/notify callback event infrastructure 1542a40c4f10SYehuda Sadeh * 1543a40c4f10SYehuda Sadeh * These callbacks are used both for watch and notify operations. 1544a40c4f10SYehuda Sadeh */ 1545a40c4f10SYehuda Sadeh static void __release_event(struct kref *kref) 1546a40c4f10SYehuda Sadeh { 1547a40c4f10SYehuda Sadeh struct ceph_osd_event *event = 1548a40c4f10SYehuda Sadeh container_of(kref, struct ceph_osd_event, kref); 1549a40c4f10SYehuda Sadeh 1550a40c4f10SYehuda Sadeh dout("__release_event %p\n", event); 1551a40c4f10SYehuda Sadeh kfree(event); 1552a40c4f10SYehuda Sadeh } 1553a40c4f10SYehuda Sadeh 1554a40c4f10SYehuda Sadeh static void get_event(struct ceph_osd_event *event) 1555a40c4f10SYehuda Sadeh { 1556a40c4f10SYehuda Sadeh kref_get(&event->kref); 1557a40c4f10SYehuda Sadeh } 1558a40c4f10SYehuda Sadeh 1559a40c4f10SYehuda Sadeh void ceph_osdc_put_event(struct ceph_osd_event *event) 1560a40c4f10SYehuda Sadeh { 1561a40c4f10SYehuda Sadeh kref_put(&event->kref, __release_event); 1562a40c4f10SYehuda Sadeh } 1563a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_put_event); 1564a40c4f10SYehuda Sadeh 1565a40c4f10SYehuda Sadeh static void __insert_event(struct ceph_osd_client *osdc, 1566a40c4f10SYehuda Sadeh struct ceph_osd_event *new) 1567a40c4f10SYehuda Sadeh { 1568a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 1569a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 1570a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 1571a40c4f10SYehuda Sadeh 1572a40c4f10SYehuda Sadeh while (*p) { 1573a40c4f10SYehuda Sadeh parent = *p; 1574a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 1575a40c4f10SYehuda Sadeh if (new->cookie < event->cookie) 1576a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 1577a40c4f10SYehuda Sadeh else if (new->cookie > event->cookie) 1578a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 1579a40c4f10SYehuda Sadeh else 1580a40c4f10SYehuda Sadeh BUG(); 1581a40c4f10SYehuda Sadeh } 1582a40c4f10SYehuda Sadeh 1583a40c4f10SYehuda Sadeh rb_link_node(&new->node, parent, p); 1584a40c4f10SYehuda Sadeh rb_insert_color(&new->node, &osdc->event_tree); 1585a40c4f10SYehuda Sadeh } 1586a40c4f10SYehuda Sadeh 1587a40c4f10SYehuda Sadeh static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc, 1588a40c4f10SYehuda Sadeh u64 cookie) 1589a40c4f10SYehuda Sadeh { 1590a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 1591a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 1592a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 1593a40c4f10SYehuda Sadeh 1594a40c4f10SYehuda Sadeh while (*p) { 1595a40c4f10SYehuda Sadeh parent = *p; 1596a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 1597a40c4f10SYehuda Sadeh if (cookie < event->cookie) 1598a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 1599a40c4f10SYehuda Sadeh else if (cookie > event->cookie) 1600a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 1601a40c4f10SYehuda Sadeh else 1602a40c4f10SYehuda Sadeh return event; 1603a40c4f10SYehuda Sadeh } 1604a40c4f10SYehuda Sadeh return NULL; 1605a40c4f10SYehuda Sadeh } 1606a40c4f10SYehuda Sadeh 1607a40c4f10SYehuda Sadeh static void __remove_event(struct ceph_osd_event *event) 1608a40c4f10SYehuda Sadeh { 1609a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 1610a40c4f10SYehuda Sadeh 1611a40c4f10SYehuda Sadeh if (!RB_EMPTY_NODE(&event->node)) { 1612a40c4f10SYehuda Sadeh dout("__remove_event removed %p\n", event); 1613a40c4f10SYehuda Sadeh rb_erase(&event->node, &osdc->event_tree); 1614a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 1615a40c4f10SYehuda Sadeh } else { 1616a40c4f10SYehuda Sadeh dout("__remove_event didn't remove %p\n", event); 1617a40c4f10SYehuda Sadeh } 1618a40c4f10SYehuda Sadeh } 1619a40c4f10SYehuda Sadeh 1620a40c4f10SYehuda Sadeh int ceph_osdc_create_event(struct ceph_osd_client *osdc, 1621a40c4f10SYehuda Sadeh void (*event_cb)(u64, u64, u8, void *), 16223c663bbdSAlex Elder void *data, struct ceph_osd_event **pevent) 1623a40c4f10SYehuda Sadeh { 1624a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 1625a40c4f10SYehuda Sadeh 1626a40c4f10SYehuda Sadeh event = kmalloc(sizeof(*event), GFP_NOIO); 1627a40c4f10SYehuda Sadeh if (!event) 1628a40c4f10SYehuda Sadeh return -ENOMEM; 1629a40c4f10SYehuda