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 */ 66e75b45cfSAlex Elder static int calc_layout(struct ceph_vino vino, 673d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 683d14c5d2SYehuda Sadeh u64 off, u64 *plen, 693d14c5d2SYehuda Sadeh struct ceph_osd_request *req, 703d14c5d2SYehuda Sadeh struct ceph_osd_req_op *op) 713d14c5d2SYehuda Sadeh { 7260e56f13SAlex Elder u64 orig_len = *plen; 7360e56f13SAlex Elder u64 bno = 0; 7460e56f13SAlex Elder u64 objoff = 0; 7560e56f13SAlex Elder u64 objlen = 0; 76d63b77f4SSage Weil int r; 773d14c5d2SYehuda Sadeh 7860e56f13SAlex Elder /* object extent? */ 7960e56f13SAlex Elder r = ceph_calc_file_object_mapping(layout, off, orig_len, &bno, 8060e56f13SAlex Elder &objoff, &objlen); 81d63b77f4SSage Weil if (r < 0) 82d63b77f4SSage Weil return r; 8360e56f13SAlex Elder if (objlen < orig_len) { 8460e56f13SAlex Elder *plen = objlen; 8560e56f13SAlex Elder dout(" skipping last %llu, final file extent %llu~%llu\n", 8660e56f13SAlex Elder orig_len - *plen, off, *plen); 8760e56f13SAlex Elder } 8860e56f13SAlex Elder 8960e56f13SAlex Elder if (op_has_extent(op->op)) { 9060e56f13SAlex Elder u32 osize = le32_to_cpu(layout->fl_object_size); 9160e56f13SAlex Elder op->extent.offset = objoff; 9260e56f13SAlex Elder op->extent.length = objlen; 9360e56f13SAlex Elder if (op->extent.truncate_size <= off - objoff) { 9460e56f13SAlex Elder op->extent.truncate_size = 0; 9560e56f13SAlex Elder } else { 9660e56f13SAlex Elder op->extent.truncate_size -= off - objoff; 9760e56f13SAlex Elder if (op->extent.truncate_size > osize) 9860e56f13SAlex Elder op->extent.truncate_size = osize; 9960e56f13SAlex Elder } 10060e56f13SAlex Elder } 10160e56f13SAlex Elder req->r_num_pages = calc_pages_for(off, *plen); 10260e56f13SAlex Elder req->r_page_alignment = off & ~PAGE_MASK; 10360e56f13SAlex Elder if (op->op == CEPH_OSD_OP_WRITE) 10460e56f13SAlex Elder op->payload_len = *plen; 10560e56f13SAlex Elder 10660e56f13SAlex Elder dout("calc_layout bno=%llx %llu~%llu (%d pages)\n", 10760e56f13SAlex Elder bno, objoff, objlen, req->r_num_pages); 1083d14c5d2SYehuda Sadeh 1092dab036bSSage Weil snprintf(req->r_oid, sizeof(req->r_oid), "%llx.%08llx", vino.ino, bno); 1103d14c5d2SYehuda Sadeh req->r_oid_len = strlen(req->r_oid); 111d63b77f4SSage Weil 112d63b77f4SSage Weil return r; 1133d14c5d2SYehuda Sadeh } 1143d14c5d2SYehuda Sadeh 1153d14c5d2SYehuda Sadeh /* 1163d14c5d2SYehuda Sadeh * requests 1173d14c5d2SYehuda Sadeh */ 1183d14c5d2SYehuda Sadeh void ceph_osdc_release_request(struct kref *kref) 1193d14c5d2SYehuda Sadeh { 1203d14c5d2SYehuda Sadeh struct ceph_osd_request *req = container_of(kref, 1213d14c5d2SYehuda Sadeh struct ceph_osd_request, 1223d14c5d2SYehuda Sadeh r_kref); 1233d14c5d2SYehuda Sadeh 1243d14c5d2SYehuda Sadeh if (req->r_request) 1253d14c5d2SYehuda Sadeh ceph_msg_put(req->r_request); 1263d14c5d2SYehuda Sadeh if (req->r_con_filling_msg) { 1279cbb1d72SAlex Elder dout("%s revoking msg %p from con %p\n", __func__, 1289cbb1d72SAlex Elder req->r_reply, req->r_con_filling_msg); 1298921d114SAlex Elder ceph_msg_revoke_incoming(req->r_reply); 1300d47766fSSage Weil req->r_con_filling_msg->ops->put(req->r_con_filling_msg); 1319cbb1d72SAlex Elder req->r_con_filling_msg = NULL; 1323d14c5d2SYehuda Sadeh } 133ab8cb34aSAlex Elder if (req->r_reply) 134ab8cb34aSAlex Elder ceph_msg_put(req->r_reply); 1353d14c5d2SYehuda Sadeh if (req->r_own_pages) 1363d14c5d2SYehuda Sadeh ceph_release_page_vector(req->r_pages, 1373d14c5d2SYehuda Sadeh req->r_num_pages); 1383d14c5d2SYehuda Sadeh ceph_put_snap_context(req->r_snapc); 139c885837fSAlex Elder ceph_pagelist_release(&req->r_trail); 1403d14c5d2SYehuda Sadeh if (req->r_mempool) 1413d14c5d2SYehuda Sadeh mempool_free(req, req->r_osdc->req_mempool); 1423d14c5d2SYehuda Sadeh else 1433d14c5d2SYehuda Sadeh kfree(req); 1443d14c5d2SYehuda Sadeh } 1453d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_release_request); 1463d14c5d2SYehuda Sadeh 1473d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc, 1483d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 149ae7ca4a3SAlex Elder unsigned int num_op, 1503d14c5d2SYehuda Sadeh bool use_mempool, 15154a54007SAlex Elder gfp_t gfp_flags) 1523d14c5d2SYehuda Sadeh { 1533d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 1543d14c5d2SYehuda Sadeh struct ceph_msg *msg; 1553d14c5d2SYehuda Sadeh size_t msg_size = sizeof(struct ceph_osd_request_head); 1563d14c5d2SYehuda Sadeh 1573d14c5d2SYehuda Sadeh msg_size += num_op*sizeof(struct ceph_osd_op); 1583d14c5d2SYehuda Sadeh 1593d14c5d2SYehuda Sadeh if (use_mempool) { 1603d14c5d2SYehuda Sadeh req = mempool_alloc(osdc->req_mempool, gfp_flags); 1613d14c5d2SYehuda Sadeh memset(req, 0, sizeof(*req)); 1623d14c5d2SYehuda Sadeh } else { 1633d14c5d2SYehuda Sadeh req = kzalloc(sizeof(*req), gfp_flags); 1643d14c5d2SYehuda Sadeh } 1653d14c5d2SYehuda Sadeh if (req == NULL) 1663d14c5d2SYehuda Sadeh return NULL; 1673d14c5d2SYehuda Sadeh 1683d14c5d2SYehuda Sadeh req->r_osdc = osdc; 1693d14c5d2SYehuda Sadeh req->r_mempool = use_mempool; 1703d14c5d2SYehuda Sadeh 1713d14c5d2SYehuda Sadeh kref_init(&req->r_kref); 1723d14c5d2SYehuda Sadeh init_completion(&req->r_completion); 1733d14c5d2SYehuda Sadeh init_completion(&req->r_safe_completion); 174a978fa20SAlex Elder RB_CLEAR_NODE(&req->r_node); 1753d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&req->r_unsafe_item); 176a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&req->r_linger_item); 177a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&req->r_linger_osd); 178935b639aSSage Weil INIT_LIST_HEAD(&req->r_req_lru_item); 179cd43045cSSage Weil INIT_LIST_HEAD(&req->r_osd_item); 180cd43045cSSage Weil 1813d14c5d2SYehuda Sadeh /* create reply message */ 1823d14c5d2SYehuda Sadeh if (use_mempool) 1833d14c5d2SYehuda Sadeh msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0); 1843d14c5d2SYehuda Sadeh else 1853d14c5d2SYehuda Sadeh msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, 186b61c2763SSage Weil OSD_OPREPLY_FRONT_LEN, gfp_flags, true); 1873d14c5d2SYehuda Sadeh if (!msg) { 1883d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 1893d14c5d2SYehuda Sadeh return NULL; 1903d14c5d2SYehuda Sadeh } 1913d14c5d2SYehuda Sadeh req->r_reply = msg; 1923d14c5d2SYehuda Sadeh 193c885837fSAlex Elder ceph_pagelist_init(&req->r_trail); 194d50b409fSSage Weil 1953d14c5d2SYehuda Sadeh /* create request message; allow space for oid */ 196224736d9SStratos Psomadakis msg_size += MAX_OBJ_NAME_SIZE; 1973d14c5d2SYehuda Sadeh if (snapc) 1983d14c5d2SYehuda Sadeh msg_size += sizeof(u64) * snapc->num_snaps; 1993d14c5d2SYehuda Sadeh if (use_mempool) 2003d14c5d2SYehuda Sadeh msg = ceph_msgpool_get(&osdc->msgpool_op, 0); 2013d14c5d2SYehuda Sadeh else 202b61c2763SSage Weil msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true); 2033d14c5d2SYehuda Sadeh if (!msg) { 2043d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 2053d14c5d2SYehuda Sadeh return NULL; 2063d14c5d2SYehuda Sadeh } 2073d14c5d2SYehuda Sadeh 2083d14c5d2SYehuda Sadeh memset(msg->front.iov_base, 0, msg->front.iov_len); 2093d14c5d2SYehuda Sadeh 2103d14c5d2SYehuda Sadeh req->r_request = msg; 2113d14c5d2SYehuda Sadeh 2123d14c5d2SYehuda Sadeh return req; 2133d14c5d2SYehuda Sadeh } 2143d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_alloc_request); 2153d14c5d2SYehuda Sadeh 2163d14c5d2SYehuda Sadeh static void osd_req_encode_op(struct ceph_osd_request *req, 2173d14c5d2SYehuda Sadeh struct ceph_osd_op *dst, 2183d14c5d2SYehuda Sadeh struct ceph_osd_req_op *src) 2193d14c5d2SYehuda Sadeh { 2203d14c5d2SYehuda Sadeh dst->op = cpu_to_le16(src->op); 2213d14c5d2SYehuda Sadeh 222065a68f9SAlex Elder switch (src->op) { 2233d14c5d2SYehuda Sadeh case CEPH_OSD_OP_READ: 2243d14c5d2SYehuda Sadeh case CEPH_OSD_OP_WRITE: 2253d14c5d2SYehuda Sadeh dst->extent.offset = 2263d14c5d2SYehuda Sadeh cpu_to_le64(src->extent.offset); 2273d14c5d2SYehuda Sadeh dst->extent.length = 2283d14c5d2SYehuda Sadeh cpu_to_le64(src->extent.length); 2293d14c5d2SYehuda Sadeh dst->extent.truncate_size = 2303d14c5d2SYehuda Sadeh cpu_to_le64(src->extent.truncate_size); 2313d14c5d2SYehuda Sadeh dst->extent.truncate_seq = 2323d14c5d2SYehuda Sadeh cpu_to_le32(src->extent.truncate_seq); 2333d14c5d2SYehuda Sadeh break; 2343d14c5d2SYehuda Sadeh 2353d14c5d2SYehuda Sadeh case CEPH_OSD_OP_GETXATTR: 2363d14c5d2SYehuda Sadeh case CEPH_OSD_OP_SETXATTR: 2373d14c5d2SYehuda Sadeh case CEPH_OSD_OP_CMPXATTR: 2383d14c5d2SYehuda Sadeh dst->xattr.name_len = cpu_to_le32(src->xattr.name_len); 2393d14c5d2SYehuda Sadeh dst->xattr.value_len = cpu_to_le32(src->xattr.value_len); 2403d14c5d2SYehuda Sadeh dst->xattr.cmp_op = src->xattr.cmp_op; 2413d14c5d2SYehuda Sadeh dst->xattr.cmp_mode = src->xattr.cmp_mode; 242c885837fSAlex Elder ceph_pagelist_append(&req->r_trail, src->xattr.name, 2433d14c5d2SYehuda Sadeh src->xattr.name_len); 244c885837fSAlex Elder ceph_pagelist_append(&req->r_trail, src->xattr.val, 2453d14c5d2SYehuda Sadeh src->xattr.value_len); 2463d14c5d2SYehuda Sadeh break; 2473d14c5d2SYehuda Sadeh case CEPH_OSD_OP_CALL: 2483d14c5d2SYehuda Sadeh dst->cls.class_len = src->cls.class_len; 2493d14c5d2SYehuda Sadeh dst->cls.method_len = src->cls.method_len; 2503d14c5d2SYehuda Sadeh dst->cls.indata_len = cpu_to_le32(src->cls.indata_len); 2513d14c5d2SYehuda Sadeh 252c885837fSAlex Elder ceph_pagelist_append(&req->r_trail, src->cls.class_name, 2533d14c5d2SYehuda Sadeh src->cls.class_len); 254c885837fSAlex Elder ceph_pagelist_append(&req->r_trail, src->cls.method_name, 2553d14c5d2SYehuda Sadeh src->cls.method_len); 256c885837fSAlex Elder ceph_pagelist_append(&req->r_trail, src->cls.indata, 2573d14c5d2SYehuda Sadeh src->cls.indata_len); 2583d14c5d2SYehuda Sadeh break; 2593d14c5d2SYehuda Sadeh case CEPH_OSD_OP_ROLLBACK: 2603d14c5d2SYehuda Sadeh dst->snap.snapid = cpu_to_le64(src->snap.snapid); 2613d14c5d2SYehuda Sadeh break; 2623d14c5d2SYehuda Sadeh case CEPH_OSD_OP_STARTSYNC: 2633d14c5d2SYehuda Sadeh break; 264a40c4f10SYehuda Sadeh case CEPH_OSD_OP_NOTIFY: 265a40c4f10SYehuda Sadeh { 266a40c4f10SYehuda Sadeh __le32 prot_ver = cpu_to_le32(src->watch.prot_ver); 267a40c4f10SYehuda Sadeh __le32 timeout = cpu_to_le32(src->watch.timeout); 268a40c4f10SYehuda Sadeh 269c885837fSAlex Elder ceph_pagelist_append(&req->r_trail, 270a40c4f10SYehuda Sadeh &prot_ver, sizeof(prot_ver)); 271c885837fSAlex Elder ceph_pagelist_append(&req->r_trail, 272a40c4f10SYehuda Sadeh &timeout, sizeof(timeout)); 273a40c4f10SYehuda Sadeh } 274a40c4f10SYehuda Sadeh case CEPH_OSD_OP_NOTIFY_ACK: 275a40c4f10SYehuda Sadeh case CEPH_OSD_OP_WATCH: 276a40c4f10SYehuda Sadeh dst->watch.cookie = cpu_to_le64(src->watch.cookie); 277a40c4f10SYehuda Sadeh dst->watch.ver = cpu_to_le64(src->watch.ver); 278a40c4f10SYehuda Sadeh dst->watch.flag = src->watch.flag; 279a40c4f10SYehuda Sadeh break; 2803d14c5d2SYehuda Sadeh default: 2813d14c5d2SYehuda Sadeh pr_err("unrecognized osd opcode %d\n", dst->op); 2823d14c5d2SYehuda Sadeh WARN_ON(1); 2833d14c5d2SYehuda Sadeh break; 2843d14c5d2SYehuda Sadeh } 2853d14c5d2SYehuda Sadeh dst->payload_len = cpu_to_le32(src->payload_len); 2863d14c5d2SYehuda Sadeh } 2873d14c5d2SYehuda Sadeh 2883d14c5d2SYehuda Sadeh /* 2893d14c5d2SYehuda Sadeh * build new request AND message 2903d14c5d2SYehuda Sadeh * 2913d14c5d2SYehuda Sadeh */ 2923d14c5d2SYehuda Sadeh void ceph_osdc_build_request(struct ceph_osd_request *req, 293ae7ca4a3SAlex Elder u64 off, u64 len, unsigned int num_op, 2943d14c5d2SYehuda Sadeh struct ceph_osd_req_op *src_ops, 2954d6b250bSAlex Elder struct ceph_snap_context *snapc, u64 snap_id, 296af77f26cSAlex Elder struct timespec *mtime) 2973d14c5d2SYehuda Sadeh { 2983d14c5d2SYehuda Sadeh struct ceph_msg *msg = req->r_request; 2993d14c5d2SYehuda Sadeh struct ceph_osd_request_head *head; 3003d14c5d2SYehuda Sadeh struct ceph_osd_req_op *src_op; 3013d14c5d2SYehuda Sadeh struct ceph_osd_op *op; 3023d14c5d2SYehuda Sadeh void *p; 3033d14c5d2SYehuda Sadeh size_t msg_size = sizeof(*head) + num_op*sizeof(*op); 3043d14c5d2SYehuda Sadeh int flags = req->r_flags; 3053d14c5d2SYehuda Sadeh u64 data_len = 0; 3063d14c5d2SYehuda Sadeh int i; 3073d14c5d2SYehuda Sadeh 308d178a9e7SAlex Elder WARN_ON((flags & (CEPH_OSD_FLAG_READ|CEPH_OSD_FLAG_WRITE)) == 0); 309d178a9e7SAlex Elder 3103d14c5d2SYehuda Sadeh head = msg->front.iov_base; 3114d6b250bSAlex Elder head->snapid = cpu_to_le64(snap_id); 3123d14c5d2SYehuda Sadeh op = (void *)(head + 1); 3133d14c5d2SYehuda Sadeh p = (void *)(op + num_op); 3143d14c5d2SYehuda Sadeh 3153d14c5d2SYehuda Sadeh req->r_snapc = ceph_get_snap_context(snapc); 3163d14c5d2SYehuda Sadeh 3173d14c5d2SYehuda Sadeh head->client_inc = cpu_to_le32(1); /* always, for now. */ 3183d14c5d2SYehuda Sadeh head->flags = cpu_to_le32(flags); 