13d14c5d2SYehuda Sadeh #include <linux/ceph/ceph_debug.h> 23d14c5d2SYehuda Sadeh 33d14c5d2SYehuda Sadeh #include <linux/module.h> 43d14c5d2SYehuda Sadeh #include <linux/err.h> 53d14c5d2SYehuda Sadeh #include <linux/highmem.h> 63d14c5d2SYehuda Sadeh #include <linux/mm.h> 73d14c5d2SYehuda Sadeh #include <linux/pagemap.h> 83d14c5d2SYehuda Sadeh #include <linux/slab.h> 93d14c5d2SYehuda Sadeh #include <linux/uaccess.h> 103d14c5d2SYehuda Sadeh #ifdef CONFIG_BLOCK 113d14c5d2SYehuda Sadeh #include <linux/bio.h> 123d14c5d2SYehuda Sadeh #endif 133d14c5d2SYehuda Sadeh 143d14c5d2SYehuda Sadeh #include <linux/ceph/libceph.h> 153d14c5d2SYehuda Sadeh #include <linux/ceph/osd_client.h> 163d14c5d2SYehuda Sadeh #include <linux/ceph/messenger.h> 173d14c5d2SYehuda Sadeh #include <linux/ceph/decode.h> 183d14c5d2SYehuda Sadeh #include <linux/ceph/auth.h> 193d14c5d2SYehuda Sadeh #include <linux/ceph/pagelist.h> 203d14c5d2SYehuda Sadeh 213d14c5d2SYehuda Sadeh #define OSD_OP_FRONT_LEN 4096 223d14c5d2SYehuda Sadeh #define OSD_OPREPLY_FRONT_LEN 512 233d14c5d2SYehuda Sadeh 243d14c5d2SYehuda Sadeh static const struct ceph_connection_operations osd_con_ops; 253d14c5d2SYehuda Sadeh 26f9d25199SAlex Elder static void __send_queued(struct ceph_osd_client *osdc); 276f6c7006SSage Weil static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd); 28a40c4f10SYehuda Sadeh static void __register_request(struct ceph_osd_client *osdc, 29a40c4f10SYehuda Sadeh struct ceph_osd_request *req); 30a40c4f10SYehuda Sadeh static void __unregister_linger_request(struct ceph_osd_client *osdc, 31a40c4f10SYehuda Sadeh struct ceph_osd_request *req); 3256e925b6SSage Weil static void __send_request(struct ceph_osd_client *osdc, 33a40c4f10SYehuda Sadeh struct ceph_osd_request *req); 343d14c5d2SYehuda Sadeh 353d14c5d2SYehuda Sadeh static int op_has_extent(int op) 363d14c5d2SYehuda Sadeh { 373d14c5d2SYehuda Sadeh return (op == CEPH_OSD_OP_READ || 383d14c5d2SYehuda Sadeh op == CEPH_OSD_OP_WRITE); 393d14c5d2SYehuda Sadeh } 403d14c5d2SYehuda Sadeh 413d14c5d2SYehuda Sadeh /* 423d14c5d2SYehuda Sadeh * Implement client access to distributed object storage cluster. 433d14c5d2SYehuda Sadeh * 443d14c5d2SYehuda Sadeh * All data objects are stored within a cluster/cloud of OSDs, or 453d14c5d2SYehuda Sadeh * "object storage devices." (Note that Ceph OSDs have _nothing_ to 463d14c5d2SYehuda Sadeh * do with the T10 OSD extensions to SCSI.) Ceph OSDs are simply 473d14c5d2SYehuda Sadeh * remote daemons serving up and coordinating consistent and safe 483d14c5d2SYehuda Sadeh * access to storage. 493d14c5d2SYehuda Sadeh * 503d14c5d2SYehuda Sadeh * Cluster membership and the mapping of data objects onto storage devices 513d14c5d2SYehuda Sadeh * are described by the osd map. 523d14c5d2SYehuda Sadeh * 533d14c5d2SYehuda Sadeh * We keep track of pending OSD requests (read, write), resubmit 543d14c5d2SYehuda Sadeh * requests to different OSDs when the cluster topology/data layout 553d14c5d2SYehuda Sadeh * change, or retry the affected requests when the communications 563d14c5d2SYehuda Sadeh * channel with an OSD is reset. 573d14c5d2SYehuda Sadeh */ 583d14c5d2SYehuda Sadeh 593d14c5d2SYehuda Sadeh /* 603d14c5d2SYehuda Sadeh * calculate the mapping of a file extent onto an object, and fill out the 613d14c5d2SYehuda Sadeh * request accordingly. shorten extent as necessary if it crosses an 623d14c5d2SYehuda Sadeh * object boundary. 633d14c5d2SYehuda Sadeh * 643d14c5d2SYehuda Sadeh * fill osd op in request message. 653d14c5d2SYehuda Sadeh */ 66dbe0fc41SAlex Elder static int calc_layout(struct ceph_file_layout *layout, u64 off, u64 *plen, 6747a05811SAlex Elder struct ceph_osd_req_op *op, u64 *bno) 683d14c5d2SYehuda Sadeh { 6960e56f13SAlex Elder u64 orig_len = *plen; 7060e56f13SAlex Elder u64 objoff = 0; 7160e56f13SAlex Elder u64 objlen = 0; 72d63b77f4SSage Weil int r; 733d14c5d2SYehuda Sadeh 7460e56f13SAlex Elder /* object extent? */ 7547a05811SAlex Elder r = ceph_calc_file_object_mapping(layout, off, orig_len, bno, 7660e56f13SAlex Elder &objoff, &objlen); 77d63b77f4SSage Weil if (r < 0) 78d63b77f4SSage Weil return r; 7960e56f13SAlex Elder if (objlen < orig_len) { 8060e56f13SAlex Elder *plen = objlen; 8160e56f13SAlex Elder dout(" skipping last %llu, final file extent %llu~%llu\n", 8260e56f13SAlex Elder orig_len - *plen, off, *plen); 8360e56f13SAlex Elder } 8460e56f13SAlex Elder 8560e56f13SAlex Elder if (op_has_extent(op->op)) { 8660e56f13SAlex Elder u32 osize = le32_to_cpu(layout->fl_object_size); 8760e56f13SAlex Elder op->extent.offset = objoff; 8860e56f13SAlex Elder op->extent.length = objlen; 8960e56f13SAlex Elder if (op->extent.truncate_size <= off - objoff) { 9060e56f13SAlex Elder op->extent.truncate_size = 0; 9160e56f13SAlex Elder } else { 9260e56f13SAlex Elder op->extent.truncate_size -= off - objoff; 9360e56f13SAlex Elder if (op->extent.truncate_size > osize) 9460e56f13SAlex Elder op->extent.truncate_size = osize; 9560e56f13SAlex Elder } 9660e56f13SAlex Elder } 9760e56f13SAlex Elder if (op->op == CEPH_OSD_OP_WRITE) 9860e56f13SAlex Elder op->payload_len = *plen; 9960e56f13SAlex Elder 10060cf5992SAlex Elder dout("calc_layout bno=%llx %llu~%llu\n", *bno, objoff, objlen); 1013d14c5d2SYehuda Sadeh 1023ff5f385SAlex Elder return 0; 1033d14c5d2SYehuda Sadeh } 1043d14c5d2SYehuda Sadeh 1053d14c5d2SYehuda Sadeh /* 1063d14c5d2SYehuda Sadeh * requests 1073d14c5d2SYehuda Sadeh */ 1083d14c5d2SYehuda Sadeh void ceph_osdc_release_request(struct kref *kref) 1093d14c5d2SYehuda Sadeh { 110e0c59487SAlex Elder int num_pages; 1113d14c5d2SYehuda Sadeh struct ceph_osd_request *req = container_of(kref, 1123d14c5d2SYehuda Sadeh struct ceph_osd_request, 1133d14c5d2SYehuda Sadeh r_kref); 1143d14c5d2SYehuda Sadeh 1153d14c5d2SYehuda Sadeh if (req->r_request) 1163d14c5d2SYehuda Sadeh ceph_msg_put(req->r_request); 1173d14c5d2SYehuda Sadeh if (req->r_con_filling_msg) { 1189cbb1d72SAlex Elder dout("%s revoking msg %p from con %p\n", __func__, 1199cbb1d72SAlex Elder req->r_reply, req->r_con_filling_msg); 1208921d114SAlex Elder ceph_msg_revoke_incoming(req->r_reply); 1210d47766fSSage Weil req->r_con_filling_msg->ops->put(req->r_con_filling_msg); 1229cbb1d72SAlex Elder req->r_con_filling_msg = NULL; 1233d14c5d2SYehuda Sadeh } 124ab8cb34aSAlex Elder if (req->r_reply) 125ab8cb34aSAlex Elder ceph_msg_put(req->r_reply); 1260fff87ecSAlex Elder 1270fff87ecSAlex Elder if (req->r_data_in.type == CEPH_OSD_DATA_TYPE_PAGES && 128e0c59487SAlex Elder req->r_data_in.own_pages) { 129e0c59487SAlex Elder num_pages = calc_pages_for((u64)req->r_data_in.alignment, 130e0c59487SAlex Elder (u64)req->r_data_in.length); 131e0c59487SAlex Elder ceph_release_page_vector(req->r_data_in.pages, num_pages); 132e0c59487SAlex Elder } 1330fff87ecSAlex Elder if (req->r_data_out.type == CEPH_OSD_DATA_TYPE_PAGES && 134e0c59487SAlex Elder req->r_data_out.own_pages) { 135e0c59487SAlex Elder num_pages = calc_pages_for((u64)req->r_data_out.alignment, 136e0c59487SAlex Elder (u64)req->r_data_out.length); 137e0c59487SAlex Elder ceph_release_page_vector(req->r_data_out.pages, num_pages); 138e0c59487SAlex Elder } 1390fff87ecSAlex Elder 1403d14c5d2SYehuda Sadeh ceph_put_snap_context(req->r_snapc); 1413d14c5d2SYehuda Sadeh if (req->r_mempool) 1423d14c5d2SYehuda Sadeh mempool_free(req, req->r_osdc->req_mempool); 1433d14c5d2SYehuda Sadeh else 1443d14c5d2SYehuda Sadeh kfree(req); 1453d14c5d2SYehuda Sadeh } 1463d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_release_request); 1473d14c5d2SYehuda Sadeh 1483d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc, 1493d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 1501b83bef2SSage Weil unsigned int num_ops, 1513d14c5d2SYehuda Sadeh bool use_mempool, 15254a54007SAlex Elder gfp_t gfp_flags) 1533d14c5d2SYehuda Sadeh { 1543d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 1553d14c5d2SYehuda Sadeh struct ceph_msg *msg; 1561b83bef2SSage Weil size_t msg_size; 1573d14c5d2SYehuda Sadeh 1581b83bef2SSage Weil msg_size = 4 + 4 + 8 + 8 + 4+8; 1591b83bef2SSage Weil msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */ 1601b83bef2SSage Weil msg_size += 1 + 8 + 4 + 4; /* pg_t */ 1611b83bef2SSage Weil msg_size += 4 + MAX_OBJ_NAME_SIZE; 1621b83bef2SSage Weil msg_size += 2 + num_ops*sizeof(struct ceph_osd_op); 1631b83bef2SSage Weil msg_size += 8; /* snapid */ 1641b83bef2SSage Weil msg_size += 8; /* snap_seq */ 1651b83bef2SSage Weil msg_size += 8 * (snapc ? snapc->num_snaps : 0); /* snaps */ 1661b83bef2SSage Weil msg_size += 4; 1673d14c5d2SYehuda Sadeh 1683d14c5d2SYehuda Sadeh if (use_mempool) { 1693d14c5d2SYehuda Sadeh req = mempool_alloc(osdc->req_mempool, gfp_flags); 1703d14c5d2SYehuda Sadeh memset(req, 0, sizeof(*req)); 1713d14c5d2SYehuda Sadeh } else { 1723d14c5d2SYehuda Sadeh req = kzalloc(sizeof(*req), gfp_flags); 1733d14c5d2SYehuda Sadeh } 1743d14c5d2SYehuda Sadeh if (req == NULL) 1753d14c5d2SYehuda Sadeh return NULL; 1763d14c5d2SYehuda Sadeh 1773d14c5d2SYehuda Sadeh req->r_osdc = osdc; 1783d14c5d2SYehuda Sadeh req->r_mempool = use_mempool; 1793d14c5d2SYehuda Sadeh 1803d14c5d2SYehuda Sadeh kref_init(&req->r_kref); 1813d14c5d2SYehuda Sadeh init_completion(&req->r_completion); 1823d14c5d2SYehuda Sadeh init_completion(&req->r_safe_completion); 183a978fa20SAlex Elder RB_CLEAR_NODE(&req->r_node); 1843d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&req->r_unsafe_item); 185a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&req->r_linger_item); 186a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&req->r_linger_osd); 187935b639aSSage Weil INIT_LIST_HEAD(&req->r_req_lru_item); 188cd43045cSSage Weil INIT_LIST_HEAD(&req->r_osd_item); 189cd43045cSSage Weil 1903d14c5d2SYehuda Sadeh /* create reply message */ 1913d14c5d2SYehuda Sadeh if (use_mempool) 1923d14c5d2SYehuda Sadeh msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0); 1933d14c5d2SYehuda Sadeh else 1943d14c5d2SYehuda Sadeh msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, 195b61c2763SSage Weil OSD_OPREPLY_FRONT_LEN, gfp_flags, true); 1963d14c5d2SYehuda Sadeh if (!msg) { 1973d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 1983d14c5d2SYehuda Sadeh return NULL; 1993d14c5d2SYehuda Sadeh } 2003d14c5d2SYehuda Sadeh req->r_reply = msg; 2013d14c5d2SYehuda Sadeh 2020fff87ecSAlex Elder req->r_data_in.type = CEPH_OSD_DATA_TYPE_NONE; 2030fff87ecSAlex Elder req->r_data_out.type = CEPH_OSD_DATA_TYPE_NONE; 204d50b409fSSage Weil 2053d14c5d2SYehuda Sadeh /* create request message; allow space for oid */ 2063d14c5d2SYehuda Sadeh if (use_mempool) 2073d14c5d2SYehuda Sadeh msg = ceph_msgpool_get(&osdc->msgpool_op, 0); 2083d14c5d2SYehuda Sadeh else 209b61c2763SSage Weil msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true); 2103d14c5d2SYehuda Sadeh if (!msg) { 2113d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 2123d14c5d2SYehuda Sadeh return NULL; 2133d14c5d2SYehuda Sadeh } 2143d14c5d2SYehuda Sadeh 2153d14c5d2SYehuda Sadeh memset(msg->front.iov_base, 0, msg->front.iov_len); 2163d14c5d2SYehuda Sadeh 2173d14c5d2SYehuda Sadeh req->r_request = msg; 2183d14c5d2SYehuda Sadeh 2193d14c5d2SYehuda Sadeh return req; 2203d14c5d2SYehuda Sadeh } 2213d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_alloc_request); 2223d14c5d2SYehuda Sadeh 223175face2SAlex Elder static u64 osd_req_encode_op(struct ceph_osd_request *req, 2243d14c5d2SYehuda Sadeh struct ceph_osd_op *dst, 2253d14c5d2SYehuda Sadeh struct ceph_osd_req_op *src) 2263d14c5d2SYehuda Sadeh { 227175face2SAlex Elder u64 out_data_len = 0; 22895e072ebSAlex Elder struct ceph_pagelist *pagelist; 229175face2SAlex Elder 2303d14c5d2SYehuda Sadeh dst->op = cpu_to_le16(src->op); 2313d14c5d2SYehuda Sadeh 232065a68f9SAlex Elder switch (src->op) { 233fbfab539SAlex Elder case CEPH_OSD_OP_STAT: 234fbfab539SAlex Elder break; 2353d14c5d2SYehuda Sadeh case CEPH_OSD_OP_READ: 2363d14c5d2SYehuda Sadeh case CEPH_OSD_OP_WRITE: 237175face2SAlex Elder if (src->op == CEPH_OSD_OP_WRITE) 238175face2SAlex Elder out_data_len = src->extent.length; 239175face2SAlex Elder dst->extent.offset = cpu_to_le64(src->extent.offset); 240175face2SAlex Elder dst->extent.length = cpu_to_le64(src->extent.length); 2413d14c5d2SYehuda Sadeh dst->extent.truncate_size = 2423d14c5d2SYehuda Sadeh cpu_to_le64(src->extent.truncate_size); 2433d14c5d2SYehuda Sadeh dst->extent.truncate_seq = 2443d14c5d2SYehuda Sadeh cpu_to_le32(src->extent.truncate_seq); 2453d14c5d2SYehuda Sadeh break; 2463d14c5d2SYehuda Sadeh case CEPH_OSD_OP_CALL: 24795e072ebSAlex Elder pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS); 24895e072ebSAlex Elder BUG_ON(!pagelist); 24995e072ebSAlex Elder ceph_pagelist_init(pagelist); 25095e072ebSAlex Elder 2513d14c5d2SYehuda Sadeh dst->cls.class_len = src->cls.class_len; 2523d14c5d2SYehuda Sadeh dst->cls.method_len = src->cls.method_len; 2533d14c5d2SYehuda Sadeh dst->cls.indata_len = cpu_to_le32(src->cls.indata_len); 25495e072ebSAlex Elder ceph_pagelist_append(pagelist, src->cls.class_name, 2553d14c5d2SYehuda Sadeh src->cls.class_len); 25695e072ebSAlex Elder ceph_pagelist_append(pagelist, src->cls.method_name, 2573d14c5d2SYehuda Sadeh src->cls.method_len); 25895e072ebSAlex Elder ceph_pagelist_append(pagelist, src->cls.indata, 2593d14c5d2SYehuda Sadeh src->cls.indata_len); 26095e072ebSAlex Elder 26195e072ebSAlex Elder req->r_data_out.type = CEPH_OSD_DATA_TYPE_PAGELIST; 26295e072ebSAlex Elder req->r_data_out.pagelist = pagelist; 26395e072ebSAlex Elder out_data_len = pagelist->length; 2643d14c5d2SYehuda Sadeh break; 2653d14c5d2SYehuda Sadeh case CEPH_OSD_OP_STARTSYNC: 2663d14c5d2SYehuda Sadeh break; 267a40c4f10SYehuda Sadeh case CEPH_OSD_OP_NOTIFY_ACK: 268a40c4f10SYehuda Sadeh case CEPH_OSD_OP_WATCH: 269a40c4f10SYehuda Sadeh dst->watch.cookie = cpu_to_le64(src->watch.cookie); 270a40c4f10SYehuda Sadeh dst->watch.ver = cpu_to_le64(src->watch.ver); 271a40c4f10SYehuda Sadeh dst->watch.flag = src->watch.flag; 272a40c4f10SYehuda Sadeh break; 2733d14c5d2SYehuda Sadeh default: 2748f63ca2dSAlex Elder pr_err("unrecognized osd opcode %d\n", src->op); 2753d14c5d2SYehuda Sadeh