Sadeh 1630a40c4f10SYehuda Sadeh dout("create_event %p\n", event); 1631a40c4f10SYehuda Sadeh event->cb = event_cb; 16323c663bbdSAlex Elder event->one_shot = 0; 1633a40c4f10SYehuda Sadeh event->data = data; 1634a40c4f10SYehuda Sadeh event->osdc = osdc; 1635a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&event->osd_node); 16363ee5234dSAlex Elder RB_CLEAR_NODE(&event->node); 1637a40c4f10SYehuda Sadeh kref_init(&event->kref); /* one ref for us */ 1638a40c4f10SYehuda Sadeh kref_get(&event->kref); /* one ref for the caller */ 1639a40c4f10SYehuda Sadeh 1640a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 1641a40c4f10SYehuda Sadeh event->cookie = ++osdc->event_count; 1642a40c4f10SYehuda Sadeh __insert_event(osdc, event); 1643a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 1644a40c4f10SYehuda Sadeh 1645a40c4f10SYehuda Sadeh *pevent = event; 1646a40c4f10SYehuda Sadeh return 0; 1647a40c4f10SYehuda Sadeh } 1648a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_create_event); 1649a40c4f10SYehuda Sadeh 1650a40c4f10SYehuda Sadeh void ceph_osdc_cancel_event(struct ceph_osd_event *event) 1651a40c4f10SYehuda Sadeh { 1652a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 1653a40c4f10SYehuda Sadeh 1654a40c4f10SYehuda Sadeh dout("cancel_event %p\n", event); 1655a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 1656a40c4f10SYehuda Sadeh __remove_event(event); 1657a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 1658a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); /* caller's */ 1659a40c4f10SYehuda Sadeh } 1660a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_cancel_event); 1661a40c4f10SYehuda Sadeh 1662a40c4f10SYehuda Sadeh 1663a40c4f10SYehuda Sadeh static void do_event_work(struct work_struct *work) 1664a40c4f10SYehuda Sadeh { 1665a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work = 1666a40c4f10SYehuda Sadeh container_of(work, struct ceph_osd_event_work, work); 1667a40c4f10SYehuda Sadeh struct ceph_osd_event *event = event_work->event; 1668a40c4f10SYehuda Sadeh u64 ver = event_work->ver; 1669a40c4f10SYehuda Sadeh u64 notify_id = event_work->notify_id; 1670a40c4f10SYehuda Sadeh u8 opcode = event_work->opcode; 1671a40c4f10SYehuda Sadeh 1672a40c4f10SYehuda Sadeh dout("do_event_work completing %p\n", event); 1673a40c4f10SYehuda Sadeh event->cb(ver, notify_id, opcode, event->data); 1674a40c4f10SYehuda Sadeh dout("do_event_work completed %p\n", event); 1675a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 1676a40c4f10SYehuda Sadeh kfree(event_work); 1677a40c4f10SYehuda Sadeh } 1678a40c4f10SYehuda Sadeh 1679a40c4f10SYehuda Sadeh 1680a40c4f10SYehuda Sadeh /* 1681a40c4f10SYehuda Sadeh * Process osd watch notifications 1682a40c4f10SYehuda Sadeh */ 16833c663bbdSAlex Elder static void handle_watch_notify(struct ceph_osd_client *osdc, 16843c663bbdSAlex Elder struct ceph_msg *msg) 1685a40c4f10SYehuda Sadeh { 1686a40c4f10SYehuda Sadeh void *p, *end; 1687a40c4f10SYehuda Sadeh u8 proto_ver; 1688a40c4f10SYehuda Sadeh u64 cookie, ver, notify_id; 1689a40c4f10SYehuda Sadeh u8 opcode; 1690a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 1691a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work; 1692a40c4f10SYehuda Sadeh 1693a40c4f10SYehuda Sadeh p = msg->front.iov_base; 1694a40c4f10SYehuda Sadeh end = p + msg->front.iov_len; 1695a40c4f10SYehuda Sadeh 1696a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, proto_ver, bad); 1697a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, opcode, bad); 1698a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, cookie, bad); 1699a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, ver, bad); 1700a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, notify_id, bad); 1701a40c4f10SYehuda Sadeh 1702a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 1703a40c4f10SYehuda Sadeh event = __find_event(osdc, cookie); 1704a40c4f10SYehuda Sadeh if (event) { 17053c663bbdSAlex Elder BUG_ON(event->one_shot); 1706a40c4f10SYehuda Sadeh get_event(event); 1707a40c4f10SYehuda Sadeh } 1708a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 1709a40c4f10SYehuda Sadeh dout("handle_watch_notify cookie %lld ver %lld event %p\n", 1710a40c4f10SYehuda Sadeh cookie, ver, event); 