3193d14c5d2SYehuda Sadeh if (flags & CEPH_OSD_FLAG_WRITE) 3203d14c5d2SYehuda Sadeh ceph_encode_timespec(&head->mtime, mtime); 321ae7ca4a3SAlex Elder BUG_ON(num_op > (unsigned int) ((u16) -1)); 3223d14c5d2SYehuda Sadeh head->num_ops = cpu_to_le16(num_op); 3233d14c5d2SYehuda Sadeh 3243d14c5d2SYehuda Sadeh /* fill in oid */ 325af77f26cSAlex Elder head->object_len = cpu_to_le32(req->r_oid_len); 326af77f26cSAlex Elder memcpy(p, req->r_oid, req->r_oid_len); 327af77f26cSAlex Elder p += req->r_oid_len; 3283d14c5d2SYehuda Sadeh 3293d14c5d2SYehuda Sadeh src_op = src_ops; 330ae7ca4a3SAlex Elder while (num_op--) 331ae7ca4a3SAlex Elder osd_req_encode_op(req, op++, src_op++); 3323d14c5d2SYehuda Sadeh 333c885837fSAlex Elder data_len += req->r_trail.length; 3343d14c5d2SYehuda Sadeh 3353d14c5d2SYehuda Sadeh if (snapc) { 3363d14c5d2SYehuda Sadeh head->snap_seq = cpu_to_le64(snapc->seq); 3373d14c5d2SYehuda Sadeh head->num_snaps = cpu_to_le32(snapc->num_snaps); 3383d14c5d2SYehuda Sadeh for (i = 0; i < snapc->num_snaps; i++) { 3393d14c5d2SYehuda Sadeh put_unaligned_le64(snapc->snaps[i], p); 3403d14c5d2SYehuda Sadeh p += sizeof(u64); 3413d14c5d2SYehuda Sadeh } 3423d14c5d2SYehuda Sadeh } 3433d14c5d2SYehuda Sadeh 3443d14c5d2SYehuda Sadeh if (flags & CEPH_OSD_FLAG_WRITE) { 3453d14c5d2SYehuda Sadeh req->r_request->hdr.data_off = cpu_to_le16(off); 3460120be3cSAlex Elder req->r_request->hdr.data_len = cpu_to_le32(len + data_len); 3473d14c5d2SYehuda Sadeh } else if (data_len) { 3483d14c5d2SYehuda Sadeh req->r_request->hdr.data_off = 0; 3493d14c5d2SYehuda Sadeh req->r_request->hdr.data_len = cpu_to_le32(data_len); 3503d14c5d2SYehuda Sadeh } 3513d14c5d2SYehuda Sadeh 352c5c6b19dSSage Weil req->r_request->page_alignment = req->r_page_alignment; 353c5c6b19dSSage Weil 3543d14c5d2SYehuda Sadeh BUG_ON(p > msg->front.iov_base + msg->front.iov_len); 3553d14c5d2SYehuda Sadeh msg_size = p - msg->front.iov_base; 3563d14c5d2SYehuda Sadeh msg->front.iov_len = msg_size; 3573d14c5d2SYehuda Sadeh msg->hdr.front_len = cpu_to_le32(msg_size); 3583d14c5d2SYehuda Sadeh return; 3593d14c5d2SYehuda Sadeh } 3603d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_build_request); 3613d14c5d2SYehuda Sadeh 3623d14c5d2SYehuda Sadeh /* 3633d14c5d2SYehuda Sadeh * build new request AND message, calculate layout, and adjust file 3643d14c5d2SYehuda Sadeh * extent as needed. 3653d14c5d2SYehuda Sadeh * 3663d14c5d2SYehuda Sadeh * if the file was recently truncated, we include information about its 3673d14c5d2SYehuda Sadeh * old and new size so that the object can be updated appropriately. (we 3683d14c5d2SYehuda Sadeh * avoid synchronously deleting truncated objects because it's slow.) 3693d14c5d2SYehuda Sadeh * 3703d14c5d2SYehuda Sadeh * if @do_sync, include a 'startsync' command so that the osd will flush 3713d14c5d2SYehuda Sadeh * data quickly. 3723d14c5d2SYehuda Sadeh */ 3733d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc, 3743d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 3753d14c5d2SYehuda Sadeh struct ceph_vino vino, 3763d14c5d2SYehuda Sadeh u64 off, u64 *plen, 3773d14c5d2SYehuda Sadeh int opcode, int flags, 3783d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 3793d14c5d2SYehuda Sadeh int do_sync, 3803d14c5d2SYehuda Sadeh u32 truncate_seq, 3813d14c5d2SYehuda Sadeh u64 truncate_size, 3823d14c5d2SYehuda Sadeh struct timespec *mtime, 383a3bea47eSAlex Elder bool use_mempool, 384b7495fc2SSage Weil int page_align) 3853d14c5d2SYehuda Sadeh { 386ae7ca4a3SAlex Elder struct ceph_osd_req_op ops[2]; 3873d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 388ae7ca4a3SAlex Elder unsigned int num_op = 1; 3896816282dSSage Weil int r; 3903d14c5d2SYehuda Sadeh 391ae7ca4a3SAlex Elder memset(&ops, 0, sizeof ops); 392ae7ca4a3SAlex Elder 3933d14c5d2SYehuda Sadeh ops[0].op = opcode; 3943d14c5d2SYehuda Sadeh ops[0].extent.truncate_seq = truncate_seq; 3953d14c5d2SYehuda Sadeh ops[0].extent.truncate_size = truncate_size; 3963d14c5d2SYehuda Sadeh 3973d14c5d2SYehuda Sadeh if (do_sync) { 3983d14c5d2SYehuda Sadeh ops[1].op = CEPH_OSD_OP_STARTSYNC; 399ae7ca4a3SAlex Elder num_op++; 400ae7ca4a3SAlex Elder } 4013d14c5d2SYehuda Sadeh 402ae7ca4a3SAlex Elder req = ceph_osdc_alloc_request(osdc, snapc, num_op, use_mempool, 403ae7ca4a3SAlex Elder GFP_NOFS); 4044ad12621SSage Weil if (!req) 4056816282dSSage Weil return ERR_PTR(-ENOMEM); 406d178a9e7SAlex Elder req->r_flags = flags; 4073d14c5d2SYehuda Sadeh 4083d14c5d2SYehuda Sadeh /* calculate max write size */ 409e75b45cfSAlex Elder r = calc_layout(vino, layout, off, plen, req, ops); 4106816282dSSage Weil if (r < 0) 4116816282dSSage Weil return ERR_PTR(r); 4123d14c5d2SYehuda Sadeh req->r_file_layout = *layout; /* keep a copy */ 4133d14c5d2SYehuda Sadeh 4149bb0ce2bSSage Weil /* in case it differs from natural (file) alignment that 4159bb0ce2bSSage Weil calc_layout filled in for us */ 4169bb0ce2bSSage Weil req->r_num_pages = calc_pages_for(page_align, *plen); 417b7495fc2SSage Weil req->r_page_alignment = page_align; 418b7495fc2SSage Weil 419ae7ca4a3SAlex Elder ceph_osdc_build_request(req, off, *plen, num_op, ops, 420ae7ca4a3SAlex Elder snapc, vino.snap, mtime); 4213d14c5d2SYehuda Sadeh 4223d14c5d2SYehuda Sadeh return req; 4233d14c5d2SYehuda Sadeh } 4243d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_new_request); 4253d14c5d2SYehuda Sadeh 4263d14c5d2SYehuda Sadeh /* 4273d14c5d2SYehuda Sadeh * We keep osd requests in an rbtree, sorted by ->r_tid. 4283d14c5d2SYehuda Sadeh */ 4293d14c5d2SYehuda Sadeh static void __insert_request(struct ceph_osd_client *osdc, 4303d14c5d2SYehuda Sadeh struct ceph_osd_request *new) 4313d14c5d2SYehuda Sadeh { 4323d14c5d2SYehuda Sadeh struct rb_node **p = &osdc->requests.rb_node; 4333d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 4343d14c5d2SYehuda Sadeh struct ceph_osd_request *req = NULL; 4353d14c5d2SYehuda Sadeh 4363d14c5d2SYehuda Sadeh while (*p) { 4373d14c5d2SYehuda Sadeh parent = *p; 4383d14c5d2SYehuda Sadeh req = rb_entry(parent, struct ceph_osd_request, r_node); 4393d14c5d2SYehuda Sadeh if (new->r_tid < req->r_tid) 4403d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 4413d14c5d2SYehuda Sadeh else if (new->r_tid > req->r_tid) 4423d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 4433d14c5d2SYehuda Sadeh else 4443d14c5d2SYehuda Sadeh BUG(); 4453d14c5d2SYehuda Sadeh } 4463d14c5d2SYehuda Sadeh 4473d14c5d2SYehuda Sadeh rb_link_node(&new->r_node, parent, p); 4483d14c5d2SYehuda Sadeh rb_insert_color(&new->r_node, &osdc->requests); 4493d14c5d2SYehuda Sadeh } 4503d14c5d2SYehuda Sadeh 4513d14c5d2SYehuda Sadeh static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc, 4523d14c5d2SYehuda Sadeh u64 tid) 4533d14c5d2SYehuda Sadeh { 4543d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 4553d14c5d2SYehuda Sadeh struct rb_node *n = osdc->requests.rb_node; 4563d14c5d2SYehuda Sadeh 4573d14c5d2SYehuda Sadeh while (n) { 4583d14c5d2SYehuda Sadeh req = rb_entry(n, struct ceph_osd_request, r_node); 4593d14c5d2SYehuda Sadeh if (tid < req->r_tid) 4603d14c5d2SYehuda Sadeh n = n->rb_left; 4613d14c5d2SYehuda Sadeh else if (tid > req->r_tid) 4623d14c5d2SYehuda Sadeh n = n->rb_right; 4633d14c5d2SYehuda Sadeh else 4643d14c5d2SYehuda Sadeh return req; 4653d14c5d2SYehuda Sadeh } 4663d14c5d2SYehuda Sadeh return NULL; 4673d14c5d2SYehuda Sadeh } 4683d14c5d2SYehuda Sadeh 4693d14c5d2SYehuda Sadeh static struct ceph_osd_request * 4703d14c5d2SYehuda Sadeh __lookup_request_ge(struct ceph_osd_client *osdc, 4713d14c5d2SYehuda Sadeh u64 tid) 4723d14c5d2SYehuda Sadeh { 4733d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 4743d14c5d2SYehuda Sadeh struct rb_node *n = osdc->requests.rb_node; 4753d14c5d2SYehuda Sadeh 4763d14c5d2SYehuda Sadeh while (n) { 4773d14c5d2SYehuda Sadeh req = rb_entry(n, struct ceph_osd_request, r_node); 4783d14c5d2SYehuda Sadeh if (tid < req->r_tid) { 4793d14c5d2SYehuda Sadeh if (!n->rb_left) 4803d14c5d2SYehuda Sadeh return req; 4813d14c5d2SYehuda Sadeh n = n->rb_left; 4823d14c5d2SYehuda Sadeh } else if (tid > req->r_tid) { 4833d14c5d2SYehuda Sadeh n = n->rb_right; 4843d14c5d2SYehuda Sadeh } else { 4853d14c5d2SYehuda Sadeh return req; 4863d14c5d2SYehuda Sadeh } 4873d14c5d2SYehuda Sadeh } 4883d14c5d2SYehuda Sadeh return NULL; 4893d14c5d2SYehuda Sadeh } 4903d14c5d2SYehuda Sadeh 4916f6c7006SSage Weil /* 4926f6c7006SSage Weil * Resubmit requests pending on the given osd. 4936f6c7006SSage Weil */ 4946f6c7006SSage Weil static void __kick_osd_requests(struct ceph_osd_client *osdc, 4956f6c7006SSage Weil struct ceph_osd *osd) 4966f6c7006SSage Weil { 497a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 4986f6c7006SSage Weil int err; 4996f6c7006SSage Weil 5006f6c7006SSage Weil dout("__kick_osd_requests osd%d\n", osd->o_osd); 5016f6c7006SSage Weil err = __reset_osd(osdc, osd); 502685a7555SAlex Elder if (err) 5036f6c7006SSage Weil return; 5046f6c7006SSage Weil 5056f6c7006SSage Weil list_for_each_entry(req, &osd->o_requests, r_osd_item) { 5066f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_unsent); 5076f6c7006SSage Weil dout("requeued %p tid %llu osd%d\n", req, req->r_tid, 5086f6c7006SSage Weil osd->o_osd); 509a40c4f10SYehuda Sadeh if (!req->r_linger) 5106f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 5116f6c7006SSage Weil } 512a40c4f10SYehuda Sadeh 513a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osd->o_linger_requests, 514a40c4f10SYehuda Sadeh r_linger_osd) { 51577f38e0eSSage Weil /* 51677f38e0eSSage Weil * reregister request prior to unregistering linger so 51777f38e0eSSage Weil * that r_osd is preserved. 51877f38e0eSSage Weil */ 51977f38e0eSSage Weil BUG_ON(!list_empty(&req->r_req_lru_item)); 520a40c4f10SYehuda Sadeh __register_request(osdc, req); 52177f38e0eSSage Weil list_add(&req->r_req_lru_item, &osdc->req_unsent); 52277f38e0eSSage Weil list_add(&req->r_osd_item, &req->r_osd->o_requests); 52377f38e0eSSage Weil __unregister_linger_request(osdc, req); 524a40c4f10SYehuda Sadeh dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid, 525a40c4f10SYehuda Sadeh osd->o_osd); 526a40c4f10SYehuda Sadeh } 5276f6c7006SSage Weil } 5286f6c7006SSage Weil 5293d14c5d2SYehuda Sadeh /* 5303d14c5d2SYehuda Sadeh * If the osd connection drops, we need to resubmit all requests. 5313d14c5d2SYehuda Sadeh */ 5323d14c5d2SYehuda Sadeh static void osd_reset(struct ceph_connection *con) 5333d14c5d2SYehuda Sadeh { 5343d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 5353d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 5363d14c5d2SYehuda Sadeh 5373d14c5d2SYehuda Sadeh if (!osd) 5383d14c5d2SYehuda Sadeh return; 5393d14c5d2SYehuda Sadeh dout("osd_reset osd%d\n", osd->o_osd); 5403d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 5413d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 54283aff95eSSage Weil mutex_lock(&osdc->request_mutex); 54383aff95eSSage Weil __kick_osd_requests(osdc, osd); 544f9d25199SAlex Elder __send_queued(osdc); 54583aff95eSSage Weil mutex_unlock(&osdc->request_mutex); 5463d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 5473d14c5d2SYehuda Sadeh } 5483d14c5d2SYehuda Sadeh 5493d14c5d2SYehuda Sadeh /* 5503d14c5d2SYehuda Sadeh * Track open sessions with osds. 5513d14c5d2SYehuda Sadeh */ 552e10006f8SAlex Elder static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum) 5533d14c5d2SYehuda Sadeh { 5543d14c5d2SYehuda Sadeh struct ceph_osd *osd; 5553d14c5d2SYehuda Sadeh 5563d14c5d2SYehuda Sadeh osd = kzalloc(sizeof(*osd), GFP_NOFS); 5573d14c5d2SYehuda Sadeh if (!osd) 5583d14c5d2SYehuda Sadeh return NULL; 5593d14c5d2SYehuda Sadeh 5603d14c5d2SYehuda Sadeh atomic_set(&osd->o_ref, 1); 5613d14c5d2SYehuda Sadeh osd->o_osdc = osdc; 562e10006f8SAlex Elder osd->o_osd = onum; 563f407731dSAlex Elder RB_CLEAR_NODE(&osd->o_node); 5643d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_requests); 565a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osd->o_linger_requests); 5663d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_osd_lru); 5673d14c5d2SYehuda Sadeh osd->o_incarnation = 1; 5683d14c5d2SYehuda Sadeh 569b7a9e5ddSSage Weil ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr); 5703d14c5d2SYehuda Sadeh 5713d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_keepalive_item); 5723d14c5d2SYehuda Sadeh return osd; 5733d14c5d2SYehuda Sadeh } 5743d14c5d2SYehuda Sadeh 5753d14c5d2SYehuda Sadeh static struct ceph_osd *get_osd(struct ceph_osd *osd) 5763d14c5d2SYehuda Sadeh { 5773d14c5d2SYehuda Sadeh if (atomic_inc_not_zero(&osd->o_ref)) { 5783d14c5d2SYehuda Sadeh dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1, 5793d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref)); 5803d14c5d2SYehuda Sadeh return osd; 5813d14c5d2SYehuda Sadeh } else { 5823d14c5d2SYehuda