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 */ 3009e94af20SIlya Dryomov static void ceph_osdc_release_request(struct kref *kref) 3013d14c5d2SYehuda Sadeh { 3029e94af20SIlya Dryomov struct ceph_osd_request *req = container_of(kref, 3039e94af20SIlya Dryomov struct ceph_osd_request, r_kref); 3045476492fSAlex Elder unsigned int which; 3053d14c5d2SYehuda Sadeh 3069e94af20SIlya Dryomov dout("%s %p (r_request %p r_reply %p)\n", __func__, req, 3079e94af20SIlya Dryomov req->r_request, req->r_reply); 3086562d661SIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&req->r_node)); 3096562d661SIlya Dryomov WARN_ON(!list_empty(&req->r_req_lru_item)); 3106562d661SIlya Dryomov WARN_ON(!list_empty(&req->r_osd_item)); 3116562d661SIlya Dryomov WARN_ON(!list_empty(&req->r_linger_item)); 3126562d661SIlya Dryomov WARN_ON(!list_empty(&req->r_linger_osd_item)); 3136562d661SIlya Dryomov WARN_ON(req->r_osd); 3149e94af20SIlya Dryomov 3153d14c5d2SYehuda Sadeh if (req->r_request) 3163d14c5d2SYehuda Sadeh ceph_msg_put(req->r_request); 317ace6d3a9SAlex Elder if (req->r_reply) { 3188921d114SAlex Elder ceph_msg_revoke_incoming(req->r_reply); 319ab8cb34aSAlex Elder ceph_msg_put(req->r_reply); 320ace6d3a9SAlex Elder } 3210fff87ecSAlex Elder 3225476492fSAlex Elder for (which = 0; which < req->r_num_ops; which++) 3235476492fSAlex Elder osd_req_op_data_release(req, which); 3240fff87ecSAlex Elder 3253d14c5d2SYehuda Sadeh ceph_put_snap_context(req->r_snapc); 3263d14c5d2SYehuda Sadeh if (req->r_mempool) 3273d14c5d2SYehuda Sadeh mempool_free(req, req->r_osdc->req_mempool); 3283d14c5d2SYehuda Sadeh else 3295522ae0bSAlex Elder kmem_cache_free(ceph_osd_request_cache, req); 3305522ae0bSAlex Elder 3313d14c5d2SYehuda Sadeh } 3329e94af20SIlya Dryomov 3339e94af20SIlya Dryomov void ceph_osdc_get_request(struct ceph_osd_request *req) 3349e94af20SIlya Dryomov { 3359e94af20SIlya Dryomov dout("%s %p (was %d)\n", __func__, req, 3369e94af20SIlya Dryomov atomic_read(&req->r_kref.refcount)); 3379e94af20SIlya Dryomov kref_get(&req->r_kref); 3389e94af20SIlya Dryomov } 3399e94af20SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_get_request); 3409e94af20SIlya Dryomov 3419e94af20SIlya Dryomov void ceph_osdc_put_request(struct ceph_osd_request *req) 3429e94af20SIlya Dryomov { 3439e94af20SIlya Dryomov dout("%s %p (was %d)\n", __func__, req, 3449e94af20SIlya Dryomov atomic_read(&req->r_kref.refcount)); 3459e94af20SIlya Dryomov kref_put(&req->r_kref, ceph_osdc_release_request); 3469e94af20SIlya Dryomov } 3479e94af20SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_put_request); 3483d14c5d2SYehuda Sadeh 3493d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc, 3503d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 3511b83bef2SSage Weil unsigned int num_ops, 3523d14c5d2SYehuda Sadeh bool use_mempool, 35354a54007SAlex Elder gfp_t gfp_flags) 3543d14c5d2SYehuda Sadeh { 3553d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 3563d14c5d2SYehuda Sadeh struct ceph_msg *msg; 3571b83bef2SSage Weil size_t msg_size; 3583d14c5d2SYehuda Sadeh 35979528734SAlex Elder BUILD_BUG_ON(CEPH_OSD_MAX_OP > U16_MAX); 36079528734SAlex Elder BUG_ON(num_ops > CEPH_OSD_MAX_OP); 36179528734SAlex Elder 3621b83bef2SSage Weil msg_size = 4 + 4 + 8 + 8 + 4+8; 3631b83bef2SSage Weil msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */ 3641b83bef2SSage Weil msg_size += 1 + 8 + 4 + 4; /* pg_t */ 3652d0ebc5dSIlya Dryomov msg_size += 4 + CEPH_MAX_OID_NAME_LEN; /* oid */ 3661b83bef2SSage Weil msg_size += 2 + num_ops*sizeof(struct ceph_osd_op); 3671b83bef2SSage Weil msg_size += 8; /* snapid */ 3681b83bef2SSage Weil msg_size += 8; /* snap_seq */ 3691b83bef2SSage Weil msg_size += 8 * (snapc ? snapc->num_snaps : 0); /* snaps */ 3701b83bef2SSage Weil msg_size += 4; 3713d14c5d2SYehuda Sadeh 3723d14c5d2SYehuda Sadeh if (use_mempool) { 3733d14c5d2SYehuda Sadeh req = mempool_alloc(osdc->req_mempool, gfp_flags); 3743d14c5d2SYehuda Sadeh memset(req, 0, sizeof(*req)); 3753d14c5d2SYehuda Sadeh } else { 3765522ae0bSAlex Elder req = kmem_cache_zalloc(ceph_osd_request_cache, gfp_flags); 3773d14c5d2SYehuda Sadeh } 3783d14c5d2SYehuda Sadeh if (req == NULL) 3793d14c5d2SYehuda Sadeh return NULL; 3803d14c5d2SYehuda Sadeh 3813d14c5d2SYehuda Sadeh req->r_osdc = osdc; 3823d14c5d2SYehuda Sadeh req->r_mempool = use_mempool; 38379528734SAlex Elder req->r_num_ops = num_ops; 3843d14c5d2SYehuda Sadeh 3853d14c5d2SYehuda Sadeh kref_init(&req->r_kref); 3863d14c5d2SYehuda Sadeh init_completion(&req->r_completion); 3873d14c5d2SYehuda Sadeh init_completion(&req->r_safe_completion); 388a978fa20SAlex Elder RB_CLEAR_NODE(&req->r_node); 3893d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&req->r_unsafe_item); 390a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&req->r_linger_item); 3911d0326b1SIlya Dryomov INIT_LIST_HEAD(&req->r_linger_osd_item); 392935b639aSSage Weil INIT_LIST_HEAD(&req->r_req_lru_item); 393cd43045cSSage Weil INIT_LIST_HEAD(&req->r_osd_item); 394cd43045cSSage Weil 3953c972c95SIlya Dryomov req->r_base_oloc.pool = -1; 396205ee118SIlya Dryomov req->r_target_oloc.pool = -1; 39722116525SIlya Dryomov 3983d14c5d2SYehuda Sadeh /* create reply message */ 3993d14c5d2SYehuda Sadeh if (use_mempool) 4003d14c5d2SYehuda Sadeh msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0); 4013d14c5d2SYehuda Sadeh else 4023d14c5d2SYehuda Sadeh msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, 403b61c2763SSage Weil OSD_OPREPLY_FRONT_LEN, gfp_flags, true); 4043d14c5d2SYehuda Sadeh if (!msg) { 4053d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 4063d14c5d2SYehuda Sadeh return NULL; 4073d14c5d2SYehuda Sadeh } 4083d14c5d2SYehuda Sadeh req->r_reply = msg; 4093d14c5d2SYehuda Sadeh 4103d14c5d2SYehuda Sadeh /* create request message; allow space for oid */ 4113d14c5d2SYehuda Sadeh if (use_mempool) 4123d14c5d2SYehuda Sadeh msg = ceph_msgpool_get(&osdc->msgpool_op, 0); 4133d14c5d2SYehuda Sadeh else 414b61c2763SSage Weil msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true); 4153d14c5d2SYehuda Sadeh if (!msg) { 4163d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 4173d14c5d2SYehuda Sadeh return NULL; 4183d14c5d2SYehuda Sadeh } 4193d14c5d2SYehuda Sadeh 4203d14c5d2SYehuda Sadeh memset(msg->front.iov_base, 0, msg->front.iov_len); 4213d14c5d2SYehuda Sadeh 4223d14c5d2SYehuda Sadeh req->r_request = msg; 4233d14c5d2SYehuda Sadeh 4243d14c5d2SYehuda Sadeh return req; 4253d14c5d2SYehuda Sadeh } 4263d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_alloc_request); 4273d14c5d2SYehuda Sadeh 428a8dd0a37SAlex Elder static bool osd_req_opcode_valid(u16 opcode) 429a8dd0a37SAlex Elder { 430a8dd0a37SAlex Elder switch (opcode) { 431a8dd0a37SAlex Elder case CEPH_OSD_OP_READ: 432a8dd0a37SAlex Elder case CEPH_OSD_OP_STAT: 433a8dd0a37SAlex Elder case CEPH_OSD_OP_MAPEXT: 434a8dd0a37SAlex Elder case CEPH_OSD_OP_MASKTRUNC: 435a8dd0a37SAlex Elder case CEPH_OSD_OP_SPARSE_READ: 436a8dd0a37SAlex Elder case CEPH_OSD_OP_NOTIFY: 437a8dd0a37SAlex Elder case CEPH_OSD_OP_NOTIFY_ACK: 438a8dd0a37SAlex Elder case CEPH_OSD_OP_ASSERT_VER: 439a8dd0a37SAlex Elder case CEPH_OSD_OP_WRITE: 440a8dd0a37SAlex Elder case CEPH_OSD_OP_WRITEFULL: 441a8dd0a37SAlex Elder case CEPH_OSD_OP_TRUNCATE: 442a8dd0a37SAlex Elder case CEPH_OSD_OP_ZERO: 443a8dd0a37SAlex Elder case CEPH_OSD_OP_DELETE: 444a8dd0a37SAlex Elder case CEPH_OSD_OP_APPEND: 445a8dd0a37SAlex Elder case CEPH_OSD_OP_STARTSYNC: 446a8dd0a37SAlex Elder case CEPH_OSD_OP_SETTRUNC: 447a8dd0a37SAlex Elder case CEPH_OSD_OP_TRIMTRUNC: 448a8dd0a37SAlex Elder case CEPH_OSD_OP_TMAPUP: 449a8dd0a37SAlex Elder case CEPH_OSD_OP_TMAPPUT: 450a8dd0a37SAlex Elder case CEPH_OSD_OP_TMAPGET: 451a8dd0a37SAlex Elder case CEPH_OSD_OP_CREATE: 452a8dd0a37SAlex Elder case CEPH_OSD_OP_ROLLBACK: 453a8dd0a37SAlex Elder case CEPH_OSD_OP_WATCH: 454a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPGETKEYS: 455a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPGETVALS: 456a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPGETHEADER: 457a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPGETVALSBYKEYS: 458a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPSETVALS: 459a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPSETHEADER: 460a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPCLEAR: 461a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAPRMKEYS: 462a8dd0a37SAlex Elder case CEPH_OSD_OP_OMAP_CMP: 463c647b8a8SIlya Dryomov case CEPH_OSD_OP_SETALLOCHINT: 464a8dd0a37SAlex Elder case CEPH_OSD_OP_CLONERANGE: 465a8dd0a37SAlex Elder case CEPH_OSD_OP_ASSERT_SRC_VERSION: 466a8dd0a37SAlex Elder case CEPH_OSD_OP_SRC_CMPXATTR: 467a8dd0a37SAlex Elder case CEPH_OSD_OP_GETXATTR: 468a8dd0a37SAlex Elder case CEPH_OSD_OP_GETXATTRS: 469a8dd0a37SAlex Elder case CEPH_OSD_OP_CMPXATTR: 470a8dd0a37SAlex Elder case CEPH_OSD_OP_SETXATTR: 471a8dd0a37SAlex Elder case CEPH_OSD_OP_SETXATTRS: 472a8dd0a37SAlex Elder case CEPH_OSD_OP_RESETXATTRS: 473a8dd0a37SAlex Elder case CEPH_OSD_OP_RMXATTR: 474a8dd0a37SAlex Elder case CEPH_OSD_OP_PULL: 475a8dd0a37SAlex Elder case CEPH_OSD_OP_PUSH: 476a8dd0a37SAlex Elder case CEPH_OSD_OP_BALANCEREADS: 477a8dd0a37SAlex Elder case CEPH_OSD_OP_UNBALANCEREADS: 478a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB: 479a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB_RESERVE: 480a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB_UNRESERVE: 481a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB_STOP: 482a8dd0a37SAlex Elder case CEPH_OSD_OP_SCRUB_MAP: 483a8dd0a37SAlex Elder case CEPH_OSD_OP_WRLOCK: 484a8dd0a37SAlex Elder case CEPH_OSD_OP_WRUNLOCK: 485a8dd0a37SAlex Elder case CEPH_OSD_OP_RDLOCK: 486a8dd0a37SAlex Elder case CEPH_OSD_OP_RDUNLOCK: 487a8dd0a37SAlex Elder case CEPH_OSD_OP_UPLOCK: 488a8dd0a37SAlex Elder case CEPH_OSD_OP_DNLOCK: 489a8dd0a37SAlex Elder case CEPH_OSD_OP_CALL: 490a8dd0a37SAlex Elder case CEPH_OSD_OP_PGLS: 491a8dd0a37SAlex Elder case CEPH_OSD_OP_PGLS_FILTER: 492a8dd0a37SAlex Elder return true; 493a8dd0a37SAlex Elder default: 494a8dd0a37SAlex Elder return false; 495a8dd0a37SAlex Elder } 496a8dd0a37SAlex Elder } 497a8dd0a37SAlex Elder 49833803f33SAlex Elder /* 49933803f33SAlex Elder * This is an osd op init function for opcodes that have no data or 50033803f33SAlex Elder * other information associated with them. It also serves as a 50133803f33SAlex Elder * common init routine for all the other init functions, below. 50233803f33SAlex Elder */ 503c99d2d4aSAlex Elder static struct ceph_osd_req_op * 50449719778SAlex Elder _osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which, 505c99d2d4aSAlex Elder u16 opcode) 50633803f33SAlex Elder { 507c99d2d4aSAlex Elder struct ceph_osd_req_op *op; 508c99d2d4aSAlex Elder 509c99d2d4aSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 51033803f33SAlex Elder BUG_ON(!osd_req_opcode_valid(opcode)); 51133803f33SAlex Elder 512c99d2d4aSAlex Elder op = &osd_req->r_ops[which]; 51333803f33SAlex Elder memset(op, 0, sizeof (*op)); 51433803f33SAlex Elder op->op = opcode; 515c99d2d4aSAlex Elder 516c99d2d4aSAlex Elder return op; 51733803f33SAlex Elder } 51833803f33SAlex Elder 51949719778SAlex Elder void osd_req_op_init(struct ceph_osd_request *osd_req, 52049719778SAlex Elder unsigned int which, u16 opcode) 52149719778SAlex Elder { 52249719778SAlex Elder (void)_osd_req_op_init(osd_req, which, opcode); 52349719778SAlex Elder } 52449719778SAlex Elder EXPORT_SYMBOL(osd_req_op_init); 52549719778SAlex Elder 526c99d2d4aSAlex Elder void osd_req_op_extent_init(struct ceph_osd_request *osd_req, 527c99d2d4aSAlex Elder unsigned int which, u16 opcode, 52833803f33SAlex Elder u64 offset, u64 length, 52933803f33SAlex Elder u64 truncate_size, u32 truncate_seq) 53033803f33SAlex Elder { 53149719778SAlex Elder struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode); 53233803f33SAlex Elder size_t payload_len = 0; 53333803f33SAlex Elder 534ad7a60deSLi Wang BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE && 535ad7a60deSLi Wang opcode != CEPH_OSD_OP_DELETE && opcode != CEPH_OSD_OP_ZERO && 536ad7a60deSLi Wang opcode != CEPH_OSD_OP_TRUNCATE); 53733803f33SAlex Elder 53833803f33SAlex Elder op->extent.offset = offset; 53933803f33SAlex Elder op->extent.length = length; 54033803f33SAlex Elder op->extent.truncate_size = truncate_size; 54133803f33SAlex Elder op->extent.truncate_seq = truncate_seq; 54233803f33SAlex Elder if (opcode == CEPH_OSD_OP_WRITE) 54333803f33SAlex Elder payload_len += length; 54433803f33SAlex Elder 54533803f33SAlex Elder op->payload_len = payload_len; 54633803f33SAlex Elder } 54733803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_init); 54833803f33SAlex Elder 549c99d2d4aSAlex Elder void osd_req_op_extent_update(struct ceph_osd_request *osd_req, 550c99d2d4aSAlex Elder unsigned int which, u64 length) 551e5975c7cSAlex Elder { 552c99d2d4aSAlex Elder struct ceph_osd_req_op *op; 553c99d2d4aSAlex Elder u64 previous; 554c99d2d4aSAlex Elder 555c99d2d4aSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 556c99d2d4aSAlex Elder op = &osd_req->r_ops[which]; 557c99d2d4aSAlex Elder previous = op->extent.length; 558e5975c7cSAlex Elder 559e5975c7cSAlex Elder if (length == previous) 560e5975c7cSAlex Elder return; /* Nothing to do */ 561e5975c7cSAlex Elder BUG_ON(length > previous); 562e5975c7cSAlex Elder 563e5975c7cSAlex Elder op->extent.length = length; 564e5975c7cSAlex Elder op->payload_len -= previous - length; 565e5975c7cSAlex Elder } 566e5975c7cSAlex Elder EXPORT_SYMBOL(osd_req_op_extent_update); 567e5975c7cSAlex Elder 568c99d2d4aSAlex Elder void osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which, 56904017e29SAlex Elder u16 opcode, const char *class, const char *method) 57033803f33SAlex Elder { 57149719778SAlex Elder struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode); 5725f562df5SAlex Elder struct ceph_pagelist *pagelist; 57333803f33SAlex Elder size_t payload_len = 0; 57433803f33SAlex Elder size_t size; 57533803f33SAlex Elder 57633803f33SAlex Elder BUG_ON(opcode != CEPH_OSD_OP_CALL); 57733803f33SAlex Elder 5785f562df5SAlex Elder pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS); 5795f562df5SAlex Elder BUG_ON(!pagelist); 5805f562df5SAlex Elder ceph_pagelist_init(pagelist); 5815f562df5SAlex Elder 58233803f33SAlex Elder op->cls.class_name = class; 58333803f33SAlex Elder size = strlen(class); 58433803f33SAlex Elder BUG_ON(size > (size_t) U8_MAX); 58533803f33SAlex Elder op->cls.class_len = size; 5865f562df5SAlex Elder ceph_pagelist_append(pagelist, class, size); 58733803f33SAlex Elder payload_len += size; 58833803f33SAlex Elder 58933803f33SAlex Elder op->cls.method_name = method; 59033803f33SAlex Elder size = strlen(method); 59133803f33SAlex Elder BUG_ON(size > (size_t) U8_MAX); 59233803f33SAlex Elder op->cls.method_len = size; 5935f562df5SAlex Elder ceph_pagelist_append(pagelist, method, size); 59433803f33SAlex Elder payload_len += size; 59533803f33SAlex Elder 596a4ce40a9SAlex Elder osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist); 5975f562df5SAlex Elder 59833803f33SAlex Elder op->cls.argc = 0; /* currently unused */ 59933803f33SAlex Elder 60033803f33SAlex Elder op->payload_len = payload_len; 60133803f33SAlex Elder } 60233803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_init); 6038c042b0dSAlex Elder 604c99d2d4aSAlex Elder void osd_req_op_watch_init(struct ceph_osd_request *osd_req, 605c99d2d4aSAlex Elder unsigned int which, u16 opcode, 60633803f33SAlex Elder u64 cookie, u64 version, int flag) 60733803f33SAlex Elder { 60849719778SAlex Elder struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode); 60933803f33SAlex Elder 610c99d2d4aSAlex Elder BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH); 61133803f33SAlex Elder 61233803f33SAlex Elder op->watch.cookie = cookie; 6139ef1ee5aSAlex Elder op->watch.ver = version; 61433803f33SAlex Elder if (opcode == CEPH_OSD_OP_WATCH && flag) 61533803f33SAlex Elder op->watch.flag = (u8)1; 61633803f33SAlex Elder } 61733803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_watch_init); 61833803f33SAlex Elder 619c647b8a8SIlya Dryomov void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req, 620c647b8a8SIlya Dryomov unsigned int which, 621c647b8a8SIlya Dryomov u64 expected_object_size, 622c647b8a8SIlya Dryomov u64 expected_write_size) 623c647b8a8SIlya Dryomov { 624c647b8a8SIlya Dryomov struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, 625c647b8a8SIlya Dryomov CEPH_OSD_OP_SETALLOCHINT); 626c647b8a8SIlya Dryomov 627c647b8a8SIlya Dryomov op->alloc_hint.expected_object_size = expected_object_size; 628c647b8a8SIlya Dryomov op->alloc_hint.expected_write_size = expected_write_size; 629c647b8a8SIlya Dryomov 630c647b8a8SIlya Dryomov /* 631c647b8a8SIlya Dryomov * CEPH_OSD_OP_SETALLOCHINT op is advisory and therefore deemed 632c647b8a8SIlya Dryomov * not worth a feature bit. Set FAILOK per-op flag to make 633c647b8a8SIlya Dryomov * sure older osds don't trip over an unsupported opcode. 