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); 3671d0326b1SIlya Dryomov INIT_LIST_HEAD(&req->r_linger_osd_item); 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: 439c647b8a8SIlya Dryomov case CEPH_OSD_OP_SETALLOCHINT: 440a8dd0a37SAlex Elder case CEPH_OSD_OP_CLONERANGE: 441a8dd0a37SAlex Elder case CEPH_OSD_OP_ASSERT_SRC_VERSION: 442a8dd0a37SAlex Elder case CEPH_OSD_OP_SRC_CMPXATTR: 443a8dd0a37SAlex Elder case CEPH_OSD_OP_GETXATTR: 444a8dd0a37SAlex Elder case CEPH_OSD_OP_GETXATTRS: 445a8dd0a37SAlex Elder case CEPH_OSD_OP_CMPXATTR: 446a8dd0a37SAlex Elder case CEPH_OSD_OP_SETXATTR: 447a8dd0a37SAlex Elder case CEPH_OSD_OP_SETXATTRS: 448a8dd0a37SAlex Elder case CEPH_OSD_OP_RESETXATTRS: 449a8dd0a37SAlex Elder case CEPH_OSD_OP_RMXATTR: 450a8dd0a37SAlex Elder case CEPH_OSD_OP_PULL: 451a8dd0a37SAlex Elder case CEPH_OSD_OP_PUSH: 452a8dd0a37SAlex Elder case CEPH_OSD_OP_BALANCEREADS: 453a8dd0a37SAlex Elder case CEPH_OSD_OP_UNBALANCEREADS: 454a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB: 455a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB_RESERVE: 456a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB_UNRESERVE: 457a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB_STOP: 458a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB_MAP: 459a8dd0a37SAlex Elder case CEPH_OSD_OP_WRLOCK: 460a8dd0a37SAlex Elder case CEPH_OSD_OP_WRUNLOCK: 461a8dd0a37SAlex Elder case CEPH_OSD_OP_RDLOCK: 462a8dd0a37SAlex Elder case CEPH_OSD_OP_RDUNLOCK: 463a8dd0a37SAlex Elder case CEPH_OSD_OP_UPLOCK: 464a8dd0a37SAlex Elder case CEPH_OSD_OP_DNLOCK: 465a8dd0a37SAlex Elder case CEPH_OSD_OP_CALL: 466a8dd0a37SAlex Elder case CEPH_OSD_OP_PGLS: 467a8dd0a37SAlex Elder case CEPH_OSD_OP_PGLS_FILTER: 468a8dd0a37SAlex Elder return true; 469a8dd0a37SAlex Elder default: 470a8dd0a37SAlex Elder return false; 471a8dd0a37SAlex Elder } 472a8dd0a37SAlex Elder } 473a8dd0a37SAlex Elder 47433803f33SAlex Elder /* 47533803f33SAlex Elder * This is an osd op init function for opcodes that have no data or 47633803f33SAlex Elder * other information associated with them. It also serves as a 47733803f33SAlex Elder * common init routine for all the other init functions, below. 47833803f33SAlex Elder */ 479c99d2d4aSAlex Elder static struct ceph_osd_req_op * 48049719778SAlex Elder _osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which, 481c99d2d4aSAlex Elder u16 opcode) 48233803f33SAlex Elder { 483c99d2d4aSAlex Elder struct ceph_osd_req_op *op; 484c99d2d4aSAlex Elder 485c99d2d4aSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 48633803f33SAlex Elder BUG_ON(!osd_req_opcode_valid(opcode)); 48733803f33SAlex Elder 488c99d2d4aSAlex Elder op = &osd_req->r_ops[which]; 48933803f33SAlex Elder memset(op, 0, sizeof (*op)); 49033803f33SAlex Elder op->op = opcode; 491c99d2d4aSAlex Elder 492c99d2d4aSAlex Elder return op; 49333803f33SAlex Elder } 49433803f33SAlex Elder 49549719778SAlex Elder void osd_req_op_init(struct ceph_osd_request *osd_req, 49649719778SAlex Elder unsigned int which, u16 opcode) 49749719778SAlex Elder { 49849719778SAlex Elder (void)_osd_req_op_init(osd_req, which, opcode); 49949719778SAlex Elder } 50049719778SAlex Elder EXPORT_SYMBOL(osd_req_op_init); 50149719778SAlex Elder 502c99d2d4aSAlex Elder void osd_req_op_extent_init(struct ceph_osd_request *osd_req, 503c99d2d4aSAlex Elder unsigned int which, u16 opcode, 50433803f33SAlex Elder u64 offset, u64 length, 50533803f33SAlex Elder u64 truncate_size, u32 truncate_seq) 50633803f33SAlex Elder { 50749719778SAlex Elder struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode); 50833803f33SAlex Elder size_t payload_len = 0; 50933803f33SAlex Elder 510ad7a60deSLi Wang BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE && 511ad7a60deSLi Wang opcode != CEPH_OSD_OP_DELETE && opcode != CEPH_OSD_OP_ZERO && 512ad7a60deSLi Wang opcode != CEPH_OSD_OP_TRUNCATE); 51333803f33SAlex Elder 51433803f33SAlex Elder op->extent.offset = offset; 51533803f33SAlex Elder op->extent.length = length; 51633803f33SAlex Elder op->extent.truncate_size = truncate_size; 51733803f33SAlex Elder op->extent.truncate_seq = truncate_seq; 51833803f33SAlex Elder if (opcode == CEPH_OSD_OP_WRITE) 51933803f33SAlex Elder payload_len += length; 52033803f33SAlex Elder 52133803f33SAlex Elder op->payload_len = payload_len; 52233803f33SAlex Elder } 52333803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_init); 52433803f33SAlex Elder 525c99d2d4aSAlex Elder void osd_req_op_extent_update(struct ceph_osd_request *osd_req, 526c99d2d4aSAlex Elder unsigned int which, u64 length) 527e5975c7cSAlex Elder { 528c99d2d4aSAlex Elder struct ceph_osd_req_op *op; 529c99d2d4aSAlex Elder u64 previous; 530c99d2d4aSAlex Elder 531c99d2d4aSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 532c99d2d4aSAlex Elder op = &osd_req->r_ops[which]; 533c99d2d4aSAlex Elder previous = op->extent.length; 534e5975c7cSAlex Elder 535e5975c7cSAlex Elder if (length == previous) 536e5975c7cSAlex Elder return; /* Nothing to do */ 537e5975c7cSAlex Elder BUG_ON(length > previous); 538e5975c7cSAlex Elder 539e5975c7cSAlex Elder op->extent.length = length; 540e5975c7cSAlex Elder op->payload_len -= previous - length; 541e5975c7cSAlex Elder } 542e5975c7cSAlex Elder EXPORT_SYMBOL(osd_req_op_extent_update); 543e5975c7cSAlex Elder 544c99d2d4aSAlex Elder void osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which, 54504017e29SAlex Elder u16 opcode, const char *class, const char *method) 54633803f33SAlex Elder { 54749719778SAlex Elder struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode); 5485f562df5SAlex Elder struct ceph_pagelist *pagelist; 54933803f33SAlex Elder size_t payload_len = 0; 55033803f33SAlex Elder size_t size; 55133803f33SAlex Elder 55233803f33SAlex Elder BUG_ON(opcode != CEPH_OSD_OP_CALL); 55333803f33SAlex Elder 5545f562df5SAlex Elder pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS); 5555f562df5SAlex Elder BUG_ON(!pagelist); 5565f562df5SAlex Elder ceph_pagelist_init(pagelist); 5575f562df5SAlex Elder 55833803f33SAlex Elder op->cls.class_name = class; 55933803f33SAlex Elder size = strlen(class); 56033803f33SAlex Elder BUG_ON(size > (size_t) U8_MAX); 56133803f33SAlex Elder op->cls.class_len = size; 5625f562df5SAlex Elder ceph_pagelist_append(pagelist, class, size); 56333803f33SAlex Elder payload_len += size; 56433803f33SAlex Elder 56533803f33SAlex Elder op->cls.method_name = method; 56633803f33SAlex Elder size = strlen(method); 56733803f33SAlex Elder BUG_ON(size > (size_t) U8_MAX); 56833803f33SAlex Elder op->cls.method_len = size; 5695f562df5SAlex Elder ceph_pagelist_append(pagelist, method, size); 57033803f33SAlex Elder payload_len += size; 57133803f33SAlex Elder 572a4ce40a9SAlex Elder osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist); 5735f562df5SAlex Elder 57433803f33SAlex Elder op->cls.argc = 0; /* currently unused */ 57533803f33SAlex Elder 57633803f33SAlex Elder op->payload_len = payload_len; 57733803f33SAlex Elder } 57833803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_init); 5798c042b0dSAlex Elder 580c99d2d4aSAlex Elder void osd_req_op_watch_init(struct ceph_osd_request *osd_req, 581c99d2d4aSAlex Elder unsigned int which, u16 opcode, 58233803f33SAlex Elder u64 cookie, u64 version, int flag) 58333803f33SAlex Elder { 58449719778SAlex Elder struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode); 58533803f33SAlex Elder 586c99d2d4aSAlex Elder BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH); 58733803f33SAlex Elder 58833803f33SAlex Elder op->watch.cookie = cookie; 5899ef1ee5aSAlex Elder op->watch.ver = version; 59033803f33SAlex Elder if (opcode == CEPH_OSD_OP_WATCH && flag) 59133803f33SAlex Elder op->watch.flag = (u8)1; 59233803f33SAlex Elder } 59333803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_watch_init); 59433803f33SAlex Elder 595c647b8a8SIlya Dryomov void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req, 596c647b8a8SIlya Dryomov unsigned int which, 597c647b8a8SIlya Dryomov u64 expected_object_size, 598c647b8a8SIlya Dryomov u64 expected_write_size) 599c647b8a8SIlya Dryomov { 600c647b8a8SIlya Dryomov struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, 601c647b8a8SIlya Dryomov CEPH_OSD_OP_SETALLOCHINT); 602c647b8a8SIlya Dryomov 603c647b8a8SIlya Dryomov op->alloc_hint.expected_object_size = expected_object_size; 604c647b8a8SIlya Dryomov op->alloc_hint.expected_write_size = expected_write_size; 605c647b8a8SIlya Dryomov 606c647b8a8SIlya Dryomov /* 607c647b8a8SIlya Dryomov * CEPH_OSD_OP_SETALLOCHINT op is advisory and therefore deemed 608c647b8a8SIlya Dryomov * not worth a feature bit. Set FAILOK per-op flag to make 609c647b8a8SIlya Dryomov * sure older osds don't trip over an unsupported opcode. 610c647b8a8SIlya Dryomov */ 611c647b8a8SIlya Dryomov op->flags |= CEPH_OSD_OP_FLAG_FAILOK; 612c647b8a8SIlya Dryomov } 613c647b8a8SIlya Dryomov EXPORT_SYMBOL(osd_req_op_alloc_hint_init); 614c647b8a8SIlya Dryomov 61590af3602SAlex Elder static void ceph_osdc_msg_data_add(struct ceph_msg *msg, 616ec9123c5SAlex Elder struct ceph_osd_data *osd_data) 617ec9123c5SAlex Elder { 618ec9123c5SAlex Elder u64 length = ceph_osd_data_length(osd_data); 619ec9123c5SAlex Elder 620ec9123c5SAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 621ec9123c5SAlex Elder BUG_ON(length > (u64) SIZE_MAX); 622ec9123c5SAlex Elder if (length) 62390af3602SAlex Elder ceph_msg_data_add_pages(msg, osd_data->pages, 624ec9123c5SAlex Elder length, osd_data->alignment); 625ec9123c5SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) { 626ec9123c5SAlex Elder BUG_ON(!length); 62790af3602SAlex Elder ceph_msg_data_add_pagelist(msg, osd_data->pagelist); 628ec9123c5SAlex Elder #ifdef CONFIG_BLOCK 629ec9123c5SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) { 63090af3602SAlex Elder ceph_msg_data_add_bio(msg, osd_data->bio, length); 631ec9123c5SAlex Elder #endif 632ec9123c5SAlex Elder } else { 633ec9123c5SAlex Elder BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE); 634ec9123c5SAlex Elder } 635ec9123c5SAlex Elder } 636ec9123c5SAlex Elder 637175face2SAlex Elder static u64 osd_req_encode_op(struct ceph_osd_request *req, 63879528734SAlex Elder struct ceph_osd_op *dst, unsigned int which) 6393d14c5d2SYehuda Sadeh { 64079528734SAlex Elder struct ceph_osd_req_op *src; 64104017e29SAlex Elder struct ceph_osd_data *osd_data; 64254d50649SAlex Elder u64 request_data_len = 0; 64304017e29SAlex Elder u64 data_length; 644175face2SAlex Elder 64579528734SAlex Elder BUG_ON(which >= req->r_num_ops); 64679528734SAlex Elder src = &req->r_ops[which]; 647a8dd0a37SAlex Elder if (WARN_ON(!osd_req_opcode_valid(src->op))) { 648a8dd0a37SAlex Elder pr_err("unrecognized osd opcode %d\n", src->op); 649a8dd0a37SAlex Elder 650a8dd0a37SAlex Elder return 0; 651a8dd0a37SAlex Elder } 6523d14c5d2SYehuda Sadeh 653065a68f9SAlex Elder switch (src->op) { 654fbfab539SAlex Elder case CEPH_OSD_OP_STAT: 65549719778SAlex Elder osd_data = &src->raw_data_in; 65649719778SAlex Elder ceph_osdc_msg_data_add(req->r_reply, osd_data); 657fbfab539SAlex Elder break; 6583d14c5d2SYehuda Sadeh case CEPH_OSD_OP_READ: 6593d14c5d2SYehuda Sadeh case CEPH_OSD_OP_WRITE: 660ad7a60deSLi Wang case CEPH_OSD_OP_ZERO: 661ad7a60deSLi Wang case CEPH_OSD_OP_DELETE: 662ad7a60deSLi Wang case CEPH_OSD_OP_TRUNCATE: 663175face2SAlex Elder if (src->op == CEPH_OSD_OP_WRITE) 66454d50649SAlex Elder request_data_len = src->extent.length; 665175face2SAlex Elder dst->extent.offset = cpu_to_le64(src->extent.offset); 666175face2SAlex Elder dst->extent.length = cpu_to_le64(src->extent.length); 6673d14c5d2SYehuda Sadeh dst->extent.truncate_size = 6683d14c5d2SYehuda Sadeh cpu_to_le64(src->extent.truncate_size); 6693d14c5d2SYehuda Sadeh dst->extent.truncate_seq = 6703d14c5d2SYehuda Sadeh cpu_to_le32(src->extent.truncate_seq); 67104017e29SAlex Elder osd_data = &src->extent.osd_data; 6725476492fSAlex Elder if (src->op == CEPH_OSD_OP_WRITE) 67304017e29SAlex Elder ceph_osdc_msg_data_add(req->r_request, osd_data); 6745476492fSAlex Elder else 67504017e29SAlex Elder ceph_osdc_msg_data_add(req->r_reply, osd_data); 6763d14c5d2SYehuda Sadeh break; 6773d14c5d2SYehuda Sadeh case CEPH_OSD_OP_CALL: 6783d14c5d2SYehuda Sadeh dst->cls.class_len = src->cls.class_len; 6793d14c5d2SYehuda Sadeh dst->cls.method_len = src->cls.method_len; 68004017e29SAlex Elder osd_data = &src->cls.request_info; 68104017e29SAlex Elder ceph_osdc_msg_data_add(req->r_request, osd_data); 68204017e29SAlex Elder BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGELIST); 68304017e29SAlex Elder request_data_len = osd_data->pagelist->length; 68404017e29SAlex Elder 68504017e29SAlex Elder osd_data = &src->cls.request_data; 68604017e29SAlex Elder data_length = ceph_osd_data_length(osd_data); 68704017e29SAlex Elder if (data_length) { 68804017e29SAlex Elder BUG_ON(osd_data->type == CEPH_OSD_DATA_TYPE_NONE); 68904017e29SAlex Elder dst->cls.indata_len = cpu_to_le32(data_length); 69004017e29SAlex Elder ceph_osdc_msg_data_add(req->r_request, osd_data); 69104017e29SAlex Elder src->payload_len += data_length; 69204017e29SAlex Elder request_data_len += data_length; 69304017e29SAlex Elder } 69404017e29SAlex Elder osd_data = &src->cls.response_data; 69504017e29SAlex Elder ceph_osdc_msg_data_add(req->r_reply, osd_data); 6963d14c5d2SYehuda Sadeh break; 6973d14c5d2SYehuda Sadeh case CEPH_OSD_OP_STARTSYNC: 6983d14c5d2SYehuda Sadeh break; 699a40c4f10SYehuda Sadeh case CEPH_OSD_OP_NOTIFY_ACK: 700a40c4f10SYehuda Sadeh case CEPH_OSD_OP_WATCH: 701a40c4f10SYehuda Sadeh dst->watch.cookie = cpu_to_le64(src->watch.cookie); 702a40c4f10SYehuda Sadeh dst->watch.ver = cpu_to_le64(src->watch.ver); 703a40c4f10SYehuda Sadeh dst->watch.flag = src->watch.flag; 704a40c4f10SYehuda Sadeh break; 705c647b8a8SIlya Dryomov case CEPH_OSD_OP_SETALLOCHINT: 706c647b8a8SIlya Dryomov dst->alloc_hint.expected_object_size = 707c647b8a8SIlya Dryomov cpu_to_le64(src->alloc_hint.expected_object_size); 708c647b8a8SIlya Dryomov dst->alloc_hint.expected_write_size = 709c647b8a8SIlya Dryomov cpu_to_le64(src->alloc_hint.expected_write_size); 710c647b8a8SIlya Dryomov break; 7113d14c5d2SYehuda Sadeh default: 7124c46459cSAlex Elder pr_err("unsupported osd opcode %s\n", 7138f63ca2dSAlex Elder ceph_osd_op_name(src->op)); 7144c46459cSAlex Elder WARN_ON(1); 715a8dd0a37SAlex Elder 716a8dd0a37SAlex Elder return 0; 7173d14c5d2SYehuda Sadeh } 7187b25bf5fSIlya Dryomov 719a8dd0a37SAlex Elder dst->op = cpu_to_le16(src->op); 7207b25bf5fSIlya Dryomov dst->flags = cpu_to_le32(src->flags); 7213d14c5d2SYehuda Sadeh dst->payload_len = cpu_to_le32(src->payload_len); 722175face2SAlex Elder 72354d50649SAlex Elder return request_data_len; 7243d14c5d2SYehuda Sadeh } 7253d14c5d2SYehuda Sadeh 7263d14c5d2SYehuda Sadeh /* 7273d14c5d2SYehuda Sadeh * build new request AND message, calculate layout, and adjust file 7283d14c5d2SYehuda Sadeh * extent as needed. 7293d14c5d2SYehuda Sadeh * 7303d14c5d2SYehuda Sadeh * if the file was recently truncated, we include information about its 7313d14c5d2SYehuda Sadeh * old and new size so that the object can be updated appropriately. (we 7323d14c5d2SYehuda Sadeh * avoid synchronously deleting truncated objects because it's slow.) 7333d14c5d2SYehuda Sadeh * 7343d14c5d2SYehuda Sadeh * if @do_sync, include a 'startsync' command so that the osd will flush 7353d14c5d2SYehuda Sadeh * data quickly. 