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 */ 3411b83bef2SSage Weil msg_size += 4 + MAX_OBJ_NAME_SIZE; 3421b83bef2SSage Weil msg_size += 2 + num_ops*sizeof(struct ceph_osd_op); 3431b83bef2SSage Weil msg_size += 8; /* snapid */ 3441b83bef2SSage Weil msg_size += 8; /* snap_seq */ 3451b83bef2SSage Weil msg_size += 8 * (snapc ? snapc->num_snaps : 0); /* snaps */ 3461b83bef2SSage Weil msg_size += 4; 3473d14c5d2SYehuda Sadeh 3483d14c5d2SYehuda Sadeh if (use_mempool) { 3493d14c5d2SYehuda Sadeh req = mempool_alloc(osdc->req_mempool, gfp_flags); 3503d14c5d2SYehuda Sadeh memset(req, 0, sizeof(*req)); 3513d14c5d2SYehuda Sadeh } else { 3525522ae0bSAlex Elder req = kmem_cache_zalloc(ceph_osd_request_cache, gfp_flags); 3533d14c5d2SYehuda Sadeh } 3543d14c5d2SYehuda Sadeh if (req == NULL) 3553d14c5d2SYehuda Sadeh return NULL; 3563d14c5d2SYehuda Sadeh 3573d14c5d2SYehuda Sadeh req->r_osdc = osdc; 3583d14c5d2SYehuda Sadeh req->r_mempool = use_mempool; 35979528734SAlex Elder req->r_num_ops = num_ops; 3603d14c5d2SYehuda Sadeh 3613d14c5d2SYehuda Sadeh kref_init(&req->r_kref); 3623d14c5d2SYehuda Sadeh init_completion(&req->r_completion); 3633d14c5d2SYehuda Sadeh init_completion(&req->r_safe_completion); 364a978fa20SAlex Elder RB_CLEAR_NODE(&req->r_node); 3653d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&req->r_unsafe_item); 366a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&req->r_linger_item); 367a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&req->r_linger_osd); 368935b639aSSage Weil INIT_LIST_HEAD(&req->r_req_lru_item); 369cd43045cSSage Weil INIT_LIST_HEAD(&req->r_osd_item); 370cd43045cSSage Weil 3713d14c5d2SYehuda Sadeh /* create reply message */ 3723d14c5d2SYehuda Sadeh if (use_mempool) 3733d14c5d2SYehuda Sadeh msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0); 3743d14c5d2SYehuda Sadeh else 3753d14c5d2SYehuda Sadeh msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, 376b61c2763SSage Weil OSD_OPREPLY_FRONT_LEN, gfp_flags, true); 3773d14c5d2SYehuda Sadeh if (!msg) { 3783d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 3793d14c5d2SYehuda Sadeh return NULL; 3803d14c5d2SYehuda Sadeh } 3813d14c5d2SYehuda Sadeh req->r_reply = msg; 3823d14c5d2SYehuda Sadeh 3833d14c5d2SYehuda Sadeh /* create request message; allow space for oid */ 3843d14c5d2SYehuda Sadeh if (use_mempool) 3853d14c5d2SYehuda Sadeh msg = ceph_msgpool_get(&osdc->msgpool_op, 0); 3863d14c5d2SYehuda Sadeh else 387b61c2763SSage Weil msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true); 3883d14c5d2SYehuda Sadeh if (!msg) { 3893d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 3903d14c5d2SYehuda Sadeh return NULL; 3913d14c5d2SYehuda Sadeh } 3923d14c5d2SYehuda Sadeh 3933d14c5d2SYehuda Sadeh memset(msg->front.iov_base, 0, msg->front.iov_len); 3943d14c5d2SYehuda Sadeh 3953d14c5d2SYehuda Sadeh req->r_request = msg; 3963d14c5d2SYehuda Sadeh 3973d14c5d2SYehuda Sadeh return req; 3983d14c5d2SYehuda Sadeh } 3993d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_alloc_request); 4003d14c5d2SYehuda Sadeh 401a8dd0a37SAlex Elder static bool osd_req_opcode_valid(u16 opcode) 402a8dd0a37SAlex Elder { 403a8dd0a37SAlex Elder switch (opcode) { 404a8dd0a37SAlex Elder case CEPH_OSD_OP_READ: 405a8dd0a37SAlex Elder case CEPH_OSD_OP_STAT: 406a8dd0a37SAlex Elder case CEPH_OSD_OP_MAPEXT: 407a8dd0a37SAlex Elder case CEPH_OSD_OP_MASKTRUNC: 408a8dd0a37SAlex Elder case CEPH_OSD_OP_SPARSE_READ: 409a8dd0a37SAlex Elder case CEPH_OSD_OP_NOTIFY: 410a8dd0a37SAlex Elder case CEPH_OSD_OP_NOTIFY_ACK: 411a8dd0a37SAlex Elder case CEPH_OSD_OP_ASSERT_VER: 412a8dd0a37SAlex Elder case CEPH_OSD_OP_WRITE: 413a8dd0a37SAlex Elder case CEPH_OSD_OP_WRITEFULL: 414a8dd0a37SAlex Elder case CEPH_OSD_OP_TRUNCATE: 415a8dd0a37SAlex Elder case CEPH_OSD_OP_ZERO: 416a8dd0a37SAlex Elder case CEPH_OSD_OP_DELETE: 417a8dd0a37SAlex Elder case CEPH_OSD_OP_APPEND: 418a8dd0a37SAlex Elder case CEPH_OSD_OP_STARTSYNC: 419a8dd0a37SAlex Elder case CEPH_OSD_OP_SETTRUNC: 420a8dd0a37SAlex Elder case CEPH_OSD_OP_TRIMTRUNC: 421a8dd0a37SAlex Elder case CEPH_OSD_OP_TMAPUP: 422a8dd0a37SAlex Elder case CEPH_OSD_OP_TMAPPUT: 423a8dd0a37SAlex Elder case CEPH_OSD_OP_TMAPGET: 424a8dd0a37SAlex Elder case CEPH_OSD_OP_CREATE: 425a8dd0a37SAlex Elder case CEPH_OSD_OP_ROLLBACK: 426a8dd0a37SAlex Elder case CEPH_OSD_OP_WATCH: 427a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPGETKEYS: 428a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPGETVALS: 429a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPGETHEADER: 430a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPGETVALSBYKEYS: 431a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPSETVALS: 432a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPSETHEADER: 433a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPCLEAR: 434a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPRMKEYS: 435a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAP_CMP: 436a8dd0a37SAlex Elder case CEPH_OSD_OP_CLONERANGE: 437a8dd0a37SAlex Elder case CEPH_OSD_OP_ASSERT_SRC_VERSION: 438a8dd0a37SAlex Elder case CEPH_OSD_OP_SRC_CMPXATTR: 439a8dd0a37SAlex Elder case CEPH_OSD_OP_GETXATTR: 440a8dd0a37SAlex Elder case CEPH_OSD_OP_GETXATTRS: 441a8dd0a37SAlex Elder case CEPH_OSD_OP_CMPXATTR: 442a8dd0a37SAlex Elder case CEPH_OSD_OP_SETXATTR: 443a8dd0a37SAlex Elder case CEPH_OSD_OP_SETXATTRS: 444a8dd0a37SAlex Elder case CEPH_OSD_OP_RESETXATTRS: 445a8dd0a37SAlex Elder case CEPH_OSD_OP_RMXATTR: 446a8dd0a37SAlex Elder case CEPH_OSD_OP_PULL: 447a8dd0a37SAlex Elder case CEPH_OSD_OP_PUSH: 448a8dd0a37SAlex Elder case CEPH_OSD_OP_BALANCEREADS: 449a8dd0a37SAlex Elder case CEPH_OSD_OP_UNBALANCEREADS: 450a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB: 451a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB_RESERVE: 452a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB_UNRESERVE: 453a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB_STOP: 454a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB_MAP: 455a8dd0a37SAlex Elder case CEPH_OSD_OP_WRLOCK: 456a8dd0a37SAlex Elder case CEPH_OSD_OP_WRUNLOCK: 457a8dd0a37SAlex Elder case CEPH_OSD_OP_RDLOCK: 458a8dd0a37SAlex Elder case CEPH_OSD_OP_RDUNLOCK: 459a8dd0a37SAlex Elder case CEPH_OSD_OP_UPLOCK: 460a8dd0a37SAlex Elder case CEPH_OSD_OP_DNLOCK: 461a8dd0a37SAlex Elder case CEPH_OSD_OP_CALL: 462a8dd0a37SAlex Elder case CEPH_OSD_OP_PGLS: 463a8dd0a37SAlex Elder case CEPH_OSD_OP_PGLS_FILTER: 464a8dd0a37SAlex Elder return true; 465a8dd0a37SAlex Elder default: 466a8dd0a37SAlex Elder return false; 467a8dd0a37SAlex Elder } 468a8dd0a37SAlex Elder } 469a8dd0a37SAlex Elder 47033803f33SAlex Elder /* 47133803f33SAlex Elder * This is an osd op init function for opcodes that have no data or 47233803f33SAlex Elder * other information associated with them. It also serves as a 47333803f33SAlex Elder * common init routine for all the other init functions, below. 47433803f33SAlex Elder */ 475c99d2d4aSAlex Elder static struct ceph_osd_req_op * 47649719778SAlex Elder _osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which, 477c99d2d4aSAlex Elder u16 opcode) 47833803f33SAlex Elder { 479c99d2d4aSAlex Elder struct ceph_osd_req_op *op; 480c99d2d4aSAlex Elder 481c99d2d4aSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 48233803f33SAlex Elder BUG_ON(!osd_req_opcode_valid(opcode)); 48333803f33SAlex Elder 484c99d2d4aSAlex Elder op = &osd_req->r_ops[which]; 48533803f33SAlex Elder memset(op, 0, sizeof (*op)); 48633803f33SAlex Elder op->op = opcode; 487c99d2d4aSAlex Elder 488c99d2d4aSAlex Elder return op; 48933803f33SAlex Elder } 49033803f33SAlex Elder 49149719778SAlex Elder void osd_req_op_init(struct ceph_osd_request *osd_req, 49249719778SAlex Elder unsigned int which, u16 opcode) 49349719778SAlex Elder { 49449719778SAlex Elder (void)_osd_req_op_init(osd_req, which, opcode); 49549719778SAlex Elder } 49649719778SAlex Elder EXPORT_SYMBOL(osd_req_op_init); 49749719778SAlex Elder 498c99d2d4aSAlex Elder void osd_req_op_extent_init(struct ceph_osd_request *osd_req, 499c99d2d4aSAlex Elder unsigned int which, u16 opcode, 50033803f33SAlex Elder u64 offset, u64 length, 50133803f33SAlex Elder u64 truncate_size, u32 truncate_seq) 50233803f33SAlex Elder { 50349719778SAlex Elder struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode); 50433803f33SAlex Elder size_t payload_len = 0; 50533803f33SAlex Elder 506ad7a60deSLi Wang BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE && 507ad7a60deSLi Wang opcode != CEPH_OSD_OP_DELETE && opcode != CEPH_OSD_OP_ZERO && 508ad7a60deSLi Wang opcode != CEPH_OSD_OP_TRUNCATE); 50933803f33SAlex Elder 51033803f33SAlex Elder op->extent.offset = offset; 51133803f33SAlex Elder op->extent.length = length; 51233803f33SAlex Elder op->extent.truncate_size = truncate_size; 51333803f33SAlex Elder op->extent.truncate_seq = truncate_seq; 51433803f33SAlex Elder if (opcode == CEPH_OSD_OP_WRITE) 51533803f33SAlex Elder payload_len += length; 51633803f33SAlex Elder 51733803f33SAlex Elder op->payload_len = payload_len; 51833803f33SAlex Elder } 51933803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_init); 52033803f33SAlex Elder 521c99d2d4aSAlex Elder void osd_req_op_extent_update(struct ceph_osd_request *osd_req, 522c99d2d4aSAlex Elder unsigned int which, u64 length) 523e5975c7cSAlex Elder { 524c99d2d4aSAlex Elder struct ceph_osd_req_op *op; 525c99d2d4aSAlex Elder u64 previous; 526c99d2d4aSAlex Elder 527c99d2d4aSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 528c99d2d4aSAlex Elder op = &osd_req->r_ops[which]; 529c99d2d4aSAlex Elder previous = op->extent.length; 530e5975c7cSAlex Elder 531e5975c7cSAlex Elder if (length == previous) 532e5975c7cSAlex Elder return; /* Nothing to do */ 533e5975c7cSAlex Elder BUG_ON(length > previous); 534e5975c7cSAlex Elder 535e5975c7cSAlex Elder op->extent.length = length; 536e5975c7cSAlex Elder op->payload_len -= previous - length; 537e5975c7cSAlex Elder } 538e5975c7cSAlex Elder EXPORT_SYMBOL(osd_req_op_extent_update); 539e5975c7cSAlex Elder 540c99d2d4aSAlex Elder void osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which, 54104017e29SAlex Elder u16 opcode, const char *class, const char *method) 54233803f33SAlex Elder { 54349719778SAlex Elder struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode); 5445f562df5SAlex Elder struct ceph_pagelist *pagelist; 54533803f33SAlex Elder size_t payload_len = 0; 54633803f33SAlex Elder size_t size; 54733803f33SAlex Elder 54833803f33SAlex Elder BUG_ON(opcode != CEPH_OSD_OP_CALL); 54933803f33SAlex Elder 5505f562df5SAlex Elder pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS); 5515f562df5SAlex Elder BUG_ON(!pagelist); 5525f562df5SAlex Elder ceph_pagelist_init(pagelist); 5535f562df5SAlex Elder 55433803f33SAlex Elder op->cls.class_name = class; 55533803f33SAlex Elder size = strlen(class); 55633803f33SAlex Elder BUG_ON(size > (size_t) U8_MAX); 55733803f33SAlex Elder op->cls.class_len = size; 5585f562df5SAlex Elder ceph_pagelist_append(pagelist, class, size); 55933803f33SAlex Elder payload_len += size; 56033803f33SAlex Elder 56133803f33SAlex Elder op->cls.method_name = method; 56233803f33SAlex Elder size = strlen(method); 56333803f33SAlex Elder BUG_ON(size > (size_t) U8_MAX); 56433803f33SAlex Elder op->cls.method_len = size; 5655f562df5SAlex Elder ceph_pagelist_append(pagelist, method, size); 56633803f33SAlex Elder payload_len += size; 56733803f33SAlex Elder 568a4ce40a9SAlex Elder osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist); 5695f562df5SAlex Elder 57033803f33SAlex Elder op->cls.argc = 0; /* currently unused */ 57133803f33SAlex Elder 57233803f33SAlex Elder op->payload_len = payload_len; 57333803f33SAlex Elder } 57433803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_init); 5758c042b0dSAlex Elder 576c99d2d4aSAlex Elder void osd_req_op_watch_init(struct ceph_osd_request *osd_req, 577c99d2d4aSAlex Elder unsigned int which, u16 opcode, 57833803f33SAlex Elder u64 cookie, u64 version, int flag) 57933803f33SAlex Elder { 58049719778SAlex Elder struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode); 58133803f33SAlex Elder 582c99d2d4aSAlex Elder BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH); 58333803f33SAlex Elder 58433803f33SAlex Elder op->watch.cookie = cookie; 5859ef1ee5aSAlex Elder op->watch.ver = version; 58633803f33SAlex Elder if (opcode == CEPH_OSD_OP_WATCH && flag) 58733803f33SAlex Elder op->watch.flag = (u8)1; 58833803f33SAlex Elder } 58933803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_watch_init); 59033803f33SAlex Elder 59190af3602SAlex Elder static void ceph_osdc_msg_data_add(struct ceph_msg *msg, 592ec9123c5SAlex Elder struct ceph_osd_data *osd_data) 593ec9123c5SAlex Elder { 594ec9123c5SAlex Elder u64 length = ceph_osd_data_length(osd_data); 595ec9123c5SAlex Elder 596ec9123c5SAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 597ec9123c5SAlex Elder BUG_ON(length > (u64) SIZE_MAX); 598ec9123c5SAlex Elder if (length) 59990af3602SAlex Elder ceph_msg_data_add_pages(msg, osd_data->pages, 600ec9123c5SAlex Elder length, osd_data->alignment); 601ec9123c5SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) { 602ec9123c5SAlex Elder BUG_ON(!length); 60390af3602SAlex Elder ceph_msg_data_add_pagelist(msg, osd_data->pagelist); 604ec9123c5SAlex Elder #ifdef CONFIG_BLOCK 605ec9123c5SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) { 60690af3602SAlex Elder ceph_msg_data_add_bio(msg, osd_data->bio, length); 607ec9123c5SAlex Elder #endif 608ec9123c5SAlex Elder } else { 609ec9123c5SAlex Elder BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE); 610ec9123c5SAlex Elder } 611ec9123c5SAlex Elder } 612ec9123c5SAlex Elder 613175face2SAlex Elder static u64 osd_req_encode_op(struct ceph_osd_request *req, 61479528734SAlex Elder struct ceph_osd_op *dst, unsigned int which) 6153d14c5d2SYehuda Sadeh { 61679528734SAlex Elder struct ceph_osd_req_op *src; 61704017e29SAlex Elder struct ceph_osd_data *osd_data; 61854d50649SAlex Elder u64 request_data_len = 0; 61904017e29SAlex Elder u64 data_length; 620175face2SAlex Elder 62179528734SAlex Elder BUG_ON(which >= req->r_num_ops); 62279528734SAlex Elder src = &req->r_ops[which]; 623a8dd0a37SAlex Elder if (WARN_ON(!osd_req_opcode_valid(src->op))) { 624a8dd0a37SAlex Elder pr_err("unrecognized osd opcode %d\n", src->op); 625a8dd0a37SAlex Elder 626a8dd0a37SAlex Elder return 0; 627a8dd0a37SAlex Elder } 6283d14c5d2SYehuda Sadeh 629065a68f9SAlex Elder switch (src->op) { 630fbfab539SAlex Elder case CEPH_OSD_OP_STAT: 63149719778SAlex Elder osd_data = &src->raw_data_in; 63249719778SAlex Elder ceph_osdc_msg_data_add(req->r_reply, osd_data); 