Sadeh dout("get_osd %p FAIL\n", osd); 5833d14c5d2SYehuda Sadeh return NULL; 5843d14c5d2SYehuda Sadeh } 5853d14c5d2SYehuda Sadeh } 5863d14c5d2SYehuda Sadeh 5873d14c5d2SYehuda Sadeh static void put_osd(struct ceph_osd *osd) 5883d14c5d2SYehuda Sadeh { 5893d14c5d2SYehuda Sadeh dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref), 5903d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref) - 1); 591a255651dSAlex Elder if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) { 5923d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth; 5933d14c5d2SYehuda Sadeh 594a255651dSAlex Elder if (ac->ops && ac->ops->destroy_authorizer) 5956c4a1915SAlex Elder ac->ops->destroy_authorizer(ac, osd->o_auth.authorizer); 5963d14c5d2SYehuda Sadeh kfree(osd); 5973d14c5d2SYehuda Sadeh } 5983d14c5d2SYehuda Sadeh } 5993d14c5d2SYehuda Sadeh 6003d14c5d2SYehuda Sadeh /* 6013d14c5d2SYehuda Sadeh * remove an osd from our map 6023d14c5d2SYehuda Sadeh */ 6033d14c5d2SYehuda Sadeh static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 6043d14c5d2SYehuda Sadeh { 6053d14c5d2SYehuda Sadeh dout("__remove_osd %p\n", osd); 6063d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_requests)); 6073d14c5d2SYehuda Sadeh rb_erase(&osd->o_node, &osdc->osds); 6083d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 6093d14c5d2SYehuda Sadeh ceph_con_close(&osd->o_con); 6103d14c5d2SYehuda Sadeh put_osd(osd); 6113d14c5d2SYehuda Sadeh } 6123d14c5d2SYehuda Sadeh 613aca420bcSSage Weil static void remove_all_osds(struct ceph_osd_client *osdc) 614aca420bcSSage Weil { 615048a9d2dSJiaju Zhang dout("%s %p\n", __func__, osdc); 616aca420bcSSage Weil mutex_lock(&osdc->request_mutex); 617aca420bcSSage Weil while (!RB_EMPTY_ROOT(&osdc->osds)) { 618aca420bcSSage Weil struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds), 619aca420bcSSage Weil struct ceph_osd, o_node); 620aca420bcSSage Weil __remove_osd(osdc, osd); 621aca420bcSSage Weil } 622aca420bcSSage Weil mutex_unlock(&osdc->request_mutex); 623aca420bcSSage Weil } 624aca420bcSSage Weil 6253d14c5d2SYehuda Sadeh static void __move_osd_to_lru(struct ceph_osd_client *osdc, 6263d14c5d2SYehuda Sadeh struct ceph_osd *osd) 6273d14c5d2SYehuda Sadeh { 6283d14c5d2SYehuda Sadeh dout("__move_osd_to_lru %p\n", osd); 6293d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_osd_lru)); 6303d14c5d2SYehuda Sadeh list_add_tail(&osd->o_osd_lru, &osdc->osd_lru); 6313d14c5d2SYehuda Sadeh osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ; 6323d14c5d2SYehuda Sadeh } 6333d14c5d2SYehuda Sadeh 6343d14c5d2SYehuda Sadeh static void __remove_osd_from_lru(struct ceph_osd *osd) 6353d14c5d2SYehuda Sadeh { 6363d14c5d2SYehuda Sadeh dout("__remove_osd_from_lru %p\n", osd); 6373d14c5d2SYehuda Sadeh if (!list_empty(&osd->o_osd_lru)) 6383d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 6393d14c5d2SYehuda Sadeh } 6403d14c5d2SYehuda Sadeh 641aca420bcSSage Weil static void remove_old_osds(struct ceph_osd_client *osdc) 6423d14c5d2SYehuda Sadeh { 6433d14c5d2SYehuda Sadeh struct ceph_osd *osd, *nosd; 6443d14c5d2SYehuda Sadeh 6453d14c5d2SYehuda Sadeh dout("__remove_old_osds %p\n", osdc); 6463d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 6473d14c5d2SYehuda Sadeh list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) { 648aca420bcSSage Weil if (time_before(jiffies, osd->lru_ttl)) 6493d14c5d2SYehuda Sadeh break; 6503d14c5d2SYehuda Sadeh __remove_osd(osdc, osd); 6513d14c5d2SYehuda Sadeh } 6523d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 6533d14c5d2SYehuda Sadeh } 6543d14c5d2SYehuda Sadeh 6553d14c5d2SYehuda Sadeh /* 6563d14c5d2SYehuda Sadeh * reset osd connect 6573d14c5d2SYehuda Sadeh */ 6583d14c5d2SYehuda Sadeh static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 6593d14c5d2SYehuda Sadeh { 660c3acb181SAlex Elder struct ceph_entity_addr *peer_addr; 6613d14c5d2SYehuda Sadeh 6623d14c5d2SYehuda Sadeh dout("__reset_osd %p osd%d\n", osd, osd->o_osd); 663a40c4f10SYehuda Sadeh if (list_empty(&osd->o_requests) && 664a40c4f10SYehuda Sadeh list_empty(&osd->o_linger_requests)) { 6653d14c5d2SYehuda Sadeh __remove_osd(osdc, osd); 666c3acb181SAlex Elder 667c3acb181SAlex Elder return -ENODEV; 668c3acb181SAlex Elder } 669c3acb181SAlex Elder 670c3acb181SAlex Elder peer_addr = &osdc->osdmap->osd_addr[osd->o_osd]; 671c3acb181SAlex Elder if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) && 6723d14c5d2SYehuda Sadeh !ceph_con_opened(&osd->o_con)) { 673c3acb181SAlex Elder struct ceph_osd_request *req; 674c3acb181SAlex Elder 6753d14c5d2SYehuda Sadeh dout(" osd addr hasn't changed and connection never opened," 6763d14c5d2SYehuda Sadeh " letting msgr retry"); 6773d14c5d2SYehuda Sadeh /* touch each r_stamp for handle_timeout()'s benfit */ 6783d14c5d2SYehuda Sadeh list_for_each_entry(req, &osd->o_requests, r_osd_item) 6793d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 680c3acb181SAlex Elder 681c3acb181SAlex Elder return -EAGAIN; 6823d14c5d2SYehuda Sadeh } 683c3acb181SAlex Elder 684c3acb181SAlex Elder ceph_con_close(&osd->o_con); 685c3acb181SAlex Elder ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr); 686c3acb181SAlex Elder osd->o_incarnation++; 687c3acb181SAlex Elder 688c3acb181SAlex Elder return 0; 6893d14c5d2SYehuda Sadeh } 6903d14c5d2SYehuda Sadeh 6913d14c5d2SYehuda Sadeh static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new) 6923d14c5d2SYehuda Sadeh { 6933d14c5d2SYehuda Sadeh struct rb_node **p = &osdc->osds.rb_node; 6943d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 6953d14c5d2SYehuda Sadeh struct ceph_osd *osd = NULL; 6963d14c5d2SYehuda Sadeh 697aca420bcSSage Weil dout("__insert_osd %p osd%d\n", new, new->o_osd); 6983d14c5d2SYehuda Sadeh while (*p) { 6993d14c5d2SYehuda Sadeh parent = *p; 7003d14c5d2SYehuda Sadeh osd = rb_entry(parent, struct ceph_osd, o_node); 7013d14c5d2SYehuda Sadeh if (new->o_osd < osd->o_osd) 7023d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 7033d14c5d2SYehuda Sadeh else if (new->o_osd > osd->o_osd) 7043d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 7053d14c5d2SYehuda Sadeh else 7063d14c5d2SYehuda Sadeh BUG(); 7073d14c5d2SYehuda Sadeh } 7083d14c5d2SYehuda Sadeh 7093d14c5d2SYehuda Sadeh rb_link_node(&new->o_node, parent, p); 7103d14c5d2SYehuda Sadeh rb_insert_color(&new->o_node, &osdc->osds); 7113d14c5d2SYehuda Sadeh } 7123d14c5d2SYehuda Sadeh 7133d14c5d2SYehuda Sadeh static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o) 7143d14c5d2SYehuda Sadeh { 7153d14c5d2SYehuda Sadeh struct ceph_osd *osd; 7163d14c5d2SYehuda Sadeh struct rb_node *n = osdc->osds.rb_node; 7173d14c5d2SYehuda Sadeh 7183d14c5d2SYehuda Sadeh while (n) { 7193d14c5d2SYehuda Sadeh osd = rb_entry(n, struct ceph_osd, o_node); 7203d14c5d2SYehuda Sadeh if (o < osd->o_osd) 7213d14c5d2SYehuda Sadeh n = n->rb_left; 7223d14c5d2SYehuda Sadeh else if (o > osd->o_osd) 7233d14c5d2SYehuda Sadeh n = n->rb_right; 7243d14c5d2SYehuda Sadeh else 7253d14c5d2SYehuda Sadeh return osd; 7263d14c5d2SYehuda Sadeh } 7273d14c5d2SYehuda Sadeh return NULL; 7283d14c5d2SYehuda Sadeh } 7293d14c5d2SYehuda Sadeh 7303d14c5d2SYehuda Sadeh static void __schedule_osd_timeout(struct ceph_osd_client *osdc) 7313d14c5d2SYehuda Sadeh { 7323d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->timeout_work, 7333d14c5d2SYehuda Sadeh osdc->client->options->osd_keepalive_timeout * HZ); 7343d14c5d2SYehuda Sadeh } 7353d14c5d2SYehuda Sadeh 7363d14c5d2SYehuda Sadeh static void __cancel_osd_timeout(struct ceph_osd_client *osdc) 7373d14c5d2SYehuda Sadeh { 7383d14c5d2SYehuda Sadeh cancel_delayed_work(&osdc->timeout_work); 7393d14c5d2SYehuda Sadeh } 7403d14c5d2SYehuda Sadeh 7413d14c5d2SYehuda Sadeh /* 7423d14c5d2SYehuda Sadeh * Register request, assign tid. If this is the first request, set up 7433d14c5d2SYehuda Sadeh * the timeout event. 7443d14c5d2SYehuda Sadeh */ 745a40c4f10SYehuda Sadeh static void __register_request(struct ceph_osd_client *osdc, 7463d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 7473d14c5d2SYehuda Sadeh { 7483d14c5d2SYehuda Sadeh req->r_tid = ++osdc->last_tid; 7493d14c5d2SYehuda Sadeh req->r_request->hdr.tid = cpu_to_le64(req->r_tid); 75077f38e0eSSage Weil dout("__register_request %p tid %lld\n", req, req->r_tid); 7513d14c5d2SYehuda Sadeh __insert_request(osdc, req); 7523d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 7533d14c5d2SYehuda Sadeh osdc->num_requests++; 7543d14c5d2SYehuda Sadeh if (osdc->num_requests == 1) { 7553d14c5d2SYehuda Sadeh dout(" first request, scheduling timeout\n"); 7563d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 7573d14c5d2SYehuda Sadeh } 758a40c4f10SYehuda Sadeh } 759a40c4f10SYehuda Sadeh 760a40c4f10SYehuda Sadeh static void register_request(struct ceph_osd_client *osdc, 761a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 762a40c4f10SYehuda Sadeh { 763a40c4f10SYehuda Sadeh mutex_lock(&osdc->request_mutex); 764a40c4f10SYehuda Sadeh __register_request(osdc, req); 7653d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 7663d14c5d2SYehuda Sadeh } 7673d14c5d2SYehuda Sadeh 7683d14c5d2SYehuda Sadeh /* 7693d14c5d2SYehuda Sadeh * called under osdc->request_mutex 7703d14c5d2SYehuda Sadeh */ 7713d14c5d2SYehuda Sadeh static void __unregister_request(struct ceph_osd_client *osdc, 7723d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 7733d14c5d2SYehuda Sadeh { 77435f9f8a0SSage Weil if (RB_EMPTY_NODE(&req->r_node)) { 77535f9f8a0SSage Weil dout("__unregister_request %p tid %lld not registered\n", 77635f9f8a0SSage Weil req, req->r_tid); 77735f9f8a0SSage Weil return; 77835f9f8a0SSage Weil } 77935f9f8a0SSage Weil 7803d14c5d2SYehuda Sadeh dout("__unregister_request %p tid %lld\n", req, req->r_tid); 7813d14c5d2SYehuda Sadeh rb_erase(&req->r_node, &osdc->requests); 7823d14c5d2SYehuda Sadeh osdc->num_requests--; 7833d14c5d2SYehuda Sadeh 7843d14c5d2SYehuda Sadeh if (req->r_osd) { 7853d14c5d2SYehuda Sadeh /* make sure the original request isn't in flight. */ 7866740a845SAlex Elder ceph_msg_revoke(req->r_request); 7873d14c5d2SYehuda Sadeh 7883d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 789a40c4f10SYehuda Sadeh if (list_empty(&req->r_osd->o_requests) && 790a40c4f10SYehuda Sadeh list_empty(&req->r_osd->o_linger_requests)) { 791a40c4f10SYehuda Sadeh dout("moving osd to %p lru\n", req->r_osd); 7923d14c5d2SYehuda Sadeh __move_osd_to_lru(osdc, req->r_osd); 793a40c4f10SYehuda Sadeh } 794fbdb9190SSage Weil if (list_empty(&req->r_linger_item)) 7953d14c5d2SYehuda Sadeh req->r_osd = NULL; 7963d14c5d2SYehuda Sadeh } 7973d14c5d2SYehuda Sadeh 7987d5f2481SAlex Elder list_del_init(&req->r_req_lru_item); 7993d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 8003d14c5d2SYehuda Sadeh 8013d14c5d2SYehuda Sadeh if (osdc->num_requests == 0) { 8023d14c5d2SYehuda Sadeh dout(" no requests, canceling timeout\n"); 8033d14c5d2SYehuda Sadeh __cancel_osd_timeout(osdc); 8043d14c5d2SYehuda Sadeh } 8053d14c5d2SYehuda Sadeh } 8063d14c5d2SYehuda Sadeh 8073d14c5d2SYehuda Sadeh /* 8083d14c5d2SYehuda Sadeh * Cancel a previously queued request message 8093d14c5d2SYehuda Sadeh */ 8103d14c5d2SYehuda Sadeh static void __cancel_request(struct ceph_osd_request *req) 8113d14c5d2SYehuda Sadeh { 8123d14c5d2SYehuda Sadeh if (req->r_sent && req->r_osd) { 8136740a845SAlex Elder ceph_msg_revoke(req->r_request); 8143d14c5d2SYehuda Sadeh req->r_sent = 0; 8153d14c5d2SYehuda Sadeh } 8163d14c5d2SYehuda Sadeh } 8173d14c5d2SYehuda Sadeh 818a40c4f10SYehuda Sadeh static void __register_linger_request(struct ceph_osd_client *osdc, 819a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 820a40c4f10SYehuda Sadeh { 821a40c4f10SYehuda Sadeh dout("__register_linger_request %p\n", req); 822a40c4f10SYehuda Sadeh list_add_tail(&req->r_linger_item, &osdc->req_linger); 8236194ea89SSage Weil if (req->r_osd) 8246194ea89SSage Weil list_add_tail(&req->r_linger_osd, 8256194ea89SSage Weil &req->r_osd->o_linger_requests); 826a40c4f10SYehuda Sadeh } 827a40c4f10SYehuda Sadeh 828a40c4f10SYehuda Sadeh static void __unregister_linger_request(struct ceph_osd_client *osdc, 829a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 830a40c4f10SYehuda Sadeh { 831a40c4f10SYehuda Sadeh dout("__unregister_linger_request %p\n", req); 832a40c4f10SYehuda Sadeh list_del_init(&req->r_linger_item); 83361c74035SAlex Elder if (req->r_osd) { 834a40c4f10SYehuda Sadeh list_del_init(&req->r_linger_osd); 835a40c4f10SYehuda Sadeh 836a40c4f10SYehuda Sadeh if (list_empty(&req->r_osd->o_requests) && 837a40c4f10SYehuda Sadeh list_empty(&req->r_osd->o_linger_requests)) { 838a40c4f10SYehuda Sadeh dout("moving osd to %p lru\n", req->r_osd); 839a40c4f10SYehuda Sadeh __move_osd_to_lru(osdc, req->r_osd); 840a40c4f10SYehuda Sadeh } 841fbdb9190SSage Weil if (list_empty(&req->r_osd_item)) 842a40c4f10SYehuda Sadeh req->r_osd = NULL; 843a40c4f10SYehuda Sadeh } 844a40c4f10SYehuda Sadeh } 845a40c4f10SYehuda Sadeh 846a40c4f10SYehuda Sadeh void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc, 847a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 848a40c4f10SYehuda Sadeh { 849a40c4f10SYehuda Sadeh mutex_lock(&osdc->request_mutex); 850a40c4f10SYehuda Sadeh if (req->r_linger) { 851a40c4f10SYehuda Sadeh __unregister_linger_request(osdc, req); 852a40c4f10SYehuda Sadeh ceph_osdc_put_request(req); 853a40c4f10SYehuda Sadeh } 854a40c4f10SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 855a40c4f10SYehuda Sadeh } 856a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_unregister_linger_request); 857a40c4f10SYehuda Sadeh 858a40c4f10SYehuda Sadeh void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc, 859a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 860a40c4f10SYehuda Sadeh { 861a40c4f10SYehuda Sadeh if (!req->r_linger) { 862a40c4f10SYehuda Sadeh dout("set_request_linger %p\n", req); 863a40c4f10SYehuda Sadeh req->r_linger = 1; 864a40c4f10SYehuda Sadeh /* 865a40c4f10SYehuda Sadeh * caller is now responsible for calling 866a40c4f10SYehuda Sadeh * unregister_linger_request 867a40c4f10SYehuda Sadeh */ 868a40c4f10SYehuda Sadeh ceph_osdc_get_request(req); 869a40c4f10SYehuda Sadeh } 870a40c4f10SYehuda Sadeh } 871a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_set_request_linger); 872a40c4f10SYehuda Sadeh 8733d14c5d2SYehuda Sadeh /* 8743d14c5d2SYehuda Sadeh * Pick an osd (the first 'up' osd in the pg), allocate the osd struct 8753d14c5d2SYehuda Sadeh * (as needed), and set the request r_osd appropriately. If there is 87625985edcSLucas De Marchi * no up osd, set r_osd to NULL. Move the request to the appropriate list 8776f6c7006SSage Weil * (unsent, homeless) or leave on in-flight lru. 8783d14c5d2SYehuda Sadeh * 8793d14c5d2SYehuda Sadeh * Return 0 if unchanged, 1 if changed, or negative on error. 8803d14c5d2SYehuda Sadeh * 8813d14c5d2SYehuda Sadeh * Caller should hold map_sem for read and request_mutex. 8823d14c5d2SYehuda Sadeh */ 8836f6c7006SSage Weil static int __map_request(struct ceph_osd_client *osdc, 88438d6453cSSage Weil struct ceph_osd_request *req, int force_resend) 8853d14c5d2SYehuda Sadeh { 8863d14c5d2SYehuda Sadeh struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base; 8873d14c5d2SYehuda Sadeh struct ceph_pg pgid; 8883d14c5d2SYehuda Sadeh int acting[CEPH_PG_MAX_SIZE]; 8893d14c5d2SYehuda Sadeh int o = -1, num = 0; 8903d14c5d2SYehuda Sadeh int err; 8913d14c5d2SYehuda Sadeh 8926f6c7006SSage Weil dout("map_request %p tid %lld\n", req, req->r_tid); 8933d14c5d2SYehuda Sadeh err = ceph_calc_object_layout(&reqhead->layout, req->r_oid, 8943d14c5d2SYehuda Sadeh &req->r_file_layout, osdc->osdmap); 8956f6c7006SSage Weil if (err) { 8966f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 8973d14c5d2SYehuda Sadeh return err; 8986f6c7006SSage Weil } 8993d14c5d2SYehuda Sadeh pgid = reqhead->layout.ol_pgid; 9003d14c5d2SYehuda Sadeh req->r_pgid = pgid; 9013d14c5d2SYehuda Sadeh 9023d14c5d2SYehuda Sadeh err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting); 9033d14c5d2SYehuda Sadeh if (err > 0) { 9043d14c5d2SYehuda Sadeh o = acting[0]; 9053d14c5d2SYehuda Sadeh num = err; 9063d14c5d2SYehuda Sadeh } 9073d14c5d2SYehuda Sadeh 90838d6453cSSage Weil if ((!force_resend && 90938d6453cSSage Weil req->r_osd && req->r_osd->o_osd == o && 9103d14c5d2SYehuda Sadeh req->r_sent >= req->r_osd->o_incarnation && 9113d14c5d2SYehuda Sadeh req->r_num_pg_osds == num && 9123d14c5d2SYehuda Sadeh memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) || 9133d14c5d2SYehuda Sadeh (req->r_osd == NULL && o == -1)) 9143d14c5d2SYehuda Sadeh return 0; /* no change */ 9153d14c5d2SYehuda Sadeh 9166f6c7006SSage Weil dout("map_request tid %llu pgid %d.%x osd%d (was osd%d)\n", 9173d14c5d2SYehuda Sadeh req->r_tid, le32_to_cpu(pgid.pool), le16_to_cpu(pgid.ps), o, 9183d14c5d2SYehuda Sadeh req->r_osd ? req->r_osd->o_osd : -1); 9193d14c5d2SYehuda Sadeh 9203d14c5d2SYehuda Sadeh /* record full pg acting set */ 9213d14c5d2SYehuda Sadeh memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num); 9223d14c5d2SYehuda Sadeh req->r_num_pg_osds = num; 9233d14c5d2SYehuda Sadeh 9243d14c5d2SYehuda Sadeh if (req->r_osd) { 9253d14c5d2SYehuda Sadeh __cancel_request(req); 9263d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 9273d14c5d2SYehuda Sadeh req->r_osd = NULL; 9283d14c5d2SYehuda Sadeh } 9293d14c5d2SYehuda Sadeh 9303d14c5d2SYehuda Sadeh req->r_osd = __lookup_osd(osdc, o); 9313d14c5d2SYehuda Sadeh if (!req->r_osd && o >= 0) { 9323d14c5d2SYehuda Sadeh err = -ENOMEM; 933e10006f8SAlex Elder req->r_osd = create_osd(osdc, o); 9346f6c7006SSage Weil if (!req->r_osd) { 9356f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 9363d14c5d2SYehuda Sadeh goto out; 9376f6c7006SSage Weil } 9383d14c5d2SYehuda Sadeh 9396f6c7006SSage Weil dout("map_request osd %p is osd%d\n", req->r_osd, o); 9403d14c5d2SYehuda Sadeh __insert_osd(osdc, req->r_osd); 9413d14c5d2SYehuda Sadeh 942b7a9e5ddSSage Weil ceph_con_open(&req->r_osd->o_con, 943b7a9e5ddSSage Weil CEPH_ENTITY_TYPE_OSD, o, 944b7a9e5ddSSage Weil &osdc->osdmap->osd_addr[o]); 9453d14c5d2SYehuda Sadeh } 9463d14c5d2SYehuda Sadeh 9473d14c5d2SYehuda Sadeh if (req->r_osd) { 9483d14c5d2SYehuda Sadeh __remove_osd_from_lru(req->r_osd); 9493d14c5d2SYehuda Sadeh list_add(&req->r_osd_item, &req->r_osd->o_requests); 9506f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_unsent); 9516f6c7006SSage Weil } else { 9526f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 9533d14c5d2SYehuda Sadeh } 9543d14c5d2SYehuda Sadeh err = 1; /* osd or pg changed */ 9553d14c5d2SYehuda Sadeh 9563d14c5d2SYehuda Sadeh out: 9573d14c5d2SYehuda Sadeh return err; 9583d14c5d2SYehuda Sadeh } 9593d14c5d2SYehuda Sadeh 9603d14c5d2SYehuda Sadeh /* 9613d14c5d2SYehuda Sadeh * caller should hold map_sem (for read) and request_mutex 9623d14c5d2SYehuda Sadeh */ 96356e925b6SSage Weil static void __send_request(struct ceph_osd_client *osdc, 9643d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 9653d14c5d2SYehuda Sadeh { 9663d14c5d2SYehuda Sadeh struct ceph_osd_request_head *reqhead; 9673d14c5d2SYehuda Sadeh 9683d14c5d2SYehuda Sadeh dout("send_request %p tid %llu to osd%d flags %d\n", 9693d14c5d2SYehuda Sadeh req, req->r_tid, req->r_osd->o_osd, req->r_flags); 9703d14c5d2SYehuda Sadeh 9713d14c5d2SYehuda Sadeh reqhead = req->r_request->front.iov_base; 9723d14c5d2SYehuda Sadeh reqhead->osdmap_epoch = cpu_to_le32(osdc->osdmap->epoch); 9733d14c5d2SYehuda Sadeh reqhead->flags |= cpu_to_le32(req->r_flags); /* e.g., RETRY */ 9743d14c5d2SYehuda Sadeh reqhead->reassert_version = req->r_reassert_version; 9753d14c5d2SYehuda Sadeh 9763d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 9773d14c5d2SYehuda Sadeh list_move_tail(&req->r_req_lru_item, &osdc->req_lru); 9783d14c5d2SYehuda Sadeh 9793d14c5d2SYehuda Sadeh ceph_msg_get(req->r_request); /* send consumes a ref */ 9803d14c5d2SYehuda Sadeh ceph_con_send(&req->r_osd->o_con, req->r_request); 9813d14c5d2SYehuda Sadeh req->r_sent = req->r_osd->o_incarnation; 9823d14c5d2SYehuda Sadeh } 9833d14c5d2SYehuda Sadeh 9843d14c5d2SYehuda Sadeh /* 9856f6c7006SSage Weil * Send any requests in the queue (req_unsent). 9866f6c7006SSage Weil */ 987f9d25199SAlex Elder static void __send_queued(struct ceph_osd_client *osdc) 9886f6c7006SSage Weil { 9896f6c7006SSage Weil struct ceph_osd_request *req, *tmp; 9906f6c7006SSage Weil 991f9d25199SAlex Elder dout("__send_queued\n"); 992f9d25199SAlex Elder list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) 9936f6c7006SSage Weil __send_request(osdc, req); 9946f6c7006SSage Weil } 9956f6c7006SSage Weil 9966f6c7006SSage Weil /* 9973d14c5d2SYehuda Sadeh * Timeout callback, called every N seconds when 1 or more osd 9983d14c5d2SYehuda Sadeh * requests has been active for more than N seconds. When this 9993d14c5d2SYehuda Sadeh * happens, we ping all OSDs with requests who have timed out to 10003d14c5d2SYehuda Sadeh * ensure any communications channel reset is detected. Reset the 10013d14c5d2SYehuda Sadeh * request timeouts another N seconds in the future as we go. 10023d14c5d2SYehuda Sadeh * Reschedule the timeout event another N seconds in future (unless 10033d14c5d2SYehuda Sadeh * there are no open requests). 10043d14c5d2SYehuda Sadeh */ 10053d14c5d2SYehuda Sadeh static void handle_timeout(struct work_struct *work) 10063d14c5d2SYehuda Sadeh { 10073d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 10083d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, timeout_work.work); 100983aff95eSSage Weil struct ceph_osd_request *req; 10103d14c5d2SYehuda Sadeh struct ceph_osd *osd; 10113d14c5d2SYehuda Sadeh unsigned long keepalive = 10123d14c5d2SYehuda Sadeh osdc->client->options->osd_keepalive_timeout * HZ; 10133d14c5d2SYehuda Sadeh struct list_head slow_osds; 10143d14c5d2SYehuda Sadeh dout("timeout\n"); 10153d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 10163d14c5d2SYehuda Sadeh 10173d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 10183d14c5d2SYehuda Sadeh 10193d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 10203d14c5d2SYehuda Sadeh 10213d14c5d2SYehuda Sadeh /* 10223d14c5d2SYehuda Sadeh * ping osds that are a bit slow. this ensures that if there 10233d14c5d2SYehuda Sadeh * is a break in the TCP connection we will notice, and reopen 10243d14c5d2SYehuda Sadeh * a connection with that osd (from the fault callback). 10253d14c5d2SYehuda Sadeh */ 10263d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&slow_osds); 10273d14c5d2SYehuda Sadeh list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) { 10283d14c5d2SYehuda Sadeh if (time_before(jiffies, req->r_stamp + keepalive)) 10293d14c5d2SYehuda Sadeh break; 10303d14c5d2SYehuda Sadeh 10313d14c5d2SYehuda Sadeh osd = req->r_osd; 10323d14c5d2SYehuda Sadeh BUG_ON(!osd); 10333d14c5d2SYehuda Sadeh dout(" tid %llu is slow, will send keepalive on osd%d\n", 10343d14c5d2SYehuda Sadeh req->r_tid, osd->o_osd); 10353d14c5d2SYehuda Sadeh list_move_tail(&osd->o_keepalive_item, &slow_osds); 10363d14c5d2SYehuda Sadeh } 10373d14c5d2SYehuda Sadeh while (!list_empty(&slow_osds)) { 10383d14c5d2SYehuda Sadeh osd = list_entry(slow_osds.next, struct ceph_osd, 10393d14c5d2SYehuda Sadeh o_keepalive_item); 10403d14c5d2SYehuda Sadeh list_del_init(&osd->o_keepalive_item); 10413d14c5d2SYehuda Sadeh ceph_con_keepalive(&osd->o_con); 10423d14c5d2SYehuda Sadeh } 10433d14c5d2SYehuda Sadeh 10443d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 1045f9d25199SAlex Elder __send_queued(osdc); 10463d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 10473d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 10483d14c5d2SYehuda Sadeh } 10493d14c5d2SYehuda Sadeh 10503d14c5d2SYehuda Sadeh static void handle_osds_timeout(struct work_struct *work) 10513d14c5d2SYehuda Sadeh { 10523d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 10533d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, 10543d14c5d2SYehuda Sadeh osds_timeout_work.work); 10553d14c5d2SYehuda Sadeh unsigned long delay = 10563d14c5d2SYehuda Sadeh osdc->client->options->osd_idle_ttl * HZ >> 2; 10573d14c5d2SYehuda Sadeh 10583d14c5d2SYehuda Sadeh dout("osds timeout\n"); 10593d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 1060aca420bcSSage Weil remove_old_osds(osdc); 10613d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 10623d14c5d2SYehuda Sadeh 10633d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 10643d14c5d2SYehuda Sadeh round_jiffies_relative(delay)); 10653d14c5d2SYehuda Sadeh } 10663d14c5d2SYehuda Sadeh 106725845472SSage Weil static void complete_request(struct ceph_osd_request *req) 106825845472SSage Weil { 106925845472SSage Weil if (req->r_safe_callback) 107025845472SSage Weil req->r_safe_callback(req, NULL); 107125845472SSage Weil complete_all(&req->r_safe_completion); /* fsync waiter */ 107225845472SSage Weil } 107325845472SSage Weil 10743d14c5d2SYehuda Sadeh /* 10753d14c5d2SYehuda Sadeh * handle osd op reply. either call the callback if it is specified, 10763d14c5d2SYehuda Sadeh * or do the completion to wake up the waiting thread. 