1a4ce40a9SAlex Elder 23d14c5d2SYehuda Sadeh #include <linux/ceph/ceph_debug.h> 33d14c5d2SYehuda Sadeh 43d14c5d2SYehuda Sadeh #include <linux/module.h> 53d14c5d2SYehuda Sadeh #include <linux/err.h> 63d14c5d2SYehuda Sadeh #include <linux/highmem.h> 73d14c5d2SYehuda Sadeh #include <linux/mm.h> 83d14c5d2SYehuda Sadeh #include <linux/pagemap.h> 93d14c5d2SYehuda Sadeh #include <linux/slab.h> 103d14c5d2SYehuda Sadeh #include <linux/uaccess.h> 113d14c5d2SYehuda Sadeh #ifdef CONFIG_BLOCK 123d14c5d2SYehuda Sadeh #include <linux/bio.h> 133d14c5d2SYehuda Sadeh #endif 143d14c5d2SYehuda Sadeh 153d14c5d2SYehuda Sadeh #include <linux/ceph/libceph.h> 163d14c5d2SYehuda Sadeh #include <linux/ceph/osd_client.h> 173d14c5d2SYehuda Sadeh #include <linux/ceph/messenger.h> 183d14c5d2SYehuda Sadeh #include <linux/ceph/decode.h> 193d14c5d2SYehuda Sadeh #include <linux/ceph/auth.h> 203d14c5d2SYehuda Sadeh #include <linux/ceph/pagelist.h> 213d14c5d2SYehuda Sadeh 223d14c5d2SYehuda Sadeh #define OSD_OP_FRONT_LEN 4096 233d14c5d2SYehuda Sadeh #define OSD_OPREPLY_FRONT_LEN 512 243d14c5d2SYehuda Sadeh 253d14c5d2SYehuda Sadeh static const struct ceph_connection_operations osd_con_ops; 263d14c5d2SYehuda Sadeh 27f9d25199SAlex Elder static void __send_queued(struct ceph_osd_client *osdc); 286f6c7006SSage Weil static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd); 29a40c4f10SYehuda Sadeh static void __register_request(struct ceph_osd_client *osdc, 30a40c4f10SYehuda Sadeh struct ceph_osd_request *req); 31a40c4f10SYehuda Sadeh static void __unregister_linger_request(struct ceph_osd_client *osdc, 32a40c4f10SYehuda Sadeh struct ceph_osd_request *req); 3356e925b6SSage Weil static void __send_request(struct ceph_osd_client *osdc, 34a40c4f10SYehuda Sadeh struct ceph_osd_request *req); 353d14c5d2SYehuda Sadeh 363d14c5d2SYehuda Sadeh /* 373d14c5d2SYehuda Sadeh * Implement client access to distributed object storage cluster. 383d14c5d2SYehuda Sadeh * 393d14c5d2SYehuda Sadeh * All data objects are stored within a cluster/cloud of OSDs, or 403d14c5d2SYehuda Sadeh * "object storage devices." (Note that Ceph OSDs have _nothing_ to 413d14c5d2SYehuda Sadeh * do with the T10 OSD extensions to SCSI.) Ceph OSDs are simply 423d14c5d2SYehuda Sadeh * remote daemons serving up and coordinating consistent and safe 433d14c5d2SYehuda Sadeh * access to storage. 443d14c5d2SYehuda Sadeh * 453d14c5d2SYehuda Sadeh * Cluster membership and the mapping of data objects onto storage devices 463d14c5d2SYehuda Sadeh * are described by the osd map. 473d14c5d2SYehuda Sadeh * 483d14c5d2SYehuda Sadeh * We keep track of pending OSD requests (read, write), resubmit 493d14c5d2SYehuda Sadeh * requests to different OSDs when the cluster topology/data layout 503d14c5d2SYehuda Sadeh * change, or retry the affected requests when the communications 513d14c5d2SYehuda Sadeh * channel with an OSD is reset. 523d14c5d2SYehuda Sadeh */ 533d14c5d2SYehuda Sadeh 543d14c5d2SYehuda Sadeh /* 553d14c5d2SYehuda Sadeh * calculate the mapping of a file extent onto an object, and fill out the 563d14c5d2SYehuda Sadeh * request accordingly. shorten extent as necessary if it crosses an 573d14c5d2SYehuda Sadeh * object boundary. 583d14c5d2SYehuda Sadeh * 593d14c5d2SYehuda Sadeh * fill osd op in request message. 603d14c5d2SYehuda Sadeh */ 61dbe0fc41SAlex Elder static int calc_layout(struct ceph_file_layout *layout, u64 off, u64 *plen, 62a19dadfbSAlex Elder u64 *objnum, u64 *objoff, u64 *objlen) 633d14c5d2SYehuda Sadeh { 6460e56f13SAlex Elder u64 orig_len = *plen; 65d63b77f4SSage Weil int r; 663d14c5d2SYehuda Sadeh 6760e56f13SAlex Elder /* object extent? */ 6875d1c941SAlex Elder r = ceph_calc_file_object_mapping(layout, off, orig_len, objnum, 6975d1c941SAlex Elder objoff, objlen); 70d63b77f4SSage Weil if (r < 0) 71d63b77f4SSage Weil return r; 7275d1c941SAlex Elder if (*objlen < orig_len) { 7375d1c941SAlex Elder *plen = *objlen; 7460e56f13SAlex Elder dout(" skipping last %llu, final file extent %llu~%llu\n", 7560e56f13SAlex Elder orig_len - *plen, off, *plen); 7660e56f13SAlex Elder } 7760e56f13SAlex Elder 7875d1c941SAlex Elder dout("calc_layout objnum=%llx %llu~%llu\n", *objnum, *objoff, *objlen); 793d14c5d2SYehuda Sadeh 803ff5f385SAlex Elder return 0; 813d14c5d2SYehuda Sadeh } 823d14c5d2SYehuda Sadeh 83c54d47bfSAlex Elder static void ceph_osd_data_init(struct ceph_osd_data *osd_data) 84c54d47bfSAlex Elder { 85c54d47bfSAlex Elder memset(osd_data, 0, sizeof (*osd_data)); 86c54d47bfSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_NONE; 87c54d47bfSAlex Elder } 88c54d47bfSAlex Elder 89a4ce40a9SAlex Elder static void ceph_osd_data_pages_init(struct ceph_osd_data *osd_data, 9043bfe5deSAlex Elder struct page **pages, u64 length, u32 alignment, 9143bfe5deSAlex Elder bool pages_from_pool, bool own_pages) 9243bfe5deSAlex Elder { 9343bfe5deSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_PAGES; 9443bfe5deSAlex Elder osd_data->pages = pages; 9543bfe5deSAlex Elder osd_data->length = length; 9643bfe5deSAlex Elder osd_data->alignment = alignment; 9743bfe5deSAlex Elder osd_data->pages_from_pool = pages_from_pool; 9843bfe5deSAlex Elder osd_data->own_pages = own_pages; 9943bfe5deSAlex Elder } 10043bfe5deSAlex Elder 101a4ce40a9SAlex Elder static void ceph_osd_data_pagelist_init(struct ceph_osd_data *osd_data, 10243bfe5deSAlex Elder struct ceph_pagelist *pagelist) 10343bfe5deSAlex Elder { 10443bfe5deSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_PAGELIST; 10543bfe5deSAlex Elder osd_data->pagelist = pagelist; 10643bfe5deSAlex Elder } 10743bfe5deSAlex Elder 10843bfe5deSAlex Elder #ifdef CONFIG_BLOCK 109a4ce40a9SAlex Elder static void ceph_osd_data_bio_init(struct ceph_osd_data *osd_data, 11043bfe5deSAlex Elder struct bio *bio, size_t bio_length) 11143bfe5deSAlex Elder { 11243bfe5deSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_BIO; 11343bfe5deSAlex Elder osd_data->bio = bio; 11443bfe5deSAlex Elder osd_data->bio_length = bio_length; 11543bfe5deSAlex Elder } 11643bfe5deSAlex Elder #endif /* CONFIG_BLOCK */ 11743bfe5deSAlex Elder 118a4ce40a9SAlex Elder struct ceph_osd_data * 119a4ce40a9SAlex Elder osd_req_op_extent_osd_data(struct ceph_osd_request *osd_req, 120a4ce40a9SAlex Elder unsigned int which, bool write_request) 121a4ce40a9SAlex Elder { 122a4ce40a9SAlex Elder BUG_ON(which >= osd_req->r_num_ops); 123a4ce40a9SAlex Elder 1245476492fSAlex Elder return &osd_req->r_ops[which].extent.osd_data; 125a4ce40a9SAlex Elder } 126a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data); 127a4ce40a9SAlex Elder 128a4ce40a9SAlex Elder struct ceph_osd_data * 129a4ce40a9SAlex Elder osd_req_op_cls_request_info(struct ceph_osd_request *osd_req, 130a4ce40a9SAlex Elder unsigned int which) 131a4ce40a9SAlex Elder { 132a4ce40a9SAlex Elder BUG_ON(which >= osd_req->r_num_ops); 133a4ce40a9SAlex Elder 1345476492fSAlex Elder return &osd_req->r_ops[which].cls.request_info; 135a4ce40a9SAlex Elder } 136a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_request_info); /* ??? */ 137a4ce40a9SAlex Elder 138a4ce40a9SAlex Elder struct ceph_osd_data * 139a4ce40a9SAlex Elder osd_req_op_cls_response_data(struct ceph_osd_request *osd_req, 140a4ce40a9SAlex Elder unsigned int which) 141a4ce40a9SAlex Elder { 142a4ce40a9SAlex Elder BUG_ON(which >= osd_req->r_num_ops); 143a4ce40a9SAlex Elder 1445476492fSAlex Elder return &osd_req->r_ops[which].cls.response_data; 145a4ce40a9SAlex Elder } 146a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_response_data); /* ??? */ 147a4ce40a9SAlex Elder 148a4ce40a9SAlex Elder void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *osd_req, 149a4ce40a9SAlex Elder unsigned int which, bool write_request, 150a4ce40a9SAlex Elder struct page **pages, u64 length, u32 alignment, 151a4ce40a9SAlex Elder bool pages_from_pool, bool own_pages) 152a4ce40a9SAlex Elder { 153a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 154a4ce40a9SAlex Elder 155a4ce40a9SAlex Elder osd_data = osd_req_op_extent_osd_data(osd_req, which, write_request); 156a4ce40a9SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 157a4ce40a9SAlex Elder pages_from_pool, own_pages); 158a4ce40a9SAlex Elder } 159a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data_pages); 160a4ce40a9SAlex Elder 161a4ce40a9SAlex Elder void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *osd_req, 162a4ce40a9SAlex Elder unsigned int which, bool write_request, 163a4ce40a9SAlex Elder struct ceph_pagelist *pagelist) 164a4ce40a9SAlex Elder { 165a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 166a4ce40a9SAlex Elder 167a4ce40a9SAlex Elder osd_data = osd_req_op_extent_osd_data(osd_req, which, write_request); 168a4ce40a9SAlex Elder ceph_osd_data_pagelist_init(osd_data, pagelist); 169a4ce40a9SAlex Elder } 170a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data_pagelist); 171a4ce40a9SAlex Elder 172a4ce40a9SAlex Elder #ifdef CONFIG_BLOCK 173a4ce40a9SAlex Elder void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *osd_req, 174a4ce40a9SAlex Elder unsigned int which, bool write_request, 175a4ce40a9SAlex Elder struct bio *bio, size_t bio_length) 176a4ce40a9SAlex Elder { 177a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 178a4ce40a9SAlex Elder 179a4ce40a9SAlex Elder osd_data = osd_req_op_extent_osd_data(osd_req, which, write_request); 180a4ce40a9SAlex Elder ceph_osd_data_bio_init(osd_data, bio, bio_length); 181a4ce40a9SAlex Elder } 182a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data_bio); 183a4ce40a9SAlex Elder #endif /* CONFIG_BLOCK */ 184a4ce40a9SAlex Elder 185a4ce40a9SAlex Elder static void osd_req_op_cls_request_info_pagelist( 186a4ce40a9SAlex Elder struct ceph_osd_request *osd_req, 187a4ce40a9SAlex Elder unsigned int which, struct ceph_pagelist *pagelist) 188a4ce40a9SAlex Elder { 189a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 190a4ce40a9SAlex Elder 191a4ce40a9SAlex Elder osd_data = osd_req_op_cls_request_info(osd_req, which); 192a4ce40a9SAlex Elder ceph_osd_data_pagelist_init(osd_data, pagelist); 193a4ce40a9SAlex Elder } 194a4ce40a9SAlex Elder 195a4ce40a9SAlex Elder void osd_req_op_cls_response_data_pages(struct ceph_osd_request *osd_req, 196a4ce40a9SAlex Elder unsigned int which, struct page **pages, u64 length, 197a4ce40a9SAlex Elder u32 alignment, bool pages_from_pool, bool own_pages) 198a4ce40a9SAlex Elder { 199a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 200a4ce40a9SAlex Elder 201a4ce40a9SAlex Elder osd_data = osd_req_op_cls_response_data(osd_req, which); 202a4ce40a9SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 203a4ce40a9SAlex Elder pages_from_pool, own_pages); 204a4ce40a9SAlex Elder } 205a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_response_data_pages); 206a4ce40a9SAlex Elder 20723c08a9cSAlex Elder static u64 ceph_osd_data_length(struct ceph_osd_data *osd_data) 20823c08a9cSAlex Elder { 20923c08a9cSAlex Elder switch (osd_data->type) { 21023c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_NONE: 21123c08a9cSAlex Elder return 0; 21223c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_PAGES: 21323c08a9cSAlex Elder return osd_data->length; 21423c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_PAGELIST: 21523c08a9cSAlex Elder return (u64)osd_data->pagelist->length; 21623c08a9cSAlex Elder #ifdef CONFIG_BLOCK 21723c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_BIO: 21823c08a9cSAlex Elder return (u64)osd_data->bio_length; 21923c08a9cSAlex Elder #endif /* CONFIG_BLOCK */ 22023c08a9cSAlex Elder default: 22123c08a9cSAlex Elder WARN(true, "unrecognized data type %d\n", (int)osd_data->type); 22223c08a9cSAlex Elder return 0; 22323c08a9cSAlex Elder } 22423c08a9cSAlex Elder } 22523c08a9cSAlex Elder 2265476492fSAlex Elder 227c54d47bfSAlex Elder static void ceph_osd_data_release(struct ceph_osd_data *osd_data) 228c54d47bfSAlex Elder { 2295476492fSAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES && osd_data->own_pages) { 230c54d47bfSAlex Elder int num_pages; 231c54d47bfSAlex Elder 232c54d47bfSAlex Elder num_pages = calc_pages_for((u64)osd_data->alignment, 233c54d47bfSAlex Elder (u64)osd_data->length); 234c54d47bfSAlex Elder ceph_release_page_vector(osd_data->pages, num_pages); 235c54d47bfSAlex Elder } 2365476492fSAlex Elder ceph_osd_data_init(osd_data); 2375476492fSAlex Elder } 2385476492fSAlex Elder 2395476492fSAlex Elder static void osd_req_op_data_release(struct ceph_osd_request *osd_req, 2405476492fSAlex Elder unsigned int which) 2415476492fSAlex Elder { 2425476492fSAlex Elder struct ceph_osd_req_op *op; 2435476492fSAlex Elder 2445476492fSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 2455476492fSAlex Elder op = &osd_req->r_ops[which]; 2465476492fSAlex Elder 2475476492fSAlex Elder switch (op->op) { 2485476492fSAlex Elder case CEPH_OSD_OP_READ: 2495476492fSAlex Elder case CEPH_OSD_OP_WRITE: 2505476492fSAlex Elder ceph_osd_data_release(&op->extent.osd_data); 2515476492fSAlex Elder break; 2525476492fSAlex Elder case CEPH_OSD_OP_CALL: 2535476492fSAlex Elder ceph_osd_data_release(&op->cls.request_info); 2545476492fSAlex Elder ceph_osd_data_release(&op->cls.response_data); 2555476492fSAlex Elder break; 2565476492fSAlex Elder default: 2575476492fSAlex Elder break; 2585476492fSAlex Elder } 259c54d47bfSAlex Elder } 260c54d47bfSAlex Elder 2613d14c5d2SYehuda Sadeh /* 2623d14c5d2SYehuda Sadeh * requests 2633d14c5d2SYehuda Sadeh */ 2643d14c5d2SYehuda Sadeh void ceph_osdc_release_request(struct kref *kref) 2653d14c5d2SYehuda Sadeh { 266c54d47bfSAlex Elder struct ceph_osd_request *req; 2675476492fSAlex Elder unsigned int which; 2683d14c5d2SYehuda Sadeh 269c54d47bfSAlex Elder req = container_of(kref, struct ceph_osd_request, r_kref); 2703d14c5d2SYehuda Sadeh if (req->r_request) 2713d14c5d2SYehuda Sadeh ceph_msg_put(req->r_request); 272ace6d3a9SAlex Elder if (req->r_reply) { 2738921d114SAlex Elder ceph_msg_revoke_incoming(req->r_reply); 274ab8cb34aSAlex Elder ceph_msg_put(req->r_reply); 275ace6d3a9SAlex Elder } 2760fff87ecSAlex Elder 2775476492fSAlex Elder for (which = 0; which < req->r_num_ops; which++) 2785476492fSAlex Elder osd_req_op_data_release(req, which); 2790fff87ecSAlex Elder 2803d14c5d2SYehuda Sadeh ceph_put_snap_context(req->r_snapc); 2813d14c5d2SYehuda Sadeh if (req->r_mempool) 2823d14c5d2SYehuda Sadeh mempool_free(req, req->r_osdc->req_mempool); 2833d14c5d2SYehuda Sadeh else 2843d14c5d2SYehuda Sadeh kfree(req); 2853d14c5d2SYehuda Sadeh } 2863d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_release_request); 2873d14c5d2SYehuda Sadeh 2883d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc, 2893d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 2901b83bef2SSage Weil unsigned int num_ops, 2913d14c5d2SYehuda Sadeh bool use_mempool, 29254a54007SAlex Elder gfp_t gfp_flags) 2933d14c5d2SYehuda Sadeh { 2943d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 2953d14c5d2SYehuda Sadeh struct ceph_msg *msg; 2961b83bef2SSage Weil size_t msg_size; 2973d14c5d2SYehuda Sadeh 29879528734SAlex Elder BUILD_BUG_ON(CEPH_OSD_MAX_OP > U16_MAX); 29979528734SAlex Elder BUG_ON(num_ops > CEPH_OSD_MAX_OP); 30079528734SAlex Elder 3011b83bef2SSage Weil msg_size = 4 + 4 + 8 + 8 + 4+8; 3021b83bef2SSage Weil msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */ 3031b83bef2SSage Weil msg_size += 1 + 8 + 4 + 4; /* pg_t */ 3041b83bef2SSage Weil msg_size += 4 + MAX_OBJ_NAME_SIZE; 3051b83bef2SSage Weil msg_size += 2 + num_ops*sizeof(struct ceph_osd_op); 3061b83bef2SSage Weil msg_size += 8; /* snapid */ 3071b83bef2SSage Weil msg_size += 8; /* snap_seq */ 3081b83bef2SSage Weil msg_size += 8 * (snapc ? snapc->num_snaps : 0); /* snaps */ 3091b83bef2SSage Weil msg_size += 4; 3103d14c5d2SYehuda Sadeh 3113d14c5d2SYehuda Sadeh if (use_mempool) { 3123d14c5d2SYehuda Sadeh req = mempool_alloc(osdc->req_mempool, gfp_flags); 