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 255522ae0bSAlex Elder static struct kmem_cache *ceph_osd_request_cache; 265522ae0bSAlex Elder 273d14c5d2SYehuda Sadeh static const struct ceph_connection_operations osd_con_ops; 283d14c5d2SYehuda Sadeh 29f9d25199SAlex Elder static void __send_queued(struct ceph_osd_client *osdc); 306f6c7006SSage Weil static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd); 31a40c4f10SYehuda Sadeh static void __register_request(struct ceph_osd_client *osdc, 32a40c4f10SYehuda Sadeh struct ceph_osd_request *req); 33a40c4f10SYehuda Sadeh static void __unregister_linger_request(struct ceph_osd_client *osdc, 34a40c4f10SYehuda Sadeh struct ceph_osd_request *req); 3556e925b6SSage Weil static void __send_request(struct ceph_osd_client *osdc, 36a40c4f10SYehuda Sadeh struct ceph_osd_request *req); 373d14c5d2SYehuda Sadeh 383d14c5d2SYehuda Sadeh /* 393d14c5d2SYehuda Sadeh * Implement client access to distributed object storage cluster. 403d14c5d2SYehuda Sadeh * 413d14c5d2SYehuda Sadeh * All data objects are stored within a cluster/cloud of OSDs, or 423d14c5d2SYehuda Sadeh * "object storage devices." (Note that Ceph OSDs have _nothing_ to 433d14c5d2SYehuda Sadeh * do with the T10 OSD extensions to SCSI.) Ceph OSDs are simply 443d14c5d2SYehuda Sadeh * remote daemons serving up and coordinating consistent and safe 453d14c5d2SYehuda Sadeh * access to storage. 463d14c5d2SYehuda Sadeh * 473d14c5d2SYehuda Sadeh * Cluster membership and the mapping of data objects onto storage devices 483d14c5d2SYehuda Sadeh * are described by the osd map. 493d14c5d2SYehuda Sadeh * 503d14c5d2SYehuda Sadeh * We keep track of pending OSD requests (read, write), resubmit 513d14c5d2SYehuda Sadeh * requests to different OSDs when the cluster topology/data layout 523d14c5d2SYehuda Sadeh * change, or retry the affected requests when the communications 533d14c5d2SYehuda Sadeh * channel with an OSD is reset. 543d14c5d2SYehuda Sadeh */ 553d14c5d2SYehuda Sadeh 563d14c5d2SYehuda Sadeh /* 573d14c5d2SYehuda Sadeh * calculate the mapping of a file extent onto an object, and fill out the 583d14c5d2SYehuda Sadeh * request accordingly. shorten extent as necessary if it crosses an 593d14c5d2SYehuda Sadeh * object boundary. 603d14c5d2SYehuda Sadeh * 613d14c5d2SYehuda Sadeh * fill osd op in request message. 623d14c5d2SYehuda Sadeh */ 63dbe0fc41SAlex Elder static int calc_layout(struct ceph_file_layout *layout, u64 off, u64 *plen, 64a19dadfbSAlex Elder u64 *objnum, u64 *objoff, u64 *objlen) 653d14c5d2SYehuda Sadeh { 6660e56f13SAlex Elder u64 orig_len = *plen; 67d63b77f4SSage Weil int r; 683d14c5d2SYehuda Sadeh 6960e56f13SAlex Elder /* object extent? */ 7075d1c941SAlex Elder r = ceph_calc_file_object_mapping(layout, off, orig_len, objnum, 7175d1c941SAlex Elder objoff, objlen); 72d63b77f4SSage Weil if (r < 0) 73d63b77f4SSage Weil return r; 7475d1c941SAlex Elder if (*objlen < orig_len) { 7575d1c941SAlex Elder *plen = *objlen; 7660e56f13SAlex Elder dout(" skipping last %llu, final file extent %llu~%llu\n", 7760e56f13SAlex Elder orig_len - *plen, off, *plen); 7860e56f13SAlex Elder } 7960e56f13SAlex Elder 8075d1c941SAlex Elder dout("calc_layout objnum=%llx %llu~%llu\n", *objnum, *objoff, *objlen); 813d14c5d2SYehuda Sadeh 823ff5f385SAlex Elder return 0; 833d14c5d2SYehuda Sadeh } 843d14c5d2SYehuda Sadeh 85c54d47bfSAlex Elder static void ceph_osd_data_init(struct ceph_osd_data *osd_data) 86c54d47bfSAlex Elder { 87c54d47bfSAlex Elder memset(osd_data, 0, sizeof (*osd_data)); 88c54d47bfSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_NONE; 89c54d47bfSAlex Elder } 90c54d47bfSAlex Elder 91a4ce40a9SAlex Elder static void ceph_osd_data_pages_init(struct ceph_osd_data *osd_data, 9243bfe5deSAlex Elder struct page **pages, u64 length, u32 alignment, 9343bfe5deSAlex Elder bool pages_from_pool, bool own_pages) 9443bfe5deSAlex Elder { 9543bfe5deSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_PAGES; 9643bfe5deSAlex Elder osd_data->pages = pages; 9743bfe5deSAlex Elder osd_data->length = length; 9843bfe5deSAlex Elder osd_data->alignment = alignment; 9943bfe5deSAlex Elder osd_data->pages_from_pool = pages_from_pool; 10043bfe5deSAlex Elder osd_data->own_pages = own_pages; 10143bfe5deSAlex Elder } 10243bfe5deSAlex Elder 103a4ce40a9SAlex Elder static void ceph_osd_data_pagelist_init(struct ceph_osd_data *osd_data, 10443bfe5deSAlex Elder struct ceph_pagelist *pagelist) 10543bfe5deSAlex Elder { 10643bfe5deSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_PAGELIST; 10743bfe5deSAlex Elder osd_data->pagelist = pagelist; 10843bfe5deSAlex Elder } 10943bfe5deSAlex Elder 11043bfe5deSAlex Elder #ifdef CONFIG_BLOCK 111a4ce40a9SAlex Elder static void ceph_osd_data_bio_init(struct ceph_osd_data *osd_data, 11243bfe5deSAlex Elder struct bio *bio, size_t bio_length) 11343bfe5deSAlex Elder { 11443bfe5deSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_BIO; 11543bfe5deSAlex Elder osd_data->bio = bio; 11643bfe5deSAlex Elder osd_data->bio_length = bio_length; 11743bfe5deSAlex Elder } 11843bfe5deSAlex Elder #endif /* CONFIG_BLOCK */ 11943bfe5deSAlex Elder 120863c7eb5SAlex Elder #define osd_req_op_data(oreq, whch, typ, fld) \ 121863c7eb5SAlex Elder ({ \ 122863c7eb5SAlex Elder BUG_ON(whch >= (oreq)->r_num_ops); \ 123863c7eb5SAlex Elder &(oreq)->r_ops[whch].typ.fld; \ 124863c7eb5SAlex Elder }) 125863c7eb5SAlex Elder 12649719778SAlex Elder static struct ceph_osd_data * 12749719778SAlex Elder osd_req_op_raw_data_in(struct ceph_osd_request *osd_req, unsigned int which) 12849719778SAlex Elder { 12949719778SAlex Elder BUG_ON(which >= osd_req->r_num_ops); 13049719778SAlex Elder 13149719778SAlex Elder return &osd_req->r_ops[which].raw_data_in; 13249719778SAlex Elder } 13349719778SAlex Elder 134a4ce40a9SAlex Elder struct ceph_osd_data * 135a4ce40a9SAlex Elder osd_req_op_extent_osd_data(struct ceph_osd_request *osd_req, 136406e2c9fSAlex Elder unsigned int which) 137a4ce40a9SAlex Elder { 138863c7eb5SAlex Elder return osd_req_op_data(osd_req, which, extent, osd_data); 139a4ce40a9SAlex Elder } 140a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data); 141a4ce40a9SAlex Elder 142a4ce40a9SAlex Elder struct ceph_osd_data * 143a4ce40a9SAlex Elder osd_req_op_cls_response_data(struct ceph_osd_request *osd_req, 144a4ce40a9SAlex Elder unsigned int which) 145a4ce40a9SAlex Elder { 146863c7eb5SAlex Elder return osd_req_op_data(osd_req, which, cls, response_data); 147a4ce40a9SAlex Elder } 148a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_response_data); /* ??? */ 149a4ce40a9SAlex Elder 15049719778SAlex Elder void osd_req_op_raw_data_in_pages(struct ceph_osd_request *osd_req, 15149719778SAlex Elder unsigned int which, struct page **pages, 15249719778SAlex Elder u64 length, u32 alignment, 15349719778SAlex Elder bool pages_from_pool, bool own_pages) 15449719778SAlex Elder { 15549719778SAlex Elder struct ceph_osd_data *osd_data; 15649719778SAlex Elder 15749719778SAlex Elder osd_data = osd_req_op_raw_data_in(osd_req, which); 15849719778SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 15949719778SAlex Elder pages_from_pool, own_pages); 16049719778SAlex Elder } 16149719778SAlex Elder EXPORT_SYMBOL(osd_req_op_raw_data_in_pages); 16249719778SAlex Elder 163a4ce40a9SAlex Elder void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *osd_req, 164406e2c9fSAlex Elder unsigned int which, struct page **pages, 165406e2c9fSAlex Elder u64 length, u32 alignment, 166a4ce40a9SAlex Elder bool pages_from_pool, bool own_pages) 167a4ce40a9SAlex Elder { 168a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 169a4ce40a9SAlex Elder 170863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, extent, osd_data); 171a4ce40a9SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 172a4ce40a9SAlex Elder pages_from_pool, own_pages); 173a4ce40a9SAlex Elder } 174a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data_pages); 175a4ce40a9SAlex Elder 176a4ce40a9SAlex Elder void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *osd_req, 177406e2c9fSAlex Elder unsigned int which, struct ceph_pagelist *pagelist) 178a4ce40a9SAlex Elder { 179a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 180a4ce40a9SAlex Elder 181863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, extent, osd_data); 182a4ce40a9SAlex Elder ceph_osd_data_pagelist_init(osd_data, pagelist); 183a4ce40a9SAlex Elder } 184a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data_pagelist); 185a4ce40a9SAlex Elder 186a4ce40a9SAlex Elder #ifdef CONFIG_BLOCK 187a4ce40a9SAlex Elder void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *osd_req, 188406e2c9fSAlex Elder unsigned int which, struct bio *bio, size_t bio_length) 189a4ce40a9SAlex Elder { 190a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 191863c7eb5SAlex Elder 192863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, extent, osd_data); 193a4ce40a9SAlex Elder ceph_osd_data_bio_init(osd_data, bio, bio_length); 194a4ce40a9SAlex Elder } 195a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data_bio); 196a4ce40a9SAlex Elder #endif /* CONFIG_BLOCK */ 197a4ce40a9SAlex Elder 198a4ce40a9SAlex Elder static void osd_req_op_cls_request_info_pagelist( 199a4ce40a9SAlex Elder struct ceph_osd_request *osd_req, 200a4ce40a9SAlex Elder unsigned int which, struct ceph_pagelist *pagelist) 201a4ce40a9SAlex Elder { 202a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 203a4ce40a9SAlex Elder 204863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, request_info); 205a4ce40a9SAlex Elder ceph_osd_data_pagelist_init(osd_data, pagelist); 206a4ce40a9SAlex Elder } 207a4ce40a9SAlex Elder 20804017e29SAlex Elder void osd_req_op_cls_request_data_pagelist( 20904017e29SAlex Elder struct ceph_osd_request *osd_req, 21004017e29SAlex Elder unsigned int which, struct ceph_pagelist *pagelist) 21104017e29SAlex Elder { 21204017e29SAlex Elder struct ceph_osd_data *osd_data; 21304017e29SAlex Elder 214863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, request_data); 21504017e29SAlex Elder ceph_osd_data_pagelist_init(osd_data, pagelist); 21604017e29SAlex Elder } 21704017e29SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_request_data_pagelist); 21804017e29SAlex Elder 2196c57b554SAlex Elder void osd_req_op_cls_request_data_pages(struct ceph_osd_request *osd_req, 2206c57b554SAlex Elder unsigned int which, struct page **pages, u64 length, 2216c57b554SAlex Elder u32 alignment, bool pages_from_pool, bool own_pages) 2226c57b554SAlex Elder { 2236c57b554SAlex Elder struct ceph_osd_data *osd_data; 2246c57b554SAlex Elder 2256c57b554SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, request_data); 2266c57b554SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 2276c57b554SAlex Elder pages_from_pool, own_pages); 2286c57b554SAlex Elder } 2296c57b554SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_request_data_pages); 2306c57b554SAlex Elder 231a4ce40a9SAlex Elder void osd_req_op_cls_response_data_pages(struct ceph_osd_request *osd_req, 232a4ce40a9SAlex Elder unsigned int which, struct page **pages, u64 length, 233a4ce40a9SAlex Elder u32 alignment, bool pages_from_pool, bool own_pages) 234a4ce40a9SAlex Elder { 235a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 236a4ce40a9SAlex Elder 237863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, response_data); 238a4ce40a9SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 239a4ce40a9SAlex Elder pages_from_pool, own_pages); 240a4ce40a9SAlex Elder } 241a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_response_data_pages); 242a4ce40a9SAlex Elder 24323c08a9cSAlex Elder static u64 ceph_osd_data_length(struct ceph_osd_data *osd_data) 24423c08a9cSAlex Elder { 24523c08a9cSAlex Elder switch (osd_data->type) { 24623c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_NONE: 24723c08a9cSAlex Elder return 0; 24823c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_PAGES: 24923c08a9cSAlex Elder return osd_data->length; 25023c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_PAGELIST: 25123c08a9cSAlex Elder return (u64)osd_data->pagelist->length; 25223c08a9cSAlex Elder #ifdef CONFIG_BLOCK 25323c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_BIO: 25423c08a9cSAlex Elder return (u64)osd_data->bio_length; 25523c08a9cSAlex Elder #endif /* CONFIG_BLOCK */ 25623c08a9cSAlex Elder default: 25723c08a9cSAlex Elder WARN(true, "unrecognized data type %d\n", (int)osd_data->type); 25823c08a9cSAlex Elder return 0; 25923c08a9cSAlex Elder } 26023c08a9cSAlex Elder } 26123c08a9cSAlex Elder 262c54d47bfSAlex Elder static void ceph_osd_data_release(struct ceph_osd_data *osd_data) 263c54d47bfSAlex Elder { 2645476492fSAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES && osd_data->own_pages) { 265c54d47bfSAlex Elder int num_pages; 266c54d47bfSAlex Elder 267c54d47bfSAlex Elder num_pages = calc_pages_for((u64)osd_data->alignment, 268c54d47bfSAlex Elder (u64)osd_data->length); 269c54d47bfSAlex Elder ceph_release_page_vector(osd_data->pages, num_pages); 270c54d47bfSAlex Elder } 2715476492fSAlex Elder ceph_osd_data_init(osd_data); 2725476492fSAlex Elder } 2735476492fSAlex Elder 2745476492fSAlex Elder static void osd_req_op_data_release(struct ceph_osd_request *osd_req, 2755476492fSAlex Elder unsigned int which) 2765476492fSAlex Elder { 2775476492fSAlex Elder struct ceph_osd_req_op *op; 2785476492fSAlex Elder 2795476492fSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 2805476492fSAlex Elder op = &osd_req->r_ops[which]; 2815476492fSAlex Elder 2825476492fSAlex Elder switch (op->op) { 2835476492fSAlex Elder case CEPH_OSD_OP_READ: 2845476492fSAlex Elder case CEPH_OSD_OP_WRITE: 2855476492fSAlex Elder ceph_osd_data_release(&op->extent.osd_data); 2865476492fSAlex Elder break; 2875476492fSAlex Elder case CEPH_OSD_OP_CALL: 2885476492fSAlex Elder ceph_osd_data_release(&op->cls.request_info); 28904017e29SAlex Elder ceph_osd_data_release(&op->cls.request_data); 2905476492fSAlex Elder ceph_osd_data_release(&op->cls.response_data); 2915476492fSAlex Elder break; 2925476492fSAlex Elder default: 2935476492fSAlex Elder break; 2945476492fSAlex Elder } 295c54d47bfSAlex Elder } 296c54d47bfSAlex Elder 2973d14c5d2SYehuda Sadeh /* 2983d14c5d2SYehuda Sadeh * requests 2993d14c5d2SYehuda Sadeh */ 3003d14c5d2SYehuda Sadeh void ceph_osdc_release_request(struct kref *kref) 3013d14c5d2SYehuda Sadeh { 302c54d47bfSAlex Elder struct ceph_osd_request *req; 3035476492fSAlex Elder unsigned int which; 3043d14c5d2SYehuda Sadeh 305c54d47bfSAlex Elder req = container_of(kref, struct ceph_osd_request, r_kref); 3063d14c5d2SYehuda Sadeh if (req->r_request) 3073d14c5d2SYehuda Sadeh ceph_msg_put(req->r_request); 308ace6d3a9SAlex Elder if (req->r_reply) { 3098921d114SAlex Elder ceph_msg_revoke_incoming(req->r_reply); 310ab8cb34aSAlex Elder ceph_msg_put(req->r_reply); 311ace6d3a9SAlex Elder } 3120fff87ecSAlex Elder 3135476492fSAlex Elder for (which = 0; which < req->r_num_ops; which++) 3145476492fSAlex Elder osd_req_op_data_release(req, which); 3150fff87ecSAlex Elder 3163d14c5d2SYehuda Sadeh ceph_put_snap_context(req->r_snapc); 3173d14c5d2SYehuda Sadeh if (req->r_mempool) 3183d14c5d2SYehuda Sadeh mempool_free(req, req->r_osdc->req_mempool); 3193d14c5d2SYehuda Sadeh else 3205522ae0bSAlex Elder kmem_cache_free(ceph_osd_request_cache, req); 3215522ae0bSAlex Elder 3223d14c5d2SYehuda Sadeh } 3233d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_release_request); 3243d14c5d2SYehuda Sadeh 3253d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc, 3263d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 3271b83bef2SSage Weil unsigned int num_ops, 3283d14c5d2SYehuda Sadeh bool use_mempool, 32954a54007SAlex Elder gfp_t gfp_flags) 3303d14c5d2SYehuda Sadeh { 3313d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 3323d14c5d2SYehuda Sadeh struct ceph_msg *msg; 3331b83bef2SSage Weil size_t msg_size; 3343d14c5d2SYehuda Sadeh 33579528734SAlex Elder BUILD_BUG_ON(CEPH_OSD_MAX_OP > U16_MAX); 33679528734SAlex Elder BUG_ON(num_ops > CEPH_OSD_MAX_OP); 33779528734SAlex Elder 3381b83bef2SSage Weil msg_size = 4 + 4 + 8 + 8 + 4+8; 3391b83bef2SSage Weil msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */ 3401b83bef2SSage Weil msg_size += 1 + 8 + 4 + 4; /* pg_t */ 3412d0ebc5dSIlya Dryomov msg_size += 4 + CEPH_MAX_OID_NAME_LEN; /* oid */ 3421b83bef2SSage Weil msg_size += 2 + num_ops*sizeof(struct ceph_osd_op); 3431b83bef2SSage Weil msg_size += 8; /* snapid */ 3441b83bef2SSage Weil msg_size += 8; /* snap_seq */ 3451b83bef2SSage Weil msg_size += 8 * (snapc ? snapc->num_snaps : 0); /* snaps */ 3461b83bef2SSage Weil msg_size += 4; 3473d14c5d2SYehuda Sadeh 3483d14c5d2SYehuda Sadeh if (use_mempool) { 3493d14c5d2SYehuda Sadeh req = mempool_alloc(osdc->req_mempool, gfp_flags); 3503d14c5d2SYehuda Sadeh memset(req, 0, sizeof(*req)); 3513d14c5d2SYehuda Sadeh } else { 3525522ae0bSAlex Elder req = kmem_cache_zalloc(ceph_osd_request_cache, gfp_flags); 3533d14c5d2SYehuda Sadeh } 3543d14c5d2SYehuda Sadeh if (req == NULL) 3553d14c5d2SYehuda Sadeh return NULL; 3563d14c5d2SYehuda Sadeh 3573d14c5d2SYehuda Sadeh req->r_osdc = osdc; 3583d14c5d2SYehuda Sadeh req->r_mempool = use_mempool; 35979528734SAlex Elder req->r_num_ops = num_ops; 3603d14c5d2SYehuda Sadeh 3613d14c5d2SYehuda Sadeh kref_init(&req->r_kref); 3623d14c5d2SYehuda Sadeh init_completion(&req->r_completion); 3633d14c5d2SYehuda Sadeh init_completion(&req->r_safe_completion); 364a978fa20SAlex Elder RB_CLEAR_NODE(&req->r_node); 3653d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&req->r_unsafe_item); 366a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&req->r_linger_item); 367a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&req->r_linger_osd); 368935b639aSSage Weil INIT_LIST_HEAD(&req->r_req_lru_item); 369cd43045cSSage Weil INIT_LIST_HEAD(&req->r_osd_item); 370cd43045cSSage Weil 