633fbfab539SAlex Elder break; 6343d14c5d2SYehuda Sadeh case CEPH_OSD_OP_READ: 6353d14c5d2SYehuda Sadeh case CEPH_OSD_OP_WRITE: 636ad7a60deSLi Wang case CEPH_OSD_OP_ZERO: 637ad7a60deSLi Wang case CEPH_OSD_OP_DELETE: 638ad7a60deSLi Wang case CEPH_OSD_OP_TRUNCATE: 639175face2SAlex Elder if (src->op == CEPH_OSD_OP_WRITE) 64054d50649SAlex Elder request_data_len = src->extent.length; 641175face2SAlex Elder dst->extent.offset = cpu_to_le64(src->extent.offset); 642175face2SAlex Elder dst->extent.length = cpu_to_le64(src->extent.length); 6433d14c5d2SYehuda Sadeh dst->extent.truncate_size = 6443d14c5d2SYehuda Sadeh cpu_to_le64(src->extent.truncate_size); 6453d14c5d2SYehuda Sadeh dst->extent.truncate_seq = 6463d14c5d2SYehuda Sadeh cpu_to_le32(src->extent.truncate_seq); 64704017e29SAlex Elder osd_data = &src->extent.osd_data; 6485476492fSAlex Elder if (src->op == CEPH_OSD_OP_WRITE) 64904017e29SAlex Elder ceph_osdc_msg_data_add(req->r_request, osd_data); 6505476492fSAlex Elder else 65104017e29SAlex Elder ceph_osdc_msg_data_add(req->r_reply, osd_data); 6523d14c5d2SYehuda Sadeh break; 6533d14c5d2SYehuda Sadeh case CEPH_OSD_OP_CALL: 6543d14c5d2SYehuda Sadeh dst->cls.class_len = src->cls.class_len; 6553d14c5d2SYehuda Sadeh dst->cls.method_len = src->cls.method_len; 65604017e29SAlex Elder osd_data = &src->cls.request_info; 65704017e29SAlex Elder ceph_osdc_msg_data_add(req->r_request, osd_data); 65804017e29SAlex Elder BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGELIST); 65904017e29SAlex Elder request_data_len = osd_data->pagelist->length; 66004017e29SAlex Elder 66104017e29SAlex Elder osd_data = &src->cls.request_data; 66204017e29SAlex Elder data_length = ceph_osd_data_length(osd_data); 66304017e29SAlex Elder if (data_length) { 66404017e29SAlex Elder BUG_ON(osd_data->type == CEPH_OSD_DATA_TYPE_NONE); 66504017e29SAlex Elder dst->cls.indata_len = cpu_to_le32(data_length); 66604017e29SAlex Elder ceph_osdc_msg_data_add(req->r_request, osd_data); 66704017e29SAlex Elder src->payload_len += data_length; 66804017e29SAlex Elder request_data_len += data_length; 66904017e29SAlex Elder } 67004017e29SAlex Elder osd_data = &src->cls.response_data; 67104017e29SAlex Elder ceph_osdc_msg_data_add(req->r_reply, osd_data); 6723d14c5d2SYehuda Sadeh break; 6733d14c5d2SYehuda Sadeh case CEPH_OSD_OP_STARTSYNC: 6743d14c5d2SYehuda Sadeh break; 675a40c4f10SYehuda Sadeh case CEPH_OSD_OP_NOTIFY_ACK: 676a40c4f10SYehuda Sadeh case CEPH_OSD_OP_WATCH: 677a40c4f10SYehuda Sadeh dst->watch.cookie = cpu_to_le64(src->watch.cookie); 678a40c4f10SYehuda Sadeh dst->watch.ver = cpu_to_le64(src->watch.ver); 679a40c4f10SYehuda Sadeh dst->watch.flag = src->watch.flag; 680a40c4f10SYehuda Sadeh break; 6813d14c5d2SYehuda Sadeh default: 6824c46459cSAlex Elder pr_err("unsupported osd opcode %s\n", 6838f63ca2dSAlex Elder ceph_osd_op_name(src->op)); 6844c46459cSAlex Elder WARN_ON(1); 685a8dd0a37SAlex Elder 686a8dd0a37SAlex Elder return 0; 6873d14c5d2SYehuda Sadeh } 688a8dd0a37SAlex Elder dst->op = cpu_to_le16(src->op); 6893d14c5d2SYehuda Sadeh dst->payload_len = cpu_to_le32(src->payload_len); 690175face2SAlex Elder 69154d50649SAlex Elder return request_data_len; 6923d14c5d2SYehuda Sadeh } 6933d14c5d2SYehuda Sadeh 6943d14c5d2SYehuda Sadeh /* 6953d14c5d2SYehuda Sadeh * build new request AND message, calculate layout, and adjust file 6963d14c5d2SYehuda Sadeh * extent as needed. 6973d14c5d2SYehuda Sadeh * 6983d14c5d2SYehuda Sadeh * if the file was recently truncated, we include information about its 6993d14c5d2SYehuda Sadeh * old and new size so that the object can be updated appropriately. (we 7003d14c5d2SYehuda Sadeh * avoid synchronously deleting truncated objects because it's slow.) 7013d14c5d2SYehuda Sadeh * 7023d14c5d2SYehuda Sadeh * if @do_sync, include a 'startsync' command so that the osd will flush 7033d14c5d2SYehuda Sadeh * data quickly. 7043d14c5d2SYehuda Sadeh */ 7053d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc, 7063d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 7073d14c5d2SYehuda Sadeh struct ceph_vino vino, 708acead002SAlex Elder u64 off, u64 *plen, int num_ops, 7093d14c5d2SYehuda Sadeh int opcode, int flags, 7103d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 7113d14c5d2SYehuda Sadeh u32 truncate_seq, 7123d14c5d2SYehuda Sadeh u64 truncate_size, 713153e5167SAlex Elder bool use_mempool) 7143d14c5d2SYehuda Sadeh { 7153d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 71675d1c941SAlex Elder u64 objnum = 0; 71775d1c941SAlex Elder u64 objoff = 0; 71875d1c941SAlex Elder u64 objlen = 0; 719d18d1e28SAlex Elder u32 object_size; 720d18d1e28SAlex Elder u64 object_base; 7216816282dSSage Weil int r; 7223d14c5d2SYehuda Sadeh 723ad7a60deSLi Wang BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE && 724ad7a60deSLi Wang opcode != CEPH_OSD_OP_DELETE && opcode != CEPH_OSD_OP_ZERO && 725ad7a60deSLi Wang opcode != CEPH_OSD_OP_TRUNCATE); 7263d14c5d2SYehuda Sadeh 727acead002SAlex Elder req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool, 728ae7ca4a3SAlex Elder GFP_NOFS); 7294ad12621SSage Weil if (!req) 7306816282dSSage Weil return ERR_PTR(-ENOMEM); 73179528734SAlex Elder 732d178a9e7SAlex Elder req->r_flags = flags; 7333d14c5d2SYehuda Sadeh 7343d14c5d2SYehuda Sadeh /* calculate max write size */ 735a19dadfbSAlex Elder r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen); 7363ff5f385SAlex Elder if (r < 0) { 7373ff5f385SAlex Elder ceph_osdc_put_request(req); 7386816282dSSage Weil return ERR_PTR(r); 7393ff5f385SAlex Elder } 740a19dadfbSAlex Elder 741d18d1e28SAlex Elder object_size = le32_to_cpu(layout->fl_object_size); 742d18d1e28SAlex Elder object_base = off - objoff; 743ccca4e37SYan, Zheng if (!(truncate_seq == 1 && truncate_size == -1ULL)) { 744d18d1e28SAlex Elder if (truncate_size <= object_base) { 745d18d1e28SAlex Elder truncate_size = 0; 746d18d1e28SAlex Elder } else { 747d18d1e28SAlex Elder truncate_size -= object_base; 748d18d1e28SAlex Elder if (truncate_size > object_size) 749d18d1e28SAlex Elder truncate_size = object_size; 750d18d1e28SAlex Elder } 751ccca4e37SYan, Zheng } 752d18d1e28SAlex Elder 753c99d2d4aSAlex Elder osd_req_op_extent_init(req, 0, opcode, objoff, objlen, 754b0270324SAlex Elder truncate_size, truncate_seq); 7558c042b0dSAlex Elder 756acead002SAlex Elder /* 757acead002SAlex Elder * A second op in the ops array means the caller wants to 758acead002SAlex Elder * also issue a include a 'startsync' command so that the 759acead002SAlex Elder * osd will flush data quickly. 760acead002SAlex Elder */ 761acead002SAlex Elder if (num_ops > 1) 762c99d2d4aSAlex Elder osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC); 763d18d1e28SAlex Elder 7643d14c5d2SYehuda Sadeh req->r_file_layout = *layout; /* keep a copy */ 7653d14c5d2SYehuda Sadeh 76675d1c941SAlex Elder snprintf(req->r_oid, sizeof(req->r_oid), "%llx.%08llx", 76775d1c941SAlex Elder vino.ino, objnum); 768dbe0fc41SAlex Elder req->r_oid_len = strlen(req->r_oid); 769dbe0fc41SAlex Elder 7703d14c5d2SYehuda Sadeh return req; 7713d14c5d2SYehuda Sadeh } 7723d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_new_request); 7733d14c5d2SYehuda Sadeh 7743d14c5d2SYehuda Sadeh /* 7753d14c5d2SYehuda Sadeh * We keep osd requests in an rbtree, sorted by ->r_tid. 7763d14c5d2SYehuda Sadeh */ 7773d14c5d2SYehuda Sadeh static void __insert_request(struct ceph_osd_client *osdc, 7783d14c5d2SYehuda Sadeh struct ceph_osd_request *new) 7793d14c5d2SYehuda Sadeh { 7803d14c5d2SYehuda Sadeh struct rb_node **p = &osdc->requests.rb_node; 7813d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 7823d14c5d2SYehuda Sadeh struct ceph_osd_request *req = NULL; 7833d14c5d2SYehuda Sadeh 7843d14c5d2SYehuda Sadeh while (*p) { 7853d14c5d2SYehuda Sadeh parent = *p; 7863d14c5d2SYehuda Sadeh req = rb_entry(parent, struct ceph_osd_request, r_node); 7873d14c5d2SYehuda Sadeh if (new->r_tid < req->r_tid) 7883d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 7893d14c5d2SYehuda Sadeh else if (new->r_tid > req->r_tid) 7903d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 7913d14c5d2SYehuda Sadeh else 7923d14c5d2SYehuda Sadeh BUG(); 7933d14c5d2SYehuda Sadeh } 7943d14c5d2SYehuda Sadeh 7953d14c5d2SYehuda Sadeh rb_link_node(&new->r_node, parent, p); 7963d14c5d2SYehuda Sadeh rb_insert_color(&new->r_node, &osdc->requests); 7973d14c5d2SYehuda Sadeh } 7983d14c5d2SYehuda Sadeh 7993d14c5d2SYehuda Sadeh static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc, 8003d14c5d2SYehuda Sadeh u64 tid) 8013d14c5d2SYehuda Sadeh { 8023d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 8033d14c5d2SYehuda Sadeh struct rb_node *n = osdc->requests.rb_node; 8043d14c5d2SYehuda Sadeh 8053d14c5d2SYehuda Sadeh while (n) { 8063d14c5d2SYehuda Sadeh req = rb_entry(n, struct ceph_osd_request, r_node); 8073d14c5d2SYehuda Sadeh if (tid < req->r_tid) 8083d14c5d2SYehuda Sadeh n = n->rb_left; 8093d14c5d2SYehuda Sadeh else if (tid > req->r_tid) 8103d14c5d2SYehuda Sadeh n = n->rb_right; 8113d14c5d2SYehuda Sadeh else 8123d14c5d2SYehuda Sadeh return req; 8133d14c5d2SYehuda Sadeh } 8143d14c5d2SYehuda Sadeh return NULL; 8153d14c5d2SYehuda Sadeh } 8163d14c5d2SYehuda Sadeh 8173d14c5d2SYehuda Sadeh static struct ceph_osd_request * 8183d14c5d2SYehuda Sadeh __lookup_request_ge(struct ceph_osd_client *osdc, 8193d14c5d2SYehuda Sadeh u64 tid) 8203d14c5d2SYehuda Sadeh { 8213d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 8223d14c5d2SYehuda Sadeh struct rb_node *n = osdc->requests.rb_node; 8233d14c5d2SYehuda Sadeh 8243d14c5d2SYehuda Sadeh while (n) { 8253d14c5d2SYehuda Sadeh req = rb_entry(n, struct ceph_osd_request, r_node); 8263d14c5d2SYehuda Sadeh if (tid < req->r_tid) { 8273d14c5d2SYehuda Sadeh if (!n->rb_left) 8283d14c5d2SYehuda Sadeh return req; 8293d14c5d2SYehuda Sadeh n = n->rb_left; 8303d14c5d2SYehuda Sadeh } else if (tid > req->r_tid) { 8313d14c5d2SYehuda Sadeh n = n->rb_right; 8323d14c5d2SYehuda Sadeh } else { 8333d14c5d2SYehuda Sadeh return req; 8343d14c5d2SYehuda Sadeh } 8353d14c5d2SYehuda Sadeh } 8363d14c5d2SYehuda Sadeh return NULL; 8373d14c5d2SYehuda Sadeh } 8383d14c5d2SYehuda Sadeh 8396f6c7006SSage Weil /* 8406f6c7006SSage Weil * Resubmit requests pending on the given osd. 8416f6c7006SSage Weil */ 8426f6c7006SSage Weil static void __kick_osd_requests(struct ceph_osd_client *osdc, 8436f6c7006SSage Weil struct ceph_osd *osd) 8446f6c7006SSage Weil { 845a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 846e02493c0SAlex Elder LIST_HEAD(resend); 8476f6c7006SSage Weil int err; 8486f6c7006SSage Weil 8496f6c7006SSage Weil dout("__kick_osd_requests osd%d\n", osd->o_osd); 8506f6c7006SSage Weil err = __reset_osd(osdc, osd); 851685a7555SAlex Elder if (err) 8526f6c7006SSage Weil return; 853e02493c0SAlex Elder /* 854e02493c0SAlex Elder * Build up a list of requests to resend by traversing the 855e02493c0SAlex Elder * osd's list of requests. Requests for a given object are 856e02493c0SAlex Elder * sent in tid order, and that is also the order they're 857e02493c0SAlex Elder * kept on this list. Therefore all requests that are in 858e02493c0SAlex Elder * flight will be found first, followed by all requests that 859e02493c0SAlex Elder * have not yet been sent. And to resend requests while 860e02493c0SAlex Elder * preserving this order we will want to put any sent 861e02493c0SAlex Elder * requests back on the front of the osd client's unsent 862e02493c0SAlex Elder * list. 863e02493c0SAlex Elder * 864e02493c0SAlex Elder * So we build a separate ordered list of already-sent 865e02493c0SAlex Elder * requests for the affected osd and splice it onto the 866e02493c0SAlex Elder * front of the osd client's unsent list. Once we've seen a 867e02493c0SAlex Elder * request that has not yet been sent we're done. Those 868e02493c0SAlex Elder * requests are already sitting right where they belong. 869e02493c0SAlex Elder */ 8706f6c7006SSage Weil list_for_each_entry(req, &osd->o_requests, r_osd_item) { 871e02493c0SAlex Elder if (!req->r_sent) 872e02493c0SAlex Elder break; 873e02493c0SAlex Elder list_move_tail(&req->r_req_lru_item, &resend); 874e02493c0SAlex Elder dout("requeueing %p tid %llu osd%d\n", req, req->r_tid, 8756f6c7006SSage Weil osd->o_osd); 876a40c4f10SYehuda Sadeh if (!req->r_linger) 8776f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 8786f6c7006SSage Weil } 879e02493c0SAlex Elder list_splice(&resend, &osdc->req_unsent); 880a40c4f10SYehuda Sadeh 881e02493c0SAlex Elder /* 882e02493c0SAlex Elder * Linger requests are re-registered before sending, which 883e02493c0SAlex Elder * sets up a new tid for each. We add them to the unsent 884e02493c0SAlex Elder * list at the end to keep things in tid order. 885e02493c0SAlex Elder */ 886a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osd->o_linger_requests, 887a40c4f10SYehuda Sadeh r_linger_osd) { 88877f38e0eSSage Weil /* 88977f38e0eSSage Weil * reregister request prior to unregistering linger so 89077f38e0eSSage Weil * that r_osd is preserved. 89177f38e0eSSage Weil */ 89277f38e0eSSage Weil BUG_ON(!list_empty(&req->r_req_lru_item)); 893a40c4f10SYehuda Sadeh __register_request(osdc, req); 894e02493c0SAlex Elder list_add_tail(&req->r_req_lru_item, &osdc->req_unsent); 895ad885927SAlex Elder list_add_tail(&req->r_osd_item, &req->r_osd->o_requests); 89677f38e0eSSage Weil __unregister_linger_request(osdc, req); 897a40c4f10SYehuda Sadeh dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid, 898a40c4f10SYehuda Sadeh osd->o_osd); 899a40c4f10SYehuda Sadeh } 9006f6c7006SSage Weil } 9016f6c7006SSage Weil 9023d14c5d2SYehuda Sadeh /* 9033d14c5d2SYehuda Sadeh * If the osd connection drops, we need to resubmit all requests. 