3133d14c5d2SYehuda Sadeh memset(req, 0, sizeof(*req)); 3143d14c5d2SYehuda Sadeh } else { 3153d14c5d2SYehuda Sadeh req = kzalloc(sizeof(*req), gfp_flags); 3163d14c5d2SYehuda Sadeh } 3173d14c5d2SYehuda Sadeh if (req == NULL) 3183d14c5d2SYehuda Sadeh return NULL; 3193d14c5d2SYehuda Sadeh 3203d14c5d2SYehuda Sadeh req->r_osdc = osdc; 3213d14c5d2SYehuda Sadeh req->r_mempool = use_mempool; 32279528734SAlex Elder req->r_num_ops = num_ops; 3233d14c5d2SYehuda Sadeh 3243d14c5d2SYehuda Sadeh kref_init(&req->r_kref); 3253d14c5d2SYehuda Sadeh init_completion(&req->r_completion); 3263d14c5d2SYehuda Sadeh init_completion(&req->r_safe_completion); 327a978fa20SAlex Elder RB_CLEAR_NODE(&req->r_node); 3283d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&req->r_unsafe_item); 329a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&req->r_linger_item); 330a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&req->r_linger_osd); 331935b639aSSage Weil INIT_LIST_HEAD(&req->r_req_lru_item); 332cd43045cSSage Weil INIT_LIST_HEAD(&req->r_osd_item); 333cd43045cSSage Weil 3343d14c5d2SYehuda Sadeh /* create reply message */ 3353d14c5d2SYehuda Sadeh if (use_mempool) 3363d14c5d2SYehuda Sadeh msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0); 3373d14c5d2SYehuda Sadeh else 3383d14c5d2SYehuda Sadeh msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, 339b61c2763SSage Weil OSD_OPREPLY_FRONT_LEN, gfp_flags, true); 3403d14c5d2SYehuda Sadeh if (!msg) { 3413d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 3423d14c5d2SYehuda Sadeh return NULL; 3433d14c5d2SYehuda Sadeh } 3443d14c5d2SYehuda Sadeh req->r_reply = msg; 3453d14c5d2SYehuda Sadeh 3463d14c5d2SYehuda Sadeh /* create request message; allow space for oid */ 3473d14c5d2SYehuda Sadeh if (use_mempool) 3483d14c5d2SYehuda Sadeh msg = ceph_msgpool_get(&osdc->msgpool_op, 0); 3493d14c5d2SYehuda Sadeh else 350b61c2763SSage Weil msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true); 3513d14c5d2SYehuda Sadeh if (!msg) { 3523d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 3533d14c5d2SYehuda Sadeh return NULL; 3543d14c5d2SYehuda Sadeh } 3553d14c5d2SYehuda Sadeh 3563d14c5d2SYehuda Sadeh memset(msg->front.iov_base, 0, msg->front.iov_len); 3573d14c5d2SYehuda Sadeh 3583d14c5d2SYehuda Sadeh req->r_request = msg; 3593d14c5d2SYehuda Sadeh 3603d14c5d2SYehuda Sadeh return req; 3613d14c5d2SYehuda Sadeh } 3623d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_alloc_request); 3633d14c5d2SYehuda Sadeh 364a8dd0a37SAlex Elder static bool osd_req_opcode_valid(u16 opcode) 365a8dd0a37SAlex Elder { 366a8dd0a37SAlex Elder switch (opcode) { 367a8dd0a37SAlex Elder case CEPH_OSD_OP_READ: 368a8dd0a37SAlex Elder case CEPH_OSD_OP_STAT: 369a8dd0a37SAlex Elder case CEPH_OSD_OP_MAPEXT: 370a8dd0a37SAlex Elder case CEPH_OSD_OP_MASKTRUNC: 371a8dd0a37SAlex Elder case CEPH_OSD_OP_SPARSE_READ: 372a8dd0a37SAlex Elder case CEPH_OSD_OP_NOTIFY: 373a8dd0a37SAlex Elder case CEPH_OSD_OP_NOTIFY_ACK: 374a8dd0a37SAlex Elder case CEPH_OSD_OP_ASSERT_VER: 375a8dd0a37SAlex Elder case CEPH_OSD_OP_WRITE: 376a8dd0a37SAlex Elder case CEPH_OSD_OP_WRITEFULL: 377a8dd0a37SAlex Elder case CEPH_OSD_OP_TRUNCATE: 378a8dd0a37SAlex Elder case CEPH_OSD_OP_ZERO: 379a8dd0a37SAlex Elder case CEPH_OSD_OP_DELETE: 380a8dd0a37SAlex Elder case CEPH_OSD_OP_APPEND: 381a8dd0a37SAlex Elder case CEPH_OSD_OP_STARTSYNC: 382a8dd0a37SAlex Elder case CEPH_OSD_OP_SETTRUNC: 383a8dd0a37SAlex Elder case CEPH_OSD_OP_TRIMTRUNC: 384a8dd0a37SAlex Elder case CEPH_OSD_OP_TMAPUP: 385a8dd0a37SAlex Elder case CEPH_OSD_OP_TMAPPUT: 386a8dd0a37SAlex Elder case CEPH_OSD_OP_TMAPGET: 387a8dd0a37SAlex Elder case CEPH_OSD_OP_CREATE: 388a8dd0a37SAlex Elder case CEPH_OSD_OP_ROLLBACK: 389a8dd0a37SAlex Elder case CEPH_OSD_OP_WATCH: 390a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPGETKEYS: 391a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPGETVALS: 392a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPGETHEADER: 393a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPGETVALSBYKEYS: 394a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPSETVALS: 395a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPSETHEADER: 396a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPCLEAR: 397a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPRMKEYS: 398a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAP_CMP: 399a8dd0a37SAlex Elder case CEPH_OSD_OP_CLONERANGE: 400a8dd0a37SAlex Elder case CEPH_OSD_OP_ASSERT_SRC_VERSION: 401a8dd0a37SAlex Elder case CEPH_OSD_OP_SRC_CMPXATTR: 402a8dd0a37SAlex Elder case CEPH_OSD_OP_GETXATTR: 403a8dd0a37SAlex Elder case CEPH_OSD_OP_GETXATTRS: 404a8dd0a37SAlex Elder case CEPH_OSD_OP_CMPXATTR: 405a8dd0a37SAlex Elder case CEPH_OSD_OP_SETXATTR: 406a8dd0a37SAlex Elder case CEPH_OSD_OP_SETXATTRS: 407a8dd0a37SAlex Elder case CEPH_OSD_OP_RESETXATTRS: 408a8dd0a37SAlex Elder case CEPH_OSD_OP_RMXATTR: 409a8dd0a37SAlex Elder case CEPH_OSD_OP_PULL: 410a8dd0a37SAlex Elder case CEPH_OSD_OP_PUSH: 411a8dd0a37SAlex Elder case CEPH_OSD_OP_BALANCEREADS: 412a8dd0a37SAlex Elder case CEPH_OSD_OP_UNBALANCEREADS: 413a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB: 414a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB_RESERVE: 415a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB_UNRESERVE: 416a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB_STOP: 417a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB_MAP: 418a8dd0a37SAlex Elder case CEPH_OSD_OP_WRLOCK: 419a8dd0a37SAlex Elder case CEPH_OSD_OP_WRUNLOCK: 420a8dd0a37SAlex Elder case CEPH_OSD_OP_RDLOCK: 421a8dd0a37SAlex Elder case CEPH_OSD_OP_RDUNLOCK: 422a8dd0a37SAlex Elder case CEPH_OSD_OP_UPLOCK: 423a8dd0a37SAlex Elder case CEPH_OSD_OP_DNLOCK: 424a8dd0a37SAlex Elder case CEPH_OSD_OP_CALL: 425a8dd0a37SAlex Elder case CEPH_OSD_OP_PGLS: 426a8dd0a37SAlex Elder case CEPH_OSD_OP_PGLS_FILTER: 427a8dd0a37SAlex Elder return true; 428a8dd0a37SAlex Elder default: 429a8dd0a37SAlex Elder return false; 430a8dd0a37SAlex Elder } 431a8dd0a37SAlex Elder } 432a8dd0a37SAlex Elder 43333803f33SAlex Elder /* 43433803f33SAlex Elder * This is an osd op init function for opcodes that have no data or 43533803f33SAlex Elder * other information associated with them. It also serves as a 43633803f33SAlex Elder * common init routine for all the other init functions, below. 43733803f33SAlex Elder */ 438c99d2d4aSAlex Elder static struct ceph_osd_req_op * 439c99d2d4aSAlex Elder osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which, 440c99d2d4aSAlex Elder u16 opcode) 44133803f33SAlex Elder { 442c99d2d4aSAlex Elder struct ceph_osd_req_op *op; 443c99d2d4aSAlex Elder 444c99d2d4aSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 44533803f33SAlex Elder BUG_ON(!osd_req_opcode_valid(opcode)); 44633803f33SAlex Elder 447c99d2d4aSAlex Elder op = &osd_req->r_ops[which]; 44833803f33SAlex Elder memset(op, 0, sizeof (*op)); 44933803f33SAlex Elder op->op = opcode; 450c99d2d4aSAlex Elder 451c99d2d4aSAlex Elder return op; 45233803f33SAlex Elder } 45333803f33SAlex Elder 454c99d2d4aSAlex Elder void osd_req_op_extent_init(struct ceph_osd_request *osd_req, 455c99d2d4aSAlex Elder unsigned int which, u16 opcode, 45633803f33SAlex Elder u64 offset, u64 length, 45733803f33SAlex Elder u64 truncate_size, u32 truncate_seq) 45833803f33SAlex Elder { 459c99d2d4aSAlex Elder struct ceph_osd_req_op *op = osd_req_op_init(osd_req, which, opcode); 46033803f33SAlex Elder size_t payload_len = 0; 46133803f33SAlex Elder 46233803f33SAlex Elder BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE); 46333803f33SAlex Elder 46433803f33SAlex Elder op->extent.offset = offset; 46533803f33SAlex Elder op->extent.length = length; 46633803f33SAlex Elder op->extent.truncate_size = truncate_size; 46733803f33SAlex Elder op->extent.truncate_seq = truncate_seq; 46833803f33SAlex Elder if (opcode == CEPH_OSD_OP_WRITE) 46933803f33SAlex Elder payload_len += length; 47033803f33SAlex Elder 47133803f33SAlex Elder op->payload_len = payload_len; 47233803f33SAlex Elder } 47333803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_init); 47433803f33SAlex Elder 475c99d2d4aSAlex Elder void osd_req_op_extent_update(struct ceph_osd_request *osd_req, 476c99d2d4aSAlex Elder unsigned int which, u64 length) 477e5975c7cSAlex Elder { 478c99d2d4aSAlex Elder struct ceph_osd_req_op *op; 479c99d2d4aSAlex Elder u64 previous; 480c99d2d4aSAlex Elder 481c99d2d4aSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 482c99d2d4aSAlex Elder op = &osd_req->r_ops[which]; 483c99d2d4aSAlex Elder previous = op->extent.length; 484e5975c7cSAlex Elder 485e5975c7cSAlex Elder if (length == previous) 486e5975c7cSAlex Elder return; /* Nothing to do */ 487e5975c7cSAlex Elder BUG_ON(length > previous); 488e5975c7cSAlex Elder 489e5975c7cSAlex Elder op->extent.length = length; 490e5975c7cSAlex Elder op->payload_len -= previous - length; 491e5975c7cSAlex Elder } 492e5975c7cSAlex Elder EXPORT_SYMBOL(osd_req_op_extent_update); 493e5975c7cSAlex Elder 494c99d2d4aSAlex Elder void osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which, 495c99d2d4aSAlex Elder u16 opcode, const char *class, const char *method, 49633803f33SAlex Elder const void *request_data, size_t request_data_size) 49733803f33SAlex Elder { 498c99d2d4aSAlex Elder struct ceph_osd_req_op *op = osd_req_op_init(osd_req, which, opcode); 4995f562df5SAlex Elder struct ceph_pagelist *pagelist; 50033803f33SAlex Elder size_t payload_len = 0; 50133803f33SAlex Elder size_t size; 50233803f33SAlex Elder 50333803f33SAlex Elder BUG_ON(opcode != CEPH_OSD_OP_CALL); 50433803f33SAlex Elder 5055f562df5SAlex Elder pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS); 5065f562df5SAlex Elder BUG_ON(!pagelist); 5075f562df5SAlex Elder ceph_pagelist_init(pagelist); 5085f562df5SAlex Elder 50933803f33SAlex Elder op->cls.class_name = class; 51033803f33SAlex Elder size = strlen(class); 51133803f33SAlex Elder BUG_ON(size > (size_t) U8_MAX); 51233803f33SAlex Elder op->cls.class_len = size; 5135f562df5SAlex Elder ceph_pagelist_append(pagelist, class, size); 51433803f33SAlex Elder payload_len += size; 51533803f33SAlex Elder 51633803f33SAlex Elder op->cls.method_name = method; 51733803f33SAlex Elder size = strlen(method); 51833803f33SAlex Elder BUG_ON(size > (size_t) U8_MAX); 51933803f33SAlex Elder op->cls.method_len = size; 5205f562df5SAlex Elder ceph_pagelist_append(pagelist, method, size); 52133803f33SAlex Elder payload_len += size; 52233803f33SAlex Elder 52354d50649SAlex Elder op->cls.request_data = request_data; 52433803f33SAlex Elder BUG_ON(request_data_size > (size_t) U32_MAX); 52554d50649SAlex Elder op->cls.request_data_len = (u32) request_data_size; 5265f562df5SAlex Elder ceph_pagelist_append(pagelist, request_data, request_data_size); 52733803f33SAlex Elder payload_len += request_data_size; 52833803f33SAlex Elder 529a4ce40a9SAlex Elder osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist); 5305f562df5SAlex Elder 53133803f33SAlex Elder op->cls.argc = 0; /* currently unused */ 53233803f33SAlex Elder 53333803f33SAlex Elder op->payload_len = payload_len; 53433803f33SAlex Elder } 53533803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_init); 5368c042b0dSAlex Elder 537c99d2d4aSAlex Elder void osd_req_op_watch_init(struct ceph_osd_request *osd_req, 538c99d2d4aSAlex Elder unsigned int which, u16 opcode, 53933803f33SAlex Elder u64 cookie, u64 version, int flag) 54033803f33SAlex Elder { 541c99d2d4aSAlex Elder struct ceph_osd_req_op *op = osd_req_op_init(osd_req, which, opcode); 54233803f33SAlex Elder 543c99d2d4aSAlex Elder BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH); 54433803f33SAlex Elder 54533803f33SAlex Elder op->watch.cookie = cookie; 54633803f33SAlex Elder /* op->watch.ver = version; */ /* XXX 3847 */ 54733803f33SAlex Elder op->watch.ver = cpu_to_le64(version); 54833803f33SAlex Elder if (opcode == CEPH_OSD_OP_WATCH && flag) 54933803f33SAlex Elder op->watch.flag = (u8)1; 55033803f33SAlex Elder } 55133803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_watch_init); 55233803f33SAlex Elder 553ec9123c5SAlex Elder static void ceph_osdc_msg_data_set(struct ceph_msg *msg, 554ec9123c5SAlex Elder struct ceph_osd_data *osd_data) 555ec9123c5SAlex Elder { 556ec9123c5SAlex Elder u64 length = ceph_osd_data_length(osd_data); 557ec9123c5SAlex Elder 558ec9123c5SAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 559ec9123c5SAlex Elder BUG_ON(length > (u64) SIZE_MAX); 560ec9123c5SAlex Elder if (length) 561ec9123c5SAlex Elder ceph_msg_data_set_pages(msg, osd_data->pages, 562ec9123c5SAlex Elder length, osd_data->alignment); 563ec9123c5SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) { 564ec9123c5SAlex Elder BUG_ON(!length); 565ec9123c5SAlex Elder ceph_msg_data_set_pagelist(msg, osd_data->pagelist); 566ec9123c5SAlex Elder #ifdef CONFIG_BLOCK 567ec9123c5SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) { 568ec9123c5SAlex Elder ceph_msg_data_set_bio(msg, osd_data->bio, length); 569ec9123c5SAlex Elder #endif 570ec9123c5SAlex Elder } else { 571ec9123c5SAlex Elder BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE); 572ec9123c5SAlex Elder } 573ec9123c5SAlex Elder } 574ec9123c5SAlex Elder 575175face2SAlex Elder static u64 osd_req_encode_op(struct ceph_osd_request *req, 57679528734SAlex Elder struct ceph_osd_op *dst, unsigned int which) 5773d14c5d2SYehuda Sadeh { 57879528734SAlex Elder struct ceph_osd_req_op *src; 57954d50649SAlex Elder u64 request_data_len = 0; 580175face2SAlex Elder 58179528734SAlex Elder BUG_ON(which >= req->r_num_ops); 58279528734SAlex Elder src = &req->r_ops[which]; 583a8dd0a37SAlex Elder if (WARN_ON(!osd_req_opcode_valid(src->op))) { 584a8dd0a37SAlex Elder pr_err("unrecognized osd opcode %d\n", src->op); 585a8dd0a37SAlex Elder 586a8dd0a37SAlex Elder return 0; 587a8dd0a37SAlex Elder } 5883d14c5d2SYehuda Sadeh 589065a68f9SAlex Elder switch (src->op) { 590fbfab539SAlex Elder case CEPH_OSD_OP_STAT: 591fbfab539SAlex Elder break; 5923d14c5d2SYehuda Sadeh case CEPH_OSD_OP_READ: 5933d14c5d2SYehuda Sadeh case CEPH_OSD_OP_WRITE: 594175face2SAlex Elder if (src->op == CEPH_OSD_OP_WRITE) 59554d50649SAlex Elder request_data_len = src->extent.length; 596175face2SAlex Elder dst->extent.offset = cpu_to_le64(src->extent.offset); 597175face2SAlex Elder dst->extent.length = cpu_to_le64(src->extent.length); 5983d14c5d2SYehuda Sadeh dst->extent.truncate_size = 5993d14c5d2SYehuda Sadeh cpu_to_le64(src->extent.truncate_size); 6003d14c5d2SYehuda Sadeh dst->extent.truncate_seq = 6013d14c5d2SYehuda Sadeh cpu_to_le32(src->extent.truncate_seq); 6025476492fSAlex Elder if (src->op == CEPH_OSD_OP_WRITE) 603ec9123c5SAlex Elder ceph_osdc_msg_data_set(req->r_request, 6045476492fSAlex Elder &src->extent.osd_data); 6055476492fSAlex Elder else 606ec9123c5SAlex Elder ceph_osdc_msg_data_set(req->r_reply, 6075476492fSAlex Elder &src->extent.osd_data); 6083d14c5d2SYehuda Sadeh break; 6093d14c5d2SYehuda Sadeh case CEPH_OSD_OP_CALL: 6103d14c5d2SYehuda Sadeh dst->cls.class_len = src->cls.class_len; 6113d14c5d2SYehuda Sadeh dst->cls.method_len = src->cls.method_len; 61254d50649SAlex Elder dst->cls.indata_len = cpu_to_le32(src->cls.request_data_len); 6135476492fSAlex Elder ceph_osdc_msg_data_set(req->r_reply, &src->cls.response_data); 6145476492fSAlex Elder ceph_osdc_msg_data_set(req->r_request, &src->cls.request_info); 6155476492fSAlex Elder BUG_ON(src->cls.request_info.type != 6165f562df5SAlex Elder CEPH_OSD_DATA_TYPE_PAGELIST); 6175476492fSAlex Elder request_data_len = src->cls.request_info.pagelist->length; 6183d14c5d2SYehuda Sadeh break; 6193d14c5d2SYehuda Sadeh case CEPH_OSD_OP_STARTSYNC: 6203d14c5d2SYehuda Sadeh break; 621a40c4f10SYehuda Sadeh case CEPH_OSD_OP_NOTIFY_ACK: 622a40c4f10SYehuda Sadeh case CEPH_OSD_OP_WATCH: 623a40c4f10SYehuda Sadeh dst->watch.cookie = cpu_to_le64(src->watch.cookie); 624a40c4f10SYehuda Sadeh dst->watch.ver = cpu_to_le64(src->watch.ver); 625a40c4f10SYehuda Sadeh dst->watch.flag = src->watch.flag; 626a40c4f10SYehuda Sadeh break; 6273d14c5d2SYehuda Sadeh default: 6284c46459cSAlex Elder pr_err("unsupported osd opcode %s\n", 6298f63ca2dSAlex Elder ceph_osd_op_name(src->op)); 6304c46459cSAlex Elder WARN_ON(1); 631a8dd0a37SAlex Elder 632a8dd0a37SAlex Elder return 0; 6333d14c5d2SYehuda Sadeh } 634a8dd0a37SAlex Elder dst->op = cpu_to_le16(src->op); 6353d14c5d2SYehuda Sadeh dst->payload_len = cpu_to_le32(src->payload_len); 636175face2SAlex Elder 63754d50649SAlex Elder return request_data_len; 6383d14c5d2SYehuda Sadeh } 6393d14c5d2SYehuda Sadeh 6403d14c5d2SYehuda Sadeh /* 6413d14c5d2SYehuda Sadeh * build new request AND message, calculate layout, and adjust file 6423d14c5d2SYehuda Sadeh * extent as needed. 