634c647b8a8SIlya Dryomov */ 635c647b8a8SIlya Dryomov op->flags |= CEPH_OSD_OP_FLAG_FAILOK; 636c647b8a8SIlya Dryomov } 637c647b8a8SIlya Dryomov EXPORT_SYMBOL(osd_req_op_alloc_hint_init); 638c647b8a8SIlya Dryomov 63990af3602SAlex Elder static void ceph_osdc_msg_data_add(struct ceph_msg *msg, 640ec9123c5SAlex Elder struct ceph_osd_data *osd_data) 641ec9123c5SAlex Elder { 642ec9123c5SAlex Elder u64 length = ceph_osd_data_length(osd_data); 643ec9123c5SAlex Elder 644ec9123c5SAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 645ec9123c5SAlex Elder BUG_ON(length > (u64) SIZE_MAX); 646ec9123c5SAlex Elder if (length) 64790af3602SAlex Elder ceph_msg_data_add_pages(msg, osd_data->pages, 648ec9123c5SAlex Elder length, osd_data->alignment); 649ec9123c5SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) { 650ec9123c5SAlex Elder BUG_ON(!length); 65190af3602SAlex Elder ceph_msg_data_add_pagelist(msg, osd_data->pagelist); 652ec9123c5SAlex Elder #ifdef CONFIG_BLOCK 653ec9123c5SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) { 65490af3602SAlex Elder ceph_msg_data_add_bio(msg, osd_data->bio, length); 655ec9123c5SAlex Elder #endif 656ec9123c5SAlex Elder } else { 657ec9123c5SAlex Elder BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE); 658ec9123c5SAlex Elder } 659ec9123c5SAlex Elder } 660ec9123c5SAlex Elder 661175face2SAlex Elder static u64 osd_req_encode_op(struct ceph_osd_request *req, 66279528734SAlex Elder struct ceph_osd_op *dst, unsigned int which) 6633d14c5d2SYehuda Sadeh { 66479528734SAlex Elder struct ceph_osd_req_op *src; 66504017e29SAlex Elder struct ceph_osd_data *osd_data; 66654d50649SAlex Elder u64 request_data_len = 0; 66704017e29SAlex Elder u64 data_length; 668175face2SAlex Elder 66979528734SAlex Elder BUG_ON(which >= req->r_num_ops); 67079528734SAlex Elder src = &req->r_ops[which]; 671a8dd0a37SAlex Elder if (WARN_ON(!osd_req_opcode_valid(src->op))) { 672a8dd0a37SAlex Elder pr_err("unrecognized osd opcode %d\n", src->op); 673a8dd0a37SAlex Elder 674a8dd0a37SAlex Elder return 0; 675a8dd0a37SAlex Elder } 6763d14c5d2SYehuda Sadeh 677065a68f9SAlex Elder switch (src->op) { 678fbfab539SAlex Elder case CEPH_OSD_OP_STAT: 67949719778SAlex Elder osd_data = &src->raw_data_in; 68049719778SAlex Elder ceph_osdc_msg_data_add(req->r_reply, osd_data); 681fbfab539SAlex Elder break; 6823d14c5d2SYehuda Sadeh case CEPH_OSD_OP_READ: 6833d14c5d2SYehuda Sadeh case CEPH_OSD_OP_WRITE: 684ad7a60deSLi Wang case CEPH_OSD_OP_ZERO: 685ad7a60deSLi Wang case CEPH_OSD_OP_DELETE: 686ad7a60deSLi Wang case CEPH_OSD_OP_TRUNCATE: 687175face2SAlex Elder if (src->op == CEPH_OSD_OP_WRITE) 68854d50649SAlex Elder request_data_len = src->extent.length; 689175face2SAlex Elder dst->extent.offset = cpu_to_le64(src->extent.offset); 690175face2SAlex Elder dst->extent.length = cpu_to_le64(src->extent.length); 6913d14c5d2SYehuda Sadeh dst->extent.truncate_size = 6923d14c5d2SYehuda Sadeh cpu_to_le64(src->extent.truncate_size); 6933d14c5d2SYehuda Sadeh dst->extent.truncate_seq = 6943d14c5d2SYehuda Sadeh cpu_to_le32(src->extent.truncate_seq); 69504017e29SAlex Elder osd_data = &src->extent.osd_data; 6965476492fSAlex Elder if (src->op == CEPH_OSD_OP_WRITE) 69704017e29SAlex Elder ceph_osdc_msg_data_add(req->r_request, osd_data); 6985476492fSAlex Elder else 69904017e29SAlex Elder ceph_osdc_msg_data_add(req->r_reply, osd_data); 7003d14c5d2SYehuda Sadeh break; 7013d14c5d2SYehuda Sadeh case CEPH_OSD_OP_CALL: 7023d14c5d2SYehuda Sadeh dst->cls.class_len = src->cls.class_len; 7033d14c5d2SYehuda Sadeh dst->cls.method_len = src->cls.method_len; 70404017e29SAlex Elder osd_data = &src->cls.request_info; 70504017e29SAlex Elder ceph_osdc_msg_data_add(req->r_request, osd_data); 70604017e29SAlex Elder BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGELIST); 70704017e29SAlex Elder request_data_len = osd_data->pagelist->length; 70804017e29SAlex Elder 70904017e29SAlex Elder osd_data = &src->cls.request_data; 71004017e29SAlex Elder data_length = ceph_osd_data_length(osd_data); 71104017e29SAlex Elder if (data_length) { 71204017e29SAlex Elder BUG_ON(osd_data->type == CEPH_OSD_DATA_TYPE_NONE); 71304017e29SAlex Elder dst->cls.indata_len = cpu_to_le32(data_length); 71404017e29SAlex Elder ceph_osdc_msg_data_add(req->r_request, osd_data); 71504017e29SAlex Elder src->payload_len += data_length; 71604017e29SAlex Elder request_data_len += data_length; 71704017e29SAlex Elder } 71804017e29SAlex Elder osd_data = &src->cls.response_data; 71904017e29SAlex Elder ceph_osdc_msg_data_add(req->r_reply, osd_data); 7203d14c5d2SYehuda Sadeh break; 7213d14c5d2SYehuda Sadeh case CEPH_OSD_OP_STARTSYNC: 7223d14c5d2SYehuda Sadeh break; 723a40c4f10SYehuda Sadeh case CEPH_OSD_OP_NOTIFY_ACK: 724a40c4f10SYehuda Sadeh case CEPH_OSD_OP_WATCH: 725a40c4f10SYehuda Sadeh dst->watch.cookie = cpu_to_le64(src->watch.cookie); 726a40c4f10SYehuda Sadeh dst->watch.ver = cpu_to_le64(src->watch.ver); 727a40c4f10SYehuda Sadeh dst->watch.flag = src->watch.flag; 728a40c4f10SYehuda Sadeh break; 729c647b8a8SIlya Dryomov case CEPH_OSD_OP_SETALLOCHINT: 730c647b8a8SIlya Dryomov dst->alloc_hint.expected_object_size = 731c647b8a8SIlya Dryomov cpu_to_le64(src->alloc_hint.expected_object_size); 732c647b8a8SIlya Dryomov dst->alloc_hint.expected_write_size = 733c647b8a8SIlya Dryomov cpu_to_le64(src->alloc_hint.expected_write_size); 734c647b8a8SIlya Dryomov break; 7353d14c5d2SYehuda Sadeh default: 7364c46459cSAlex Elder pr_err("unsupported osd opcode %s\n", 7378f63ca2dSAlex Elder ceph_osd_op_name(src->op)); 7384c46459cSAlex Elder WARN_ON(1); 739a8dd0a37SAlex Elder 740a8dd0a37SAlex Elder return 0; 7413d14c5d2SYehuda Sadeh } 7427b25bf5fSIlya Dryomov 743a8dd0a37SAlex Elder dst->op = cpu_to_le16(src->op); 7447b25bf5fSIlya Dryomov dst->flags = cpu_to_le32(src->flags); 7453d14c5d2SYehuda Sadeh dst->payload_len = cpu_to_le32(src->payload_len); 746175face2SAlex Elder 74754d50649SAlex Elder return request_data_len; 7483d14c5d2SYehuda Sadeh } 7493d14c5d2SYehuda Sadeh 7503d14c5d2SYehuda Sadeh /* 7513d14c5d2SYehuda Sadeh * build new request AND message, calculate layout, and adjust file 7523d14c5d2SYehuda Sadeh * extent as needed. 7533d14c5d2SYehuda Sadeh * 7543d14c5d2SYehuda Sadeh * if the file was recently truncated, we include information about its 7553d14c5d2SYehuda Sadeh * old and new size so that the object can be updated appropriately. (we 7563d14c5d2SYehuda Sadeh * avoid synchronously deleting truncated objects because it's slow.) 7573d14c5d2SYehuda Sadeh * 7583d14c5d2SYehuda Sadeh * if @do_sync, include a 'startsync' command so that the osd will flush 7593d14c5d2SYehuda Sadeh * data quickly. 7603d14c5d2SYehuda Sadeh */ 7613d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc, 7623d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 7633d14c5d2SYehuda Sadeh struct ceph_vino vino, 764acead002SAlex Elder u64 off, u64 *plen, int num_ops, 7653d14c5d2SYehuda Sadeh int opcode, int flags, 7663d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 7673d14c5d2SYehuda Sadeh u32 truncate_seq, 7683d14c5d2SYehuda Sadeh u64 truncate_size, 769153e5167SAlex Elder bool use_mempool) 7703d14c5d2SYehuda Sadeh { 7713d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 77275d1c941SAlex Elder u64 objnum = 0; 77375d1c941SAlex Elder u64 objoff = 0; 77475d1c941SAlex Elder u64 objlen = 0; 775d18d1e28SAlex Elder u32 object_size; 776d18d1e28SAlex Elder u64 object_base; 7776816282dSSage Weil int r; 7783d14c5d2SYehuda Sadeh 779ad7a60deSLi Wang BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE && 780ad7a60deSLi Wang opcode != CEPH_OSD_OP_DELETE && opcode != CEPH_OSD_OP_ZERO && 781ad7a60deSLi Wang opcode != CEPH_OSD_OP_TRUNCATE); 7823d14c5d2SYehuda Sadeh 783acead002SAlex Elder req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool, 784ae7ca4a3SAlex Elder GFP_NOFS); 7854ad12621SSage Weil if (!req) 7866816282dSSage Weil return ERR_PTR(-ENOMEM); 78779528734SAlex Elder 788d178a9e7SAlex Elder req->r_flags = flags; 7893d14c5d2SYehuda Sadeh 7903d14c5d2SYehuda Sadeh /* calculate max write size */ 791a19dadfbSAlex Elder r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen); 7923ff5f385SAlex Elder if (r < 0) { 7933ff5f385SAlex Elder ceph_osdc_put_request(req); 7946816282dSSage Weil return ERR_PTR(r); 7953ff5f385SAlex Elder } 796a19dadfbSAlex Elder 797d18d1e28SAlex Elder object_size = le32_to_cpu(layout->fl_object_size); 798d18d1e28SAlex Elder object_base = off - objoff; 799ccca4e37SYan, Zheng if (!(truncate_seq == 1 && truncate_size == -1ULL)) { 800d18d1e28SAlex Elder if (truncate_size <= object_base) { 801d18d1e28SAlex Elder truncate_size = 0; 802d18d1e28SAlex Elder } else { 803d18d1e28SAlex Elder truncate_size -= object_base; 804d18d1e28SAlex Elder if (truncate_size > object_size) 805d18d1e28SAlex Elder truncate_size = object_size; 806d18d1e28SAlex Elder } 807ccca4e37SYan, Zheng } 808d18d1e28SAlex Elder 809c99d2d4aSAlex Elder osd_req_op_extent_init(req, 0, opcode, objoff, objlen, 810b0270324SAlex Elder truncate_size, truncate_seq); 8118c042b0dSAlex Elder 812acead002SAlex Elder /* 813acead002SAlex Elder * A second op in the ops array means the caller wants to 814acead002SAlex Elder * also issue a include a 'startsync' command so that the 815acead002SAlex Elder * osd will flush data quickly. 816acead002SAlex Elder */ 817acead002SAlex Elder if (num_ops > 1) 818c99d2d4aSAlex Elder osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC); 819d18d1e28SAlex Elder 8203c972c95SIlya Dryomov req->r_base_oloc.pool = ceph_file_layout_pg_pool(*layout); 8213d14c5d2SYehuda Sadeh 8223c972c95SIlya Dryomov snprintf(req->r_base_oid.name, sizeof(req->r_base_oid.name), 8234295f221SIlya Dryomov "%llx.%08llx", vino.ino, objnum); 8243c972c95SIlya Dryomov req->r_base_oid.name_len = strlen(req->r_base_oid.name); 825dbe0fc41SAlex Elder 8263d14c5d2SYehuda Sadeh return req; 8273d14c5d2SYehuda Sadeh } 8283d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_new_request); 8293d14c5d2SYehuda Sadeh 8303d14c5d2SYehuda Sadeh /* 8313d14c5d2SYehuda Sadeh * We keep osd requests in an rbtree, sorted by ->r_tid. 8323d14c5d2SYehuda Sadeh */ 8333d14c5d2SYehuda Sadeh static void __insert_request(struct ceph_osd_client *osdc, 8343d14c5d2SYehuda Sadeh struct ceph_osd_request *new) 8353d14c5d2SYehuda Sadeh { 8363d14c5d2SYehuda Sadeh struct rb_node **p = &osdc->requests.rb_node; 8373d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 8383d14c5d2SYehuda Sadeh struct ceph_osd_request *req = NULL; 8393d14c5d2SYehuda Sadeh 8403d14c5d2SYehuda Sadeh while (*p) { 8413d14c5d2SYehuda Sadeh parent = *p; 8423d14c5d2SYehuda Sadeh req = rb_entry(parent, struct ceph_osd_request, r_node); 8433d14c5d2SYehuda Sadeh if (new->r_tid < req->r_tid) 8443d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 8453d14c5d2SYehuda Sadeh else if (new->r_tid > req->r_tid) 8463d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 8473d14c5d2SYehuda Sadeh else 8483d14c5d2SYehuda Sadeh BUG(); 8493d14c5d2SYehuda Sadeh } 8503d14c5d2SYehuda Sadeh 8513d14c5d2SYehuda Sadeh rb_link_node(&new->r_node, parent, p); 8523d14c5d2SYehuda Sadeh rb_insert_color(&new->r_node, &osdc->requests); 8533d14c5d2SYehuda Sadeh } 8543d14c5d2SYehuda Sadeh 8553d14c5d2SYehuda Sadeh static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc, 8563d14c5d2SYehuda Sadeh u64 tid) 8573d14c5d2SYehuda Sadeh { 8583d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 8593d14c5d2SYehuda Sadeh struct rb_node *n = osdc->requests.rb_node; 8603d14c5d2SYehuda Sadeh 8613d14c5d2SYehuda Sadeh while (n) { 8623d14c5d2SYehuda Sadeh req = rb_entry(n, struct ceph_osd_request, r_node); 8633d14c5d2SYehuda Sadeh if (tid < req->r_tid) 8643d14c5d2SYehuda Sadeh n = n->rb_left; 8653d14c5d2SYehuda Sadeh else if (tid > req->r_tid) 8663d14c5d2SYehuda Sadeh n = n->rb_right; 8673d14c5d2SYehuda Sadeh else 8683d14c5d2SYehuda Sadeh return req; 8693d14c5d2SYehuda Sadeh } 8703d14c5d2SYehuda Sadeh return NULL; 8713d14c5d2SYehuda Sadeh } 8723d14c5d2SYehuda Sadeh 8733d14c5d2SYehuda Sadeh static struct ceph_osd_request * 8743d14c5d2SYehuda Sadeh __lookup_request_ge(struct ceph_osd_client *osdc, 8753d14c5d2SYehuda Sadeh u64 tid) 8763d14c5d2SYehuda Sadeh { 8773d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 8783d14c5d2SYehuda Sadeh struct rb_node *n = osdc->requests.rb_node; 8793d14c5d2SYehuda Sadeh 8803d14c5d2SYehuda Sadeh while (n) { 8813d14c5d2SYehuda Sadeh req = rb_entry(n, struct ceph_osd_request, r_node); 8823d14c5d2SYehuda Sadeh if (tid < req->r_tid) { 8833d14c5d2SYehuda Sadeh if (!n->rb_left) 8843d14c5d2SYehuda Sadeh return req; 8853d14c5d2SYehuda Sadeh n = n->rb_left; 8863d14c5d2SYehuda Sadeh } else if (tid > req->r_tid) { 8873d14c5d2SYehuda Sadeh n = n->rb_right; 8883d14c5d2SYehuda Sadeh } else { 8893d14c5d2SYehuda Sadeh return req; 8903d14c5d2SYehuda Sadeh } 8913d14c5d2SYehuda Sadeh } 8923d14c5d2SYehuda Sadeh return NULL; 8933d14c5d2SYehuda Sadeh } 8943d14c5d2SYehuda Sadeh 8956f6c7006SSage Weil /* 8966f6c7006SSage Weil * Resubmit requests pending on the given osd. 8976f6c7006SSage Weil */ 8986f6c7006SSage Weil static void __kick_osd_requests(struct ceph_osd_client *osdc, 8996f6c7006SSage Weil struct ceph_osd *osd) 9006f6c7006SSage Weil { 901a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 902e02493c0SAlex Elder LIST_HEAD(resend); 9036f6c7006SSage Weil int err; 9046f6c7006SSage Weil 9056f6c7006SSage Weil dout("__kick_osd_requests osd%d\n", osd->o_osd); 9066f6c7006SSage Weil err = __reset_osd(osdc, osd); 907685a7555SAlex Elder if (err) 9086f6c7006SSage Weil return; 909e02493c0SAlex Elder /* 910e02493c0SAlex Elder * Build up a list of requests to resend by traversing the 911e02493c0SAlex Elder * osd's list of requests. Requests for a given object are 912e02493c0SAlex Elder * sent in tid order, and that is also the order they're 913e02493c0SAlex Elder * kept on this list. Therefore all requests that are in 914e02493c0SAlex Elder * flight will be found first, followed by all requests that 915e02493c0SAlex Elder * have not yet been sent. And to resend requests while 916e02493c0SAlex Elder * preserving this order we will want to put any sent 917e02493c0SAlex Elder * requests back on the front of the osd client's unsent 918e02493c0SAlex Elder * list. 919e02493c0SAlex Elder * 920e02493c0SAlex Elder * So we build a separate ordered list of already-sent 921e02493c0SAlex Elder * requests for the affected osd and splice it onto the 922e02493c0SAlex Elder * front of the osd client's unsent list. Once we've seen a 923e02493c0SAlex Elder * request that has not yet been sent we're done. Those 924e02493c0SAlex Elder * requests are already sitting right where they belong. 925e02493c0SAlex Elder */ 9266f6c7006SSage Weil list_for_each_entry(req, &osd->o_requests, r_osd_item) { 927e02493c0SAlex Elder if (!req->r_sent) 928e02493c0SAlex Elder break; 929e02493c0SAlex Elder list_move_tail(&req->r_req_lru_item, &resend); 930e02493c0SAlex Elder dout("requeueing %p tid %llu osd%d\n", req, req->r_tid, 9316f6c7006SSage Weil osd->o_osd); 932a40c4f10SYehuda Sadeh if (!req->r_linger) 9336f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 9346f6c7006SSage Weil } 935e02493c0SAlex Elder list_splice(&resend, &osdc->req_unsent); 936a40c4f10SYehuda Sadeh 937e02493c0SAlex Elder /* 938e02493c0SAlex Elder * Linger requests are re-registered before sending, which 939e02493c0SAlex Elder * sets up a new tid for each. We add them to the unsent 940e02493c0SAlex Elder * list at the end to keep things in tid order. 941e02493c0SAlex Elder */ 942a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osd->o_linger_requests, 9431d0326b1SIlya Dryomov r_linger_osd_item) { 94477f38e0eSSage Weil /* 94577f38e0eSSage Weil * reregister request prior to unregistering linger so 94677f38e0eSSage Weil * that r_osd is preserved. 94777f38e0eSSage Weil */ 94877f38e0eSSage Weil BUG_ON(!list_empty(&req->r_req_lru_item)); 949a40c4f10SYehuda Sadeh __register_request(osdc, req); 950e02493c0SAlex Elder list_add_tail(&req->r_req_lru_item, &osdc->req_unsent); 951ad885927SAlex Elder list_add_tail(&req->r_osd_item, &req->r_osd->o_requests); 95277f38e0eSSage Weil __unregister_linger_request(osdc, req); 953a40c4f10SYehuda Sadeh dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid, 954a40c4f10SYehuda Sadeh osd->o_osd); 955a40c4f10SYehuda Sadeh } 9566f6c7006SSage Weil } 9576f6c7006SSage Weil 9583d14c5d2SYehuda Sadeh /* 9593d14c5d2SYehuda Sadeh * If the osd connection drops, we need to resubmit all requests. 