7363d14c5d2SYehuda Sadeh */ 7373d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc, 7383d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 7393d14c5d2SYehuda Sadeh struct ceph_vino vino, 740acead002SAlex Elder u64 off, u64 *plen, int num_ops, 7413d14c5d2SYehuda Sadeh int opcode, int flags, 7423d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 7433d14c5d2SYehuda Sadeh u32 truncate_seq, 7443d14c5d2SYehuda Sadeh u64 truncate_size, 745153e5167SAlex Elder bool use_mempool) 7463d14c5d2SYehuda Sadeh { 7473d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 74875d1c941SAlex Elder u64 objnum = 0; 74975d1c941SAlex Elder u64 objoff = 0; 75075d1c941SAlex Elder u64 objlen = 0; 751d18d1e28SAlex Elder u32 object_size; 752d18d1e28SAlex Elder u64 object_base; 7536816282dSSage Weil int r; 7543d14c5d2SYehuda Sadeh 755ad7a60deSLi Wang BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE && 756ad7a60deSLi Wang opcode != CEPH_OSD_OP_DELETE && opcode != CEPH_OSD_OP_ZERO && 757ad7a60deSLi Wang opcode != CEPH_OSD_OP_TRUNCATE); 7583d14c5d2SYehuda Sadeh 759acead002SAlex Elder req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool, 760ae7ca4a3SAlex Elder GFP_NOFS); 7614ad12621SSage Weil if (!req) 7626816282dSSage Weil return ERR_PTR(-ENOMEM); 76379528734SAlex Elder 764d178a9e7SAlex Elder req->r_flags = flags; 7653d14c5d2SYehuda Sadeh 7663d14c5d2SYehuda Sadeh /* calculate max write size */ 767a19dadfbSAlex Elder r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen); 7683ff5f385SAlex Elder if (r < 0) { 7693ff5f385SAlex Elder ceph_osdc_put_request(req); 7706816282dSSage Weil return ERR_PTR(r); 7713ff5f385SAlex Elder } 772a19dadfbSAlex Elder 773d18d1e28SAlex Elder object_size = le32_to_cpu(layout->fl_object_size); 774d18d1e28SAlex Elder object_base = off - objoff; 775ccca4e37SYan, Zheng if (!(truncate_seq == 1 && truncate_size == -1ULL)) { 776d18d1e28SAlex Elder if (truncate_size <= object_base) { 777d18d1e28SAlex Elder truncate_size = 0; 778d18d1e28SAlex Elder } else { 779d18d1e28SAlex Elder truncate_size -= object_base; 780d18d1e28SAlex Elder if (truncate_size > object_size) 781d18d1e28SAlex Elder truncate_size = object_size; 782d18d1e28SAlex Elder } 783ccca4e37SYan, Zheng } 784d18d1e28SAlex Elder 785c99d2d4aSAlex Elder osd_req_op_extent_init(req, 0, opcode, objoff, objlen, 786b0270324SAlex Elder truncate_size, truncate_seq); 7878c042b0dSAlex Elder 788acead002SAlex Elder /* 789acead002SAlex Elder * A second op in the ops array means the caller wants to 790acead002SAlex Elder * also issue a include a 'startsync' command so that the 791acead002SAlex Elder * osd will flush data quickly. 792acead002SAlex Elder */ 793acead002SAlex Elder if (num_ops > 1) 794c99d2d4aSAlex Elder osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC); 795d18d1e28SAlex Elder 7963c972c95SIlya Dryomov req->r_base_oloc.pool = ceph_file_layout_pg_pool(*layout); 7973d14c5d2SYehuda Sadeh 7983c972c95SIlya Dryomov snprintf(req->r_base_oid.name, sizeof(req->r_base_oid.name), 7994295f221SIlya Dryomov "%llx.%08llx", vino.ino, objnum); 8003c972c95SIlya Dryomov req->r_base_oid.name_len = strlen(req->r_base_oid.name); 801dbe0fc41SAlex Elder 8023d14c5d2SYehuda Sadeh return req; 8033d14c5d2SYehuda Sadeh } 8043d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_new_request); 8053d14c5d2SYehuda Sadeh 8063d14c5d2SYehuda Sadeh /* 8073d14c5d2SYehuda Sadeh * We keep osd requests in an rbtree, sorted by ->r_tid. 8083d14c5d2SYehuda Sadeh */ 8093d14c5d2SYehuda Sadeh static void __insert_request(struct ceph_osd_client *osdc, 8103d14c5d2SYehuda Sadeh struct ceph_osd_request *new) 8113d14c5d2SYehuda Sadeh { 8123d14c5d2SYehuda Sadeh struct rb_node **p = &osdc->requests.rb_node; 8133d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 8143d14c5d2SYehuda Sadeh struct ceph_osd_request *req = NULL; 8153d14c5d2SYehuda Sadeh 8163d14c5d2SYehuda Sadeh while (*p) { 8173d14c5d2SYehuda Sadeh parent = *p; 8183d14c5d2SYehuda Sadeh req = rb_entry(parent, struct ceph_osd_request, r_node); 8193d14c5d2SYehuda Sadeh if (new->r_tid < req->r_tid) 8203d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 8213d14c5d2SYehuda Sadeh else if (new->r_tid > req->r_tid) 8223d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 8233d14c5d2SYehuda Sadeh else 8243d14c5d2SYehuda Sadeh BUG(); 8253d14c5d2SYehuda Sadeh } 8263d14c5d2SYehuda Sadeh 8273d14c5d2SYehuda Sadeh rb_link_node(&new->r_node, parent, p); 8283d14c5d2SYehuda Sadeh rb_insert_color(&new->r_node, &osdc->requests); 8293d14c5d2SYehuda Sadeh } 8303d14c5d2SYehuda Sadeh 8313d14c5d2SYehuda Sadeh static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc, 8323d14c5d2SYehuda Sadeh u64 tid) 8333d14c5d2SYehuda Sadeh { 8343d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 8353d14c5d2SYehuda Sadeh struct rb_node *n = osdc->requests.rb_node; 8363d14c5d2SYehuda Sadeh 8373d14c5d2SYehuda Sadeh while (n) { 8383d14c5d2SYehuda Sadeh req = rb_entry(n, struct ceph_osd_request, r_node); 8393d14c5d2SYehuda Sadeh if (tid < req->r_tid) 8403d14c5d2SYehuda Sadeh n = n->rb_left; 8413d14c5d2SYehuda Sadeh else if (tid > req->r_tid) 8423d14c5d2SYehuda Sadeh n = n->rb_right; 8433d14c5d2SYehuda Sadeh else 8443d14c5d2SYehuda Sadeh return req; 8453d14c5d2SYehuda Sadeh } 8463d14c5d2SYehuda Sadeh return NULL; 8473d14c5d2SYehuda Sadeh } 8483d14c5d2SYehuda Sadeh 8493d14c5d2SYehuda Sadeh static struct ceph_osd_request * 8503d14c5d2SYehuda Sadeh __lookup_request_ge(struct ceph_osd_client *osdc, 8513d14c5d2SYehuda Sadeh u64 tid) 8523d14c5d2SYehuda Sadeh { 8533d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 8543d14c5d2SYehuda Sadeh struct rb_node *n = osdc->requests.rb_node; 8553d14c5d2SYehuda Sadeh 8563d14c5d2SYehuda Sadeh while (n) { 8573d14c5d2SYehuda Sadeh req = rb_entry(n, struct ceph_osd_request, r_node); 8583d14c5d2SYehuda Sadeh if (tid < req->r_tid) { 8593d14c5d2SYehuda Sadeh if (!n->rb_left) 8603d14c5d2SYehuda Sadeh return req; 8613d14c5d2SYehuda Sadeh n = n->rb_left; 8623d14c5d2SYehuda Sadeh } else if (tid > req->r_tid) { 8633d14c5d2SYehuda Sadeh n = n->rb_right; 8643d14c5d2SYehuda Sadeh } else { 8653d14c5d2SYehuda Sadeh return req; 8663d14c5d2SYehuda Sadeh } 8673d14c5d2SYehuda Sadeh } 8683d14c5d2SYehuda Sadeh return NULL; 8693d14c5d2SYehuda Sadeh } 8703d14c5d2SYehuda Sadeh 8716f6c7006SSage Weil /* 8726f6c7006SSage Weil * Resubmit requests pending on the given osd. 8736f6c7006SSage Weil */ 8746f6c7006SSage Weil static void __kick_osd_requests(struct ceph_osd_client *osdc, 8756f6c7006SSage Weil struct ceph_osd *osd) 8766f6c7006SSage Weil { 877a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 878e02493c0SAlex Elder LIST_HEAD(resend); 8796f6c7006SSage Weil int err; 8806f6c7006SSage Weil 8816f6c7006SSage Weil dout("__kick_osd_requests osd%d\n", osd->o_osd); 8826f6c7006SSage Weil err = __reset_osd(osdc, osd); 883685a7555SAlex Elder if (err) 8846f6c7006SSage Weil return; 885e02493c0SAlex Elder /* 886e02493c0SAlex Elder * Build up a list of requests to resend by traversing the 887e02493c0SAlex Elder * osd's list of requests. Requests for a given object are 888e02493c0SAlex Elder * sent in tid order, and that is also the order they're 889e02493c0SAlex Elder * kept on this list. Therefore all requests that are in 890e02493c0SAlex Elder * flight will be found first, followed by all requests that 891e02493c0SAlex Elder * have not yet been sent. And to resend requests while 892e02493c0SAlex Elder * preserving this order we will want to put any sent 893e02493c0SAlex Elder * requests back on the front of the osd client's unsent 894e02493c0SAlex Elder * list. 895e02493c0SAlex Elder * 896e02493c0SAlex Elder * So we build a separate ordered list of already-sent 897e02493c0SAlex Elder * requests for the affected osd and splice it onto the 898e02493c0SAlex Elder * front of the osd client's unsent list. Once we've seen a 899e02493c0SAlex Elder * request that has not yet been sent we're done. Those 900e02493c0SAlex Elder * requests are already sitting right where they belong. 901e02493c0SAlex Elder */ 9026f6c7006SSage Weil list_for_each_entry(req, &osd->o_requests, r_osd_item) { 903e02493c0SAlex Elder if (!req->r_sent) 904e02493c0SAlex Elder break; 905e02493c0SAlex Elder list_move_tail(&req->r_req_lru_item, &resend); 906e02493c0SAlex Elder dout("requeueing %p tid %llu osd%d\n", req, req->r_tid, 9076f6c7006SSage Weil osd->o_osd); 908a40c4f10SYehuda Sadeh if (!req->r_linger) 9096f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 9106f6c7006SSage Weil } 911e02493c0SAlex Elder list_splice(&resend, &osdc->req_unsent); 912a40c4f10SYehuda Sadeh 913e02493c0SAlex Elder /* 914e02493c0SAlex Elder * Linger requests are re-registered before sending, which 915e02493c0SAlex Elder * sets up a new tid for each. We add them to the unsent 916e02493c0SAlex Elder * list at the end to keep things in tid order. 917e02493c0SAlex Elder */ 918a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osd->o_linger_requests, 9191d0326b1SIlya Dryomov r_linger_osd_item) { 92077f38e0eSSage Weil /* 92177f38e0eSSage Weil * reregister request prior to unregistering linger so 92277f38e0eSSage Weil * that r_osd is preserved. 92377f38e0eSSage Weil */ 92477f38e0eSSage Weil BUG_ON(!list_empty(&req->r_req_lru_item)); 925a40c4f10SYehuda Sadeh __register_request(osdc, req); 926e02493c0SAlex Elder list_add_tail(&req->r_req_lru_item, &osdc->req_unsent); 927ad885927SAlex Elder list_add_tail(&req->r_osd_item, &req->r_osd->o_requests); 92877f38e0eSSage Weil __unregister_linger_request(osdc, req); 929a40c4f10SYehuda Sadeh dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid, 930a40c4f10SYehuda Sadeh osd->o_osd); 931a40c4f10SYehuda Sadeh } 9326f6c7006SSage Weil } 9336f6c7006SSage Weil 9343d14c5d2SYehuda Sadeh /* 9353d14c5d2SYehuda Sadeh * If the osd connection drops, we need to resubmit all requests. 9363d14c5d2SYehuda Sadeh */ 9373d14c5d2SYehuda Sadeh static void osd_reset(struct ceph_connection *con) 9383d14c5d2SYehuda Sadeh { 9393d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 9403d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 9413d14c5d2SYehuda Sadeh 9423d14c5d2SYehuda Sadeh if (!osd) 9433d14c5d2SYehuda Sadeh return; 9443d14c5d2SYehuda Sadeh dout("osd_reset osd%d\n", osd->o_osd); 9453d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 9463d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 94783aff95eSSage Weil mutex_lock(&osdc->request_mutex); 94883aff95eSSage Weil __kick_osd_requests(osdc, osd); 949f9d25199SAlex Elder __send_queued(osdc); 95083aff95eSSage Weil mutex_unlock(&osdc->request_mutex); 9513d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 9523d14c5d2SYehuda Sadeh } 9533d14c5d2SYehuda Sadeh 9543d14c5d2SYehuda Sadeh /* 9553d14c5d2SYehuda Sadeh * Track open sessions with osds. 9563d14c5d2SYehuda Sadeh */ 957e10006f8SAlex Elder static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum) 9583d14c5d2SYehuda Sadeh { 9593d14c5d2SYehuda Sadeh struct ceph_osd *osd; 9603d14c5d2SYehuda Sadeh 9613d14c5d2SYehuda Sadeh osd = kzalloc(sizeof(*osd), GFP_NOFS); 9623d14c5d2SYehuda Sadeh if (!osd) 9633d14c5d2SYehuda Sadeh return NULL; 9643d14c5d2SYehuda Sadeh 9653d14c5d2SYehuda Sadeh atomic_set(&osd->o_ref, 1); 9663d14c5d2SYehuda Sadeh osd->o_osdc = osdc; 967e10006f8SAlex Elder osd->o_osd = onum; 968f407731dSAlex Elder RB_CLEAR_NODE(&osd->o_node); 9693d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_requests); 970a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osd->o_linger_requests); 9713d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_osd_lru); 9723d14c5d2SYehuda Sadeh osd->o_incarnation = 1; 9733d14c5d2SYehuda Sadeh 974b7a9e5ddSSage Weil ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr); 9753d14c5d2SYehuda Sadeh 9763d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_keepalive_item); 9773d14c5d2SYehuda Sadeh return osd; 9783d14c5d2SYehuda Sadeh } 9793d14c5d2SYehuda Sadeh 9803d14c5d2SYehuda Sadeh static struct ceph_osd *get_osd(struct ceph_osd *osd) 9813d14c5d2SYehuda Sadeh { 9823d14c5d2SYehuda Sadeh if (atomic_inc_not_zero(&osd->o_ref)) { 9833d14c5d2SYehuda Sadeh dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1, 9843d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref)); 9853d14c5d2SYehuda Sadeh return osd; 9863d14c5d2SYehuda Sadeh } else { 9873d14c5d2SYehuda Sadeh dout("get_osd %p FAIL\n", osd); 9883d14c5d2SYehuda Sadeh return NULL; 9893d14c5d2SYehuda Sadeh } 9903d14c5d2SYehuda Sadeh } 9913d14c5d2SYehuda Sadeh 9923d14c5d2SYehuda Sadeh static void put_osd(struct ceph_osd *osd) 9933d14c5d2SYehuda Sadeh { 9943d14c5d2SYehuda Sadeh dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref), 9953d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref) - 1); 996a255651dSAlex Elder if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) { 9973d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth; 9983d14c5d2SYehuda Sadeh 99927859f97SSage Weil ceph_auth_destroy_authorizer(ac, osd->o_auth.authorizer); 10003d14c5d2SYehuda Sadeh kfree(osd); 10013d14c5d2SYehuda Sadeh } 10023d14c5d2SYehuda Sadeh } 10033d14c5d2SYehuda Sadeh 10043d14c5d2SYehuda Sadeh /* 10053d14c5d2SYehuda Sadeh * remove an osd from our map 10063d14c5d2SYehuda Sadeh */ 10073d14c5d2SYehuda Sadeh static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 10083d14c5d2SYehuda Sadeh { 10093d14c5d2SYehuda Sadeh dout("__remove_osd %p\n", osd); 10103d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_requests)); 10113d14c5d2SYehuda Sadeh rb_erase(&osd->o_node, &osdc->osds); 10123d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 10133d14c5d2SYehuda Sadeh ceph_con_close(&osd->o_con); 10143d14c5d2SYehuda Sadeh put_osd(osd); 10153d14c5d2SYehuda Sadeh } 10163d14c5d2SYehuda Sadeh 1017aca420bcSSage Weil static void remove_all_osds(struct ceph_osd_client *osdc) 1018aca420bcSSage Weil { 1019048a9d2dSJiaju Zhang dout("%s %p\n", __func__, osdc); 1020aca420bcSSage Weil mutex_lock(&osdc->request_mutex); 1021aca420bcSSage Weil while (!RB_EMPTY_ROOT(&osdc->osds)) { 1022aca420bcSSage Weil struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds), 1023aca420bcSSage Weil struct ceph_osd, o_node); 1024aca420bcSSage Weil __remove_osd(osdc, osd); 1025aca420bcSSage Weil } 1026aca420bcSSage Weil mutex_unlock(&osdc->request_mutex); 1027aca420bcSSage Weil } 1028aca420bcSSage Weil 10293d14c5d2SYehuda Sadeh static void __move_osd_to_lru(struct ceph_osd_client *osdc, 10303d14c5d2SYehuda Sadeh struct ceph_osd *osd) 10313d14c5d2SYehuda Sadeh { 1032bbf37ec3SIlya Dryomov dout("%s %p\n", __func__, osd); 10333d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_osd_lru)); 1034bbf37ec3SIlya Dryomov 10353d14c5d2SYehuda Sadeh list_add_tail(&osd->o_osd_lru, &osdc->osd_lru); 10363d14c5d2SYehuda Sadeh osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ; 10373d14c5d2SYehuda Sadeh } 10383d14c5d2SYehuda Sadeh 1039bbf37ec3SIlya Dryomov static void maybe_move_osd_to_lru(struct ceph_osd_client *osdc, 1040bbf37ec3SIlya Dryomov struct ceph_osd *osd) 1041bbf37ec3SIlya Dryomov { 1042bbf37ec3SIlya Dryomov dout("%s %p\n", __func__, osd); 1043bbf37ec3SIlya Dryomov 1044bbf37ec3SIlya Dryomov if (list_empty(&osd->o_requests) && 1045bbf37ec3SIlya Dryomov list_empty(&osd->o_linger_requests)) 1046bbf37ec3SIlya Dryomov __move_osd_to_lru(osdc, osd); 1047bbf37ec3SIlya Dryomov } 1048bbf37ec3SIlya Dryomov 10493d14c5d2SYehuda Sadeh static void __remove_osd_from_lru(struct ceph_osd *osd) 10503d14c5d2SYehuda Sadeh { 