1711a40c4f10SYehuda Sadeh if (event) { 1712a40c4f10SYehuda Sadeh event_work = kmalloc(sizeof(*event_work), GFP_NOIO); 1713a40c4f10SYehuda Sadeh if (!event_work) { 1714a40c4f10SYehuda Sadeh dout("ERROR: could not allocate event_work\n"); 1715a40c4f10SYehuda Sadeh goto done_err; 1716a40c4f10SYehuda Sadeh } 17176b0ae409SMariusz Kozlowski INIT_WORK(&event_work->work, do_event_work); 1718a40c4f10SYehuda Sadeh event_work->event = event; 1719a40c4f10SYehuda Sadeh event_work->ver = ver; 1720a40c4f10SYehuda Sadeh event_work->notify_id = notify_id; 1721a40c4f10SYehuda Sadeh event_work->opcode = opcode; 1722a40c4f10SYehuda Sadeh if (!queue_work(osdc->notify_wq, &event_work->work)) { 1723a40c4f10SYehuda Sadeh dout("WARNING: failed to queue notify event work\n"); 1724a40c4f10SYehuda Sadeh goto done_err; 1725a40c4f10SYehuda Sadeh } 1726a40c4f10SYehuda Sadeh } 1727a40c4f10SYehuda Sadeh 1728a40c4f10SYehuda Sadeh return; 1729a40c4f10SYehuda Sadeh 1730a40c4f10SYehuda Sadeh done_err: 1731a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 1732a40c4f10SYehuda Sadeh return; 1733a40c4f10SYehuda Sadeh 1734a40c4f10SYehuda Sadeh bad: 1735a40c4f10SYehuda Sadeh pr_err("osdc handle_watch_notify corrupt msg\n"); 1736a40c4f10SYehuda Sadeh return; 1737a40c4f10SYehuda Sadeh } 1738a40c4f10SYehuda Sadeh 1739a40c4f10SYehuda Sadeh /* 17403d14c5d2SYehuda Sadeh * Register request, send initial attempt. 17413d14c5d2SYehuda Sadeh */ 17423d14c5d2SYehuda Sadeh int ceph_osdc_start_request(struct ceph_osd_client *osdc, 17433d14c5d2SYehuda Sadeh struct ceph_osd_request *req, 17443d14c5d2SYehuda Sadeh bool nofail) 17453d14c5d2SYehuda Sadeh { 17463d14c5d2SYehuda Sadeh int rc = 0; 17473d14c5d2SYehuda Sadeh 17483d14c5d2SYehuda Sadeh req->r_request->pages = req->r_pages; 17493d14c5d2SYehuda Sadeh req->r_request->nr_pages = req->r_num_pages; 17503d14c5d2SYehuda Sadeh #ifdef CONFIG_BLOCK 17513d14c5d2SYehuda Sadeh req->r_request->bio = req->r_bio; 17523d14c5d2SYehuda Sadeh #endif 1753c885837fSAlex Elder req->r_request->trail = &req->r_trail; 17543d14c5d2SYehuda Sadeh 17553d14c5d2SYehuda Sadeh register_request(osdc, req); 17563d14c5d2SYehuda Sadeh 17573d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 17583d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 17593d14c5d2SYehuda Sadeh /* 17603d14c5d2SYehuda Sadeh * a racing kick_requests() may have sent the message for us 17613d14c5d2SYehuda Sadeh * while we dropped request_mutex above, so only send now if 17623d14c5d2SYehuda Sadeh * the request still han't been touched yet. 17633d14c5d2SYehuda Sadeh */ 17643d14c5d2SYehuda Sadeh if (req->r_sent == 0) { 176538d6453cSSage Weil rc = __map_request(osdc, req, 0); 17669d6fcb08SSage Weil if (rc < 0) { 17679d6fcb08SSage Weil if (nofail) { 17689d6fcb08SSage Weil dout("osdc_start_request failed map, " 17699d6fcb08SSage Weil " will retry %lld\n", req->r_tid); 17709d6fcb08SSage Weil rc = 0; 17719d6fcb08SSage Weil } 1772234af26fSDan Carpenter goto out_unlock; 17739d6fcb08SSage Weil } 17746f6c7006SSage Weil if (req->r_osd == NULL) { 17756f6c7006SSage Weil dout("send_request %p no up osds in pg\n", req); 17766f6c7006SSage Weil ceph_monc_request_next_osdmap(&osdc->client->monc); 17776f6c7006SSage Weil } else { 177856e925b6SSage Weil __send_request(osdc, req); 177956e925b6SSage Weil } 17803d14c5d2SYehuda Sadeh rc = 0; 17816f6c7006SSage Weil } 1782234af26fSDan Carpenter 1783234af26fSDan Carpenter out_unlock: 17843d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 17853d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 17863d14c5d2SYehuda Sadeh return rc; 17873d14c5d2SYehuda Sadeh } 17883d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_start_request); 17893d14c5d2SYehuda Sadeh 17903d14c5d2SYehuda Sadeh /* 17913d14c5d2SYehuda Sadeh * wait for a request to complete 17923d14c5d2SYehuda Sadeh */ 17933d14c5d2SYehuda Sadeh int ceph_osdc_wait_request(struct ceph_osd_client *osdc, 17943d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 17953d14c5d2SYehuda Sadeh { 17963d14c5d2SYehuda Sadeh int rc; 17973d14c5d2SYehuda Sadeh 17983d14c5d2SYehuda Sadeh rc = wait_for_completion_interruptible(&req->r_completion); 17993d14c5d2SYehuda Sadeh if (rc < 0) { 18003d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 