WARN_ON(1); 2763d14c5d2SYehuda Sadeh break; 2774c46459cSAlex Elder case CEPH_OSD_OP_MAPEXT: 2784c46459cSAlex Elder case CEPH_OSD_OP_MASKTRUNC: 2794c46459cSAlex Elder case CEPH_OSD_OP_SPARSE_READ: 280a9f36c3eSAlex Elder case CEPH_OSD_OP_NOTIFY: 2814c46459cSAlex Elder case CEPH_OSD_OP_ASSERT_VER: 2824c46459cSAlex Elder case CEPH_OSD_OP_WRITEFULL: 2834c46459cSAlex Elder case CEPH_OSD_OP_TRUNCATE: 2844c46459cSAlex Elder case CEPH_OSD_OP_ZERO: 2854c46459cSAlex Elder case CEPH_OSD_OP_DELETE: 2864c46459cSAlex Elder case CEPH_OSD_OP_APPEND: 2874c46459cSAlex Elder case CEPH_OSD_OP_SETTRUNC: 2884c46459cSAlex Elder case CEPH_OSD_OP_TRIMTRUNC: 2894c46459cSAlex Elder case CEPH_OSD_OP_TMAPUP: 2904c46459cSAlex Elder case CEPH_OSD_OP_TMAPPUT: 2914c46459cSAlex Elder case CEPH_OSD_OP_TMAPGET: 2924c46459cSAlex Elder case CEPH_OSD_OP_CREATE: 293a9f36c3eSAlex Elder case CEPH_OSD_OP_ROLLBACK: 2944c46459cSAlex Elder case CEPH_OSD_OP_OMAPGETKEYS: 2954c46459cSAlex Elder case CEPH_OSD_OP_OMAPGETVALS: 2964c46459cSAlex Elder case CEPH_OSD_OP_OMAPGETHEADER: 2974c46459cSAlex Elder case CEPH_OSD_OP_OMAPGETVALSBYKEYS: 2984c46459cSAlex Elder case CEPH_OSD_OP_MODE_RD: 2994c46459cSAlex Elder case CEPH_OSD_OP_OMAPSETVALS: 3004c46459cSAlex Elder case CEPH_OSD_OP_OMAPSETHEADER: 3014c46459cSAlex Elder case CEPH_OSD_OP_OMAPCLEAR: 3024c46459cSAlex Elder case CEPH_OSD_OP_OMAPRMKEYS: 3034c46459cSAlex Elder case CEPH_OSD_OP_OMAP_CMP: 3044c46459cSAlex Elder case CEPH_OSD_OP_CLONERANGE: 3054c46459cSAlex Elder case CEPH_OSD_OP_ASSERT_SRC_VERSION: 3064c46459cSAlex Elder case CEPH_OSD_OP_SRC_CMPXATTR: 307a9f36c3eSAlex Elder case CEPH_OSD_OP_GETXATTR: 3084c46459cSAlex Elder case CEPH_OSD_OP_GETXATTRS: 309a9f36c3eSAlex Elder case CEPH_OSD_OP_CMPXATTR: 310a9f36c3eSAlex Elder case CEPH_OSD_OP_SETXATTR: 3114c46459cSAlex Elder case CEPH_OSD_OP_SETXATTRS: 3124c46459cSAlex Elder case CEPH_OSD_OP_RESETXATTRS: 3134c46459cSAlex Elder case CEPH_OSD_OP_RMXATTR: 3144c46459cSAlex Elder case CEPH_OSD_OP_PULL: 3154c46459cSAlex Elder case CEPH_OSD_OP_PUSH: 3164c46459cSAlex Elder case CEPH_OSD_OP_BALANCEREADS: 3174c46459cSAlex Elder case CEPH_OSD_OP_UNBALANCEREADS: 3184c46459cSAlex Elder case CEPH_OSD_OP_SCRUB: 3194c46459cSAlex Elder case CEPH_OSD_OP_SCRUB_RESERVE: 3204c46459cSAlex Elder case CEPH_OSD_OP_SCRUB_UNRESERVE: 3214c46459cSAlex Elder case CEPH_OSD_OP_SCRUB_STOP: 3224c46459cSAlex Elder case CEPH_OSD_OP_SCRUB_MAP: 3234c46459cSAlex Elder case CEPH_OSD_OP_WRLOCK: 3244c46459cSAlex Elder case CEPH_OSD_OP_WRUNLOCK: 3254c46459cSAlex Elder case CEPH_OSD_OP_RDLOCK: 3264c46459cSAlex Elder case CEPH_OSD_OP_RDUNLOCK: 3274c46459cSAlex Elder case CEPH_OSD_OP_UPLOCK: 3284c46459cSAlex Elder case CEPH_OSD_OP_DNLOCK: 3294c46459cSAlex Elder case CEPH_OSD_OP_PGLS: 3304c46459cSAlex Elder case CEPH_OSD_OP_PGLS_FILTER: 3314c46459cSAlex Elder pr_err("unsupported osd opcode %s\n", 3328f63ca2dSAlex Elder ceph_osd_op_name(src->op)); 3334c46459cSAlex Elder WARN_ON(1); 3344c46459cSAlex Elder break; 3353d14c5d2SYehuda Sadeh } 3363d14c5d2SYehuda Sadeh dst->payload_len = cpu_to_le32(src->payload_len); 337175face2SAlex Elder 338175face2SAlex Elder return out_data_len; 3393d14c5d2SYehuda Sadeh } 3403d14c5d2SYehuda Sadeh 3413d14c5d2SYehuda Sadeh /* 3423d14c5d2SYehuda Sadeh * build new request AND message 3433d14c5d2SYehuda Sadeh * 3443d14c5d2SYehuda Sadeh */ 3453d14c5d2SYehuda Sadeh void ceph_osdc_build_request(struct ceph_osd_request *req, 346175face2SAlex Elder u64 off, unsigned int num_ops, 3473d14c5d2SYehuda Sadeh struct ceph_osd_req_op *src_ops, 3484d6b250bSAlex Elder struct ceph_snap_context *snapc, u64 snap_id, 349af77f26cSAlex Elder struct timespec *mtime) 3503d14c5d2SYehuda Sadeh { 3513d14c5d2SYehuda Sadeh struct ceph_msg *msg = req->r_request; 3523d14c5d2SYehuda Sadeh struct ceph_osd_req_op *src_op; 3533d14c5d2SYehuda Sadeh void *p; 3541b83bef2SSage Weil size_t msg_size; 3553d14c5d2SYehuda Sadeh int flags = req->r_flags; 356f44246e3SAlex Elder u64 data_len; 3573d14c5d2SYehuda Sadeh int i; 3583d14c5d2SYehuda Sadeh 3591b83bef2SSage Weil req->r_num_ops = num_ops; 3601b83bef2SSage Weil req->r_snapid = snap_id; 3613d14c5d2SYehuda Sadeh req->r_snapc = ceph_get_snap_context(snapc); 3623d14c5d2SYehuda Sadeh 3631b83bef2SSage Weil /* encode request */ 3641b83bef2SSage Weil msg->hdr.version = cpu_to_le16(4); 3653d14c5d2SYehuda Sadeh 3661b83bef2SSage Weil p = msg->front.iov_base; 3671b83bef2SSage Weil ceph_encode_32(&p, 1); /* client_inc is always 1 */ 3681b83bef2SSage Weil req->r_request_osdmap_epoch = p; 3691b83bef2SSage Weil p += 4; 3701b83bef2SSage Weil req->r_request_flags = p; 3711b83bef2SSage Weil p += 4; 3721b83bef2SSage Weil if (req->r_flags & CEPH_OSD_FLAG_WRITE) 3731b83bef2SSage Weil ceph_encode_timespec(p, mtime); 3741b83bef2SSage Weil p += sizeof(struct ceph_timespec); 3751b83bef2SSage Weil req->r_request_reassert_version = p; 3761b83bef2SSage Weil p += sizeof(struct ceph_eversion); /* will get filled in */ 3771b83bef2SSage Weil 3781b83bef2SSage Weil /* oloc */ 3791b83bef2SSage Weil ceph_encode_8(&p, 4); 3801b83bef2SSage Weil ceph_encode_8(&p, 4); 3811b83bef2SSage Weil ceph_encode_32(&p, 8 + 4 + 4); 3821b83bef2SSage Weil req->r_request_pool = p; 3831b83bef2SSage Weil p += 8; 3841b83bef2SSage Weil ceph_encode_32(&p, -1); /* preferred */ 3851b83bef2SSage Weil ceph_encode_32(&p, 0); /* key len */ 3861b83bef2SSage Weil 3871b83bef2SSage Weil ceph_encode_8(&p, 1); 3881b83bef2SSage Weil req->r_request_pgid = p; 3891b83bef2SSage Weil p += 8 + 4; 3901b83bef2SSage Weil ceph_encode_32(&p, -1); /* preferred */ 3911b83bef2SSage Weil 3921b83bef2SSage Weil /* oid */ 3931b83bef2SSage Weil ceph_encode_32(&p, req->r_oid_len); 394af77f26cSAlex Elder memcpy(p, req->r_oid, req->r_oid_len); 3951b83bef2SSage Weil dout("oid '%.*s' len %d\n", req->r_oid_len, req->r_oid, req->r_oid_len); 396af77f26cSAlex Elder p += req->r_oid_len; 3973d14c5d2SYehuda Sadeh 398175face2SAlex Elder /* ops--can imply data */ 3991b83bef2SSage Weil ceph_encode_16(&p, num_ops); 4003d14c5d2SYehuda Sadeh src_op = src_ops; 4011b83bef2SSage Weil req->r_request_ops = p; 402175face2SAlex Elder data_len = 0; 4031b83bef2SSage Weil for (i = 0; i < num_ops; i++, src_op++) { 404175face2SAlex Elder data_len += osd_req_encode_op(req, p, src_op); 4051b83bef2SSage Weil p += sizeof(struct ceph_osd_op); 4061b83bef2SSage Weil } 4073d14c5d2SYehuda Sadeh 4081b83bef2SSage Weil /* snaps */ 4091b83bef2SSage Weil ceph_encode_64(&p, req->r_snapid); 4101b83bef2SSage Weil ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0); 4111b83bef2SSage Weil ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0); 4121b83bef2SSage Weil if (req->r_snapc) { 4133d14c5d2SYehuda Sadeh for (i = 0; i < snapc->num_snaps; i++) { 4141b83bef2SSage Weil ceph_encode_64(&p, req->r_snapc->snaps[i]); 4153d14c5d2SYehuda Sadeh } 4163d14c5d2SYehuda Sadeh } 4173d14c5d2SYehuda Sadeh 4181b83bef2SSage Weil req->r_request_attempts = p; 4191b83bef2SSage Weil p += 4; 4201b83bef2SSage Weil 421175face2SAlex Elder /* data */ 422175face2SAlex Elder if (flags & CEPH_OSD_FLAG_WRITE) 4233d14c5d2SYehuda Sadeh req->r_request->hdr.data_off = cpu_to_le16(off); 424f44246e3SAlex Elder req->r_request->hdr.data_len = cpu_to_le32(data_len); 425c5c6b19dSSage Weil 4263d14c5d2SYehuda Sadeh BUG_ON(p > msg->front.iov_base + msg->front.iov_len); 4273d14c5d2SYehuda Sadeh msg_size = p - msg->front.iov_base; 4283d14c5d2SYehuda Sadeh msg->front.iov_len = msg_size; 4293d14c5d2SYehuda Sadeh msg->hdr.front_len = cpu_to_le32(msg_size); 4301b83bef2SSage Weil 4311b83bef2SSage Weil dout("build_request msg_size was %d num_ops %d\n", (int)msg_size, 4321b83bef2SSage Weil num_ops); 4333d14c5d2SYehuda Sadeh return; 4343d14c5d2SYehuda Sadeh } 4353d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_build_request); 4363d14c5d2SYehuda Sadeh 4373d14c5d2SYehuda Sadeh /* 4383d14c5d2SYehuda Sadeh * build new request AND message, calculate layout, and adjust file 4393d14c5d2SYehuda Sadeh * extent as needed. 4403d14c5d2SYehuda Sadeh * 4413d14c5d2SYehuda Sadeh * if the file was recently truncated, we include information about its 4423d14c5d2SYehuda Sadeh * old and new size so that the object can be updated appropriately. (we 4433d14c5d2SYehuda Sadeh * avoid synchronously deleting truncated objects because it's slow.) 4443d14c5d2SYehuda Sadeh * 4453d14c5d2SYehuda Sadeh * if @do_sync, include a 'startsync' command so that the osd will flush 4463d14c5d2SYehuda Sadeh * data quickly. 4473d14c5d2SYehuda Sadeh */ 4483d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc, 4493d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 4503d14c5d2SYehuda Sadeh struct ceph_vino vino, 4513d14c5d2SYehuda Sadeh u64 off, u64 *plen, 4523d14c5d2SYehuda Sadeh int opcode, int flags, 4533d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 4543d14c5d2SYehuda Sadeh int do_sync, 4553d14c5d2SYehuda Sadeh u32 truncate_seq, 4563d14c5d2SYehuda Sadeh u64 truncate_size, 4573d14c5d2SYehuda Sadeh struct timespec *mtime, 458153e5167SAlex Elder bool use_mempool) 4593d14c5d2SYehuda Sadeh { 460ae7ca4a3SAlex Elder struct ceph_osd_req_op ops[2]; 4613d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 462ae7ca4a3SAlex Elder unsigned int num_op = 1; 46347a05811SAlex Elder u64 bno = 0; 4646816282dSSage Weil int r; 4653d14c5d2SYehuda Sadeh 466ae7ca4a3SAlex Elder memset(&ops, 0, sizeof ops); 467ae7ca4a3SAlex Elder 4683d14c5d2SYehuda Sadeh ops[0].op = opcode; 4693d14c5d2SYehuda Sadeh ops[0].extent.truncate_seq = truncate_seq; 4703d14c5d2SYehuda Sadeh ops[0].extent.truncate_size = truncate_size; 4713d14c5d2SYehuda Sadeh 4723d14c5d2SYehuda Sadeh if (do_sync) { 4733d14c5d2SYehuda Sadeh ops[1].op = CEPH_OSD_OP_STARTSYNC; 474ae7ca4a3SAlex Elder num_op++; 475ae7ca4a3SAlex Elder } 4763d14c5d2SYehuda Sadeh 477ae7ca4a3SAlex Elder req = ceph_osdc_alloc_request(osdc, snapc, num_op, use_mempool, 478ae7ca4a3SAlex Elder GFP_NOFS); 4794ad12621SSage Weil if (!req) 4806816282dSSage Weil return ERR_PTR(-ENOMEM); 481d178a9e7SAlex Elder req->r_flags = flags; 4823d14c5d2SYehuda Sadeh 4833d14c5d2SYehuda Sadeh /* calculate max write size */ 48460cf5992SAlex Elder r = calc_layout(layout, off, plen, ops, &bno); 4853ff5f385SAlex Elder if (r < 0) { 4863ff5f385SAlex Elder ceph_osdc_put_request(req); 4876816282dSSage Weil return ERR_PTR(r); 4883ff5f385SAlex Elder } 4893d14c5d2SYehuda Sadeh req->r_file_layout = *layout; /* keep a copy */ 4903d14c5d2SYehuda Sadeh 491dbe0fc41SAlex Elder snprintf(req->r_oid, sizeof(req->r_oid), "%llx.%08llx", vino.ino, bno); 492dbe0fc41SAlex Elder req->r_oid_len = strlen(req->r_oid); 493dbe0fc41SAlex Elder 494175face2SAlex Elder ceph_osdc_build_request(req, off, num_op, ops, 495ae7ca4a3SAlex Elder snapc, vino.snap, mtime); 4963d14c5d2SYehuda Sadeh 4973d14c5d2SYehuda Sadeh return req; 4983d14c5d2SYehuda Sadeh } 4993d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_new_request); 5003d14c5d2SYehuda Sadeh 5013d14c5d2SYehuda Sadeh /* 5023d14c5d2SYehuda Sadeh * We keep osd requests in an rbtree, sorted by ->r_tid. 5033d14c5d2SYehuda Sadeh */ 5043d14c5d2SYehuda Sadeh static void __insert_request(struct ceph_osd_client *osdc, 5053d14c5d2SYehuda Sadeh struct ceph_osd_request *new) 5063d14c5d2SYehuda Sadeh { 5073d14c5d2SYehuda Sadeh struct rb_node **p = &osdc->requests.rb_node; 5083d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 5093d14c5d2SYehuda Sadeh struct ceph_osd_request *req = NULL; 5103d14c5d2SYehuda Sadeh 5113d14c5d2SYehuda Sadeh while (*p) { 5123d14c5d2SYehuda Sadeh parent = *p; 5133d14c5d2SYehuda Sadeh req = rb_entry(parent, struct ceph_osd_request, r_node); 5143d14c5d2SYehuda Sadeh if (new->r_tid < req->r_tid) 5153d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 5163d14c5d2SYehuda Sadeh else if (new->r_tid > req->r_tid) 5173d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 5183d14c5d2SYehuda Sadeh else 5193d14c5d2SYehuda Sadeh BUG(); 5203d14c5d2SYehuda Sadeh } 5213d14c5d2SYehuda Sadeh 5223d14c5d2SYehuda Sadeh rb_link_node(&new->r_node, parent, p); 5233d14c5d2SYehuda Sadeh rb_insert_color(&new->r_node, &osdc->requests); 5243d14c5d2SYehuda Sadeh } 5253d14c5d2SYehuda Sadeh 5263d14c5d2SYehuda Sadeh static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc, 5273d14c5d2SYehuda Sadeh u64 tid) 5283d14c5d2SYehuda Sadeh { 5293d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 5303d14c5d2SYehuda Sadeh struct rb_node *n = osdc->requests.rb_node; 5313d14c5d2SYehuda Sadeh 5323d14c5d2SYehuda Sadeh while (n) { 5333d14c5d2SYehuda Sadeh req = rb_entry(n, struct ceph_osd_request, r_node); 5343d14c5d2SYehuda Sadeh if (tid < req->r_tid) 5353d14c5d2SYehuda Sadeh n = n->rb_left; 5363d14c5d2SYehuda Sadeh else if (tid > req->r_tid) 5373d14c5d2SYehuda Sadeh n = n->rb_right; 5383d14c5d2SYehuda Sadeh else 5393d14c5d2SYehuda Sadeh return req; 5403d14c5d2SYehuda Sadeh } 5413d14c5d2SYehuda Sadeh return NULL; 5423d14c5d2SYehuda Sadeh } 5433d14c5d2SYehuda Sadeh 5443d14c5d2SYehuda Sadeh static struct ceph_osd_request * 5453d14c5d2SYehuda Sadeh __lookup_request_ge(struct ceph_osd_client *osdc, 5463d14c5d2SYehuda Sadeh u64 tid) 5473d14c5d2SYehuda Sadeh { 5483d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 5493d14c5d2SYehuda Sadeh struct rb_node *n = osdc->requests.rb_node; 5503d14c5d2SYehuda Sadeh 5513d14c5d2SYehuda Sadeh while (n) { 5523d14c5d2SYehuda Sadeh req = rb_entry(n, struct ceph_osd_request, r_node); 5533d14c5d2SYehuda Sadeh if (tid < req->r_tid) { 5543d14c5d2SYehuda Sadeh if (!n->rb_left) 5553d14c5d2SYehuda Sadeh return req; 5563d14c5d2SYehuda Sadeh n = n->rb_left; 5573d14c5d2SYehuda Sadeh } else if (tid > req->r_tid) { 5583d14c5d2SYehuda Sadeh n = n->rb_right; 5593d14c5d2SYehuda Sadeh } else { 5603d14c5d2SYehuda Sadeh return req; 5613d14c5d2SYehuda Sadeh } 5623d14c5d2SYehuda Sadeh } 5633d14c5d2SYehuda Sadeh return NULL; 5643d14c5d2SYehuda Sadeh } 5653d14c5d2SYehuda Sadeh 5666f6c7006SSage Weil /* 5676f6c7006SSage Weil * Resubmit requests pending on the given osd. 5686f6c7006SSage Weil */ 5696f6c7006SSage Weil static void __kick_osd_requests(struct ceph_osd_client *osdc, 5706f6c7006SSage Weil struct ceph_osd *osd) 5716f6c7006SSage Weil { 572a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 5736f6c7006SSage Weil int err; 5746f6c7006SSage Weil 5756f6c7006SSage Weil dout("__kick_osd_requests osd%d\n", osd->o_osd); 5766f6c7006SSage Weil err = __reset_osd(osdc, osd); 577685a7555SAlex Elder if (err) 5786f6c7006SSage Weil return; 5796f6c7006SSage Weil 5806f6c7006SSage Weil list_for_each_entry(req, &osd->o_requests, r_osd_item) { 5816f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_unsent); 5826f6c7006SSage Weil dout("requeued %p tid %llu osd%d\n", req, req->r_tid, 5836f6c7006SSage Weil osd->o_osd); 584a40c4f10SYehuda Sadeh if (!req->r_linger) 5856f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 5866f6c7006SSage Weil } 587a40c4f10SYehuda Sadeh 588a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osd->o_linger_requests, 589a40c4f10SYehuda Sadeh r_linger_osd) { 59077f38e0eSSage Weil /* 59177f38e0eSSage Weil * reregister request prior to unregistering linger so 59277f38e0eSSage Weil * that r_osd is preserved. 