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); 332cc6128aSIlya Dryomov static void __unregister_request(struct ceph_osd_client *osdc, 342cc6128aSIlya Dryomov struct ceph_osd_request *req); 35a40c4f10SYehuda Sadeh static void __unregister_linger_request(struct ceph_osd_client *osdc, 36a40c4f10SYehuda Sadeh struct ceph_osd_request *req); 372cc6128aSIlya Dryomov static void __enqueue_request(struct ceph_osd_request *req); 3856e925b6SSage Weil static void __send_request(struct ceph_osd_client *osdc, 39a40c4f10SYehuda Sadeh struct ceph_osd_request *req); 403d14c5d2SYehuda Sadeh 413d14c5d2SYehuda Sadeh /* 423d14c5d2SYehuda Sadeh * Implement client access to distributed object storage cluster. 433d14c5d2SYehuda Sadeh * 443d14c5d2SYehuda Sadeh * All data objects are stored within a cluster/cloud of OSDs, or 453d14c5d2SYehuda Sadeh * "object storage devices." (Note that Ceph OSDs have _nothing_ to 463d14c5d2SYehuda Sadeh * do with the T10 OSD extensions to SCSI.) Ceph OSDs are simply 473d14c5d2SYehuda Sadeh * remote daemons serving up and coordinating consistent and safe 483d14c5d2SYehuda Sadeh * access to storage. 493d14c5d2SYehuda Sadeh * 503d14c5d2SYehuda Sadeh * Cluster membership and the mapping of data objects onto storage devices 513d14c5d2SYehuda Sadeh * are described by the osd map. 523d14c5d2SYehuda Sadeh * 533d14c5d2SYehuda Sadeh * We keep track of pending OSD requests (read, write), resubmit 543d14c5d2SYehuda Sadeh * requests to different OSDs when the cluster topology/data layout 553d14c5d2SYehuda Sadeh * change, or retry the affected requests when the communications 563d14c5d2SYehuda Sadeh * channel with an OSD is reset. 573d14c5d2SYehuda Sadeh */ 583d14c5d2SYehuda Sadeh 593d14c5d2SYehuda Sadeh /* 603d14c5d2SYehuda Sadeh * calculate the mapping of a file extent onto an object, and fill out the 613d14c5d2SYehuda Sadeh * request accordingly. shorten extent as necessary if it crosses an 623d14c5d2SYehuda Sadeh * object boundary. 633d14c5d2SYehuda Sadeh * 643d14c5d2SYehuda Sadeh * fill osd op in request message. 653d14c5d2SYehuda Sadeh */ 66dbe0fc41SAlex Elder static int calc_layout(struct ceph_file_layout *layout, u64 off, u64 *plen, 67a19dadfbSAlex Elder u64 *objnum, u64 *objoff, u64 *objlen) 683d14c5d2SYehuda Sadeh { 6960e56f13SAlex Elder u64 orig_len = *plen; 70d63b77f4SSage Weil int r; 713d14c5d2SYehuda Sadeh 7260e56f13SAlex Elder /* object extent? */ 7375d1c941SAlex Elder r = ceph_calc_file_object_mapping(layout, off, orig_len, objnum, 7475d1c941SAlex Elder objoff, objlen); 75d63b77f4SSage Weil if (r < 0) 76d63b77f4SSage Weil return r; 7775d1c941SAlex Elder if (*objlen < orig_len) { 7875d1c941SAlex Elder *plen = *objlen; 7960e56f13SAlex Elder dout(" skipping last %llu, final file extent %llu~%llu\n", 8060e56f13SAlex Elder orig_len - *plen, off, *plen); 8160e56f13SAlex Elder } 8260e56f13SAlex Elder 8375d1c941SAlex Elder dout("calc_layout objnum=%llx %llu~%llu\n", *objnum, *objoff, *objlen); 843d14c5d2SYehuda Sadeh 853ff5f385SAlex Elder return 0; 863d14c5d2SYehuda Sadeh } 873d14c5d2SYehuda Sadeh 88c54d47bfSAlex Elder static void ceph_osd_data_init(struct ceph_osd_data *osd_data) 89c54d47bfSAlex Elder { 90c54d47bfSAlex Elder memset(osd_data, 0, sizeof (*osd_data)); 91c54d47bfSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_NONE; 92c54d47bfSAlex Elder } 93c54d47bfSAlex Elder 94a4ce40a9SAlex Elder static void ceph_osd_data_pages_init(struct ceph_osd_data *osd_data, 9543bfe5deSAlex Elder struct page **pages, u64 length, u32 alignment, 9643bfe5deSAlex Elder bool pages_from_pool, bool own_pages) 9743bfe5deSAlex Elder { 9843bfe5deSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_PAGES; 9943bfe5deSAlex Elder osd_data->pages = pages; 10043bfe5deSAlex Elder osd_data->length = length; 10143bfe5deSAlex Elder osd_data->alignment = alignment; 10243bfe5deSAlex Elder osd_data->pages_from_pool = pages_from_pool; 10343bfe5deSAlex Elder osd_data->own_pages = own_pages; 10443bfe5deSAlex Elder } 10543bfe5deSAlex Elder 106a4ce40a9SAlex Elder static void ceph_osd_data_pagelist_init(struct ceph_osd_data *osd_data, 10743bfe5deSAlex Elder struct ceph_pagelist *pagelist) 10843bfe5deSAlex Elder { 10943bfe5deSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_PAGELIST; 11043bfe5deSAlex Elder osd_data->pagelist = pagelist; 11143bfe5deSAlex Elder } 11243bfe5deSAlex Elder 11343bfe5deSAlex Elder #ifdef CONFIG_BLOCK 114a4ce40a9SAlex Elder static void ceph_osd_data_bio_init(struct ceph_osd_data *osd_data, 11543bfe5deSAlex Elder struct bio *bio, size_t bio_length) 11643bfe5deSAlex Elder { 11743bfe5deSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_BIO; 11843bfe5deSAlex Elder osd_data->bio = bio; 11943bfe5deSAlex Elder osd_data->bio_length = bio_length; 12043bfe5deSAlex Elder } 12143bfe5deSAlex Elder #endif /* CONFIG_BLOCK */ 12243bfe5deSAlex Elder 123863c7eb5SAlex Elder #define osd_req_op_data(oreq, whch, typ, fld) \ 124863c7eb5SAlex Elder ({ \ 125863c7eb5SAlex Elder BUG_ON(whch >= (oreq)->r_num_ops); \ 126863c7eb5SAlex Elder &(oreq)->r_ops[whch].typ.fld; \ 127863c7eb5SAlex Elder }) 128863c7eb5SAlex Elder 12949719778SAlex Elder static struct ceph_osd_data * 13049719778SAlex Elder osd_req_op_raw_data_in(struct ceph_osd_request *osd_req, unsigned int which) 13149719778SAlex Elder { 13249719778SAlex Elder BUG_ON(which >= osd_req->r_num_ops); 13349719778SAlex Elder 13449719778SAlex Elder return &osd_req->r_ops[which].raw_data_in; 13549719778SAlex Elder } 13649719778SAlex Elder 137a4ce40a9SAlex Elder struct ceph_osd_data * 138a4ce40a9SAlex Elder osd_req_op_extent_osd_data(struct ceph_osd_request *osd_req, 139406e2c9fSAlex Elder unsigned int which) 140a4ce40a9SAlex Elder { 141863c7eb5SAlex Elder return osd_req_op_data(osd_req, which, extent, osd_data); 142a4ce40a9SAlex Elder } 143a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data); 144a4ce40a9SAlex Elder 145a4ce40a9SAlex Elder struct ceph_osd_data * 146a4ce40a9SAlex Elder osd_req_op_cls_response_data(struct ceph_osd_request *osd_req, 147a4ce40a9SAlex Elder unsigned int which) 148a4ce40a9SAlex Elder { 149863c7eb5SAlex Elder return osd_req_op_data(osd_req, which, cls, response_data); 150a4ce40a9SAlex Elder } 151a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_response_data); /* ??? */ 152a4ce40a9SAlex Elder 15349719778SAlex Elder void osd_req_op_raw_data_in_pages(struct ceph_osd_request *osd_req, 15449719778SAlex Elder unsigned int which, struct page **pages, 15549719778SAlex Elder u64 length, u32 alignment, 15649719778SAlex Elder bool pages_from_pool, bool own_pages) 15749719778SAlex Elder { 15849719778SAlex Elder struct ceph_osd_data *osd_data; 15949719778SAlex Elder 16049719778SAlex Elder osd_data = osd_req_op_raw_data_in(osd_req, which); 16149719778SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 16249719778SAlex Elder pages_from_pool, own_pages); 16349719778SAlex Elder } 16449719778SAlex Elder EXPORT_SYMBOL(osd_req_op_raw_data_in_pages); 16549719778SAlex Elder 166a4ce40a9SAlex Elder void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *osd_req, 167406e2c9fSAlex Elder unsigned int which, struct page **pages, 168406e2c9fSAlex Elder u64 length, u32 alignment, 169a4ce40a9SAlex Elder bool pages_from_pool, bool own_pages) 170a4ce40a9SAlex Elder { 171a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 172a4ce40a9SAlex Elder 173863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, extent, osd_data); 174a4ce40a9SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 175a4ce40a9SAlex Elder pages_from_pool, own_pages); 176a4ce40a9SAlex Elder } 177a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data_pages); 178a4ce40a9SAlex Elder 179a4ce40a9SAlex Elder void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *osd_req, 180406e2c9fSAlex Elder unsigned int which, struct ceph_pagelist *pagelist) 181a4ce40a9SAlex Elder { 182a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 183a4ce40a9SAlex Elder 184863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, extent, osd_data); 185a4ce40a9SAlex Elder ceph_osd_data_pagelist_init(osd_data, pagelist); 186a4ce40a9SAlex Elder } 187a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data_pagelist); 188a4ce40a9SAlex Elder 189a4ce40a9SAlex Elder #ifdef CONFIG_BLOCK 190a4ce40a9SAlex Elder void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *osd_req, 191406e2c9fSAlex Elder unsigned int