10513d14c5d2SYehuda Sadeh dout("__remove_osd_from_lru %p\n", osd); 10523d14c5d2SYehuda Sadeh if (!list_empty(&osd->o_osd_lru)) 10533d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 10543d14c5d2SYehuda Sadeh } 10553d14c5d2SYehuda Sadeh 1056aca420bcSSage Weil static void remove_old_osds(struct ceph_osd_client *osdc) 10573d14c5d2SYehuda Sadeh { 10583d14c5d2SYehuda Sadeh struct ceph_osd *osd, *nosd; 10593d14c5d2SYehuda Sadeh 10603d14c5d2SYehuda Sadeh dout("__remove_old_osds %p\n", osdc); 10613d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 10623d14c5d2SYehuda Sadeh list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) { 1063aca420bcSSage Weil if (time_before(jiffies, osd->lru_ttl)) 10643d14c5d2SYehuda Sadeh break; 10653d14c5d2SYehuda Sadeh __remove_osd(osdc, osd); 10663d14c5d2SYehuda Sadeh } 10673d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 10683d14c5d2SYehuda Sadeh } 10693d14c5d2SYehuda Sadeh 10703d14c5d2SYehuda Sadeh /* 10713d14c5d2SYehuda Sadeh * reset osd connect 10723d14c5d2SYehuda Sadeh */ 10733d14c5d2SYehuda Sadeh static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 10743d14c5d2SYehuda Sadeh { 1075c3acb181SAlex Elder struct ceph_entity_addr *peer_addr; 10763d14c5d2SYehuda Sadeh 10773d14c5d2SYehuda Sadeh dout("__reset_osd %p osd%d\n", osd, osd->o_osd); 1078a40c4f10SYehuda Sadeh if (list_empty(&osd->o_requests) && 1079a40c4f10SYehuda Sadeh list_empty(&osd->o_linger_requests)) { 10803d14c5d2SYehuda Sadeh __remove_osd(osdc, osd); 1081c3acb181SAlex Elder 1082c3acb181SAlex Elder return -ENODEV; 1083c3acb181SAlex Elder } 1084c3acb181SAlex Elder 1085c3acb181SAlex Elder peer_addr = &osdc->osdmap->osd_addr[osd->o_osd]; 1086c3acb181SAlex Elder if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) && 10873d14c5d2SYehuda Sadeh !ceph_con_opened(&osd->o_con)) { 1088c3acb181SAlex Elder struct ceph_osd_request *req; 1089c3acb181SAlex Elder 10903d14c5d2SYehuda Sadeh dout("osd addr hasn't changed and connection never opened, " 10910b4af2e8SIlya Dryomov "letting msgr retry\n"); 10923d14c5d2SYehuda Sadeh /* touch each r_stamp for handle_timeout()'s benfit */ 10933d14c5d2SYehuda Sadeh list_for_each_entry(req, &osd->o_requests, r_osd_item) 10943d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 1095c3acb181SAlex Elder 1096c3acb181SAlex Elder return -EAGAIN; 10973d14c5d2SYehuda Sadeh } 1098c3acb181SAlex Elder 1099c3acb181SAlex Elder ceph_con_close(&osd->o_con); 1100c3acb181SAlex Elder ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr); 1101c3acb181SAlex Elder osd->o_incarnation++; 1102c3acb181SAlex Elder 1103c3acb181SAlex Elder return 0; 11043d14c5d2SYehuda Sadeh } 11053d14c5d2SYehuda Sadeh 11063d14c5d2SYehuda Sadeh static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new) 11073d14c5d2SYehuda Sadeh { 11083d14c5d2SYehuda Sadeh struct rb_node **p = &osdc->osds.rb_node; 11093d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 11103d14c5d2SYehuda Sadeh struct ceph_osd *osd = NULL; 11113d14c5d2SYehuda Sadeh 1112aca420bcSSage Weil dout("__insert_osd %p osd%d\n", new, new->o_osd); 11133d14c5d2SYehuda Sadeh while (*p) { 11143d14c5d2SYehuda Sadeh parent = *p; 11153d14c5d2SYehuda Sadeh osd = rb_entry(parent, struct ceph_osd, o_node); 11163d14c5d2SYehuda Sadeh if (new->o_osd < osd->o_osd) 11173d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 11183d14c5d2SYehuda Sadeh else if (new->o_osd > osd->o_osd) 11193d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 11203d14c5d2SYehuda Sadeh else 11213d14c5d2SYehuda Sadeh BUG(); 11223d14c5d2SYehuda Sadeh } 11233d14c5d2SYehuda Sadeh 11243d14c5d2SYehuda Sadeh rb_link_node(&new->o_node, parent, p); 11253d14c5d2SYehuda Sadeh rb_insert_color(&new->o_node, &osdc->osds); 11263d14c5d2SYehuda Sadeh } 11273d14c5d2SYehuda Sadeh 11283d14c5d2SYehuda Sadeh static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o) 11293d14c5d2SYehuda Sadeh { 11303d14c5d2SYehuda Sadeh struct ceph_osd *osd; 11313d14c5d2SYehuda Sadeh struct rb_node *n = osdc->osds.rb_node; 11323d14c5d2SYehuda Sadeh 11333d14c5d2SYehuda Sadeh while (n) { 11343d14c5d2SYehuda Sadeh osd = rb_entry(n, struct ceph_osd, o_node); 11353d14c5d2SYehuda Sadeh if (o < osd->o_osd) 11363d14c5d2SYehuda Sadeh n = n->rb_left; 11373d14c5d2SYehuda Sadeh else if (o > osd->o_osd) 11383d14c5d2SYehuda Sadeh n = n->rb_right; 11393d14c5d2SYehuda Sadeh else 11403d14c5d2SYehuda Sadeh return osd; 11413d14c5d2SYehuda Sadeh } 11423d14c5d2SYehuda Sadeh return NULL; 11433d14c5d2SYehuda Sadeh } 11443d14c5d2SYehuda Sadeh 11453d14c5d2SYehuda Sadeh static void __schedule_osd_timeout(struct ceph_osd_client *osdc) 11463d14c5d2SYehuda Sadeh { 11473d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->timeout_work, 11483d14c5d2SYehuda Sadeh osdc->client->options->osd_keepalive_timeout * HZ); 11493d14c5d2SYehuda Sadeh } 11503d14c5d2SYehuda Sadeh 11513d14c5d2SYehuda Sadeh static void __cancel_osd_timeout(struct ceph_osd_client *osdc) 11523d14c5d2SYehuda Sadeh { 11533d14c5d2SYehuda Sadeh cancel_delayed_work(&osdc->timeout_work); 11543d14c5d2SYehuda Sadeh } 11553d14c5d2SYehuda Sadeh 11563d14c5d2SYehuda Sadeh /* 11573d14c5d2SYehuda Sadeh * Register request, assign tid. If this is the first request, set up 11583d14c5d2SYehuda Sadeh * the timeout event. 11593d14c5d2SYehuda Sadeh */ 1160a40c4f10SYehuda Sadeh static void __register_request(struct ceph_osd_client *osdc, 11613d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 11623d14c5d2SYehuda Sadeh { 11633d14c5d2SYehuda Sadeh req->r_tid = ++osdc->last_tid; 11643d14c5d2SYehuda Sadeh req->r_request->hdr.tid = cpu_to_le64(req->r_tid); 116577f38e0eSSage Weil dout("__register_request %p tid %lld\n", req, req->r_tid); 11663d14c5d2SYehuda Sadeh __insert_request(osdc, req); 11673d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 11683d14c5d2SYehuda Sadeh osdc->num_requests++; 11693d14c5d2SYehuda Sadeh if (osdc->num_requests == 1) { 11703d14c5d2SYehuda Sadeh dout(" first request, scheduling timeout\n"); 11713d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 11723d14c5d2SYehuda Sadeh } 1173a40c4f10SYehuda Sadeh } 1174a40c4f10SYehuda Sadeh 11753d14c5d2SYehuda Sadeh /* 11763d14c5d2SYehuda Sadeh * called under osdc->request_mutex 11773d14c5d2SYehuda Sadeh */ 11783d14c5d2SYehuda Sadeh static void __unregister_request(struct ceph_osd_client *osdc, 11793d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 11803d14c5d2SYehuda Sadeh { 118135f9f8a0SSage Weil if (RB_EMPTY_NODE(&req->r_node)) { 118235f9f8a0SSage Weil dout("__unregister_request %p tid %lld not registered\n", 118335f9f8a0SSage Weil req, req->r_tid); 118435f9f8a0SSage Weil return; 118535f9f8a0SSage Weil } 118635f9f8a0SSage Weil 11873d14c5d2SYehuda Sadeh dout("__unregister_request %p tid %lld\n", req, req->r_tid); 11883d14c5d2SYehuda Sadeh rb_erase(&req->r_node, &osdc->requests); 11893d14c5d2SYehuda Sadeh osdc->num_requests--; 11903d14c5d2SYehuda Sadeh 11913d14c5d2SYehuda Sadeh if (req->r_osd) { 11923d14c5d2SYehuda Sadeh /* make sure the original request isn't in flight. */ 11936740a845SAlex Elder ceph_msg_revoke(req->r_request); 11943d14c5d2SYehuda Sadeh 11953d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 1196bbf37ec3SIlya Dryomov maybe_move_osd_to_lru(osdc, req->r_osd); 1197fbdb9190SSage Weil if (list_empty(&req->r_linger_item)) 11983d14c5d2SYehuda Sadeh req->r_osd = NULL; 11993d14c5d2SYehuda Sadeh } 12003d14c5d2SYehuda Sadeh 12017d5f2481SAlex Elder list_del_init(&req->r_req_lru_item); 12023d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 12033d14c5d2SYehuda Sadeh 12043d14c5d2SYehuda Sadeh if (osdc->num_requests == 0) { 12053d14c5d2SYehuda Sadeh dout(" no requests, canceling timeout\n"); 12063d14c5d2SYehuda Sadeh __cancel_osd_timeout(osdc); 12073d14c5d2SYehuda Sadeh } 12083d14c5d2SYehuda Sadeh } 12093d14c5d2SYehuda Sadeh 12103d14c5d2SYehuda Sadeh /* 12113d14c5d2SYehuda Sadeh * Cancel a previously queued request message 12123d14c5d2SYehuda Sadeh */ 12133d14c5d2SYehuda Sadeh static void __cancel_request(struct ceph_osd_request *req) 12143d14c5d2SYehuda Sadeh { 12153d14c5d2SYehuda Sadeh if (req->r_sent && req->r_osd) { 12166740a845SAlex Elder ceph_msg_revoke(req->r_request); 12173d14c5d2SYehuda Sadeh req->r_sent = 0; 12183d14c5d2SYehuda Sadeh } 12193d14c5d2SYehuda Sadeh } 12203d14c5d2SYehuda Sadeh 1221a40c4f10SYehuda Sadeh static void __register_linger_request(struct ceph_osd_client *osdc, 1222a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1223a40c4f10SYehuda Sadeh { 1224a40c4f10SYehuda Sadeh dout("__register_linger_request %p\n", req); 122596e4dac6SAlex Elder ceph_osdc_get_request(req); 1226a40c4f10SYehuda Sadeh list_add_tail(&req->r_linger_item, &osdc->req_linger); 12276194ea89SSage Weil if (req->r_osd) 12281d0326b1SIlya Dryomov list_add_tail(&req->r_linger_osd_item, 12296194ea89SSage Weil &req->r_osd->o_linger_requests); 1230a40c4f10SYehuda Sadeh } 1231a40c4f10SYehuda Sadeh 1232a40c4f10SYehuda Sadeh static void __unregister_linger_request(struct ceph_osd_client *osdc, 1233a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1234a40c4f10SYehuda Sadeh { 1235a40c4f10SYehuda Sadeh dout("__unregister_linger_request %p\n", req); 1236a40c4f10SYehuda Sadeh list_del_init(&req->r_linger_item); 123761c74035SAlex Elder if (req->r_osd) { 12381d0326b1SIlya Dryomov list_del_init(&req->r_linger_osd_item); 1239bbf37ec3SIlya Dryomov maybe_move_osd_to_lru(osdc, req->r_osd); 1240fbdb9190SSage Weil if (list_empty(&req->r_osd_item)) 1241a40c4f10SYehuda Sadeh req->r_osd = NULL; 1242a40c4f10SYehuda Sadeh } 124396e4dac6SAlex Elder ceph_osdc_put_request(req); 1244a40c4f10SYehuda Sadeh } 1245a40c4f10SYehuda Sadeh 1246a40c4f10SYehuda Sadeh void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc, 1247a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1248a40c4f10SYehuda Sadeh { 1249a40c4f10SYehuda Sadeh mutex_lock(&osdc->request_mutex); 1250a40c4f10SYehuda Sadeh if (req->r_linger) { 1251c10ebbf5SAlex Elder req->r_linger = 0; 125296e4dac6SAlex Elder __unregister_linger_request(osdc, req); 1253a40c4f10SYehuda Sadeh } 1254a40c4f10SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 1255a40c4f10SYehuda Sadeh } 1256a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_unregister_linger_request); 1257a40c4f10SYehuda Sadeh 1258a40c4f10SYehuda Sadeh void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc, 1259a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1260a40c4f10SYehuda Sadeh { 1261a40c4f10SYehuda Sadeh if (!req->r_linger) { 1262a40c4f10SYehuda Sadeh dout("set_request_linger %p\n", req); 1263a40c4f10SYehuda Sadeh req->r_linger = 1; 1264a40c4f10SYehuda Sadeh } 1265a40c4f10SYehuda Sadeh } 1266a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_set_request_linger); 1267a40c4f10SYehuda Sadeh 12683d14c5d2SYehuda Sadeh /* 1269d29adb34SJosh Durgin * Returns whether a request should be blocked from being sent 1270d29adb34SJosh Durgin * based on the current osdmap and osd_client settings. 1271d29adb34SJosh Durgin * 1272d29adb34SJosh Durgin * Caller should hold map_sem for read. 1273d29adb34SJosh Durgin */ 1274d29adb34SJosh Durgin static bool __req_should_be_paused(struct ceph_osd_client *osdc, 1275d29adb34SJosh Durgin struct ceph_osd_request *req) 1276d29adb34SJosh Durgin { 1277d29adb34SJosh Durgin bool pauserd = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD); 1278d29adb34SJosh Durgin bool pausewr = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR) || 1279d29adb34SJosh Durgin ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL); 1280d29adb34SJosh Durgin return (req->r_flags & CEPH_OSD_FLAG_READ && pauserd) || 1281d29adb34SJosh Durgin (req->r_flags & CEPH_OSD_FLAG_WRITE && pausewr); 1282d29adb34SJosh Durgin } 1283d29adb34SJosh Durgin 1284d29adb34SJosh Durgin /* 128517a13e40SIlya Dryomov * Calculate mapping of a request to a PG. Takes tiering into account. 128617a13e40SIlya Dryomov */ 128717a13e40SIlya Dryomov static int __calc_request_pg(struct ceph_osdmap *osdmap, 128817a13e40SIlya Dryomov struct ceph_osd_request *req, 128917a13e40SIlya Dryomov struct ceph_pg *pg_out) 129017a13e40SIlya Dryomov { 1291205ee118SIlya Dryomov bool need_check_tiering; 1292205ee118SIlya Dryomov 1293205ee118SIlya Dryomov need_check_tiering = false; 1294205ee118SIlya Dryomov if (req->r_target_oloc.pool == -1) { 1295205ee118SIlya Dryomov req->r_target_oloc = req->r_base_oloc; /* struct */ 1296205ee118SIlya Dryomov need_check_tiering = true; 1297205ee118SIlya Dryomov } 1298205ee118SIlya Dryomov if (req->r_target_oid.name_len == 0) { 1299205ee118SIlya Dryomov ceph_oid_copy(&req->r_target_oid, &req->r_base_oid); 1300205ee118SIlya Dryomov need_check_tiering = true; 1301205ee118SIlya Dryomov } 1302205ee118SIlya Dryomov 1303205ee118SIlya Dryomov if (need_check_tiering && 1304205ee118SIlya Dryomov (req->r_flags & CEPH_OSD_FLAG_IGNORE_OVERLAY) == 0) { 130517a13e40SIlya Dryomov struct ceph_pg_pool_info *pi; 130617a13e40SIlya Dryomov 1307205ee118SIlya Dryomov pi = ceph_pg_pool_by_id(osdmap, req->r_target_oloc.pool); 130817a13e40SIlya Dryomov if (pi) { 130917a13e40SIlya Dryomov if ((req->r_flags & CEPH_OSD_FLAG_READ) && 131017a13e40SIlya Dryomov pi->read_tier >= 0) 1311205ee118SIlya Dryomov req->r_target_oloc.pool = pi->read_tier; 131217a13e40SIlya Dryomov if ((req->r_flags & CEPH_OSD_FLAG_WRITE) && 131317a13e40SIlya Dryomov pi->write_tier >= 0) 1314205ee118SIlya Dryomov req->r_target_oloc.pool = pi->write_tier; 131517a13e40SIlya Dryomov } 131617a13e40SIlya Dryomov /* !pi is caught in ceph_oloc_oid_to_pg() */ 131717a13e40SIlya Dryomov } 131817a13e40SIlya Dryomov 1319205ee118SIlya Dryomov return ceph_oloc_oid_to_pg(osdmap, &req->r_target_oloc, 1320205ee118SIlya Dryomov &req->r_target_oid, pg_out); 132117a13e40SIlya Dryomov } 132217a13e40SIlya Dryomov 132317a13e40SIlya Dryomov /* 13243d14c5d2SYehuda Sadeh * Pick an osd (the first 'up' osd in the pg), allocate the osd struct 13253d14c5d2SYehuda Sadeh * (as needed), and set the request r_osd appropriately. If there is 132625985edcSLucas De Marchi * no up osd, set r_osd to NULL. Move the request to the appropriate list 13276f6c7006SSage Weil * (unsent, homeless) or leave on in-flight lru. 13283d14c5d2SYehuda Sadeh * 13293d14c5d2SYehuda Sadeh * Return 0 if unchanged, 1 if changed, or negative on error. 13303d14c5d2SYehuda Sadeh * 13313d14c5d2SYehuda Sadeh * Caller should hold map_sem for read and request_mutex. 