3713c972c95SIlya Dryomov req->r_base_oloc.pool = -1; 372205ee118SIlya Dryomov req->r_target_oloc.pool = -1; 37322116525SIlya Dryomov 3743d14c5d2SYehuda Sadeh /* create reply message */ 3753d14c5d2SYehuda Sadeh if (use_mempool) 3763d14c5d2SYehuda Sadeh msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0); 3773d14c5d2SYehuda Sadeh else 3783d14c5d2SYehuda Sadeh msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, 379b61c2763SSage Weil OSD_OPREPLY_FRONT_LEN, gfp_flags, true); 3803d14c5d2SYehuda Sadeh if (!msg) { 3813d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 3823d14c5d2SYehuda Sadeh return NULL; 3833d14c5d2SYehuda Sadeh } 3843d14c5d2SYehuda Sadeh req->r_reply = msg; 3853d14c5d2SYehuda Sadeh 3863d14c5d2SYehuda Sadeh /* create request message; allow space for oid */ 3873d14c5d2SYehuda Sadeh if (use_mempool) 3883d14c5d2SYehuda Sadeh msg = ceph_msgpool_get(&osdc->msgpool_op, 0); 3893d14c5d2SYehuda Sadeh else 390b61c2763SSage Weil msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true); 3913d14c5d2SYehuda Sadeh if (!msg) { 3923d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 3933d14c5d2SYehuda Sadeh return NULL; 3943d14c5d2SYehuda Sadeh } 3953d14c5d2SYehuda Sadeh 3963d14c5d2SYehuda Sadeh memset(msg->front.iov_base, 0, msg->front.iov_len); 3973d14c5d2SYehuda Sadeh 3983d14c5d2SYehuda Sadeh req->r_request = msg; 3993d14c5d2SYehuda Sadeh 4003d14c5d2SYehuda Sadeh return req; 4013d14c5d2SYehuda Sadeh } 4023d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_alloc_request); 4033d14c5d2SYehuda Sadeh 404a8dd0a37SAlex Elder static bool osd_req_opcode_valid(u16 opcode) 405a8dd0a37SAlex Elder { 406a8dd0a37SAlex Elder switch (opcode) { 407a8dd0a37SAlex Elder case CEPH_OSD_OP_READ: 408a8dd0a37SAlex Elder case CEPH_OSD_OP_STAT: 409a8dd0a37SAlex Elder case CEPH_OSD_OP_MAPEXT: 410a8dd0a37SAlex Elder case CEPH_OSD_OP_MASKTRUNC: 411a8dd0a37SAlex Elder case CEPH_OSD_OP_SPARSE_READ: 412a8dd0a37SAlex Elder case CEPH_OSD_OP_NOTIFY: 413a8dd0a37SAlex Elder case CEPH_OSD_OP_NOTIFY_ACK: 414a8dd0a37SAlex Elder case CEPH_OSD_OP_ASSERT_VER: 415a8dd0a37SAlex Elder case CEPH_OSD_OP_WRITE: 416a8dd0a37SAlex Elder case CEPH_OSD_OP_WRITEFULL: 417a8dd0a37SAlex Elder case CEPH_OSD_OP_TRUNCATE: 418a8dd0a37SAlex Elder case CEPH_OSD_OP_ZERO: 419a8dd0a37SAlex Elder case CEPH_OSD_OP_DELETE: 420a8dd0a37SAlex Elder case CEPH_OSD_OP_APPEND: 421a8dd0a37SAlex Elder case CEPH_OSD_OP_STARTSYNC: 422a8dd0a37SAlex Elder case CEPH_OSD_OP_SETTRUNC: 423a8dd0a37SAlex Elder case CEPH_OSD_OP_TRIMTRUNC: 424a8dd0a37SAlex Elder case CEPH_OSD_OP_TMAPUP: 425a8dd0a37SAlex Elder case CEPH_OSD_OP_TMAPPUT: 426a8dd0a37SAlex Elder case CEPH_OSD_OP_TMAPGET: 427a8dd0a37SAlex Elder case CEPH_OSD_OP_CREATE: 428a8dd0a37SAlex Elder case CEPH_OSD_OP_ROLLBACK: 429a8dd0a37SAlex Elder case CEPH_OSD_OP_WATCH: 430a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPGETKEYS: 431a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPGETVALS: 432a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPGETHEADER: 433a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPGETVALSBYKEYS: 434a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPSETVALS: 435a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPSETHEADER: 436a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPCLEAR: 437a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPRMKEYS: 438a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAP_CMP: 439a8dd0a37SAlex Elder case CEPH_OSD_OP_CLONERANGE: 440a8dd0a37SAlex Elder case CEPH_OSD_OP_ASSERT_SRC_VERSION: 441a8dd0a37SAlex Elder case CEPH_OSD_OP_SRC_CMPXATTR: 442a8dd0a37SAlex Elder case CEPH_OSD_OP_GETXATTR: 443a8dd0a37SAlex Elder case CEPH_OSD_OP_GETXATTRS: 444a8dd0a37SAlex Elder case CEPH_OSD_OP_CMPXATTR: 445a8dd0a37SAlex Elder case CEPH_OSD_OP_SETXATTR: 446a8dd0a37SAlex Elder case CEPH_OSD_OP_SETXATTRS: 447a8dd0a37SAlex Elder case CEPH_OSD_OP_RESETXATTRS: 448a8dd0a37SAlex Elder case CEPH_OSD_OP_RMXATTR: 449a8dd0a37SAlex Elder case CEPH_OSD_OP_PULL: 450a8dd0a37SAlex Elder case CEPH_OSD_OP_PUSH: 451a8dd0a37SAlex Elder case CEPH_OSD_OP_BALANCEREADS: 452a8dd0a37SAlex Elder case CEPH_OSD_OP_UNBALANCEREADS: 453a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB: 454a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB_RESERVE: 455a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB_UNRESERVE: 456a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB_STOP: 457a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB_MAP: 458a8dd0a37SAlex Elder case CEPH_OSD_OP_WRLOCK: 459a8dd0a37SAlex Elder case CEPH_OSD_OP_WRUNLOCK: 460a8dd0a37SAlex Elder case CEPH_OSD_OP_RDLOCK: 461a8dd0a37SAlex Elder case CEPH_OSD_OP_RDUNLOCK: 462a8dd0a37SAlex Elder case CEPH_OSD_OP_UPLOCK: 463a8dd0a37SAlex Elder case CEPH_OSD_OP_DNLOCK: 464a8dd0a37SAlex Elder case CEPH_OSD_OP_CALL: 465a8dd0a37SAlex Elder case CEPH_OSD_OP_PGLS: 466a8dd0a37SAlex Elder case CEPH_OSD_OP_PGLS_FILTER: 467a8dd0a37SAlex Elder return true; 468a8dd0a37SAlex Elder default: 469a8dd0a37SAlex Elder return false; 470a8dd0a37SAlex Elder } 471a8dd0a37SAlex Elder } 472a8dd0a37SAlex Elder 47333803f33SAlex Elder /* 47433803f33SAlex Elder * This is an osd op init function for opcodes that have no data or 47533803f33SAlex Elder * other information associated with them. It also serves as a 47633803f33SAlex Elder * common init routine for all the other init functions, below. 47733803f33SAlex Elder */ 478c99d2d4aSAlex Elder static struct ceph_osd_req_op * 47949719778SAlex Elder _osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which, 480c99d2d4aSAlex Elder u16 opcode) 48133803f33SAlex Elder { 482c99d2d4aSAlex Elder struct ceph_osd_req_op *op; 483c99d2d4aSAlex Elder 484c99d2d4aSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 48533803f33SAlex Elder BUG_ON(!osd_req_opcode_valid(opcode)); 48633803f33SAlex Elder 487c99d2d4aSAlex Elder op = &osd_req->r_ops[which]; 48833803f33SAlex Elder memset(op, 0, sizeof (*op)); 48933803f33SAlex Elder op->op = opcode; 490c99d2d4aSAlex Elder 491c99d2d4aSAlex Elder return op; 49233803f33SAlex Elder } 49333803f33SAlex Elder 49449719778SAlex Elder void osd_req_op_init(struct ceph_osd_request *osd_req, 49549719778SAlex Elder unsigned int which, u16 opcode) 49649719778SAlex Elder { 49749719778SAlex Elder (void)_osd_req_op_init(osd_req, which, opcode); 49849719778SAlex Elder } 49949719778SAlex Elder EXPORT_SYMBOL(osd_req_op_init); 50049719778SAlex Elder 501c99d2d4aSAlex Elder void osd_req_op_extent_init(struct ceph_osd_request *osd_req, 502c99d2d4aSAlex Elder unsigned int which, u16 opcode, 50333803f33SAlex Elder u64 offset, u64 length, 50433803f33SAlex Elder u64 truncate_size, u32 truncate_seq) 50533803f33SAlex Elder { 50649719778SAlex Elder struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode); 50733803f33SAlex Elder size_t payload_len = 0; 50833803f33SAlex Elder 509ad7a60deSLi Wang BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE && 510ad7a60deSLi Wang opcode != CEPH_OSD_OP_DELETE && opcode != CEPH_OSD_OP_ZERO && 511ad7a60deSLi Wang opcode != CEPH_OSD_OP_TRUNCATE); 51233803f33SAlex Elder 51333803f33SAlex Elder op->extent.offset = offset; 51433803f33SAlex Elder op->extent.length = length; 51533803f33SAlex Elder op->extent.truncate_size = truncate_size; 51633803f33SAlex Elder op->extent.truncate_seq = truncate_seq; 51733803f33SAlex Elder if (opcode == CEPH_OSD_OP_WRITE) 51833803f33SAlex Elder payload_len += length; 51933803f33SAlex Elder 52033803f33SAlex Elder op->payload_len = payload_len; 52133803f33SAlex Elder } 52233803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_init); 52333803f33SAlex Elder 524c99d2d4aSAlex Elder void osd_req_op_extent_update(struct ceph_osd_request *osd_req, 525c99d2d4aSAlex Elder unsigned int which, u64 length) 526e5975c7cSAlex Elder { 527c99d2d4aSAlex Elder struct ceph_osd_req_op *op; 528c99d2d4aSAlex Elder u64 previous; 529c99d2d4aSAlex Elder 530c99d2d4aSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 531c99d2d4aSAlex Elder op = &osd_req->r_ops[which]; 532c99d2d4aSAlex Elder previous = op->extent.length; 533e5975c7cSAlex Elder 534e5975c7cSAlex Elder if (length == previous) 535e5975c7cSAlex Elder return; /* Nothing to do */ 536e5975c7cSAlex Elder BUG_ON(length > previous); 537e5975c7cSAlex Elder 538e5975c7cSAlex Elder op->extent.length = length; 539e5975c7cSAlex Elder op->payload_len -= previous - length; 540e5975c7cSAlex Elder } 541e5975c7cSAlex Elder EXPORT_SYMBOL(osd_req_op_extent_update); 542e5975c7cSAlex Elder 543c99d2d4aSAlex Elder void osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which, 54404017e29SAlex Elder u16 opcode, const char *class, const char *method) 54533803f33SAlex Elder { 54649719778SAlex Elder struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode); 5475f562df5SAlex Elder struct ceph_pagelist *pagelist; 54833803f33SAlex Elder size_t payload_len = 0; 54933803f33SAlex Elder size_t size; 55033803f33SAlex Elder 55133803f33SAlex Elder BUG_ON(opcode != CEPH_OSD_OP_CALL); 55233803f33SAlex Elder 5535f562df5SAlex Elder pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS); 5545f562df5SAlex Elder BUG_ON(!pagelist); 5555f562df5SAlex Elder ceph_pagelist_init(pagelist); 5565f562df5SAlex Elder 55733803f33SAlex Elder op->cls.class_name = class; 55833803f33SAlex Elder size = strlen(class); 55933803f33SAlex Elder BUG_ON(size > (size_t) U8_MAX); 56033803f33SAlex Elder op->cls.class_len = size; 5615f562df5SAlex Elder ceph_pagelist_append(pagelist, class, size); 56233803f33SAlex Elder payload_len += size; 56333803f33SAlex Elder 56433803f33SAlex Elder op->cls.method_name = method; 56533803f33SAlex Elder size = strlen(method); 56633803f33SAlex Elder BUG_ON(size > (size_t) U8_MAX); 56733803f33SAlex Elder op->cls.method_len = size; 5685f562df5SAlex Elder ceph_pagelist_append(pagelist, method, size); 56933803f33SAlex Elder payload_len += size; 57033803f33SAlex Elder 571a4ce40a9SAlex Elder osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist); 5725f562df5SAlex Elder 57333803f33SAlex Elder op->cls.argc = 0; /* currently unused */ 57433803f33SAlex Elder 57533803f33SAlex Elder op->payload_len = payload_len; 57633803f33SAlex Elder } 57733803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_init); 5788c042b0dSAlex Elder 579c99d2d4aSAlex Elder void osd_req_op_watch_init(struct ceph_osd_request *osd_req, 580c99d2d4aSAlex Elder unsigned int which, u16 opcode, 58133803f33SAlex Elder u64 cookie, u64 version, int flag) 58233803f33SAlex Elder { 58349719778SAlex Elder struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode); 58433803f33SAlex Elder 585c99d2d4aSAlex Elder BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH); 58633803f33SAlex Elder 58733803f33SAlex Elder op->watch.cookie = cookie; 5889ef1ee5aSAlex Elder op->watch.ver = version; 58933803f33SAlex Elder if (opcode == CEPH_OSD_OP_WATCH && flag) 59033803f33SAlex Elder op->watch.flag = (u8)1; 59133803f33SAlex Elder } 59233803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_watch_init); 59333803f33SAlex Elder 59490af3602SAlex Elder static void ceph_osdc_msg_data_add(struct ceph_msg *msg, 595ec9123c5SAlex Elder struct ceph_osd_data *osd_data) 596ec9123c5SAlex Elder { 597ec9123c5SAlex Elder u64 length = ceph_osd_data_length(osd_data); 598ec9123c5SAlex Elder 599ec9123c5SAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 600ec9123c5SAlex Elder BUG_ON(length > (u64) SIZE_MAX); 601ec9123c5SAlex Elder if (length) 60290af3602SAlex Elder ceph_msg_data_add_pages(msg, osd_data->pages, 603ec9123c5SAlex Elder length, osd_data->alignment); 604ec9123c5SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) { 605ec9123c5SAlex Elder BUG_ON(!length); 60690af3602SAlex Elder ceph_msg_data_add_pagelist(msg, osd_data->pagelist); 607ec9123c5SAlex Elder #ifdef CONFIG_BLOCK 608ec9123c5SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) { 60990af3602SAlex Elder ceph_msg_data_add_bio(msg, osd_data->bio, length); 610ec9123c5SAlex Elder #endif 611ec9123c5SAlex Elder } else { 612ec9123c5SAlex Elder BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE); 613ec9123c5SAlex Elder } 614ec9123c5SAlex Elder } 615ec9123c5SAlex Elder 616175face2SAlex Elder static u64 osd_req_encode_op(struct ceph_osd_request *req, 61779528734SAlex Elder struct ceph_osd_op *dst, unsigned int which) 6183d14c5d2SYehuda Sadeh { 61979528734SAlex Elder struct ceph_osd_req_op *src; 62004017e29SAlex Elder struct ceph_osd_data *osd_data; 62154d50649SAlex Elder u64 request_data_len = 0; 62204017e29SAlex Elder u64 data_length; 623175face2SAlex Elder 62479528734SAlex Elder BUG_ON(which >= req->r_num_ops); 62579528734SAlex Elder src = &req->r_ops[which]; 626a8dd0a37SAlex Elder if (WARN_ON(!osd_req_opcode_valid(src->op))) { 627a8dd0a37SAlex Elder pr_err("unrecognized osd opcode %d\n", src->op); 628a8dd0a37SAlex Elder 629a8dd0a37SAlex Elder return 0; 630a8dd0a37SAlex Elder } 6313d14c5d2SYehuda Sadeh 632065a68f9SAlex Elder switch (src->op) { 633fbfab539SAlex Elder case CEPH_OSD_OP_STAT: 63449719778SAlex Elder osd_data = &src->raw_data_in; 63549719778SAlex Elder ceph_osdc_msg_data_add(req->r_reply, osd_data); 636fbfab539SAlex Elder break; 6373d14c5d2SYehuda Sadeh case CEPH_OSD_OP_READ: 6383d14c5d2SYehuda Sadeh case CEPH_OSD_OP_WRITE: 639ad7a60deSLi Wang case CEPH_OSD_OP_ZERO: 640ad7a60deSLi Wang case CEPH_OSD_OP_DELETE: 641ad7a60deSLi Wang case CEPH_OSD_OP_TRUNCATE: 642175face2SAlex Elder if (src->op == CEPH_OSD_OP_WRITE) 64354d50649SAlex Elder request_data_len = src->extent.length; 644175face2SAlex Elder dst->extent.offset = cpu_to_le64(src->extent.offset); 645175face2SAlex Elder dst->extent.length = cpu_to_le64(src->extent.length); 6463d14c5d2SYehuda Sadeh dst->extent.truncate_size = 6473d14c5d2SYehuda Sadeh cpu_to_le64(src->extent.truncate_size); 6483d14c5d2SYehuda Sadeh dst->extent.truncate_seq = 6493d14c5d2SYehuda Sadeh cpu_to_le32(src->extent.truncate_seq); 65004017e29SAlex Elder osd_data = &src->extent.osd_data; 6515476492fSAlex Elder if (src->op == CEPH_OSD_OP_WRITE) 65204017e29SAlex Elder ceph_osdc_msg_data_add(req->r_request, osd_data); 6535476492fSAlex Elder else 65404017e29SAlex Elder ceph_osdc_msg_data_add(req->r_reply, osd_data); 6553d14c5d2SYehuda Sadeh break; 6563d14c5d2SYehuda Sadeh case CEPH_OSD_OP_CALL: 6573d14c5d2SYehuda Sadeh dst->cls.class_len = src->cls.class_len; 6583d14c5d2SYehuda Sadeh dst->cls.method_len = src->cls.method_len; 65904017e29SAlex Elder osd_data = &src->cls.request_info; 66004017e29SAlex Elder ceph_osdc_msg_data_add(req->r_request, osd_data); 66104017e29SAlex Elder BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGELIST); 66204017e29SAlex Elder request_data_len = osd_data->pagelist->length; 66304017e29SAlex Elder 66404017e29SAlex Elder osd_data = &src->cls.request_data; 66504017e29SAlex Elder data_length = ceph_osd_data_length(osd_data); 66604017e29SAlex Elder if (data_length) { 66704017e29SAlex Elder BUG_ON(osd_data->type == CEPH_OSD_DATA_TYPE_NONE); 66804017e29SAlex Elder dst->cls.indata_len = cpu_to_le32(data_length); 66904017e29SAlex Elder ceph_osdc_msg_data_add(req->r_request, osd_data); 67004017e29SAlex Elder src->payload_len += data_length; 67104017e29SAlex Elder request_data_len += data_length; 67204017e29SAlex Elder } 67304017e29SAlex Elder osd_data = &src->cls.response_data; 67404017e29SAlex Elder ceph_osdc_msg_data_add(req->r_reply, osd_data); 6753d14c5d2SYehuda Sadeh break; 6763d14c5d2SYehuda Sadeh case CEPH_OSD_OP_STARTSYNC: 6773d14c5d2SYehuda Sadeh break; 678a40c4f10SYehuda Sadeh case CEPH_OSD_OP_NOTIFY_ACK: 679a40c4f10SYehuda Sadeh case CEPH_OSD_OP_WATCH: 680a40c4f10SYehuda Sadeh dst->watch.cookie = cpu_to_le64(src->watch.cookie); 681a40c4f10SYehuda Sadeh dst->watch.ver = cpu_to_le64(src->watch.ver); 682a40c4f10SYehuda Sadeh dst->watch.flag = src->watch.flag; 683a40c4f10SYehuda Sadeh break; 6843d14c5d2SYehuda Sadeh default: 6854c46459cSAlex Elder pr_err("unsupported osd opcode %s\n", 6868f63ca2dSAlex Elder ceph_osd_op_name(src->op)); 6874c46459cSAlex Elder WARN_ON(1); 688a8dd0a37SAlex Elder 689a8dd0a37SAlex Elder return 0; 6903d14c5d2SYehuda Sadeh } 6917b25bf5fSIlya Dryomov 692a8dd0a37SAlex Elder dst->op = cpu_to_le16(src->op); 6937b25bf5fSIlya Dryomov dst->flags = cpu_to_le32(src->flags); 6943d14c5d2SYehuda Sadeh dst->payload_len = cpu_to_le32(src->payload_len); 695175face2SAlex Elder 69654d50649SAlex Elder return request_data_len; 6973d14c5d2SYehuda Sadeh } 6983d14c5d2SYehuda Sadeh 6993d14c5d2SYehuda Sadeh /* 7003d14c5d2SYehuda Sadeh * build new request AND message, calculate layout, and adjust file 7013d14c5d2SYehuda Sadeh * extent as needed. 7023d14c5d2SYehuda Sadeh * 7033d14c5d2SYehuda Sadeh * if the file was recently truncated, we include information about its 7043d14c5d2SYehuda Sadeh * old and new size so that the object can be updated appropriately. (we 7053d14c5d2SYehuda Sadeh * avoid synchronously deleting truncated objects because it's slow.) 7063d14c5d2SYehuda Sadeh * 7073d14c5d2SYehuda Sadeh * if @do_sync, include a 'startsync' command so that the osd will flush 7083d14c5d2SYehuda Sadeh * data quickly. 