which, struct bio *bio, size_t bio_length) 192a4ce40a9SAlex Elder { 193a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 194863c7eb5SAlex Elder 195863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, extent, osd_data); 196a4ce40a9SAlex Elder ceph_osd_data_bio_init(osd_data, bio, bio_length); 197a4ce40a9SAlex Elder } 198a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data_bio); 199a4ce40a9SAlex Elder #endif /* CONFIG_BLOCK */ 200a4ce40a9SAlex Elder 201a4ce40a9SAlex Elder static void osd_req_op_cls_request_info_pagelist( 202a4ce40a9SAlex Elder struct ceph_osd_request *osd_req, 203a4ce40a9SAlex Elder unsigned int which, struct ceph_pagelist *pagelist) 204a4ce40a9SAlex Elder { 205a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 206a4ce40a9SAlex Elder 207863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, request_info); 208a4ce40a9SAlex Elder ceph_osd_data_pagelist_init(osd_data, pagelist); 209a4ce40a9SAlex Elder } 210a4ce40a9SAlex Elder 21104017e29SAlex Elder void osd_req_op_cls_request_data_pagelist( 21204017e29SAlex Elder struct ceph_osd_request *osd_req, 21304017e29SAlex Elder unsigned int which, struct ceph_pagelist *pagelist) 21404017e29SAlex Elder { 21504017e29SAlex Elder struct ceph_osd_data *osd_data; 21604017e29SAlex Elder 217863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, request_data); 21804017e29SAlex Elder ceph_osd_data_pagelist_init(osd_data, pagelist); 21904017e29SAlex Elder } 22004017e29SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_request_data_pagelist); 22104017e29SAlex Elder 2226c57b554SAlex Elder void osd_req_op_cls_request_data_pages(struct ceph_osd_request *osd_req, 2236c57b554SAlex Elder unsigned int which, struct page **pages, u64 length, 2246c57b554SAlex Elder u32 alignment, bool pages_from_pool, bool own_pages) 2256c57b554SAlex Elder { 2266c57b554SAlex Elder struct ceph_osd_data *osd_data; 2276c57b554SAlex Elder 2286c57b554SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, request_data); 2296c57b554SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 2306c57b554SAlex Elder pages_from_pool, own_pages); 2316c57b554SAlex Elder } 2326c57b554SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_request_data_pages); 2336c57b554SAlex Elder 234a4ce40a9SAlex Elder void osd_req_op_cls_response_data_pages(struct ceph_osd_request *osd_req, 235a4ce40a9SAlex Elder unsigned int which, struct page **pages, u64 length, 236a4ce40a9SAlex Elder u32 alignment, bool pages_from_pool, bool own_pages) 237a4ce40a9SAlex Elder { 238a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 239a4ce40a9SAlex Elder 240863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, response_data); 241a4ce40a9SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 242a4ce40a9SAlex Elder pages_from_pool, own_pages); 243a4ce40a9SAlex Elder } 244a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_response_data_pages); 245a4ce40a9SAlex Elder 24623c08a9cSAlex Elder static u64 ceph_osd_data_length(struct ceph_osd_data *osd_data) 24723c08a9cSAlex Elder { 24823c08a9cSAlex Elder switch (osd_data->type) { 24923c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_NONE: 25023c08a9cSAlex Elder return 0; 25123c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_PAGES: 25223c08a9cSAlex Elder return osd_data->length; 25323c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_PAGELIST: 25423c08a9cSAlex Elder return (u64)osd_data->pagelist->length; 25523c08a9cSAlex Elder #ifdef CONFIG_BLOCK 25623c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_BIO: 25723c08a9cSAlex Elder return (u64)osd_data->bio_length; 25823c08a9cSAlex Elder #endif /* CONFIG_BLOCK */ 25923c08a9cSAlex Elder default: 26023c08a9cSAlex Elder WARN(true, "unrecognized data type %d\n", (int)osd_data->type); 26123c08a9cSAlex Elder return 0; 26223c08a9cSAlex Elder } 26323c08a9cSAlex Elder } 26423c08a9cSAlex Elder 265c54d47bfSAlex Elder static void ceph_osd_data_release(struct ceph_osd_data *osd_data) 266c54d47bfSAlex Elder { 2675476492fSAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES && osd_data->own_pages) { 268c54d47bfSAlex Elder int num_pages; 269c54d47bfSAlex Elder 270c54d47bfSAlex Elder num_pages = calc_pages_for((u64)osd_data->alignment, 271c54d47bfSAlex Elder (u64)osd_data->length); 272c54d47bfSAlex Elder ceph_release_page_vector(osd_data->pages, num_pages); 273c54d47bfSAlex Elder } 2745476492fSAlex Elder ceph_osd_data_init(osd_data); 2755476492fSAlex Elder } 2765476492fSAlex Elder 2775476492fSAlex Elder static void osd_req_op_data_release(struct ceph_osd_request *osd_req, 2785476492fSAlex Elder unsigned int which) 2795476492fSAlex Elder { 2805476492fSAlex Elder struct ceph_osd_req_op *op; 2815476492fSAlex Elder 2825476492fSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 2835476492fSAlex Elder op = &osd_req->r_ops[which]; 2845476492fSAlex Elder 2855476492fSAlex Elder switch (op->op) { 2865476492fSAlex Elder case CEPH_OSD_OP_READ: 2875476492fSAlex Elder case CEPH_OSD_OP_WRITE: 2885476492fSAlex Elder ceph_osd_data_release(&op->extent.osd_data); 2895476492fSAlex Elder break; 2905476492fSAlex Elder case CEPH_OSD_OP_CALL: 2915476492fSAlex Elder ceph_osd_data_release(&op->cls.request_info); 29204017e29SAlex Elder ceph_osd_data_release(&op->cls.request_data); 2935476492fSAlex Elder ceph_osd_data_release(&op->cls.response_data); 2945476492fSAlex Elder break; 2955476492fSAlex Elder default: 2965476492fSAlex Elder break; 2975476492fSAlex Elder } 298c54d47bfSAlex Elder } 299c54d47bfSAlex Elder 3003d14c5d2SYehuda Sadeh /* 3013d14c5d2SYehuda Sadeh * requests 3023d14c5d2SYehuda Sadeh */ 3039e94af20SIlya Dryomov static void ceph_osdc_release_request(struct kref *kref) 3043d14c5d2SYehuda Sadeh { 3059e94af20SIlya Dryomov struct ceph_osd_request *req = container_of(kref, 3069e94af20SIlya Dryomov struct ceph_osd_request, r_kref); 3075476492fSAlex Elder unsigned int which; 3083d14c5d2SYehuda Sadeh 3099e94af20SIlya Dryomov dout("%s %p (r_request %p r_reply %p)\n", __func__, req, 3109e94af20SIlya Dryomov req->r_request, req->r_reply); 3116562d661SIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&req->r_node)); 3126562d661SIlya Dryomov WARN_ON(!list_empty(&req->r_req_lru_item)); 3136562d661SIlya Dryomov WARN_ON(!list_empty(&req->r_osd_item)); 3146562d661SIlya Dryomov WARN_ON(!list_empty(&req->r_linger_item)); 3156562d661SIlya Dryomov WARN_ON(!list_empty(&req->r_linger_osd_item)); 3166562d661SIlya Dryomov WARN_ON(req->r_osd); 3179e94af20SIlya Dryomov 3183d14c5d2SYehuda Sadeh if (req->r_request) 3193d14c5d2SYehuda Sadeh ceph_msg_put(req->r_request); 320ace6d3a9SAlex Elder if (req->r_reply) { 3218921d114SAlex Elder ceph_msg_revoke_incoming(req->r_reply); 322ab8cb34aSAlex Elder ceph_msg_put(req->r_reply); 323ace6d3a9SAlex Elder } 3240fff87ecSAlex Elder 3255476492fSAlex Elder for (which = 0; which < req->r_num_ops; which++) 3265476492fSAlex Elder osd_req_op_data_release(req, which); 3270fff87ecSAlex Elder 3283d14c5d2SYehuda Sadeh ceph_put_snap_context(req->r_snapc); 3293d14c5d2SYehuda Sadeh if (req->r_mempool) 3303d14c5d2SYehuda Sadeh mempool_free(req, req->r_osdc->req_mempool); 3313d14c5d2SYehuda Sadeh else 3325522ae0bSAlex Elder kmem_cache_free(ceph_osd_request_cache, req); 3335522ae0bSAlex Elder 3343d14c5d2SYehuda Sadeh } 3359e94af20SIlya Dryomov 3369e94af20SIlya Dryomov void ceph_osdc_get_request(struct ceph_osd_request *req) 3379e94af20SIlya Dryomov { 3389e94af20SIlya Dryomov dout("%s %p (was %d)\n", __func__, req, 3399e94af20SIlya Dryomov atomic_read(&req->r_kref.refcount)); 3409e94af20SIlya Dryomov kref_get(&req->r_kref); 3419e94af20SIlya Dryomov } 3429e94af20SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_get_request); 3439e94af20SIlya Dryomov 3449e94af20SIlya Dryomov void ceph_osdc_put_request(struct ceph_osd_request *req) 3459e94af20SIlya Dryomov { 3469e94af20SIlya Dryomov dout("%s %p (was %d)\n", __func__, req, 3479e94af20SIlya Dryomov atomic_read(&req->r_kref.refcount)); 3489e94af20SIlya Dryomov kref_put(&req->r_kref, ceph_osdc_release_request); 3499e94af20SIlya Dryomov } 3509e94af20SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_put_request); 3513d14c5d2SYehuda Sadeh 3523d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc, 3533d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 3541b83bef2SSage Weil unsigned int num_ops, 3553d14c5d2SYehuda Sadeh bool use_mempool, 35654a54007SAlex Elder