13323d14c5d2SYehuda Sadeh */ 13336f6c7006SSage Weil static int __map_request(struct ceph_osd_client *osdc, 133438d6453cSSage Weil struct ceph_osd_request *req, int force_resend) 13353d14c5d2SYehuda Sadeh { 13365b191d99SSage Weil struct ceph_pg pgid; 13373d14c5d2SYehuda Sadeh int acting[CEPH_PG_MAX_SIZE]; 13388008ab10SIlya Dryomov int num, o; 13393d14c5d2SYehuda Sadeh int err; 1340d29adb34SJosh Durgin bool was_paused; 13413d14c5d2SYehuda Sadeh 13426f6c7006SSage Weil dout("map_request %p tid %lld\n", req, req->r_tid); 134317a13e40SIlya Dryomov 134417a13e40SIlya Dryomov err = __calc_request_pg(osdc->osdmap, req, &pgid); 13456f6c7006SSage Weil if (err) { 13466f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 13473d14c5d2SYehuda Sadeh return err; 13486f6c7006SSage Weil } 13493d14c5d2SYehuda Sadeh req->r_pgid = pgid; 13503d14c5d2SYehuda Sadeh 13518008ab10SIlya Dryomov num = ceph_calc_pg_acting(osdc->osdmap, pgid, acting, &o); 13528008ab10SIlya Dryomov if (num < 0) 13538008ab10SIlya Dryomov num = 0; 13543d14c5d2SYehuda Sadeh 1355d29adb34SJosh Durgin was_paused = req->r_paused; 1356d29adb34SJosh Durgin req->r_paused = __req_should_be_paused(osdc, req); 1357d29adb34SJosh Durgin if (was_paused && !req->r_paused) 1358d29adb34SJosh Durgin force_resend = 1; 1359d29adb34SJosh Durgin 136038d6453cSSage Weil if ((!force_resend && 136138d6453cSSage Weil req->r_osd && req->r_osd->o_osd == o && 13623d14c5d2SYehuda Sadeh req->r_sent >= req->r_osd->o_incarnation && 13633d14c5d2SYehuda Sadeh req->r_num_pg_osds == num && 13643d14c5d2SYehuda Sadeh memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) || 1365d29adb34SJosh Durgin (req->r_osd == NULL && o == -1) || 1366d29adb34SJosh Durgin req->r_paused) 13673d14c5d2SYehuda Sadeh return 0; /* no change */ 13683d14c5d2SYehuda Sadeh 13695b191d99SSage Weil dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n", 13705b191d99SSage Weil req->r_tid, pgid.pool, pgid.seed, o, 13713d14c5d2SYehuda Sadeh req->r_osd ? req->r_osd->o_osd : -1); 13723d14c5d2SYehuda Sadeh 13733d14c5d2SYehuda Sadeh /* record full pg acting set */ 13743d14c5d2SYehuda Sadeh memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num); 13753d14c5d2SYehuda Sadeh req->r_num_pg_osds = num; 13763d14c5d2SYehuda Sadeh 13773d14c5d2SYehuda Sadeh if (req->r_osd) { 13783d14c5d2SYehuda Sadeh __cancel_request(req); 13793d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 13803d14c5d2SYehuda Sadeh req->r_osd = NULL; 13813d14c5d2SYehuda Sadeh } 13823d14c5d2SYehuda Sadeh 13833d14c5d2SYehuda Sadeh req->r_osd = __lookup_osd(osdc, o); 13843d14c5d2SYehuda Sadeh if (!req->r_osd && o >= 0) { 13853d14c5d2SYehuda Sadeh err = -ENOMEM; 1386e10006f8SAlex Elder req->r_osd = create_osd(osdc, o); 13876f6c7006SSage Weil if (!req->r_osd) { 13886f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 13893d14c5d2SYehuda Sadeh goto out; 13906f6c7006SSage Weil } 13913d14c5d2SYehuda Sadeh 13926f6c7006SSage Weil dout("map_request osd %p is osd%d\n", req->r_osd, o); 13933d14c5d2SYehuda Sadeh __insert_osd(osdc, req->r_osd); 13943d14c5d2SYehuda Sadeh 1395b7a9e5ddSSage Weil ceph_con_open(&req->r_osd->o_con, 1396b7a9e5ddSSage Weil CEPH_ENTITY_TYPE_OSD, o, 1397b7a9e5ddSSage Weil &osdc->osdmap->osd_addr[o]); 13983d14c5d2SYehuda Sadeh } 13993d14c5d2SYehuda Sadeh 14003d14c5d2SYehuda Sadeh if (req->r_osd) { 14013d14c5d2SYehuda Sadeh __remove_osd_from_lru(req->r_osd); 1402ad885927SAlex Elder list_add_tail(&req->r_osd_item, &req->r_osd->o_requests); 1403ad885927SAlex Elder list_move_tail(&req->r_req_lru_item, &osdc->req_unsent); 14046f6c7006SSage Weil } else { 1405ad885927SAlex Elder list_move_tail(&req->r_req_lru_item, &osdc->req_notarget); 14063d14c5d2SYehuda Sadeh } 14073d14c5d2SYehuda Sadeh err = 1; /* osd or pg changed */ 14083d14c5d2SYehuda Sadeh 14093d14c5d2SYehuda Sadeh out: 14103d14c5d2SYehuda Sadeh return err; 14113d14c5d2SYehuda Sadeh } 14123d14c5d2SYehuda Sadeh 14133d14c5d2SYehuda Sadeh /* 14143d14c5d2SYehuda Sadeh * caller should hold map_sem (for read) and request_mutex 14153d14c5d2SYehuda Sadeh */ 141656e925b6SSage Weil static void __send_request(struct ceph_osd_client *osdc, 14173d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 14183d14c5d2SYehuda Sadeh { 14191b83bef2SSage Weil void *p; 14203d14c5d2SYehuda Sadeh 14211b83bef2SSage Weil dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n", 14221b83bef2SSage Weil req, req->r_tid, req->r_osd->o_osd, req->r_flags, 14231b83bef2SSage Weil (unsigned long long)req->r_pgid.pool, req->r_pgid.seed); 14243d14c5d2SYehuda Sadeh 14251b83bef2SSage Weil /* fill in message content that changes each time we send it */ 14261b83bef2SSage Weil put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch); 14271b83bef2SSage Weil put_unaligned_le32(req->r_flags, req->r_request_flags); 1428205ee118SIlya Dryomov put_unaligned_le64(req->r_target_oloc.pool, req->r_request_pool); 14291b83bef2SSage Weil p = req->r_request_pgid; 14301b83bef2SSage Weil ceph_encode_64(&p, req->r_pgid.pool); 14311b83bef2SSage Weil ceph_encode_32(&p, req->r_pgid.seed); 14321b83bef2SSage Weil put_unaligned_le64(1, req->r_request_attempts); /* FIXME */ 14331b83bef2SSage Weil memcpy(req->r_request_reassert_version, &req->r_reassert_version, 14341b83bef2SSage Weil sizeof(req->r_reassert_version)); 14352169aea6SSage Weil 14363d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 14373d14c5d2SYehuda Sadeh list_move_tail(&req->r_req_lru_item, &osdc->req_lru); 14383d14c5d2SYehuda Sadeh 14393d14c5d2SYehuda Sadeh ceph_msg_get(req->r_request); /* send consumes a ref */ 144026be8808SAlex Elder 14413d14c5d2SYehuda Sadeh req->r_sent = req->r_osd->o_incarnation; 144226be8808SAlex Elder 144326be8808SAlex Elder ceph_con_send(&req->r_osd->o_con, req->r_request); 14443d14c5d2SYehuda Sadeh } 14453d14c5d2SYehuda Sadeh 14463d14c5d2SYehuda Sadeh /* 14476f6c7006SSage Weil * Send any requests in the queue (req_unsent). 14486f6c7006SSage Weil */ 1449f9d25199SAlex Elder static void __send_queued(struct ceph_osd_client *osdc) 14506f6c7006SSage Weil { 14516f6c7006SSage Weil struct ceph_osd_request *req, *tmp; 14526f6c7006SSage Weil 1453f9d25199SAlex Elder dout("__send_queued\n"); 1454f9d25199SAlex Elder list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) 14556f6c7006SSage Weil __send_request(osdc, req); 14566f6c7006SSage Weil } 14576f6c7006SSage Weil 14586f6c7006SSage Weil /* 14590bbfdfe8SIlya Dryomov * Caller should hold map_sem for read and request_mutex. 14600bbfdfe8SIlya Dryomov */ 14610bbfdfe8SIlya Dryomov static int __ceph_osdc_start_request(struct ceph_osd_client *osdc, 14620bbfdfe8SIlya Dryomov struct ceph_osd_request *req, 14630bbfdfe8SIlya Dryomov bool nofail) 14640bbfdfe8SIlya Dryomov { 14650bbfdfe8SIlya Dryomov int rc; 14660bbfdfe8SIlya Dryomov 14670bbfdfe8SIlya Dryomov __register_request(osdc, req); 14680bbfdfe8SIlya Dryomov req->r_sent = 0; 14690bbfdfe8SIlya Dryomov req->r_got_reply = 0; 14700bbfdfe8SIlya Dryomov rc = __map_request(osdc, req, 0); 14710bbfdfe8SIlya Dryomov if (rc < 0) { 14720bbfdfe8SIlya Dryomov if (nofail) { 14730bbfdfe8SIlya Dryomov dout("osdc_start_request failed map, " 14740bbfdfe8SIlya Dryomov " will retry %lld\n", req->r_tid); 14750bbfdfe8SIlya Dryomov rc = 0; 14760bbfdfe8SIlya Dryomov } else { 14770bbfdfe8SIlya Dryomov __unregister_request(osdc, req); 14780bbfdfe8SIlya Dryomov } 14790bbfdfe8SIlya Dryomov return rc; 14800bbfdfe8SIlya Dryomov } 14810bbfdfe8SIlya Dryomov 14820bbfdfe8SIlya Dryomov if (req->r_osd == NULL) { 14830bbfdfe8SIlya Dryomov dout("send_request %p no up osds in pg\n", req); 14840bbfdfe8SIlya Dryomov ceph_monc_request_next_osdmap(&osdc->client->monc); 14850bbfdfe8SIlya Dryomov } else { 14860bbfdfe8SIlya Dryomov __send_queued(osdc); 14870bbfdfe8SIlya Dryomov } 14880bbfdfe8SIlya Dryomov 14890bbfdfe8SIlya Dryomov return 0; 14900bbfdfe8SIlya Dryomov } 14910bbfdfe8SIlya Dryomov 14920bbfdfe8SIlya Dryomov /* 14933d14c5d2SYehuda Sadeh * Timeout callback, called every N seconds when 1 or more osd 14943d14c5d2SYehuda Sadeh * requests has been active for more than N seconds. When this 14953d14c5d2SYehuda Sadeh * happens, we ping all OSDs with requests who have timed out to 14963d14c5d2SYehuda Sadeh * ensure any communications channel reset is detected. Reset the 14973d14c5d2SYehuda Sadeh * request timeouts another N seconds in the future as we go. 14983d14c5d2SYehuda Sadeh * Reschedule the timeout event another N seconds in future (unless 14993d14c5d2SYehuda Sadeh * there are no open requests). 15003d14c5d2SYehuda Sadeh */ 15013d14c5d2SYehuda Sadeh static void handle_timeout(struct work_struct *work) 15023d14c5d2SYehuda Sadeh { 15033d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 15043d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, timeout_work.work); 150583aff95eSSage Weil struct ceph_osd_request *req; 15063d14c5d2SYehuda Sadeh struct ceph_osd *osd; 15073d14c5d2SYehuda Sadeh unsigned long keepalive = 15083d14c5d2SYehuda Sadeh osdc->client->options->osd_keepalive_timeout * HZ; 15093d14c5d2SYehuda Sadeh struct list_head slow_osds; 15103d14c5d2SYehuda Sadeh dout("timeout\n"); 15113d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 15123d14c5d2SYehuda Sadeh 15133d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 15143d14c5d2SYehuda Sadeh 15153d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 15163d14c5d2SYehuda Sadeh 15173d14c5d2SYehuda Sadeh /* 15183d14c5d2SYehuda Sadeh * ping osds that are a bit slow. this ensures that if there 15193d14c5d2SYehuda Sadeh * is a break in the TCP connection we will notice, and reopen 15203d14c5d2SYehuda Sadeh * a connection with that osd (from the fault callback). 15213d14c5d2SYehuda Sadeh */ 15223d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&slow_osds); 15233d14c5d2SYehuda Sadeh list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) { 15243d14c5d2SYehuda Sadeh if (time_before(jiffies, req->r_stamp + keepalive)) 15253d14c5d2SYehuda Sadeh break; 15263d14c5d2SYehuda Sadeh 15273d14c5d2SYehuda Sadeh osd = req->r_osd; 15283d14c5d2SYehuda Sadeh BUG_ON(!osd); 15293d14c5d2SYehuda Sadeh dout(" tid %llu is slow, will send keepalive on osd%d\n", 15303d14c5d2SYehuda Sadeh req->r_tid, osd->o_osd); 15313d14c5d2SYehuda Sadeh list_move_tail(&osd->o_keepalive_item, &slow_osds); 15323d14c5d2SYehuda Sadeh } 15333d14c5d2SYehuda Sadeh while (!list_empty(&slow_osds)) { 15343d14c5d2SYehuda Sadeh osd = list_entry(slow_osds.next, struct ceph_osd, 15353d14c5d2SYehuda Sadeh o_keepalive_item); 15363d14c5d2SYehuda Sadeh list_del_init(&osd->o_keepalive_item); 15373d14c5d2SYehuda Sadeh ceph_con_keepalive(&osd->o_con); 15383d14c5d2SYehuda Sadeh } 15393d14c5d2SYehuda Sadeh 15403d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 1541f9d25199SAlex Elder __send_queued(osdc); 15423d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 15433d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 15443d14c5d2SYehuda Sadeh } 15453d14c5d2SYehuda Sadeh 15463d14c5d2SYehuda Sadeh static void handle_osds_timeout(struct work_struct *work) 15473d14c5d2SYehuda Sadeh { 15483d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 15493d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, 15503d14c5d2SYehuda Sadeh osds_timeout_work.work); 15513d14c5d2SYehuda Sadeh unsigned long delay = 15523d14c5d2SYehuda Sadeh osdc->client->options->osd_idle_ttl * HZ >> 2; 15533d14c5d2SYehuda Sadeh 15543d14c5d2SYehuda Sadeh dout("osds timeout\n"); 15553d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 1556aca420bcSSage Weil remove_old_osds(osdc); 15573d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 15583d14c5d2SYehuda Sadeh 15593d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 15603d14c5d2SYehuda Sadeh round_jiffies_relative(delay)); 15613d14c5d2SYehuda Sadeh } 15623d14c5d2SYehuda Sadeh 1563205ee118SIlya Dryomov static int ceph_oloc_decode(void **p, void *end, 1564205ee118SIlya Dryomov struct ceph_object_locator *oloc) 1565205ee118SIlya Dryomov { 1566205ee118SIlya Dryomov u8 struct_v, struct_cv; 1567205ee118SIlya Dryomov u32 len; 1568205ee118SIlya Dryomov void *struct_end; 1569205ee118SIlya Dryomov int ret = 0; 1570205ee118SIlya Dryomov 1571205ee118SIlya Dryomov ceph_decode_need(p, end, 1 + 1 + 4, e_inval); 1572205ee118SIlya Dryomov struct_v = ceph_decode_8(p); 1573205ee118SIlya Dryomov struct_cv = ceph_decode_8(p); 1574205ee118SIlya Dryomov if (struct_v < 3) { 1575205ee118SIlya Dryomov pr_warn("got v %d < 3 cv %d of ceph_object_locator\n", 1576205ee118SIlya Dryomov struct_v, struct_cv); 1577205ee118SIlya Dryomov goto e_inval; 1578205ee118SIlya Dryomov } 1579205ee118SIlya Dryomov if (struct_cv > 6) { 1580205ee118SIlya Dryomov pr_warn("got v %d cv %d > 6 of ceph_object_locator\n", 1581205ee118SIlya Dryomov struct_v, struct_cv); 1582205ee118SIlya Dryomov goto e_inval; 1583205ee118SIlya Dryomov } 1584205ee118SIlya Dryomov len = ceph_decode_32(p); 1585205ee118SIlya Dryomov ceph_decode_need(p, end, len, e_inval); 1586205ee118SIlya Dryomov struct_end = *p + len; 1587205ee118SIlya Dryomov 1588205ee118SIlya Dryomov oloc->pool = ceph_decode_64(p); 1589205ee118SIlya Dryomov *p += 4; /* skip preferred */ 1590205ee118SIlya Dryomov 1591205ee118SIlya Dryomov len = ceph_decode_32(p); 1592205ee118SIlya Dryomov if (len > 0) { 1593205ee118SIlya Dryomov pr_warn("ceph_object_locator::key is set\n"); 1594205ee118SIlya Dryomov goto e_inval; 1595205ee118SIlya Dryomov } 1596205ee118SIlya Dryomov 1597205ee118SIlya Dryomov if (struct_v >= 5) { 1598205ee118SIlya Dryomov len = ceph_decode_32(p); 1599205ee118SIlya Dryomov if (len > 0) { 1600205ee118SIlya Dryomov pr_warn("ceph_object_locator::nspace is set\n"); 1601205ee118SIlya Dryomov goto e_inval; 1602205ee118SIlya Dryomov } 1603205ee118SIlya Dryomov } 1604205ee118SIlya Dryomov 1605205ee118SIlya Dryomov if (struct_v >= 6) { 1606205ee118SIlya Dryomov s64 hash = ceph_decode_64(p); 1607205ee118SIlya Dryomov if (hash != -1) { 1608205ee118SIlya Dryomov pr_warn("ceph_object_locator::hash is set\n"); 1609205ee118SIlya Dryomov goto e_inval; 1610205ee118SIlya Dryomov } 1611205ee118SIlya Dryomov } 1612205ee118SIlya Dryomov 1613205ee118SIlya Dryomov /* skip the rest */ 1614205ee118SIlya Dryomov *p = struct_end; 1615205ee118SIlya Dryomov out: 1616205ee118SIlya Dryomov return ret; 1617205ee118SIlya Dryomov 1618205ee118SIlya Dryomov e_inval: 1619205ee118SIlya Dryomov ret = -EINVAL; 1620205ee118SIlya Dryomov goto out; 1621205ee118SIlya Dryomov } 1622205ee118SIlya Dryomov 1623205ee118SIlya Dryomov static int ceph_redirect_decode(void **p, void *end, 1624205ee118SIlya Dryomov struct ceph_request_redirect *redir) 1625205ee118SIlya Dryomov { 1626205ee118SIlya Dryomov u8 struct_v, struct_cv; 1627205ee118SIlya Dryomov u32 len; 1628205ee118SIlya Dryomov void *struct_end; 1629205ee118SIlya Dryomov int ret; 1630205ee118SIlya Dryomov 1631205ee118SIlya Dryomov ceph_decode_need(p, end, 1 + 1 + 4, e_inval); 1632205ee118SIlya Dryomov struct_v = ceph_decode_8(p); 1633205ee118SIlya Dryomov struct_cv = ceph_decode_8(p); 1634205ee118SIlya Dryomov if (struct_cv > 1) { 1635205ee118SIlya Dryomov pr_warn("got v %d cv %d > 1 of ceph_request_redirect\n", 1636205ee118SIlya Dryomov struct_v, struct_cv); 1637205ee118SIlya Dryomov goto e_inval; 1638205ee118SIlya Dryomov } 1639205ee118SIlya Dryomov len = ceph_decode_32(p); 1640205ee118SIlya Dryomov ceph_decode_need(p, end, len, e_inval); 1641205ee118SIlya Dryomov struct_end = *p + len; 1642205ee118SIlya Dryomov 1643205ee118SIlya Dryomov ret = ceph_oloc_decode(p, end, &redir->oloc); 1644205ee118SIlya Dryomov if (ret) 1645205ee118SIlya Dryomov goto out; 1646205ee118SIlya Dryomov 1647205ee118SIlya Dryomov len = ceph_decode_32(p); 1648205ee118SIlya Dryomov if (len > 0) { 1649205ee118SIlya Dryomov pr_warn("ceph_request_redirect::object_name is set\n"); 1650205ee118SIlya Dryomov goto e_inval; 1651205ee118SIlya Dryomov } 1652205ee118SIlya Dryomov 1653205ee118SIlya Dryomov len = ceph_decode_32(p); 1654205ee118SIlya Dryomov *p += len; /* skip osd_instructions */ 1655205ee118SIlya Dryomov 1656205ee118SIlya Dryomov /* skip the rest */ 1657205ee118SIlya Dryomov *p = struct_end; 1658205ee118SIlya Dryomov out: 1659205ee118SIlya Dryomov return ret; 1660205ee118SIlya Dryomov 1661205ee118SIlya Dryomov e_inval: 1662205ee118SIlya Dryomov ret = -EINVAL; 1663205ee118SIlya Dryomov goto out; 1664205ee118SIlya Dryomov } 1665205ee118SIlya Dryomov 166625845472SSage Weil static void complete_request(struct ceph_osd_request *req) 166725845472SSage Weil { 166825845472SSage Weil complete_all(&req->r_safe_completion); /* fsync waiter */ 166925845472SSage Weil } 167025845472SSage Weil 16713d14c5d2SYehuda Sadeh /* 16723d14c5d2SYehuda Sadeh * handle osd op reply. either call the callback if it is specified, 16733d14c5d2SYehuda Sadeh * or do the completion to wake up the waiting thread. 