6433d14c5d2SYehuda Sadeh * 6443d14c5d2SYehuda Sadeh * if the file was recently truncated, we include information about its 6453d14c5d2SYehuda Sadeh * old and new size so that the object can be updated appropriately. (we 6463d14c5d2SYehuda Sadeh * avoid synchronously deleting truncated objects because it's slow.) 6473d14c5d2SYehuda Sadeh * 6483d14c5d2SYehuda Sadeh * if @do_sync, include a 'startsync' command so that the osd will flush 6493d14c5d2SYehuda Sadeh * data quickly. 6503d14c5d2SYehuda Sadeh */ 6513d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc, 6523d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 6533d14c5d2SYehuda Sadeh struct ceph_vino vino, 654acead002SAlex Elder u64 off, u64 *plen, int num_ops, 6553d14c5d2SYehuda Sadeh int opcode, int flags, 6563d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 6573d14c5d2SYehuda Sadeh u32 truncate_seq, 6583d14c5d2SYehuda Sadeh u64 truncate_size, 659153e5167SAlex Elder bool use_mempool) 6603d14c5d2SYehuda Sadeh { 6613d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 66275d1c941SAlex Elder u64 objnum = 0; 66375d1c941SAlex Elder u64 objoff = 0; 66475d1c941SAlex Elder u64 objlen = 0; 665d18d1e28SAlex Elder u32 object_size; 666d18d1e28SAlex Elder u64 object_base; 6676816282dSSage Weil int r; 6683d14c5d2SYehuda Sadeh 669d18d1e28SAlex Elder BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE); 6703d14c5d2SYehuda Sadeh 671acead002SAlex Elder req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool, 672ae7ca4a3SAlex Elder GFP_NOFS); 6734ad12621SSage Weil if (!req) 6746816282dSSage Weil return ERR_PTR(-ENOMEM); 67579528734SAlex Elder 676d178a9e7SAlex Elder req->r_flags = flags; 6773d14c5d2SYehuda Sadeh 6783d14c5d2SYehuda Sadeh /* calculate max write size */ 679a19dadfbSAlex Elder r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen); 6803ff5f385SAlex Elder if (r < 0) { 6813ff5f385SAlex Elder ceph_osdc_put_request(req); 6826816282dSSage Weil return ERR_PTR(r); 6833ff5f385SAlex Elder } 684a19dadfbSAlex Elder 685d18d1e28SAlex Elder object_size = le32_to_cpu(layout->fl_object_size); 686d18d1e28SAlex Elder object_base = off - objoff; 687d18d1e28SAlex Elder if (truncate_size <= object_base) { 688d18d1e28SAlex Elder truncate_size = 0; 689d18d1e28SAlex Elder } else { 690d18d1e28SAlex Elder truncate_size -= object_base; 691d18d1e28SAlex Elder if (truncate_size > object_size) 692d18d1e28SAlex Elder truncate_size = object_size; 693d18d1e28SAlex Elder } 694d18d1e28SAlex Elder 695c99d2d4aSAlex Elder osd_req_op_extent_init(req, 0, opcode, objoff, objlen, 696b0270324SAlex Elder truncate_size, truncate_seq); 6978c042b0dSAlex Elder 698acead002SAlex Elder /* 699acead002SAlex Elder * A second op in the ops array means the caller wants to 700acead002SAlex Elder * also issue a include a 'startsync' command so that the 701acead002SAlex Elder * osd will flush data quickly. 702acead002SAlex Elder */ 703acead002SAlex Elder if (num_ops > 1) 704c99d2d4aSAlex Elder osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC); 705d18d1e28SAlex Elder 7063d14c5d2SYehuda Sadeh req->r_file_layout = *layout; /* keep a copy */ 7073d14c5d2SYehuda Sadeh 70875d1c941SAlex Elder snprintf(req->r_oid, sizeof(req->r_oid), "%llx.%08llx", 70975d1c941SAlex Elder vino.ino, objnum); 710dbe0fc41SAlex Elder req->r_oid_len = strlen(req->r_oid); 711dbe0fc41SAlex Elder 7123d14c5d2SYehuda Sadeh return req; 7133d14c5d2SYehuda Sadeh } 7143d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_new_request); 7153d14c5d2SYehuda Sadeh 7163d14c5d2SYehuda Sadeh /* 7173d14c5d2SYehuda Sadeh * We keep osd requests in an rbtree, sorted by ->r_tid. 7183d14c5d2SYehuda Sadeh */ 7193d14c5d2SYehuda Sadeh static void __insert_request(struct ceph_osd_client *osdc, 7203d14c5d2SYehuda Sadeh struct ceph_osd_request *new) 7213d14c5d2SYehuda Sadeh { 7223d14c5d2SYehuda Sadeh struct rb_node **p = &osdc->requests.rb_node; 7233d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 7243d14c5d2SYehuda Sadeh struct ceph_osd_request *req = NULL; 7253d14c5d2SYehuda Sadeh 7263d14c5d2SYehuda Sadeh while (*p) { 7273d14c5d2SYehuda Sadeh parent = *p; 7283d14c5d2SYehuda Sadeh req = rb_entry(parent, struct ceph_osd_request, r_node); 7293d14c5d2SYehuda Sadeh if (new->r_tid < req->r_tid) 7303d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 7313d14c5d2SYehuda Sadeh else if (new->r_tid > req->r_tid) 7323d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 7333d14c5d2SYehuda Sadeh else 7343d14c5d2SYehuda Sadeh BUG(); 7353d14c5d2SYehuda Sadeh } 7363d14c5d2SYehuda Sadeh 7373d14c5d2SYehuda Sadeh rb_link_node(&new->r_node, parent, p); 7383d14c5d2SYehuda Sadeh rb_insert_color(&new->r_node, &osdc->requests); 7393d14c5d2SYehuda Sadeh } 7403d14c5d2SYehuda Sadeh 7413d14c5d2SYehuda Sadeh static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc, 7423d14c5d2SYehuda Sadeh u64 tid) 7433d14c5d2SYehuda Sadeh { 7443d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 7453d14c5d2SYehuda Sadeh struct rb_node *n = osdc->requests.rb_node; 7463d14c5d2SYehuda Sadeh 7473d14c5d2SYehuda Sadeh while (n) { 7483d14c5d2SYehuda Sadeh req = rb_entry(n, struct ceph_osd_request, r_node); 7493d14c5d2SYehuda Sadeh if (tid < req->r_tid) 7503d14c5d2SYehuda Sadeh n = n->rb_left; 7513d14c5d2SYehuda Sadeh else if (tid > req->r_tid) 7523d14c5d2SYehuda Sadeh n = n->rb_right; 7533d14c5d2SYehuda Sadeh else 7543d14c5d2SYehuda Sadeh return req; 7553d14c5d2SYehuda Sadeh } 7563d14c5d2SYehuda Sadeh return NULL; 7573d14c5d2SYehuda Sadeh } 7583d14c5d2SYehuda Sadeh 7593d14c5d2SYehuda Sadeh static struct ceph_osd_request * 7603d14c5d2SYehuda Sadeh __lookup_request_ge(struct ceph_osd_client *osdc, 7613d14c5d2SYehuda Sadeh u64 tid) 7623d14c5d2SYehuda Sadeh { 7633d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 7643d14c5d2SYehuda Sadeh struct rb_node *n = osdc->requests.rb_node; 7653d14c5d2SYehuda Sadeh 7663d14c5d2SYehuda Sadeh while (n) { 7673d14c5d2SYehuda Sadeh req = rb_entry(n, struct ceph_osd_request, r_node); 7683d14c5d2SYehuda Sadeh if (tid < req->r_tid) { 7693d14c5d2SYehuda Sadeh if (!n->rb_left) 7703d14c5d2SYehuda Sadeh return req; 7713d14c5d2SYehuda Sadeh n = n->rb_left; 7723d14c5d2SYehuda Sadeh } else if (tid > req->r_tid) { 7733d14c5d2SYehuda Sadeh n = n->rb_right; 7743d14c5d2SYehuda Sadeh } else { 7753d14c5d2SYehuda Sadeh return req; 7763d14c5d2SYehuda Sadeh } 7773d14c5d2SYehuda Sadeh } 7783d14c5d2SYehuda Sadeh return NULL; 7793d14c5d2SYehuda Sadeh } 7803d14c5d2SYehuda Sadeh 7816f6c7006SSage Weil /* 7826f6c7006SSage Weil * Resubmit requests pending on the given osd. 7836f6c7006SSage Weil */ 7846f6c7006SSage Weil static void __kick_osd_requests(struct ceph_osd_client *osdc, 7856f6c7006SSage Weil struct ceph_osd *osd) 7866f6c7006SSage Weil { 787a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 788e02493c0SAlex Elder LIST_HEAD(resend); 7896f6c7006SSage Weil int err; 7906f6c7006SSage Weil 7916f6c7006SSage Weil dout("__kick_osd_requests osd%d\n", osd->o_osd); 7926f6c7006SSage Weil err = __reset_osd(osdc, osd); 793685a7555SAlex Elder if (err) 7946f6c7006SSage Weil return; 795e02493c0SAlex Elder /* 796e02493c0SAlex Elder * Build up a list of requests to resend by traversing the 797e02493c0SAlex Elder * osd's list of requests. Requests for a given object are 798e02493c0SAlex Elder * sent in tid order, and that is also the order they're 799e02493c0SAlex Elder * kept on this list. Therefore all requests that are in 800e02493c0SAlex Elder * flight will be found first, followed by all requests that 801e02493c0SAlex Elder * have not yet been sent. And to resend requests while 802e02493c0SAlex Elder * preserving this order we will want to put any sent 803e02493c0SAlex Elder * requests back on the front of the osd client's unsent 804e02493c0SAlex Elder * list. 805e02493c0SAlex Elder * 806e02493c0SAlex Elder * So we build a separate ordered list of already-sent 807e02493c0SAlex Elder * requests for the affected osd and splice it onto the 808e02493c0SAlex Elder * front of the osd client's unsent list. Once we've seen a 809e02493c0SAlex Elder * request that has not yet been sent we're done. Those 810e02493c0SAlex Elder * requests are already sitting right where they belong. 811e02493c0SAlex Elder */ 8126f6c7006SSage Weil list_for_each_entry(req, &osd->o_requests, r_osd_item) { 813e02493c0SAlex Elder if (!req->r_sent) 814e02493c0SAlex Elder break; 815e02493c0SAlex Elder list_move_tail(&req->r_req_lru_item, &resend); 816e02493c0SAlex Elder dout("requeueing %p tid %llu osd%d\n", req, req->r_tid, 8176f6c7006SSage Weil osd->o_osd); 818a40c4f10SYehuda Sadeh if (!req->r_linger) 8196f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 8206f6c7006SSage Weil } 821e02493c0SAlex Elder list_splice(&resend, &osdc->req_unsent); 822a40c4f10SYehuda Sadeh 823e02493c0SAlex Elder /* 824e02493c0SAlex Elder * Linger requests are re-registered before sending, which 825e02493c0SAlex Elder * sets up a new tid for each. We add them to the unsent 826e02493c0SAlex Elder * list at the end to keep things in tid order. 827e02493c0SAlex Elder */ 828a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osd->o_linger_requests, 829a40c4f10SYehuda Sadeh r_linger_osd) { 83077f38e0eSSage Weil /* 83177f38e0eSSage Weil * reregister request prior to unregistering linger so 83277f38e0eSSage Weil * that r_osd is preserved. 83377f38e0eSSage Weil */ 83477f38e0eSSage Weil BUG_ON(!list_empty(&req->r_req_lru_item)); 835a40c4f10SYehuda Sadeh __register_request(osdc, req); 836e02493c0SAlex Elder list_add_tail(&req->r_req_lru_item, &osdc->req_unsent); 837ad885927SAlex Elder list_add_tail(&req->r_osd_item, &req->r_osd->o_requests); 83877f38e0eSSage Weil __unregister_linger_request(osdc, req); 839a40c4f10SYehuda Sadeh dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid, 840a40c4f10SYehuda Sadeh osd->o_osd); 841a40c4f10SYehuda Sadeh } 8426f6c7006SSage Weil } 8436f6c7006SSage Weil 8443d14c5d2SYehuda Sadeh /* 8453d14c5d2SYehuda Sadeh * If the osd connection drops, we need to resubmit all requests. 8463d14c5d2SYehuda Sadeh */ 8473d14c5d2SYehuda Sadeh static void osd_reset(struct ceph_connection *con) 8483d14c5d2SYehuda Sadeh { 8493d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 8503d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 8513d14c5d2SYehuda Sadeh 8523d14c5d2SYehuda Sadeh if (!osd) 8533d14c5d2SYehuda Sadeh return; 8543d14c5d2SYehuda Sadeh dout("osd_reset osd%d\n", osd->o_osd); 8553d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 8563d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 85783aff95eSSage Weil mutex_lock(&osdc->request_mutex); 85883aff95eSSage Weil __kick_osd_requests(osdc, osd); 859f9d25199SAlex Elder __send_queued(osdc); 86083aff95eSSage Weil mutex_unlock(&osdc->request_mutex); 8613d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 8623d14c5d2SYehuda Sadeh } 8633d14c5d2SYehuda Sadeh 8643d14c5d2SYehuda Sadeh /* 8653d14c5d2SYehuda Sadeh * Track open sessions with osds. 8663d14c5d2SYehuda Sadeh */ 867e10006f8SAlex Elder static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum) 8683d14c5d2SYehuda Sadeh { 8693d14c5d2SYehuda Sadeh struct ceph_osd *osd; 8703d14c5d2SYehuda Sadeh 8713d14c5d2SYehuda Sadeh osd = kzalloc(sizeof(*osd), GFP_NOFS); 8723d14c5d2SYehuda Sadeh if (!osd) 8733d14c5d2SYehuda Sadeh return NULL; 8743d14c5d2SYehuda Sadeh 8753d14c5d2SYehuda Sadeh atomic_set(&osd->o_ref, 1); 8763d14c5d2SYehuda Sadeh osd->o_osdc = osdc; 877e10006f8SAlex Elder osd->o_osd = onum; 878f407731dSAlex Elder RB_CLEAR_NODE(&osd->o_node); 8793d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_requests); 880a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osd->o_linger_requests); 8813d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_osd_lru); 8823d14c5d2SYehuda Sadeh osd->o_incarnation = 1; 8833d14c5d2SYehuda Sadeh 884b7a9e5ddSSage Weil ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr); 8853d14c5d2SYehuda Sadeh 8863d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_keepalive_item); 8873d14c5d2SYehuda Sadeh return osd; 8883d14c5d2SYehuda Sadeh } 8893d14c5d2SYehuda Sadeh 8903d14c5d2SYehuda Sadeh static struct ceph_osd *get_osd(struct ceph_osd *osd) 8913d14c5d2SYehuda Sadeh { 8923d14c5d2SYehuda Sadeh if (atomic_inc_not_zero(&osd->o_ref)) { 8933d14c5d2SYehuda Sadeh dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1, 8943d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref)); 8953d14c5d2SYehuda Sadeh return osd; 8963d14c5d2SYehuda Sadeh } else { 8973d14c5d2SYehuda Sadeh dout("get_osd %p FAIL\n", osd); 8983d14c5d2SYehuda Sadeh return NULL; 8993d14c5d2SYehuda Sadeh } 9003d14c5d2SYehuda Sadeh } 9013d14c5d2SYehuda Sadeh 9023d14c5d2SYehuda Sadeh static void put_osd(struct ceph_osd *osd) 9033d14c5d2SYehuda Sadeh { 9043d14c5d2SYehuda Sadeh dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref), 9053d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref) - 1); 906a255651dSAlex Elder if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) { 9073d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth; 9083d14c5d2SYehuda Sadeh 90927859f97SSage Weil ceph_auth_destroy_authorizer(ac, osd->o_auth.authorizer); 9103d14c5d2SYehuda Sadeh kfree(osd); 9113d14c5d2SYehuda Sadeh } 9123d14c5d2SYehuda Sadeh } 9133d14c5d2SYehuda Sadeh 9143d14c5d2SYehuda Sadeh /* 9153d14c5d2SYehuda Sadeh * remove an osd from our map 9163d14c5d2SYehuda Sadeh */ 9173d14c5d2SYehuda Sadeh static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 9183d14c5d2SYehuda Sadeh { 9193d14c5d2SYehuda Sadeh dout("__remove_osd %p\n", osd); 9203d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_requests)); 9213d14c5d2SYehuda Sadeh rb_erase(&osd->o_node, &osdc->osds); 9223d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 9233d14c5d2SYehuda Sadeh ceph_con_close(&osd->o_con); 9243d14c5d2SYehuda Sadeh put_osd(osd); 9253d14c5d2SYehuda Sadeh } 9263d14c5d2SYehuda Sadeh 927aca420bcSSage Weil static void remove_all_osds(struct ceph_osd_client *osdc) 928aca420bcSSage Weil { 929048a9d2dSJiaju Zhang dout("%s %p\n", __func__, osdc); 930aca420bcSSage Weil mutex_lock(&osdc->request_mutex); 931aca420bcSSage Weil while (!RB_EMPTY_ROOT(&osdc->osds)) { 932aca420bcSSage Weil struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds), 933aca420bcSSage Weil struct ceph_osd, o_node); 934aca420bcSSage Weil __remove_osd(osdc, osd); 935aca420bcSSage Weil } 936aca420bcSSage Weil mutex_unlock(&osdc->request_mutex); 937aca420bcSSage Weil } 938aca420bcSSage Weil 9393d14c5d2SYehuda Sadeh static void __move_osd_to_lru(struct ceph_osd_client *osdc, 9403d14c5d2SYehuda Sadeh struct ceph_osd *osd) 9413d14c5d2SYehuda Sadeh { 9423d14c5d2SYehuda Sadeh dout("__move_osd_to_lru %p\n", osd); 9433d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_osd_lru)); 9443d14c5d2SYehuda Sadeh list_add_tail(&osd->o_osd_lru, &osdc->osd_lru); 9453d14c5d2SYehuda Sadeh osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ; 9463d14c5d2SYehuda Sadeh } 