gfp_t gfp_flags) 3573d14c5d2SYehuda Sadeh { 3583d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 3593d14c5d2SYehuda Sadeh struct ceph_msg *msg; 3601b83bef2SSage Weil size_t msg_size; 3613d14c5d2SYehuda Sadeh 36279528734SAlex Elder BUILD_BUG_ON(CEPH_OSD_MAX_OP > U16_MAX); 36379528734SAlex Elder BUG_ON(num_ops > CEPH_OSD_MAX_OP); 36479528734SAlex Elder 3651b83bef2SSage Weil msg_size = 4 + 4 + 8 + 8 + 4+8; 3661b83bef2SSage Weil msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */ 3671b83bef2SSage Weil msg_size += 1 + 8 + 4 + 4; /* pg_t */ 3682d0ebc5dSIlya Dryomov msg_size += 4 + CEPH_MAX_OID_NAME_LEN; /* oid */ 3691b83bef2SSage Weil msg_size += 2 + num_ops*sizeof(struct ceph_osd_op); 3701b83bef2SSage Weil msg_size += 8; /* snapid */ 3711b83bef2SSage Weil msg_size += 8; /* snap_seq */ 3721b83bef2SSage Weil msg_size += 8 * (snapc ? snapc->num_snaps : 0); /* snaps */ 3731b83bef2SSage Weil msg_size += 4; 3743d14c5d2SYehuda Sadeh 3753d14c5d2SYehuda Sadeh if (use_mempool) { 3763d14c5d2SYehuda Sadeh req = mempool_alloc(osdc->req_mempool, gfp_flags); 3773d14c5d2SYehuda Sadeh memset(req, 0, sizeof(*req)); 3783d14c5d2SYehuda Sadeh } else { 3795522ae0bSAlex Elder req = kmem_cache_zalloc(ceph_osd_request_cache, gfp_flags); 3803d14c5d2SYehuda Sadeh } 3813d14c5d2SYehuda Sadeh if (req == NULL) 3823d14c5d2SYehuda Sadeh return NULL; 3833d14c5d2SYehuda Sadeh 3843d14c5d2SYehuda Sadeh req->r_osdc = osdc; 3853d14c5d2SYehuda Sadeh req->r_mempool = use_mempool; 38679528734SAlex Elder req->r_num_ops = num_ops; 3873d14c5d2SYehuda Sadeh 3883d14c5d2SYehuda Sadeh kref_init(&req->r_kref); 3893d14c5d2SYehuda Sadeh init_completion(&req->r_completion); 3903d14c5d2SYehuda Sadeh init_completion(&req->r_safe_completion); 391a978fa20SAlex Elder RB_CLEAR_NODE(&req->r_node); 3923d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&req->r_unsafe_item); 393a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&req->r_linger_item); 3941d0326b1SIlya Dryomov INIT_LIST_HEAD(&req->r_linger_osd_item); 395935b639aSSage Weil INIT_LIST_HEAD(&req->r_req_lru_item); 396cd43045cSSage Weil INIT_LIST_HEAD(&req->r_osd_item); 397cd43045cSSage Weil 3983c972c95SIlya Dryomov req->r_base_oloc.pool = -1; 399205ee118SIlya Dryomov req->r_target_oloc.pool = -1; 40022116525SIlya Dryomov 4013d14c5d2SYehuda Sadeh /* create reply message */ 4023d14c5d2SYehuda Sadeh if (use_mempool) 4033d14c5d2SYehuda Sadeh msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0); 4043d14c5d2SYehuda Sadeh else 4053d14c5d2SYehuda Sadeh msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, 406b61c2763SSage Weil OSD_OPREPLY_FRONT_LEN, gfp_flags, true); 4073d14c5d2SYehuda Sadeh if (!msg) { 4083d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 4093d14c5d2SYehuda Sadeh return NULL; 4103d14c5d2SYehuda Sadeh } 4113d14c5d2SYehuda Sadeh req->r_reply = msg; 4123d14c5d2SYehuda Sadeh 4133d14c5d2SYehuda Sadeh /* create request message; allow space for oid */ 4143d14c5d2SYehuda Sadeh if (use_mempool) 4153d14c5d2SYehuda Sadeh msg = ceph_msgpool_get(&osdc->msgpool_op, 0); 4163d14c5d2SYehuda Sadeh else 417b61c2763SSage Weil msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true); 4183d14c5d2SYehuda Sadeh if (!msg) { 4193d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 4203d14c5d2SYehuda Sadeh return NULL; 4213d14c5d2SYehuda Sadeh } 4223d14c5d2SYehuda Sadeh 4233d14c5d2SYehuda Sadeh memset(msg->front.iov_base, 0, msg->front.iov_len); 4243d14c5d2SYehuda Sadeh 4253d14c5d2SYehuda Sadeh req->r_request = msg; 4263d14c5d2SYehuda Sadeh 4273d14c5d2SYehuda Sadeh return req; 4283d14c5d2SYehuda Sadeh } 4293d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_alloc_request); 4303d14c5d2SYehuda Sadeh 431a8dd0a37SAlex Elder static bool osd_req_opcode_valid(u16 opcode) 432a8dd0a37SAlex Elder { 433a8dd0a37SAlex Elder switch (opcode) { 43470b5bfa3SIlya Dryomov #define GENERATE_CASE(op, opcode, str) case CEPH_OSD_OP_##op: return true; 43570b5bfa3SIlya Dryomov __CEPH_FORALL_OSD_OPS(GENERATE_CASE) 43670b5bfa3SIlya Dryomov #undef GENERATE_CASE 437a8dd0a37SAlex Elder default: 438a8dd0a37SAlex Elder return false; 439a8dd0a37SAlex Elder } 440a8dd0a37SAlex Elder } 441a8dd0a37SAlex Elder 44233803f33SAlex Elder /* 44333803f33SAlex Elder * This is an osd op init function for opcodes that have no data or 44433803f33SAlex Elder * other information associated with them. It also serves as a 44533803f33SAlex Elder * common init routine for all the other init functions, below. 44633803f33SAlex Elder */ 447c99d2d4aSAlex Elder static struct ceph_osd_req_op * 44849719778SAlex Elder _osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which, 449c99d2d4aSAlex Elder u16 opcode) 45033803f33SAlex Elder { 451c99d2d4aSAlex Elder struct ceph_osd_req_op *op; 452c99d2d4aSAlex Elder 453c99d2d4aSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 45433803f33SAlex Elder BUG_ON(!osd_req_opcode_valid(opcode)); 45533803f33SAlex Elder 456c99d2d4aSAlex Elder op = &osd_req->r_ops[which]; 45733803f33SAlex Elder memset(op, 0, sizeof (*op)); 45833803f33SAlex Elder op->op = opcode; 459c99d2d4aSAlex Elder 460c99d2d4aSAlex Elder return op; 46133803f33SAlex Elder } 46233803f33SAlex Elder 46349719778SAlex Elder void osd_req_op_init(struct ceph_osd_request *osd_req, 46449719778SAlex Elder unsigned int which, u16 opcode) 46549719778SAlex Elder { 46649719778SAlex Elder (void)_osd_req_op_init(osd_req, which, opcode); 46749719778SAlex Elder } 46849719778SAlex Elder EXPORT_SYMBOL(osd_req_op_init); 46949719778SAlex Elder 470c99d2d4aSAlex Elder void osd_req_op_extent_init(struct ceph_osd_request *osd_req, 471c99d2d4aSAlex Elder unsigned int which, u16 opcode, 47233803f33SAlex Elder u64 offset, u64 length, 47333803f33SAlex Elder u64 truncate_size, u32 truncate_seq) 47433803f33SAlex Elder { 47549719778SAlex Elder struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode); 47633803f33SAlex Elder size_t payload_len = 0; 47733803f33SAlex Elder 478ad7a60deSLi Wang BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE && 479ad7a60deSLi Wang opcode != CEPH_OSD_OP_DELETE && opcode != CEPH_OSD_OP_ZERO && 480ad7a60deSLi Wang opcode != CEPH_OSD_OP_TRUNCATE); 48133803f33SAlex Elder 48233803f33SAlex Elder op->extent.offset = offset; 48333803f33SAlex Elder op->extent.length = length; 48433803f33SAlex Elder op->extent.truncate_size = truncate_size; 48533803f33SAlex Elder op->extent.truncate_seq = truncate_seq; 48633803f33SAlex Elder if (opcode == CEPH_OSD_OP_WRITE) 48733803f33SAlex Elder payload_len += length; 48833803f33SAlex Elder 48933803f33SAlex Elder op->payload_len = payload_len; 49033803f33SAlex Elder } 49133803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_init); 49233803f33SAlex Elder 493c99d2d4aSAlex Elder void osd_req_op_extent_update(struct ceph_osd_request *osd_req, 494c99d2d4aSAlex Elder unsigned int which, u64 length) 495e5975c7cSAlex Elder { 496c99d2d4aSAlex Elder struct ceph_osd_req_op *op; 497c99d2d4aSAlex Elder u64 previous; 498c99d2d4aSAlex Elder 499c99d2d4aSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 500c99d2d4aSAlex Elder op = &osd_req->r_ops[which]; 501c99d2d4aSAlex Elder previous = op->extent.length; 502e5975c7cSAlex Elder 503e5975c7cSAlex Elder if (length == previous) 504e5975c7cSAlex Elder return; /* Nothing to do */ 505e5975c7cSAlex Elder BUG_ON(length > previous); 506e5975c7cSAlex Elder 507e5975c7cSAlex Elder op->extent.length = length; 508e5975c7cSAlex Elder op->payload_len -= previous - length; 509e5975c7cSAlex Elder } 510e5975c7cSAlex Elder EXPORT_SYMBOL(osd_req_op_extent_update); 511e5975c7cSAlex Elder 512c99d2d4aSAlex Elder void osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which, 51304017e29SAlex Elder u16 opcode, const char *class, const char *method) 51433803f33SAlex Elder { 51549719778SAlex Elder struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode); 5165f562df5SAlex Elder struct ceph_pagelist *pagelist; 51733803f33SAlex Elder size_t payload_len = 0; 51833803f33SAlex Elder size_t size; 51933803f33SAlex Elder 52033803f33SAlex Elder BUG_ON(opcode != CEPH_OSD_OP_CALL); 52133803f33SAlex Elder 5225f562df5SAlex Elder pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS); 5235f562df5SAlex Elder BUG_ON(!pagelist); 5245f562df5SAlex Elder ceph_pagelist_init(pagelist); 5255f562df5SAlex Elder 52633803f33SAlex Elder op->cls.class_name = class; 52733803f33SAlex Elder size = strlen(class); 