9603d14c5d2SYehuda Sadeh */ 9613d14c5d2SYehuda Sadeh static void osd_reset(struct ceph_connection *con) 9623d14c5d2SYehuda Sadeh { 9633d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 9643d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 9653d14c5d2SYehuda Sadeh 9663d14c5d2SYehuda Sadeh if (!osd) 9673d14c5d2SYehuda Sadeh return; 9683d14c5d2SYehuda Sadeh dout("osd_reset osd%d\n", osd->o_osd); 9693d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 9703d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 97183aff95eSSage Weil mutex_lock(&osdc->request_mutex); 97283aff95eSSage Weil __kick_osd_requests(osdc, osd); 973f9d25199SAlex Elder __send_queued(osdc); 97483aff95eSSage Weil mutex_unlock(&osdc->request_mutex); 9753d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 9763d14c5d2SYehuda Sadeh } 9773d14c5d2SYehuda Sadeh 9783d14c5d2SYehuda Sadeh /* 9793d14c5d2SYehuda Sadeh * Track open sessions with osds. 9803d14c5d2SYehuda Sadeh */ 981e10006f8SAlex Elder static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum) 9823d14c5d2SYehuda Sadeh { 9833d14c5d2SYehuda Sadeh struct ceph_osd *osd; 9843d14c5d2SYehuda Sadeh 9853d14c5d2SYehuda Sadeh osd = kzalloc(sizeof(*osd), GFP_NOFS); 9863d14c5d2SYehuda Sadeh if (!osd) 9873d14c5d2SYehuda Sadeh return NULL; 9883d14c5d2SYehuda Sadeh 9893d14c5d2SYehuda Sadeh atomic_set(&osd->o_ref, 1); 9903d14c5d2SYehuda Sadeh osd->o_osdc = osdc; 991e10006f8SAlex Elder osd->o_osd = onum; 992f407731dSAlex Elder RB_CLEAR_NODE(&osd->o_node); 9933d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_requests); 994a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osd->o_linger_requests); 9953d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_osd_lru); 9963d14c5d2SYehuda Sadeh osd->o_incarnation = 1; 9973d14c5d2SYehuda Sadeh 998b7a9e5ddSSage Weil ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr); 9993d14c5d2SYehuda Sadeh 10003d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_keepalive_item); 10013d14c5d2SYehuda Sadeh return osd; 10023d14c5d2SYehuda Sadeh } 10033d14c5d2SYehuda Sadeh 10043d14c5d2SYehuda Sadeh static struct ceph_osd *get_osd(struct ceph_osd *osd) 10053d14c5d2SYehuda Sadeh { 10063d14c5d2SYehuda Sadeh if (atomic_inc_not_zero(&osd->o_ref)) { 10073d14c5d2SYehuda Sadeh dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1, 10083d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref)); 10093d14c5d2SYehuda Sadeh return osd; 10103d14c5d2SYehuda Sadeh } else { 10113d14c5d2SYehuda Sadeh dout("get_osd %p FAIL\n", osd); 10123d14c5d2SYehuda Sadeh return NULL; 10133d14c5d2SYehuda Sadeh } 10143d14c5d2SYehuda Sadeh } 10153d14c5d2SYehuda Sadeh 10163d14c5d2SYehuda Sadeh static void put_osd(struct ceph_osd *osd) 10173d14c5d2SYehuda Sadeh { 10183d14c5d2SYehuda Sadeh dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref), 10193d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref) - 1); 1020a255651dSAlex Elder if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) { 10213d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth; 10223d14c5d2SYehuda Sadeh 102327859f97SSage Weil ceph_auth_destroy_authorizer(ac, osd->o_auth.authorizer); 10243d14c5d2SYehuda Sadeh kfree(osd); 10253d14c5d2SYehuda Sadeh } 10263d14c5d2SYehuda Sadeh } 10273d14c5d2SYehuda Sadeh 10283d14c5d2SYehuda Sadeh /* 10293d14c5d2SYehuda Sadeh * remove an osd from our map 10303d14c5d2SYehuda Sadeh */ 10313d14c5d2SYehuda Sadeh static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 10323d14c5d2SYehuda Sadeh { 10333d14c5d2SYehuda Sadeh dout("__remove_osd %p\n", osd); 10343d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_requests)); 10357c6e6fc5SIlya Dryomov BUG_ON(!list_empty(&osd->o_linger_requests)); 10367c6e6fc5SIlya Dryomov 10373d14c5d2SYehuda Sadeh rb_erase(&osd->o_node, &osdc->osds); 10383d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 10393d14c5d2SYehuda Sadeh ceph_con_close(&osd->o_con); 10403d14c5d2SYehuda Sadeh put_osd(osd); 10413d14c5d2SYehuda Sadeh } 10423d14c5d2SYehuda Sadeh 1043aca420bcSSage Weil static void remove_all_osds(struct ceph_osd_client *osdc) 1044aca420bcSSage Weil { 1045048a9d2dSJiaju Zhang dout("%s %p\n", __func__, osdc); 1046aca420bcSSage Weil mutex_lock(&osdc->request_mutex); 1047aca420bcSSage Weil while (!RB_EMPTY_ROOT(&osdc->osds)) { 1048aca420bcSSage Weil struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds), 1049aca420bcSSage Weil struct ceph_osd, o_node); 1050aca420bcSSage Weil __remove_osd(osdc, osd); 1051aca420bcSSage Weil } 1052aca420bcSSage Weil mutex_unlock(&osdc->request_mutex); 1053aca420bcSSage Weil } 1054aca420bcSSage Weil 10553d14c5d2SYehuda Sadeh static void __move_osd_to_lru(struct ceph_osd_client *osdc, 10563d14c5d2SYehuda Sadeh struct ceph_osd *osd) 10573d14c5d2SYehuda Sadeh { 1058bbf37ec3SIlya Dryomov dout("%s %p\n", __func__, osd); 10593d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_osd_lru)); 1060bbf37ec3SIlya Dryomov 10613d14c5d2SYehuda Sadeh list_add_tail(&osd->o_osd_lru, &osdc->osd_lru); 10623d14c5d2SYehuda Sadeh osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ; 10633d14c5d2SYehuda Sadeh } 10643d14c5d2SYehuda Sadeh 1065bbf37ec3SIlya Dryomov static void maybe_move_osd_to_lru(struct ceph_osd_client *osdc, 1066bbf37ec3SIlya Dryomov struct ceph_osd *osd) 1067bbf37ec3SIlya Dryomov { 1068bbf37ec3SIlya Dryomov dout("%s %p\n", __func__, osd); 1069bbf37ec3SIlya Dryomov 1070bbf37ec3SIlya Dryomov if (list_empty(&osd->o_requests) && 1071bbf37ec3SIlya Dryomov list_empty(&osd->o_linger_requests)) 1072bbf37ec3SIlya Dryomov __move_osd_to_lru(osdc, osd); 1073bbf37ec3SIlya Dryomov } 1074bbf37ec3SIlya Dryomov 10753d14c5d2SYehuda Sadeh static void __remove_osd_from_lru(struct ceph_osd *osd) 10763d14c5d2SYehuda Sadeh { 10773d14c5d2SYehuda Sadeh dout("__remove_osd_from_lru %p\n", osd); 10783d14c5d2SYehuda Sadeh if (!list_empty(&osd->o_osd_lru)) 10793d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 10803d14c5d2SYehuda Sadeh } 10813d14c5d2SYehuda Sadeh 1082aca420bcSSage Weil static void remove_old_osds(struct ceph_osd_client *osdc) 10833d14c5d2SYehuda Sadeh { 10843d14c5d2SYehuda Sadeh struct ceph_osd *osd, *nosd; 10853d14c5d2SYehuda Sadeh 10863d14c5d2SYehuda Sadeh dout("__remove_old_osds %p\n", osdc); 10873d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 10883d14c5d2SYehuda Sadeh list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) { 1089aca420bcSSage Weil if (time_before(jiffies, osd->lru_ttl)) 10903d14c5d2SYehuda Sadeh break; 10913d14c5d2SYehuda Sadeh __remove_osd(osdc, osd); 10923d14c5d2SYehuda Sadeh } 10933d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 10943d14c5d2SYehuda Sadeh } 10953d14c5d2SYehuda Sadeh 10963d14c5d2SYehuda Sadeh /* 10973d14c5d2SYehuda Sadeh * reset osd connect 10983d14c5d2SYehuda Sadeh */ 10993d14c5d2SYehuda Sadeh static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 11003d14c5d2SYehuda Sadeh { 1101c3acb181SAlex Elder struct ceph_entity_addr *peer_addr; 11023d14c5d2SYehuda Sadeh 11033d14c5d2SYehuda Sadeh dout("__reset_osd %p osd%d\n", osd, osd->o_osd); 1104a40c4f10SYehuda Sadeh if (list_empty(&osd->o_requests) && 1105a40c4f10SYehuda Sadeh list_empty(&osd->o_linger_requests)) { 11063d14c5d2SYehuda Sadeh __remove_osd(osdc, osd); 1107c3acb181SAlex Elder 1108c3acb181SAlex Elder return -ENODEV; 1109c3acb181SAlex Elder } 1110c3acb181SAlex Elder 1111c3acb181SAlex Elder peer_addr = &osdc->osdmap->osd_addr[osd->o_osd]; 1112c3acb181SAlex Elder if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) && 11133d14c5d2SYehuda Sadeh !ceph_con_opened(&osd->o_con)) { 1114c3acb181SAlex Elder struct ceph_osd_request *req; 1115c3acb181SAlex Elder 11163d14c5d2SYehuda Sadeh dout("osd addr hasn't changed and connection never opened, " 11170b4af2e8SIlya Dryomov "letting msgr retry\n"); 11183d14c5d2SYehuda Sadeh /* touch each r_stamp for handle_timeout()'s benfit */ 11193d14c5d2SYehuda Sadeh list_for_each_entry(req, &osd->o_requests, r_osd_item) 11203d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 1121c3acb181SAlex Elder 1122c3acb181SAlex Elder return -EAGAIN; 11233d14c5d2SYehuda Sadeh } 1124c3acb181SAlex Elder 1125c3acb181SAlex Elder ceph_con_close(&osd->o_con); 1126c3acb181SAlex Elder ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr); 1127c3acb181SAlex Elder osd->o_incarnation++; 1128c3acb181SAlex Elder 1129c3acb181SAlex Elder return 0; 11303d14c5d2SYehuda Sadeh } 11313d14c5d2SYehuda Sadeh 11323d14c5d2SYehuda Sadeh static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new) 11333d14c5d2SYehuda Sadeh { 11343d14c5d2SYehuda Sadeh struct rb_node **p = &osdc->osds.rb_node; 11353d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 11363d14c5d2SYehuda Sadeh struct ceph_osd *osd = NULL; 11373d14c5d2SYehuda Sadeh 1138aca420bcSSage Weil dout("__insert_osd %p osd%d\n", new, new->o_osd); 11393d14c5d2SYehuda Sadeh while (*p) { 11403d14c5d2SYehuda Sadeh parent = *p; 11413d14c5d2SYehuda Sadeh osd = rb_entry(parent, struct ceph_osd, o_node); 11423d14c5d2SYehuda Sadeh if (new->o_osd < osd->o_osd) 11433d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 11443d14c5d2SYehuda Sadeh else if (new->o_osd > osd->o_osd) 11453d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 11463d14c5d2SYehuda Sadeh else 11473d14c5d2SYehuda Sadeh BUG(); 11483d14c5d2SYehuda Sadeh } 11493d14c5d2SYehuda Sadeh 11503d14c5d2SYehuda Sadeh rb_link_node(&new->o_node, parent, p); 11513d14c5d2SYehuda Sadeh rb_insert_color(&new->o_node, &osdc->osds); 11523d14c5d2SYehuda Sadeh } 11533d14c5d2SYehuda Sadeh 11543d14c5d2SYehuda Sadeh static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o) 11553d14c5d2SYehuda Sadeh { 11563d14c5d2SYehuda Sadeh struct ceph_osd *osd; 11573d14c5d2SYehuda Sadeh struct rb_node *n = osdc->osds.rb_node; 11583d14c5d2SYehuda Sadeh 11593d14c5d2SYehuda Sadeh while (n) { 11603d14c5d2SYehuda Sadeh osd = rb_entry(n, struct ceph_osd, o_node); 11613d14c5d2SYehuda Sadeh if (o < osd->o_osd) 11623d14c5d2SYehuda Sadeh n = n->rb_left; 11633d14c5d2SYehuda Sadeh else if (o > osd->o_osd) 11643d14c5d2SYehuda Sadeh n = n->rb_right; 11653d14c5d2SYehuda Sadeh else 11663d14c5d2SYehuda Sadeh return osd; 11673d14c5d2SYehuda Sadeh } 11683d14c5d2SYehuda Sadeh return NULL; 11693d14c5d2SYehuda Sadeh } 11703d14c5d2SYehuda Sadeh 11713d14c5d2SYehuda Sadeh static void __schedule_osd_timeout(struct ceph_osd_client *osdc) 11723d14c5d2SYehuda Sadeh { 11733d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->timeout_work, 11743d14c5d2SYehuda Sadeh osdc->client->options->osd_keepalive_timeout * HZ); 11753d14c5d2SYehuda Sadeh } 11763d14c5d2SYehuda Sadeh 11773d14c5d2SYehuda Sadeh static void __cancel_osd_timeout(struct ceph_osd_client *osdc) 11783d14c5d2SYehuda Sadeh { 11793d14c5d2SYehuda Sadeh cancel_delayed_work(&osdc->timeout_work); 11803d14c5d2SYehuda Sadeh } 11813d14c5d2SYehuda Sadeh 11823d14c5d2SYehuda Sadeh /* 11833d14c5d2SYehuda Sadeh * Register request, assign tid. If this is the first request, set up 11843d14c5d2SYehuda Sadeh * the timeout event. 11853d14c5d2SYehuda Sadeh */ 1186a40c4f10SYehuda Sadeh static void __register_request(struct ceph_osd_client *osdc, 11873d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 11883d14c5d2SYehuda Sadeh { 11893d14c5d2SYehuda Sadeh req->r_tid = ++osdc->last_tid; 11903d14c5d2SYehuda Sadeh req->r_request->hdr.tid = cpu_to_le64(req->r_tid); 119177f38e0eSSage Weil dout("__register_request %p tid %lld\n", req, req->r_tid); 11923d14c5d2SYehuda Sadeh __insert_request(osdc, req); 11933d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 11943d14c5d2SYehuda Sadeh osdc->num_requests++; 11953d14c5d2SYehuda Sadeh if (osdc->num_requests == 1) { 11963d14c5d2SYehuda Sadeh dout(" first request, scheduling timeout\n"); 11973d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 11983d14c5d2SYehuda Sadeh } 1199a40c4f10SYehuda Sadeh } 1200a40c4f10SYehuda Sadeh 12013d14c5d2SYehuda Sadeh /* 12023d14c5d2SYehuda Sadeh * called under osdc->request_mutex 12033d14c5d2SYehuda Sadeh */ 12043d14c5d2SYehuda Sadeh static void __unregister_request(struct ceph_osd_client *osdc, 12053d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 12063d14c5d2SYehuda Sadeh { 120735f9f8a0SSage Weil if (RB_EMPTY_NODE(&req->r_node)) { 120835f9f8a0SSage Weil dout("__unregister_request %p tid %lld not registered\n", 120935f9f8a0SSage Weil req, req->r_tid); 121035f9f8a0SSage Weil return; 121135f9f8a0SSage Weil } 121235f9f8a0SSage Weil 12133d14c5d2SYehuda Sadeh dout("__unregister_request %p tid %lld\n", req, req->r_tid); 12143d14c5d2SYehuda Sadeh rb_erase(&req->r_node, &osdc->requests); 12156562d661SIlya Dryomov RB_CLEAR_NODE(&req->r_node); 12163d14c5d2SYehuda Sadeh osdc->num_requests--; 12173d14c5d2SYehuda Sadeh 12183d14c5d2SYehuda Sadeh if (req->r_osd) { 12193d14c5d2SYehuda Sadeh /* make sure the original request isn't in flight. */ 12206740a845SAlex Elder ceph_msg_revoke(req->r_request); 12213d14c5d2SYehuda Sadeh 12223d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 1223bbf37ec3SIlya Dryomov maybe_move_osd_to_lru(osdc, req->r_osd); 12244f23409eSIlya Dryomov if (list_empty(&req->r_linger_osd_item)) 12253d14c5d2SYehuda Sadeh req->r_osd = NULL; 12263d14c5d2SYehuda Sadeh } 12273d14c5d2SYehuda Sadeh 12287d5f2481SAlex Elder list_del_init(&req->r_req_lru_item); 12293d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 12303d14c5d2SYehuda Sadeh 12313d14c5d2SYehuda Sadeh if (osdc->num_requests == 0) { 12323d14c5d2SYehuda Sadeh dout(" no requests, canceling timeout\n"); 12333d14c5d2SYehuda Sadeh __cancel_osd_timeout(osdc); 12343d14c5d2SYehuda Sadeh } 12353d14c5d2SYehuda Sadeh } 12363d14c5d2SYehuda Sadeh 12373d14c5d2SYehuda Sadeh /* 12383d14c5d2SYehuda Sadeh * Cancel a previously queued request message 12393d14c5d2SYehuda Sadeh */ 12403d14c5d2SYehuda Sadeh static void __cancel_request(struct ceph_osd_request *req) 12413d14c5d2SYehuda Sadeh { 12423d14c5d2SYehuda Sadeh if (req->r_sent && req->r_osd) { 12436740a845SAlex Elder ceph_msg_revoke(req->r_request); 12443d14c5d2SYehuda Sadeh req->r_sent = 0; 12453d14c5d2SYehuda Sadeh } 12463d14c5d2SYehuda Sadeh } 12473d14c5d2SYehuda Sadeh 1248a40c4f10SYehuda Sadeh static void __register_linger_request(struct ceph_osd_client *osdc, 1249a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1250a40c4f10SYehuda Sadeh { 1251af593064SIlya Dryomov dout("%s %p tid %llu\n", __func__, req, req->r_tid); 1252af593064SIlya Dryomov WARN_ON(!req->r_linger); 1253af593064SIlya Dryomov 125496e4dac6SAlex Elder ceph_osdc_get_request(req); 1255a40c4f10SYehuda Sadeh list_add_tail(&req->r_linger_item, &osdc->req_linger); 12566194ea89SSage Weil if (req->r_osd) 12571d0326b1SIlya Dryomov list_add_tail(&req->r_linger_osd_item, 12586194ea89SSage Weil &req->r_osd->o_linger_requests); 1259a40c4f10SYehuda Sadeh } 1260a40c4f10SYehuda Sadeh 1261a40c4f10SYehuda Sadeh static void __unregister_linger_request(struct ceph_osd_client *osdc, 1262a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1263a40c4f10SYehuda Sadeh { 1264af593064SIlya Dryomov WARN_ON(!req->r_linger); 1265af593064SIlya Dryomov 1266af593064SIlya Dryomov if (list_empty(&req->r_linger_item)) { 1267af593064SIlya Dryomov dout("%s %p tid %llu not registered\n", __func__, req, 1268af593064SIlya Dryomov req->r_tid); 1269af593064SIlya Dryomov return; 1270af593064SIlya Dryomov } 1271af593064SIlya Dryomov 1272af593064SIlya Dryomov dout("%s %p tid %llu\n", __func__, req, req->r_tid); 1273a40c4f10SYehuda Sadeh list_del_init(&req->r_linger_item); 1274af593064SIlya Dryomov 127561c74035SAlex Elder if (req->r_osd) { 12761d0326b1SIlya Dryomov list_del_init(&req->r_linger_osd_item); 1277bbf37ec3SIlya Dryomov maybe_move_osd_to_lru(osdc, req->r_osd); 1278fbdb9190SSage Weil if (list_empty(&req->r_osd_item)) 1279a40c4f10SYehuda Sadeh req->r_osd = NULL; 1280a40c4f10SYehuda Sadeh } 128196e4dac6SAlex Elder ceph_osdc_put_request(req); 1282a40c4f10SYehuda Sadeh } 1283a40c4f10SYehuda Sadeh 1284a40c4f10SYehuda Sadeh void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc, 1285a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1286a40c4f10SYehuda Sadeh { 1287a40c4f10SYehuda Sadeh if (!req->r_linger) { 1288a40c4f10SYehuda Sadeh dout("set_request_linger %p\n", req); 1289a40c4f10SYehuda Sadeh req->r_linger = 1; 1290a40c4f10SYehuda Sadeh } 1291a40c4f10SYehuda Sadeh } 1292a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_set_request_linger); 1293a40c4f10SYehuda Sadeh 12943d14c5d2SYehuda Sadeh /* 1295d29adb34SJosh Durgin * Returns whether a request should be blocked from being sent 1296d29adb34SJosh Durgin * based on the current osdmap and osd_client settings. 