9473d14c5d2SYehuda Sadeh 9483d14c5d2SYehuda Sadeh static void __remove_osd_from_lru(struct ceph_osd *osd) 9493d14c5d2SYehuda Sadeh { 9503d14c5d2SYehuda Sadeh dout("__remove_osd_from_lru %p\n", osd); 9513d14c5d2SYehuda Sadeh if (!list_empty(&osd->o_osd_lru)) 9523d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 9533d14c5d2SYehuda Sadeh } 9543d14c5d2SYehuda Sadeh 955aca420bcSSage Weil static void remove_old_osds(struct ceph_osd_client *osdc) 9563d14c5d2SYehuda Sadeh { 9573d14c5d2SYehuda Sadeh struct ceph_osd *osd, *nosd; 9583d14c5d2SYehuda Sadeh 9593d14c5d2SYehuda Sadeh dout("__remove_old_osds %p\n", osdc); 9603d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 9613d14c5d2SYehuda Sadeh list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) { 962aca420bcSSage Weil if (time_before(jiffies, osd->lru_ttl)) 9633d14c5d2SYehuda Sadeh break; 9643d14c5d2SYehuda Sadeh __remove_osd(osdc, osd); 9653d14c5d2SYehuda Sadeh } 9663d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 9673d14c5d2SYehuda Sadeh } 9683d14c5d2SYehuda Sadeh 9693d14c5d2SYehuda Sadeh /* 9703d14c5d2SYehuda Sadeh * reset osd connect 9713d14c5d2SYehuda Sadeh */ 9723d14c5d2SYehuda Sadeh static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 9733d14c5d2SYehuda Sadeh { 974c3acb181SAlex Elder struct ceph_entity_addr *peer_addr; 9753d14c5d2SYehuda Sadeh 9763d14c5d2SYehuda Sadeh dout("__reset_osd %p osd%d\n", osd, osd->o_osd); 977a40c4f10SYehuda Sadeh if (list_empty(&osd->o_requests) && 978a40c4f10SYehuda Sadeh list_empty(&osd->o_linger_requests)) { 9793d14c5d2SYehuda Sadeh __remove_osd(osdc, osd); 980c3acb181SAlex Elder 981c3acb181SAlex Elder return -ENODEV; 982c3acb181SAlex Elder } 983c3acb181SAlex Elder 984c3acb181SAlex Elder peer_addr = &osdc->osdmap->osd_addr[osd->o_osd]; 985c3acb181SAlex Elder if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) && 9863d14c5d2SYehuda Sadeh !ceph_con_opened(&osd->o_con)) { 987c3acb181SAlex Elder struct ceph_osd_request *req; 988c3acb181SAlex Elder 9893d14c5d2SYehuda Sadeh dout(" osd addr hasn't changed and connection never opened," 9903d14c5d2SYehuda Sadeh " letting msgr retry"); 9913d14c5d2SYehuda Sadeh /* touch each r_stamp for handle_timeout()'s benfit */ 9923d14c5d2SYehuda Sadeh list_for_each_entry(req, &osd->o_requests, r_osd_item) 9933d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 994c3acb181SAlex Elder 995c3acb181SAlex Elder return -EAGAIN; 9963d14c5d2SYehuda Sadeh } 997c3acb181SAlex Elder 998c3acb181SAlex Elder ceph_con_close(&osd->o_con); 999c3acb181SAlex Elder ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr); 1000c3acb181SAlex Elder osd->o_incarnation++; 1001c3acb181SAlex Elder 1002c3acb181SAlex Elder return 0; 10033d14c5d2SYehuda Sadeh } 10043d14c5d2SYehuda Sadeh 10053d14c5d2SYehuda Sadeh static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new) 10063d14c5d2SYehuda Sadeh { 10073d14c5d2SYehuda Sadeh struct rb_node **p = &osdc->osds.rb_node; 10083d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 10093d14c5d2SYehuda Sadeh struct ceph_osd *osd = NULL; 10103d14c5d2SYehuda Sadeh 1011aca420bcSSage Weil dout("__insert_osd %p osd%d\n", new, new->o_osd); 10123d14c5d2SYehuda Sadeh while (*p) { 10133d14c5d2SYehuda Sadeh parent = *p; 10143d14c5d2SYehuda Sadeh osd = rb_entry(parent, struct ceph_osd, o_node); 10153d14c5d2SYehuda Sadeh if (new->o_osd < osd->o_osd) 10163d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 10173d14c5d2SYehuda Sadeh else if (new->o_osd > osd->o_osd) 10183d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 10193d14c5d2SYehuda Sadeh else 10203d14c5d2SYehuda Sadeh BUG(); 10213d14c5d2SYehuda Sadeh } 10223d14c5d2SYehuda Sadeh 10233d14c5d2SYehuda Sadeh rb_link_node(&new->o_node, parent, p); 10243d14c5d2SYehuda Sadeh rb_insert_color(&new->o_node, &osdc->osds); 10253d14c5d2SYehuda Sadeh } 10263d14c5d2SYehuda Sadeh 10273d14c5d2SYehuda Sadeh static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o) 10283d14c5d2SYehuda Sadeh { 10293d14c5d2SYehuda Sadeh struct ceph_osd *osd; 10303d14c5d2SYehuda Sadeh struct rb_node *n = osdc->osds.rb_node; 10313d14c5d2SYehuda Sadeh 10323d14c5d2SYehuda Sadeh while (n) { 10333d14c5d2SYehuda Sadeh osd = rb_entry(n, struct ceph_osd, o_node); 10343d14c5d2SYehuda Sadeh if (o < osd->o_osd) 10353d14c5d2SYehuda Sadeh n = n->rb_left; 10363d14c5d2SYehuda Sadeh else if (o > osd->o_osd) 10373d14c5d2SYehuda Sadeh n = n->rb_right; 10383d14c5d2SYehuda Sadeh else 10393d14c5d2SYehuda Sadeh return osd; 10403d14c5d2SYehuda Sadeh } 10413d14c5d2SYehuda Sadeh return NULL; 10423d14c5d2SYehuda Sadeh } 10433d14c5d2SYehuda Sadeh 10443d14c5d2SYehuda Sadeh static void __schedule_osd_timeout(struct ceph_osd_client *osdc) 10453d14c5d2SYehuda Sadeh { 10463d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->timeout_work, 10473d14c5d2SYehuda Sadeh osdc->client->options->osd_keepalive_timeout * HZ); 10483d14c5d2SYehuda Sadeh } 10493d14c5d2SYehuda Sadeh 10503d14c5d2SYehuda Sadeh static void __cancel_osd_timeout(struct ceph_osd_client *osdc) 10513d14c5d2SYehuda Sadeh { 10523d14c5d2SYehuda Sadeh cancel_delayed_work(&osdc->timeout_work); 10533d14c5d2SYehuda Sadeh } 10543d14c5d2SYehuda Sadeh 10553d14c5d2SYehuda Sadeh /* 10563d14c5d2SYehuda Sadeh * Register request, assign tid. If this is the first request, set up 10573d14c5d2SYehuda Sadeh * the timeout event. 10583d14c5d2SYehuda Sadeh */ 1059a40c4f10SYehuda Sadeh static void __register_request(struct ceph_osd_client *osdc, 10603d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 10613d14c5d2SYehuda Sadeh { 10623d14c5d2SYehuda Sadeh req->r_tid = ++osdc->last_tid; 10633d14c5d2SYehuda Sadeh req->r_request->hdr.tid = cpu_to_le64(req->r_tid); 106477f38e0eSSage Weil dout("__register_request %p tid %lld\n", req, req->r_tid); 10653d14c5d2SYehuda Sadeh __insert_request(osdc, req); 10663d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 10673d14c5d2SYehuda Sadeh osdc->num_requests++; 10683d14c5d2SYehuda Sadeh if (osdc->num_requests == 1) { 10693d14c5d2SYehuda Sadeh dout(" first request, scheduling timeout\n"); 10703d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 10713d14c5d2SYehuda Sadeh } 1072a40c4f10SYehuda Sadeh } 1073a40c4f10SYehuda Sadeh 10743d14c5d2SYehuda Sadeh /* 10753d14c5d2SYehuda Sadeh * called under osdc->request_mutex 10763d14c5d2SYehuda Sadeh */ 10773d14c5d2SYehuda Sadeh static void __unregister_request(struct ceph_osd_client *osdc, 10783d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 10793d14c5d2SYehuda Sadeh { 108035f9f8a0SSage Weil if (RB_EMPTY_NODE(&req->r_node)) { 108135f9f8a0SSage Weil dout("__unregister_request %p tid %lld not registered\n", 108235f9f8a0SSage Weil req, req->r_tid); 108335f9f8a0SSage Weil return; 108435f9f8a0SSage Weil } 108535f9f8a0SSage Weil 10863d14c5d2SYehuda Sadeh dout("__unregister_request %p tid %lld\n", req, req->r_tid); 10873d14c5d2SYehuda Sadeh rb_erase(&req->r_node, &osdc->requests); 10883d14c5d2SYehuda Sadeh osdc->num_requests--; 10893d14c5d2SYehuda Sadeh 10903d14c5d2SYehuda Sadeh if (req->r_osd) { 10913d14c5d2SYehuda Sadeh /* make sure the original request isn't in flight. */ 10926740a845SAlex Elder ceph_msg_revoke(req->r_request); 10933d14c5d2SYehuda Sadeh 10943d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 1095a40c4f10SYehuda Sadeh if (list_empty(&req->r_osd->o_requests) && 1096a40c4f10SYehuda Sadeh list_empty(&req->r_osd->o_linger_requests)) { 1097a40c4f10SYehuda Sadeh dout("moving osd to %p lru\n", req->r_osd); 10983d14c5d2SYehuda Sadeh __move_osd_to_lru(osdc, req->r_osd); 1099a40c4f10SYehuda Sadeh } 1100fbdb9190SSage Weil if (list_empty(&req->r_linger_item)) 11013d14c5d2SYehuda Sadeh req->r_osd = NULL; 11023d14c5d2SYehuda Sadeh } 11033d14c5d2SYehuda Sadeh 11047d5f2481SAlex Elder list_del_init(&req->r_req_lru_item); 11053d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 11063d14c5d2SYehuda Sadeh 11073d14c5d2SYehuda Sadeh if (osdc->num_requests == 0) { 11083d14c5d2SYehuda Sadeh dout(" no requests, canceling timeout\n"); 11093d14c5d2SYehuda Sadeh __cancel_osd_timeout(osdc); 11103d14c5d2SYehuda Sadeh } 11113d14c5d2SYehuda Sadeh } 11123d14c5d2SYehuda Sadeh 11133d14c5d2SYehuda Sadeh /* 11143d14c5d2SYehuda Sadeh * Cancel a previously queued request message 11153d14c5d2SYehuda Sadeh */ 11163d14c5d2SYehuda Sadeh static void __cancel_request(struct ceph_osd_request *req) 11173d14c5d2SYehuda Sadeh { 11183d14c5d2SYehuda Sadeh if (req->r_sent && req->r_osd) { 11196740a845SAlex Elder ceph_msg_revoke(req->r_request); 11203d14c5d2SYehuda Sadeh req->r_sent = 0; 11213d14c5d2SYehuda Sadeh } 11223d14c5d2SYehuda Sadeh } 11233d14c5d2SYehuda Sadeh 1124a40c4f10SYehuda Sadeh static void __register_linger_request(struct ceph_osd_client *osdc, 1125a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1126a40c4f10SYehuda Sadeh { 1127a40c4f10SYehuda Sadeh dout("__register_linger_request %p\n", req); 1128a40c4f10SYehuda Sadeh list_add_tail(&req->r_linger_item, &osdc->req_linger); 11296194ea89SSage Weil if (req->r_osd) 11306194ea89SSage Weil list_add_tail(&req->r_linger_osd, 11316194ea89SSage Weil &req->r_osd->o_linger_requests); 1132a40c4f10SYehuda Sadeh } 1133a40c4f10SYehuda Sadeh 1134a40c4f10SYehuda Sadeh static void __unregister_linger_request(struct ceph_osd_client *osdc, 1135a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1136a40c4f10SYehuda Sadeh { 1137a40c4f10SYehuda Sadeh dout("__unregister_linger_request %p\n", req); 1138a40c4f10SYehuda Sadeh list_del_init(&req->r_linger_item); 113961c74035SAlex Elder if (req->r_osd) { 1140a40c4f10SYehuda Sadeh list_del_init(&req->r_linger_osd); 1141a40c4f10SYehuda Sadeh 1142a40c4f10SYehuda Sadeh if (list_empty(&req->r_osd->o_requests) && 1143a40c4f10SYehuda Sadeh list_empty(&req->r_osd->o_linger_requests)) { 1144a40c4f10SYehuda Sadeh dout("moving osd to %p lru\n", req->r_osd); 1145a40c4f10SYehuda Sadeh __move_osd_to_lru(osdc, req->r_osd); 1146a40c4f10SYehuda Sadeh } 1147fbdb9190SSage Weil if (list_empty(&req->r_osd_item)) 1148a40c4f10SYehuda Sadeh req->r_osd = NULL; 1149a40c4f10SYehuda Sadeh } 1150a40c4f10SYehuda Sadeh } 1151a40c4f10SYehuda Sadeh 1152a40c4f10SYehuda Sadeh void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc, 1153a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1154a40c4f10SYehuda Sadeh { 1155a40c4f10SYehuda Sadeh mutex_lock(&osdc->request_mutex); 1156a40c4f10SYehuda Sadeh if (req->r_linger) { 1157a40c4f10SYehuda Sadeh __unregister_linger_request(osdc, req); 1158a40c4f10SYehuda Sadeh ceph_osdc_put_request(req); 1159a40c4f10SYehuda Sadeh } 1160a40c4f10SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 1161a40c4f10SYehuda Sadeh } 1162a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_unregister_linger_request); 1163a40c4f10SYehuda Sadeh 1164a40c4f10SYehuda Sadeh void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc, 1165a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1166a40c4f10SYehuda Sadeh { 1167a40c4f10SYehuda Sadeh if (!req->r_linger) { 1168a40c4f10SYehuda Sadeh dout("set_request_linger %p\n", req); 1169a40c4f10SYehuda Sadeh req->r_linger = 1; 1170a40c4f10SYehuda Sadeh /* 1171a40c4f10SYehuda Sadeh * caller is now responsible for calling 1172a40c4f10SYehuda Sadeh * unregister_linger_request 1173a40c4f10SYehuda Sadeh */ 1174a40c4f10SYehuda Sadeh ceph_osdc_get_request(req); 1175a40c4f10SYehuda Sadeh } 1176a40c4f10SYehuda Sadeh } 1177a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_set_request_linger); 1178a40c4f10SYehuda Sadeh 11793d14c5d2SYehuda Sadeh /* 11803d14c5d2SYehuda Sadeh * Pick an osd (the first 'up' osd in the pg), allocate the osd struct 11813d14c5d2SYehuda Sadeh * (as needed), and set the request r_osd appropriately. If there is 118225985edcSLucas De Marchi * no up osd, set r_osd to NULL. Move the request to the appropriate list 11836f6c7006SSage Weil * (unsent, homeless) or leave on in-flight lru. 11843d14c5d2SYehuda Sadeh * 11853d14c5d2SYehuda Sadeh * Return 0 if unchanged, 1 if changed, or negative on error. 11863d14c5d2SYehuda Sadeh * 11873d14c5d2SYehuda Sadeh * Caller should hold map_sem for read and request_mutex. 11883d14c5d2SYehuda Sadeh */ 11896f6c7006SSage Weil static int __map_request(struct ceph_osd_client *osdc, 119038d6453cSSage Weil struct ceph_osd_request *req, int force_resend) 11913d14c5d2SYehuda Sadeh { 11925b191d99SSage Weil struct ceph_pg pgid; 11933d14c5d2SYehuda Sadeh int acting[CEPH_PG_MAX_SIZE]; 11943d14c5d2SYehuda Sadeh int o = -1, num = 0; 11953d14c5d2SYehuda Sadeh int err; 11963d14c5d2SYehuda Sadeh 11976f6c7006SSage Weil dout("map_request %p tid %lld\n", req, req->r_tid); 119841766f87SAlex Elder err = ceph_calc_ceph_pg(&pgid, req->r_oid, osdc->osdmap, 119941766f87SAlex Elder ceph_file_layout_pg_pool(req->r_file_layout)); 12006f6c7006SSage Weil if (err) { 12016f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 12023d14c5d2SYehuda Sadeh return err; 12036f6c7006SSage Weil } 12043d14c5d2SYehuda Sadeh req->r_pgid = pgid; 12053d14c5d2SYehuda Sadeh 12063d14c5d2SYehuda Sadeh err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting); 12073d14c5d2SYehuda Sadeh if (err > 0) { 12083d14c5d2SYehuda Sadeh o = acting[0]; 12093d14c5d2SYehuda Sadeh num = err; 12103d14c5d2SYehuda Sadeh } 12113d14c5d2SYehuda Sadeh 121238d6453cSSage Weil if ((!force_resend && 121338d6453cSSage Weil req->r_osd && req->r_osd->o_osd == o && 12143d14c5d2SYehuda Sadeh req->r_sent >= req->r_osd->o_incarnation && 12153d14c5d2SYehuda Sadeh req->r_num_pg_osds == num && 12163d14c5d2SYehuda Sadeh memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) || 12173d14c5d2SYehuda Sadeh (req->r_osd == NULL && o == -1)) 12183d14c5d2SYehuda Sadeh return 0; /* no change */ 12193d14c5d2SYehuda Sadeh 12205b191d99SSage Weil dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n", 12215b191d99SSage Weil req->r_tid, pgid.pool, pgid.seed, o, 12223d14c5d2SYehuda Sadeh req->r_osd ? req->r_osd->o_osd : -1); 12233d14c5d2SYehuda Sadeh 12243d14c5d2SYehuda Sadeh /* record full pg acting set */ 12253d14c5d2SYehuda Sadeh memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num); 12263d14c5d2SYehuda Sadeh req->r_num_pg_osds = num; 12273d14c5d2SYehuda Sadeh 12283d14c5d2SYehuda Sadeh if (req->r_osd) { 12293d14c5d2SYehuda Sadeh __cancel_request(req); 12303d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 12313d14c5d2SYehuda Sadeh req->r_osd = NULL; 12323d14c5d2SYehuda Sadeh } 12333d14c5d2SYehuda Sadeh 12343d14c5d2SYehuda Sadeh req->r_osd = __lookup_osd(osdc, o); 12353d14c5d2SYehuda Sadeh if (!req->r_osd && o >= 0) { 12363d14c5d2SYehuda Sadeh err = -ENOMEM; 1237e10006f8SAlex Elder req->r_osd = create_osd(osdc, o); 12386f6c7006SSage Weil if (!req->r_osd) { 12396f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 12403d14c5d2SYehuda Sadeh goto out; 12416f6c7006SSage Weil } 12423d14c5d2SYehuda Sadeh 12436f6c7006SSage Weil dout("map_request osd %p is osd%d\n", req->r_osd, o); 12443d14c5d2SYehuda Sadeh __insert_osd(osdc, req->r_osd); 12453d14c5d2SYehuda Sadeh 1246b7a9e5ddSSage Weil ceph_con_open(&req->r_osd->o_con, 1247b7a9e5ddSSage Weil CEPH_ENTITY_TYPE_OSD, o, 1248b7a9e5ddSSage Weil &osdc->osdmap->osd_addr[o]); 12493d14c5d2SYehuda Sadeh } 12503d14c5d2SYehuda Sadeh 12513d14c5d2SYehuda Sadeh if (req->r_osd) { 12523d14c5d2SYehuda Sadeh __remove_osd_from_lru(req->r_osd); 1253ad885927SAlex Elder list_add_tail(&req->r_osd_item, &req->r_osd->o_requests); 1254ad885927SAlex Elder list_move_tail(&req->r_req_lru_item, &osdc->req_unsent); 12556f6c7006SSage Weil } else { 1256ad885927SAlex Elder list_move_tail(&req->r_req_lru_item, &osdc->req_notarget); 12573d14c5d2SYehuda Sadeh } 12583d14c5d2SYehuda Sadeh err = 1; /* osd or pg changed */ 12593d14c5d2SYehuda Sadeh 12603d14c5d2SYehuda Sadeh out: 12613d14c5d2SYehuda Sadeh return err; 12623d14c5d2SYehuda Sadeh } 12633d14c5d2SYehuda Sadeh 12643d14c5d2SYehuda Sadeh /* 12653d14c5d2SYehuda Sadeh * caller should hold map_sem (for read) and request_mutex 12663d14c5d2SYehuda Sadeh */ 126756e925b6SSage Weil static void __send_request(struct ceph_osd_client *osdc, 12683d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 12693d14c5d2SYehuda Sadeh { 12701b83bef2SSage Weil void *p; 12713d14c5d2SYehuda Sadeh 12721b83bef2SSage Weil dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n", 12731b83bef2SSage Weil req, req->r_tid, req->r_osd->o_osd, req->r_flags, 12741b83bef2SSage Weil (unsigned long long)req->r_pgid.pool, req->r_pgid.seed); 12753d14c5d2SYehuda Sadeh 12761b83bef2SSage Weil /* fill in message content that changes each time we send it */ 12771b83bef2SSage Weil put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch); 12781b83bef2SSage Weil put_unaligned_le32(req->r_flags, req->r_request_flags); 12791b83bef2SSage Weil put_unaligned_le64(req->r_pgid.pool, req->r_request_pool); 12801b83bef2SSage Weil p = req->r_request_pgid; 12811b83bef2SSage Weil ceph_encode_64(&p, req->r_pgid.pool); 12821b83bef2SSage Weil ceph_encode_32(&p, req->r_pgid.seed); 12831b83bef2SSage Weil put_unaligned_le64(1, req->r_request_attempts); /* FIXME */ 12841b83bef2SSage Weil memcpy(req->r_request_reassert_version, &req->r_reassert_version, 12851b83bef2SSage Weil sizeof(req->r_reassert_version)); 12862169aea6SSage Weil 12873d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 12883d14c5d2SYehuda Sadeh list_move_tail(&req->r_req_lru_item, &osdc->req_lru); 12893d14c5d2SYehuda Sadeh 12903d14c5d2SYehuda Sadeh ceph_msg_get(req->r_request); /* send consumes a ref */ 12913d14c5d2SYehuda Sadeh ceph_con_send(&req->r_osd->o_con, req->r_request); 12923d14c5d2SYehuda Sadeh req->r_sent = req->r_osd->o_incarnation; 12933d14c5d2SYehuda Sadeh } 12943d14c5d2SYehuda Sadeh 12953d14c5d2SYehuda Sadeh /* 12966f6c7006SSage Weil * Send any requests in the queue (req_unsent). 