16743d14c5d2SYehuda Sadeh */ 16753d14c5d2SYehuda Sadeh static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg, 16763d14c5d2SYehuda Sadeh struct ceph_connection *con) 16773d14c5d2SYehuda Sadeh { 16781b83bef2SSage Weil void *p, *end; 16793d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 1680205ee118SIlya Dryomov struct ceph_request_redirect redir; 16813d14c5d2SYehuda Sadeh u64 tid; 16821b83bef2SSage Weil int object_len; 168379528734SAlex Elder unsigned int numops; 168479528734SAlex Elder int payload_len, flags; 16853d14c5d2SYehuda Sadeh s32 result; 16861b83bef2SSage Weil s32 retry_attempt; 16871b83bef2SSage Weil struct ceph_pg pg; 16881b83bef2SSage Weil int err; 16891b83bef2SSage Weil u32 reassert_epoch; 16901b83bef2SSage Weil u64 reassert_version; 16911b83bef2SSage Weil u32 osdmap_epoch; 16920d5af164SAlex Elder int already_completed; 16939fc6e064SAlex Elder u32 bytes; 169479528734SAlex Elder unsigned int i; 16953d14c5d2SYehuda Sadeh 16963d14c5d2SYehuda Sadeh tid = le64_to_cpu(msg->hdr.tid); 16971b83bef2SSage Weil dout("handle_reply %p tid %llu\n", msg, tid); 16981b83bef2SSage Weil 16991b83bef2SSage Weil p = msg->front.iov_base; 17001b83bef2SSage Weil end = p + msg->front.iov_len; 17011b83bef2SSage Weil 17021b83bef2SSage Weil ceph_decode_need(&p, end, 4, bad); 17031b83bef2SSage Weil object_len = ceph_decode_32(&p); 17041b83bef2SSage Weil ceph_decode_need(&p, end, object_len, bad); 17051b83bef2SSage Weil p += object_len; 17061b83bef2SSage Weil 1707ef4859d6SAlex Elder err = ceph_decode_pgid(&p, end, &pg); 17081b83bef2SSage Weil if (err) 17093d14c5d2SYehuda Sadeh goto bad; 17101b83bef2SSage Weil 17111b83bef2SSage Weil ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad); 17121b83bef2SSage Weil flags = ceph_decode_64(&p); 17131b83bef2SSage Weil result = ceph_decode_32(&p); 17141b83bef2SSage Weil reassert_epoch = ceph_decode_32(&p); 17151b83bef2SSage Weil reassert_version = ceph_decode_64(&p); 17161b83bef2SSage Weil osdmap_epoch = ceph_decode_32(&p); 17171b83bef2SSage Weil 17183d14c5d2SYehuda Sadeh /* lookup */ 1719ff513aceSIlya Dryomov down_read(&osdc->map_sem); 17203d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 17213d14c5d2SYehuda Sadeh req = __lookup_request(osdc, tid); 17223d14c5d2SYehuda Sadeh if (req == NULL) { 17233d14c5d2SYehuda Sadeh dout("handle_reply tid %llu dne\n", tid); 17248058fd45SAlex Elder goto bad_mutex; 17253d14c5d2SYehuda Sadeh } 17263d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 17271b83bef2SSage Weil 17281b83bef2SSage Weil dout("handle_reply %p tid %llu req %p result %d\n", msg, tid, 17291b83bef2SSage Weil req, result); 17301b83bef2SSage Weil 173118741196SDan Carpenter ceph_decode_need(&p, end, 4, bad_put); 17321b83bef2SSage Weil numops = ceph_decode_32(&p); 17331b83bef2SSage Weil if (numops > CEPH_OSD_MAX_OP) 17341b83bef2SSage Weil goto bad_put; 17351b83bef2SSage Weil if (numops != req->r_num_ops) 17361b83bef2SSage Weil goto bad_put; 17371b83bef2SSage Weil payload_len = 0; 173818741196SDan Carpenter ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad_put); 17391b83bef2SSage Weil for (i = 0; i < numops; i++) { 17401b83bef2SSage Weil struct ceph_osd_op *op = p; 17411b83bef2SSage Weil int len; 17421b83bef2SSage Weil 17431b83bef2SSage Weil len = le32_to_cpu(op->payload_len); 17441b83bef2SSage Weil req->r_reply_op_len[i] = len; 17451b83bef2SSage Weil dout(" op %d has %d bytes\n", i, len); 17461b83bef2SSage Weil payload_len += len; 17471b83bef2SSage Weil p += sizeof(*op); 17481b83bef2SSage Weil } 17499fc6e064SAlex Elder bytes = le32_to_cpu(msg->hdr.data_len); 17509fc6e064SAlex Elder if (payload_len != bytes) { 17511b83bef2SSage Weil pr_warning("sum of op payload lens %d != data_len %d", 17529fc6e064SAlex Elder payload_len, bytes); 17531b83bef2SSage Weil goto bad_put; 17541b83bef2SSage Weil } 17551b83bef2SSage Weil 175618741196SDan Carpenter ceph_decode_need(&p, end, 4 + numops * 4, bad_put); 17571b83bef2SSage Weil retry_attempt = ceph_decode_32(&p); 17581b83bef2SSage Weil for (i = 0; i < numops; i++) 17591b83bef2SSage Weil req->r_reply_op_result[i] = ceph_decode_32(&p); 17603d14c5d2SYehuda Sadeh 1761205ee118SIlya Dryomov if (le16_to_cpu(msg->hdr.version) >= 6) { 1762205ee118SIlya Dryomov p += 8 + 4; /* skip replay_version */ 1763205ee118SIlya Dryomov p += 8; /* skip user_version */ 1764205ee118SIlya Dryomov 1765205ee118SIlya Dryomov err = ceph_redirect_decode(&p, end, &redir); 1766205ee118SIlya Dryomov if (err) 1767205ee118SIlya Dryomov goto bad_put; 1768205ee118SIlya Dryomov } else { 1769205ee118SIlya Dryomov redir.oloc.pool = -1; 1770205ee118SIlya Dryomov } 1771205ee118SIlya Dryomov 1772205ee118SIlya Dryomov if (redir.oloc.pool != -1) { 1773205ee118SIlya Dryomov dout("redirect pool %lld\n", redir.oloc.pool); 1774205ee118SIlya Dryomov 1775205ee118SIlya Dryomov __unregister_request(osdc, req); 1776205ee118SIlya Dryomov 1777205ee118SIlya Dryomov req->r_target_oloc = redir.oloc; /* struct */ 1778205ee118SIlya Dryomov 1779205ee118SIlya Dryomov /* 1780205ee118SIlya Dryomov * Start redirect requests with nofail=true. If 1781205ee118SIlya Dryomov * mapping fails, request will end up on the notarget 1782205ee118SIlya Dryomov * list, waiting for the new osdmap (which can take 1783205ee118SIlya Dryomov * a while), even though the original request mapped 1784205ee118SIlya Dryomov * successfully. In the future we might want to follow 1785205ee118SIlya Dryomov * original request's nofail setting here. 1786205ee118SIlya Dryomov */ 1787ff513aceSIlya Dryomov err = __ceph_osdc_start_request(osdc, req, true); 1788205ee118SIlya Dryomov BUG_ON(err); 1789205ee118SIlya Dryomov 1790ff513aceSIlya Dryomov goto out_unlock; 1791205ee118SIlya Dryomov } 1792205ee118SIlya Dryomov 1793eb845ff1SYan, Zheng already_completed = req->r_got_reply; 17943d14c5d2SYehuda Sadeh if (!req->r_got_reply) { 17951b83bef2SSage Weil req->r_result = result; 17963d14c5d2SYehuda Sadeh dout("handle_reply result %d bytes %d\n", req->r_result, 17973d14c5d2SYehuda Sadeh bytes); 17983d14c5d2SYehuda Sadeh if (req->r_result == 0) 17993d14c5d2SYehuda Sadeh req->r_result = bytes; 18003d14c5d2SYehuda Sadeh 18013d14c5d2SYehuda Sadeh /* in case this is a write and we need to replay, */ 18021b83bef2SSage Weil req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch); 18031b83bef2SSage Weil req->r_reassert_version.version = cpu_to_le64(reassert_version); 18043d14c5d2SYehuda Sadeh 18053d14c5d2SYehuda Sadeh req->r_got_reply = 1; 18063d14c5d2SYehuda Sadeh } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) { 18073d14c5d2SYehuda Sadeh dout("handle_reply tid %llu dup ack\n", tid); 1808ff513aceSIlya Dryomov goto out_unlock; 18093d14c5d2SYehuda Sadeh } 18103d14c5d2SYehuda Sadeh 18113d14c5d2SYehuda Sadeh dout("handle_reply tid %llu flags %d\n", tid, flags); 18123d14c5d2SYehuda Sadeh 1813a40c4f10SYehuda Sadeh if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK)) 1814a40c4f10SYehuda Sadeh __register_linger_request(osdc, req); 1815a40c4f10SYehuda Sadeh 18163d14c5d2SYehuda Sadeh /* either this is a read, or we got the safe response */ 18173d14c5d2SYehuda Sadeh if (result < 0 || 18183d14c5d2SYehuda Sadeh (flags & CEPH_OSD_FLAG_ONDISK) || 18193d14c5d2SYehuda Sadeh ((flags & CEPH_OSD_FLAG_WRITE) == 0)) 18203d14c5d2SYehuda Sadeh __unregister_request(osdc, req); 18213d14c5d2SYehuda Sadeh 18223d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 1823ff513aceSIlya Dryomov up_read(&osdc->map_sem); 18243d14c5d2SYehuda Sadeh 1825eb845ff1SYan, Zheng if (!already_completed) { 182661c5d6bfSYan, Zheng if (req->r_unsafe_callback && 182761c5d6bfSYan, Zheng result >= 0 && !(flags & CEPH_OSD_FLAG_ONDISK)) 182861c5d6bfSYan, Zheng req->r_unsafe_callback(req, true); 18293d14c5d2SYehuda Sadeh if (req->r_callback) 18303d14c5d2SYehuda Sadeh req->r_callback(req, msg); 18313d14c5d2SYehuda Sadeh else 18323d14c5d2SYehuda Sadeh complete_all(&req->r_completion); 1833eb845ff1SYan, Zheng } 18343d14c5d2SYehuda Sadeh 183561c5d6bfSYan, Zheng if (flags & CEPH_OSD_FLAG_ONDISK) { 183661c5d6bfSYan, Zheng if (req->r_unsafe_callback && already_completed) 183761c5d6bfSYan, Zheng req->r_unsafe_callback(req, false); 183825845472SSage Weil complete_request(req); 183961c5d6bfSYan, Zheng } 18403d14c5d2SYehuda Sadeh 1841ff513aceSIlya Dryomov out: 1842a40c4f10SYehuda Sadeh dout("req=%p req->r_linger=%d\n", req, req->r_linger); 18433d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 18443d14c5d2SYehuda Sadeh return; 1845ff513aceSIlya Dryomov out_unlock: 1846ff513aceSIlya Dryomov mutex_unlock(&osdc->request_mutex); 1847ff513aceSIlya Dryomov up_read(&osdc->map_sem); 1848ff513aceSIlya Dryomov goto out; 18493d14c5d2SYehuda Sadeh 18501b83bef2SSage Weil bad_put: 185137c89bdeSLi Wang req->r_result = -EIO; 185237c89bdeSLi Wang __unregister_request(osdc, req); 185337c89bdeSLi Wang if (req->r_callback) 185437c89bdeSLi Wang req->r_callback(req, msg); 185537c89bdeSLi Wang else 185637c89bdeSLi Wang complete_all(&req->r_completion); 185737c89bdeSLi Wang complete_request(req); 18581b83bef2SSage Weil ceph_osdc_put_request(req); 18598058fd45SAlex Elder bad_mutex: 18608058fd45SAlex Elder mutex_unlock(&osdc->request_mutex); 1861ff513aceSIlya Dryomov up_read(&osdc->map_sem); 18623d14c5d2SYehuda Sadeh bad: 18631b83bef2SSage Weil pr_err("corrupt osd_op_reply got %d %d\n", 18641b83bef2SSage Weil (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len)); 18653d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 18663d14c5d2SYehuda Sadeh } 18673d14c5d2SYehuda Sadeh 18686f6c7006SSage Weil static void reset_changed_osds(struct ceph_osd_client *osdc) 18693d14c5d2SYehuda Sadeh { 18703d14c5d2SYehuda Sadeh struct rb_node *p, *n; 18713d14c5d2SYehuda Sadeh 18723d14c5d2SYehuda Sadeh for (p = rb_first(&osdc->osds); p; p = n) { 18736f6c7006SSage Weil struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node); 18743d14c5d2SYehuda Sadeh 18753d14c5d2SYehuda Sadeh n = rb_next(p); 18763d14c5d2SYehuda Sadeh if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) || 18773d14c5d2SYehuda Sadeh memcmp(&osd->o_con.peer_addr, 18783d14c5d2SYehuda Sadeh ceph_osd_addr(osdc->osdmap, 18793d14c5d2SYehuda Sadeh osd->o_osd), 18803d14c5d2SYehuda Sadeh sizeof(struct ceph_entity_addr)) != 0) 18813d14c5d2SYehuda Sadeh __reset_osd(osdc, osd); 18823d14c5d2SYehuda Sadeh } 18833d14c5d2SYehuda Sadeh } 18843d14c5d2SYehuda Sadeh 18853d14c5d2SYehuda Sadeh /* 18866f6c7006SSage Weil * Requeue requests whose mapping to an OSD has changed. If requests map to 18876f6c7006SSage Weil * no osd, request a new map. 18883d14c5d2SYehuda Sadeh * 1889e6d50f67SAlex Elder * Caller should hold map_sem for read. 18903d14c5d2SYehuda Sadeh */ 18919a1ea2dbSJosh Durgin static void kick_requests(struct ceph_osd_client *osdc, bool force_resend, 18929a1ea2dbSJosh Durgin bool force_resend_writes) 18933d14c5d2SYehuda Sadeh { 1894a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 18956f6c7006SSage Weil struct rb_node *p; 18966f6c7006SSage Weil int needmap = 0; 18976f6c7006SSage Weil int err; 18989a1ea2dbSJosh Durgin bool force_resend_req; 18993d14c5d2SYehuda Sadeh 19009a1ea2dbSJosh Durgin dout("kick_requests %s %s\n", force_resend ? " (force resend)" : "", 19019a1ea2dbSJosh Durgin force_resend_writes ? " (force resend writes)" : ""); 19023d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 19036194ea89SSage Weil for (p = rb_first(&osdc->requests); p; ) { 19046f6c7006SSage Weil req = rb_entry(p, struct ceph_osd_request, r_node); 19056194ea89SSage Weil p = rb_next(p); 1906ab60b16dSAlex Elder 1907ab60b16dSAlex Elder /* 1908ab60b16dSAlex Elder * For linger requests that have not yet been 1909ab60b16dSAlex Elder * registered, move them to the linger list; they'll 1910ab60b16dSAlex Elder * be sent to the osd in the loop below. Unregister 1911ab60b16dSAlex Elder * the request before re-registering it as a linger 1912ab60b16dSAlex Elder * request to ensure the __map_request() below 1913ab60b16dSAlex Elder * will decide it needs to be sent. 1914ab60b16dSAlex Elder */ 1915ab60b16dSAlex Elder if (req->r_linger && list_empty(&req->r_linger_item)) { 1916ab60b16dSAlex Elder dout("%p tid %llu restart on osd%d\n", 1917ab60b16dSAlex Elder req, req->r_tid, 1918ab60b16dSAlex Elder req->r_osd ? req->r_osd->o_osd : -1); 191996e4dac6SAlex Elder ceph_osdc_get_request(req); 1920ab60b16dSAlex Elder __unregister_request(osdc, req); 1921ab60b16dSAlex Elder __register_linger_request(osdc, req); 192296e4dac6SAlex Elder ceph_osdc_put_request(req); 1923ab60b16dSAlex Elder continue; 1924ab60b16dSAlex Elder } 1925ab60b16dSAlex Elder 19269a1ea2dbSJosh Durgin force_resend_req = force_resend || 19279a1ea2dbSJosh Durgin (force_resend_writes && 19289a1ea2dbSJosh Durgin req->r_flags & CEPH_OSD_FLAG_WRITE); 19299a1ea2dbSJosh Durgin err = __map_request(osdc, req, force_resend_req); 19306f6c7006SSage Weil if (err < 0) 19316f6c7006SSage Weil continue; /* error */ 19326f6c7006SSage Weil if (req->r_osd == NULL) { 19336f6c7006SSage Weil dout("%p tid %llu maps to no osd\n", req, req->r_tid); 19346f6c7006SSage Weil needmap++; /* request a newer map */ 19356f6c7006SSage Weil } else if (err > 0) { 19366194ea89SSage Weil if (!req->r_linger) { 19376194ea89SSage Weil dout("%p tid %llu requeued on osd%d\n", req, 19386194ea89SSage Weil req->r_tid, 19396f6c7006SSage Weil req->r_osd ? req->r_osd->o_osd : -1); 19406f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 19416f6c7006SSage Weil } 19426f6c7006SSage Weil } 19436194ea89SSage Weil } 1944a40c4f10SYehuda Sadeh 1945a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osdc->req_linger, 1946a40c4f10SYehuda Sadeh r_linger_item) { 1947a40c4f10SYehuda Sadeh dout("linger req=%p req->r_osd=%p\n", req, req->r_osd); 1948a40c4f10SYehuda Sadeh 19499a1ea2dbSJosh Durgin err = __map_request(osdc, req, 19509a1ea2dbSJosh Durgin force_resend || force_resend_writes); 1951ab60b16dSAlex Elder dout("__map_request returned %d\n", err); 1952a40c4f10SYehuda Sadeh if (err == 0) 1953a40c4f10SYehuda Sadeh continue; /* no change and no osd was specified */ 1954a40c4f10SYehuda Sadeh if (err < 0) 1955a40c4f10SYehuda Sadeh continue; /* hrm! */ 1956a40c4f10SYehuda Sadeh if (req->r_osd == NULL) { 1957a40c4f10SYehuda Sadeh dout("tid %llu maps to no valid osd\n", req->r_tid); 1958a40c4f10SYehuda Sadeh needmap++; /* request a newer map */ 1959a40c4f10SYehuda Sadeh continue; 1960a40c4f10SYehuda Sadeh } 1961a40c4f10SYehuda Sadeh 1962a40c4f10SYehuda Sadeh dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid, 1963a40c4f10SYehuda Sadeh req->r_osd ? req->r_osd->o_osd : -1); 1964a40c4f10SYehuda Sadeh __register_request(osdc, req); 1965c89ce05eSAlex Elder __unregister_linger_request(osdc, req); 1966a40c4f10SYehuda Sadeh } 196714d2f38dSAlex Elder reset_changed_osds(osdc); 19683d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 19693d14c5d2SYehuda Sadeh 19703d14c5d2SYehuda Sadeh if (needmap) { 19713d14c5d2SYehuda Sadeh dout("%d requests for down osds, need new map\n", needmap); 19723d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 19733d14c5d2SYehuda Sadeh } 19743d14c5d2SYehuda Sadeh } 19756f6c7006SSage Weil 19766f6c7006SSage Weil 19773d14c5d2SYehuda Sadeh /* 19783d14c5d2SYehuda Sadeh * Process updated osd map. 19793d14c5d2SYehuda Sadeh * 19803d14c5d2SYehuda Sadeh * The message contains any number of incremental and full maps, normally 19813d14c5d2SYehuda Sadeh * indicating some sort of topology change in the cluster. Kick requests 19823d14c5d2SYehuda Sadeh * off to different OSDs as needed. 