52833803f33SAlex Elder BUG_ON(size > (size_t) U8_MAX); 52933803f33SAlex Elder op->cls.class_len = size; 5305f562df5SAlex Elder ceph_pagelist_append(pagelist, class, size); 53133803f33SAlex Elder payload_len += size; 53233803f33SAlex Elder 53333803f33SAlex Elder op->cls.method_name = method; 53433803f33SAlex Elder size = strlen(method); 53533803f33SAlex Elder BUG_ON(size > (size_t) U8_MAX); 53633803f33SAlex Elder op->cls.method_len = size; 5375f562df5SAlex Elder ceph_pagelist_append(pagelist, method, size); 53833803f33SAlex Elder payload_len += size; 53933803f33SAlex Elder 540a4ce40a9SAlex Elder osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist); 5415f562df5SAlex Elder 54233803f33SAlex Elder op->cls.argc = 0; /* currently unused */ 54333803f33SAlex Elder 54433803f33SAlex Elder op->payload_len = payload_len; 54533803f33SAlex Elder } 54633803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_init); 5478c042b0dSAlex Elder 548c99d2d4aSAlex Elder void osd_req_op_watch_init(struct ceph_osd_request *osd_req, 549c99d2d4aSAlex Elder unsigned int which, u16 opcode, 55033803f33SAlex Elder u64 cookie, u64 version, int flag) 55133803f33SAlex Elder { 55249719778SAlex Elder struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode); 55333803f33SAlex Elder 554c99d2d4aSAlex Elder BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH); 55533803f33SAlex Elder 55633803f33SAlex Elder op->watch.cookie = cookie; 5579ef1ee5aSAlex Elder op->watch.ver = version; 55833803f33SAlex Elder if (opcode == CEPH_OSD_OP_WATCH && flag) 55933803f33SAlex Elder op->watch.flag = (u8)1; 56033803f33SAlex Elder } 56133803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_watch_init); 56233803f33SAlex Elder 563c647b8a8SIlya Dryomov void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req, 564c647b8a8SIlya Dryomov unsigned int which, 565c647b8a8SIlya Dryomov u64 expected_object_size, 566c647b8a8SIlya Dryomov u64 expected_write_size) 567c647b8a8SIlya Dryomov { 568c647b8a8SIlya Dryomov struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, 569c647b8a8SIlya Dryomov CEPH_OSD_OP_SETALLOCHINT); 570c647b8a8SIlya Dryomov 571c647b8a8SIlya Dryomov op->alloc_hint.expected_object_size = expected_object_size; 572c647b8a8SIlya Dryomov op->alloc_hint.expected_write_size = expected_write_size; 573c647b8a8SIlya Dryomov 574c647b8a8SIlya Dryomov /* 575c647b8a8SIlya Dryomov * CEPH_OSD_OP_SETALLOCHINT op is advisory and therefore deemed 576c647b8a8SIlya Dryomov * not worth a feature bit. Set FAILOK per-op flag to make 577c647b8a8SIlya Dryomov * sure older osds don't trip over an unsupported opcode. 578c647b8a8SIlya Dryomov */ 579c647b8a8SIlya Dryomov op->flags |= CEPH_OSD_OP_FLAG_FAILOK; 580c647b8a8SIlya Dryomov } 581c647b8a8SIlya Dryomov EXPORT_SYMBOL(osd_req_op_alloc_hint_init); 582c647b8a8SIlya Dryomov 58390af3602SAlex Elder static void ceph_osdc_msg_data_add(struct ceph_msg *msg, 584ec9123c5SAlex Elder struct ceph_osd_data *osd_data) 585ec9123c5SAlex Elder { 586ec9123c5SAlex Elder u64 length = ceph_osd_data_length(osd_data); 587ec9123c5SAlex Elder 588ec9123c5SAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 589ec9123c5SAlex Elder BUG_ON(length > (u64) SIZE_MAX); 590ec9123c5SAlex Elder if (length) 59190af3602SAlex Elder ceph_msg_data_add_pages(msg, osd_data->pages, 592ec9123c5SAlex Elder length, osd_data->alignment); 593ec9123c5SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) { 594ec9123c5SAlex Elder BUG_ON(!length); 59590af3602SAlex Elder ceph_msg_data_add_pagelist(msg, osd_data->pagelist); 596ec9123c5SAlex Elder #ifdef CONFIG_BLOCK 597ec9123c5SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) { 59890af3602SAlex Elder ceph_msg_data_add_bio(msg, osd_data->bio, length); 599ec9123c5SAlex Elder #endif 600ec9123c5SAlex Elder } else { 601ec9123c5SAlex Elder BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE); 602ec9123c5SAlex Elder } 603ec9123c5SAlex Elder } 604ec9123c5SAlex Elder 605175face2SAlex Elder static u64 osd_req_encode_op(struct ceph_osd_request *req, 60679528734SAlex Elder struct ceph_osd_op *dst, unsigned int which) 6073d14c5d2SYehuda Sadeh { 60879528734SAlex Elder struct ceph_osd_req_op *src; 60904017e29SAlex Elder struct ceph_osd_data *osd_data; 61054d50649SAlex Elder u64 request_data_len = 0; 61104017e29SAlex Elder u64 data_length; 612175face2SAlex Elder 61379528734SAlex Elder BUG_ON(which >= req->r_num_ops); 61479528734SAlex Elder src = &req->r_ops[which]; 615a8dd0a37SAlex Elder if (WARN_ON(!osd_req_opcode_valid(src->op))) { 616a8dd0a37SAlex Elder pr_err("unrecognized osd opcode %d\n", src->op); 617a8dd0a37SAlex Elder 618a8dd0a37SAlex Elder return 0; 619a8dd0a37SAlex Elder } 6203d14c5d2SYehuda Sadeh 621065a68f9SAlex Elder switch (src->op) { 622fbfab539SAlex Elder case CEPH_OSD_OP_STAT: 62349719778SAlex Elder osd_data = &src->raw_data_in; 62449719778SAlex Elder ceph_osdc_msg_data_add(req->r_reply, osd_data); 625fbfab539SAlex Elder break; 6263d14c5d2SYehuda Sadeh case CEPH_OSD_OP_READ: 6273d14c5d2SYehuda Sadeh case CEPH_OSD_OP_WRITE: 628ad7a60deSLi Wang case CEPH_OSD_OP_ZERO: 629ad7a60deSLi Wang case CEPH_OSD_OP_DELETE: 630ad7a60deSLi Wang case CEPH_OSD_OP_TRUNCATE: 631175face2SAlex Elder if (src->op == CEPH_OSD_OP_WRITE) 63254d50649SAlex Elder request_data_len = src->extent.length; 633175face2SAlex Elder dst->extent.offset = cpu_to_le64(src->extent.offset); 634175face2SAlex Elder dst->extent.length = cpu_to_le64(src->extent.length); 6353d14c5d2SYehuda Sadeh dst->extent.truncate_size = 6363d14c5d2SYehuda Sadeh cpu_to_le64(src->extent.truncate_size); 6373d14c5d2SYehuda Sadeh dst->extent.truncate_seq = 6383d14c5d2SYehuda Sadeh cpu_to_le32(src->extent.truncate_seq); 63904017e29SAlex Elder osd_data = &src->extent.osd_data; 6405476492fSAlex Elder if (src->op == CEPH_OSD_OP_WRITE) 64104017e29SAlex Elder ceph_osdc_msg_data_add(req->r_request, osd_data); 6425476492fSAlex Elder else 64304017e29SAlex Elder ceph_osdc_msg_data_add(req->r_reply, osd_data); 6443d14c5d2SYehuda Sadeh break; 6453d14c5d2SYehuda Sadeh case CEPH_OSD_OP_CALL: 6463d14c5d2SYehuda Sadeh dst->cls.class_len = src->cls.class_len; 6473d14c5d2SYehuda Sadeh dst->cls.method_len = src->cls.method_len; 64804017e29SAlex Elder osd_data = &src->cls.request_info; 64904017e29SAlex Elder ceph_osdc_msg_data_add(req->r_request, osd_data); 65004017e29SAlex Elder BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGELIST); 65104017e29SAlex Elder request_data_len = osd_data->pagelist->length; 65204017e29SAlex Elder 65304017e29SAlex Elder osd_data = &src->cls.request_data; 65404017e29SAlex Elder data_length = ceph_osd_data_length(osd_data); 65504017e29SAlex Elder if (data_length) { 65604017e29SAlex Elder BUG_ON(osd_data->type == CEPH_OSD_DATA_TYPE_NONE); 65704017e29SAlex Elder dst->cls.indata_len = cpu_to_le32(data_length); 65804017e29SAlex Elder ceph_osdc_msg_data_add(req->r_request, osd_data); 65904017e29SAlex Elder src->payload_len += data_length; 66004017e29SAlex Elder request_data_len += data_length; 66104017e29SAlex Elder } 66204017e29SAlex Elder osd_data = &src->cls.response_data; 66304017e29SAlex Elder ceph_osdc_msg_data_add(req->r_reply, osd_data); 6643d14c5d2SYehuda Sadeh break; 6653d14c5d2SYehuda Sadeh case CEPH_OSD_OP_STARTSYNC: 6663d14c5d2SYehuda Sadeh break; 667a40c4f10SYehuda Sadeh case CEPH_OSD_OP_NOTIFY_ACK: 668a40c4f10SYehuda Sadeh case CEPH_OSD_OP_WATCH: 669a40c4f10SYehuda Sadeh dst->watch.cookie = cpu_to_le64(src->watch.cookie); 670a40c4f10SYehuda Sadeh dst->watch.ver = cpu_to_le64(src->watch.ver); 671a40c4f10SYehuda Sadeh dst->watch.flag = src->watch.flag; 672a40c4f10SYehuda Sadeh break; 673c647b8a8SIlya Dryomov case CEPH_OSD_OP_SETALLOCHINT: 674c647b8a8SIlya Dryomov dst->alloc_hint.expected_object_size = 675c647b8a8SIlya Dryomov cpu_to_le64(src->alloc_hint.expected_object_size); 676c647b8a8SIlya Dryomov dst->alloc_hint.expected_write_size = 677c647b8a8SIlya Dryomov cpu_to_le64(src->alloc_hint.expected_write_size); 678c647b8a8SIlya Dryomov break; 6793d14c5d2SYehuda Sadeh default: 6804c46459cSAlex Elder pr_err("unsupported osd opcode %s\n", 6818f63ca2dSAlex Elder ceph_osd_op_name(src->op)); 6824c46459cSAlex Elder WARN_ON(1); 683a8dd0a37SAlex Elder 684a8dd0a37SAlex Elder return 0; 6853d14c5d2SYehuda Sadeh } 6867b25bf5fSIlya Dryomov 687a8dd0a37SAlex Elder dst->op = cpu_to_le16(src->op); 6887b25bf5fSIlya Dryomov dst->flags = cpu_to_le32(src->flags); 6893d14c5d2SYehuda Sadeh dst->payload_len = cpu_to_le32(src->payload_len); 690175face2SAlex Elder 69154d50649SAlex Elder return request_data_len; 6923d14c5d2SYehuda Sadeh } 6933d14c5d2SYehuda Sadeh 6943d14c5d2SYehuda Sadeh /* 6953d14c5d2SYehuda Sadeh * build new request AND message, calculate layout, and adjust file 6963d14c5d2SYehuda Sadeh * extent as needed. 