7093d14c5d2SYehuda Sadeh */ 7103d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc, 7113d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 7123d14c5d2SYehuda Sadeh struct ceph_vino vino, 713acead002SAlex Elder u64 off, u64 *plen, int num_ops, 7143d14c5d2SYehuda Sadeh int opcode, int flags, 7153d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 7163d14c5d2SYehuda Sadeh u32 truncate_seq, 7173d14c5d2SYehuda Sadeh u64 truncate_size, 718153e5167SAlex Elder bool use_mempool) 7193d14c5d2SYehuda Sadeh { 7203d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 72175d1c941SAlex Elder u64 objnum = 0; 72275d1c941SAlex Elder u64 objoff = 0; 72375d1c941SAlex Elder u64 objlen = 0; 724d18d1e28SAlex Elder u32 object_size; 725d18d1e28SAlex Elder u64 object_base; 7266816282dSSage Weil int r; 7273d14c5d2SYehuda Sadeh 728ad7a60deSLi Wang BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE && 729ad7a60deSLi Wang opcode != CEPH_OSD_OP_DELETE && opcode != CEPH_OSD_OP_ZERO && 730ad7a60deSLi Wang opcode != CEPH_OSD_OP_TRUNCATE); 7313d14c5d2SYehuda Sadeh 732acead002SAlex Elder req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool, 733ae7ca4a3SAlex Elder GFP_NOFS); 7344ad12621SSage Weil if (!req) 7356816282dSSage Weil return ERR_PTR(-ENOMEM); 73679528734SAlex Elder 737d178a9e7SAlex Elder req->r_flags = flags; 7383d14c5d2SYehuda Sadeh 7393d14c5d2SYehuda Sadeh /* calculate max write size */ 740a19dadfbSAlex Elder r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen); 7413ff5f385SAlex Elder if (r < 0) { 7423ff5f385SAlex Elder ceph_osdc_put_request(req); 7436816282dSSage Weil return ERR_PTR(r); 7443ff5f385SAlex Elder } 745a19dadfbSAlex Elder 746d18d1e28SAlex Elder object_size = le32_to_cpu(layout->fl_object_size); 747d18d1e28SAlex Elder object_base = off - objoff; 748ccca4e37SYan, Zheng if (!(truncate_seq == 1 && truncate_size == -1ULL)) { 749d18d1e28SAlex Elder if (truncate_size <= object_base) { 750d18d1e28SAlex Elder truncate_size = 0; 751d18d1e28SAlex Elder } else { 752d18d1e28SAlex Elder truncate_size -= object_base; 753d18d1e28SAlex Elder if (truncate_size > object_size) 754d18d1e28SAlex Elder truncate_size = object_size; 755d18d1e28SAlex Elder } 756ccca4e37SYan, Zheng } 757d18d1e28SAlex Elder 758c99d2d4aSAlex Elder osd_req_op_extent_init(req, 0, opcode, objoff, objlen, 759b0270324SAlex Elder truncate_size, truncate_seq); 7608c042b0dSAlex Elder 761acead002SAlex Elder /* 762acead002SAlex Elder * A second op in the ops array means the caller wants to 763acead002SAlex Elder * also issue a include a 'startsync' command so that the 764acead002SAlex Elder * osd will flush data quickly. 765acead002SAlex Elder */ 766acead002SAlex Elder if (num_ops > 1) 767c99d2d4aSAlex Elder osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC); 768d18d1e28SAlex Elder 7693c972c95SIlya Dryomov req->r_base_oloc.pool = ceph_file_layout_pg_pool(*layout); 7703d14c5d2SYehuda Sadeh 7713c972c95SIlya Dryomov snprintf(req->r_base_oid.name, sizeof(req->r_base_oid.name), 7724295f221SIlya Dryomov "%llx.%08llx", vino.ino, objnum); 7733c972c95SIlya Dryomov req->r_base_oid.name_len = strlen(req->r_base_oid.name); 774dbe0fc41SAlex Elder 7753d14c5d2SYehuda Sadeh return req; 7763d14c5d2SYehuda Sadeh } 7773d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_new_request); 7783d14c5d2SYehuda Sadeh 7793d14c5d2SYehuda Sadeh /* 7803d14c5d2SYehuda Sadeh * We keep osd requests in an rbtree, sorted by ->r_tid. 7813d14c5d2SYehuda Sadeh */ 7823d14c5d2SYehuda Sadeh static void __insert_request(struct ceph_osd_client *osdc, 7833d14c5d2SYehuda Sadeh struct ceph_osd_request *new) 7843d14c5d2SYehuda Sadeh { 7853d14c5d2SYehuda Sadeh struct rb_node **p = &osdc->requests.rb_node; 7863d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 7873d14c5d2SYehuda Sadeh struct ceph_osd_request *req = NULL; 7883d14c5d2SYehuda Sadeh 7893d14c5d2SYehuda Sadeh while (*p) { 7903d14c5d2SYehuda Sadeh parent = *p; 7913d14c5d2SYehuda Sadeh req = rb_entry(parent, struct ceph_osd_request, r_node); 7923d14c5d2SYehuda Sadeh if (new->r_tid < req->r_tid) 7933d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 7943d14c5d2SYehuda Sadeh else if (new->r_tid > req->r_tid) 7953d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 7963d14c5d2SYehuda Sadeh else 7973d14c5d2SYehuda Sadeh BUG(); 7983d14c5d2SYehuda Sadeh } 7993d14c5d2SYehuda Sadeh 8003d14c5d2SYehuda Sadeh rb_link_node(&new->r_node, parent, p); 8013d14c5d2SYehuda Sadeh rb_insert_color(&new->r_node, &osdc->requests); 8023d14c5d2SYehuda Sadeh } 8033d14c5d2SYehuda Sadeh 8043d14c5d2SYehuda Sadeh static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc, 8053d14c5d2SYehuda Sadeh u64 tid) 8063d14c5d2SYehuda Sadeh { 8073d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 8083d14c5d2SYehuda Sadeh struct rb_node *n = osdc->requests.rb_node; 8093d14c5d2SYehuda Sadeh 8103d14c5d2SYehuda Sadeh while (n) { 8113d14c5d2SYehuda Sadeh req = rb_entry(n, struct ceph_osd_request, r_node); 8123d14c5d2SYehuda Sadeh if (tid < req->r_tid) 8133d14c5d2SYehuda Sadeh n = n->rb_left; 8143d14c5d2SYehuda Sadeh else if (tid > req->r_tid) 8153d14c5d2SYehuda Sadeh n = n->rb_right; 8163d14c5d2SYehuda Sadeh else 8173d14c5d2SYehuda Sadeh return req; 8183d14c5d2SYehuda Sadeh } 8193d14c5d2SYehuda Sadeh return NULL; 8203d14c5d2SYehuda Sadeh } 8213d14c5d2SYehuda Sadeh 8223d14c5d2SYehuda Sadeh static struct ceph_osd_request * 8233d14c5d2SYehuda Sadeh __lookup_request_ge(struct ceph_osd_client *osdc, 8243d14c5d2SYehuda Sadeh u64 tid) 8253d14c5d2SYehuda Sadeh { 8263d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 8273d14c5d2SYehuda Sadeh struct rb_node *n = osdc->requests.rb_node; 8283d14c5d2SYehuda Sadeh 8293d14c5d2SYehuda Sadeh while (n) { 8303d14c5d2SYehuda Sadeh req = rb_entry(n, struct ceph_osd_request, r_node); 8313d14c5d2SYehuda Sadeh if (tid < req->r_tid) { 8323d14c5d2SYehuda Sadeh if (!n->rb_left) 8333d14c5d2SYehuda Sadeh return req; 8343d14c5d2SYehuda Sadeh n = n->rb_left; 8353d14c5d2SYehuda Sadeh } else if (tid > req->r_tid) { 8363d14c5d2SYehuda Sadeh n = n->rb_right; 8373d14c5d2SYehuda Sadeh } else { 8383d14c5d2SYehuda Sadeh return req; 8393d14c5d2SYehuda Sadeh } 8403d14c5d2SYehuda Sadeh } 8413d14c5d2SYehuda Sadeh return NULL; 8423d14c5d2SYehuda Sadeh } 8433d14c5d2SYehuda Sadeh 8446f6c7006SSage Weil /* 8456f6c7006SSage Weil * Resubmit requests pending on the given osd. 8466f6c7006SSage Weil */ 8476f6c7006SSage Weil static void __kick_osd_requests(struct ceph_osd_client *osdc, 8486f6c7006SSage Weil struct ceph_osd *osd) 8496f6c7006SSage Weil { 850a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 851e02493c0SAlex Elder LIST_HEAD(resend); 8526f6c7006SSage Weil int err; 8536f6c7006SSage Weil 8546f6c7006SSage Weil dout("__kick_osd_requests osd%d\n", osd->o_osd); 8556f6c7006SSage Weil err = __reset_osd(osdc, osd); 856685a7555SAlex Elder if (err) 8576f6c7006SSage Weil return; 858e02493c0SAlex Elder /* 859e02493c0SAlex Elder * Build up a list of requests to resend by traversing the 860e02493c0SAlex Elder * osd's list of requests. Requests for a given object are 861e02493c0SAlex Elder * sent in tid order, and that is also the order they're 862e02493c0SAlex Elder * kept on this list. Therefore all requests that are in 863e02493c0SAlex Elder * flight will be found first, followed by all requests that 864e02493c0SAlex Elder * have not yet been sent. And to resend requests while 865e02493c0SAlex Elder * preserving this order we will want to put any sent 866e02493c0SAlex Elder * requests back on the front of the osd client's unsent 867e02493c0SAlex Elder * list. 868e02493c0SAlex Elder * 869e02493c0SAlex Elder * So we build a separate ordered list of already-sent 870e02493c0SAlex Elder * requests for the affected osd and splice it onto the 871e02493c0SAlex Elder * front of the osd client's unsent list. Once we've seen a 872e02493c0SAlex Elder * request that has not yet been sent we're done. Those 873e02493c0SAlex Elder * requests are already sitting right where they belong. 874e02493c0SAlex Elder */ 8756f6c7006SSage Weil list_for_each_entry(req, &osd->o_requests, r_osd_item) { 876e02493c0SAlex Elder if (!req->r_sent) 877e02493c0SAlex Elder break; 878e02493c0SAlex Elder list_move_tail(&req->r_req_lru_item, &resend); 879e02493c0SAlex Elder dout("requeueing %p tid %llu osd%d\n", req, req->r_tid, 8806f6c7006SSage Weil osd->o_osd); 881a40c4f10SYehuda Sadeh if (!req->r_linger) 8826f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 8836f6c7006SSage Weil } 884e02493c0SAlex Elder list_splice(&resend, &osdc->req_unsent); 885a40c4f10SYehuda Sadeh 886e02493c0SAlex Elder /* 887e02493c0SAlex Elder * Linger requests are re-registered before sending, which 888e02493c0SAlex Elder * sets up a new tid for each. We add them to the unsent 889e02493c0SAlex Elder * list at the end to keep things in tid order. 890e02493c0SAlex Elder */ 891a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osd->o_linger_requests, 892a40c4f10SYehuda Sadeh r_linger_osd) { 89377f38e0eSSage Weil /* 89477f38e0eSSage Weil * reregister request prior to unregistering linger so 89577f38e0eSSage Weil * that r_osd is preserved. 89677f38e0eSSage Weil */ 89777f38e0eSSage Weil BUG_ON(!list_empty(&req->r_req_lru_item)); 898a40c4f10SYehuda Sadeh __register_request(osdc, req); 899e02493c0SAlex Elder list_add_tail(&req->r_req_lru_item, &osdc->req_unsent); 900ad885927SAlex Elder list_add_tail(&req->r_osd_item, &req->r_osd->o_requests); 90177f38e0eSSage Weil __unregister_linger_request(osdc, req); 902a40c4f10SYehuda Sadeh dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid, 903a40c4f10SYehuda Sadeh osd->o_osd); 904a40c4f10SYehuda Sadeh } 9056f6c7006SSage Weil } 9066f6c7006SSage Weil 9073d14c5d2SYehuda Sadeh /* 9083d14c5d2SYehuda Sadeh * If the osd connection drops, we need to resubmit all requests. 9093d14c5d2SYehuda Sadeh */ 9103d14c5d2SYehuda Sadeh static void osd_reset(struct ceph_connection *con) 9113d14c5d2SYehuda Sadeh { 9123d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 9133d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 9143d14c5d2SYehuda Sadeh 9153d14c5d2SYehuda Sadeh if (!osd) 9163d14c5d2SYehuda Sadeh return; 9173d14c5d2SYehuda Sadeh dout("osd_reset osd%d\n", osd->o_osd); 9183d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 9193d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 92083aff95eSSage Weil mutex_lock(&osdc->request_mutex); 92183aff95eSSage Weil __kick_osd_requests(osdc, osd); 922f9d25199SAlex Elder __send_queued(osdc); 92383aff95eSSage Weil mutex_unlock(&osdc->request_mutex); 9243d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 9253d14c5d2SYehuda Sadeh } 9263d14c5d2SYehuda Sadeh 9273d14c5d2SYehuda Sadeh /* 9283d14c5d2SYehuda Sadeh * Track open sessions with osds. 9293d14c5d2SYehuda Sadeh */ 930e10006f8SAlex Elder static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum) 9313d14c5d2SYehuda Sadeh { 9323d14c5d2SYehuda Sadeh struct ceph_osd *osd; 9333d14c5d2SYehuda Sadeh 9343d14c5d2SYehuda Sadeh osd = kzalloc(sizeof(*osd), GFP_NOFS); 9353d14c5d2SYehuda Sadeh if (!osd) 9363d14c5d2SYehuda Sadeh return NULL; 9373d14c5d2SYehuda Sadeh 9383d14c5d2SYehuda Sadeh atomic_set(&osd->o_ref, 1); 9393d14c5d2SYehuda Sadeh osd->o_osdc = osdc; 940e10006f8SAlex Elder osd->o_osd = onum; 941f407731dSAlex Elder RB_CLEAR_NODE(&osd->o_node); 9423d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_requests); 943a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osd->o_linger_requests); 9443d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_osd_lru); 9453d14c5d2SYehuda Sadeh osd->o_incarnation = 1; 9463d14c5d2SYehuda Sadeh 947b7a9e5ddSSage Weil ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr); 9483d14c5d2SYehuda Sadeh 9493d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_keepalive_item); 9503d14c5d2SYehuda Sadeh return osd; 9513d14c5d2SYehuda Sadeh } 9523d14c5d2SYehuda Sadeh 9533d14c5d2SYehuda Sadeh static struct ceph_osd *get_osd(struct ceph_osd *osd) 9543d14c5d2SYehuda Sadeh { 9553d14c5d2SYehuda Sadeh if (atomic_inc_not_zero(&osd->o_ref)) { 9563d14c5d2SYehuda Sadeh dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1, 9573d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref)); 9583d14c5d2SYehuda Sadeh return osd; 9593d14c5d2SYehuda Sadeh } else { 9603d14c5d2SYehuda Sadeh dout("get_osd %p FAIL\n", osd); 9613d14c5d2SYehuda Sadeh return NULL; 9623d14c5d2SYehuda Sadeh } 9633d14c5d2SYehuda Sadeh } 9643d14c5d2SYehuda Sadeh 9653d14c5d2SYehuda Sadeh static void put_osd(struct ceph_osd *osd) 9663d14c5d2SYehuda Sadeh { 9673d14c5d2SYehuda Sadeh dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref), 9683d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref) - 1); 969a255651dSAlex Elder if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) { 9703d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth; 9713d14c5d2SYehuda Sadeh 97227859f97SSage Weil ceph_auth_destroy_authorizer(ac, osd->o_auth.authorizer); 9733d14c5d2SYehuda Sadeh kfree(osd); 9743d14c5d2SYehuda Sadeh } 9753d14c5d2SYehuda Sadeh } 9763d14c5d2SYehuda Sadeh 9773d14c5d2SYehuda Sadeh /* 9783d14c5d2SYehuda Sadeh * remove an osd from our map 9793d14c5d2SYehuda Sadeh */ 9803d14c5d2SYehuda Sadeh static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 9813d14c5d2SYehuda Sadeh { 9823d14c5d2SYehuda Sadeh dout("__remove_osd %p\n", osd); 9833d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_requests)); 9843d14c5d2SYehuda Sadeh rb_erase(&osd->o_node, &osdc->osds); 9853d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 9863d14c5d2SYehuda Sadeh ceph_con_close(&osd->o_con); 9873d14c5d2SYehuda Sadeh put_osd(osd); 9883d14c5d2SYehuda Sadeh } 9893d14c5d2SYehuda Sadeh 990aca420bcSSage Weil static void remove_all_osds(struct ceph_osd_client *osdc) 991aca420bcSSage Weil { 992048a9d2dSJiaju Zhang dout("%s %p\n", __func__, osdc); 993aca420bcSSage Weil mutex_lock(&osdc->request_mutex); 994aca420bcSSage Weil while (!RB_EMPTY_ROOT(&osdc->osds)) { 995aca420bcSSage Weil struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds), 996aca420bcSSage Weil struct ceph_osd, o_node); 997aca420bcSSage Weil __remove_osd(osdc, osd); 998aca420bcSSage Weil } 999aca420bcSSage Weil mutex_unlock(&osdc->request_mutex); 1000aca420bcSSage Weil } 1001aca420bcSSage Weil 10023d14c5d2SYehuda Sadeh static void __move_osd_to_lru(struct ceph_osd_client *osdc, 10033d14c5d2SYehuda Sadeh struct ceph_osd *osd) 10043d14c5d2SYehuda Sadeh { 10053d14c5d2SYehuda Sadeh dout("__move_osd_to_lru %p\n", osd); 10063d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_osd_lru)); 10073d14c5d2SYehuda Sadeh list_add_tail(&osd->o_osd_lru, &osdc->osd_lru); 10083d14c5d2SYehuda Sadeh osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ; 10093d14c5d2SYehuda Sadeh } 10103d14c5d2SYehuda Sadeh 10113d14c5d2SYehuda Sadeh static void __remove_osd_from_lru(struct ceph_osd *osd) 10123d14c5d2SYehuda Sadeh { 10133d14c5d2SYehuda Sadeh dout("__remove_osd_from_lru %p\n", osd); 10143d14c5d2SYehuda Sadeh if (!list_empty(&osd->o_osd_lru)) 10153d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 10163d14c5d2SYehuda Sadeh } 10173d14c5d2SYehuda Sadeh 1018aca420bcSSage Weil static void remove_old_osds(struct ceph_osd_client *osdc) 10193d14c5d2SYehuda Sadeh { 10203d14c5d2SYehuda Sadeh struct ceph_osd *osd, *nosd; 10213d14c5d2SYehuda Sadeh 10223d14c5d2SYehuda Sadeh dout("__remove_old_osds %p\n", osdc); 10233d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 10243d14c5d2SYehuda Sadeh list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) { 1025aca420bcSSage Weil if (time_before(jiffies, osd->lru_ttl)) 10263d14c5d2SYehuda Sadeh break; 10273d14c5d2SYehuda Sadeh __remove_osd(osdc, osd); 10283d14c5d2SYehuda Sadeh } 10293d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 10303d14c5d2SYehuda Sadeh } 10313d14c5d2SYehuda Sadeh 10323d14c5d2SYehuda Sadeh /* 10333d14c5d2SYehuda Sadeh * reset osd connect 10343d14c5d2SYehuda Sadeh */ 10353d14c5d2SYehuda Sadeh static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 10363d14c5d2SYehuda Sadeh { 1037c3acb181SAlex Elder struct ceph_entity_addr *peer_addr; 10383d14c5d2SYehuda Sadeh 10393d14c5d2SYehuda Sadeh dout("__reset_osd %p osd%d\n", osd, osd->o_osd); 1040a40c4f10SYehuda Sadeh if (list_empty(&osd->o_requests) && 1041a40c4f10SYehuda Sadeh list_empty(&osd->o_linger_requests)) { 10423d14c5d2SYehuda Sadeh __remove_osd(osdc, osd); 1043c3acb181SAlex Elder 1044c3acb181SAlex Elder return -ENODEV; 1045c3acb181SAlex Elder } 1046c3acb181SAlex Elder 1047c3acb181SAlex Elder peer_addr = &osdc->osdmap->osd_addr[osd->o_osd]; 1048c3acb181SAlex Elder if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) && 10493d14c5d2SYehuda Sadeh !ceph_con_opened(&osd->o_con)) { 1050c3acb181SAlex Elder struct ceph_osd_request *req; 1051c3acb181SAlex Elder 10523d14c5d2SYehuda Sadeh dout("osd addr hasn't changed and connection never opened, " 10530b4af2e8SIlya Dryomov "letting msgr retry\n"); 10543d14c5d2SYehuda Sadeh /* touch each r_stamp for handle_timeout()'s benfit */ 10553d14c5d2SYehuda Sadeh list_for_each_entry(req, &osd->o_requests, r_osd_item) 10563d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 1057c3acb181SAlex Elder 1058c3acb181SAlex Elder return -EAGAIN; 10593d14c5d2SYehuda Sadeh } 1060c3acb181SAlex Elder 1061c3acb181SAlex Elder ceph_con_close(&osd->o_con); 1062c3acb181SAlex Elder ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr); 1063c3acb181SAlex Elder osd->o_incarnation++; 1064c3acb181SAlex Elder 1065c3acb181SAlex Elder return 0; 10663d14c5d2SYehuda Sadeh } 10673d14c5d2SYehuda Sadeh 10683d14c5d2SYehuda Sadeh static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new) 10693d14c5d2SYehuda Sadeh { 10703d14c5d2SYehuda Sadeh struct rb_node **p = &osdc->osds.rb_node; 10713d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 10723d14c5d2SYehuda Sadeh struct ceph_osd *osd = NULL; 10733d14c5d2SYehuda Sadeh 1074aca420bcSSage Weil dout("__insert_osd %p osd%d\n", new, new->o_osd); 10753d14c5d2SYehuda Sadeh while (*p) { 10763d14c5d2SYehuda Sadeh parent = *p; 10773d14c5d2SYehuda Sadeh osd = rb_entry(parent, struct ceph_osd, o_node); 10783d14c5d2SYehuda Sadeh if (new->o_osd < osd->o_osd) 10793d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 10803d14c5d2SYehuda Sadeh else if (new->o_osd > osd->o_osd) 10813d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 10823d14c5d2SYehuda Sadeh else 10833d14c5d2SYehuda Sadeh BUG(); 10843d14c5d2SYehuda Sadeh } 10853d14c5d2SYehuda Sadeh 10863d14c5d2SYehuda Sadeh rb_link_node(&new->o_node, parent, p); 10873d14c5d2SYehuda Sadeh rb_insert_color(&new->o_node, &osdc->osds); 10883d14c5d2SYehuda Sadeh } 10893d14c5d2SYehuda Sadeh 10903d14c5d2SYehuda Sadeh static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o) 10913d14c5d2SYehuda Sadeh { 10923d14c5d2SYehuda Sadeh struct ceph_osd *osd; 10933d14c5d2SYehuda Sadeh struct rb_node *n = osdc->osds.rb_node; 10943d14c5d2SYehuda Sadeh 10953d14c5d2SYehuda Sadeh while (n) { 10963d14c5d2SYehuda Sadeh osd = rb_entry(n, struct ceph_osd, o_node); 10973d14c5d2SYehuda Sadeh if (o < osd->o_osd) 10983d14c5d2SYehuda Sadeh n = n->rb_left; 10993d14c5d2SYehuda Sadeh else if (o > osd->o_osd) 11003d14c5d2SYehuda Sadeh n = n->rb_right; 11013d14c5d2SYehuda Sadeh else 11023d14c5d2SYehuda Sadeh return osd; 11033d14c5d2SYehuda Sadeh } 11043d14c5d2SYehuda Sadeh return NULL; 11053d14c5d2SYehuda Sadeh } 11063d14c5d2SYehuda Sadeh 11073d14c5d2SYehuda Sadeh static void __schedule_osd_timeout(struct ceph_osd_client *osdc) 11083d14c5d2SYehuda Sadeh { 11093d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->timeout_work, 11103d14c5d2SYehuda Sadeh osdc->client->options->osd_keepalive_timeout * HZ); 11113d14c5d2SYehuda Sadeh } 11123d14c5d2SYehuda Sadeh 11133d14c5d2SYehuda Sadeh static void __cancel_osd_timeout(struct ceph_osd_client *osdc) 11143d14c5d2SYehuda Sadeh { 11153d14c5d2SYehuda Sadeh cancel_delayed_work(&osdc->timeout_work); 11163d14c5d2SYehuda Sadeh } 11173d14c5d2SYehuda Sadeh 11183d14c5d2SYehuda Sadeh /* 11193d14c5d2SYehuda Sadeh * Register request, assign tid. If this is the first request, set up 11203d14c5d2SYehuda Sadeh * the timeout event. 