19833d14c5d2SYehuda Sadeh */ 19843d14c5d2SYehuda Sadeh void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg) 19853d14c5d2SYehuda Sadeh { 19863d14c5d2SYehuda Sadeh void *p, *end, *next; 19873d14c5d2SYehuda Sadeh u32 nr_maps, maplen; 19883d14c5d2SYehuda Sadeh u32 epoch; 19893d14c5d2SYehuda Sadeh struct ceph_osdmap *newmap = NULL, *oldmap; 19903d14c5d2SYehuda Sadeh int err; 19913d14c5d2SYehuda Sadeh struct ceph_fsid fsid; 19929a1ea2dbSJosh Durgin bool was_full; 19933d14c5d2SYehuda Sadeh 19943d14c5d2SYehuda Sadeh dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0); 19953d14c5d2SYehuda Sadeh p = msg->front.iov_base; 19963d14c5d2SYehuda Sadeh end = p + msg->front.iov_len; 19973d14c5d2SYehuda Sadeh 19983d14c5d2SYehuda Sadeh /* verify fsid */ 19993d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, sizeof(fsid), bad); 20003d14c5d2SYehuda Sadeh ceph_decode_copy(&p, &fsid, sizeof(fsid)); 20013d14c5d2SYehuda Sadeh if (ceph_check_fsid(osdc->client, &fsid) < 0) 20023d14c5d2SYehuda Sadeh return; 20033d14c5d2SYehuda Sadeh 20043d14c5d2SYehuda Sadeh down_write(&osdc->map_sem); 20053d14c5d2SYehuda Sadeh 20069a1ea2dbSJosh Durgin was_full = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL); 20079a1ea2dbSJosh Durgin 20083d14c5d2SYehuda Sadeh /* incremental maps */ 20093d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 20103d14c5d2SYehuda Sadeh dout(" %d inc maps\n", nr_maps); 20113d14c5d2SYehuda Sadeh while (nr_maps > 0) { 20123d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 20133d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 20143d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 20153d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 20163d14c5d2SYehuda Sadeh next = p + maplen; 20173d14c5d2SYehuda Sadeh if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) { 20183d14c5d2SYehuda Sadeh dout("applying incremental map %u len %d\n", 20193d14c5d2SYehuda Sadeh epoch, maplen); 20203d14c5d2SYehuda Sadeh newmap = osdmap_apply_incremental(&p, next, 20213d14c5d2SYehuda Sadeh osdc->osdmap, 202215d9882cSAlex Elder &osdc->client->msgr); 20233d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 20243d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 20253d14c5d2SYehuda Sadeh goto bad; 20263d14c5d2SYehuda Sadeh } 20273d14c5d2SYehuda Sadeh BUG_ON(!newmap); 20283d14c5d2SYehuda Sadeh if (newmap != osdc->osdmap) { 20293d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 20303d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 20313d14c5d2SYehuda Sadeh } 20329a1ea2dbSJosh Durgin was_full = was_full || 20339a1ea2dbSJosh Durgin ceph_osdmap_flag(osdc->osdmap, 20349a1ea2dbSJosh Durgin CEPH_OSDMAP_FULL); 20359a1ea2dbSJosh Durgin kick_requests(osdc, 0, was_full); 20363d14c5d2SYehuda Sadeh } else { 20373d14c5d2SYehuda Sadeh dout("ignoring incremental map %u len %d\n", 20383d14c5d2SYehuda Sadeh epoch, maplen); 20393d14c5d2SYehuda Sadeh } 20403d14c5d2SYehuda Sadeh p = next; 20413d14c5d2SYehuda Sadeh nr_maps--; 20423d14c5d2SYehuda Sadeh } 20433d14c5d2SYehuda Sadeh if (newmap) 20443d14c5d2SYehuda Sadeh goto done; 20453d14c5d2SYehuda Sadeh 20463d14c5d2SYehuda Sadeh /* full maps */ 20473d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 20483d14c5d2SYehuda Sadeh dout(" %d full maps\n", nr_maps); 20493d14c5d2SYehuda Sadeh while (nr_maps) { 20503d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 20513d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 20523d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 20533d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 20543d14c5d2SYehuda Sadeh if (nr_maps > 1) { 20553d14c5d2SYehuda Sadeh dout("skipping non-latest full map %u len %d\n", 20563d14c5d2SYehuda Sadeh epoch, maplen); 20573d14c5d2SYehuda Sadeh } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) { 20583d14c5d2SYehuda Sadeh dout("skipping full map %u len %d, " 20593d14c5d2SYehuda Sadeh "older than our %u\n", epoch, maplen, 20603d14c5d2SYehuda Sadeh osdc->osdmap->epoch); 20613d14c5d2SYehuda Sadeh } else { 206238d6453cSSage Weil int skipped_map = 0; 206338d6453cSSage Weil 20643d14c5d2SYehuda Sadeh dout("taking full map %u len %d\n", epoch, maplen); 2065a2505d63SIlya Dryomov newmap = ceph_osdmap_decode(&p, p+maplen); 20663d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 20673d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 20683d14c5d2SYehuda Sadeh goto bad; 20693d14c5d2SYehuda Sadeh } 20703d14c5d2SYehuda Sadeh BUG_ON(!newmap); 20713d14c5d2SYehuda Sadeh oldmap = osdc->osdmap; 20723d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 207338d6453cSSage Weil if (oldmap) { 207438d6453cSSage Weil if (oldmap->epoch + 1 < newmap->epoch) 207538d6453cSSage Weil skipped_map = 1; 20763d14c5d2SYehuda Sadeh ceph_osdmap_destroy(oldmap); 207738d6453cSSage Weil } 20789a1ea2dbSJosh Durgin was_full = was_full || 20799a1ea2dbSJosh Durgin ceph_osdmap_flag(osdc->osdmap, 20809a1ea2dbSJosh Durgin CEPH_OSDMAP_FULL); 20819a1ea2dbSJosh Durgin kick_requests(osdc, skipped_map, was_full); 20823d14c5d2SYehuda Sadeh } 20833d14c5d2SYehuda Sadeh p += maplen; 20843d14c5d2SYehuda Sadeh nr_maps--; 20853d14c5d2SYehuda Sadeh } 20863d14c5d2SYehuda Sadeh 2087b72e19b9SDan Carpenter if (!osdc->osdmap) 2088b72e19b9SDan Carpenter goto bad; 20893d14c5d2SYehuda Sadeh done: 20903d14c5d2SYehuda Sadeh downgrade_write(&osdc->map_sem); 20913d14c5d2SYehuda Sadeh ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch); 2092cd634fb6SSage Weil 2093cd634fb6SSage Weil /* 2094cd634fb6SSage Weil * subscribe to subsequent osdmap updates if full to ensure 2095cd634fb6SSage Weil * we find out when we are no longer full and stop returning 2096cd634fb6SSage Weil * ENOSPC. 2097cd634fb6SSage Weil */ 2098d29adb34SJosh Durgin if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) || 2099d29adb34SJosh Durgin ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD) || 2100d29adb34SJosh Durgin ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR)) 2101cd634fb6SSage Weil ceph_monc_request_next_osdmap(&osdc->client->monc); 2102cd634fb6SSage Weil 2103f9d25199SAlex Elder mutex_lock(&osdc->request_mutex); 2104f9d25199SAlex Elder __send_queued(osdc); 2105f9d25199SAlex Elder mutex_unlock(&osdc->request_mutex); 21063d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 21073d14c5d2SYehuda Sadeh wake_up_all(&osdc->client->auth_wq); 21083d14c5d2SYehuda Sadeh return; 21093d14c5d2SYehuda Sadeh 21103d14c5d2SYehuda Sadeh bad: 21113d14c5d2SYehuda Sadeh pr_err("osdc handle_map corrupt msg\n"); 21123d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 21133d14c5d2SYehuda Sadeh up_write(&osdc->map_sem); 21143d14c5d2SYehuda Sadeh } 21153d14c5d2SYehuda Sadeh 21163d14c5d2SYehuda Sadeh /* 2117a40c4f10SYehuda Sadeh * watch/notify callback event infrastructure 2118a40c4f10SYehuda Sadeh * 2119a40c4f10SYehuda Sadeh * These callbacks are used both for watch and notify operations. 2120a40c4f10SYehuda Sadeh */ 2121a40c4f10SYehuda Sadeh static void __release_event(struct kref *kref) 2122a40c4f10SYehuda Sadeh { 2123a40c4f10SYehuda Sadeh struct ceph_osd_event *event = 2124a40c4f10SYehuda Sadeh container_of(kref, struct ceph_osd_event, kref); 2125a40c4f10SYehuda Sadeh 2126a40c4f10SYehuda Sadeh dout("__release_event %p\n", event); 2127a40c4f10SYehuda Sadeh kfree(event); 2128a40c4f10SYehuda Sadeh } 2129a40c4f10SYehuda Sadeh 2130a40c4f10SYehuda Sadeh static void get_event(struct ceph_osd_event *event) 2131a40c4f10SYehuda Sadeh { 2132a40c4f10SYehuda Sadeh kref_get(&event->kref); 2133a40c4f10SYehuda Sadeh } 2134a40c4f10SYehuda Sadeh 2135a40c4f10SYehuda Sadeh void ceph_osdc_put_event(struct ceph_osd_event *event) 2136a40c4f10SYehuda Sadeh { 2137a40c4f10SYehuda Sadeh kref_put(&event->kref, __release_event); 2138a40c4f10SYehuda Sadeh } 2139a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_put_event); 2140a40c4f10SYehuda Sadeh 2141a40c4f10SYehuda Sadeh static void __insert_event(struct ceph_osd_client *osdc, 2142a40c4f10SYehuda Sadeh struct ceph_osd_event *new) 2143a40c4f10SYehuda Sadeh { 2144a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 2145a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 2146a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 2147a40c4f10SYehuda Sadeh 2148a40c4f10SYehuda Sadeh while (*p) { 2149a40c4f10SYehuda Sadeh parent = *p; 2150a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 2151a40c4f10SYehuda Sadeh if (new->cookie < event->cookie) 2152a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 2153a40c4f10SYehuda Sadeh else if (new->cookie > event->cookie) 2154a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 2155a40c4f10SYehuda Sadeh else 2156a40c4f10SYehuda Sadeh BUG(); 2157a40c4f10SYehuda Sadeh } 2158a40c4f10SYehuda Sadeh 2159a40c4f10SYehuda Sadeh rb_link_node(&new->node, parent, p); 2160a40c4f10SYehuda Sadeh rb_insert_color(&new->node, &osdc->event_tree); 2161a40c4f10SYehuda Sadeh } 2162a40c4f10SYehuda Sadeh 2163a40c4f10SYehuda Sadeh static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc, 2164a40c4f10SYehuda Sadeh u64 cookie) 2165a40c4f10SYehuda Sadeh { 2166a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 2167a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 2168a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 2169a40c4f10SYehuda Sadeh 2170a40c4f10SYehuda Sadeh while (*p) { 2171a40c4f10SYehuda Sadeh parent = *p; 2172a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 2173a40c4f10SYehuda Sadeh if (cookie < event->cookie) 2174a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 2175a40c4f10SYehuda Sadeh else if (cookie > event->cookie) 2176a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 2177a40c4f10SYehuda Sadeh else 2178a40c4f10SYehuda Sadeh return event; 2179a40c4f10SYehuda Sadeh } 2180a40c4f10SYehuda Sadeh return NULL; 2181a40c4f10SYehuda Sadeh } 2182a40c4f10SYehuda Sadeh 2183a40c4f10SYehuda Sadeh static void __remove_event(struct ceph_osd_event *event) 2184a40c4f10SYehuda Sadeh { 2185a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 2186a40c4f10SYehuda Sadeh 2187a40c4f10SYehuda Sadeh if (!RB_EMPTY_NODE(&event->node)) { 2188a40c4f10SYehuda Sadeh dout("__remove_event removed %p\n", event); 2189a40c4f10SYehuda Sadeh rb_erase(&event->node, &osdc->event_tree); 2190a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 2191a40c4f10SYehuda Sadeh } else { 2192a40c4f10SYehuda Sadeh dout("__remove_event didn't remove %p\n", event); 2193a40c4f10SYehuda Sadeh } 2194a40c4f10SYehuda Sadeh } 2195a40c4f10SYehuda Sadeh 2196a40c4f10SYehuda Sadeh int ceph_osdc_create_event(struct ceph_osd_client *osdc, 2197a40c4f10SYehuda Sadeh void (*event_cb)(u64, u64, u8, void *), 21983c663bbdSAlex Elder void *data, struct ceph_osd_event **pevent) 2199a40c4f10SYehuda Sadeh { 2200a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 2201a40c4f10SYehuda Sadeh 2202a40c4f10SYehuda Sadeh event = kmalloc(sizeof(*event), GFP_NOIO); 2203a40c4f10SYehuda Sadeh if (!event) 2204a40c4f10SYehuda Sadeh return -ENOMEM; 2205a40c4f10SYehuda Sadeh 2206a40c4f10SYehuda Sadeh dout("create_event %p\n", event); 2207a40c4f10SYehuda Sadeh event->cb = event_cb; 22083c663bbdSAlex Elder event->one_shot = 0; 2209a40c4f10SYehuda Sadeh event->data = data; 2210a40c4f10SYehuda Sadeh event->osdc = osdc; 2211a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&event->osd_node); 22123ee5234dSAlex Elder RB_CLEAR_NODE(&event->node); 2213a40c4f10SYehuda Sadeh kref_init(&event->kref); /* one ref for us */ 2214a40c4f10SYehuda Sadeh kref_get(&event->kref); /* one ref for the caller */ 2215a40c4f10SYehuda Sadeh 2216a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 2217a40c4f10SYehuda Sadeh event->cookie = ++osdc->event_count; 2218a40c4f10SYehuda Sadeh __insert_event(osdc, event); 2219a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 2220a40c4f10SYehuda Sadeh 2221a40c4f10SYehuda Sadeh *pevent = event; 2222a40c4f10SYehuda Sadeh return 0; 2223a40c4f10SYehuda Sadeh } 2224a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_create_event); 2225a40c4f10SYehuda Sadeh 2226a40c4f10SYehuda Sadeh void ceph_osdc_cancel_event(struct ceph_osd_event *event) 2227a40c4f10SYehuda Sadeh { 2228a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 2229a40c4f10SYehuda Sadeh 2230a40c4f10SYehuda Sadeh dout("cancel_event %p\n", event); 2231a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 2232a40c4f10SYehuda Sadeh __remove_event(event); 2233a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 2234a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); /* caller's */ 2235a40c4f10SYehuda Sadeh } 2236a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_cancel_event); 2237a40c4f10SYehuda Sadeh 2238a40c4f10SYehuda Sadeh 2239a40c4f10SYehuda Sadeh static void do_event_work(struct work_struct *work) 2240a40c4f10SYehuda Sadeh { 2241a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work = 2242a40c4f10SYehuda Sadeh container_of(work, struct ceph_osd_event_work, work); 2243a40c4f10SYehuda Sadeh struct ceph_osd_event *event = event_work->event; 2244a40c4f10SYehuda Sadeh u64 ver = event_work->ver; 2245a40c4f10SYehuda Sadeh u64 notify_id = event_work->notify_id; 2246a40c4f10SYehuda Sadeh u8 opcode = event_work->opcode; 2247a40c4f10SYehuda Sadeh 2248a40c4f10SYehuda Sadeh dout("do_event_work completing %p\n", event); 2249a40c4f10SYehuda Sadeh event->cb(ver, notify_id, opcode, event->data); 2250a40c4f10SYehuda Sadeh dout("do_event_work completed %p\n", event); 2251a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 2252a40c4f10SYehuda Sadeh kfree(event_work); 2253a40c4f10SYehuda Sadeh } 2254a40c4f10SYehuda Sadeh 2255a40c4f10SYehuda Sadeh 2256a40c4f10SYehuda Sadeh /* 2257a40c4f10SYehuda Sadeh * Process osd watch notifications 2258a40c4f10SYehuda Sadeh */ 22593c663bbdSAlex Elder static void handle_watch_notify(struct ceph_osd_client *osdc, 22603c663bbdSAlex Elder struct ceph_msg *msg) 2261a40c4f10SYehuda Sadeh { 2262a40c4f10SYehuda Sadeh void *p, *end; 2263a40c4f10SYehuda Sadeh u8 proto_ver; 2264a40c4f10SYehuda Sadeh u64 cookie, ver, notify_id; 2265a40c4f10SYehuda Sadeh u8 opcode; 2266a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 2267a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work; 2268a40c4f10SYehuda Sadeh 2269a40c4f10SYehuda Sadeh p = msg->front.iov_base; 2270a40c4f10SYehuda Sadeh end = p + msg->front.iov_len; 2271a40c4f10SYehuda Sadeh 2272a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, proto_ver, bad); 2273a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, opcode, bad); 2274a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, cookie, bad); 2275a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, ver, bad); 2276a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, notify_id, bad); 2277a40c4f10SYehuda Sadeh 2278a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 2279a40c4f10SYehuda Sadeh event = __find_event(osdc, cookie); 2280a40c4f10SYehuda Sadeh if (event) { 22813c663bbdSAlex Elder BUG_ON(event->one_shot); 2282a40c4f10SYehuda Sadeh get_event(event); 