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 } 691a8dd0a37SAlex Elder dst->op = cpu_to_le16(src->op); 6923d14c5d2SYehuda Sadeh dst->payload_len = cpu_to_le32(src->payload_len); 693175face2SAlex Elder 69454d50649SAlex Elder return request_data_len; 6953d14c5d2SYehuda Sadeh } 6963d14c5d2SYehuda Sadeh 6973d14c5d2SYehuda Sadeh /* 6983d14c5d2SYehuda Sadeh * build new request AND message, calculate layout, and adjust file 6993d14c5d2SYehuda Sadeh * extent as needed. 7003d14c5d2SYehuda Sadeh * 7013d14c5d2SYehuda Sadeh * if the file was recently truncated, we include information about its 7023d14c5d2SYehuda Sadeh * old and new size so that the object can be updated appropriately. (we 7033d14c5d2SYehuda Sadeh * avoid synchronously deleting truncated objects because it's slow.) 7043d14c5d2SYehuda Sadeh * 7053d14c5d2SYehuda Sadeh * if @do_sync, include a 'startsync' command so that the osd will flush 7063d14c5d2SYehuda Sadeh * data quickly. 7073d14c5d2SYehuda Sadeh */ 7083d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc, 7093d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 7103d14c5d2SYehuda Sadeh struct ceph_vino vino, 711acead002SAlex Elder u64 off, u64 *plen, int num_ops, 7123d14c5d2SYehuda Sadeh int opcode, int flags, 7133d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 7143d14c5d2SYehuda Sadeh u32 truncate_seq, 7153d14c5d2SYehuda Sadeh u64 truncate_size, 716153e5167SAlex Elder bool use_mempool) 7173d14c5d2SYehuda Sadeh { 7183d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 71975d1c941SAlex Elder u64 objnum = 0; 72075d1c941SAlex Elder u64 objoff = 0; 72175d1c941SAlex Elder u64 objlen = 0; 722d18d1e28SAlex Elder u32 object_size; 723d18d1e28SAlex Elder u64 object_base; 7246816282dSSage Weil int r; 7253d14c5d2SYehuda Sadeh 726ad7a60deSLi Wang BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE && 727ad7a60deSLi Wang opcode != CEPH_OSD_OP_DELETE && opcode != CEPH_OSD_OP_ZERO && 728ad7a60deSLi Wang opcode != CEPH_OSD_OP_TRUNCATE); 7293d14c5d2SYehuda Sadeh 730acead002SAlex Elder req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool, 731ae7ca4a3SAlex Elder GFP_NOFS); 7324ad12621SSage Weil if (!req) 7336816282dSSage Weil return ERR_PTR(-ENOMEM); 73479528734SAlex Elder 735d178a9e7SAlex Elder req->r_flags = flags; 7363d14c5d2SYehuda Sadeh 7373d14c5d2SYehuda Sadeh /* calculate max write size */ 738a19dadfbSAlex Elder r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen); 7393ff5f385SAlex Elder if (r < 0) { 7403ff5f385SAlex Elder ceph_osdc_put_request(req); 7416816282dSSage Weil return ERR_PTR(r); 7423ff5f385SAlex Elder } 743a19dadfbSAlex Elder 744d18d1e28SAlex Elder object_size = le32_to_cpu(layout->fl_object_size); 745d18d1e28SAlex Elder object_base = off - objoff; 746ccca4e37SYan, Zheng if (!(truncate_seq == 1 && truncate_size == -1ULL)) { 747d18d1e28SAlex Elder if (truncate_size <= object_base) { 748d18d1e28SAlex Elder truncate_size = 0; 749d18d1e28SAlex Elder } else { 750d18d1e28SAlex Elder truncate_size -= object_base; 751d18d1e28SAlex Elder if (truncate_size > object_size) 752d18d1e28SAlex Elder truncate_size = object_size; 753d18d1e28SAlex Elder } 754ccca4e37SYan, Zheng } 755d18d1e28SAlex Elder 756c99d2d4aSAlex Elder osd_req_op_extent_init(req, 0, opcode, objoff, objlen, 757b0270324SAlex Elder truncate_size, truncate_seq); 7588c042b0dSAlex Elder 759acead002SAlex Elder /* 760acead002SAlex Elder * A second op in the ops array means the caller wants to 761acead002SAlex Elder * also issue a include a 'startsync' command so that the 762acead002SAlex Elder * osd will flush data quickly. 763acead002SAlex Elder */ 764acead002SAlex Elder if (num_ops > 1) 765c99d2d4aSAlex Elder osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC); 766d18d1e28SAlex Elder 7673c972c95SIlya Dryomov req->r_base_oloc.pool = ceph_file_layout_pg_pool(*layout); 7683d14c5d2SYehuda Sadeh 7693c972c95SIlya Dryomov snprintf(req->r_base_oid.name, sizeof(req->r_base_oid.name), 7704295f221SIlya Dryomov "%llx.%08llx", vino.ino, objnum); 7713c972c95SIlya Dryomov req->r_base_oid.name_len = strlen(req->r_base_oid.name); 772dbe0fc41SAlex Elder 7733d14c5d2SYehuda Sadeh return req; 7743d14c5d2SYehuda Sadeh } 7753d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_new_request); 7763d14c5d2SYehuda Sadeh 7773d14c5d2SYehuda Sadeh /* 7783d14c5d2SYehuda Sadeh * We keep osd requests in an rbtree, sorted by ->r_tid. 7793d14c5d2SYehuda Sadeh */ 7803d14c5d2SYehuda Sadeh static void __insert_request(struct ceph_osd_client *osdc, 7813d14c5d2SYehuda Sadeh struct ceph_osd_request *new) 7823d14c5d2SYehuda Sadeh { 7833d14c5d2SYehuda Sadeh struct rb_node **p = &osdc->requests.rb_node; 7843d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 7853d14c5d2SYehuda Sadeh struct ceph_osd_request *req = NULL; 7863d14c5d2SYehuda Sadeh 7873d14c5d2SYehuda Sadeh while (*p) { 7883d14c5d2SYehuda Sadeh parent = *p; 7893d14c5d2SYehuda Sadeh req = rb_entry(parent, struct ceph_osd_request, r_node); 7903d14c5d2SYehuda Sadeh if (new->r_tid < req->r_tid) 7913d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 7923d14c5d2SYehuda Sadeh else if (new->r_tid > req->r_tid) 7933d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 7943d14c5d2SYehuda Sadeh else 7953d14c5d2SYehuda Sadeh BUG(); 7963d14c5d2SYehuda Sadeh } 7973d14c5d2SYehuda Sadeh 7983d14c5d2SYehuda Sadeh rb_link_node(&new->r_node, parent, p); 7993d14c5d2SYehuda Sadeh rb_insert_color(&new->r_node, &osdc->requests); 8003d14c5d2SYehuda Sadeh } 8013d14c5d2SYehuda Sadeh 8023d14c5d2SYehuda Sadeh static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc, 8033d14c5d2SYehuda Sadeh u64 tid) 8043d14c5d2SYehuda Sadeh { 8053d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 8063d14c5d2SYehuda Sadeh struct rb_node *n = osdc->requests.rb_node; 8073d14c5d2SYehuda Sadeh 8083d14c5d2SYehuda Sadeh while (n) { 8093d14c5d2SYehuda Sadeh req = rb_entry(n, struct ceph_osd_request, r_node); 8103d14c5d2SYehuda Sadeh if (tid < req->r_tid) 8113d14c5d2SYehuda Sadeh n = n->rb_left; 8123d14c5d2SYehuda Sadeh else if (tid > req->r_tid) 8133d14c5d2SYehuda Sadeh n = n->rb_right; 8143d14c5d2SYehuda Sadeh else 8153d14c5d2SYehuda Sadeh return req; 8163d14c5d2SYehuda Sadeh } 8173d14c5d2SYehuda Sadeh return NULL; 8183d14c5d2SYehuda Sadeh } 8193d14c5d2SYehuda Sadeh 8203d14c5d2SYehuda Sadeh static struct ceph_osd_request * 8213d14c5d2SYehuda Sadeh __lookup_request_ge(struct ceph_osd_client *osdc, 8223d14c5d2SYehuda Sadeh u64 tid) 8233d14c5d2SYehuda Sadeh { 8243d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 8253d14c5d2SYehuda Sadeh struct rb_node *n = osdc->requests.rb_node; 8263d14c5d2SYehuda Sadeh 8273d14c5d2SYehuda Sadeh while (n) { 8283d14c5d2SYehuda Sadeh req = rb_entry(n, struct ceph_osd_request, r_node); 8293d14c5d2SYehuda Sadeh if (tid < req->r_tid) { 8303d14c5d2SYehuda Sadeh if (!n->rb_left) 8313d14c5d2SYehuda Sadeh return req; 8323d14c5d2SYehuda Sadeh n = n->rb_left; 8333d14c5d2SYehuda Sadeh } else if (tid > req->r_tid) { 8343d14c5d2SYehuda Sadeh n = n->rb_right; 8353d14c5d2SYehuda Sadeh } else { 8363d14c5d2SYehuda Sadeh return req; 8373d14c5d2SYehuda Sadeh } 8383d14c5d2SYehuda Sadeh } 8393d14c5d2SYehuda Sadeh return NULL; 8403d14c5d2SYehuda Sadeh } 8413d14c5d2SYehuda Sadeh 8426f6c7006SSage Weil /* 8436f6c7006SSage Weil * Resubmit requests pending on the given osd. 8446f6c7006SSage Weil */ 8456f6c7006SSage Weil static void __kick_osd_requests(struct ceph_osd_client *osdc, 8466f6c7006SSage Weil struct ceph_osd *osd) 8476f6c7006SSage Weil { 848a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 849e02493c0SAlex Elder LIST_HEAD(resend); 8506f6c7006SSage Weil int err; 8516f6c7006SSage Weil 8526f6c7006SSage Weil dout("__kick_osd_requests osd%d\n", osd->o_osd); 8536f6c7006SSage Weil err = __reset_osd(osdc, osd); 854685a7555SAlex Elder if (err) 8556f6c7006SSage Weil return; 856e02493c0SAlex Elder /* 857e02493c0SAlex Elder * Build up a list of requests to resend by traversing the 858e02493c0SAlex Elder * osd's list of requests. Requests for a given object are 859e02493c0SAlex Elder * sent in tid order, and that is also the order they're 860e02493c0SAlex Elder * kept on this list. Therefore all requests that are in 861e02493c0SAlex Elder * flight will be found first, followed by all requests that 862e02493c0SAlex Elder * have not yet been sent. And to resend requests while 863e02493c0SAlex Elder * preserving this order we will want to put any sent 864e02493c0SAlex Elder * requests back on the front of the osd client's unsent 865e02493c0SAlex Elder * list. 866e02493c0SAlex Elder * 867e02493c0SAlex Elder * So we build a separate ordered list of already-sent 868e02493c0SAlex Elder * requests for the affected osd and splice it onto the 869e02493c0SAlex Elder * front of the osd client's unsent list. Once we've seen a 870e02493c0SAlex Elder * request that has not yet been sent we're done. Those 871e02493c0SAlex Elder * requests are already sitting right where they belong. 872e02493c0SAlex Elder */ 8736f6c7006SSage Weil list_for_each_entry(req, &osd->o_requests, r_osd_item) { 874e02493c0SAlex Elder if (!req->r_sent) 875e02493c0SAlex Elder break; 876e02493c0SAlex Elder list_move_tail(&req->r_req_lru_item, &resend); 877e02493c0SAlex Elder dout("requeueing %p tid %llu osd%d\n", req, req->r_tid, 8786f6c7006SSage Weil osd->o_osd); 879a40c4f10SYehuda Sadeh if (!req->r_linger) 8806f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 8816f6c7006SSage Weil } 882e02493c0SAlex Elder list_splice(&resend, &osdc->req_unsent); 883a40c4f10SYehuda Sadeh 884e02493c0SAlex Elder /* 885e02493c0SAlex Elder * Linger requests are re-registered before sending, which 886e02493c0SAlex Elder * sets up a new tid for each. We add them to the unsent 887e02493c0SAlex Elder * list at the end to keep things in tid order. 888e02493c0SAlex Elder */ 889a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osd->o_linger_requests, 890a40c4f10SYehuda Sadeh r_linger_osd) { 89177f38e0eSSage Weil /* 89277f38e0eSSage Weil * reregister request prior to unregistering linger so 89377f38e0eSSage Weil * that r_osd is preserved. 89477f38e0eSSage Weil */ 89577f38e0eSSage Weil BUG_ON(!list_empty(&req->r_req_lru_item)); 896a40c4f10SYehuda Sadeh __register_request(osdc, req); 897e02493c0SAlex Elder list_add_tail(&req->r_req_lru_item, &osdc->req_unsent); 898ad885927SAlex Elder list_add_tail(&req->r_osd_item, &req->r_osd->o_requests); 89977f38e0eSSage Weil __unregister_linger_request(osdc, req); 900a40c4f10SYehuda Sadeh dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid, 901a40c4f10SYehuda Sadeh osd->o_osd); 902a40c4f10SYehuda Sadeh } 9036f6c7006SSage Weil } 9046f6c7006SSage Weil 9053d14c5d2SYehuda Sadeh /* 9063d14c5d2SYehuda Sadeh * If the osd connection drops, we need to resubmit all requests. 9073d14c5d2SYehuda Sadeh */ 9083d14c5d2SYehuda Sadeh static void osd_reset(struct ceph_connection *con) 9093d14c5d2SYehuda Sadeh { 9103d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 9113d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 9123d14c5d2SYehuda Sadeh 9133d14c5d2SYehuda Sadeh if (!osd) 9143d14c5d2SYehuda Sadeh return; 9153d14c5d2SYehuda Sadeh dout("osd_reset osd%d\n", osd->o_osd); 9163d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 9173d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 91883aff95eSSage Weil mutex_lock(&osdc->request_mutex); 91983aff95eSSage Weil __kick_osd_requests(osdc, osd); 920f9d25199SAlex Elder __send_queued(osdc); 92183aff95eSSage Weil mutex_unlock(&osdc->request_mutex); 9223d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 9233d14c5d2SYehuda Sadeh } 9243d14c5d2SYehuda Sadeh 9253d14c5d2SYehuda Sadeh /* 9263d14c5d2SYehuda Sadeh * Track open sessions with osds. 9273d14c5d2SYehuda Sadeh */ 928e10006f8SAlex Elder static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum) 9293d14c5d2SYehuda Sadeh { 9303d14c5d2SYehuda Sadeh struct ceph_osd *osd; 9313d14c5d2SYehuda Sadeh 9323d14c5d2SYehuda Sadeh osd = kzalloc(sizeof(*osd), GFP_NOFS); 9333d14c5d2SYehuda Sadeh if (!osd) 9343d14c5d2SYehuda Sadeh return NULL; 9353d14c5d2SYehuda Sadeh 9363d14c5d2SYehuda Sadeh atomic_set(&osd->o_ref, 1); 9373d14c5d2SYehuda Sadeh osd->o_osdc = osdc; 938e10006f8SAlex Elder osd->o_osd = onum; 939f407731dSAlex Elder RB_CLEAR_NODE(&osd->o_node); 9403d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_requests); 941a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osd->o_linger_requests); 9423d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_osd_lru); 9433d14c5d2SYehuda Sadeh osd->o_incarnation = 1; 9443d14c5d2SYehuda Sadeh 945b7a9e5ddSSage Weil ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr); 9463d14c5d2SYehuda Sadeh 9473d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_keepalive_item); 9483d14c5d2SYehuda Sadeh return osd; 9493d14c5d2SYehuda Sadeh } 9503d14c5d2SYehuda Sadeh 9513d14c5d2SYehuda Sadeh static struct ceph_osd *get_osd(struct ceph_osd *osd) 9523d14c5d2SYehuda Sadeh { 9533d14c5d2SYehuda Sadeh if (atomic_inc_not_zero(&osd->o_ref)) { 9543d14c5d2SYehuda Sadeh dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1, 9553d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref)); 9563d14c5d2SYehuda Sadeh return osd; 9573d14c5d2SYehuda Sadeh } else { 9583d14c5d2SYehuda Sadeh dout("get_osd %p FAIL\n", osd); 9593d14c5d2SYehuda Sadeh return NULL; 9603d14c5d2SYehuda Sadeh } 9613d14c5d2SYehuda Sadeh } 9623d14c5d2SYehuda Sadeh 9633d14c5d2SYehuda Sadeh static void put_osd(struct ceph_osd *osd) 9643d14c5d2SYehuda Sadeh { 9653d14c5d2SYehuda Sadeh dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref), 9663d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref) - 1); 967a255651dSAlex Elder if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) { 9683d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth; 9693d14c5d2SYehuda Sadeh 97027859f97SSage Weil ceph_auth_destroy_authorizer(ac, osd->o_auth.authorizer); 9713d14c5d2SYehuda Sadeh kfree(osd); 9723d14c5d2SYehuda Sadeh } 9733d14c5d2SYehuda Sadeh } 9743d14c5d2SYehuda Sadeh 9753d14c5d2SYehuda Sadeh /* 9763d14c5d2SYehuda Sadeh * remove an osd from our map 9773d14c5d2SYehuda Sadeh */ 9783d14c5d2SYehuda Sadeh static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 9793d14c5d2SYehuda Sadeh { 9803d14c5d2SYehuda Sadeh dout("__remove_osd %p\n", osd); 9813d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_requests)); 9823d14c5d2SYehuda Sadeh rb_erase(&osd->o_node, &osdc->osds); 9833d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 9843d14c5d2SYehuda Sadeh ceph_con_close(&osd->o_con); 9853d14c5d2SYehuda Sadeh put_osd(osd); 9863d14c5d2SYehuda Sadeh } 9873d14c5d2SYehuda Sadeh 988aca420bcSSage Weil static void remove_all_osds(struct ceph_osd_client *osdc) 989aca420bcSSage Weil { 990048a9d2dSJiaju Zhang dout("%s %p\n", __func__, osdc); 991aca420bcSSage Weil mutex_lock(&osdc->request_mutex); 992aca420bcSSage Weil while (!RB_EMPTY_ROOT(&osdc->osds)) { 993aca420bcSSage Weil struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds), 994aca420bcSSage Weil struct ceph_osd, o_node); 995aca420bcSSage Weil __remove_osd(osdc, osd); 996aca420bcSSage Weil } 997aca420bcSSage Weil mutex_unlock(&osdc->request_mutex); 998aca420bcSSage Weil } 999aca420bcSSage Weil 10003d14c5d2SYehuda Sadeh static void __move_osd_to_lru(struct ceph_osd_client *osdc, 10013d14c5d2SYehuda Sadeh struct ceph_osd *osd) 10023d14c5d2SYehuda Sadeh { 10033d14c5d2SYehuda Sadeh dout("__move_osd_to_lru %p\n", osd); 10043d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_osd_lru)); 10053d14c5d2SYehuda Sadeh list_add_tail(&osd->o_osd_lru, &osdc->osd_lru); 10063d14c5d2SYehuda Sadeh osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ; 10073d14c5d2SYehuda Sadeh } 10083d14c5d2SYehuda Sadeh 10093d14c5d2SYehuda Sadeh static void __remove_osd_from_lru(struct ceph_osd *osd) 10103d14c5d2SYehuda Sadeh { 10113d14c5d2SYehuda Sadeh dout("__remove_osd_from_lru %p\n", osd); 10123d14c5d2SYehuda Sadeh if (!list_empty(&osd->o_osd_lru)) 10133d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 10143d14c5d2SYehuda Sadeh } 10153d14c5d2SYehuda Sadeh 1016aca420bcSSage Weil static void remove_old_osds(struct ceph_osd_client *osdc) 10173d14c5d2SYehuda Sadeh { 10183d14c5d2SYehuda Sadeh struct ceph_osd *osd, *nosd; 10193d14c5d2SYehuda Sadeh 10203d14c5d2SYehuda Sadeh dout("__remove_old_osds %p\n", osdc); 10213d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 10223d14c5d2SYehuda Sadeh list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) { 1023aca420bcSSage Weil if (time_before(jiffies, osd->lru_ttl)) 10243d14c5d2SYehuda Sadeh break; 10253d14c5d2SYehuda Sadeh __remove_osd(osdc, osd); 10263d14c5d2SYehuda Sadeh } 10273d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 10283d14c5d2SYehuda Sadeh } 10293d14c5d2SYehuda Sadeh 10303d14c5d2SYehuda Sadeh /* 10313d14c5d2SYehuda Sadeh * reset osd connect 10323d14c5d2SYehuda Sadeh */ 10333d14c5d2SYehuda Sadeh static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 10343d14c5d2SYehuda Sadeh { 1035c3acb181SAlex Elder struct ceph_entity_addr *peer_addr; 10363d14c5d2SYehuda Sadeh 10373d14c5d2SYehuda Sadeh dout("__reset_osd %p osd%d\n", osd, osd->o_osd); 1038a40c4f10SYehuda Sadeh if (list_empty(&osd->o_requests) && 1039a40c4f10SYehuda Sadeh list_empty(&osd->o_linger_requests)) { 10403d14c5d2SYehuda Sadeh __remove_osd(osdc, osd); 1041c3acb181SAlex Elder 1042c3acb181SAlex Elder return -ENODEV; 1043c3acb181SAlex Elder } 1044c3acb181SAlex Elder 1045c3acb181SAlex Elder peer_addr = &osdc->osdmap->osd_addr[osd->o_osd]; 1046c3acb181SAlex Elder if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) && 10473d14c5d2SYehuda Sadeh !ceph_con_opened(&osd->o_con)) { 1048c3acb181SAlex Elder struct ceph_osd_request *req; 1049c3acb181SAlex Elder 10503d14c5d2SYehuda Sadeh dout("osd addr hasn't changed and connection never opened, " 10510b4af2e8SIlya Dryomov "letting msgr retry\n"); 10523d14c5d2SYehuda Sadeh /* touch each r_stamp for handle_timeout()'s benfit */ 10533d14c5d2SYehuda Sadeh list_for_each_entry(req, &osd->o_requests, r_osd_item) 10543d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 1055c3acb181SAlex Elder 1056c3acb181SAlex Elder return -EAGAIN; 10573d14c5d2SYehuda Sadeh } 1058c3acb181SAlex Elder 1059c3acb181SAlex Elder ceph_con_close(&osd->o_con); 1060c3acb181SAlex Elder ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr); 1061c3acb181SAlex Elder osd->o_incarnation++; 1062c3acb181SAlex Elder 1063c3acb181SAlex Elder return 0; 10643d14c5d2SYehuda Sadeh } 10653d14c5d2SYehuda Sadeh 10663d14c5d2SYehuda Sadeh static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new) 10673d14c5d2SYehuda Sadeh { 10683d14c5d2SYehuda Sadeh struct rb_node **p = &osdc->osds.rb_node; 10693d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 10703d14c5d2SYehuda Sadeh struct ceph_osd *osd = NULL; 10713d14c5d2SYehuda Sadeh 1072aca420bcSSage Weil dout("__insert_osd %p osd%d\n", new, new->o_osd); 10733d14c5d2SYehuda Sadeh while (*p) { 10743d14c5d2SYehuda Sadeh parent = *p; 10753d14c5d2SYehuda Sadeh osd = rb_entry(parent, struct ceph_osd, o_node); 10763d14c5d2SYehuda Sadeh if (new->o_osd < osd->o_osd) 10773d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 10783d14c5d2SYehuda Sadeh else if (new->o_osd > osd->o_osd) 10793d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 10803d14c5d2SYehuda Sadeh else 10813d14c5d2SYehuda Sadeh BUG(); 10823d14c5d2SYehuda Sadeh } 10833d14c5d2SYehuda Sadeh 10843d14c5d2SYehuda Sadeh rb_link_node(&new->o_node, parent, p); 10853d14c5d2SYehuda Sadeh rb_insert_color(&new->o_node, &osdc->osds); 10863d14c5d2SYehuda Sadeh } 10873d14c5d2SYehuda Sadeh 10883d14c5d2SYehuda Sadeh static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o) 10893d14c5d2SYehuda Sadeh { 10903d14c5d2SYehuda Sadeh struct ceph_osd *osd; 10913d14c5d2SYehuda Sadeh struct rb_node *n = osdc->osds.rb_node; 10923d14c5d2SYehuda Sadeh 10933d14c5d2SYehuda Sadeh while (n) { 10943d14c5d2SYehuda Sadeh osd = rb_entry(n, struct ceph_osd, o_node); 10953d14c5d2SYehuda Sadeh if (o < osd->o_osd) 10963d14c5d2SYehuda Sadeh n = n->rb_left; 10973d14c5d2SYehuda Sadeh else if (o > osd->o_osd) 10983d14c5d2SYehuda Sadeh n = n->rb_right; 10993d14c5d2SYehuda Sadeh else 11003d14c5d2SYehuda Sadeh return osd; 11013d14c5d2SYehuda Sadeh } 11023d14c5d2SYehuda Sadeh return NULL; 11033d14c5d2SYehuda Sadeh } 11043d14c5d2SYehuda Sadeh 11053d14c5d2SYehuda Sadeh static void __schedule_osd_timeout(struct ceph_osd_client *osdc) 11063d14c5d2SYehuda Sadeh { 11073d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->timeout_work, 11083d14c5d2SYehuda Sadeh osdc->client->options->osd_keepalive_timeout * HZ); 11093d14c5d2SYehuda Sadeh } 11103d14c5d2SYehuda Sadeh 11113d14c5d2SYehuda Sadeh static void __cancel_osd_timeout(struct ceph_osd_client *osdc) 11123d14c5d2SYehuda Sadeh { 11133d14c5d2SYehuda Sadeh cancel_delayed_work(&osdc->timeout_work); 11143d14c5d2SYehuda Sadeh } 11153d14c5d2SYehuda Sadeh 11163d14c5d2SYehuda Sadeh /* 11173d14c5d2SYehuda Sadeh * Register request, assign tid. If this is the first request, set up 11183d14c5d2SYehuda Sadeh * the timeout event. 11193d14c5d2SYehuda Sadeh */ 1120a40c4f10SYehuda Sadeh static void __register_request(struct ceph_osd_client *osdc, 11213d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 11223d14c5d2SYehuda Sadeh { 11233d14c5d2SYehuda Sadeh req->r_tid = ++osdc->last_tid; 11243d14c5d2SYehuda Sadeh req->r_request->hdr.tid = cpu_to_le64(req->r_tid); 112577f38e0eSSage Weil dout("__register_request %p tid %lld\n", req, req->r_tid); 11263d14c5d2SYehuda Sadeh __insert_request(osdc, req); 11273d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 11283d14c5d2SYehuda Sadeh osdc->num_requests++; 11293d14c5d2SYehuda Sadeh if (osdc->num_requests == 1) { 11303d14c5d2SYehuda Sadeh dout(" first request, scheduling timeout\n"); 11313d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 11323d14c5d2SYehuda Sadeh } 1133a40c4f10SYehuda Sadeh } 1134a40c4f10SYehuda Sadeh 11353d14c5d2SYehuda Sadeh /* 11363d14c5d2SYehuda Sadeh * called under osdc->request_mutex 11373d14c5d2SYehuda Sadeh */ 11383d14c5d2SYehuda Sadeh static void __unregister_request(struct ceph_osd_client *osdc, 11393d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 11403d14c5d2SYehuda Sadeh { 114135f9f8a0SSage Weil if (RB_EMPTY_NODE(&req->r_node)) { 114235f9f8a0SSage Weil dout("__unregister_request %p tid %lld not registered\n", 114335f9f8a0SSage Weil req, req->r_tid); 114435f9f8a0SSage Weil return; 114535f9f8a0SSage Weil } 114635f9f8a0SSage Weil 11473d14c5d2SYehuda Sadeh dout("__unregister_request %p tid %lld\n", req, req->r_tid); 11483d14c5d2SYehuda Sadeh rb_erase(&req->r_node, &osdc->requests); 11493d14c5d2SYehuda Sadeh osdc->num_requests--; 11503d14c5d2SYehuda Sadeh 11513d14c5d2SYehuda Sadeh if (req->r_osd) { 11523d14c5d2SYehuda Sadeh /* make sure the original request isn't in flight. */ 11536740a845SAlex Elder ceph_msg_revoke(req->r_request); 11543d14c5d2SYehuda Sadeh 11553d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 1156a40c4f10SYehuda Sadeh if (list_empty(&req->r_osd->o_requests) && 1157a40c4f10SYehuda Sadeh list_empty(&req->r_osd->o_linger_requests)) { 1158a40c4f10SYehuda Sadeh dout("moving osd to %p lru\n", req->r_osd); 11593d14c5d2SYehuda Sadeh __move_osd_to_lru(osdc, req->r_osd); 1160a40c4f10SYehuda Sadeh } 1161fbdb9190SSage Weil if (list_empty(&req->r_linger_item)) 11623d14c5d2SYehuda Sadeh req->r_osd = NULL; 11633d14c5d2SYehuda Sadeh } 11643d14c5d2SYehuda Sadeh 11657d5f2481SAlex Elder list_del_init(&req->r_req_lru_item); 11663d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 11673d14c5d2SYehuda Sadeh 11683d14c5d2SYehuda Sadeh if (osdc->num_requests == 0) { 11693d14c5d2SYehuda Sadeh dout(" no requests, canceling timeout\n"); 11703d14c5d2SYehuda Sadeh __cancel_osd_timeout(osdc); 11713d14c5d2SYehuda Sadeh } 11723d14c5d2SYehuda Sadeh } 11733d14c5d2SYehuda Sadeh 11743d14c5d2SYehuda Sadeh /* 11753d14c5d2SYehuda Sadeh * Cancel a previously queued request message 11763d14c5d2SYehuda Sadeh */ 11773d14c5d2SYehuda Sadeh static void __cancel_request(struct ceph_osd_request *req) 11783d14c5d2SYehuda Sadeh { 11793d14c5d2SYehuda Sadeh if (req->r_sent && req->r_osd) { 11806740a845SAlex Elder ceph_msg_revoke(req->r_request); 11813d14c5d2SYehuda Sadeh req->r_sent = 0; 11823d14c5d2SYehuda Sadeh } 11833d14c5d2SYehuda Sadeh } 11843d14c5d2SYehuda Sadeh 1185a40c4f10SYehuda Sadeh static void __register_linger_request(struct ceph_osd_client *osdc, 1186a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1187a40c4f10SYehuda Sadeh { 1188a40c4f10SYehuda Sadeh dout("__register_linger_request %p\n", req); 118996e4dac6SAlex Elder ceph_osdc_get_request(req); 1190a40c4f10SYehuda Sadeh list_add_tail(&req->r_linger_item, &osdc->req_linger); 11916194ea89SSage Weil if (req->r_osd) 11926194ea89SSage Weil list_add_tail(&req->r_linger_osd, 11936194ea89SSage Weil &req->r_osd->o_linger_requests); 1194a40c4f10SYehuda Sadeh } 1195a40c4f10SYehuda Sadeh 1196a40c4f10SYehuda Sadeh static void __unregister_linger_request(struct ceph_osd_client *osdc, 1197a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1198a40c4f10SYehuda Sadeh { 1199a40c4f10SYehuda Sadeh dout("__unregister_linger_request %p\n", req); 1200a40c4f10SYehuda Sadeh list_del_init(&req->r_linger_item); 120161c74035SAlex Elder if (req->r_osd) { 1202a40c4f10SYehuda Sadeh list_del_init(&req->r_linger_osd); 1203a40c4f10SYehuda Sadeh 1204a40c4f10SYehuda Sadeh if (list_empty(&req->r_osd->o_requests) && 1205a40c4f10SYehuda Sadeh list_empty(&req->r_osd->o_linger_requests)) { 1206a40c4f10SYehuda Sadeh dout("moving osd to %p lru\n", req->r_osd); 1207a40c4f10SYehuda Sadeh __move_osd_to_lru(osdc, req->r_osd); 1208a40c4f10SYehuda Sadeh } 1209fbdb9190SSage Weil if (list_empty(&req->r_osd_item)) 1210a40c4f10SYehuda Sadeh req->r_osd = NULL; 1211a40c4f10SYehuda Sadeh } 121296e4dac6SAlex Elder ceph_osdc_put_request(req); 1213a40c4f10SYehuda Sadeh } 1214a40c4f10SYehuda Sadeh 1215a40c4f10SYehuda Sadeh void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc, 1216a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1217a40c4f10SYehuda Sadeh { 1218a40c4f10SYehuda Sadeh mutex_lock(&osdc->request_mutex); 1219a40c4f10SYehuda Sadeh if (req->r_linger) { 1220c10ebbf5SAlex Elder req->r_linger = 0; 122196e4dac6SAlex Elder __unregister_linger_request(osdc, req); 1222a40c4f10SYehuda Sadeh } 1223a40c4f10SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 1224a40c4f10SYehuda Sadeh } 1225a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_unregister_linger_request); 1226a40c4f10SYehuda Sadeh 1227a40c4f10SYehuda Sadeh void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc, 1228a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1229a40c4f10SYehuda Sadeh { 1230a40c4f10SYehuda Sadeh if (!req->r_linger) { 1231a40c4f10SYehuda Sadeh dout("set_request_linger %p\n", req); 1232a40c4f10SYehuda Sadeh req->r_linger = 1; 1233a40c4f10SYehuda Sadeh } 1234a40c4f10SYehuda Sadeh } 1235a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_set_request_linger); 1236a40c4f10SYehuda Sadeh 12373d14c5d2SYehuda Sadeh /* 1238d29adb34SJosh Durgin * Returns whether a request should be blocked from being sent 1239d29adb34SJosh Durgin * based on the current osdmap and osd_client settings. 