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, 120406e2c9fSAlex Elder unsigned int which) 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 * 13904017e29SAlex Elder osd_req_op_cls_request_data(struct ceph_osd_request *osd_req, 14004017e29SAlex Elder unsigned int which) 14104017e29SAlex Elder { 14204017e29SAlex Elder BUG_ON(which >= osd_req->r_num_ops); 14304017e29SAlex Elder 14404017e29SAlex Elder return &osd_req->r_ops[which].cls.request_data; 14504017e29SAlex Elder } 14604017e29SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_request_data); /* ??? */ 14704017e29SAlex Elder 14804017e29SAlex Elder struct ceph_osd_data * 149a4ce40a9SAlex Elder osd_req_op_cls_response_data(struct ceph_osd_request *osd_req, 150a4ce40a9SAlex Elder unsigned int which) 151a4ce40a9SAlex Elder { 152a4ce40a9SAlex Elder BUG_ON(which >= osd_req->r_num_ops); 153a4ce40a9SAlex Elder 1545476492fSAlex Elder return &osd_req->r_ops[which].cls.response_data; 155a4ce40a9SAlex Elder } 156a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_response_data); /* ??? */ 157a4ce40a9SAlex Elder 158a4ce40a9SAlex Elder void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *osd_req, 159406e2c9fSAlex Elder unsigned int which, struct page **pages, 160406e2c9fSAlex Elder u64 length, u32 alignment, 161a4ce40a9SAlex Elder bool pages_from_pool, bool own_pages) 162a4ce40a9SAlex Elder { 163a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 164a4ce40a9SAlex Elder 165406e2c9fSAlex Elder osd_data = osd_req_op_extent_osd_data(osd_req, which); 166a4ce40a9SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 167a4ce40a9SAlex Elder pages_from_pool, own_pages); 168a4ce40a9SAlex Elder } 169a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data_pages); 170a4ce40a9SAlex Elder 171a4ce40a9SAlex Elder void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *osd_req, 172406e2c9fSAlex Elder unsigned int which, struct ceph_pagelist *pagelist) 173a4ce40a9SAlex Elder { 174a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 175a4ce40a9SAlex Elder 176406e2c9fSAlex Elder osd_data = osd_req_op_extent_osd_data(osd_req, which); 177a4ce40a9SAlex Elder ceph_osd_data_pagelist_init(osd_data, pagelist); 178a4ce40a9SAlex Elder } 179a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data_pagelist); 180a4ce40a9SAlex Elder 181a4ce40a9SAlex Elder #ifdef CONFIG_BLOCK 182a4ce40a9SAlex Elder void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *osd_req, 183406e2c9fSAlex Elder unsigned int which, struct bio *bio, size_t bio_length) 184a4ce40a9SAlex Elder { 185a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 186406e2c9fSAlex Elder osd_data = osd_req_op_extent_osd_data(osd_req, which); 187a4ce40a9SAlex Elder ceph_osd_data_bio_init(osd_data, bio, bio_length); 188a4ce40a9SAlex Elder } 189a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data_bio); 190a4ce40a9SAlex Elder #endif /* CONFIG_BLOCK */ 191a4ce40a9SAlex Elder 192a4ce40a9SAlex Elder static void osd_req_op_cls_request_info_pagelist( 193a4ce40a9SAlex Elder struct ceph_osd_request *osd_req, 194a4ce40a9SAlex Elder unsigned int which, struct ceph_pagelist *pagelist) 195a4ce40a9SAlex Elder { 196a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 197a4ce40a9SAlex Elder 198a4ce40a9SAlex Elder osd_data = osd_req_op_cls_request_info(osd_req, which); 199a4ce40a9SAlex Elder ceph_osd_data_pagelist_init(osd_data, pagelist); 200a4ce40a9SAlex Elder } 201a4ce40a9SAlex Elder 20204017e29SAlex Elder void osd_req_op_cls_request_data_pagelist( 20304017e29SAlex Elder struct ceph_osd_request *osd_req, 20404017e29SAlex Elder unsigned int which, struct ceph_pagelist *pagelist) 20504017e29SAlex Elder { 20604017e29SAlex Elder struct ceph_osd_data *osd_data; 20704017e29SAlex Elder 20804017e29SAlex Elder osd_data = osd_req_op_cls_request_data(osd_req, which); 20904017e29SAlex Elder ceph_osd_data_pagelist_init(osd_data, pagelist); 21004017e29SAlex Elder } 21104017e29SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_request_data_pagelist); 21204017e29SAlex Elder 213a4ce40a9SAlex Elder void osd_req_op_cls_response_data_pages(struct ceph_osd_request *osd_req, 214a4ce40a9SAlex Elder unsigned int which, struct page **pages, u64 length, 215a4ce40a9SAlex Elder u32 alignment, bool pages_from_pool, bool own_pages) 216a4ce40a9SAlex Elder { 217a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 218a4ce40a9SAlex Elder 219a4ce40a9SAlex Elder osd_data = osd_req_op_cls_response_data(osd_req, which); 220a4ce40a9SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 221a4ce40a9SAlex Elder pages_from_pool, own_pages); 222a4ce40a9SAlex Elder } 223a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_response_data_pages); 224a4ce40a9SAlex Elder 22523c08a9cSAlex Elder static u64 ceph_osd_data_length(struct ceph_osd_data *osd_data) 22623c08a9cSAlex Elder { 22723c08a9cSAlex Elder switch (osd_data->type) { 22823c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_NONE: 22923c08a9cSAlex Elder return 0; 23023c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_PAGES: 23123c08a9cSAlex Elder return osd_data->length; 23223c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_PAGELIST: 23323c08a9cSAlex Elder return (u64)osd_data->pagelist->length; 23423c08a9cSAlex Elder #ifdef CONFIG_BLOCK 23523c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_BIO: 23623c08a9cSAlex Elder return (u64)osd_data->bio_length; 23723c08a9cSAlex Elder #endif /* CONFIG_BLOCK */ 23823c08a9cSAlex Elder default: 23923c08a9cSAlex Elder WARN(true, "unrecognized data type %d\n", (int)osd_data->type); 24023c08a9cSAlex Elder return 0; 24123c08a9cSAlex Elder } 24223c08a9cSAlex Elder } 24323c08a9cSAlex Elder 2445476492fSAlex Elder 245c54d47bfSAlex Elder static void ceph_osd_data_release(struct ceph_osd_data *osd_data) 246c54d47bfSAlex Elder { 2475476492fSAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES && osd_data->own_pages) { 248c54d47bfSAlex Elder int num_pages; 249c54d47bfSAlex Elder 250c54d47bfSAlex Elder num_pages = calc_pages_for((u64)osd_data->alignment, 251c54d47bfSAlex Elder (u64)osd_data->length); 252c54d47bfSAlex Elder ceph_release_page_vector(osd_data->pages, num_pages); 253c54d47bfSAlex Elder } 2545476492fSAlex Elder ceph_osd_data_init(osd_data); 2555476492fSAlex Elder } 2565476492fSAlex Elder 2575476492fSAlex Elder static void osd_req_op_data_release(struct ceph_osd_request *osd_req, 2585476492fSAlex Elder unsigned int which) 2595476492fSAlex Elder { 2605476492fSAlex Elder struct ceph_osd_req_op *op; 2615476492fSAlex Elder 2625476492fSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 2635476492fSAlex Elder op = &osd_req->r_ops[which]; 2645476492fSAlex Elder 2655476492fSAlex Elder switch (op->op) { 2665476492fSAlex Elder case CEPH_OSD_OP_READ: 2675476492fSAlex Elder case CEPH_OSD_OP_WRITE: 2685476492fSAlex Elder ceph_osd_data_release(&op->extent.osd_data); 2695476492fSAlex Elder break; 2705476492fSAlex Elder case CEPH_OSD_OP_CALL: 2715476492fSAlex Elder ceph_osd_data_release(&op->cls.request_info); 27204017e29SAlex Elder ceph_osd_data_release(&op->cls.request_data); 2735476492fSAlex Elder ceph_osd_data_release(&op->cls.response_data); 2745476492fSAlex Elder break; 2755476492fSAlex Elder default: 2765476492fSAlex Elder break; 2775476492fSAlex Elder } 278c54d47bfSAlex Elder } 279c54d47bfSAlex Elder 2803d14c5d2SYehuda Sadeh /* 2813d14c5d2SYehuda Sadeh * requests 2823d14c5d2SYehuda Sadeh */ 2833d14c5d2SYehuda Sadeh void ceph_osdc_release_request(struct kref *kref) 2843d14c5d2SYehuda Sadeh { 285c54d47bfSAlex Elder struct ceph_osd_request *req; 2865476492fSAlex Elder unsigned int which; 2873d14c5d2SYehuda Sadeh 288c54d47bfSAlex Elder req = container_of(kref, struct ceph_osd_request, r_kref); 2893d14c5d2SYehuda Sadeh if (req->r_request) 2903d14c5d2SYehuda Sadeh ceph_msg_put(req->r_request); 291ace6d3a9SAlex Elder if (req->r_reply) { 2928921d114SAlex Elder ceph_msg_revoke_incoming(req->r_reply); 293ab8cb34aSAlex Elder ceph_msg_put(req->r_reply); 294ace6d3a9SAlex Elder } 2950fff87ecSAlex Elder 2965476492fSAlex Elder for (which = 0; which < req->r_num_ops; which++) 2975476492fSAlex Elder osd_req_op_data_release(req, which); 2980fff87ecSAlex Elder 2993d14c5d2SYehuda Sadeh ceph_put_snap_context(req->r_snapc); 3003d14c5d2SYehuda Sadeh if (req->r_mempool) 3013d14c5d2SYehuda Sadeh mempool_free(req, req->r_osdc->req_mempool); 3023d14c5d2SYehuda Sadeh else 3033d14c5d2SYehuda Sadeh kfree(req); 3043d14c5d2SYehuda Sadeh } 3053d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_release_request); 3063d14c5d2SYehuda Sadeh 3073d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc, 3083d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 3091b83bef2SSage Weil unsigned int num_ops, 3103d14c5d2SYehuda Sadeh bool use_mempool, 31154a54007SAlex Elder gfp_t gfp_flags) 3123d14c5d2SYehuda Sadeh { 3133d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 3143d14c5d2SYehuda Sadeh struct ceph_msg *msg; 3151b83bef2SSage Weil size_t msg_size; 3163d14c5d2SYehuda Sadeh 31779528734SAlex Elder BUILD_BUG_ON(CEPH_OSD_MAX_OP > U16_MAX); 31879528734SAlex Elder BUG_ON(num_ops > CEPH_OSD_MAX_OP); 31979528734SAlex Elder 3201b83bef2SSage Weil msg_size = 4 + 4 + 8 + 8 + 4+8; 3211b83bef2SSage Weil msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */ 3221b83bef2SSage Weil msg_size += 1 + 8 + 4 + 4; /* pg_t */ 3231b83bef2SSage Weil msg_size += 4 + MAX_OBJ_NAME_SIZE; 3241b83bef2SSage Weil msg_size += 2 + num_ops*sizeof(struct ceph_osd_op); 3251b83bef2SSage Weil msg_size += 8; /* snapid */ 3261b83bef2SSage Weil msg_size += 8; /* snap_seq */ 3271b83bef2SSage Weil msg_size += 8 * (snapc ? snapc->num_snaps : 0); /* snaps */ 3281b83bef2SSage Weil msg_size += 4; 3293d14c5d2SYehuda Sadeh 3303d14c5d2SYehuda Sadeh if (use_mempool) { 3313d14c5d2SYehuda Sadeh req = mempool_alloc(osdc->req_mempool, gfp_flags); 3323d14c5d2SYehuda Sadeh memset(req, 0, sizeof(*req)); 3333d14c5d2SYehuda Sadeh } else { 3343d14c5d2SYehuda Sadeh req = kzalloc(sizeof(*req), gfp_flags); 3353d14c5d2SYehuda Sadeh } 3363d14c5d2SYehuda Sadeh if (req == NULL) 3373d14c5d2SYehuda Sadeh return NULL; 3383d14c5d2SYehuda Sadeh 3393d14c5d2SYehuda Sadeh req->r_osdc = osdc; 3403d14c5d2SYehuda Sadeh req->r_mempool = use_mempool; 34179528734SAlex Elder req->r_num_ops = num_ops; 3423d14c5d2SYehuda Sadeh 3433d14c5d2SYehuda Sadeh kref_init(&req->r_kref); 3443d14c5d2SYehuda Sadeh init_completion(&req->r_completion); 3453d14c5d2SYehuda Sadeh init_completion(&req->r_safe_completion); 346a978fa20SAlex Elder RB_CLEAR_NODE(&req->r_node); 3473d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&req->r_unsafe_item); 348a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&req->r_linger_item); 349a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&req->r_linger_osd); 350935b639aSSage Weil INIT_LIST_HEAD(&req->r_req_lru_item); 351cd43045cSSage Weil INIT_LIST_HEAD(&req->r_osd_item); 352cd43045cSSage Weil 3533d14c5d2SYehuda Sadeh /* create reply message */ 3543d14c5d2SYehuda Sadeh if (use_mempool) 3553d14c5d2SYehuda Sadeh msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0); 3563d14c5d2SYehuda Sadeh else 3573d14c5d2SYehuda Sadeh msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, 358b61c2763SSage Weil OSD_OPREPLY_FRONT_LEN, gfp_flags, true); 3593d14c5d2SYehuda Sadeh if (!msg) { 3603d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 3613d14c5d2SYehuda Sadeh return NULL; 3623d14c5d2SYehuda Sadeh } 3633d14c5d2SYehuda Sadeh req->r_reply = msg; 3643d14c5d2SYehuda Sadeh 3653d14c5d2SYehuda Sadeh /* create request message; allow space for oid */ 3663d14c5d2SYehuda Sadeh if (use_mempool) 3673d14c5d2SYehuda Sadeh msg = ceph_msgpool_get(&osdc->msgpool_op, 0); 3683d14c5d2SYehuda Sadeh else 369b61c2763SSage Weil msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true); 3703d14c5d2SYehuda Sadeh if (!msg) { 3713d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 3723d14c5d2SYehuda Sadeh return NULL; 3733d14c5d2SYehuda Sadeh } 3743d14c5d2SYehuda Sadeh 3753d14c5d2SYehuda Sadeh memset(msg->front.iov_base, 0, msg->front.iov_len); 3763d14c5d2SYehuda Sadeh 3773d14c5d2SYehuda Sadeh req->r_request = msg; 3783d14c5d2SYehuda Sadeh 3793d14c5d2SYehuda Sadeh return req; 3803d14c5d2SYehuda Sadeh } 3813d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_alloc_request); 3823d14c5d2SYehuda Sadeh 383a8dd0a37SAlex Elder static bool osd_req_opcode_valid(u16 opcode) 384a8dd0a37SAlex Elder { 385a8dd0a37SAlex Elder switch (opcode) { 386a8dd0a37SAlex Elder case CEPH_OSD_OP_READ: 387a8dd0a37SAlex Elder case CEPH_OSD_OP_STAT: 388a8dd0a37SAlex Elder case CEPH_OSD_OP_MAPEXT: 389a8dd0a37SAlex Elder case CEPH_OSD_OP_MASKTRUNC: 390a8dd0a37SAlex Elder case CEPH_OSD_OP_SPARSE_READ: 391a8dd0a37SAlex Elder case CEPH_OSD_OP_NOTIFY: 392a8dd0a37SAlex Elder case CEPH_OSD_OP_NOTIFY_ACK: 393a8dd0a37SAlex Elder case CEPH_OSD_OP_ASSERT_VER: 394a8dd0a37SAlex Elder case CEPH_OSD_OP_WRITE: 395a8dd0a37SAlex Elder case CEPH_OSD_OP_WRITEFULL: 396a8dd0a37SAlex Elder case CEPH_OSD_OP_TRUNCATE: 397a8dd0a37SAlex Elder case CEPH_OSD_OP_ZERO: 398a8dd0a37SAlex Elder case CEPH_OSD_OP_DELETE: 399a8dd0a37SAlex Elder case CEPH_OSD_OP_APPEND: 400a8dd0a37SAlex Elder case CEPH_OSD_OP_STARTSYNC: 401a8dd0a37SAlex Elder case CEPH_OSD_OP_SETTRUNC: 402a8dd0a37SAlex Elder case CEPH_OSD_OP_TRIMTRUNC: 403a8dd0a37SAlex Elder case CEPH_OSD_OP_TMAPUP: 404a8dd0a37SAlex Elder case CEPH_OSD_OP_TMAPPUT: 405a8dd0a37SAlex Elder case CEPH_OSD_OP_TMAPGET: 406a8dd0a37SAlex Elder case CEPH_OSD_OP_CREATE: 407a8dd0a37SAlex Elder case CEPH_OSD_OP_ROLLBACK: 408a8dd0a37SAlex Elder case CEPH_OSD_OP_WATCH: 409a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPGETKEYS: 410a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPGETVALS: 411a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPGETHEADER: 412a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPGETVALSBYKEYS: 413a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPSETVALS: 414a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPSETHEADER: 415a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPCLEAR: 416a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPRMKEYS: 417a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAP_CMP: 418a8dd0a37SAlex Elder case CEPH_OSD_OP_CLONERANGE: 419a8dd0a37SAlex Elder case CEPH_OSD_OP_ASSERT_SRC_VERSION: 420a8dd0a37SAlex Elder case CEPH_OSD_OP_SRC_CMPXATTR: 421a8dd0a37SAlex Elder case CEPH_OSD_OP_GETXATTR: 422a8dd0a37SAlex Elder case CEPH_OSD_OP_GETXATTRS: 423a8dd0a37SAlex Elder case CEPH_OSD_OP_CMPXATTR: 424a8dd0a37SAlex Elder case CEPH_OSD_OP_SETXATTR: 425a8dd0a37SAlex Elder case CEPH_OSD_OP_SETXATTRS: 426a8dd0a37SAlex Elder case CEPH_OSD_OP_RESETXATTRS: 427a8dd0a37SAlex Elder case CEPH_OSD_OP_RMXATTR: 428a8dd0a37SAlex Elder case CEPH_OSD_OP_PULL: 429a8dd0a37SAlex Elder case CEPH_OSD_OP_PUSH: 430a8dd0a37SAlex Elder case CEPH_OSD_OP_BALANCEREADS: 431a8dd0a37SAlex Elder case CEPH_OSD_OP_UNBALANCEREADS: 432a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB: 433a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB_RESERVE: 434a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB_UNRESERVE: 435a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB_STOP: 436a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB_MAP: 437a8dd0a37SAlex Elder case CEPH_OSD_OP_WRLOCK: 438a8dd0a37SAlex Elder case CEPH_OSD_OP_WRUNLOCK: 439a8dd0a37SAlex Elder case CEPH_OSD_OP_RDLOCK: 440a8dd0a37SAlex Elder case CEPH_OSD_OP_RDUNLOCK: 441a8dd0a37SAlex Elder case CEPH_OSD_OP_UPLOCK: 442a8dd0a37SAlex Elder case CEPH_OSD_OP_DNLOCK: 443a8dd0a37SAlex Elder case CEPH_OSD_OP_CALL: 444a8dd0a37SAlex Elder case CEPH_OSD_OP_PGLS: 445a8dd0a37SAlex Elder case CEPH_OSD_OP_PGLS_FILTER: 446a8dd0a37SAlex Elder return true; 447a8dd0a37SAlex Elder default: 448a8dd0a37SAlex Elder return false; 449a8dd0a37SAlex Elder } 450a8dd0a37SAlex Elder } 451a8dd0a37SAlex Elder 45233803f33SAlex Elder /* 45333803f33SAlex Elder * This is an osd op init function for opcodes that have no data or 45433803f33SAlex Elder * other information associated with them. It also serves as a 45533803f33SAlex Elder * common init routine for all the other init functions, below. 