2283a40c4f10SYehuda Sadeh } 2284a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 2285a40c4f10SYehuda Sadeh dout("handle_watch_notify cookie %lld ver %lld event %p\n", 2286a40c4f10SYehuda Sadeh cookie, ver, event); 2287a40c4f10SYehuda Sadeh if (event) { 2288a40c4f10SYehuda Sadeh event_work = kmalloc(sizeof(*event_work), GFP_NOIO); 2289a40c4f10SYehuda Sadeh if (!event_work) { 2290a40c4f10SYehuda Sadeh dout("ERROR: could not allocate event_work\n"); 2291a40c4f10SYehuda Sadeh goto done_err; 2292a40c4f10SYehuda Sadeh } 22936b0ae409SMariusz Kozlowski INIT_WORK(&event_work->work, do_event_work); 2294a40c4f10SYehuda Sadeh event_work->event = event; 2295a40c4f10SYehuda Sadeh event_work->ver = ver; 2296a40c4f10SYehuda Sadeh event_work->notify_id = notify_id; 2297a40c4f10SYehuda Sadeh event_work->opcode = opcode; 2298a40c4f10SYehuda Sadeh if (!queue_work(osdc->notify_wq, &event_work->work)) { 2299a40c4f10SYehuda Sadeh dout("WARNING: failed to queue notify event work\n"); 2300a40c4f10SYehuda Sadeh goto done_err; 2301a40c4f10SYehuda Sadeh } 2302a40c4f10SYehuda Sadeh } 2303a40c4f10SYehuda Sadeh 2304a40c4f10SYehuda Sadeh return; 2305a40c4f10SYehuda Sadeh 2306a40c4f10SYehuda Sadeh done_err: 2307a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 2308a40c4f10SYehuda Sadeh return; 2309a40c4f10SYehuda Sadeh 2310a40c4f10SYehuda Sadeh bad: 2311a40c4f10SYehuda Sadeh pr_err("osdc handle_watch_notify corrupt msg\n"); 2312a40c4f10SYehuda Sadeh } 2313a40c4f10SYehuda Sadeh 2314a40c4f10SYehuda Sadeh /* 2315e65550fdSAlex Elder * build new request AND message 2316e65550fdSAlex Elder * 2317e65550fdSAlex Elder */ 2318e65550fdSAlex Elder void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off, 2319e65550fdSAlex Elder struct ceph_snap_context *snapc, u64 snap_id, 2320e65550fdSAlex Elder struct timespec *mtime) 2321e65550fdSAlex Elder { 2322e65550fdSAlex Elder struct ceph_msg *msg = req->r_request; 2323e65550fdSAlex Elder void *p; 2324e65550fdSAlex Elder size_t msg_size; 2325e65550fdSAlex Elder int flags = req->r_flags; 2326e65550fdSAlex Elder u64 data_len; 2327e65550fdSAlex Elder unsigned int i; 2328e65550fdSAlex Elder 2329e65550fdSAlex Elder req->r_snapid = snap_id; 2330e65550fdSAlex Elder req->r_snapc = ceph_get_snap_context(snapc); 2331e65550fdSAlex Elder 2332e65550fdSAlex Elder /* encode request */ 2333e65550fdSAlex Elder msg->hdr.version = cpu_to_le16(4); 2334e65550fdSAlex Elder 2335e65550fdSAlex Elder p = msg->front.iov_base; 2336e65550fdSAlex Elder ceph_encode_32(&p, 1); /* client_inc is always 1 */ 2337e65550fdSAlex Elder req->r_request_osdmap_epoch = p; 2338e65550fdSAlex Elder p += 4; 2339e65550fdSAlex Elder req->r_request_flags = p; 2340e65550fdSAlex Elder p += 4; 2341e65550fdSAlex Elder if (req->r_flags & CEPH_OSD_FLAG_WRITE) 2342e65550fdSAlex Elder ceph_encode_timespec(p, mtime); 2343e65550fdSAlex Elder p += sizeof(struct ceph_timespec); 2344e65550fdSAlex Elder req->r_request_reassert_version = p; 2345e65550fdSAlex Elder p += sizeof(struct ceph_eversion); /* will get filled in */ 2346e65550fdSAlex Elder 2347e65550fdSAlex Elder /* oloc */ 2348e65550fdSAlex Elder ceph_encode_8(&p, 4); 2349e65550fdSAlex Elder ceph_encode_8(&p, 4); 2350e65550fdSAlex Elder ceph_encode_32(&p, 8 + 4 + 4); 2351e65550fdSAlex Elder req->r_request_pool = p; 2352e65550fdSAlex Elder p += 8; 2353e65550fdSAlex Elder ceph_encode_32(&p, -1); /* preferred */ 2354e65550fdSAlex Elder ceph_encode_32(&p, 0); /* key len */ 2355e65550fdSAlex Elder 2356e65550fdSAlex Elder ceph_encode_8(&p, 1); 2357e65550fdSAlex Elder req->r_request_pgid = p; 2358e65550fdSAlex Elder p += 8 + 4; 2359e65550fdSAlex Elder ceph_encode_32(&p, -1); /* preferred */ 2360e65550fdSAlex Elder 2361e65550fdSAlex Elder /* oid */ 23623c972c95SIlya Dryomov ceph_encode_32(&p, req->r_base_oid.name_len); 23633c972c95SIlya Dryomov memcpy(p, req->r_base_oid.name, req->r_base_oid.name_len); 23643c972c95SIlya Dryomov dout("oid '%.*s' len %d\n", req->r_base_oid.name_len, 23653c972c95SIlya Dryomov req->r_base_oid.name, req->r_base_oid.name_len); 23663c972c95SIlya Dryomov p += req->r_base_oid.name_len; 2367e65550fdSAlex Elder 2368e65550fdSAlex Elder /* ops--can imply data */ 2369e65550fdSAlex Elder ceph_encode_16(&p, (u16)req->r_num_ops); 2370e65550fdSAlex Elder data_len = 0; 2371e65550fdSAlex Elder for (i = 0; i < req->r_num_ops; i++) { 2372e65550fdSAlex Elder data_len += osd_req_encode_op(req, p, i); 2373e65550fdSAlex Elder p += sizeof(struct ceph_osd_op); 2374e65550fdSAlex Elder } 2375e65550fdSAlex Elder 2376e65550fdSAlex Elder /* snaps */ 2377e65550fdSAlex Elder ceph_encode_64(&p, req->r_snapid); 2378e65550fdSAlex Elder ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0); 2379e65550fdSAlex Elder ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0); 2380e65550fdSAlex Elder if (req->r_snapc) { 2381e65550fdSAlex Elder for (i = 0; i < snapc->num_snaps; i++) { 2382e65550fdSAlex Elder ceph_encode_64(&p, req->r_snapc->snaps[i]); 2383e65550fdSAlex Elder } 2384e65550fdSAlex Elder } 2385e65550fdSAlex Elder 2386e65550fdSAlex Elder req->r_request_attempts = p; 2387e65550fdSAlex Elder p += 4; 2388e65550fdSAlex Elder 2389e65550fdSAlex Elder /* data */ 2390e65550fdSAlex Elder if (flags & CEPH_OSD_FLAG_WRITE) { 2391e65550fdSAlex Elder u16 data_off; 2392e65550fdSAlex Elder 2393e65550fdSAlex Elder /* 2394e65550fdSAlex Elder * The header "data_off" is a hint to the receiver 2395e65550fdSAlex Elder * allowing it to align received data into its 2396e65550fdSAlex Elder * buffers such that there's no need to re-copy 2397e65550fdSAlex Elder * it before writing it to disk (direct I/O). 2398e65550fdSAlex Elder */ 2399e65550fdSAlex Elder data_off = (u16) (off & 0xffff); 2400e65550fdSAlex Elder req->r_request->hdr.data_off = cpu_to_le16(data_off); 2401e65550fdSAlex Elder } 2402e65550fdSAlex Elder req->r_request->hdr.data_len = cpu_to_le32(data_len); 2403e65550fdSAlex Elder 2404e65550fdSAlex Elder BUG_ON(p > msg->front.iov_base + msg->front.iov_len); 2405e65550fdSAlex Elder msg_size = p - msg->front.iov_base; 2406e65550fdSAlex Elder msg->front.iov_len = msg_size; 2407e65550fdSAlex Elder msg->hdr.front_len = cpu_to_le32(msg_size); 2408e65550fdSAlex Elder 2409e65550fdSAlex Elder dout("build_request msg_size was %d\n", (int)msg_size); 2410e65550fdSAlex Elder } 2411e65550fdSAlex Elder EXPORT_SYMBOL(ceph_osdc_build_request); 2412e65550fdSAlex Elder 2413e65550fdSAlex Elder /* 24143d14c5d2SYehuda Sadeh * Register request, send initial attempt. 24153d14c5d2SYehuda Sadeh */ 24163d14c5d2SYehuda Sadeh int ceph_osdc_start_request(struct ceph_osd_client *osdc, 24173d14c5d2SYehuda Sadeh struct ceph_osd_request *req, 24183d14c5d2SYehuda Sadeh bool nofail) 24193d14c5d2SYehuda Sadeh { 24200bbfdfe8SIlya Dryomov int rc; 24213d14c5d2SYehuda Sadeh 24223d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 24233d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 24240bbfdfe8SIlya Dryomov 24250bbfdfe8SIlya Dryomov rc = __ceph_osdc_start_request(osdc, req, nofail); 24260bbfdfe8SIlya Dryomov 24273d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 24283d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 24290bbfdfe8SIlya Dryomov 24303d14c5d2SYehuda Sadeh return rc; 24313d14c5d2SYehuda Sadeh } 24323d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_start_request); 24333d14c5d2SYehuda Sadeh 24343d14c5d2SYehuda Sadeh /* 24353d14c5d2SYehuda Sadeh * wait for a request to complete 24363d14c5d2SYehuda Sadeh */ 24373d14c5d2SYehuda Sadeh int ceph_osdc_wait_request(struct ceph_osd_client *osdc, 24383d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 24393d14c5d2SYehuda Sadeh { 24403d14c5d2SYehuda Sadeh int rc; 24413d14c5d2SYehuda Sadeh 24423d14c5d2SYehuda Sadeh rc = wait_for_completion_interruptible(&req->r_completion); 24433d14c5d2SYehuda Sadeh if (rc < 0) { 24443d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 24453d14c5d2SYehuda Sadeh __cancel_request(req); 24463d14c5d2SYehuda Sadeh __unregister_request(osdc, req); 24473d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 244825845472SSage Weil complete_request(req); 24493d14c5d2SYehuda Sadeh dout("wait_request tid %llu canceled/timed out\n", req->r_tid); 24503d14c5d2SYehuda Sadeh return rc; 24513d14c5d2SYehuda Sadeh } 24523d14c5d2SYehuda Sadeh 24533d14c5d2SYehuda Sadeh dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result); 24543d14c5d2SYehuda Sadeh return req->r_result; 24553d14c5d2SYehuda Sadeh } 24563d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_wait_request); 24573d14c5d2SYehuda Sadeh 24583d14c5d2SYehuda Sadeh /* 24593d14c5d2SYehuda Sadeh * sync - wait for all in-flight requests to flush. avoid starvation. 24603d14c5d2SYehuda Sadeh */ 24613d14c5d2SYehuda Sadeh void ceph_osdc_sync(struct ceph_osd_client *osdc) 24623d14c5d2SYehuda Sadeh { 24633d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 24643d14c5d2SYehuda Sadeh u64 last_tid, next_tid = 0; 24653d14c5d2SYehuda Sadeh 24663d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 24673d14c5d2SYehuda Sadeh last_tid = osdc->last_tid; 24683d14c5d2SYehuda Sadeh while (1) { 24693d14c5d2SYehuda Sadeh req = __lookup_request_ge(osdc, next_tid); 24703d14c5d2SYehuda Sadeh if (!req) 24713d14c5d2SYehuda Sadeh break; 24723d14c5d2SYehuda Sadeh if (req->r_tid > last_tid) 24733d14c5d2SYehuda Sadeh break; 24743d14c5d2SYehuda Sadeh 24753d14c5d2SYehuda Sadeh next_tid = req->r_tid + 1; 24763d14c5d2SYehuda Sadeh if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0) 24773d14c5d2SYehuda Sadeh continue; 24783d14c5d2SYehuda Sadeh 24793d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 24803d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 24813d14c5d2SYehuda Sadeh dout("sync waiting on tid %llu (last is %llu)\n", 24823d14c5d2SYehuda Sadeh req->r_tid, last_tid); 24833d14c5d2SYehuda Sadeh wait_for_completion(&req->r_safe_completion); 24843d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 24853d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 24863d14c5d2SYehuda Sadeh } 24873d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 24883d14c5d2SYehuda Sadeh dout("sync done (thru tid %llu)\n", last_tid); 24893d14c5d2SYehuda Sadeh } 24903d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_sync); 24913d14c5d2SYehuda Sadeh 24923d14c5d2SYehuda Sadeh /* 2493dd935f44SJosh Durgin * Call all pending notify callbacks - for use after a watch is 2494dd935f44SJosh Durgin * unregistered, to make sure no more callbacks for it will be invoked 2495dd935f44SJosh Durgin */ 2496f6479449Sstephen hemminger void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc) 2497dd935f44SJosh Durgin { 2498dd935f44SJosh Durgin flush_workqueue(osdc->notify_wq); 2499dd935f44SJosh Durgin } 2500dd935f44SJosh Durgin EXPORT_SYMBOL(ceph_osdc_flush_notifies); 2501dd935f44SJosh Durgin 2502dd935f44SJosh Durgin 2503dd935f44SJosh Durgin /* 25043d14c5d2SYehuda Sadeh * init, shutdown 25053d14c5d2SYehuda Sadeh */ 25063d14c5d2SYehuda Sadeh int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client) 25073d14c5d2SYehuda Sadeh { 25083d14c5d2SYehuda Sadeh int err; 25093d14c5d2SYehuda Sadeh 25103d14c5d2SYehuda Sadeh dout("init\n"); 25113d14c5d2SYehuda Sadeh osdc->client = client; 25123d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 25133d14c5d2SYehuda Sadeh init_rwsem(&osdc->map_sem); 25143d14c5d2SYehuda Sadeh init_completion(&osdc->map_waiters); 25153d14c5d2SYehuda Sadeh osdc->last_requested_map = 0; 25163d14c5d2SYehuda Sadeh mutex_init(&osdc->request_mutex); 25173d14c5d2SYehuda Sadeh osdc->last_tid = 0; 25183d14c5d2SYehuda Sadeh osdc->osds = RB_ROOT; 25193d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->osd_lru); 25203d14c5d2SYehuda Sadeh osdc->requests = RB_ROOT; 25213d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_lru); 25226f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_unsent); 25236f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_notarget); 2524a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_linger); 25253d14c5d2SYehuda Sadeh osdc->num_requests = 0; 25263d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout); 25273d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout); 2528a40c4f10SYehuda Sadeh spin_lock_init(&osdc->event_lock); 2529a40c4f10SYehuda Sadeh osdc->event_tree = RB_ROOT; 2530a40c4f10SYehuda Sadeh osdc->event_count = 0; 25313d14c5d2SYehuda Sadeh 25323d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 25333d14c5d2SYehuda Sadeh round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ)); 25343d14c5d2SYehuda Sadeh 25353d14c5d2SYehuda Sadeh err = -ENOMEM; 25363d14c5d2SYehuda Sadeh osdc->req_mempool = mempool_create_kmalloc_pool(10, 25373d14c5d2SYehuda Sadeh sizeof(struct ceph_osd_request)); 25383d14c5d2SYehuda Sadeh if (!osdc->req_mempool) 25393d14c5d2SYehuda Sadeh goto out; 25403d14c5d2SYehuda Sadeh 2541d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP, 2542d50b409fSSage Weil OSD_OP_FRONT_LEN, 10, true, 25433d14c5d2SYehuda Sadeh "osd_op"); 25443d14c5d2SYehuda Sadeh if (err < 0) 25453d14c5d2SYehuda Sadeh goto out_mempool; 2546d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY, 25473d14c5d2SYehuda Sadeh OSD_OPREPLY_FRONT_LEN, 10, true, 25483d14c5d2SYehuda Sadeh "osd_op_reply"); 25493d14c5d2SYehuda Sadeh if (err < 0) 25503d14c5d2SYehuda Sadeh goto out_msgpool; 2551a40c4f10SYehuda Sadeh 2552dbcae088SDan Carpenter err = -ENOMEM; 2553a40c4f10SYehuda Sadeh osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify"); 2554dbcae088SDan Carpenter if (!osdc->notify_wq) 2555c172ec5cSIlya Dryomov goto out_msgpool_reply; 2556c172ec5cSIlya Dryomov 25573d14c5d2SYehuda Sadeh return 0; 25583d14c5d2SYehuda Sadeh 2559c172ec5cSIlya Dryomov out_msgpool_reply: 2560c172ec5cSIlya Dryomov ceph_msgpool_destroy(&osdc->msgpool_op_reply); 25613d14c5d2SYehuda Sadeh out_msgpool: 25623d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 25633d14c5d2SYehuda Sadeh out_mempool: 25643d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 25653d14c5d2SYehuda Sadeh out: 25663d14c5d2SYehuda Sadeh return err; 25673d14c5d2SYehuda Sadeh } 25683d14c5d2SYehuda Sadeh 25693d14c5d2SYehuda Sadeh void ceph_osdc_stop(struct ceph_osd_client *osdc) 25703d14c5d2SYehuda Sadeh { 2571a40c4f10SYehuda Sadeh flush_workqueue(osdc->notify_wq); 2572a40c4f10SYehuda Sadeh destroy_workqueue(osdc->notify_wq); 25733d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->timeout_work); 25743d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->osds_timeout_work); 25753d14c5d2SYehuda Sadeh if (osdc->osdmap) { 25763d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 25773d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 25783d14c5d2SYehuda Sadeh } 2579aca420bcSSage Weil remove_all_osds(osdc); 25803d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 25813d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 25823d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op_reply); 25833d14c5d2SYehuda Sadeh } 25843d14c5d2SYehuda Sadeh 25853d14c5d2SYehuda Sadeh /* 25863d14c5d2SYehuda Sadeh * Read some contiguous pages. If we cross a stripe boundary, shorten 25873d14c5d2SYehuda Sadeh * *plen. Return number of bytes read, or error. 