59377f38e0eSSage Weil */ 59477f38e0eSSage Weil BUG_ON(!list_empty(&req->r_req_lru_item)); 595a40c4f10SYehuda Sadeh __register_request(osdc, req); 59677f38e0eSSage Weil list_add(&req->r_req_lru_item, &osdc->req_unsent); 59777f38e0eSSage Weil list_add(&req->r_osd_item, &req->r_osd->o_requests); 59877f38e0eSSage Weil __unregister_linger_request(osdc, req); 599a40c4f10SYehuda Sadeh dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid, 600a40c4f10SYehuda Sadeh osd->o_osd); 601a40c4f10SYehuda Sadeh } 6026f6c7006SSage Weil } 6036f6c7006SSage Weil 6043d14c5d2SYehuda Sadeh /* 6053d14c5d2SYehuda Sadeh * If the osd connection drops, we need to resubmit all requests. 6063d14c5d2SYehuda Sadeh */ 6073d14c5d2SYehuda Sadeh static void osd_reset(struct ceph_connection *con) 6083d14c5d2SYehuda Sadeh { 6093d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 6103d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 6113d14c5d2SYehuda Sadeh 6123d14c5d2SYehuda Sadeh if (!osd) 6133d14c5d2SYehuda Sadeh return; 6143d14c5d2SYehuda Sadeh dout("osd_reset osd%d\n", osd->o_osd); 6153d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 6163d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 61783aff95eSSage Weil mutex_lock(&osdc->request_mutex); 61883aff95eSSage Weil __kick_osd_requests(osdc, osd); 619f9d25199SAlex Elder __send_queued(osdc); 62083aff95eSSage Weil mutex_unlock(&osdc->request_mutex); 6213d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 6223d14c5d2SYehuda Sadeh } 6233d14c5d2SYehuda Sadeh 6243d14c5d2SYehuda Sadeh /* 6253d14c5d2SYehuda Sadeh * Track open sessions with osds. 6263d14c5d2SYehuda Sadeh */ 627e10006f8SAlex Elder static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum) 6283d14c5d2SYehuda Sadeh { 6293d14c5d2SYehuda Sadeh struct ceph_osd *osd; 6303d14c5d2SYehuda Sadeh 6313d14c5d2SYehuda Sadeh osd = kzalloc(sizeof(*osd), GFP_NOFS); 6323d14c5d2SYehuda Sadeh if (!osd) 6333d14c5d2SYehuda Sadeh return NULL; 6343d14c5d2SYehuda Sadeh 6353d14c5d2SYehuda Sadeh atomic_set(&osd->o_ref, 1); 6363d14c5d2SYehuda Sadeh osd->o_osdc = osdc; 637e10006f8SAlex Elder osd->o_osd = onum; 638f407731dSAlex Elder RB_CLEAR_NODE(&osd->o_node); 6393d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_requests); 640a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osd->o_linger_requests); 6413d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_osd_lru); 6423d14c5d2SYehuda Sadeh osd->o_incarnation = 1; 6433d14c5d2SYehuda Sadeh 644b7a9e5ddSSage Weil ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr); 6453d14c5d2SYehuda Sadeh 6463d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_keepalive_item); 6473d14c5d2SYehuda Sadeh return osd; 6483d14c5d2SYehuda Sadeh } 6493d14c5d2SYehuda Sadeh 6503d14c5d2SYehuda Sadeh static struct ceph_osd *get_osd(struct ceph_osd *osd) 6513d14c5d2SYehuda Sadeh { 6523d14c5d2SYehuda Sadeh if (atomic_inc_not_zero(&osd->o_ref)) { 6533d14c5d2SYehuda Sadeh dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1, 6543d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref)); 6553d14c5d2SYehuda Sadeh return osd; 6563d14c5d2SYehuda Sadeh } else { 6573d14c5d2SYehuda Sadeh dout("get_osd %p FAIL\n", osd); 6583d14c5d2SYehuda Sadeh return NULL; 6593d14c5d2SYehuda Sadeh } 6603d14c5d2SYehuda Sadeh } 6613d14c5d2SYehuda Sadeh 6623d14c5d2SYehuda Sadeh static void put_osd(struct ceph_osd *osd) 6633d14c5d2SYehuda Sadeh { 6643d14c5d2SYehuda Sadeh dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref), 6653d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref) - 1); 666a255651dSAlex Elder if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) { 6673d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth; 6683d14c5d2SYehuda Sadeh 66927859f97SSage Weil ceph_auth_destroy_authorizer(ac, osd->o_auth.authorizer); 6703d14c5d2SYehuda Sadeh kfree(osd); 6713d14c5d2SYehuda Sadeh } 6723d14c5d2SYehuda Sadeh } 6733d14c5d2SYehuda Sadeh 6743d14c5d2SYehuda Sadeh /* 6753d14c5d2SYehuda Sadeh * remove an osd from our map 6763d14c5d2SYehuda Sadeh */ 6773d14c5d2SYehuda Sadeh static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 6783d14c5d2SYehuda Sadeh { 6793d14c5d2SYehuda Sadeh dout("__remove_osd %p\n", osd); 6803d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_requests)); 6813d14c5d2SYehuda Sadeh rb_erase(&osd->o_node, &osdc->osds); 6823d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 6833d14c5d2SYehuda Sadeh ceph_con_close(&osd->o_con); 6843d14c5d2SYehuda Sadeh put_osd(osd); 6853d14c5d2SYehuda Sadeh } 6863d14c5d2SYehuda Sadeh 687aca420bcSSage Weil static void remove_all_osds(struct ceph_osd_client *osdc) 688aca420bcSSage Weil { 689048a9d2dSJiaju Zhang dout("%s %p\n", __func__, osdc); 690aca420bcSSage Weil mutex_lock(&osdc->request_mutex); 691aca420bcSSage Weil while (!RB_EMPTY_ROOT(&osdc->osds)) { 692aca420bcSSage Weil struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds), 693aca420bcSSage Weil struct ceph_osd, o_node); 694aca420bcSSage Weil __remove_osd(osdc, osd); 695aca420bcSSage Weil } 696aca420bcSSage Weil mutex_unlock(&osdc->request_mutex); 697aca420bcSSage Weil } 698aca420bcSSage Weil 6993d14c5d2SYehuda Sadeh static void __move_osd_to_lru(struct ceph_osd_client *osdc, 7003d14c5d2SYehuda Sadeh struct ceph_osd *osd) 7013d14c5d2SYehuda Sadeh { 7023d14c5d2SYehuda Sadeh dout("__move_osd_to_lru %p\n", osd); 7033d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_osd_lru)); 7043d14c5d2SYehuda Sadeh list_add_tail(&osd->o_osd_lru, &osdc->osd_lru); 7053d14c5d2SYehuda Sadeh osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ; 7063d14c5d2SYehuda Sadeh } 7073d14c5d2SYehuda Sadeh 7083d14c5d2SYehuda Sadeh static void __remove_osd_from_lru(struct ceph_osd *osd) 7093d14c5d2SYehuda Sadeh { 7103d14c5d2SYehuda Sadeh dout("__remove_osd_from_lru %p\n", osd); 7113d14c5d2SYehuda Sadeh if (!list_empty(&osd->o_osd_lru)) 7123d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 7133d14c5d2SYehuda Sadeh } 7143d14c5d2SYehuda Sadeh 715aca420bcSSage Weil static void remove_old_osds(struct ceph_osd_client *osdc) 7163d14c5d2SYehuda Sadeh { 7173d14c5d2SYehuda Sadeh struct ceph_osd *osd, *nosd; 7183d14c5d2SYehuda Sadeh 7193d14c5d2SYehuda Sadeh dout("__remove_old_osds %p\n", osdc); 7203d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 7213d14c5d2SYehuda Sadeh list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) { 722aca420bcSSage Weil if (time_before(jiffies, osd->lru_ttl)) 7233d14c5d2SYehuda Sadeh break; 7243d14c5d2SYehuda Sadeh __remove_osd(osdc, osd); 7253d14c5d2SYehuda Sadeh } 7263d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 7273d14c5d2SYehuda Sadeh } 7283d14c5d2SYehuda Sadeh 7293d14c5d2SYehuda Sadeh /* 7303d14c5d2SYehuda Sadeh * reset osd connect 7313d14c5d2SYehuda Sadeh */ 7323d14c5d2SYehuda Sadeh static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 7333d14c5d2SYehuda Sadeh { 734c3acb181SAlex Elder struct ceph_entity_addr *peer_addr; 7353d14c5d2SYehuda Sadeh 7363d14c5d2SYehuda Sadeh dout("__reset_osd %p osd%d\n", osd, osd->o_osd); 737a40c4f10SYehuda Sadeh if (list_empty(&osd->o_requests) && 738a40c4f10SYehuda Sadeh list_empty(&osd->o_linger_requests)) { 7393d14c5d2SYehuda Sadeh __remove_osd(osdc, osd); 740c3acb181SAlex Elder 741c3acb181SAlex Elder return -ENODEV; 742c3acb181SAlex Elder } 743c3acb181SAlex Elder 744c3acb181SAlex Elder peer_addr = &osdc->osdmap->osd_addr[osd->o_osd]; 745c3acb181SAlex Elder if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) && 7463d14c5d2SYehuda Sadeh !ceph_con_opened(&osd->o_con)) { 747c3acb181SAlex Elder struct ceph_osd_request *req; 748c3acb181SAlex Elder 7493d14c5d2SYehuda Sadeh dout(" osd addr hasn't changed and connection never opened," 7503d14c5d2SYehuda Sadeh " letting msgr retry"); 7513d14c5d2SYehuda Sadeh /* touch each r_stamp for handle_timeout()'s benfit */ 7523d14c5d2SYehuda Sadeh list_for_each_entry(req, &osd->o_requests, r_osd_item) 7533d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 754c3acb181SAlex Elder 755c3acb181SAlex Elder return -EAGAIN; 7563d14c5d2SYehuda Sadeh } 757c3acb181SAlex Elder 758c3acb181SAlex Elder ceph_con_close(&osd->o_con); 759c3acb181SAlex Elder ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr); 760c3acb181SAlex Elder osd->o_incarnation++; 761c3acb181SAlex Elder 762c3acb181SAlex Elder return 0; 7633d14c5d2SYehuda Sadeh } 7643d14c5d2SYehuda Sadeh 7653d14c5d2SYehuda Sadeh static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new) 7663d14c5d2SYehuda Sadeh { 7673d14c5d2SYehuda Sadeh struct rb_node **p = &osdc->osds.rb_node; 7683d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 7693d14c5d2SYehuda Sadeh struct ceph_osd *osd = NULL; 7703d14c5d2SYehuda Sadeh 771aca420bcSSage Weil dout("__insert_osd %p osd%d\n", new, new->o_osd); 7723d14c5d2SYehuda Sadeh while (*p) { 7733d14c5d2SYehuda Sadeh parent = *p; 7743d14c5d2SYehuda Sadeh osd = rb_entry(parent, struct ceph_osd, o_node); 7753d14c5d2SYehuda Sadeh if (new->o_osd < osd->o_osd) 7763d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 7773d14c5d2SYehuda Sadeh else if (new->o_osd > osd->o_osd) 7783d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 7793d14c5d2SYehuda Sadeh else 7803d14c5d2SYehuda Sadeh BUG(); 7813d14c5d2SYehuda Sadeh } 7823d14c5d2SYehuda Sadeh 7833d14c5d2SYehuda Sadeh rb_link_node(&new->o_node, parent, p); 7843d14c5d2SYehuda Sadeh rb_insert_color(&new->o_node, &osdc->osds); 7853d14c5d2SYehuda Sadeh } 7863d14c5d2SYehuda Sadeh 7873d14c5d2SYehuda Sadeh static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o) 7883d14c5d2SYehuda Sadeh { 7893d14c5d2SYehuda Sadeh struct ceph_osd *osd; 7903d14c5d2SYehuda Sadeh struct rb_node *n = osdc->osds.rb_node; 7913d14c5d2SYehuda Sadeh 7923d14c5d2SYehuda Sadeh while (n) { 7933d14c5d2SYehuda Sadeh osd = rb_entry(n, struct ceph_osd, o_node); 7943d14c5d2SYehuda Sadeh if (o < osd->o_osd) 7953d14c5d2SYehuda Sadeh n = n->rb_left; 7963d14c5d2SYehuda Sadeh else if (o > osd->o_osd) 7973d14c5d2SYehuda Sadeh n = n->rb_right; 7983d14c5d2SYehuda Sadeh else 7993d14c5d2SYehuda Sadeh return osd; 8003d14c5d2SYehuda Sadeh } 8013d14c5d2SYehuda Sadeh return NULL; 8023d14c5d2SYehuda Sadeh } 8033d14c5d2SYehuda Sadeh 8043d14c5d2SYehuda Sadeh static void __schedule_osd_timeout(struct ceph_osd_client *osdc) 8053d14c5d2SYehuda Sadeh { 8063d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->timeout_work, 8073d14c5d2SYehuda Sadeh osdc->client->options->osd_keepalive_timeout * HZ); 8083d14c5d2SYehuda Sadeh } 8093d14c5d2SYehuda Sadeh 8103d14c5d2SYehuda Sadeh static void __cancel_osd_timeout(struct ceph_osd_client *osdc) 8113d14c5d2SYehuda Sadeh { 8123d14c5d2SYehuda Sadeh cancel_delayed_work(&osdc->timeout_work); 8133d14c5d2SYehuda Sadeh } 8143d14c5d2SYehuda Sadeh 8153d14c5d2SYehuda Sadeh /* 8163d14c5d2SYehuda Sadeh * Register request, assign tid. If this is the first request, set up 8173d14c5d2SYehuda Sadeh * the timeout event. 8183d14c5d2SYehuda Sadeh */ 819a40c4f10SYehuda Sadeh static void __register_request(struct ceph_osd_client *osdc, 8203d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 8213d14c5d2SYehuda Sadeh { 8223d14c5d2SYehuda Sadeh req->r_tid = ++osdc->last_tid; 8233d14c5d2SYehuda Sadeh req->r_request->hdr.tid = cpu_to_le64(req->r_tid); 82477f38e0eSSage Weil dout("__register_request %p tid %lld\n", req, req->r_tid); 8253d14c5d2SYehuda Sadeh __insert_request(osdc, req); 8263d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 8273d14c5d2SYehuda Sadeh osdc->num_requests++; 8283d14c5d2SYehuda Sadeh if (osdc->num_requests == 1) { 8293d14c5d2SYehuda Sadeh dout(" first request, scheduling timeout\n"); 8303d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 8313d14c5d2SYehuda Sadeh } 832a40c4f10SYehuda Sadeh } 833a40c4f10SYehuda Sadeh 834a40c4f10SYehuda Sadeh static void register_request(struct ceph_osd_client *osdc, 835a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 836a40c4f10SYehuda Sadeh { 837a40c4f10SYehuda Sadeh mutex_lock(&osdc->request_mutex); 838a40c4f10SYehuda Sadeh __register_request(osdc, req); 8393d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 8403d14c5d2SYehuda Sadeh } 8413d14c5d2SYehuda Sadeh 8423d14c5d2SYehuda Sadeh /* 8433d14c5d2SYehuda Sadeh * called under osdc->request_mutex 8443d14c5d2SYehuda Sadeh */ 8453d14c5d2SYehuda Sadeh static void __unregister_request(struct ceph_osd_client *osdc, 8463d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 8473d14c5d2SYehuda Sadeh { 84835f9f8a0SSage Weil if (RB_EMPTY_NODE(&req->r_node)) { 84935f9f8a0SSage Weil dout("__unregister_request %p tid %lld not registered\n", 85035f9f8a0SSage Weil req, req->r_tid); 85135f9f8a0SSage Weil return; 85235f9f8a0SSage Weil } 85335f9f8a0SSage Weil 8543d14c5d2SYehuda Sadeh dout("__unregister_request %p tid %lld\n", req, req->r_tid); 8553d14c5d2SYehuda Sadeh rb_erase(&req->r_node, &osdc->requests); 8563d14c5d2SYehuda Sadeh osdc->num_requests--; 8573d14c5d2SYehuda Sadeh 8583d14c5d2SYehuda Sadeh if (req->r_osd) { 8593d14c5d2SYehuda Sadeh /* make sure the original request isn't in