1240d29adb34SJosh Durgin * 1241d29adb34SJosh Durgin * Caller should hold map_sem for read. 1242d29adb34SJosh Durgin */ 1243d29adb34SJosh Durgin static bool __req_should_be_paused(struct ceph_osd_client *osdc, 1244d29adb34SJosh Durgin struct ceph_osd_request *req) 1245d29adb34SJosh Durgin { 1246d29adb34SJosh Durgin bool pauserd = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD); 1247d29adb34SJosh Durgin bool pausewr = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR) || 1248d29adb34SJosh Durgin ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL); 1249d29adb34SJosh Durgin return (req->r_flags & CEPH_OSD_FLAG_READ && pauserd) || 1250d29adb34SJosh Durgin (req->r_flags & CEPH_OSD_FLAG_WRITE && pausewr); 1251d29adb34SJosh Durgin } 1252d29adb34SJosh Durgin 1253d29adb34SJosh Durgin /* 125417a13e40SIlya Dryomov * Calculate mapping of a request to a PG. Takes tiering into account. 125517a13e40SIlya Dryomov */ 125617a13e40SIlya Dryomov static int __calc_request_pg(struct ceph_osdmap *osdmap, 125717a13e40SIlya Dryomov struct ceph_osd_request *req, 125817a13e40SIlya Dryomov struct ceph_pg *pg_out) 125917a13e40SIlya Dryomov { 1260205ee118SIlya Dryomov bool need_check_tiering; 1261205ee118SIlya Dryomov 1262205ee118SIlya Dryomov need_check_tiering = false; 1263205ee118SIlya Dryomov if (req->r_target_oloc.pool == -1) { 1264205ee118SIlya Dryomov req->r_target_oloc = req->r_base_oloc; /* struct */ 1265205ee118SIlya Dryomov need_check_tiering = true; 1266205ee118SIlya Dryomov } 1267205ee118SIlya Dryomov if (req->r_target_oid.name_len == 0) { 1268205ee118SIlya Dryomov ceph_oid_copy(&req->r_target_oid, &req->r_base_oid); 1269205ee118SIlya Dryomov need_check_tiering = true; 1270205ee118SIlya Dryomov } 1271205ee118SIlya Dryomov 1272205ee118SIlya Dryomov if (need_check_tiering && 1273205ee118SIlya Dryomov (req->r_flags & CEPH_OSD_FLAG_IGNORE_OVERLAY) == 0) { 127417a13e40SIlya Dryomov struct ceph_pg_pool_info *pi; 127517a13e40SIlya Dryomov 1276205ee118SIlya Dryomov pi = ceph_pg_pool_by_id(osdmap, req->r_target_oloc.pool); 127717a13e40SIlya Dryomov if (pi) { 127817a13e40SIlya Dryomov if ((req->r_flags & CEPH_OSD_FLAG_READ) && 127917a13e40SIlya Dryomov pi->read_tier >= 0) 1280205ee118SIlya Dryomov req->r_target_oloc.pool = pi->read_tier; 128117a13e40SIlya Dryomov if ((req->r_flags & CEPH_OSD_FLAG_WRITE) && 128217a13e40SIlya Dryomov pi->write_tier >= 0) 1283205ee118SIlya Dryomov req->r_target_oloc.pool = pi->write_tier; 128417a13e40SIlya Dryomov } 128517a13e40SIlya Dryomov /* !pi is caught in ceph_oloc_oid_to_pg() */ 128617a13e40SIlya Dryomov } 128717a13e40SIlya Dryomov 1288205ee118SIlya Dryomov return ceph_oloc_oid_to_pg(osdmap, &req->r_target_oloc, 1289205ee118SIlya Dryomov &req->r_target_oid, pg_out); 129017a13e40SIlya Dryomov } 129117a13e40SIlya Dryomov 129217a13e40SIlya Dryomov /* 12933d14c5d2SYehuda Sadeh * Pick an osd (the first 'up' osd in the pg), allocate the osd struct 12943d14c5d2SYehuda Sadeh * (as needed), and set the request r_osd appropriately. If there is 129525985edcSLucas De Marchi * no up osd, set r_osd to NULL. Move the request to the appropriate list 12966f6c7006SSage Weil * (unsent, homeless) or leave on in-flight lru. 12973d14c5d2SYehuda Sadeh * 12983d14c5d2SYehuda Sadeh * Return 0 if unchanged, 1 if changed, or negative on error. 12993d14c5d2SYehuda Sadeh * 13003d14c5d2SYehuda Sadeh * Caller should hold map_sem for read and request_mutex. 13013d14c5d2SYehuda Sadeh */ 13026f6c7006SSage Weil static int __map_request(struct ceph_osd_client *osdc, 130338d6453cSSage Weil struct ceph_osd_request *req, int force_resend) 13043d14c5d2SYehuda Sadeh { 13055b191d99SSage Weil struct ceph_pg pgid; 13063d14c5d2SYehuda Sadeh int acting[CEPH_PG_MAX_SIZE]; 13073d14c5d2SYehuda Sadeh int o = -1, num = 0; 13083d14c5d2SYehuda Sadeh int err; 1309d29adb34SJosh Durgin bool was_paused; 13103d14c5d2SYehuda Sadeh 13116f6c7006SSage Weil dout("map_request %p tid %lld\n", req, req->r_tid); 131217a13e40SIlya Dryomov 131317a13e40SIlya Dryomov err = __calc_request_pg(osdc->osdmap, req, &pgid); 13146f6c7006SSage Weil if (err) { 13156f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 13163d14c5d2SYehuda Sadeh return err; 13176f6c7006SSage Weil } 13183d14c5d2SYehuda Sadeh req->r_pgid = pgid; 13193d14c5d2SYehuda Sadeh 13203d14c5d2SYehuda Sadeh err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting); 13213d14c5d2SYehuda Sadeh if (err > 0) { 13223d14c5d2SYehuda Sadeh o = acting[0]; 13233d14c5d2SYehuda Sadeh num = err; 13243d14c5d2SYehuda Sadeh } 13253d14c5d2SYehuda Sadeh 1326d29adb34SJosh Durgin was_paused = req->r_paused; 1327d29adb34SJosh Durgin req->r_paused = __req_should_be_paused(osdc, req); 1328d29adb34SJosh Durgin if (was_paused && !req->r_paused) 1329d29adb34SJosh Durgin force_resend = 1; 1330d29adb34SJosh Durgin 133138d6453cSSage Weil if ((!force_resend && 133238d6453cSSage Weil req->r_osd && req->r_osd->o_osd == o && 13333d14c5d2SYehuda Sadeh req->r_sent >= req->r_osd->o_incarnation && 13343d14c5d2SYehuda Sadeh req->r_num_pg_osds == num && 13353d14c5d2SYehuda Sadeh memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) || 1336d29adb34SJosh Durgin (req->r_osd == NULL && o == -1) || 1337d29adb34SJosh Durgin req->r_paused) 13383d14c5d2SYehuda Sadeh return 0; /* no change */ 13393d14c5d2SYehuda Sadeh 13405b191d99SSage Weil dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n", 13415b191d99SSage Weil req->r_tid, pgid.pool, pgid.seed, o, 13423d14c5d2SYehuda Sadeh req->r_osd ? req->r_osd->o_osd : -1); 13433d14c5d2SYehuda Sadeh 13443d14c5d2SYehuda Sadeh /* record full pg acting set */ 13453d14c5d2SYehuda Sadeh memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num); 13463d14c5d2SYehuda Sadeh req->r_num_pg_osds = num; 13473d14c5d2SYehuda Sadeh 13483d14c5d2SYehuda Sadeh if (req->r_osd) { 13493d14c5d2SYehuda Sadeh __cancel_request(req); 13503d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 13513d14c5d2SYehuda Sadeh req->r_osd = NULL; 13523d14c5d2SYehuda Sadeh } 13533d14c5d2SYehuda Sadeh 13543d14c5d2SYehuda Sadeh req->r_osd = __lookup_osd(osdc, o); 13553d14c5d2SYehuda Sadeh if (!req->r_osd && o >= 0) { 13563d14c5d2SYehuda Sadeh err = -ENOMEM; 1357e10006f8SAlex Elder req->r_osd = create_osd(osdc, o); 13586f6c7006SSage Weil if (!req->r_osd) { 13596f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 13603d14c5d2SYehuda Sadeh goto out; 13616f6c7006SSage Weil } 13623d14c5d2SYehuda Sadeh 13636f6c7006SSage Weil dout("map_request osd %p is osd%d\n", req->r_osd, o); 13643d14c5d2SYehuda Sadeh __insert_osd(osdc, req->r_osd); 13653d14c5d2SYehuda Sadeh 1366b7a9e5ddSSage Weil ceph_con_open(&req->r_osd->o_con, 1367b7a9e5ddSSage Weil CEPH_ENTITY_TYPE_OSD, o, 1368b7a9e5ddSSage Weil &osdc->osdmap->osd_addr[o]); 13693d14c5d2SYehuda Sadeh } 13703d14c5d2SYehuda Sadeh 13713d14c5d2SYehuda Sadeh if (req->r_osd) { 13723d14c5d2SYehuda Sadeh __remove_osd_from_lru(req->r_osd); 1373ad885927SAlex Elder list_add_tail(&req->r_osd_item, &req->r_osd->o_requests); 1374ad885927SAlex Elder list_move_tail(&req->r_req_lru_item, &osdc->req_unsent); 13756f6c7006SSage Weil } else { 1376ad885927SAlex Elder list_move_tail(&req->r_req_lru_item, &osdc->req_notarget); 13773d14c5d2SYehuda Sadeh } 13783d14c5d2SYehuda Sadeh err = 1; /* osd or pg changed */ 13793d14c5d2SYehuda Sadeh 13803d14c5d2SYehuda Sadeh out: 13813d14c5d2SYehuda Sadeh return err; 13823d14c5d2SYehuda Sadeh } 13833d14c5d2SYehuda Sadeh 13843d14c5d2SYehuda Sadeh /* 13853d14c5d2SYehuda Sadeh * caller should hold map_sem (for read) and request_mutex 13863d14c5d2SYehuda Sadeh */ 138756e925b6SSage Weil static void __send_request(struct ceph_osd_client *osdc, 13883d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 13893d14c5d2SYehuda Sadeh { 13901b83bef2SSage Weil void *p; 13913d14c5d2SYehuda Sadeh 13921b83bef2SSage Weil dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n", 13931b83bef2SSage Weil req, req->r_tid, req->r_osd->o_osd, req->r_flags, 13941b83bef2SSage Weil (unsigned long long)req->r_pgid.pool, req->r_pgid.seed); 13953d14c5d2SYehuda Sadeh 13961b83bef2SSage Weil /* fill in message content that changes each time we send it */ 13971b83bef2SSage Weil put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch); 13981b83bef2SSage Weil put_unaligned_le32(req->r_flags, req->r_request_flags); 1399205ee118SIlya Dryomov put_unaligned_le64(req->r_target_oloc.pool, req->r_request_pool); 14001b83bef2SSage Weil p = req->r_request_pgid; 14011b83bef2SSage Weil ceph_encode_64(&p, req->r_pgid.pool); 14021b83bef2SSage Weil ceph_encode_32(&p, req->r_pgid.seed); 14031b83bef2SSage Weil put_unaligned_le64(1, req->r_request_attempts); /* FIXME */ 14041b83bef2SSage Weil memcpy(req->r_request_reassert_version, &req->r_reassert_version, 14051b83bef2SSage Weil sizeof(req->r_reassert_version)); 14062169aea6SSage Weil 14073d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 14083d14c5d2SYehuda Sadeh list_move_tail(&req->r_req_lru_item, &osdc->req_lru); 14093d14c5d2SYehuda Sadeh 14103d14c5d2SYehuda Sadeh ceph_msg_get(req->r_request); /* send consumes a ref */ 141126be8808SAlex Elder 14123d14c5d2SYehuda Sadeh req->r_sent = req->r_osd->o_incarnation; 141326be8808SAlex Elder 141426be8808SAlex Elder ceph_con_send(&req->r_osd->o_con, req->r_request); 14153d14c5d2SYehuda Sadeh } 14163d14c5d2SYehuda Sadeh 14173d14c5d2SYehuda Sadeh /* 14186f6c7006SSage Weil * Send any requests in the queue (req_unsent). 14196f6c7006SSage Weil */ 1420f9d25199SAlex Elder static void __send_queued(struct ceph_osd_client *osdc) 14216f6c7006SSage Weil { 14226f6c7006SSage Weil struct ceph_osd_request *req, *tmp; 14236f6c7006SSage Weil 1424f9d25199SAlex Elder dout("__send_queued\n"); 1425f9d25199SAlex Elder list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) 14266f6c7006SSage Weil __send_request(osdc, req); 14276f6c7006SSage Weil } 14286f6c7006SSage Weil 14296f6c7006SSage Weil /* 14303d14c5d2SYehuda Sadeh * Timeout callback, called every N seconds when 1 or more osd 14313d14c5d2SYehuda Sadeh * requests has been active for more than N seconds. When this 14323d14c5d2SYehuda Sadeh * happens, we ping all OSDs with requests who have timed out to 14333d14c5d2SYehuda Sadeh * ensure any communications channel reset is detected. Reset the 14343d14c5d2SYehuda Sadeh * request timeouts another N seconds in the future as we go. 14353d14c5d2SYehuda Sadeh * Reschedule the timeout event another N seconds in future (unless 14363d14c5d2SYehuda Sadeh * there are no open requests). 14373d14c5d2SYehuda Sadeh */ 14383d14c5d2SYehuda Sadeh static void handle_timeout(struct work_struct *work) 14393d14c5d2SYehuda Sadeh { 14403d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 14413d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, timeout_work.work); 144283aff95eSSage Weil struct ceph_osd_request *req; 14433d14c5d2SYehuda Sadeh struct ceph_osd *osd; 14443d14c5d2SYehuda Sadeh unsigned long keepalive = 14453d14c5d2SYehuda Sadeh osdc->client->options->osd_keepalive_timeout * HZ; 14463d14c5d2SYehuda Sadeh struct list_head slow_osds; 14473d14c5d2SYehuda Sadeh dout("timeout\n"); 14483d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 14493d14c5d2SYehuda Sadeh 14503d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 14513d14c5d2SYehuda Sadeh 14523d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 14533d14c5d2SYehuda Sadeh 14543d14c5d2SYehuda Sadeh /* 14553d14c5d2SYehuda Sadeh * ping osds that are a bit slow. this ensures that if there 14563d14c5d2SYehuda Sadeh * is a break in the TCP connection we will notice, and reopen 14573d14c5d2SYehuda Sadeh * a connection with that osd (from the fault callback). 