12976f6c7006SSage Weil */ 1298f9d25199SAlex Elder static void __send_queued(struct ceph_osd_client *osdc) 12996f6c7006SSage Weil { 13006f6c7006SSage Weil struct ceph_osd_request *req, *tmp; 13016f6c7006SSage Weil 1302f9d25199SAlex Elder dout("__send_queued\n"); 1303f9d25199SAlex Elder list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) 13046f6c7006SSage Weil __send_request(osdc, req); 13056f6c7006SSage Weil } 13066f6c7006SSage Weil 13076f6c7006SSage Weil /* 13083d14c5d2SYehuda Sadeh * Timeout callback, called every N seconds when 1 or more osd 13093d14c5d2SYehuda Sadeh * requests has been active for more than N seconds. When this 13103d14c5d2SYehuda Sadeh * happens, we ping all OSDs with requests who have timed out to 13113d14c5d2SYehuda Sadeh * ensure any communications channel reset is detected. Reset the 13123d14c5d2SYehuda Sadeh * request timeouts another N seconds in the future as we go. 13133d14c5d2SYehuda Sadeh * Reschedule the timeout event another N seconds in future (unless 13143d14c5d2SYehuda Sadeh * there are no open requests). 13153d14c5d2SYehuda Sadeh */ 13163d14c5d2SYehuda Sadeh static void handle_timeout(struct work_struct *work) 13173d14c5d2SYehuda Sadeh { 13183d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 13193d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, timeout_work.work); 132083aff95eSSage Weil struct ceph_osd_request *req; 13213d14c5d2SYehuda Sadeh struct ceph_osd *osd; 13223d14c5d2SYehuda Sadeh unsigned long keepalive = 13233d14c5d2SYehuda Sadeh osdc->client->options->osd_keepalive_timeout * HZ; 13243d14c5d2SYehuda Sadeh struct list_head slow_osds; 13253d14c5d2SYehuda Sadeh dout("timeout\n"); 13263d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 13273d14c5d2SYehuda Sadeh 13283d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 13293d14c5d2SYehuda Sadeh 13303d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 13313d14c5d2SYehuda Sadeh 13323d14c5d2SYehuda Sadeh /* 13333d14c5d2SYehuda Sadeh * ping osds that are a bit slow. this ensures that if there 13343d14c5d2SYehuda Sadeh * is a break in the TCP connection we will notice, and reopen 13353d14c5d2SYehuda Sadeh * a connection with that osd (from the fault callback). 13363d14c5d2SYehuda Sadeh */ 13373d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&slow_osds); 13383d14c5d2SYehuda Sadeh list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) { 13393d14c5d2SYehuda Sadeh if (time_before(jiffies, req->r_stamp + keepalive)) 13403d14c5d2SYehuda Sadeh break; 13413d14c5d2SYehuda Sadeh 13423d14c5d2SYehuda Sadeh osd = req->r_osd; 13433d14c5d2SYehuda Sadeh BUG_ON(!osd); 13443d14c5d2SYehuda Sadeh dout(" tid %llu is slow, will send keepalive on osd%d\n", 13453d14c5d2SYehuda Sadeh req->r_tid, osd->o_osd); 13463d14c5d2SYehuda Sadeh list_move_tail(&osd->o_keepalive_item, &slow_osds); 13473d14c5d2SYehuda Sadeh } 13483d14c5d2SYehuda Sadeh while (!list_empty(&slow_osds)) { 13493d14c5d2SYehuda Sadeh osd = list_entry(slow_osds.next, struct ceph_osd, 13503d14c5d2SYehuda Sadeh o_keepalive_item); 13513d14c5d2SYehuda Sadeh list_del_init(&osd->o_keepalive_item); 13523d14c5d2SYehuda Sadeh ceph_con_keepalive(&osd->o_con); 13533d14c5d2SYehuda Sadeh } 13543d14c5d2SYehuda Sadeh 13553d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 1356f9d25199SAlex Elder __send_queued(osdc); 13573d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 13583d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 13593d14c5d2SYehuda Sadeh } 13603d14c5d2SYehuda Sadeh 13613d14c5d2SYehuda Sadeh static void handle_osds_timeout(struct work_struct *work) 13623d14c5d2SYehuda Sadeh { 13633d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 13643d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, 13653d14c5d2SYehuda Sadeh osds_timeout_work.work); 13663d14c5d2SYehuda Sadeh unsigned long delay = 13673d14c5d2SYehuda Sadeh osdc->client->options->osd_idle_ttl * HZ >> 2; 13683d14c5d2SYehuda Sadeh 13693d14c5d2SYehuda Sadeh dout("osds timeout\n"); 13703d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 1371aca420bcSSage Weil remove_old_osds(osdc); 13723d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 13733d14c5d2SYehuda Sadeh 13743d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 13753d14c5d2SYehuda Sadeh round_jiffies_relative(delay)); 13763d14c5d2SYehuda Sadeh } 13773d14c5d2SYehuda Sadeh 137825845472SSage Weil static void complete_request(struct ceph_osd_request *req) 137925845472SSage Weil { 138025845472SSage Weil if (req->r_safe_callback) 138125845472SSage Weil req->r_safe_callback(req, NULL); 138225845472SSage Weil complete_all(&req->r_safe_completion); /* fsync waiter */ 138325845472SSage Weil } 138425845472SSage Weil 13853d14c5d2SYehuda Sadeh /* 13863d14c5d2SYehuda Sadeh * handle osd op reply. either call the callback if it is specified, 13873d14c5d2SYehuda Sadeh * or do the completion to wake up the waiting thread. 13883d14c5d2SYehuda Sadeh */ 13893d14c5d2SYehuda Sadeh static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg, 13903d14c5d2SYehuda Sadeh struct ceph_connection *con) 13913d14c5d2SYehuda Sadeh { 13921b83bef2SSage Weil void *p, *end; 13933d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 13943d14c5d2SYehuda Sadeh u64 tid; 13951b83bef2SSage Weil int object_len; 139679528734SAlex Elder unsigned int numops; 139779528734SAlex Elder int payload_len, flags; 13983d14c5d2SYehuda Sadeh s32 result; 13991b83bef2SSage Weil s32 retry_attempt; 14001b83bef2SSage Weil struct ceph_pg pg; 14011b83bef2SSage Weil int err; 14021b83bef2SSage Weil u32 reassert_epoch; 14031b83bef2SSage Weil u64 reassert_version; 14041b83bef2SSage Weil u32 osdmap_epoch; 14050d5af164SAlex Elder int already_completed; 14069fc6e064SAlex Elder u32 bytes; 140779528734SAlex Elder unsigned int i; 14083d14c5d2SYehuda Sadeh 14093d14c5d2SYehuda Sadeh tid = le64_to_cpu(msg->hdr.tid); 14101b83bef2SSage Weil dout("handle_reply %p tid %llu\n", msg, tid); 14111b83bef2SSage Weil 14121b83bef2SSage Weil p = msg->front.iov_base; 14131b83bef2SSage Weil end = p + msg->front.iov_len; 14141b83bef2SSage Weil 14151b83bef2SSage Weil ceph_decode_need(&p, end, 4, bad); 14161b83bef2SSage Weil object_len = ceph_decode_32(&p); 14171b83bef2SSage Weil ceph_decode_need(&p, end, object_len, bad); 14181b83bef2SSage Weil p += object_len; 14191b83bef2SSage Weil 1420ef4859d6SAlex Elder err = ceph_decode_pgid(&p, end, &pg); 14211b83bef2SSage Weil if (err) 14223d14c5d2SYehuda Sadeh goto bad; 14231b83bef2SSage Weil 14241b83bef2SSage Weil ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad); 14251b83bef2SSage Weil flags = ceph_decode_64(&p); 14261b83bef2SSage Weil result = ceph_decode_32(&p); 14271b83bef2SSage Weil reassert_epoch = ceph_decode_32(&p); 14281b83bef2SSage Weil reassert_version = ceph_decode_64(&p); 14291b83bef2SSage Weil osdmap_epoch = ceph_decode_32(&p); 14301b83bef2SSage Weil 14313d14c5d2SYehuda Sadeh /* lookup */ 14323d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 14333d14c5d2SYehuda Sadeh req = __lookup_request(osdc, tid); 14343d14c5d2SYehuda Sadeh if (req == NULL) { 14353d14c5d2SYehuda Sadeh dout("handle_reply tid %llu dne\n", tid); 14368058fd45SAlex Elder goto bad_mutex; 14373d14c5d2SYehuda Sadeh } 14383d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 14391b83bef2SSage Weil 14401b83bef2SSage Weil dout("handle_reply %p tid %llu req %p result %d\n", msg, tid, 14411b83bef2SSage Weil req, result); 14421b83bef2SSage Weil 14431b83bef2SSage Weil ceph_decode_need(&p, end, 4, bad); 14441b83bef2SSage Weil numops = ceph_decode_32(&p); 14451b83bef2SSage Weil if (numops > CEPH_OSD_MAX_OP) 14461b83bef2SSage Weil goto bad_put; 14471b83bef2SSage Weil if (numops != req->r_num_ops) 14481b83bef2SSage Weil goto bad_put; 14491b83bef2SSage Weil payload_len = 0; 14501b83bef2SSage Weil ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad); 14511b83bef2SSage Weil for (i = 0; i < numops; i++) { 14521b83bef2SSage Weil struct ceph_osd_op *op = p; 14531b83bef2SSage Weil int len; 14541b83bef2SSage Weil 14551b83bef2SSage Weil len = le32_to_cpu(op->payload_len); 14561b83bef2SSage Weil req->r_reply_op_len[i] = len; 14571b83bef2SSage Weil dout(" op %d has %d bytes\n", i, len); 14581b83bef2SSage Weil payload_len += len; 14591b83bef2SSage Weil p += sizeof(*op); 14601b83bef2SSage Weil } 14619fc6e064SAlex Elder bytes = le32_to_cpu(msg->hdr.data_len); 14629fc6e064SAlex Elder if (payload_len != bytes) { 14631b83bef2SSage Weil pr_warning("sum of op payload lens %d != data_len %d", 14649fc6e064SAlex Elder payload_len, bytes); 14651b83bef2SSage Weil goto bad_put; 14661b83bef2SSage Weil } 14671b83bef2SSage Weil 14681b83bef2SSage Weil ceph_decode_need(&p, end, 4 + numops * 4, bad); 14691b83bef2SSage Weil retry_attempt = ceph_decode_32(&p); 14701b83bef2SSage Weil for (i = 0; i < numops; i++) 14711b83bef2SSage Weil req->r_reply_op_result[i] = ceph_decode_32(&p); 14723d14c5d2SYehuda Sadeh 14733d14c5d2SYehuda Sadeh if (!req->r_got_reply) { 14743d14c5d2SYehuda Sadeh 14751b83bef2SSage Weil req->r_result = result; 14763d14c5d2SYehuda Sadeh dout("handle_reply result %d bytes %d\n", req->r_result, 14773d14c5d2SYehuda Sadeh bytes); 14783d14c5d2SYehuda Sadeh if (req->r_result == 0) 14793d14c5d2SYehuda Sadeh req->r_result = bytes; 14803d14c5d2SYehuda Sadeh 14813d14c5d2SYehuda Sadeh /* in case this is a write and we need to replay, */ 14821b83bef2SSage Weil req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch); 14831b83bef2SSage Weil req->r_reassert_version.version = cpu_to_le64(reassert_version); 14843d14c5d2SYehuda Sadeh 14853d14c5d2SYehuda Sadeh req->r_got_reply = 1; 14863d14c5d2SYehuda Sadeh } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) { 14873d14c5d2SYehuda Sadeh dout("handle_reply tid %llu dup ack\n", tid); 14883d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 14893d14c5d2SYehuda Sadeh goto done; 14903d14c5d2SYehuda Sadeh } 14913d14c5d2SYehuda Sadeh 14923d14c5d2SYehuda Sadeh dout("handle_reply tid %llu flags %d\n", tid, flags); 14933d14c5d2SYehuda Sadeh 1494a40c4f10SYehuda Sadeh if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK)) 1495a40c4f10SYehuda Sadeh __register_linger_request(osdc, req); 1496a40c4f10SYehuda Sadeh 14973d14c5d2SYehuda Sadeh /* either this is a read, or we got the safe response */ 14983d14c5d2SYehuda Sadeh if (result < 0 || 14993d14c5d2SYehuda Sadeh (flags & CEPH_OSD_FLAG_ONDISK) || 15003d14c5d2SYehuda Sadeh ((flags & CEPH_OSD_FLAG_WRITE) == 0)) 15013d14c5d2SYehuda Sadeh __unregister_request(osdc, req); 15023d14c5d2SYehuda Sadeh 15030d5af164SAlex Elder already_completed = req->r_completed; 15040d5af164SAlex Elder req->r_completed = 1; 15053d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 15060d5af164SAlex Elder if (already_completed) 15070d5af164SAlex Elder goto done; 15083d14c5d2SYehuda Sadeh 15093d14c5d2SYehuda Sadeh if (req->r_callback) 15103d14c5d2SYehuda Sadeh req->r_callback(req, msg); 15113d14c5d2SYehuda Sadeh else 15123d14c5d2SYehuda Sadeh complete_all(&req->r_completion); 15133d14c5d2SYehuda Sadeh 151425845472SSage Weil if (flags & CEPH_OSD_FLAG_ONDISK) 151525845472SSage Weil complete_request(req); 15163d14c5d2SYehuda Sadeh 15173d14c5d2SYehuda Sadeh done: 1518a40c4f10SYehuda Sadeh dout("req=%p req->r_linger=%d\n", req, req->r_linger); 15193d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 15203d14c5d2SYehuda Sadeh return; 15213d14c5d2SYehuda Sadeh 15221b83bef2SSage Weil bad_put: 15231b83bef2SSage Weil ceph_osdc_put_request(req); 15248058fd45SAlex Elder bad_mutex: 15258058fd45SAlex Elder mutex_unlock(&osdc->request_mutex); 15263d14c5d2SYehuda Sadeh bad: 15271b83bef2SSage Weil pr_err("corrupt osd_op_reply got %d %d\n", 15281b83bef2SSage Weil (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len)); 15293d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 15303d14c5d2SYehuda Sadeh } 15313d14c5d2SYehuda Sadeh 15326f6c7006SSage Weil static void reset_changed_osds(struct ceph_osd_client *osdc) 15333d14c5d2SYehuda Sadeh { 15343d14c5d2SYehuda Sadeh struct rb_node *p, *n; 15353d14c5d2SYehuda Sadeh 15363d14c5d2SYehuda Sadeh for (p = rb_first(&osdc->osds); p; p = n) { 15376f6c7006SSage Weil struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node); 15383d14c5d2SYehuda Sadeh 15393d14c5d2SYehuda Sadeh n = rb_next(p); 15403d14c5d2SYehuda Sadeh if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) || 15413d14c5d2SYehuda Sadeh memcmp(&osd->o_con.peer_addr, 15423d14c5d2SYehuda Sadeh ceph_osd_addr(osdc->osdmap, 15433d14c5d2SYehuda Sadeh osd->o_osd), 15443d14c5d2SYehuda Sadeh sizeof(struct ceph_entity_addr)) != 0) 15453d14c5d2SYehuda Sadeh __reset_osd(osdc, osd); 15463d14c5d2SYehuda Sadeh } 15473d14c5d2SYehuda Sadeh } 15483d14c5d2SYehuda Sadeh 15493d14c5d2SYehuda Sadeh /* 15506f6c7006SSage Weil * Requeue requests whose mapping to an OSD has changed. If requests map to 15516f6c7006SSage Weil * no osd, request a new map. 15523d14c5d2SYehuda Sadeh * 1553e6d50f67SAlex Elder * Caller should hold map_sem for read. 15543d14c5d2SYehuda Sadeh */ 155538d6453cSSage Weil static void kick_requests(struct ceph_osd_client *osdc, int force_resend) 15563d14c5d2SYehuda Sadeh { 1557a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 15586f6c7006SSage Weil struct rb_node *p; 15596f6c7006SSage Weil int needmap = 0; 15606f6c7006SSage Weil int err; 15613d14c5d2SYehuda Sadeh 156238d6453cSSage Weil dout("kick_requests %s\n", force_resend ? " (force resend)" : ""); 15633d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 15646194ea89SSage Weil for (p = rb_first(&osdc->requests); p; ) { 15656f6c7006SSage Weil req = rb_entry(p, struct ceph_osd_request, r_node); 15666194ea89SSage Weil p = rb_next(p); 1567ab60b16dSAlex Elder 1568ab60b16dSAlex Elder /* 1569ab60b16dSAlex Elder * For linger requests that have not yet been 1570ab60b16dSAlex Elder * registered, move them to the linger list; they'll 1571ab60b16dSAlex Elder * be sent to the osd in the loop