11213d14c5d2SYehuda Sadeh */ 1122a40c4f10SYehuda Sadeh static void __register_request(struct ceph_osd_client *osdc, 11233d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 11243d14c5d2SYehuda Sadeh { 11253d14c5d2SYehuda Sadeh req->r_tid = ++osdc->last_tid; 11263d14c5d2SYehuda Sadeh req->r_request->hdr.tid = cpu_to_le64(req->r_tid); 112777f38e0eSSage Weil dout("__register_request %p tid %lld\n", req, req->r_tid); 11283d14c5d2SYehuda Sadeh __insert_request(osdc, req); 11293d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 11303d14c5d2SYehuda Sadeh osdc->num_requests++; 11313d14c5d2SYehuda Sadeh if (osdc->num_requests == 1) { 11323d14c5d2SYehuda Sadeh dout(" first request, scheduling timeout\n"); 11333d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 11343d14c5d2SYehuda Sadeh } 1135a40c4f10SYehuda Sadeh } 1136a40c4f10SYehuda Sadeh 11373d14c5d2SYehuda Sadeh /* 11383d14c5d2SYehuda Sadeh * called under osdc->request_mutex 11393d14c5d2SYehuda Sadeh */ 11403d14c5d2SYehuda Sadeh static void __unregister_request(struct ceph_osd_client *osdc, 11413d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 11423d14c5d2SYehuda Sadeh { 114335f9f8a0SSage Weil if (RB_EMPTY_NODE(&req->r_node)) { 114435f9f8a0SSage Weil dout("__unregister_request %p tid %lld not registered\n", 114535f9f8a0SSage Weil req, req->r_tid); 114635f9f8a0SSage Weil return; 114735f9f8a0SSage Weil } 114835f9f8a0SSage Weil 11493d14c5d2SYehuda Sadeh dout("__unregister_request %p tid %lld\n", req, req->r_tid); 11503d14c5d2SYehuda Sadeh rb_erase(&req->r_node, &osdc->requests); 11513d14c5d2SYehuda Sadeh osdc->num_requests--; 11523d14c5d2SYehuda Sadeh 11533d14c5d2SYehuda Sadeh if (req->r_osd) { 11543d14c5d2SYehuda Sadeh /* make sure the original request isn't in flight. */ 11556740a845SAlex Elder ceph_msg_revoke(req->r_request); 11563d14c5d2SYehuda Sadeh 11573d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 1158a40c4f10SYehuda Sadeh if (list_empty(&req->r_osd->o_requests) && 1159a40c4f10SYehuda Sadeh list_empty(&req->r_osd->o_linger_requests)) { 1160a40c4f10SYehuda Sadeh dout("moving osd to %p lru\n", req->r_osd); 11613d14c5d2SYehuda Sadeh __move_osd_to_lru(osdc, req->r_osd); 1162a40c4f10SYehuda Sadeh } 1163fbdb9190SSage Weil if (list_empty(&req->r_linger_item)) 11643d14c5d2SYehuda Sadeh req->r_osd = NULL; 11653d14c5d2SYehuda Sadeh } 11663d14c5d2SYehuda Sadeh 11677d5f2481SAlex Elder list_del_init(&req->r_req_lru_item); 11683d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 11693d14c5d2SYehuda Sadeh 11703d14c5d2SYehuda Sadeh if (osdc->num_requests == 0) { 11713d14c5d2SYehuda Sadeh dout(" no requests, canceling timeout\n"); 11723d14c5d2SYehuda Sadeh __cancel_osd_timeout(osdc); 11733d14c5d2SYehuda Sadeh } 11743d14c5d2SYehuda Sadeh } 11753d14c5d2SYehuda Sadeh 11763d14c5d2SYehuda Sadeh /* 11773d14c5d2SYehuda Sadeh * Cancel a previously queued request message 11783d14c5d2SYehuda Sadeh */ 11793d14c5d2SYehuda Sadeh static void __cancel_request(struct ceph_osd_request *req) 11803d14c5d2SYehuda Sadeh { 11813d14c5d2SYehuda Sadeh if (req->r_sent && req->r_osd) { 11826740a845SAlex Elder ceph_msg_revoke(req->r_request); 11833d14c5d2SYehuda Sadeh req->r_sent = 0; 11843d14c5d2SYehuda Sadeh } 11853d14c5d2SYehuda Sadeh } 11863d14c5d2SYehuda Sadeh 1187a40c4f10SYehuda Sadeh static void __register_linger_request(struct ceph_osd_client *osdc, 1188a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1189a40c4f10SYehuda Sadeh { 1190a40c4f10SYehuda Sadeh dout("__register_linger_request %p\n", req); 119196e4dac6SAlex Elder ceph_osdc_get_request(req); 1192a40c4f10SYehuda Sadeh list_add_tail(&req->r_linger_item, &osdc->req_linger); 11936194ea89SSage Weil if (req->r_osd) 11946194ea89SSage Weil list_add_tail(&req->r_linger_osd, 11956194ea89SSage Weil &req->r_osd->o_linger_requests); 1196a40c4f10SYehuda Sadeh } 1197a40c4f10SYehuda Sadeh 1198a40c4f10SYehuda Sadeh static void __unregister_linger_request(struct ceph_osd_client *osdc, 1199a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1200a40c4f10SYehuda Sadeh { 1201a40c4f10SYehuda Sadeh dout("__unregister_linger_request %p\n", req); 1202a40c4f10SYehuda Sadeh list_del_init(&req->r_linger_item); 120361c74035SAlex Elder if (req->r_osd) { 1204a40c4f10SYehuda Sadeh list_del_init(&req->r_linger_osd); 1205a40c4f10SYehuda Sadeh 1206a40c4f10SYehuda Sadeh if (list_empty(&req->r_osd->o_requests) && 1207a40c4f10SYehuda Sadeh list_empty(&req->r_osd->o_linger_requests)) { 1208a40c4f10SYehuda Sadeh dout("moving osd to %p lru\n", req->r_osd); 1209a40c4f10SYehuda Sadeh __move_osd_to_lru(osdc, req->r_osd); 1210a40c4f10SYehuda Sadeh } 1211fbdb9190SSage Weil if (list_empty(&req->r_osd_item)) 1212a40c4f10SYehuda Sadeh req->r_osd = NULL; 1213a40c4f10SYehuda Sadeh } 121496e4dac6SAlex Elder ceph_osdc_put_request(req); 1215a40c4f10SYehuda Sadeh } 1216a40c4f10SYehuda Sadeh 1217a40c4f10SYehuda Sadeh void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc, 1218a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1219a40c4f10SYehuda Sadeh { 1220a40c4f10SYehuda Sadeh mutex_lock(&osdc->request_mutex); 1221a40c4f10SYehuda Sadeh if (req->r_linger) { 1222c10ebbf5SAlex Elder req->r_linger = 0; 122396e4dac6SAlex Elder __unregister_linger_request(osdc, req); 1224a40c4f10SYehuda Sadeh } 1225a40c4f10SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 1226a40c4f10SYehuda Sadeh } 1227a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_unregister_linger_request); 1228a40c4f10SYehuda Sadeh 1229a40c4f10SYehuda Sadeh void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc, 1230a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1231a40c4f10SYehuda Sadeh { 1232a40c4f10SYehuda Sadeh if (!req->r_linger) { 1233a40c4f10SYehuda Sadeh dout("set_request_linger %p\n", req); 1234a40c4f10SYehuda Sadeh req->r_linger = 1; 1235a40c4f10SYehuda Sadeh } 1236a40c4f10SYehuda Sadeh } 1237a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_set_request_linger); 1238a40c4f10SYehuda Sadeh 12393d14c5d2SYehuda Sadeh /* 1240d29adb34SJosh Durgin * Returns whether a request should be blocked from being sent 1241d29adb34SJosh Durgin * based on the current osdmap and osd_client settings. 1242d29adb34SJosh Durgin * 1243d29adb34SJosh Durgin * Caller should hold map_sem for read. 1244d29adb34SJosh Durgin */ 1245d29adb34SJosh Durgin static bool __req_should_be_paused(struct ceph_osd_client *osdc, 1246d29adb34SJosh Durgin struct ceph_osd_request *req) 1247d29adb34SJosh Durgin { 1248d29adb34SJosh Durgin bool pauserd = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD); 1249d29adb34SJosh Durgin bool pausewr = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR) || 1250d29adb34SJosh Durgin ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL); 1251d29adb34SJosh Durgin return (req->r_flags & CEPH_OSD_FLAG_READ && pauserd) || 1252d29adb34SJosh Durgin (req->r_flags & CEPH_OSD_FLAG_WRITE && pausewr); 1253d29adb34SJosh Durgin } 1254d29adb34SJosh Durgin 1255d29adb34SJosh Durgin /* 125617a13e40SIlya Dryomov * Calculate mapping of a request to a PG. Takes tiering into account. 125717a13e40SIlya Dryomov */ 125817a13e40SIlya Dryomov static int __calc_request_pg(struct ceph_osdmap *osdmap, 125917a13e40SIlya Dryomov struct ceph_osd_request *req, 126017a13e40SIlya Dryomov struct ceph_pg *pg_out) 126117a13e40SIlya Dryomov { 1262205ee118SIlya Dryomov bool need_check_tiering; 1263205ee118SIlya Dryomov 1264205ee118SIlya Dryomov need_check_tiering = false; 1265205ee118SIlya Dryomov if (req->r_target_oloc.pool == -1) { 1266205ee118SIlya Dryomov req->r_target_oloc = req->r_base_oloc; /* struct */ 1267205ee118SIlya Dryomov need_check_tiering = true; 1268205ee118SIlya Dryomov } 1269205ee118SIlya Dryomov if (req->r_target_oid.name_len == 0) { 1270205ee118SIlya Dryomov ceph_oid_copy(&req->r_target_oid, &req->r_base_oid); 1271205ee118SIlya Dryomov need_check_tiering = true; 1272205ee118SIlya Dryomov } 1273205ee118SIlya Dryomov 1274205ee118SIlya Dryomov if (need_check_tiering && 1275205ee118SIlya Dryomov (req->r_flags & CEPH_OSD_FLAG_IGNORE_OVERLAY) == 0) { 127617a13e40SIlya Dryomov struct ceph_pg_pool_info *pi; 127717a13e40SIlya Dryomov 1278205ee118SIlya Dryomov pi = ceph_pg_pool_by_id(osdmap, req->r_target_oloc.pool); 127917a13e40SIlya Dryomov if (pi) { 128017a13e40SIlya Dryomov if ((req->r_flags & CEPH_OSD_FLAG_READ) && 128117a13e40SIlya Dryomov pi->read_tier >= 0) 1282205ee118SIlya Dryomov req->r_target_oloc.pool = pi->read_tier; 128317a13e40SIlya Dryomov if ((req->r_flags & CEPH_OSD_FLAG_WRITE) && 128417a13e40SIlya Dryomov pi->write_tier >= 0) 1285205ee118SIlya Dryomov req->r_target_oloc.pool = pi->write_tier; 128617a13e40SIlya Dryomov } 128717a13e40SIlya Dryomov /* !pi is caught in ceph_oloc_oid_to_pg() */ 128817a13e40SIlya Dryomov } 128917a13e40SIlya Dryomov 1290205ee118SIlya Dryomov return ceph_oloc_oid_to_pg(osdmap, &req->r_target_oloc, 1291205ee118SIlya Dryomov &req->r_target_oid, pg_out); 129217a13e40SIlya Dryomov } 129317a13e40SIlya Dryomov 129417a13e40SIlya Dryomov /* 12953d14c5d2SYehuda Sadeh * Pick an osd (the first 'up' osd in the pg), allocate the osd struct 12963d14c5d2SYehuda Sadeh * (as needed), and set the request r_osd appropriately. If there is 129725985edcSLucas De Marchi * no up osd, set r_osd to NULL. Move the request to the appropriate list 12986f6c7006SSage Weil * (unsent, homeless) or leave on in-flight lru. 12993d14c5d2SYehuda Sadeh * 13003d14c5d2SYehuda Sadeh * Return 0 if unchanged, 1 if changed, or negative on error. 13013d14c5d2SYehuda Sadeh * 13023d14c5d2SYehuda Sadeh * Caller should hold map_sem for read and request_mutex. 13033d14c5d2SYehuda Sadeh */ 13046f6c7006SSage Weil static int __map_request(struct ceph_osd_client *osdc, 130538d6453cSSage Weil struct ceph_osd_request *req, int force_resend) 13063d14c5d2SYehuda Sadeh { 13075b191d99SSage Weil struct ceph_pg pgid; 13083d14c5d2SYehuda Sadeh int acting[CEPH_PG_MAX_SIZE]; 13093d14c5d2SYehuda Sadeh int o = -1, num = 0; 13103d14c5d2SYehuda Sadeh int err; 1311d29adb34SJosh Durgin bool was_paused; 13123d14c5d2SYehuda Sadeh 13136f6c7006SSage Weil dout("map_request %p tid %lld\n", req, req->r_tid); 131417a13e40SIlya Dryomov 131517a13e40SIlya Dryomov err = __calc_request_pg(osdc->osdmap, req, &pgid); 13166f6c7006SSage Weil if (err) { 13176f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 13183d14c5d2SYehuda Sadeh return err; 13196f6c7006SSage Weil } 13203d14c5d2SYehuda Sadeh req->r_pgid = pgid; 13213d14c5d2SYehuda Sadeh 13223d14c5d2SYehuda Sadeh err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting); 13233d14c5d2SYehuda Sadeh if (err > 0) { 13243d14c5d2SYehuda Sadeh o = acting[0]; 13253d14c5d2SYehuda Sadeh num = err; 13263d14c5d2SYehuda Sadeh } 13273d14c5d2SYehuda Sadeh 1328d29adb34SJosh Durgin was_paused = req->r_paused; 1329d29adb34SJosh Durgin req->r_paused = __req_should_be_paused(osdc, req); 1330d29adb34SJosh Durgin if (was_paused && !req->r_paused) 1331d29adb34SJosh Durgin force_resend = 1; 1332d29adb34SJosh Durgin 133338d6453cSSage Weil if ((!force_resend && 133438d6453cSSage Weil req->r_osd && req->r_osd->o_osd == o && 13353d14c5d2SYehuda Sadeh req->r_sent >= req->r_osd->o_incarnation && 13363d14c5d2SYehuda Sadeh req->r_num_pg_osds == num && 13373d14c5d2SYehuda Sadeh memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) || 1338d29adb34SJosh Durgin (req->r_osd == NULL && o == -1) || 1339d29adb34SJosh Durgin req->r_paused) 13403d14c5d2SYehuda Sadeh return 0; /* no change */ 13413d14c5d2SYehuda Sadeh 13425b191d99SSage Weil dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n", 13435b191d99SSage Weil req->r_tid, pgid.pool, pgid.seed, o, 13443d14c5d2SYehuda Sadeh req->r_osd ? req->r_osd->o_osd : -1); 13453d14c5d2SYehuda Sadeh 13463d14c5d2SYehuda Sadeh /* record full pg acting set */ 13473d14c5d2SYehuda Sadeh memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num); 13483d14c5d2SYehuda Sadeh req->r_num_pg_osds = num; 13493d14c5d2SYehuda Sadeh 13503d14c5d2SYehuda Sadeh if (req->r_osd) { 13513d14c5d2SYehuda Sadeh __cancel_request(req); 13523d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 13533d14c5d2SYehuda Sadeh req->r_osd = NULL; 13543d14c5d2SYehuda Sadeh } 13553d14c5d2SYehuda Sadeh 13563d14c5d2SYehuda Sadeh req->r_osd = __lookup_osd(osdc, o); 13573d14c5d2SYehuda Sadeh if (!req->r_osd && o >= 0) { 13583d14c5d2SYehuda Sadeh err = -ENOMEM; 1359e10006f8SAlex Elder req->r_osd = create_osd(osdc, o); 13606f6c7006SSage Weil if (!req->r_osd) { 13616f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 13623d14c5d2SYehuda Sadeh goto out; 13636f6c7006SSage Weil } 13643d14c5d2SYehuda Sadeh 13656f6c7006SSage Weil dout("map_request osd %p is osd%d\n", req->r_osd, o); 13663d14c5d2SYehuda Sadeh __insert_osd(osdc, req->r_osd); 13673d14c5d2SYehuda Sadeh 1368b7a9e5ddSSage Weil ceph_con_open(&req->r_osd->o_con, 1369b7a9e5ddSSage Weil CEPH_ENTITY_TYPE_OSD, o, 1370b7a9e5ddSSage Weil &osdc->osdmap->osd_addr[o]); 13713d14c5d2SYehuda Sadeh } 13723d14c5d2SYehuda Sadeh 13733d14c5d2SYehuda Sadeh if (req->r_osd) { 13743d14c5d2SYehuda Sadeh __remove_osd_from_lru(req->r_osd); 1375ad885927SAlex Elder list_add_tail(&req->r_osd_item, &req->r_osd->o_requests); 1376ad885927SAlex Elder list_move_tail(&req->r_req_lru_item, &osdc->req_unsent); 13776f6c7006SSage Weil } else { 1378ad885927SAlex Elder list_move_tail(&req->r_req_lru_item, &osdc->req_notarget); 13793d14c5d2SYehuda Sadeh } 13803d14c5d2SYehuda Sadeh err = 1; /* osd or pg changed */ 13813d14c5d2SYehuda Sadeh 13823d14c5d2SYehuda Sadeh out: 13833d14c5d2SYehuda Sadeh return err; 13843d14c5d2SYehuda Sadeh } 13853d14c5d2SYehuda Sadeh 13863d14c5d2SYehuda Sadeh /* 13873d14c5d2SYehuda Sadeh * caller should hold map_sem (for read) and request_mutex 13883d14c5d2SYehuda Sadeh */ 138956e925b6SSage Weil static void __send_request(struct ceph_osd_client *osdc, 13903d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 13913d14c5d2SYehuda Sadeh { 13921b83bef2SSage Weil void *p; 13933d14c5d2SYehuda Sadeh 13941b83bef2SSage Weil dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n", 13951b83bef2SSage Weil req, req->r_tid, req->r_osd->o_osd, req->r_flags, 13961b83bef2SSage Weil (unsigned long long)req->r_pgid.pool, req->r_pgid.seed); 13973d14c5d2SYehuda Sadeh 13981b83bef2SSage Weil /* fill in message content that changes each time we send it */ 13991b83bef2SSage Weil put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch); 14001b83bef2SSage Weil put_unaligned_le32(req->r_flags, req->r_request_flags); 1401205ee118SIlya Dryomov put_unaligned_le64(req->r_target_oloc.pool, req->r_request_pool); 14021b83bef2SSage Weil p = req->r_request_pgid; 14031b83bef2SSage Weil ceph_encode_64(&p, req->r_pgid.pool); 14041b83bef2SSage Weil ceph_encode_32(&p, req->r_pgid.seed); 14051b83bef2SSage Weil put_unaligned_le64(1, req->r_request_attempts); /* FIXME */ 14061b83bef2SSage Weil memcpy(req->r_request_reassert_version, &req->r_reassert_version, 14071b83bef2SSage Weil sizeof(req->r_reassert_version)); 14082169aea6SSage Weil 14093d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 14103d14c5d2SYehuda Sadeh list_move_tail(&req->r_req_lru_item, &osdc->req_lru); 14113d14c5d2SYehuda Sadeh 14123d14c5d2SYehuda Sadeh ceph_msg_get(req->r_request); /* send consumes a ref */ 141326be8808SAlex Elder 14143d14c5d2SYehuda Sadeh req->r_sent = req->r_osd->o_incarnation; 141526be8808SAlex Elder 141626be8808SAlex Elder ceph_con_send(&req->r_osd->o_con, req->r_request); 14173d14c5d2SYehuda Sadeh } 14183d14c5d2SYehuda Sadeh 14193d14c5d2SYehuda Sadeh /* 14206f6c7006SSage Weil * Send any requests in the queue (req_unsent). 14216f6c7006SSage Weil */ 1422f9d25199SAlex Elder static void __send_queued(struct ceph_osd_client *osdc) 14236f6c7006SSage Weil { 14246f6c7006SSage Weil struct ceph_osd_request *req, *tmp; 14256f6c7006SSage Weil 1426f9d25199SAlex Elder dout("__send_queued\n"); 1427f9d25199SAlex Elder list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) 14286f6c7006SSage Weil __send_request(osdc, req); 14296f6c7006SSage Weil } 14306f6c7006SSage Weil 14316f6c7006SSage Weil /* 14320bbfdfe8SIlya Dryomov * Caller should hold map_sem for read and request_mutex. 