6973d14c5d2SYehuda Sadeh * 6983d14c5d2SYehuda Sadeh * if the file was recently truncated, we include information about its 6993d14c5d2SYehuda Sadeh * old and new size so that the object can be updated appropriately. (we 7003d14c5d2SYehuda Sadeh * avoid synchronously deleting truncated objects because it's slow.) 7013d14c5d2SYehuda Sadeh * 7023d14c5d2SYehuda Sadeh * if @do_sync, include a 'startsync' command so that the osd will flush 7033d14c5d2SYehuda Sadeh * data quickly. 7043d14c5d2SYehuda Sadeh */ 7053d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc, 7063d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 7073d14c5d2SYehuda Sadeh struct ceph_vino vino, 708acead002SAlex Elder u64 off, u64 *plen, int num_ops, 7093d14c5d2SYehuda Sadeh int opcode, int flags, 7103d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 7113d14c5d2SYehuda Sadeh u32 truncate_seq, 7123d14c5d2SYehuda Sadeh u64 truncate_size, 713153e5167SAlex Elder bool use_mempool) 7143d14c5d2SYehuda Sadeh { 7153d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 71675d1c941SAlex Elder u64 objnum = 0; 71775d1c941SAlex Elder u64 objoff = 0; 71875d1c941SAlex Elder u64 objlen = 0; 719d18d1e28SAlex Elder u32 object_size; 720d18d1e28SAlex Elder u64 object_base; 7216816282dSSage Weil int r; 7223d14c5d2SYehuda Sadeh 723ad7a60deSLi Wang BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE && 724ad7a60deSLi Wang opcode != CEPH_OSD_OP_DELETE && opcode != CEPH_OSD_OP_ZERO && 725ad7a60deSLi Wang opcode != CEPH_OSD_OP_TRUNCATE); 7263d14c5d2SYehuda Sadeh 727acead002SAlex Elder req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool, 728ae7ca4a3SAlex Elder GFP_NOFS); 7294ad12621SSage Weil if (!req) 7306816282dSSage Weil return ERR_PTR(-ENOMEM); 73179528734SAlex Elder 732d178a9e7SAlex Elder req->r_flags = flags; 7333d14c5d2SYehuda Sadeh 7343d14c5d2SYehuda Sadeh /* calculate max write size */ 735a19dadfbSAlex Elder r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen); 7363ff5f385SAlex Elder if (r < 0) { 7373ff5f385SAlex Elder ceph_osdc_put_request(req); 7386816282dSSage Weil return ERR_PTR(r); 7393ff5f385SAlex Elder } 740a19dadfbSAlex Elder 741d18d1e28SAlex Elder object_size = le32_to_cpu(layout->fl_object_size); 742d18d1e28SAlex Elder object_base = off - objoff; 743ccca4e37SYan, Zheng if (!(truncate_seq == 1 && truncate_size == -1ULL)) { 744d18d1e28SAlex Elder if (truncate_size <= object_base) { 745d18d1e28SAlex Elder truncate_size = 0; 746d18d1e28SAlex Elder } else { 747d18d1e28SAlex Elder truncate_size -= object_base; 748d18d1e28SAlex Elder if (truncate_size > object_size) 749d18d1e28SAlex Elder truncate_size = object_size; 750d18d1e28SAlex Elder } 751ccca4e37SYan, Zheng } 752d18d1e28SAlex Elder 753c99d2d4aSAlex Elder osd_req_op_extent_init(req, 0, opcode, objoff, objlen, 754b0270324SAlex Elder truncate_size, truncate_seq); 7558c042b0dSAlex Elder 756acead002SAlex Elder /* 757acead002SAlex Elder * A second op in the ops array means the caller wants to 758acead002SAlex Elder * also issue a include a 'startsync' command so that the 759acead002SAlex Elder * osd will flush data quickly. 760acead002SAlex Elder */ 761acead002SAlex Elder if (num_ops > 1) 762c99d2d4aSAlex Elder osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC); 763d18d1e28SAlex Elder 7643c972c95SIlya Dryomov req->r_base_oloc.pool = ceph_file_layout_pg_pool(*layout); 7653d14c5d2SYehuda Sadeh 7663c972c95SIlya Dryomov snprintf(req->r_base_oid.name, sizeof(req->r_base_oid.name), 7674295f221SIlya Dryomov "%llx.%08llx", vino.ino, objnum); 7683c972c95SIlya Dryomov req->r_base_oid.name_len = strlen(req->r_base_oid.name); 769dbe0fc41SAlex Elder 7703d14c5d2SYehuda Sadeh return req; 7713d14c5d2SYehuda Sadeh } 7723d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_new_request); 7733d14c5d2SYehuda Sadeh 7743d14c5d2SYehuda Sadeh /* 7753d14c5d2SYehuda Sadeh * We keep osd requests in an rbtree, sorted by ->r_tid. 7763d14c5d2SYehuda Sadeh */ 7773d14c5d2SYehuda Sadeh static void __insert_request(struct ceph_osd_client *osdc, 7783d14c5d2SYehuda Sadeh struct ceph_osd_request *new) 7793d14c5d2SYehuda Sadeh { 7803d14c5d2SYehuda Sadeh struct rb_node **p = &osdc->requests.rb_node; 7813d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 7823d14c5d2SYehuda Sadeh struct ceph_osd_request *req = NULL; 7833d14c5d2SYehuda Sadeh 7843d14c5d2SYehuda Sadeh while (*p) { 7853d14c5d2SYehuda Sadeh parent = *p; 7863d14c5d2SYehuda Sadeh req = rb_entry(parent, struct ceph_osd_request, r_node); 7873d14c5d2SYehuda Sadeh if (new->r_tid < req->r_tid) 7883d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 7893d14c5d2SYehuda Sadeh else if (new->r_tid > req->r_tid) 7903d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 7913d14c5d2SYehuda Sadeh else 7923d14c5d2SYehuda Sadeh BUG(); 7933d14c5d2SYehuda Sadeh } 7943d14c5d2SYehuda Sadeh 7953d14c5d2SYehuda Sadeh rb_link_node(&new->r_node, parent, p); 7963d14c5d2SYehuda Sadeh rb_insert_color(&new->r_node, &osdc->requests); 7973d14c5d2SYehuda Sadeh } 7983d14c5d2SYehuda Sadeh 7993d14c5d2SYehuda Sadeh static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc, 8003d14c5d2SYehuda Sadeh u64 tid) 8013d14c5d2SYehuda Sadeh { 8023d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 8033d14c5d2SYehuda Sadeh struct rb_node *n = osdc->requests.rb_node; 8043d14c5d2SYehuda Sadeh 8053d14c5d2SYehuda Sadeh while (n) { 8063d14c5d2SYehuda Sadeh req = rb_entry(n, struct ceph_osd_request, r_node); 8073d14c5d2SYehuda Sadeh if (tid < req->r_tid) 8083d14c5d2SYehuda Sadeh n = n->rb_left; 8093d14c5d2SYehuda Sadeh else if (tid > req->r_tid) 8103d14c5d2SYehuda Sadeh n = n->rb_right; 8113d14c5d2SYehuda Sadeh else 8123d14c5d2SYehuda Sadeh return req; 8133d14c5d2SYehuda Sadeh } 8143d14c5d2SYehuda Sadeh return NULL; 8153d14c5d2SYehuda Sadeh } 8163d14c5d2SYehuda Sadeh 8173d14c5d2SYehuda Sadeh static struct ceph_osd_request * 8183d14c5d2SYehuda Sadeh __lookup_request_ge(struct ceph_osd_client *osdc, 8193d14c5d2SYehuda Sadeh u64 tid) 8203d14c5d2SYehuda Sadeh { 8213d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 8223d14c5d2SYehuda Sadeh struct rb_node *n = osdc->requests.rb_node; 8233d14c5d2SYehuda Sadeh 8243d14c5d2SYehuda Sadeh while (n) { 8253d14c5d2SYehuda Sadeh req = rb_entry(n, struct ceph_osd_request, r_node); 8263d14c5d2SYehuda Sadeh if (tid < req->r_tid) { 8273d14c5d2SYehuda Sadeh if (!n->rb_left) 8283d14c5d2SYehuda Sadeh return req; 8293d14c5d2SYehuda Sadeh n = n->rb_left; 8303d14c5d2SYehuda Sadeh } else if (tid > req->r_tid) { 8313d14c5d2SYehuda Sadeh n = n->rb_right; 8323d14c5d2SYehuda Sadeh } else { 8333d14c5d2SYehuda Sadeh return req; 8343d14c5d2SYehuda Sadeh } 8353d14c5d2SYehuda Sadeh } 8363d14c5d2SYehuda Sadeh return NULL; 8373d14c5d2SYehuda Sadeh } 8383d14c5d2SYehuda Sadeh 8392cc6128aSIlya Dryomov static void __kick_linger_request(struct ceph_osd_request *req) 8402cc6128aSIlya Dryomov { 8412cc6128aSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 8422cc6128aSIlya Dryomov struct ceph_osd *osd = req->r_osd; 8432cc6128aSIlya Dryomov 8442cc6128aSIlya Dryomov /* 8452cc6128aSIlya Dryomov * Linger requests need to be resent with a new tid to avoid 8462cc6128aSIlya Dryomov * the dup op detection logic on the OSDs. Achieve this with 8472cc6128aSIlya Dryomov * a re-register dance instead of open-coding. 