1297d29adb34SJosh Durgin * 1298d29adb34SJosh Durgin * Caller should hold map_sem for read. 1299d29adb34SJosh Durgin */ 1300d29adb34SJosh Durgin static bool __req_should_be_paused(struct ceph_osd_client *osdc, 1301d29adb34SJosh Durgin struct ceph_osd_request *req) 1302d29adb34SJosh Durgin { 1303d29adb34SJosh Durgin bool pauserd = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD); 1304d29adb34SJosh Durgin bool pausewr = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR) || 1305d29adb34SJosh Durgin ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL); 1306d29adb34SJosh Durgin return (req->r_flags & CEPH_OSD_FLAG_READ && pauserd) || 1307d29adb34SJosh Durgin (req->r_flags & CEPH_OSD_FLAG_WRITE && pausewr); 1308d29adb34SJosh Durgin } 1309d29adb34SJosh Durgin 1310d29adb34SJosh Durgin /* 131117a13e40SIlya Dryomov * Calculate mapping of a request to a PG. Takes tiering into account. 131217a13e40SIlya Dryomov */ 131317a13e40SIlya Dryomov static int __calc_request_pg(struct ceph_osdmap *osdmap, 131417a13e40SIlya Dryomov struct ceph_osd_request *req, 131517a13e40SIlya Dryomov struct ceph_pg *pg_out) 131617a13e40SIlya Dryomov { 1317205ee118SIlya Dryomov bool need_check_tiering; 1318205ee118SIlya Dryomov 1319205ee118SIlya Dryomov need_check_tiering = false; 1320205ee118SIlya Dryomov if (req->r_target_oloc.pool == -1) { 1321205ee118SIlya Dryomov req->r_target_oloc = req->r_base_oloc; /* struct */ 1322205ee118SIlya Dryomov need_check_tiering = true; 1323205ee118SIlya Dryomov } 1324205ee118SIlya Dryomov if (req->r_target_oid.name_len == 0) { 1325205ee118SIlya Dryomov ceph_oid_copy(&req->r_target_oid, &req->r_base_oid); 1326205ee118SIlya Dryomov need_check_tiering = true; 1327205ee118SIlya Dryomov } 1328205ee118SIlya Dryomov 1329205ee118SIlya Dryomov if (need_check_tiering && 1330205ee118SIlya Dryomov (req->r_flags & CEPH_OSD_FLAG_IGNORE_OVERLAY) == 0) { 133117a13e40SIlya Dryomov struct ceph_pg_pool_info *pi; 133217a13e40SIlya Dryomov 1333205ee118SIlya Dryomov pi = ceph_pg_pool_by_id(osdmap, req->r_target_oloc.pool); 133417a13e40SIlya Dryomov if (pi) { 133517a13e40SIlya Dryomov if ((req->r_flags & CEPH_OSD_FLAG_READ) && 133617a13e40SIlya Dryomov pi->read_tier >= 0) 1337205ee118SIlya Dryomov req->r_target_oloc.pool = pi->read_tier; 133817a13e40SIlya Dryomov if ((req->r_flags & CEPH_OSD_FLAG_WRITE) && 133917a13e40SIlya Dryomov pi->write_tier >= 0) 1340205ee118SIlya Dryomov req->r_target_oloc.pool = pi->write_tier; 134117a13e40SIlya Dryomov } 134217a13e40SIlya Dryomov /* !pi is caught in ceph_oloc_oid_to_pg() */ 134317a13e40SIlya Dryomov } 134417a13e40SIlya Dryomov 1345205ee118SIlya Dryomov return ceph_oloc_oid_to_pg(osdmap, &req->r_target_oloc, 1346205ee118SIlya Dryomov &req->r_target_oid, pg_out); 134717a13e40SIlya Dryomov } 134817a13e40SIlya Dryomov 1349f671b581SIlya Dryomov static void __enqueue_request(struct ceph_osd_request *req) 1350f671b581SIlya Dryomov { 1351f671b581SIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 1352f671b581SIlya Dryomov 1353f671b581SIlya Dryomov dout("%s %p tid %llu to osd%d\n", __func__, req, req->r_tid, 1354f671b581SIlya Dryomov req->r_osd ? req->r_osd->o_osd : -1); 1355f671b581SIlya Dryomov 1356f671b581SIlya Dryomov if (req->r_osd) { 1357f671b581SIlya Dryomov __remove_osd_from_lru(req->r_osd); 1358f671b581SIlya Dryomov list_add_tail(&req->r_osd_item, &req->r_osd->o_requests); 1359f671b581SIlya Dryomov list_move_tail(&req->r_req_lru_item, &osdc->req_unsent); 1360f671b581SIlya Dryomov } else { 1361f671b581SIlya Dryomov list_move_tail(&req->r_req_lru_item, &osdc->req_notarget); 1362f671b581SIlya Dryomov } 1363f671b581SIlya Dryomov } 1364f671b581SIlya Dryomov 136517a13e40SIlya Dryomov /* 13663d14c5d2SYehuda Sadeh * Pick an osd (the first 'up' osd in the pg), allocate the osd struct 13673d14c5d2SYehuda Sadeh * (as needed), and set the request r_osd appropriately. If there is 136825985edcSLucas De Marchi * no up osd, set r_osd to NULL. Move the request to the appropriate list 13696f6c7006SSage Weil * (unsent, homeless) or leave on in-flight lru. 13703d14c5d2SYehuda Sadeh * 13713d14c5d2SYehuda Sadeh * Return 0 if unchanged, 1 if changed, or negative on error. 13723d14c5d2SYehuda Sadeh * 13733d14c5d2SYehuda Sadeh * Caller should hold map_sem for read and request_mutex. 13743d14c5d2SYehuda Sadeh */ 13756f6c7006SSage Weil static int __map_request(struct ceph_osd_client *osdc, 137638d6453cSSage Weil struct ceph_osd_request *req, int force_resend) 13773d14c5d2SYehuda Sadeh { 13785b191d99SSage Weil struct ceph_pg pgid; 13793d14c5d2SYehuda Sadeh int acting[CEPH_PG_MAX_SIZE]; 13808008ab10SIlya Dryomov int num, o; 13813d14c5d2SYehuda Sadeh int err; 1382d29adb34SJosh Durgin bool was_paused; 13833d14c5d2SYehuda Sadeh 13846f6c7006SSage Weil dout("map_request %p tid %lld\n", req, req->r_tid); 138517a13e40SIlya Dryomov 138617a13e40SIlya Dryomov err = __calc_request_pg(osdc->osdmap, req, &pgid); 13876f6c7006SSage Weil if (err) { 13886f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 13893d14c5d2SYehuda Sadeh return err; 13906f6c7006SSage Weil } 13913d14c5d2SYehuda Sadeh req->r_pgid = pgid; 13923d14c5d2SYehuda Sadeh 13938008ab10SIlya Dryomov num = ceph_calc_pg_acting(osdc->osdmap, pgid, acting, &o); 13948008ab10SIlya Dryomov if (num < 0) 13958008ab10SIlya Dryomov num = 0; 13963d14c5d2SYehuda Sadeh 1397d29adb34SJosh Durgin was_paused = req->r_paused; 1398d29adb34SJosh Durgin req->r_paused = __req_should_be_paused(osdc, req); 1399d29adb34SJosh Durgin if (was_paused && !req->r_paused) 1400d29adb34SJosh Durgin force_resend = 1; 1401d29adb34SJosh Durgin 140238d6453cSSage Weil if ((!force_resend && 140338d6453cSSage Weil req->r_osd && req->r_osd->o_osd == o && 14043d14c5d2SYehuda Sadeh req->r_sent >= req->r_osd->o_incarnation && 14053d14c5d2SYehuda Sadeh req->r_num_pg_osds == num && 14063d14c5d2SYehuda Sadeh memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) || 1407d29adb34SJosh Durgin (req->r_osd == NULL && o == -1) || 1408d29adb34SJosh Durgin req->r_paused) 14093d14c5d2SYehuda Sadeh return 0; /* no change */ 14103d14c5d2SYehuda Sadeh 14115b191d99SSage Weil dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n", 14125b191d99SSage Weil req->r_tid, pgid.pool, pgid.seed, o, 14133d14c5d2SYehuda Sadeh req->r_osd ? req->r_osd->o_osd : -1); 14143d14c5d2SYehuda Sadeh 14153d14c5d2SYehuda Sadeh /* record full pg acting set */ 14163d14c5d2SYehuda Sadeh memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num); 14173d14c5d2SYehuda Sadeh req->r_num_pg_osds = num; 14183d14c5d2SYehuda Sadeh 14193d14c5d2SYehuda Sadeh if (req->r_osd) { 14203d14c5d2SYehuda Sadeh __cancel_request(req); 14213d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 14223d14c5d2SYehuda Sadeh req->r_osd = NULL; 14233d14c5d2SYehuda Sadeh } 14243d14c5d2SYehuda Sadeh 14253d14c5d2SYehuda Sadeh req->r_osd = __lookup_osd(osdc, o); 14263d14c5d2SYehuda Sadeh if (!req->r_osd && o >= 0) { 14273d14c5d2SYehuda Sadeh err = -ENOMEM; 1428e10006f8SAlex Elder req->r_osd = create_osd(osdc, o); 14296f6c7006SSage Weil if (!req->r_osd) { 14306f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 14313d14c5d2SYehuda Sadeh goto out; 14326f6c7006SSage Weil } 14333d14c5d2SYehuda Sadeh 14346f6c7006SSage Weil dout("map_request osd %p is osd%d\n", req->r_osd, o); 14353d14c5d2SYehuda Sadeh __insert_osd(osdc, req->r_osd); 14363d14c5d2SYehuda Sadeh 1437b7a9e5ddSSage Weil ceph_con_open(&req->r_osd->o_con, 1438b7a9e5ddSSage Weil CEPH_ENTITY_TYPE_OSD, o, 1439b7a9e5ddSSage Weil &osdc->osdmap->osd_addr[o]); 14403d14c5d2SYehuda Sadeh } 14413d14c5d2SYehuda Sadeh 1442f671b581SIlya Dryomov __enqueue_request(req); 14433d14c5d2SYehuda Sadeh err = 1; /* osd or pg changed */ 14443d14c5d2SYehuda Sadeh 14453d14c5d2SYehuda Sadeh out: 14463d14c5d2SYehuda Sadeh return err; 14473d14c5d2SYehuda Sadeh } 14483d14c5d2SYehuda Sadeh 14493d14c5d2SYehuda Sadeh /* 14503d14c5d2SYehuda Sadeh * caller should hold map_sem (for read) and request_mutex 14513d14c5d2SYehuda Sadeh */ 145256e925b6SSage Weil static void __send_request(struct ceph_osd_client *osdc, 14533d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 14543d14c5d2SYehuda Sadeh { 14551b83bef2SSage Weil void *p; 14563d14c5d2SYehuda Sadeh 14571b83bef2SSage Weil dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n", 14581b83bef2SSage Weil req, req->r_tid, req->r_osd->o_osd, req->r_flags, 14591b83bef2SSage Weil (unsigned long long)req->r_pgid.pool, req->r_pgid.seed); 14603d14c5d2SYehuda Sadeh 14611b83bef2SSage Weil /* fill in message content that changes each time we send it */ 14621b83bef2SSage Weil put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch); 14631b83bef2SSage Weil put_unaligned_le32(req->r_flags, req->r_request_flags); 1464205ee118SIlya Dryomov put_unaligned_le64(req->r_target_oloc.pool, req->r_request_pool); 14651b83bef2SSage Weil p = req->r_request_pgid; 14661b83bef2SSage Weil ceph_encode_64(&p, req->r_pgid.pool); 14671b83bef2SSage Weil ceph_encode_32(&p, req->r_pgid.seed); 14681b83bef2SSage Weil put_unaligned_le64(1, req->r_request_attempts); /* FIXME */ 14691b83bef2SSage Weil memcpy(req->r_request_reassert_version, &req->r_reassert_version, 14701b83bef2SSage Weil sizeof(req->r_reassert_version)); 14712169aea6SSage Weil 14723d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 14733d14c5d2SYehuda Sadeh list_move_tail(&req->r_req_lru_item, &osdc->req_lru); 14743d14c5d2SYehuda Sadeh 14753d14c5d2SYehuda Sadeh ceph_msg_get(req->r_request); /* send consumes a ref */ 147626be8808SAlex Elder 14773d14c5d2SYehuda Sadeh req->r_sent = req->r_osd->o_incarnation; 147826be8808SAlex Elder 147926be8808SAlex Elder ceph_con_send(&req->r_osd->o_con, req->r_request); 14803d14c5d2SYehuda Sadeh } 14813d14c5d2SYehuda Sadeh 14823d14c5d2SYehuda Sadeh /* 14836f6c7006SSage Weil * Send any requests in the queue (req_unsent). 14846f6c7006SSage Weil */ 1485f9d25199SAlex Elder static void __send_queued(struct ceph_osd_client *osdc) 14866f6c7006SSage Weil { 14876f6c7006SSage Weil struct ceph_osd_request *req, *tmp; 14886f6c7006SSage Weil 1489f9d25199SAlex Elder dout("__send_queued\n"); 1490f9d25199SAlex Elder list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) 14916f6c7006SSage Weil __send_request(osdc, req); 14926f6c7006SSage Weil } 14936f6c7006SSage Weil 14946f6c7006SSage Weil /* 14950bbfdfe8SIlya Dryomov * Caller should hold map_sem for read and request_mutex. 14960bbfdfe8SIlya Dryomov */ 14970bbfdfe8SIlya Dryomov static int __ceph_osdc_start_request(struct ceph_osd_client *osdc, 14980bbfdfe8SIlya Dryomov struct ceph_osd_request *req, 14990bbfdfe8SIlya Dryomov bool nofail) 15000bbfdfe8SIlya Dryomov { 15010bbfdfe8SIlya Dryomov int rc; 15020bbfdfe8SIlya Dryomov 15030bbfdfe8SIlya Dryomov __register_request(osdc, req); 15040bbfdfe8SIlya Dryomov req->r_sent = 0; 15050bbfdfe8SIlya Dryomov req->r_got_reply = 0; 15060bbfdfe8SIlya Dryomov rc = __map_request(osdc, req, 0); 15070bbfdfe8SIlya Dryomov if (rc < 0) { 15080bbfdfe8SIlya Dryomov if (nofail) { 15090bbfdfe8SIlya Dryomov dout("osdc_start_request failed map, " 15100bbfdfe8SIlya Dryomov " will retry %lld\n", req->r_tid); 15110bbfdfe8SIlya Dryomov rc = 0; 15120bbfdfe8SIlya Dryomov } else { 15130bbfdfe8SIlya Dryomov __unregister_request(osdc, req); 15140bbfdfe8SIlya Dryomov } 15150bbfdfe8SIlya Dryomov return rc; 15160bbfdfe8SIlya Dryomov } 15170bbfdfe8SIlya Dryomov 15180bbfdfe8SIlya Dryomov if (req->r_osd == NULL) { 15190bbfdfe8SIlya Dryomov dout("send_request %p no up osds in pg\n", req); 15200bbfdfe8SIlya Dryomov ceph_monc_request_next_osdmap(&osdc->client->monc); 15210bbfdfe8SIlya Dryomov } else { 15220bbfdfe8SIlya Dryomov __send_queued(osdc); 15230bbfdfe8SIlya Dryomov } 15240bbfdfe8SIlya Dryomov 15250bbfdfe8SIlya Dryomov return 0; 15260bbfdfe8SIlya Dryomov } 15270bbfdfe8SIlya Dryomov 15280bbfdfe8SIlya Dryomov /* 15293d14c5d2SYehuda Sadeh * Timeout callback, called every N seconds when 1 or more osd 15303d14c5d2SYehuda Sadeh * requests has been active for more than N seconds. When this 15313d14c5d2SYehuda Sadeh * happens, we ping all OSDs with requests who have timed out to 15323d14c5d2SYehuda Sadeh * ensure any communications channel reset is detected. Reset the 15333d14c5d2SYehuda Sadeh * request timeouts another N seconds in the future as we go. 15343d14c5d2SYehuda Sadeh * Reschedule the timeout event another N seconds in future (unless 15353d14c5d2SYehuda Sadeh * there are no open requests). 15363d14c5d2SYehuda Sadeh */ 15373d14c5d2SYehuda Sadeh static void handle_timeout(struct work_struct *work) 15383d14c5d2SYehuda Sadeh { 15393d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 15403d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, timeout_work.work); 154183aff95eSSage Weil struct ceph_osd_request *req; 15423d14c5d2SYehuda Sadeh struct ceph_osd *osd; 15433d14c5d2SYehuda Sadeh unsigned long keepalive = 15443d14c5d2SYehuda Sadeh osdc->client->options->osd_keepalive_timeout * HZ; 15453d14c5d2SYehuda Sadeh struct list_head slow_osds; 15463d14c5d2SYehuda Sadeh dout("timeout\n"); 15473d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 15483d14c5d2SYehuda Sadeh 15493d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 15503d14c5d2SYehuda Sadeh 15513d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 15523d14c5d2SYehuda Sadeh 15533d14c5d2SYehuda Sadeh /* 15543d14c5d2SYehuda Sadeh * ping osds that are a bit slow. this ensures that if there 15553d14c5d2SYehuda Sadeh * is a break in the TCP connection we will notice, and reopen 15563d14c5d2SYehuda Sadeh * a connection with that osd (from the fault callback). 15573d14c5d2SYehuda Sadeh */ 15583d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&slow_osds); 15593d14c5d2SYehuda Sadeh list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) { 15603d14c5d2SYehuda Sadeh if (time_before(jiffies, req->r_stamp + keepalive)) 15613d14c5d2SYehuda Sadeh break; 15623d14c5d2SYehuda Sadeh 15633d14c5d2SYehuda Sadeh osd = req->r_osd; 15643d14c5d2SYehuda Sadeh BUG_ON(!osd); 15653d14c5d2SYehuda Sadeh dout(" tid %llu is slow, will send keepalive on osd%d\n", 15663d14c5d2SYehuda Sadeh req->r_tid, osd->o_osd); 15673d14c5d2SYehuda Sadeh list_move_tail(&osd->o_keepalive_item, &slow_osds); 15683d14c5d2SYehuda Sadeh } 15693d14c5d2SYehuda Sadeh while (!list_empty(&slow_osds)) { 15703d14c5d2SYehuda Sadeh osd = list_entry(slow_osds.next, struct ceph_osd, 15713d14c5d2SYehuda Sadeh o_keepalive_item); 15723d14c5d2SYehuda Sadeh list_del_init(&osd->o_keepalive_item); 15733d14c5d2SYehuda Sadeh ceph_con_keepalive(&osd->o_con); 15743d14c5d2SYehuda Sadeh } 15753d14c5d2SYehuda Sadeh 15763d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 1577f9d25199SAlex Elder __send_queued(osdc); 15783d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 15793d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 15803d14c5d2SYehuda Sadeh } 15813d14c5d2SYehuda Sadeh 15823d14c5d2SYehuda Sadeh static void handle_osds_timeout(struct work_struct *work) 15833d14c5d2SYehuda Sadeh { 15843d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 15853d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, 15863d14c5d2SYehuda Sadeh osds_timeout_work.work); 15873d14c5d2SYehuda Sadeh unsigned long delay = 15883d14c5d2SYehuda Sadeh osdc->client->options->osd_idle_ttl * HZ >> 2; 15893d14c5d2SYehuda Sadeh 15903d14c5d2SYehuda Sadeh dout("osds timeout\n"); 15913d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 1592aca420bcSSage Weil remove_old_osds(osdc); 15933d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 15943d14c5d2SYehuda Sadeh 15953d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 15963d14c5d2SYehuda Sadeh round_jiffies_relative(delay)); 15973d14c5d2SYehuda Sadeh } 15983d14c5d2SYehuda Sadeh 1599205ee118SIlya Dryomov static int ceph_oloc_decode(void **p, void *end, 1600205ee118SIlya Dryomov struct ceph_object_locator *oloc) 1601205ee118SIlya Dryomov { 1602205ee118SIlya Dryomov u8 struct_v, struct_cv; 1603205ee118SIlya Dryomov u32 len; 1604205ee118SIlya Dryomov void *struct_end; 1605205ee118SIlya Dryomov int ret = 0; 1606205ee118SIlya Dryomov 1607205ee118SIlya Dryomov ceph_decode_need(p, end, 1 + 1 + 4, e_inval); 1608205ee118SIlya Dryomov struct_v = ceph_decode_8(p); 1609205ee118SIlya Dryomov struct_cv = ceph_decode_8(p); 1610205ee118SIlya Dryomov if (struct_v < 3) { 1611205ee118SIlya Dryomov pr_warn("got v %d < 3 cv %d of ceph_object_locator\n", 1612205ee118SIlya Dryomov struct_v, struct_cv); 1613205ee118SIlya Dryomov goto e_inval; 1614205ee118SIlya Dryomov } 1615205ee118SIlya Dryomov