45633803f33SAlex Elder */ 457c99d2d4aSAlex Elder static struct ceph_osd_req_op * 458c99d2d4aSAlex Elder osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which, 459c99d2d4aSAlex Elder u16 opcode) 46033803f33SAlex Elder { 461c99d2d4aSAlex Elder struct ceph_osd_req_op *op; 462c99d2d4aSAlex Elder 463c99d2d4aSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 46433803f33SAlex Elder BUG_ON(!osd_req_opcode_valid(opcode)); 46533803f33SAlex Elder 466c99d2d4aSAlex Elder op = &osd_req->r_ops[which]; 46733803f33SAlex Elder memset(op, 0, sizeof (*op)); 46833803f33SAlex Elder op->op = opcode; 469c99d2d4aSAlex Elder 470c99d2d4aSAlex Elder return op; 47133803f33SAlex Elder } 47233803f33SAlex Elder 473c99d2d4aSAlex Elder void osd_req_op_extent_init(struct ceph_osd_request *osd_req, 474c99d2d4aSAlex Elder unsigned int which, u16 opcode, 47533803f33SAlex Elder u64 offset, u64 length, 47633803f33SAlex Elder u64 truncate_size, u32 truncate_seq) 47733803f33SAlex Elder { 478c99d2d4aSAlex Elder struct ceph_osd_req_op *op = osd_req_op_init(osd_req, which, opcode); 47933803f33SAlex Elder size_t payload_len = 0; 48033803f33SAlex Elder 48133803f33SAlex Elder BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE); 48233803f33SAlex Elder 48333803f33SAlex Elder op->extent.offset = offset; 48433803f33SAlex Elder op->extent.length = length; 48533803f33SAlex Elder op->extent.truncate_size = truncate_size; 48633803f33SAlex Elder op->extent.truncate_seq = truncate_seq; 48733803f33SAlex Elder if (opcode == CEPH_OSD_OP_WRITE) 48833803f33SAlex Elder payload_len += length; 48933803f33SAlex Elder 49033803f33SAlex Elder op->payload_len = payload_len; 49133803f33SAlex Elder } 49233803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_init); 49333803f33SAlex Elder 494c99d2d4aSAlex Elder void osd_req_op_extent_update(struct ceph_osd_request *osd_req, 495c99d2d4aSAlex Elder unsigned int which, u64 length) 496e5975c7cSAlex Elder { 497c99d2d4aSAlex Elder struct ceph_osd_req_op *op; 498c99d2d4aSAlex Elder u64 previous; 499c99d2d4aSAlex Elder 500c99d2d4aSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 501c99d2d4aSAlex Elder op = &osd_req->r_ops[which]; 502c99d2d4aSAlex Elder previous = op->extent.length; 503e5975c7cSAlex Elder 504e5975c7cSAlex Elder if (length == previous) 505e5975c7cSAlex Elder return; /* Nothing to do */ 506e5975c7cSAlex Elder BUG_ON(length > previous); 507e5975c7cSAlex Elder 508e5975c7cSAlex Elder op->extent.length = length; 509e5975c7cSAlex Elder op->payload_len -= previous - length; 510e5975c7cSAlex Elder } 511e5975c7cSAlex Elder EXPORT_SYMBOL(osd_req_op_extent_update); 512e5975c7cSAlex Elder 513c99d2d4aSAlex Elder void osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which, 51404017e29SAlex Elder u16 opcode, const char *class, const char *method) 51533803f33SAlex Elder { 516c99d2d4aSAlex Elder struct ceph_osd_req_op *op = osd_req_op_init(osd_req, which, opcode); 5175f562df5SAlex Elder struct ceph_pagelist *pagelist; 51833803f33SAlex Elder size_t payload_len = 0; 51933803f33SAlex Elder size_t size; 52033803f33SAlex Elder 52133803f33SAlex Elder BUG_ON(opcode != CEPH_OSD_OP_CALL); 52233803f33SAlex Elder 5235f562df5SAlex Elder pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS); 5245f562df5SAlex Elder BUG_ON(!pagelist); 5255f562df5SAlex Elder ceph_pagelist_init(pagelist); 5265f562df5SAlex Elder 52733803f33SAlex Elder op->cls.class_name = class; 52833803f33SAlex Elder size = strlen(class); 52933803f33SAlex Elder BUG_ON(size > (size_t) U8_MAX); 53033803f33SAlex Elder op->cls.class_len = size; 5315f562df5SAlex Elder ceph_pagelist_append(pagelist, class, size); 53233803f33SAlex Elder payload_len += size; 53333803f33SAlex Elder 53433803f33SAlex Elder op->cls.method_name = method; 53533803f33SAlex Elder size = strlen(method); 53633803f33SAlex Elder BUG_ON(size > (size_t) U8_MAX); 53733803f33SAlex Elder op->cls.method_len = size; 5385f562df5SAlex Elder ceph_pagelist_append(pagelist, method, size); 53933803f33SAlex Elder payload_len += size; 54033803f33SAlex Elder 541a4ce40a9SAlex Elder osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist); 5425f562df5SAlex Elder 54333803f33SAlex Elder op->cls.argc = 0; /* currently unused */ 54433803f33SAlex Elder 54533803f33SAlex Elder op->payload_len = payload_len; 54633803f33SAlex Elder } 54733803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_init); 5488c042b0dSAlex Elder 549c99d2d4aSAlex Elder void osd_req_op_watch_init(struct ceph_osd_request *osd_req, 550c99d2d4aSAlex Elder unsigned int which, u16 opcode, 55133803f33SAlex Elder u64 cookie, u64 version, int flag) 55233803f33SAlex Elder { 553c99d2d4aSAlex Elder struct ceph_osd_req_op *op = osd_req_op_init(osd_req, which, opcode); 55433803f33SAlex Elder 555c99d2d4aSAlex Elder BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH); 55633803f33SAlex Elder 55733803f33SAlex Elder op->watch.cookie = cookie; 55833803f33SAlex Elder /* op->watch.ver = version; */ /* XXX 3847 */ 55933803f33SAlex Elder op->watch.ver = cpu_to_le64(version); 56033803f33SAlex Elder if (opcode == CEPH_OSD_OP_WATCH && flag) 56133803f33SAlex Elder op->watch.flag = (u8)1; 56233803f33SAlex Elder } 56333803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_watch_init); 56433803f33SAlex Elder 56590af3602SAlex Elder static void ceph_osdc_msg_data_add(struct ceph_msg *msg, 566ec9123c5SAlex Elder struct ceph_osd_data *osd_data) 567ec9123c5SAlex Elder { 568ec9123c5SAlex Elder u64 length = ceph_osd_data_length(osd_data); 569ec9123c5SAlex Elder 570ec9123c5SAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 571ec9123c5SAlex Elder BUG_ON(length > (u64) SIZE_MAX); 572ec9123c5SAlex Elder if (length) 57390af3602SAlex Elder ceph_msg_data_add_pages(msg, osd_data->pages, 574ec9123c5SAlex Elder length, osd_data->alignment); 575ec9123c5SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) { 576ec9123c5SAlex Elder BUG_ON(!length); 57790af3602SAlex Elder ceph_msg_data_add_pagelist(msg, osd_data->pagelist); 578ec9123c5SAlex Elder #ifdef CONFIG_BLOCK 579ec9123c5SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) { 58090af3602SAlex Elder ceph_msg_data_add_bio(msg, osd_data->bio, length); 581ec9123c5SAlex Elder #endif 582ec9123c5SAlex Elder } else { 583ec9123c5SAlex Elder BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE); 584ec9123c5SAlex Elder } 585ec9123c5SAlex Elder } 586ec9123c5SAlex Elder 587175face2SAlex Elder static u64 osd_req_encode_op(struct ceph_osd_request *req, 58879528734SAlex Elder struct ceph_osd_op *dst, unsigned int which) 5893d14c5d2SYehuda Sadeh { 59079528734SAlex Elder struct ceph_osd_req_op *src; 59104017e29SAlex Elder struct ceph_osd_data *osd_data; 59254d50649SAlex Elder u64 request_data_len = 0; 59304017e29SAlex Elder u64 data_length; 594175face2SAlex Elder 59579528734SAlex Elder BUG_ON(which >= req->r_num_ops); 59679528734SAlex Elder src = &req->r_ops[which]; 597a8dd0a37SAlex Elder if (WARN_ON(!osd_req_opcode_valid(src->op))) { 598a8dd0a37SAlex Elder pr_err("unrecognized osd opcode %d\n", src->op); 599a8dd0a37SAlex Elder 600a8dd0a37SAlex Elder return 0; 601a8dd0a37SAlex Elder } 6023d14c5d2SYehuda Sadeh 603065a68f9SAlex Elder switch (src->op) { 604fbfab539SAlex Elder case CEPH_OSD_OP_STAT: 605fbfab539SAlex Elder break; 6063d14c5d2SYehuda Sadeh case CEPH_OSD_OP_READ: 6073d14c5d2SYehuda Sadeh case CEPH_OSD_OP_WRITE: 608175face2SAlex Elder if (src->op == CEPH_OSD_OP_WRITE) 60954d50649SAlex Elder request_data_len = src->extent.length; 610175face2SAlex Elder dst->extent.offset = cpu_to_le64(src->extent.offset); 611175face2SAlex Elder dst->extent.length = cpu_to_le64(src->extent.length); 6123d14c5d2SYehuda Sadeh dst->extent.truncate_size = 6133d14c5d2SYehuda Sadeh cpu_to_le64(src->extent.truncate_size); 6143d14c5d2SYehuda Sadeh dst->extent.truncate_seq = 6153d14c5d2SYehuda Sadeh cpu_to_le32(src->extent.truncate_seq); 61604017e29SAlex Elder osd_data = &src->extent.osd_data; 6175476492fSAlex Elder if (src->op == CEPH_OSD_OP_WRITE) 61804017e29SAlex Elder ceph_osdc_msg_data_add(req->r_request, osd_data); 6195476492fSAlex Elder else 62004017e29SAlex Elder ceph_osdc_msg_data_add(req->r_reply, osd_data); 6213d14c5d2SYehuda Sadeh break; 6223d14c5d2SYehuda Sadeh case CEPH_OSD_OP_CALL: 6233d14c5d2SYehuda Sadeh dst->cls.class_len = src->cls.class_len; 6243d14c5d2SYehuda Sadeh dst->cls.method_len = src->cls.method_len; 62504017e29SAlex Elder osd_data = &src->cls.request_info; 62604017e29SAlex Elder ceph_osdc_msg_data_add(req->r_request, osd_data); 62704017e29SAlex Elder BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGELIST); 62804017e29SAlex Elder request_data_len = osd_data->pagelist->length; 62904017e29SAlex Elder 63004017e29SAlex Elder osd_data = &src->cls.request_data; 63104017e29SAlex Elder data_length = ceph_osd_data_length(osd_data); 63204017e29SAlex Elder if (data_length) { 63304017e29SAlex Elder BUG_ON(osd_data->type == CEPH_OSD_DATA_TYPE_NONE); 63404017e29SAlex Elder dst->cls.indata_len = cpu_to_le32(data_length); 63504017e29SAlex Elder ceph_osdc_msg_data_add(req->r_request, osd_data); 63604017e29SAlex Elder src->payload_len += data_length; 63704017e29SAlex Elder request_data_len += data_length; 63804017e29SAlex Elder } 63904017e29SAlex Elder osd_data = &src->cls.response_data; 64004017e29SAlex Elder ceph_osdc_msg_data_add(req->r_reply, osd_data); 6413d14c5d2SYehuda Sadeh break; 6423d14c5d2SYehuda Sadeh case CEPH_OSD_OP_STARTSYNC: 6433d14c5d2SYehuda Sadeh break; 644a40c4f10SYehuda Sadeh case CEPH_OSD_OP_NOTIFY_ACK: 645a40c4f10SYehuda Sadeh case CEPH_OSD_OP_WATCH: 646a40c4f10SYehuda Sadeh dst->watch.cookie = cpu_to_le64(src->watch.cookie); 647a40c4f10SYehuda Sadeh dst->watch.ver = cpu_to_le64(src->watch.ver); 648a40c4f10SYehuda Sadeh dst->watch.flag = src->watch.flag; 649a40c4f10SYehuda Sadeh break; 6503d14c5d2SYehuda Sadeh default: 6514c46459cSAlex Elder pr_err("unsupported osd opcode %s\n", 6528f63ca2dSAlex Elder ceph_osd_op_name(src->op)); 6534c46459cSAlex Elder WARN_ON(1); 654a8dd0a37SAlex Elder 655a8dd0a37SAlex Elder return 0; 6563d14c5d2SYehuda Sadeh } 657a8dd0a37SAlex Elder dst->op = cpu_to_le16(src->op); 6583d14c5d2SYehuda Sadeh dst->payload_len = cpu_to_le32(src->payload_len); 659175face2SAlex Elder 66054d50649SAlex Elder return request_data_len; 6613d14c5d2SYehuda Sadeh } 6623d14c5d2SYehuda Sadeh 6633d14c5d2SYehuda Sadeh /* 6643d14c5d2SYehuda Sadeh * build new request AND message, calculate layout, and adjust file 6653d14c5d2SYehuda Sadeh * extent as needed. 6663d14c5d2SYehuda Sadeh * 6673d14c5d2SYehuda Sadeh * if the file was recently truncated, we include information about its 6683d14c5d2SYehuda Sadeh * old and new size so that the object can be updated appropriately. (we 6693d14c5d2SYehuda Sadeh * avoid synchronously deleting truncated objects because it's slow.) 6703d14c5d2SYehuda Sadeh * 6713d14c5d2SYehuda Sadeh * if @do_sync, include a 'startsync' command so that the osd will flush 6723d14c5d2SYehuda Sadeh * data quickly. 6733d14c5d2SYehuda Sadeh */ 6743d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc, 6753d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 6763d14c5d2SYehuda Sadeh struct ceph_vino vino, 677acead002SAlex Elder u64 off, u64 *plen, int num_ops, 6783d14c5d2SYehuda Sadeh int opcode, int flags, 6793d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 6803d14c5d2SYehuda Sadeh u32 truncate_seq, 6813d14c5d2SYehuda Sadeh u64 truncate_size, 682153e5167SAlex Elder bool use_mempool) 6833d14c5d2SYehuda Sadeh { 6843d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 68575d1c941SAlex Elder u64 objnum = 0; 68675d1c941SAlex Elder u64 objoff = 0; 68775d1c941SAlex Elder u64 objlen = 0; 688d18d1e28SAlex Elder u32 object_size; 689d18d1e28SAlex Elder u64 object_base; 6906816282dSSage Weil int r; 6913d14c5d2SYehuda Sadeh 692d18d1e28SAlex Elder BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE); 6933d14c5d2SYehuda Sadeh 694acead002SAlex Elder req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool, 695ae7ca4a3SAlex Elder GFP_NOFS); 6964ad12621SSage Weil if (!req) 6976816282dSSage Weil return ERR_PTR(-ENOMEM); 69879528734SAlex Elder 699d178a9e7SAlex Elder req->r_flags = flags; 7003d14c5d2SYehuda Sadeh 7013d14c5d2SYehuda Sadeh /* calculate max write size */ 702a19dadfbSAlex Elder r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen); 7033ff5f385SAlex Elder if (r < 0) { 7043ff5f385SAlex Elder ceph_osdc_put_request(req); 7056816282dSSage Weil return ERR_PTR(r); 7063ff5f385SAlex Elder } 707a19dadfbSAlex Elder 708d18d1e28SAlex Elder object_size = le32_to_cpu(layout->fl_object_size); 709d18d1e28SAlex Elder object_base = off - objoff; 710d18d1e28SAlex Elder if (truncate_size <= object_base) { 711d18d1e28SAlex Elder truncate_size = 0; 712d18d1e28SAlex Elder } else { 713d18d1e28SAlex Elder truncate_size -= object_base; 714d18d1e28SAlex Elder if (truncate_size > object_size) 715d18d1e28SAlex Elder truncate_size = object_size; 716d18d1e28SAlex Elder } 717d18d1e28SAlex Elder 718c99d2d4aSAlex Elder osd_req_op_extent_init(req, 0, opcode, objoff, objlen, 719b0270324SAlex Elder truncate_size, truncate_seq); 7208c042b0dSAlex Elder 721acead002SAlex Elder /* 722acead002SAlex Elder * A second op in the ops array means the caller wants to 723acead002SAlex Elder * also issue a include a 'startsync' command so that the 724acead002SAlex Elder * osd will flush data quickly. 725acead002SAlex Elder */ 726acead002SAlex Elder if (num_ops > 1) 727c99d2d4aSAlex Elder osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC); 728d18d1e28SAlex Elder 7293d14c5d2SYehuda Sadeh req->r_file_layout = *layout; /* keep a copy */ 7303d14c5d2SYehuda Sadeh 73175d1c941SAlex Elder snprintf(req->r_oid, sizeof(req->r_oid), "%llx.