14330bbfdfe8SIlya Dryomov */ 14340bbfdfe8SIlya Dryomov static int __ceph_osdc_start_request(struct ceph_osd_client *osdc, 14350bbfdfe8SIlya Dryomov struct ceph_osd_request *req, 14360bbfdfe8SIlya Dryomov bool nofail) 14370bbfdfe8SIlya Dryomov { 14380bbfdfe8SIlya Dryomov int rc; 14390bbfdfe8SIlya Dryomov 14400bbfdfe8SIlya Dryomov __register_request(osdc, req); 14410bbfdfe8SIlya Dryomov req->r_sent = 0; 14420bbfdfe8SIlya Dryomov req->r_got_reply = 0; 14430bbfdfe8SIlya Dryomov rc = __map_request(osdc, req, 0); 14440bbfdfe8SIlya Dryomov if (rc < 0) { 14450bbfdfe8SIlya Dryomov if (nofail) { 14460bbfdfe8SIlya Dryomov dout("osdc_start_request failed map, " 14470bbfdfe8SIlya Dryomov " will retry %lld\n", req->r_tid); 14480bbfdfe8SIlya Dryomov rc = 0; 14490bbfdfe8SIlya Dryomov } else { 14500bbfdfe8SIlya Dryomov __unregister_request(osdc, req); 14510bbfdfe8SIlya Dryomov } 14520bbfdfe8SIlya Dryomov return rc; 14530bbfdfe8SIlya Dryomov } 14540bbfdfe8SIlya Dryomov 14550bbfdfe8SIlya Dryomov if (req->r_osd == NULL) { 14560bbfdfe8SIlya Dryomov dout("send_request %p no up osds in pg\n", req); 14570bbfdfe8SIlya Dryomov ceph_monc_request_next_osdmap(&osdc->client->monc); 14580bbfdfe8SIlya Dryomov } else { 14590bbfdfe8SIlya Dryomov __send_queued(osdc); 14600bbfdfe8SIlya Dryomov } 14610bbfdfe8SIlya Dryomov 14620bbfdfe8SIlya Dryomov return 0; 14630bbfdfe8SIlya Dryomov } 14640bbfdfe8SIlya Dryomov 14650bbfdfe8SIlya Dryomov /* 14663d14c5d2SYehuda Sadeh * Timeout callback, called every N seconds when 1 or more osd 14673d14c5d2SYehuda Sadeh * requests has been active for more than N seconds. When this 14683d14c5d2SYehuda Sadeh * happens, we ping all OSDs with requests who have timed out to 14693d14c5d2SYehuda Sadeh * ensure any communications channel reset is detected. Reset the 14703d14c5d2SYehuda Sadeh * request timeouts another N seconds in the future as we go. 14713d14c5d2SYehuda Sadeh * Reschedule the timeout event another N seconds in future (unless 14723d14c5d2SYehuda Sadeh * there are no open requests). 14733d14c5d2SYehuda Sadeh */ 14743d14c5d2SYehuda Sadeh static void handle_timeout(struct work_struct *work) 14753d14c5d2SYehuda Sadeh { 14763d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 14773d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, timeout_work.work); 147883aff95eSSage Weil struct ceph_osd_request *req; 14793d14c5d2SYehuda Sadeh struct ceph_osd *osd; 14803d14c5d2SYehuda Sadeh unsigned long keepalive = 14813d14c5d2SYehuda Sadeh osdc->client->options->osd_keepalive_timeout * HZ; 14823d14c5d2SYehuda Sadeh struct list_head slow_osds; 14833d14c5d2SYehuda Sadeh dout("timeout\n"); 14843d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 14853d14c5d2SYehuda Sadeh 14863d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 14873d14c5d2SYehuda Sadeh 14883d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 14893d14c5d2SYehuda Sadeh 14903d14c5d2SYehuda Sadeh /* 14913d14c5d2SYehuda Sadeh * ping osds that are a bit slow. this ensures that if there 14923d14c5d2SYehuda Sadeh * is a break in the TCP connection we will notice, and reopen 14933d14c5d2SYehuda Sadeh * a connection with that osd (from the fault callback). 14943d14c5d2SYehuda Sadeh */ 14953d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&slow_osds); 14963d14c5d2SYehuda Sadeh list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) { 14973d14c5d2SYehuda Sadeh if (time_before(jiffies, req->r_stamp + keepalive)) 14983d14c5d2SYehuda Sadeh break; 14993d14c5d2SYehuda Sadeh 15003d14c5d2SYehuda Sadeh osd = req->r_osd; 15013d14c5d2SYehuda Sadeh BUG_ON(!osd); 15023d14c5d2SYehuda Sadeh dout(" tid %llu is slow, will send keepalive on osd%d\n", 15033d14c5d2SYehuda Sadeh req->r_tid, osd->o_osd); 15043d14c5d2SYehuda Sadeh list_move_tail(&osd->o_keepalive_item, &slow_osds); 15053d14c5d2SYehuda Sadeh } 15063d14c5d2SYehuda Sadeh while (!list_empty(&slow_osds)) { 15073d14c5d2SYehuda Sadeh osd = list_entry(slow_osds.next, struct ceph_osd, 15083d14c5d2SYehuda Sadeh o_keepalive_item); 15093d14c5d2SYehuda Sadeh list_del_init(&osd->o_keepalive_item); 15103d14c5d2SYehuda Sadeh ceph_con_keepalive(&osd->o_con); 15113d14c5d2SYehuda Sadeh } 15123d14c5d2SYehuda Sadeh 15133d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 1514f9d25199SAlex Elder __send_queued(osdc); 15153d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 15163d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 15173d14c5d2SYehuda Sadeh } 15183d14c5d2SYehuda Sadeh 15193d14c5d2SYehuda Sadeh static void handle_osds_timeout(struct work_struct *work) 15203d14c5d2SYehuda Sadeh { 15213d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 15223d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, 15233d14c5d2SYehuda Sadeh osds_timeout_work.work); 15243d14c5d2SYehuda Sadeh unsigned long delay = 15253d14c5d2SYehuda Sadeh osdc->client->options->osd_idle_ttl * HZ >> 2; 15263d14c5d2SYehuda Sadeh 15273d14c5d2SYehuda Sadeh dout("osds timeout\n"); 15283d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 1529aca420bcSSage Weil remove_old_osds(osdc); 15303d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 15313d14c5d2SYehuda Sadeh 15323d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 15333d14c5d2SYehuda Sadeh round_jiffies_relative(delay)); 15343d14c5d2SYehuda Sadeh } 15353d14c5d2SYehuda Sadeh 1536205ee118SIlya Dryomov static int ceph_oloc_decode(void **p, void *end, 1537205ee118SIlya Dryomov struct ceph_object_locator *oloc) 1538205ee118SIlya Dryomov { 1539205ee118SIlya Dryomov u8 struct_v, struct_cv; 1540205ee118SIlya Dryomov u32 len; 1541205ee118SIlya Dryomov void *struct_end; 1542205ee118SIlya Dryomov int ret = 0; 1543205ee118SIlya Dryomov 1544205ee118SIlya Dryomov ceph_decode_need(p, end, 1 + 1 + 4, e_inval); 1545205ee118SIlya Dryomov struct_v = ceph_decode_8(p); 1546205ee118SIlya Dryomov struct_cv = ceph_decode_8(p); 1547205ee118SIlya Dryomov if (struct_v < 3) { 1548205ee118SIlya Dryomov pr_warn("got v %d < 3 cv %d of ceph_object_locator\n", 1549205ee118SIlya Dryomov struct_v, struct_cv); 1550205ee118SIlya Dryomov goto e_inval; 1551205ee118SIlya Dryomov } 1552205ee118SIlya Dryomov if (struct_cv > 6) { 1553205ee118SIlya Dryomov pr_warn("got v %d cv %d > 6 of ceph_object_locator\n", 1554205ee118SIlya Dryomov struct_v, struct_cv); 1555205ee118SIlya Dryomov goto e_inval; 1556205ee118SIlya Dryomov } 1557205ee118SIlya Dryomov len = ceph_decode_32(p); 1558205ee118SIlya Dryomov ceph_decode_need(p, end, len, e_inval); 1559205ee118SIlya Dryomov struct_end = *p + len; 1560205ee118SIlya Dryomov 1561205ee118SIlya Dryomov oloc->pool = ceph_decode_64(p); 1562205ee118SIlya Dryomov *p += 4; /* skip preferred */ 1563205ee118SIlya Dryomov 1564205ee118SIlya Dryomov len = ceph_decode_32(p); 1565205ee118SIlya Dryomov if (len > 0) { 1566205ee118SIlya Dryomov pr_warn("ceph_object_locator::key is set\n"); 1567205ee118SIlya Dryomov goto e_inval; 1568205ee118SIlya Dryomov } 1569205ee118SIlya Dryomov 1570205ee118SIlya Dryomov if (struct_v >= 5) { 1571205ee118SIlya Dryomov len = ceph_decode_32(p); 1572205ee118SIlya Dryomov if (len > 0) { 1573205ee118SIlya Dryomov pr_warn("ceph_object_locator::nspace is set\n"); 1574205ee118SIlya Dryomov goto e_inval; 1575205ee118SIlya Dryomov } 1576205ee118SIlya Dryomov } 1577205ee118SIlya Dryomov 1578205ee118SIlya Dryomov if (struct_v >= 6) { 1579205ee118SIlya Dryomov s64 hash = ceph_decode_64(p); 1580205ee118SIlya Dryomov if (hash != -1) { 1581205ee118SIlya Dryomov pr_warn("ceph_object_locator::hash is set\n"); 1582205ee118SIlya Dryomov goto e_inval; 1583205ee118SIlya Dryomov } 1584205ee118SIlya Dryomov } 1585205ee118SIlya Dryomov 1586205ee118SIlya Dryomov /* skip the rest */ 1587205ee118SIlya Dryomov *p = struct_end; 1588205ee118SIlya Dryomov out: 1589205ee118SIlya Dryomov return ret; 1590205ee118SIlya Dryomov 1591205ee118SIlya Dryomov e_inval: 1592205ee118SIlya Dryomov ret = -EINVAL; 1593205ee118SIlya Dryomov goto out; 1594205ee118SIlya Dryomov } 1595205ee118SIlya Dryomov 1596205ee118SIlya Dryomov static int ceph_redirect_decode(void **p, void *end, 1597205ee118SIlya Dryomov struct ceph_request_redirect *redir) 1598205ee118SIlya Dryomov { 1599205ee118SIlya Dryomov u8 struct_v, struct_cv; 1600205ee118SIlya Dryomov u32 len; 1601205ee118SIlya Dryomov void *struct_end; 1602205ee118SIlya Dryomov int ret; 1603205ee118SIlya Dryomov 1604205ee118SIlya Dryomov ceph_decode_need(p, end, 1 + 1 + 4, e_inval); 1605205ee118SIlya Dryomov struct_v = ceph_decode_8(p); 1606205ee118SIlya Dryomov struct_cv = ceph_decode_8(p); 1607205ee118SIlya Dryomov if (struct_cv > 1) { 1608205ee118SIlya Dryomov pr_warn("got v %d cv %d > 1 of ceph_request_redirect\n", 1609205ee118SIlya Dryomov struct_v, struct_cv); 1610205ee118SIlya Dryomov goto e_inval; 1611205ee118SIlya Dryomov } 1612205ee118SIlya Dryomov len = ceph_decode_32(p); 1613205ee118SIlya Dryomov ceph_decode_need(p, end, len, e_inval); 1614205ee118SIlya Dryomov struct_end = *p + len; 1615205ee118SIlya Dryomov 1616205ee118SIlya Dryomov ret = ceph_oloc_decode(p, end, &redir->oloc); 1617205ee118SIlya Dryomov if (ret) 1618205ee118SIlya Dryomov goto out; 1619205ee118SIlya Dryomov 1620205ee118SIlya Dryomov len = ceph_decode_32(p); 1621205ee118SIlya Dryomov if (len > 0) { 1622205ee118SIlya Dryomov pr_warn("ceph_request_redirect::object_name is set\n"); 1623205ee118SIlya Dryomov goto e_inval; 1624205ee118SIlya Dryomov } 1625205ee118SIlya Dryomov 1626205ee118SIlya Dryomov len = ceph_decode_32(p); 1627205ee118SIlya Dryomov *p += len; /* skip osd_instructions */ 1628205ee118SIlya Dryomov 1629205ee118SIlya Dryomov /* skip the rest */ 1630205ee118SIlya Dryomov *p = struct_end; 1631205ee118SIlya Dryomov out: 1632205ee118SIlya Dryomov return ret; 1633205ee118SIlya Dryomov 1634205ee118SIlya Dryomov e_inval: 1635205ee118SIlya Dryomov ret = -EINVAL; 1636205ee118SIlya Dryomov goto out; 1637205ee118SIlya Dryomov } 1638205ee118SIlya Dryomov 163925845472SSage Weil static void complete_request(struct ceph_osd_request *req) 164025845472SSage Weil { 164125845472SSage Weil complete_all(&req->r_safe_completion); /* fsync waiter */ 164225845472SSage Weil } 164325845472SSage Weil 16443d14c5d2SYehuda Sadeh /* 16453d14c5d2SYehuda Sadeh * handle osd op reply. either call the callback if it is specified, 16463d14c5d2SYehuda Sadeh * or do the completion to wake up the waiting thread. 16473d14c5d2SYehuda Sadeh */ 16483d14c5d2SYehuda Sadeh static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg, 16493d14c5d2SYehuda Sadeh struct ceph_connection *con) 16503d14c5d2SYehuda Sadeh { 16511b83bef2SSage Weil void *p, *end; 16523d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 1653205ee118SIlya Dryomov struct ceph_request_redirect redir; 16543d14c5d2SYehuda Sadeh u64 tid; 16551b83bef2SSage Weil int object_len; 165679528734SAlex Elder unsigned int numops; 165779528734SAlex Elder int payload_len, flags; 16583d14c5d2SYehuda Sadeh s32 result; 16591b83bef2SSage Weil s32 retry_attempt; 16601b83bef2SSage Weil struct ceph_pg pg; 16611b83bef2SSage Weil int err; 16621b83bef2SSage Weil u32 reassert_epoch; 16631b83bef2SSage Weil u64 reassert_version; 16641b83bef2SSage Weil u32 osdmap_epoch; 16650d5af164SAlex Elder int already_completed; 16669fc6e064SAlex Elder u32 bytes; 166779528734SAlex Elder unsigned int i; 16683d14c5d2SYehuda Sadeh 16693d14c5d2SYehuda Sadeh tid = le64_to_cpu(msg->hdr.tid); 16701b83bef2SSage Weil dout("handle_reply %p tid %llu\n", msg, tid); 16711b83bef2SSage Weil 16721b83bef2SSage Weil p = msg->front.iov_base; 16731b83bef2SSage Weil end = p + msg->front.iov_len; 16741b83bef2SSage Weil 16751b83bef2SSage Weil ceph_decode_need(&p, end, 4, bad); 16761b83bef2SSage Weil object_len = ceph_decode_32(&p); 16771b83bef2SSage Weil ceph_decode_need(&p, end, object_len, bad); 16781b83bef2SSage Weil p += object_len; 16791b83bef2SSage Weil 1680ef4859d6SAlex Elder err = ceph_decode_pgid(&p, end, &pg); 16811b83bef2SSage Weil if (err) 16823d14c5d2SYehuda Sadeh goto bad; 16831b83bef2SSage Weil 16841b83bef2SSage Weil ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad); 16851b83bef2SSage Weil flags = ceph_decode_64(&p); 16861b83bef2SSage Weil result = ceph_decode_32(&p); 16871b83bef2SSage Weil reassert_epoch = ceph_decode_32(&p); 16881b83bef2SSage Weil reassert_version = ceph_decode_64(&p); 16891b83bef2SSage Weil osdmap_epoch = ceph_decode_32(&p); 16901b83bef2SSage Weil 16913d14c5d2SYehuda Sadeh /* lookup */ 1692ff513aceSIlya Dryomov down_read(&osdc->map_sem); 16933d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 16943d14c5d2SYehuda Sadeh req = __lookup_request(osdc, tid); 16953d14c5d2SYehuda Sadeh if (req == NULL) { 16963d14c5d2SYehuda Sadeh dout("handle_reply tid %llu dne\n", tid); 16978058fd45SAlex Elder goto bad_mutex; 16983d14c5d2SYehuda Sadeh } 16993d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 17001b83bef2SSage Weil 17011b83bef2SSage Weil dout("handle_reply %p tid %llu req %p result %d\n", msg, tid, 17021b83bef2SSage Weil req, result); 17031b83bef2SSage Weil 170418741196SDan Carpenter ceph_decode_need(&p, end, 4, bad_put); 17051b83bef2SSage Weil numops = ceph_decode_32(&p); 17061b83bef2SSage Weil if (numops > CEPH_OSD_MAX_OP) 17071b83bef2SSage Weil goto bad_put; 17081b83bef2SSage Weil if (numops != req->r_num_ops) 17091b83bef2SSage Weil goto bad_put; 17101b83bef2SSage Weil payload_len = 0; 171118741196SDan Carpenter ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad_put); 17121b83bef2SSage Weil for (i = 0; i < numops; i++) { 17131b83bef2SSage Weil struct ceph_osd_op *op = p; 17141b83bef2SSage Weil int len; 17151b83bef2SSage Weil 17161b83bef2SSage Weil len = le32_to_cpu(op->payload_len); 17171b83bef2SSage Weil req->r_reply_op_len[i] = len; 17181b83bef2SSage Weil dout(" op %d has %d bytes\n", i, len); 17191b83bef2SSage Weil payload_len += len; 17201b83bef2SSage Weil p += sizeof(*op); 17211b83bef2SSage Weil } 17229fc6e064SAlex Elder bytes = le32_to_cpu(msg->hdr.data_len); 17239fc6e064SAlex Elder if (payload_len != bytes) { 17241b83bef2SSage Weil pr_warning("sum of op payload lens %d != data_len %d", 17259fc6e064SAlex Elder payload_len, bytes); 17261b83bef2SSage Weil goto bad_put; 17271b83bef2SSage Weil } 17281b83bef2SSage Weil 172918741196SDan Carpenter ceph_decode_need(&p, end, 4 + numops * 4, bad_put); 17301b83bef2SSage Weil retry_attempt = ceph_decode_32(&p); 17311b83bef2SSage Weil for (i = 0; i < numops; i++) 17321b83bef2SSage Weil req->r_reply_op_result[i] = ceph_decode_32(&p); 17333d14c5d2SYehuda Sadeh 1734205ee118SIlya Dryomov if (le16_to_cpu(msg->hdr.version) >= 6) { 1735205ee118SIlya Dryomov p += 8 + 4; /* skip replay_version */ 1736205ee118SIlya Dryomov p += 8; /* skip user_version */ 1737205ee118SIlya Dryomov 1738205ee118SIlya Dryomov err = ceph_redirect_decode(&p, end, &redir); 1739205ee118SIlya Dryomov if (err) 1740205ee118SIlya Dryomov goto bad_put; 1741205ee118SIlya Dryomov } else { 1742205ee118SIlya Dryomov redir.oloc.pool = -1; 1743205ee118SIlya Dryomov } 1744205ee118SIlya Dryomov 1745205ee118SIlya Dryomov if (redir.oloc.pool != -1) { 1746205ee118SIlya Dryomov dout("redirect pool %lld\n", redir.oloc.pool); 1747205ee118SIlya Dryomov 1748205ee118SIlya Dryomov __unregister_request(osdc, req); 1749205ee118SIlya Dryomov 1750205ee118SIlya Dryomov req->r_target_oloc = redir.oloc; /* struct */ 1751205ee118SIlya Dryomov 1752205ee118SIlya Dryomov /* 1753205ee118SIlya Dryomov * Start redirect requests with nofail=true. If 1754205ee118SIlya Dryomov * mapping fails, request will end up on the notarget 1755205ee118SIlya Dryomov * list, waiting for the new osdmap (which can take 1756205ee118SIlya Dryomov * a while), even though the original request mapped 1757205ee118SIlya Dryomov * successfully. In the future we might want to follow 1758205ee118SIlya Dryomov * original request's nofail setting here. 1759205ee118SIlya Dryomov */ 1760ff513aceSIlya Dryomov err = __ceph_osdc_start_request(osdc, req, true); 1761205ee118SIlya Dryomov BUG_ON(err); 1762205ee118SIlya Dryomov 1763ff513aceSIlya Dryomov goto out_unlock; 1764205ee118SIlya Dryomov } 1765205ee118SIlya Dryomov 1766eb845ff1SYan, Zheng already_completed = req->r_got_reply; 17673d14c5d2SYehuda Sadeh if (!req->r_got_reply) { 17681b83bef2SSage Weil req->r_result = result; 17693d14c5d2SYehuda Sadeh dout("handle_reply result %d bytes %d\n", req->r_result, 17703d14c5d2SYehuda Sadeh bytes); 17713d14c5d2SYehuda Sadeh if (req->r_result == 0) 17723d14c5d2SYehuda Sadeh req->r_result = bytes; 17733d14c5d2SYehuda Sadeh 17743d14c5d2SYehuda Sadeh /* in case this is a write and we need to replay, */ 17751b83bef2SSage Weil req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch); 17761b83bef2SSage Weil req->r_reassert_version.version = cpu_to_le64(reassert_version); 17773d14c5d2SYehuda Sadeh 17783d14c5d2SYehuda Sadeh req->r_got_reply = 1; 17793d14c5d2SYehuda Sadeh } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) { 17803d14c5d2SYehuda Sadeh dout("handle_reply tid %llu dup ack\n", tid); 1781ff513aceSIlya Dryomov goto out_unlock; 17823d14c5d2SYehuda Sadeh } 17833d14c5d2SYehuda Sadeh 17843d14c5d2SYehuda Sadeh dout("handle_reply tid %llu flags %d\n", tid, flags); 17853d14c5d2SYehuda Sadeh 1786a40c4f10SYehuda Sadeh if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK)) 1787a40c4f10SYehuda Sadeh __register_linger_request(osdc, req); 1788a40c4f10SYehuda Sadeh 17893d14c5d2SYehuda Sadeh /* either this is a read, or we got the safe response */ 17903d14c5d2SYehuda Sadeh if (result < 0 || 17913d14c5d2SYehuda Sadeh (flags & CEPH_OSD_FLAG_ONDISK) || 17923d14c5d2SYehuda Sadeh ((flags & CEPH_OSD_FLAG_WRITE) == 0)) 17933d14c5d2SYehuda Sadeh __unregister_request(osdc, req); 17943d14c5d2SYehuda Sadeh 17953d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 1796ff513aceSIlya Dryomov up_read(&osdc->map_sem); 17973d14c5d2SYehuda Sadeh 1798eb845ff1SYan, Zheng if (!already_completed) { 179961c5d6bfSYan, Zheng if (req->r_unsafe_callback && 180061c5d6bfSYan, Zheng result >= 0 && !