if (struct_cv > 6) { 1616205ee118SIlya Dryomov pr_warn("got v %d cv %d > 6 of ceph_object_locator\n", 1617205ee118SIlya Dryomov struct_v, struct_cv); 1618205ee118SIlya Dryomov goto e_inval; 1619205ee118SIlya Dryomov } 1620205ee118SIlya Dryomov len = ceph_decode_32(p); 1621205ee118SIlya Dryomov ceph_decode_need(p, end, len, e_inval); 1622205ee118SIlya Dryomov struct_end = *p + len; 1623205ee118SIlya Dryomov 1624205ee118SIlya Dryomov oloc->pool = ceph_decode_64(p); 1625205ee118SIlya Dryomov *p += 4; /* skip preferred */ 1626205ee118SIlya Dryomov 1627205ee118SIlya Dryomov len = ceph_decode_32(p); 1628205ee118SIlya Dryomov if (len > 0) { 1629205ee118SIlya Dryomov pr_warn("ceph_object_locator::key is set\n"); 1630205ee118SIlya Dryomov goto e_inval; 1631205ee118SIlya Dryomov } 1632205ee118SIlya Dryomov 1633205ee118SIlya Dryomov if (struct_v >= 5) { 1634205ee118SIlya Dryomov len = ceph_decode_32(p); 1635205ee118SIlya Dryomov if (len > 0) { 1636205ee118SIlya Dryomov pr_warn("ceph_object_locator::nspace is set\n"); 1637205ee118SIlya Dryomov goto e_inval; 1638205ee118SIlya Dryomov } 1639205ee118SIlya Dryomov } 1640205ee118SIlya Dryomov 1641205ee118SIlya Dryomov if (struct_v >= 6) { 1642205ee118SIlya Dryomov s64 hash = ceph_decode_64(p); 1643205ee118SIlya Dryomov if (hash != -1) { 1644205ee118SIlya Dryomov pr_warn("ceph_object_locator::hash is set\n"); 1645205ee118SIlya Dryomov goto e_inval; 1646205ee118SIlya Dryomov } 1647205ee118SIlya Dryomov } 1648205ee118SIlya Dryomov 1649205ee118SIlya Dryomov /* skip the rest */ 1650205ee118SIlya Dryomov *p = struct_end; 1651205ee118SIlya Dryomov out: 1652205ee118SIlya Dryomov return ret; 1653205ee118SIlya Dryomov 1654205ee118SIlya Dryomov e_inval: 1655205ee118SIlya Dryomov ret = -EINVAL; 1656205ee118SIlya Dryomov goto out; 1657205ee118SIlya Dryomov } 1658205ee118SIlya Dryomov 1659205ee118SIlya Dryomov static int ceph_redirect_decode(void **p, void *end, 1660205ee118SIlya Dryomov struct ceph_request_redirect *redir) 1661205ee118SIlya Dryomov { 1662205ee118SIlya Dryomov u8 struct_v, struct_cv; 1663205ee118SIlya Dryomov u32 len; 1664205ee118SIlya Dryomov void *struct_end; 1665205ee118SIlya Dryomov int ret; 1666205ee118SIlya Dryomov 1667205ee118SIlya Dryomov ceph_decode_need(p, end, 1 + 1 + 4, e_inval); 1668205ee118SIlya Dryomov struct_v = ceph_decode_8(p); 1669205ee118SIlya Dryomov struct_cv = ceph_decode_8(p); 1670205ee118SIlya Dryomov if (struct_cv > 1) { 1671205ee118SIlya Dryomov pr_warn("got v %d cv %d > 1 of ceph_request_redirect\n", 1672205ee118SIlya Dryomov struct_v, struct_cv); 1673205ee118SIlya Dryomov goto e_inval; 1674205ee118SIlya Dryomov } 1675205ee118SIlya Dryomov len = ceph_decode_32(p); 1676205ee118SIlya Dryomov ceph_decode_need(p, end, len, e_inval); 1677205ee118SIlya Dryomov struct_end = *p + len; 1678205ee118SIlya Dryomov 1679205ee118SIlya Dryomov ret = ceph_oloc_decode(p, end, &redir->oloc); 1680205ee118SIlya Dryomov if (ret) 1681205ee118SIlya Dryomov goto out; 1682205ee118SIlya Dryomov 1683205ee118SIlya Dryomov len = ceph_decode_32(p); 1684205ee118SIlya Dryomov if (len > 0) { 1685205ee118SIlya Dryomov pr_warn("ceph_request_redirect::object_name is set\n"); 1686205ee118SIlya Dryomov goto e_inval; 1687205ee118SIlya Dryomov } 1688205ee118SIlya Dryomov 1689205ee118SIlya Dryomov len = ceph_decode_32(p); 1690205ee118SIlya Dryomov *p += len; /* skip osd_instructions */ 1691205ee118SIlya Dryomov 1692205ee118SIlya Dryomov /* skip the rest */ 1693205ee118SIlya Dryomov *p = struct_end; 1694205ee118SIlya Dryomov out: 1695205ee118SIlya Dryomov return ret; 1696205ee118SIlya Dryomov 1697205ee118SIlya Dryomov e_inval: 1698205ee118SIlya Dryomov ret = -EINVAL; 1699205ee118SIlya Dryomov goto out; 1700205ee118SIlya Dryomov } 1701205ee118SIlya Dryomov 170225845472SSage Weil static void complete_request(struct ceph_osd_request *req) 170325845472SSage Weil { 170425845472SSage Weil complete_all(&req->r_safe_completion); /* fsync waiter */ 170525845472SSage Weil } 170625845472SSage Weil 17073d14c5d2SYehuda Sadeh /* 17083d14c5d2SYehuda Sadeh * handle osd op reply. either call the callback if it is specified, 17093d14c5d2SYehuda Sadeh * or do the completion to wake up the waiting thread. 17103d14c5d2SYehuda Sadeh */ 17113d14c5d2SYehuda Sadeh static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg, 17123d14c5d2SYehuda Sadeh struct ceph_connection *con) 17133d14c5d2SYehuda Sadeh { 17141b83bef2SSage Weil void *p, *end; 17153d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 1716205ee118SIlya Dryomov struct ceph_request_redirect redir; 17173d14c5d2SYehuda Sadeh u64 tid; 17181b83bef2SSage Weil int object_len; 171979528734SAlex Elder unsigned int numops; 172079528734SAlex Elder int payload_len, flags; 17213d14c5d2SYehuda Sadeh s32 result; 17221b83bef2SSage Weil s32 retry_attempt; 17231b83bef2SSage Weil struct ceph_pg pg; 17241b83bef2SSage Weil int err; 17251b83bef2SSage Weil u32 reassert_epoch; 17261b83bef2SSage Weil u64 reassert_version; 17271b83bef2SSage Weil u32 osdmap_epoch; 17280d5af164SAlex Elder int already_completed; 17299fc6e064SAlex Elder u32 bytes; 173079528734SAlex Elder unsigned int i; 17313d14c5d2SYehuda Sadeh 17323d14c5d2SYehuda Sadeh tid = le64_to_cpu(msg->hdr.tid); 17331b83bef2SSage Weil dout("handle_reply %p tid %llu\n", msg, tid); 17341b83bef2SSage Weil 17351b83bef2SSage Weil p = msg->front.iov_base; 17361b83bef2SSage Weil end = p + msg->front.iov_len; 17371b83bef2SSage Weil 17381b83bef2SSage Weil ceph_decode_need(&p, end, 4, bad); 17391b83bef2SSage Weil object_len = ceph_decode_32(&p); 17401b83bef2SSage Weil ceph_decode_need(&p, end, object_len, bad); 17411b83bef2SSage Weil p += object_len; 17421b83bef2SSage Weil 1743ef4859d6SAlex Elder err = ceph_decode_pgid(&p, end, &pg); 17441b83bef2SSage Weil if (err) 17453d14c5d2SYehuda Sadeh goto bad; 17461b83bef2SSage Weil 17471b83bef2SSage Weil ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad); 17481b83bef2SSage Weil flags = ceph_decode_64(&p); 17491b83bef2SSage Weil result = ceph_decode_32(&p); 17501b83bef2SSage Weil reassert_epoch = ceph_decode_32(&p); 17511b83bef2SSage Weil reassert_version = ceph_decode_64(&p); 17521b83bef2SSage Weil osdmap_epoch = ceph_decode_32(&p); 17531b83bef2SSage Weil 17543d14c5d2SYehuda Sadeh /* lookup */ 1755ff513aceSIlya Dryomov down_read(&osdc->map_sem); 17563d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 17573d14c5d2SYehuda Sadeh req = __lookup_request(osdc, tid); 17583d14c5d2SYehuda Sadeh if (req == NULL) { 17593d14c5d2SYehuda Sadeh dout("handle_reply tid %llu dne\n", tid); 17608058fd45SAlex Elder goto bad_mutex; 17613d14c5d2SYehuda Sadeh } 17623d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 17631b83bef2SSage Weil 17641b83bef2SSage Weil dout("handle_reply %p tid %llu req %p result %d\n", msg, tid, 17651b83bef2SSage Weil req, result); 17661b83bef2SSage Weil 176718741196SDan Carpenter ceph_decode_need(&p, end, 4, bad_put); 17681b83bef2SSage Weil numops = ceph_decode_32(&p); 17691b83bef2SSage Weil if (numops > CEPH_OSD_MAX_OP) 17701b83bef2SSage Weil goto bad_put; 17711b83bef2SSage Weil if (numops != req->r_num_ops) 17721b83bef2SSage Weil goto bad_put; 17731b83bef2SSage Weil payload_len = 0; 177418741196SDan Carpenter ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad_put); 17751b83bef2SSage Weil for (i = 0; i < numops; i++) { 17761b83bef2SSage Weil struct ceph_osd_op *op = p; 17771b83bef2SSage Weil int len; 17781b83bef2SSage Weil 17791b83bef2SSage Weil len = le32_to_cpu(op->payload_len); 17801b83bef2SSage Weil req->r_reply_op_len[i] = len; 17811b83bef2SSage Weil dout(" op %d has %d bytes\n", i, len); 17821b83bef2SSage Weil payload_len += len; 17831b83bef2SSage Weil p += sizeof(*op); 17841b83bef2SSage Weil } 17859fc6e064SAlex Elder bytes = le32_to_cpu(msg->hdr.data_len); 17869fc6e064SAlex Elder if (payload_len != bytes) { 17871b83bef2SSage Weil pr_warning("sum of op payload lens %d != data_len %d", 17889fc6e064SAlex Elder payload_len, bytes); 17891b83bef2SSage Weil goto bad_put; 17901b83bef2SSage Weil } 17911b83bef2SSage Weil 179218741196SDan Carpenter ceph_decode_need(&p, end, 4 + numops * 4, bad_put); 17931b83bef2SSage Weil retry_attempt = ceph_decode_32(&p); 17941b83bef2SSage Weil for (i = 0; i < numops; i++) 17951b83bef2SSage Weil req->r_reply_op_result[i] = ceph_decode_32(&p); 17963d14c5d2SYehuda Sadeh 1797205ee118SIlya Dryomov if (le16_to_cpu(msg->hdr.version) >= 6) { 1798205ee118SIlya Dryomov p += 8 + 4; /* skip replay_version */ 1799205ee118SIlya Dryomov p += 8; /* skip user_version */ 1800205ee118SIlya Dryomov 1801205ee118SIlya Dryomov err = ceph_redirect_decode(&p, end, &redir); 1802205ee118SIlya Dryomov if (err) 1803205ee118SIlya Dryomov goto bad_put; 1804205ee118SIlya Dryomov } else { 1805205ee118SIlya Dryomov redir.oloc.pool = -1; 1806205ee118SIlya Dryomov } 1807205ee118SIlya Dryomov 1808205ee118SIlya Dryomov if (redir.oloc.pool != -1) { 1809205ee118SIlya Dryomov dout("redirect pool %lld\n", redir.oloc.pool); 1810205ee118SIlya Dryomov 1811205ee118SIlya Dryomov __unregister_request(osdc, req); 1812205ee118SIlya Dryomov 1813205ee118SIlya Dryomov req->r_target_oloc = redir.oloc; /* struct */ 1814205ee118SIlya Dryomov 1815205ee118SIlya Dryomov /* 1816205ee118SIlya Dryomov * Start redirect requests with nofail=true. If 1817205ee118SIlya Dryomov * mapping fails, request will end up on the notarget 1818205ee118SIlya Dryomov * list, waiting for the new osdmap (which can take 1819205ee118SIlya Dryomov * a while), even though the original request mapped 1820205ee118SIlya Dryomov * successfully. In the future we might want to follow 1821205ee118SIlya Dryomov * original request's nofail setting here. 1822205ee118SIlya Dryomov */ 1823ff513aceSIlya Dryomov err = __ceph_osdc_start_request(osdc, req, true); 1824205ee118SIlya Dryomov BUG_ON(err); 1825205ee118SIlya Dryomov 1826ff513aceSIlya Dryomov goto out_unlock; 1827205ee118SIlya Dryomov } 1828205ee118SIlya Dryomov 1829eb845ff1SYan, Zheng already_completed = req->r_got_reply; 18303d14c5d2SYehuda Sadeh if (!req->r_got_reply) { 18311b83bef2SSage Weil req->r_result = result; 18323d14c5d2SYehuda Sadeh dout("handle_reply result %d bytes %d\n", req->r_result, 18333d14c5d2SYehuda Sadeh bytes); 18343d14c5d2SYehuda Sadeh if (req->r_result == 0) 18353d14c5d2SYehuda Sadeh req->r_result = bytes; 18363d14c5d2SYehuda Sadeh 18373d14c5d2SYehuda Sadeh /* in case this is a write and we need to replay, */ 18381b83bef2SSage Weil req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch); 18391b83bef2SSage Weil req->r_reassert_version.version = cpu_to_le64(reassert_version); 18403d14c5d2SYehuda Sadeh 18413d14c5d2SYehuda Sadeh req->r_got_reply = 1; 18423d14c5d2SYehuda Sadeh } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) { 18433d14c5d2SYehuda Sadeh dout("handle_reply tid %llu dup ack\n", tid); 1844ff513aceSIlya Dryomov goto out_unlock; 18453d14c5d2SYehuda Sadeh } 18463d14c5d2SYehuda Sadeh 18473d14c5d2SYehuda Sadeh dout("handle_reply tid %llu flags %d\n", tid, flags); 18483d14c5d2SYehuda Sadeh 1849a40c4f10SYehuda Sadeh if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK)) 1850a40c4f10SYehuda Sadeh __register_linger_request(osdc, req); 1851a40c4f10SYehuda Sadeh 18523d14c5d2SYehuda Sadeh /* either this is a read, or we got the safe response */ 18533d14c5d2SYehuda Sadeh if (result < 0 || 18543d14c5d2SYehuda Sadeh (flags & CEPH_OSD_FLAG_ONDISK) || 18553d14c5d2SYehuda Sadeh ((flags & CEPH_OSD_FLAG_WRITE) == 0)) 18563d14c5d2SYehuda Sadeh __unregister_request(osdc, req); 18573d14c5d2SYehuda Sadeh 18583d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 1859ff513aceSIlya Dryomov up_read(&osdc->map_sem); 18603d14c5d2SYehuda Sadeh 1861eb845ff1SYan, Zheng if (!already_completed) { 186261c5d6bfSYan, Zheng if (req->r_unsafe_callback && 186361c5d6bfSYan, Zheng result >= 0 && !(flags & CEPH_OSD_FLAG_ONDISK)) 186461c5d6bfSYan, Zheng req->r_unsafe_callback(req, true); 18653d14c5d2SYehuda Sadeh if (req->r_callback) 18663d14c5d2SYehuda Sadeh req->r_callback(req, msg); 18673d14c5d2SYehuda Sadeh else 18683d14c5d2SYehuda Sadeh complete_all(&req->r_completion); 1869eb845ff1SYan, Zheng } 18703d14c5d2SYehuda Sadeh 187161c5d6bfSYan, Zheng if (flags & CEPH_OSD_FLAG_ONDISK) { 187261c5d6bfSYan, Zheng if (req->r_unsafe_callback && already_completed) 187361c5d6bfSYan, Zheng req->r_unsafe_callback(req, false); 187425845472SSage Weil complete_request(req); 187561c5d6bfSYan, Zheng } 18763d14c5d2SYehuda Sadeh 1877ff513aceSIlya Dryomov out: 1878a40c4f10SYehuda Sadeh dout("req=%p req->r_linger=%d\n", req, req->r_linger); 18793d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 18803d14c5d2SYehuda Sadeh return; 1881ff513aceSIlya Dryomov out_unlock: 1882ff513aceSIlya Dryomov mutex_unlock(&osdc->request_mutex); 1883ff513aceSIlya Dryomov up_read(&osdc->map_sem); 1884ff513aceSIlya Dryomov goto out; 18853d14c5d2SYehuda Sadeh 18861b83bef2SSage Weil bad_put: 188737c89bdeSLi Wang req->r_result = -EIO; 188837c89bdeSLi Wang __unregister_request(osdc, req); 188937c89bdeSLi Wang if (req->r_callback) 189037c89bdeSLi Wang req->r_callback(req, msg); 189137c89bdeSLi Wang else 189237c89bdeSLi Wang complete_all(&req->r_completion); 189337c89bdeSLi Wang complete_request(req); 18941b83bef2SSage Weil ceph_osdc_put_request(req); 18958058fd45SAlex Elder bad_mutex: 18968058fd45SAlex Elder mutex_unlock(&osdc->request_mutex); 1897ff513aceSIlya Dryomov up_read(&osdc->map_sem); 18983d14c5d2SYehuda Sadeh bad: 18991b83bef2SSage Weil pr_err("corrupt osd_op_reply got %d %d\n", 19001b83bef2SSage Weil (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len)); 19013d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 19023d14c5d2SYehuda Sadeh } 19033d14c5d2SYehuda Sadeh 19046f6c7006SSage Weil static void reset_changed_osds(struct ceph_osd_client *osdc) 19053d14c5d2SYehuda Sadeh { 19063d14c5d2SYehuda Sadeh struct rb_node *p, *n; 19073d14c5d2SYehuda Sadeh 19083d14c5d2SYehuda Sadeh for (p = rb_first(&osdc->osds); p; p = n) { 19096f6c7006SSage Weil struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node); 19103d14c5d2SYehuda Sadeh 19113d14c5d2SYehuda Sadeh n = rb_next(p); 19123d14c5d2SYehuda Sadeh if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) || 19133d14c5d2SYehuda Sadeh memcmp(&osd->o_con.peer_addr, 19143d14c5d2SYehuda Sadeh ceph_osd_addr(osdc->osdmap, 19153d14c5d2SYehuda Sadeh osd->o_osd), 19163d14c5d2SYehuda Sadeh sizeof(struct ceph_entity_addr)) != 0) 19173d14c5d2SYehuda Sadeh __reset_osd(osdc, osd); 19183d14c5d2SYehuda Sadeh } 19193d14c5d2SYehuda Sadeh } 19203d14c5d2SYehuda Sadeh 19213d14c5d2SYehuda Sadeh /* 19226f6c7006SSage Weil * Requeue requests whose mapping to an OSD has changed. If requests map to 19236f6c7006SSage Weil * no osd, request a new map. 19243d14c5d2SYehuda Sadeh * 1925e6d50f67SAlex Elder * Caller should hold map_sem for read. 19263d14c5d2SYehuda Sadeh */ 19279a1ea2dbSJosh Durgin static void kick_requests(struct ceph_osd_client *osdc, bool force_resend, 19289a1ea2dbSJosh Durgin bool force_resend_writes) 19293d14c5d2SYehuda Sadeh { 1930a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 19316f6c7006SSage Weil struct rb_node *p; 19326f6c7006SSage Weil int needmap = 0; 19336f6c7006SSage Weil int err; 19349a1ea2dbSJosh Durgin bool force_resend_req; 19353d14c5d2SYehuda Sadeh 19369a1ea2dbSJosh Durgin dout("kick_requests %s %s\n", force_resend ? " (force resend)" : "", 19379a1ea2dbSJosh Durgin force_resend_writes ? " (force resend writes)" : ""); 19383d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 19396194ea89SSage Weil for (p = rb_first(&osdc->requests); p; ) { 19406f6c7006SSage Weil req = rb_entry(p, struct ceph_osd_request, r_node); 19416194ea89SSage Weil p = rb_next(p); 1942ab60b16dSAlex Elder 1943ab60b16dSAlex Elder /* 1944ab60b16dSAlex Elder * For linger requests that have not yet been 1945ab60b16dSAlex Elder * registered, move them to the linger list; they'll 1946ab60b16dSAlex Elder * be sent to the osd in the loop below. Unregister 1947ab60b16dSAlex Elder * the request before re-registering it as a linger 1948ab60b16dSAlex Elder * request to ensure the __map_request() below 1949ab60b16dSAlex Elder * will decide it needs to be sent. 1950ab60b16dSAlex Elder */ 1951ab60b16dSAlex Elder if (req->r_linger && list_empty(&req->r_linger_item)) { 1952ab60b16dSAlex Elder dout("%p tid %llu restart on osd%d\n", 1953ab60b16dSAlex Elder req, req->r_tid, 1954ab60b16dSAlex Elder req->r_osd ? req->r_osd->o_osd : -1); 195596e4dac6SAlex Elder ceph_osdc_get_request(req); 1956ab60b16dSAlex Elder __unregister_request(osdc, req); 1957ab60b16dSAlex Elder __register_linger_request(osdc, req); 