%08llx", 73275d1c941SAlex Elder vino.ino, objnum); 733dbe0fc41SAlex Elder req->r_oid_len = strlen(req->r_oid); 734dbe0fc41SAlex Elder 7353d14c5d2SYehuda Sadeh return req; 7363d14c5d2SYehuda Sadeh } 7373d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_new_request); 7383d14c5d2SYehuda Sadeh 7393d14c5d2SYehuda Sadeh /* 7403d14c5d2SYehuda Sadeh * We keep osd requests in an rbtree, sorted by ->r_tid. 7413d14c5d2SYehuda Sadeh */ 7423d14c5d2SYehuda Sadeh static void __insert_request(struct ceph_osd_client *osdc, 7433d14c5d2SYehuda Sadeh struct ceph_osd_request *new) 7443d14c5d2SYehuda Sadeh { 7453d14c5d2SYehuda Sadeh struct rb_node **p = &osdc->requests.rb_node; 7463d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 7473d14c5d2SYehuda Sadeh struct ceph_osd_request *req = NULL; 7483d14c5d2SYehuda Sadeh 7493d14c5d2SYehuda Sadeh while (*p) { 7503d14c5d2SYehuda Sadeh parent = *p; 7513d14c5d2SYehuda Sadeh req = rb_entry(parent, struct ceph_osd_request, r_node); 7523d14c5d2SYehuda Sadeh if (new->r_tid < req->r_tid) 7533d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 7543d14c5d2SYehuda Sadeh else if (new->r_tid > req->r_tid) 7553d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 7563d14c5d2SYehuda Sadeh else 7573d14c5d2SYehuda Sadeh BUG(); 7583d14c5d2SYehuda Sadeh } 7593d14c5d2SYehuda Sadeh 7603d14c5d2SYehuda Sadeh rb_link_node(&new->r_node, parent, p); 7613d14c5d2SYehuda Sadeh rb_insert_color(&new->r_node, &osdc->requests); 7623d14c5d2SYehuda Sadeh } 7633d14c5d2SYehuda Sadeh 7643d14c5d2SYehuda Sadeh static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc, 7653d14c5d2SYehuda Sadeh u64 tid) 7663d14c5d2SYehuda Sadeh { 7673d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 7683d14c5d2SYehuda Sadeh struct rb_node *n = osdc->requests.rb_node; 7693d14c5d2SYehuda Sadeh 7703d14c5d2SYehuda Sadeh while (n) { 7713d14c5d2SYehuda Sadeh req = rb_entry(n, struct ceph_osd_request, r_node); 7723d14c5d2SYehuda Sadeh if (tid < req->r_tid) 7733d14c5d2SYehuda Sadeh n = n->rb_left; 7743d14c5d2SYehuda Sadeh else if (tid > req->r_tid) 7753d14c5d2SYehuda Sadeh n = n->rb_right; 7763d14c5d2SYehuda Sadeh else 7773d14c5d2SYehuda Sadeh return req; 7783d14c5d2SYehuda Sadeh } 7793d14c5d2SYehuda Sadeh return NULL; 7803d14c5d2SYehuda Sadeh } 7813d14c5d2SYehuda Sadeh 7823d14c5d2SYehuda Sadeh static struct ceph_osd_request * 7833d14c5d2SYehuda Sadeh __lookup_request_ge(struct ceph_osd_client *osdc, 7843d14c5d2SYehuda Sadeh u64 tid) 7853d14c5d2SYehuda Sadeh { 7863d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 7873d14c5d2SYehuda Sadeh struct rb_node *n = osdc->requests.rb_node; 7883d14c5d2SYehuda Sadeh 7893d14c5d2SYehuda Sadeh while (n) { 7903d14c5d2SYehuda Sadeh req = rb_entry(n, struct ceph_osd_request, r_node); 7913d14c5d2SYehuda Sadeh if (tid < req->r_tid) { 7923d14c5d2SYehuda Sadeh if (!n->rb_left) 7933d14c5d2SYehuda Sadeh return req; 7943d14c5d2SYehuda Sadeh n = n->rb_left; 7953d14c5d2SYehuda Sadeh } else if (tid > req->r_tid) { 7963d14c5d2SYehuda Sadeh n = n->rb_right; 7973d14c5d2SYehuda Sadeh } else { 7983d14c5d2SYehuda Sadeh return req; 7993d14c5d2SYehuda Sadeh } 8003d14c5d2SYehuda Sadeh } 8013d14c5d2SYehuda Sadeh return NULL; 8023d14c5d2SYehuda Sadeh } 8033d14c5d2SYehuda Sadeh 8046f6c7006SSage Weil /* 8056f6c7006SSage Weil * Resubmit requests pending on the given osd. 8066f6c7006SSage Weil */ 8076f6c7006SSage Weil static void __kick_osd_requests(struct ceph_osd_client *osdc, 8086f6c7006SSage Weil struct ceph_osd *osd) 8096f6c7006SSage Weil { 810a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 811e02493c0SAlex Elder LIST_HEAD(resend); 8126f6c7006SSage Weil int err; 8136f6c7006SSage Weil 8146f6c7006SSage Weil dout("__kick_osd_requests osd%d\n", osd->o_osd); 8156f6c7006SSage Weil err = __reset_osd(osdc, osd); 816685a7555SAlex Elder if (err) 8176f6c7006SSage Weil return; 818e02493c0SAlex Elder /* 819e02493c0SAlex Elder * Build up a list of requests to resend by traversing the 820e02493c0SAlex Elder * osd's list of requests. Requests for a given object are 821e02493c0SAlex Elder * sent in tid order, and that is also the order they're 822e02493c0SAlex Elder * kept on this list. Therefore all requests that are in 823e02493c0SAlex Elder * flight will be found first, followed by all requests that 824e02493c0SAlex Elder * have not yet been sent. And to resend requests while 825e02493c0SAlex Elder * preserving this order we will want to put any sent 826e02493c0SAlex Elder * requests back on the front of the osd client's unsent 827e02493c0SAlex Elder * list. 828e02493c0SAlex Elder * 829e02493c0SAlex Elder * So we build a separate ordered list of already-sent 830e02493c0SAlex Elder * requests for the affected osd and splice it onto the 831e02493c0SAlex Elder * front of the osd client's unsent list. Once we've seen a 832e02493c0SAlex Elder * request that has not yet been sent we're done. Those 833e02493c0SAlex Elder * requests are already sitting right where they belong. 834e02493c0SAlex Elder */ 8356f6c7006SSage Weil list_for_each_entry(req, &osd->o_requests, r_osd_item) { 836e02493c0SAlex Elder if (!req->r_sent) 837e02493c0SAlex Elder break; 838e02493c0SAlex Elder list_move_tail(&req->r_req_lru_item, &resend); 839e02493c0SAlex Elder dout("requeueing %p tid %llu osd%d\n", req, req->r_tid, 8406f6c7006SSage Weil osd->o_osd); 841a40c4f10SYehuda Sadeh if (!req->r_linger) 8426f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 8436f6c7006SSage Weil } 844e02493c0SAlex Elder list_splice(&resend, &osdc->req_unsent); 845a40c4f10SYehuda Sadeh 846e02493c0SAlex Elder /* 847e02493c0SAlex Elder * Linger requests are re-registered before sending, which 848e02493c0SAlex Elder * sets up a new tid for each. We add them to the unsent 849e02493c0SAlex Elder * list at the end to keep things in tid order. 850e02493c0SAlex Elder */ 851a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osd->o_linger_requests, 852a40c4f10SYehuda Sadeh r_linger_osd) { 85377f38e0eSSage Weil /* 85477f38e0eSSage Weil * reregister request prior to unregistering linger so 85577f38e0eSSage Weil * that r_osd is preserved. 85677f38e0eSSage Weil */ 85777f38e0eSSage Weil BUG_ON(!list_empty(&req->r_req_lru_item)); 858a40c4f10SYehuda Sadeh __register_request(osdc, req); 859e02493c0SAlex Elder list_add_tail(&req->r_req_lru_item, &osdc->req_unsent); 860ad885927SAlex Elder list_add_tail(&req->r_osd_item, &req->r_osd->o_requests); 86177f38e0eSSage Weil __unregister_linger_request(osdc, req); 862a40c4f10SYehuda Sadeh dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid, 863a40c4f10SYehuda Sadeh osd->o_osd); 864a40c4f10SYehuda Sadeh } 8656f6c7006SSage Weil } 8666f6c7006SSage Weil 8673d14c5d2SYehuda Sadeh /* 8683d14c5d2SYehuda Sadeh * If the osd connection drops, we need to resubmit all requests. 8693d14c5d2SYehuda Sadeh */ 8703d14c5d2SYehuda Sadeh static void osd_reset(struct ceph_connection *con) 8713d14c5d2SYehuda Sadeh { 8723d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 8733d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 8743d14c5d2SYehuda Sadeh 8753d14c5d2SYehuda Sadeh if (!osd) 8763d14c5d2SYehuda Sadeh return; 8773d14c5d2SYehuda Sadeh dout("osd_reset osd%d\n", osd->o_osd); 8783d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 8793d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 88083aff95eSSage Weil mutex_lock(&osdc->request_mutex); 88183aff95eSSage Weil __kick_osd_requests(osdc, osd); 882f9d25199SAlex Elder __send_queued(osdc); 88383aff95eSSage Weil mutex_unlock(&osdc->request_mutex); 8843d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 8853d14c5d2SYehuda Sadeh } 8863d14c5d2SYehuda Sadeh 8873d14c5d2SYehuda Sadeh /* 8883d14c5d2SYehuda Sadeh * Track open sessions with osds. 8893d14c5d2SYehuda Sadeh */ 890e10006f8SAlex Elder static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum) 8913d14c5d2SYehuda Sadeh { 8923d14c5d2SYehuda Sadeh struct ceph_osd *osd; 8933d14c5d2SYehuda Sadeh 8943d14c5d2SYehuda Sadeh osd = kzalloc(sizeof(*osd), GFP_NOFS); 8953d14c5d2SYehuda Sadeh if (!osd) 8963d14c5d2SYehuda Sadeh return NULL; 8973d14c5d2SYehuda Sadeh 8983d14c5d2SYehuda Sadeh atomic_set(&osd->o_ref, 1); 8993d14c5d2SYehuda Sadeh osd->o_osdc = osdc; 900e10006f8SAlex Elder osd->o_osd = onum; 901f407731dSAlex Elder RB_CLEAR_NODE(&osd->o_node); 9023d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_requests); 903a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osd->o_linger_requests); 9043d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_osd_lru); 9053d14c5d2SYehuda Sadeh osd->o_incarnation = 1; 9063d14c5d2SYehuda Sadeh 907b7a9e5ddSSage Weil ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr); 9083d14c5d2SYehuda Sadeh 9093d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_keepalive_item); 9103d14c5d2SYehuda Sadeh return osd; 9113d14c5d2SYehuda Sadeh } 9123d14c5d2SYehuda Sadeh 9133d14c5d2SYehuda Sadeh static struct ceph_osd *get_osd(struct ceph_osd *osd) 9143d14c5d2SYehuda Sadeh { 9153d14c5d2SYehuda Sadeh if (atomic_inc_not_zero(&osd->o_ref)) { 9163d14c5d2SYehuda Sadeh dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1, 9173d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref)); 9183d14c5d2SYehuda Sadeh return osd; 9193d14c5d2SYehuda Sadeh } else { 9203d14c5d2SYehuda Sadeh dout("get_osd %p FAIL\n", osd); 9213d14c5d2SYehuda Sadeh return NULL; 9223d14c5d2SYehuda Sadeh } 9233d14c5d2SYehuda Sadeh } 9243d14c5d2SYehuda Sadeh 9253d14c5d2SYehuda Sadeh static void put_osd(struct ceph_osd *osd) 9263d14c5d2SYehuda Sadeh { 9273d14c5d2SYehuda Sadeh dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref), 9283d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref) - 1); 929a255651dSAlex Elder if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) { 9303d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth; 9313d14c5d2SYehuda Sadeh 93227859f97SSage Weil ceph_auth_destroy_authorizer(ac, osd->o_auth.authorizer); 9333d14c5d2SYehuda Sadeh kfree(osd); 9343d14c5d2SYehuda Sadeh } 9353d14c5d2SYehuda Sadeh } 9363d14c5d2SYehuda Sadeh 9373d14c5d2SYehuda Sadeh /* 9383d14c5d2SYehuda Sadeh * remove an osd from our map 9393d14c5d2SYehuda Sadeh */ 9403d14c5d2SYehuda Sadeh static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 9413d14c5d2SYehuda Sadeh { 9423d14c5d2SYehuda Sadeh dout("__remove_osd %p\n", osd); 9433d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_requests)); 9443d14c5d2SYehuda Sadeh rb_erase(&osd->o_node, &osdc->osds); 9453d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 9463d14c5d2SYehuda Sadeh ceph_con_close(&osd->o_con); 9473d14c5d2SYehuda Sadeh put_osd(osd); 9483d14c5d2SYehuda Sadeh } 9493d14c5d2SYehuda Sadeh 950aca420bcSSage Weil static void remove_all_osds(struct ceph_osd_client *osdc) 951aca420bcSSage Weil { 952048a9d2dSJiaju Zhang dout("%s %p\n", __func__, osdc); 953aca420bcSSage Weil mutex_lock(&osdc->request_mutex); 954aca420bcSSage Weil while (!RB_EMPTY_ROOT(&osdc->osds)) { 955aca420bcSSage Weil struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds), 956aca420bcSSage Weil struct ceph_osd, o_node); 957aca420bcSSage Weil __remove_osd(osdc, osd); 958aca420bcSSage Weil } 959aca420bcSSage Weil mutex_unlock(&osdc->request_mutex); 960aca420bcSSage Weil } 961aca420bcSSage Weil 9623d14c5d2SYehuda Sadeh static void __move_osd_to_lru(struct ceph_osd_client *osdc, 9633d14c5d2SYehuda Sadeh struct ceph_osd *osd) 9643d14c5d2SYehuda Sadeh { 9653d14c5d2SYehuda Sadeh dout("__move_osd_to_lru %p\n", osd); 9663d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_osd_lru)); 9673d14c5d2SYehuda Sadeh list_add_tail(&osd->o_osd_lru, &osdc->osd_lru); 9683d14c5d2SYehuda Sadeh osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ; 9693d14c5d2SYehuda Sadeh } 9703d14c5d2SYehuda Sadeh 9713d14c5d2SYehuda Sadeh static void __remove_osd_from_lru(struct ceph_osd *osd) 9723d14c5d2SYehuda Sadeh { 9733d14c5d2SYehuda Sadeh dout("__remove_osd_from_lru %p\n", osd); 9743d14c5d2SYehuda Sadeh if (!list_empty(&osd->o_osd_lru)) 9753d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 9763d14c5d2SYehuda Sadeh } 9773d14c5d2SYehuda Sadeh 978aca420bcSSage Weil static void remove_old_osds(struct ceph_osd_client *osdc) 9793d14c5d2SYehuda Sadeh { 9803d14c5d2SYehuda Sadeh struct ceph_osd *osd, *nosd; 9813d14c5d2SYehuda Sadeh 9823d14c5d2SYehuda Sadeh dout("__remove_old_osds %p\n", osdc); 9833d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 9843d14c5d2SYehuda Sadeh list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) { 985aca420bcSSage Weil if (time_before(jiffies, osd->lru_ttl)) 9863d14c5d2SYehuda Sadeh break; 9873d14c5d2SYehuda Sadeh __remove_osd(osdc, osd); 9883d14c5d2SYehuda Sadeh } 9893d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 9903d14c5d2SYehuda Sadeh } 9913d14c5d2SYehuda Sadeh 9923d14c5d2SYehuda Sadeh /* 9933d14c5d2SYehuda Sadeh * reset osd connect 9943d14c5d2SYehuda Sadeh */ 9953d14c5d2SYehuda Sadeh static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 9963d14c5d2SYehuda Sadeh { 997c3acb181SAlex Elder struct ceph_entity_addr *peer_addr; 9983d14c5d2SYehuda Sadeh 9993d14c5d2SYehuda Sadeh dout("__reset_osd %p osd%d\n", osd, osd->o_osd); 1000a40c4f10SYehuda Sadeh if (list_empty(&osd->o_requests) && 1001a40c4f10SYehuda Sadeh list_empty(&osd->o_linger_requests)) { 10023d14c5d2SYehuda Sadeh __remove_osd(osdc, osd); 1003c3acb181SAlex Elder 1004c3acb181SAlex Elder return -ENODEV; 1005c3acb181SAlex Elder } 1006c3acb181SAlex Elder 1007c3acb181SAlex Elder peer_addr = &osdc->osdmap->osd_addr[osd->o_osd]; 1008c3acb181SAlex Elder if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) && 10093d14c5d2SYehuda Sadeh !ceph_con_opened(&osd->o_con)) { 1010c3acb181SAlex Elder struct ceph_osd_request *req; 1011c3acb181SAlex Elder 10123d14c5d2SYehuda Sadeh dout(" osd addr hasn't changed and connection never opened," 10133d14c5d2SYehuda Sadeh " letting msgr retry"); 10143d14c5d2SYehuda Sadeh /* touch each r_stamp for handle_timeout()'s benfit */ 10153d14c5d2SYehuda Sadeh list_for_each_entry(req, &osd->o_requests, r_osd_item) 10163d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 1017c3acb181SAlex Elder 1018c3acb181SAlex Elder return -EAGAIN; 10193d14c5d2SYehuda Sadeh } 1020c3acb181SAlex Elder 1021c3acb181SAlex Elder ceph_con_close(&osd->o_con); 1022c3acb181SAlex Elder ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr); 1023c3acb181SAlex Elder osd->o_incarnation++; 1024c3acb181SAlex Elder 1025c3acb181SAlex Elder return 0; 10263d14c5d2SYehuda Sadeh } 10273d14c5d2SYehuda Sadeh 10283d14c5d2SYehuda Sadeh static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new) 10293d14c5d2SYehuda Sadeh { 10303d14c5d2SYehuda Sadeh struct rb_node **p = &osdc->osds.rb_node; 10313d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 10323d14c5d2SYehuda Sadeh struct ceph_osd *osd = NULL; 10333d14c5d2SYehuda Sadeh 1034aca420bcSSage Weil dout("__insert_osd %p osd%d\n", new, new->o_osd); 10353d14c5d2SYehuda Sadeh while (*p) { 10363d14c5d2SYehuda Sadeh parent = *p; 10373d14c5d2SYehuda Sadeh osd = rb_entry(parent, struct ceph_osd, o_node); 10383d14c5d2SYehuda Sadeh if (new->o_osd < osd->o_osd) 10393d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 10403d14c5d2SYehuda Sadeh else if (new->o_osd > osd->o_osd) 10413d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 10423d14c5d2SYehuda Sadeh else 10433d14c5d2SYehuda Sadeh BUG(); 10443d14c5d2SYehuda Sadeh } 10453d14c5d2SYehuda Sadeh 10463d14c5d2SYehuda Sadeh rb_link_node(&new->o_node, parent, p); 10473d14c5d2SYehuda Sadeh rb_insert_color(&new->o_node, &osdc->osds); 10483d14c5d2SYehuda Sadeh } 10493d14c5d2SYehuda Sadeh 10503d14c5d2SYehuda Sadeh static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o) 10513d14c5d2SYehuda Sadeh { 10523d14c5d2SYehuda Sadeh struct ceph_osd *osd; 10533d14c5d2SYehuda Sadeh struct rb_node *n = osdc->osds.rb_node; 10543d14c5d2SYehuda Sadeh 10553d14c5d2SYehuda Sadeh while (n) { 10563d14c5d2SYehuda Sadeh osd = rb_entry(n, struct ceph_osd, o_node); 10573d14c5d2SYehuda Sadeh if (o < osd->o_osd) 10583d14c5d2SYehuda Sadeh n = n->rb_left; 10593d14c5d2SYehuda Sadeh else if (o > osd->o_osd) 10603d14c5d2SYehuda Sadeh n = n->rb_right; 10613d14c5d2SYehuda Sadeh else 10623d14c5d2SYehuda Sadeh return osd; 10633d14c5d2SYehuda Sadeh } 10643d14c5d2SYehuda Sadeh return NULL; 10653d14c5d2SYehuda Sadeh } 10663d14c5d2SYehuda Sadeh 10673d14c5d2SYehuda Sadeh static void __schedule_osd_timeout(struct ceph_osd_client *osdc) 10683d14c5d2SYehuda Sadeh { 10693d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->timeout_work, 10703d14c5d2SYehuda Sadeh osdc->client->options->osd_keepalive_timeout * HZ); 10713d14c5d2SYehuda Sadeh } 10723d14c5d2SYehuda Sadeh 10733d14c5d2SYehuda Sadeh static void __cancel_osd_timeout(struct ceph_osd_client *osdc) 10743d14c5d2SYehuda Sadeh { 10753d14c5d2SYehuda Sadeh cancel_delayed_work(&osdc->timeout_work); 10763d14c5d2SYehuda Sadeh } 10773d14c5d2SYehuda Sadeh 10783d14c5d2SYehuda Sadeh /* 10793d14c5d2SYehuda Sadeh * Register request, assign tid. If this is the first request, set up 10803d14c5d2SYehuda Sadeh * the timeout event. 