8482cc6128aSIlya Dryomov */ 8492cc6128aSIlya Dryomov ceph_osdc_get_request(req); 8502cc6128aSIlya Dryomov if (!list_empty(&req->r_linger_item)) 8512cc6128aSIlya Dryomov __unregister_linger_request(osdc, req); 8522cc6128aSIlya Dryomov else 8532cc6128aSIlya Dryomov __unregister_request(osdc, req); 8542cc6128aSIlya Dryomov __register_request(osdc, req); 8552cc6128aSIlya Dryomov ceph_osdc_put_request(req); 8562cc6128aSIlya Dryomov 8572cc6128aSIlya Dryomov /* 8582cc6128aSIlya Dryomov * Unless request has been registered as both normal and 8592cc6128aSIlya Dryomov * lingering, __unregister{,_linger}_request clears r_osd. 8602cc6128aSIlya Dryomov * However, here we need to preserve r_osd to make sure we 8612cc6128aSIlya Dryomov * requeue on the same OSD. 8622cc6128aSIlya Dryomov */ 8632cc6128aSIlya Dryomov WARN_ON(req->r_osd || !osd); 8642cc6128aSIlya Dryomov req->r_osd = osd; 8652cc6128aSIlya Dryomov 8662cc6128aSIlya Dryomov dout("%s requeueing %p tid %llu\n", __func__, req, req->r_tid); 8672cc6128aSIlya Dryomov __enqueue_request(req); 8682cc6128aSIlya Dryomov } 8692cc6128aSIlya Dryomov 8706f6c7006SSage Weil /* 8716f6c7006SSage Weil * Resubmit requests pending on the given osd. 8726f6c7006SSage Weil */ 8736f6c7006SSage Weil static void __kick_osd_requests(struct ceph_osd_client *osdc, 8746f6c7006SSage Weil struct ceph_osd *osd) 8756f6c7006SSage Weil { 876a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 877e02493c0SAlex Elder LIST_HEAD(resend); 8782cc6128aSIlya Dryomov LIST_HEAD(resend_linger); 8796f6c7006SSage Weil int err; 8806f6c7006SSage Weil 8812cc6128aSIlya Dryomov dout("%s osd%d\n", __func__, osd->o_osd); 8826f6c7006SSage Weil err = __reset_osd(osdc, osd); 883685a7555SAlex Elder if (err) 8846f6c7006SSage Weil return; 8852cc6128aSIlya Dryomov 886e02493c0SAlex Elder /* 887e02493c0SAlex Elder * Build up a list of requests to resend by traversing the 888e02493c0SAlex Elder * osd's list of requests. Requests for a given object are 889e02493c0SAlex Elder * sent in tid order, and that is also the order they're 890e02493c0SAlex Elder * kept on this list. Therefore all requests that are in 891e02493c0SAlex Elder * flight will be found first, followed by all requests that 892e02493c0SAlex Elder * have not yet been sent. And to resend requests while 893e02493c0SAlex Elder * preserving this order we will want to put any sent 894e02493c0SAlex Elder * requests back on the front of the osd client's unsent 895e02493c0SAlex Elder * list. 896e02493c0SAlex Elder * 897e02493c0SAlex Elder * So we build a separate ordered list of already-sent 898e02493c0SAlex Elder * requests for the affected osd and splice it onto the 899e02493c0SAlex Elder * front of the osd client's unsent list. Once we've seen a 900e02493c0SAlex Elder * request that has not yet been sent we're done. Those 901e02493c0SAlex Elder * requests are already sitting right where they belong. 902e02493c0SAlex Elder */ 9036f6c7006SSage Weil list_for_each_entry(req, &osd->o_requests, r_osd_item) { 904e02493c0SAlex Elder if (!req->r_sent) 905e02493c0SAlex Elder break; 9062cc6128aSIlya Dryomov 9072cc6128aSIlya Dryomov if (!req->r_linger) { 9082cc6128aSIlya Dryomov dout("%s requeueing %p tid %llu\n", __func__, req, 9092cc6128aSIlya Dryomov req->r_tid); 910e02493c0SAlex Elder list_move_tail(&req->r_req_lru_item, &resend); 9116f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 9122cc6128aSIlya Dryomov } else { 9132cc6128aSIlya Dryomov list_move_tail(&req->r_req_lru_item, &resend_linger); 9142cc6128aSIlya Dryomov } 9156f6c7006SSage Weil } 916e02493c0SAlex Elder list_splice(&resend, &osdc->req_unsent); 917a40c4f10SYehuda Sadeh 918e02493c0SAlex Elder /* 9192cc6128aSIlya Dryomov * Both registered and not yet registered linger requests are 9202cc6128aSIlya Dryomov * enqueued with a new tid on the same OSD. We add/move them 9212cc6128aSIlya Dryomov * to req_unsent/o_requests at the end to keep things in tid 9222cc6128aSIlya Dryomov * order. 923e02493c0SAlex Elder */ 924a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osd->o_linger_requests, 9251d0326b1SIlya Dryomov r_linger_osd_item) { 9262cc6128aSIlya Dryomov WARN_ON(!list_empty(&req->r_req_lru_item)); 9272cc6128aSIlya Dryomov __kick_linger_request(req); 928a40c4f10SYehuda Sadeh } 9292cc6128aSIlya Dryomov 9302cc6128aSIlya Dryomov list_for_each_entry_safe(req, nreq, &resend_linger, r_req_lru_item) 9312cc6128aSIlya Dryomov __kick_linger_request(req); 9326f6c7006SSage Weil } 9336f6c7006SSage Weil 9343d14c5d2SYehuda Sadeh /* 9353d14c5d2SYehuda Sadeh * If the osd connection drops, we need to resubmit all requests. 9363d14c5d2SYehuda Sadeh */ 9373d14c5d2SYehuda Sadeh static void osd_reset(struct ceph_connection *con) 9383d14c5d2SYehuda Sadeh { 9393d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 9403d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 9413d14c5d2SYehuda Sadeh 9423d14c5d2SYehuda Sadeh if (!osd) 9433d14c5d2SYehuda Sadeh return; 9443d14c5d2SYehuda Sadeh dout("osd_reset osd%d\n", osd->o_osd); 9453d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 9463d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 94783aff95eSSage Weil mutex_lock(&osdc->request_mutex); 94883aff95eSSage Weil __kick_osd_requests(osdc, osd); 949f9d25199SAlex Elder __send_queued(osdc); 95083aff95eSSage Weil mutex_unlock(&osdc->request_mutex); 9513d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 9523d14c5d2SYehuda Sadeh } 9533d14c5d2SYehuda Sadeh 9543d14c5d2SYehuda Sadeh /* 9553d14c5d2SYehuda Sadeh * Track open sessions with osds. 9563d14c5d2SYehuda Sadeh */ 957e10006f8SAlex Elder static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum) 9583d14c5d2SYehuda Sadeh { 9593d14c5d2SYehuda Sadeh struct ceph_osd *osd; 9603d14c5d2SYehuda Sadeh 9613d14c5d2SYehuda Sadeh osd = kzalloc(sizeof(*osd), GFP_NOFS); 9623d14c5d2SYehuda Sadeh if (!osd) 9633d14c5d2SYehuda Sadeh return NULL; 9643d14c5d2SYehuda Sadeh 9653d14c5d2SYehuda Sadeh atomic_set(&osd->o_ref, 1); 9663d14c5d2SYehuda Sadeh osd->o_osdc = osdc; 967e10006f8SAlex Elder osd->o_osd = onum; 968f407731dSAlex Elder RB_CLEAR_NODE(&osd->o_node); 9693d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_requests); 970a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osd->o_linger_requests); 9713d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_osd_lru); 9723d14c5d2SYehuda Sadeh osd->o_incarnation = 1; 9733d14c5d2SYehuda Sadeh 974b7a9e5ddSSage Weil ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr); 9753d14c5d2SYehuda Sadeh 9763d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_keepalive_item); 9773d14c5d2SYehuda Sadeh return osd; 9783d14c5d2SYehuda Sadeh } 9793d14c5d2SYehuda Sadeh 9803d14c5d2SYehuda Sadeh static struct ceph_osd *get_osd(struct ceph_osd *osd) 9813d14c5d2SYehuda Sadeh { 9823d14c5d2SYehuda Sadeh if (atomic_inc_not_zero(&osd->o_ref)) { 9833d14c5d2SYehuda Sadeh dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1, 9843d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref)); 9853d14c5d2SYehuda Sadeh return osd; 9863d14c5d2SYehuda Sadeh } else { 9873d14c5d2SYehuda Sadeh dout("get_osd %p FAIL\n", osd); 9883d14c5d2SYehuda Sadeh return NULL; 9893d14c5d2SYehuda Sadeh } 9903d14c5d2SYehuda Sadeh } 9913d14c5d2SYehuda Sadeh 9923d14c5d2SYehuda Sadeh static void put_osd(struct ceph_osd *osd) 9933d14c5d2SYehuda Sadeh { 9943d14c5d2SYehuda Sadeh dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref), 9953d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref) - 1); 996a255651dSAlex Elder if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) { 9973d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth; 9983d14c5d2SYehuda Sadeh 99927859f97SSage Weil ceph_auth_destroy_authorizer(ac, osd->o_auth.authorizer); 10003d14c5d2SYehuda Sadeh kfree(osd); 10013d14c5d2SYehuda Sadeh } 10023d14c5d2SYehuda Sadeh } 10033d14c5d2SYehuda Sadeh 10043d14c5d2SYehuda Sadeh /* 10053d14c5d2SYehuda Sadeh * remove an osd from our map 10063d14c5d2SYehuda Sadeh */ 10073d14c5d2SYehuda Sadeh static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 10083d14c5d2SYehuda Sadeh { 10093d14c5d2SYehuda Sadeh dout("__remove_osd %p\n", osd); 1010cc9f1f51SIlya Dryomov WARN_ON(!list_empty(&osd->o_requests)); 1011cc9f1f51SIlya Dryomov WARN_ON(!list_empty(&osd->o_linger_requests)); 10127c6e6fc5SIlya Dryomov 10133d14c5d2SYehuda Sadeh rb_erase(&osd->o_node, &osdc->osds); 10143d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 10153d14c5d2SYehuda Sadeh ceph_con_close(&osd->o_con); 10163d14c5d2SYehuda Sadeh put_osd(osd); 10173d14c5d2SYehuda Sadeh } 10183d14c5d2SYehuda Sadeh 1019aca420bcSSage Weil static void remove_all_osds(struct ceph_osd_client *osdc) 1020aca420bcSSage Weil { 1021048a9d2dSJiaju Zhang dout("%s %p\n", __func__, osdc); 1022aca420bcSSage Weil mutex_lock(&osdc->request_mutex); 1023aca420bcSSage Weil while (!RB_EMPTY_ROOT(&osdc->osds)) { 1024aca420bcSSage Weil struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds), 1025aca420bcSSage Weil struct ceph_osd, o_node); 1026aca420bcSSage Weil __remove_osd(osdc, osd); 1027aca420bcSSage Weil } 1028aca420bcSSage Weil mutex_unlock(&osdc->request_mutex); 