14583d14c5d2SYehuda Sadeh */ 14593d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&slow_osds); 14603d14c5d2SYehuda Sadeh list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) { 14613d14c5d2SYehuda Sadeh if (time_before(jiffies, req->r_stamp + keepalive)) 14623d14c5d2SYehuda Sadeh break; 14633d14c5d2SYehuda Sadeh 14643d14c5d2SYehuda Sadeh osd = req->r_osd; 14653d14c5d2SYehuda Sadeh BUG_ON(!osd); 14663d14c5d2SYehuda Sadeh dout(" tid %llu is slow, will send keepalive on osd%d\n", 14673d14c5d2SYehuda Sadeh req->r_tid, osd->o_osd); 14683d14c5d2SYehuda Sadeh list_move_tail(&osd->o_keepalive_item, &slow_osds); 14693d14c5d2SYehuda Sadeh } 14703d14c5d2SYehuda Sadeh while (!list_empty(&slow_osds)) { 14713d14c5d2SYehuda Sadeh osd = list_entry(slow_osds.next, struct ceph_osd, 14723d14c5d2SYehuda Sadeh o_keepalive_item); 14733d14c5d2SYehuda Sadeh list_del_init(&osd->o_keepalive_item); 14743d14c5d2SYehuda Sadeh ceph_con_keepalive(&osd->o_con); 14753d14c5d2SYehuda Sadeh } 14763d14c5d2SYehuda Sadeh 14773d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 1478f9d25199SAlex Elder __send_queued(osdc); 14793d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 14803d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 14813d14c5d2SYehuda Sadeh } 14823d14c5d2SYehuda Sadeh 14833d14c5d2SYehuda Sadeh static void handle_osds_timeout(struct work_struct *work) 14843d14c5d2SYehuda Sadeh { 14853d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 14863d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, 14873d14c5d2SYehuda Sadeh osds_timeout_work.work); 14883d14c5d2SYehuda Sadeh unsigned long delay = 14893d14c5d2SYehuda Sadeh osdc->client->options->osd_idle_ttl * HZ >> 2; 14903d14c5d2SYehuda Sadeh 14913d14c5d2SYehuda Sadeh dout("osds timeout\n"); 14923d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 1493aca420bcSSage Weil remove_old_osds(osdc); 14943d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 14953d14c5d2SYehuda Sadeh 14963d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 14973d14c5d2SYehuda Sadeh round_jiffies_relative(delay)); 14983d14c5d2SYehuda Sadeh } 14993d14c5d2SYehuda Sadeh 1500205ee118SIlya Dryomov static int ceph_oloc_decode(void **p, void *end, 1501205ee118SIlya Dryomov struct ceph_object_locator *oloc) 1502205ee118SIlya Dryomov { 1503205ee118SIlya Dryomov u8 struct_v, struct_cv; 1504205ee118SIlya Dryomov u32 len; 1505205ee118SIlya Dryomov void *struct_end; 1506205ee118SIlya Dryomov int ret = 0; 1507205ee118SIlya Dryomov 1508205ee118SIlya Dryomov ceph_decode_need(p, end, 1 + 1 + 4, e_inval); 1509205ee118SIlya Dryomov struct_v = ceph_decode_8(p); 1510205ee118SIlya Dryomov struct_cv = ceph_decode_8(p); 1511205ee118SIlya Dryomov if (struct_v < 3) { 1512205ee118SIlya Dryomov pr_warn("got v %d < 3 cv %d of ceph_object_locator\n", 1513205ee118SIlya Dryomov struct_v, struct_cv); 1514205ee118SIlya Dryomov goto e_inval; 1515205ee118SIlya Dryomov } 1516205ee118SIlya Dryomov if (struct_cv > 6) { 1517205ee118SIlya Dryomov pr_warn("got v %d cv %d > 6 of ceph_object_locator\n", 1518205ee118SIlya Dryomov struct_v, struct_cv); 1519205ee118SIlya Dryomov goto e_inval; 1520205ee118SIlya Dryomov } 1521205ee118SIlya Dryomov len = ceph_decode_32(p); 1522205ee118SIlya Dryomov ceph_decode_need(p, end, len, e_inval); 1523205ee118SIlya Dryomov struct_end = *p + len; 1524205ee118SIlya Dryomov 1525205ee118SIlya Dryomov oloc->pool = ceph_decode_64(p); 1526205ee118SIlya Dryomov *p += 4; /* skip preferred */ 1527205ee118SIlya Dryomov 1528205ee118SIlya Dryomov len = ceph_decode_32(p); 1529205ee118SIlya Dryomov if (len > 0) { 1530205ee118SIlya Dryomov pr_warn("ceph_object_locator::key is set\n"); 1531205ee118SIlya Dryomov goto e_inval; 1532205ee118SIlya Dryomov } 1533205ee118SIlya Dryomov 1534205ee118SIlya Dryomov if (struct_v >= 5) { 1535205ee118SIlya Dryomov len = ceph_decode_32(p); 1536205ee118SIlya Dryomov if (len > 0) { 1537205ee118SIlya Dryomov pr_warn("ceph_object_locator::nspace is set\n"); 1538205ee118SIlya Dryomov goto e_inval; 1539205ee118SIlya Dryomov } 1540205ee118SIlya Dryomov } 1541205ee118SIlya Dryomov 1542205ee118SIlya Dryomov if (struct_v >= 6) { 1543205ee118SIlya Dryomov s64 hash = ceph_decode_64(p); 1544205ee118SIlya Dryomov if (hash != -1) { 1545205ee118SIlya Dryomov pr_warn("ceph_object_locator::hash is set\n"); 1546205ee118SIlya Dryomov goto e_inval; 1547205ee118SIlya Dryomov } 1548205ee118SIlya Dryomov } 1549205ee118SIlya Dryomov 1550205ee118SIlya Dryomov /* skip the rest */ 1551205ee118SIlya Dryomov *p = struct_end; 1552205ee118SIlya Dryomov out: 1553205ee118SIlya Dryomov return ret; 1554205ee118SIlya Dryomov 1555205ee118SIlya Dryomov e_inval: 1556205ee118SIlya Dryomov ret = -EINVAL; 1557205ee118SIlya Dryomov goto out; 1558205ee118SIlya Dryomov } 1559205ee118SIlya Dryomov 1560205ee118SIlya Dryomov static int ceph_redirect_decode(void **p, void *end, 1561205ee118SIlya Dryomov struct ceph_request_redirect *redir) 1562205ee118SIlya Dryomov { 1563205ee118SIlya Dryomov u8 struct_v, struct_cv; 1564205ee118SIlya Dryomov u32 len; 1565205ee118SIlya Dryomov void *struct_end; 1566205ee118SIlya Dryomov int ret; 1567205ee118SIlya Dryomov 1568205ee118SIlya Dryomov ceph_decode_need(p, end, 1 + 1 + 4, e_inval); 1569205ee118SIlya Dryomov struct_v = ceph_decode_8(p); 1570205ee118SIlya Dryomov struct_cv = ceph_decode_8(p); 1571205ee118SIlya Dryomov if (struct_cv > 1) { 1572205ee118SIlya Dryomov pr_warn("got v %d cv %d > 1 of ceph_request_redirect\n", 1573205ee118SIlya Dryomov struct_v, struct_cv); 1574205ee118SIlya Dryomov goto e_inval; 1575205ee118SIlya Dryomov } 1576205ee118SIlya Dryomov len = ceph_decode_32(p); 1577205ee118SIlya Dryomov ceph_decode_need(p, end, len, e_inval); 1578205ee118SIlya Dryomov struct_end = *p + len; 1579205ee118SIlya Dryomov 1580205ee118SIlya Dryomov ret = ceph_oloc_decode(p, end, &redir->oloc); 1581205ee118SIlya Dryomov if (ret) 1582205ee118SIlya Dryomov goto out; 1583205ee118SIlya Dryomov 1584205ee118SIlya Dryomov len = ceph_decode_32(p); 1585205ee118SIlya Dryomov if (len > 0) { 1586205ee118SIlya Dryomov pr_warn("ceph_request_redirect::object_name is set\n"); 1587205ee118SIlya Dryomov goto e_inval; 1588205ee118SIlya Dryomov } 1589205ee118SIlya Dryomov 1590205ee118SIlya Dryomov len = ceph_decode_32(p); 1591205ee118SIlya Dryomov *p += len; /* skip osd_instructions */ 1592205ee118SIlya Dryomov 1593205ee118SIlya Dryomov /* skip the rest */ 1594205ee118SIlya Dryomov *p = struct_end; 1595205ee118SIlya Dryomov out: 1596205ee118SIlya Dryomov return ret; 1597205ee118SIlya Dryomov 1598205ee118SIlya Dryomov e_inval: 1599205ee118SIlya Dryomov ret = -EINVAL; 1600205ee118SIlya Dryomov goto out; 1601205ee118SIlya Dryomov } 1602205ee118SIlya Dryomov 160325845472SSage Weil static void complete_request(struct ceph_osd_request *req) 160425845472SSage Weil { 160525845472SSage Weil complete_all(&req->r_safe_completion); /* fsync waiter */ 160625845472SSage Weil } 160725845472SSage Weil 16083d14c5d2SYehuda Sadeh /* 16093d14c5d2SYehuda Sadeh * handle osd op reply. either call the callback if it is specified, 16103d14c5d2SYehuda Sadeh * or do the completion to wake up the waiting thread. 16113d14c5d2SYehuda Sadeh */ 16123d14c5d2SYehuda Sadeh static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg, 16133d14c5d2SYehuda Sadeh struct ceph_connection *con) 16143d14c5d2SYehuda Sadeh { 16151b83bef2SSage Weil void *p, *end; 16163d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 1617205ee118SIlya Dryomov struct ceph_request_redirect redir; 16183d14c5d2SYehuda Sadeh u64 tid; 16191b83bef2SSage Weil int object_len; 162079528734SAlex Elder unsigned int numops; 162179528734SAlex Elder int payload_len, flags; 16223d14c5d2SYehuda Sadeh s32 result; 16231b83bef2SSage Weil s32 retry_attempt; 16241b83bef2SSage Weil struct ceph_pg pg; 16251b83bef2SSage Weil int err; 16261b83bef2SSage Weil u32 reassert_epoch; 16271b83bef2SSage Weil u64 reassert_version; 16281b83bef2SSage Weil u32 osdmap_epoch; 16290d5af164SAlex Elder int already_completed; 16309fc6e064SAlex Elder u32 bytes; 163179528734SAlex Elder unsigned int i; 16323d14c5d2SYehuda Sadeh 16333d14c5d2SYehuda Sadeh tid = le64_to_cpu(msg->hdr.tid); 16341b83bef2SSage Weil dout("handle_reply %p tid %llu\n", msg, tid); 16351b83bef2SSage Weil 16361b83bef2SSage Weil p = msg->front.iov_base; 16371b83bef2SSage Weil end = p + msg->front.iov_len; 16381b83bef2SSage Weil 16391b83bef2SSage Weil ceph_decode_need(&p, end, 4, bad); 16401b83bef2SSage Weil object_len = ceph_decode_32(&p); 16411b83bef2SSage Weil ceph_decode_need(&p, end, object_len, bad); 16421b83bef2SSage Weil p += object_len; 16431b83bef2SSage Weil 1644ef4859d6SAlex Elder err = ceph_decode_pgid(&p, end, &pg); 16451b83bef2SSage Weil if (err) 16463d14c5d2SYehuda Sadeh goto bad; 16471b83bef2SSage Weil 16481b83bef2SSage Weil ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad); 16491b83bef2SSage Weil flags = ceph_decode_64(&p); 16501b83bef2SSage Weil result = ceph_decode_32(&p); 16511b83bef2SSage Weil reassert_epoch = ceph_decode_32(&p); 16521b83bef2SSage Weil reassert_version = ceph_decode_64(&p); 16531b83bef2SSage Weil osdmap_epoch = ceph_decode_32(&p); 16541b83bef2SSage Weil 16553d14c5d2SYehuda Sadeh /* lookup */ 16563d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 16573d14c5d2SYehuda Sadeh req = __lookup_request(osdc, tid); 16583d14c5d2SYehuda Sadeh if (req == NULL) { 16593d14c5d2SYehuda Sadeh dout("handle_reply tid %llu dne\n", tid); 16608058fd45SAlex Elder goto bad_mutex; 16613d14c5d2SYehuda Sadeh } 16623d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 16631b83bef2SSage Weil 16641b83bef2SSage Weil dout("handle_reply %p tid %llu req %p result %d\n", msg, tid, 16651b83bef2SSage Weil req, result); 16661b83bef2SSage Weil 166718741196SDan Carpenter ceph_decode_need(&p, end, 4, bad_put); 16681b83bef2SSage Weil numops = ceph_decode_32(&p); 16691b83bef2SSage Weil if (numops > CEPH_OSD_MAX_OP) 16701b83bef2SSage Weil goto bad_put; 16711b83bef2SSage Weil if (numops != req->r_num_ops) 16721b83bef2SSage Weil goto bad_put; 16731b83bef2SSage Weil payload_len = 0; 167418741196SDan Carpenter ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad_put); 16751b83bef2SSage Weil for (i = 0; i < numops; i++) { 16761b83bef2SSage Weil struct ceph_osd_op *op = p; 16771b83bef2SSage Weil int len; 16781b83bef2SSage Weil 16791b83bef2SSage Weil len = le32_to_cpu(op->payload_len); 16801b83bef2SSage Weil req->r_reply_op_len[i] = len; 16811b83bef2SSage Weil dout(" op %d has %d bytes\n", i, len); 16821b83bef2SSage Weil payload_len += len; 16831b83bef2SSage Weil p += sizeof(*op); 16841b83bef2SSage Weil } 16859fc6e064SAlex Elder bytes = le32_to_cpu(msg->hdr.data_len); 16869fc6e064SAlex Elder if (payload_len != bytes) { 16871b83bef2SSage Weil pr_warning("sum of op payload lens %d != data_len %d", 16889fc6e064SAlex Elder payload_len, bytes); 16891b83bef2SSage Weil goto bad_put; 16901b83bef2SSage Weil } 16911b83bef2SSage Weil 169218741196SDan Carpenter ceph_decode_need(&p, end, 4 + numops * 4, bad_put); 16931b83bef2SSage Weil retry_attempt = ceph_decode_32(&p); 16941b83bef2SSage Weil for (i = 0; i < numops; i++) 16951b83bef2SSage Weil req->r_reply_op_result[i] = ceph_decode_32(&p); 16963d14c5d2SYehuda Sadeh 1697205ee118SIlya Dryomov if (le16_to_cpu(msg->hdr.version) >= 6) { 1698205ee118SIlya Dryomov p += 8 + 4; /* skip replay_version */ 1699205ee118SIlya Dryomov p += 8; /* skip user_version */ 1700205ee118SIlya Dryomov 1701205ee118SIlya Dryomov err = ceph_redirect_decode(&p, end, &redir); 1702205ee118SIlya Dryomov if (err) 1703205ee118SIlya Dryomov goto bad_put; 1704205ee118SIlya Dryomov } else { 1705205ee118SIlya Dryomov redir.oloc.pool = -1; 1706205ee118SIlya Dryomov } 1707205ee118SIlya Dryomov 1708205ee118SIlya Dryomov if (redir.oloc.pool != -1) { 1709205ee118SIlya Dryomov dout("redirect pool %lld\n", redir.oloc.pool); 1710205ee118SIlya Dryomov 1711205ee118SIlya Dryomov __unregister_request(osdc, req); 1712205ee118SIlya Dryomov mutex_unlock(&osdc->request_mutex); 1713205ee118SIlya Dryomov 1714205ee118SIlya Dryomov req->r_target_oloc = redir.oloc; /* struct */ 1715205ee118SIlya Dryomov 1716205ee118SIlya Dryomov /* 1717205ee118SIlya Dryomov * Start redirect requests with nofail=true. If 1718205ee118SIlya Dryomov * mapping fails, request will end up on the notarget 1719205ee118SIlya Dryomov * list, waiting for the new osdmap (which can take 1720205ee118SIlya Dryomov * a while), even though the original request mapped 1721205ee118SIlya Dryomov * successfully. In the future we might want to follow 1722205ee118SIlya Dryomov * original request's nofail setting here. 1723205ee118SIlya Dryomov */ 1724205ee118SIlya Dryomov err = ceph_osdc_start_request(osdc, req, true); 1725205ee118SIlya Dryomov BUG_ON(err); 1726205ee118SIlya Dryomov 1727205ee118SIlya Dryomov goto done; 1728205ee118SIlya Dryomov } 1729205ee118SIlya Dryomov 1730eb845ff1SYan, Zheng already_completed = req->r_got_reply; 17313d14c5d2SYehuda Sadeh if (!req->r_got_reply) { 17321b83bef2SSage Weil req->r_result = result; 17333d14c5d2SYehuda Sadeh dout("handle_reply result %d bytes %d\n", req->r_result, 17343d14c5d2SYehuda Sadeh bytes); 17353d14c5d2SYehuda Sadeh if (req->r_result == 0) 17363d14c5d2SYehuda Sadeh req->r_result = bytes; 17373d14c5d2SYehuda Sadeh 17383d14c5d2SYehuda Sadeh /* in case this is a write and we need to replay, */ 17391b83bef2SSage Weil req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch); 17401b83bef2SSage Weil req->r_reassert_version.version = cpu_to_le64(reassert_version); 17413d14c5d2SYehuda Sadeh 17423d14c5d2SYehuda Sadeh req->r_got_reply = 1; 17433d14c5d2SYehuda Sadeh } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) { 17443d14c5d2SYehuda Sadeh dout("handle_reply tid %llu dup ack\n", tid); 17453d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 17463d14c5d2SYehuda Sadeh goto done; 17473d14c5d2SYehuda Sadeh } 17483d14c5d2SYehuda Sadeh 17493d14c5d2SYehuda Sadeh dout("handle_reply tid %llu flags %d\n", tid, flags); 17503d14c5d2SYehuda Sadeh 1751a40c4f10SYehuda Sadeh if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK)) 1752a40c4f10SYehuda Sadeh __register_linger_request(osdc, req); 1753a40c4f10SYehuda Sadeh 17543d14c5d2SYehuda Sadeh /* either this is a read, or we got the safe response */ 17553d14c5d2SYehuda Sadeh if (result < 0 || 17563d14c5d2SYehuda Sadeh (flags & CEPH_OSD_FLAG_ONDISK) || 17573d14c5d2SYehuda Sadeh ((flags & CEPH_OSD_FLAG_WRITE) == 0)) 17583d14c5d2SYehuda Sadeh __unregister_request(osdc, req); 17593d14c5d2SYehuda Sadeh 17603d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 17613d14c5d2SYehuda Sadeh 1762eb845ff1SYan, Zheng if (!already_completed) { 176361c5d6bfSYan, Zheng if (req->r_unsafe_callback && 176461c5d6bfSYan, Zheng result >= 0 && !