flight. */ 8606740a845SAlex Elder ceph_msg_revoke(req->r_request); 8613d14c5d2SYehuda Sadeh 8623d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 863a40c4f10SYehuda Sadeh if (list_empty(&req->r_osd->o_requests) && 864a40c4f10SYehuda Sadeh list_empty(&req->r_osd->o_linger_requests)) { 865a40c4f10SYehuda Sadeh dout("moving osd to %p lru\n", req->r_osd); 8663d14c5d2SYehuda Sadeh __move_osd_to_lru(osdc, req->r_osd); 867a40c4f10SYehuda Sadeh } 868fbdb9190SSage Weil if (list_empty(&req->r_linger_item)) 8693d14c5d2SYehuda Sadeh req->r_osd = NULL; 8703d14c5d2SYehuda Sadeh } 8713d14c5d2SYehuda Sadeh 8727d5f2481SAlex Elder list_del_init(&req->r_req_lru_item); 8733d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 8743d14c5d2SYehuda Sadeh 8753d14c5d2SYehuda Sadeh if (osdc->num_requests == 0) { 8763d14c5d2SYehuda Sadeh dout(" no requests, canceling timeout\n"); 8773d14c5d2SYehuda Sadeh __cancel_osd_timeout(osdc); 8783d14c5d2SYehuda Sadeh } 8793d14c5d2SYehuda Sadeh } 8803d14c5d2SYehuda Sadeh 8813d14c5d2SYehuda Sadeh /* 8823d14c5d2SYehuda Sadeh * Cancel a previously queued request message 8833d14c5d2SYehuda Sadeh */ 8843d14c5d2SYehuda Sadeh static void __cancel_request(struct ceph_osd_request *req) 8853d14c5d2SYehuda Sadeh { 8863d14c5d2SYehuda Sadeh if (req->r_sent && req->r_osd) { 8876740a845SAlex Elder ceph_msg_revoke(req->r_request); 8883d14c5d2SYehuda Sadeh req->r_sent = 0; 8893d14c5d2SYehuda Sadeh } 8903d14c5d2SYehuda Sadeh } 8913d14c5d2SYehuda Sadeh 892a40c4f10SYehuda Sadeh static void __register_linger_request(struct ceph_osd_client *osdc, 893a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 894a40c4f10SYehuda Sadeh { 895a40c4f10SYehuda Sadeh dout("__register_linger_request %p\n", req); 896a40c4f10SYehuda Sadeh list_add_tail(&req->r_linger_item, &osdc->req_linger); 8976194ea89SSage Weil if (req->r_osd) 8986194ea89SSage Weil list_add_tail(&req->r_linger_osd, 8996194ea89SSage Weil &req->r_osd->o_linger_requests); 900a40c4f10SYehuda Sadeh } 901a40c4f10SYehuda Sadeh 902a40c4f10SYehuda Sadeh static void __unregister_linger_request(struct ceph_osd_client *osdc, 903a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 904a40c4f10SYehuda Sadeh { 905a40c4f10SYehuda Sadeh dout("__unregister_linger_request %p\n", req); 906a40c4f10SYehuda Sadeh list_del_init(&req->r_linger_item); 90761c74035SAlex Elder if (req->r_osd) { 908a40c4f10SYehuda Sadeh list_del_init(&req->r_linger_osd); 909a40c4f10SYehuda Sadeh 910a40c4f10SYehuda Sadeh if (list_empty(&req->r_osd->o_requests) && 911a40c4f10SYehuda Sadeh list_empty(&req->r_osd->o_linger_requests)) { 912a40c4f10SYehuda Sadeh dout("moving osd to %p lru\n", req->r_osd); 913a40c4f10SYehuda Sadeh __move_osd_to_lru(osdc, req->r_osd); 914a40c4f10SYehuda Sadeh } 915fbdb9190SSage Weil if (list_empty(&req->r_osd_item)) 916a40c4f10SYehuda Sadeh req->r_osd = NULL; 917a40c4f10SYehuda Sadeh } 918a40c4f10SYehuda Sadeh } 919a40c4f10SYehuda Sadeh 920a40c4f10SYehuda Sadeh void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc, 921a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 922a40c4f10SYehuda Sadeh { 923a40c4f10SYehuda Sadeh mutex_lock(&osdc->request_mutex); 924a40c4f10SYehuda Sadeh if (req->r_linger) { 925a40c4f10SYehuda Sadeh __unregister_linger_request(osdc, req); 926a40c4f10SYehuda Sadeh ceph_osdc_put_request(req); 927a40c4f10SYehuda Sadeh } 928a40c4f10SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 929a40c4f10SYehuda Sadeh } 930a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_unregister_linger_request); 931a40c4f10SYehuda Sadeh 932a40c4f10SYehuda Sadeh void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc, 933a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 934a40c4f10SYehuda Sadeh { 935a40c4f10SYehuda Sadeh if (!req->r_linger) { 936a40c4f10SYehuda Sadeh dout("set_request_linger %p\n", req); 937a40c4f10SYehuda Sadeh req->r_linger = 1; 938a40c4f10SYehuda Sadeh /* 939a40c4f10SYehuda Sadeh * caller is now responsible for calling 940a40c4f10SYehuda Sadeh * unregister_linger_request 941a40c4f10SYehuda Sadeh */ 942a40c4f10SYehuda Sadeh ceph_osdc_get_request(req); 943a40c4f10SYehuda Sadeh } 944a40c4f10SYehuda Sadeh } 945a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_set_request_linger); 946a40c4f10SYehuda Sadeh 9473d14c5d2SYehuda Sadeh /* 9483d14c5d2SYehuda Sadeh * Pick an osd (the first 'up' osd in the pg), allocate the osd struct 9493d14c5d2SYehuda Sadeh * (as needed), and set the request r_osd appropriately. If there is 95025985edcSLucas De Marchi * no up osd, set r_osd to NULL. Move the request to the appropriate list 9516f6c7006SSage Weil * (unsent, homeless) or leave on in-flight lru. 9523d14c5d2SYehuda Sadeh * 9533d14c5d2SYehuda Sadeh * Return 0 if unchanged, 1 if changed, or negative on error. 9543d14c5d2SYehuda Sadeh * 9553d14c5d2SYehuda Sadeh * Caller should hold map_sem for read and request_mutex. 9563d14c5d2SYehuda Sadeh */ 9576f6c7006SSage Weil static int __map_request(struct ceph_osd_client *osdc, 95838d6453cSSage Weil struct ceph_osd_request *req, int force_resend) 9593d14c5d2SYehuda Sadeh { 9605b191d99SSage Weil struct ceph_pg pgid; 9613d14c5d2SYehuda Sadeh int acting[CEPH_PG_MAX_SIZE]; 9623d14c5d2SYehuda Sadeh int o = -1, num = 0; 9633d14c5d2SYehuda Sadeh int err; 9643d14c5d2SYehuda Sadeh 9656f6c7006SSage Weil dout("map_request %p tid %lld\n", req, req->r_tid); 96641766f87SAlex Elder err = ceph_calc_ceph_pg(&pgid, req->r_oid, osdc->osdmap, 96741766f87SAlex Elder ceph_file_layout_pg_pool(req->r_file_layout)); 9686f6c7006SSage Weil if (err) { 9696f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 9703d14c5d2SYehuda Sadeh return err; 9716f6c7006SSage Weil } 9723d14c5d2SYehuda Sadeh req->r_pgid = pgid; 9733d14c5d2SYehuda Sadeh 9743d14c5d2SYehuda Sadeh err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting); 9753d14c5d2SYehuda Sadeh if (err > 0) { 9763d14c5d2SYehuda Sadeh o = acting[0]; 9773d14c5d2SYehuda Sadeh num = err; 9783d14c5d2SYehuda Sadeh } 9793d14c5d2SYehuda Sadeh 98038d6453cSSage Weil if ((!force_resend && 98138d6453cSSage Weil req->r_osd && req->r_osd->o_osd == o && 9823d14c5d2SYehuda Sadeh req->r_sent >= req->r_osd->o_incarnation && 9833d14c5d2SYehuda Sadeh req->r_num_pg_osds == num && 9843d14c5d2SYehuda Sadeh memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) || 9853d14c5d2SYehuda Sadeh (req->r_osd == NULL && o == -1)) 9863d14c5d2SYehuda Sadeh return 0; /* no change */ 9873d14c5d2SYehuda Sadeh 9885b191d99SSage Weil dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n", 9895b191d99SSage Weil req->r_tid, pgid.pool, pgid.seed, o, 9903d14c5d2SYehuda Sadeh req->r_osd ? req->r_osd->o_osd : -1); 9913d14c5d2SYehuda Sadeh 9923d14c5d2SYehuda Sadeh /* record full pg acting set */ 9933d14c5d2SYehuda Sadeh memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num); 9943d14c5d2SYehuda Sadeh req->r_num_pg_osds = num; 9953d14c5d2SYehuda Sadeh 9963d14c5d2SYehuda Sadeh if (req->r_osd) { 9973d14c5d2SYehuda Sadeh __cancel_request(req); 9983d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 9993d14c5d2SYehuda Sadeh req->r_osd = NULL; 10003d14c5d2SYehuda Sadeh } 10013d14c5d2SYehuda Sadeh 10023d14c5d2SYehuda Sadeh req->r_osd = __lookup_osd(osdc, o); 10033d14c5d2SYehuda Sadeh if (!req->r_osd && o >= 0) { 10043d14c5d2SYehuda Sadeh err = -ENOMEM; 1005e10006f8SAlex Elder req->r_osd = create_osd(osdc, o); 10066f6c7006SSage Weil if (!req->r_osd) { 10076f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 10083d14c5d2SYehuda Sadeh goto out; 10096f6c7006SSage Weil } 10103d14c5d2SYehuda Sadeh 10116f6c7006SSage Weil dout("map_request osd %p is osd%d\n", req->r_osd, o); 10123d14c5d2SYehuda Sadeh __insert_osd(osdc, req->r_osd); 10133d14c5d2SYehuda Sadeh 1014b7a9e5ddSSage Weil ceph_con_open(&req->r_osd->o_con, 1015b7a9e5ddSSage Weil CEPH_ENTITY_TYPE_OSD, o, 1016b7a9e5ddSSage Weil &osdc->osdmap->osd_addr[o]); 10173d14c5d2SYehuda Sadeh } 10183d14c5d2SYehuda Sadeh 10193d14c5d2SYehuda Sadeh if (req->r_osd) { 10203d14c5d2SYehuda Sadeh __remove_osd_from_lru(req->r_osd); 10213d14c5d2SYehuda Sadeh list_add(&req->r_osd_item, &req->r_osd->o_requests); 10226f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_unsent); 10236f6c7006SSage Weil } else { 10246f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 10253d14c5d2SYehuda Sadeh } 10263d14c5d2SYehuda Sadeh err = 1; /* osd or pg changed */ 10273d14c5d2SYehuda Sadeh 10283d14c5d2SYehuda Sadeh out: 10293d14c5d2SYehuda Sadeh return err; 10303d14c5d2SYehuda Sadeh } 10313d14c5d2SYehuda Sadeh 10323d14c5d2SYehuda Sadeh /* 10333d14c5d2SYehuda Sadeh * caller should hold map_sem (for read) and request_mutex 10343d14c5d2SYehuda Sadeh */ 103556e925b6SSage Weil static void __send_request(struct ceph_osd_client *osdc, 10363d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 10373d14c5d2SYehuda Sadeh { 10381b83bef2SSage Weil void *p; 10393d14c5d2SYehuda Sadeh 10401b83bef2SSage Weil dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n", 10411b83bef2SSage Weil req, req->r_tid, req->r_osd->o_osd, req->r_flags, 10421b83bef2SSage Weil (unsigned long long)req->r_pgid.pool, req->r_pgid.seed); 10433d14c5d2SYehuda Sadeh 10441b83bef2SSage Weil /* fill in message content that changes each time we send it */ 10451b83bef2SSage Weil put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch); 10461b83bef2SSage Weil put_unaligned_le32(req->r_flags, req->r_request_flags); 10471b83bef2SSage Weil put_unaligned_le64(req->r_pgid.pool, req->r_request_pool); 10481b83bef2SSage Weil p = req->r_request_pgid; 10491b83bef2SSage Weil ceph_encode_64(&p, req->r_pgid.pool); 10501b83bef2SSage Weil ceph_encode_32(&p, req->r_pgid.seed); 10511b83bef2SSage Weil put_unaligned_le64(1, req->r_request_attempts); /* FIXME */ 10521b83bef2SSage Weil memcpy(req->r_request_reassert_version, &req->r_reassert_version, 10531b83bef2SSage Weil sizeof(req->r_reassert_version)); 10542169aea6SSage Weil 10553d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 10563d14c5d2SYehuda Sadeh list_move_tail(&req->r_req_lru_item, &osdc->req_lru); 10573d14c5d2SYehuda Sadeh 10583d14c5d2SYehuda Sadeh ceph_msg_get(req->r_request); /* send consumes a ref */ 10593d14c5d2SYehuda Sadeh ceph_con_send(&req->r_osd->o_con, req->r_request); 10603d14c5d2SYehuda Sadeh req->r_sent = req->r_osd->o_incarnation; 10613d14c5d2SYehuda Sadeh } 10623d14c5d2SYehuda Sadeh 10633d14c5d2SYehuda Sadeh /* 10646f6c7006SSage Weil * Send any requests in the queue (req_unsent). 10656f6c7006SSage Weil */ 1066f9d25199SAlex Elder static void __send_queued(struct ceph_osd_client *osdc) 10676f6c7006SSage Weil { 10686f6c7006SSage Weil struct ceph_osd_request *req, *tmp; 10696f6c7006SSage Weil 1070f9d25199SAlex Elder dout("__send_queued\n"); 1071f9d25199SAlex Elder list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) 10726f6c7006SSage Weil __send_request(osdc, req); 10736f6c7006SSage Weil } 10746f6c7006SSage Weil 10756f6c7006SSage Weil /* 10763d14c5d2SYehuda Sadeh * Timeout callback, called every N seconds when 1 or more osd 10773d14c5d2SYehuda Sadeh * requests has been active for more than N seconds. When this 10783d14c5d2SYehuda Sadeh * happens, we ping all OSDs with requests who have timed out to 10793d14c5d2SYehuda Sadeh * ensure any communications channel reset is detected. Reset the 10803d14c5d2SYehuda Sadeh * request timeouts another N seconds in the future as we go. 10813d14c5d2SYehuda Sadeh * Reschedule the timeout event another N seconds in future (unless 10823d14c5d2SYehuda Sadeh * there are no open requests). 10833d14c5d2SYehuda Sadeh */ 10843d14c5d2SYehuda Sadeh static void handle_timeout(struct work_struct *work) 10853d14c5d2SYehuda Sadeh { 10863d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 10873d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, timeout_work.work); 108883aff95eSSage Weil struct ceph_osd_request *req; 10893d14c5d2SYehuda Sadeh struct ceph_osd *osd; 10903d14c5d2SYehuda Sadeh unsigned long keepalive = 10913d14c5d2SYehuda Sadeh osdc->client->options->osd_keepalive_timeout * HZ; 10923d14c5d2SYehuda Sadeh struct list_head slow_osds; 10933d14c5d2SYehuda Sadeh dout("timeout\n"); 10943d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 10953d14c5d2SYehuda Sadeh 10963d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 10973d14c5d2SYehuda Sadeh 10983d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 10993d14c5d2SYehuda Sadeh 11003d14c5d2SYehuda Sadeh /* 11013d14c5d2SYehuda Sadeh * ping osds that are a bit slow. this ensures that if there 11023d14c5d2SYehuda Sadeh * is a break in the TCP connection we will notice, and reopen 11033d14c5d2SYehuda Sadeh * a connection with that osd (from the fault callback). 