9043d14c5d2SYehuda Sadeh */ 9053d14c5d2SYehuda Sadeh static void osd_reset(struct ceph_connection *con) 9063d14c5d2SYehuda Sadeh { 9073d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 9083d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 9093d14c5d2SYehuda Sadeh 9103d14c5d2SYehuda Sadeh if (!osd) 9113d14c5d2SYehuda Sadeh return; 9123d14c5d2SYehuda Sadeh dout("osd_reset osd%d\n", osd->o_osd); 9133d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 9143d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 91583aff95eSSage Weil mutex_lock(&osdc->request_mutex); 91683aff95eSSage Weil __kick_osd_requests(osdc, osd); 917f9d25199SAlex Elder __send_queued(osdc); 91883aff95eSSage Weil mutex_unlock(&osdc->request_mutex); 9193d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 9203d14c5d2SYehuda Sadeh } 9213d14c5d2SYehuda Sadeh 9223d14c5d2SYehuda Sadeh /* 9233d14c5d2SYehuda Sadeh * Track open sessions with osds. 9243d14c5d2SYehuda Sadeh */ 925e10006f8SAlex Elder static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum) 9263d14c5d2SYehuda Sadeh { 9273d14c5d2SYehuda Sadeh struct ceph_osd *osd; 9283d14c5d2SYehuda Sadeh 9293d14c5d2SYehuda Sadeh osd = kzalloc(sizeof(*osd), GFP_NOFS); 9303d14c5d2SYehuda Sadeh if (!osd) 9313d14c5d2SYehuda Sadeh return NULL; 9323d14c5d2SYehuda Sadeh 9333d14c5d2SYehuda Sadeh atomic_set(&osd->o_ref, 1); 9343d14c5d2SYehuda Sadeh osd->o_osdc = osdc; 935e10006f8SAlex Elder osd->o_osd = onum; 936f407731dSAlex Elder RB_CLEAR_NODE(&osd->o_node); 9373d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_requests); 938a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osd->o_linger_requests); 9393d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_osd_lru); 9403d14c5d2SYehuda Sadeh osd->o_incarnation = 1; 9413d14c5d2SYehuda Sadeh 942b7a9e5ddSSage Weil ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr); 9433d14c5d2SYehuda Sadeh 9443d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_keepalive_item); 9453d14c5d2SYehuda Sadeh return osd; 9463d14c5d2SYehuda Sadeh } 9473d14c5d2SYehuda Sadeh 9483d14c5d2SYehuda Sadeh static struct ceph_osd *get_osd(struct ceph_osd *osd) 9493d14c5d2SYehuda Sadeh { 9503d14c5d2SYehuda Sadeh if (atomic_inc_not_zero(&osd->o_ref)) { 9513d14c5d2SYehuda Sadeh dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1, 9523d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref)); 9533d14c5d2SYehuda Sadeh return osd; 9543d14c5d2SYehuda Sadeh } else { 9553d14c5d2SYehuda Sadeh dout("get_osd %p FAIL\n", osd); 9563d14c5d2SYehuda Sadeh return NULL; 9573d14c5d2SYehuda Sadeh } 9583d14c5d2SYehuda Sadeh } 9593d14c5d2SYehuda Sadeh 9603d14c5d2SYehuda Sadeh static void put_osd(struct ceph_osd *osd) 9613d14c5d2SYehuda Sadeh { 9623d14c5d2SYehuda Sadeh dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref), 9633d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref) - 1); 964a255651dSAlex Elder if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) { 9653d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth; 9663d14c5d2SYehuda Sadeh 96727859f97SSage Weil ceph_auth_destroy_authorizer(ac, osd->o_auth.authorizer); 9683d14c5d2SYehuda Sadeh kfree(osd); 9693d14c5d2SYehuda Sadeh } 9703d14c5d2SYehuda Sadeh } 9713d14c5d2SYehuda Sadeh 9723d14c5d2SYehuda Sadeh /* 9733d14c5d2SYehuda Sadeh * remove an osd from our map 9743d14c5d2SYehuda Sadeh */ 9753d14c5d2SYehuda Sadeh static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 9763d14c5d2SYehuda Sadeh { 9773d14c5d2SYehuda Sadeh dout("__remove_osd %p\n", osd); 9783d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_requests)); 9793d14c5d2SYehuda Sadeh rb_erase(&osd->o_node, &osdc->osds); 9803d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 9813d14c5d2SYehuda Sadeh ceph_con_close(&osd->o_con); 9823d14c5d2SYehuda Sadeh put_osd(osd); 9833d14c5d2SYehuda Sadeh } 9843d14c5d2SYehuda Sadeh 985aca420bcSSage Weil static void remove_all_osds(struct ceph_osd_client *osdc) 986aca420bcSSage Weil { 987048a9d2dSJiaju Zhang dout("%s %p\n", __func__, osdc); 988aca420bcSSage Weil mutex_lock(&osdc->request_mutex); 989aca420bcSSage Weil while (!RB_EMPTY_ROOT(&osdc->osds)) { 990aca420bcSSage Weil struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds), 991aca420bcSSage Weil struct ceph_osd, o_node); 992aca420bcSSage Weil __remove_osd(osdc, osd); 993aca420bcSSage Weil } 994aca420bcSSage Weil mutex_unlock(&osdc->request_mutex); 995aca420bcSSage Weil } 996aca420bcSSage Weil 9973d14c5d2SYehuda Sadeh static void __move_osd_to_lru(struct ceph_osd_client *osdc, 9983d14c5d2SYehuda Sadeh struct ceph_osd *osd) 9993d14c5d2SYehuda Sadeh { 10003d14c5d2SYehuda Sadeh dout("__move_osd_to_lru %p\n", osd); 10013d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_osd_lru)); 10023d14c5d2SYehuda Sadeh list_add_tail(&osd->o_osd_lru, &osdc->osd_lru); 10033d14c5d2SYehuda Sadeh osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ; 10043d14c5d2SYehuda Sadeh } 10053d14c5d2SYehuda Sadeh 10063d14c5d2SYehuda Sadeh static void __remove_osd_from_lru(struct ceph_osd *osd) 10073d14c5d2SYehuda Sadeh { 10083d14c5d2SYehuda Sadeh dout("__remove_osd_from_lru %p\n", osd); 10093d14c5d2SYehuda Sadeh if (!list_empty(&osd->o_osd_lru)) 10103d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 10113d14c5d2SYehuda Sadeh } 10123d14c5d2SYehuda Sadeh 1013aca420bcSSage Weil static void remove_old_osds(struct ceph_osd_client *osdc) 10143d14c5d2SYehuda Sadeh { 10153d14c5d2SYehuda Sadeh struct ceph_osd *osd, *nosd; 10163d14c5d2SYehuda Sadeh 10173d14c5d2SYehuda Sadeh dout("__remove_old_osds %p\n", osdc); 10183d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 10193d14c5d2SYehuda Sadeh list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) { 1020aca420bcSSage Weil if (time_before(jiffies, osd->lru_ttl)) 10213d14c5d2SYehuda Sadeh break; 10223d14c5d2SYehuda Sadeh __remove_osd(osdc, osd); 10233d14c5d2SYehuda Sadeh } 10243d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 10253d14c5d2SYehuda Sadeh } 10263d14c5d2SYehuda Sadeh 10273d14c5d2SYehuda Sadeh /* 10283d14c5d2SYehuda Sadeh * reset osd connect 10293d14c5d2SYehuda Sadeh */ 10303d14c5d2SYehuda Sadeh static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 10313d14c5d2SYehuda Sadeh { 1032c3acb181SAlex Elder struct ceph_entity_addr *peer_addr; 10333d14c5d2SYehuda Sadeh 10343d14c5d2SYehuda Sadeh dout("__reset_osd %p osd%d\n", osd, osd->o_osd); 1035a40c4f10SYehuda Sadeh if (list_empty(&osd->o_requests) && 1036a40c4f10SYehuda Sadeh list_empty(&osd->o_linger_requests)) { 10373d14c5d2SYehuda Sadeh __remove_osd(osdc, osd); 1038c3acb181SAlex Elder 1039c3acb181SAlex Elder return -ENODEV; 1040c3acb181SAlex Elder } 1041c3acb181SAlex Elder 1042c3acb181SAlex Elder peer_addr = &osdc->osdmap->osd_addr[osd->o_osd]; 1043c3acb181SAlex Elder if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) && 10443d14c5d2SYehuda Sadeh !ceph_con_opened(&osd->o_con)) { 1045c3acb181SAlex Elder struct ceph_osd_request *req; 1046c3acb181SAlex Elder 10473d14c5d2SYehuda Sadeh dout(" osd addr hasn't changed and connection never opened," 10483d14c5d2SYehuda Sadeh " letting msgr retry"); 10493d14c5d2SYehuda Sadeh /* touch each r_stamp for handle_timeout()'s benfit */ 10503d14c5d2SYehuda Sadeh list_for_each_entry(req, &osd->o_requests, r_osd_item) 10513d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 1052c3acb181SAlex Elder 1053c3acb181SAlex Elder return -EAGAIN; 10543d14c5d2SYehuda Sadeh } 1055c3acb181SAlex Elder 1056c3acb181SAlex Elder ceph_con_close(&osd->o_con); 1057c3acb181SAlex Elder ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr); 1058c3acb181SAlex Elder osd->o_incarnation++; 1059c3acb181SAlex Elder 1060c3acb181SAlex Elder return 0; 10613d14c5d2SYehuda Sadeh } 10623d14c5d2SYehuda Sadeh 10633d14c5d2SYehuda Sadeh static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new) 10643d14c5d2SYehuda Sadeh { 10653d14c5d2SYehuda Sadeh struct rb_node **p = &osdc->osds.rb_node; 10663d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 10673d14c5d2SYehuda Sadeh struct ceph_osd *osd = NULL; 10683d14c5d2SYehuda Sadeh 1069aca420bcSSage Weil dout("__insert_osd %p osd%d\n", new, new->o_osd); 10703d14c5d2SYehuda Sadeh while (*p) { 10713d14c5d2SYehuda Sadeh parent = *p; 10723d14c5d2SYehuda Sadeh osd = rb_entry(parent, struct ceph_osd, o_node); 10733d14c5d2SYehuda Sadeh if (new->o_osd < osd->o_osd) 10743d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 10753d14c5d2SYehuda Sadeh else if (new->o_osd > osd->o_osd) 10763d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 10773d14c5d2SYehuda Sadeh else 10783d14c5d2SYehuda Sadeh BUG(); 10793d14c5d2SYehuda Sadeh } 10803d14c5d2SYehuda Sadeh 10813d14c5d2SYehuda Sadeh rb_link_node(&new->o_node, parent, p); 10823d14c5d2SYehuda Sadeh rb_insert_color(&new->o_node, &osdc->osds); 10833d14c5d2SYehuda Sadeh } 10843d14c5d2SYehuda Sadeh 10853d14c5d2SYehuda Sadeh static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o) 10863d14c5d2SYehuda Sadeh { 10873d14c5d2SYehuda Sadeh struct ceph_osd *osd; 10883d14c5d2SYehuda Sadeh struct rb_node *n = osdc->osds.rb_node; 10893d14c5d2SYehuda Sadeh 10903d14c5d2SYehuda Sadeh while (n) { 10913d14c5d2SYehuda Sadeh osd = rb_entry(n, struct ceph_osd, o_node); 10923d14c5d2SYehuda Sadeh if (o < osd->o_osd) 10933d14c5d2SYehuda Sadeh n = n->rb_left; 10943d14c5d2SYehuda Sadeh else if (o > osd->o_osd) 10953d14c5d2SYehuda Sadeh n = n->rb_right; 10963d14c5d2SYehuda Sadeh else 10973d14c5d2SYehuda Sadeh return osd; 10983d14c5d2SYehuda Sadeh } 10993d14c5d2SYehuda Sadeh return NULL; 11003d14c5d2SYehuda Sadeh } 11013d14c5d2SYehuda Sadeh 11023d14c5d2SYehuda Sadeh static void __schedule_osd_timeout(struct ceph_osd_client *osdc) 11033d14c5d2SYehuda Sadeh { 11043d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->timeout_work, 11053d14c5d2SYehuda Sadeh osdc->client->options->osd_keepalive_timeout * HZ); 11063d14c5d2SYehuda Sadeh } 11073d14c5d2SYehuda Sadeh 11083d14c5d2SYehuda Sadeh static void __cancel_osd_timeout(struct ceph_osd_client *osdc) 11093d14c5d2SYehuda Sadeh { 11103d14c5d2SYehuda Sadeh cancel_delayed_work(&osdc->timeout_work); 11113d14c5d2SYehuda Sadeh } 11123d14c5d2SYehuda Sadeh 11133d14c5d2SYehuda Sadeh /* 11143d14c5d2SYehuda Sadeh * Register request, assign tid. If this is the first request, set up 11153d14c5d2SYehuda Sadeh * the timeout event. 11163d14c5d2SYehuda Sadeh */ 1117a40c4f10SYehuda Sadeh static void __register_request(struct ceph_osd_client *osdc, 11183d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 11193d14c5d2SYehuda Sadeh { 11203d14c5d2SYehuda Sadeh req->r_tid = ++osdc->last_tid; 11213d14c5d2SYehuda Sadeh req->r_request->hdr.tid = cpu_to_le64(req->r_tid); 112277f38e0eSSage Weil dout("__register_request %p tid %lld\n", req, req->r_tid); 11233d14c5d2SYehuda Sadeh __insert_request(osdc, req); 11243d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 11253d14c5d2SYehuda Sadeh osdc->num_requests++; 11263d14c5d2SYehuda Sadeh if (osdc->num_requests == 1) { 11273d14c5d2SYehuda Sadeh dout(" first request, scheduling timeout\n"); 11283d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 11293d14c5d2SYehuda Sadeh } 1130a40c4f10SYehuda Sadeh } 1131a40c4f10SYehuda Sadeh 11323d14c5d2SYehuda Sadeh /* 11333d14c5d2SYehuda Sadeh * called under osdc->request_mutex 11343d14c5d2SYehuda Sadeh */ 11353d14c5d2SYehuda Sadeh static void __unregister_request(struct ceph_osd_client *osdc, 11363d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 11373d14c5d2SYehuda Sadeh { 113835f9f8a0SSage Weil if (RB_EMPTY_NODE(&req->r_node)) { 113935f9f8a0SSage Weil dout("__unregister_request %p tid %lld not registered\n", 114035f9f8a0SSage Weil req, req->r_tid); 114135f9f8a0SSage Weil return; 114235f9f8a0SSage Weil } 114335f9f8a0SSage Weil 11443d14c5d2SYehuda Sadeh dout("__unregister_request %p tid %lld\n", req, req->r_tid); 11453d14c5d2SYehuda Sadeh rb_erase(&req->r_node, &osdc->requests); 11463d14c5d2SYehuda Sadeh osdc->num_requests--; 11473d14c5d2SYehuda Sadeh 11483d14c5d2SYehuda Sadeh if (req->r_osd) { 11493d14c5d2SYehuda Sadeh /* make sure the original request isn't in flight. */ 11506740a845SAlex Elder ceph_msg_revoke(req->r_request); 11513d14c5d2SYehuda Sadeh 11523d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 1153a40c4f10SYehuda Sadeh if (list_empty(&req->r_osd->o_requests) && 1154a40c4f10SYehuda Sadeh list_empty(&req->r_osd->o_linger_requests)) { 1155a40c4f10SYehuda Sadeh dout("moving osd to %p lru\n", req->r_osd); 11563d14c5d2SYehuda Sadeh __move_osd_to_lru(osdc, req->r_osd); 1157a40c4f10SYehuda Sadeh } 1158fbdb9190SSage Weil if (list_empty(&req->r_linger_item)) 11593d14c5d2SYehuda Sadeh req->r_osd = NULL; 11603d14c5d2SYehuda Sadeh } 11613d14c5d2SYehuda Sadeh 11627d5f2481SAlex Elder list_del_init(&req->r_req_lru_item); 11633d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 11643d14c5d2SYehuda Sadeh 11653d14c5d2SYehuda Sadeh if (osdc->num_requests == 0) { 11663d14c5d2SYehuda Sadeh dout(" no requests, canceling timeout\n"); 11673d14c5d2SYehuda Sadeh __cancel_osd_timeout(osdc); 11683d14c5d2SYehuda Sadeh } 11693d14c5d2SYehuda Sadeh } 11703d14c5d2SYehuda Sadeh 11713d14c5d2SYehuda Sadeh /* 11723d14c5d2SYehuda Sadeh * Cancel a previously queued request message 11733d14c5d2SYehuda Sadeh */ 11743d14c5d2SYehuda Sadeh static void __cancel_request(struct ceph_osd_request *req) 11753d14c5d2SYehuda Sadeh { 11763d14c5d2SYehuda Sadeh if (req->r_sent && req->r_osd) { 11776740a845SAlex Elder ceph_msg_revoke(req->r_request); 11783d14c5d2SYehuda Sadeh req->r_sent = 0; 11793d14c5d2SYehuda Sadeh } 11803d14c5d2SYehuda Sadeh } 11813d14c5d2SYehuda Sadeh 1182a40c4f10SYehuda Sadeh static void __register_linger_request(struct ceph_osd_client *osdc, 1183a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1184a40c4f10SYehuda Sadeh { 1185a40c4f10SYehuda Sadeh dout("__register_linger_request %p\n", req); 118696e4dac6SAlex Elder ceph_osdc_get_request(req); 1187a40c4f10SYehuda Sadeh list_add_tail(&req->r_linger_item, &osdc->req_linger); 11886194ea89SSage Weil if (req->r_osd) 11896194ea89SSage Weil list_add_tail(&req->r_linger_osd, 11906194ea89SSage Weil &req->r_osd->o_linger_requests); 1191a40c4f10SYehuda Sadeh } 1192a40c4f10SYehuda Sadeh 1193a40c4f10SYehuda Sadeh static void __unregister_linger_request(struct ceph_osd_client *osdc, 1194a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1195a40c4f10SYehuda Sadeh { 1196a40c4f10SYehuda Sadeh dout("__unregister_linger_request %p\n", req); 1197a40c4f10SYehuda Sadeh list_del_init(&req->r_linger_item); 119861c74035SAlex Elder if (req->r_osd) { 1199a40c4f10SYehuda Sadeh list_del_init(&req->r_linger_osd); 1200a40c4f10SYehuda Sadeh 1201a40c4f10SYehuda Sadeh if (list_empty(&req->r_osd->o_requests) && 1202a40c4f10SYehuda Sadeh list_empty(&req->r_osd->o_linger_requests)) { 1203a40c4f10SYehuda Sadeh dout("moving osd to %p lru\n", req->r_osd); 1204a40c4f10SYehuda Sadeh __move_osd_to_lru(osdc, req->r_osd); 1205a40c4f10SYehuda Sadeh } 1206fbdb9190SSage Weil if (list_empty(&req->r_osd_item)) 1207a40c4f10SYehuda Sadeh req->r_osd = NULL; 1208a40c4f10SYehuda Sadeh } 120996e4dac6SAlex Elder ceph_osdc_put_request(req); 1210a40c4f10SYehuda Sadeh } 1211a40c4f10SYehuda Sadeh 1212a40c4f10SYehuda Sadeh void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc, 1213a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1214a40c4f10SYehuda Sadeh { 1215a40c4f10SYehuda Sadeh mutex_lock(&osdc->request_mutex); 1216a40c4f10SYehuda Sadeh if (req->r_linger) { 1217c10ebbf5SAlex Elder req->r_linger = 0; 121896e4dac6SAlex Elder __unregister_linger_request(osdc, req); 1219a40c4f10SYehuda Sadeh } 1220a40c4f10SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 1221a40c4f10SYehuda Sadeh } 1222a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_unregister_linger_request); 1223a40c4f10SYehuda Sadeh 1224a40c4f10SYehuda Sadeh void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc, 1225a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1226a40c4f10SYehuda Sadeh { 1227a40c4f10SYehuda Sadeh if (!req->r_linger) { 1228a40c4f10SYehuda Sadeh dout("set_request_linger %p\n", req); 1229a40c4f10SYehuda Sadeh req->r_linger = 1; 1230a40c4f10SYehuda Sadeh } 1231a40c4f10SYehuda Sadeh } 1232a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_set_request_linger); 1233a40c4f10SYehuda Sadeh 12343d14c5d2SYehuda Sadeh /* 1235d29adb34SJosh Durgin * Returns whether a request should be blocked from being sent 1236d29adb34SJosh Durgin * based on the current osdmap and osd_client settings. 