(flags & CEPH_OSD_FLAG_ONDISK)) 176561c5d6bfSYan, Zheng req->r_unsafe_callback(req, true); 17663d14c5d2SYehuda Sadeh if (req->r_callback) 17673d14c5d2SYehuda Sadeh req->r_callback(req, msg); 17683d14c5d2SYehuda Sadeh else 17693d14c5d2SYehuda Sadeh complete_all(&req->r_completion); 1770eb845ff1SYan, Zheng } 17713d14c5d2SYehuda Sadeh 177261c5d6bfSYan, Zheng if (flags & CEPH_OSD_FLAG_ONDISK) { 177361c5d6bfSYan, Zheng if (req->r_unsafe_callback && already_completed) 177461c5d6bfSYan, Zheng req->r_unsafe_callback(req, false); 177525845472SSage Weil complete_request(req); 177661c5d6bfSYan, Zheng } 17773d14c5d2SYehuda Sadeh 17783d14c5d2SYehuda Sadeh done: 1779a40c4f10SYehuda Sadeh dout("req=%p req->r_linger=%d\n", req, req->r_linger); 17803d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 17813d14c5d2SYehuda Sadeh return; 17823d14c5d2SYehuda Sadeh 17831b83bef2SSage Weil bad_put: 178437c89bdeSLi Wang req->r_result = -EIO; 178537c89bdeSLi Wang __unregister_request(osdc, req); 178637c89bdeSLi Wang if (req->r_callback) 178737c89bdeSLi Wang req->r_callback(req, msg); 178837c89bdeSLi Wang else 178937c89bdeSLi Wang complete_all(&req->r_completion); 179037c89bdeSLi Wang complete_request(req); 17911b83bef2SSage Weil ceph_osdc_put_request(req); 17928058fd45SAlex Elder bad_mutex: 17938058fd45SAlex Elder mutex_unlock(&osdc->request_mutex); 17943d14c5d2SYehuda Sadeh bad: 17951b83bef2SSage Weil pr_err("corrupt osd_op_reply got %d %d\n", 17961b83bef2SSage Weil (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len)); 17973d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 17983d14c5d2SYehuda Sadeh } 17993d14c5d2SYehuda Sadeh 18006f6c7006SSage Weil static void reset_changed_osds(struct ceph_osd_client *osdc) 18013d14c5d2SYehuda Sadeh { 18023d14c5d2SYehuda Sadeh struct rb_node *p, *n; 18033d14c5d2SYehuda Sadeh 18043d14c5d2SYehuda Sadeh for (p = rb_first(&osdc->osds); p; p = n) { 18056f6c7006SSage Weil struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node); 18063d14c5d2SYehuda Sadeh 18073d14c5d2SYehuda Sadeh n = rb_next(p); 18083d14c5d2SYehuda Sadeh if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) || 18093d14c5d2SYehuda Sadeh memcmp(&osd->o_con.peer_addr, 18103d14c5d2SYehuda Sadeh ceph_osd_addr(osdc->osdmap, 18113d14c5d2SYehuda Sadeh osd->o_osd), 18123d14c5d2SYehuda Sadeh sizeof(struct ceph_entity_addr)) != 0) 18133d14c5d2SYehuda Sadeh __reset_osd(osdc, osd); 18143d14c5d2SYehuda Sadeh } 18153d14c5d2SYehuda Sadeh } 18163d14c5d2SYehuda Sadeh 18173d14c5d2SYehuda Sadeh /* 18186f6c7006SSage Weil * Requeue requests whose mapping to an OSD has changed. If requests map to 18196f6c7006SSage Weil * no osd, request a new map. 18203d14c5d2SYehuda Sadeh * 1821e6d50f67SAlex Elder * Caller should hold map_sem for read. 18223d14c5d2SYehuda Sadeh */ 18239a1ea2dbSJosh Durgin static void kick_requests(struct ceph_osd_client *osdc, bool force_resend, 18249a1ea2dbSJosh Durgin bool force_resend_writes) 18253d14c5d2SYehuda Sadeh { 1826a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 18276f6c7006SSage Weil struct rb_node *p; 18286f6c7006SSage Weil int needmap = 0; 18296f6c7006SSage Weil int err; 18309a1ea2dbSJosh Durgin bool force_resend_req; 18313d14c5d2SYehuda Sadeh 18329a1ea2dbSJosh Durgin dout("kick_requests %s %s\n", force_resend ? " (force resend)" : "", 18339a1ea2dbSJosh Durgin force_resend_writes ? " (force resend writes)" : ""); 18343d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 18356194ea89SSage Weil for (p = rb_first(&osdc->requests); p; ) { 18366f6c7006SSage Weil req = rb_entry(p, struct ceph_osd_request, r_node); 18376194ea89SSage Weil p = rb_next(p); 1838ab60b16dSAlex Elder 1839ab60b16dSAlex Elder /* 1840ab60b16dSAlex Elder * For linger requests that have not yet been 1841ab60b16dSAlex Elder * registered, move them to the linger list; they'll 1842ab60b16dSAlex Elder * be sent to the osd in the loop below. Unregister 1843ab60b16dSAlex Elder * the request before re-registering it as a linger 1844ab60b16dSAlex Elder * request to ensure the __map_request() below 1845ab60b16dSAlex Elder * will decide it needs to be sent. 1846ab60b16dSAlex Elder */ 1847ab60b16dSAlex Elder if (req->r_linger && list_empty(&req->r_linger_item)) { 1848ab60b16dSAlex Elder dout("%p tid %llu restart on osd%d\n", 1849ab60b16dSAlex Elder req, req->r_tid, 1850ab60b16dSAlex Elder req->r_osd ? req->r_osd->o_osd : -1); 185196e4dac6SAlex Elder ceph_osdc_get_request(req); 1852ab60b16dSAlex Elder __unregister_request(osdc, req); 1853ab60b16dSAlex Elder __register_linger_request(osdc, req); 185496e4dac6SAlex Elder ceph_osdc_put_request(req); 1855ab60b16dSAlex Elder continue; 1856ab60b16dSAlex Elder } 1857ab60b16dSAlex Elder 18589a1ea2dbSJosh Durgin force_resend_req = force_resend || 18599a1ea2dbSJosh Durgin (force_resend_writes && 18609a1ea2dbSJosh Durgin req->r_flags & CEPH_OSD_FLAG_WRITE); 18619a1ea2dbSJosh Durgin err = __map_request(osdc, req, force_resend_req); 18626f6c7006SSage Weil if (err < 0) 18636f6c7006SSage Weil continue; /* error */ 18646f6c7006SSage Weil if (req->r_osd == NULL) { 18656f6c7006SSage Weil dout("%p tid %llu maps to no osd\n", req, req->r_tid); 18666f6c7006SSage Weil needmap++; /* request a newer map */ 18676f6c7006SSage Weil } else if (err > 0) { 18686194ea89SSage Weil if (!req->r_linger) { 18696194ea89SSage Weil dout("%p tid %llu requeued on osd%d\n", req, 18706194ea89SSage Weil req->r_tid, 18716f6c7006SSage Weil req->r_osd ? req->r_osd->o_osd : -1); 18726f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 18736f6c7006SSage Weil } 18746f6c7006SSage Weil } 18756194ea89SSage Weil } 1876a40c4f10SYehuda Sadeh 1877a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osdc->req_linger, 1878a40c4f10SYehuda Sadeh r_linger_item) { 1879a40c4f10SYehuda Sadeh dout("linger req=%p req->r_osd=%p\n", req, req->r_osd); 1880a40c4f10SYehuda Sadeh 18819a1ea2dbSJosh Durgin err = __map_request(osdc, req, 18829a1ea2dbSJosh Durgin force_resend || force_resend_writes); 1883ab60b16dSAlex Elder dout("__map_request returned %d\n", err); 1884a40c4f10SYehuda Sadeh if (err == 0) 1885a40c4f10SYehuda Sadeh continue; /* no change and no osd was specified */ 1886a40c4f10SYehuda Sadeh if (err < 0) 1887a40c4f10SYehuda Sadeh continue; /* hrm! */ 1888a40c4f10SYehuda Sadeh if (req->r_osd == NULL) { 1889a40c4f10SYehuda Sadeh dout("tid %llu maps to no valid osd\n", req->r_tid); 1890a40c4f10SYehuda Sadeh needmap++; /* request a newer map */ 1891a40c4f10SYehuda Sadeh continue; 1892a40c4f10SYehuda Sadeh } 1893a40c4f10SYehuda Sadeh 1894a40c4f10SYehuda Sadeh dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid, 1895a40c4f10SYehuda Sadeh req->r_osd ? req->r_osd->o_osd : -1); 1896a40c4f10SYehuda Sadeh __register_request(osdc, req); 1897c89ce05eSAlex Elder __unregister_linger_request(osdc, req); 1898a40c4f10SYehuda Sadeh } 189914d2f38dSAlex Elder reset_changed_osds(osdc); 19003d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 19013d14c5d2SYehuda Sadeh 19023d14c5d2SYehuda Sadeh if (needmap) { 19033d14c5d2SYehuda Sadeh dout("%d requests for down osds, need new map\n", needmap); 19043d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 19053d14c5d2SYehuda Sadeh } 19063d14c5d2SYehuda Sadeh } 19076f6c7006SSage Weil 19086f6c7006SSage Weil 19093d14c5d2SYehuda Sadeh /* 19103d14c5d2SYehuda Sadeh * Process updated osd map. 19113d14c5d2SYehuda Sadeh * 19123d14c5d2SYehuda Sadeh * The message contains any number of incremental and full maps, normally 19133d14c5d2SYehuda Sadeh * indicating some sort of topology change in the cluster. Kick requests 19143d14c5d2SYehuda Sadeh * off to different OSDs as needed. 19153d14c5d2SYehuda Sadeh */ 19163d14c5d2SYehuda Sadeh void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg) 19173d14c5d2SYehuda Sadeh { 19183d14c5d2SYehuda Sadeh void *p, *end, *next; 19193d14c5d2SYehuda Sadeh u32 nr_maps, maplen; 19203d14c5d2SYehuda Sadeh u32 epoch; 19213d14c5d2SYehuda Sadeh struct ceph_osdmap *newmap = NULL, *oldmap; 19223d14c5d2SYehuda Sadeh int err; 19233d14c5d2SYehuda Sadeh struct ceph_fsid fsid; 19249a1ea2dbSJosh Durgin bool was_full; 19253d14c5d2SYehuda Sadeh 19263d14c5d2SYehuda Sadeh dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0); 19273d14c5d2SYehuda Sadeh p = msg->front.iov_base; 19283d14c5d2SYehuda Sadeh end = p + msg->front.iov_len; 19293d14c5d2SYehuda Sadeh 19303d14c5d2SYehuda Sadeh /* verify fsid */ 19313d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, sizeof(fsid), bad); 19323d14c5d2SYehuda Sadeh ceph_decode_copy(&p, &fsid, sizeof(fsid)); 19333d14c5d2SYehuda Sadeh if (ceph_check_fsid(osdc->client, &fsid) < 0) 19343d14c5d2SYehuda Sadeh return; 19353d14c5d2SYehuda Sadeh 19363d14c5d2SYehuda Sadeh down_write(&osdc->map_sem); 19373d14c5d2SYehuda Sadeh 19389a1ea2dbSJosh Durgin was_full = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL); 19399a1ea2dbSJosh Durgin 19403d14c5d2SYehuda Sadeh /* incremental maps */ 19413d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 19423d14c5d2SYehuda Sadeh dout(" %d inc maps\n", nr_maps); 19433d14c5d2SYehuda Sadeh while (nr_maps > 0) { 19443d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 19453d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 19463d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 19473d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 19483d14c5d2SYehuda Sadeh next = p + maplen; 19493d14c5d2SYehuda Sadeh if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) { 19503d14c5d2SYehuda Sadeh dout("applying incremental map %u len %d\n", 19513d14c5d2SYehuda Sadeh epoch, maplen); 19523d14c5d2SYehuda Sadeh newmap = osdmap_apply_incremental(&p, next, 19533d14c5d2SYehuda Sadeh osdc->osdmap, 195415d9882cSAlex Elder &osdc->client->msgr); 19553d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 19563d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 19573d14c5d2SYehuda Sadeh goto bad; 19583d14c5d2SYehuda Sadeh } 19593d14c5d2SYehuda Sadeh BUG_ON(!newmap); 19603d14c5d2SYehuda Sadeh if (newmap != osdc->osdmap) { 19613d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 19623d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 19633d14c5d2SYehuda Sadeh } 19649a1ea2dbSJosh Durgin was_full = was_full || 19659a1ea2dbSJosh Durgin ceph_osdmap_flag(osdc->osdmap, 19669a1ea2dbSJosh Durgin CEPH_OSDMAP_FULL); 19679a1ea2dbSJosh Durgin kick_requests(osdc, 0, was_full); 19683d14c5d2SYehuda Sadeh } else { 19693d14c5d2SYehuda Sadeh dout("ignoring incremental map %u len %d\n", 19703d14c5d2SYehuda Sadeh epoch, maplen); 19713d14c5d2SYehuda Sadeh } 19723d14c5d2SYehuda Sadeh p = next; 19733d14c5d2SYehuda Sadeh nr_maps--; 19743d14c5d2SYehuda Sadeh } 19753d14c5d2SYehuda Sadeh if (newmap) 19763d14c5d2SYehuda Sadeh goto done; 19773d14c5d2SYehuda Sadeh 19783d14c5d2SYehuda Sadeh /* full maps */ 19793d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 19803d14c5d2SYehuda Sadeh dout(" %d full maps\n", nr_maps); 19813d14c5d2SYehuda Sadeh while (nr_maps) { 19823d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 19833d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 19843d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 19853d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 19863d14c5d2SYehuda Sadeh if (nr_maps > 1) { 19873d14c5d2SYehuda Sadeh dout("skipping non-latest full map %u len %d\n", 19883d14c5d2SYehuda Sadeh epoch, maplen); 19893d14c5d2SYehuda Sadeh } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) { 19903d14c5d2SYehuda Sadeh dout("skipping full map %u len %d, " 19913d14c5d2SYehuda Sadeh "older than our %u\n", epoch, maplen, 19923d14c5d2SYehuda Sadeh osdc->osdmap->epoch); 19933d14c5d2SYehuda Sadeh } else { 199438d6453cSSage Weil int skipped_map = 0; 199538d6453cSSage Weil 19963d14c5d2SYehuda Sadeh dout("taking full map %u len %d\n", epoch, maplen); 19973d14c5d2SYehuda Sadeh newmap = osdmap_decode(&p, p+maplen); 19983d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 19993d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 20003d14c5d2SYehuda Sadeh goto bad; 20013d14c5d2SYehuda Sadeh } 20023d14c5d2SYehuda Sadeh BUG_ON(!newmap); 20033d14c5d2SYehuda Sadeh oldmap = osdc->osdmap; 20043d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 200538d6453cSSage Weil if (oldmap) { 200638d6453cSSage Weil if (oldmap->epoch + 1 < newmap->epoch) 200738d6453cSSage Weil skipped_map = 1; 20083d14c5d2SYehuda Sadeh ceph_osdmap_destroy(oldmap); 200938d6453cSSage Weil } 20109a1ea2dbSJosh Durgin was_full = was_full || 20119a1ea2dbSJosh Durgin ceph_osdmap_flag(osdc->osdmap, 20129a1ea2dbSJosh Durgin CEPH_OSDMAP_FULL); 20139a1ea2dbSJosh Durgin kick_requests(osdc, skipped_map, was_full); 20143d14c5d2SYehuda Sadeh } 20153d14c5d2SYehuda Sadeh p += maplen; 20163d14c5d2SYehuda Sadeh nr_maps--; 20173d14c5d2SYehuda Sadeh } 20183d14c5d2SYehuda Sadeh 2019b72e19b9SDan Carpenter if (!osdc->osdmap) 2020b72e19b9SDan Carpenter goto bad; 20213d14c5d2SYehuda Sadeh done: 20223d14c5d2SYehuda Sadeh downgrade_write(&osdc->map_sem); 20233d14c5d2SYehuda Sadeh ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch); 2024cd634fb6SSage Weil 2025cd634fb6SSage Weil /* 2026cd634fb6SSage Weil * subscribe to subsequent osdmap updates if full to ensure 2027cd634fb6SSage Weil * we find out when we are no longer full and stop returning 2028cd634fb6SSage Weil * ENOSPC. 