18013d14c5d2SYehuda Sadeh __cancel_request(req); 18023d14c5d2SYehuda Sadeh __unregister_request(osdc, req); 18033d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 180425845472SSage Weil complete_request(req); 18053d14c5d2SYehuda Sadeh dout("wait_request tid %llu canceled/timed out\n", req->r_tid); 18063d14c5d2SYehuda Sadeh return rc; 18073d14c5d2SYehuda Sadeh } 18083d14c5d2SYehuda Sadeh 18093d14c5d2SYehuda Sadeh dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result); 18103d14c5d2SYehuda Sadeh return req->r_result; 18113d14c5d2SYehuda Sadeh } 18123d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_wait_request); 18133d14c5d2SYehuda Sadeh 18143d14c5d2SYehuda Sadeh /* 18153d14c5d2SYehuda Sadeh * sync - wait for all in-flight requests to flush. avoid starvation. 18163d14c5d2SYehuda Sadeh */ 18173d14c5d2SYehuda Sadeh void ceph_osdc_sync(struct ceph_osd_client *osdc) 18183d14c5d2SYehuda Sadeh { 18193d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 18203d14c5d2SYehuda Sadeh u64 last_tid, next_tid = 0; 18213d14c5d2SYehuda Sadeh 18223d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 18233d14c5d2SYehuda Sadeh last_tid = osdc->last_tid; 18243d14c5d2SYehuda Sadeh while (1) { 18253d14c5d2SYehuda Sadeh req = __lookup_request_ge(osdc, next_tid); 18263d14c5d2SYehuda Sadeh if (!req) 18273d14c5d2SYehuda Sadeh break; 18283d14c5d2SYehuda Sadeh if (req->r_tid > last_tid) 18293d14c5d2SYehuda Sadeh break; 18303d14c5d2SYehuda Sadeh 18313d14c5d2SYehuda Sadeh next_tid = req->r_tid + 1; 18323d14c5d2SYehuda Sadeh if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0) 18333d14c5d2SYehuda Sadeh continue; 18343d14c5d2SYehuda Sadeh 18353d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 18363d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 18373d14c5d2SYehuda Sadeh dout("sync waiting on tid %llu (last is %llu)\n", 18383d14c5d2SYehuda Sadeh req->r_tid, last_tid); 18393d14c5d2SYehuda Sadeh wait_for_completion(&req->r_safe_completion); 18403d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 18413d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 18423d14c5d2SYehuda Sadeh } 18433d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 18443d14c5d2SYehuda Sadeh dout("sync done (thru tid %llu)\n", last_tid); 18453d14c5d2SYehuda Sadeh } 18463d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_sync); 18473d14c5d2SYehuda Sadeh 18483d14c5d2SYehuda Sadeh /* 18493d14c5d2SYehuda Sadeh * init, shutdown 18503d14c5d2SYehuda Sadeh */ 18513d14c5d2SYehuda Sadeh int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client) 18523d14c5d2SYehuda Sadeh { 18533d14c5d2SYehuda Sadeh int err; 18543d14c5d2SYehuda Sadeh 18553d14c5d2SYehuda Sadeh dout("init\n"); 18563d14c5d2SYehuda Sadeh osdc->client = client; 18573d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 18583d14c5d2SYehuda Sadeh init_rwsem(&osdc->map_sem); 18593d14c5d2SYehuda Sadeh init_completion(&osdc->map_waiters); 18603d14c5d2SYehuda Sadeh osdc->last_requested_map = 0; 18613d14c5d2SYehuda Sadeh mutex_init(&osdc->request_mutex); 18623d14c5d2SYehuda Sadeh osdc->last_tid = 0; 18633d14c5d2SYehuda Sadeh osdc->osds = RB_ROOT; 18643d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->osd_lru); 18653d14c5d2SYehuda Sadeh osdc->requests = RB_ROOT; 18663d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_lru); 18676f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_unsent); 18686f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_notarget); 1869a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_linger); 18703d14c5d2SYehuda Sadeh osdc->num_requests = 0; 18713d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout); 18723d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout); 1873a40c4f10SYehuda Sadeh spin_lock_init(&osdc->event_lock); 1874a40c4f10SYehuda Sadeh osdc->event_tree = RB_ROOT; 1875a40c4f10SYehuda Sadeh osdc->event_count = 0; 18763d14c5d2SYehuda Sadeh 18773d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 18783d14c5d2SYehuda Sadeh round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ)); 18793d14c5d2SYehuda Sadeh 18803d14c5d2SYehuda Sadeh err = -ENOMEM; 18813d14c5d2SYehuda Sadeh osdc->req_mempool = mempool_create_kmalloc_pool(10, 18823d14c5d2SYehuda Sadeh