below. Unregister 1572ab60b16dSAlex Elder * the request before re-registering it as a linger 1573ab60b16dSAlex Elder * request to ensure the __map_request() below 1574ab60b16dSAlex Elder * will decide it needs to be sent. 1575ab60b16dSAlex Elder */ 1576ab60b16dSAlex Elder if (req->r_linger && list_empty(&req->r_linger_item)) { 1577ab60b16dSAlex Elder dout("%p tid %llu restart on osd%d\n", 1578ab60b16dSAlex Elder req, req->r_tid, 1579ab60b16dSAlex Elder req->r_osd ? req->r_osd->o_osd : -1); 1580ab60b16dSAlex Elder __unregister_request(osdc, req); 1581ab60b16dSAlex Elder __register_linger_request(osdc, req); 1582ab60b16dSAlex Elder continue; 1583ab60b16dSAlex Elder } 1584ab60b16dSAlex Elder 158538d6453cSSage Weil err = __map_request(osdc, req, force_resend); 15866f6c7006SSage Weil if (err < 0) 15876f6c7006SSage Weil continue; /* error */ 15886f6c7006SSage Weil if (req->r_osd == NULL) { 15896f6c7006SSage Weil dout("%p tid %llu maps to no osd\n", req, req->r_tid); 15906f6c7006SSage Weil needmap++; /* request a newer map */ 15916f6c7006SSage Weil } else if (err > 0) { 15926194ea89SSage Weil if (!req->r_linger) { 15936194ea89SSage Weil dout("%p tid %llu requeued on osd%d\n", req, 15946194ea89SSage Weil req->r_tid, 15956f6c7006SSage Weil req->r_osd ? req->r_osd->o_osd : -1); 15966f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 15976f6c7006SSage Weil } 15986f6c7006SSage Weil } 15996194ea89SSage Weil } 1600a40c4f10SYehuda Sadeh 1601a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osdc->req_linger, 1602a40c4f10SYehuda Sadeh r_linger_item) { 1603a40c4f10SYehuda Sadeh dout("linger req=%p req->r_osd=%p\n", req, req->r_osd); 1604a40c4f10SYehuda Sadeh 160538d6453cSSage Weil err = __map_request(osdc, req, force_resend); 1606ab60b16dSAlex Elder dout("__map_request returned %d\n", err); 1607a40c4f10SYehuda Sadeh if (err == 0) 1608a40c4f10SYehuda Sadeh continue; /* no change and no osd was specified */ 1609a40c4f10SYehuda Sadeh if (err < 0) 1610a40c4f10SYehuda Sadeh continue; /* hrm! */ 1611a40c4f10SYehuda Sadeh if (req->r_osd == NULL) { 1612a40c4f10SYehuda Sadeh dout("tid %llu maps to no valid osd\n", req->r_tid); 1613a40c4f10SYehuda Sadeh needmap++; /* request a newer map */ 1614a40c4f10SYehuda Sadeh continue; 1615a40c4f10SYehuda Sadeh } 1616a40c4f10SYehuda Sadeh 1617a40c4f10SYehuda Sadeh dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid, 1618a40c4f10SYehuda Sadeh req->r_osd ? req->r_osd->o_osd : -1); 1619a40c4f10SYehuda Sadeh __register_request(osdc, req); 1620c89ce05eSAlex Elder __unregister_linger_request(osdc, req); 1621a40c4f10SYehuda Sadeh } 16223d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 16233d14c5d2SYehuda Sadeh 16243d14c5d2SYehuda Sadeh if (needmap) { 16253d14c5d2SYehuda Sadeh dout("%d requests for down osds, need new map\n", needmap); 16263d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 16273d14c5d2SYehuda Sadeh } 1628e6d50f67SAlex Elder reset_changed_osds(osdc); 16293d14c5d2SYehuda Sadeh } 16306f6c7006SSage Weil 16316f6c7006SSage Weil 16323d14c5d2SYehuda Sadeh /* 16333d14c5d2SYehuda Sadeh * Process updated osd map. 16343d14c5d2SYehuda Sadeh * 16353d14c5d2SYehuda Sadeh * The message contains any number of incremental and full maps, normally 16363d14c5d2SYehuda Sadeh * indicating some sort of topology change in the cluster. Kick requests 16373d14c5d2SYehuda Sadeh * off to different OSDs as needed. 16383d14c5d2SYehuda Sadeh */ 16393d14c5d2SYehuda Sadeh void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg) 16403d14c5d2SYehuda Sadeh { 16413d14c5d2SYehuda Sadeh void *p, *end, *next; 16423d14c5d2SYehuda Sadeh u32 nr_maps, maplen; 16433d14c5d2SYehuda Sadeh u32 epoch; 16443d14c5d2SYehuda Sadeh struct ceph_osdmap *newmap = NULL, *oldmap; 16453d14c5d2SYehuda Sadeh int err; 16463d14c5d2SYehuda Sadeh struct ceph_fsid fsid; 16473d14c5d2SYehuda Sadeh 16483d14c5d2SYehuda Sadeh dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0); 16493d14c5d2SYehuda Sadeh p = msg->front.iov_base; 16503d14c5d2SYehuda Sadeh end = p + msg->front.iov_len; 16513d14c5d2SYehuda Sadeh 16523d14c5d2SYehuda Sadeh /* verify fsid */ 16533d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, sizeof(fsid), bad); 16543d14c5d2SYehuda Sadeh ceph_decode_copy(&p, &fsid, sizeof(fsid)); 16553d14c5d2SYehuda Sadeh if (ceph_check_fsid(osdc->client, &fsid) < 0) 16563d14c5d2SYehuda Sadeh return; 16573d14c5d2SYehuda Sadeh 16583d14c5d2SYehuda Sadeh down_write(&osdc->map_sem); 16593d14c5d2SYehuda Sadeh 16603d14c5d2SYehuda Sadeh /* incremental maps */ 16613d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 16623d14c5d2SYehuda Sadeh dout(" %d inc maps\n", nr_maps); 16633d14c5d2SYehuda Sadeh while (nr_maps > 0) { 16643d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 16653d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 16663d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 16673d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 16683d14c5d2SYehuda Sadeh next = p + maplen; 16693d14c5d2SYehuda Sadeh if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) { 16703d14c5d2SYehuda Sadeh dout("applying incremental map %u len %d\n", 16713d14c5d2SYehuda Sadeh epoch, maplen); 16723d14c5d2SYehuda Sadeh newmap = osdmap_apply_incremental(&p, next, 16733d14c5d2SYehuda Sadeh osdc->osdmap, 167415d9882cSAlex Elder &osdc->client->msgr); 16753d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 16763d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 16773d14c5d2SYehuda Sadeh goto bad; 16783d14c5d2SYehuda Sadeh } 16793d14c5d2SYehuda Sadeh BUG_ON(!newmap); 16803d14c5d2SYehuda Sadeh if (newmap != osdc->osdmap) { 16813d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 16823d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 16833d14c5d2SYehuda Sadeh } 168438d6453cSSage Weil kick_requests(osdc, 0); 16853d14c5d2SYehuda Sadeh } else { 16863d14c5d2SYehuda Sadeh dout("ignoring incremental map %u len %d\n", 16873d14c5d2SYehuda Sadeh epoch, maplen); 16883d14c5d2SYehuda Sadeh } 16893d14c5d2SYehuda Sadeh p = next; 16903d14c5d2SYehuda Sadeh nr_maps--; 16913d14c5d2SYehuda Sadeh } 16923d14c5d2SYehuda Sadeh if (newmap) 16933d14c5d2SYehuda Sadeh goto done; 16943d14c5d2SYehuda Sadeh 16953d14c5d2SYehuda Sadeh /* full maps */ 16963d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 16973d14c5d2SYehuda Sadeh dout(" %d full maps\n", nr_maps); 16983d14c5d2SYehuda Sadeh while (nr_maps) { 16993d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 17003d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 17013d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 17023d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 17033d14c5d2SYehuda Sadeh if (nr_maps > 1) { 17043d14c5d2SYehuda Sadeh dout("skipping non-latest full map %u len %d\n", 17053d14c5d2SYehuda Sadeh epoch, maplen); 17063d14c5d2SYehuda Sadeh } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) { 17073d14c5d2SYehuda Sadeh dout("skipping full map %u len %d, " 17083d14c5d2SYehuda Sadeh "older than our %u\n", epoch, maplen, 17093d14c5d2SYehuda Sadeh osdc->osdmap->epoch); 17103d14c5d2SYehuda Sadeh } else { 171138d6453cSSage Weil int skipped_map = 0; 171238d6453cSSage Weil 17133d14c5d2SYehuda Sadeh dout("taking full map %u len %d\n", epoch, maplen); 17143d14c5d2SYehuda Sadeh newmap = osdmap_decode(&p, p+maplen); 17153d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 17163d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 17173d14c5d2SYehuda Sadeh goto bad; 17183d14c5d2SYehuda Sadeh } 17193d14c5d2SYehuda Sadeh BUG_ON(!newmap); 17203d14c5d2SYehuda Sadeh oldmap = osdc->osdmap; 17213d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 172238d6453cSSage Weil if (oldmap) { 172338d6453cSSage Weil if (oldmap->epoch + 1 < newmap->epoch) 172438d6453cSSage Weil skipped_map = 1; 17253d14c5d2SYehuda Sadeh ceph_osdmap_destroy(oldmap); 172638d6453cSSage Weil } 172738d6453cSSage Weil kick_requests(osdc, skipped_map); 17283d14c5d2SYehuda Sadeh } 17293d14c5d2SYehuda Sadeh p += maplen; 17303d14c5d2SYehuda Sadeh nr_maps--; 17313d14c5d2SYehuda Sadeh } 17323d14c5d2SYehuda Sadeh 17333d14c5d2SYehuda Sadeh done: 17343d14c5d2SYehuda Sadeh downgrade_write(&osdc->map_sem); 17353d14c5d2SYehuda Sadeh ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch); 1736cd634fb6SSage Weil 1737cd634fb6SSage Weil /* 1738cd634fb6SSage Weil * subscribe to subsequent osdmap updates if full to ensure 1739cd634fb6SSage Weil * we find out when we are no longer full and stop returning 1740cd634fb6SSage Weil * ENOSPC. 1741cd634fb6SSage Weil */ 1742cd634fb6SSage Weil if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL)) 1743cd634fb6SSage Weil ceph_monc_request_next_osdmap(&osdc->client->monc); 1744cd634fb6SSage Weil 1745f9d25199SAlex Elder mutex_lock(&osdc->request_mutex); 1746f9d25199SAlex Elder __send_queued(osdc); 1747f9d25199SAlex Elder mutex_unlock(&osdc->request_mutex); 17483d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 17493d14c5d2SYehuda Sadeh wake_up_all(&osdc->client->auth_wq); 17503d14c5d2SYehuda Sadeh return; 17513d14c5d2SYehuda Sadeh 17523d14c5d2SYehuda Sadeh bad: 17533d14c5d2SYehuda Sadeh pr_err("osdc handle_map corrupt msg\n"); 17543d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 17553d14c5d2SYehuda Sadeh up_write(&osdc->map_sem); 17563d14c5d2SYehuda Sadeh return; 17573d14c5d2SYehuda Sadeh } 17583d14c5d2SYehuda Sadeh 17593d14c5d2SYehuda Sadeh /* 1760a40c4f10SYehuda Sadeh * watch/notify callback event infrastructure 1761a40c4f10SYehuda Sadeh * 1762a40c4f10SYehuda Sadeh * These callbacks are used both for watch and notify operations. 1763a40c4f10SYehuda Sadeh */ 1764a40c4f10SYehuda Sadeh static void __release_event(struct kref *kref) 1765a40c4f10SYehuda Sadeh { 1766a40c4f10SYehuda Sadeh struct ceph_osd_event *event = 1767a40c4f10SYehuda Sadeh container_of(kref, struct ceph_osd_event, kref); 1768a40c4f10SYehuda Sadeh 1769a40c4f10SYehuda Sadeh dout("__release_event %p\n", event); 1770a40c4f10SYehuda Sadeh kfree(event); 1771a40c4f10SYehuda Sadeh } 1772a40c4f10SYehuda Sadeh 1773a40c4f10SYehuda Sadeh static void get_event(struct ceph_osd_event *event) 1774a40c4f10SYehuda Sadeh { 1775a40c4f10SYehuda Sadeh kref_get(&event->kref); 1776a40c4f10SYehuda Sadeh } 1777a40c4f10SYehuda Sadeh 1778a40c4f10SYehuda Sadeh void ceph_osdc_put_event(struct ceph_osd_event *event) 1779a40c4f10SYehuda Sadeh { 1780a40c4f10SYehuda Sadeh kref_put(&event->kref, __release_event); 1781a40c4f10SYehuda Sadeh } 1782a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_put_event); 1783a40c4f10SYehuda Sadeh 1784a40c4f10SYehuda Sadeh static void __insert_event(struct ceph_osd_client *osdc, 1785a40c4f10SYehuda Sadeh struct ceph_osd_event *new) 1786a40c4f10SYehuda Sadeh { 1787a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 1788a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 1789a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 1790a40c4f10SYehuda Sadeh 1791a40c4f10SYehuda Sadeh while (*p) { 1792a40c4f10SYehuda Sadeh parent = *p; 1793a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 1794a40c4f10SYehuda Sadeh if (new->cookie < event->cookie) 1795a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 1796a40c4f10SYehuda Sadeh else if (new->cookie > event->cookie) 1797a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 1798a40c4f10SYehuda Sadeh else 1799a40c4f10SYehuda Sadeh BUG(); 1800a40c4f10SYehuda Sadeh } 1801a40c4f10SYehuda Sadeh 1802a40c4f10SYehuda Sadeh rb_link_node(&new->node, parent, p); 1803a40c4f10SYehuda Sadeh rb_insert_color(&new->node, &osdc->event_tree); 1804a40c4f10SYehuda Sadeh } 1805a40c4f10SYehuda Sadeh 1806a40c4f10SYehuda Sadeh static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc, 1807a40c4f10SYehuda Sadeh u64 cookie) 1808a40c4f10SYehuda Sadeh { 1809a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 1810a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 1811a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 1812a40c4f10SYehuda Sadeh 1813a40c4f10SYehuda Sadeh while (*p) { 1814a40c4f10SYehuda Sadeh parent = *p; 1815a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 1816a40c4f10SYehuda Sadeh if (cookie < event->cookie) 1817a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 1818a40c4f10SYehuda Sadeh else if (cookie > event->cookie) 1819a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 1820a40c4f10SYehuda Sadeh else 1821a40c4f10SYehuda Sadeh return event; 1822a40c4f10SYehuda Sadeh } 1823a40c4f10SYehuda Sadeh return NULL; 1824a40c4f10SYehuda Sadeh } 1825a40c4f10SYehuda Sadeh 1826a40c4f10SYehuda Sadeh static void __remove_event(struct ceph_osd_event *event) 1827a40c4f10SYehuda Sadeh { 1828a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 1829a40c4f10SYehuda Sadeh 1830a40c4f10SYehuda Sadeh if (!RB_EMPTY_NODE(&event->node)) { 1831a40c4f10SYehuda Sadeh dout("__remove_event removed %p\n", event); 1832a40c4f10SYehuda Sadeh rb_erase(&event->node, &osdc->event_tree); 1833a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 1834a40c4f10SYehuda Sadeh } else { 1835a40c4f10SYehuda Sadeh dout("__remove_event didn't remove %p\n", event); 1836a40c4f10SYehuda Sadeh } 1837a40c4f10SYehuda Sadeh } 1838a40c4f10SYehuda Sadeh 1839a40c4f10SYehuda Sadeh int ceph_osdc_create_event(struct ceph_osd_client *osdc, 1840a40c4f10SYehuda Sadeh void (*event_cb)(u64, u64, u8, void *), 18413c663bbdSAlex Elder void *data, struct ceph_osd_event **pevent) 1842a40c4f10SYehuda Sadeh { 1843a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 1844a40c4f10SYehuda Sadeh 1845a40c4f10SYehuda Sadeh event = kmalloc(sizeof(*event), GFP_NOIO); 1846a40c4f10SYehuda Sadeh if (!event) 1847a40c4f10SYehuda Sadeh return -ENOMEM; 1848a40c4f10SYehuda Sadeh 1849a40c4f10SYehuda Sadeh dout("create_event %p\n", event); 1850a40c4f10SYehuda Sadeh event->cb = event_cb; 18513c663bbdSAlex Elder event->one_shot = 0; 1852a40c4f10SYehuda Sadeh event->data = data; 1853a40c4f10SYehuda Sadeh event->osdc = osdc; 1854a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&event->osd_node); 18553ee5234dSAlex Elder RB_CLEAR_NODE(&event->node); 1856a40c4f10SYehuda Sadeh kref_init(&event->kref); /* one ref for us */ 1857a40c4f10SYehuda Sadeh kref_get(&event->kref); /* one ref for the caller */ 1858a40c4f10SYehuda Sadeh 1859a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 1860a40c4f10SYehuda Sadeh event->cookie = ++osdc->event_count; 1861a40c4f10SYehuda Sadeh __insert_event(osdc, event); 1862a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 1863a40c4f10SYehuda Sadeh 1864a40c4f10SYehuda Sadeh *pevent = event; 1865a40c4f10SYehuda Sadeh return 0; 1866a40c4f10SYehuda Sadeh } 1867a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_create_event); 1868a40c4f10SYehuda Sadeh 1869a40c4f10SYehuda Sadeh void ceph_osdc_cancel_event(struct ceph_osd_event *event) 1870a40c4f10SYehuda Sadeh { 1871a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 1872a40c4f10SYehuda Sadeh 1873a40c4f10SYehuda Sadeh dout("cancel_event %p\n", event); 1874a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 1875a40c4f10SYehuda Sadeh __remove_event(event); 1876a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 1877a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); /* caller's */ 1878a40c4f10SYehuda Sadeh } 1879a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_cancel_event); 1880a40c4f10SYehuda Sadeh 1881a40c4f10SYehuda Sadeh 1882a40c4f10SYehuda Sadeh static void do_event_work(struct work_struct *work) 1883a40c4f10SYehuda Sadeh { 1884a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work = 1885a40c4f10SYehuda Sadeh container_of(work, struct ceph_osd_event_work, work); 1886a40c4f10SYehuda Sadeh struct ceph_osd_event *event = event_work->event; 1887a40c4f10SYehuda Sadeh u64 ver = event_work->ver; 1888a40c4f10SYehuda Sadeh u64 notify_id = event_work->notify_id; 