1237d29adb34SJosh Durgin * 1238d29adb34SJosh Durgin * Caller should hold map_sem for read. 1239d29adb34SJosh Durgin */ 1240d29adb34SJosh Durgin static bool __req_should_be_paused(struct ceph_osd_client *osdc, 1241d29adb34SJosh Durgin struct ceph_osd_request *req) 1242d29adb34SJosh Durgin { 1243d29adb34SJosh Durgin bool pauserd = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD); 1244d29adb34SJosh Durgin bool pausewr = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR) || 1245d29adb34SJosh Durgin ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL); 1246d29adb34SJosh Durgin return (req->r_flags & CEPH_OSD_FLAG_READ && pauserd) || 1247d29adb34SJosh Durgin (req->r_flags & CEPH_OSD_FLAG_WRITE && pausewr); 1248d29adb34SJosh Durgin } 1249d29adb34SJosh Durgin 1250d29adb34SJosh Durgin /* 12513d14c5d2SYehuda Sadeh * Pick an osd (the first 'up' osd in the pg), allocate the osd struct 12523d14c5d2SYehuda Sadeh * (as needed), and set the request r_osd appropriately. If there is 125325985edcSLucas De Marchi * no up osd, set r_osd to NULL. Move the request to the appropriate list 12546f6c7006SSage Weil * (unsent, homeless) or leave on in-flight lru. 12553d14c5d2SYehuda Sadeh * 12563d14c5d2SYehuda Sadeh * Return 0 if unchanged, 1 if changed, or negative on error. 12573d14c5d2SYehuda Sadeh * 12583d14c5d2SYehuda Sadeh * Caller should hold map_sem for read and request_mutex. 12593d14c5d2SYehuda Sadeh */ 12606f6c7006SSage Weil static int __map_request(struct ceph_osd_client *osdc, 126138d6453cSSage Weil struct ceph_osd_request *req, int force_resend) 12623d14c5d2SYehuda Sadeh { 12635b191d99SSage Weil struct ceph_pg pgid; 12643d14c5d2SYehuda Sadeh int acting[CEPH_PG_MAX_SIZE]; 12653d14c5d2SYehuda Sadeh int o = -1, num = 0; 12663d14c5d2SYehuda Sadeh int err; 1267d29adb34SJosh Durgin bool was_paused; 12683d14c5d2SYehuda Sadeh 12696f6c7006SSage Weil dout("map_request %p tid %lld\n", req, req->r_tid); 127041766f87SAlex Elder err = ceph_calc_ceph_pg(&pgid, req->r_oid, osdc->osdmap, 127141766f87SAlex Elder ceph_file_layout_pg_pool(req->r_file_layout)); 12726f6c7006SSage Weil if (err) { 12736f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 12743d14c5d2SYehuda Sadeh return err; 12756f6c7006SSage Weil } 12763d14c5d2SYehuda Sadeh req->r_pgid = pgid; 12773d14c5d2SYehuda Sadeh 12783d14c5d2SYehuda Sadeh err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting); 12793d14c5d2SYehuda Sadeh if (err > 0) { 12803d14c5d2SYehuda Sadeh o = acting[0]; 12813d14c5d2SYehuda Sadeh num = err; 12823d14c5d2SYehuda Sadeh } 12833d14c5d2SYehuda Sadeh 1284d29adb34SJosh Durgin was_paused = req->r_paused; 1285d29adb34SJosh Durgin req->r_paused = __req_should_be_paused(osdc, req); 1286d29adb34SJosh Durgin if (was_paused && !req->r_paused) 1287d29adb34SJosh Durgin force_resend = 1; 1288d29adb34SJosh Durgin 128938d6453cSSage Weil if ((!force_resend && 129038d6453cSSage Weil req->r_osd && req->r_osd->o_osd == o && 12913d14c5d2SYehuda Sadeh req->r_sent >= req->r_osd->o_incarnation && 12923d14c5d2SYehuda Sadeh req->r_num_pg_osds == num && 12933d14c5d2SYehuda Sadeh memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) || 1294d29adb34SJosh Durgin (req->r_osd == NULL && o == -1) || 1295d29adb34SJosh Durgin req->r_paused) 12963d14c5d2SYehuda Sadeh return 0; /* no change */ 12973d14c5d2SYehuda Sadeh 12985b191d99SSage Weil dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n", 12995b191d99SSage Weil req->r_tid, pgid.pool, pgid.seed, o, 13003d14c5d2SYehuda Sadeh req->r_osd ? req->r_osd->o_osd : -1); 13013d14c5d2SYehuda Sadeh 13023d14c5d2SYehuda Sadeh /* record full pg acting set */ 13033d14c5d2SYehuda Sadeh memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num); 13043d14c5d2SYehuda Sadeh req->r_num_pg_osds = num; 13053d14c5d2SYehuda Sadeh 13063d14c5d2SYehuda Sadeh if (req->r_osd) { 13073d14c5d2SYehuda Sadeh __cancel_request(req); 13083d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 13093d14c5d2SYehuda Sadeh req->r_osd = NULL; 13103d14c5d2SYehuda Sadeh } 13113d14c5d2SYehuda Sadeh 13123d14c5d2SYehuda Sadeh req->r_osd = __lookup_osd(osdc, o); 13133d14c5d2SYehuda Sadeh if (!req->r_osd && o >= 0) { 13143d14c5d2SYehuda Sadeh err = -ENOMEM; 1315e10006f8SAlex Elder req->r_osd = create_osd(osdc, o); 13166f6c7006SSage Weil if (!req->r_osd) { 13176f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 13183d14c5d2SYehuda Sadeh goto out; 13196f6c7006SSage Weil } 13203d14c5d2SYehuda Sadeh 13216f6c7006SSage Weil dout("map_request osd %p is osd%d\n", req->r_osd, o); 13223d14c5d2SYehuda Sadeh __insert_osd(osdc, req->r_osd); 13233d14c5d2SYehuda Sadeh 1324b7a9e5ddSSage Weil ceph_con_open(&req->r_osd->o_con, 1325b7a9e5ddSSage Weil CEPH_ENTITY_TYPE_OSD, o, 1326b7a9e5ddSSage Weil &osdc->osdmap->osd_addr[o]); 13273d14c5d2SYehuda Sadeh } 13283d14c5d2SYehuda Sadeh 13293d14c5d2SYehuda Sadeh if (req->r_osd) { 13303d14c5d2SYehuda Sadeh __remove_osd_from_lru(req->r_osd); 1331ad885927SAlex Elder list_add_tail(&req->r_osd_item, &req->r_osd->o_requests); 1332ad885927SAlex Elder list_move_tail(&req->r_req_lru_item, &osdc->req_unsent); 13336f6c7006SSage Weil } else { 1334ad885927SAlex Elder list_move_tail(&req->r_req_lru_item, &osdc->req_notarget); 13353d14c5d2SYehuda Sadeh } 13363d14c5d2SYehuda Sadeh err = 1; /* osd or pg changed */ 13373d14c5d2SYehuda Sadeh 13383d14c5d2SYehuda Sadeh out: 13393d14c5d2SYehuda Sadeh return err; 13403d14c5d2SYehuda Sadeh } 13413d14c5d2SYehuda Sadeh 13423d14c5d2SYehuda Sadeh /* 13433d14c5d2SYehuda Sadeh * caller should hold map_sem (for read) and request_mutex 13443d14c5d2SYehuda Sadeh */ 134556e925b6SSage Weil static void __send_request(struct ceph_osd_client *osdc, 13463d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 13473d14c5d2SYehuda Sadeh { 13481b83bef2SSage Weil void *p; 13493d14c5d2SYehuda Sadeh 13501b83bef2SSage Weil dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n", 13511b83bef2SSage Weil req, req->r_tid, req->r_osd->o_osd, req->r_flags, 13521b83bef2SSage Weil (unsigned long long)req->r_pgid.pool, req->r_pgid.seed); 13533d14c5d2SYehuda Sadeh 13541b83bef2SSage Weil /* fill in message content that changes each time we send it */ 13551b83bef2SSage Weil put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch); 13561b83bef2SSage Weil put_unaligned_le32(req->r_flags, req->r_request_flags); 13571b83bef2SSage Weil put_unaligned_le64(req->r_pgid.pool, req->r_request_pool); 13581b83bef2SSage Weil p = req->r_request_pgid; 13591b83bef2SSage Weil ceph_encode_64(&p, req->r_pgid.pool); 13601b83bef2SSage Weil ceph_encode_32(&p, req->r_pgid.seed); 13611b83bef2SSage Weil put_unaligned_le64(1, req->r_request_attempts); /* FIXME */ 13621b83bef2SSage Weil memcpy(req->r_request_reassert_version, &req->r_reassert_version, 13631b83bef2SSage Weil sizeof(req->r_reassert_version)); 13642169aea6SSage Weil 13653d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 13663d14c5d2SYehuda Sadeh list_move_tail(&req->r_req_lru_item, &osdc->req_lru); 13673d14c5d2SYehuda Sadeh 13683d14c5d2SYehuda Sadeh ceph_msg_get(req->r_request); /* send consumes a ref */ 136926be8808SAlex Elder 13703d14c5d2SYehuda Sadeh req->r_sent = req->r_osd->o_incarnation; 137126be8808SAlex Elder 137226be8808SAlex Elder ceph_con_send(&req->r_osd->o_con, req->r_request); 13733d14c5d2SYehuda Sadeh } 13743d14c5d2SYehuda Sadeh 13753d14c5d2SYehuda Sadeh /* 13766f6c7006SSage Weil * Send any requests in the queue (req_unsent). 13776f6c7006SSage Weil */ 1378f9d25199SAlex Elder static void __send_queued(struct ceph_osd_client *osdc) 13796f6c7006SSage Weil { 13806f6c7006SSage Weil struct ceph_osd_request *req, *tmp; 13816f6c7006SSage Weil 1382f9d25199SAlex Elder dout("__send_queued\n"); 1383f9d25199SAlex Elder list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) 13846f6c7006SSage Weil __send_request(osdc, req); 13856f6c7006SSage Weil } 13866f6c7006SSage Weil 13876f6c7006SSage Weil /* 13883d14c5d2SYehuda Sadeh * Timeout callback, called every N seconds when 1 or more osd 13893d14c5d2SYehuda Sadeh * requests has been active for more than N seconds. When this 13903d14c5d2SYehuda Sadeh * happens, we ping all OSDs with requests who have timed out to 13913d14c5d2SYehuda Sadeh * ensure any communications channel reset is detected. Reset the 13923d14c5d2SYehuda Sadeh * request timeouts another N seconds in the future as we go. 13933d14c5d2SYehuda Sadeh * Reschedule the timeout event another N seconds in future (unless 13943d14c5d2SYehuda Sadeh * there are no open requests). 13953d14c5d2SYehuda Sadeh */ 13963d14c5d2SYehuda Sadeh static void handle_timeout(struct work_struct *work) 13973d14c5d2SYehuda Sadeh { 13983d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 13993d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, timeout_work.work); 140083aff95eSSage Weil struct ceph_osd_request *req; 14013d14c5d2SYehuda Sadeh struct ceph_osd *osd; 14023d14c5d2SYehuda Sadeh unsigned long keepalive = 14033d14c5d2SYehuda Sadeh osdc->client->options->osd_keepalive_timeout * HZ; 14043d14c5d2SYehuda Sadeh struct list_head slow_osds; 14053d14c5d2SYehuda Sadeh dout("timeout\n"); 14063d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 14073d14c5d2SYehuda Sadeh 14083d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 14093d14c5d2SYehuda Sadeh 14103d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 14113d14c5d2SYehuda Sadeh 14123d14c5d2SYehuda Sadeh /* 14133d14c5d2SYehuda Sadeh * ping osds that are a bit slow. this ensures that if there 14143d14c5d2SYehuda Sadeh * is a break in the TCP connection we will notice, and reopen 14153d14c5d2SYehuda Sadeh * a connection with that osd (from the fault callback). 14163d14c5d2SYehuda Sadeh */ 14173d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&slow_osds); 14183d14c5d2SYehuda Sadeh list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) { 14193d14c5d2SYehuda Sadeh if (time_before(jiffies, req->r_stamp + keepalive)) 14203d14c5d2SYehuda Sadeh break; 14213d14c5d2SYehuda Sadeh 14223d14c5d2SYehuda Sadeh osd = req->r_osd; 14233d14c5d2SYehuda Sadeh BUG_ON(!osd); 14243d14c5d2SYehuda Sadeh dout(" tid %llu is slow, will send keepalive on osd%d\n", 14253d14c5d2SYehuda Sadeh req->r_tid, osd->o_osd); 14263d14c5d2SYehuda Sadeh list_move_tail(&osd->o_keepalive_item, &slow_osds); 14273d14c5d2SYehuda Sadeh } 14283d14c5d2SYehuda Sadeh while (!list_empty(&slow_osds)) { 14293d14c5d2SYehuda Sadeh osd = list_entry(slow_osds.next, struct ceph_osd, 14303d14c5d2SYehuda Sadeh o_keepalive_item); 14313d14c5d2SYehuda Sadeh list_del_init(&osd->o_keepalive_item); 14323d14c5d2SYehuda Sadeh ceph_con_keepalive(&osd->o_con); 14333d14c5d2SYehuda Sadeh } 14343d14c5d2SYehuda Sadeh 14353d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 1436f9d25199SAlex Elder __send_queued(osdc); 14373d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 14383d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 14393d14c5d2SYehuda Sadeh } 14403d14c5d2SYehuda Sadeh 14413d14c5d2SYehuda Sadeh static void handle_osds_timeout(struct work_struct *work) 14423d14c5d2SYehuda Sadeh { 14433d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 14443d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, 14453d14c5d2SYehuda Sadeh osds_timeout_work.work); 14463d14c5d2SYehuda Sadeh unsigned long delay = 14473d14c5d2SYehuda Sadeh osdc->client->options->osd_idle_ttl * HZ >> 2; 14483d14c5d2SYehuda Sadeh 14493d14c5d2SYehuda Sadeh dout("osds timeout\n"); 14503d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 1451aca420bcSSage Weil remove_old_osds(osdc); 14523d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 14533d14c5d2SYehuda Sadeh 14543d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 14553d14c5d2SYehuda Sadeh round_jiffies_relative(delay)); 14563d14c5d2SYehuda Sadeh } 14573d14c5d2SYehuda Sadeh 145825845472SSage Weil static void complete_request(struct ceph_osd_request *req) 145925845472SSage Weil { 146025845472SSage Weil complete_all(&req->r_safe_completion); /* fsync waiter */ 146125845472SSage Weil } 146225845472SSage Weil 14633d14c5d2SYehuda Sadeh /* 14643d14c5d2SYehuda Sadeh * handle osd op reply. either call the callback if it is specified, 14653d14c5d2SYehuda Sadeh * or do the completion to wake up the waiting thread. 