2029cd634fb6SSage Weil */ 2030d29adb34SJosh Durgin if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) || 2031d29adb34SJosh Durgin ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD) || 2032d29adb34SJosh Durgin ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR)) 2033cd634fb6SSage Weil ceph_monc_request_next_osdmap(&osdc->client->monc); 2034cd634fb6SSage Weil 2035f9d25199SAlex Elder mutex_lock(&osdc->request_mutex); 2036f9d25199SAlex Elder __send_queued(osdc); 2037f9d25199SAlex Elder mutex_unlock(&osdc->request_mutex); 20383d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 20393d14c5d2SYehuda Sadeh wake_up_all(&osdc->client->auth_wq); 20403d14c5d2SYehuda Sadeh return; 20413d14c5d2SYehuda Sadeh 20423d14c5d2SYehuda Sadeh bad: 20433d14c5d2SYehuda Sadeh pr_err("osdc handle_map corrupt msg\n"); 20443d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 20453d14c5d2SYehuda Sadeh up_write(&osdc->map_sem); 20463d14c5d2SYehuda Sadeh return; 20473d14c5d2SYehuda Sadeh } 20483d14c5d2SYehuda Sadeh 20493d14c5d2SYehuda Sadeh /* 2050a40c4f10SYehuda Sadeh * watch/notify callback event infrastructure 2051a40c4f10SYehuda Sadeh * 2052a40c4f10SYehuda Sadeh * These callbacks are used both for watch and notify operations. 2053a40c4f10SYehuda Sadeh */ 2054a40c4f10SYehuda Sadeh static void __release_event(struct kref *kref) 2055a40c4f10SYehuda Sadeh { 2056a40c4f10SYehuda Sadeh struct ceph_osd_event *event = 2057a40c4f10SYehuda Sadeh container_of(kref, struct ceph_osd_event, kref); 2058a40c4f10SYehuda Sadeh 2059a40c4f10SYehuda Sadeh dout("__release_event %p\n", event); 2060a40c4f10SYehuda Sadeh kfree(event); 2061a40c4f10SYehuda Sadeh } 2062a40c4f10SYehuda Sadeh 2063a40c4f10SYehuda Sadeh static void get_event(struct ceph_osd_event *event) 2064a40c4f10SYehuda Sadeh { 2065a40c4f10SYehuda Sadeh kref_get(&event->kref); 2066a40c4f10SYehuda Sadeh } 2067a40c4f10SYehuda Sadeh 2068a40c4f10SYehuda Sadeh void ceph_osdc_put_event(struct ceph_osd_event *event) 2069a40c4f10SYehuda Sadeh { 2070a40c4f10SYehuda Sadeh kref_put(&event->kref, __release_event); 2071a40c4f10SYehuda Sadeh } 2072a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_put_event); 2073a40c4f10SYehuda Sadeh 2074a40c4f10SYehuda Sadeh static void __insert_event(struct ceph_osd_client *osdc, 2075a40c4f10SYehuda Sadeh struct ceph_osd_event *new) 2076a40c4f10SYehuda Sadeh { 2077a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 2078a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 2079a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 2080a40c4f10SYehuda Sadeh 2081a40c4f10SYehuda Sadeh while (*p) { 2082a40c4f10SYehuda Sadeh parent = *p; 2083a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 2084a40c4f10SYehuda Sadeh if (new->cookie < event->cookie) 2085a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 2086a40c4f10SYehuda Sadeh else if (new->cookie > event->cookie) 2087a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 2088a40c4f10SYehuda Sadeh else 2089a40c4f10SYehuda Sadeh BUG(); 2090a40c4f10SYehuda Sadeh } 2091a40c4f10SYehuda Sadeh 2092a40c4f10SYehuda Sadeh rb_link_node(&new->node, parent, p); 2093a40c4f10SYehuda Sadeh rb_insert_color(&new->node, &osdc->event_tree); 2094a40c4f10SYehuda Sadeh } 2095a40c4f10SYehuda Sadeh 2096a40c4f10SYehuda Sadeh static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc, 2097a40c4f10SYehuda Sadeh u64 cookie) 2098a40c4f10SYehuda Sadeh { 2099a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 2100a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 2101a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 2102a40c4f10SYehuda Sadeh 2103a40c4f10SYehuda Sadeh while (*p) { 2104a40c4f10SYehuda Sadeh parent = *p; 2105a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 2106a40c4f10SYehuda Sadeh if (cookie < event->cookie) 2107a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 2108a40c4f10SYehuda Sadeh else if (cookie > event->cookie) 2109a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 2110a40c4f10SYehuda Sadeh else 2111a40c4f10SYehuda Sadeh return event; 2112a40c4f10SYehuda Sadeh } 2113a40c4f10SYehuda Sadeh return NULL; 2114a40c4f10SYehuda Sadeh } 2115a40c4f10SYehuda Sadeh 2116a40c4f10SYehuda Sadeh static void __remove_event(struct ceph_osd_event *event) 2117a40c4f10SYehuda Sadeh { 2118a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 2119a40c4f10SYehuda Sadeh 2120a40c4f10SYehuda Sadeh if (!RB_EMPTY_NODE(&event->node)) { 2121a40c4f10SYehuda Sadeh dout("__remove_event removed %p\n", event); 2122a40c4f10SYehuda Sadeh rb_erase(&event->node, &osdc->event_tree); 2123a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 2124a40c4f10SYehuda Sadeh } else { 2125a40c4f10SYehuda Sadeh dout("__remove_event didn't remove %p\n", event); 2126a40c4f10SYehuda Sadeh } 2127a40c4f10SYehuda Sadeh } 2128a40c4f10SYehuda Sadeh 2129a40c4f10SYehuda Sadeh int ceph_osdc_create_event(struct ceph_osd_client *osdc, 2130a40c4f10SYehuda Sadeh void (*event_cb)(u64, u64, u8, void *), 21313c663bbdSAlex Elder void *data, struct ceph_osd_event **pevent) 2132a40c4f10SYehuda Sadeh { 2133a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 2134a40c4f10SYehuda Sadeh 2135a40c4f10SYehuda Sadeh event = kmalloc(sizeof(*event), GFP_NOIO); 2136a40c4f10SYehuda Sadeh if (!event) 2137a40c4f10SYehuda Sadeh return -ENOMEM; 2138a40c4f10SYehuda Sadeh 2139a40c4f10SYehuda Sadeh dout("create_event %p\n", event); 2140a40c4f10SYehuda Sadeh event->cb = event_cb; 21413c663bbdSAlex Elder event->one_shot = 0; 2142a40c4f10SYehuda Sadeh event->data = data; 2143a40c4f10SYehuda Sadeh event->osdc = osdc; 2144a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&event->osd_node); 21453ee5234dSAlex Elder RB_CLEAR_NODE(&event->node); 2146a40c4f10SYehuda Sadeh kref_init(&event->kref); /* one ref for us */ 2147a40c4f10SYehuda Sadeh kref_get(&event->kref); /* one ref for the caller */ 2148a40c4f10SYehuda Sadeh 2149a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 2150a40c4f10SYehuda Sadeh event->cookie = ++osdc->event_count; 2151a40c4f10SYehuda Sadeh __insert_event(osdc, event); 2152a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 2153a40c4f10SYehuda Sadeh 2154a40c4f10SYehuda Sadeh *pevent = event; 2155a40c4f10SYehuda Sadeh return 0; 2156a40c4f10SYehuda Sadeh } 2157a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_create_event); 2158a40c4f10SYehuda Sadeh 2159a40c4f10SYehuda Sadeh void ceph_osdc_cancel_event(struct ceph_osd_event *event) 2160a40c4f10SYehuda Sadeh { 2161a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 2162a40c4f10SYehuda Sadeh 2163a40c4f10SYehuda Sadeh dout("cancel_event %p\n", event); 2164a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 2165a40c4f10SYehuda Sadeh __remove_event(event); 2166a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 2167a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); /* caller's */ 2168a40c4f10SYehuda Sadeh } 2169a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_cancel_event); 2170a40c4f10SYehuda Sadeh 2171a40c4f10SYehuda Sadeh 2172a40c4f10SYehuda Sadeh static void do_event_work(struct work_struct *work) 2173a40c4f10SYehuda Sadeh { 2174a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work = 2175a40c4f10SYehuda Sadeh container_of(work, struct ceph_osd_event_work, work); 2176a40c4f10SYehuda Sadeh struct ceph_osd_event *event = event_work->event; 2177a40c4f10SYehuda Sadeh u64 ver = event_work->ver; 2178a40c4f10SYehuda Sadeh u64 notify_id = event_work->notify_id; 2179a40c4f10SYehuda Sadeh u8 opcode = event_work->opcode; 2180a40c4f10SYehuda Sadeh 2181a40c4f10SYehuda Sadeh dout("do_event_work completing %p\n", event); 2182a40c4f10SYehuda Sadeh event->cb(ver, notify_id, opcode, event->data); 2183a40c4f10SYehuda Sadeh dout("do_event_work completed %p\n", event); 2184a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 2185a40c4f10SYehuda Sadeh kfree(event_work); 2186a40c4f10SYehuda Sadeh } 2187a40c4f10SYehuda Sadeh 2188a40c4f10SYehuda Sadeh 2189a40c4f10SYehuda Sadeh /* 2190a40c4f10SYehuda Sadeh * Process osd watch notifications 2191a40c4f10SYehuda Sadeh */ 21923c663bbdSAlex Elder static void handle_watch_notify(struct ceph_osd_client *osdc, 21933c663bbdSAlex Elder struct ceph_msg *msg) 2194a40c4f10SYehuda Sadeh { 2195a40c4f10SYehuda Sadeh void *p, *end; 2196a40c4f10SYehuda Sadeh u8 proto_ver; 2197a40c4f10SYehuda Sadeh u64 cookie, ver, notify_id; 2198a40c4f10SYehuda Sadeh u8 opcode; 2199a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 2200a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work; 2201a40c4f10SYehuda Sadeh 2202a40c4f10SYehuda Sadeh p = msg->front.iov_base; 2203a40c4f10SYehuda Sadeh end = p + msg->front.iov_len; 2204a40c4f10SYehuda Sadeh 2205a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, proto_ver, bad); 2206a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, opcode, bad); 2207a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, cookie, bad); 2208a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, ver, bad); 2209a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, notify_id, bad); 2210a40c4f10SYehuda Sadeh 2211a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 2212a40c4f10SYehuda Sadeh event = __find_event(osdc, cookie); 2213a40c4f10SYehuda Sadeh if (event) { 22143c663bbdSAlex Elder BUG_ON(event->one_shot); 2215a40c4f10SYehuda Sadeh get_event(event); 2216a40c4f10SYehuda Sadeh } 2217a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 2218a40c4f10SYehuda Sadeh dout("handle_watch_notify cookie %lld ver %lld event %p\n", 2219a40c4f10SYehuda Sadeh cookie, ver, event); 2220a40c4f10SYehuda Sadeh if (event) { 2221a40c4f10SYehuda Sadeh event_work = kmalloc(sizeof(*event_work), GFP_NOIO); 2222a40c4f10SYehuda Sadeh if (!event_work) { 2223a40c4f10SYehuda Sadeh dout("ERROR: could not allocate event_work\n"); 2224a40c4f10SYehuda Sadeh goto done_err; 2225a40c4f10SYehuda Sadeh } 22266b0ae409SMariusz Kozlowski INIT_WORK(&event_work->work, do_event_work); 2227a40c4f10SYehuda Sadeh event_work->event = event; 2228a40c4f10SYehuda Sadeh event_work->ver = ver; 2229a40c4f10SYehuda Sadeh event_work->notify_id = notify_id; 2230a40c4f10SYehuda Sadeh event_work->opcode = opcode; 2231a40c4f10SYehuda Sadeh if (!queue_work(osdc->notify_wq, &event_work->work)) { 2232a40c4f10SYehuda Sadeh dout("WARNING: failed to queue notify event work\n"); 2233a40c4f10SYehuda Sadeh goto done_err; 2234a40c4f10SYehuda Sadeh } 2235a40c4f10SYehuda Sadeh } 2236a40c4f10SYehuda Sadeh 2237a40c4f10SYehuda Sadeh return; 2238a40c4f10SYehuda Sadeh 2239a40c4f10SYehuda Sadeh done_err: 2240a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 2241a40c4f10SYehuda Sadeh return; 2242a40c4f10SYehuda Sadeh 2243a40c4f10SYehuda Sadeh bad: 2244a40c4f10SYehuda Sadeh pr_err("osdc handle_watch_notify corrupt msg\n"); 2245a40c4f10SYehuda Sadeh return; 2246a40c4f10SYehuda Sadeh } 2247a40c4f10SYehuda Sadeh 2248a40c4f10SYehuda Sadeh /* 2249e65550fdSAlex Elder * build new request AND message 2250e65550fdSAlex Elder * 2251e65550fdSAlex Elder */ 2252e65550fdSAlex Elder void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off, 2253e65550fdSAlex Elder struct ceph_snap_context *snapc, u64 snap_id, 2254e65550fdSAlex Elder struct timespec *mtime) 2255e65550fdSAlex Elder { 2256e65550fdSAlex Elder struct ceph_msg *msg = req->r_request; 2257e65550fdSAlex Elder void *p; 2258e65550fdSAlex Elder size_t msg_size; 2259e65550fdSAlex Elder int flags = req->r_flags; 2260e65550fdSAlex Elder u64 data_len; 2261e65550fdSAlex Elder unsigned int i; 2262e65550fdSAlex Elder 2263e65550fdSAlex Elder req->r_snapid = snap_id; 2264e65550fdSAlex Elder req->r_snapc = ceph_get_snap_context(snapc); 2265e65550fdSAlex Elder 2266e65550fdSAlex Elder /* encode request */ 2267e65550fdSAlex Elder msg->hdr.version = cpu_to_le16(4); 2268e65550fdSAlex Elder 2269e65550fdSAlex Elder p = msg->front.iov_base; 2270e65550fdSAlex Elder ceph_encode_32(&p, 1); /* client_inc is always 1 */ 2271e65550fdSAlex Elder req->r_request_osdmap_epoch = p; 2272e65550fdSAlex Elder p += 4; 2273e65550fdSAlex Elder req->r_request_flags = p; 2274e65550fdSAlex Elder p += 4; 2275e65550fdSAlex Elder if (req->r_flags & CEPH_OSD_FLAG_WRITE) 2276e65550fdSAlex Elder ceph_encode_timespec(p, mtime); 2277e65550fdSAlex Elder p += sizeof(struct ceph_timespec); 2278e65550fdSAlex Elder req->r_request_reassert_version = p; 2279e65550fdSAlex Elder p += sizeof(struct ceph_eversion); /* will get filled in */ 2280e65550fdSAlex Elder 2281e65550fdSAlex Elder /* oloc */ 2282e65550fdSAlex Elder ceph_encode_8(&p, 4); 2283e65550fdSAlex Elder ceph_encode_8(&p, 4); 2284e65550fdSAlex Elder ceph_encode_32(&p, 8 + 4 + 4); 2285e65550fdSAlex Elder req->r_request_pool = p; 2286e65550fdSAlex Elder p += 8; 2287e65550fdSAlex Elder ceph_encode_32(&p, -1); /* preferred */ 2288e65550fdSAlex Elder ceph_encode_32(&p, 0); /* key len */ 2289e65550fdSAlex Elder 2290e65550fdSAlex Elder ceph_encode_8(&p, 1); 2291e65550fdSAlex Elder req->r_request_pgid = p; 2292e65550fdSAlex Elder p += 8 + 4; 2293e65550fdSAlex Elder ceph_encode_32(&p, -1); /* preferred */ 2294e65550fdSAlex Elder 2295e65550fdSAlex Elder /* oid */ 22963c972c95SIlya Dryomov ceph_encode_32(&p, req->r_base_oid.name_len); 22973c972c95SIlya Dryomov memcpy(p, req->r_base_oid.name, req->r_base_oid.name_len); 22983c972c95SIlya Dryomov dout("oid '%.