11043d14c5d2SYehuda Sadeh */ 11053d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&slow_osds); 11063d14c5d2SYehuda Sadeh list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) { 11073d14c5d2SYehuda Sadeh if (time_before(jiffies, req->r_stamp + keepalive)) 11083d14c5d2SYehuda Sadeh break; 11093d14c5d2SYehuda Sadeh 11103d14c5d2SYehuda Sadeh osd = req->r_osd; 11113d14c5d2SYehuda Sadeh BUG_ON(!osd); 11123d14c5d2SYehuda Sadeh dout(" tid %llu is slow, will send keepalive on osd%d\n", 11133d14c5d2SYehuda Sadeh req->r_tid, osd->o_osd); 11143d14c5d2SYehuda Sadeh list_move_tail(&osd->o_keepalive_item, &slow_osds); 11153d14c5d2SYehuda Sadeh } 11163d14c5d2SYehuda Sadeh while (!list_empty(&slow_osds)) { 11173d14c5d2SYehuda Sadeh osd = list_entry(slow_osds.next, struct ceph_osd, 11183d14c5d2SYehuda Sadeh o_keepalive_item); 11193d14c5d2SYehuda Sadeh list_del_init(&osd->o_keepalive_item); 11203d14c5d2SYehuda Sadeh ceph_con_keepalive(&osd->o_con); 11213d14c5d2SYehuda Sadeh } 11223d14c5d2SYehuda Sadeh 11233d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 1124f9d25199SAlex Elder __send_queued(osdc); 11253d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 11263d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 11273d14c5d2SYehuda Sadeh } 11283d14c5d2SYehuda Sadeh 11293d14c5d2SYehuda Sadeh static void handle_osds_timeout(struct work_struct *work) 11303d14c5d2SYehuda Sadeh { 11313d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 11323d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, 11333d14c5d2SYehuda Sadeh osds_timeout_work.work); 11343d14c5d2SYehuda Sadeh unsigned long delay = 11353d14c5d2SYehuda Sadeh osdc->client->options->osd_idle_ttl * HZ >> 2; 11363d14c5d2SYehuda Sadeh 11373d14c5d2SYehuda Sadeh dout("osds timeout\n"); 11383d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 1139aca420bcSSage Weil remove_old_osds(osdc); 11403d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 11413d14c5d2SYehuda Sadeh 11423d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 11433d14c5d2SYehuda Sadeh round_jiffies_relative(delay)); 11443d14c5d2SYehuda Sadeh } 11453d14c5d2SYehuda Sadeh 114625845472SSage Weil static void complete_request(struct ceph_osd_request *req) 114725845472SSage Weil { 114825845472SSage Weil if (req->r_safe_callback) 114925845472SSage Weil req->r_safe_callback(req, NULL); 115025845472SSage Weil complete_all(&req->r_safe_completion); /* fsync waiter */ 115125845472SSage Weil } 115225845472SSage Weil 11531b83bef2SSage Weil static int __decode_pgid(void **p, void *end, struct ceph_pg *pgid) 11541b83bef2SSage Weil { 11551b83bef2SSage Weil __u8 v; 11561b83bef2SSage Weil 11571b83bef2SSage Weil ceph_decode_need(p, end, 1 + 8 + 4 + 4, bad); 11581b83bef2SSage Weil v = ceph_decode_8(p); 11591b83bef2SSage Weil if (v > 1) { 11601b83bef2SSage Weil pr_warning("do not understand pg encoding %d > 1", v); 11611b83bef2SSage Weil return -EINVAL; 11621b83bef2SSage Weil } 11631b83bef2SSage Weil pgid->pool = ceph_decode_64(p); 11641b83bef2SSage Weil pgid->seed = ceph_decode_32(p); 11651b83bef2SSage Weil *p += 4; 11661b83bef2SSage Weil return 0; 11671b83bef2SSage Weil 11681b83bef2SSage Weil bad: 11691b83bef2SSage Weil pr_warning("incomplete pg encoding"); 11701b83bef2SSage Weil return -EINVAL; 11711b83bef2SSage Weil } 11721b83bef2SSage Weil 11733d14c5d2SYehuda Sadeh /* 11743d14c5d2SYehuda Sadeh * handle osd op reply. either call the callback if it is specified, 11753d14c5d2SYehuda Sadeh * or do the completion to wake up the waiting thread. 11763d14c5d2SYehuda Sadeh */ 11773d14c5d2SYehuda Sadeh static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg, 11783d14c5d2SYehuda Sadeh struct ceph_connection *con) 11793d14c5d2SYehuda Sadeh { 11801b83bef2SSage Weil void *p, *end; 11813d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 11823d14c5d2SYehuda Sadeh u64 tid; 11831b83bef2SSage Weil int object_len; 11841b83bef2SSage Weil int numops, payload_len, flags; 11853d14c5d2SYehuda Sadeh s32 result; 11861b83bef2SSage Weil s32 retry_attempt; 11871b83bef2SSage Weil struct ceph_pg pg; 11881b83bef2SSage Weil int err; 11891b83bef2SSage Weil u32 reassert_epoch; 11901b83bef2SSage Weil u64 reassert_version; 11911b83bef2SSage Weil u32 osdmap_epoch; 11920d5af164SAlex Elder int already_completed; 11931b83bef2SSage Weil int i; 11943d14c5d2SYehuda Sadeh 11953d14c5d2SYehuda Sadeh tid = le64_to_cpu(msg->hdr.tid); 11961b83bef2SSage Weil dout("handle_reply %p tid %llu\n", msg, tid); 11971b83bef2SSage Weil 11981b83bef2SSage Weil p = msg->front.iov_base; 11991b83bef2SSage Weil end = p + msg->front.iov_len; 12001b83bef2SSage Weil 12011b83bef2SSage Weil ceph_decode_need(&p, end, 4, bad); 12021b83bef2SSage Weil object_len = ceph_decode_32(&p); 12031b83bef2SSage Weil ceph_decode_need(&p, end, object_len, bad); 12041b83bef2SSage Weil p += object_len; 12051b83bef2SSage Weil 12061b83bef2SSage Weil err = __decode_pgid(&p, end, &pg); 12071b83bef2SSage Weil if (err) 12083d14c5d2SYehuda Sadeh goto bad; 12091b83bef2SSage Weil 12101b83bef2SSage Weil ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad); 12111b83bef2SSage Weil flags = ceph_decode_64(&p); 12121b83bef2SSage Weil result = ceph_decode_32(&p); 12131b83bef2SSage Weil reassert_epoch = ceph_decode_32(&p); 12141b83bef2SSage Weil reassert_version = ceph_decode_64(&p); 12151b83bef2SSage Weil osdmap_epoch = ceph_decode_32(&p); 12161b83bef2SSage Weil 12173d14c5d2SYehuda Sadeh /* lookup */ 12183d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 12193d14c5d2SYehuda Sadeh req = __lookup_request(osdc, tid); 12203d14c5d2SYehuda Sadeh if (req == NULL) { 12213d14c5d2SYehuda Sadeh dout("handle_reply tid %llu dne\n", tid); 12223d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 12233d14c5d2SYehuda Sadeh return; 12243d14c5d2SYehuda Sadeh } 12253d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 12261b83bef2SSage Weil 12271b83bef2SSage Weil dout("handle_reply %p tid %llu req %p result %d\n", msg, tid, 12281b83bef2SSage Weil req, result); 12291b83bef2SSage Weil 12301b83bef2SSage Weil ceph_decode_need(&p, end, 4, bad); 12311b83bef2SSage Weil numops = ceph_decode_32(&p); 12321b83bef2SSage Weil if (numops > CEPH_OSD_MAX_OP) 12331b83bef2SSage Weil goto bad_put; 12341b83bef2SSage Weil if (numops != req->r_num_ops) 12351b83bef2SSage Weil goto bad_put; 12361b83bef2SSage Weil payload_len = 0; 12371b83bef2SSage Weil ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad); 12381b83bef2SSage Weil for (i = 0; i < numops; i++) { 12391b83bef2SSage Weil struct ceph_osd_op *op = p; 12401b83bef2SSage Weil int len; 12411b83bef2SSage Weil 12421b83bef2SSage Weil len = le32_to_cpu(op->payload_len); 12431b83bef2SSage Weil req->r_reply_op_len[i] = len; 12441b83bef2SSage Weil dout(" op %d has %d bytes\n", i, len); 12451b83bef2SSage Weil payload_len += len; 12461b83bef2SSage Weil p += sizeof(*op); 12471b83bef2SSage Weil } 12481b83bef2SSage Weil if (payload_len != le32_to_cpu(msg->hdr.data_len)) { 12491b83bef2SSage Weil pr_warning("sum of op payload lens %d != data_len %d", 12501b83bef2SSage Weil payload_len, le32_to_cpu(msg->hdr.data_len)); 12511b83bef2SSage Weil goto bad_put; 12521b83bef2SSage Weil } 12531b83bef2SSage Weil 12541b83bef2SSage Weil ceph_decode_need(&p, end, 4 + numops * 4, bad); 12551b83bef2SSage Weil retry_attempt = ceph_decode_32(&p); 12561b83bef2SSage Weil for (i = 0; i < numops; i++) 12571b83bef2SSage Weil req->r_reply_op_result[i] = ceph_decode_32(&p); 12583d14c5d2SYehuda Sadeh 12593d14c5d2SYehuda Sadeh /* 12603d14c5d2SYehuda Sadeh * if this connection filled our message, drop our reference now, to 12613d14c5d2SYehuda Sadeh * avoid a (safe but slower) revoke later. 12623d14c5d2SYehuda Sadeh */ 12633d14c5d2SYehuda Sadeh if (req->r_con_filling_msg == con && req->r_reply == msg) { 12643d14c5d2SYehuda Sadeh dout(" dropping con_filling_msg ref %p\n", con); 12653d14c5d2SYehuda Sadeh req->r_con_filling_msg = NULL; 12660d47766fSSage Weil con->ops->put(con); 12673d14c5d2SYehuda Sadeh } 12683d14c5d2SYehuda Sadeh 12693d14c5d2SYehuda Sadeh if (!req->r_got_reply) { 127095c96174SEric Dumazet unsigned int bytes; 12713d14c5d2SYehuda Sadeh 12721b83bef2SSage Weil req->r_result = result; 12733d14c5d2SYehuda Sadeh bytes = le32_to_cpu(msg->hdr.data_len); 12743d14c5d2SYehuda Sadeh dout("handle_reply result %d bytes %d\n", req->r_result, 12753d14c5d2SYehuda Sadeh bytes); 12763d14c5d2SYehuda Sadeh if (req->r_result == 0) 12773d14c5d2SYehuda Sadeh req->r_result = bytes; 12783d14c5d2SYehuda Sadeh 12793d14c5d2SYehuda Sadeh /* in case this is a write and we need to replay, */ 12801b83bef2SSage Weil req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch); 12811b83bef2SSage Weil req->r_reassert_version.version = cpu_to_le64(reassert_version); 12823d14c5d2SYehuda Sadeh 12833d14c5d2SYehuda Sadeh req->r_got_reply = 1; 12843d14c5d2SYehuda Sadeh } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) { 12853d14c5d2SYehuda Sadeh dout("handle_reply tid %llu dup ack\n", tid); 12863d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 12873d14c5d2SYehuda Sadeh goto done; 12883d14c5d2SYehuda Sadeh } 12893d14c5d2SYehuda Sadeh 12903d14c5d2SYehuda Sadeh dout("handle_reply tid %llu flags %d\n", tid, flags); 12913d14c5d2SYehuda Sadeh 1292a40c4f10SYehuda Sadeh if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK)) 1293a40c4f10SYehuda Sadeh __register_linger_request(osdc, req); 1294a40c4f10SYehuda Sadeh 12953d14c5d2SYehuda Sadeh /* either this is a read, or we got the safe response */ 12963d14c5d2SYehuda Sadeh if (result < 0 || 12973d14c5d2SYehuda Sadeh (flags & CEPH_OSD_FLAG_ONDISK) || 12983d14c5d2SYehuda Sadeh ((flags & CEPH_OSD_FLAG_WRITE) == 0)) 12993d14c5d2SYehuda Sadeh __unregister_request(osdc, req); 13003d14c5d2SYehuda Sadeh 13010d5af164SAlex Elder already_completed = req->r_completed; 13020d5af164SAlex Elder req->r_completed = 1; 13033d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 13040d5af164SAlex Elder if (already_completed) 13050d5af164SAlex Elder goto done; 13063d14c5d2SYehuda Sadeh 13073d14c5d2SYehuda Sadeh if (req->r_callback) 13083d14c5d2SYehuda Sadeh req->r_callback(req, msg); 13093d14c5d2SYehuda Sadeh else 13103d14c5d2SYehuda Sadeh complete_all(&req->r_completion); 13113d14c5d2SYehuda Sadeh 131225845472SSage Weil if (flags & CEPH_OSD_FLAG_ONDISK) 131325845472SSage Weil complete_request(req); 13143d14c5d2SYehuda Sadeh 13153d14c5d2SYehuda Sadeh done: 1316a40c4f10SYehuda Sadeh dout("req=%p req->r_linger=%d\n", req, req->r_linger); 13173d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 13183d14c5d2SYehuda Sadeh return; 13193d14c5d2SYehuda Sadeh 13201b83bef2SSage Weil bad_put: 13211b83bef2SSage Weil ceph_osdc_put_request(req); 13223d14c5d2SYehuda Sadeh bad: 13231b83bef2SSage Weil pr_err("corrupt osd_op_reply got %d %d\n", 13241b83bef2SSage Weil (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len)); 13253d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 13263d14c5d2SYehuda Sadeh } 13273d14c5d2SYehuda Sadeh 13286f6c7006SSage Weil static void reset_changed_osds(struct ceph_osd_client *osdc) 13293d14c5d2SYehuda Sadeh { 13303d14c5d2SYehuda Sadeh struct rb_node *p, *n; 13313d14c5d2SYehuda Sadeh 13323d14c5d2SYehuda Sadeh for (p = rb_first(&osdc->osds); p; p = n) { 13336f6c7006SSage Weil struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node); 13343d14c5d2SYehuda Sadeh 13353d14c5d2SYehuda Sadeh n = rb_next(p); 13363d14c5d2SYehuda Sadeh if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) || 13373d14c5d2SYehuda Sadeh memcmp(&osd->o_con.peer_addr, 13383d14c5d2SYehuda Sadeh ceph_osd_addr(osdc->osdmap, 13393d14c5d2SYehuda Sadeh osd->o_osd), 13403d14c5d2SYehuda Sadeh sizeof(struct ceph_entity_addr)) != 0) 13413d14c5d2SYehuda Sadeh __reset_osd(osdc, osd); 13423d14c5d2SYehuda Sadeh } 13433d14c5d2SYehuda Sadeh } 13443d14c5d2SYehuda Sadeh 13453d14c5d2SYehuda Sadeh /* 13466f6c7006SSage Weil * Requeue requests whose mapping to an OSD has changed. If requests map to 13476f6c7006SSage Weil * no osd, request a new map. 13483d14c5d2SYehuda Sadeh * 1349e6d50f67SAlex Elder * Caller should hold map_sem for read. 13503d14c5d2SYehuda Sadeh */ 135138d6453cSSage Weil static void kick_requests(struct ceph_osd_client *osdc, int force_resend) 13523d14c5d2SYehuda Sadeh { 1353a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 13546f6c7006SSage Weil struct rb_node *p; 13556f6c7006SSage Weil int needmap = 0; 13566f6c7006SSage Weil int err; 13573d14c5d2SYehuda Sadeh 135838d6453cSSage Weil dout("kick_requests %s\n", force_resend ? " (force resend)" : ""); 13593d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 13606194ea89SSage Weil for (p = rb_first(&osdc->requests); p; ) { 13616f6c7006SSage Weil req = rb_entry(p, struct ceph_osd_request, r_node); 13626194ea89SSage Weil p = rb_next(p); 1363ab60b16dSAlex Elder 1364ab60b16dSAlex Elder /* 1365ab60b16dSAlex Elder * For linger requests that have not yet been 1366ab60b16dSAlex Elder * registered, move them to the linger list; they'll 1367ab60b16dSAlex Elder * be sent to the osd in the loop below. Unregister 1368ab60b16dSAlex Elder * the request before re-registering it as a linger 1369ab60b16dSAlex Elder * request to ensure the __map_request() below 1370ab60b16dSAlex Elder * will decide it needs to be sent. 1371ab60b16dSAlex Elder */ 1372ab60b16dSAlex Elder if (req->r_linger && list_empty(&req->r_linger_item)) { 1373ab60b16dSAlex Elder dout("%p tid %llu restart on osd%d\n", 1374ab60b16dSAlex Elder req, req->r_tid, 1375ab60b16dSAlex Elder req->r_osd ? req->r_osd->o_osd : -1); 1376ab60b16dSAlex Elder __unregister_request(osdc, req); 1377ab60b16dSAlex Elder __register_linger_request(osdc, req); 1378ab60b16dSAlex Elder continue; 1379ab60b16dSAlex Elder } 1380ab60b16dSAlex Elder 138138d6453cSSage Weil err = __map_request(osdc, req, force_resend); 13826f6c7006SSage Weil if (err < 0) 13836f6c7006SSage Weil continue; /* error */ 13846f6c7006SSage Weil if (req->r_osd == NULL) { 13856f6c7006SSage Weil dout("%p tid %llu maps to no osd\n", req, req->r_tid); 13866f6c7006SSage Weil needmap++; /* request a newer map */ 13876f6c7006SSage Weil } else if (err > 0) { 13886194ea89SSage Weil if (!req->r_linger) { 13896194ea89SSage Weil dout("%p tid %llu requeued on osd%d\n", req, 13906194ea89SSage Weil req->r_tid, 13916f6c7006SSage Weil req->r_osd ? req->r_osd->o_osd : -1); 13926f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 13936f6c7006SSage Weil } 13946f6c7006SSage Weil } 13956194ea89SSage Weil } 1396a40c4f10SYehuda Sadeh 1397a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osdc->req_linger, 1398a40c4f10SYehuda Sadeh r_linger_item) { 1399a40c4f10SYehuda Sadeh dout("linger req=%p req->r_osd=%p\n", req, req->r_osd); 1400a40c4f10SYehuda Sadeh 140138d6453cSSage Weil err = __map_request(osdc, req, force_resend); 1402ab60b16dSAlex Elder dout("__map_request returned %d\n", err); 1403a40c4f10SYehuda Sadeh if (err == 0) 1404a40c4f10SYehuda Sadeh continue; /* no change and no osd was specified */ 1405a40c4f10SYehuda Sadeh if (err < 0) 1406a40c4f10SYehuda Sadeh continue; /* hrm! */ 1407a40c4f10SYehuda Sadeh if (req->r_osd == NULL) { 1408a40c4f10SYehuda Sadeh dout("tid %llu maps to no valid osd\n", req->r_tid); 1409a40c4f10SYehuda Sadeh needmap++; /* request a newer map */ 1410a40c4f10SYehuda Sadeh continue; 1411a40c4f10SYehuda Sadeh } 1412a40c4f10SYehuda Sadeh 1413a40c4f10SYehuda Sadeh dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid, 1414a40c4f10SYehuda Sadeh req->r_osd ? req->r_osd->o_osd : -1); 1415a40c4f10SYehuda Sadeh __register_request(osdc, req); 1416c89ce05eSAlex Elder __unregister_linger_request(osdc, req); 1417a40c4f10SYehuda Sadeh } 14183d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 14193d14c5d2SYehuda Sadeh 14203d14c5d2SYehuda Sadeh if (needmap) { 14213d14c5d2SYehuda Sadeh dout("%d requests for down osds, need new map\n", needmap); 14223d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 14233d14c5d2SYehuda Sadeh } 1424e6d50f67SAlex Elder reset_changed_osds(osdc); 14253d14c5d2SYehuda Sadeh } 14266f6c7006SSage Weil 14276f6c7006SSage Weil 14283d14c5d2SYehuda Sadeh /* 14293d14c5d2SYehuda Sadeh * Process updated osd map. 14303d14c5d2SYehuda Sadeh * 14313d14c5d2SYehuda Sadeh * The message contains any number of incremental and full maps, normally 14323d14c5d2SYehuda Sadeh * indicating some sort of topology change in the cluster. Kick requests 14333d14c5d2SYehuda Sadeh * off to different OSDs as needed. 