10813d14c5d2SYehuda Sadeh */ 1082a40c4f10SYehuda Sadeh static void __register_request(struct ceph_osd_client *osdc, 10833d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 10843d14c5d2SYehuda Sadeh { 10853d14c5d2SYehuda Sadeh req->r_tid = ++osdc->last_tid; 10863d14c5d2SYehuda Sadeh req->r_request->hdr.tid = cpu_to_le64(req->r_tid); 108777f38e0eSSage Weil dout("__register_request %p tid %lld\n", req, req->r_tid); 10883d14c5d2SYehuda Sadeh __insert_request(osdc, req); 10893d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 10903d14c5d2SYehuda Sadeh osdc->num_requests++; 10913d14c5d2SYehuda Sadeh if (osdc->num_requests == 1) { 10923d14c5d2SYehuda Sadeh dout(" first request, scheduling timeout\n"); 10933d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 10943d14c5d2SYehuda Sadeh } 1095a40c4f10SYehuda Sadeh } 1096a40c4f10SYehuda Sadeh 10973d14c5d2SYehuda Sadeh /* 10983d14c5d2SYehuda Sadeh * called under osdc->request_mutex 10993d14c5d2SYehuda Sadeh */ 11003d14c5d2SYehuda Sadeh static void __unregister_request(struct ceph_osd_client *osdc, 11013d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 11023d14c5d2SYehuda Sadeh { 110335f9f8a0SSage Weil if (RB_EMPTY_NODE(&req->r_node)) { 110435f9f8a0SSage Weil dout("__unregister_request %p tid %lld not registered\n", 110535f9f8a0SSage Weil req, req->r_tid); 110635f9f8a0SSage Weil return; 110735f9f8a0SSage Weil } 110835f9f8a0SSage Weil 11093d14c5d2SYehuda Sadeh dout("__unregister_request %p tid %lld\n", req, req->r_tid); 11103d14c5d2SYehuda Sadeh rb_erase(&req->r_node, &osdc->requests); 11113d14c5d2SYehuda Sadeh osdc->num_requests--; 11123d14c5d2SYehuda Sadeh 11133d14c5d2SYehuda Sadeh if (req->r_osd) { 11143d14c5d2SYehuda Sadeh /* make sure the original request isn't in flight. */ 11156740a845SAlex Elder ceph_msg_revoke(req->r_request); 11163d14c5d2SYehuda Sadeh 11173d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 1118a40c4f10SYehuda Sadeh if (list_empty(&req->r_osd->o_requests) && 1119a40c4f10SYehuda Sadeh list_empty(&req->r_osd->o_linger_requests)) { 1120a40c4f10SYehuda Sadeh dout("moving osd to %p lru\n", req->r_osd); 11213d14c5d2SYehuda Sadeh __move_osd_to_lru(osdc, req->r_osd); 1122a40c4f10SYehuda Sadeh } 1123fbdb9190SSage Weil if (list_empty(&req->r_linger_item)) 11243d14c5d2SYehuda Sadeh req->r_osd = NULL; 11253d14c5d2SYehuda Sadeh } 11263d14c5d2SYehuda Sadeh 11277d5f2481SAlex Elder list_del_init(&req->r_req_lru_item); 11283d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 11293d14c5d2SYehuda Sadeh 11303d14c5d2SYehuda Sadeh if (osdc->num_requests == 0) { 11313d14c5d2SYehuda Sadeh dout(" no requests, canceling timeout\n"); 11323d14c5d2SYehuda Sadeh __cancel_osd_timeout(osdc); 11333d14c5d2SYehuda Sadeh } 11343d14c5d2SYehuda Sadeh } 11353d14c5d2SYehuda Sadeh 11363d14c5d2SYehuda Sadeh /* 11373d14c5d2SYehuda Sadeh * Cancel a previously queued request message 11383d14c5d2SYehuda Sadeh */ 11393d14c5d2SYehuda Sadeh static void __cancel_request(struct ceph_osd_request *req) 11403d14c5d2SYehuda Sadeh { 11413d14c5d2SYehuda Sadeh if (req->r_sent && req->r_osd) { 11426740a845SAlex Elder ceph_msg_revoke(req->r_request); 11433d14c5d2SYehuda Sadeh req->r_sent = 0; 11443d14c5d2SYehuda Sadeh } 11453d14c5d2SYehuda Sadeh } 11463d14c5d2SYehuda Sadeh 1147a40c4f10SYehuda Sadeh static void __register_linger_request(struct ceph_osd_client *osdc, 1148a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1149a40c4f10SYehuda Sadeh { 1150a40c4f10SYehuda Sadeh dout("__register_linger_request %p\n", req); 1151a40c4f10SYehuda Sadeh list_add_tail(&req->r_linger_item, &osdc->req_linger); 11526194ea89SSage Weil if (req->r_osd) 11536194ea89SSage Weil list_add_tail(&req->r_linger_osd, 11546194ea89SSage Weil &req->r_osd->o_linger_requests); 1155a40c4f10SYehuda Sadeh } 1156a40c4f10SYehuda Sadeh 1157a40c4f10SYehuda Sadeh static void __unregister_linger_request(struct ceph_osd_client *osdc, 1158a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1159a40c4f10SYehuda Sadeh { 1160a40c4f10SYehuda Sadeh dout("__unregister_linger_request %p\n", req); 1161a40c4f10SYehuda Sadeh list_del_init(&req->r_linger_item); 116261c74035SAlex Elder if (req->r_osd) { 1163a40c4f10SYehuda Sadeh list_del_init(&req->r_linger_osd); 1164a40c4f10SYehuda Sadeh 1165a40c4f10SYehuda Sadeh if (list_empty(&req->r_osd->o_requests) && 1166a40c4f10SYehuda Sadeh list_empty(&req->r_osd->o_linger_requests)) { 1167a40c4f10SYehuda Sadeh dout("moving osd to %p lru\n", req->r_osd); 1168a40c4f10SYehuda Sadeh __move_osd_to_lru(osdc, req->r_osd); 1169a40c4f10SYehuda Sadeh } 1170fbdb9190SSage Weil if (list_empty(&req->r_osd_item)) 1171a40c4f10SYehuda Sadeh req->r_osd = NULL; 1172a40c4f10SYehuda Sadeh } 1173a40c4f10SYehuda Sadeh } 1174a40c4f10SYehuda Sadeh 1175a40c4f10SYehuda Sadeh void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc, 1176a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1177a40c4f10SYehuda Sadeh { 1178a40c4f10SYehuda Sadeh mutex_lock(&osdc->request_mutex); 1179a40c4f10SYehuda Sadeh if (req->r_linger) { 1180a40c4f10SYehuda Sadeh __unregister_linger_request(osdc, req); 1181a40c4f10SYehuda Sadeh ceph_osdc_put_request(req); 1182a40c4f10SYehuda Sadeh } 1183a40c4f10SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 1184a40c4f10SYehuda Sadeh } 1185a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_unregister_linger_request); 1186a40c4f10SYehuda Sadeh 1187a40c4f10SYehuda Sadeh void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc, 1188a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1189a40c4f10SYehuda Sadeh { 1190a40c4f10SYehuda Sadeh if (!req->r_linger) { 1191a40c4f10SYehuda Sadeh dout("set_request_linger %p\n", req); 1192a40c4f10SYehuda Sadeh req->r_linger = 1; 1193a40c4f10SYehuda Sadeh /* 1194a40c4f10SYehuda Sadeh * caller is now responsible for calling 1195a40c4f10SYehuda Sadeh * unregister_linger_request 1196a40c4f10SYehuda Sadeh */ 1197a40c4f10SYehuda Sadeh ceph_osdc_get_request(req); 1198a40c4f10SYehuda Sadeh } 1199a40c4f10SYehuda Sadeh } 1200a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_set_request_linger); 1201a40c4f10SYehuda Sadeh 12023d14c5d2SYehuda Sadeh /* 12033d14c5d2SYehuda Sadeh * Pick an osd (the first 'up' osd in the pg), allocate the osd struct 12043d14c5d2SYehuda Sadeh * (as needed), and set the request r_osd appropriately. If there is 120525985edcSLucas De Marchi * no up osd, set r_osd to NULL. Move the request to the appropriate list 12066f6c7006SSage Weil * (unsent, homeless) or leave on in-flight lru. 12073d14c5d2SYehuda Sadeh * 12083d14c5d2SYehuda Sadeh * Return 0 if unchanged, 1 if changed, or negative on error. 12093d14c5d2SYehuda Sadeh * 12103d14c5d2SYehuda Sadeh * Caller should hold map_sem for read and request_mutex. 12113d14c5d2SYehuda Sadeh */ 12126f6c7006SSage Weil static int __map_request(struct ceph_osd_client *osdc, 121338d6453cSSage Weil struct ceph_osd_request *req, int force_resend) 12143d14c5d2SYehuda Sadeh { 12155b191d99SSage Weil struct ceph_pg pgid; 12163d14c5d2SYehuda Sadeh int acting[CEPH_PG_MAX_SIZE]; 12173d14c5d2SYehuda Sadeh int o = -1, num = 0; 12183d14c5d2SYehuda Sadeh int err; 12193d14c5d2SYehuda Sadeh 12206f6c7006SSage Weil dout("map_request %p tid %lld\n", req, req->r_tid); 122141766f87SAlex Elder err = ceph_calc_ceph_pg(&pgid, req->r_oid, osdc->osdmap, 122241766f87SAlex Elder ceph_file_layout_pg_pool(req->r_file_layout)); 12236f6c7006SSage Weil if (err) { 12246f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 12253d14c5d2SYehuda Sadeh return err; 12266f6c7006SSage Weil } 12273d14c5d2SYehuda Sadeh req->r_pgid = pgid; 12283d14c5d2SYehuda Sadeh 12293d14c5d2SYehuda Sadeh err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting); 12303d14c5d2SYehuda Sadeh if (err > 0) { 12313d14c5d2SYehuda Sadeh o = acting[0]; 12323d14c5d2SYehuda Sadeh num = err; 12333d14c5d2SYehuda Sadeh } 12343d14c5d2SYehuda Sadeh 123538d6453cSSage Weil if ((!force_resend && 123638d6453cSSage Weil req->r_osd && req->r_osd->o_osd == o && 12373d14c5d2SYehuda Sadeh req->r_sent >= req->r_osd->o_incarnation && 12383d14c5d2SYehuda Sadeh req->r_num_pg_osds == num && 12393d14c5d2SYehuda Sadeh memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) || 12403d14c5d2SYehuda Sadeh (req->r_osd == NULL && o == -1)) 12413d14c5d2SYehuda Sadeh return 0; /* no change */ 12423d14c5d2SYehuda Sadeh 12435b191d99SSage Weil dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n", 12445b191d99SSage Weil req->r_tid, pgid.pool, pgid.seed, o, 12453d14c5d2SYehuda Sadeh req->r_osd ? req->r_osd->o_osd : -1); 12463d14c5d2SYehuda Sadeh 12473d14c5d2SYehuda Sadeh /* record full pg acting set */ 12483d14c5d2SYehuda Sadeh memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num); 12493d14c5d2SYehuda Sadeh req->r_num_pg_osds = num; 12503d14c5d2SYehuda Sadeh 12513d14c5d2SYehuda Sadeh if (req->r_osd) { 12523d14c5d2SYehuda Sadeh __cancel_request(req); 12533d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 12543d14c5d2SYehuda Sadeh req->r_osd = NULL; 12553d14c5d2SYehuda Sadeh } 12563d14c5d2SYehuda Sadeh 12573d14c5d2SYehuda Sadeh req->r_osd = __lookup_osd(osdc, o); 12583d14c5d2SYehuda Sadeh if (!req->r_osd && o >= 0) { 12593d14c5d2SYehuda Sadeh err = -ENOMEM; 1260e10006f8SAlex Elder req->r_osd = create_osd(osdc, o); 12616f6c7006SSage Weil if (!req->r_osd) { 12626f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 12633d14c5d2SYehuda Sadeh goto out; 12646f6c7006SSage Weil } 12653d14c5d2SYehuda Sadeh 12666f6c7006SSage Weil dout("map_request osd %p is osd%d\n", req->r_osd, o); 12673d14c5d2SYehuda Sadeh __insert_osd(osdc, req->r_osd); 12683d14c5d2SYehuda Sadeh 1269b7a9e5ddSSage Weil ceph_con_open(&req->r_osd->o_con, 1270b7a9e5ddSSage Weil CEPH_ENTITY_TYPE_OSD, o, 1271b7a9e5ddSSage Weil &osdc->osdmap->osd_addr[o]); 12723d14c5d2SYehuda Sadeh } 12733d14c5d2SYehuda Sadeh 12743d14c5d2SYehuda Sadeh if (req->r_osd) { 12753d14c5d2SYehuda Sadeh __remove_osd_from_lru(req->r_osd); 1276ad885927SAlex Elder list_add_tail(&req->r_osd_item, &req->r_osd->o_requests); 1277ad885927SAlex Elder list_move_tail(&req->r_req_lru_item, &osdc->req_unsent); 12786f6c7006SSage Weil } else { 1279ad885927SAlex Elder list_move_tail(&req->r_req_lru_item, &osdc->req_notarget); 12803d14c5d2SYehuda Sadeh } 12813d14c5d2SYehuda Sadeh err = 1; /* osd or pg changed */ 12823d14c5d2SYehuda Sadeh 12833d14c5d2SYehuda Sadeh out: 12843d14c5d2SYehuda Sadeh return err; 12853d14c5d2SYehuda Sadeh } 12863d14c5d2SYehuda Sadeh 12873d14c5d2SYehuda Sadeh /* 12883d14c5d2SYehuda Sadeh * caller should hold map_sem (for read) and request_mutex 12893d14c5d2SYehuda Sadeh */ 129056e925b6SSage Weil static void __send_request(struct ceph_osd_client *osdc, 12913d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 12923d14c5d2SYehuda Sadeh { 12931b83bef2SSage Weil void *p; 12943d14c5d2SYehuda Sadeh 12951b83bef2SSage Weil dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n", 12961b83bef2SSage Weil req, req->r_tid, req->r_osd->o_osd, req->r_flags, 12971b83bef2SSage Weil (unsigned long long)req->r_pgid.pool, req->r_pgid.seed); 12983d14c5d2SYehuda Sadeh 12991b83bef2SSage Weil /* fill in message content that changes each time we send it */ 13001b83bef2SSage Weil put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch); 13011b83bef2SSage Weil put_unaligned_le32(req->r_flags, req->r_request_flags); 13021b83bef2SSage Weil put_unaligned_le64(req->r_pgid.pool, req->r_request_pool); 13031b83bef2SSage Weil p = req->r_request_pgid; 13041b83bef2SSage Weil ceph_encode_64(&p, req->r_pgid.pool); 13051b83bef2SSage Weil ceph_encode_32(&p, req->r_pgid.seed); 13061b83bef2SSage Weil put_unaligned_le64(1, req->r_request_attempts); /* FIXME */ 13071b83bef2SSage Weil memcpy(req->r_request_reassert_version, &req->r_reassert_version, 13081b83bef2SSage Weil sizeof(req->r_reassert_version)); 13092169aea6SSage Weil 13103d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 13113d14c5d2SYehuda Sadeh list_move_tail(&req->r_req_lru_item, &osdc->req_lru); 13123d14c5d2SYehuda Sadeh 13133d14c5d2SYehuda Sadeh ceph_msg_get(req->r_request); /* send consumes a ref */ 131426be8808SAlex Elder 131526be8808SAlex Elder /* Mark the request unsafe if this is the first timet's being sent. */ 131626be8808SAlex Elder 131726be8808SAlex Elder if (!req->r_sent && req->r_unsafe_callback) 131826be8808SAlex Elder req->r_unsafe_callback(req, true); 13193d14c5d2SYehuda Sadeh req->r_sent = req->r_osd->o_incarnation; 132026be8808SAlex Elder 132126be8808SAlex Elder ceph_con_send(&req->r_osd->o_con, req->r_request); 13223d14c5d2SYehuda Sadeh } 13233d14c5d2SYehuda Sadeh 13243d14c5d2SYehuda Sadeh /* 13256f6c7006SSage Weil * Send any requests in the queue (req_unsent). 13266f6c7006SSage Weil */ 1327f9d25199SAlex Elder static void __send_queued(struct ceph_osd_client *osdc) 13286f6c7006SSage Weil { 13296f6c7006SSage Weil struct ceph_osd_request *req, *tmp; 13306f6c7006SSage Weil 1331f9d25199SAlex Elder dout("__send_queued\n"); 1332f9d25199SAlex Elder list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) 13336f6c7006SSage Weil __send_request(osdc, req); 13346f6c7006SSage Weil } 13356f6c7006SSage Weil 13366f6c7006SSage Weil /* 13373d14c5d2SYehuda Sadeh * Timeout callback, called every N seconds when 1 or more osd 13383d14c5d2SYehuda Sadeh * requests has been active for more than N seconds. When this 13393d14c5d2SYehuda Sadeh * happens, we ping all OSDs with requests who have timed out to 13403d14c5d2SYehuda Sadeh * ensure any communications channel reset is detected. Reset the 13413d14c5d2SYehuda Sadeh * request timeouts another N seconds in the future as we go. 13423d14c5d2SYehuda Sadeh * Reschedule the timeout event another N seconds in future (unless 13433d14c5d2SYehuda Sadeh * there are no open requests). 