10773d14c5d2SYehuda Sadeh */ 10783d14c5d2SYehuda Sadeh static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg, 10793d14c5d2SYehuda Sadeh struct ceph_connection *con) 10803d14c5d2SYehuda Sadeh { 10813d14c5d2SYehuda Sadeh struct ceph_osd_reply_head *rhead = msg->front.iov_base; 10823d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 10833d14c5d2SYehuda Sadeh u64 tid; 10843d14c5d2SYehuda Sadeh int numops, object_len, flags; 10853d14c5d2SYehuda Sadeh s32 result; 10863d14c5d2SYehuda Sadeh 10873d14c5d2SYehuda Sadeh tid = le64_to_cpu(msg->hdr.tid); 10883d14c5d2SYehuda Sadeh if (msg->front.iov_len < sizeof(*rhead)) 10893d14c5d2SYehuda Sadeh goto bad; 10903d14c5d2SYehuda Sadeh numops = le32_to_cpu(rhead->num_ops); 10913d14c5d2SYehuda Sadeh object_len = le32_to_cpu(rhead->object_len); 10923d14c5d2SYehuda Sadeh result = le32_to_cpu(rhead->result); 10933d14c5d2SYehuda Sadeh if (msg->front.iov_len != sizeof(*rhead) + object_len + 10943d14c5d2SYehuda Sadeh numops * sizeof(struct ceph_osd_op)) 10953d14c5d2SYehuda Sadeh goto bad; 10963d14c5d2SYehuda Sadeh dout("handle_reply %p tid %llu result %d\n", msg, tid, (int)result); 10973d14c5d2SYehuda Sadeh /* lookup */ 10983d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 10993d14c5d2SYehuda Sadeh req = __lookup_request(osdc, tid); 11003d14c5d2SYehuda Sadeh if (req == NULL) { 11013d14c5d2SYehuda Sadeh dout("handle_reply tid %llu dne\n", tid); 11023d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 11033d14c5d2SYehuda Sadeh return; 11043d14c5d2SYehuda Sadeh } 11053d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 11063d14c5d2SYehuda Sadeh flags = le32_to_cpu(rhead->flags); 11073d14c5d2SYehuda Sadeh 11083d14c5d2SYehuda Sadeh /* 11093d14c5d2SYehuda Sadeh * if this connection filled our message, drop our reference now, to 11103d14c5d2SYehuda Sadeh * avoid a (safe but slower) revoke later. 11113d14c5d2SYehuda Sadeh */ 11123d14c5d2SYehuda Sadeh if (req->r_con_filling_msg == con && req->r_reply == msg) { 11133d14c5d2SYehuda Sadeh dout(" dropping con_filling_msg ref %p\n", con); 11143d14c5d2SYehuda Sadeh req->r_con_filling_msg = NULL; 11150d47766fSSage Weil con->ops->put(con); 11163d14c5d2SYehuda Sadeh } 11173d14c5d2SYehuda Sadeh 11183d14c5d2SYehuda Sadeh if (!req->r_got_reply) { 111995c96174SEric Dumazet unsigned int bytes; 11203d14c5d2SYehuda Sadeh 11213d14c5d2SYehuda Sadeh req->r_result = le32_to_cpu(rhead->result); 11223d14c5d2SYehuda Sadeh bytes = le32_to_cpu(msg->hdr.data_len); 11233d14c5d2SYehuda Sadeh dout("handle_reply result %d bytes %d\n", req->r_result, 11243d14c5d2SYehuda Sadeh bytes); 11253d14c5d2SYehuda Sadeh if (req->r_result == 0) 11263d14c5d2SYehuda Sadeh req->r_result = bytes; 11273d14c5d2SYehuda Sadeh 11283d14c5d2SYehuda Sadeh /* in case this is a write and we need to replay, */ 11293d14c5d2SYehuda Sadeh req->r_reassert_version = rhead->reassert_version; 11303d14c5d2SYehuda Sadeh 11313d14c5d2SYehuda Sadeh req->r_got_reply = 1; 11323d14c5d2SYehuda Sadeh } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) { 11333d14c5d2SYehuda Sadeh dout("handle_reply tid %llu dup ack\n", tid); 11343d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 11353d14c5d2SYehuda Sadeh goto done; 11363d14c5d2SYehuda Sadeh } 11373d14c5d2SYehuda Sadeh 11383d14c5d2SYehuda Sadeh dout("handle_reply tid %llu flags %d\n", tid, flags); 11393d14c5d2SYehuda Sadeh 1140a40c4f10SYehuda Sadeh if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK)) 1141a40c4f10SYehuda Sadeh __register_linger_request(osdc, req); 1142a40c4f10SYehuda Sadeh 11433d14c5d2SYehuda Sadeh /* either this is a read, or we got the safe response */ 11443d14c5d2SYehuda Sadeh if (result < 0 || 11453d14c5d2SYehuda Sadeh (flags & CEPH_OSD_FLAG_ONDISK) || 11463d14c5d2SYehuda Sadeh ((flags & CEPH_OSD_FLAG_WRITE) == 0)) 11473d14c5d2SYehuda Sadeh __unregister_request(osdc, req); 11483d14c5d2SYehuda Sadeh 11493d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 11503d14c5d2SYehuda Sadeh 11513d14c5d2SYehuda Sadeh if (req->r_callback) 11523d14c5d2SYehuda Sadeh req->r_callback(req, msg); 11533d14c5d2SYehuda Sadeh else 11543d14c5d2SYehuda Sadeh complete_all(&req->r_completion); 11553d14c5d2SYehuda Sadeh 115625845472SSage Weil if (flags & CEPH_OSD_FLAG_ONDISK) 115725845472SSage Weil complete_request(req); 11583d14c5d2SYehuda Sadeh 11593d14c5d2SYehuda Sadeh done: 1160a40c4f10SYehuda Sadeh dout("req=%p req->r_linger=%d\n", req, req->r_linger); 11613d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 11623d14c5d2SYehuda Sadeh return; 11633d14c5d2SYehuda Sadeh 11643d14c5d2SYehuda Sadeh bad: 11653d14c5d2SYehuda Sadeh pr_err("corrupt osd_op_reply got %d %d expected %d\n", 11663d14c5d2SYehuda Sadeh (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len), 11673d14c5d2SYehuda Sadeh (int)sizeof(*rhead)); 11683d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 11693d14c5d2SYehuda Sadeh } 11703d14c5d2SYehuda Sadeh 11716f6c7006SSage Weil static void reset_changed_osds(struct ceph_osd_client *osdc) 11723d14c5d2SYehuda Sadeh { 11733d14c5d2SYehuda Sadeh struct rb_node *p, *n; 11743d14c5d2SYehuda Sadeh 11753d14c5d2SYehuda Sadeh for (p = rb_first(&osdc->osds); p; p = n) { 11766f6c7006SSage Weil struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node); 11773d14c5d2SYehuda Sadeh 11783d14c5d2SYehuda Sadeh n = rb_next(p); 11793d14c5d2SYehuda Sadeh if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) || 11803d14c5d2SYehuda Sadeh memcmp(&osd->o_con.peer_addr, 11813d14c5d2SYehuda Sadeh ceph_osd_addr(osdc->osdmap, 11823d14c5d2SYehuda Sadeh osd->o_osd), 11833d14c5d2SYehuda Sadeh sizeof(struct ceph_entity_addr)) != 0) 11843d14c5d2SYehuda Sadeh __reset_osd(osdc, osd); 11853d14c5d2SYehuda Sadeh } 11863d14c5d2SYehuda Sadeh } 11873d14c5d2SYehuda Sadeh 11883d14c5d2SYehuda Sadeh /* 11896f6c7006SSage Weil * Requeue requests whose mapping to an OSD has changed. If requests map to 11906f6c7006SSage Weil * no osd, request a new map. 11913d14c5d2SYehuda Sadeh * 1192e6d50f67SAlex Elder * Caller should hold map_sem for read. 11933d14c5d2SYehuda Sadeh */ 119438d6453cSSage Weil static void kick_requests(struct ceph_osd_client *osdc, int force_resend) 11953d14c5d2SYehuda Sadeh { 1196a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 11976f6c7006SSage Weil struct rb_node *p; 11986f6c7006SSage Weil int needmap = 0; 11996f6c7006SSage Weil int err; 12003d14c5d2SYehuda Sadeh 120138d6453cSSage Weil dout("kick_requests %s\n", force_resend ? " (force resend)" : ""); 12023d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 12036194ea89SSage Weil for (p = rb_first(&osdc->requests); p; ) { 12046f6c7006SSage Weil req = rb_entry(p, struct ceph_osd_request, r_node); 12056194ea89SSage Weil p = rb_next(p); 1206ab60b16dSAlex Elder 1207ab60b16dSAlex Elder /* 1208ab60b16dSAlex Elder * For linger requests that have not yet been 1209ab60b16dSAlex Elder * registered, move them to the linger list; they'll 1210ab60b16dSAlex Elder * be sent to the osd in the loop below. Unregister 1211ab60b16dSAlex Elder * the request before re-registering it as a linger 1212ab60b16dSAlex Elder * request to ensure the __map_request() below 1213ab60b16dSAlex Elder * will decide it needs to be sent. 1214ab60b16dSAlex Elder */ 1215ab60b16dSAlex Elder if (req->r_linger && list_empty(&req->r_linger_item)) { 1216ab60b16dSAlex Elder dout("%p tid %llu restart on osd%d\n", 1217ab60b16dSAlex Elder req, req->r_tid, 1218ab60b16dSAlex Elder req->r_osd ? req->r_osd->o_osd : -1); 1219ab60b16dSAlex Elder __unregister_request(osdc, req); 1220ab60b16dSAlex Elder __register_linger_request(osdc, req); 1221ab60b16dSAlex Elder continue; 1222ab60b16dSAlex Elder } 1223ab60b16dSAlex Elder 122438d6453cSSage Weil err = __map_request(osdc, req, force_resend); 12256f6c7006SSage Weil if (err < 0) 12266f6c7006SSage Weil continue; /* error */ 12276f6c7006SSage Weil if (req->r_osd == NULL) { 12286f6c7006SSage Weil dout("%p tid %llu maps to no osd\n", req, req->r_tid); 12296f6c7006SSage Weil needmap++; /* request a newer map */ 12306f6c7006SSage Weil } else if (err > 0) { 12316194ea89SSage Weil if (!req->r_linger) { 12326194ea89SSage Weil dout("%p tid %llu requeued on osd%d\n", req, 12336194ea89SSage Weil req->r_tid, 12346f6c7006SSage Weil req->r_osd ? req->r_osd->o_osd : -1); 12356f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 12366f6c7006SSage Weil } 12376f6c7006SSage Weil } 12386194ea89SSage Weil } 1239a40c4f10SYehuda Sadeh 1240a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osdc->req_linger, 1241a40c4f10SYehuda Sadeh r_linger_item) { 1242a40c4f10SYehuda Sadeh dout("linger req=%p req->r_osd=%p\n", req, req->r_osd); 1243a40c4f10SYehuda Sadeh 124438d6453cSSage Weil err = __map_request(osdc, req, force_resend); 1245ab60b16dSAlex Elder dout("__map_request returned %d\n", err); 1246a40c4f10SYehuda Sadeh if (err == 0) 1247a40c4f10SYehuda Sadeh continue; /* no change and no osd was specified */ 1248a40c4f10SYehuda Sadeh if (err < 0) 1249a40c4f10SYehuda Sadeh continue; /* hrm! */ 1250a40c4f10SYehuda Sadeh if (req->r_osd == NULL) { 1251a40c4f10SYehuda Sadeh dout("tid %llu maps to no valid osd\n", req->r_tid); 1252a40c4f10SYehuda Sadeh needmap++; /* request a newer map */ 1253a40c4f10SYehuda Sadeh continue; 1254a40c4f10SYehuda Sadeh } 1255a40c4f10SYehuda Sadeh 1256a40c4f10SYehuda Sadeh dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid, 1257a40c4f10SYehuda Sadeh req->r_osd ? req->r_osd->o_osd : -1); 1258a40c4f10SYehuda Sadeh __register_request(osdc, req); 1259c89ce05eSAlex Elder __unregister_linger_request(osdc, req); 1260a40c4f10SYehuda Sadeh } 12613d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 12623d14c5d2SYehuda Sadeh 12633d14c5d2SYehuda Sadeh if (needmap) { 12643d14c5d2SYehuda Sadeh dout("%d requests for down osds, need new map\n", needmap); 12653d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 12663d14c5d2SYehuda Sadeh } 1267e6d50f67SAlex Elder reset_changed_osds(osdc); 12683d14c5d2SYehuda Sadeh } 12696f6c7006SSage Weil 12706f6c7006SSage Weil 12713d14c5d2SYehuda Sadeh /* 12723d14c5d2SYehuda Sadeh * Process updated osd map. 12733d14c5d2SYehuda Sadeh * 12743d14c5d2SYehuda Sadeh * The message contains any number of incremental and full maps, normally 12753d14c5d2SYehuda Sadeh * indicating some sort of topology change in the cluster. Kick requests 12763d14c5d2SYehuda Sadeh * off to different OSDs as needed. 12773d14c5d2SYehuda Sadeh */ 12783d14c5d2SYehuda Sadeh void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg) 12793d14c5d2SYehuda Sadeh { 12803d14c5d2SYehuda Sadeh void *p, *end, *next; 12813d14c5d2SYehuda Sadeh u32 nr_maps, maplen; 12823d14c5d2SYehuda Sadeh u32 epoch; 12833d14c5d2SYehuda Sadeh struct ceph_osdmap *newmap = NULL, *oldmap; 12843d14c5d2SYehuda Sadeh int err; 12853d14c5d2SYehuda Sadeh struct ceph_fsid fsid; 12863d14c5d2SYehuda Sadeh 12873d14c5d2SYehuda Sadeh dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0); 12883d14c5d2SYehuda Sadeh p = msg->front.iov_base; 12893d14c5d2SYehuda Sadeh end = p + msg->front.iov_len; 12903d14c5d2SYehuda Sadeh 12913d14c5d2SYehuda Sadeh /* verify fsid */ 12923d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, sizeof(fsid), bad); 12933d14c5d2SYehuda Sadeh ceph_decode_copy(&p, &fsid, sizeof(fsid)); 12943d14c5d2SYehuda Sadeh if (ceph_check_fsid(osdc->client, &fsid) < 0) 12953d14c5d2SYehuda Sadeh return; 12963d14c5d2SYehuda Sadeh 12973d14c5d2SYehuda Sadeh down_write(&osdc->map_sem); 12983d14c5d2SYehuda Sadeh 12993d14c5d2SYehuda Sadeh /* incremental maps */ 13003d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 13013d14c5d2SYehuda Sadeh dout(" %d inc maps\n", nr_maps); 13023d14c5d2SYehuda Sadeh while (nr_maps > 0) { 13033d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 13043d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 13053d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 13063d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 13073d14c5d2SYehuda Sadeh next = p + maplen; 13083d14c5d2SYehuda Sadeh if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) { 13093d14c5d2SYehuda Sadeh dout("applying incremental map %u len %d\n", 13103d14c5d2SYehuda Sadeh epoch, maplen); 13113d14c5d2SYehuda Sadeh newmap = osdmap_apply_incremental(&p, next, 13123d14c5d2SYehuda Sadeh osdc->osdmap, 131315d9882cSAlex Elder &osdc->client->msgr); 13143d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 13153d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 13163d14c5d2SYehuda Sadeh goto bad; 13173d14c5d2SYehuda Sadeh } 13183d14c5d2SYehuda Sadeh BUG_ON(!newmap); 13193d14c5d2SYehuda Sadeh if (newmap != osdc->osdmap) { 13203d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 13213d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 13223d14c5d2SYehuda Sadeh } 132338d6453cSSage Weil kick_requests(osdc, 0); 13243d14c5d2SYehuda Sadeh } else { 13253d14c5d2SYehuda Sadeh dout("ignoring incremental map %u len %d\n", 13263d14c5d2SYehuda Sadeh epoch, maplen); 13273d14c5d2SYehuda Sadeh } 13283d14c5d2SYehuda Sadeh p = next; 13293d14c5d2SYehuda Sadeh nr_maps--; 13303d14c5d2SYehuda Sadeh } 13313d14c5d2SYehuda Sadeh if (newmap) 13323d14c5d2SYehuda Sadeh goto done; 13333d14c5d2SYehuda Sadeh 13343d14c5d2SYehuda Sadeh /* full maps */ 