14343d14c5d2SYehuda Sadeh */ 14353d14c5d2SYehuda Sadeh void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg) 14363d14c5d2SYehuda Sadeh { 14373d14c5d2SYehuda Sadeh void *p, *end, *next; 14383d14c5d2SYehuda Sadeh u32 nr_maps, maplen; 14393d14c5d2SYehuda Sadeh u32 epoch; 14403d14c5d2SYehuda Sadeh struct ceph_osdmap *newmap = NULL, *oldmap; 14413d14c5d2SYehuda Sadeh int err; 14423d14c5d2SYehuda Sadeh struct ceph_fsid fsid; 14433d14c5d2SYehuda Sadeh 14443d14c5d2SYehuda Sadeh dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0); 14453d14c5d2SYehuda Sadeh p = msg->front.iov_base; 14463d14c5d2SYehuda Sadeh end = p + msg->front.iov_len; 14473d14c5d2SYehuda Sadeh 14483d14c5d2SYehuda Sadeh /* verify fsid */ 14493d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, sizeof(fsid), bad); 14503d14c5d2SYehuda Sadeh ceph_decode_copy(&p, &fsid, sizeof(fsid)); 14513d14c5d2SYehuda Sadeh if (ceph_check_fsid(osdc->client, &fsid) < 0) 14523d14c5d2SYehuda Sadeh return; 14533d14c5d2SYehuda Sadeh 14543d14c5d2SYehuda Sadeh down_write(&osdc->map_sem); 14553d14c5d2SYehuda Sadeh 14563d14c5d2SYehuda Sadeh /* incremental maps */ 14573d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 14583d14c5d2SYehuda Sadeh dout(" %d inc maps\n", nr_maps); 14593d14c5d2SYehuda Sadeh while (nr_maps > 0) { 14603d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 14613d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 14623d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 14633d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 14643d14c5d2SYehuda Sadeh next = p + maplen; 14653d14c5d2SYehuda Sadeh if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) { 14663d14c5d2SYehuda Sadeh dout("applying incremental map %u len %d\n", 14673d14c5d2SYehuda Sadeh epoch, maplen); 14683d14c5d2SYehuda Sadeh newmap = osdmap_apply_incremental(&p, next, 14693d14c5d2SYehuda Sadeh osdc->osdmap, 147015d9882cSAlex Elder &osdc->client->msgr); 14713d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 14723d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 14733d14c5d2SYehuda Sadeh goto bad; 14743d14c5d2SYehuda Sadeh } 14753d14c5d2SYehuda Sadeh BUG_ON(!newmap); 14763d14c5d2SYehuda Sadeh if (newmap != osdc->osdmap) { 14773d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 14783d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 14793d14c5d2SYehuda Sadeh } 148038d6453cSSage Weil kick_requests(osdc, 0); 14813d14c5d2SYehuda Sadeh } else { 14823d14c5d2SYehuda Sadeh dout("ignoring incremental map %u len %d\n", 14833d14c5d2SYehuda Sadeh epoch, maplen); 14843d14c5d2SYehuda Sadeh } 14853d14c5d2SYehuda Sadeh p = next; 14863d14c5d2SYehuda Sadeh nr_maps--; 14873d14c5d2SYehuda Sadeh } 14883d14c5d2SYehuda Sadeh if (newmap) 14893d14c5d2SYehuda Sadeh goto done; 14903d14c5d2SYehuda Sadeh 14913d14c5d2SYehuda Sadeh /* full maps */ 14923d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 14933d14c5d2SYehuda Sadeh dout(" %d full maps\n", nr_maps); 14943d14c5d2SYehuda Sadeh while (nr_maps) { 14953d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 14963d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 14973d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 14983d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 14993d14c5d2SYehuda Sadeh if (nr_maps > 1) { 15003d14c5d2SYehuda Sadeh dout("skipping non-latest full map %u len %d\n", 15013d14c5d2SYehuda Sadeh epoch, maplen); 15023d14c5d2SYehuda Sadeh } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) { 15033d14c5d2SYehuda Sadeh dout("skipping full map %u len %d, " 15043d14c5d2SYehuda Sadeh "older than our %u\n", epoch, maplen, 15053d14c5d2SYehuda Sadeh osdc->osdmap->epoch); 15063d14c5d2SYehuda Sadeh } else { 150738d6453cSSage Weil int skipped_map = 0; 150838d6453cSSage Weil 15093d14c5d2SYehuda Sadeh dout("taking full map %u len %d\n", epoch, maplen); 15103d14c5d2SYehuda Sadeh newmap = osdmap_decode(&p, p+maplen); 15113d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 15123d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 15133d14c5d2SYehuda Sadeh goto bad; 15143d14c5d2SYehuda Sadeh } 15153d14c5d2SYehuda Sadeh BUG_ON(!newmap); 15163d14c5d2SYehuda Sadeh oldmap = osdc->osdmap; 15173d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 151838d6453cSSage Weil if (oldmap) { 151938d6453cSSage Weil if (oldmap->epoch + 1 < newmap->epoch) 152038d6453cSSage Weil skipped_map = 1; 15213d14c5d2SYehuda Sadeh ceph_osdmap_destroy(oldmap); 152238d6453cSSage Weil } 152338d6453cSSage Weil kick_requests(osdc, skipped_map); 15243d14c5d2SYehuda Sadeh } 15253d14c5d2SYehuda Sadeh p += maplen; 15263d14c5d2SYehuda Sadeh nr_maps--; 15273d14c5d2SYehuda Sadeh } 15283d14c5d2SYehuda Sadeh 15293d14c5d2SYehuda Sadeh done: 15303d14c5d2SYehuda Sadeh downgrade_write(&osdc->map_sem); 15313d14c5d2SYehuda Sadeh ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch); 1532cd634fb6SSage Weil 1533cd634fb6SSage Weil /* 1534cd634fb6SSage Weil * subscribe to subsequent osdmap updates if full to ensure 1535cd634fb6SSage Weil * we find out when we are no longer full and stop returning 1536cd634fb6SSage Weil * ENOSPC. 1537cd634fb6SSage Weil */ 1538cd634fb6SSage Weil if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL)) 1539cd634fb6SSage Weil ceph_monc_request_next_osdmap(&osdc->client->monc); 1540cd634fb6SSage Weil 1541f9d25199SAlex Elder mutex_lock(&osdc->request_mutex); 1542f9d25199SAlex Elder __send_queued(osdc); 1543f9d25199SAlex Elder mutex_unlock(&osdc->request_mutex); 15443d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 15453d14c5d2SYehuda Sadeh wake_up_all(&osdc->client->auth_wq); 15463d14c5d2SYehuda Sadeh return; 15473d14c5d2SYehuda Sadeh 15483d14c5d2SYehuda Sadeh bad: 15493d14c5d2SYehuda Sadeh pr_err("osdc handle_map corrupt msg\n"); 15503d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 15513d14c5d2SYehuda Sadeh up_write(&osdc->map_sem); 15523d14c5d2SYehuda Sadeh return; 15533d14c5d2SYehuda Sadeh } 15543d14c5d2SYehuda Sadeh 15553d14c5d2SYehuda Sadeh /* 1556a40c4f10SYehuda Sadeh * watch/notify callback event infrastructure 1557a40c4f10SYehuda Sadeh * 1558a40c4f10SYehuda Sadeh * These callbacks are used both for watch and notify operations. 1559a40c4f10SYehuda Sadeh */ 1560a40c4f10SYehuda Sadeh static void __release_event(struct kref *kref) 1561a40c4f10SYehuda Sadeh { 1562a40c4f10SYehuda Sadeh struct ceph_osd_event *event = 1563a40c4f10SYehuda Sadeh container_of(kref, struct ceph_osd_event, kref); 1564a40c4f10SYehuda Sadeh 1565a40c4f10SYehuda Sadeh dout("__release_event %p\n", event); 1566a40c4f10SYehuda Sadeh kfree(event); 1567a40c4f10SYehuda Sadeh } 1568a40c4f10SYehuda Sadeh 1569a40c4f10SYehuda Sadeh static void get_event(struct ceph_osd_event *event) 1570a40c4f10SYehuda Sadeh { 1571a40c4f10SYehuda Sadeh kref_get(&event->kref); 1572a40c4f10SYehuda Sadeh } 1573a40c4f10SYehuda Sadeh 1574a40c4f10SYehuda Sadeh void ceph_osdc_put_event(struct ceph_osd_event *event) 1575a40c4f10SYehuda Sadeh { 1576a40c4f10SYehuda Sadeh kref_put(&event->kref, __release_event); 1577a40c4f10SYehuda Sadeh } 1578a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_put_event); 1579a40c4f10SYehuda Sadeh 1580a40c4f10SYehuda Sadeh static void __insert_event(struct ceph_osd_client *osdc, 1581a40c4f10SYehuda Sadeh struct ceph_osd_event *new) 1582a40c4f10SYehuda Sadeh { 1583a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 1584a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 1585a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 1586a40c4f10SYehuda Sadeh 1587a40c4f10SYehuda Sadeh while (*p) { 1588a40c4f10SYehuda Sadeh parent = *p; 1589a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 1590a40c4f10SYehuda Sadeh if (new->cookie < event->cookie) 1591a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 1592a40c4f10SYehuda Sadeh else if (new->cookie > event->cookie) 1593a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 1594a40c4f10SYehuda Sadeh else 1595a40c4f10SYehuda Sadeh BUG(); 1596a40c4f10SYehuda Sadeh } 1597a40c4f10SYehuda Sadeh 1598a40c4f10SYehuda Sadeh rb_link_node(&new->node, parent, p); 1599a40c4f10SYehuda Sadeh rb_insert_color(&new->node, &osdc->event_tree); 1600a40c4f10SYehuda Sadeh } 1601a40c4f10SYehuda Sadeh 1602a40c4f10SYehuda Sadeh static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc, 1603a40c4f10SYehuda Sadeh u64 cookie) 1604a40c4f10SYehuda Sadeh { 1605a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 1606a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 1607a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 1608a40c4f10SYehuda Sadeh 1609a40c4f10SYehuda Sadeh while (*p) { 1610a40c4f10SYehuda Sadeh parent = *p; 1611a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 1612a40c4f10SYehuda Sadeh if (cookie < event->cookie) 1613a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 1614a40c4f10SYehuda Sadeh else if (cookie > event->cookie) 1615a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 1616a40c4f10SYehuda Sadeh else 1617a40c4f10SYehuda Sadeh return event; 1618a40c4f10SYehuda Sadeh } 1619a40c4f10SYehuda Sadeh return NULL; 1620a40c4f10SYehuda Sadeh } 1621a40c4f10SYehuda Sadeh 1622a40c4f10SYehuda Sadeh static void __remove_event(struct ceph_osd_event *event) 1623a40c4f10SYehuda Sadeh { 1624a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 1625a40c4f10SYehuda Sadeh 1626a40c4f10SYehuda Sadeh if (!RB_EMPTY_NODE(&event->node)) { 1627a40c4f10SYehuda Sadeh dout("__remove_event removed %p\n", event); 1628a40c4f10SYehuda Sadeh rb_erase(&event->node, &osdc->event_tree); 1629a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 1630a40c4f10SYehuda Sadeh } else { 1631a40c4f10SYehuda Sadeh dout("__remove_event didn't remove %p\n", event); 1632a40c4f10SYehuda Sadeh } 1633a40c4f10SYehuda Sadeh } 1634a40c4f10SYehuda Sadeh 1635a40c4f10SYehuda Sadeh int ceph_osdc_create_event(struct ceph_osd_client *osdc, 1636a40c4f10SYehuda Sadeh void (*event_cb)(u64, u64, u8, void *), 16373c663bbdSAlex Elder void *data, struct ceph_osd_event **pevent) 1638a40c4f10SYehuda Sadeh { 1639a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 1640a40c4f10SYehuda Sadeh 1641a40c4f10SYehuda Sadeh event = kmalloc(sizeof(*event), GFP_NOIO); 1642a40c4f10SYehuda Sadeh if (!event) 1643a40c4f10SYehuda Sadeh return -ENOMEM; 1644a40c4f10SYehuda Sadeh 1645a40c4f10SYehuda Sadeh dout("create_event %p\n", event); 1646a40c4f10SYehuda Sadeh event->cb = event_cb; 16473c663bbdSAlex Elder event->one_shot = 0; 1648a40c4f10SYehuda Sadeh event->data = data; 1649a40c4f10SYehuda Sadeh event->osdc = osdc; 1650a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&event->osd_node); 16513ee5234dSAlex Elder RB_CLEAR_NODE(&event->node); 1652a40c4f10SYehuda Sadeh kref_init(&event->kref); /* one ref for us */ 1653a40c4f10SYehuda Sadeh kref_get(&event->kref); /* one ref for the caller */ 1654a40c4f10SYehuda Sadeh 1655a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 1656a40c4f10SYehuda Sadeh event->cookie = ++osdc->event_count; 1657a40c4f10SYehuda Sadeh __insert_event(osdc, event); 1658a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 1659a40c4f10SYehuda Sadeh 1660a40c4f10SYehuda Sadeh *pevent = event; 1661a40c4f10SYehuda Sadeh return 0; 1662a40c4f10SYehuda Sadeh } 1663a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_create_event); 1664a40c4f10SYehuda Sadeh 1665a40c4f10SYehuda Sadeh void ceph_osdc_cancel_event(struct ceph_osd_event *event) 1666a40c4f10SYehuda Sadeh { 1667a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 1668a40c4f10SYehuda Sadeh 1669a40c4f10SYehuda Sadeh dout("cancel_event %p\n", event); 1670a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 1671a40c4f10SYehuda Sadeh __remove_event(event); 1672a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 1673a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); /* caller's */ 1674a40c4f10SYehuda Sadeh } 1675a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_cancel_event); 1676a40c4f10SYehuda Sadeh 1677a40c4f10SYehuda Sadeh 1678a40c4f10SYehuda Sadeh static void do_event_work(struct work_struct *work) 1679a40c4f10SYehuda Sadeh { 1680a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work = 1681a40c4f10SYehuda Sadeh container_of(work, struct ceph_osd_event_work, work); 1682a40c4f10SYehuda Sadeh struct ceph_osd_event *event = event_work->event; 1683a40c4f10SYehuda Sadeh u64 ver = event_work->ver; 1684a40c4f10SYehuda Sadeh u64 notify_id = event_work->notify_id; 1685a40c4f10SYehuda Sadeh u8 opcode = event_work->opcode; 1686a40c4f10SYehuda Sadeh 1687a40c4f10SYehuda Sadeh dout("do_event_work completing %p\n", event); 1688a40c4f10SYehuda Sadeh event->cb(ver, notify_id, opcode, event->data); 1689a40c4f10SYehuda Sadeh dout("do_event_work completed %p\n", event); 1690a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 1691a40c4f10SYehuda Sadeh kfree(event_work); 1692a40c4f10SYehuda Sadeh } 1693a40c4f10SYehuda Sadeh 1694a40c4f10SYehuda Sadeh 1695a40c4f10SYehuda Sadeh /* 1696a40c4f10SYehuda Sadeh * Process osd watch notifications 1697a40c4f10SYehuda Sadeh */ 16983c663bbdSAlex Elder static void