sizeof(struct ceph_osd_request)); 18833d14c5d2SYehuda Sadeh if (!osdc->req_mempool) 18843d14c5d2SYehuda Sadeh goto out; 18853d14c5d2SYehuda Sadeh 1886d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP, 1887d50b409fSSage Weil OSD_OP_FRONT_LEN, 10, true, 18883d14c5d2SYehuda Sadeh "osd_op"); 18893d14c5d2SYehuda Sadeh if (err < 0) 18903d14c5d2SYehuda Sadeh goto out_mempool; 1891d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY, 18923d14c5d2SYehuda Sadeh OSD_OPREPLY_FRONT_LEN, 10, true, 18933d14c5d2SYehuda Sadeh "osd_op_reply"); 18943d14c5d2SYehuda Sadeh if (err < 0) 18953d14c5d2SYehuda Sadeh goto out_msgpool; 1896a40c4f10SYehuda Sadeh 1897a40c4f10SYehuda Sadeh osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify"); 1898a40c4f10SYehuda Sadeh if (IS_ERR(osdc->notify_wq)) { 1899a40c4f10SYehuda Sadeh err = PTR_ERR(osdc->notify_wq); 1900a40c4f10SYehuda Sadeh osdc->notify_wq = NULL; 1901a40c4f10SYehuda Sadeh goto out_msgpool; 1902a40c4f10SYehuda Sadeh } 19033d14c5d2SYehuda Sadeh return 0; 19043d14c5d2SYehuda Sadeh 19053d14c5d2SYehuda Sadeh out_msgpool: 19063d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 19073d14c5d2SYehuda Sadeh out_mempool: 19083d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 19093d14c5d2SYehuda Sadeh out: 19103d14c5d2SYehuda Sadeh return err; 19113d14c5d2SYehuda Sadeh } 19123d14c5d2SYehuda Sadeh 19133d14c5d2SYehuda Sadeh void ceph_osdc_stop(struct ceph_osd_client *osdc) 19143d14c5d2SYehuda Sadeh { 1915a40c4f10SYehuda Sadeh flush_workqueue(osdc->notify_wq); 1916a40c4f10SYehuda Sadeh destroy_workqueue(osdc->notify_wq); 19173d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->timeout_work); 19183d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->osds_timeout_work); 19193d14c5d2SYehuda Sadeh if (osdc->osdmap) { 19203d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 19213d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 19223d14c5d2SYehuda Sadeh } 1923aca420bcSSage Weil remove_all_osds(osdc); 19243d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 19253d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 19263d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op_reply); 19273d14c5d2SYehuda Sadeh } 19283d14c5d2SYehuda Sadeh 19293d14c5d2SYehuda Sadeh /* 19303d14c5d2SYehuda Sadeh * Read some contiguous pages. If we cross a stripe boundary, shorten 19313d14c5d2SYehuda Sadeh * *plen. Return number of bytes read, or error. 19323d14c5d2SYehuda Sadeh */ 19333d14c5d2SYehuda Sadeh int ceph_osdc_readpages(struct ceph_osd_client *osdc, 19343d14c5d2SYehuda Sadeh struct ceph_vino vino, struct ceph_file_layout *layout, 19353d14c5d2SYehuda Sadeh u64 off, u64 *plen, 19363d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 1937b7495fc2SSage Weil struct page **pages, int num_pages, int page_align) 19383d14c5d2SYehuda Sadeh { 19393d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 19403d14c5d2SYehuda Sadeh int rc = 0; 19413d14c5d2SYehuda Sadeh 19423d14c5d2SYehuda Sadeh dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino, 19433d14c5d2SYehuda Sadeh vino.snap, off, *plen); 19443d14c5d2SYehuda Sadeh req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 19453d14c5d2SYehuda Sadeh CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ, 19463d14c5d2SYehuda Sadeh NULL, 0, truncate_seq, truncate_size, NULL, 1947a3bea47eSAlex Elder false, page_align); 19486816282dSSage Weil if (IS_ERR(req)) 19496816282dSSage Weil return PTR_ERR(req); 19503d14c5d2SYehuda Sadeh 19513d14c5d2SYehuda Sadeh /* it may be a short read due to an object boundary */ 19523d14c5d2SYehuda Sadeh req->r_pages = pages; 19533d14c5d2SYehuda Sadeh 1954b7495fc2SSage Weil dout("readpages final extent is %llu~%llu (%d pages align %d)\n", 1955b7495fc2SSage Weil off, *plen, req->r_num_pages, page_align); 19563d14c5d2SYehuda Sadeh 19573d14c5d2SYehuda Sadeh rc = ceph_osdc_start_request(osdc, req, false); 19583d14c5d2SYehuda Sadeh if (!rc) 19593d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 19603d14c5d2SYehuda Sadeh 19613d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 19623d14c5d2SYehuda Sadeh dout("readpages result %d\n", rc); 19633d14c5d2SYehuda Sadeh return rc; 19643d14c5d2SYehuda Sadeh } 19653d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_readpages); 