13443d14c5d2SYehuda Sadeh */ 13453d14c5d2SYehuda Sadeh static void handle_timeout(struct work_struct *work) 13463d14c5d2SYehuda Sadeh { 13473d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 13483d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, timeout_work.work); 134983aff95eSSage Weil struct ceph_osd_request *req; 13503d14c5d2SYehuda Sadeh struct ceph_osd *osd; 13513d14c5d2SYehuda Sadeh unsigned long keepalive = 13523d14c5d2SYehuda Sadeh osdc->client->options->osd_keepalive_timeout * HZ; 13533d14c5d2SYehuda Sadeh struct list_head slow_osds; 13543d14c5d2SYehuda Sadeh dout("timeout\n"); 13553d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 13563d14c5d2SYehuda Sadeh 13573d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 13583d14c5d2SYehuda Sadeh 13593d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 13603d14c5d2SYehuda Sadeh 13613d14c5d2SYehuda Sadeh /* 13623d14c5d2SYehuda Sadeh * ping osds that are a bit slow. this ensures that if there 13633d14c5d2SYehuda Sadeh * is a break in the TCP connection we will notice, and reopen 13643d14c5d2SYehuda Sadeh * a connection with that osd (from the fault callback). 13653d14c5d2SYehuda Sadeh */ 13663d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&slow_osds); 13673d14c5d2SYehuda Sadeh list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) { 13683d14c5d2SYehuda Sadeh if (time_before(jiffies, req->r_stamp + keepalive)) 13693d14c5d2SYehuda Sadeh break; 13703d14c5d2SYehuda Sadeh 13713d14c5d2SYehuda Sadeh osd = req->r_osd; 13723d14c5d2SYehuda Sadeh BUG_ON(!osd); 13733d14c5d2SYehuda Sadeh dout(" tid %llu is slow, will send keepalive on osd%d\n", 13743d14c5d2SYehuda Sadeh req->r_tid, osd->o_osd); 13753d14c5d2SYehuda Sadeh list_move_tail(&osd->o_keepalive_item, &slow_osds); 13763d14c5d2SYehuda Sadeh } 13773d14c5d2SYehuda Sadeh while (!list_empty(&slow_osds)) { 13783d14c5d2SYehuda Sadeh osd = list_entry(slow_osds.next, struct ceph_osd, 13793d14c5d2SYehuda Sadeh o_keepalive_item); 13803d14c5d2SYehuda Sadeh list_del_init(&osd->o_keepalive_item); 13813d14c5d2SYehuda Sadeh ceph_con_keepalive(&osd->o_con); 13823d14c5d2SYehuda Sadeh } 13833d14c5d2SYehuda Sadeh 13843d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 1385f9d25199SAlex Elder __send_queued(osdc); 13863d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 13873d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 13883d14c5d2SYehuda Sadeh } 13893d14c5d2SYehuda Sadeh 13903d14c5d2SYehuda Sadeh static void handle_osds_timeout(struct work_struct *work) 13913d14c5d2SYehuda Sadeh { 13923d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 13933d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, 13943d14c5d2SYehuda Sadeh osds_timeout_work.work); 13953d14c5d2SYehuda Sadeh unsigned long delay = 13963d14c5d2SYehuda Sadeh osdc->client->options->osd_idle_ttl * HZ >> 2; 13973d14c5d2SYehuda Sadeh 13983d14c5d2SYehuda Sadeh dout("osds timeout\n"); 13993d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 1400aca420bcSSage Weil remove_old_osds(osdc); 14013d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 14023d14c5d2SYehuda Sadeh 14033d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 14043d14c5d2SYehuda Sadeh round_jiffies_relative(delay)); 14053d14c5d2SYehuda Sadeh } 14063d14c5d2SYehuda Sadeh 140725845472SSage Weil static void complete_request(struct ceph_osd_request *req) 140825845472SSage Weil { 140926be8808SAlex Elder if (req->r_unsafe_callback) 141026be8808SAlex Elder req->r_unsafe_callback(req, false); 141125845472SSage Weil complete_all(&req->r_safe_completion); /* fsync waiter */ 141225845472SSage Weil } 141325845472SSage Weil 14143d14c5d2SYehuda Sadeh /* 14153d14c5d2SYehuda Sadeh * handle osd op reply. either call the callback if it is specified, 14163d14c5d2SYehuda Sadeh * or do the completion to wake up the waiting thread. 14173d14c5d2SYehuda Sadeh */ 14183d14c5d2SYehuda Sadeh static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg, 14193d14c5d2SYehuda Sadeh struct ceph_connection *con) 14203d14c5d2SYehuda Sadeh { 14211b83bef2SSage Weil void *p, *end; 14223d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 14233d14c5d2SYehuda Sadeh u64 tid; 14241b83bef2SSage Weil int object_len; 142579528734SAlex Elder unsigned int numops; 142679528734SAlex Elder int payload_len, flags; 14273d14c5d2SYehuda Sadeh s32 result; 14281b83bef2SSage Weil s32 retry_attempt; 14291b83bef2SSage Weil struct ceph_pg pg; 14301b83bef2SSage Weil int err; 14311b83bef2SSage Weil u32 reassert_epoch; 14321b83bef2SSage Weil u64 reassert_version; 14331b83bef2SSage Weil u32 osdmap_epoch; 14340d5af164SAlex Elder int already_completed; 14359fc6e064SAlex Elder u32 bytes; 143679528734SAlex Elder unsigned int i; 14373d14c5d2SYehuda Sadeh 14383d14c5d2SYehuda Sadeh tid = le64_to_cpu(msg->hdr.tid); 14391b83bef2SSage Weil dout("handle_reply %p tid %llu\n", msg, tid); 14401b83bef2SSage Weil 14411b83bef2SSage Weil p = msg->front.iov_base; 14421b83bef2SSage Weil end = p + msg->front.iov_len; 14431b83bef2SSage Weil 14441b83bef2SSage Weil ceph_decode_need(&p, end, 4, bad); 14451b83bef2SSage Weil object_len = ceph_decode_32(&p); 14461b83bef2SSage Weil ceph_decode_need(&p, end, object_len, bad); 14471b83bef2SSage Weil p += object_len; 14481b83bef2SSage Weil 1449ef4859d6SAlex Elder err = ceph_decode_pgid(&p, end, &pg); 14501b83bef2SSage Weil if (err) 14513d14c5d2SYehuda Sadeh goto bad; 14521b83bef2SSage Weil 14531b83bef2SSage Weil ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad); 14541b83bef2SSage Weil flags = ceph_decode_64(&p); 14551b83bef2SSage Weil result = ceph_decode_32(&p); 14561b83bef2SSage Weil reassert_epoch = ceph_decode_32(&p); 14571b83bef2SSage Weil reassert_version = ceph_decode_64(&p); 14581b83bef2SSage Weil osdmap_epoch = ceph_decode_32(&p); 14591b83bef2SSage Weil 14603d14c5d2SYehuda Sadeh /* lookup */ 14613d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 14623d14c5d2SYehuda Sadeh req = __lookup_request(osdc, tid); 14633d14c5d2SYehuda Sadeh if (req == NULL) { 14643d14c5d2SYehuda Sadeh dout("handle_reply tid %llu dne\n", tid); 14658058fd45SAlex Elder goto bad_mutex; 14663d14c5d2SYehuda Sadeh } 14673d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 14681b83bef2SSage Weil 14691b83bef2SSage Weil dout("handle_reply %p tid %llu req %p result %d\n", msg, tid, 14701b83bef2SSage Weil req, result); 14711b83bef2SSage Weil 14721b83bef2SSage Weil ceph_decode_need(&p, end, 4, bad); 14731b83bef2SSage Weil numops = ceph_decode_32(&p); 14741b83bef2SSage Weil if (numops > CEPH_OSD_MAX_OP) 14751b83bef2SSage Weil goto bad_put; 14761b83bef2SSage Weil if (numops != req->r_num_ops) 14771b83bef2SSage Weil goto bad_put; 14781b83bef2SSage Weil payload_len = 0; 14791b83bef2SSage Weil ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad); 14801b83bef2SSage Weil for (i = 0; i < numops; i++) { 14811b83bef2SSage Weil struct ceph_osd_op *op = p; 14821b83bef2SSage Weil int len; 14831b83bef2SSage Weil 14841b83bef2SSage Weil len = le32_to_cpu(op->payload_len); 14851b83bef2SSage Weil req->r_reply_op_len[i] = len; 14861b83bef2SSage Weil dout(" op %d has %d bytes\n", i, len); 14871b83bef2SSage Weil payload_len += len; 14881b83bef2SSage Weil p += sizeof(*op); 14891b83bef2SSage Weil } 14909fc6e064SAlex Elder bytes = le32_to_cpu(msg->hdr.data_len); 14919fc6e064SAlex Elder if (payload_len != bytes) { 14921b83bef2SSage Weil pr_warning("sum of op payload lens %d != data_len %d", 14939fc6e064SAlex Elder payload_len, bytes); 14941b83bef2SSage Weil goto bad_put; 14951b83bef2SSage Weil } 14961b83bef2SSage Weil 14971b83bef2SSage Weil ceph_decode_need(&p, end, 4 + numops * 4, bad); 14981b83bef2SSage Weil retry_attempt = ceph_decode_32(&p); 14991b83bef2SSage Weil for (i = 0; i < numops; i++) 15001b83bef2SSage Weil req->r_reply_op_result[i] = ceph_decode_32(&p); 15013d14c5d2SYehuda Sadeh 15023d14c5d2SYehuda Sadeh if (!req->r_got_reply) { 15033d14c5d2SYehuda Sadeh 15041b83bef2SSage Weil req->r_result = result; 15053d14c5d2SYehuda Sadeh dout("handle_reply result %d bytes %d\n", req->r_result, 15063d14c5d2SYehuda Sadeh bytes); 15073d14c5d2SYehuda Sadeh if (req->r_result == 0) 15083d14c5d2SYehuda Sadeh req->r_result = bytes; 15093d14c5d2SYehuda Sadeh 15103d14c5d2SYehuda Sadeh /* in case this is a write and we need to replay, */ 15111b83bef2SSage Weil req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch); 15121b83bef2SSage Weil req->r_reassert_version.version = cpu_to_le64(reassert_version); 15133d14c5d2SYehuda Sadeh 15143d14c5d2SYehuda Sadeh req->r_got_reply = 1; 15153d14c5d2SYehuda Sadeh } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) { 15163d14c5d2SYehuda Sadeh dout("handle_reply tid %llu dup ack\n", tid); 15173d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 15183d14c5d2SYehuda Sadeh goto done; 15193d14c5d2SYehuda Sadeh } 15203d14c5d2SYehuda Sadeh 15213d14c5d2SYehuda Sadeh dout("handle_reply tid %llu flags %d\n", tid, flags); 15223d14c5d2SYehuda Sadeh 1523a40c4f10SYehuda Sadeh if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK)) 1524a40c4f10SYehuda Sadeh __register_linger_request(osdc, req); 1525a40c4f10SYehuda Sadeh 15263d14c5d2SYehuda Sadeh /* either this is a read, or we got the safe response */ 15273d14c5d2SYehuda Sadeh if (result < 0 || 15283d14c5d2SYehuda Sadeh (flags & CEPH_OSD_FLAG_ONDISK) || 15293d14c5d2SYehuda Sadeh ((flags & CEPH_OSD_FLAG_WRITE) == 0)) 15303d14c5d2SYehuda Sadeh __unregister_request(osdc, req); 15313d14c5d2SYehuda Sadeh 15320d5af164SAlex Elder already_completed = req->r_completed; 15330d5af164SAlex Elder req->r_completed = 1; 15343d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 15350d5af164SAlex Elder if (already_completed) 15360d5af164SAlex Elder goto done; 15373d14c5d2SYehuda Sadeh 15383d14c5d2SYehuda Sadeh if (req->r_callback) 15393d14c5d2SYehuda Sadeh req->r_callback(req, msg); 15403d14c5d2SYehuda Sadeh else 15413d14c5d2SYehuda Sadeh complete_all(&req->r_completion); 15423d14c5d2SYehuda Sadeh 154325845472SSage Weil if (flags & CEPH_OSD_FLAG_ONDISK) 154425845472SSage Weil complete_request(req); 15453d14c5d2SYehuda Sadeh 15463d14c5d2SYehuda Sadeh done: 1547a40c4f10SYehuda Sadeh dout("req=%p req->r_linger=%d\n", req, req->r_linger); 15483d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 15493d14c5d2SYehuda Sadeh return; 15503d14c5d2SYehuda Sadeh 15511b83bef2SSage Weil bad_put: 15521b83bef2SSage Weil ceph_osdc_put_request(req); 15538058fd45SAlex Elder bad_mutex: 15548058fd45SAlex Elder mutex_unlock(&osdc->request_mutex); 15553d14c5d2SYehuda Sadeh bad: 15561b83bef2SSage Weil pr_err("corrupt osd_op_reply got %d %d\n", 15571b83bef2SSage Weil (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len)); 15583d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 15593d14c5d2SYehuda Sadeh } 15603d14c5d2SYehuda Sadeh 15616f6c7006SSage Weil static void reset_changed_osds(struct ceph_osd_client *osdc) 15623d14c5d2SYehuda Sadeh { 15633d14c5d2SYehuda Sadeh struct rb_node *p, *n; 15643d14c5d2SYehuda Sadeh 15653d14c5d2SYehuda Sadeh for (p = rb_first(&osdc->osds); p; p = n) { 15666f6c7006SSage Weil struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node); 15673d14c5d2SYehuda Sadeh 15683d14c5d2SYehuda Sadeh n = rb_next(p); 15693d14c5d2SYehuda Sadeh if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) || 15703d14c5d2SYehuda Sadeh memcmp(&osd->o_con.peer_addr, 15713d14c5d2SYehuda Sadeh ceph_osd_addr(osdc->osdmap, 15723d14c5d2SYehuda Sadeh osd->o_osd), 15733d14c5d2SYehuda Sadeh sizeof(struct ceph_entity_addr)) != 0) 15743d14c5d2SYehuda Sadeh __reset_osd(osdc, osd); 15753d14c5d2SYehuda Sadeh } 15763d14c5d2SYehuda Sadeh } 15773d14c5d2SYehuda Sadeh 15783d14c5d2SYehuda Sadeh /* 15796f6c7006SSage Weil * Requeue requests whose mapping to an OSD has changed. If requests map to 15806f6c7006SSage Weil * no osd, request a new map. 15813d14c5d2SYehuda Sadeh * 1582e6d50f67SAlex Elder * Caller should hold map_sem for read. 15833d14c5d2SYehuda Sadeh */ 158438d6453cSSage Weil static void kick_requests(struct ceph_osd_client *osdc, int force_resend) 15853d14c5d2SYehuda Sadeh { 1586a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 15876f6c7006SSage Weil struct rb_node *p; 15886f6c7006SSage Weil int needmap = 0; 15896f6c7006SSage Weil int err; 15903d14c5d2SYehuda Sadeh 159138d6453cSSage Weil dout("kick_requests %s\n", force_resend ? " (force resend)" : ""); 15923d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 15936194ea89SSage Weil for (p = rb_first(&osdc->requests); p; ) { 15946f6c7006SSage Weil req = rb_entry(p, struct ceph_osd_request, r_node); 15956194ea89SSage Weil p = rb_next(p); 1596ab60b16dSAlex Elder 1597ab60b16dSAlex Elder /* 1598ab60b16dSAlex Elder * For linger requests that have not yet been 1599ab60b16dSAlex Elder * registered, move them to the linger list; they'll 1600ab60b16dSAlex Elder * be sent to the osd in the loop below. Unregister 1601ab60b16dSAlex Elder * the request before re-registering it as a linger 1602ab60b16dSAlex Elder * request to ensure the __map_request() below 1603ab60b16dSAlex Elder * will decide it needs to be sent. 1604ab60b16dSAlex Elder */ 1605ab60b16dSAlex Elder if (req->r_linger && list_empty(&req->r_linger_item)) { 1606ab60b16dSAlex Elder dout("%p tid %llu restart on osd%d\n", 1607ab60b16dSAlex Elder req, req->r_tid, 1608ab60b16dSAlex Elder req->r_osd ? req->r_osd->o_osd : -1); 1609ab60b16dSAlex Elder __unregister_request(osdc, req); 1610ab60b16dSAlex Elder __register_linger_request(osdc, req); 1611ab60b16dSAlex Elder continue; 1612ab60b16dSAlex Elder } 1613ab60b16dSAlex Elder 161438d6453cSSage Weil err = __map_request(osdc, req, force_resend); 16156f6c7006SSage Weil if (err < 0) 16166f6c7006SSage Weil continue; /* error */ 16176f6c7006SSage Weil if (req->r_osd == NULL) { 16186f6c7006SSage Weil dout("%p tid %llu maps to no osd\n", req, req->r_tid); 16196f6c7006SSage Weil needmap++; /* request a newer map */ 16206f6c7006SSage Weil } else if (err > 0) { 16216194ea89SSage Weil if (!req->r_linger) { 16226194ea89SSage Weil dout("%p tid %llu requeued on osd%d\n", req, 16236194ea89SSage Weil req->r_tid, 16246f6c7006SSage Weil req->r_osd ? req->r_osd->o_osd : -1); 16256f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 16266f6c7006SSage Weil } 16276f6c7006SSage Weil } 16286194ea89SSage Weil } 1629a40c4f10SYehuda Sadeh 1630a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osdc->req_linger, 1631a40c4f10SYehuda Sadeh r_linger_item) { 1632a40c4f10SYehuda Sadeh dout("linger req=%p req->r_osd=%p\n", req, req->r_osd); 1633a40c4f10SYehuda Sadeh 163438d6453cSSage Weil err = __map_request(osdc, req, force_resend); 1635ab60b16dSAlex Elder dout("__map_request returned %d\n", err); 1636a40c4f10SYehuda Sadeh if (err == 0) 1637a40c4f10SYehuda Sadeh continue; /* no change and no osd was specified */ 1638a40c4f10SYehuda Sadeh if (err < 0) 1639a40c4f10SYehuda Sadeh continue; /* hrm! */ 1640a40c4f10SYehuda Sadeh if (req->r_osd == NULL) { 1641a40c4f10SYehuda Sadeh dout("tid %llu maps to no valid osd\n", req->r_tid); 1642a40c4f10SYehuda Sadeh needmap++; /* request a newer map */ 1643a40c4f10SYehuda Sadeh continue; 1644a40c4f10SYehuda Sadeh } 1645a40c4f10SYehuda Sadeh 1646a40c4f10SYehuda Sadeh dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid, 1647a40c4f10SYehuda Sadeh req->r_osd ? req->r_osd->o_osd : -1); 1648a40c4f10SYehuda Sadeh __register_request(osdc, req); 1649c89ce05eSAlex Elder __unregister_linger_request(osdc, req); 1650a40c4f10SYehuda Sadeh } 16513d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 16523d14c5d2SYehuda Sadeh 16533d14c5d2SYehuda Sadeh if (needmap) { 16543d14c5d2SYehuda Sadeh dout("%d requests for down osds, need new map\n", needmap); 16553d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 16563d14c5d2SYehuda Sadeh } 1657e6d50f67SAlex Elder reset_changed_osds(osdc); 16583d14c5d2SYehuda Sadeh } 16596f6c7006SSage Weil 16606f6c7006SSage Weil 16613d14c5d2SYehuda Sadeh /* 16623d14c5d2SYehuda Sadeh * Process updated osd map. 16633d14c5d2SYehuda Sadeh * 16643d14c5d2SYehuda Sadeh * The message contains any number of incremental and full maps, normally 16653d14c5d2SYehuda Sadeh * indicating some sort of topology change in the cluster. Kick requests 16663d14c5d2SYehuda Sadeh * off to different OSDs as needed. 