13353d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 13363d14c5d2SYehuda Sadeh dout(" %d full maps\n", nr_maps); 13373d14c5d2SYehuda Sadeh while (nr_maps) { 13383d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 13393d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 13403d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 13413d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 13423d14c5d2SYehuda Sadeh if (nr_maps > 1) { 13433d14c5d2SYehuda Sadeh dout("skipping non-latest full map %u len %d\n", 13443d14c5d2SYehuda Sadeh epoch, maplen); 13453d14c5d2SYehuda Sadeh } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) { 13463d14c5d2SYehuda Sadeh dout("skipping full map %u len %d, " 13473d14c5d2SYehuda Sadeh "older than our %u\n", epoch, maplen, 13483d14c5d2SYehuda Sadeh osdc->osdmap->epoch); 13493d14c5d2SYehuda Sadeh } else { 135038d6453cSSage Weil int skipped_map = 0; 135138d6453cSSage Weil 13523d14c5d2SYehuda Sadeh dout("taking full map %u len %d\n", epoch, maplen); 13533d14c5d2SYehuda Sadeh newmap = osdmap_decode(&p, p+maplen); 13543d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 13553d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 13563d14c5d2SYehuda Sadeh goto bad; 13573d14c5d2SYehuda Sadeh } 13583d14c5d2SYehuda Sadeh BUG_ON(!newmap); 13593d14c5d2SYehuda Sadeh oldmap = osdc->osdmap; 13603d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 136138d6453cSSage Weil if (oldmap) { 136238d6453cSSage Weil if (oldmap->epoch + 1 < newmap->epoch) 136338d6453cSSage Weil skipped_map = 1; 13643d14c5d2SYehuda Sadeh ceph_osdmap_destroy(oldmap); 136538d6453cSSage Weil } 136638d6453cSSage Weil kick_requests(osdc, skipped_map); 13673d14c5d2SYehuda Sadeh } 13683d14c5d2SYehuda Sadeh p += maplen; 13693d14c5d2SYehuda Sadeh nr_maps--; 13703d14c5d2SYehuda Sadeh } 13713d14c5d2SYehuda Sadeh 13723d14c5d2SYehuda Sadeh done: 13733d14c5d2SYehuda Sadeh downgrade_write(&osdc->map_sem); 13743d14c5d2SYehuda Sadeh ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch); 1375cd634fb6SSage Weil 1376cd634fb6SSage Weil /* 1377cd634fb6SSage Weil * subscribe to subsequent osdmap updates if full to ensure 1378cd634fb6SSage Weil * we find out when we are no longer full and stop returning 1379cd634fb6SSage Weil * ENOSPC. 1380cd634fb6SSage Weil */ 1381cd634fb6SSage Weil if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL)) 1382cd634fb6SSage Weil ceph_monc_request_next_osdmap(&osdc->client->monc); 1383cd634fb6SSage Weil 1384f9d25199SAlex Elder mutex_lock(&osdc->request_mutex); 1385f9d25199SAlex Elder __send_queued(osdc); 1386f9d25199SAlex Elder mutex_unlock(&osdc->request_mutex); 13873d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 13883d14c5d2SYehuda Sadeh wake_up_all(&osdc->client->auth_wq); 13893d14c5d2SYehuda Sadeh return; 13903d14c5d2SYehuda Sadeh 13913d14c5d2SYehuda Sadeh bad: 13923d14c5d2SYehuda Sadeh pr_err("osdc handle_map corrupt msg\n"); 13933d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 13943d14c5d2SYehuda Sadeh up_write(&osdc->map_sem); 13953d14c5d2SYehuda Sadeh return; 13963d14c5d2SYehuda Sadeh } 13973d14c5d2SYehuda Sadeh 13983d14c5d2SYehuda Sadeh /* 1399a40c4f10SYehuda Sadeh * watch/notify callback event infrastructure 1400a40c4f10SYehuda Sadeh * 1401a40c4f10SYehuda Sadeh * These callbacks are used both for watch and notify operations. 1402a40c4f10SYehuda Sadeh */ 1403a40c4f10SYehuda Sadeh static void __release_event(struct kref *kref) 1404a40c4f10SYehuda Sadeh { 1405a40c4f10SYehuda Sadeh struct ceph_osd_event *event = 1406a40c4f10SYehuda Sadeh container_of(kref, struct ceph_osd_event, kref); 1407a40c4f10SYehuda Sadeh 1408a40c4f10SYehuda Sadeh dout("__release_event %p\n", event); 1409a40c4f10SYehuda Sadeh kfree(event); 1410a40c4f10SYehuda Sadeh } 1411a40c4f10SYehuda Sadeh 1412a40c4f10SYehuda Sadeh static void get_event(struct ceph_osd_event *event) 1413a40c4f10SYehuda Sadeh { 1414a40c4f10SYehuda Sadeh kref_get(&event->kref); 1415a40c4f10SYehuda Sadeh } 1416a40c4f10SYehuda Sadeh 1417a40c4f10SYehuda Sadeh void ceph_osdc_put_event(struct ceph_osd_event *event) 1418a40c4f10SYehuda Sadeh { 1419a40c4f10SYehuda Sadeh kref_put(&event->kref, __release_event); 1420a40c4f10SYehuda Sadeh } 1421a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_put_event); 1422a40c4f10SYehuda Sadeh 1423a40c4f10SYehuda Sadeh static void __insert_event(struct ceph_osd_client *osdc, 1424a40c4f10SYehuda Sadeh struct ceph_osd_event *new) 1425a40c4f10SYehuda Sadeh { 1426a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 1427a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 1428a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 1429a40c4f10SYehuda Sadeh 1430a40c4f10SYehuda Sadeh while (*p) { 1431a40c4f10SYehuda Sadeh parent = *p; 1432a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 1433a40c4f10SYehuda Sadeh if (new->cookie < event->cookie) 1434a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 1435a40c4f10SYehuda Sadeh else if (new->cookie > event->cookie) 1436a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 1437a40c4f10SYehuda Sadeh else 1438a40c4f10SYehuda Sadeh BUG(); 1439a40c4f10SYehuda Sadeh } 1440a40c4f10SYehuda Sadeh 1441a40c4f10SYehuda Sadeh rb_link_node(&new->node, parent, p); 1442a40c4f10SYehuda Sadeh rb_insert_color(&new->node, &osdc->event_tree); 1443a40c4f10SYehuda Sadeh } 1444a40c4f10SYehuda Sadeh 1445a40c4f10SYehuda Sadeh static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc, 1446a40c4f10SYehuda Sadeh u64 cookie) 1447a40c4f10SYehuda Sadeh { 1448a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 1449a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 1450a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 1451a40c4f10SYehuda Sadeh 1452a40c4f10SYehuda Sadeh while (*p) { 1453a40c4f10SYehuda Sadeh parent = *p; 1454a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 1455a40c4f10SYehuda Sadeh if (cookie < event->cookie) 1456a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 1457a40c4f10SYehuda Sadeh else if (cookie > event->cookie) 1458a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 1459a40c4f10SYehuda Sadeh else 1460a40c4f10SYehuda Sadeh return event; 1461a40c4f10SYehuda Sadeh } 1462a40c4f10SYehuda Sadeh return NULL; 1463a40c4f10SYehuda Sadeh } 1464a40c4f10SYehuda Sadeh 1465a40c4f10SYehuda Sadeh static void __remove_event(struct ceph_osd_event *event) 1466a40c4f10SYehuda Sadeh { 1467a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 1468a40c4f10SYehuda Sadeh 1469a40c4f10SYehuda Sadeh if (!RB_EMPTY_NODE(&event->node)) { 1470a40c4f10SYehuda Sadeh dout("__remove_event removed %p\n", event); 1471a40c4f10SYehuda Sadeh rb_erase(&event->node, &osdc->event_tree); 1472a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 1473a40c4f10SYehuda Sadeh } else { 1474a40c4f10SYehuda Sadeh dout("__remove_event didn't remove %p\n", event); 1475a40c4f10SYehuda Sadeh } 1476a40c4f10SYehuda Sadeh } 1477a40c4f10SYehuda Sadeh 1478a40c4f10SYehuda Sadeh int ceph_osdc_create_event(struct ceph_osd_client *osdc, 1479a40c4f10SYehuda Sadeh void (*event_cb)(u64, u64, u8, void *), 1480a40c4f10SYehuda Sadeh int one_shot, void *data, 1481a40c4f10SYehuda Sadeh struct ceph_osd_event **pevent) 1482a40c4f10SYehuda Sadeh { 1483a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 1484a40c4f10SYehuda Sadeh 1485a40c4f10SYehuda Sadeh event = kmalloc(sizeof(*event), GFP_NOIO); 1486a40c4f10SYehuda Sadeh if (!event) 1487a40c4f10SYehuda Sadeh return -ENOMEM; 1488a40c4f10SYehuda Sadeh 1489a40c4f10SYehuda Sadeh dout("create_event %p\n", event); 1490a40c4f10SYehuda Sadeh event->cb = event_cb; 1491a40c4f10SYehuda Sadeh event->one_shot = one_shot; 1492a40c4f10SYehuda Sadeh event->data = data; 1493a40c4f10SYehuda Sadeh event->osdc = osdc; 1494a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&event->osd_node); 14953ee5234dSAlex Elder RB_CLEAR_NODE(&event->node); 1496a40c4f10SYehuda Sadeh kref_init(&event->kref); /* one ref for us */ 1497a40c4f10SYehuda Sadeh kref_get(&event->kref); /* one ref for the caller */ 1498a40c4f10SYehuda Sadeh init_completion(&event->completion); 1499a40c4f10SYehuda Sadeh 1500a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 1501a40c4f10SYehuda Sadeh event->cookie = ++osdc->event_count; 1502a40c4f10SYehuda Sadeh __insert_event(osdc, event); 1503a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 1504a40c4f10SYehuda Sadeh 1505a40c4f10SYehuda Sadeh *pevent = event; 1506a40c4f10SYehuda Sadeh return 0; 1507a40c4f10SYehuda Sadeh } 1508a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_create_event); 1509a40c4f10SYehuda Sadeh 1510a40c4f10SYehuda Sadeh void ceph_osdc_cancel_event(struct ceph_osd_event *event) 1511a40c4f10SYehuda Sadeh { 1512a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 1513a40c4f10SYehuda Sadeh 1514a40c4f10SYehuda Sadeh dout("cancel_event %p\n", event); 1515a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 1516a40c4f10SYehuda Sadeh __remove_event(event); 1517a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 1518a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); /* caller's */ 1519a40c4f10SYehuda Sadeh } 1520a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_cancel_event); 1521a40c4f10SYehuda Sadeh 1522a40c4f10SYehuda Sadeh 1523a40c4f10SYehuda Sadeh static void do_event_work(struct work_struct *work) 1524a40c4f10SYehuda Sadeh { 1525a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work = 1526a40c4f10SYehuda Sadeh container_of(work, struct ceph_osd_event_work, work); 1527a40c4f10SYehuda Sadeh struct ceph_osd_event *event = event_work->event; 1528a40c4f10SYehuda Sadeh u64 ver = event_work->ver; 1529a40c4f10SYehuda Sadeh u64 notify_id = event_work->notify_id; 1530a40c4f10SYehuda Sadeh u8 opcode = event_work->opcode; 1531a40c4f10SYehuda Sadeh 1532a40c4f10SYehuda Sadeh dout("do_event_work completing %p\n", event); 1533a40c4f10SYehuda Sadeh event->cb(ver, notify_id, opcode, event->data); 1534a40c4f10SYehuda Sadeh complete(&event->completion); 1535a40c4f10SYehuda Sadeh dout("do_event_work completed %p\n", event); 1536a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 1537a40c4f10SYehuda Sadeh kfree(event_work); 1538a40c4f10SYehuda Sadeh } 1539a40c4f10SYehuda Sadeh 1540a40c4f10SYehuda Sadeh 1541a40c4f10SYehuda Sadeh /* 1542a40c4f10SYehuda Sadeh * Process osd watch notifications 1543a40c4f10SYehuda Sadeh */ 1544a40c4f10SYehuda Sadeh void handle_watch_notify(struct ceph_osd_client *osdc, struct ceph_msg *msg) 1545a40c4f10SYehuda Sadeh { 1546a40c4f10SYehuda Sadeh void *p, *end; 1547a40c4f10SYehuda Sadeh u8 proto_ver; 1548a40c4f10SYehuda Sadeh u64 cookie, ver, notify_id; 1549a40c4f10SYehuda Sadeh u8 opcode; 1550a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 1551a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work; 1552a40c4f10SYehuda Sadeh 1553a40c4f10SYehuda Sadeh p = msg->front.iov_base; 1554a40c4f10SYehuda Sadeh end = p + msg->front.iov_len; 1555a40c4f10SYehuda Sadeh 1556a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, proto_ver, bad); 1557a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, opcode, bad); 1558a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, cookie, bad); 1559a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, ver, bad); 1560a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, notify_id, bad); 1561a40c4f10SYehuda Sadeh 1562a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 1563a40c4f10SYehuda Sadeh event = __find_event(osdc, cookie); 1564a40c4f10SYehuda Sadeh if (event) { 1565a40c4f10SYehuda Sadeh get_event(event); 1566a40c4f10SYehuda Sadeh if (event->one_shot) 1567a40c4f10SYehuda Sadeh __remove_event(event); 1568a40c4f10SYehuda Sadeh } 1569a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 1570a40c4f10SYehuda Sadeh dout("handle_watch_notify cookie %lld ver %lld event %p\n", 1571a40c4f10SYehuda Sadeh cookie, ver, event); 1572a40c4f10SYehuda Sadeh if (event) { 1573a40c4f10SYehuda Sadeh event_work = kmalloc(sizeof(*event_work), GFP_NOIO); 1574a40c4f10SYehuda Sadeh if (!event_work) { 1575a40c4f10SYehuda Sadeh dout("ERROR: could not allocate event_work\n"); 1576a40c4f10SYehuda Sadeh goto done_err; 1577a40c4f10SYehuda Sadeh } 15786b0ae409SMariusz Kozlowski INIT_WORK(&event_work->work, do_event_work); 1579a40c4f10SYehuda Sadeh event_work->event = event; 1580a40c4f10SYehuda Sadeh event_work->ver = ver; 1581a40c4f10SYehuda Sadeh event_work->notify_id = notify_id; 1582a40c4f10SYehuda Sadeh event_work->opcode = opcode; 1583a40c4f10SYehuda Sadeh if (!queue_work(osdc->notify_wq, &event_work->work)) { 1584a40c4f10SYehuda Sadeh dout("WARNING: failed to queue notify event work\n"); 1585a40c4f10SYehuda Sadeh goto done_err; 1586a40c4f10SYehuda Sadeh } 1587a40c4f10SYehuda Sadeh } 1588a40c4f10SYehuda Sadeh 1589a40c4f10SYehuda Sadeh return; 1590a40c4f10SYehuda Sadeh 1591a40c4f10SYehuda Sadeh done_err: 1592a40c4f10SYehuda Sadeh complete(&event->completion); 1593a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 1594a40c4f10SYehuda Sadeh return; 1595a40c4f10SYehuda Sadeh 1596a40c4f10SYehuda Sadeh bad: 1597a40c4f10SYehuda Sadeh pr_err("osdc handle_watch_notify corrupt msg\n"); 1598a40c4f10SYehuda Sadeh return; 1599a40c4f10SYehuda Sadeh } 1600a40c4f10SYehuda Sadeh 1601a40c4f10SYehuda Sadeh int ceph_osdc_wait_event(struct ceph_osd_event *event, unsigned long timeout) 1602a40c4f10SYehuda Sadeh { 1603a40c4f10SYehuda Sadeh int err; 1604a40c4f10SYehuda Sadeh 1605a40c4f10SYehuda Sadeh dout("wait_event %p\n", event); 1606a40c4f10SYehuda Sadeh err = wait_for_completion_interruptible_timeout(&event->completion, 1607a40c4f10SYehuda Sadeh timeout * HZ); 1608a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 1609a40c4f10SYehuda Sadeh if (err > 0) 1610a40c4f10SYehuda Sadeh err = 0; 1611a40c4f10SYehuda Sadeh dout("wait_event %p returns %d\n", event, err); 1612a40c4f10SYehuda Sadeh return err; 1613a40c4f10SYehuda Sadeh } 1614a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_wait_event); 1615a40c4f10SYehuda Sadeh 1616a40c4f10SYehuda Sadeh /* 16173d14c5d2SYehuda Sadeh * Register request, send initial attempt. 