19663d14c5d2SYehuda Sadeh 19673d14c5d2SYehuda Sadeh /* 19683d14c5d2SYehuda Sadeh * do a synchronous write on N pages 19693d14c5d2SYehuda Sadeh */ 19703d14c5d2SYehuda Sadeh int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino, 19713d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 19723d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 19733d14c5d2SYehuda Sadeh u64 off, u64 len, 19743d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 19753d14c5d2SYehuda Sadeh struct timespec *mtime, 197624808826SAlex Elder struct page **pages, int num_pages) 19773d14c5d2SYehuda Sadeh { 19783d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 19793d14c5d2SYehuda Sadeh int rc = 0; 1980b7495fc2SSage Weil int page_align = off & ~PAGE_MASK; 19813d14c5d2SYehuda Sadeh 19823d14c5d2SYehuda Sadeh BUG_ON(vino.snap != CEPH_NOSNAP); 19833d14c5d2SYehuda Sadeh req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 19843d14c5d2SYehuda Sadeh CEPH_OSD_OP_WRITE, 198524808826SAlex Elder CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE, 1986fbf8685fSAlex Elder snapc, 0, 19873d14c5d2SYehuda Sadeh truncate_seq, truncate_size, mtime, 1988a3bea47eSAlex Elder true, page_align); 19896816282dSSage Weil if (IS_ERR(req)) 19906816282dSSage Weil return PTR_ERR(req); 19913d14c5d2SYehuda Sadeh 19923d14c5d2SYehuda Sadeh /* it may be a short write due to an object boundary */ 19933d14c5d2SYehuda Sadeh req->r_pages = pages; 19943d14c5d2SYehuda Sadeh dout("writepages %llu~%llu (%d pages)\n", off, len, 19953d14c5d2SYehuda Sadeh req->r_num_pages); 19963d14c5d2SYehuda Sadeh 199787f979d3SAlex Elder rc = ceph_osdc_start_request(osdc, req, true); 19983d14c5d2SYehuda Sadeh if (!rc) 19993d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 20003d14c5d2SYehuda Sadeh 20013d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 20023d14c5d2SYehuda Sadeh if (rc == 0) 20033d14c5d2SYehuda Sadeh rc = len; 20043d14c5d2SYehuda Sadeh dout("writepages result %d\n", rc); 20053d14c5d2SYehuda Sadeh return rc; 20063d14c5d2SYehuda Sadeh } 20073d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_writepages); 20083d14c5d2SYehuda Sadeh 20093d14c5d2SYehuda Sadeh /* 20103d14c5d2SYehuda Sadeh * handle incoming message 20113d14c5d2SYehuda Sadeh */ 20123d14c5d2SYehuda Sadeh static void dispatch(struct ceph_connection *con, struct ceph_msg *msg) 20133d14c5d2SYehuda Sadeh { 20143d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 20153d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 20163d14c5d2SYehuda Sadeh int type = le16_to_cpu(msg->hdr.type); 20173d14c5d2SYehuda Sadeh 20183d14c5d2SYehuda Sadeh if (!osd) 20193d14c5d2SYehuda Sadeh goto out; 20203d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 20213d14c5d2SYehuda Sadeh 20223d14c5d2SYehuda Sadeh switch (type) { 20233d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 20243d14c5d2SYehuda Sadeh ceph_osdc_handle_map(osdc, msg); 20253d14c5d2SYehuda Sadeh break; 20263d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 20273d14c5d2SYehuda Sadeh handle_reply(osdc, msg, con); 20283d14c5d2SYehuda Sadeh break; 2029a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 2030a40c4f10SYehuda Sadeh handle_watch_notify(osdc, msg); 2031a40c4f10SYehuda Sadeh break; 20323d14c5d2SYehuda Sadeh 20333d14c5d2SYehuda Sadeh default: 20343d14c5d2SYehuda Sadeh pr_err("received unknown message type %d %s\n", type, 20353d14c5d2SYehuda Sadeh ceph_msg_type_name(type)); 20363d14c5d2SYehuda Sadeh } 20373d14c5d2SYehuda Sadeh out: 20383d14c5d2SYehuda Sadeh ceph_msg_put(msg); 20393d14c5d2SYehuda Sadeh } 20403d14c5d2SYehuda Sadeh 20413d14c5d2SYehuda Sadeh /* 20423d14c5d2SYehuda Sadeh * lookup and return message for incoming reply. set up reply message 20433d14c5d2SYehuda Sadeh * pages. 