*s' len %d\n", req->r_base_oid.name_len, 22993c972c95SIlya Dryomov req->r_base_oid.name, req->r_base_oid.name_len); 23003c972c95SIlya Dryomov p += req->r_base_oid.name_len; 2301e65550fdSAlex Elder 2302e65550fdSAlex Elder /* ops--can imply data */ 2303e65550fdSAlex Elder ceph_encode_16(&p, (u16)req->r_num_ops); 2304e65550fdSAlex Elder data_len = 0; 2305e65550fdSAlex Elder for (i = 0; i < req->r_num_ops; i++) { 2306e65550fdSAlex Elder data_len += osd_req_encode_op(req, p, i); 2307e65550fdSAlex Elder p += sizeof(struct ceph_osd_op); 2308e65550fdSAlex Elder } 2309e65550fdSAlex Elder 2310e65550fdSAlex Elder /* snaps */ 2311e65550fdSAlex Elder ceph_encode_64(&p, req->r_snapid); 2312e65550fdSAlex Elder ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0); 2313e65550fdSAlex Elder ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0); 2314e65550fdSAlex Elder if (req->r_snapc) { 2315e65550fdSAlex Elder for (i = 0; i < snapc->num_snaps; i++) { 2316e65550fdSAlex Elder ceph_encode_64(&p, req->r_snapc->snaps[i]); 2317e65550fdSAlex Elder } 2318e65550fdSAlex Elder } 2319e65550fdSAlex Elder 2320e65550fdSAlex Elder req->r_request_attempts = p; 2321e65550fdSAlex Elder p += 4; 2322e65550fdSAlex Elder 2323e65550fdSAlex Elder /* data */ 2324e65550fdSAlex Elder if (flags & CEPH_OSD_FLAG_WRITE) { 2325e65550fdSAlex Elder u16 data_off; 2326e65550fdSAlex Elder 2327e65550fdSAlex Elder /* 2328e65550fdSAlex Elder * The header "data_off" is a hint to the receiver 2329e65550fdSAlex Elder * allowing it to align received data into its 2330e65550fdSAlex Elder * buffers such that there's no need to re-copy 2331e65550fdSAlex Elder * it before writing it to disk (direct I/O). 2332e65550fdSAlex Elder */ 2333e65550fdSAlex Elder data_off = (u16) (off & 0xffff); 2334e65550fdSAlex Elder req->r_request->hdr.data_off = cpu_to_le16(data_off); 2335e65550fdSAlex Elder } 2336e65550fdSAlex Elder req->r_request->hdr.data_len = cpu_to_le32(data_len); 2337e65550fdSAlex Elder 2338e65550fdSAlex Elder BUG_ON(p > msg->front.iov_base + msg->front.iov_len); 2339e65550fdSAlex Elder msg_size = p - msg->front.iov_base; 2340e65550fdSAlex Elder msg->front.iov_len = msg_size; 2341e65550fdSAlex Elder msg->hdr.front_len = cpu_to_le32(msg_size); 2342e65550fdSAlex Elder 2343e65550fdSAlex Elder dout("build_request msg_size was %d\n", (int)msg_size); 2344e65550fdSAlex Elder } 2345e65550fdSAlex Elder EXPORT_SYMBOL(ceph_osdc_build_request); 2346e65550fdSAlex Elder 2347e65550fdSAlex Elder /* 23483d14c5d2SYehuda Sadeh * Register request, send initial attempt. 23493d14c5d2SYehuda Sadeh */ 23503d14c5d2SYehuda Sadeh int ceph_osdc_start_request(struct ceph_osd_client *osdc, 23513d14c5d2SYehuda Sadeh struct ceph_osd_request *req, 23523d14c5d2SYehuda Sadeh bool nofail) 23533d14c5d2SYehuda Sadeh { 23543d14c5d2SYehuda Sadeh int rc = 0; 23553d14c5d2SYehuda Sadeh 23563d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 23573d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 2358dc4b870cSAlex Elder __register_request(osdc, req); 2359c10ebbf5SAlex Elder req->r_sent = 0; 2360c10ebbf5SAlex Elder req->r_got_reply = 0; 236138d6453cSSage Weil rc = __map_request(osdc, req, 0); 23629d6fcb08SSage Weil if (rc < 0) { 23639d6fcb08SSage Weil if (nofail) { 23649d6fcb08SSage Weil dout("osdc_start_request failed map, " 23659d6fcb08SSage Weil " will retry %lld\n", req->r_tid); 23669d6fcb08SSage Weil rc = 0; 236773d9f7eeSmajianpeng } else { 236873d9f7eeSmajianpeng __unregister_request(osdc, req); 23699d6fcb08SSage Weil } 2370234af26fSDan Carpenter goto out_unlock; 23719d6fcb08SSage Weil } 23726f6c7006SSage Weil if (req->r_osd == NULL) { 23736f6c7006SSage Weil dout("send_request %p no up osds in pg\n", req); 23746f6c7006SSage Weil ceph_monc_request_next_osdmap(&osdc->client->monc); 23756f6c7006SSage Weil } else { 23767e2766a1SAlex Elder __send_queued(osdc); 237756e925b6SSage Weil } 23783d14c5d2SYehuda Sadeh rc = 0; 2379234af26fSDan Carpenter out_unlock: 23803d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 23813d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 23823d14c5d2SYehuda Sadeh return rc; 23833d14c5d2SYehuda Sadeh } 23843d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_start_request); 23853d14c5d2SYehuda Sadeh 23863d14c5d2SYehuda Sadeh /* 23873d14c5d2SYehuda Sadeh * wait for a request to complete 23883d14c5d2SYehuda Sadeh */ 23893d14c5d2SYehuda Sadeh int ceph_osdc_wait_request(struct ceph_osd_client *osdc, 23903d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 23913d14c5d2SYehuda Sadeh { 23923d14c5d2SYehuda Sadeh int rc; 23933d14c5d2SYehuda Sadeh 23943d14c5d2SYehuda Sadeh rc = wait_for_completion_interruptible(&req->r_completion); 23953d14c5d2SYehuda Sadeh if (rc < 0) { 23963d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 23973d14c5d2SYehuda Sadeh __cancel_request(req); 23983d14c5d2SYehuda Sadeh __unregister_request(osdc, req); 23993d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 240025845472SSage Weil complete_request(req); 24013d14c5d2SYehuda Sadeh dout("wait_request tid %llu canceled/timed out\n", req->r_tid); 24023d14c5d2SYehuda Sadeh return rc; 24033d14c5d2SYehuda Sadeh } 24043d14c5d2SYehuda Sadeh 24053d14c5d2SYehuda Sadeh dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result); 24063d14c5d2SYehuda Sadeh return req->r_result; 24073d14c5d2SYehuda Sadeh } 24083d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_wait_request); 24093d14c5d2SYehuda Sadeh 24103d14c5d2SYehuda Sadeh /* 24113d14c5d2SYehuda Sadeh * sync - wait for all in-flight requests to flush. avoid starvation. 24123d14c5d2SYehuda Sadeh */ 24133d14c5d2SYehuda Sadeh void ceph_osdc_sync(struct ceph_osd_client *osdc) 24143d14c5d2SYehuda Sadeh { 24153d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 24163d14c5d2SYehuda Sadeh u64 last_tid, next_tid = 0; 24173d14c5d2SYehuda Sadeh 24183d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 24193d14c5d2SYehuda Sadeh last_tid = osdc->last_tid; 24203d14c5d2SYehuda Sadeh while (1) { 24213d14c5d2SYehuda Sadeh req = __lookup_request_ge(osdc, next_tid); 24223d14c5d2SYehuda Sadeh if (!req) 24233d14c5d2SYehuda Sadeh break; 24243d14c5d2SYehuda Sadeh if (req->r_tid > last_tid) 24253d14c5d2SYehuda Sadeh break; 24263d14c5d2SYehuda Sadeh 24273d14c5d2SYehuda Sadeh next_tid = req->r_tid + 1; 24283d14c5d2SYehuda Sadeh if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0) 24293d14c5d2SYehuda Sadeh continue; 24303d14c5d2SYehuda Sadeh 24313d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 24323d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 24333d14c5d2SYehuda Sadeh dout("sync waiting on tid %llu (last is %llu)\n", 24343d14c5d2SYehuda Sadeh req->r_tid, last_tid); 24353d14c5d2SYehuda Sadeh wait_for_completion(&req->r_safe_completion); 24363d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 24373d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 24383d14c5d2SYehuda Sadeh } 24393d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 24403d14c5d2SYehuda Sadeh dout("sync done (thru tid %llu)\n", last_tid); 24413d14c5d2SYehuda Sadeh } 24423d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_sync); 24433d14c5d2SYehuda Sadeh 24443d14c5d2SYehuda Sadeh /* 2445dd935f44SJosh Durgin * Call all pending notify callbacks - for use after a watch is 2446dd935f44SJosh Durgin * unregistered, to make sure no more callbacks for it will be invoked 2447dd935f44SJosh Durgin */ 2448dd935f44SJosh Durgin extern void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc) 2449dd935f44SJosh Durgin { 2450dd935f44SJosh Durgin flush_workqueue(osdc->notify_wq); 2451dd935f44SJosh Durgin } 2452dd935f44SJosh Durgin EXPORT_SYMBOL(ceph_osdc_flush_notifies); 2453dd935f44SJosh Durgin 2454dd935f44SJosh Durgin 2455dd935f44SJosh Durgin /* 24563d14c5d2SYehuda Sadeh * init, shutdown 24573d14c5d2SYehuda Sadeh */ 24583d14c5d2SYehuda Sadeh int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client) 24593d14c5d2SYehuda Sadeh { 24603d14c5d2SYehuda Sadeh int err; 24613d14c5d2SYehuda Sadeh 24623d14c5d2SYehuda Sadeh dout("init\n"); 24633d14c5d2SYehuda Sadeh osdc->client = client; 24643d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 24653d14c5d2SYehuda Sadeh init_rwsem(&osdc->map_sem); 24663d14c5d2SYehuda Sadeh init_completion(&osdc->map_waiters); 24673d14c5d2SYehuda Sadeh osdc->last_requested_map = 0; 24683d14c5d2SYehuda Sadeh mutex_init(&osdc->request_mutex); 24693d14c5d2SYehuda Sadeh osdc->last_tid = 0; 24703d14c5d2SYehuda Sadeh osdc->osds = RB_ROOT; 24713d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->osd_lru); 24723d14c5d2SYehuda Sadeh osdc->requests = RB_ROOT; 24733d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_lru); 24746f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_unsent); 24756f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_notarget); 2476a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_linger); 24773d14c5d2SYehuda Sadeh osdc->num_requests = 0; 24783d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout); 24793d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout); 2480a40c4f10SYehuda Sadeh spin_lock_init(&osdc->event_lock); 2481a40c4f10SYehuda Sadeh osdc->event_tree = RB_ROOT; 2482a40c4f10SYehuda Sadeh osdc->event_count = 0; 24833d14c5d2SYehuda Sadeh 24843d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 24853d14c5d2SYehuda Sadeh round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ)); 24863d14c5d2SYehuda Sadeh 24873d14c5d2SYehuda Sadeh err = -ENOMEM; 24883d14c5d2SYehuda Sadeh osdc->req_mempool = mempool_create_kmalloc_pool(10, 24893d14c5d2SYehuda Sadeh sizeof(struct ceph_osd_request)); 24903d14c5d2SYehuda Sadeh if (!osdc->req_mempool) 24913d14c5d2SYehuda Sadeh goto out; 24923d14c5d2SYehuda Sadeh 2493d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP, 2494d50b409fSSage Weil OSD_OP_FRONT_LEN, 10, true, 24953d14c5d2SYehuda Sadeh "osd_op"); 24963d14c5d2SYehuda Sadeh if (err < 0) 24973d14c5d2SYehuda Sadeh goto out_mempool; 2498d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY, 24993d14c5d2SYehuda Sadeh OSD_OPREPLY_FRONT_LEN, 10, true, 25003d14c5d2SYehuda Sadeh "osd_op_reply"); 25013d14c5d2SYehuda Sadeh if (err < 0) 25023d14c5d2SYehuda Sadeh goto out_msgpool; 2503a40c4f10SYehuda Sadeh 2504dbcae088SDan Carpenter err = -ENOMEM; 2505a40c4f10SYehuda Sadeh osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify"); 2506dbcae088SDan Carpenter if (!osdc->notify_wq) 2507a40c4f10SYehuda Sadeh goto out_msgpool; 25083d14c5d2SYehuda Sadeh return 0; 25093d14c5d2SYehuda Sadeh 25103d14c5d2SYehuda Sadeh out_msgpool: 25113d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 25123d14c5d2SYehuda Sadeh out_mempool: 25133d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 25143d14c5d2SYehuda Sadeh out: 25153d14c5d2SYehuda Sadeh return err; 25163d14c5d2SYehuda Sadeh } 25173d14c5d2SYehuda Sadeh 25183d14c5d2SYehuda Sadeh void ceph_osdc_stop(struct ceph_osd_client *osdc) 25193d14c5d2SYehuda Sadeh { 2520a40c4f10SYehuda Sadeh flush_workqueue(osdc->notify_wq); 2521a40c4f10SYehuda Sadeh destroy_workqueue(osdc->notify_wq); 25223d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->timeout_work); 25233d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->osds_timeout_work); 25243d14c5d2SYehuda Sadeh if (osdc->osdmap) { 25253d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 25263d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 25273d14c5d2SYehuda Sadeh } 2528aca420bcSSage Weil remove_all_osds(osdc); 25293d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 25303d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 25313d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op_reply); 25323d14c5d2SYehuda Sadeh } 25333d14c5d2SYehuda Sadeh 25343d14c5d2SYehuda Sadeh /* 25353d14c5d2SYehuda Sadeh * Read some contiguous pages. If we cross a stripe boundary, shorten 25363d14c5d2SYehuda Sadeh * *plen. Return number of bytes read, or error. 25373d14c5d2SYehuda Sadeh */ 25383d14c5d2SYehuda Sadeh int ceph_osdc_readpages(struct ceph_osd_client *osdc, 25393d14c5d2SYehuda Sadeh struct ceph_vino vino, struct ceph_file_layout *layout, 25403d14c5d2SYehuda Sadeh u64 off, u64 *plen, 25413d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 2542b7495fc2SSage Weil struct page **pages, int num_pages, int page_align) 25433d14c5d2SYehuda Sadeh { 25443d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 25453d14c5d2SYehuda Sadeh int rc = 0; 25463d14c5d2SYehuda Sadeh 25473d14c5d2SYehuda Sadeh dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino, 25483d14c5d2SYehuda Sadeh vino.snap, off, *plen); 254979528734SAlex Elder req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 1, 25503d14c5d2SYehuda Sadeh CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ, 2551acead002SAlex Elder NULL, truncate_seq, truncate_size, 2552153e5167SAlex Elder false); 25536816282dSSage Weil if (IS_ERR(req)) 25546816282dSSage Weil return PTR_ERR(req); 25553d14c5d2SYehuda Sadeh 25563d14c5d2SYehuda Sadeh /* it may be a short read due to an object boundary */ 25570fff87ecSAlex Elder 2558406e2c9fSAlex Elder osd_req_op_extent_osd_data_pages(req, 0, 2559a4ce40a9SAlex Elder pages, *plen, page_align, false, false); 25603d14c5d2SYehuda Sadeh 2561e0c59487SAlex Elder dout("readpages final extent is %llu~%llu (%llu bytes align %d)\n", 256243bfe5deSAlex Elder off, *plen, *plen, page_align); 25633d14c5d2SYehuda Sadeh 256479528734SAlex Elder ceph_osdc_build_request(req, off, NULL, vino.snap, NULL); 256502ee07d3SAlex Elder 25663d14c5d2SYehuda Sadeh rc = ceph_osdc_start_request(osdc, req, false); 25673d14c5d2SYehuda Sadeh if (!rc) 25683d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 25693d14c5d2SYehuda Sadeh 25703d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 25713d14c5d2SYehuda Sadeh dout("readpages result %d\n", rc); 25723d14c5d2SYehuda Sadeh return rc; 25733d14c5d2SYehuda Sadeh } 25743d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_readpages); 25753d14c5d2SYehuda Sadeh 25763d14c5d2SYehuda Sadeh /* 25773d14c5d2SYehuda Sadeh * do a synchronous write on N pages 25783d14c5d2SYehuda Sadeh */ 