handle_watch_notify(struct ceph_osd_client *osdc, 16993c663bbdSAlex Elder struct ceph_msg *msg) 1700a40c4f10SYehuda Sadeh { 1701a40c4f10SYehuda Sadeh void *p, *end; 1702a40c4f10SYehuda Sadeh u8 proto_ver; 1703a40c4f10SYehuda Sadeh u64 cookie, ver, notify_id; 1704a40c4f10SYehuda Sadeh u8 opcode; 1705a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 1706a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work; 1707a40c4f10SYehuda Sadeh 1708a40c4f10SYehuda Sadeh p = msg->front.iov_base; 1709a40c4f10SYehuda Sadeh end = p + msg->front.iov_len; 1710a40c4f10SYehuda Sadeh 1711a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, proto_ver, bad); 1712a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, opcode, bad); 1713a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, cookie, bad); 1714a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, ver, bad); 1715a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, notify_id, bad); 1716a40c4f10SYehuda Sadeh 1717a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 1718a40c4f10SYehuda Sadeh event = __find_event(osdc, cookie); 1719a40c4f10SYehuda Sadeh if (event) { 17203c663bbdSAlex Elder BUG_ON(event->one_shot); 1721a40c4f10SYehuda Sadeh get_event(event); 1722a40c4f10SYehuda Sadeh } 1723a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 1724a40c4f10SYehuda Sadeh dout("handle_watch_notify cookie %lld ver %lld event %p\n", 1725a40c4f10SYehuda Sadeh cookie, ver, event); 1726a40c4f10SYehuda Sadeh if (event) { 1727a40c4f10SYehuda Sadeh event_work = kmalloc(sizeof(*event_work), GFP_NOIO); 1728a40c4f10SYehuda Sadeh if (!event_work) { 1729a40c4f10SYehuda Sadeh dout("ERROR: could not allocate event_work\n"); 1730a40c4f10SYehuda Sadeh goto done_err; 1731a40c4f10SYehuda Sadeh } 17326b0ae409SMariusz Kozlowski INIT_WORK(&event_work->work, do_event_work); 1733a40c4f10SYehuda Sadeh event_work->event = event; 1734a40c4f10SYehuda Sadeh event_work->ver = ver; 1735a40c4f10SYehuda Sadeh event_work->notify_id = notify_id; 1736a40c4f10SYehuda Sadeh event_work->opcode = opcode; 1737a40c4f10SYehuda Sadeh if (!queue_work(osdc->notify_wq, &event_work->work)) { 1738a40c4f10SYehuda Sadeh dout("WARNING: failed to queue notify event work\n"); 1739a40c4f10SYehuda Sadeh goto done_err; 1740a40c4f10SYehuda Sadeh } 1741a40c4f10SYehuda Sadeh } 1742a40c4f10SYehuda Sadeh 1743a40c4f10SYehuda Sadeh return; 1744a40c4f10SYehuda Sadeh 1745a40c4f10SYehuda Sadeh done_err: 1746a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 1747a40c4f10SYehuda Sadeh return; 1748a40c4f10SYehuda Sadeh 1749a40c4f10SYehuda Sadeh bad: 1750a40c4f10SYehuda Sadeh pr_err("osdc handle_watch_notify corrupt msg\n"); 1751a40c4f10SYehuda Sadeh return; 1752a40c4f10SYehuda Sadeh } 1753a40c4f10SYehuda Sadeh 175470636773SAlex Elder static void ceph_osdc_msg_data_set(struct ceph_msg *msg, 175570636773SAlex Elder struct ceph_osd_data *osd_data) 175670636773SAlex Elder { 175770636773SAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 175870636773SAlex Elder BUG_ON(osd_data->length > (u64) SIZE_MAX); 175970636773SAlex Elder if (osd_data->length) 176070636773SAlex Elder ceph_msg_data_set_pages(msg, osd_data->pages, 176170636773SAlex Elder osd_data->length, osd_data->alignment); 17629a5e6d09SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) { 17639a5e6d09SAlex Elder BUG_ON(!osd_data->pagelist->length); 17649a5e6d09SAlex Elder ceph_msg_data_set_pagelist(msg, osd_data->pagelist); 176570636773SAlex Elder #ifdef CONFIG_BLOCK 176670636773SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) { 176770636773SAlex Elder ceph_msg_data_set_bio(msg, osd_data->bio); 176870636773SAlex Elder #endif 176970636773SAlex Elder } else { 177070636773SAlex Elder BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE); 177170636773SAlex Elder } 177270636773SAlex Elder } 177370636773SAlex Elder 1774a40c4f10SYehuda Sadeh /* 17753d14c5d2SYehuda Sadeh * Register request, send initial attempt. 17763d14c5d2SYehuda Sadeh */ 17773d14c5d2SYehuda Sadeh int ceph_osdc_start_request(struct ceph_osd_client *osdc, 17783d14c5d2SYehuda Sadeh struct ceph_osd_request *req, 17793d14c5d2SYehuda Sadeh bool nofail) 17803d14c5d2SYehuda Sadeh { 17813d14c5d2SYehuda Sadeh int rc = 0; 17823d14c5d2SYehuda Sadeh 178370636773SAlex Elder /* Set up response incoming data and request outgoing data fields */ 17840fff87ecSAlex Elder 178570636773SAlex Elder ceph_osdc_msg_data_set(req->r_reply, &req->r_data_in); 178670636773SAlex Elder ceph_osdc_msg_data_set(req->r_request, &req->r_data_out); 17873d14c5d2SYehuda Sadeh 17883d14c5d2SYehuda Sadeh register_request(osdc, req); 17893d14c5d2SYehuda Sadeh 17903d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 17913d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 17923d14c5d2SYehuda Sadeh /* 17933d14c5d2SYehuda Sadeh * a racing kick_requests() may have sent the message for us 17943d14c5d2SYehuda Sadeh * while we dropped request_mutex above, so only send now if 17953d14c5d2SYehuda Sadeh * the request still han't been touched yet. 17963d14c5d2SYehuda Sadeh */ 17973d14c5d2SYehuda Sadeh if (req->r_sent == 0) { 179838d6453cSSage Weil rc = __map_request(osdc, req, 0); 17999d6fcb08SSage Weil if (rc < 0) { 18009d6fcb08SSage Weil if (nofail) { 18019d6fcb08SSage Weil dout("osdc_start_request failed map, " 18029d6fcb08SSage Weil " will retry %lld\n", req->r_tid); 18039d6fcb08SSage Weil rc = 0; 18049d6fcb08SSage Weil } 1805234af26fSDan Carpenter goto out_unlock; 18069d6fcb08SSage Weil } 18076f6c7006SSage Weil if (req->r_osd == NULL) { 18086f6c7006SSage Weil dout("send_request %p no up osds in pg\n", req); 18096f6c7006SSage Weil ceph_monc_request_next_osdmap(&osdc->client->monc); 18106f6c7006SSage Weil } else { 181156e925b6SSage Weil __send_request(osdc, req); 181256e925b6SSage Weil } 18133d14c5d2SYehuda Sadeh rc = 0; 18146f6c7006SSage Weil } 1815234af26fSDan Carpenter 1816234af26fSDan Carpenter out_unlock: 18173d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 18183d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 18193d14c5d2SYehuda Sadeh return rc; 18203d14c5d2SYehuda Sadeh } 18213d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_start_request); 18223d14c5d2SYehuda Sadeh 18233d14c5d2SYehuda Sadeh /* 18243d14c5d2SYehuda Sadeh * wait for a request to complete 18253d14c5d2SYehuda Sadeh */ 18263d14c5d2SYehuda Sadeh int ceph_osdc_wait_request(struct ceph_osd_client *osdc, 18273d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 18283d14c5d2SYehuda Sadeh { 18293d14c5d2SYehuda Sadeh int rc; 18303d14c5d2SYehuda Sadeh 18313d14c5d2SYehuda Sadeh rc = wait_for_completion_interruptible(&req->r_completion); 18323d14c5d2SYehuda Sadeh if (rc < 0) { 18333d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 18343d14c5d2SYehuda Sadeh __cancel_request(req); 18353d14c5d2SYehuda Sadeh __unregister_request(osdc, req); 18363d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 183725845472SSage Weil complete_request(req); 18383d14c5d2SYehuda Sadeh dout("wait_request tid %llu canceled/timed out\n", req->r_tid); 18393d14c5d2SYehuda Sadeh return rc; 18403d14c5d2SYehuda Sadeh } 18413d14c5d2SYehuda Sadeh 18423d14c5d2SYehuda Sadeh dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result); 18433d14c5d2SYehuda Sadeh return req->r_result; 18443d14c5d2SYehuda Sadeh } 18453d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_wait_request); 18463d14c5d2SYehuda Sadeh 18473d14c5d2SYehuda Sadeh /* 18483d14c5d2SYehuda Sadeh * sync - wait for all in-flight requests to flush. avoid starvation. 18493d14c5d2SYehuda Sadeh */ 18503d14c5d2SYehuda Sadeh void ceph_osdc_sync(struct ceph_osd_client *osdc) 18513d14c5d2SYehuda Sadeh { 18523d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 18533d14c5d2SYehuda Sadeh u64 last_tid, next_tid = 0; 18543d14c5d2SYehuda Sadeh 18553d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 18563d14c5d2SYehuda Sadeh last_tid = osdc->last_tid; 18573d14c5d2SYehuda Sadeh while (1) { 18583d14c5d2SYehuda Sadeh req = __lookup_request_ge(osdc, next_tid); 18593d14c5d2SYehuda Sadeh if (!req) 18603d14c5d2SYehuda Sadeh break; 18613d14c5d2SYehuda Sadeh if (req->r_tid > last_tid) 18623d14c5d2SYehuda Sadeh break; 18633d14c5d2SYehuda Sadeh 18643d14c5d2SYehuda Sadeh next_tid = req->r_tid + 1; 18653d14c5d2SYehuda Sadeh if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0) 18663d14c5d2SYehuda Sadeh continue; 18673d14c5d2SYehuda Sadeh 18683d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 18693d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 18703d14c5d2SYehuda Sadeh dout("sync waiting on tid %llu (last is %llu)\n", 18713d14c5d2SYehuda Sadeh req->r_tid, last_tid); 18723d14c5d2SYehuda Sadeh wait_for_completion(&req->r_safe_completion); 18733d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 18743d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 18753d14c5d2SYehuda Sadeh } 18763d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 18773d14c5d2SYehuda Sadeh dout("sync done (thru tid %llu)\n", last_tid); 18783d14c5d2SYehuda Sadeh } 18793d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_sync); 18803d14c5d2SYehuda Sadeh 18813d14c5d2SYehuda Sadeh /* 18823d14c5d2SYehuda Sadeh * init, shutdown 18833d14c5d2SYehuda Sadeh */ 18843d14c5d2SYehuda Sadeh int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client) 18853d14c5d2SYehuda Sadeh { 18863d14c5d2SYehuda Sadeh int err; 18873d14c5d2SYehuda Sadeh 18883d14c5d2SYehuda Sadeh dout("init\n"); 18893d14c5d2SYehuda Sadeh osdc->client = client; 18903d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 18913d14c5d2SYehuda Sadeh init_rwsem(&osdc->map_sem); 18923d14c5d2SYehuda Sadeh init_completion(&osdc->map_waiters); 18933d14c5d2SYehuda Sadeh osdc->last_requested_map = 0; 18943d14c5d2SYehuda Sadeh mutex_init(&osdc->request_mutex); 18953d14c5d2SYehuda Sadeh osdc->last_tid = 0; 18963d14c5d2SYehuda Sadeh osdc->osds = RB_ROOT; 18973d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->osd_lru); 18983d14c5d2SYehuda Sadeh osdc->requests = RB_ROOT; 18993d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_lru); 19006f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_unsent); 19016f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_notarget); 1902a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_linger); 19033d14c5d2SYehuda Sadeh osdc->num_requests = 0; 19043d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout); 19053d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout); 1906a40c4f10SYehuda Sadeh spin_lock_init(&osdc->event_lock); 1907a40c4f10SYehuda Sadeh osdc->event_tree = RB_ROOT; 1908a40c4f10SYehuda Sadeh osdc->event_count = 0; 19093d14c5d2SYehuda Sadeh 19103d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 19113d14c5d2SYehuda Sadeh round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ)); 19123d14c5d2SYehuda Sadeh 19133d14c5d2SYehuda Sadeh err = -ENOMEM; 19143d14c5d2SYehuda Sadeh osdc->req_mempool = mempool_create_kmalloc_pool(10, 19153d14c5d2SYehuda Sadeh sizeof(struct ceph_osd_request)); 19163d14c5d2SYehuda Sadeh if (!osdc->req_mempool) 19173d14c5d2SYehuda Sadeh goto out; 19183d14c5d2SYehuda Sadeh 1919d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP, 1920d50b409fSSage Weil OSD_OP_FRONT_LEN, 10, true, 19213d14c5d2SYehuda Sadeh "osd_op"); 19223d14c5d2SYehuda Sadeh if (err < 0) 19233d14c5d2SYehuda Sadeh goto out_mempool; 1924d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY, 19253d14c5d2SYehuda Sadeh OSD_OPREPLY_FRONT_LEN, 10, true, 19263d14c5d2SYehuda Sadeh "osd_op_reply"); 19273d14c5d2SYehuda Sadeh if (err < 0) 19283d14c5d2SYehuda Sadeh goto out_msgpool; 1929a40c4f10SYehuda Sadeh 1930a40c4f10SYehuda Sadeh osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify"); 1931a40c4f10SYehuda Sadeh if (IS_ERR(osdc->notify_wq)) { 1932a40c4f10SYehuda Sadeh err = PTR_ERR(osdc->notify_wq); 1933a40c4f10SYehuda Sadeh osdc->notify_wq = NULL; 1934a40c4f10SYehuda Sadeh goto out_msgpool; 1935a40c4f10SYehuda Sadeh } 19363d14c5d2SYehuda Sadeh return 0; 19373d14c5d2SYehuda Sadeh 19383d14c5d2SYehuda Sadeh out_msgpool: 19393d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 19403d14c5d2SYehuda Sadeh out_mempool: 19413d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 19423d14c5d2SYehuda Sadeh out: 19433d14c5d2SYehuda Sadeh return err; 19443d14c5d2SYehuda Sadeh } 19453d14c5d2SYehuda Sadeh 19463d14c5d2SYehuda Sadeh void ceph_osdc_stop(struct ceph_osd_client *osdc) 19473d14c5d2SYehuda Sadeh { 1948a40c4f10SYehuda Sadeh flush_workqueue(osdc->notify_wq); 1949a40c4f10SYehuda Sadeh destroy_workqueue(osdc->notify_wq); 19503d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->timeout_work); 19513d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->osds_timeout_work); 19523d14c5d2SYehuda Sadeh if (osdc->osdmap) { 19533d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 19543d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 19553d14c5d2SYehuda Sadeh } 1956aca420bcSSage Weil remove_all_osds(osdc); 19573d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 19583d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 19593d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op_reply); 19603d14c5d2SYehuda Sadeh } 19613d14c5d2SYehuda Sadeh 19623d14c5d2SYehuda Sadeh /* 19633d14c5d2SYehuda Sadeh * Read some contiguous pages. If we cross a stripe boundary, shorten 19643d14c5d2SYehuda Sadeh * *plen. Return number of bytes read, or error. 