16673d14c5d2SYehuda Sadeh */ 16683d14c5d2SYehuda Sadeh void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg) 16693d14c5d2SYehuda Sadeh { 16703d14c5d2SYehuda Sadeh void *p, *end, *next; 16713d14c5d2SYehuda Sadeh u32 nr_maps, maplen; 16723d14c5d2SYehuda Sadeh u32 epoch; 16733d14c5d2SYehuda Sadeh struct ceph_osdmap *newmap = NULL, *oldmap; 16743d14c5d2SYehuda Sadeh int err; 16753d14c5d2SYehuda Sadeh struct ceph_fsid fsid; 16763d14c5d2SYehuda Sadeh 16773d14c5d2SYehuda Sadeh dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0); 16783d14c5d2SYehuda Sadeh p = msg->front.iov_base; 16793d14c5d2SYehuda Sadeh end = p + msg->front.iov_len; 16803d14c5d2SYehuda Sadeh 16813d14c5d2SYehuda Sadeh /* verify fsid */ 16823d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, sizeof(fsid), bad); 16833d14c5d2SYehuda Sadeh ceph_decode_copy(&p, &fsid, sizeof(fsid)); 16843d14c5d2SYehuda Sadeh if (ceph_check_fsid(osdc->client, &fsid) < 0) 16853d14c5d2SYehuda Sadeh return; 16863d14c5d2SYehuda Sadeh 16873d14c5d2SYehuda Sadeh down_write(&osdc->map_sem); 16883d14c5d2SYehuda Sadeh 16893d14c5d2SYehuda Sadeh /* incremental maps */ 16903d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 16913d14c5d2SYehuda Sadeh dout(" %d inc maps\n", nr_maps); 16923d14c5d2SYehuda Sadeh while (nr_maps > 0) { 16933d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 16943d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 16953d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 16963d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 16973d14c5d2SYehuda Sadeh next = p + maplen; 16983d14c5d2SYehuda Sadeh if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) { 16993d14c5d2SYehuda Sadeh dout("applying incremental map %u len %d\n", 17003d14c5d2SYehuda Sadeh epoch, maplen); 17013d14c5d2SYehuda Sadeh newmap = osdmap_apply_incremental(&p, next, 17023d14c5d2SYehuda Sadeh osdc->osdmap, 170315d9882cSAlex Elder &osdc->client->msgr); 17043d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 17053d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 17063d14c5d2SYehuda Sadeh goto bad; 17073d14c5d2SYehuda Sadeh } 17083d14c5d2SYehuda Sadeh BUG_ON(!newmap); 17093d14c5d2SYehuda Sadeh if (newmap != osdc->osdmap) { 17103d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 17113d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 17123d14c5d2SYehuda Sadeh } 171338d6453cSSage Weil kick_requests(osdc, 0); 17143d14c5d2SYehuda Sadeh } else { 17153d14c5d2SYehuda Sadeh dout("ignoring incremental map %u len %d\n", 17163d14c5d2SYehuda Sadeh epoch, maplen); 17173d14c5d2SYehuda Sadeh } 17183d14c5d2SYehuda Sadeh p = next; 17193d14c5d2SYehuda Sadeh nr_maps--; 17203d14c5d2SYehuda Sadeh } 17213d14c5d2SYehuda Sadeh if (newmap) 17223d14c5d2SYehuda Sadeh goto done; 17233d14c5d2SYehuda Sadeh 17243d14c5d2SYehuda Sadeh /* full maps */ 17253d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 17263d14c5d2SYehuda Sadeh dout(" %d full maps\n", nr_maps); 17273d14c5d2SYehuda Sadeh while (nr_maps) { 17283d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 17293d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 17303d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 17313d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 17323d14c5d2SYehuda Sadeh if (nr_maps > 1) { 17333d14c5d2SYehuda Sadeh dout("skipping non-latest full map %u len %d\n", 17343d14c5d2SYehuda Sadeh epoch, maplen); 17353d14c5d2SYehuda Sadeh } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) { 17363d14c5d2SYehuda Sadeh dout("skipping full map %u len %d, " 17373d14c5d2SYehuda Sadeh "older than our %u\n", epoch, maplen, 17383d14c5d2SYehuda Sadeh osdc->osdmap->epoch); 17393d14c5d2SYehuda Sadeh } else { 174038d6453cSSage Weil int skipped_map = 0; 174138d6453cSSage Weil 17423d14c5d2SYehuda Sadeh dout("taking full map %u len %d\n", epoch, maplen); 17433d14c5d2SYehuda Sadeh newmap = osdmap_decode(&p, p+maplen); 17443d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 17453d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 17463d14c5d2SYehuda Sadeh goto bad; 17473d14c5d2SYehuda Sadeh } 17483d14c5d2SYehuda Sadeh BUG_ON(!newmap); 17493d14c5d2SYehuda Sadeh oldmap = osdc->osdmap; 17503d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 175138d6453cSSage Weil if (oldmap) { 175238d6453cSSage Weil if (oldmap->epoch + 1 < newmap->epoch) 175338d6453cSSage Weil skipped_map = 1; 17543d14c5d2SYehuda Sadeh ceph_osdmap_destroy(oldmap); 175538d6453cSSage Weil } 175638d6453cSSage Weil kick_requests(osdc, skipped_map); 17573d14c5d2SYehuda Sadeh } 17583d14c5d2SYehuda Sadeh p += maplen; 17593d14c5d2SYehuda Sadeh nr_maps--; 17603d14c5d2SYehuda Sadeh } 17613d14c5d2SYehuda Sadeh 17623d14c5d2SYehuda Sadeh done: 17633d14c5d2SYehuda Sadeh downgrade_write(&osdc->map_sem); 17643d14c5d2SYehuda Sadeh ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch); 1765cd634fb6SSage Weil 1766cd634fb6SSage Weil /* 1767cd634fb6SSage Weil * subscribe to subsequent osdmap updates if full to ensure 1768cd634fb6SSage Weil * we find out when we are no longer full and stop returning 1769cd634fb6SSage Weil * ENOSPC. 1770cd634fb6SSage Weil */ 1771cd634fb6SSage Weil if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL)) 1772cd634fb6SSage Weil ceph_monc_request_next_osdmap(&osdc->client->monc); 1773cd634fb6SSage Weil 1774f9d25199SAlex Elder mutex_lock(&osdc->request_mutex); 1775f9d25199SAlex Elder __send_queued(osdc); 1776f9d25199SAlex Elder mutex_unlock(&osdc->request_mutex); 17773d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 17783d14c5d2SYehuda Sadeh wake_up_all(&osdc->client->auth_wq); 17793d14c5d2SYehuda Sadeh return; 17803d14c5d2SYehuda Sadeh 17813d14c5d2SYehuda Sadeh bad: 17823d14c5d2SYehuda Sadeh pr_err("osdc handle_map corrupt msg\n"); 17833d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 17843d14c5d2SYehuda Sadeh up_write(&osdc->map_sem); 17853d14c5d2SYehuda Sadeh return; 17863d14c5d2SYehuda Sadeh } 17873d14c5d2SYehuda Sadeh 17883d14c5d2SYehuda Sadeh /* 1789a40c4f10SYehuda Sadeh * watch/notify callback event infrastructure 1790a40c4f10SYehuda Sadeh * 1791a40c4f10SYehuda Sadeh * These callbacks are used both for watch and notify operations. 1792a40c4f10SYehuda Sadeh */ 1793a40c4f10SYehuda Sadeh static void __release_event(struct kref *kref) 1794a40c4f10SYehuda Sadeh { 1795a40c4f10SYehuda Sadeh struct ceph_osd_event *event = 1796a40c4f10SYehuda Sadeh container_of(kref, struct ceph_osd_event, kref); 1797a40c4f10SYehuda Sadeh 1798a40c4f10SYehuda Sadeh dout("__release_event %p\n", event); 1799a40c4f10SYehuda Sadeh kfree(event); 1800a40c4f10SYehuda Sadeh } 1801a40c4f10SYehuda Sadeh 1802a40c4f10SYehuda Sadeh static void get_event(struct ceph_osd_event *event) 1803a40c4f10SYehuda Sadeh { 1804a40c4f10SYehuda Sadeh kref_get(&event->kref); 1805a40c4f10SYehuda Sadeh } 1806a40c4f10SYehuda Sadeh 1807a40c4f10SYehuda Sadeh void ceph_osdc_put_event(struct ceph_osd_event *event) 1808a40c4f10SYehuda Sadeh { 1809a40c4f10SYehuda Sadeh kref_put(&event->kref, __release_event); 1810a40c4f10SYehuda Sadeh } 1811a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_put_event); 1812a40c4f10SYehuda Sadeh 1813a40c4f10SYehuda Sadeh static void __insert_event(struct ceph_osd_client *osdc, 1814a40c4f10SYehuda Sadeh struct ceph_osd_event *new) 1815a40c4f10SYehuda Sadeh { 1816a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 1817a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 1818a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 1819a40c4f10SYehuda Sadeh 1820a40c4f10SYehuda Sadeh while (*p) { 1821a40c4f10SYehuda Sadeh parent = *p; 1822a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 1823a40c4f10SYehuda Sadeh if (new->cookie < event->cookie) 1824a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 1825a40c4f10SYehuda Sadeh else if (new->cookie > event->cookie) 1826a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 1827a40c4f10SYehuda Sadeh else 1828a40c4f10SYehuda Sadeh BUG(); 1829a40c4f10SYehuda Sadeh } 1830a40c4f10SYehuda Sadeh 1831a40c4f10SYehuda Sadeh rb_link_node(&new->node, parent, p); 1832a40c4f10SYehuda Sadeh rb_insert_color(&new->node, &osdc->event_tree); 1833a40c4f10SYehuda Sadeh } 1834a40c4f10SYehuda Sadeh 1835a40c4f10SYehuda Sadeh static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc, 1836a40c4f10SYehuda Sadeh u64 cookie) 1837a40c4f10SYehuda Sadeh { 1838a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 1839a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 1840a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 1841a40c4f10SYehuda Sadeh 1842a40c4f10SYehuda Sadeh while (*p) { 1843a40c4f10SYehuda Sadeh parent = *p; 1844a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 1845a40c4f10SYehuda Sadeh if (cookie < event->cookie) 1846a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 1847a40c4f10SYehuda Sadeh else if (cookie > event->cookie) 1848a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 1849a40c4f10SYehuda Sadeh else 1850a40c4f10SYehuda Sadeh return event; 1851a40c4f10SYehuda Sadeh } 1852a40c4f10SYehuda Sadeh return NULL; 1853a40c4f10SYehuda Sadeh } 1854a40c4f10SYehuda Sadeh 1855a40c4f10SYehuda Sadeh static void __remove_event(struct ceph_osd_event *event) 1856a40c4f10SYehuda Sadeh { 1857a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 1858a40c4f10SYehuda Sadeh 1859a40c4f10SYehuda Sadeh if (!RB_EMPTY_NODE(&event->node)) { 1860a40c4f10SYehuda Sadeh dout("__remove_event removed %p\n", event); 1861a40c4f10SYehuda Sadeh rb_erase(&event->node, &osdc->event_tree); 1862a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 1863a40c4f10SYehuda Sadeh } else { 1864a40c4f10SYehuda Sadeh dout("__remove_event didn't remove %p\n", event); 1865a40c4f10SYehuda Sadeh } 1866a40c4f10SYehuda Sadeh } 1867a40c4f10SYehuda Sadeh 1868a40c4f10SYehuda Sadeh int ceph_osdc_create_event(struct ceph_osd_client *osdc, 1869a40c4f10SYehuda Sadeh void (*event_cb)(u64, u64, u8, void *), 18703c663bbdSAlex Elder void *data, struct ceph_osd_event **pevent) 1871a40c4f10SYehuda Sadeh { 1872a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 1873a40c4f10SYehuda Sadeh 1874a40c4f10SYehuda Sadeh event = kmalloc(sizeof(*event), GFP_NOIO); 1875a40c4f10SYehuda Sadeh if (!event) 1876a40c4f10SYehuda Sadeh return -ENOMEM; 1877a40c4f10SYehuda Sadeh 1878a40c4f10SYehuda Sadeh dout("create_event %p\n", event); 1879a40c4f10SYehuda Sadeh event->cb = event_cb; 18803c663bbdSAlex Elder event->one_shot = 0; 1881a40c4f10SYehuda Sadeh event->data = data; 1882a40c4f10SYehuda Sadeh event->osdc = osdc; 1883a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&event->osd_node); 18843ee5234dSAlex Elder RB_CLEAR_NODE(&event->node); 1885a40c4f10SYehuda Sadeh kref_init(&event->kref); /* one ref for us */ 1886a40c4f10SYehuda Sadeh kref_get(&event->kref); /* one ref for the caller */ 1887a40c4f10SYehuda Sadeh 1888a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 1889a40c4f10SYehuda Sadeh event->cookie = ++osdc->event_count; 1890a40c4f10SYehuda Sadeh __insert_event(osdc, event); 1891a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 1892a40c4f10SYehuda Sadeh 1893a40c4f10SYehuda Sadeh *pevent = event; 1894a40c4f10SYehuda Sadeh return 0; 1895a40c4f10SYehuda Sadeh } 1896a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_create_event); 1897a40c4f10SYehuda Sadeh 1898a40c4f10SYehuda Sadeh void ceph_osdc_cancel_event(struct ceph_osd_event *event) 1899a40c4f10SYehuda Sadeh { 1900a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 1901a40c4f10SYehuda Sadeh 1902a40c4f10SYehuda Sadeh dout("cancel_event %p\n", event); 1903a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 1904a40c4f10SYehuda Sadeh __remove_event(event); 1905a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 1906a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); /* caller's */ 1907a40c4f10SYehuda Sadeh } 1908a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_cancel_event); 1909a40c4f10SYehuda Sadeh 1910a40c4f10SYehuda Sadeh 1911a40c4f10SYehuda Sadeh static void do_event_work(struct work_struct *work) 1912a40c4f10SYehuda Sadeh { 1913a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work = 1914a40c4f10SYehuda Sadeh container_of(work, struct ceph_osd_event_work, work); 1915a40c4f10SYehuda Sadeh struct ceph_osd_event *event = event_work->event; 1916a40c4f10SYehuda Sadeh u64 ver = event_work->ver; 1917a40c4f10SYehuda Sadeh u64 notify_id = event_work->notify_id; 1918a40c4f10SYehuda Sadeh u8 opcode = event_work->opcode; 1919a40c4f10SYehuda Sadeh 1920a40c4f10SYehuda Sadeh dout("do_event_work completing %p\n", event); 1921a40c4f10SYehuda Sadeh event->cb(ver, notify_id, opcode, event->data); 1922a40c4f10SYehuda Sadeh dout("do_event_work completed %p\n", event); 1923a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 1924a40c4f10SYehuda Sadeh kfree(event_work); 1925a40c4f10SYehuda Sadeh } 1926a40c4f10SYehuda Sadeh 1927a40c4f10SYehuda Sadeh 1928a40c4f10SYehuda Sadeh /* 1929a40c4f10SYehuda Sadeh * Process osd watch notifications 1930a40c4f10SYehuda Sadeh */ 19313c663bbdSAlex Elder static void handle_watch_notify(struct ceph_osd_client *osdc, 19323c663bbdSAlex Elder struct ceph_msg *msg) 1933a40c4f10SYehuda Sadeh { 1934a40c4f10SYehuda Sadeh void *p, *end; 1935a40c4f10SYehuda Sadeh u8 proto_ver; 1936a40c4f10SYehuda Sadeh u64 cookie, ver, notify_id; 1937a40c4f10SYehuda Sadeh u8 opcode; 1938a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 1939a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work; 1940a40c4f10SYehuda Sadeh 1941a40c4f10SYehuda Sadeh p = msg->front.iov_base; 1942a40c4f10SYehuda Sadeh end = p + msg->front.iov_len; 1943a40c4f10SYehuda Sadeh 1944a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, proto_ver, bad); 1945a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, opcode, bad); 