(flags & CEPH_OSD_FLAG_ONDISK)) 180161c5d6bfSYan, Zheng req->r_unsafe_callback(req, true); 18023d14c5d2SYehuda Sadeh if (req->r_callback) 18033d14c5d2SYehuda Sadeh req->r_callback(req, msg); 18043d14c5d2SYehuda Sadeh else 18053d14c5d2SYehuda Sadeh complete_all(&req->r_completion); 1806eb845ff1SYan, Zheng } 18073d14c5d2SYehuda Sadeh 180861c5d6bfSYan, Zheng if (flags & CEPH_OSD_FLAG_ONDISK) { 180961c5d6bfSYan, Zheng if (req->r_unsafe_callback && already_completed) 181061c5d6bfSYan, Zheng req->r_unsafe_callback(req, false); 181125845472SSage Weil complete_request(req); 181261c5d6bfSYan, Zheng } 18133d14c5d2SYehuda Sadeh 1814ff513aceSIlya Dryomov out: 1815a40c4f10SYehuda Sadeh dout("req=%p req->r_linger=%d\n", req, req->r_linger); 18163d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 18173d14c5d2SYehuda Sadeh return; 1818ff513aceSIlya Dryomov out_unlock: 1819ff513aceSIlya Dryomov mutex_unlock(&osdc->request_mutex); 1820ff513aceSIlya Dryomov up_read(&osdc->map_sem); 1821ff513aceSIlya Dryomov goto out; 18223d14c5d2SYehuda Sadeh 18231b83bef2SSage Weil bad_put: 182437c89bdeSLi Wang req->r_result = -EIO; 182537c89bdeSLi Wang __unregister_request(osdc, req); 182637c89bdeSLi Wang if (req->r_callback) 182737c89bdeSLi Wang req->r_callback(req, msg); 182837c89bdeSLi Wang else 182937c89bdeSLi Wang complete_all(&req->r_completion); 183037c89bdeSLi Wang complete_request(req); 18311b83bef2SSage Weil ceph_osdc_put_request(req); 18328058fd45SAlex Elder bad_mutex: 18338058fd45SAlex Elder mutex_unlock(&osdc->request_mutex); 1834ff513aceSIlya Dryomov up_read(&osdc->map_sem); 18353d14c5d2SYehuda Sadeh bad: 18361b83bef2SSage Weil pr_err("corrupt osd_op_reply got %d %d\n", 18371b83bef2SSage Weil (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len)); 18383d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 18393d14c5d2SYehuda Sadeh } 18403d14c5d2SYehuda Sadeh 18416f6c7006SSage Weil static void reset_changed_osds(struct ceph_osd_client *osdc) 18423d14c5d2SYehuda Sadeh { 18433d14c5d2SYehuda Sadeh struct rb_node *p, *n; 18443d14c5d2SYehuda Sadeh 18453d14c5d2SYehuda Sadeh for (p = rb_first(&osdc->osds); p; p = n) { 18466f6c7006SSage Weil struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node); 18473d14c5d2SYehuda Sadeh 18483d14c5d2SYehuda Sadeh n = rb_next(p); 18493d14c5d2SYehuda Sadeh if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) || 18503d14c5d2SYehuda Sadeh memcmp(&osd->o_con.peer_addr, 18513d14c5d2SYehuda Sadeh ceph_osd_addr(osdc->osdmap, 18523d14c5d2SYehuda Sadeh osd->o_osd), 18533d14c5d2SYehuda Sadeh sizeof(struct ceph_entity_addr)) != 0) 18543d14c5d2SYehuda Sadeh __reset_osd(osdc, osd); 18553d14c5d2SYehuda Sadeh } 18563d14c5d2SYehuda Sadeh } 18573d14c5d2SYehuda Sadeh 18583d14c5d2SYehuda Sadeh /* 18596f6c7006SSage Weil * Requeue requests whose mapping to an OSD has changed. If requests map to 18606f6c7006SSage Weil * no osd, request a new map. 18613d14c5d2SYehuda Sadeh * 1862e6d50f67SAlex Elder * Caller should hold map_sem for read. 18633d14c5d2SYehuda Sadeh */ 18649a1ea2dbSJosh Durgin static void kick_requests(struct ceph_osd_client *osdc, bool force_resend, 18659a1ea2dbSJosh Durgin bool force_resend_writes) 18663d14c5d2SYehuda Sadeh { 1867a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 18686f6c7006SSage Weil struct rb_node *p; 18696f6c7006SSage Weil int needmap = 0; 18706f6c7006SSage Weil int err; 18719a1ea2dbSJosh Durgin bool force_resend_req; 18723d14c5d2SYehuda Sadeh 18739a1ea2dbSJosh Durgin dout("kick_requests %s %s\n", force_resend ? " (force resend)" : "", 18749a1ea2dbSJosh Durgin force_resend_writes ? " (force resend writes)" : ""); 18753d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 18766194ea89SSage Weil for (p = rb_first(&osdc->requests); p; ) { 18776f6c7006SSage Weil req = rb_entry(p, struct ceph_osd_request, r_node); 18786194ea89SSage Weil p = rb_next(p); 1879ab60b16dSAlex Elder 1880ab60b16dSAlex Elder /* 1881ab60b16dSAlex Elder * For linger requests that have not yet been 1882ab60b16dSAlex Elder * registered, move them to the linger list; they'll 1883ab60b16dSAlex Elder * be sent to the osd in the loop below. Unregister 1884ab60b16dSAlex Elder * the request before re-registering it as a linger 1885ab60b16dSAlex Elder * request to ensure the __map_request() below 1886ab60b16dSAlex Elder * will decide it needs to be sent. 1887ab60b16dSAlex Elder */ 1888ab60b16dSAlex Elder if (req->r_linger && list_empty(&req->r_linger_item)) { 1889ab60b16dSAlex Elder dout("%p tid %llu restart on osd%d\n", 1890ab60b16dSAlex Elder req, req->r_tid, 1891ab60b16dSAlex Elder req->r_osd ? req->r_osd->o_osd : -1); 189296e4dac6SAlex Elder ceph_osdc_get_request(req); 1893ab60b16dSAlex Elder __unregister_request(osdc, req); 1894ab60b16dSAlex Elder __register_linger_request(osdc, req); 189596e4dac6SAlex Elder ceph_osdc_put_request(req); 1896ab60b16dSAlex Elder continue; 1897ab60b16dSAlex Elder } 1898ab60b16dSAlex Elder 18999a1ea2dbSJosh Durgin force_resend_req = force_resend || 19009a1ea2dbSJosh Durgin (force_resend_writes && 19019a1ea2dbSJosh Durgin req->r_flags & CEPH_OSD_FLAG_WRITE); 19029a1ea2dbSJosh Durgin err = __map_request(osdc, req, force_resend_req); 19036f6c7006SSage Weil if (err < 0) 19046f6c7006SSage Weil continue; /* error */ 19056f6c7006SSage Weil if (req->r_osd == NULL) { 19066f6c7006SSage Weil dout("%p tid %llu maps to no osd\n", req, req->r_tid); 19076f6c7006SSage Weil needmap++; /* request a newer map */ 19086f6c7006SSage Weil } else if (err > 0) { 19096194ea89SSage Weil if (!req->r_linger) { 19106194ea89SSage Weil dout("%p tid %llu requeued on osd%d\n", req, 19116194ea89SSage Weil req->r_tid, 19126f6c7006SSage Weil req->r_osd ? req->r_osd->o_osd : -1); 19136f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 19146f6c7006SSage Weil } 19156f6c7006SSage Weil } 19166194ea89SSage Weil } 1917a40c4f10SYehuda Sadeh 1918a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osdc->req_linger, 1919a40c4f10SYehuda Sadeh r_linger_item) { 1920a40c4f10SYehuda Sadeh dout("linger req=%p req->r_osd=%p\n", req, req->r_osd); 1921a40c4f10SYehuda Sadeh 19229a1ea2dbSJosh Durgin err = __map_request(osdc, req, 19239a1ea2dbSJosh Durgin force_resend || force_resend_writes); 1924ab60b16dSAlex Elder dout("__map_request returned %d\n", err); 1925a40c4f10SYehuda Sadeh if (err == 0) 1926a40c4f10SYehuda Sadeh continue; /* no change and no osd was specified */ 1927a40c4f10SYehuda Sadeh if (err < 0) 1928a40c4f10SYehuda Sadeh continue; /* hrm! */ 1929a40c4f10SYehuda Sadeh if (req->r_osd == NULL) { 1930a40c4f10SYehuda Sadeh dout("tid %llu maps to no valid osd\n", req->r_tid); 1931a40c4f10SYehuda Sadeh needmap++; /* request a newer map */ 1932a40c4f10SYehuda Sadeh continue; 1933a40c4f10SYehuda Sadeh } 1934a40c4f10SYehuda Sadeh 1935a40c4f10SYehuda Sadeh dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid, 1936a40c4f10SYehuda Sadeh req->r_osd ? req->r_osd->o_osd : -1); 1937a40c4f10SYehuda Sadeh __register_request(osdc, req); 1938c89ce05eSAlex Elder __unregister_linger_request(osdc, req); 1939a40c4f10SYehuda Sadeh } 194014d2f38dSAlex Elder reset_changed_osds(osdc); 19413d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 19423d14c5d2SYehuda Sadeh 19433d14c5d2SYehuda Sadeh if (needmap) { 19443d14c5d2SYehuda Sadeh dout("%d requests for down osds, need new map\n", needmap); 19453d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 19463d14c5d2SYehuda Sadeh } 19473d14c5d2SYehuda Sadeh } 19486f6c7006SSage Weil 19496f6c7006SSage Weil 19503d14c5d2SYehuda Sadeh /* 19513d14c5d2SYehuda Sadeh * Process updated osd map. 19523d14c5d2SYehuda Sadeh * 19533d14c5d2SYehuda Sadeh * The message contains any number of incremental and full maps, normally 19543d14c5d2SYehuda Sadeh * indicating some sort of topology change in the cluster. Kick requests 19553d14c5d2SYehuda Sadeh * off to different OSDs as needed. 19563d14c5d2SYehuda Sadeh */ 19573d14c5d2SYehuda Sadeh void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg) 19583d14c5d2SYehuda Sadeh { 19593d14c5d2SYehuda Sadeh void *p, *end, *next; 19603d14c5d2SYehuda Sadeh u32 nr_maps, maplen; 19613d14c5d2SYehuda Sadeh u32 epoch; 19623d14c5d2SYehuda Sadeh struct ceph_osdmap *newmap = NULL, *oldmap; 19633d14c5d2SYehuda Sadeh int err; 19643d14c5d2SYehuda Sadeh struct ceph_fsid fsid; 19659a1ea2dbSJosh Durgin bool was_full; 19663d14c5d2SYehuda Sadeh 19673d14c5d2SYehuda Sadeh dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0); 19683d14c5d2SYehuda Sadeh p = msg->front.iov_base; 19693d14c5d2SYehuda Sadeh end = p + msg->front.iov_len; 19703d14c5d2SYehuda Sadeh 19713d14c5d2SYehuda Sadeh /* verify fsid */ 19723d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, sizeof(fsid), bad); 19733d14c5d2SYehuda Sadeh ceph_decode_copy(&p, &fsid, sizeof(fsid)); 19743d14c5d2SYehuda Sadeh if (ceph_check_fsid(osdc->client, &fsid) < 0) 19753d14c5d2SYehuda Sadeh return; 19763d14c5d2SYehuda Sadeh 19773d14c5d2SYehuda Sadeh down_write(&osdc->map_sem); 19783d14c5d2SYehuda Sadeh 19799a1ea2dbSJosh Durgin was_full = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL); 19809a1ea2dbSJosh Durgin 19813d14c5d2SYehuda Sadeh /* incremental maps */ 19823d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 19833d14c5d2SYehuda Sadeh dout(" %d inc maps\n", nr_maps); 19843d14c5d2SYehuda Sadeh while (nr_maps > 0) { 19853d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 19863d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 19873d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 19883d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 19893d14c5d2SYehuda Sadeh next = p + maplen; 19903d14c5d2SYehuda Sadeh if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) { 19913d14c5d2SYehuda Sadeh dout("applying incremental map %u len %d\n", 19923d14c5d2SYehuda Sadeh epoch, maplen); 19933d14c5d2SYehuda Sadeh newmap = osdmap_apply_incremental(&p, next, 19943d14c5d2SYehuda Sadeh osdc->osdmap, 199515d9882cSAlex Elder &osdc->client->msgr); 19963d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 19973d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 19983d14c5d2SYehuda Sadeh goto bad; 19993d14c5d2SYehuda Sadeh } 20003d14c5d2SYehuda Sadeh BUG_ON(!newmap); 20013d14c5d2SYehuda Sadeh if (newmap != osdc->osdmap) { 20023d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 20033d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 20043d14c5d2SYehuda Sadeh } 20059a1ea2dbSJosh Durgin was_full = was_full || 20069a1ea2dbSJosh Durgin ceph_osdmap_flag(osdc->osdmap, 20079a1ea2dbSJosh Durgin CEPH_OSDMAP_FULL); 20089a1ea2dbSJosh Durgin kick_requests(osdc, 0, was_full); 20093d14c5d2SYehuda Sadeh } else { 20103d14c5d2SYehuda Sadeh dout("ignoring incremental map %u len %d\n", 20113d14c5d2SYehuda Sadeh epoch, maplen); 20123d14c5d2SYehuda Sadeh } 20133d14c5d2SYehuda Sadeh p = next; 20143d14c5d2SYehuda Sadeh nr_maps--; 20153d14c5d2SYehuda Sadeh } 20163d14c5d2SYehuda Sadeh if (newmap) 20173d14c5d2SYehuda Sadeh goto done; 20183d14c5d2SYehuda Sadeh 20193d14c5d2SYehuda Sadeh /* full maps */ 20203d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 20213d14c5d2SYehuda Sadeh dout(" %d full maps\n", nr_maps); 20223d14c5d2SYehuda Sadeh while (nr_maps) { 20233d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 20243d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 20253d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 20263d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 20273d14c5d2SYehuda Sadeh if (nr_maps > 1) { 20283d14c5d2SYehuda Sadeh dout("skipping non-latest full map %u len %d\n", 20293d14c5d2SYehuda Sadeh epoch, maplen); 20303d14c5d2SYehuda Sadeh } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) { 20313d14c5d2SYehuda Sadeh dout("skipping full map %u len %d, " 20323d14c5d2SYehuda Sadeh "older than our %u\n", epoch, maplen, 20333d14c5d2SYehuda Sadeh osdc->osdmap->epoch); 20343d14c5d2SYehuda Sadeh } else { 203538d6453cSSage Weil int skipped_map = 0; 203638d6453cSSage Weil 20373d14c5d2SYehuda Sadeh dout("taking full map %u len %d\n", epoch, maplen); 20383d14c5d2SYehuda Sadeh newmap = osdmap_decode(&p, p+maplen); 20393d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 20403d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 20413d14c5d2SYehuda Sadeh goto bad; 20423d14c5d2SYehuda Sadeh } 20433d14c5d2SYehuda Sadeh BUG_ON(!newmap); 20443d14c5d2SYehuda Sadeh oldmap = osdc->osdmap; 20453d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 204638d6453cSSage Weil if (oldmap) { 204738d6453cSSage Weil if (oldmap->epoch + 1 < newmap->epoch) 204838d6453cSSage Weil skipped_map = 1; 20493d14c5d2SYehuda Sadeh ceph_osdmap_destroy(oldmap); 205038d6453cSSage Weil } 20519a1ea2dbSJosh Durgin was_full = was_full || 20529a1ea2dbSJosh Durgin ceph_osdmap_flag(osdc->osdmap, 20539a1ea2dbSJosh Durgin CEPH_OSDMAP_FULL); 20549a1ea2dbSJosh Durgin kick_requests(osdc, skipped_map, was_full); 20553d14c5d2SYehuda Sadeh } 20563d14c5d2SYehuda Sadeh p += maplen; 20573d14c5d2SYehuda Sadeh nr_maps--; 20583d14c5d2SYehuda Sadeh } 20593d14c5d2SYehuda Sadeh 2060b72e19b9SDan Carpenter if (!osdc->osdmap) 2061b72e19b9SDan Carpenter goto bad; 20623d14c5d2SYehuda Sadeh done: 20633d14c5d2SYehuda Sadeh downgrade_write(&osdc->map_sem); 20643d14c5d2SYehuda Sadeh ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch); 2065cd634fb6SSage Weil 2066cd634fb6SSage Weil /* 2067cd634fb6SSage Weil * subscribe to subsequent osdmap updates if full to ensure 2068cd634fb6SSage Weil * we find out when we are no longer full and stop returning 2069cd634fb6SSage Weil * ENOSPC. 2070cd634fb6SSage Weil */ 2071d29adb34SJosh Durgin if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) || 2072d29adb34SJosh Durgin ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD) || 2073d29adb34SJosh Durgin ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR)) 2074cd634fb6SSage Weil ceph_monc_request_next_osdmap(&osdc->client->monc); 2075cd634fb6SSage Weil 2076f9d25199SAlex Elder mutex_lock(&osdc->request_mutex); 2077f9d25199SAlex Elder __send_queued(osdc); 2078f9d25199SAlex Elder mutex_unlock(&osdc->request_mutex); 20793d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 20803d14c5d2SYehuda Sadeh wake_up_all(&osdc->client->auth_wq); 20813d14c5d2SYehuda Sadeh return; 20823d14c5d2SYehuda Sadeh 20833d14c5d2SYehuda Sadeh bad: 20843d14c5d2SYehuda Sadeh pr_err("osdc handle_map corrupt msg\n"); 20853d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 20863d14c5d2SYehuda Sadeh up_write(&osdc->map_sem); 20873d14c5d2SYehuda Sadeh return; 20883d14c5d2SYehuda Sadeh } 20893d14c5d2SYehuda Sadeh 20903d14c5d2SYehuda Sadeh /* 2091a40c4f10SYehuda Sadeh * watch/notify callback event infrastructure 2092a40c4f10SYehuda Sadeh * 2093a40c4f10SYehuda Sadeh * These callbacks are used both for watch and notify operations. 2094a40c4f10SYehuda Sadeh */ 2095a40c4f10SYehuda Sadeh static void __release_event(struct kref *kref) 2096a40c4f10SYehuda Sadeh { 2097a40c4f10SYehuda Sadeh struct ceph_osd_event *event = 2098a40c4f10SYehuda Sadeh container_of(kref, struct ceph_osd_event, kref); 2099a40c4f10SYehuda Sadeh 2100a40c4f10SYehuda Sadeh dout("__release_event %p\n", event); 2101a40c4f10SYehuda Sadeh kfree(event); 2102a40c4f10SYehuda Sadeh } 2103a40c4f10SYehuda Sadeh 2104a40c4f10SYehuda Sadeh static void get_event(struct ceph_osd_event *event) 2105a40c4f10SYehuda Sadeh { 2106a40c4f10SYehuda Sadeh kref_get(&event->kref); 2107a40c4f10SYehuda Sadeh } 2108a40c4f10SYehuda Sadeh 2109a40c4f10SYehuda Sadeh void ceph_osdc_put_event(struct ceph_osd_event *event) 2110a40c4f10SYehuda Sadeh { 2111a40c4f10SYehuda Sadeh kref_put(&event->kref, __release_event); 2112a40c4f10SYehuda Sadeh } 2113a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_put_event); 2114a40c4f10SYehuda Sadeh 2115a40c4f10SYehuda Sadeh static void __insert_event(struct ceph_osd_client *osdc, 2116a40c4f10SYehuda Sadeh struct ceph_osd_event *new) 2117a40c4f10SYehuda Sadeh { 2118a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 2119a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 2120a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 2121a40c4f10SYehuda Sadeh 2122a40c4f10SYehuda Sadeh while (*p) { 2123a40c4f10SYehuda Sadeh parent = *p; 2124a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 2125a40c4f10SYehuda Sadeh if (new->cookie < event->cookie) 2126a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 2127a40c4f10SYehuda Sadeh else if (new->cookie > event->cookie) 2128a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 2129a40c4f10SYehuda Sadeh else 2130a40c4f10SYehuda Sadeh BUG(); 2131a40c4f10SYehuda Sadeh } 2132a40c4f10SYehuda Sadeh 2133a40c4f10SYehuda Sadeh rb_link_node(&new->node, parent, p); 2134a40c4f10SYehuda Sadeh rb_insert_color(&new->node, &osdc->event_tree); 2135a40c4f10SYehuda Sadeh } 2136a40c4f10SYehuda Sadeh 2137a40c4f10SYehuda Sadeh static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc, 2138a40c4f10SYehuda Sadeh u64 cookie) 2139a40c4f10SYehuda Sadeh { 2140a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 2141a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 2142a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 2143a40c4f10SYehuda Sadeh 2144a40c4f10SYehuda Sadeh while (*p) { 2145a40c4f10SYehuda Sadeh parent = *p; 2146a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 2147a40c4f10SYehuda Sadeh if (cookie < event->cookie) 2148a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 2149a40c4f10SYehuda Sadeh else if (cookie > event->cookie) 2150a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 2151a40c4f10SYehuda Sadeh else 2152a40c4f10SYehuda Sadeh return event; 2153a40c4f10SYehuda Sadeh } 2154a40c4f10SYehuda Sadeh return NULL; 2155a40c4f10SYehuda Sadeh } 2156a40c4f10SYehuda Sadeh 2157a40c4f10SYehuda Sadeh static void __remove_event(struct ceph_osd_event *event) 2158a40c4f10SYehuda Sadeh { 2159a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 2160a40c4f10SYehuda Sadeh 2161a40c4f10SYehuda Sadeh if (!RB_EMPTY_NODE(&event->node)) { 2162a40c4f10SYehuda