1029aca420bcSSage Weil } 1030aca420bcSSage Weil 10313d14c5d2SYehuda Sadeh static void __move_osd_to_lru(struct ceph_osd_client *osdc, 10323d14c5d2SYehuda Sadeh struct ceph_osd *osd) 10333d14c5d2SYehuda Sadeh { 1034bbf37ec3SIlya Dryomov dout("%s %p\n", __func__, osd); 10353d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_osd_lru)); 1036bbf37ec3SIlya Dryomov 10373d14c5d2SYehuda Sadeh list_add_tail(&osd->o_osd_lru, &osdc->osd_lru); 10383d14c5d2SYehuda Sadeh osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ; 10393d14c5d2SYehuda Sadeh } 10403d14c5d2SYehuda Sadeh 1041bbf37ec3SIlya Dryomov static void maybe_move_osd_to_lru(struct ceph_osd_client *osdc, 1042bbf37ec3SIlya Dryomov struct ceph_osd *osd) 1043bbf37ec3SIlya Dryomov { 1044bbf37ec3SIlya Dryomov dout("%s %p\n", __func__, osd); 1045bbf37ec3SIlya Dryomov 1046bbf37ec3SIlya Dryomov if (list_empty(&osd->o_requests) && 1047bbf37ec3SIlya Dryomov list_empty(&osd->o_linger_requests)) 1048bbf37ec3SIlya Dryomov __move_osd_to_lru(osdc, osd); 1049bbf37ec3SIlya Dryomov } 1050bbf37ec3SIlya Dryomov 10513d14c5d2SYehuda Sadeh static void __remove_osd_from_lru(struct ceph_osd *osd) 10523d14c5d2SYehuda Sadeh { 10533d14c5d2SYehuda Sadeh dout("__remove_osd_from_lru %p\n", osd); 10543d14c5d2SYehuda Sadeh if (!list_empty(&osd->o_osd_lru)) 10553d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 10563d14c5d2SYehuda Sadeh } 10573d14c5d2SYehuda Sadeh 1058aca420bcSSage Weil static void remove_old_osds(struct ceph_osd_client *osdc) 10593d14c5d2SYehuda Sadeh { 10603d14c5d2SYehuda Sadeh struct ceph_osd *osd, *nosd; 10613d14c5d2SYehuda Sadeh 10623d14c5d2SYehuda Sadeh dout("__remove_old_osds %p\n", osdc); 10633d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 10643d14c5d2SYehuda Sadeh list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) { 1065aca420bcSSage Weil if (time_before(jiffies, osd->lru_ttl)) 10663d14c5d2SYehuda Sadeh break; 10673d14c5d2SYehuda Sadeh __remove_osd(osdc, osd); 10683d14c5d2SYehuda Sadeh } 10693d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 10703d14c5d2SYehuda Sadeh } 10713d14c5d2SYehuda Sadeh 10723d14c5d2SYehuda Sadeh /* 10733d14c5d2SYehuda Sadeh * reset osd connect 10743d14c5d2SYehuda Sadeh */ 10753d14c5d2SYehuda Sadeh static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 10763d14c5d2SYehuda Sadeh { 1077c3acb181SAlex Elder struct ceph_entity_addr *peer_addr; 10783d14c5d2SYehuda Sadeh 10793d14c5d2SYehuda Sadeh dout("__reset_osd %p osd%d\n", osd, osd->o_osd); 1080a40c4f10SYehuda Sadeh if (list_empty(&osd->o_requests) && 1081a40c4f10SYehuda Sadeh list_empty(&osd->o_linger_requests)) { 10823d14c5d2SYehuda Sadeh __remove_osd(osdc, osd); 1083c3acb181SAlex Elder 1084c3acb181SAlex Elder return -ENODEV; 1085c3acb181SAlex Elder } 1086c3acb181SAlex Elder 1087c3acb181SAlex Elder peer_addr = &osdc->osdmap->osd_addr[osd->o_osd]; 1088c3acb181SAlex Elder if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) && 10893d14c5d2SYehuda Sadeh !ceph_con_opened(&osd->o_con)) { 1090c3acb181SAlex Elder struct ceph_osd_request *req; 1091c3acb181SAlex Elder 10923d14c5d2SYehuda Sadeh dout("osd addr hasn't changed and connection never opened, " 10930b4af2e8SIlya Dryomov "letting msgr retry\n"); 10943d14c5d2SYehuda Sadeh /* touch each r_stamp for handle_timeout()'s benfit */ 10953d14c5d2SYehuda Sadeh list_for_each_entry(req, &osd->o_requests, r_osd_item) 10963d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 1097c3acb181SAlex Elder 1098c3acb181SAlex Elder return -EAGAIN; 10993d14c5d2SYehuda Sadeh } 1100c3acb181SAlex Elder 1101c3acb181SAlex Elder ceph_con_close(&osd->o_con); 1102c3acb181SAlex Elder ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr); 1103c3acb181SAlex Elder osd->o_incarnation++; 1104c3acb181SAlex Elder 1105c3acb181SAlex Elder return 0; 11063d14c5d2SYehuda Sadeh } 11073d14c5d2SYehuda Sadeh 11083d14c5d2SYehuda Sadeh static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new) 11093d14c5d2SYehuda Sadeh { 11103d14c5d2SYehuda Sadeh struct rb_node **p = &osdc->osds.rb_node; 11113d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 11123d14c5d2SYehuda Sadeh struct ceph_osd *osd = NULL; 11133d14c5d2SYehuda Sadeh 1114aca420bcSSage Weil dout("__insert_osd %p osd%d\n", new, new->o_osd); 11153d14c5d2SYehuda Sadeh while (*p) { 11163d14c5d2SYehuda Sadeh parent = *p; 11173d14c5d2SYehuda Sadeh osd = rb_entry(parent, struct ceph_osd, o_node); 11183d14c5d2SYehuda Sadeh if (new->o_osd < osd->o_osd) 11193d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 11203d14c5d2SYehuda Sadeh else if (new->o_osd > osd->o_osd) 11213d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 11223d14c5d2SYehuda Sadeh else 11233d14c5d2SYehuda Sadeh BUG(); 11243d14c5d2SYehuda Sadeh } 11253d14c5d2SYehuda Sadeh 11263d14c5d2SYehuda Sadeh rb_link_node(&new->o_node, parent, p); 11273d14c5d2SYehuda Sadeh rb_insert_color(&new->o_node, &osdc->osds); 11283d14c5d2SYehuda Sadeh } 11293d14c5d2SYehuda Sadeh 11303d14c5d2SYehuda Sadeh static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o) 11313d14c5d2SYehuda Sadeh { 11323d14c5d2SYehuda Sadeh struct ceph_osd *osd; 11333d14c5d2SYehuda Sadeh struct rb_node *n = osdc->osds.rb_node; 11343d14c5d2SYehuda Sadeh 11353d14c5d2SYehuda Sadeh while (n) { 11363d14c5d2SYehuda Sadeh osd = rb_entry(n, struct ceph_osd, o_node); 11373d14c5d2SYehuda Sadeh if (o < osd->o_osd) 11383d14c5d2SYehuda Sadeh n = n->rb_left; 11393d14c5d2SYehuda Sadeh else if (o > osd->o_osd) 11403d14c5d2SYehuda Sadeh n = n->rb_right; 11413d14c5d2SYehuda Sadeh else 11423d14c5d2SYehuda Sadeh return osd; 11433d14c5d2SYehuda Sadeh } 11443d14c5d2SYehuda Sadeh return NULL; 11453d14c5d2SYehuda Sadeh } 11463d14c5d2SYehuda Sadeh 11473d14c5d2SYehuda Sadeh static void __schedule_osd_timeout(struct ceph_osd_client *osdc) 11483d14c5d2SYehuda Sadeh { 11493d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->timeout_work, 11503d14c5d2SYehuda Sadeh osdc->client->options->osd_keepalive_timeout * HZ); 11513d14c5d2SYehuda Sadeh } 11523d14c5d2SYehuda Sadeh 11533d14c5d2SYehuda Sadeh static void __cancel_osd_timeout(struct ceph_osd_client *osdc) 11543d14c5d2SYehuda Sadeh { 11553d14c5d2SYehuda Sadeh cancel_delayed_work(&osdc->timeout_work); 11563d14c5d2SYehuda Sadeh } 11573d14c5d2SYehuda Sadeh 11583d14c5d2SYehuda Sadeh /* 11593d14c5d2SYehuda Sadeh * Register request, assign tid. If this is the first request, set up 11603d14c5d2SYehuda Sadeh * the timeout event. 11613d14c5d2SYehuda Sadeh */ 1162a40c4f10SYehuda Sadeh static void __register_request(struct ceph_osd_client *osdc, 11633d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 11643d14c5d2SYehuda Sadeh { 11653d14c5d2SYehuda Sadeh req->r_tid = ++osdc->last_tid; 11663d14c5d2SYehuda Sadeh req->r_request->hdr.tid = cpu_to_le64(req->r_tid); 116777f38e0eSSage Weil dout("__register_request %p tid %lld\n", req, req->r_tid); 11683d14c5d2SYehuda Sadeh __insert_request(osdc, req); 11693d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 11703d14c5d2SYehuda Sadeh osdc->num_requests++; 11713d14c5d2SYehuda Sadeh if (osdc->num_requests == 1) { 11723d14c5d2SYehuda Sadeh dout(" first request, scheduling timeout\n"); 11733d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 11743d14c5d2SYehuda Sadeh } 1175a40c4f10SYehuda Sadeh } 1176a40c4f10SYehuda Sadeh 11773d14c5d2SYehuda Sadeh /* 11783d14c5d2SYehuda Sadeh * called under osdc->request_mutex 11793d14c5d2SYehuda Sadeh */ 11803d14c5d2SYehuda Sadeh static void __unregister_request(struct ceph_osd_client *osdc, 11813d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 11823d14c5d2SYehuda Sadeh { 118335f9f8a0SSage Weil if (RB_EMPTY_NODE(&req->r_node)) { 118435f9f8a0SSage Weil dout("__unregister_request %p tid %lld not registered\n", 118535f9f8a0SSage Weil req, req->r_tid); 118635f9f8a0SSage Weil return; 118735f9f8a0SSage Weil } 118835f9f8a0SSage Weil 11893d14c5d2SYehuda Sadeh dout("__unregister_request %p tid %lld\n", req, req->r_tid); 11903d14c5d2SYehuda Sadeh rb_erase(&req->r_node, &osdc->requests); 11916562d661SIlya Dryomov RB_CLEAR_NODE(&req->r_node); 11923d14c5d2SYehuda Sadeh osdc->num_requests--; 11933d14c5d2SYehuda Sadeh 11943d14c5d2SYehuda Sadeh if (req->r_osd) { 11953d14c5d2SYehuda Sadeh /* make sure the original request isn't in flight. */ 11966740a845SAlex Elder ceph_msg_revoke(req->r_request); 11973d14c5d2SYehuda Sadeh 11983d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 