1946a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, cookie, bad); 1947a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, ver, bad); 1948a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, notify_id, bad); 1949a40c4f10SYehuda Sadeh 1950a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 1951a40c4f10SYehuda Sadeh event = __find_event(osdc, cookie); 1952a40c4f10SYehuda Sadeh if (event) { 19533c663bbdSAlex Elder BUG_ON(event->one_shot); 1954a40c4f10SYehuda Sadeh get_event(event); 1955a40c4f10SYehuda Sadeh } 1956a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 1957a40c4f10SYehuda Sadeh dout("handle_watch_notify cookie %lld ver %lld event %p\n", 1958a40c4f10SYehuda Sadeh cookie, ver, event); 1959a40c4f10SYehuda Sadeh if (event) { 1960a40c4f10SYehuda Sadeh event_work = kmalloc(sizeof(*event_work), GFP_NOIO); 1961a40c4f10SYehuda Sadeh if (!event_work) { 1962a40c4f10SYehuda Sadeh dout("ERROR: could not allocate event_work\n"); 1963a40c4f10SYehuda Sadeh goto done_err; 1964a40c4f10SYehuda Sadeh } 19656b0ae409SMariusz Kozlowski INIT_WORK(&event_work->work, do_event_work); 1966a40c4f10SYehuda Sadeh event_work->event = event; 1967a40c4f10SYehuda Sadeh event_work->ver = ver; 1968a40c4f10SYehuda Sadeh event_work->notify_id = notify_id; 1969a40c4f10SYehuda Sadeh event_work->opcode = opcode; 1970a40c4f10SYehuda Sadeh if (!queue_work(osdc->notify_wq, &event_work->work)) { 1971a40c4f10SYehuda Sadeh dout("WARNING: failed to queue notify event work\n"); 1972a40c4f10SYehuda Sadeh goto done_err; 1973a40c4f10SYehuda Sadeh } 1974a40c4f10SYehuda Sadeh } 1975a40c4f10SYehuda Sadeh 1976a40c4f10SYehuda Sadeh return; 1977a40c4f10SYehuda Sadeh 1978a40c4f10SYehuda Sadeh done_err: 1979a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 1980a40c4f10SYehuda Sadeh return; 1981a40c4f10SYehuda Sadeh 1982a40c4f10SYehuda Sadeh bad: 1983a40c4f10SYehuda Sadeh pr_err("osdc handle_watch_notify corrupt msg\n"); 1984a40c4f10SYehuda Sadeh return; 1985a40c4f10SYehuda Sadeh } 1986a40c4f10SYehuda Sadeh 1987a40c4f10SYehuda Sadeh /* 1988e65550fdSAlex Elder * build new request AND message 1989e65550fdSAlex Elder * 1990e65550fdSAlex Elder */ 1991e65550fdSAlex Elder void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off, 1992e65550fdSAlex Elder struct ceph_snap_context *snapc, u64 snap_id, 1993e65550fdSAlex Elder struct timespec *mtime) 1994e65550fdSAlex Elder { 1995e65550fdSAlex Elder struct ceph_msg *msg = req->r_request; 1996e65550fdSAlex Elder void *p; 1997e65550fdSAlex Elder size_t msg_size; 1998e65550fdSAlex Elder int flags = req->r_flags; 1999e65550fdSAlex Elder u64 data_len; 2000e65550fdSAlex Elder unsigned int i; 2001e65550fdSAlex Elder 2002e65550fdSAlex Elder req->r_snapid = snap_id; 2003e65550fdSAlex Elder req->r_snapc = ceph_get_snap_context(snapc); 2004e65550fdSAlex Elder 2005e65550fdSAlex Elder /* encode request */ 2006e65550fdSAlex Elder msg->hdr.version = cpu_to_le16(4); 2007e65550fdSAlex Elder 2008e65550fdSAlex Elder p = msg->front.iov_base; 2009e65550fdSAlex Elder ceph_encode_32(&p, 1); /* client_inc is always 1 */ 2010e65550fdSAlex Elder req->r_request_osdmap_epoch = p; 2011e65550fdSAlex Elder p += 4; 2012e65550fdSAlex Elder req->r_request_flags = p; 2013e65550fdSAlex Elder p += 4; 2014e65550fdSAlex Elder if (req->r_flags & CEPH_OSD_FLAG_WRITE) 2015e65550fdSAlex Elder ceph_encode_timespec(p, mtime); 2016e65550fdSAlex Elder p += sizeof(struct ceph_timespec); 2017e65550fdSAlex Elder req->r_request_reassert_version = p; 2018e65550fdSAlex Elder p += sizeof(struct ceph_eversion); /* will get filled in */ 2019e65550fdSAlex Elder 2020e65550fdSAlex Elder /* oloc */ 2021e65550fdSAlex Elder ceph_encode_8(&p, 4); 2022e65550fdSAlex Elder ceph_encode_8(&p, 4); 2023e65550fdSAlex Elder ceph_encode_32(&p, 8 + 4 + 4); 2024e65550fdSAlex Elder req->r_request_pool = p; 2025e65550fdSAlex Elder p += 8; 2026e65550fdSAlex Elder ceph_encode_32(&p, -1); /* preferred */ 2027e65550fdSAlex Elder ceph_encode_32(&p, 0); /* key len */ 2028e65550fdSAlex Elder 2029e65550fdSAlex Elder ceph_encode_8(&p, 1); 2030e65550fdSAlex Elder req->r_request_pgid = p; 2031e65550fdSAlex Elder p += 8 + 4; 2032e65550fdSAlex Elder ceph_encode_32(&p, -1); /* preferred */ 2033e65550fdSAlex Elder 2034e65550fdSAlex Elder /* oid */ 2035e65550fdSAlex Elder ceph_encode_32(&p, req->r_oid_len); 2036e65550fdSAlex Elder memcpy(p, req->r_oid, req->r_oid_len); 2037e65550fdSAlex Elder dout("oid '%.*s' len %d\n", req->r_oid_len, req->r_oid, req->r_oid_len); 2038e65550fdSAlex Elder p += req->r_oid_len; 2039e65550fdSAlex Elder 2040e65550fdSAlex Elder /* ops--can imply data */ 2041e65550fdSAlex Elder ceph_encode_16(&p, (u16)req->r_num_ops); 2042e65550fdSAlex Elder data_len = 0; 2043e65550fdSAlex Elder for (i = 0; i < req->r_num_ops; i++) { 2044e65550fdSAlex Elder data_len += osd_req_encode_op(req, p, i); 2045e65550fdSAlex Elder p += sizeof(struct ceph_osd_op); 2046e65550fdSAlex Elder } 2047e65550fdSAlex Elder 2048e65550fdSAlex Elder /* snaps */ 2049e65550fdSAlex Elder ceph_encode_64(&p, req->r_snapid); 2050e65550fdSAlex Elder ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0); 2051e65550fdSAlex Elder ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0); 2052e65550fdSAlex Elder if (req->r_snapc) { 2053e65550fdSAlex Elder for (i = 0; i < snapc->num_snaps; i++) { 2054e65550fdSAlex Elder ceph_encode_64(&p, req->r_snapc->snaps[i]); 2055e65550fdSAlex Elder } 2056e65550fdSAlex Elder } 2057e65550fdSAlex Elder 2058e65550fdSAlex Elder req->r_request_attempts = p; 2059e65550fdSAlex Elder p += 4; 2060e65550fdSAlex Elder 2061e65550fdSAlex Elder /* data */ 2062e65550fdSAlex Elder if (flags & CEPH_OSD_FLAG_WRITE) { 2063e65550fdSAlex Elder u16 data_off; 2064e65550fdSAlex Elder 2065e65550fdSAlex Elder /* 2066e65550fdSAlex Elder * The header "data_off" is a hint to the receiver 2067e65550fdSAlex Elder * allowing it to align received data into its 2068e65550fdSAlex Elder * buffers such that there's no need to re-copy 2069e65550fdSAlex Elder * it before writing it to disk (direct I/O). 2070e65550fdSAlex Elder */ 2071e65550fdSAlex Elder data_off = (u16) (off & 0xffff); 2072e65550fdSAlex Elder req->r_request->hdr.data_off = cpu_to_le16(data_off); 2073e65550fdSAlex Elder } 2074e65550fdSAlex Elder req->r_request->hdr.data_len = cpu_to_le32(data_len); 2075e65550fdSAlex Elder 2076e65550fdSAlex Elder BUG_ON(p > msg->front.iov_base + msg->front.iov_len); 2077e65550fdSAlex Elder msg_size = p - msg->front.iov_base; 2078e65550fdSAlex Elder msg->front.iov_len = msg_size; 2079e65550fdSAlex Elder msg->hdr.front_len = cpu_to_le32(msg_size); 2080e65550fdSAlex Elder 2081e65550fdSAlex Elder dout("build_request msg_size was %d\n", (int)msg_size); 2082e65550fdSAlex Elder } 2083e65550fdSAlex Elder EXPORT_SYMBOL(ceph_osdc_build_request); 2084e65550fdSAlex Elder 2085e65550fdSAlex Elder /* 20863d14c5d2SYehuda Sadeh * Register request, send initial attempt. 20873d14c5d2SYehuda Sadeh */ 20883d14c5d2SYehuda Sadeh int ceph_osdc_start_request(struct ceph_osd_client *osdc, 20893d14c5d2SYehuda Sadeh struct ceph_osd_request *req, 20903d14c5d2SYehuda Sadeh bool nofail) 20913d14c5d2SYehuda Sadeh { 20923d14c5d2SYehuda Sadeh int rc = 0; 20933d14c5d2SYehuda Sadeh 20943d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 20953d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 2096dc4b870cSAlex Elder __register_request(osdc, req); 209792451b49SAlex Elder WARN_ON(req->r_sent); 209838d6453cSSage Weil rc = __map_request(osdc, req, 0); 20999d6fcb08SSage Weil if (rc < 0) { 21009d6fcb08SSage Weil if (nofail) { 21019d6fcb08SSage Weil dout("osdc_start_request failed map, " 21029d6fcb08SSage Weil " will retry %lld\n", req->r_tid); 21039d6fcb08SSage Weil rc = 0; 21049d6fcb08SSage Weil } 2105234af26fSDan Carpenter goto out_unlock; 21069d6fcb08SSage Weil } 21076f6c7006SSage Weil if (req->r_osd == NULL) { 21086f6c7006SSage Weil dout("send_request %p no up osds in pg\n", req); 21096f6c7006SSage Weil ceph_monc_request_next_osdmap(&osdc->client->monc); 21106f6c7006SSage Weil } else { 21117e2766a1SAlex Elder __send_queued(osdc); 211256e925b6SSage Weil } 21133d14c5d2SYehuda Sadeh rc = 0; 2114234af26fSDan Carpenter out_unlock: 21153d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 21163d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 21173d14c5d2SYehuda Sadeh return rc; 21183d14c5d2SYehuda Sadeh } 21193d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_start_request); 21203d14c5d2SYehuda Sadeh 21213d14c5d2SYehuda Sadeh /* 21223d14c5d2SYehuda Sadeh * wait for a request to complete 21233d14c5d2SYehuda Sadeh */ 21243d14c5d2SYehuda Sadeh int ceph_osdc_wait_request(struct ceph_osd_client *osdc, 21253d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 21263d14c5d2SYehuda Sadeh { 21273d14c5d2SYehuda Sadeh int rc; 21283d14c5d2SYehuda Sadeh 21293d14c5d2SYehuda Sadeh rc = wait_for_completion_interruptible(&req->r_completion); 21303d14c5d2SYehuda Sadeh if (rc < 0) { 21313d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 21323d14c5d2SYehuda Sadeh __cancel_request(req); 21333d14c5d2SYehuda Sadeh __unregister_request(osdc, req); 21343d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 213525845472SSage Weil complete_request(req); 21363d14c5d2SYehuda Sadeh dout("wait_request tid %llu canceled/timed out\n", req->r_tid); 21373d14c5d2SYehuda Sadeh return rc; 21383d14c5d2SYehuda Sadeh } 21393d14c5d2SYehuda Sadeh 21403d14c5d2SYehuda Sadeh dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result); 21413d14c5d2SYehuda Sadeh return req->r_result; 21423d14c5d2SYehuda Sadeh } 21433d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_wait_request); 21443d14c5d2SYehuda Sadeh 21453d14c5d2SYehuda Sadeh /* 21463d14c5d2SYehuda Sadeh * sync - wait for all in-flight requests to flush. avoid starvation. 21473d14c5d2SYehuda Sadeh */ 21483d14c5d2SYehuda Sadeh void ceph_osdc_sync(struct ceph_osd_client *osdc) 21493d14c5d2SYehuda Sadeh { 21503d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 21513d14c5d2SYehuda Sadeh u64 last_tid, next_tid = 0; 21523d14c5d2SYehuda Sadeh 21533d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 21543d14c5d2SYehuda Sadeh last_tid = osdc->last_tid; 21553d14c5d2SYehuda Sadeh while (1) { 21563d14c5d2SYehuda Sadeh req = __lookup_request_ge(osdc, next_tid); 21573d14c5d2SYehuda Sadeh if (!req) 21583d14c5d2SYehuda Sadeh break; 21593d14c5d2SYehuda Sadeh if (req->r_tid > last_tid) 21603d14c5d2SYehuda Sadeh break; 21613d14c5d2SYehuda Sadeh 21623d14c5d2SYehuda Sadeh next_tid = req->r_tid + 1; 21633d14c5d2SYehuda Sadeh if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0) 21643d14c5d2SYehuda Sadeh continue; 21653d14c5d2SYehuda Sadeh 21663d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 21673d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 21683d14c5d2SYehuda Sadeh dout("sync waiting on tid %llu (last is %llu)\n", 21693d14c5d2SYehuda Sadeh req->r_tid, last_tid); 21703d14c5d2SYehuda Sadeh wait_for_completion(&req->r_safe_completion); 21713d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 21723d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 21733d14c5d2SYehuda Sadeh } 21743d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 21753d14c5d2SYehuda Sadeh dout("sync done (thru tid %llu)\n", last_tid); 21763d14c5d2SYehuda Sadeh } 21773d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_sync); 21783d14c5d2SYehuda Sadeh 21793d14c5d2SYehuda Sadeh /* 21803d14c5d2SYehuda Sadeh * init, shutdown 21813d14c5d2SYehuda Sadeh */ 21823d14c5d2SYehuda Sadeh int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client) 21833d14c5d2SYehuda Sadeh { 21843d14c5d2SYehuda Sadeh int err; 21853d14c5d2SYehuda Sadeh 21863d14c5d2SYehuda Sadeh dout("init\n"); 21873d14c5d2SYehuda Sadeh osdc->client = client; 21883d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 21893d14c5d2SYehuda Sadeh init_rwsem(&osdc->map_sem); 21903d14c5d2SYehuda Sadeh init_completion(&osdc->map_waiters); 21913d14c5d2SYehuda Sadeh osdc->last_requested_map = 0; 21923d14c5d2SYehuda Sadeh mutex_init(&osdc->request_mutex); 21933d14c5d2SYehuda Sadeh osdc->last_tid = 0; 21943d14c5d2SYehuda Sadeh osdc->osds = RB_ROOT; 21953d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->osd_lru); 21963d14c5d2SYehuda Sadeh osdc->requests = RB_ROOT; 21973d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_lru); 21986f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_unsent); 21996f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_notarget); 2200a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_linger); 22013d14c5d2SYehuda Sadeh osdc->num_requests = 0; 22023d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout); 22033d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout); 2204a40c4f10SYehuda Sadeh spin_lock_init(&osdc->event_lock); 2205a40c4f10SYehuda Sadeh osdc->event_tree = RB_ROOT; 2206a40c4f10SYehuda Sadeh osdc->event_count = 0; 22073d14c5d2SYehuda Sadeh 22083d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 22093d14c5d2SYehuda Sadeh round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ)); 22103d14c5d2SYehuda Sadeh 22113d14c5d2SYehuda Sadeh err = -ENOMEM; 22123d14c5d2SYehuda Sadeh osdc->req_mempool = mempool_create_kmalloc_pool(10, 22133d14c5d2SYehuda Sadeh sizeof(struct ceph_osd_request)); 22143d14c5d2SYehuda Sadeh if (!osdc->req_mempool) 22153d14c5d2SYehuda Sadeh goto out; 22163d14c5d2SYehuda Sadeh 2217d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP, 2218d50b409fSSage Weil OSD_OP_FRONT_LEN, 10, true, 22193d14c5d2SYehuda Sadeh "osd_op"); 22203d14c5d2SYehuda Sadeh if (err < 0) 22213d14c5d2SYehuda Sadeh goto out_mempool; 2222d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY, 22233d14c5d2SYehuda Sadeh OSD_OPREPLY_FRONT_LEN, 10, true, 22243d14c5d2SYehuda Sadeh "osd_op_reply"); 22253d14c5d2SYehuda Sadeh if (err < 0) 22263d14c5d2SYehuda Sadeh goto out_msgpool; 2227a40c4f10SYehuda Sadeh 2228a40c4f10SYehuda Sadeh osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify"); 2229a40c4f10SYehuda Sadeh if (IS_ERR(osdc->notify_wq)) { 2230a40c4f10SYehuda Sadeh err = PTR_ERR(osdc->notify_wq); 2231a40c4f10SYehuda Sadeh osdc->notify_wq = NULL; 2232a40c4f10SYehuda Sadeh goto out_msgpool; 2233a40c4f10SYehuda Sadeh } 22343d14c5d2SYehuda Sadeh return 0; 22353d14c5d2SYehuda Sadeh 22363d14c5d2SYehuda Sadeh out_msgpool: 22373d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 22383d14c5d2SYehuda Sadeh out_mempool: 22393d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 22403d14c5d2SYehuda Sadeh out: 22413d14c5d2SYehuda Sadeh return err; 22423d14c5d2SYehuda Sadeh } 22433d14c5d2SYehuda Sadeh 22443d14c5d2SYehuda Sadeh void ceph_osdc_stop(struct ceph_osd_client *osdc) 22453d14c5d2SYehuda Sadeh { 2246a40c4f10SYehuda Sadeh flush_workqueue(osdc->notify_wq); 2247a40c4f10SYehuda Sadeh destroy_workqueue(osdc->notify_wq); 22483d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->timeout_work); 22493d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->osds_timeout_work); 22503d14c5d2SYehuda Sadeh if (osdc->osdmap) { 22513d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 22523d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 22533d14c5d2SYehuda Sadeh } 2254aca420bcSSage Weil remove_all_osds(osdc); 22553d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 22563d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 22573d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op_reply); 22583d14c5d2SYehuda Sadeh } 22593d14c5d2SYehuda Sadeh 22603d14c5d2SYehuda Sadeh /* 22613d14c5d2SYehuda Sadeh * Read some contiguous pages. If we cross a stripe boundary, shorten 22623d14c5d2SYehuda Sadeh * *plen. Return number of bytes read, or error. 