20443d14c5d2SYehuda Sadeh */ 20453d14c5d2SYehuda Sadeh static struct ceph_msg *get_reply(struct ceph_connection *con, 20463d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 20473d14c5d2SYehuda Sadeh int *skip) 20483d14c5d2SYehuda Sadeh { 20493d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 20503d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = osd->o_osdc; 20513d14c5d2SYehuda Sadeh struct ceph_msg *m; 20523d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 20533d14c5d2SYehuda Sadeh int front = le32_to_cpu(hdr->front_len); 20543d14c5d2SYehuda Sadeh int data_len = le32_to_cpu(hdr->data_len); 20553d14c5d2SYehuda Sadeh u64 tid; 20563d14c5d2SYehuda Sadeh 20573d14c5d2SYehuda Sadeh tid = le64_to_cpu(hdr->tid); 20583d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 20593d14c5d2SYehuda Sadeh req = __lookup_request(osdc, tid); 20603d14c5d2SYehuda Sadeh if (!req) { 20613d14c5d2SYehuda Sadeh *skip = 1; 20623d14c5d2SYehuda Sadeh m = NULL; 2063756a16a5SSage Weil dout("get_reply unknown tid %llu from osd%d\n", tid, 20643d14c5d2SYehuda Sadeh osd->o_osd); 20653d14c5d2SYehuda Sadeh goto out; 20663d14c5d2SYehuda Sadeh } 20673d14c5d2SYehuda Sadeh 20683d14c5d2SYehuda Sadeh if (req->r_con_filling_msg) { 20698921d114SAlex Elder dout("%s revoking msg %p from old con %p\n", __func__, 20703d14c5d2SYehuda Sadeh req->r_reply, req->r_con_filling_msg); 20718921d114SAlex Elder ceph_msg_revoke_incoming(req->r_reply); 20720d47766fSSage Weil req->r_con_filling_msg->ops->put(req->r_con_filling_msg); 20733d14c5d2SYehuda Sadeh req->r_con_filling_msg = NULL; 20743d14c5d2SYehuda Sadeh } 20753d14c5d2SYehuda Sadeh 20763d14c5d2SYehuda Sadeh if (front > req->r_reply->front.iov_len) { 20773d14c5d2SYehuda Sadeh pr_warning("get_reply front %d > preallocated %d\n", 20783d14c5d2SYehuda Sadeh front, (int)req->r_reply->front.iov_len); 2079b61c2763SSage Weil m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS, false); 20803d14c5d2SYehuda Sadeh if (!m) 20813d14c5d2SYehuda Sadeh goto out; 20823d14c5d2SYehuda Sadeh ceph_msg_put(req->r_reply); 20833d14c5d2SYehuda Sadeh req->r_reply = m; 20843d14c5d2SYehuda Sadeh } 20853d14c5d2SYehuda Sadeh m = ceph_msg_get(req->r_reply); 20863d14c5d2SYehuda Sadeh 20873d14c5d2SYehuda Sadeh if (data_len > 0) { 2088b7495fc2SSage Weil int want = calc_pages_for(req->r_page_alignment, data_len); 20893d14c5d2SYehuda Sadeh 20909cbb1d72SAlex Elder if (req->r_pages && unlikely(req->r_num_pages < want)) { 20919bb0ce2bSSage Weil pr_warning("tid %lld reply has %d bytes %d pages, we" 20929bb0ce2bSSage Weil " had only %d pages ready\n", tid, data_len, 20939bb0ce2bSSage Weil want, req->r_num_pages); 20943d14c5d2SYehuda Sadeh *skip = 1; 20953d14c5d2SYehuda Sadeh ceph_msg_put(m); 20963d14c5d2SYehuda Sadeh m = NULL; 20973d14c5d2SYehuda Sadeh goto out; 20983d14c5d2SYehuda Sadeh } 20993d14c5d2SYehuda Sadeh m->pages = req->r_pages; 21003d14c5d2SYehuda Sadeh m->nr_pages = req->r_num_pages; 2101c5c6b19dSSage Weil m->page_alignment = req->r_page_alignment; 21023d14c5d2SYehuda Sadeh #ifdef CONFIG_BLOCK 21033d14c5d2SYehuda Sadeh m->bio = req->r_bio; 21043d14c5d2SYehuda Sadeh #endif 21053d14c5d2SYehuda Sadeh } 21063d14c5d2SYehuda Sadeh *skip = 0; 21070d47766fSSage Weil req->r_con_filling_msg = con->ops->get(con); 21083d14c5d2SYehuda Sadeh dout("get_reply tid %lld %p\n", tid, m); 21093d14c5d2SYehuda Sadeh 21103d14c5d2SYehuda Sadeh out: 21113d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 21123d14c5d2SYehuda Sadeh return m; 21133d14c5d2SYehuda Sadeh 21143d14c5d2SYehuda Sadeh } 21153d14c5d2SYehuda Sadeh 21163d14c5d2SYehuda Sadeh static struct ceph_msg *alloc_msg(struct ceph_connection *con, 21173d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 21183d14c5d2SYehuda Sadeh int *skip) 21193d14c5d2SYehuda Sadeh { 21203d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 21213d14c5d2SYehuda Sadeh int type = le16_to_cpu(hdr->type); 21223d14c5d2SYehuda Sadeh int front = le32_to_cpu(hdr->front_len); 21233d14c5d2SYehuda Sadeh 21241c20f2d2SAlex Elder *skip = 0; 21253d14c5d2SYehuda Sadeh switch (type) { 21263d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 2127a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 2128b61c2763SSage Weil return ceph_msg_new(type, front, GFP_NOFS, false); 21293d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 21303d14c5d2SYehuda Sadeh return get_reply(con, hdr, skip); 21313d14c5d2SYehuda Sadeh default: 21323d14c5d2SYehuda Sadeh pr_info("alloc_msg unexpected msg type %d from osd%d\n", type, 21333d14c5d2SYehuda Sadeh osd->o_osd); 21343d14c5d2SYehuda Sadeh *skip = 1; 21353d14c5d2SYehuda Sadeh return NULL; 21363d14c5d2SYehuda Sadeh } 21373d14c5d2SYehuda Sadeh } 21383d14c5d2SYehuda