Sadeh dout("__remove_event removed %p\n", event); 2163a40c4f10SYehuda Sadeh rb_erase(&event->node, &osdc->event_tree); 2164a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 2165a40c4f10SYehuda Sadeh } else { 2166a40c4f10SYehuda Sadeh dout("__remove_event didn't remove %p\n", event); 2167a40c4f10SYehuda Sadeh } 2168a40c4f10SYehuda Sadeh } 2169a40c4f10SYehuda Sadeh 2170a40c4f10SYehuda Sadeh int ceph_osdc_create_event(struct ceph_osd_client *osdc, 2171a40c4f10SYehuda Sadeh void (*event_cb)(u64, u64, u8, void *), 21723c663bbdSAlex Elder void *data, struct ceph_osd_event **pevent) 2173a40c4f10SYehuda Sadeh { 2174a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 2175a40c4f10SYehuda Sadeh 2176a40c4f10SYehuda Sadeh event = kmalloc(sizeof(*event), GFP_NOIO); 2177a40c4f10SYehuda Sadeh if (!event) 2178a40c4f10SYehuda Sadeh return -ENOMEM; 2179a40c4f10SYehuda Sadeh 2180a40c4f10SYehuda Sadeh dout("create_event %p\n", event); 2181a40c4f10SYehuda Sadeh event->cb = event_cb; 21823c663bbdSAlex Elder event->one_shot = 0; 2183a40c4f10SYehuda Sadeh event->data = data; 2184a40c4f10SYehuda Sadeh event->osdc = osdc; 2185a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&event->osd_node); 21863ee5234dSAlex Elder RB_CLEAR_NODE(&event->node); 2187a40c4f10SYehuda Sadeh kref_init(&event->kref); /* one ref for us */ 2188a40c4f10SYehuda Sadeh kref_get(&event->kref); /* one ref for the caller */ 2189a40c4f10SYehuda Sadeh 2190a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 2191a40c4f10SYehuda Sadeh event->cookie = ++osdc->event_count; 2192a40c4f10SYehuda Sadeh __insert_event(osdc, event); 2193a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 2194a40c4f10SYehuda Sadeh 2195a40c4f10SYehuda Sadeh *pevent = event; 2196a40c4f10SYehuda Sadeh return 0; 2197a40c4f10SYehuda Sadeh } 2198a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_create_event); 2199a40c4f10SYehuda Sadeh 2200a40c4f10SYehuda Sadeh void ceph_osdc_cancel_event(struct ceph_osd_event *event) 2201a40c4f10SYehuda Sadeh { 2202a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 2203a40c4f10SYehuda Sadeh 2204a40c4f10SYehuda Sadeh dout("cancel_event %p\n", event); 2205a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 2206a40c4f10SYehuda Sadeh __remove_event(event); 2207a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 2208a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); /* caller's */ 2209a40c4f10SYehuda Sadeh } 2210a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_cancel_event); 2211a40c4f10SYehuda Sadeh 2212a40c4f10SYehuda Sadeh 2213a40c4f10SYehuda Sadeh static void do_event_work(struct work_struct *work) 2214a40c4f10SYehuda Sadeh { 2215a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work = 2216a40c4f10SYehuda Sadeh container_of(work, struct ceph_osd_event_work, work); 2217a40c4f10SYehuda Sadeh struct ceph_osd_event *event = event_work->event; 2218a40c4f10SYehuda Sadeh u64 ver = event_work->ver; 2219a40c4f10SYehuda Sadeh u64 notify_id = event_work->notify_id; 2220a40c4f10SYehuda Sadeh u8 opcode = event_work->opcode; 2221a40c4f10SYehuda Sadeh 2222a40c4f10SYehuda Sadeh dout("do_event_work completing %p\n", event); 2223a40c4f10SYehuda Sadeh event->cb(ver, notify_id, opcode, event->data); 2224a40c4f10SYehuda Sadeh dout("do_event_work completed %p\n", event); 2225a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 2226a40c4f10SYehuda Sadeh kfree(event_work); 2227a40c4f10SYehuda Sadeh } 2228a40c4f10SYehuda Sadeh 2229a40c4f10SYehuda Sadeh 2230a40c4f10SYehuda Sadeh /* 2231a40c4f10SYehuda Sadeh * Process osd watch notifications 2232a40c4f10SYehuda Sadeh */ 22333c663bbdSAlex Elder static void handle_watch_notify(struct ceph_osd_client *osdc, 22343c663bbdSAlex Elder struct ceph_msg *msg) 2235a40c4f10SYehuda Sadeh { 2236a40c4f10SYehuda Sadeh void *p, *end; 2237a40c4f10SYehuda Sadeh u8 proto_ver; 2238a40c4f10SYehuda Sadeh u64 cookie, ver, notify_id; 2239a40c4f10SYehuda Sadeh u8 opcode; 2240a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 2241a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work; 2242a40c4f10SYehuda Sadeh 2243a40c4f10SYehuda Sadeh p = msg->front.iov_base; 2244a40c4f10SYehuda Sadeh end = p + msg->front.iov_len; 2245a40c4f10SYehuda Sadeh 2246a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, proto_ver, bad); 2247a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, opcode, bad); 2248a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, cookie, bad); 2249a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, ver, bad); 2250a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, notify_id, bad); 2251a40c4f10SYehuda Sadeh 2252a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 2253a40c4f10SYehuda Sadeh event = __find_event(osdc, cookie); 2254a40c4f10SYehuda Sadeh if (event) { 22553c663bbdSAlex Elder BUG_ON(event->one_shot); 2256a40c4f10SYehuda Sadeh get_event(event); 2257a40c4f10SYehuda Sadeh } 2258a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 2259a40c4f10SYehuda Sadeh dout("handle_watch_notify cookie %lld ver %lld event %p\n", 2260a40c4f10SYehuda Sadeh cookie, ver, event); 2261a40c4f10SYehuda Sadeh if (event) { 2262a40c4f10SYehuda Sadeh event_work = kmalloc(sizeof(*event_work), GFP_NOIO); 2263a40c4f10SYehuda Sadeh if (!event_work) { 2264a40c4f10SYehuda Sadeh dout("ERROR: could not allocate event_work\n"); 2265a40c4f10SYehuda Sadeh goto done_err; 2266a40c4f10SYehuda Sadeh } 22676b0ae409SMariusz Kozlowski INIT_WORK(&event_work->work, do_event_work); 2268a40c4f10SYehuda Sadeh event_work->event = event; 2269a40c4f10SYehuda Sadeh event_work->ver = ver; 2270a40c4f10SYehuda Sadeh event_work->notify_id = notify_id; 2271a40c4f10SYehuda Sadeh event_work->opcode = opcode; 2272a40c4f10SYehuda Sadeh if (!queue_work(osdc->notify_wq, &event_work->work)) { 2273a40c4f10SYehuda Sadeh dout("WARNING: failed to queue notify event work\n"); 2274a40c4f10SYehuda Sadeh goto done_err; 2275a40c4f10SYehuda Sadeh } 2276a40c4f10SYehuda Sadeh } 2277a40c4f10SYehuda Sadeh 2278a40c4f10SYehuda Sadeh return; 2279a40c4f10SYehuda Sadeh 2280a40c4f10SYehuda Sadeh done_err: 2281a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 2282a40c4f10SYehuda Sadeh return; 2283a40c4f10SYehuda Sadeh 2284a40c4f10SYehuda Sadeh bad: 2285a40c4f10SYehuda Sadeh pr_err("osdc handle_watch_notify corrupt msg\n"); 2286a40c4f10SYehuda Sadeh return; 2287a40c4f10SYehuda Sadeh } 2288a40c4f10SYehuda Sadeh 2289a40c4f10SYehuda Sadeh /* 2290e65550fdSAlex Elder * build new request AND message 2291e65550fdSAlex Elder * 2292e65550fdSAlex Elder */ 2293e65550fdSAlex Elder void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off, 2294e65550fdSAlex Elder struct ceph_snap_context *snapc, u64 snap_id, 2295e65550fdSAlex Elder struct timespec *mtime) 2296e65550fdSAlex Elder { 2297e65550fdSAlex Elder struct ceph_msg *msg = req->r_request; 2298e65550fdSAlex Elder void *p; 2299e65550fdSAlex Elder size_t msg_size; 2300e65550fdSAlex Elder int flags = req->r_flags; 2301e65550fdSAlex Elder u64 data_len; 2302e65550fdSAlex Elder unsigned int i; 2303e65550fdSAlex Elder 2304e65550fdSAlex Elder req->r_snapid = snap_id; 2305e65550fdSAlex Elder req->r_snapc = ceph_get_snap_context(snapc); 2306e65550fdSAlex Elder 2307e65550fdSAlex Elder /* encode request */ 2308e65550fdSAlex Elder msg->hdr.version = cpu_to_le16(4); 2309e65550fdSAlex Elder 2310e65550fdSAlex Elder p = msg->front.iov_base; 2311e65550fdSAlex Elder ceph_encode_32(&p, 1); /* client_inc is always 1 */ 2312e65550fdSAlex Elder req->r_request_osdmap_epoch = p; 2313e65550fdSAlex Elder p += 4; 2314e65550fdSAlex Elder req->r_request_flags = p; 2315e65550fdSAlex Elder p += 4; 2316e65550fdSAlex Elder if (req->r_flags & CEPH_OSD_FLAG_WRITE) 2317e65550fdSAlex Elder ceph_encode_timespec(p, mtime); 2318e65550fdSAlex Elder p += sizeof(struct ceph_timespec); 2319e65550fdSAlex Elder req->r_request_reassert_version = p; 2320e65550fdSAlex Elder p += sizeof(struct ceph_eversion); /* will get filled in */ 2321e65550fdSAlex Elder 2322e65550fdSAlex Elder /* oloc */ 2323e65550fdSAlex Elder ceph_encode_8(&p, 4); 2324e65550fdSAlex Elder ceph_encode_8(&p, 4); 2325e65550fdSAlex Elder ceph_encode_32(&p, 8 + 4 + 4); 2326e65550fdSAlex Elder req->r_request_pool = p; 2327e65550fdSAlex Elder p += 8; 2328e65550fdSAlex Elder ceph_encode_32(&p, -1); /* preferred */ 2329e65550fdSAlex Elder ceph_encode_32(&p, 0); /* key len */ 2330e65550fdSAlex Elder 2331e65550fdSAlex Elder ceph_encode_8(&p, 1); 2332e65550fdSAlex Elder req->r_request_pgid = p; 2333e65550fdSAlex Elder p += 8 + 4; 2334e65550fdSAlex Elder ceph_encode_32(&p, -1); /* preferred */ 2335e65550fdSAlex Elder 2336e65550fdSAlex Elder /* oid */ 23373c972c95SIlya Dryomov ceph_encode_32(&p, req->r_base_oid.name_len); 23383c972c95SIlya Dryomov memcpy(p, req->r_base_oid.name, req->r_base_oid.name_len); 23393c972c95SIlya Dryomov dout("oid '%.*s' len %d\n", req->r_base_oid.name_len, 23403c972c95SIlya Dryomov req->r_base_oid.name, req->r_base_oid.name_len); 23413c972c95SIlya Dryomov p += req->r_base_oid.name_len; 2342e65550fdSAlex Elder 2343e65550fdSAlex Elder /* ops--can imply data */ 2344e65550fdSAlex Elder ceph_encode_16(&p, (u16)req->r_num_ops); 2345e65550fdSAlex Elder data_len = 0; 2346e65550fdSAlex Elder for (i = 0; i < req->r_num_ops; i++) { 2347e65550fdSAlex Elder data_len += osd_req_encode_op(req, p, i); 2348e65550fdSAlex Elder p += sizeof(struct ceph_osd_op); 2349e65550fdSAlex Elder } 2350e65550fdSAlex Elder 2351e65550fdSAlex Elder /* snaps */ 2352e65550fdSAlex Elder ceph_encode_64(&p, req->r_snapid); 2353e65550fdSAlex Elder ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0); 2354e65550fdSAlex Elder ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0); 2355e65550fdSAlex Elder if (req->r_snapc) { 2356e65550fdSAlex Elder for (i = 0; i < snapc->num_snaps; i++) { 2357e65550fdSAlex Elder ceph_encode_64(&p, req->r_snapc->snaps[i]); 2358e65550fdSAlex Elder } 2359e65550fdSAlex Elder } 2360e65550fdSAlex Elder 2361e65550fdSAlex Elder req->r_request_attempts = p; 2362e65550fdSAlex Elder p += 4; 2363e65550fdSAlex Elder 2364e65550fdSAlex Elder /* data */ 2365e65550fdSAlex Elder if (flags & CEPH_OSD_FLAG_WRITE) { 2366e65550fdSAlex Elder u16 data_off; 2367e65550fdSAlex Elder 2368e65550fdSAlex Elder /* 2369e65550fdSAlex Elder * The header "data_off" is a hint to the receiver 2370e65550fdSAlex Elder * allowing it to align received data into its 2371e65550fdSAlex Elder * buffers such that there's no need to re-copy 2372e65550fdSAlex Elder * it before writing it to disk (direct I/O). 2373e65550fdSAlex Elder */ 2374e65550fdSAlex Elder data_off = (u16) (off & 0xffff); 2375e65550fdSAlex Elder req->r_request->hdr.data_off = cpu_to_le16(data_off); 2376e65550fdSAlex Elder } 2377e65550fdSAlex Elder req->r_request->hdr.data_len = cpu_to_le32(data_len); 2378e65550fdSAlex Elder 2379e65550fdSAlex Elder BUG_ON(p > msg->front.iov_base + msg->front.iov_len); 2380e65550fdSAlex Elder msg_size = p - msg->front.iov_base; 2381e65550fdSAlex Elder msg->front.iov_len = msg_size; 2382e65550fdSAlex Elder msg->hdr.front_len = cpu_to_le32(msg_size); 2383e65550fdSAlex Elder 2384e65550fdSAlex Elder dout("build_request msg_size was %d\n", (int)msg_size); 2385e65550fdSAlex Elder } 2386e65550fdSAlex Elder EXPORT_SYMBOL(ceph_osdc_build_request); 2387e65550fdSAlex Elder 2388e65550fdSAlex Elder /* 23893d14c5d2SYehuda Sadeh * Register request, send initial attempt. 23903d14c5d2SYehuda Sadeh */ 23913d14c5d2SYehuda Sadeh int ceph_osdc_start_request(struct ceph_osd_client *osdc, 23923d14c5d2SYehuda Sadeh struct ceph_osd_request *req, 23933d14c5d2SYehuda Sadeh bool nofail) 23943d14c5d2SYehuda Sadeh { 23950bbfdfe8SIlya Dryomov int rc; 23963d14c5d2SYehuda Sadeh 23973d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 23983d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 23990bbfdfe8SIlya Dryomov 24000bbfdfe8SIlya Dryomov rc = __ceph_osdc_start_request(osdc, req, nofail); 24010bbfdfe8SIlya Dryomov 24023d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 24033d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 24040bbfdfe8SIlya Dryomov 24053d14c5d2SYehuda Sadeh return rc; 24063d14c5d2SYehuda Sadeh } 24073d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_start_request); 24083d14c5d2SYehuda Sadeh 24093d14c5d2SYehuda Sadeh /* 24103d14c5d2SYehuda Sadeh * wait for a request to complete 24113d14c5d2SYehuda Sadeh */ 24123d14c5d2SYehuda Sadeh int ceph_osdc_wait_request(struct ceph_osd_client *osdc, 24133d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 24143d14c5d2SYehuda Sadeh { 24153d14c5d2SYehuda Sadeh int rc; 24163d14c5d2SYehuda Sadeh 24173d14c5d2SYehuda Sadeh rc = wait_for_completion_interruptible(&req->r_completion); 24183d14c5d2SYehuda Sadeh if (rc < 0) { 24193d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 24203d14c5d2SYehuda Sadeh __cancel_request(req); 24213d14c5d2SYehuda Sadeh __unregister_request(osdc, req); 24223d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 242325845472SSage Weil complete_request(req); 24243d14c5d2SYehuda Sadeh dout("wait_request tid %llu canceled/timed out\n", req->r_tid); 24253d14c5d2SYehuda Sadeh return rc; 24263d14c5d2SYehuda Sadeh } 24273d14c5d2SYehuda Sadeh 24283d14c5d2SYehuda Sadeh dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result); 24293d14c5d2SYehuda Sadeh return req->r_result; 24303d14c5d2SYehuda Sadeh } 24313d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_wait_request); 24323d14c5d2SYehuda Sadeh 24333d14c5d2SYehuda Sadeh /* 24343d14c5d2SYehuda Sadeh * sync - wait for all in-flight requests to flush. avoid starvation. 24353d14c5d2SYehuda Sadeh */ 24363d14c5d2SYehuda Sadeh void ceph_osdc_sync(struct ceph_osd_client *osdc) 24373d14c5d2SYehuda Sadeh { 24383d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 24393d14c5d2SYehuda Sadeh u64 last_tid, next_tid = 0; 24403d14c5d2SYehuda Sadeh 24413d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 24423d14c5d2SYehuda Sadeh last_tid = osdc->last_tid; 24433d14c5d2SYehuda Sadeh while (1) { 24443d14c5d2SYehuda Sadeh req = __lookup_request_ge(osdc, next_tid); 24453d14c5d2SYehuda Sadeh if (!req) 24463d14c5d2SYehuda Sadeh break; 24473d14c5d2SYehuda Sadeh if (req->r_tid > last_tid) 24483d14c5d2SYehuda Sadeh break; 24493d14c5d2SYehuda Sadeh 24503d14c5d2SYehuda Sadeh next_tid = req->r_tid + 1; 24513d14c5d2SYehuda Sadeh if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0) 24523d14c5d2SYehuda Sadeh continue; 24533d14c5d2SYehuda Sadeh 24543d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 24553d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 24563d14c5d2SYehuda Sadeh dout("sync waiting on tid %llu (last is %llu)\n", 24573d14c5d2SYehuda Sadeh req->r_tid, last_tid); 24583d14c5d2SYehuda Sadeh wait_for_completion(&req->r_safe_completion); 24593d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 24603d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 24613d14c5d2SYehuda Sadeh } 24623d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 24633d14c5d2SYehuda Sadeh dout("sync done (thru tid %llu)\n", last_tid); 24643d14c5d2SYehuda Sadeh } 24653d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_sync); 24663d14c5d2SYehuda Sadeh 24673d14c5d2SYehuda Sadeh /* 2468dd935f44SJosh Durgin * Call all pending notify callbacks - for use after a watch is 2469dd935f44SJosh Durgin * unregistered, to make sure no more callbacks for it will be invoked 2470dd935f44SJosh Durgin */ 2471dd935f44SJosh Durgin extern void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc) 2472dd935f44SJosh Durgin { 2473dd935f44SJosh Durgin flush_workqueue(osdc->notify_wq); 2474dd935f44SJosh Durgin } 2475dd935f44SJosh Durgin EXPORT_SYMBOL(ceph_osdc_flush_notifies); 2476dd935f44SJosh Durgin 2477dd935f44SJosh Durgin 2478dd935f44SJosh Durgin /* 24793d14c5d2SYehuda Sadeh * init, shutdown 24803d14c5d2SYehuda Sadeh */ 24813d14c5d2SYehuda Sadeh int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client) 24823d14c5d2SYehuda Sadeh { 24833d14c5d2SYehuda Sadeh int err; 24843d14c5d2SYehuda Sadeh 24853d14c5d2SYehuda Sadeh dout("init\n"); 24863d14c5d2SYehuda Sadeh osdc->client = client; 24873d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 24883d14c5d2SYehuda Sadeh init_rwsem(&osdc->map_sem); 24893d14c5d2SYehuda Sadeh init_completion(&osdc->map_waiters); 24903d14c5d2SYehuda Sadeh osdc->last_requested_map = 0; 24913d14c5d2SYehuda Sadeh mutex_init(&osdc->request_mutex); 24923d14c5d2SYehuda Sadeh osdc->last_tid = 0; 24933d14c5d2SYehuda Sadeh osdc->osds = RB_ROOT; 24943d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->osd_lru); 24953d14c5d2SYehuda Sadeh osdc->requests = RB_ROOT; 24963d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_lru); 24976f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_unsent); 24986f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_notarget); 2499a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_linger); 25003d14c5d2SYehuda Sadeh osdc->num_requests = 0; 25013d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout); 25023d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout); 2503a40c4f10SYehuda Sadeh spin_lock_init(&osdc->event_lock); 2504a40c4f10SYehuda Sadeh osdc->event_tree = RB_ROOT; 2505a40c4f10SYehuda Sadeh osdc->event_count = 0; 25063d14c5d2SYehuda Sadeh 25073d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 25083d14c5d2SYehuda Sadeh round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ)); 25093d14c5d2SYehuda Sadeh 25103d14c5d2SYehuda Sadeh err = -ENOMEM; 25113d14c5d2SYehuda Sadeh osdc->req_mempool = mempool_create_kmalloc_pool(10, 25123d14c5d2SYehuda Sadeh sizeof(struct ceph_osd_request)); 25133d14c5d2SYehuda Sadeh if (!osdc->req_mempool) 25143d14c5d2SYehuda Sadeh goto out; 25153d14c5d2SYehuda Sadeh 2516d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP, 2517d50b409fSSage Weil OSD_OP_FRONT_LEN, 10, true, 25183d14c5d2SYehuda Sadeh "osd_op"); 25193d14c5d2SYehuda Sadeh if (err < 0) 25203d14c5d2SYehuda Sadeh goto out_mempool; 2521d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY, 25223d14c5d2SYehuda Sadeh OSD_OPREPLY_FRONT_LEN, 10, true, 25233d14c5d2SYehuda Sadeh "osd_op_reply"); 25243d14c5d2SYehuda Sadeh if (err < 0) 25253d14c5d2SYehuda Sadeh goto out_msgpool; 2526a40c4f10SYehuda Sadeh 2527dbcae088SDan Carpenter err = -ENOMEM; 2528a40c4f10SYehuda Sadeh osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify"); 2529dbcae088SDan Carpenter if (!osdc->notify_wq) 2530c172ec5cSIlya Dryomov goto out_msgpool_reply; 2531c172ec5cSIlya Dryomov 25323d14c5d2SYehuda Sadeh return 0; 25333d14c5d2SYehuda Sadeh 2534c172ec5cSIlya Dryomov out_msgpool_reply: 2535c172ec5cSIlya Dryomov ceph_msgpool_destroy(&osdc->msgpool_op_reply); 25363d14c5d2SYehuda Sadeh out_msgpool: 25373d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 25383d14c5d2SYehuda Sadeh out_mempool: 25393d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 25403d14c5d2SYehuda Sadeh out: 25413d14c5d2SYehuda Sadeh return err; 25423d14c5d2SYehuda Sadeh } 25433d14c5d2SYehuda Sadeh 25443d14c5d2SYehuda Sadeh void ceph_osdc_stop(struct ceph_osd_client *osdc) 25453d14c5d2SYehuda Sadeh { 2546a40c4f10SYehuda Sadeh flush_workqueue(osdc->notify_wq); 2547a40c4f10SYehuda Sadeh destroy_workqueue(osdc->notify_wq); 25483d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->timeout_work); 25493d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->osds_timeout_work); 25503d14c5d2SYehuda Sadeh if (osdc->osdmap) { 25513d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 25523d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 25533d14c5d2SYehuda Sadeh } 2554aca420bcSSage Weil remove_all_osds(osdc); 25553d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 25563d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 25573d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op_reply); 25583d14c5d2SYehuda Sadeh } 25593d14c5d2SYehuda Sadeh 25603d14c5d2SYehuda Sadeh /* 25613d14c5d2SYehuda Sadeh * Read some contiguous pages. If we cross a stripe boundary, shorten 25623d14c5d2SYehuda Sadeh * *plen. Return number of bytes read, or error. 