1889a40c4f10SYehuda Sadeh u8 opcode = event_work->opcode; 1890a40c4f10SYehuda Sadeh 1891a40c4f10SYehuda Sadeh dout("do_event_work completing %p\n", event); 1892a40c4f10SYehuda Sadeh event->cb(ver, notify_id, opcode, event->data); 1893a40c4f10SYehuda Sadeh dout("do_event_work completed %p\n", event); 1894a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 1895a40c4f10SYehuda Sadeh kfree(event_work); 1896a40c4f10SYehuda Sadeh } 1897a40c4f10SYehuda Sadeh 1898a40c4f10SYehuda Sadeh 1899a40c4f10SYehuda Sadeh /* 1900a40c4f10SYehuda Sadeh * Process osd watch notifications 1901a40c4f10SYehuda Sadeh */ 19023c663bbdSAlex Elder static void handle_watch_notify(struct ceph_osd_client *osdc, 19033c663bbdSAlex Elder struct ceph_msg *msg) 1904a40c4f10SYehuda Sadeh { 1905a40c4f10SYehuda Sadeh void *p, *end; 1906a40c4f10SYehuda Sadeh u8 proto_ver; 1907a40c4f10SYehuda Sadeh u64 cookie, ver, notify_id; 1908a40c4f10SYehuda Sadeh u8 opcode; 1909a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 1910a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work; 1911a40c4f10SYehuda Sadeh 1912a40c4f10SYehuda Sadeh p = msg->front.iov_base; 1913a40c4f10SYehuda Sadeh end = p + msg->front.iov_len; 1914a40c4f10SYehuda Sadeh 1915a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, proto_ver, bad); 1916a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, opcode, bad); 1917a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, cookie, bad); 1918a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, ver, bad); 1919a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, notify_id, bad); 1920a40c4f10SYehuda Sadeh 1921a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 1922a40c4f10SYehuda Sadeh event = __find_event(osdc, cookie); 1923a40c4f10SYehuda Sadeh if (event) { 19243c663bbdSAlex Elder BUG_ON(event->one_shot); 1925a40c4f10SYehuda Sadeh get_event(event); 1926a40c4f10SYehuda Sadeh } 1927a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 1928a40c4f10SYehuda Sadeh dout("handle_watch_notify cookie %lld ver %lld event %p\n", 1929a40c4f10SYehuda Sadeh cookie, ver, event); 1930a40c4f10SYehuda Sadeh if (event) { 1931a40c4f10SYehuda Sadeh event_work = kmalloc(sizeof(*event_work), GFP_NOIO); 1932a40c4f10SYehuda Sadeh if (!event_work) { 1933a40c4f10SYehuda Sadeh dout("ERROR: could not allocate event_work\n"); 1934a40c4f10SYehuda Sadeh goto done_err; 1935a40c4f10SYehuda Sadeh } 19366b0ae409SMariusz Kozlowski INIT_WORK(&event_work->work, do_event_work); 1937a40c4f10SYehuda Sadeh event_work->event = event; 1938a40c4f10SYehuda Sadeh event_work->ver = ver; 1939a40c4f10SYehuda Sadeh event_work->notify_id = notify_id; 1940a40c4f10SYehuda Sadeh event_work->opcode = opcode; 1941a40c4f10SYehuda Sadeh if (!queue_work(osdc->notify_wq, &event_work->work)) { 1942a40c4f10SYehuda Sadeh dout("WARNING: failed to queue notify event work\n"); 1943a40c4f10SYehuda Sadeh goto done_err; 1944a40c4f10SYehuda Sadeh } 1945a40c4f10SYehuda Sadeh } 1946a40c4f10SYehuda Sadeh 1947a40c4f10SYehuda Sadeh return; 1948a40c4f10SYehuda Sadeh 1949a40c4f10SYehuda Sadeh done_err: 1950a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 1951a40c4f10SYehuda Sadeh return; 1952a40c4f10SYehuda Sadeh 1953a40c4f10SYehuda Sadeh bad: 1954a40c4f10SYehuda Sadeh pr_err("osdc handle_watch_notify corrupt msg\n"); 1955a40c4f10SYehuda Sadeh return; 1956a40c4f10SYehuda Sadeh } 1957a40c4f10SYehuda Sadeh 1958a40c4f10SYehuda Sadeh /* 1959e65550fdSAlex Elder * build new request AND message 1960e65550fdSAlex Elder * 1961e65550fdSAlex Elder */ 1962e65550fdSAlex Elder void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off, 1963e65550fdSAlex Elder struct ceph_snap_context *snapc, u64 snap_id, 1964e65550fdSAlex Elder struct timespec *mtime) 1965e65550fdSAlex Elder { 1966e65550fdSAlex Elder struct ceph_msg *msg = req->r_request; 1967e65550fdSAlex Elder void *p; 1968e65550fdSAlex Elder size_t msg_size; 1969e65550fdSAlex Elder int flags = req->r_flags; 1970e65550fdSAlex Elder u64 data_len; 1971e65550fdSAlex Elder unsigned int i; 1972e65550fdSAlex Elder 1973e65550fdSAlex Elder req->r_snapid = snap_id; 1974e65550fdSAlex Elder req->r_snapc = ceph_get_snap_context(snapc); 1975e65550fdSAlex Elder 1976e65550fdSAlex Elder /* encode request */ 1977e65550fdSAlex Elder msg->hdr.version = cpu_to_le16(4); 1978e65550fdSAlex Elder 1979e65550fdSAlex Elder p = msg->front.iov_base; 1980e65550fdSAlex Elder ceph_encode_32(&p, 1); /* client_inc is always 1 */ 1981e65550fdSAlex Elder req->r_request_osdmap_epoch = p; 1982e65550fdSAlex Elder p += 4; 1983e65550fdSAlex Elder req->r_request_flags = p; 1984e65550fdSAlex Elder p += 4; 1985e65550fdSAlex Elder if (req->r_flags & CEPH_OSD_FLAG_WRITE) 1986e65550fdSAlex Elder ceph_encode_timespec(p, mtime); 1987e65550fdSAlex Elder p += sizeof(struct ceph_timespec); 1988e65550fdSAlex Elder req->r_request_reassert_version = p; 1989e65550fdSAlex Elder p += sizeof(struct ceph_eversion); /* will get filled in */ 1990e65550fdSAlex Elder 1991e65550fdSAlex Elder /* oloc */ 1992e65550fdSAlex Elder ceph_encode_8(&p, 4); 1993e65550fdSAlex Elder ceph_encode_8(&p, 4); 1994e65550fdSAlex Elder ceph_encode_32(&p, 8 + 4 + 4); 1995e65550fdSAlex Elder req->r_request_pool = p; 1996e65550fdSAlex Elder p += 8; 1997e65550fdSAlex Elder ceph_encode_32(&p, -1); /* preferred */ 1998e65550fdSAlex Elder ceph_encode_32(&p, 0); /* key len */ 1999e65550fdSAlex Elder 2000e65550fdSAlex Elder ceph_encode_8(&p, 1); 2001e65550fdSAlex Elder req->r_request_pgid = p; 2002e65550fdSAlex Elder p += 8 + 4; 2003e65550fdSAlex Elder ceph_encode_32(&p, -1); /* preferred */ 2004e65550fdSAlex Elder 2005e65550fdSAlex Elder /* oid */ 2006e65550fdSAlex Elder ceph_encode_32(&p, req->r_oid_len); 2007e65550fdSAlex Elder memcpy(p, req->r_oid, req->r_oid_len); 2008e65550fdSAlex Elder dout("oid '%.*s' len %d\n", req->r_oid_len, req->r_oid, req->r_oid_len); 2009e65550fdSAlex Elder p += req->r_oid_len; 2010e65550fdSAlex Elder 2011e65550fdSAlex Elder /* ops--can imply data */ 2012e65550fdSAlex Elder ceph_encode_16(&p, (u16)req->r_num_ops); 2013e65550fdSAlex Elder data_len = 0; 2014e65550fdSAlex Elder for (i = 0; i < req->r_num_ops; i++) { 2015e65550fdSAlex Elder data_len += osd_req_encode_op(req, p, i); 2016e65550fdSAlex Elder p += sizeof(struct ceph_osd_op); 2017e65550fdSAlex Elder } 2018e65550fdSAlex Elder 2019e65550fdSAlex Elder /* snaps */ 2020e65550fdSAlex Elder ceph_encode_64(&p, req->r_snapid); 2021e65550fdSAlex Elder ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0); 2022e65550fdSAlex Elder ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0); 2023e65550fdSAlex Elder if (req->r_snapc) { 2024e65550fdSAlex Elder for (i = 0; i < snapc->num_snaps; i++) { 2025e65550fdSAlex Elder ceph_encode_64(&p, req->r_snapc->snaps[i]); 2026e65550fdSAlex Elder } 2027e65550fdSAlex Elder } 2028e65550fdSAlex Elder 2029e65550fdSAlex Elder req->r_request_attempts = p; 2030e65550fdSAlex Elder p += 4; 2031e65550fdSAlex Elder 2032e65550fdSAlex Elder /* data */ 2033e65550fdSAlex Elder if (flags & CEPH_OSD_FLAG_WRITE) { 2034e65550fdSAlex Elder u16 data_off; 2035e65550fdSAlex Elder 2036e65550fdSAlex Elder /* 2037e65550fdSAlex Elder * The header "data_off" is a hint to the receiver 2038e65550fdSAlex Elder * allowing it to align received data into its 2039e65550fdSAlex Elder * buffers such that there's no need to re-copy 2040e65550fdSAlex Elder * it before writing it to disk (direct I/O). 2041e65550fdSAlex Elder */ 2042e65550fdSAlex Elder data_off = (u16) (off & 0xffff); 2043e65550fdSAlex Elder req->r_request->hdr.data_off = cpu_to_le16(data_off); 2044e65550fdSAlex Elder } 2045e65550fdSAlex Elder req->r_request->hdr.data_len = cpu_to_le32(data_len); 2046e65550fdSAlex Elder 2047e65550fdSAlex Elder BUG_ON(p > msg->front.iov_base + msg->front.iov_len); 2048e65550fdSAlex Elder msg_size = p - msg->front.iov_base; 2049e65550fdSAlex Elder msg->front.iov_len = msg_size; 2050e65550fdSAlex Elder msg->hdr.front_len = cpu_to_le32(msg_size); 2051e65550fdSAlex Elder 2052e65550fdSAlex Elder dout("build_request msg_size was %d\n", (int)msg_size); 2053e65550fdSAlex Elder } 2054e65550fdSAlex Elder EXPORT_SYMBOL(ceph_osdc_build_request); 2055e65550fdSAlex Elder 2056e65550fdSAlex Elder /* 20573d14c5d2SYehuda Sadeh * Register request, send initial attempt. 20583d14c5d2SYehuda Sadeh */ 20593d14c5d2SYehuda Sadeh int ceph_osdc_start_request(struct ceph_osd_client *osdc, 20603d14c5d2SYehuda Sadeh struct ceph_osd_request *req, 20613d14c5d2SYehuda Sadeh bool nofail) 20623d14c5d2SYehuda Sadeh { 20633d14c5d2SYehuda Sadeh int rc = 0; 20643d14c5d2SYehuda Sadeh 20653d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 20663d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 2067dc4b870cSAlex Elder __register_request(osdc, req); 206892451b49SAlex Elder WARN_ON(req->r_sent); 206938d6453cSSage Weil rc = __map_request(osdc, req, 0); 20709d6fcb08SSage Weil if (rc < 0) { 20719d6fcb08SSage Weil if (nofail) { 20729d6fcb08SSage Weil dout("osdc_start_request failed map, " 20739d6fcb08SSage Weil " will retry %lld\n", req->r_tid); 20749d6fcb08SSage Weil rc = 0; 20759d6fcb08SSage Weil } 2076234af26fSDan Carpenter goto out_unlock; 20779d6fcb08SSage Weil } 20786f6c7006SSage Weil if (req->r_osd == NULL) { 20796f6c7006SSage Weil dout("send_request %p no up osds in pg\n", req); 20806f6c7006SSage Weil ceph_monc_request_next_osdmap(&osdc->client->monc); 20816f6c7006SSage Weil } else { 20827e2766a1SAlex Elder __send_queued(osdc); 208356e925b6SSage Weil } 20843d14c5d2SYehuda Sadeh rc = 0; 2085234af26fSDan Carpenter out_unlock: 20863d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 20873d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 20883d14c5d2SYehuda Sadeh return rc; 20893d14c5d2SYehuda Sadeh } 20903d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_start_request); 20913d14c5d2SYehuda Sadeh 20923d14c5d2SYehuda Sadeh /* 20933d14c5d2SYehuda Sadeh * wait for a request to complete 20943d14c5d2SYehuda Sadeh */ 20953d14c5d2SYehuda Sadeh int ceph_osdc_wait_request(struct ceph_osd_client *osdc, 20963d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 20973d14c5d2SYehuda Sadeh { 20983d14c5d2SYehuda Sadeh int rc; 20993d14c5d2SYehuda Sadeh 21003d14c5d2SYehuda Sadeh rc = wait_for_completion_interruptible(&req->r_completion); 21013d14c5d2SYehuda Sadeh if (rc < 0) { 21023d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 21033d14c5d2SYehuda Sadeh __cancel_request(req); 21043d14c5d2SYehuda Sadeh __unregister_request(osdc, req); 21053d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 210625845472SSage Weil complete_request(req); 21073d14c5d2SYehuda Sadeh dout("wait_request tid %llu canceled/timed out\n", req->r_tid); 21083d14c5d2SYehuda Sadeh return rc; 21093d14c5d2SYehuda Sadeh } 21103d14c5d2SYehuda Sadeh 21113d14c5d2SYehuda Sadeh dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result); 21123d14c5d2SYehuda Sadeh return req->r_result; 21133d14c5d2SYehuda Sadeh } 21143d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_wait_request); 21153d14c5d2SYehuda Sadeh 21163d14c5d2SYehuda Sadeh /* 21173d14c5d2SYehuda Sadeh * sync - wait for all in-flight requests to flush. avoid starvation. 21183d14c5d2SYehuda Sadeh */ 21193d14c5d2SYehuda Sadeh void ceph_osdc_sync(struct ceph_osd_client *osdc) 21203d14c5d2SYehuda Sadeh { 21213d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 21223d14c5d2SYehuda Sadeh u64 last_tid, next_tid = 0; 21233d14c5d2SYehuda Sadeh 21243d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 21253d14c5d2SYehuda Sadeh last_tid = osdc->last_tid; 21263d14c5d2SYehuda Sadeh while (1) { 21273d14c5d2SYehuda Sadeh req = __lookup_request_ge(osdc, next_tid); 21283d14c5d2SYehuda Sadeh if (!req) 21293d14c5d2SYehuda Sadeh break; 21303d14c5d2SYehuda Sadeh if (req->r_tid > last_tid) 21313d14c5d2SYehuda Sadeh break; 21323d14c5d2SYehuda Sadeh 21333d14c5d2SYehuda Sadeh next_tid = req->r_tid + 1; 21343d14c5d2SYehuda Sadeh if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0) 21353d14c5d2SYehuda Sadeh continue; 21363d14c5d2SYehuda Sadeh 21373d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 21383d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 21393d14c5d2SYehuda Sadeh dout("sync waiting on tid %llu (last is %llu)\n", 21403d14c5d2SYehuda Sadeh req->r_tid, last_tid); 21413d14c5d2SYehuda Sadeh wait_for_completion(&req->r_safe_completion); 21423d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 21433d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 21443d14c5d2SYehuda Sadeh } 21453d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 21463d14c5d2SYehuda Sadeh dout("sync done (thru tid %llu)\n", last_tid); 21473d14c5d2SYehuda Sadeh } 21483d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_sync); 21493d14c5d2SYehuda Sadeh 21503d14c5d2SYehuda Sadeh /* 21513d14c5d2SYehuda Sadeh * init, shutdown 21523d14c5d2SYehuda Sadeh */ 21533d14c5d2SYehuda Sadeh int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client) 21543d14c5d2SYehuda Sadeh { 21553d14c5d2SYehuda Sadeh int err; 21563d14c5d2SYehuda Sadeh 21573d14c5d2SYehuda Sadeh dout("init\n"); 21583d14c5d2SYehuda Sadeh osdc->client = client; 21593d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 21603d14c5d2SYehuda Sadeh init_rwsem(&osdc->map_sem); 21613d14c5d2SYehuda Sadeh init_completion(&osdc->map_waiters); 21623d14c5d2SYehuda Sadeh osdc->last_requested_map = 0; 21633d14c5d2SYehuda Sadeh mutex_init(&osdc->request_mutex); 21643d14c5d2SYehuda Sadeh osdc->last_tid = 0; 21653d14c5d2SYehuda Sadeh osdc->osds = RB_ROOT; 21663d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->osd_lru); 21673d14c5d2SYehuda Sadeh osdc->requests = RB_ROOT; 21683d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_lru); 21696f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_unsent); 21706f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_notarget); 2171a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_linger); 21723d14c5d2SYehuda Sadeh osdc->num_requests = 0; 21733d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout); 21743d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout); 2175a40c4f10SYehuda Sadeh spin_lock_init(&osdc->event_lock); 2176a40c4f10SYehuda Sadeh osdc->event_tree = RB_ROOT; 2177a40c4f10SYehuda Sadeh osdc->event_count = 0; 21783d14c5d2SYehuda Sadeh 21793d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 21803d14c5d2SYehuda Sadeh round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ)); 21813d14c5d2SYehuda Sadeh 21823d14c5d2SYehuda Sadeh err = -ENOMEM; 21833d14c5d2SYehuda Sadeh osdc->req_mempool = mempool_create_kmalloc_pool(10, 21843d14c5d2SYehuda Sadeh sizeof(struct ceph_osd_request)); 21853d14c5d2SYehuda Sadeh if (!osdc->req_mempool) 21863d14c5d2SYehuda Sadeh goto out; 21873d14c5d2SYehuda Sadeh 2188d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP, 2189d50b409fSSage Weil OSD_OP_FRONT_LEN, 10, true, 21903d14c5d2SYehuda Sadeh "osd_op"); 21913d14c5d2SYehuda Sadeh if (err < 0) 21923d14c5d2SYehuda Sadeh goto out_mempool; 2193d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY, 21943d14c5d2SYehuda Sadeh OSD_OPREPLY_FRONT_LEN, 10, true, 21953d14c5d2SYehuda Sadeh "osd_op_reply"); 21963d14c5d2SYehuda Sadeh if (err < 0) 21973d14c5d2SYehuda Sadeh goto out_msgpool; 2198a40c4f10SYehuda Sadeh 2199a40c4f10SYehuda Sadeh osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify"); 2200a40c4f10SYehuda Sadeh if (IS_ERR(osdc->notify_wq)) { 2201a40c4f10SYehuda Sadeh err = PTR_ERR(osdc->notify_wq); 2202a40c4f10SYehuda Sadeh osdc->notify_wq = NULL; 2203a40c4f10SYehuda Sadeh goto out_msgpool; 2204a40c4f10SYehuda Sadeh } 22053d14c5d2SYehuda Sadeh return 0; 22063d14c5d2SYehuda Sadeh 22073d14c5d2SYehuda Sadeh out_msgpool: 22083d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 22093d14c5d2SYehuda Sadeh out_mempool: 22103d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 22113d14c5d2SYehuda Sadeh out: 22123d14c5d2SYehuda Sadeh return err; 22133d14c5d2SYehuda Sadeh } 22143d14c5d2SYehuda Sadeh 22153d14c5d2SYehuda Sadeh void ceph_osdc_stop(struct ceph_osd_client *osdc) 22163d14c5d2SYehuda Sadeh { 2217a40c4f10SYehuda Sadeh flush_workqueue(osdc->notify_wq); 2218a40c4f10SYehuda Sadeh destroy_workqueue(osdc->notify_wq); 22193d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->timeout_work); 22203d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->osds_timeout_work); 22213d14c5d2SYehuda Sadeh if (osdc->osdmap) { 22223d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 22233d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 22243d14c5d2SYehuda Sadeh } 2225aca420bcSSage Weil remove_all_osds(osdc); 22263d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 22273d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 22283d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op_reply); 22293d14c5d2SYehuda Sadeh } 22303d14c5d2SYehuda Sadeh 22313d14c5d2SYehuda Sadeh /* 22323d14c5d2SYehuda Sadeh * Read some contiguous pages. If we cross a stripe boundary, shorten 22333d14c5d2SYehuda Sadeh * *plen. Return number of bytes read, or error. 