19653d14c5d2SYehuda Sadeh */ 19663d14c5d2SYehuda Sadeh int ceph_osdc_readpages(struct ceph_osd_client *osdc, 19673d14c5d2SYehuda Sadeh struct ceph_vino vino, struct ceph_file_layout *layout, 19683d14c5d2SYehuda Sadeh u64 off, u64 *plen, 19693d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 1970b7495fc2SSage Weil struct page **pages, int num_pages, int page_align) 19713d14c5d2SYehuda Sadeh { 19723d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 19730fff87ecSAlex Elder struct ceph_osd_data *osd_data; 19743d14c5d2SYehuda Sadeh int rc = 0; 19753d14c5d2SYehuda Sadeh 19763d14c5d2SYehuda Sadeh dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino, 19773d14c5d2SYehuda Sadeh vino.snap, off, *plen); 19783d14c5d2SYehuda Sadeh req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 19793d14c5d2SYehuda Sadeh CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ, 19803d14c5d2SYehuda Sadeh NULL, 0, truncate_seq, truncate_size, NULL, 1981153e5167SAlex Elder false); 19826816282dSSage Weil if (IS_ERR(req)) 19836816282dSSage Weil return PTR_ERR(req); 19843d14c5d2SYehuda Sadeh 19853d14c5d2SYehuda Sadeh /* it may be a short read due to an object boundary */ 19860fff87ecSAlex Elder 19870fff87ecSAlex Elder osd_data = &req->r_data_in; 19880fff87ecSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_PAGES; 19890fff87ecSAlex Elder osd_data->pages = pages; 1990e0c59487SAlex Elder osd_data->length = *plen; 19910fff87ecSAlex Elder osd_data->alignment = page_align; 19923d14c5d2SYehuda Sadeh 1993e0c59487SAlex Elder dout("readpages final extent is %llu~%llu (%llu bytes align %d)\n", 1994e0c59487SAlex Elder off, *plen, osd_data->length, page_align); 19953d14c5d2SYehuda Sadeh 19963d14c5d2SYehuda Sadeh rc = ceph_osdc_start_request(osdc, req, false); 19973d14c5d2SYehuda Sadeh if (!rc) 19983d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 19993d14c5d2SYehuda Sadeh 20003d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 20013d14c5d2SYehuda Sadeh dout("readpages result %d\n", rc); 20023d14c5d2SYehuda Sadeh return rc; 20033d14c5d2SYehuda Sadeh } 20043d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_readpages); 20053d14c5d2SYehuda Sadeh 20063d14c5d2SYehuda Sadeh /* 20073d14c5d2SYehuda Sadeh * do a synchronous write on N pages 20083d14c5d2SYehuda Sadeh */ 20093d14c5d2SYehuda Sadeh int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino, 20103d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 20113d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 20123d14c5d2SYehuda Sadeh u64 off, u64 len, 20133d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 20143d14c5d2SYehuda Sadeh struct timespec *mtime, 201524808826SAlex Elder struct page **pages, int num_pages) 20163d14c5d2SYehuda Sadeh { 20173d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 20180fff87ecSAlex Elder struct ceph_osd_data *osd_data; 20193d14c5d2SYehuda Sadeh int rc = 0; 2020b7495fc2SSage Weil int page_align = off & ~PAGE_MASK; 20213d14c5d2SYehuda Sadeh 20223d14c5d2SYehuda Sadeh BUG_ON(vino.snap != CEPH_NOSNAP); 20233d14c5d2SYehuda Sadeh req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 20243d14c5d2SYehuda Sadeh CEPH_OSD_OP_WRITE, 202524808826SAlex Elder CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE, 2026fbf8685fSAlex Elder snapc, 0, 20273d14c5d2SYehuda Sadeh truncate_seq, truncate_size, mtime, 2028153e5167SAlex Elder true); 20296816282dSSage Weil if (IS_ERR(req)) 20306816282dSSage Weil return PTR_ERR(req); 20313d14c5d2SYehuda Sadeh 20323d14c5d2SYehuda Sadeh /* it may be a short write due to an object boundary */ 20330fff87ecSAlex Elder osd_data = &req->r_data_out; 20340fff87ecSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_PAGES; 20350fff87ecSAlex Elder osd_data->pages = pages; 2036e0c59487SAlex Elder osd_data->length = len; 20370fff87ecSAlex Elder osd_data->alignment = page_align; 2038e0c59487SAlex Elder dout("writepages %llu~%llu (%llu bytes)\n", off, len, osd_data->length); 20393d14c5d2SYehuda Sadeh 204087f979d3SAlex Elder rc = ceph_osdc_start_request(osdc, req, true); 20413d14c5d2SYehuda Sadeh if (!rc) 20423d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 20433d14c5d2SYehuda Sadeh 20443d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 20453d14c5d2SYehuda Sadeh if (rc == 0) 20463d14c5d2SYehuda Sadeh rc = len; 20473d14c5d2SYehuda Sadeh dout("writepages result %d\n", rc); 20483d14c5d2SYehuda Sadeh return rc; 20493d14c5d2SYehuda Sadeh } 20503d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_writepages); 20513d14c5d2SYehuda Sadeh 20523d14c5d2SYehuda Sadeh /* 20533d14c5d2SYehuda Sadeh * handle incoming message 20543d14c5d2SYehuda Sadeh */ 20553d14c5d2SYehuda Sadeh static void dispatch(struct ceph_connection *con, struct ceph_msg *msg) 20563d14c5d2SYehuda Sadeh { 20573d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 20583d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 20593d14c5d2SYehuda Sadeh int type = le16_to_cpu(msg->hdr.type); 20603d14c5d2SYehuda Sadeh 20613d14c5d2SYehuda Sadeh if (!osd) 20623d14c5d2SYehuda Sadeh goto out; 20633d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 20643d14c5d2SYehuda Sadeh 20653d14c5d2SYehuda Sadeh switch (type) { 20663d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 20673d14c5d2SYehuda Sadeh ceph_osdc_handle_map(osdc, msg); 20683d14c5d2SYehuda Sadeh break; 20693d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 20703d14c5d2SYehuda Sadeh handle_reply(osdc, msg, con); 20713d14c5d2SYehuda Sadeh break; 2072a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 2073a40c4f10SYehuda Sadeh handle_watch_notify(osdc, msg); 2074a40c4f10SYehuda Sadeh break; 20753d14c5d2SYehuda Sadeh 20763d14c5d2SYehuda Sadeh default: 20773d14c5d2SYehuda Sadeh pr_err("received unknown message type %d %s\n", type, 20783d14c5d2SYehuda Sadeh ceph_msg_type_name(type)); 20793d14c5d2SYehuda Sadeh } 20803d14c5d2SYehuda Sadeh out: 20813d14c5d2SYehuda Sadeh ceph_msg_put(msg); 20823d14c5d2SYehuda Sadeh } 20833d14c5d2SYehuda Sadeh 20843d14c5d2SYehuda Sadeh /* 20853d14c5d2SYehuda Sadeh * lookup and return message for incoming reply. set up reply message 20863d14c5d2SYehuda Sadeh * pages. 20873d14c5d2SYehuda Sadeh */ 20883d14c5d2SYehuda Sadeh static struct ceph_msg *get_reply(struct ceph_connection *con, 20893d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 20903d14c5d2SYehuda Sadeh int *skip) 20913d14c5d2SYehuda Sadeh { 20923d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 20933d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = osd->o_osdc; 20943d14c5d2SYehuda Sadeh struct ceph_msg *m; 20953d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 20963d14c5d2SYehuda Sadeh int front = le32_to_cpu(hdr->front_len); 20973d14c5d2SYehuda Sadeh int data_len = le32_to_cpu(hdr->data_len); 20983d14c5d2SYehuda Sadeh u64 tid; 20993d14c5d2SYehuda Sadeh 21003d14c5d2SYehuda Sadeh tid = le64_to_cpu(hdr->tid); 21013d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 21023d14c5d2SYehuda Sadeh req = __lookup_request(osdc, tid); 21033d14c5d2SYehuda Sadeh if (!req) { 21043d14c5d2SYehuda Sadeh *skip = 1; 21053d14c5d2SYehuda Sadeh m = NULL; 2106756a16a5SSage Weil dout("get_reply unknown tid %llu from osd%d\n", tid, 21073d14c5d2SYehuda Sadeh osd->o_osd); 21083d14c5d2SYehuda Sadeh goto out; 21093d14c5d2SYehuda Sadeh } 21103d14c5d2SYehuda Sadeh 21113d14c5d2SYehuda Sadeh if (req->r_con_filling_msg) { 21128921d114SAlex Elder dout("%s revoking msg %p from old con %p\n", __func__, 21133d14c5d2SYehuda Sadeh req->r_reply, req->r_con_filling_msg); 21148921d114SAlex Elder ceph_msg_revoke_incoming(req->r_reply); 21150d47766fSSage Weil req->r_con_filling_msg->ops->put(req->r_con_filling_msg); 21163d14c5d2SYehuda Sadeh req->r_con_filling_msg = NULL; 21173d14c5d2SYehuda Sadeh } 21183d14c5d2SYehuda Sadeh 21193d14c5d2SYehuda Sadeh if (front > req->r_reply->front.iov_len) { 21203d14c5d2SYehuda Sadeh pr_warning("get_reply front %d > preallocated %d\n", 21213d14c5d2SYehuda Sadeh front, (int)req->r_reply->front.iov_len); 2122b61c2763SSage Weil m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS, false); 21233d14c5d2SYehuda Sadeh if (!m) 21243d14c5d2SYehuda Sadeh goto out; 21253d14c5d2SYehuda Sadeh ceph_msg_put(req->r_reply); 21263d14c5d2SYehuda Sadeh req->r_reply = m; 21273d14c5d2SYehuda Sadeh } 21283d14c5d2SYehuda Sadeh m = ceph_msg_get(req->r_reply); 21293d14c5d2SYehuda Sadeh 21303d14c5d2SYehuda Sadeh if (data_len > 0) { 21310fff87ecSAlex Elder struct ceph_osd_data *osd_data = &req->r_data_in; 21320fff87ecSAlex Elder 21330fff87ecSAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 21340fff87ecSAlex Elder if (osd_data->pages && 2135e0c59487SAlex Elder unlikely(osd_data->length < data_len)) { 21362ac2b7a6SAlex Elder 2137e0c59487SAlex Elder pr_warning("tid %lld reply has %d bytes " 2138e0c59487SAlex Elder "we had only %llu bytes ready\n", 2139e0c59487SAlex Elder tid, data_len, osd_data->length); 21403d14c5d2SYehuda Sadeh *skip = 1; 21413d14c5d2SYehuda Sadeh ceph_msg_put(m); 21423d14c5d2SYehuda Sadeh m = NULL; 21433d14c5d2SYehuda Sadeh goto out; 21443d14c5d2SYehuda Sadeh } 21453d14c5d2SYehuda Sadeh } 21462ac2b7a6SAlex Elder } 21473d14c5d2SYehuda Sadeh *skip = 0; 21480d47766fSSage Weil req->r_con_filling_msg = con->ops->get(con); 21493d14c5d2SYehuda Sadeh dout("get_reply tid %lld %p\n", tid, m); 21503d14c5d2SYehuda Sadeh 21513d14c5d2SYehuda Sadeh out: 21523d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 21533d14c5d2SYehuda Sadeh return m; 21543d14c5d2SYehuda Sadeh 21553d14c5d2SYehuda Sadeh } 21563d14c5d2SYehuda Sadeh 21573d14c5d2SYehuda Sadeh static struct ceph_msg *alloc_msg(struct ceph_connection *con, 21583d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 21593d14c5d2SYehuda Sadeh int *skip) 21603d14c5d2SYehuda Sadeh { 21613d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 21623d14c5d2SYehuda Sadeh int type = le16_to_cpu(hdr->type); 21633d14c5d2SYehuda Sadeh int front = le32_to_cpu(hdr->front_len); 21643d14c5d2SYehuda Sadeh 21651c20f2d2SAlex Elder *skip = 0; 21663d14c5d2SYehuda Sadeh switch (type) { 21673d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 2168a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 2169b61c2763SSage Weil return ceph_msg_new(type, front, GFP_NOFS, false); 21703d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 21713d14c5d2SYehuda Sadeh return get_reply(con, hdr, skip); 21723d14c5d2SYehuda Sadeh default: 21733d14c5d2SYehuda Sadeh pr_info("alloc_msg unexpected msg type %d from osd%d\n", type, 21743d14c5d2SYehuda Sadeh osd->o_osd); 21753d14c5d2SYehuda Sadeh *skip = 1; 21763d14c5d2SYehuda Sadeh return NULL; 21773d14c5d2SYehuda Sadeh } 21783d14c5d2SYehuda Sadeh } 21793d14c5d2SYehuda Sadeh 21803d14c5d2SYehuda Sadeh /* 21813d14c5d2SYehuda Sadeh * Wrappers to refcount containing ceph_osd struct 21823d14c5d2SYehuda Sadeh */ 21833d14c5d2SYehuda Sadeh static struct ceph_connection *get_osd_con(struct ceph_connection *con) 21843d14c5d2SYehuda Sadeh { 21853d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 21863d14c5d2SYehuda Sadeh if (get_osd(osd)) 21873d14c5d2SYehuda Sadeh return con; 21883d14c5d2SYehuda Sadeh return NULL; 21893d14c5d2SYehuda Sadeh } 21903d14c5d2SYehuda Sadeh 21913d14c5d2SYehuda Sadeh static void put_osd_con(struct ceph_connection *con) 21923d14c5d2SYehuda Sadeh { 21933d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 21943d14c5d2SYehuda Sadeh put_osd(osd); 21953d14c5d2SYehuda Sadeh } 21963d14c5d2SYehuda Sadeh 21973d14c5d2SYehuda Sadeh /* 21983d14c5d2SYehuda Sadeh * authentication 21993d14c5d2SYehuda Sadeh */ 2200a3530df3SAlex Elder /* 2201a3530df3SAlex Elder * Note: returned pointer is the address of a structure that's 2202a3530df3SAlex Elder * managed separately. Caller must *not* attempt to free it. 2203a3530df3SAlex Elder */ 2204a3530df3SAlex Elder static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con, 22058f43fb53SAlex Elder int *proto, int force_new) 22063d14c5d2SYehuda Sadeh { 22073d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 22083d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 22093d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 221074f1869fSAlex Elder struct ceph_auth_handshake *auth = &o->o_auth; 22113d14c5d2SYehuda Sadeh 221274f1869fSAlex Elder if (force_new && auth->authorizer) { 221327859f97SSage Weil ceph_auth_destroy_authorizer(ac, auth->authorizer); 221474f1869fSAlex Elder auth->authorizer = NULL; 22153d14c5d2SYehuda Sadeh } 221627859f97SSage Weil if (!auth->authorizer) { 221727859f97SSage Weil int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 2218a3530df3SAlex Elder auth); 22193d14c5d2SYehuda Sadeh if (ret) 2220a3530df3SAlex Elder return ERR_PTR(ret); 222127859f97SSage Weil } else { 222227859f97SSage Weil int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 22230bed9b5cSSage Weil auth); 22240bed9b5cSSage Weil if (ret) 22250bed9b5cSSage Weil return ERR_PTR(ret); 22263d14c5d2SYehuda Sadeh } 22273d14c5d2SYehuda Sadeh *proto = ac->protocol; 222874f1869fSAlex Elder 2229a3530df3SAlex Elder return auth; 22303d14c5d2SYehuda Sadeh } 22313d14c5d2SYehuda Sadeh 22323d14c5d2SYehuda Sadeh 22333d14c5d2SYehuda Sadeh static int verify_authorizer_reply(struct ceph_connection *con, int len) 22343d14c5d2SYehuda Sadeh { 22353d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 22363d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 22373d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 22383d14c5d2SYehuda Sadeh 223927859f97SSage Weil return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len); 22403d14c5d2SYehuda Sadeh } 22413d14c5d2SYehuda Sadeh 22423d14c5d2SYehuda Sadeh static int invalidate_authorizer(struct ceph_connection *con) 22433d14c5d2SYehuda Sadeh { 22443d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 22453d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 22463d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 22473d14c5d2SYehuda Sadeh 224827859f97SSage Weil ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD); 22493d14c5d2SYehuda Sadeh return ceph_monc_validate_auth(&osdc->client->monc); 22503d14c5d2SYehuda Sadeh } 22513d14c5d2SYehuda Sadeh 22523d14c5d2SYehuda Sadeh static const struct ceph_connection_operations osd_con_ops = { 22533d14c5d2SYehuda Sadeh .get = get_osd_con, 22543d14c5d2SYehuda Sadeh .put = put_osd_con, 22553d14c5d2SYehuda Sadeh .dispatch = dispatch, 22563d14c5d2SYehuda Sadeh .get_authorizer = get_authorizer, 22573d14c5d2SYehuda Sadeh .verify_authorizer_reply = verify_authorizer_reply, 22583d14c5d2SYehuda Sadeh .invalidate_authorizer = invalidate_authorizer, 22593d14c5d2SYehuda Sadeh .alloc_msg = alloc_msg, 22603d14c5d2SYehuda Sadeh .fault = osd_reset, 22613d14c5d2SYehuda Sadeh }; 2262