14663d14c5d2SYehuda Sadeh */ 14673d14c5d2SYehuda Sadeh static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg, 14683d14c5d2SYehuda Sadeh struct ceph_connection *con) 14693d14c5d2SYehuda Sadeh { 14701b83bef2SSage Weil void *p, *end; 14713d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 14723d14c5d2SYehuda Sadeh u64 tid; 14731b83bef2SSage Weil int object_len; 147479528734SAlex Elder unsigned int numops; 147579528734SAlex Elder int payload_len, flags; 14763d14c5d2SYehuda Sadeh s32 result; 14771b83bef2SSage Weil s32 retry_attempt; 14781b83bef2SSage Weil struct ceph_pg pg; 14791b83bef2SSage Weil int err; 14801b83bef2SSage Weil u32 reassert_epoch; 14811b83bef2SSage Weil u64 reassert_version; 14821b83bef2SSage Weil u32 osdmap_epoch; 14830d5af164SAlex Elder int already_completed; 14849fc6e064SAlex Elder u32 bytes; 148579528734SAlex Elder unsigned int i; 14863d14c5d2SYehuda Sadeh 14873d14c5d2SYehuda Sadeh tid = le64_to_cpu(msg->hdr.tid); 14881b83bef2SSage Weil dout("handle_reply %p tid %llu\n", msg, tid); 14891b83bef2SSage Weil 14901b83bef2SSage Weil p = msg->front.iov_base; 14911b83bef2SSage Weil end = p + msg->front.iov_len; 14921b83bef2SSage Weil 14931b83bef2SSage Weil ceph_decode_need(&p, end, 4, bad); 14941b83bef2SSage Weil object_len = ceph_decode_32(&p); 14951b83bef2SSage Weil ceph_decode_need(&p, end, object_len, bad); 14961b83bef2SSage Weil p += object_len; 14971b83bef2SSage Weil 1498ef4859d6SAlex Elder err = ceph_decode_pgid(&p, end, &pg); 14991b83bef2SSage Weil if (err) 15003d14c5d2SYehuda Sadeh goto bad; 15011b83bef2SSage Weil 15021b83bef2SSage Weil ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad); 15031b83bef2SSage Weil flags = ceph_decode_64(&p); 15041b83bef2SSage Weil result = ceph_decode_32(&p); 15051b83bef2SSage Weil reassert_epoch = ceph_decode_32(&p); 15061b83bef2SSage Weil reassert_version = ceph_decode_64(&p); 15071b83bef2SSage Weil osdmap_epoch = ceph_decode_32(&p); 15081b83bef2SSage Weil 15093d14c5d2SYehuda Sadeh /* lookup */ 15103d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 15113d14c5d2SYehuda Sadeh req = __lookup_request(osdc, tid); 15123d14c5d2SYehuda Sadeh if (req == NULL) { 15133d14c5d2SYehuda Sadeh dout("handle_reply tid %llu dne\n", tid); 15148058fd45SAlex Elder goto bad_mutex; 15153d14c5d2SYehuda Sadeh } 15163d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 15171b83bef2SSage Weil 15181b83bef2SSage Weil dout("handle_reply %p tid %llu req %p result %d\n", msg, tid, 15191b83bef2SSage Weil req, result); 15201b83bef2SSage Weil 152118741196SDan Carpenter ceph_decode_need(&p, end, 4, bad_put); 15221b83bef2SSage Weil numops = ceph_decode_32(&p); 15231b83bef2SSage Weil if (numops > CEPH_OSD_MAX_OP) 15241b83bef2SSage Weil goto bad_put; 15251b83bef2SSage Weil if (numops != req->r_num_ops) 15261b83bef2SSage Weil goto bad_put; 15271b83bef2SSage Weil payload_len = 0; 152818741196SDan Carpenter ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad_put); 15291b83bef2SSage Weil for (i = 0; i < numops; i++) { 15301b83bef2SSage Weil struct ceph_osd_op *op = p; 15311b83bef2SSage Weil int len; 15321b83bef2SSage Weil 15331b83bef2SSage Weil len = le32_to_cpu(op->payload_len); 15341b83bef2SSage Weil req->r_reply_op_len[i] = len; 15351b83bef2SSage Weil dout(" op %d has %d bytes\n", i, len); 15361b83bef2SSage Weil payload_len += len; 15371b83bef2SSage Weil p += sizeof(*op); 15381b83bef2SSage Weil } 15399fc6e064SAlex Elder bytes = le32_to_cpu(msg->hdr.data_len); 15409fc6e064SAlex Elder if (payload_len != bytes) { 15411b83bef2SSage Weil pr_warning("sum of op payload lens %d != data_len %d", 15429fc6e064SAlex Elder payload_len, bytes); 15431b83bef2SSage Weil goto bad_put; 15441b83bef2SSage Weil } 15451b83bef2SSage Weil 154618741196SDan Carpenter ceph_decode_need(&p, end, 4 + numops * 4, bad_put); 15471b83bef2SSage Weil retry_attempt = ceph_decode_32(&p); 15481b83bef2SSage Weil for (i = 0; i < numops; i++) 15491b83bef2SSage Weil req->r_reply_op_result[i] = ceph_decode_32(&p); 15503d14c5d2SYehuda Sadeh 1551eb845ff1SYan, Zheng already_completed = req->r_got_reply; 1552eb845ff1SYan, Zheng 15533d14c5d2SYehuda Sadeh if (!req->r_got_reply) { 15543d14c5d2SYehuda Sadeh 15551b83bef2SSage Weil req->r_result = result; 15563d14c5d2SYehuda Sadeh dout("handle_reply result %d bytes %d\n", req->r_result, 15573d14c5d2SYehuda Sadeh bytes); 15583d14c5d2SYehuda Sadeh if (req->r_result == 0) 15593d14c5d2SYehuda Sadeh req->r_result = bytes; 15603d14c5d2SYehuda Sadeh 15613d14c5d2SYehuda Sadeh /* in case this is a write and we need to replay, */ 15621b83bef2SSage Weil req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch); 15631b83bef2SSage Weil req->r_reassert_version.version = cpu_to_le64(reassert_version); 15643d14c5d2SYehuda Sadeh 15653d14c5d2SYehuda Sadeh req->r_got_reply = 1; 15663d14c5d2SYehuda Sadeh } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) { 15673d14c5d2SYehuda Sadeh dout("handle_reply tid %llu dup ack\n", tid); 15683d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 15693d14c5d2SYehuda Sadeh goto done; 15703d14c5d2SYehuda Sadeh } 15713d14c5d2SYehuda Sadeh 15723d14c5d2SYehuda Sadeh dout("handle_reply tid %llu flags %d\n", tid, flags); 15733d14c5d2SYehuda Sadeh 1574a40c4f10SYehuda Sadeh if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK)) 1575a40c4f10SYehuda Sadeh __register_linger_request(osdc, req); 1576a40c4f10SYehuda Sadeh 15773d14c5d2SYehuda Sadeh /* either this is a read, or we got the safe response */ 15783d14c5d2SYehuda Sadeh if (result < 0 || 15793d14c5d2SYehuda Sadeh (flags & CEPH_OSD_FLAG_ONDISK) || 15803d14c5d2SYehuda Sadeh ((flags & CEPH_OSD_FLAG_WRITE) == 0)) 15813d14c5d2SYehuda Sadeh __unregister_request(osdc, req); 15823d14c5d2SYehuda Sadeh 15833d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 15843d14c5d2SYehuda Sadeh 1585eb845ff1SYan, Zheng if (!already_completed) { 158661c5d6bfSYan, Zheng if (req->r_unsafe_callback && 158761c5d6bfSYan, Zheng result >= 0 && !(flags & CEPH_OSD_FLAG_ONDISK)) 158861c5d6bfSYan, Zheng req->r_unsafe_callback(req, true); 15893d14c5d2SYehuda Sadeh if (req->r_callback) 15903d14c5d2SYehuda Sadeh req->r_callback(req, msg); 15913d14c5d2SYehuda Sadeh else 15923d14c5d2SYehuda Sadeh complete_all(&req->r_completion); 1593eb845ff1SYan, Zheng } 15943d14c5d2SYehuda Sadeh 159561c5d6bfSYan, Zheng if (flags & CEPH_OSD_FLAG_ONDISK) { 159661c5d6bfSYan, Zheng if (req->r_unsafe_callback && already_completed) 159761c5d6bfSYan, Zheng req->r_unsafe_callback(req, false); 159825845472SSage Weil complete_request(req); 159961c5d6bfSYan, Zheng } 16003d14c5d2SYehuda Sadeh 16013d14c5d2SYehuda Sadeh done: 1602a40c4f10SYehuda Sadeh dout("req=%p req->r_linger=%d\n", req, req->r_linger); 16033d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 16043d14c5d2SYehuda Sadeh return; 16053d14c5d2SYehuda Sadeh 16061b83bef2SSage Weil bad_put: 160737c89bdeSLi Wang req->r_result = -EIO; 160837c89bdeSLi Wang __unregister_request(osdc, req); 160937c89bdeSLi Wang if (req->r_callback) 161037c89bdeSLi Wang req->r_callback(req, msg); 161137c89bdeSLi Wang else 161237c89bdeSLi Wang complete_all(&req->r_completion); 161337c89bdeSLi Wang complete_request(req); 16141b83bef2SSage Weil ceph_osdc_put_request(req); 16158058fd45SAlex Elder bad_mutex: 16168058fd45SAlex Elder mutex_unlock(&osdc->request_mutex); 16173d14c5d2SYehuda Sadeh bad: 16181b83bef2SSage Weil pr_err("corrupt osd_op_reply got %d %d\n", 16191b83bef2SSage Weil (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len)); 16203d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 16213d14c5d2SYehuda Sadeh } 16223d14c5d2SYehuda Sadeh 16236f6c7006SSage Weil static void reset_changed_osds(struct ceph_osd_client *osdc) 16243d14c5d2SYehuda Sadeh { 16253d14c5d2SYehuda Sadeh struct rb_node *p, *n; 16263d14c5d2SYehuda Sadeh 16273d14c5d2SYehuda Sadeh for (p = rb_first(&osdc->osds); p; p = n) { 16286f6c7006SSage Weil struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node); 16293d14c5d2SYehuda Sadeh 16303d14c5d2SYehuda Sadeh n = rb_next(p); 16313d14c5d2SYehuda Sadeh if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) || 16323d14c5d2SYehuda Sadeh memcmp(&osd->o_con.peer_addr, 16333d14c5d2SYehuda Sadeh ceph_osd_addr(osdc->osdmap, 16343d14c5d2SYehuda Sadeh osd->o_osd), 16353d14c5d2SYehuda Sadeh sizeof(struct ceph_entity_addr)) != 0) 16363d14c5d2SYehuda Sadeh __reset_osd(osdc, osd); 16373d14c5d2SYehuda Sadeh } 16383d14c5d2SYehuda Sadeh } 16393d14c5d2SYehuda Sadeh 16403d14c5d2SYehuda Sadeh /* 16416f6c7006SSage Weil * Requeue requests whose mapping to an OSD has changed. If requests map to 16426f6c7006SSage Weil * no osd, request a new map. 16433d14c5d2SYehuda Sadeh * 1644e6d50f67SAlex Elder * Caller should hold map_sem for read. 16453d14c5d2SYehuda Sadeh */ 16469a1ea2dbSJosh Durgin static void kick_requests(struct ceph_osd_client *osdc, bool force_resend, 16479a1ea2dbSJosh Durgin bool force_resend_writes) 16483d14c5d2SYehuda Sadeh { 1649a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 16506f6c7006SSage Weil struct rb_node *p; 16516f6c7006SSage Weil int needmap = 0; 16526f6c7006SSage Weil int err; 16539a1ea2dbSJosh Durgin bool force_resend_req; 16543d14c5d2SYehuda Sadeh 16559a1ea2dbSJosh Durgin dout("kick_requests %s %s\n", force_resend ? " (force resend)" : "", 16569a1ea2dbSJosh Durgin force_resend_writes ? " (force resend writes)" : ""); 16573d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 16586194ea89SSage Weil for (p = rb_first(&osdc->requests); p; ) { 16596f6c7006SSage Weil req = rb_entry(p, struct ceph_osd_request, r_node); 16606194ea89SSage Weil p = rb_next(p); 1661ab60b16dSAlex Elder 1662ab60b16dSAlex Elder /* 1663ab60b16dSAlex Elder * For linger requests that have not yet been 1664ab60b16dSAlex Elder * registered, move them to the linger list; they'll 1665ab60b16dSAlex Elder * be sent to the osd in the loop below. Unregister 1666ab60b16dSAlex Elder * the request before re-registering it as a linger 1667ab60b16dSAlex Elder * request to ensure the __map_request() below 1668ab60b16dSAlex Elder * will decide it needs to be sent. 1669ab60b16dSAlex Elder */ 1670ab60b16dSAlex Elder if (req->r_linger && list_empty(&req->r_linger_item)) { 1671ab60b16dSAlex Elder dout("%p tid %llu restart on osd%d\n", 1672ab60b16dSAlex Elder req, req->r_tid, 1673ab60b16dSAlex Elder req->r_osd ? req->r_osd->o_osd : -1); 167496e4dac6SAlex Elder ceph_osdc_get_request(req); 1675ab60b16dSAlex Elder __unregister_request(osdc, req); 1676ab60b16dSAlex Elder __register_linger_request(osdc, req); 167796e4dac6SAlex Elder ceph_osdc_put_request(req); 1678ab60b16dSAlex Elder continue; 1679ab60b16dSAlex Elder } 1680ab60b16dSAlex Elder 16819a1ea2dbSJosh Durgin force_resend_req = force_resend || 16829a1ea2dbSJosh Durgin (force_resend_writes && 16839a1ea2dbSJosh Durgin req->r_flags & CEPH_OSD_FLAG_WRITE); 16849a1ea2dbSJosh Durgin err = __map_request(osdc, req, force_resend_req); 16856f6c7006SSage Weil if (err < 0) 16866f6c7006SSage Weil continue; /* error */ 16876f6c7006SSage Weil if (req->r_osd == NULL) { 16886f6c7006SSage Weil dout("%p tid %llu maps to no osd\n", req, req->r_tid); 16896f6c7006SSage Weil needmap++; /* request a newer map */ 16906f6c7006SSage Weil } else if (err > 0) { 16916194ea89SSage Weil if (!req->r_linger) { 16926194ea89SSage Weil dout("%p tid %llu requeued on osd%d\n", req, 16936194ea89SSage Weil req->r_tid, 16946f6c7006SSage Weil req->r_osd ? req->r_osd->o_osd : -1); 16956f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 16966f6c7006SSage Weil } 16976f6c7006SSage Weil } 16986194ea89SSage Weil } 1699a40c4f10SYehuda Sadeh 1700a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osdc->req_linger, 1701a40c4f10SYehuda Sadeh r_linger_item) { 1702a40c4f10SYehuda Sadeh dout("linger req=%p req->r_osd=%p\n", req, req->r_osd); 1703a40c4f10SYehuda Sadeh 17049a1ea2dbSJosh Durgin err = __map_request(osdc, req, 17059a1ea2dbSJosh Durgin force_resend || force_resend_writes); 1706ab60b16dSAlex Elder dout("__map_request returned %d\n", err); 1707a40c4f10SYehuda Sadeh if (err == 0) 1708a40c4f10SYehuda Sadeh continue; /* no change and no osd was specified */ 1709a40c4f10SYehuda Sadeh if (err < 0) 1710a40c4f10SYehuda Sadeh continue; /* hrm! */ 1711a40c4f10SYehuda Sadeh if (req->r_osd == NULL) { 1712a40c4f10SYehuda Sadeh dout("tid %llu maps to no valid osd\n", req->r_tid); 1713a40c4f10SYehuda Sadeh needmap++; /* request a newer map */ 1714a40c4f10SYehuda Sadeh continue; 1715a40c4f10SYehuda Sadeh } 1716a40c4f10SYehuda Sadeh 1717a40c4f10SYehuda Sadeh dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid, 1718a40c4f10SYehuda Sadeh req->r_osd ? req->r_osd->o_osd : -1); 1719a40c4f10SYehuda Sadeh __register_request(osdc, req); 1720c89ce05eSAlex Elder __unregister_linger_request(osdc, req); 1721a40c4f10SYehuda Sadeh } 172214d2f38dSAlex Elder reset_changed_osds(osdc); 17233d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 17243d14c5d2SYehuda Sadeh 17253d14c5d2SYehuda Sadeh if (needmap) { 17263d14c5d2SYehuda Sadeh dout("%d requests for down osds, need new map\n", needmap); 17273d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 17283d14c5d2SYehuda Sadeh } 17293d14c5d2SYehuda Sadeh } 17306f6c7006SSage Weil 17316f6c7006SSage Weil 17323d14c5d2SYehuda Sadeh /* 17333d14c5d2SYehuda Sadeh * Process updated osd map. 17343d14c5d2SYehuda Sadeh * 17353d14c5d2SYehuda Sadeh * The message contains any number of incremental and full maps, normally 17363d14c5d2SYehuda Sadeh * indicating some sort of topology change in the cluster. Kick requests 17373d14c5d2SYehuda Sadeh * off to different OSDs as needed. 17383d14c5d2SYehuda Sadeh */ 17393d14c5d2SYehuda Sadeh void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg) 17403d14c5d2SYehuda Sadeh { 17413d14c5d2SYehuda Sadeh void *p, *end, *next; 17423d14c5d2SYehuda Sadeh u32 nr_maps, maplen; 17433d14c5d2SYehuda Sadeh u32 epoch; 17443d14c5d2SYehuda Sadeh struct ceph_osdmap *newmap = NULL, *oldmap; 17453d14c5d2SYehuda Sadeh int err; 17463d14c5d2SYehuda Sadeh struct ceph_fsid fsid; 17479a1ea2dbSJosh Durgin bool was_full; 17483d14c5d2SYehuda Sadeh 17493d14c5d2SYehuda Sadeh dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0); 17503d14c5d2SYehuda Sadeh p = msg->front.iov_base; 17513d14c5d2SYehuda Sadeh end = p + msg->front.iov_len; 17523d14c5d2SYehuda Sadeh 17533d14c5d2SYehuda Sadeh /* verify fsid */ 17543d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, sizeof(fsid), bad); 17553d14c5d2SYehuda Sadeh ceph_decode_copy(&p, &fsid, sizeof(fsid)); 17563d14c5d2SYehuda Sadeh if (ceph_check_fsid(osdc->client, &fsid) < 0) 17573d14c5d2SYehuda Sadeh return; 17583d14c5d2SYehuda Sadeh 17593d14c5d2SYehuda Sadeh down_write(&osdc->map_sem); 17603d14c5d2SYehuda Sadeh 17619a1ea2dbSJosh Durgin was_full = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL); 17629a1ea2dbSJosh Durgin 17633d14c5d2SYehuda Sadeh /* incremental maps */ 17643d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 17653d14c5d2SYehuda Sadeh dout(" %d inc maps\n", nr_maps); 17663d14c5d2SYehuda Sadeh while (nr_maps > 0) { 17673d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 17683d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 17693d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 17703d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 17713d14c5d2SYehuda Sadeh next = p + maplen; 17723d14c5d2SYehuda Sadeh if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) { 17733d14c5d2SYehuda Sadeh dout("applying incremental map %u len %d\n", 17743d14c5d2SYehuda Sadeh epoch, maplen); 17753d14c5d2SYehuda Sadeh newmap = osdmap_apply_incremental(&p, next, 17763d14c5d2SYehuda Sadeh osdc->osdmap, 177715d9882cSAlex Elder &osdc->client->msgr); 17783d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 17793d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 17803d14c5d2SYehuda Sadeh goto bad; 17813d14c5d2SYehuda Sadeh } 17823d14c5d2SYehuda Sadeh BUG_ON(!newmap); 17833d14c5d2SYehuda Sadeh if (newmap != osdc->osdmap) { 17843d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 17853d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 17863d14c5d2SYehuda Sadeh } 17879a1ea2dbSJosh Durgin was_full = was_full || 17889a1ea2dbSJosh Durgin ceph_osdmap_flag(osdc->osdmap, 17899a1ea2dbSJosh Durgin CEPH_OSDMAP_FULL); 17909a1ea2dbSJosh Durgin kick_requests(osdc, 0, was_full); 17913d14c5d2SYehuda Sadeh } else { 17923d14c5d2SYehuda Sadeh dout("ignoring incremental map %u len %d\n", 17933d14c5d2SYehuda Sadeh epoch, maplen); 17943d14c5d2SYehuda Sadeh } 17953d14c5d2SYehuda Sadeh p = next; 17963d14c5d2SYehuda Sadeh nr_maps--; 17973d14c5d2SYehuda Sadeh } 17983d14c5d2SYehuda Sadeh if (newmap) 17993d14c5d2SYehuda Sadeh goto done; 18003d14c5d2SYehuda Sadeh 18013d14c5d2SYehuda Sadeh /* full maps */ 18023d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 18033d14c5d2SYehuda Sadeh dout(" %d full maps\n", nr_maps); 18043d14c5d2SYehuda Sadeh while (nr_maps) { 18053d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 18063d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 18073d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 18083d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 18093d14c5d2SYehuda Sadeh if (nr_maps > 1) { 18103d14c5d2SYehuda Sadeh dout("skipping non-latest full map %u len %d\n", 18113d14c5d2SYehuda Sadeh epoch, maplen); 18123d14c5d2SYehuda Sadeh } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) { 18133d14c5d2SYehuda Sadeh dout("skipping full map %u len %d, " 18143d14c5d2SYehuda Sadeh "older than our %u\n", epoch, maplen, 18153d14c5d2SYehuda Sadeh osdc->osdmap->epoch); 18163d14c5d2SYehuda Sadeh } else { 181738d6453cSSage Weil int skipped_map = 0; 181838d6453cSSage Weil 18193d14c5d2SYehuda Sadeh dout("taking full map %u len %d\n", epoch, maplen); 18203d14c5d2SYehuda Sadeh newmap = osdmap_decode(&p, p+maplen); 18213d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 18223d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 18233d14c5d2SYehuda Sadeh goto bad; 18243d14c5d2SYehuda Sadeh } 18253d14c5d2SYehuda Sadeh BUG_ON(!newmap); 18263d14c5d2SYehuda Sadeh oldmap = osdc->osdmap; 18273d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 182838d6453cSSage Weil if (oldmap) { 182938d6453cSSage Weil if (oldmap->epoch + 1 < newmap->epoch) 183038d6453cSSage Weil skipped_map = 1; 18313d14c5d2SYehuda Sadeh ceph_osdmap_destroy(oldmap); 183238d6453cSSage Weil } 18339a1ea2dbSJosh Durgin was_full = was_full || 18349a1ea2dbSJosh Durgin ceph_osdmap_flag(osdc->osdmap, 18359a1ea2dbSJosh Durgin CEPH_OSDMAP_FULL); 18369a1ea2dbSJosh Durgin kick_requests(osdc, skipped_map, was_full); 18373d14c5d2SYehuda Sadeh } 18383d14c5d2SYehuda Sadeh p += maplen; 18393d14c5d2SYehuda Sadeh nr_maps--; 18403d14c5d2SYehuda Sadeh } 18413d14c5d2SYehuda Sadeh 1842b72e19b9SDan Carpenter if (!osdc->osdmap) 1843b72e19b9SDan Carpenter goto bad; 18443d14c5d2SYehuda Sadeh done: 18453d14c5d2SYehuda Sadeh downgrade_write(&osdc->map_sem); 18463d14c5d2SYehuda Sadeh ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch); 1847cd634fb6SSage Weil 1848cd634fb6SSage Weil /* 1849cd634fb6SSage Weil * subscribe to subsequent osdmap updates if full to ensure 1850cd634fb6SSage Weil * we find out when we are no longer full and stop returning 1851cd634fb6SSage Weil * ENOSPC. 1852cd634fb6SSage Weil */ 1853d29adb34SJosh Durgin if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) || 1854d29adb34SJosh Durgin ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD) || 1855d29adb34SJosh Durgin ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR)) 1856cd634fb6SSage Weil ceph_monc_request_next_osdmap(&osdc->client->monc); 1857cd634fb6SSage Weil 1858f9d25199SAlex Elder mutex_lock(&osdc->request_mutex); 1859f9d25199SAlex Elder __send_queued(osdc); 1860f9d25199SAlex Elder mutex_unlock(&osdc->request_mutex); 18613d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 18623d14c5d2SYehuda Sadeh wake_up_all(&osdc->client->auth_wq); 18633d14c5d2SYehuda Sadeh return; 18643d14c5d2SYehuda Sadeh 18653d14c5d2SYehuda Sadeh bad: 18663d14c5d2SYehuda Sadeh pr_err("osdc handle_map corrupt msg\n"); 18673d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 18683d14c5d2SYehuda Sadeh up_write(&osdc->map_sem); 18693d14c5d2SYehuda Sadeh return; 18703d14c5d2SYehuda Sadeh } 18713d14c5d2SYehuda Sadeh 18723d14c5d2SYehuda Sadeh /* 1873a40c4f10SYehuda Sadeh * watch/notify callback event infrastructure 1874a40c4f10SYehuda Sadeh * 1875a40c4f10SYehuda Sadeh * These callbacks are used both for watch and notify operations. 1876a40c4f10SYehuda Sadeh */ 1877a40c4f10SYehuda Sadeh static void __release_event(struct kref *kref) 1878a40c4f10SYehuda Sadeh { 1879a40c4f10SYehuda Sadeh struct ceph_osd_event *event = 1880a40c4f10SYehuda Sadeh container_of(kref, struct ceph_osd_event, kref); 1881a40c4f10SYehuda Sadeh 1882a40c4f10SYehuda Sadeh dout("__release_event %p\n", event); 1883a40c4f10SYehuda Sadeh kfree(event); 1884a40c4f10SYehuda Sadeh } 1885a40c4f10SYehuda Sadeh 1886a40c4f10SYehuda Sadeh static void get_event(struct ceph_osd_event *event) 1887a40c4f10SYehuda Sadeh { 1888a40c4f10SYehuda Sadeh kref_get(&event->kref); 1889a40c4f10SYehuda Sadeh } 1890a40c4f10SYehuda Sadeh 1891a40c4f10SYehuda Sadeh void ceph_osdc_put_event(struct ceph_osd_event *event) 1892a40c4f10SYehuda Sadeh { 1893a40c4f10SYehuda Sadeh kref_put(&event->kref, __release_event); 1894a40c4f10SYehuda Sadeh } 1895a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_put_event); 1896a40c4f10SYehuda Sadeh 1897a40c4f10SYehuda Sadeh static void __insert_event(struct ceph_osd_client *osdc, 1898a40c4f10SYehuda Sadeh struct ceph_osd_event *new) 1899a40c4f10SYehuda Sadeh { 1900a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 1901a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 1902a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 1903a40c4f10SYehuda Sadeh 1904a40c4f10SYehuda Sadeh while (*p) { 1905a40c4f10SYehuda Sadeh parent = *p; 1906a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 1907a40c4f10SYehuda Sadeh if (new->cookie < event->cookie) 1908a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 1909a40c4f10SYehuda Sadeh else if (new->cookie > event->cookie) 1910a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 1911a40c4f10SYehuda Sadeh else 1912a40c4f10SYehuda Sadeh BUG(); 1913a40c4f10SYehuda Sadeh } 1914a40c4f10SYehuda Sadeh 1915a40c4f10SYehuda Sadeh rb_link_node(&new->node, parent, p); 1916a40c4f10SYehuda Sadeh rb_insert_color(&new->node, &osdc->event_tree); 1917a40c4f10SYehuda Sadeh } 1918a40c4f10SYehuda Sadeh 1919a40c4f10SYehuda Sadeh static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc, 1920a40c4f10SYehuda Sadeh u64 cookie) 1921a40c4f10SYehuda Sadeh { 1922a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 1923a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 1924a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 1925a40c4f10SYehuda Sadeh 1926a40c4f10SYehuda Sadeh while (*p) { 1927a40c4f10SYehuda Sadeh parent = *p; 1928a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 1929a40c4f10SYehuda Sadeh if (cookie < event->cookie) 1930a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 1931a40c4f10SYehuda Sadeh else if (cookie > event->cookie) 1932a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 1933a40c4f10SYehuda Sadeh else 1934a40c4f10SYehuda Sadeh return event; 1935a40c4f10SYehuda Sadeh } 1936a40c4f10SYehuda Sadeh return NULL; 1937a40c4f10SYehuda Sadeh } 1938a40c4f10SYehuda Sadeh 1939a40c4f10SYehuda Sadeh static void __remove_event(struct ceph_osd_event *event) 1940a40c4f10SYehuda Sadeh { 1941a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 1942a40c4f10SYehuda Sadeh 1943a40c4f10SYehuda Sadeh if (!RB_EMPTY_NODE(&event->node)) { 1944a40c4f10SYehuda Sadeh dout("__remove_event removed %p\n", event); 1945a40c4f10SYehuda Sadeh rb_erase(&event->node, &osdc->event_tree); 1946a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 1947a40c4f10SYehuda Sadeh } else { 1948a40c4f10SYehuda Sadeh dout("__remove_event didn't remove %p\n", event); 1949a40c4f10SYehuda Sadeh } 1950a40c4f10SYehuda Sadeh } 1951a40c4f10SYehuda Sadeh 1952a40c4f10SYehuda Sadeh int ceph_osdc_create_event(struct ceph_osd_client *osdc, 1953a40c4f10SYehuda Sadeh void (*event_cb)(u64, u64, u8, void *), 19543c663bbdSAlex Elder void *data, struct ceph_osd_event **pevent) 1955a40c4f10SYehuda Sadeh { 1956a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 1957a40c4f10SYehuda Sadeh 1958a40c4f10SYehuda Sadeh event = kmalloc(sizeof(*event), GFP_NOIO); 1959a40c4f10SYehuda Sadeh if (!event) 1960a40c4f10SYehuda Sadeh return -ENOMEM; 1961a40c4f10SYehuda Sadeh 1962a40c4f10SYehuda Sadeh dout("create_event %p\n", event); 1963a40c4f10SYehuda Sadeh event->cb = event_cb; 19643c663bbdSAlex Elder event->one_shot = 0; 1965a40c4f10SYehuda Sadeh event->data = data; 1966a40c4f10SYehuda Sadeh event->osdc = osdc; 1967a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&event->osd_node); 19683ee5234dSAlex Elder RB_CLEAR_NODE(&event->node); 1969a40c4f10SYehuda Sadeh kref_init(&event->kref); /* one ref for us */ 1970a40c4f10SYehuda Sadeh kref_get(&event->kref); /* one ref for the caller */ 1971a40c4f10SYehuda Sadeh 1972a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 1973a40c4f10SYehuda Sadeh event->cookie = ++osdc->event_count; 1974a40c4f10SYehuda Sadeh __insert_event(osdc, event); 1975a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 1976a40c4f10SYehuda Sadeh 1977a40c4f10SYehuda Sadeh *pevent = event; 1978a40c4f10SYehuda Sadeh return 0; 1979a40c4f10SYehuda Sadeh } 1980a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_create_event); 1981a40c4f10SYehuda Sadeh 1982a40c4f10SYehuda Sadeh void ceph_osdc_cancel_event(struct ceph_osd_event *event) 1983a40c4f10SYehuda Sadeh { 1984a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 1985a40c4f10SYehuda Sadeh 1986a40c4f10SYehuda Sadeh dout("cancel_event %p\n", event); 1987a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 1988a40c4f10SYehuda Sadeh __remove_event(event); 1989a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 1990a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); /* caller's */ 1991a40c4f10SYehuda Sadeh } 1992a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_cancel_event); 1993a40c4f10SYehuda Sadeh 1994a40c4f10SYehuda Sadeh 1995a40c4f10SYehuda Sadeh static void do_event_work(struct work_struct *work) 1996a40c4f10SYehuda Sadeh { 1997a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work = 1998a40c4f10SYehuda Sadeh container_of(work, struct ceph_osd_event_work, work); 1999a40c4f10SYehuda Sadeh struct ceph_osd_event *event = event_work->event; 2000a40c4f10SYehuda Sadeh u64 ver = event_work->ver; 2001a40c4f10SYehuda Sadeh u64 notify_id = event_work->notify_id; 2002a40c4f10SYehuda Sadeh u8 opcode = event_work->opcode; 2003a40c4f10SYehuda Sadeh 2004a40c4f10SYehuda Sadeh dout("do_event_work completing %p\n", event); 2005a40c4f10SYehuda Sadeh event->cb(ver, notify_id, opcode, event->data); 2006a40c4f10SYehuda Sadeh dout("do_event_work completed %p\n", event); 2007a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 2008a40c4f10SYehuda Sadeh kfree(event_work); 2009a40c4f10SYehuda Sadeh } 2010a40c4f10SYehuda Sadeh 2011a40c4f10SYehuda Sadeh 2012a40c4f10SYehuda Sadeh /* 2013a40c4f10SYehuda Sadeh * Process osd watch notifications 2014a40c4f10SYehuda Sadeh */ 20153c663bbdSAlex Elder static void handle_watch_notify(struct ceph_osd_client *osdc, 20163c663bbdSAlex Elder struct ceph_msg *msg) 2017a40c4f10SYehuda Sadeh { 2018a40c4f10SYehuda Sadeh void *p, *end; 2019a40c4f10SYehuda Sadeh u8 proto_ver; 2020a40c4f10SYehuda Sadeh u64 cookie, ver, notify_id; 2021a40c4f10SYehuda Sadeh u8 opcode; 2022a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 2023a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work; 2024a40c4f10SYehuda Sadeh 2025a40c4f10SYehuda Sadeh p = msg->front.iov_base; 2026a40c4f10SYehuda Sadeh end = p + msg->front.iov_len; 2027a40c4f10SYehuda Sadeh 2028a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, proto_ver, bad); 2029a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, opcode, bad); 2030a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, cookie, bad); 2031a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, ver, bad); 2032a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, notify_id, bad); 2033a40c4f10SYehuda Sadeh 2034a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 2035a40c4f10SYehuda Sadeh event = __find_event(osdc, cookie); 2036a40c4f10SYehuda Sadeh if (event) { 20373c663bbdSAlex Elder BUG_ON(event->one_shot); 2038a40c4f10SYehuda Sadeh get_event(event); 2039a40c4f10SYehuda Sadeh } 2040a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 2041a40c4f10SYehuda Sadeh dout("handle_watch_notify cookie %lld ver %lld event %p\n", 2042a40c4f10SYehuda Sadeh cookie, ver, event); 2043a40c4f10SYehuda Sadeh if (event) { 2044a40c4f10SYehuda Sadeh event_work = kmalloc(sizeof(*event_work), GFP_NOIO); 2045a40c4f10SYehuda Sadeh if (!event_work) { 2046a40c4f10SYehuda Sadeh dout("ERROR: could not allocate event_work\n"); 2047a40c4f10SYehuda Sadeh goto done_err; 2048a40c4f10SYehuda Sadeh } 20496b0ae409SMariusz Kozlowski INIT_WORK(&event_work->work, do_event_work); 2050a40c4f10SYehuda Sadeh event_work->event = event; 2051a40c4f10SYehuda Sadeh event_work->ver = ver; 2052a40c4f10SYehuda Sadeh event_work->notify_id = notify_id; 2053a40c4f10SYehuda Sadeh event_work->opcode = opcode; 2054a40c4f10SYehuda Sadeh if (!queue_work(osdc->notify_wq, &event_work->work)) { 2055a40c4f10SYehuda Sadeh dout("WARNING: failed to queue notify event work\n"); 2056a40c4f10SYehuda Sadeh goto done_err; 2057a40c4f10SYehuda Sadeh } 2058a40c4f10SYehuda Sadeh } 2059a40c4f10SYehuda Sadeh 2060a40c4f10SYehuda Sadeh return; 2061a40c4f10SYehuda Sadeh 2062a40c4f10SYehuda Sadeh done_err: 2063a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 2064a40c4f10SYehuda Sadeh return; 2065a40c4f10SYehuda Sadeh 2066a40c4f10SYehuda Sadeh bad: 2067a40c4f10SYehuda Sadeh pr_err("osdc handle_watch_notify corrupt msg\n"); 2068a40c4f10SYehuda Sadeh return; 2069a40c4f10SYehuda Sadeh } 2070a40c4f10SYehuda Sadeh 2071a40c4f10SYehuda Sadeh /* 2072e65550fdSAlex Elder * build new request AND message 2073e65550fdSAlex Elder * 2074e65550fdSAlex Elder */ 2075e65550fdSAlex Elder void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off, 2076e65550fdSAlex Elder struct ceph_snap_context *snapc, u64 snap_id, 2077e65550fdSAlex Elder struct timespec *mtime) 2078e65550fdSAlex Elder { 2079e65550fdSAlex Elder struct ceph_msg *msg = req->r_request; 2080e65550fdSAlex Elder void *p; 2081e65550fdSAlex Elder size_t msg_size; 2082e65550fdSAlex Elder int flags = req->r_flags; 2083e65550fdSAlex Elder u64 data_len; 2084e65550fdSAlex Elder unsigned int i; 2085e65550fdSAlex Elder 2086e65550fdSAlex Elder req->r_snapid = snap_id; 2087e65550fdSAlex Elder req->r_snapc = ceph_get_snap_context(snapc); 2088e65550fdSAlex Elder 2089e65550fdSAlex Elder /* encode request */ 2090e65550fdSAlex Elder msg->hdr.version = cpu_to_le16(4); 2091e65550fdSAlex Elder 2092e65550fdSAlex Elder p = msg->front.iov_base; 2093e65550fdSAlex Elder ceph_encode_32(&p, 1); /* client_inc is always 1 */ 2094e65550fdSAlex Elder req->r_request_osdmap_epoch = p; 2095e65550fdSAlex Elder p += 4; 2096e65550fdSAlex Elder req->r_request_flags = p; 2097e65550fdSAlex Elder p += 4; 2098e65550fdSAlex Elder if (req->r_flags & CEPH_OSD_FLAG_WRITE) 2099e65550fdSAlex Elder ceph_encode_timespec(p, mtime); 2100e65550fdSAlex Elder p += sizeof(struct ceph_timespec); 2101e65550fdSAlex Elder req->r_request_reassert_version = p; 2102e65550fdSAlex Elder p += sizeof(struct ceph_eversion); /* will get filled in */ 2103e65550fdSAlex Elder 2104e65550fdSAlex Elder /* oloc */ 2105e65550fdSAlex Elder ceph_encode_8(&p, 4); 2106e65550fdSAlex Elder ceph_encode_8(&p, 4); 2107e65550fdSAlex Elder ceph_encode_32(&p, 8 + 4 + 4); 2108e65550fdSAlex Elder req->r_request_pool = p; 2109e65550fdSAlex Elder p += 8; 2110e65550fdSAlex Elder ceph_encode_32(&p, -1); /* preferred */ 2111e65550fdSAlex Elder ceph_encode_32(&p, 0); /* key len */ 2112e65550fdSAlex Elder 2113e65550fdSAlex Elder ceph_encode_8(&p, 1); 2114e65550fdSAlex Elder req->r_request_pgid = p; 2115e65550fdSAlex Elder p += 8 + 4; 2116e65550fdSAlex Elder ceph_encode_32(&p, -1); /* preferred */ 2117e65550fdSAlex Elder 2118e65550fdSAlex Elder /* oid */ 2119e65550fdSAlex Elder ceph_encode_32(&p, req->r_oid_len); 2120e65550fdSAlex Elder memcpy(p, req->r_oid, req->r_oid_len); 2121e65550fdSAlex Elder dout("oid '%.