22343d14c5d2SYehuda Sadeh */ 22353d14c5d2SYehuda Sadeh int ceph_osdc_readpages(struct ceph_osd_client *osdc, 22363d14c5d2SYehuda Sadeh struct ceph_vino vino, struct ceph_file_layout *layout, 22373d14c5d2SYehuda Sadeh u64 off, u64 *plen, 22383d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 2239b7495fc2SSage Weil struct page **pages, int num_pages, int page_align) 22403d14c5d2SYehuda Sadeh { 22413d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 22423d14c5d2SYehuda Sadeh int rc = 0; 22433d14c5d2SYehuda Sadeh 22443d14c5d2SYehuda Sadeh dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino, 22453d14c5d2SYehuda Sadeh vino.snap, off, *plen); 224679528734SAlex Elder req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 1, 22473d14c5d2SYehuda Sadeh CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ, 2248acead002SAlex Elder NULL, truncate_seq, truncate_size, 2249153e5167SAlex Elder false); 22506816282dSSage Weil if (IS_ERR(req)) 22516816282dSSage Weil return PTR_ERR(req); 22523d14c5d2SYehuda Sadeh 22533d14c5d2SYehuda Sadeh /* it may be a short read due to an object boundary */ 22540fff87ecSAlex Elder 2255a4ce40a9SAlex Elder osd_req_op_extent_osd_data_pages(req, 0, false, 2256a4ce40a9SAlex Elder pages, *plen, page_align, false, false); 22573d14c5d2SYehuda Sadeh 2258e0c59487SAlex Elder dout("readpages final extent is %llu~%llu (%llu bytes align %d)\n", 225943bfe5deSAlex Elder off, *plen, *plen, page_align); 22603d14c5d2SYehuda Sadeh 226179528734SAlex Elder ceph_osdc_build_request(req, off, NULL, vino.snap, NULL); 226202ee07d3SAlex Elder 22633d14c5d2SYehuda Sadeh rc = ceph_osdc_start_request(osdc, req, false); 22643d14c5d2SYehuda Sadeh if (!rc) 22653d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 22663d14c5d2SYehuda Sadeh 22673d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 22683d14c5d2SYehuda Sadeh dout("readpages result %d\n", rc); 22693d14c5d2SYehuda Sadeh return rc; 22703d14c5d2SYehuda Sadeh } 22713d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_readpages); 22723d14c5d2SYehuda Sadeh 22733d14c5d2SYehuda Sadeh /* 22743d14c5d2SYehuda Sadeh * do a synchronous write on N pages 22753d14c5d2SYehuda Sadeh */ 22763d14c5d2SYehuda Sadeh int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino, 22773d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 22783d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 22793d14c5d2SYehuda Sadeh u64 off, u64 len, 22803d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 22813d14c5d2SYehuda Sadeh struct timespec *mtime, 228224808826SAlex Elder struct page **pages, int num_pages) 22833d14c5d2SYehuda Sadeh { 22843d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 22853d14c5d2SYehuda Sadeh int rc = 0; 2286b7495fc2SSage Weil int page_align = off & ~PAGE_MASK; 22873d14c5d2SYehuda Sadeh 2288acead002SAlex Elder BUG_ON(vino.snap != CEPH_NOSNAP); /* snapshots aren't writeable */ 228979528734SAlex Elder req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 1, 22903d14c5d2SYehuda Sadeh CEPH_OSD_OP_WRITE, 229124808826SAlex Elder CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE, 2292acead002SAlex Elder snapc, truncate_seq, truncate_size, 2293153e5167SAlex Elder true); 22946816282dSSage Weil if (IS_ERR(req)) 22956816282dSSage Weil return PTR_ERR(req); 22963d14c5d2SYehuda Sadeh 22973d14c5d2SYehuda Sadeh /* it may be a short write due to an object boundary */ 2298a4ce40a9SAlex Elder osd_req_op_extent_osd_data_pages(req, 0, true, pages, len, page_align, 229943bfe5deSAlex Elder false, false); 230043bfe5deSAlex Elder dout("writepages %llu~%llu (%llu bytes)\n", off, len, len); 23013d14c5d2SYehuda Sadeh 230279528734SAlex Elder ceph_osdc_build_request(req, off, snapc, CEPH_NOSNAP, mtime); 230302ee07d3SAlex Elder 230487f979d3SAlex Elder rc = ceph_osdc_start_request(osdc, req, true); 23053d14c5d2SYehuda Sadeh if (!rc) 23063d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 23073d14c5d2SYehuda Sadeh 23083d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 23093d14c5d2SYehuda Sadeh if (rc == 0) 23103d14c5d2SYehuda Sadeh rc = len; 23113d14c5d2SYehuda Sadeh dout("writepages result %d\n", rc); 23123d14c5d2SYehuda Sadeh return rc; 23133d14c5d2SYehuda Sadeh } 23143d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_writepages); 23153d14c5d2SYehuda Sadeh 23163d14c5d2SYehuda Sadeh /* 23173d14c5d2SYehuda Sadeh * handle incoming message 23183d14c5d2SYehuda Sadeh */ 23193d14c5d2SYehuda Sadeh static void dispatch(struct ceph_connection *con, struct ceph_msg *msg) 23203d14c5d2SYehuda Sadeh { 23213d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 23223d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 23233d14c5d2SYehuda Sadeh int type = le16_to_cpu(msg->hdr.type); 23243d14c5d2SYehuda Sadeh 23253d14c5d2SYehuda Sadeh if (!osd) 23263d14c5d2SYehuda Sadeh goto out; 23273d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 23283d14c5d2SYehuda Sadeh 23293d14c5d2SYehuda Sadeh switch (type) { 23303d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 23313d14c5d2SYehuda Sadeh ceph_osdc_handle_map(osdc, msg); 23323d14c5d2SYehuda Sadeh break; 23333d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 23343d14c5d2SYehuda Sadeh handle_reply(osdc, msg, con); 23353d14c5d2SYehuda Sadeh break; 2336a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 2337a40c4f10SYehuda Sadeh handle_watch_notify(osdc, msg); 2338a40c4f10SYehuda Sadeh break; 23393d14c5d2SYehuda Sadeh 23403d14c5d2SYehuda Sadeh default: 23413d14c5d2SYehuda Sadeh pr_err("received unknown message type %d %s\n", type, 23423d14c5d2SYehuda Sadeh ceph_msg_type_name(type)); 23433d14c5d2SYehuda Sadeh } 23443d14c5d2SYehuda Sadeh out: 23453d14c5d2SYehuda Sadeh ceph_msg_put(msg); 23463d14c5d2SYehuda Sadeh } 23473d14c5d2SYehuda Sadeh 23483d14c5d2SYehuda Sadeh /* 23493d14c5d2SYehuda Sadeh * lookup and return message for incoming reply. set up reply message 23503d14c5d2SYehuda Sadeh * pages. 23513d14c5d2SYehuda Sadeh */ 23523d14c5d2SYehuda Sadeh static struct ceph_msg *get_reply(struct ceph_connection *con, 23533d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 23543d14c5d2SYehuda Sadeh int *skip) 23553d14c5d2SYehuda Sadeh { 23563d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 23573d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = osd->o_osdc; 23583d14c5d2SYehuda Sadeh struct ceph_msg *m; 23593d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 23603d14c5d2SYehuda Sadeh int front = le32_to_cpu(hdr->front_len); 23613d14c5d2SYehuda Sadeh int data_len = le32_to_cpu(hdr->data_len); 23623d14c5d2SYehuda Sadeh u64 tid; 23633d14c5d2SYehuda Sadeh 23643d14c5d2SYehuda Sadeh tid = le64_to_cpu(hdr->tid); 23653d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 23663d14c5d2SYehuda Sadeh req = __lookup_request(osdc, tid); 23673d14c5d2SYehuda Sadeh if (!req) { 23683d14c5d2SYehuda Sadeh *skip = 1; 23693d14c5d2SYehuda Sadeh m = NULL; 2370756a16a5SSage Weil dout("get_reply unknown tid %llu from osd%d\n", tid, 23713d14c5d2SYehuda Sadeh osd->o_osd); 23723d14c5d2SYehuda Sadeh goto out; 23733d14c5d2SYehuda Sadeh } 23743d14c5d2SYehuda Sadeh 2375ace6d3a9SAlex Elder if (req->r_reply->con) 23768921d114SAlex Elder dout("%s revoking msg %p from old con %p\n", __func__, 2377ace6d3a9SAlex Elder req->r_reply, req->r_reply->con); 23788921d114SAlex Elder ceph_msg_revoke_incoming(req->r_reply); 23793d14c5d2SYehuda Sadeh 23803d14c5d2SYehuda Sadeh if (front > req->r_reply->front.iov_len) { 23813d14c5d2SYehuda Sadeh pr_warning("get_reply front %d > preallocated %d\n", 23823d14c5d2SYehuda Sadeh front, (int)req->r_reply->front.iov_len); 2383b61c2763SSage Weil m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS, false); 23843d14c5d2SYehuda Sadeh if (!m) 23853d14c5d2SYehuda Sadeh goto out; 23863d14c5d2SYehuda Sadeh ceph_msg_put(req->r_reply); 23873d14c5d2SYehuda Sadeh req->r_reply = m; 23883d14c5d2SYehuda Sadeh } 23893d14c5d2SYehuda Sadeh m = ceph_msg_get(req->r_reply); 23903d14c5d2SYehuda Sadeh 23913d14c5d2SYehuda Sadeh if (data_len > 0) { 2392a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 23930fff87ecSAlex Elder 2394a4ce40a9SAlex Elder /* 2395a4ce40a9SAlex Elder * XXX This is assuming there is only one op containing 2396a4ce40a9SAlex Elder * XXX page data. Probably OK for reads, but this 2397a4ce40a9SAlex Elder * XXX ought to be done more generally. 2398a4ce40a9SAlex Elder */ 2399a4ce40a9SAlex Elder osd_data = osd_req_op_extent_osd_data(req, 0, false); 24000fff87ecSAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 24010fff87ecSAlex Elder if (osd_data->pages && 2402e0c59487SAlex Elder unlikely(osd_data->length < data_len)) { 24032ac2b7a6SAlex Elder 2404e0c59487SAlex Elder pr_warning("tid %lld reply has %d bytes " 2405e0c59487SAlex Elder "we had only %llu bytes ready\n", 2406e0c59487SAlex Elder tid, data_len, osd_data->length); 24073d14c5d2SYehuda Sadeh *skip = 1; 24083d14c5d2SYehuda Sadeh ceph_msg_put(m); 24093d14c5d2SYehuda Sadeh m = NULL; 24103d14c5d2SYehuda Sadeh goto out; 24113d14c5d2SYehuda Sadeh } 24123d14c5d2SYehuda Sadeh } 24132ac2b7a6SAlex Elder } 24143d14c5d2SYehuda Sadeh *skip = 0; 24153d14c5d2SYehuda Sadeh dout("get_reply tid %lld %p\n", tid, m); 24163d14c5d2SYehuda Sadeh 24173d14c5d2SYehuda Sadeh out: 24183d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 24193d14c5d2SYehuda Sadeh return m; 24203d14c5d2SYehuda Sadeh 24213d14c5d2SYehuda Sadeh } 24223d14c5d2SYehuda Sadeh 24233d14c5d2SYehuda Sadeh static struct ceph_msg *alloc_msg(struct ceph_connection *con, 24243d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 24253d14c5d2SYehuda Sadeh int *skip) 24263d14c5d2SYehuda Sadeh { 24273d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 24283d14c5d2SYehuda Sadeh int type = le16_to_cpu(hdr->type); 24293d14c5d2SYehuda Sadeh int front = le32_to_cpu(hdr->front_len); 24303d14c5d2SYehuda Sadeh 24311c20f2d2SAlex Elder *skip = 0; 24323d14c5d2SYehuda Sadeh switch (type) { 24333d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 2434a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 2435b61c2763SSage Weil return ceph_msg_new(type, front, GFP_NOFS, false); 24363d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 24373d14c5d2SYehuda Sadeh return get_reply(con, hdr, skip); 24383d14c5d2SYehuda Sadeh default: 24393d14c5d2SYehuda Sadeh pr_info("alloc_msg unexpected msg type %d from osd%d\n", type, 24403d14c5d2SYehuda Sadeh osd->o_osd); 24413d14c5d2SYehuda Sadeh *skip = 1; 24423d14c5d2SYehuda Sadeh return NULL; 24433d14c5d2SYehuda Sadeh } 24443d14c5d2SYehuda Sadeh } 24453d14c5d2SYehuda Sadeh 24463d14c5d2SYehuda Sadeh /* 24473d14c5d2SYehuda Sadeh * Wrappers to refcount containing ceph_osd struct 24483d14c5d2SYehuda Sadeh */ 24493d14c5d2SYehuda Sadeh static struct ceph_connection *get_osd_con(struct ceph_connection *con) 24503d14c5d2SYehuda Sadeh { 24513d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 24523d14c5d2SYehuda Sadeh if (get_osd(osd)) 24533d14c5d2SYehuda Sadeh return con; 24543d14c5d2SYehuda Sadeh return NULL; 24553d14c5d2SYehuda Sadeh } 24563d14c5d2SYehuda Sadeh 24573d14c5d2SYehuda Sadeh static void put_osd_con(struct ceph_connection *con) 24583d14c5d2SYehuda Sadeh { 24593d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 24603d14c5d2SYehuda Sadeh put_osd(osd); 24613d14c5d2SYehuda Sadeh } 24623d14c5d2SYehuda Sadeh 24633d14c5d2SYehuda Sadeh /* 24643d14c5d2SYehuda Sadeh * authentication 24653d14c5d2SYehuda Sadeh */ 2466a3530df3SAlex Elder /* 2467a3530df3SAlex Elder * Note: returned pointer is the address of a structure that's 2468a3530df3SAlex Elder * managed separately. Caller must *not* attempt to free it. 2469a3530df3SAlex Elder */ 2470a3530df3SAlex Elder static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con, 24718f43fb53SAlex Elder int *proto, int force_new) 24723d14c5d2SYehuda Sadeh { 24733d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 24743d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 24753d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 247674f1869fSAlex Elder struct ceph_auth_handshake *auth = &o->o_auth; 24773d14c5d2SYehuda Sadeh 247874f1869fSAlex Elder if (force_new && auth->authorizer) { 247927859f97SSage Weil ceph_auth_destroy_authorizer(ac, auth->authorizer); 248074f1869fSAlex Elder auth->authorizer = NULL; 24813d14c5d2SYehuda Sadeh } 248227859f97SSage Weil if (!auth->authorizer) { 248327859f97SSage Weil int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 2484a3530df3SAlex Elder auth); 24853d14c5d2SYehuda Sadeh if (ret) 2486a3530df3SAlex Elder return ERR_PTR(ret); 248727859f97SSage Weil } else { 248827859f97SSage Weil int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 24890bed9b5cSSage Weil auth); 24900bed9b5cSSage Weil if (ret) 24910bed9b5cSSage Weil return ERR_PTR(ret); 24923d14c5d2SYehuda Sadeh } 24933d14c5d2SYehuda Sadeh *proto = ac->protocol; 249474f1869fSAlex Elder 2495a3530df3SAlex Elder return auth; 24963d14c5d2SYehuda Sadeh } 24973d14c5d2SYehuda Sadeh 24983d14c5d2SYehuda Sadeh 24993d14c5d2SYehuda Sadeh static int verify_authorizer_reply(struct ceph_connection *con, int len) 25003d14c5d2SYehuda Sadeh { 25013d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 25023d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 25033d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 25043d14c5d2SYehuda Sadeh 250527859f97SSage Weil return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len); 25063d14c5d2SYehuda Sadeh } 25073d14c5d2SYehuda Sadeh 25083d14c5d2SYehuda Sadeh static int invalidate_authorizer(struct ceph_connection *con) 25093d14c5d2SYehuda Sadeh { 25103d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 25113d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 25123d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 25133d14c5d2SYehuda Sadeh 251427859f97SSage Weil ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD); 25153d14c5d2SYehuda Sadeh return ceph_monc_validate_auth(&osdc->client->monc); 25163d14c5d2SYehuda Sadeh } 25173d14c5d2SYehuda Sadeh 25183d14c5d2SYehuda Sadeh static const struct ceph_connection_operations osd_con_ops = { 25193d14c5d2SYehuda Sadeh .get = get_osd_con, 25203d14c5d2SYehuda Sadeh .put = put_osd_con, 25213d14c5d2SYehuda Sadeh .dispatch = dispatch, 25223d14c5d2SYehuda Sadeh .get_authorizer = get_authorizer, 25233d14c5d2SYehuda Sadeh .verify_authorizer_reply = verify_authorizer_reply, 25243d14c5d2SYehuda Sadeh .invalidate_authorizer = invalidate_authorizer, 25253d14c5d2SYehuda Sadeh .alloc_msg = alloc_msg, 25263d14c5d2SYehuda Sadeh .fault = osd_reset, 25273d14c5d2SYehuda Sadeh }; 2528