22633d14c5d2SYehuda Sadeh */ 22643d14c5d2SYehuda Sadeh int ceph_osdc_readpages(struct ceph_osd_client *osdc, 22653d14c5d2SYehuda Sadeh struct ceph_vino vino, struct ceph_file_layout *layout, 22663d14c5d2SYehuda Sadeh u64 off, u64 *plen, 22673d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 2268b7495fc2SSage Weil struct page **pages, int num_pages, int page_align) 22693d14c5d2SYehuda Sadeh { 22703d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 22713d14c5d2SYehuda Sadeh int rc = 0; 22723d14c5d2SYehuda Sadeh 22733d14c5d2SYehuda Sadeh dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino, 22743d14c5d2SYehuda Sadeh vino.snap, off, *plen); 227579528734SAlex Elder req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 1, 22763d14c5d2SYehuda Sadeh CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ, 2277acead002SAlex Elder NULL, truncate_seq, truncate_size, 2278153e5167SAlex Elder false); 22796816282dSSage Weil if (IS_ERR(req)) 22806816282dSSage Weil return PTR_ERR(req); 22813d14c5d2SYehuda Sadeh 22823d14c5d2SYehuda Sadeh /* it may be a short read due to an object boundary */ 22830fff87ecSAlex Elder 2284406e2c9fSAlex Elder osd_req_op_extent_osd_data_pages(req, 0, 2285a4ce40a9SAlex Elder pages, *plen, page_align, false, false); 22863d14c5d2SYehuda Sadeh 2287e0c59487SAlex Elder dout("readpages final extent is %llu~%llu (%llu bytes align %d)\n", 228843bfe5deSAlex Elder off, *plen, *plen, page_align); 22893d14c5d2SYehuda Sadeh 229079528734SAlex Elder ceph_osdc_build_request(req, off, NULL, vino.snap, NULL); 229102ee07d3SAlex Elder 22923d14c5d2SYehuda Sadeh rc = ceph_osdc_start_request(osdc, req, false); 22933d14c5d2SYehuda Sadeh if (!rc) 22943d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 22953d14c5d2SYehuda Sadeh 22963d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 22973d14c5d2SYehuda Sadeh dout("readpages result %d\n", rc); 22983d14c5d2SYehuda Sadeh return rc; 22993d14c5d2SYehuda Sadeh } 23003d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_readpages); 23013d14c5d2SYehuda Sadeh 23023d14c5d2SYehuda Sadeh /* 23033d14c5d2SYehuda Sadeh * do a synchronous write on N pages 23043d14c5d2SYehuda Sadeh */ 23053d14c5d2SYehuda Sadeh int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino, 23063d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 23073d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 23083d14c5d2SYehuda Sadeh u64 off, u64 len, 23093d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 23103d14c5d2SYehuda Sadeh struct timespec *mtime, 231124808826SAlex Elder struct page **pages, int num_pages) 23123d14c5d2SYehuda Sadeh { 23133d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 23143d14c5d2SYehuda Sadeh int rc = 0; 2315b7495fc2SSage Weil int page_align = off & ~PAGE_MASK; 23163d14c5d2SYehuda Sadeh 2317acead002SAlex Elder BUG_ON(vino.snap != CEPH_NOSNAP); /* snapshots aren't writeable */ 231879528734SAlex Elder req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 1, 23193d14c5d2SYehuda Sadeh CEPH_OSD_OP_WRITE, 232024808826SAlex Elder CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE, 2321acead002SAlex Elder snapc, truncate_seq, truncate_size, 2322153e5167SAlex Elder true); 23236816282dSSage Weil if (IS_ERR(req)) 23246816282dSSage Weil return PTR_ERR(req); 23253d14c5d2SYehuda Sadeh 23263d14c5d2SYehuda Sadeh /* it may be a short write due to an object boundary */ 2327406e2c9fSAlex Elder osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align, 232843bfe5deSAlex Elder false, false); 232943bfe5deSAlex Elder dout("writepages %llu~%llu (%llu bytes)\n", off, len, len); 23303d14c5d2SYehuda Sadeh 233179528734SAlex Elder ceph_osdc_build_request(req, off, snapc, CEPH_NOSNAP, mtime); 233202ee07d3SAlex Elder 233387f979d3SAlex Elder rc = ceph_osdc_start_request(osdc, req, true); 23343d14c5d2SYehuda Sadeh if (!rc) 23353d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 23363d14c5d2SYehuda Sadeh 23373d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 23383d14c5d2SYehuda Sadeh if (rc == 0) 23393d14c5d2SYehuda Sadeh rc = len; 23403d14c5d2SYehuda Sadeh dout("writepages result %d\n", rc); 23413d14c5d2SYehuda Sadeh return rc; 23423d14c5d2SYehuda Sadeh } 23433d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_writepages); 23443d14c5d2SYehuda Sadeh 23453d14c5d2SYehuda Sadeh /* 23463d14c5d2SYehuda Sadeh * handle incoming message 23473d14c5d2SYehuda Sadeh */ 23483d14c5d2SYehuda Sadeh static void dispatch(struct ceph_connection *con, struct ceph_msg *msg) 23493d14c5d2SYehuda Sadeh { 23503d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 23513d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 23523d14c5d2SYehuda Sadeh int type = le16_to_cpu(msg->hdr.type); 23533d14c5d2SYehuda Sadeh 23543d14c5d2SYehuda Sadeh if (!osd) 23553d14c5d2SYehuda Sadeh goto out; 23563d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 23573d14c5d2SYehuda Sadeh 23583d14c5d2SYehuda Sadeh switch (type) { 23593d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 23603d14c5d2SYehuda Sadeh ceph_osdc_handle_map(osdc, msg); 23613d14c5d2SYehuda Sadeh break; 23623d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 23633d14c5d2SYehuda Sadeh handle_reply(osdc, msg, con); 23643d14c5d2SYehuda Sadeh break; 2365a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 2366a40c4f10SYehuda Sadeh handle_watch_notify(osdc, msg); 2367a40c4f10SYehuda Sadeh break; 23683d14c5d2SYehuda Sadeh 23693d14c5d2SYehuda Sadeh default: 23703d14c5d2SYehuda Sadeh pr_err("received unknown message type %d %s\n", type, 23713d14c5d2SYehuda Sadeh ceph_msg_type_name(type)); 23723d14c5d2SYehuda Sadeh } 23733d14c5d2SYehuda Sadeh out: 23743d14c5d2SYehuda Sadeh ceph_msg_put(msg); 23753d14c5d2SYehuda Sadeh } 23763d14c5d2SYehuda Sadeh 23773d14c5d2SYehuda Sadeh /* 23783d14c5d2SYehuda Sadeh * lookup and return message for incoming reply. set up reply message 23793d14c5d2SYehuda Sadeh * pages. 23803d14c5d2SYehuda Sadeh */ 23813d14c5d2SYehuda Sadeh static struct ceph_msg *get_reply(struct ceph_connection *con, 23823d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 23833d14c5d2SYehuda Sadeh int *skip) 23843d14c5d2SYehuda Sadeh { 23853d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 23863d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = osd->o_osdc; 23873d14c5d2SYehuda Sadeh struct ceph_msg *m; 23883d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 23893d14c5d2SYehuda Sadeh int front = le32_to_cpu(hdr->front_len); 23903d14c5d2SYehuda Sadeh int data_len = le32_to_cpu(hdr->data_len); 23913d14c5d2SYehuda Sadeh u64 tid; 23923d14c5d2SYehuda Sadeh 23933d14c5d2SYehuda Sadeh tid = le64_to_cpu(hdr->tid); 23943d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 23953d14c5d2SYehuda Sadeh req = __lookup_request(osdc, tid); 23963d14c5d2SYehuda Sadeh if (!req) { 23973d14c5d2SYehuda Sadeh *skip = 1; 23983d14c5d2SYehuda Sadeh m = NULL; 2399756a16a5SSage Weil dout("get_reply unknown tid %llu from osd%d\n", tid, 24003d14c5d2SYehuda Sadeh osd->o_osd); 24013d14c5d2SYehuda Sadeh goto out; 24023d14c5d2SYehuda Sadeh } 24033d14c5d2SYehuda Sadeh 2404ace6d3a9SAlex Elder if (req->r_reply->con) 24058921d114SAlex Elder dout("%s revoking msg %p from old con %p\n", __func__, 2406ace6d3a9SAlex Elder req->r_reply, req->r_reply->con); 24078921d114SAlex Elder ceph_msg_revoke_incoming(req->r_reply); 24083d14c5d2SYehuda Sadeh 24093d14c5d2SYehuda Sadeh if (front > req->r_reply->front.iov_len) { 24103d14c5d2SYehuda Sadeh pr_warning("get_reply front %d > preallocated %d\n", 24113d14c5d2SYehuda Sadeh front, (int)req->r_reply->front.iov_len); 2412b61c2763SSage Weil m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS, false); 24133d14c5d2SYehuda Sadeh if (!m) 24143d14c5d2SYehuda Sadeh goto out; 24153d14c5d2SYehuda Sadeh ceph_msg_put(req->r_reply); 24163d14c5d2SYehuda Sadeh req->r_reply = m; 24173d14c5d2SYehuda Sadeh } 24183d14c5d2SYehuda Sadeh m = ceph_msg_get(req->r_reply); 24193d14c5d2SYehuda Sadeh 24203d14c5d2SYehuda Sadeh if (data_len > 0) { 2421a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 24220fff87ecSAlex Elder 2423a4ce40a9SAlex Elder /* 2424a4ce40a9SAlex Elder * XXX This is assuming there is only one op containing 2425a4ce40a9SAlex Elder * XXX page data. Probably OK for reads, but this 2426a4ce40a9SAlex Elder * XXX ought to be done more generally. 2427a4ce40a9SAlex Elder */ 2428406e2c9fSAlex Elder osd_data = osd_req_op_extent_osd_data(req, 0); 24290fff87ecSAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 24300fff87ecSAlex Elder if (osd_data->pages && 2431e0c59487SAlex Elder unlikely(osd_data->length < data_len)) { 24322ac2b7a6SAlex Elder 2433e0c59487SAlex Elder pr_warning("tid %lld reply has %d bytes " 2434e0c59487SAlex Elder "we had only %llu bytes ready\n", 2435e0c59487SAlex Elder tid, data_len, osd_data->length); 24363d14c5d2SYehuda Sadeh *skip = 1; 24373d14c5d2SYehuda Sadeh ceph_msg_put(m); 24383d14c5d2SYehuda Sadeh m = NULL; 24393d14c5d2SYehuda Sadeh goto out; 24403d14c5d2SYehuda Sadeh } 24413d14c5d2SYehuda Sadeh } 24422ac2b7a6SAlex Elder } 24433d14c5d2SYehuda Sadeh *skip = 0; 24443d14c5d2SYehuda Sadeh dout("get_reply tid %lld %p\n", tid, m); 24453d14c5d2SYehuda Sadeh 24463d14c5d2SYehuda Sadeh out: 24473d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 24483d14c5d2SYehuda Sadeh return m; 24493d14c5d2SYehuda Sadeh 24503d14c5d2SYehuda Sadeh } 24513d14c5d2SYehuda Sadeh 24523d14c5d2SYehuda Sadeh static struct ceph_msg *alloc_msg(struct ceph_connection *con, 24533d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 24543d14c5d2SYehuda Sadeh int *skip) 24553d14c5d2SYehuda Sadeh { 24563d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 24573d14c5d2SYehuda Sadeh int type = le16_to_cpu(hdr->type); 24583d14c5d2SYehuda Sadeh int front = le32_to_cpu(hdr->front_len); 24593d14c5d2SYehuda Sadeh 24601c20f2d2SAlex Elder *skip = 0; 24613d14c5d2SYehuda Sadeh switch (type) { 24623d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 2463a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 2464b61c2763SSage Weil return ceph_msg_new(type, front, GFP_NOFS, false); 24653d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 24663d14c5d2SYehuda Sadeh return get_reply(con, hdr, skip); 24673d14c5d2SYehuda Sadeh default: 24683d14c5d2SYehuda Sadeh pr_info("alloc_msg unexpected msg type %d from osd%d\n", type, 24693d14c5d2SYehuda Sadeh osd->o_osd); 24703d14c5d2SYehuda Sadeh *skip = 1; 24713d14c5d2SYehuda Sadeh return NULL; 24723d14c5d2SYehuda Sadeh } 24733d14c5d2SYehuda Sadeh } 24743d14c5d2SYehuda Sadeh 24753d14c5d2SYehuda Sadeh /* 24763d14c5d2SYehuda Sadeh * Wrappers to refcount containing ceph_osd struct 24773d14c5d2SYehuda Sadeh */ 24783d14c5d2SYehuda Sadeh static struct ceph_connection *get_osd_con(struct ceph_connection *con) 24793d14c5d2SYehuda Sadeh { 24803d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 24813d14c5d2SYehuda Sadeh if (get_osd(osd)) 24823d14c5d2SYehuda Sadeh return con; 24833d14c5d2SYehuda Sadeh return NULL; 24843d14c5d2SYehuda Sadeh } 24853d14c5d2SYehuda Sadeh 24863d14c5d2SYehuda Sadeh static void put_osd_con(struct ceph_connection *con) 24873d14c5d2SYehuda Sadeh { 24883d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 24893d14c5d2SYehuda Sadeh put_osd(osd); 24903d14c5d2SYehuda Sadeh } 24913d14c5d2SYehuda Sadeh 24923d14c5d2SYehuda Sadeh /* 24933d14c5d2SYehuda Sadeh * authentication 24943d14c5d2SYehuda Sadeh */ 2495a3530df3SAlex Elder /* 2496a3530df3SAlex Elder * Note: returned pointer is the address of a structure that's 2497a3530df3SAlex Elder * managed separately. Caller must *not* attempt to free it. 2498a3530df3SAlex Elder */ 2499a3530df3SAlex Elder static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con, 25008f43fb53SAlex Elder int *proto, int force_new) 25013d14c5d2SYehuda Sadeh { 25023d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 25033d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 25043d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 250574f1869fSAlex Elder struct ceph_auth_handshake *auth = &o->o_auth; 25063d14c5d2SYehuda Sadeh 250774f1869fSAlex Elder if (force_new && auth->authorizer) { 250827859f97SSage Weil ceph_auth_destroy_authorizer(ac, auth->authorizer); 250974f1869fSAlex Elder auth->authorizer = NULL; 25103d14c5d2SYehuda Sadeh } 251127859f97SSage Weil if (!auth->authorizer) { 251227859f97SSage Weil int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 2513a3530df3SAlex Elder auth); 25143d14c5d2SYehuda Sadeh if (ret) 2515a3530df3SAlex Elder return ERR_PTR(ret); 251627859f97SSage Weil } else { 251727859f97SSage Weil int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 25180bed9b5cSSage Weil auth); 25190bed9b5cSSage Weil if (ret) 25200bed9b5cSSage Weil return ERR_PTR(ret); 25213d14c5d2SYehuda Sadeh } 25223d14c5d2SYehuda Sadeh *proto = ac->protocol; 252374f1869fSAlex Elder 2524a3530df3SAlex Elder return auth; 25253d14c5d2SYehuda Sadeh } 25263d14c5d2SYehuda Sadeh 25273d14c5d2SYehuda Sadeh 25283d14c5d2SYehuda Sadeh static int verify_authorizer_reply(struct ceph_connection *con, int len) 25293d14c5d2SYehuda Sadeh { 25303d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 25313d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 25323d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 25333d14c5d2SYehuda Sadeh 253427859f97SSage Weil return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len); 25353d14c5d2SYehuda Sadeh } 25363d14c5d2SYehuda Sadeh 25373d14c5d2SYehuda Sadeh static int invalidate_authorizer(struct ceph_connection *con) 25383d14c5d2SYehuda Sadeh { 25393d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 25403d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 25413d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 25423d14c5d2SYehuda Sadeh 254327859f97SSage Weil ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD); 25443d14c5d2SYehuda Sadeh return ceph_monc_validate_auth(&osdc->client->monc); 25453d14c5d2SYehuda Sadeh } 25463d14c5d2SYehuda Sadeh 25473d14c5d2SYehuda Sadeh static const struct ceph_connection_operations osd_con_ops = { 25483d14c5d2SYehuda Sadeh .get = get_osd_con, 25493d14c5d2SYehuda Sadeh .put = put_osd_con, 25503d14c5d2SYehuda Sadeh .dispatch = dispatch, 25513d14c5d2SYehuda Sadeh .get_authorizer = get_authorizer, 25523d14c5d2SYehuda Sadeh .verify_authorizer_reply = verify_authorizer_reply, 25533d14c5d2SYehuda Sadeh .invalidate_authorizer = invalidate_authorizer, 25543d14c5d2SYehuda Sadeh .alloc_msg = alloc_msg, 25553d14c5d2SYehuda Sadeh .fault = osd_reset, 25563d14c5d2SYehuda Sadeh }; 2557