16183d14c5d2SYehuda Sadeh */ 16193d14c5d2SYehuda Sadeh int ceph_osdc_start_request(struct ceph_osd_client *osdc, 16203d14c5d2SYehuda Sadeh struct ceph_osd_request *req, 16213d14c5d2SYehuda Sadeh bool nofail) 16223d14c5d2SYehuda Sadeh { 16233d14c5d2SYehuda Sadeh int rc = 0; 16243d14c5d2SYehuda Sadeh 16253d14c5d2SYehuda Sadeh req->r_request->pages = req->r_pages; 16263d14c5d2SYehuda Sadeh req->r_request->nr_pages = req->r_num_pages; 16273d14c5d2SYehuda Sadeh #ifdef CONFIG_BLOCK 16283d14c5d2SYehuda Sadeh req->r_request->bio = req->r_bio; 16293d14c5d2SYehuda Sadeh #endif 1630c885837fSAlex Elder req->r_request->trail = &req->r_trail; 16313d14c5d2SYehuda Sadeh 16323d14c5d2SYehuda Sadeh register_request(osdc, req); 16333d14c5d2SYehuda Sadeh 16343d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 16353d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 16363d14c5d2SYehuda Sadeh /* 16373d14c5d2SYehuda Sadeh * a racing kick_requests() may have sent the message for us 16383d14c5d2SYehuda Sadeh * while we dropped request_mutex above, so only send now if 16393d14c5d2SYehuda Sadeh * the request still han't been touched yet. 16403d14c5d2SYehuda Sadeh */ 16413d14c5d2SYehuda Sadeh if (req->r_sent == 0) { 164238d6453cSSage Weil rc = __map_request(osdc, req, 0); 16439d6fcb08SSage Weil if (rc < 0) { 16449d6fcb08SSage Weil if (nofail) { 16459d6fcb08SSage Weil dout("osdc_start_request failed map, " 16469d6fcb08SSage Weil " will retry %lld\n", req->r_tid); 16479d6fcb08SSage Weil rc = 0; 16489d6fcb08SSage Weil } 1649234af26fSDan Carpenter goto out_unlock; 16509d6fcb08SSage Weil } 16516f6c7006SSage Weil if (req->r_osd == NULL) { 16526f6c7006SSage Weil dout("send_request %p no up osds in pg\n", req); 16536f6c7006SSage Weil ceph_monc_request_next_osdmap(&osdc->client->monc); 16546f6c7006SSage Weil } else { 165556e925b6SSage Weil __send_request(osdc, req); 165656e925b6SSage Weil } 16573d14c5d2SYehuda Sadeh rc = 0; 16586f6c7006SSage Weil } 1659234af26fSDan Carpenter 1660234af26fSDan Carpenter out_unlock: 16613d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 16623d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 16633d14c5d2SYehuda Sadeh return rc; 16643d14c5d2SYehuda Sadeh } 16653d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_start_request); 16663d14c5d2SYehuda Sadeh 16673d14c5d2SYehuda Sadeh /* 16683d14c5d2SYehuda Sadeh * wait for a request to complete 16693d14c5d2SYehuda Sadeh */ 16703d14c5d2SYehuda Sadeh int ceph_osdc_wait_request(struct ceph_osd_client *osdc, 16713d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 16723d14c5d2SYehuda Sadeh { 16733d14c5d2SYehuda Sadeh int rc; 16743d14c5d2SYehuda Sadeh 16753d14c5d2SYehuda Sadeh rc = wait_for_completion_interruptible(&req->r_completion); 16763d14c5d2SYehuda Sadeh if (rc < 0) { 16773d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 16783d14c5d2SYehuda Sadeh __cancel_request(req); 16793d14c5d2SYehuda Sadeh __unregister_request(osdc, req); 16803d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 168125845472SSage Weil complete_request(req); 16823d14c5d2SYehuda Sadeh dout("wait_request tid %llu canceled/timed out\n", req->r_tid); 16833d14c5d2SYehuda Sadeh return rc; 16843d14c5d2SYehuda Sadeh } 16853d14c5d2SYehuda Sadeh 16863d14c5d2SYehuda Sadeh dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result); 16873d14c5d2SYehuda Sadeh return req->r_result; 16883d14c5d2SYehuda Sadeh } 16893d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_wait_request); 16903d14c5d2SYehuda Sadeh 16913d14c5d2SYehuda Sadeh /* 16923d14c5d2SYehuda Sadeh * sync - wait for all in-flight requests to flush. avoid starvation. 16933d14c5d2SYehuda Sadeh */ 16943d14c5d2SYehuda Sadeh void ceph_osdc_sync(struct ceph_osd_client *osdc) 16953d14c5d2SYehuda Sadeh { 16963d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 16973d14c5d2SYehuda Sadeh u64 last_tid, next_tid = 0; 16983d14c5d2SYehuda Sadeh 16993d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 17003d14c5d2SYehuda Sadeh last_tid = osdc->last_tid; 17013d14c5d2SYehuda Sadeh while (1) { 17023d14c5d2SYehuda Sadeh req = __lookup_request_ge(osdc, next_tid); 17033d14c5d2SYehuda Sadeh if (!req) 17043d14c5d2SYehuda Sadeh break; 17053d14c5d2SYehuda Sadeh if (req->r_tid > last_tid) 17063d14c5d2SYehuda Sadeh break; 17073d14c5d2SYehuda Sadeh 17083d14c5d2SYehuda Sadeh next_tid = req->r_tid + 1; 17093d14c5d2SYehuda Sadeh if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0) 17103d14c5d2SYehuda Sadeh continue; 17113d14c5d2SYehuda Sadeh 17123d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 17133d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 17143d14c5d2SYehuda Sadeh dout("sync waiting on tid %llu (last is %llu)\n", 17153d14c5d2SYehuda Sadeh req->r_tid, last_tid); 17163d14c5d2SYehuda Sadeh wait_for_completion(&req->r_safe_completion); 17173d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 17183d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 17193d14c5d2SYehuda Sadeh } 17203d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 17213d14c5d2SYehuda Sadeh dout("sync done (thru tid %llu)\n", last_tid); 17223d14c5d2SYehuda Sadeh } 17233d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_sync); 17243d14c5d2SYehuda Sadeh 17253d14c5d2SYehuda Sadeh /* 17263d14c5d2SYehuda Sadeh * init, shutdown 17273d14c5d2SYehuda Sadeh */ 17283d14c5d2SYehuda Sadeh int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client) 17293d14c5d2SYehuda Sadeh { 17303d14c5d2SYehuda Sadeh int err; 17313d14c5d2SYehuda Sadeh 17323d14c5d2SYehuda Sadeh dout("init\n"); 17333d14c5d2SYehuda Sadeh osdc->client = client; 17343d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 17353d14c5d2SYehuda Sadeh init_rwsem(&osdc->map_sem); 17363d14c5d2SYehuda Sadeh init_completion(&osdc->map_waiters); 17373d14c5d2SYehuda Sadeh osdc->last_requested_map = 0; 17383d14c5d2SYehuda Sadeh mutex_init(&osdc->request_mutex); 17393d14c5d2SYehuda Sadeh osdc->last_tid = 0; 17403d14c5d2SYehuda Sadeh osdc->osds = RB_ROOT; 17413d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->osd_lru); 17423d14c5d2SYehuda Sadeh osdc->requests = RB_ROOT; 17433d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_lru); 17446f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_unsent); 17456f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_notarget); 1746a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_linger); 17473d14c5d2SYehuda Sadeh osdc->num_requests = 0; 17483d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout); 17493d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout); 1750a40c4f10SYehuda Sadeh spin_lock_init(&osdc->event_lock); 1751a40c4f10SYehuda Sadeh osdc->event_tree = RB_ROOT; 1752a40c4f10SYehuda Sadeh osdc->event_count = 0; 17533d14c5d2SYehuda Sadeh 17543d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 17553d14c5d2SYehuda Sadeh round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ)); 17563d14c5d2SYehuda Sadeh 17573d14c5d2SYehuda Sadeh err = -ENOMEM; 17583d14c5d2SYehuda Sadeh osdc->req_mempool = mempool_create_kmalloc_pool(10, 17593d14c5d2SYehuda Sadeh sizeof(struct ceph_osd_request)); 17603d14c5d2SYehuda Sadeh if (!osdc->req_mempool) 17613d14c5d2SYehuda Sadeh goto out; 17623d14c5d2SYehuda Sadeh 1763d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP, 1764d50b409fSSage Weil OSD_OP_FRONT_LEN, 10, true, 17653d14c5d2SYehuda Sadeh "osd_op"); 17663d14c5d2SYehuda Sadeh if (err < 0) 17673d14c5d2SYehuda Sadeh goto out_mempool; 1768d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY, 17693d14c5d2SYehuda Sadeh OSD_OPREPLY_FRONT_LEN, 10, true, 17703d14c5d2SYehuda Sadeh "osd_op_reply"); 17713d14c5d2SYehuda Sadeh if (err < 0) 17723d14c5d2SYehuda Sadeh goto out_msgpool; 1773a40c4f10SYehuda Sadeh 1774a40c4f10SYehuda Sadeh osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify"); 1775a40c4f10SYehuda Sadeh if (IS_ERR(osdc->notify_wq)) { 1776a40c4f10SYehuda Sadeh err = PTR_ERR(osdc->notify_wq); 1777a40c4f10SYehuda Sadeh osdc->notify_wq = NULL; 1778a40c4f10SYehuda Sadeh goto out_msgpool; 1779a40c4f10SYehuda Sadeh } 17803d14c5d2SYehuda Sadeh return 0; 17813d14c5d2SYehuda Sadeh 17823d14c5d2SYehuda Sadeh out_msgpool: 17833d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 17843d14c5d2SYehuda Sadeh out_mempool: 17853d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 17863d14c5d2SYehuda Sadeh out: 17873d14c5d2SYehuda Sadeh return err; 17883d14c5d2SYehuda Sadeh } 17893d14c5d2SYehuda Sadeh 17903d14c5d2SYehuda Sadeh void ceph_osdc_stop(struct ceph_osd_client *osdc) 17913d14c5d2SYehuda Sadeh { 1792a40c4f10SYehuda Sadeh flush_workqueue(osdc->notify_wq); 1793a40c4f10SYehuda Sadeh destroy_workqueue(osdc->notify_wq); 17943d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->timeout_work); 17953d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->osds_timeout_work); 17963d14c5d2SYehuda Sadeh if (osdc->osdmap) { 17973d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 17983d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 17993d14c5d2SYehuda Sadeh } 1800aca420bcSSage Weil remove_all_osds(osdc); 18013d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 18023d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 18033d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op_reply); 18043d14c5d2SYehuda Sadeh } 18053d14c5d2SYehuda Sadeh 18063d14c5d2SYehuda Sadeh /* 18073d14c5d2SYehuda Sadeh * Read some contiguous pages. If we cross a stripe boundary, shorten 18083d14c5d2SYehuda Sadeh * *plen. Return number of bytes read, or error. 