195896e4dac6SAlex Elder ceph_osdc_put_request(req); 1959ab60b16dSAlex Elder continue; 1960ab60b16dSAlex Elder } 1961ab60b16dSAlex Elder 19629a1ea2dbSJosh Durgin force_resend_req = force_resend || 19639a1ea2dbSJosh Durgin (force_resend_writes && 19649a1ea2dbSJosh Durgin req->r_flags & CEPH_OSD_FLAG_WRITE); 19659a1ea2dbSJosh Durgin err = __map_request(osdc, req, force_resend_req); 19666f6c7006SSage Weil if (err < 0) 19676f6c7006SSage Weil continue; /* error */ 19686f6c7006SSage Weil if (req->r_osd == NULL) { 19696f6c7006SSage Weil dout("%p tid %llu maps to no osd\n", req, req->r_tid); 19706f6c7006SSage Weil needmap++; /* request a newer map */ 19716f6c7006SSage Weil } else if (err > 0) { 19726194ea89SSage Weil if (!req->r_linger) { 19736194ea89SSage Weil dout("%p tid %llu requeued on osd%d\n", req, 19746194ea89SSage Weil req->r_tid, 19756f6c7006SSage Weil req->r_osd ? req->r_osd->o_osd : -1); 19766f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 19776f6c7006SSage Weil } 19786f6c7006SSage Weil } 19796194ea89SSage Weil } 1980a40c4f10SYehuda Sadeh 1981a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osdc->req_linger, 1982a40c4f10SYehuda Sadeh r_linger_item) { 1983a40c4f10SYehuda Sadeh dout("linger req=%p req->r_osd=%p\n", req, req->r_osd); 1984a40c4f10SYehuda Sadeh 19859a1ea2dbSJosh Durgin err = __map_request(osdc, req, 19869a1ea2dbSJosh Durgin force_resend || force_resend_writes); 1987ab60b16dSAlex Elder dout("__map_request returned %d\n", err); 1988a40c4f10SYehuda Sadeh if (err == 0) 1989a40c4f10SYehuda Sadeh continue; /* no change and no osd was specified */ 1990a40c4f10SYehuda Sadeh if (err < 0) 1991a40c4f10SYehuda Sadeh continue; /* hrm! */ 1992a40c4f10SYehuda Sadeh if (req->r_osd == NULL) { 1993a40c4f10SYehuda Sadeh dout("tid %llu maps to no valid osd\n", req->r_tid); 1994a40c4f10SYehuda Sadeh needmap++; /* request a newer map */ 1995a40c4f10SYehuda Sadeh continue; 1996a40c4f10SYehuda Sadeh } 1997a40c4f10SYehuda Sadeh 1998a40c4f10SYehuda Sadeh dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid, 1999a40c4f10SYehuda Sadeh req->r_osd ? req->r_osd->o_osd : -1); 2000a40c4f10SYehuda Sadeh __register_request(osdc, req); 2001c89ce05eSAlex Elder __unregister_linger_request(osdc, req); 2002a40c4f10SYehuda Sadeh } 200314d2f38dSAlex Elder reset_changed_osds(osdc); 20043d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 20053d14c5d2SYehuda Sadeh 20063d14c5d2SYehuda Sadeh if (needmap) { 20073d14c5d2SYehuda Sadeh dout("%d requests for down osds, need new map\n", needmap); 20083d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 20093d14c5d2SYehuda Sadeh } 20103d14c5d2SYehuda Sadeh } 20116f6c7006SSage Weil 20126f6c7006SSage Weil 20133d14c5d2SYehuda Sadeh /* 20143d14c5d2SYehuda Sadeh * Process updated osd map. 20153d14c5d2SYehuda Sadeh * 20163d14c5d2SYehuda Sadeh * The message contains any number of incremental and full maps, normally 20173d14c5d2SYehuda Sadeh * indicating some sort of topology change in the cluster. Kick requests 20183d14c5d2SYehuda Sadeh * off to different OSDs as needed. 20193d14c5d2SYehuda Sadeh */ 20203d14c5d2SYehuda Sadeh void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg) 20213d14c5d2SYehuda Sadeh { 20223d14c5d2SYehuda Sadeh void *p, *end, *next; 20233d14c5d2SYehuda Sadeh u32 nr_maps, maplen; 20243d14c5d2SYehuda Sadeh u32 epoch; 20253d14c5d2SYehuda Sadeh struct ceph_osdmap *newmap = NULL, *oldmap; 20263d14c5d2SYehuda Sadeh int err; 20273d14c5d2SYehuda Sadeh struct ceph_fsid fsid; 20289a1ea2dbSJosh Durgin bool was_full; 20293d14c5d2SYehuda Sadeh 20303d14c5d2SYehuda Sadeh dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0); 20313d14c5d2SYehuda Sadeh p = msg->front.iov_base; 20323d14c5d2SYehuda Sadeh end = p + msg->front.iov_len; 20333d14c5d2SYehuda Sadeh 20343d14c5d2SYehuda Sadeh /* verify fsid */ 20353d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, sizeof(fsid), bad); 20363d14c5d2SYehuda Sadeh ceph_decode_copy(&p, &fsid, sizeof(fsid)); 20373d14c5d2SYehuda Sadeh if (ceph_check_fsid(osdc->client, &fsid) < 0) 20383d14c5d2SYehuda Sadeh return; 20393d14c5d2SYehuda Sadeh 20403d14c5d2SYehuda Sadeh down_write(&osdc->map_sem); 20413d14c5d2SYehuda Sadeh 20429a1ea2dbSJosh Durgin was_full = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL); 20439a1ea2dbSJosh Durgin 20443d14c5d2SYehuda Sadeh /* incremental maps */ 20453d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 20463d14c5d2SYehuda Sadeh dout(" %d inc maps\n", nr_maps); 20473d14c5d2SYehuda Sadeh while (nr_maps > 0) { 20483d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 20493d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 20503d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 20513d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 20523d14c5d2SYehuda Sadeh next = p + maplen; 20533d14c5d2SYehuda Sadeh if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) { 20543d14c5d2SYehuda Sadeh dout("applying incremental map %u len %d\n", 20553d14c5d2SYehuda Sadeh epoch, maplen); 20563d14c5d2SYehuda Sadeh newmap = osdmap_apply_incremental(&p, next, 20573d14c5d2SYehuda Sadeh osdc->osdmap, 205815d9882cSAlex Elder &osdc->client->msgr); 20593d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 20603d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 20613d14c5d2SYehuda Sadeh goto bad; 20623d14c5d2SYehuda Sadeh } 20633d14c5d2SYehuda Sadeh BUG_ON(!newmap); 20643d14c5d2SYehuda Sadeh if (newmap != osdc->osdmap) { 20653d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 20663d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 20673d14c5d2SYehuda Sadeh } 20689a1ea2dbSJosh Durgin was_full = was_full || 20699a1ea2dbSJosh Durgin ceph_osdmap_flag(osdc->osdmap, 20709a1ea2dbSJosh Durgin CEPH_OSDMAP_FULL); 20719a1ea2dbSJosh Durgin kick_requests(osdc, 0, was_full); 20723d14c5d2SYehuda Sadeh } else { 20733d14c5d2SYehuda Sadeh dout("ignoring incremental map %u len %d\n", 20743d14c5d2SYehuda Sadeh epoch, maplen); 20753d14c5d2SYehuda Sadeh } 20763d14c5d2SYehuda Sadeh p = next; 20773d14c5d2SYehuda Sadeh nr_maps--; 20783d14c5d2SYehuda Sadeh } 20793d14c5d2SYehuda Sadeh if (newmap) 20803d14c5d2SYehuda Sadeh goto done; 20813d14c5d2SYehuda Sadeh 20823d14c5d2SYehuda Sadeh /* full maps */ 20833d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 20843d14c5d2SYehuda Sadeh dout(" %d full maps\n", nr_maps); 20853d14c5d2SYehuda Sadeh while (nr_maps) { 20863d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 20873d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 20883d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 20893d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 20903d14c5d2SYehuda Sadeh if (nr_maps > 1) { 20913d14c5d2SYehuda Sadeh dout("skipping non-latest full map %u len %d\n", 20923d14c5d2SYehuda Sadeh epoch, maplen); 20933d14c5d2SYehuda Sadeh } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) { 20943d14c5d2SYehuda Sadeh dout("skipping full map %u len %d, " 20953d14c5d2SYehuda Sadeh "older than our %u\n", epoch, maplen, 20963d14c5d2SYehuda Sadeh osdc->osdmap->epoch); 20973d14c5d2SYehuda Sadeh } else { 209838d6453cSSage Weil int skipped_map = 0; 209938d6453cSSage Weil 21003d14c5d2SYehuda Sadeh dout("taking full map %u len %d\n", epoch, maplen); 2101a2505d63SIlya Dryomov newmap = ceph_osdmap_decode(&p, p+maplen); 21023d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 21033d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 21043d14c5d2SYehuda Sadeh goto bad; 21053d14c5d2SYehuda Sadeh } 21063d14c5d2SYehuda Sadeh BUG_ON(!newmap); 21073d14c5d2SYehuda Sadeh oldmap = osdc->osdmap; 21083d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 210938d6453cSSage Weil if (oldmap) { 211038d6453cSSage Weil if (oldmap->epoch + 1 < newmap->epoch) 211138d6453cSSage Weil skipped_map = 1; 21123d14c5d2SYehuda Sadeh ceph_osdmap_destroy(oldmap); 211338d6453cSSage Weil } 21149a1ea2dbSJosh Durgin was_full = was_full || 21159a1ea2dbSJosh Durgin ceph_osdmap_flag(osdc->osdmap, 21169a1ea2dbSJosh Durgin CEPH_OSDMAP_FULL); 21179a1ea2dbSJosh Durgin kick_requests(osdc, skipped_map, was_full); 21183d14c5d2SYehuda Sadeh } 21193d14c5d2SYehuda Sadeh p += maplen; 21203d14c5d2SYehuda Sadeh nr_maps--; 21213d14c5d2SYehuda Sadeh } 21223d14c5d2SYehuda Sadeh 2123b72e19b9SDan Carpenter if (!osdc->osdmap) 2124b72e19b9SDan Carpenter goto bad; 21253d14c5d2SYehuda Sadeh done: 21263d14c5d2SYehuda Sadeh downgrade_write(&osdc->map_sem); 21273d14c5d2SYehuda Sadeh ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch); 2128cd634fb6SSage Weil 2129cd634fb6SSage Weil /* 2130cd634fb6SSage Weil * subscribe to subsequent osdmap updates if full to ensure 2131cd634fb6SSage Weil * we find out when we are no longer full and stop returning 2132cd634fb6SSage Weil * ENOSPC. 2133cd634fb6SSage Weil */ 2134d29adb34SJosh Durgin if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) || 2135d29adb34SJosh Durgin ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD) || 2136d29adb34SJosh Durgin ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR)) 2137cd634fb6SSage Weil ceph_monc_request_next_osdmap(&osdc->client->monc); 2138cd634fb6SSage Weil 2139f9d25199SAlex Elder mutex_lock(&osdc->request_mutex); 2140f9d25199SAlex Elder __send_queued(osdc); 2141f9d25199SAlex Elder mutex_unlock(&osdc->request_mutex); 21423d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 21433d14c5d2SYehuda Sadeh wake_up_all(&osdc->client->auth_wq); 21443d14c5d2SYehuda Sadeh return; 21453d14c5d2SYehuda Sadeh 21463d14c5d2SYehuda Sadeh bad: 21473d14c5d2SYehuda Sadeh pr_err("osdc handle_map corrupt msg\n"); 21483d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 21493d14c5d2SYehuda Sadeh up_write(&osdc->map_sem); 21503d14c5d2SYehuda Sadeh } 21513d14c5d2SYehuda Sadeh 21523d14c5d2SYehuda Sadeh /* 2153a40c4f10SYehuda Sadeh * watch/notify callback event infrastructure 2154a40c4f10SYehuda Sadeh * 2155a40c4f10SYehuda Sadeh * These callbacks are used both for watch and notify operations. 2156a40c4f10SYehuda Sadeh */ 2157a40c4f10SYehuda Sadeh static void __release_event(struct kref *kref) 2158a40c4f10SYehuda Sadeh { 2159a40c4f10SYehuda Sadeh struct ceph_osd_event *event = 2160a40c4f10SYehuda Sadeh container_of(kref, struct ceph_osd_event, kref); 2161a40c4f10SYehuda Sadeh 2162a40c4f10SYehuda Sadeh dout("__release_event %p\n", event); 2163a40c4f10SYehuda Sadeh kfree(event); 2164a40c4f10SYehuda Sadeh } 2165a40c4f10SYehuda Sadeh 2166a40c4f10SYehuda Sadeh static void get_event(struct ceph_osd_event *event) 2167a40c4f10SYehuda Sadeh { 2168a40c4f10SYehuda Sadeh kref_get(&event->kref); 2169a40c4f10SYehuda Sadeh } 2170a40c4f10SYehuda Sadeh 2171a40c4f10SYehuda Sadeh void ceph_osdc_put_event(struct ceph_osd_event *event) 2172a40c4f10SYehuda Sadeh { 2173a40c4f10SYehuda Sadeh kref_put(&event->kref, __release_event); 2174a40c4f10SYehuda Sadeh } 2175a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_put_event); 2176a40c4f10SYehuda Sadeh 2177a40c4f10SYehuda Sadeh static void __insert_event(struct ceph_osd_client *osdc, 2178a40c4f10SYehuda Sadeh struct ceph_osd_event *new) 2179a40c4f10SYehuda Sadeh { 2180a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 2181a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 2182a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 2183a40c4f10SYehuda Sadeh 2184a40c4f10SYehuda Sadeh while (*p) { 2185a40c4f10SYehuda Sadeh parent = *p; 2186a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 2187a40c4f10SYehuda Sadeh if (new->cookie < event->cookie) 2188a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 2189a40c4f10SYehuda Sadeh else if (new->cookie > event->cookie) 2190a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 2191a40c4f10SYehuda Sadeh else 2192a40c4f10SYehuda Sadeh BUG(); 2193a40c4f10SYehuda Sadeh } 2194a40c4f10SYehuda Sadeh 2195a40c4f10SYehuda Sadeh rb_link_node(&new->node, parent, p); 2196a40c4f10SYehuda Sadeh rb_insert_color(&new->node, &osdc->event_tree); 2197a40c4f10SYehuda Sadeh } 2198a40c4f10SYehuda Sadeh 2199a40c4f10SYehuda Sadeh static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc, 2200a40c4f10SYehuda Sadeh u64 cookie) 2201a40c4f10SYehuda Sadeh { 2202a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 2203a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 2204a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 2205a40c4f10SYehuda Sadeh 2206a40c4f10SYehuda Sadeh while (*p) { 2207a40c4f10SYehuda Sadeh parent = *p; 2208a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 2209a40c4f10SYehuda Sadeh if (cookie < event->cookie) 2210a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 2211a40c4f10SYehuda Sadeh else if (cookie > event->cookie) 2212a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 2213a40c4f10SYehuda Sadeh else 2214a40c4f10SYehuda Sadeh return event; 2215a40c4f10SYehuda Sadeh } 2216a40c4f10SYehuda Sadeh return NULL; 2217a40c4f10SYehuda Sadeh } 2218a40c4f10SYehuda Sadeh 2219a40c4f10SYehuda Sadeh static void __remove_event(struct ceph_osd_event *event) 2220a40c4f10SYehuda Sadeh { 2221a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 2222a40c4f10SYehuda Sadeh 2223a40c4f10SYehuda Sadeh if (!RB_EMPTY_NODE(&event->node)) { 2224a40c4f10SYehuda Sadeh dout("__remove_event removed %p\n", event); 2225a40c4f10SYehuda Sadeh rb_erase(&event->node, &osdc->event_tree); 2226a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 2227a40c4f10SYehuda Sadeh } else { 2228a40c4f10SYehuda Sadeh dout("__remove_event didn't remove %p\n", event); 2229a40c4f10SYehuda Sadeh } 2230a40c4f10SYehuda Sadeh } 2231a40c4f10SYehuda Sadeh 2232a40c4f10SYehuda Sadeh int ceph_osdc_create_event(struct ceph_osd_client *osdc, 2233a40c4f10SYehuda Sadeh void (*event_cb)(u64, u64, u8, void *), 22343c663bbdSAlex Elder void *data, struct ceph_osd_event **pevent) 2235a40c4f10SYehuda Sadeh { 2236a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 2237a40c4f10SYehuda Sadeh 2238a40c4f10SYehuda Sadeh event = kmalloc(sizeof(*event), GFP_NOIO); 2239a40c4f10SYehuda Sadeh if (!event) 2240a40c4f10SYehuda Sadeh return -ENOMEM; 2241a40c4f10SYehuda Sadeh 2242a40c4f10SYehuda Sadeh dout("create_event %p\n", event); 2243a40c4f10SYehuda Sadeh event->cb = event_cb; 22443c663bbdSAlex Elder event->one_shot = 0; 2245a40c4f10SYehuda Sadeh event->data = data; 2246a40c4f10SYehuda Sadeh event->osdc = osdc; 2247a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&event->osd_node); 22483ee5234dSAlex Elder RB_CLEAR_NODE(&event->node); 2249a40c4f10SYehuda Sadeh kref_init(&event->kref); /* one ref for us */ 2250a40c4f10SYehuda Sadeh kref_get(&event->kref); /* one ref for the caller */ 2251a40c4f10SYehuda Sadeh 2252a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 2253a40c4f10SYehuda Sadeh event->cookie = ++osdc->event_count; 2254a40c4f10SYehuda Sadeh __insert_event(osdc, event); 2255a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 2256a40c4f10SYehuda Sadeh 2257a40c4f10SYehuda Sadeh *pevent = event; 2258a40c4f10SYehuda Sadeh return 0; 2259a40c4f10SYehuda Sadeh } 2260a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_create_event); 2261a40c4f10SYehuda Sadeh 2262a40c4f10SYehuda Sadeh void ceph_osdc_cancel_event(struct ceph_osd_event *event) 2263a40c4f10SYehuda Sadeh { 2264a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 2265a40c4f10SYehuda Sadeh 2266a40c4f10SYehuda Sadeh dout("cancel_event %p\n", event); 2267a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 2268a40c4f10SYehuda Sadeh __remove_event(event); 2269a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 2270a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); /* caller's */ 2271a40c4f10SYehuda Sadeh } 2272a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_cancel_event); 2273a40c4f10SYehuda Sadeh 2274a40c4f10SYehuda Sadeh 2275a40c4f10SYehuda Sadeh static void do_event_work(struct work_struct *work) 2276a40c4f10SYehuda Sadeh { 2277a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work = 2278a40c4f10SYehuda Sadeh container_of(work, struct ceph_osd_event_work, work); 2279a40c4f10SYehuda Sadeh struct ceph_osd_event *event = event_work->event; 2280a40c4f10SYehuda Sadeh u64 ver = event_work->ver; 2281a40c4f10SYehuda Sadeh u64 notify_id = event_work->notify_id; 2282a40c4f10SYehuda Sadeh u8 opcode = event_work->opcode; 2283a40c4f10SYehuda Sadeh 2284a40c4f10SYehuda Sadeh dout("do_event_work completing %p\n", event); 2285a40c4f10SYehuda Sadeh event->cb(ver, notify_id, opcode, event->data); 2286a40c4f10SYehuda Sadeh dout("do_event_work completed %p\n", event); 2287a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 2288a40c4f10SYehuda Sadeh kfree(event_work); 2289a40c4f10SYehuda Sadeh } 2290a40c4f10SYehuda