25793d14c5d2SYehuda Sadeh int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino, 25803d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 25813d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 25823d14c5d2SYehuda Sadeh u64 off, u64 len, 25833d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 25843d14c5d2SYehuda Sadeh struct timespec *mtime, 258524808826SAlex Elder struct page **pages, int num_pages) 25863d14c5d2SYehuda Sadeh { 25873d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 25883d14c5d2SYehuda Sadeh int rc = 0; 2589b7495fc2SSage Weil int page_align = off & ~PAGE_MASK; 25903d14c5d2SYehuda Sadeh 2591acead002SAlex Elder BUG_ON(vino.snap != CEPH_NOSNAP); /* snapshots aren't writeable */ 259279528734SAlex Elder req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 1, 25933d14c5d2SYehuda Sadeh CEPH_OSD_OP_WRITE, 259424808826SAlex Elder CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE, 2595acead002SAlex Elder snapc, truncate_seq, truncate_size, 2596153e5167SAlex Elder true); 25976816282dSSage Weil if (IS_ERR(req)) 25986816282dSSage Weil return PTR_ERR(req); 25993d14c5d2SYehuda Sadeh 26003d14c5d2SYehuda Sadeh /* it may be a short write due to an object boundary */ 2601406e2c9fSAlex Elder osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align, 260243bfe5deSAlex Elder false, false); 260343bfe5deSAlex Elder dout("writepages %llu~%llu (%llu bytes)\n", off, len, len); 26043d14c5d2SYehuda Sadeh 260579528734SAlex Elder ceph_osdc_build_request(req, off, snapc, CEPH_NOSNAP, mtime); 260602ee07d3SAlex Elder 260787f979d3SAlex Elder rc = ceph_osdc_start_request(osdc, req, true); 26083d14c5d2SYehuda Sadeh if (!rc) 26093d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 26103d14c5d2SYehuda Sadeh 26113d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 26123d14c5d2SYehuda Sadeh if (rc == 0) 26133d14c5d2SYehuda Sadeh rc = len; 26143d14c5d2SYehuda Sadeh dout("writepages result %d\n", rc); 26153d14c5d2SYehuda Sadeh return rc; 26163d14c5d2SYehuda Sadeh } 26173d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_writepages); 26183d14c5d2SYehuda Sadeh 26195522ae0bSAlex Elder int ceph_osdc_setup(void) 26205522ae0bSAlex Elder { 26215522ae0bSAlex Elder BUG_ON(ceph_osd_request_cache); 26225522ae0bSAlex Elder ceph_osd_request_cache = kmem_cache_create("ceph_osd_request", 26235522ae0bSAlex Elder sizeof (struct ceph_osd_request), 26245522ae0bSAlex Elder __alignof__(struct ceph_osd_request), 26255522ae0bSAlex Elder 0, NULL); 26265522ae0bSAlex Elder 26275522ae0bSAlex Elder return ceph_osd_request_cache ? 0 : -ENOMEM; 26285522ae0bSAlex Elder } 26295522ae0bSAlex Elder EXPORT_SYMBOL(ceph_osdc_setup); 26305522ae0bSAlex Elder 26315522ae0bSAlex Elder void ceph_osdc_cleanup(void) 26325522ae0bSAlex Elder { 26335522ae0bSAlex Elder BUG_ON(!ceph_osd_request_cache); 26345522ae0bSAlex Elder kmem_cache_destroy(ceph_osd_request_cache); 26355522ae0bSAlex Elder ceph_osd_request_cache = NULL; 26365522ae0bSAlex Elder } 26375522ae0bSAlex Elder EXPORT_SYMBOL(ceph_osdc_cleanup); 26385522ae0bSAlex Elder 26393d14c5d2SYehuda Sadeh /* 26403d14c5d2SYehuda Sadeh * handle incoming message 26413d14c5d2SYehuda Sadeh */ 26423d14c5d2SYehuda Sadeh static void dispatch(struct ceph_connection *con, struct ceph_msg *msg) 26433d14c5d2SYehuda Sadeh { 26443d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 26453d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 26463d14c5d2SYehuda Sadeh int type = le16_to_cpu(msg->hdr.type); 26473d14c5d2SYehuda Sadeh 26483d14c5d2SYehuda Sadeh if (!osd) 26493d14c5d2SYehuda Sadeh goto out; 26503d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 26513d14c5d2SYehuda Sadeh 26523d14c5d2SYehuda Sadeh switch (type) { 26533d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 26543d14c5d2SYehuda Sadeh ceph_osdc_handle_map(osdc, msg); 26553d14c5d2SYehuda Sadeh break; 26563d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 26573d14c5d2SYehuda Sadeh handle_reply(osdc, msg, con); 26583d14c5d2SYehuda Sadeh break; 2659a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 2660a40c4f10SYehuda Sadeh handle_watch_notify(osdc, msg); 2661a40c4f10SYehuda Sadeh break; 26623d14c5d2SYehuda Sadeh 26633d14c5d2SYehuda Sadeh default: 26643d14c5d2SYehuda Sadeh pr_err("received unknown message type %d %s\n", type, 26653d14c5d2SYehuda Sadeh ceph_msg_type_name(type)); 26663d14c5d2SYehuda Sadeh } 26673d14c5d2SYehuda Sadeh out: 26683d14c5d2SYehuda Sadeh ceph_msg_put(msg); 26693d14c5d2SYehuda Sadeh } 26703d14c5d2SYehuda Sadeh 26713d14c5d2SYehuda Sadeh /* 26723d14c5d2SYehuda Sadeh * lookup and return message for incoming reply. set up reply message 26733d14c5d2SYehuda Sadeh * pages. 26743d14c5d2SYehuda Sadeh */ 26753d14c5d2SYehuda Sadeh static struct ceph_msg *get_reply(struct ceph_connection *con, 26763d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 26773d14c5d2SYehuda Sadeh int *skip) 26783d14c5d2SYehuda Sadeh { 26793d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 26803d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = osd->o_osdc; 26813d14c5d2SYehuda Sadeh struct ceph_msg *m; 26823d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 26833f0a4ac5SIlya Dryomov int front_len = le32_to_cpu(hdr->front_len); 26843d14c5d2SYehuda Sadeh int data_len = le32_to_cpu(hdr->data_len); 26853d14c5d2SYehuda Sadeh u64 tid; 26863d14c5d2SYehuda Sadeh 26873d14c5d2SYehuda Sadeh tid = le64_to_cpu(hdr->tid); 26883d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 26893d14c5d2SYehuda Sadeh req = __lookup_request(osdc, tid); 26903d14c5d2SYehuda Sadeh if (!req) { 26913d14c5d2SYehuda Sadeh *skip = 1; 26923d14c5d2SYehuda Sadeh m = NULL; 2693756a16a5SSage Weil dout("get_reply unknown tid %llu from osd%d\n", tid, 26943d14c5d2SYehuda Sadeh osd->o_osd); 26953d14c5d2SYehuda Sadeh goto out; 26963d14c5d2SYehuda Sadeh } 26973d14c5d2SYehuda Sadeh 2698ace6d3a9SAlex Elder if (req->r_reply->con) 26998921d114SAlex Elder dout("%s revoking msg %p from old con %p\n", __func__, 2700ace6d3a9SAlex Elder req->r_reply, req->r_reply->con); 27018921d114SAlex Elder ceph_msg_revoke_incoming(req->r_reply); 27023d14c5d2SYehuda Sadeh 2703f2be82b0SIlya Dryomov if (front_len > req->r_reply->front_alloc_len) { 27044974341eSAlex Elder pr_warning("get_reply front %d > preallocated %d (%u#%llu)\n", 2705f2be82b0SIlya Dryomov front_len, req->r_reply->front_alloc_len, 27064974341eSAlex Elder (unsigned int)con->peer_name.type, 27074974341eSAlex Elder le64_to_cpu(con->peer_name.num)); 27083f0a4ac5SIlya Dryomov m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front_len, GFP_NOFS, 27093f0a4ac5SIlya Dryomov false); 27103d14c5d2SYehuda Sadeh if (!m) 27113d14c5d2SYehuda Sadeh goto out; 27123d14c5d2SYehuda Sadeh ceph_msg_put(req->r_reply); 27133d14c5d2SYehuda Sadeh req->r_reply = m; 27143d14c5d2SYehuda Sadeh } 27153d14c5d2SYehuda Sadeh m = ceph_msg_get(req->r_reply); 27163d14c5d2SYehuda Sadeh 27173d14c5d2SYehuda Sadeh if (data_len > 0) { 2718a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 27190fff87ecSAlex Elder 2720a4ce40a9SAlex Elder /* 2721a4ce40a9SAlex Elder * XXX This is assuming there is only one op containing 2722a4ce40a9SAlex Elder * XXX page data. Probably OK for reads, but this 2723a4ce40a9SAlex Elder * XXX ought to be done more generally. 2724a4ce40a9SAlex Elder */ 2725406e2c9fSAlex Elder osd_data = osd_req_op_extent_osd_data(req, 0); 27260fff87ecSAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 27270fff87ecSAlex Elder if (osd_data->pages && 2728e0c59487SAlex Elder unlikely(osd_data->length < data_len)) { 27292ac2b7a6SAlex Elder 2730e0c59487SAlex Elder pr_warning("tid %lld reply has %d bytes " 2731e0c59487SAlex Elder "we had only %llu bytes ready\n", 2732e0c59487SAlex Elder tid, data_len, osd_data->length); 27333d14c5d2SYehuda Sadeh *skip = 1; 27343d14c5d2SYehuda Sadeh ceph_msg_put(m); 27353d14c5d2SYehuda Sadeh m = NULL; 27363d14c5d2SYehuda Sadeh goto out; 27373d14c5d2SYehuda Sadeh } 27383d14c5d2SYehuda Sadeh } 27392ac2b7a6SAlex Elder } 27403d14c5d2SYehuda Sadeh *skip = 0; 27413d14c5d2SYehuda Sadeh dout("get_reply tid %lld %p\n", tid, m); 27423d14c5d2SYehuda Sadeh 27433d14c5d2SYehuda Sadeh out: 27443d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 27453d14c5d2SYehuda Sadeh return m; 27463d14c5d2SYehuda Sadeh 27473d14c5d2SYehuda Sadeh } 27483d14c5d2SYehuda Sadeh 27493d14c5d2SYehuda Sadeh static struct ceph_msg *alloc_msg(struct ceph_connection *con, 27503d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 27513d14c5d2SYehuda Sadeh int *skip) 27523d14c5d2SYehuda Sadeh { 27533d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 27543d14c5d2SYehuda Sadeh int type = le16_to_cpu(hdr->type); 27553d14c5d2SYehuda Sadeh int front = le32_to_cpu(hdr->front_len); 27563d14c5d2SYehuda Sadeh 27571c20f2d2SAlex Elder *skip = 0; 27583d14c5d2SYehuda Sadeh switch (type) { 27593d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 2760a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 2761b61c2763SSage Weil return ceph_msg_new(type, front, GFP_NOFS, false); 27623d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 27633d14c5d2SYehuda Sadeh return get_reply(con, hdr, skip); 27643d14c5d2SYehuda Sadeh default: 27653d14c5d2SYehuda Sadeh pr_info("alloc_msg unexpected msg type %d from osd%d\n", type, 27663d14c5d2SYehuda Sadeh osd->o_osd); 27673d14c5d2SYehuda Sadeh *skip = 1; 27683d14c5d2SYehuda Sadeh return NULL; 27693d14c5d2SYehuda Sadeh } 27703d14c5d2SYehuda Sadeh } 27713d14c5d2SYehuda Sadeh 27723d14c5d2SYehuda Sadeh /* 27733d14c5d2SYehuda Sadeh * Wrappers to refcount containing ceph_osd struct 27743d14c5d2SYehuda Sadeh */ 27753d14c5d2SYehuda Sadeh static struct ceph_connection *get_osd_con(struct ceph_connection *con) 27763d14c5d2SYehuda Sadeh { 27773d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 27783d14c5d2SYehuda Sadeh if (get_osd(osd)) 27793d14c5d2SYehuda Sadeh return con; 27803d14c5d2SYehuda Sadeh return NULL; 27813d14c5d2SYehuda Sadeh } 27823d14c5d2SYehuda Sadeh 27833d14c5d2SYehuda Sadeh static void put_osd_con(struct ceph_connection *con) 27843d14c5d2SYehuda Sadeh { 27853d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 27863d14c5d2SYehuda Sadeh put_osd(osd); 27873d14c5d2SYehuda Sadeh } 27883d14c5d2SYehuda Sadeh 27893d14c5d2SYehuda Sadeh /* 27903d14c5d2SYehuda Sadeh * authentication 27913d14c5d2SYehuda Sadeh */ 2792a3530df3SAlex Elder /* 2793a3530df3SAlex Elder * Note: returned pointer is the address of a structure that's 2794a3530df3SAlex Elder * managed separately. Caller must *not* attempt to free it. 2795a3530df3SAlex Elder */ 2796a3530df3SAlex Elder static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con, 27978f43fb53SAlex Elder int *proto, int force_new) 27983d14c5d2SYehuda Sadeh { 27993d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 28003d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 28013d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 280274f1869fSAlex Elder struct ceph_auth_handshake *auth = &o->o_auth; 28033d14c5d2SYehuda Sadeh 280474f1869fSAlex Elder if (force_new && auth->authorizer) { 280527859f97SSage Weil ceph_auth_destroy_authorizer(ac, auth->authorizer); 280674f1869fSAlex Elder auth->authorizer = NULL; 28073d14c5d2SYehuda Sadeh } 280827859f97SSage Weil if (!auth->authorizer) { 280927859f97SSage Weil int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 2810a3530df3SAlex Elder auth); 28113d14c5d2SYehuda Sadeh if (ret) 2812a3530df3SAlex Elder return ERR_PTR(ret); 281327859f97SSage Weil } else { 281427859f97SSage Weil int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 28150bed9b5cSSage Weil auth); 28160bed9b5cSSage Weil if (ret) 28170bed9b5cSSage Weil return ERR_PTR(ret); 28183d14c5d2SYehuda Sadeh } 28193d14c5d2SYehuda Sadeh *proto = ac->protocol; 282074f1869fSAlex Elder 2821a3530df3SAlex Elder return auth; 28223d14c5d2SYehuda Sadeh } 28233d14c5d2SYehuda Sadeh 28243d14c5d2SYehuda Sadeh 28253d14c5d2SYehuda Sadeh static int verify_authorizer_reply(struct ceph_connection *con, int len) 28263d14c5d2SYehuda Sadeh { 28273d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 28283d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 28293d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 28303d14c5d2SYehuda Sadeh 283127859f97SSage Weil return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len); 28323d14c5d2SYehuda Sadeh } 28333d14c5d2SYehuda Sadeh 28343d14c5d2SYehuda Sadeh static int invalidate_authorizer(struct ceph_connection *con) 28353d14c5d2SYehuda Sadeh { 28363d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 28373d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 28383d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 28393d14c5d2SYehuda Sadeh 284027859f97SSage Weil ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD); 28413d14c5d2SYehuda Sadeh return ceph_monc_validate_auth(&osdc->client->monc); 28423d14c5d2SYehuda Sadeh } 28433d14c5d2SYehuda Sadeh 28443d14c5d2SYehuda Sadeh static const struct ceph_connection_operations osd_con_ops = { 28453d14c5d2SYehuda Sadeh .get = get_osd_con, 28463d14c5d2SYehuda Sadeh .put = put_osd_con, 28473d14c5d2SYehuda Sadeh .dispatch = dispatch, 28483d14c5d2SYehuda Sadeh .get_authorizer = get_authorizer, 28493d14c5d2SYehuda Sadeh .verify_authorizer_reply = verify_authorizer_reply, 28503d14c5d2SYehuda Sadeh .invalidate_authorizer = invalidate_authorizer, 28513d14c5d2SYehuda Sadeh .alloc_msg = alloc_msg, 28523d14c5d2SYehuda Sadeh .fault = osd_reset, 28533d14c5d2SYehuda Sadeh }; 2854