25633d14c5d2SYehuda Sadeh */ 25643d14c5d2SYehuda Sadeh int ceph_osdc_readpages(struct ceph_osd_client *osdc, 25653d14c5d2SYehuda Sadeh struct ceph_vino vino, struct ceph_file_layout *layout, 25663d14c5d2SYehuda Sadeh u64 off, u64 *plen, 25673d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 2568b7495fc2SSage Weil struct page **pages, int num_pages, int page_align) 25693d14c5d2SYehuda Sadeh { 25703d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 25713d14c5d2SYehuda Sadeh int rc = 0; 25723d14c5d2SYehuda Sadeh 25733d14c5d2SYehuda Sadeh dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino, 25743d14c5d2SYehuda Sadeh vino.snap, off, *plen); 257579528734SAlex Elder req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 1, 25763d14c5d2SYehuda Sadeh CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ, 2577acead002SAlex Elder NULL, truncate_seq, truncate_size, 2578153e5167SAlex Elder false); 25796816282dSSage Weil if (IS_ERR(req)) 25806816282dSSage Weil return PTR_ERR(req); 25813d14c5d2SYehuda Sadeh 25823d14c5d2SYehuda Sadeh /* it may be a short read due to an object boundary */ 25830fff87ecSAlex Elder 2584406e2c9fSAlex Elder osd_req_op_extent_osd_data_pages(req, 0, 2585a4ce40a9SAlex Elder pages, *plen, page_align, false, false); 25863d14c5d2SYehuda Sadeh 2587e0c59487SAlex Elder dout("readpages final extent is %llu~%llu (%llu bytes align %d)\n", 258843bfe5deSAlex Elder off, *plen, *plen, page_align); 25893d14c5d2SYehuda Sadeh 259079528734SAlex Elder ceph_osdc_build_request(req, off, NULL, vino.snap, NULL); 259102ee07d3SAlex Elder 25923d14c5d2SYehuda Sadeh rc = ceph_osdc_start_request(osdc, req, false); 25933d14c5d2SYehuda Sadeh if (!rc) 25943d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 25953d14c5d2SYehuda Sadeh 25963d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 25973d14c5d2SYehuda Sadeh dout("readpages result %d\n", rc); 25983d14c5d2SYehuda Sadeh return rc; 25993d14c5d2SYehuda Sadeh } 26003d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_readpages); 26013d14c5d2SYehuda Sadeh 26023d14c5d2SYehuda Sadeh /* 26033d14c5d2SYehuda Sadeh * do a synchronous write on N pages 26043d14c5d2SYehuda Sadeh */ 26053d14c5d2SYehuda Sadeh int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino, 26063d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 26073d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 26083d14c5d2SYehuda Sadeh u64 off, u64 len, 26093d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 26103d14c5d2SYehuda Sadeh struct timespec *mtime, 261124808826SAlex Elder struct page **pages, int num_pages) 26123d14c5d2SYehuda Sadeh { 26133d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 26143d14c5d2SYehuda Sadeh int rc = 0; 2615b7495fc2SSage Weil int page_align = off & ~PAGE_MASK; 26163d14c5d2SYehuda Sadeh 2617acead002SAlex Elder BUG_ON(vino.snap != CEPH_NOSNAP); /* snapshots aren't writeable */ 261879528734SAlex Elder req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 1, 26193d14c5d2SYehuda Sadeh CEPH_OSD_OP_WRITE, 262024808826SAlex Elder CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE, 2621acead002SAlex Elder snapc, truncate_seq, truncate_size, 2622153e5167SAlex Elder true); 26236816282dSSage Weil if (IS_ERR(req)) 26246816282dSSage Weil return PTR_ERR(req); 26253d14c5d2SYehuda Sadeh 26263d14c5d2SYehuda Sadeh /* it may be a short write due to an object boundary */ 2627406e2c9fSAlex Elder osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align, 262843bfe5deSAlex Elder false, false); 262943bfe5deSAlex Elder dout("writepages %llu~%llu (%llu bytes)\n", off, len, len); 26303d14c5d2SYehuda Sadeh 263179528734SAlex Elder ceph_osdc_build_request(req, off, snapc, CEPH_NOSNAP, mtime); 263202ee07d3SAlex Elder 263387f979d3SAlex Elder rc = ceph_osdc_start_request(osdc, req, true); 26343d14c5d2SYehuda Sadeh if (!rc) 26353d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 26363d14c5d2SYehuda Sadeh 26373d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 26383d14c5d2SYehuda Sadeh if (rc == 0) 26393d14c5d2SYehuda Sadeh rc = len; 26403d14c5d2SYehuda Sadeh dout("writepages result %d\n", rc); 26413d14c5d2SYehuda Sadeh return rc; 26423d14c5d2SYehuda Sadeh } 26433d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_writepages); 26443d14c5d2SYehuda Sadeh 26455522ae0bSAlex Elder int ceph_osdc_setup(void) 26465522ae0bSAlex Elder { 26475522ae0bSAlex Elder BUG_ON(ceph_osd_request_cache); 26485522ae0bSAlex Elder ceph_osd_request_cache = kmem_cache_create("ceph_osd_request", 26495522ae0bSAlex Elder sizeof (struct ceph_osd_request), 26505522ae0bSAlex Elder __alignof__(struct ceph_osd_request), 26515522ae0bSAlex Elder 0, NULL); 26525522ae0bSAlex Elder 26535522ae0bSAlex Elder return ceph_osd_request_cache ? 0 : -ENOMEM; 26545522ae0bSAlex Elder } 26555522ae0bSAlex Elder EXPORT_SYMBOL(ceph_osdc_setup); 26565522ae0bSAlex Elder 26575522ae0bSAlex Elder void ceph_osdc_cleanup(void) 26585522ae0bSAlex Elder { 26595522ae0bSAlex Elder BUG_ON(!ceph_osd_request_cache); 26605522ae0bSAlex Elder kmem_cache_destroy(ceph_osd_request_cache); 26615522ae0bSAlex Elder ceph_osd_request_cache = NULL; 26625522ae0bSAlex Elder } 26635522ae0bSAlex Elder EXPORT_SYMBOL(ceph_osdc_cleanup); 26645522ae0bSAlex Elder 26653d14c5d2SYehuda Sadeh /* 26663d14c5d2SYehuda Sadeh * handle incoming message 26673d14c5d2SYehuda Sadeh */ 26683d14c5d2SYehuda Sadeh static void dispatch(struct ceph_connection *con, struct ceph_msg *msg) 26693d14c5d2SYehuda Sadeh { 26703d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 26713d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 26723d14c5d2SYehuda Sadeh int type = le16_to_cpu(msg->hdr.type); 26733d14c5d2SYehuda Sadeh 26743d14c5d2SYehuda Sadeh if (!osd) 26753d14c5d2SYehuda Sadeh goto out; 26763d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 26773d14c5d2SYehuda Sadeh 26783d14c5d2SYehuda Sadeh switch (type) { 26793d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 26803d14c5d2SYehuda Sadeh ceph_osdc_handle_map(osdc, msg); 26813d14c5d2SYehuda Sadeh break; 26823d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 26833d14c5d2SYehuda Sadeh handle_reply(osdc, msg, con); 26843d14c5d2SYehuda Sadeh break; 2685a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 2686a40c4f10SYehuda Sadeh handle_watch_notify(osdc, msg); 2687a40c4f10SYehuda Sadeh break; 26883d14c5d2SYehuda Sadeh 26893d14c5d2SYehuda Sadeh default: 26903d14c5d2SYehuda Sadeh pr_err("received unknown message type %d %s\n", type, 26913d14c5d2SYehuda Sadeh ceph_msg_type_name(type)); 26923d14c5d2SYehuda Sadeh } 26933d14c5d2SYehuda Sadeh out: 26943d14c5d2SYehuda Sadeh ceph_msg_put(msg); 26953d14c5d2SYehuda Sadeh } 26963d14c5d2SYehuda Sadeh 26973d14c5d2SYehuda Sadeh /* 26983d14c5d2SYehuda Sadeh * lookup and return message for incoming reply. set up reply message 26993d14c5d2SYehuda Sadeh * pages. 27003d14c5d2SYehuda Sadeh */ 27013d14c5d2SYehuda Sadeh static struct ceph_msg *get_reply(struct ceph_connection *con, 27023d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 27033d14c5d2SYehuda Sadeh int *skip) 27043d14c5d2SYehuda Sadeh { 27053d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 27063d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = osd->o_osdc; 27073d14c5d2SYehuda Sadeh struct ceph_msg *m; 27083d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 27093f0a4ac5SIlya Dryomov int front_len = le32_to_cpu(hdr->front_len); 27103d14c5d2SYehuda Sadeh int data_len = le32_to_cpu(hdr->data_len); 27113d14c5d2SYehuda Sadeh u64 tid; 27123d14c5d2SYehuda Sadeh 27133d14c5d2SYehuda Sadeh tid = le64_to_cpu(hdr->tid); 27143d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 27153d14c5d2SYehuda Sadeh req = __lookup_request(osdc, tid); 27163d14c5d2SYehuda Sadeh if (!req) { 27173d14c5d2SYehuda Sadeh *skip = 1; 27183d14c5d2SYehuda Sadeh m = NULL; 2719756a16a5SSage Weil dout("get_reply unknown tid %llu from osd%d\n", tid, 27203d14c5d2SYehuda Sadeh osd->o_osd); 27213d14c5d2SYehuda Sadeh goto out; 27223d14c5d2SYehuda Sadeh } 27233d14c5d2SYehuda Sadeh 2724ace6d3a9SAlex Elder if (req->r_reply->con) 27258921d114SAlex Elder dout("%s revoking msg %p from old con %p\n", __func__, 2726ace6d3a9SAlex Elder req->r_reply, req->r_reply->con); 27278921d114SAlex Elder ceph_msg_revoke_incoming(req->r_reply); 27283d14c5d2SYehuda Sadeh 2729f2be82b0SIlya Dryomov if (front_len > req->r_reply->front_alloc_len) { 27304974341eSAlex Elder pr_warning("get_reply front %d > preallocated %d (%u#%llu)\n", 2731f2be82b0SIlya Dryomov front_len, req->r_reply->front_alloc_len, 27324974341eSAlex Elder (unsigned int)con->peer_name.type, 27334974341eSAlex Elder le64_to_cpu(con->peer_name.num)); 27343f0a4ac5SIlya Dryomov m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front_len, GFP_NOFS, 27353f0a4ac5SIlya Dryomov false); 27363d14c5d2SYehuda Sadeh if (!m) 27373d14c5d2SYehuda Sadeh goto out; 27383d14c5d2SYehuda Sadeh ceph_msg_put(req->r_reply); 27393d14c5d2SYehuda Sadeh req->r_reply = m; 27403d14c5d2SYehuda Sadeh } 27413d14c5d2SYehuda Sadeh m = ceph_msg_get(req->r_reply); 27423d14c5d2SYehuda Sadeh 27433d14c5d2SYehuda Sadeh if (data_len > 0) { 2744a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 27450fff87ecSAlex Elder 2746a4ce40a9SAlex Elder /* 2747a4ce40a9SAlex Elder * XXX This is assuming there is only one op containing 2748a4ce40a9SAlex Elder * XXX page data. Probably OK for reads, but this 2749a4ce40a9SAlex Elder * XXX ought to be done more generally. 2750a4ce40a9SAlex Elder */ 2751406e2c9fSAlex Elder osd_data = osd_req_op_extent_osd_data(req, 0); 27520fff87ecSAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 27530fff87ecSAlex Elder if (osd_data->pages && 2754e0c59487SAlex Elder unlikely(osd_data->length < data_len)) { 27552ac2b7a6SAlex Elder 2756e0c59487SAlex Elder pr_warning("tid %lld reply has %d bytes " 2757e0c59487SAlex Elder "we had only %llu bytes ready\n", 2758e0c59487SAlex Elder tid, data_len, osd_data->length); 27593d14c5d2SYehuda Sadeh *skip = 1; 27603d14c5d2SYehuda Sadeh ceph_msg_put(m); 27613d14c5d2SYehuda Sadeh m = NULL; 27623d14c5d2SYehuda Sadeh goto out; 27633d14c5d2SYehuda Sadeh } 27643d14c5d2SYehuda Sadeh } 27652ac2b7a6SAlex Elder } 27663d14c5d2SYehuda Sadeh *skip = 0; 27673d14c5d2SYehuda Sadeh dout("get_reply tid %lld %p\n", tid, m); 27683d14c5d2SYehuda Sadeh 27693d14c5d2SYehuda Sadeh out: 27703d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 27713d14c5d2SYehuda Sadeh return m; 27723d14c5d2SYehuda Sadeh 27733d14c5d2SYehuda Sadeh } 27743d14c5d2SYehuda Sadeh 27753d14c5d2SYehuda Sadeh static struct ceph_msg *alloc_msg(struct ceph_connection *con, 27763d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 27773d14c5d2SYehuda Sadeh int *skip) 27783d14c5d2SYehuda Sadeh { 27793d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 27803d14c5d2SYehuda Sadeh int type = le16_to_cpu(hdr->type); 27813d14c5d2SYehuda Sadeh int front = le32_to_cpu(hdr->front_len); 27823d14c5d2SYehuda Sadeh 27831c20f2d2SAlex Elder *skip = 0; 27843d14c5d2SYehuda Sadeh switch (type) { 27853d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 2786a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 2787b61c2763SSage Weil return ceph_msg_new(type, front, GFP_NOFS, false); 27883d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 27893d14c5d2SYehuda Sadeh return get_reply(con, hdr, skip); 27903d14c5d2SYehuda Sadeh default: 27913d14c5d2SYehuda Sadeh pr_info("alloc_msg unexpected msg type %d from osd%d\n", type, 27923d14c5d2SYehuda Sadeh osd->o_osd); 27933d14c5d2SYehuda Sadeh *skip = 1; 27943d14c5d2SYehuda Sadeh return NULL; 27953d14c5d2SYehuda Sadeh } 27963d14c5d2SYehuda Sadeh } 27973d14c5d2SYehuda Sadeh 27983d14c5d2SYehuda Sadeh /* 27993d14c5d2SYehuda Sadeh * Wrappers to refcount containing ceph_osd struct 28003d14c5d2SYehuda Sadeh */ 28013d14c5d2SYehuda Sadeh static struct ceph_connection *get_osd_con(struct ceph_connection *con) 28023d14c5d2SYehuda Sadeh { 28033d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 28043d14c5d2SYehuda Sadeh if (get_osd(osd)) 28053d14c5d2SYehuda Sadeh return con; 28063d14c5d2SYehuda Sadeh return NULL; 28073d14c5d2SYehuda Sadeh } 28083d14c5d2SYehuda Sadeh 28093d14c5d2SYehuda Sadeh static void put_osd_con(struct ceph_connection *con) 28103d14c5d2SYehuda Sadeh { 28113d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 28123d14c5d2SYehuda Sadeh put_osd(osd); 28133d14c5d2SYehuda Sadeh } 28143d14c5d2SYehuda Sadeh 28153d14c5d2SYehuda Sadeh /* 28163d14c5d2SYehuda Sadeh * authentication 28173d14c5d2SYehuda Sadeh */ 2818a3530df3SAlex Elder /* 2819a3530df3SAlex Elder * Note: returned pointer is the address of a structure that's 2820a3530df3SAlex Elder * managed separately. Caller must *not* attempt to free it. 2821a3530df3SAlex Elder */ 2822a3530df3SAlex Elder static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con, 28238f43fb53SAlex Elder int *proto, int force_new) 28243d14c5d2SYehuda Sadeh { 28253d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 28263d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 28273d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 282874f1869fSAlex Elder struct ceph_auth_handshake *auth = &o->o_auth; 28293d14c5d2SYehuda Sadeh 283074f1869fSAlex Elder if (force_new && auth->authorizer) { 283127859f97SSage Weil ceph_auth_destroy_authorizer(ac, auth->authorizer); 283274f1869fSAlex Elder auth->authorizer = NULL; 28333d14c5d2SYehuda Sadeh } 283427859f97SSage Weil if (!auth->authorizer) { 283527859f97SSage Weil int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 2836a3530df3SAlex Elder auth); 28373d14c5d2SYehuda Sadeh if (ret) 2838a3530df3SAlex Elder return ERR_PTR(ret); 283927859f97SSage Weil } else { 284027859f97SSage Weil int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 28410bed9b5cSSage Weil auth); 28420bed9b5cSSage Weil if (ret) 28430bed9b5cSSage Weil return ERR_PTR(ret); 28443d14c5d2SYehuda Sadeh } 28453d14c5d2SYehuda Sadeh *proto = ac->protocol; 284674f1869fSAlex Elder 2847a3530df3SAlex Elder return auth; 28483d14c5d2SYehuda Sadeh } 28493d14c5d2SYehuda Sadeh 28503d14c5d2SYehuda Sadeh 28513d14c5d2SYehuda Sadeh static int verify_authorizer_reply(struct ceph_connection *con, int len) 28523d14c5d2SYehuda Sadeh { 28533d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 28543d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 28553d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 28563d14c5d2SYehuda Sadeh 285727859f97SSage Weil return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len); 28583d14c5d2SYehuda Sadeh } 28593d14c5d2SYehuda Sadeh 28603d14c5d2SYehuda Sadeh static int invalidate_authorizer(struct ceph_connection *con) 28613d14c5d2SYehuda Sadeh { 28623d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 28633d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 28643d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 28653d14c5d2SYehuda Sadeh 286627859f97SSage Weil ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD); 28673d14c5d2SYehuda Sadeh return ceph_monc_validate_auth(&osdc->client->monc); 28683d14c5d2SYehuda Sadeh } 28693d14c5d2SYehuda Sadeh 28703d14c5d2SYehuda Sadeh static const struct ceph_connection_operations osd_con_ops = { 28713d14c5d2SYehuda Sadeh .get = get_osd_con, 28723d14c5d2SYehuda Sadeh .put = put_osd_con, 28733d14c5d2SYehuda Sadeh .dispatch = dispatch, 28743d14c5d2SYehuda Sadeh .get_authorizer = get_authorizer, 28753d14c5d2SYehuda Sadeh .verify_authorizer_reply = verify_authorizer_reply, 28763d14c5d2SYehuda Sadeh .invalidate_authorizer = invalidate_authorizer, 28773d14c5d2SYehuda Sadeh .alloc_msg = alloc_msg, 28783d14c5d2SYehuda Sadeh .fault = osd_reset, 28793d14c5d2SYehuda Sadeh }; 2880