Sadeh 2291a40c4f10SYehuda Sadeh 2292a40c4f10SYehuda Sadeh /* 2293a40c4f10SYehuda Sadeh * Process osd watch notifications 2294a40c4f10SYehuda Sadeh */ 22953c663bbdSAlex Elder static void handle_watch_notify(struct ceph_osd_client *osdc, 22963c663bbdSAlex Elder struct ceph_msg *msg) 2297a40c4f10SYehuda Sadeh { 2298a40c4f10SYehuda Sadeh void *p, *end; 2299a40c4f10SYehuda Sadeh u8 proto_ver; 2300a40c4f10SYehuda Sadeh u64 cookie, ver, notify_id; 2301a40c4f10SYehuda Sadeh u8 opcode; 2302a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 2303a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work; 2304a40c4f10SYehuda Sadeh 2305a40c4f10SYehuda Sadeh p = msg->front.iov_base; 2306a40c4f10SYehuda Sadeh end = p + msg->front.iov_len; 2307a40c4f10SYehuda Sadeh 2308a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, proto_ver, bad); 2309a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, opcode, bad); 2310a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, cookie, bad); 2311a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, ver, bad); 2312a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, notify_id, bad); 2313a40c4f10SYehuda Sadeh 2314a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 2315a40c4f10SYehuda Sadeh event = __find_event(osdc, cookie); 2316a40c4f10SYehuda Sadeh if (event) { 23173c663bbdSAlex Elder BUG_ON(event->one_shot); 2318a40c4f10SYehuda Sadeh get_event(event); 2319a40c4f10SYehuda Sadeh } 2320a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 2321a40c4f10SYehuda Sadeh dout("handle_watch_notify cookie %lld ver %lld event %p\n", 2322a40c4f10SYehuda Sadeh cookie, ver, event); 2323a40c4f10SYehuda Sadeh if (event) { 2324a40c4f10SYehuda Sadeh event_work = kmalloc(sizeof(*event_work), GFP_NOIO); 2325a40c4f10SYehuda Sadeh if (!event_work) { 2326a40c4f10SYehuda Sadeh dout("ERROR: could not allocate event_work\n"); 2327a40c4f10SYehuda Sadeh goto done_err; 2328a40c4f10SYehuda Sadeh } 23296b0ae409SMariusz Kozlowski INIT_WORK(&event_work->work, do_event_work); 2330a40c4f10SYehuda Sadeh event_work->event = event; 2331a40c4f10SYehuda Sadeh event_work->ver = ver; 2332a40c4f10SYehuda Sadeh event_work->notify_id = notify_id; 2333a40c4f10SYehuda Sadeh event_work->opcode = opcode; 2334a40c4f10SYehuda Sadeh if (!queue_work(osdc->notify_wq, &event_work->work)) { 2335a40c4f10SYehuda Sadeh dout("WARNING: failed to queue notify event work\n"); 2336a40c4f10SYehuda Sadeh goto done_err; 2337a40c4f10SYehuda Sadeh } 2338a40c4f10SYehuda Sadeh } 2339a40c4f10SYehuda Sadeh 2340a40c4f10SYehuda Sadeh return; 2341a40c4f10SYehuda Sadeh 2342a40c4f10SYehuda Sadeh done_err: 2343a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 2344a40c4f10SYehuda Sadeh return; 2345a40c4f10SYehuda Sadeh 2346a40c4f10SYehuda Sadeh bad: 2347a40c4f10SYehuda Sadeh pr_err("osdc handle_watch_notify corrupt msg\n"); 2348a40c4f10SYehuda Sadeh } 2349a40c4f10SYehuda Sadeh 2350a40c4f10SYehuda Sadeh /* 2351e65550fdSAlex Elder * build new request AND message 2352e65550fdSAlex Elder * 2353e65550fdSAlex Elder */ 2354e65550fdSAlex Elder void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off, 2355e65550fdSAlex Elder struct ceph_snap_context *snapc, u64 snap_id, 2356e65550fdSAlex Elder struct timespec *mtime) 2357e65550fdSAlex Elder { 2358e65550fdSAlex Elder struct ceph_msg *msg = req->r_request; 2359e65550fdSAlex Elder void *p; 2360e65550fdSAlex Elder size_t msg_size; 2361e65550fdSAlex Elder int flags = req->r_flags; 2362e65550fdSAlex Elder u64 data_len; 2363e65550fdSAlex Elder unsigned int i; 2364e65550fdSAlex Elder 2365e65550fdSAlex Elder req->r_snapid = snap_id; 2366e65550fdSAlex Elder req->r_snapc = ceph_get_snap_context(snapc); 2367e65550fdSAlex Elder 2368e65550fdSAlex Elder /* encode request */ 2369e65550fdSAlex Elder msg->hdr.version = cpu_to_le16(4); 2370e65550fdSAlex Elder 2371e65550fdSAlex Elder p = msg->front.iov_base; 2372e65550fdSAlex Elder ceph_encode_32(&p, 1); /* client_inc is always 1 */ 2373e65550fdSAlex Elder req->r_request_osdmap_epoch = p; 2374e65550fdSAlex Elder p += 4; 2375e65550fdSAlex Elder req->r_request_flags = p; 2376e65550fdSAlex Elder p += 4; 2377e65550fdSAlex Elder if (req->r_flags & CEPH_OSD_FLAG_WRITE) 2378e65550fdSAlex Elder ceph_encode_timespec(p, mtime); 2379e65550fdSAlex Elder p += sizeof(struct ceph_timespec); 2380e65550fdSAlex Elder req->r_request_reassert_version = p; 2381e65550fdSAlex Elder p += sizeof(struct ceph_eversion); /* will get filled in */ 2382e65550fdSAlex Elder 2383e65550fdSAlex Elder /* oloc */ 2384e65550fdSAlex Elder ceph_encode_8(&p, 4); 2385e65550fdSAlex Elder ceph_encode_8(&p, 4); 2386e65550fdSAlex Elder ceph_encode_32(&p, 8 + 4 + 4); 2387e65550fdSAlex Elder req->r_request_pool = p; 2388e65550fdSAlex Elder p += 8; 2389e65550fdSAlex Elder ceph_encode_32(&p, -1); /* preferred */ 2390e65550fdSAlex Elder ceph_encode_32(&p, 0); /* key len */ 2391e65550fdSAlex Elder 2392e65550fdSAlex Elder ceph_encode_8(&p, 1); 2393e65550fdSAlex Elder req->r_request_pgid = p; 2394e65550fdSAlex Elder p += 8 + 4; 2395e65550fdSAlex Elder ceph_encode_32(&p, -1); /* preferred */ 2396e65550fdSAlex Elder 2397e65550fdSAlex Elder /* oid */ 23983c972c95SIlya Dryomov ceph_encode_32(&p, req->r_base_oid.name_len); 23993c972c95SIlya Dryomov memcpy(p, req->r_base_oid.name, req->r_base_oid.name_len); 24003c972c95SIlya Dryomov dout("oid '%.*s' len %d\n", req->r_base_oid.name_len, 24013c972c95SIlya Dryomov req->r_base_oid.name, req->r_base_oid.name_len); 24023c972c95SIlya Dryomov p += req->r_base_oid.name_len; 2403e65550fdSAlex Elder 2404e65550fdSAlex Elder /* ops--can imply data */ 2405e65550fdSAlex Elder ceph_encode_16(&p, (u16)req->r_num_ops); 2406e65550fdSAlex Elder data_len = 0; 2407e65550fdSAlex Elder for (i = 0; i < req->r_num_ops; i++) { 2408e65550fdSAlex Elder data_len += osd_req_encode_op(req, p, i); 2409e65550fdSAlex Elder p += sizeof(struct ceph_osd_op); 2410e65550fdSAlex Elder } 2411e65550fdSAlex Elder 2412e65550fdSAlex Elder /* snaps */ 2413e65550fdSAlex Elder ceph_encode_64(&p, req->r_snapid); 2414e65550fdSAlex Elder ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0); 2415e65550fdSAlex Elder ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0); 2416e65550fdSAlex Elder if (req->r_snapc) { 2417e65550fdSAlex Elder for (i = 0; i < snapc->num_snaps; i++) { 2418e65550fdSAlex Elder ceph_encode_64(&p, req->r_snapc->snaps[i]); 2419e65550fdSAlex Elder } 2420e65550fdSAlex Elder } 2421e65550fdSAlex Elder 2422e65550fdSAlex Elder req->r_request_attempts = p; 2423e65550fdSAlex Elder p += 4; 2424e65550fdSAlex Elder 2425e65550fdSAlex Elder /* data */ 2426e65550fdSAlex Elder if (flags & CEPH_OSD_FLAG_WRITE) { 2427e65550fdSAlex Elder u16 data_off; 2428e65550fdSAlex Elder 2429e65550fdSAlex Elder /* 2430e65550fdSAlex Elder * The header "data_off" is a hint to the receiver 2431e65550fdSAlex Elder * allowing it to align received data into its 2432e65550fdSAlex Elder * buffers such that there's no need to re-copy 2433e65550fdSAlex Elder * it before writing it to disk (direct I/O). 2434e65550fdSAlex Elder */ 2435e65550fdSAlex Elder data_off = (u16) (off & 0xffff); 2436e65550fdSAlex Elder req->r_request->hdr.data_off = cpu_to_le16(data_off); 2437e65550fdSAlex Elder } 2438e65550fdSAlex Elder req->r_request->hdr.data_len = cpu_to_le32(data_len); 2439e65550fdSAlex Elder 2440e65550fdSAlex Elder BUG_ON(p > msg->front.iov_base + msg->front.iov_len); 2441e65550fdSAlex Elder msg_size = p - msg->front.iov_base; 2442e65550fdSAlex Elder msg->front.iov_len = msg_size; 2443e65550fdSAlex Elder msg->hdr.front_len = cpu_to_le32(msg_size); 2444e65550fdSAlex Elder 2445e65550fdSAlex Elder dout("build_request msg_size was %d\n", (int)msg_size); 2446e65550fdSAlex Elder } 2447e65550fdSAlex Elder EXPORT_SYMBOL(ceph_osdc_build_request); 2448e65550fdSAlex Elder 2449e65550fdSAlex Elder /* 24503d14c5d2SYehuda Sadeh * Register request, send initial attempt. 24513d14c5d2SYehuda Sadeh */ 24523d14c5d2SYehuda Sadeh int ceph_osdc_start_request(struct ceph_osd_client *osdc, 24533d14c5d2SYehuda Sadeh struct ceph_osd_request *req, 24543d14c5d2SYehuda Sadeh bool nofail) 24553d14c5d2SYehuda Sadeh { 24560bbfdfe8SIlya Dryomov int rc; 24573d14c5d2SYehuda Sadeh 24583d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 24593d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 24600bbfdfe8SIlya Dryomov 24610bbfdfe8SIlya Dryomov rc = __ceph_osdc_start_request(osdc, req, nofail); 24620bbfdfe8SIlya Dryomov 24633d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 24643d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 24650bbfdfe8SIlya Dryomov 24663d14c5d2SYehuda Sadeh return rc; 24673d14c5d2SYehuda Sadeh } 24683d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_start_request); 24693d14c5d2SYehuda Sadeh 24703d14c5d2SYehuda Sadeh /* 2471c9f9b93dSIlya Dryomov * Unregister a registered request. The request is not completed (i.e. 2472c9f9b93dSIlya Dryomov * no callbacks or wakeups) - higher layers are supposed to know what 2473c9f9b93dSIlya Dryomov * they are canceling. 2474c9f9b93dSIlya Dryomov */ 2475c9f9b93dSIlya Dryomov void ceph_osdc_cancel_request(struct ceph_osd_request *req) 2476c9f9b93dSIlya Dryomov { 2477c9f9b93dSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 2478c9f9b93dSIlya Dryomov 2479c9f9b93dSIlya Dryomov mutex_lock(&osdc->request_mutex); 2480c9f9b93dSIlya Dryomov if (req->r_linger) 2481c9f9b93dSIlya Dryomov __unregister_linger_request(osdc, req); 2482c9f9b93dSIlya Dryomov __unregister_request(osdc, req); 2483c9f9b93dSIlya Dryomov mutex_unlock(&osdc->request_mutex); 2484c9f9b93dSIlya Dryomov 2485c9f9b93dSIlya Dryomov dout("%s %p tid %llu canceled\n", __func__, req, req->r_tid); 2486c9f9b93dSIlya Dryomov } 2487c9f9b93dSIlya Dryomov EXPORT_SYMBOL(ceph_osdc_cancel_request); 2488c9f9b93dSIlya Dryomov 2489c9f9b93dSIlya Dryomov /* 24903d14c5d2SYehuda Sadeh * wait for a request to complete 24913d14c5d2SYehuda Sadeh */ 24923d14c5d2SYehuda Sadeh int ceph_osdc_wait_request(struct ceph_osd_client *osdc, 24933d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 24943d14c5d2SYehuda Sadeh { 24953d14c5d2SYehuda Sadeh int rc; 24963d14c5d2SYehuda Sadeh 2497c9f9b93dSIlya Dryomov dout("%s %p tid %llu\n", __func__, req, req->r_tid); 2498c9f9b93dSIlya Dryomov 24993d14c5d2SYehuda Sadeh rc = wait_for_completion_interruptible(&req->r_completion); 25003d14c5d2SYehuda Sadeh if (rc < 0) { 2501c9f9b93dSIlya Dryomov dout("%s %p tid %llu interrupted\n", __func__, req, req->r_tid); 2502c9f9b93dSIlya Dryomov ceph_osdc_cancel_request(req); 250325845472SSage Weil complete_request(req); 25043d14c5d2SYehuda Sadeh return rc; 25053d14c5d2SYehuda Sadeh } 25063d14c5d2SYehuda Sadeh 2507c9f9b93dSIlya Dryomov dout("%s %p tid %llu result %d\n", __func__, req, req->r_tid, 2508c9f9b93dSIlya Dryomov req->r_result); 25093d14c5d2SYehuda Sadeh return req->r_result; 25103d14c5d2SYehuda Sadeh } 25113d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_wait_request); 25123d14c5d2SYehuda Sadeh 25133d14c5d2SYehuda Sadeh /* 25143d14c5d2SYehuda Sadeh * sync - wait for all in-flight requests to flush. avoid starvation. 25153d14c5d2SYehuda Sadeh */ 25163d14c5d2SYehuda Sadeh void ceph_osdc_sync(struct ceph_osd_client *osdc) 25173d14c5d2SYehuda Sadeh { 25183d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 25193d14c5d2SYehuda Sadeh u64 last_tid, next_tid = 0; 25203d14c5d2SYehuda Sadeh 25213d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 25223d14c5d2SYehuda Sadeh last_tid = osdc->last_tid; 25233d14c5d2SYehuda Sadeh while (1) { 25243d14c5d2SYehuda Sadeh req = __lookup_request_ge(osdc, next_tid); 25253d14c5d2SYehuda Sadeh if (!req) 25263d14c5d2SYehuda Sadeh break; 25273d14c5d2SYehuda Sadeh if (req->r_tid > last_tid) 25283d14c5d2SYehuda Sadeh break; 25293d14c5d2SYehuda Sadeh 25303d14c5d2SYehuda Sadeh next_tid = req->r_tid + 1; 25313d14c5d2SYehuda Sadeh if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0) 25323d14c5d2SYehuda Sadeh continue; 25333d14c5d2SYehuda Sadeh 25343d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 25353d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 25363d14c5d2SYehuda Sadeh dout("sync waiting on tid %llu (last is %llu)\n", 25373d14c5d2SYehuda Sadeh req->r_tid, last_tid); 25383d14c5d2SYehuda Sadeh wait_for_completion(&req->r_safe_completion); 25393d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 25403d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 25413d14c5d2SYehuda Sadeh } 25423d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 25433d14c5d2SYehuda Sadeh dout("sync done (thru tid %llu)\n", last_tid); 25443d14c5d2SYehuda Sadeh } 25453d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_sync); 25463d14c5d2SYehuda Sadeh 25473d14c5d2SYehuda Sadeh /* 2548dd935f44SJosh Durgin * Call all pending notify callbacks - for use after a watch is 2549dd935f44SJosh Durgin * unregistered, to make sure no more callbacks for it will be invoked 2550dd935f44SJosh Durgin */ 2551f6479449Sstephen hemminger void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc) 2552dd935f44SJosh Durgin { 2553dd935f44SJosh Durgin flush_workqueue(osdc->notify_wq); 2554dd935f44SJosh Durgin } 2555dd935f44SJosh Durgin EXPORT_SYMBOL(ceph_osdc_flush_notifies); 2556dd935f44SJosh Durgin 2557dd935f44SJosh Durgin 2558dd935f44SJosh Durgin /* 25593d14c5d2SYehuda Sadeh * init, shutdown 25603d14c5d2SYehuda Sadeh */ 25613d14c5d2SYehuda Sadeh int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client) 25623d14c5d2SYehuda Sadeh { 25633d14c5d2SYehuda Sadeh int err; 25643d14c5d2SYehuda Sadeh 25653d14c5d2SYehuda Sadeh dout("init\n"); 25663d14c5d2SYehuda Sadeh osdc->client = client; 25673d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 25683d14c5d2SYehuda Sadeh init_rwsem(&osdc->map_sem); 25693d14c5d2SYehuda Sadeh init_completion(&osdc->map_waiters); 25703d14c5d2SYehuda Sadeh osdc->last_requested_map = 0; 25713d14c5d2SYehuda Sadeh mutex_init(&osdc->request_mutex); 25723d14c5d2SYehuda Sadeh osdc->last_tid = 0; 25733d14c5d2SYehuda Sadeh osdc->osds = RB_ROOT; 25743d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->osd_lru); 25753d14c5d2SYehuda Sadeh osdc->requests = RB_ROOT; 25763d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_lru); 25776f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_unsent); 25786f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_notarget); 2579a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_linger); 25803d14c5d2SYehuda Sadeh osdc->num_requests = 0; 25813d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout); 25823d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout); 2583a40c4f10SYehuda Sadeh spin_lock_init(&osdc->event_lock); 2584a40c4f10SYehuda Sadeh osdc->event_tree = RB_ROOT; 2585a40c4f10SYehuda Sadeh osdc->event_count = 0; 25863d14c5d2SYehuda Sadeh 25873d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 25883d14c5d2SYehuda Sadeh round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ)); 25893d14c5d2SYehuda Sadeh 25903d14c5d2SYehuda Sadeh err = -ENOMEM; 25913d14c5d2SYehuda Sadeh osdc->req_mempool = mempool_create_kmalloc_pool(10, 25923d14c5d2SYehuda Sadeh sizeof(struct ceph_osd_request)); 25933d14c5d2SYehuda Sadeh if (!osdc->req_mempool) 25943d14c5d2SYehuda Sadeh goto out; 25953d14c5d2SYehuda Sadeh 2596d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP, 2597d50b409fSSage Weil OSD_OP_FRONT_LEN, 10, true, 25983d14c5d2SYehuda Sadeh "osd_op"); 25993d14c5d2SYehuda Sadeh if (err < 0) 26003d14c5d2SYehuda Sadeh goto out_mempool; 2601d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY, 26023d14c5d2SYehuda Sadeh OSD_OPREPLY_FRONT_LEN, 10, true, 26033d14c5d2SYehuda Sadeh "osd_op_reply"); 26043d14c5d2SYehuda Sadeh if (err < 0) 26053d14c5d2SYehuda Sadeh goto out_msgpool; 2606a40c4f10SYehuda Sadeh 2607dbcae088SDan Carpenter err = -ENOMEM; 2608a40c4f10SYehuda Sadeh osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify"); 2609dbcae088SDan Carpenter if (!osdc->notify_wq) 2610c172ec5cSIlya Dryomov goto out_msgpool_reply; 2611c172ec5cSIlya Dryomov 26123d14c5d2SYehuda Sadeh return 0; 26133d14c5d2SYehuda Sadeh 2614c172ec5cSIlya Dryomov out_msgpool_reply: 2615c172ec5cSIlya Dryomov ceph_msgpool_destroy(&osdc->msgpool_op_reply); 26163d14c5d2SYehuda Sadeh out_msgpool: 26173d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 26183d14c5d2SYehuda Sadeh out_mempool: 26193d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 26203d14c5d2SYehuda Sadeh out: 26213d14c5d2SYehuda Sadeh return err; 26223d14c5d2SYehuda Sadeh } 26233d14c5d2SYehuda Sadeh 26243d14c5d2SYehuda Sadeh void ceph_osdc_stop(struct ceph_osd_client *osdc) 26253d14c5d2SYehuda Sadeh { 2626a40c4f10SYehuda Sadeh flush_workqueue(osdc->notify_wq); 2627a40c4f10SYehuda Sadeh destroy_workqueue(osdc->notify_wq); 26283d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->timeout_work); 26293d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->osds_timeout_work); 26303d14c5d2SYehuda Sadeh if (osdc->osdmap) { 26313d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 26323d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 26333d14c5d2SYehuda Sadeh } 2634aca420bcSSage Weil remove_all_osds(osdc); 26353d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 26363d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 26373d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op_reply); 26383d14c5d2SYehuda Sadeh } 26393d14c5d2SYehuda Sadeh 26403d14c5d2SYehuda Sadeh /* 26413d14c5d2SYehuda Sadeh * Read some contiguous pages. If we cross a stripe boundary, shorten 26423d14c5d2SYehuda Sadeh * *plen. Return number of bytes read, or error. 