18093d14c5d2SYehuda Sadeh */ 18103d14c5d2SYehuda Sadeh int ceph_osdc_readpages(struct ceph_osd_client *osdc, 18113d14c5d2SYehuda Sadeh struct ceph_vino vino, struct ceph_file_layout *layout, 18123d14c5d2SYehuda Sadeh u64 off, u64 *plen, 18133d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 1814b7495fc2SSage Weil struct page **pages, int num_pages, int page_align) 18153d14c5d2SYehuda Sadeh { 18163d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 18173d14c5d2SYehuda Sadeh int rc = 0; 18183d14c5d2SYehuda Sadeh 18193d14c5d2SYehuda Sadeh dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino, 18203d14c5d2SYehuda Sadeh vino.snap, off, *plen); 18213d14c5d2SYehuda Sadeh req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 18223d14c5d2SYehuda Sadeh CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ, 18233d14c5d2SYehuda Sadeh NULL, 0, truncate_seq, truncate_size, NULL, 1824a3bea47eSAlex Elder false, page_align); 18256816282dSSage Weil if (IS_ERR(req)) 18266816282dSSage Weil return PTR_ERR(req); 18273d14c5d2SYehuda Sadeh 18283d14c5d2SYehuda Sadeh /* it may be a short read due to an object boundary */ 18293d14c5d2SYehuda Sadeh req->r_pages = pages; 18303d14c5d2SYehuda Sadeh 1831b7495fc2SSage Weil dout("readpages final extent is %llu~%llu (%d pages align %d)\n", 1832b7495fc2SSage Weil off, *plen, req->r_num_pages, page_align); 18333d14c5d2SYehuda Sadeh 18343d14c5d2SYehuda Sadeh rc = ceph_osdc_start_request(osdc, req, false); 18353d14c5d2SYehuda Sadeh if (!rc) 18363d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 18373d14c5d2SYehuda Sadeh 18383d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 18393d14c5d2SYehuda Sadeh dout("readpages result %d\n", rc); 18403d14c5d2SYehuda Sadeh return rc; 18413d14c5d2SYehuda Sadeh } 18423d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_readpages); 18433d14c5d2SYehuda Sadeh 18443d14c5d2SYehuda Sadeh /* 18453d14c5d2SYehuda Sadeh * do a synchronous write on N pages 18463d14c5d2SYehuda Sadeh */ 18473d14c5d2SYehuda Sadeh int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino, 18483d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 18493d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 18503d14c5d2SYehuda Sadeh u64 off, u64 len, 18513d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 18523d14c5d2SYehuda Sadeh struct timespec *mtime, 185324808826SAlex Elder struct page **pages, int num_pages) 18543d14c5d2SYehuda Sadeh { 18553d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 18563d14c5d2SYehuda Sadeh int rc = 0; 1857b7495fc2SSage Weil int page_align = off & ~PAGE_MASK; 18583d14c5d2SYehuda Sadeh 18593d14c5d2SYehuda Sadeh BUG_ON(vino.snap != CEPH_NOSNAP); 18603d14c5d2SYehuda Sadeh req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 18613d14c5d2SYehuda Sadeh CEPH_OSD_OP_WRITE, 186224808826SAlex Elder CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE, 1863fbf8685fSAlex Elder snapc, 0, 18643d14c5d2SYehuda Sadeh truncate_seq, truncate_size, mtime, 1865a3bea47eSAlex Elder true, page_align); 18666816282dSSage Weil if (IS_ERR(req)) 18676816282dSSage Weil return PTR_ERR(req); 18683d14c5d2SYehuda Sadeh 18693d14c5d2SYehuda Sadeh /* it may be a short write due to an object boundary */ 18703d14c5d2SYehuda Sadeh req->r_pages = pages; 18713d14c5d2SYehuda Sadeh dout("writepages %llu~%llu (%d pages)\n", off, len, 18723d14c5d2SYehuda Sadeh req->r_num_pages); 18733d14c5d2SYehuda Sadeh 187487f979d3SAlex Elder rc = ceph_osdc_start_request(osdc, req, true); 18753d14c5d2SYehuda Sadeh if (!rc) 18763d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 18773d14c5d2SYehuda Sadeh 18783d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 18793d14c5d2SYehuda Sadeh if (rc == 0) 18803d14c5d2SYehuda Sadeh rc = len; 18813d14c5d2SYehuda Sadeh dout("writepages result %d\n", rc); 18823d14c5d2SYehuda Sadeh return rc; 18833d14c5d2SYehuda Sadeh } 18843d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_writepages); 18853d14c5d2SYehuda Sadeh 18863d14c5d2SYehuda Sadeh /* 18873d14c5d2SYehuda Sadeh * handle incoming message 18883d14c5d2SYehuda Sadeh */ 18893d14c5d2SYehuda Sadeh static void dispatch(struct ceph_connection *con, struct ceph_msg *msg) 18903d14c5d2SYehuda Sadeh { 18913d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 18923d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 18933d14c5d2SYehuda Sadeh int type = le16_to_cpu(msg->hdr.type); 18943d14c5d2SYehuda Sadeh 18953d14c5d2SYehuda Sadeh if (!osd) 18963d14c5d2SYehuda Sadeh goto out; 18973d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 18983d14c5d2SYehuda Sadeh 18993d14c5d2SYehuda Sadeh switch (type) { 19003d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 19013d14c5d2SYehuda Sadeh ceph_osdc_handle_map(osdc, msg); 19023d14c5d2SYehuda Sadeh break; 19033d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 19043d14c5d2SYehuda Sadeh handle_reply(osdc, msg, con); 19053d14c5d2SYehuda Sadeh break; 1906a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 1907a40c4f10SYehuda Sadeh handle_watch_notify(osdc, msg); 1908a40c4f10SYehuda Sadeh break; 19093d14c5d2SYehuda Sadeh 19103d14c5d2SYehuda Sadeh default: 19113d14c5d2SYehuda Sadeh pr_err("received unknown message type %d %s\n", type, 19123d14c5d2SYehuda Sadeh ceph_msg_type_name(type)); 19133d14c5d2SYehuda Sadeh } 19143d14c5d2SYehuda Sadeh out: 19153d14c5d2SYehuda Sadeh ceph_msg_put(msg); 19163d14c5d2SYehuda Sadeh } 19173d14c5d2SYehuda Sadeh 19183d14c5d2SYehuda Sadeh /* 19193d14c5d2SYehuda Sadeh * lookup and return message for incoming reply. set up reply message 19203d14c5d2SYehuda Sadeh * pages. 19213d14c5d2SYehuda Sadeh */ 19223d14c5d2SYehuda Sadeh static struct ceph_msg *get_reply(struct ceph_connection *con, 19233d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 19243d14c5d2SYehuda Sadeh int *skip) 19253d14c5d2SYehuda Sadeh { 19263d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 19273d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = osd->o_osdc; 19283d14c5d2SYehuda Sadeh struct ceph_msg *m; 19293d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 19303d14c5d2SYehuda Sadeh int front = le32_to_cpu(hdr->front_len); 19313d14c5d2SYehuda Sadeh int data_len = le32_to_cpu(hdr->data_len); 19323d14c5d2SYehuda Sadeh u64 tid; 19333d14c5d2SYehuda Sadeh 19343d14c5d2SYehuda Sadeh tid = le64_to_cpu(hdr->tid); 19353d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 19363d14c5d2SYehuda Sadeh req = __lookup_request(osdc, tid); 19373d14c5d2SYehuda Sadeh if (!req) { 19383d14c5d2SYehuda Sadeh *skip = 1; 19393d14c5d2SYehuda Sadeh m = NULL; 1940756a16a5SSage Weil dout("get_reply unknown tid %llu from osd%d\n", tid, 19413d14c5d2SYehuda Sadeh osd->o_osd); 19423d14c5d2SYehuda Sadeh goto out; 19433d14c5d2SYehuda Sadeh } 19443d14c5d2SYehuda Sadeh 19453d14c5d2SYehuda Sadeh if (req->r_con_filling_msg) { 19468921d114SAlex Elder dout("%s revoking msg %p from old con %p\n", __func__, 19473d14c5d2SYehuda Sadeh req->r_reply, req->r_con_filling_msg); 19488921d114SAlex Elder ceph_msg_revoke_incoming(req->r_reply); 19490d47766fSSage Weil req->r_con_filling_msg->ops->put(req->r_con_filling_msg); 19503d14c5d2SYehuda Sadeh req->r_con_filling_msg = NULL; 19513d14c5d2SYehuda Sadeh } 19523d14c5d2SYehuda Sadeh 19533d14c5d2SYehuda Sadeh if (front > req->r_reply->front.iov_len) { 19543d14c5d2SYehuda Sadeh pr_warning("get_reply front %d > preallocated %d\n", 19553d14c5d2SYehuda Sadeh front, (int)req->r_reply->front.iov_len); 1956b61c2763SSage Weil m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS, false); 19573d14c5d2SYehuda Sadeh if (!m) 19583d14c5d2SYehuda Sadeh goto out; 19593d14c5d2SYehuda Sadeh ceph_msg_put(req->r_reply); 19603d14c5d2SYehuda Sadeh req->r_reply = m; 19613d14c5d2SYehuda Sadeh } 19623d14c5d2SYehuda Sadeh m = ceph_msg_get(req->r_reply); 19633d14c5d2SYehuda Sadeh 19643d14c5d2SYehuda Sadeh if (data_len > 0) { 1965b7495fc2SSage Weil int want = calc_pages_for(req->r_page_alignment, data_len); 19663d14c5d2SYehuda Sadeh 19679cbb1d72SAlex Elder if (req->r_pages && unlikely(req->r_num_pages < want)) { 19689bb0ce2bSSage Weil pr_warning("tid %lld reply has %d bytes %d pages, we" 19699bb0ce2bSSage Weil " had only %d pages ready\n", tid, data_len, 19709bb0ce2bSSage Weil want, req->r_num_pages); 19713d14c5d2SYehuda Sadeh *skip = 1; 19723d14c5d2SYehuda Sadeh ceph_msg_put(m); 19733d14c5d2SYehuda Sadeh m = NULL; 19743d14c5d2SYehuda Sadeh goto out; 19753d14c5d2SYehuda Sadeh } 19763d14c5d2SYehuda Sadeh m->pages = req->r_pages; 19773d14c5d2SYehuda Sadeh m->nr_pages = req->r_num_pages; 1978c5c6b19dSSage Weil m->page_alignment = req->r_page_alignment; 19793d14c5d2SYehuda Sadeh #ifdef CONFIG_BLOCK 19803d14c5d2SYehuda Sadeh m->bio = req->r_bio; 19813d14c5d2SYehuda Sadeh #endif 19823d14c5d2SYehuda Sadeh } 19833d14c5d2SYehuda Sadeh *skip = 0; 19840d47766fSSage Weil req->r_con_filling_msg = con->ops->get(con); 19853d14c5d2SYehuda Sadeh dout("get_reply tid %lld %p\n", tid, m); 19863d14c5d2SYehuda Sadeh 19873d14c5d2SYehuda Sadeh out: 19883d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 19893d14c5d2SYehuda Sadeh return m; 19903d14c5d2SYehuda Sadeh 19913d14c5d2SYehuda Sadeh } 19923d14c5d2SYehuda Sadeh 19933d14c5d2SYehuda Sadeh static struct ceph_msg *alloc_msg(struct ceph_connection *con, 19943d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 19953d14c5d2SYehuda Sadeh int *skip) 19963d14c5d2SYehuda Sadeh { 19973d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 19983d14c5d2SYehuda Sadeh int type = le16_to_cpu(hdr->type); 19993d14c5d2SYehuda Sadeh int front = le32_to_cpu(hdr->front_len); 20003d14c5d2SYehuda Sadeh 20011c20f2d2SAlex Elder *skip = 0; 20023d14c5d2SYehuda Sadeh switch (type) { 20033d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 2004a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 2005b61c2763SSage Weil return ceph_msg_new(type, front, GFP_NOFS, false); 20063d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 20073d14c5d2SYehuda Sadeh return get_reply(con, hdr, skip); 20083d14c5d2SYehuda Sadeh default: 20093d14c5d2SYehuda Sadeh pr_info("alloc_msg unexpected msg type %d from osd%d\n", type, 20103d14c5d2SYehuda Sadeh osd->o_osd); 20113d14c5d2SYehuda Sadeh *skip = 1; 20123d14c5d2SYehuda Sadeh return NULL; 20133d14c5d2SYehuda Sadeh } 20143d14c5d2SYehuda Sadeh } 20153d14c5d2SYehuda Sadeh 20163d14c5d2SYehuda Sadeh /* 20173d14c5d2SYehuda Sadeh * Wrappers to refcount containing ceph_osd struct 20183d14c5d2SYehuda Sadeh */ 20193d14c5d2SYehuda Sadeh static struct ceph_connection *get_osd_con(struct ceph_connection *con) 20203d14c5d2SYehuda Sadeh { 20213d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 20223d14c5d2SYehuda Sadeh if (get_osd(osd)) 20233d14c5d2SYehuda Sadeh return con; 20243d14c5d2SYehuda Sadeh return NULL; 20253d14c5d2SYehuda Sadeh } 20263d14c5d2SYehuda Sadeh 20273d14c5d2SYehuda Sadeh static void put_osd_con(struct ceph_connection *con) 20283d14c5d2SYehuda Sadeh { 20293d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 20303d14c5d2SYehuda Sadeh put_osd(osd); 20313d14c5d2SYehuda Sadeh } 20323d14c5d2SYehuda Sadeh 20333d14c5d2SYehuda Sadeh /* 20343d14c5d2SYehuda Sadeh * authentication 20353d14c5d2SYehuda Sadeh */ 2036a3530df3SAlex Elder /* 2037a3530df3SAlex Elder * Note: returned pointer is the address of a structure that's 2038a3530df3SAlex Elder * managed separately. Caller must *not* attempt to free it. 2039a3530df3SAlex Elder */ 2040a3530df3SAlex Elder static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con, 20418f43fb53SAlex Elder int *proto, int force_new) 20423d14c5d2SYehuda Sadeh { 20433d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 20443d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 20453d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 204674f1869fSAlex Elder struct ceph_auth_handshake *auth = &o->o_auth; 20473d14c5d2SYehuda Sadeh 204874f1869fSAlex Elder if (force_new && auth->authorizer) { 2049a255651dSAlex Elder if (ac->ops && ac->ops->destroy_authorizer) 205074f1869fSAlex Elder ac->ops->destroy_authorizer(ac, auth->authorizer); 205174f1869fSAlex Elder auth->authorizer = NULL; 20523d14c5d2SYehuda Sadeh } 2053a255651dSAlex Elder if (!auth->authorizer && ac->ops && ac->ops->create_authorizer) { 2054a3530df3SAlex Elder int ret = ac->ops->create_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 2055a3530df3SAlex Elder auth); 20563d14c5d2SYehuda Sadeh if (ret) 2057a3530df3SAlex Elder return ERR_PTR(ret); 20583d14c5d2SYehuda Sadeh } 20593d14c5d2SYehuda Sadeh *proto = ac->protocol; 206074f1869fSAlex Elder 2061a3530df3SAlex Elder return auth; 20623d14c5d2SYehuda Sadeh } 20633d14c5d2SYehuda Sadeh 20643d14c5d2SYehuda Sadeh 20653d14c5d2SYehuda Sadeh static int verify_authorizer_reply(struct ceph_connection *con, int len) 20663d14c5d2SYehuda Sadeh { 20673d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 20683d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 20693d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 20703d14c5d2SYehuda Sadeh 2071a255651dSAlex Elder /* 2072a255651dSAlex Elder * XXX If ac->ops or ac->ops->verify_authorizer_reply is null, 2073a255651dSAlex Elder * XXX which do we do: succeed or fail? 2074a255651dSAlex Elder */ 20756c4a1915SAlex Elder return ac->ops->verify_authorizer_reply(ac, o->o_auth.authorizer, len); 20763d14c5d2SYehuda Sadeh } 20773d14c5d2SYehuda Sadeh 20783d14c5d2SYehuda Sadeh static int invalidate_authorizer(struct ceph_connection *con) 20793d14c5d2SYehuda Sadeh { 20803d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 20813d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 20823d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 20833d14c5d2SYehuda Sadeh 2084a255651dSAlex Elder if (ac->ops && ac->ops->invalidate_authorizer) 20853d14c5d2SYehuda Sadeh ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD); 20863d14c5d2SYehuda Sadeh 20873d14c5d2SYehuda Sadeh return ceph_monc_validate_auth(&osdc->client->monc); 20883d14c5d2SYehuda Sadeh } 20893d14c5d2SYehuda Sadeh 20903d14c5d2SYehuda Sadeh static const struct ceph_connection_operations osd_con_ops = { 20913d14c5d2SYehuda Sadeh .get = get_osd_con, 20923d14c5d2SYehuda Sadeh .put = put_osd_con, 20933d14c5d2SYehuda Sadeh .dispatch = dispatch, 20943d14c5d2SYehuda Sadeh .get_authorizer = get_authorizer, 20953d14c5d2SYehuda Sadeh .verify_authorizer_reply = verify_authorizer_reply, 20963d14c5d2SYehuda Sadeh .invalidate_authorizer = invalidate_authorizer, 20973d14c5d2SYehuda Sadeh .alloc_msg = alloc_msg, 20983d14c5d2SYehuda Sadeh .fault = osd_reset, 20993d14c5d2SYehuda Sadeh }; 2100