*s' len %d\n", req->r_oid_len, req->r_oid, req->r_oid_len); 2122e65550fdSAlex Elder p += req->r_oid_len; 2123e65550fdSAlex Elder 2124e65550fdSAlex Elder /* ops--can imply data */ 2125e65550fdSAlex Elder ceph_encode_16(&p, (u16)req->r_num_ops); 2126e65550fdSAlex Elder data_len = 0; 2127e65550fdSAlex Elder for (i = 0; i < req->r_num_ops; i++) { 2128e65550fdSAlex Elder data_len += osd_req_encode_op(req, p, i); 2129e65550fdSAlex Elder p += sizeof(struct ceph_osd_op); 2130e65550fdSAlex Elder } 2131e65550fdSAlex Elder 2132e65550fdSAlex Elder /* snaps */ 2133e65550fdSAlex Elder ceph_encode_64(&p, req->r_snapid); 2134e65550fdSAlex Elder ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0); 2135e65550fdSAlex Elder ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0); 2136e65550fdSAlex Elder if (req->r_snapc) { 2137e65550fdSAlex Elder for (i = 0; i < snapc->num_snaps; i++) { 2138e65550fdSAlex Elder ceph_encode_64(&p, req->r_snapc->snaps[i]); 2139e65550fdSAlex Elder } 2140e65550fdSAlex Elder } 2141e65550fdSAlex Elder 2142e65550fdSAlex Elder req->r_request_attempts = p; 2143e65550fdSAlex Elder p += 4; 2144e65550fdSAlex Elder 2145e65550fdSAlex Elder /* data */ 2146e65550fdSAlex Elder if (flags & CEPH_OSD_FLAG_WRITE) { 2147e65550fdSAlex Elder u16 data_off; 2148e65550fdSAlex Elder 2149e65550fdSAlex Elder /* 2150e65550fdSAlex Elder * The header "data_off" is a hint to the receiver 2151e65550fdSAlex Elder * allowing it to align received data into its 2152e65550fdSAlex Elder * buffers such that there's no need to re-copy 2153e65550fdSAlex Elder * it before writing it to disk (direct I/O). 2154e65550fdSAlex Elder */ 2155e65550fdSAlex Elder data_off = (u16) (off & 0xffff); 2156e65550fdSAlex Elder req->r_request->hdr.data_off = cpu_to_le16(data_off); 2157e65550fdSAlex Elder } 2158e65550fdSAlex Elder req->r_request->hdr.data_len = cpu_to_le32(data_len); 2159e65550fdSAlex Elder 2160e65550fdSAlex Elder BUG_ON(p > msg->front.iov_base + msg->front.iov_len); 2161e65550fdSAlex Elder msg_size = p - msg->front.iov_base; 2162e65550fdSAlex Elder msg->front.iov_len = msg_size; 2163e65550fdSAlex Elder msg->hdr.front_len = cpu_to_le32(msg_size); 2164e65550fdSAlex Elder 2165e65550fdSAlex Elder dout("build_request msg_size was %d\n", (int)msg_size); 2166e65550fdSAlex Elder } 2167e65550fdSAlex Elder EXPORT_SYMBOL(ceph_osdc_build_request); 2168e65550fdSAlex Elder 2169e65550fdSAlex Elder /* 21703d14c5d2SYehuda Sadeh * Register request, send initial attempt. 21713d14c5d2SYehuda Sadeh */ 21723d14c5d2SYehuda Sadeh int ceph_osdc_start_request(struct ceph_osd_client *osdc, 21733d14c5d2SYehuda Sadeh struct ceph_osd_request *req, 21743d14c5d2SYehuda Sadeh bool nofail) 21753d14c5d2SYehuda Sadeh { 21763d14c5d2SYehuda Sadeh int rc = 0; 21773d14c5d2SYehuda Sadeh 21783d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 21793d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 2180dc4b870cSAlex Elder __register_request(osdc, req); 2181c10ebbf5SAlex Elder req->r_sent = 0; 2182c10ebbf5SAlex Elder req->r_got_reply = 0; 218338d6453cSSage Weil rc = __map_request(osdc, req, 0); 21849d6fcb08SSage Weil if (rc < 0) { 21859d6fcb08SSage Weil if (nofail) { 21869d6fcb08SSage Weil dout("osdc_start_request failed map, " 21879d6fcb08SSage Weil " will retry %lld\n", req->r_tid); 21889d6fcb08SSage Weil rc = 0; 218973d9f7eeSmajianpeng } else { 219073d9f7eeSmajianpeng __unregister_request(osdc, req); 21919d6fcb08SSage Weil } 2192234af26fSDan Carpenter goto out_unlock; 21939d6fcb08SSage Weil } 21946f6c7006SSage Weil if (req->r_osd == NULL) { 21956f6c7006SSage Weil dout("send_request %p no up osds in pg\n", req); 21966f6c7006SSage Weil ceph_monc_request_next_osdmap(&osdc->client->monc); 21976f6c7006SSage Weil } else { 21987e2766a1SAlex Elder __send_queued(osdc); 219956e925b6SSage Weil } 22003d14c5d2SYehuda Sadeh rc = 0; 2201234af26fSDan Carpenter out_unlock: 22023d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 22033d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 22043d14c5d2SYehuda Sadeh return rc; 22053d14c5d2SYehuda Sadeh } 22063d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_start_request); 22073d14c5d2SYehuda Sadeh 22083d14c5d2SYehuda Sadeh /* 22093d14c5d2SYehuda Sadeh * wait for a request to complete 22103d14c5d2SYehuda Sadeh */ 22113d14c5d2SYehuda Sadeh int ceph_osdc_wait_request(struct ceph_osd_client *osdc, 22123d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 22133d14c5d2SYehuda Sadeh { 22143d14c5d2SYehuda Sadeh int rc; 22153d14c5d2SYehuda Sadeh 22163d14c5d2SYehuda Sadeh rc = wait_for_completion_interruptible(&req->r_completion); 22173d14c5d2SYehuda Sadeh if (rc < 0) { 22183d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 22193d14c5d2SYehuda Sadeh __cancel_request(req); 22203d14c5d2SYehuda Sadeh __unregister_request(osdc, req); 22213d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 222225845472SSage Weil complete_request(req); 22233d14c5d2SYehuda Sadeh dout("wait_request tid %llu canceled/timed out\n", req->r_tid); 22243d14c5d2SYehuda Sadeh return rc; 22253d14c5d2SYehuda Sadeh } 22263d14c5d2SYehuda Sadeh 22273d14c5d2SYehuda Sadeh dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result); 22283d14c5d2SYehuda Sadeh return req->r_result; 22293d14c5d2SYehuda Sadeh } 22303d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_wait_request); 22313d14c5d2SYehuda Sadeh 22323d14c5d2SYehuda Sadeh /* 22333d14c5d2SYehuda Sadeh * sync - wait for all in-flight requests to flush. avoid starvation. 22343d14c5d2SYehuda Sadeh */ 22353d14c5d2SYehuda Sadeh void ceph_osdc_sync(struct ceph_osd_client *osdc) 22363d14c5d2SYehuda Sadeh { 22373d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 22383d14c5d2SYehuda Sadeh u64 last_tid, next_tid = 0; 22393d14c5d2SYehuda Sadeh 22403d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 22413d14c5d2SYehuda Sadeh last_tid = osdc->last_tid; 22423d14c5d2SYehuda Sadeh while (1) { 22433d14c5d2SYehuda Sadeh req = __lookup_request_ge(osdc, next_tid); 22443d14c5d2SYehuda Sadeh if (!req) 22453d14c5d2SYehuda Sadeh break; 22463d14c5d2SYehuda Sadeh if (req->r_tid > last_tid) 22473d14c5d2SYehuda Sadeh break; 22483d14c5d2SYehuda Sadeh 22493d14c5d2SYehuda Sadeh next_tid = req->r_tid + 1; 22503d14c5d2SYehuda Sadeh if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0) 22513d14c5d2SYehuda Sadeh continue; 22523d14c5d2SYehuda Sadeh 22533d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 22543d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 22553d14c5d2SYehuda Sadeh dout("sync waiting on tid %llu (last is %llu)\n", 22563d14c5d2SYehuda Sadeh req->r_tid, last_tid); 22573d14c5d2SYehuda Sadeh wait_for_completion(&req->r_safe_completion); 22583d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 22593d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 22603d14c5d2SYehuda Sadeh } 22613d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 22623d14c5d2SYehuda Sadeh dout("sync done (thru tid %llu)\n", last_tid); 22633d14c5d2SYehuda Sadeh } 22643d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_sync); 22653d14c5d2SYehuda Sadeh 22663d14c5d2SYehuda Sadeh /* 2267dd935f44SJosh Durgin * Call all pending notify callbacks - for use after a watch is 2268dd935f44SJosh Durgin * unregistered, to make sure no more callbacks for it will be invoked 2269dd935f44SJosh Durgin */ 2270dd935f44SJosh Durgin extern void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc) 2271dd935f44SJosh Durgin { 2272dd935f44SJosh Durgin flush_workqueue(osdc->notify_wq); 2273dd935f44SJosh Durgin } 2274dd935f44SJosh Durgin EXPORT_SYMBOL(ceph_osdc_flush_notifies); 2275dd935f44SJosh Durgin 2276dd935f44SJosh Durgin 2277dd935f44SJosh Durgin /* 22783d14c5d2SYehuda Sadeh * init, shutdown 22793d14c5d2SYehuda Sadeh */ 22803d14c5d2SYehuda Sadeh int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client) 22813d14c5d2SYehuda Sadeh { 22823d14c5d2SYehuda Sadeh int err; 22833d14c5d2SYehuda Sadeh 22843d14c5d2SYehuda Sadeh dout("init\n"); 22853d14c5d2SYehuda Sadeh osdc->client = client; 22863d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 22873d14c5d2SYehuda Sadeh init_rwsem(&osdc->map_sem); 22883d14c5d2SYehuda Sadeh init_completion(&osdc->map_waiters); 22893d14c5d2SYehuda Sadeh osdc->last_requested_map = 0; 22903d14c5d2SYehuda Sadeh mutex_init(&osdc->request_mutex); 22913d14c5d2SYehuda Sadeh osdc->last_tid = 0; 22923d14c5d2SYehuda Sadeh osdc->osds = RB_ROOT; 22933d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->osd_lru); 22943d14c5d2SYehuda Sadeh osdc->requests = RB_ROOT; 22953d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_lru); 22966f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_unsent); 22976f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_notarget); 2298a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_linger); 22993d14c5d2SYehuda Sadeh osdc->num_requests = 0; 23003d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout); 23013d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout); 2302a40c4f10SYehuda Sadeh spin_lock_init(&osdc->event_lock); 2303a40c4f10SYehuda Sadeh osdc->event_tree = RB_ROOT; 2304a40c4f10SYehuda Sadeh osdc->event_count = 0; 23053d14c5d2SYehuda Sadeh 23063d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 23073d14c5d2SYehuda Sadeh round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ)); 23083d14c5d2SYehuda Sadeh 23093d14c5d2SYehuda Sadeh err = -ENOMEM; 23103d14c5d2SYehuda Sadeh osdc->req_mempool = mempool_create_kmalloc_pool(10, 23113d14c5d2SYehuda Sadeh sizeof(struct ceph_osd_request)); 23123d14c5d2SYehuda Sadeh if (!osdc->req_mempool) 23133d14c5d2SYehuda Sadeh goto out; 23143d14c5d2SYehuda Sadeh 2315d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP, 2316d50b409fSSage Weil OSD_OP_FRONT_LEN, 10, true, 23173d14c5d2SYehuda Sadeh "osd_op"); 23183d14c5d2SYehuda Sadeh if (err < 0) 23193d14c5d2SYehuda Sadeh goto out_mempool; 2320d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY, 23213d14c5d2SYehuda Sadeh OSD_OPREPLY_FRONT_LEN, 10, true, 23223d14c5d2SYehuda Sadeh "osd_op_reply"); 23233d14c5d2SYehuda Sadeh if (err < 0) 23243d14c5d2SYehuda Sadeh goto out_msgpool; 2325a40c4f10SYehuda Sadeh 2326dbcae088SDan Carpenter err = -ENOMEM; 2327a40c4f10SYehuda Sadeh osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify"); 2328dbcae088SDan Carpenter if (!osdc->notify_wq) 2329a40c4f10SYehuda Sadeh goto out_msgpool; 23303d14c5d2SYehuda Sadeh return 0; 23313d14c5d2SYehuda Sadeh 23323d14c5d2SYehuda Sadeh out_msgpool: 23333d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 23343d14c5d2SYehuda Sadeh out_mempool: 23353d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 23363d14c5d2SYehuda Sadeh out: 23373d14c5d2SYehuda Sadeh return err; 23383d14c5d2SYehuda Sadeh } 23393d14c5d2SYehuda Sadeh 23403d14c5d2SYehuda Sadeh void ceph_osdc_stop(struct ceph_osd_client *osdc) 23413d14c5d2SYehuda Sadeh { 2342a40c4f10SYehuda Sadeh flush_workqueue(osdc->notify_wq); 2343a40c4f10SYehuda Sadeh destroy_workqueue(osdc->notify_wq); 23443d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->timeout_work); 23453d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->osds_timeout_work); 23463d14c5d2SYehuda Sadeh if (osdc->osdmap) { 23473d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 23483d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 23493d14c5d2SYehuda Sadeh } 2350aca420bcSSage Weil remove_all_osds(osdc); 23513d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 23523d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 23533d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op_reply); 23543d14c5d2SYehuda Sadeh } 23553d14c5d2SYehuda Sadeh 23563d14c5d2SYehuda Sadeh /* 23573d14c5d2SYehuda Sadeh * Read some contiguous pages. If we cross a stripe boundary, shorten 23583d14c5d2SYehuda Sadeh * *plen. Return number of bytes read, or error. 