25883d14c5d2SYehuda Sadeh */ 25893d14c5d2SYehuda Sadeh int ceph_osdc_readpages(struct ceph_osd_client *osdc, 25903d14c5d2SYehuda Sadeh struct ceph_vino vino, struct ceph_file_layout *layout, 25913d14c5d2SYehuda Sadeh u64 off, u64 *plen, 25923d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 2593b7495fc2SSage Weil struct page **pages, int num_pages, int page_align) 25943d14c5d2SYehuda Sadeh { 25953d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 25963d14c5d2SYehuda Sadeh int rc = 0; 25973d14c5d2SYehuda Sadeh 25983d14c5d2SYehuda Sadeh dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino, 25993d14c5d2SYehuda Sadeh vino.snap, off, *plen); 260079528734SAlex Elder req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 1, 26013d14c5d2SYehuda Sadeh CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ, 2602acead002SAlex Elder NULL, truncate_seq, truncate_size, 2603153e5167SAlex Elder false); 26046816282dSSage Weil if (IS_ERR(req)) 26056816282dSSage Weil return PTR_ERR(req); 26063d14c5d2SYehuda Sadeh 26073d14c5d2SYehuda Sadeh /* it may be a short read due to an object boundary */ 26080fff87ecSAlex Elder 2609406e2c9fSAlex Elder osd_req_op_extent_osd_data_pages(req, 0, 2610a4ce40a9SAlex Elder pages, *plen, page_align, false, false); 26113d14c5d2SYehuda Sadeh 2612e0c59487SAlex Elder dout("readpages final extent is %llu~%llu (%llu bytes align %d)\n", 261343bfe5deSAlex Elder off, *plen, *plen, page_align); 26143d14c5d2SYehuda Sadeh 261579528734SAlex Elder ceph_osdc_build_request(req, off, NULL, vino.snap, NULL); 261602ee07d3SAlex Elder 26173d14c5d2SYehuda Sadeh rc = ceph_osdc_start_request(osdc, req, false); 26183d14c5d2SYehuda Sadeh if (!rc) 26193d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 26203d14c5d2SYehuda Sadeh 26213d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 26223d14c5d2SYehuda Sadeh dout("readpages result %d\n", rc); 26233d14c5d2SYehuda Sadeh return rc; 26243d14c5d2SYehuda Sadeh } 26253d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_readpages); 26263d14c5d2SYehuda Sadeh 26273d14c5d2SYehuda Sadeh /* 26283d14c5d2SYehuda Sadeh * do a synchronous write on N pages 26293d14c5d2SYehuda Sadeh */ 26303d14c5d2SYehuda Sadeh int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino, 26313d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 26323d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 26333d14c5d2SYehuda Sadeh u64 off, u64 len, 26343d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 26353d14c5d2SYehuda Sadeh struct timespec *mtime, 263624808826SAlex Elder struct page **pages, int num_pages) 26373d14c5d2SYehuda Sadeh { 26383d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 26393d14c5d2SYehuda Sadeh int rc = 0; 2640b7495fc2SSage Weil int page_align = off & ~PAGE_MASK; 26413d14c5d2SYehuda Sadeh 2642acead002SAlex Elder BUG_ON(vino.snap != CEPH_NOSNAP); /* snapshots aren't writeable */ 264379528734SAlex Elder req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 1, 26443d14c5d2SYehuda Sadeh CEPH_OSD_OP_WRITE, 264524808826SAlex Elder CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE, 2646acead002SAlex Elder snapc, truncate_seq, truncate_size, 2647153e5167SAlex Elder true); 26486816282dSSage Weil if (IS_ERR(req)) 26496816282dSSage Weil return PTR_ERR(req); 26503d14c5d2SYehuda Sadeh 26513d14c5d2SYehuda Sadeh /* it may be a short write due to an object boundary */ 2652406e2c9fSAlex Elder osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align, 265343bfe5deSAlex Elder false, false); 265443bfe5deSAlex Elder dout("writepages %llu~%llu (%llu bytes)\n", off, len, len); 26553d14c5d2SYehuda Sadeh 265679528734SAlex Elder ceph_osdc_build_request(req, off, snapc, CEPH_NOSNAP, mtime); 265702ee07d3SAlex Elder 265887f979d3SAlex Elder rc = ceph_osdc_start_request(osdc, req, true); 26593d14c5d2SYehuda Sadeh if (!rc) 26603d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 26613d14c5d2SYehuda Sadeh 26623d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 26633d14c5d2SYehuda Sadeh if (rc == 0) 26643d14c5d2SYehuda Sadeh rc = len; 26653d14c5d2SYehuda Sadeh dout("writepages result %d\n", rc); 26663d14c5d2SYehuda Sadeh return rc; 26673d14c5d2SYehuda Sadeh } 26683d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_writepages); 26693d14c5d2SYehuda Sadeh 26705522ae0bSAlex Elder int ceph_osdc_setup(void) 26715522ae0bSAlex Elder { 26725522ae0bSAlex Elder BUG_ON(ceph_osd_request_cache); 26735522ae0bSAlex Elder ceph_osd_request_cache = kmem_cache_create("ceph_osd_request", 26745522ae0bSAlex Elder sizeof (struct ceph_osd_request), 26755522ae0bSAlex Elder __alignof__(struct ceph_osd_request), 26765522ae0bSAlex Elder 0, NULL); 26775522ae0bSAlex Elder 26785522ae0bSAlex Elder return ceph_osd_request_cache ? 0 : -ENOMEM; 26795522ae0bSAlex Elder } 26805522ae0bSAlex Elder EXPORT_SYMBOL(ceph_osdc_setup); 26815522ae0bSAlex Elder 26825522ae0bSAlex Elder void ceph_osdc_cleanup(void) 26835522ae0bSAlex Elder { 26845522ae0bSAlex Elder BUG_ON(!ceph_osd_request_cache); 26855522ae0bSAlex Elder kmem_cache_destroy(ceph_osd_request_cache); 26865522ae0bSAlex Elder ceph_osd_request_cache = NULL; 26875522ae0bSAlex Elder } 26885522ae0bSAlex Elder EXPORT_SYMBOL(ceph_osdc_cleanup); 26895522ae0bSAlex Elder 26903d14c5d2SYehuda Sadeh /* 26913d14c5d2SYehuda Sadeh * handle incoming message 26923d14c5d2SYehuda Sadeh */ 26933d14c5d2SYehuda Sadeh static void dispatch(struct ceph_connection *con, struct ceph_msg *msg) 26943d14c5d2SYehuda Sadeh { 26953d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 26963d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 26973d14c5d2SYehuda Sadeh int type = le16_to_cpu(msg->hdr.type); 26983d14c5d2SYehuda Sadeh 26993d14c5d2SYehuda Sadeh if (!osd) 27003d14c5d2SYehuda Sadeh goto out; 27013d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 27023d14c5d2SYehuda Sadeh 27033d14c5d2SYehuda Sadeh switch (type) { 27043d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 27053d14c5d2SYehuda Sadeh ceph_osdc_handle_map(osdc, msg); 27063d14c5d2SYehuda Sadeh break; 27073d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 27083d14c5d2SYehuda Sadeh handle_reply(osdc, msg, con); 27093d14c5d2SYehuda Sadeh break; 2710a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 2711a40c4f10SYehuda Sadeh handle_watch_notify(osdc, msg); 2712a40c4f10SYehuda Sadeh break; 27133d14c5d2SYehuda Sadeh 27143d14c5d2SYehuda Sadeh default: 27153d14c5d2SYehuda Sadeh pr_err("received unknown message type %d %s\n", type, 27163d14c5d2SYehuda Sadeh ceph_msg_type_name(type)); 27173d14c5d2SYehuda Sadeh } 27183d14c5d2SYehuda Sadeh out: 27193d14c5d2SYehuda Sadeh ceph_msg_put(msg); 27203d14c5d2SYehuda Sadeh } 27213d14c5d2SYehuda Sadeh 27223d14c5d2SYehuda Sadeh /* 27233d14c5d2SYehuda Sadeh * lookup and return message for incoming reply. set up reply message 27243d14c5d2SYehuda Sadeh * pages. 27253d14c5d2SYehuda Sadeh */ 27263d14c5d2SYehuda Sadeh static struct ceph_msg *get_reply(struct ceph_connection *con, 27273d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 27283d14c5d2SYehuda Sadeh int *skip) 27293d14c5d2SYehuda Sadeh { 27303d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 27313d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = osd->o_osdc; 27323d14c5d2SYehuda Sadeh struct ceph_msg *m; 27333d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 27343f0a4ac5SIlya Dryomov int front_len = le32_to_cpu(hdr->front_len); 27353d14c5d2SYehuda Sadeh int data_len = le32_to_cpu(hdr->data_len); 27363d14c5d2SYehuda Sadeh u64 tid; 27373d14c5d2SYehuda Sadeh 27383d14c5d2SYehuda Sadeh tid = le64_to_cpu(hdr->tid); 27393d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 27403d14c5d2SYehuda Sadeh req = __lookup_request(osdc, tid); 27413d14c5d2SYehuda Sadeh if (!req) { 27423d14c5d2SYehuda Sadeh *skip = 1; 27433d14c5d2SYehuda Sadeh m = NULL; 2744756a16a5SSage Weil dout("get_reply unknown tid %llu from osd%d\n", tid, 27453d14c5d2SYehuda Sadeh osd->o_osd); 27463d14c5d2SYehuda Sadeh goto out; 27473d14c5d2SYehuda Sadeh } 27483d14c5d2SYehuda Sadeh 2749ace6d3a9SAlex Elder if (req->r_reply->con) 27508921d114SAlex Elder dout("%s revoking msg %p from old con %p\n", __func__, 2751ace6d3a9SAlex Elder req->r_reply, req->r_reply->con); 27528921d114SAlex Elder ceph_msg_revoke_incoming(req->r_reply); 27533d14c5d2SYehuda Sadeh 2754f2be82b0SIlya Dryomov if (front_len > req->r_reply->front_alloc_len) { 27554974341eSAlex Elder pr_warning("get_reply front %d > preallocated %d (%u#%llu)\n", 2756f2be82b0SIlya Dryomov front_len, req->r_reply->front_alloc_len, 27574974341eSAlex Elder (unsigned int)con->peer_name.type, 27584974341eSAlex Elder le64_to_cpu(con->peer_name.num)); 27593f0a4ac5SIlya Dryomov m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front_len, GFP_NOFS, 27603f0a4ac5SIlya Dryomov false); 27613d14c5d2SYehuda Sadeh if (!m) 27623d14c5d2SYehuda Sadeh goto out; 27633d14c5d2SYehuda Sadeh ceph_msg_put(req->r_reply); 27643d14c5d2SYehuda Sadeh req->r_reply = m; 27653d14c5d2SYehuda Sadeh } 27663d14c5d2SYehuda Sadeh m = ceph_msg_get(req->r_reply); 27673d14c5d2SYehuda Sadeh 27683d14c5d2SYehuda Sadeh if (data_len > 0) { 2769a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 27700fff87ecSAlex Elder 2771a4ce40a9SAlex Elder /* 2772a4ce40a9SAlex Elder * XXX This is assuming there is only one op containing 2773a4ce40a9SAlex Elder * XXX page data. Probably OK for reads, but this 2774a4ce40a9SAlex Elder * XXX ought to be done more generally. 2775a4ce40a9SAlex Elder */ 2776406e2c9fSAlex Elder osd_data = osd_req_op_extent_osd_data(req, 0); 27770fff87ecSAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 27780fff87ecSAlex Elder if (osd_data->pages && 2779e0c59487SAlex Elder unlikely(osd_data->length < data_len)) { 27802ac2b7a6SAlex Elder 2781e0c59487SAlex Elder pr_warning("tid %lld reply has %d bytes " 2782e0c59487SAlex Elder "we had only %llu bytes ready\n", 2783e0c59487SAlex Elder tid, data_len, osd_data->length); 27843d14c5d2SYehuda Sadeh *skip = 1; 27853d14c5d2SYehuda Sadeh ceph_msg_put(m); 27863d14c5d2SYehuda Sadeh m = NULL; 27873d14c5d2SYehuda Sadeh goto out; 27883d14c5d2SYehuda Sadeh } 27893d14c5d2SYehuda Sadeh } 27902ac2b7a6SAlex Elder } 27913d14c5d2SYehuda Sadeh *skip = 0; 27923d14c5d2SYehuda Sadeh dout("get_reply tid %lld %p\n", tid, m); 27933d14c5d2SYehuda Sadeh 27943d14c5d2SYehuda Sadeh out: 27953d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 27963d14c5d2SYehuda Sadeh return m; 27973d14c5d2SYehuda Sadeh 27983d14c5d2SYehuda Sadeh } 27993d14c5d2SYehuda Sadeh 28003d14c5d2SYehuda Sadeh static struct ceph_msg *alloc_msg(struct ceph_connection *con, 28013d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 28023d14c5d2SYehuda Sadeh int *skip) 28033d14c5d2SYehuda Sadeh { 28043d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 28053d14c5d2SYehuda Sadeh int type = le16_to_cpu(hdr->type); 28063d14c5d2SYehuda Sadeh int front = le32_to_cpu(hdr->front_len); 28073d14c5d2SYehuda Sadeh 28081c20f2d2SAlex Elder *skip = 0; 28093d14c5d2SYehuda Sadeh switch (type) { 28103d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 2811a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 2812b61c2763SSage Weil return ceph_msg_new(type, front, GFP_NOFS, false); 28133d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 28143d14c5d2SYehuda Sadeh return get_reply(con, hdr, skip); 28153d14c5d2SYehuda Sadeh default: 28163d14c5d2SYehuda Sadeh pr_info("alloc_msg unexpected msg type %d from osd%d\n", type, 28173d14c5d2SYehuda Sadeh osd->o_osd); 28183d14c5d2SYehuda Sadeh *skip = 1; 28193d14c5d2SYehuda Sadeh return NULL; 28203d14c5d2SYehuda Sadeh } 28213d14c5d2SYehuda Sadeh } 28223d14c5d2SYehuda Sadeh 28233d14c5d2SYehuda Sadeh /* 28243d14c5d2SYehuda Sadeh * Wrappers to refcount containing ceph_osd struct 28253d14c5d2SYehuda Sadeh */ 28263d14c5d2SYehuda Sadeh static struct ceph_connection *get_osd_con(struct ceph_connection *con) 28273d14c5d2SYehuda Sadeh { 28283d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 28293d14c5d2SYehuda Sadeh if (get_osd(osd)) 28303d14c5d2SYehuda Sadeh return con; 28313d14c5d2SYehuda Sadeh return NULL; 28323d14c5d2SYehuda Sadeh } 28333d14c5d2SYehuda Sadeh 28343d14c5d2SYehuda Sadeh static void put_osd_con(struct ceph_connection *con) 28353d14c5d2SYehuda Sadeh { 28363d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 28373d14c5d2SYehuda Sadeh put_osd(osd); 28383d14c5d2SYehuda Sadeh } 28393d14c5d2SYehuda Sadeh 28403d14c5d2SYehuda Sadeh /* 28413d14c5d2SYehuda Sadeh * authentication 28423d14c5d2SYehuda Sadeh */ 2843a3530df3SAlex Elder /* 2844a3530df3SAlex Elder * Note: returned pointer is the address of a structure that's 2845a3530df3SAlex Elder * managed separately. Caller must *not* attempt to free it. 2846a3530df3SAlex Elder */ 2847a3530df3SAlex Elder static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con, 28488f43fb53SAlex Elder int *proto, int force_new) 28493d14c5d2SYehuda Sadeh { 28503d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 28513d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 28523d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 285374f1869fSAlex Elder struct ceph_auth_handshake *auth = &o->o_auth; 28543d14c5d2SYehuda Sadeh 285574f1869fSAlex Elder if (force_new && auth->authorizer) { 285627859f97SSage Weil ceph_auth_destroy_authorizer(ac, auth->authorizer); 285774f1869fSAlex Elder auth->authorizer = NULL; 28583d14c5d2SYehuda Sadeh } 285927859f97SSage Weil if (!auth->authorizer) { 286027859f97SSage Weil int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 2861a3530df3SAlex Elder auth); 28623d14c5d2SYehuda Sadeh if (ret) 2863a3530df3SAlex Elder return ERR_PTR(ret); 286427859f97SSage Weil } else { 286527859f97SSage Weil int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 28660bed9b5cSSage Weil auth); 28670bed9b5cSSage Weil if (ret) 28680bed9b5cSSage Weil return ERR_PTR(ret); 28693d14c5d2SYehuda Sadeh } 28703d14c5d2SYehuda Sadeh *proto = ac->protocol; 287174f1869fSAlex Elder 2872a3530df3SAlex Elder return auth; 28733d14c5d2SYehuda Sadeh } 28743d14c5d2SYehuda Sadeh 28753d14c5d2SYehuda Sadeh 28763d14c5d2SYehuda Sadeh static int verify_authorizer_reply(struct ceph_connection *con, int len) 28773d14c5d2SYehuda Sadeh { 28783d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 28793d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 28803d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 28813d14c5d2SYehuda Sadeh 288227859f97SSage Weil return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len); 28833d14c5d2SYehuda Sadeh } 28843d14c5d2SYehuda Sadeh 28853d14c5d2SYehuda Sadeh static int invalidate_authorizer(struct ceph_connection *con) 28863d14c5d2SYehuda Sadeh { 28873d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 28883d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 28893d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 28903d14c5d2SYehuda Sadeh 289127859f97SSage Weil ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD); 28923d14c5d2SYehuda Sadeh return ceph_monc_validate_auth(&osdc->client->monc); 28933d14c5d2SYehuda Sadeh } 28943d14c5d2SYehuda Sadeh 28953d14c5d2SYehuda Sadeh static const struct ceph_connection_operations osd_con_ops = { 28963d14c5d2SYehuda Sadeh .get = get_osd_con, 28973d14c5d2SYehuda Sadeh .put = put_osd_con, 28983d14c5d2SYehuda Sadeh .dispatch = dispatch, 28993d14c5d2SYehuda Sadeh .get_authorizer = get_authorizer, 29003d14c5d2SYehuda Sadeh .verify_authorizer_reply = verify_authorizer_reply, 29013d14c5d2SYehuda Sadeh .invalidate_authorizer = invalidate_authorizer, 29023d14c5d2SYehuda Sadeh .alloc_msg = alloc_msg, 29033d14c5d2SYehuda Sadeh .fault = osd_reset, 29043d14c5d2SYehuda Sadeh }; 2905