26433d14c5d2SYehuda Sadeh */ 26443d14c5d2SYehuda Sadeh int ceph_osdc_readpages(struct ceph_osd_client *osdc, 26453d14c5d2SYehuda Sadeh struct ceph_vino vino, struct ceph_file_layout *layout, 26463d14c5d2SYehuda Sadeh u64 off, u64 *plen, 26473d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 2648b7495fc2SSage Weil struct page **pages, int num_pages, int page_align) 26493d14c5d2SYehuda Sadeh { 26503d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 26513d14c5d2SYehuda Sadeh int rc = 0; 26523d14c5d2SYehuda Sadeh 26533d14c5d2SYehuda Sadeh dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino, 26543d14c5d2SYehuda Sadeh vino.snap, off, *plen); 265579528734SAlex Elder req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 1, 26563d14c5d2SYehuda Sadeh CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ, 2657acead002SAlex Elder NULL, truncate_seq, truncate_size, 2658153e5167SAlex Elder false); 26596816282dSSage Weil if (IS_ERR(req)) 26606816282dSSage Weil return PTR_ERR(req); 26613d14c5d2SYehuda Sadeh 26623d14c5d2SYehuda Sadeh /* it may be a short read due to an object boundary */ 26630fff87ecSAlex Elder 2664406e2c9fSAlex Elder osd_req_op_extent_osd_data_pages(req, 0, 2665a4ce40a9SAlex Elder pages, *plen, page_align, false, false); 26663d14c5d2SYehuda Sadeh 2667e0c59487SAlex Elder dout("readpages final extent is %llu~%llu (%llu bytes align %d)\n", 266843bfe5deSAlex Elder off, *plen, *plen, page_align); 26693d14c5d2SYehuda Sadeh 267079528734SAlex Elder ceph_osdc_build_request(req, off, NULL, vino.snap, NULL); 267102ee07d3SAlex Elder 26723d14c5d2SYehuda Sadeh rc = ceph_osdc_start_request(osdc, req, false); 26733d14c5d2SYehuda Sadeh if (!rc) 26743d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 26753d14c5d2SYehuda Sadeh 26763d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 26773d14c5d2SYehuda Sadeh dout("readpages result %d\n", rc); 26783d14c5d2SYehuda Sadeh return rc; 26793d14c5d2SYehuda Sadeh } 26803d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_readpages); 26813d14c5d2SYehuda Sadeh 26823d14c5d2SYehuda Sadeh /* 26833d14c5d2SYehuda Sadeh * do a synchronous write on N pages 26843d14c5d2SYehuda Sadeh */ 26853d14c5d2SYehuda Sadeh int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino, 26863d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 26873d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 26883d14c5d2SYehuda Sadeh u64 off, u64 len, 26893d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 26903d14c5d2SYehuda Sadeh struct timespec *mtime, 269124808826SAlex Elder struct page **pages, int num_pages) 26923d14c5d2SYehuda Sadeh { 26933d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 26943d14c5d2SYehuda Sadeh int rc = 0; 2695b7495fc2SSage Weil int page_align = off & ~PAGE_MASK; 26963d14c5d2SYehuda Sadeh 2697acead002SAlex Elder BUG_ON(vino.snap != CEPH_NOSNAP); /* snapshots aren't writeable */ 269879528734SAlex Elder req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 1, 26993d14c5d2SYehuda Sadeh CEPH_OSD_OP_WRITE, 270024808826SAlex Elder CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE, 2701acead002SAlex Elder snapc, truncate_seq, truncate_size, 2702153e5167SAlex Elder true); 27036816282dSSage Weil if (IS_ERR(req)) 27046816282dSSage Weil return PTR_ERR(req); 27053d14c5d2SYehuda Sadeh 27063d14c5d2SYehuda Sadeh /* it may be a short write due to an object boundary */ 2707406e2c9fSAlex Elder osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align, 270843bfe5deSAlex Elder false, false); 270943bfe5deSAlex Elder dout("writepages %llu~%llu (%llu bytes)\n", off, len, len); 27103d14c5d2SYehuda Sadeh 271179528734SAlex Elder ceph_osdc_build_request(req, off, snapc, CEPH_NOSNAP, mtime); 271202ee07d3SAlex Elder 271387f979d3SAlex Elder rc = ceph_osdc_start_request(osdc, req, true); 27143d14c5d2SYehuda Sadeh if (!rc) 27153d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 27163d14c5d2SYehuda Sadeh 27173d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 27183d14c5d2SYehuda Sadeh if (rc == 0) 27193d14c5d2SYehuda Sadeh rc = len; 27203d14c5d2SYehuda Sadeh dout("writepages result %d\n", rc); 27213d14c5d2SYehuda Sadeh return rc; 27223d14c5d2SYehuda Sadeh } 27233d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_writepages); 27243d14c5d2SYehuda Sadeh 27255522ae0bSAlex Elder int ceph_osdc_setup(void) 27265522ae0bSAlex Elder { 27275522ae0bSAlex Elder BUG_ON(ceph_osd_request_cache); 27285522ae0bSAlex Elder ceph_osd_request_cache = kmem_cache_create("ceph_osd_request", 27295522ae0bSAlex Elder sizeof (struct ceph_osd_request), 27305522ae0bSAlex Elder __alignof__(struct ceph_osd_request), 27315522ae0bSAlex Elder 0, NULL); 27325522ae0bSAlex Elder 27335522ae0bSAlex Elder return ceph_osd_request_cache ? 0 : -ENOMEM; 27345522ae0bSAlex Elder } 27355522ae0bSAlex Elder EXPORT_SYMBOL(ceph_osdc_setup); 27365522ae0bSAlex Elder 27375522ae0bSAlex Elder void ceph_osdc_cleanup(void) 27385522ae0bSAlex Elder { 27395522ae0bSAlex Elder BUG_ON(!ceph_osd_request_cache); 27405522ae0bSAlex Elder kmem_cache_destroy(ceph_osd_request_cache); 27415522ae0bSAlex Elder ceph_osd_request_cache = NULL; 27425522ae0bSAlex Elder } 27435522ae0bSAlex Elder EXPORT_SYMBOL(ceph_osdc_cleanup); 27445522ae0bSAlex Elder 27453d14c5d2SYehuda Sadeh /* 27463d14c5d2SYehuda Sadeh * handle incoming message 27473d14c5d2SYehuda Sadeh */ 27483d14c5d2SYehuda Sadeh static void dispatch(struct ceph_connection *con, struct ceph_msg *msg) 27493d14c5d2SYehuda Sadeh { 27503d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 27513d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 27523d14c5d2SYehuda Sadeh int type = le16_to_cpu(msg->hdr.type); 27533d14c5d2SYehuda Sadeh 27543d14c5d2SYehuda Sadeh if (!osd) 27553d14c5d2SYehuda Sadeh goto out; 27563d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 27573d14c5d2SYehuda Sadeh 27583d14c5d2SYehuda Sadeh switch (type) { 27593d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 27603d14c5d2SYehuda Sadeh ceph_osdc_handle_map(osdc, msg); 27613d14c5d2SYehuda Sadeh break; 27623d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 27633d14c5d2SYehuda Sadeh handle_reply(osdc, msg, con); 27643d14c5d2SYehuda Sadeh break; 2765a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 2766a40c4f10SYehuda Sadeh handle_watch_notify(osdc, msg); 2767a40c4f10SYehuda Sadeh break; 27683d14c5d2SYehuda Sadeh 27693d14c5d2SYehuda Sadeh default: 27703d14c5d2SYehuda Sadeh pr_err("received unknown message type %d %s\n", type, 27713d14c5d2SYehuda Sadeh ceph_msg_type_name(type)); 27723d14c5d2SYehuda Sadeh } 27733d14c5d2SYehuda Sadeh out: 27743d14c5d2SYehuda Sadeh ceph_msg_put(msg); 27753d14c5d2SYehuda Sadeh } 27763d14c5d2SYehuda Sadeh 27773d14c5d2SYehuda Sadeh /* 27783d14c5d2SYehuda Sadeh * lookup and return message for incoming reply. set up reply message 27793d14c5d2SYehuda Sadeh * pages. 27803d14c5d2SYehuda Sadeh */ 27813d14c5d2SYehuda Sadeh static struct ceph_msg *get_reply(struct ceph_connection *con, 27823d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 27833d14c5d2SYehuda Sadeh int *skip) 27843d14c5d2SYehuda Sadeh { 27853d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 27863d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = osd->o_osdc; 27873d14c5d2SYehuda Sadeh struct ceph_msg *m; 27883d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 27893f0a4ac5SIlya Dryomov int front_len = le32_to_cpu(hdr->front_len); 27903d14c5d2SYehuda Sadeh int data_len = le32_to_cpu(hdr->data_len); 27913d14c5d2SYehuda Sadeh u64 tid; 27923d14c5d2SYehuda Sadeh 27933d14c5d2SYehuda Sadeh tid = le64_to_cpu(hdr->tid); 27943d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 27953d14c5d2SYehuda Sadeh req = __lookup_request(osdc, tid); 27963d14c5d2SYehuda Sadeh if (!req) { 27973d14c5d2SYehuda Sadeh *skip = 1; 27983d14c5d2SYehuda Sadeh m = NULL; 2799756a16a5SSage Weil dout("get_reply unknown tid %llu from osd%d\n", tid, 28003d14c5d2SYehuda Sadeh osd->o_osd); 28013d14c5d2SYehuda Sadeh goto out; 28023d14c5d2SYehuda Sadeh } 28033d14c5d2SYehuda Sadeh 2804ace6d3a9SAlex Elder if (req->r_reply->con) 28058921d114SAlex Elder dout("%s revoking msg %p from old con %p\n", __func__, 2806ace6d3a9SAlex Elder req->r_reply, req->r_reply->con); 28078921d114SAlex Elder ceph_msg_revoke_incoming(req->r_reply); 28083d14c5d2SYehuda Sadeh 2809f2be82b0SIlya Dryomov if (front_len > req->r_reply->front_alloc_len) { 28104974341eSAlex Elder pr_warning("get_reply front %d > preallocated %d (%u#%llu)\n", 2811f2be82b0SIlya Dryomov front_len, req->r_reply->front_alloc_len, 28124974341eSAlex Elder (unsigned int)con->peer_name.type, 28134974341eSAlex Elder le64_to_cpu(con->peer_name.num)); 28143f0a4ac5SIlya Dryomov m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front_len, GFP_NOFS, 28153f0a4ac5SIlya Dryomov false); 28163d14c5d2SYehuda Sadeh if (!m) 28173d14c5d2SYehuda Sadeh goto out; 28183d14c5d2SYehuda Sadeh ceph_msg_put(req->r_reply); 28193d14c5d2SYehuda Sadeh req->r_reply = m; 28203d14c5d2SYehuda Sadeh } 28213d14c5d2SYehuda Sadeh m = ceph_msg_get(req->r_reply); 28223d14c5d2SYehuda Sadeh 28233d14c5d2SYehuda Sadeh if (data_len > 0) { 2824a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 28250fff87ecSAlex Elder 2826a4ce40a9SAlex Elder /* 2827a4ce40a9SAlex Elder * XXX This is assuming there is only one op containing 2828a4ce40a9SAlex Elder * XXX page data. Probably OK for reads, but this 2829a4ce40a9SAlex Elder * XXX ought to be done more generally. 2830a4ce40a9SAlex Elder */ 2831406e2c9fSAlex Elder osd_data = osd_req_op_extent_osd_data(req, 0); 28320fff87ecSAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 28330fff87ecSAlex Elder if (osd_data->pages && 2834e0c59487SAlex Elder unlikely(osd_data->length < data_len)) { 28352ac2b7a6SAlex Elder 2836e0c59487SAlex Elder pr_warning("tid %lld reply has %d bytes " 2837e0c59487SAlex Elder "we had only %llu bytes ready\n", 2838e0c59487SAlex Elder tid, data_len, osd_data->length); 28393d14c5d2SYehuda Sadeh *skip = 1; 28403d14c5d2SYehuda Sadeh ceph_msg_put(m); 28413d14c5d2SYehuda Sadeh m = NULL; 28423d14c5d2SYehuda Sadeh goto out; 28433d14c5d2SYehuda Sadeh } 28443d14c5d2SYehuda Sadeh } 28452ac2b7a6SAlex Elder } 28463d14c5d2SYehuda Sadeh *skip = 0; 28473d14c5d2SYehuda Sadeh dout("get_reply tid %lld %p\n", tid, m); 28483d14c5d2SYehuda Sadeh 28493d14c5d2SYehuda Sadeh out: 28503d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 28513d14c5d2SYehuda Sadeh return m; 28523d14c5d2SYehuda Sadeh 28533d14c5d2SYehuda Sadeh } 28543d14c5d2SYehuda Sadeh 28553d14c5d2SYehuda Sadeh static struct ceph_msg *alloc_msg(struct ceph_connection *con, 28563d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 28573d14c5d2SYehuda Sadeh int *skip) 28583d14c5d2SYehuda Sadeh { 28593d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 28603d14c5d2SYehuda Sadeh int type = le16_to_cpu(hdr->type); 28613d14c5d2SYehuda Sadeh int front = le32_to_cpu(hdr->front_len); 28623d14c5d2SYehuda Sadeh 28631c20f2d2SAlex Elder *skip = 0; 28643d14c5d2SYehuda Sadeh switch (type) { 28653d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 2866a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 2867b61c2763SSage Weil return ceph_msg_new(type, front, GFP_NOFS, false); 28683d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 28693d14c5d2SYehuda Sadeh return get_reply(con, hdr, skip); 28703d14c5d2SYehuda Sadeh default: 28713d14c5d2SYehuda Sadeh pr_info("alloc_msg unexpected msg type %d from osd%d\n", type, 28723d14c5d2SYehuda Sadeh osd->o_osd); 28733d14c5d2SYehuda Sadeh *skip = 1; 28743d14c5d2SYehuda Sadeh return NULL; 28753d14c5d2SYehuda Sadeh } 28763d14c5d2SYehuda Sadeh } 28773d14c5d2SYehuda Sadeh 28783d14c5d2SYehuda Sadeh /* 28793d14c5d2SYehuda Sadeh * Wrappers to refcount containing ceph_osd struct 28803d14c5d2SYehuda Sadeh */ 28813d14c5d2SYehuda Sadeh static struct ceph_connection *get_osd_con(struct ceph_connection *con) 28823d14c5d2SYehuda Sadeh { 28833d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 28843d14c5d2SYehuda Sadeh if (get_osd(osd)) 28853d14c5d2SYehuda Sadeh return con; 28863d14c5d2SYehuda Sadeh return NULL; 28873d14c5d2SYehuda Sadeh } 28883d14c5d2SYehuda Sadeh 28893d14c5d2SYehuda Sadeh static void put_osd_con(struct ceph_connection *con) 28903d14c5d2SYehuda Sadeh { 28913d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 28923d14c5d2SYehuda Sadeh put_osd(osd); 28933d14c5d2SYehuda Sadeh } 28943d14c5d2SYehuda Sadeh 28953d14c5d2SYehuda Sadeh /* 28963d14c5d2SYehuda Sadeh * authentication 28973d14c5d2SYehuda Sadeh */ 2898a3530df3SAlex Elder /* 2899a3530df3SAlex Elder * Note: returned pointer is the address of a structure that's 2900a3530df3SAlex Elder * managed separately. Caller must *not* attempt to free it. 2901a3530df3SAlex Elder */ 2902a3530df3SAlex Elder static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con, 29038f43fb53SAlex Elder int *proto, int force_new) 29043d14c5d2SYehuda Sadeh { 29053d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 29063d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 29073d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 290874f1869fSAlex Elder struct ceph_auth_handshake *auth = &o->o_auth; 29093d14c5d2SYehuda Sadeh 291074f1869fSAlex Elder if (force_new && auth->authorizer) { 291127859f97SSage Weil ceph_auth_destroy_authorizer(ac, auth->authorizer); 291274f1869fSAlex Elder auth->authorizer = NULL; 29133d14c5d2SYehuda Sadeh } 291427859f97SSage Weil if (!auth->authorizer) { 291527859f97SSage Weil int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 2916a3530df3SAlex Elder auth); 29173d14c5d2SYehuda Sadeh if (ret) 2918a3530df3SAlex Elder return ERR_PTR(ret); 291927859f97SSage Weil } else { 292027859f97SSage Weil int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 29210bed9b5cSSage Weil auth); 29220bed9b5cSSage Weil if (ret) 29230bed9b5cSSage Weil return ERR_PTR(ret); 29243d14c5d2SYehuda Sadeh } 29253d14c5d2SYehuda Sadeh *proto = ac->protocol; 292674f1869fSAlex Elder 2927a3530df3SAlex Elder return auth; 29283d14c5d2SYehuda Sadeh } 29293d14c5d2SYehuda Sadeh 29303d14c5d2SYehuda Sadeh 29313d14c5d2SYehuda Sadeh static int verify_authorizer_reply(struct ceph_connection *con, int len) 29323d14c5d2SYehuda Sadeh { 29333d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 29343d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 29353d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 29363d14c5d2SYehuda Sadeh 293727859f97SSage Weil return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len); 29383d14c5d2SYehuda Sadeh } 29393d14c5d2SYehuda Sadeh 29403d14c5d2SYehuda Sadeh static int invalidate_authorizer(struct ceph_connection *con) 29413d14c5d2SYehuda Sadeh { 29423d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 29433d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 29443d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 29453d14c5d2SYehuda Sadeh 294627859f97SSage Weil ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD); 29473d14c5d2SYehuda Sadeh return ceph_monc_validate_auth(&osdc->client->monc); 29483d14c5d2SYehuda Sadeh } 29493d14c5d2SYehuda Sadeh 29503d14c5d2SYehuda Sadeh static const struct ceph_connection_operations osd_con_ops = { 29513d14c5d2SYehuda Sadeh .get = get_osd_con, 29523d14c5d2SYehuda Sadeh .put = put_osd_con, 29533d14c5d2SYehuda Sadeh .dispatch = dispatch, 29543d14c5d2SYehuda Sadeh .get_authorizer = get_authorizer, 29553d14c5d2SYehuda Sadeh .verify_authorizer_reply = verify_authorizer_reply, 29563d14c5d2SYehuda Sadeh .invalidate_authorizer = invalidate_authorizer, 29573d14c5d2SYehuda Sadeh .alloc_msg = alloc_msg, 29583d14c5d2SYehuda Sadeh .fault = osd_reset, 29593d14c5d2SYehuda Sadeh }; 2960