Sadeh 21393d14c5d2SYehuda Sadeh /* 21403d14c5d2SYehuda Sadeh * Wrappers to refcount containing ceph_osd struct 21413d14c5d2SYehuda Sadeh */ 21423d14c5d2SYehuda Sadeh static struct ceph_connection *get_osd_con(struct ceph_connection *con) 21433d14c5d2SYehuda Sadeh { 21443d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 21453d14c5d2SYehuda Sadeh if (get_osd(osd)) 21463d14c5d2SYehuda Sadeh return con; 21473d14c5d2SYehuda Sadeh return NULL; 21483d14c5d2SYehuda Sadeh } 21493d14c5d2SYehuda Sadeh 21503d14c5d2SYehuda Sadeh static void put_osd_con(struct ceph_connection *con) 21513d14c5d2SYehuda Sadeh { 21523d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 21533d14c5d2SYehuda Sadeh put_osd(osd); 21543d14c5d2SYehuda Sadeh } 21553d14c5d2SYehuda Sadeh 21563d14c5d2SYehuda Sadeh /* 21573d14c5d2SYehuda Sadeh * authentication 21583d14c5d2SYehuda Sadeh */ 2159a3530df3SAlex Elder /* 2160a3530df3SAlex Elder * Note: returned pointer is the address of a structure that's 2161a3530df3SAlex Elder * managed separately. Caller must *not* attempt to free it. 2162a3530df3SAlex Elder */ 2163a3530df3SAlex Elder static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con, 21648f43fb53SAlex Elder int *proto, int force_new) 21653d14c5d2SYehuda Sadeh { 21663d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 21673d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 21683d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 216974f1869fSAlex Elder struct ceph_auth_handshake *auth = &o->o_auth; 21703d14c5d2SYehuda Sadeh 217174f1869fSAlex Elder if (force_new && auth->authorizer) { 2172a255651dSAlex Elder if (ac->ops && ac->ops->destroy_authorizer) 217374f1869fSAlex Elder ac->ops->destroy_authorizer(ac, auth->authorizer); 217474f1869fSAlex Elder auth->authorizer = NULL; 21753d14c5d2SYehuda Sadeh } 2176a255651dSAlex Elder if (!auth->authorizer && ac->ops && ac->ops->create_authorizer) { 2177a3530df3SAlex Elder int ret = ac->ops->create_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 2178a3530df3SAlex Elder auth); 21793d14c5d2SYehuda Sadeh if (ret) 2180a3530df3SAlex Elder return ERR_PTR(ret); 21813d14c5d2SYehuda Sadeh } 21823d14c5d2SYehuda Sadeh *proto = ac->protocol; 218374f1869fSAlex Elder 2184a3530df3SAlex Elder return auth; 21853d14c5d2SYehuda Sadeh } 21863d14c5d2SYehuda Sadeh 21873d14c5d2SYehuda Sadeh 21883d14c5d2SYehuda Sadeh static int verify_authorizer_reply(struct ceph_connection *con, int len) 21893d14c5d2SYehuda Sadeh { 21903d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 21913d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 21923d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 21933d14c5d2SYehuda Sadeh 2194a255651dSAlex Elder /* 2195a255651dSAlex Elder * XXX If ac->ops or ac->ops->verify_authorizer_reply is null, 2196a255651dSAlex Elder * XXX which do we do: succeed or fail? 2197a255651dSAlex Elder */ 21986c4a1915SAlex Elder return ac->ops->verify_authorizer_reply(ac, o->o_auth.authorizer, len); 21993d14c5d2SYehuda Sadeh } 22003d14c5d2SYehuda Sadeh 22013d14c5d2SYehuda Sadeh static int invalidate_authorizer(struct ceph_connection *con) 22023d14c5d2SYehuda Sadeh { 22033d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 22043d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 22053d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 22063d14c5d2SYehuda Sadeh 2207a255651dSAlex Elder if (ac->ops && ac->ops->invalidate_authorizer) 22083d14c5d2SYehuda Sadeh ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD); 22093d14c5d2SYehuda Sadeh 22103d14c5d2SYehuda Sadeh return ceph_monc_validate_auth(&osdc->client->monc); 22113d14c5d2SYehuda Sadeh } 22123d14c5d2SYehuda Sadeh 22133d14c5d2SYehuda Sadeh static const struct ceph_connection_operations osd_con_ops = { 22143d14c5d2SYehuda Sadeh .get = get_osd_con, 22153d14c5d2SYehuda Sadeh .put = put_osd_con, 22163d14c5d2SYehuda Sadeh .dispatch = dispatch, 22173d14c5d2SYehuda Sadeh .get_authorizer = get_authorizer, 22183d14c5d2SYehuda Sadeh .verify_authorizer_reply = verify_authorizer_reply, 22193d14c5d2SYehuda Sadeh .invalidate_authorizer = invalidate_authorizer, 22203d14c5d2SYehuda Sadeh .alloc_msg = alloc_msg, 22213d14c5d2SYehuda Sadeh .fault = osd_reset, 22223d14c5d2SYehuda Sadeh }; 2223