23593d14c5d2SYehuda Sadeh */ 23603d14c5d2SYehuda Sadeh int ceph_osdc_readpages(struct ceph_osd_client *osdc, 23613d14c5d2SYehuda Sadeh struct ceph_vino vino, struct ceph_file_layout *layout, 23623d14c5d2SYehuda Sadeh u64 off, u64 *plen, 23633d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 2364b7495fc2SSage Weil struct page **pages, int num_pages, int page_align) 23653d14c5d2SYehuda Sadeh { 23663d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 23673d14c5d2SYehuda Sadeh int rc = 0; 23683d14c5d2SYehuda Sadeh 23693d14c5d2SYehuda Sadeh dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino, 23703d14c5d2SYehuda Sadeh vino.snap, off, *plen); 237179528734SAlex Elder req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 1, 23723d14c5d2SYehuda Sadeh CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ, 2373acead002SAlex Elder NULL, truncate_seq, truncate_size, 2374153e5167SAlex Elder false); 23756816282dSSage Weil if (IS_ERR(req)) 23766816282dSSage Weil return PTR_ERR(req); 23773d14c5d2SYehuda Sadeh 23783d14c5d2SYehuda Sadeh /* it may be a short read due to an object boundary */ 23790fff87ecSAlex Elder 2380406e2c9fSAlex Elder osd_req_op_extent_osd_data_pages(req, 0, 2381a4ce40a9SAlex Elder pages, *plen, page_align, false, false); 23823d14c5d2SYehuda Sadeh 2383e0c59487SAlex Elder dout("readpages final extent is %llu~%llu (%llu bytes align %d)\n", 238443bfe5deSAlex Elder off, *plen, *plen, page_align); 23853d14c5d2SYehuda Sadeh 238679528734SAlex Elder ceph_osdc_build_request(req, off, NULL, vino.snap, NULL); 238702ee07d3SAlex Elder 23883d14c5d2SYehuda Sadeh rc = ceph_osdc_start_request(osdc, req, false); 23893d14c5d2SYehuda Sadeh if (!rc) 23903d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 23913d14c5d2SYehuda Sadeh 23923d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 23933d14c5d2SYehuda Sadeh dout("readpages result %d\n", rc); 23943d14c5d2SYehuda Sadeh return rc; 23953d14c5d2SYehuda Sadeh } 23963d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_readpages); 23973d14c5d2SYehuda Sadeh 23983d14c5d2SYehuda Sadeh /* 23993d14c5d2SYehuda Sadeh * do a synchronous write on N pages 24003d14c5d2SYehuda Sadeh */ 24013d14c5d2SYehuda Sadeh int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino, 24023d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 24033d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 24043d14c5d2SYehuda Sadeh u64 off, u64 len, 24053d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 24063d14c5d2SYehuda Sadeh struct timespec *mtime, 240724808826SAlex Elder struct page **pages, int num_pages) 24083d14c5d2SYehuda Sadeh { 24093d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 24103d14c5d2SYehuda Sadeh int rc = 0; 2411b7495fc2SSage Weil int page_align = off & ~PAGE_MASK; 24123d14c5d2SYehuda Sadeh 2413acead002SAlex Elder BUG_ON(vino.snap != CEPH_NOSNAP); /* snapshots aren't writeable */ 241479528734SAlex Elder req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 1, 24153d14c5d2SYehuda Sadeh CEPH_OSD_OP_WRITE, 241624808826SAlex Elder CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE, 2417acead002SAlex Elder snapc, truncate_seq, truncate_size, 2418153e5167SAlex Elder true); 24196816282dSSage Weil if (IS_ERR(req)) 24206816282dSSage Weil return PTR_ERR(req); 24213d14c5d2SYehuda Sadeh 24223d14c5d2SYehuda Sadeh /* it may be a short write due to an object boundary */ 2423406e2c9fSAlex Elder osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align, 242443bfe5deSAlex Elder false, false); 242543bfe5deSAlex Elder dout("writepages %llu~%llu (%llu bytes)\n", off, len, len); 24263d14c5d2SYehuda Sadeh 242779528734SAlex Elder ceph_osdc_build_request(req, off, snapc, CEPH_NOSNAP, mtime); 242802ee07d3SAlex Elder 242987f979d3SAlex Elder rc = ceph_osdc_start_request(osdc, req, true); 24303d14c5d2SYehuda Sadeh if (!rc) 24313d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 24323d14c5d2SYehuda Sadeh 24333d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 24343d14c5d2SYehuda Sadeh if (rc == 0) 24353d14c5d2SYehuda Sadeh rc = len; 24363d14c5d2SYehuda Sadeh dout("writepages result %d\n", rc); 24373d14c5d2SYehuda Sadeh return rc; 24383d14c5d2SYehuda Sadeh } 24393d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_writepages); 24403d14c5d2SYehuda Sadeh 24415522ae0bSAlex Elder int ceph_osdc_setup(void) 24425522ae0bSAlex Elder { 24435522ae0bSAlex Elder BUG_ON(ceph_osd_request_cache); 24445522ae0bSAlex Elder ceph_osd_request_cache = kmem_cache_create("ceph_osd_request", 24455522ae0bSAlex Elder sizeof (struct ceph_osd_request), 24465522ae0bSAlex Elder __alignof__(struct ceph_osd_request), 24475522ae0bSAlex Elder 0, NULL); 24485522ae0bSAlex Elder 24495522ae0bSAlex Elder return ceph_osd_request_cache ? 0 : -ENOMEM; 24505522ae0bSAlex Elder } 24515522ae0bSAlex Elder EXPORT_SYMBOL(ceph_osdc_setup); 24525522ae0bSAlex Elder 24535522ae0bSAlex Elder void ceph_osdc_cleanup(void) 24545522ae0bSAlex Elder { 24555522ae0bSAlex Elder BUG_ON(!ceph_osd_request_cache); 24565522ae0bSAlex Elder kmem_cache_destroy(ceph_osd_request_cache); 24575522ae0bSAlex Elder ceph_osd_request_cache = NULL; 24585522ae0bSAlex Elder } 24595522ae0bSAlex Elder EXPORT_SYMBOL(ceph_osdc_cleanup); 24605522ae0bSAlex Elder 24613d14c5d2SYehuda Sadeh /* 24623d14c5d2SYehuda Sadeh * handle incoming message 24633d14c5d2SYehuda Sadeh */ 24643d14c5d2SYehuda Sadeh static void dispatch(struct ceph_connection *con, struct ceph_msg *msg) 24653d14c5d2SYehuda Sadeh { 24663d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 24673d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 24683d14c5d2SYehuda Sadeh int type = le16_to_cpu(msg->hdr.type); 24693d14c5d2SYehuda Sadeh 24703d14c5d2SYehuda Sadeh if (!osd) 24713d14c5d2SYehuda Sadeh goto out; 24723d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 24733d14c5d2SYehuda Sadeh 24743d14c5d2SYehuda Sadeh switch (type) { 24753d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 24763d14c5d2SYehuda Sadeh ceph_osdc_handle_map(osdc, msg); 24773d14c5d2SYehuda Sadeh break; 24783d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 24793d14c5d2SYehuda Sadeh handle_reply(osdc, msg, con); 24803d14c5d2SYehuda Sadeh break; 2481a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 2482a40c4f10SYehuda Sadeh handle_watch_notify(osdc, msg); 2483a40c4f10SYehuda Sadeh break; 24843d14c5d2SYehuda Sadeh 24853d14c5d2SYehuda Sadeh default: 24863d14c5d2SYehuda Sadeh pr_err("received unknown message type %d %s\n", type, 24873d14c5d2SYehuda Sadeh ceph_msg_type_name(type)); 24883d14c5d2SYehuda Sadeh } 24893d14c5d2SYehuda Sadeh out: 24903d14c5d2SYehuda Sadeh ceph_msg_put(msg); 24913d14c5d2SYehuda Sadeh } 24923d14c5d2SYehuda Sadeh 24933d14c5d2SYehuda Sadeh /* 24943d14c5d2SYehuda Sadeh * lookup and return message for incoming reply. set up reply message 24953d14c5d2SYehuda Sadeh * pages. 24963d14c5d2SYehuda Sadeh */ 24973d14c5d2SYehuda Sadeh static struct ceph_msg *get_reply(struct ceph_connection *con, 24983d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 24993d14c5d2SYehuda Sadeh int *skip) 25003d14c5d2SYehuda Sadeh { 25013d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 25023d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = osd->o_osdc; 25033d14c5d2SYehuda Sadeh struct ceph_msg *m; 25043d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 25053f0a4ac5SIlya Dryomov int front_len = le32_to_cpu(hdr->front_len); 25063d14c5d2SYehuda Sadeh int data_len = le32_to_cpu(hdr->data_len); 25073d14c5d2SYehuda Sadeh u64 tid; 25083d14c5d2SYehuda Sadeh 25093d14c5d2SYehuda Sadeh tid = le64_to_cpu(hdr->tid); 25103d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 25113d14c5d2SYehuda Sadeh req = __lookup_request(osdc, tid); 25123d14c5d2SYehuda Sadeh if (!req) { 25133d14c5d2SYehuda Sadeh *skip = 1; 25143d14c5d2SYehuda Sadeh m = NULL; 2515756a16a5SSage Weil dout("get_reply unknown tid %llu from osd%d\n", tid, 25163d14c5d2SYehuda Sadeh osd->o_osd); 25173d14c5d2SYehuda Sadeh goto out; 25183d14c5d2SYehuda Sadeh } 25193d14c5d2SYehuda Sadeh 2520ace6d3a9SAlex Elder if (req->r_reply->con) 25218921d114SAlex Elder dout("%s revoking msg %p from old con %p\n", __func__, 2522ace6d3a9SAlex Elder req->r_reply, req->r_reply->con); 25238921d114SAlex Elder ceph_msg_revoke_incoming(req->r_reply); 25243d14c5d2SYehuda Sadeh 2525f2be82b0SIlya Dryomov if (front_len > req->r_reply->front_alloc_len) { 25264974341eSAlex Elder pr_warning("get_reply front %d > preallocated %d (%u#%llu)\n", 2527f2be82b0SIlya Dryomov front_len, req->r_reply->front_alloc_len, 25284974341eSAlex Elder (unsigned int)con->peer_name.type, 25294974341eSAlex Elder le64_to_cpu(con->peer_name.num)); 25303f0a4ac5SIlya Dryomov m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front_len, GFP_NOFS, 25313f0a4ac5SIlya Dryomov false); 25323d14c5d2SYehuda Sadeh if (!m) 25333d14c5d2SYehuda Sadeh goto out; 25343d14c5d2SYehuda Sadeh ceph_msg_put(req->r_reply); 25353d14c5d2SYehuda Sadeh req->r_reply = m; 25363d14c5d2SYehuda Sadeh } 25373d14c5d2SYehuda Sadeh m = ceph_msg_get(req->r_reply); 25383d14c5d2SYehuda Sadeh 25393d14c5d2SYehuda Sadeh if (data_len > 0) { 2540a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 25410fff87ecSAlex Elder 2542a4ce40a9SAlex Elder /* 2543a4ce40a9SAlex Elder * XXX This is assuming there is only one op containing 2544a4ce40a9SAlex Elder * XXX page data. Probably OK for reads, but this 2545a4ce40a9SAlex Elder * XXX ought to be done more generally. 2546a4ce40a9SAlex Elder */ 2547406e2c9fSAlex Elder osd_data = osd_req_op_extent_osd_data(req, 0); 25480fff87ecSAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 25490fff87ecSAlex Elder if (osd_data->pages && 2550e0c59487SAlex Elder unlikely(osd_data->length < data_len)) { 25512ac2b7a6SAlex Elder 2552e0c59487SAlex Elder pr_warning("tid %lld reply has %d bytes " 2553e0c59487SAlex Elder "we had only %llu bytes ready\n", 2554e0c59487SAlex Elder tid, data_len, osd_data->length); 25553d14c5d2SYehuda Sadeh *skip = 1; 25563d14c5d2SYehuda Sadeh ceph_msg_put(m); 25573d14c5d2SYehuda Sadeh m = NULL; 25583d14c5d2SYehuda Sadeh goto out; 25593d14c5d2SYehuda Sadeh } 25603d14c5d2SYehuda Sadeh } 25612ac2b7a6SAlex Elder } 25623d14c5d2SYehuda Sadeh *skip = 0; 25633d14c5d2SYehuda Sadeh dout("get_reply tid %lld %p\n", tid, m); 25643d14c5d2SYehuda Sadeh 25653d14c5d2SYehuda Sadeh out: 25663d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 25673d14c5d2SYehuda Sadeh return m; 25683d14c5d2SYehuda Sadeh 25693d14c5d2SYehuda Sadeh } 25703d14c5d2SYehuda Sadeh 25713d14c5d2SYehuda Sadeh static struct ceph_msg *alloc_msg(struct ceph_connection *con, 25723d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 25733d14c5d2SYehuda Sadeh int *skip) 25743d14c5d2SYehuda Sadeh { 25753d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 25763d14c5d2SYehuda Sadeh int type = le16_to_cpu(hdr->type); 25773d14c5d2SYehuda Sadeh int front = le32_to_cpu(hdr->front_len); 25783d14c5d2SYehuda Sadeh 25791c20f2d2SAlex Elder *skip = 0; 25803d14c5d2SYehuda Sadeh switch (type) { 25813d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 2582a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 2583b61c2763SSage Weil return ceph_msg_new(type, front, GFP_NOFS, false); 25843d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 25853d14c5d2SYehuda Sadeh return get_reply(con, hdr, skip); 25863d14c5d2SYehuda Sadeh default: 25873d14c5d2SYehuda Sadeh pr_info("alloc_msg unexpected msg type %d from osd%d\n", type, 25883d14c5d2SYehuda Sadeh osd->o_osd); 25893d14c5d2SYehuda Sadeh *skip = 1; 25903d14c5d2SYehuda Sadeh return NULL; 25913d14c5d2SYehuda Sadeh } 25923d14c5d2SYehuda Sadeh } 25933d14c5d2SYehuda Sadeh 25943d14c5d2SYehuda Sadeh /* 25953d14c5d2SYehuda Sadeh * Wrappers to refcount containing ceph_osd struct 25963d14c5d2SYehuda Sadeh */ 25973d14c5d2SYehuda Sadeh static struct ceph_connection *get_osd_con(struct ceph_connection *con) 25983d14c5d2SYehuda Sadeh { 25993d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 26003d14c5d2SYehuda Sadeh if (get_osd(osd)) 26013d14c5d2SYehuda Sadeh return con; 26023d14c5d2SYehuda Sadeh return NULL; 26033d14c5d2SYehuda Sadeh } 26043d14c5d2SYehuda Sadeh 26053d14c5d2SYehuda Sadeh static void put_osd_con(struct ceph_connection *con) 26063d14c5d2SYehuda Sadeh { 26073d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 26083d14c5d2SYehuda Sadeh put_osd(osd); 26093d14c5d2SYehuda Sadeh } 26103d14c5d2SYehuda Sadeh 26113d14c5d2SYehuda Sadeh /* 26123d14c5d2SYehuda Sadeh * authentication 26133d14c5d2SYehuda Sadeh */ 2614a3530df3SAlex Elder /* 2615a3530df3SAlex Elder * Note: returned pointer is the address of a structure that's 2616a3530df3SAlex Elder * managed separately. Caller must *not* attempt to free it. 2617a3530df3SAlex Elder */ 2618a3530df3SAlex Elder static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con, 26198f43fb53SAlex Elder int *proto, int force_new) 26203d14c5d2SYehuda Sadeh { 26213d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 26223d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 26233d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 262474f1869fSAlex Elder struct ceph_auth_handshake *auth = &o->o_auth; 26253d14c5d2SYehuda Sadeh 262674f1869fSAlex Elder if (force_new && auth->authorizer) { 262727859f97SSage Weil ceph_auth_destroy_authorizer(ac, auth->authorizer); 262874f1869fSAlex Elder auth->authorizer = NULL; 26293d14c5d2SYehuda Sadeh } 263027859f97SSage Weil if (!auth->authorizer) { 263127859f97SSage Weil int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 2632a3530df3SAlex Elder auth); 26333d14c5d2SYehuda Sadeh if (ret) 2634a3530df3SAlex Elder return ERR_PTR(ret); 263527859f97SSage Weil } else { 263627859f97SSage Weil int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 26370bed9b5cSSage Weil auth); 26380bed9b5cSSage Weil if (ret) 26390bed9b5cSSage Weil return ERR_PTR(ret); 26403d14c5d2SYehuda Sadeh } 26413d14c5d2SYehuda Sadeh *proto = ac->protocol; 264274f1869fSAlex Elder 2643a3530df3SAlex Elder return auth; 26443d14c5d2SYehuda Sadeh } 26453d14c5d2SYehuda Sadeh 26463d14c5d2SYehuda Sadeh 26473d14c5d2SYehuda Sadeh static int verify_authorizer_reply(struct ceph_connection *con, int len) 26483d14c5d2SYehuda Sadeh { 26493d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 26503d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 26513d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 26523d14c5d2SYehuda Sadeh 265327859f97SSage Weil return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len); 26543d14c5d2SYehuda Sadeh } 26553d14c5d2SYehuda Sadeh 26563d14c5d2SYehuda Sadeh static int invalidate_authorizer(struct ceph_connection *con) 26573d14c5d2SYehuda Sadeh { 26583d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 26593d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 26603d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 26613d14c5d2SYehuda Sadeh 266227859f97SSage Weil ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD); 26633d14c5d2SYehuda Sadeh return ceph_monc_validate_auth(&osdc->client->monc); 26643d14c5d2SYehuda Sadeh } 26653d14c5d2SYehuda Sadeh 26663d14c5d2SYehuda Sadeh static const struct ceph_connection_operations osd_con_ops = { 26673d14c5d2SYehuda Sadeh .get = get_osd_con, 26683d14c5d2SYehuda Sadeh .put = put_osd_con, 26693d14c5d2SYehuda Sadeh .dispatch = dispatch, 26703d14c5d2SYehuda Sadeh .get_authorizer = get_authorizer, 26713d14c5d2SYehuda Sadeh .verify_authorizer_reply = verify_authorizer_reply, 26723d14c5d2SYehuda Sadeh .invalidate_authorizer = invalidate_authorizer, 26733d14c5d2SYehuda Sadeh .alloc_msg = alloc_msg, 26743d14c5d2SYehuda Sadeh .fault = osd_reset, 26753d14c5d2SYehuda Sadeh }; 2676