1199bbf37ec3SIlya Dryomov maybe_move_osd_to_lru(osdc, req->r_osd); 12004f23409eSIlya Dryomov if (list_empty(&req->r_linger_osd_item)) 12013d14c5d2SYehuda Sadeh req->r_osd = NULL; 12023d14c5d2SYehuda Sadeh } 12033d14c5d2SYehuda Sadeh 12047d5f2481SAlex Elder list_del_init(&req->r_req_lru_item); 12053d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 12063d14c5d2SYehuda Sadeh 12073d14c5d2SYehuda Sadeh if (osdc->num_requests == 0) { 12083d14c5d2SYehuda Sadeh dout(" no requests, canceling timeout\n"); 12093d14c5d2SYehuda Sadeh __cancel_osd_timeout(osdc); 12103d14c5d2SYehuda Sadeh } 12113d14c5d2SYehuda Sadeh } 12123d14c5d2SYehuda Sadeh 12133d14c5d2SYehuda Sadeh /* 12143d14c5d2SYehuda Sadeh * Cancel a previously queued request message 12153d14c5d2SYehuda Sadeh */ 12163d14c5d2SYehuda Sadeh static void __cancel_request(struct ceph_osd_request *req) 12173d14c5d2SYehuda Sadeh { 12183d14c5d2SYehuda Sadeh if (req->r_sent && req->r_osd) { 12196740a845SAlex Elder ceph_msg_revoke(req->r_request); 12203d14c5d2SYehuda Sadeh req->r_sent = 0; 12213d14c5d2SYehuda Sadeh } 12223d14c5d2SYehuda Sadeh } 12233d14c5d2SYehuda Sadeh 1224a40c4f10SYehuda Sadeh static void __register_linger_request(struct ceph_osd_client *osdc, 1225a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1226a40c4f10SYehuda Sadeh { 1227af593064SIlya Dryomov dout("%s %p tid %llu\n", __func__, req, req->r_tid); 1228af593064SIlya Dryomov WARN_ON(!req->r_linger); 1229af593064SIlya Dryomov 123096e4dac6SAlex Elder ceph_osdc_get_request(req); 1231a40c4f10SYehuda Sadeh list_add_tail(&req->r_linger_item, &osdc->req_linger); 12326194ea89SSage Weil if (req->r_osd) 12331d0326b1SIlya Dryomov list_add_tail(&req->r_linger_osd_item, 12346194ea89SSage Weil &req->r_osd->o_linger_requests); 1235a40c4f10SYehuda Sadeh } 1236a40c4f10SYehuda Sadeh 1237a40c4f10SYehuda Sadeh static void __unregister_linger_request(struct ceph_osd_client *osdc, 1238a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1239a40c4f10SYehuda Sadeh { 1240af593064SIlya Dryomov WARN_ON(!req->r_linger); 1241af593064SIlya Dryomov 1242af593064SIlya Dryomov if (list_empty(&req->r_linger_item)) { 1243af593064SIlya Dryomov dout("%s %p tid %llu not registered\n", __func__, req, 1244af593064SIlya Dryomov req->r_tid); 1245af593064SIlya Dryomov return; 1246af593064SIlya Dryomov } 1247af593064SIlya Dryomov 1248af593064SIlya Dryomov dout("%s %p tid %llu\n", __func__, req, req->r_tid); 1249a40c4f10SYehuda Sadeh list_del_init(&req->r_linger_item); 1250af593064SIlya Dryomov 125161c74035SAlex Elder if (req->r_osd) { 12521d0326b1SIlya Dryomov list_del_init(&req->r_linger_osd_item); 1253bbf37ec3SIlya Dryomov maybe_move_osd_to_lru(osdc, req->r_osd); 1254fbdb9190SSage Weil if (list_empty(&req->r_osd_item)) 1255a40c4f10SYehuda Sadeh req->r_osd = NULL; 1256a40c4f10SYehuda Sadeh } 1257ba9d114eSIlya Dryomov 1258ba9d114eSIlya Dryomov list_del_init(&req->r_req_lru_item); /* can be on notarget */ 125996e4dac6SAlex Elder ceph_osdc_put_request(req); 1260a40c4f10SYehuda Sadeh } 1261a40c4f10SYehuda Sadeh 1262a40c4f10SYehuda Sadeh void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc, 1263a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1264a40c4f10SYehuda Sadeh { 1265a40c4f10SYehuda Sadeh if (!req->r_linger) { 1266a40c4f10SYehuda Sadeh dout("set_request_linger %p\n", req); 1267a40c4f10SYehuda Sadeh req->r_linger = 1; 1268a40c4f10SYehuda Sadeh } 1269a40c4f10SYehuda Sadeh } 1270a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_set_request_linger); 1271a40c4f10SYehuda Sadeh 12723d14c5d2SYehuda Sadeh /* 1273d29adb34SJosh Durgin * Returns whether a request should be blocked from being sent 1274d29adb34SJosh Durgin * based on the current osdmap and osd_client settings. 1275d29adb34SJosh Durgin * 1276d29adb34SJosh Durgin * Caller should hold map_sem for read. 1277d29adb34SJosh Durgin */ 1278d29adb34SJosh Durgin static bool __req_should_be_paused(struct ceph_osd_client *osdc, 1279d29adb34SJosh Durgin struct ceph_osd_request *req) 1280d29adb34SJosh Durgin { 1281d29adb34SJosh Durgin bool pauserd = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD); 1282d29adb34SJosh Durgin bool pausewr = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR) || 1283d29adb34SJosh Durgin ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL); 1284d29adb34SJosh Durgin return (req->r_flags & CEPH_OSD_FLAG_READ && pauserd) || 1285d29adb34SJosh Durgin (req->r_flags & CEPH_OSD_FLAG_WRITE && pausewr); 1286d29adb34SJosh Durgin } 1287d29adb34SJosh Durgin 1288d29adb34SJosh Durgin /* 128917a13e40SIlya Dryomov * Calculate mapping of a request to a PG. Takes tiering into account. 129017a13e40SIlya Dryomov */ 129117a13e40SIlya Dryomov static int __calc_request_pg(struct ceph_osdmap *osdmap, 129217a13e40SIlya Dryomov struct ceph_osd_request *req, 129317a13e40SIlya Dryomov struct ceph_pg *pg_out) 129417a13e40SIlya Dryomov { 1295205ee118SIlya Dryomov bool need_check_tiering; 1296205ee118SIlya Dryomov 1297205ee118SIlya Dryomov need_check_tiering = false; 1298205ee118SIlya Dryomov if (req->r_target_oloc.pool == -1) { 1299205ee118SIlya Dryomov req->r_target_oloc = req->r_base_oloc; /* struct */ 1300205ee118SIlya Dryomov need_check_tiering = true; 1301205ee118SIlya Dryomov } 1302205ee118SIlya Dryomov if (req->r_target_oid.name_len == 0) { 1303205ee118SIlya Dryomov ceph_oid_copy(&req->r_target_oid, &req->r_base_oid); 1304205ee118SIlya Dryomov need_check_tiering = true; 1305205ee118SIlya Dryomov } 1306205ee118SIlya Dryomov 1307205ee118SIlya Dryomov if (need_check_tiering && 1308205ee118SIlya Dryomov (req->r_flags & CEPH_OSD_FLAG_IGNORE_OVERLAY) == 0) { 130917a13e40SIlya Dryomov struct ceph_pg_pool_info *pi; 131017a13e40SIlya Dryomov 1311205ee118SIlya Dryomov pi = ceph_pg_pool_by_id(osdmap, req->r_target_oloc.pool); 131217a13e40SIlya Dryomov if (pi) { 131317a13e40SIlya Dryomov if ((req->r_flags & CEPH_OSD_FLAG_READ) && 131417a13e40SIlya Dryomov pi->read_tier >= 0) 1315205ee118SIlya Dryomov req->r_target_oloc.pool = pi->read_tier; 131617a13e40SIlya Dryomov if ((req->r_flags & CEPH_OSD_FLAG_WRITE) && 131717a13e40SIlya Dryomov pi->write_tier >= 0) 1318205ee118SIlya Dryomov req->r_target_oloc.pool = pi->write_tier; 131917a13e40SIlya Dryomov } 132017a13e40SIlya Dryomov /* !pi is caught in ceph_oloc_oid_to_pg() */ 132117a13e40SIlya Dryomov } 132217a13e40SIlya Dryomov 1323205ee118SIlya Dryomov return ceph_oloc_oid_to_pg(osdmap, &req->r_target_oloc, 1324205ee118SIlya Dryomov &req->r_target_oid, pg_out); 132517a13e40SIlya Dryomov } 132617a13e40SIlya Dryomov 1327f671b581SIlya Dryomov static void __enqueue_request(struct ceph_osd_request *req) 1328f671b581SIlya Dryomov { 1329f671b581SIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 1330f671b581SIlya Dryomov 1331f671b581SIlya Dryomov dout("%s %p tid %llu to osd%d\n", __func__, req, req->r_tid, 1332f671b581SIlya Dryomov req->r_osd ? req->r_osd->o_osd : -1); 1333f671b581SIlya Dryomov 1334f671b581SIlya Dryomov if (req->r_osd) { 1335f671b581SIlya Dryomov __remove_osd_from_lru(req->r_osd); 1336f671b581SIlya Dryomov list_add_tail(&req->r_osd_item, &req->r_osd->o_requests); 1337f671b581SIlya Dryomov list_move_tail(&req->r_req_lru_item, &osdc->req_unsent); 1338f671b581SIlya Dryomov } else { 1339f671b581SIlya Dryomov list_move_tail(&req->r_req_lru_item, &osdc->req_notarget); 1340f671b581SIlya Dryomov } 1341f671b581SIlya Dryomov } 1342f671b581SIlya Dryomov 134317a13e40SIlya Dryomov /* 13443d14c5d2SYehuda Sadeh * Pick an osd (the first 'up' osd in the pg), allocate the osd struct 13453d14c5d2SYehuda Sadeh * (as needed), and set the request r_osd appropriately. If there is 134625985edcSLucas De Marchi * no up osd, set r_osd to NULL. Move the request to the appropriate list 13476f6c7006SSage Weil * (unsent, homeless) or leave on in-flight lru. 13483d14c5d2SYehuda Sadeh * 13493d14c5d2SYehuda Sadeh * Return 0 if unchanged, 1 if changed, or negative on error. 13503d14c5d2SYehuda Sadeh * 13513d14c5d2SYehuda Sadeh * Caller should hold map_sem for read and request_mutex. 13523d14c5d2SYehuda Sadeh */ 13536f6c7006SSage Weil static int __map_request(struct ceph_osd_client *osdc, 135438d6453cSSage Weil struct ceph_osd_request *req, int force_resend) 13553d14c5d2SYehuda Sadeh { 13565b191d99SSage Weil struct ceph_pg pgid; 13573d14c5d2SYehuda Sadeh int acting[CEPH_PG_MAX_SIZE]; 13588008ab10SIlya Dryomov int num, o; 13593d14c5d2SYehuda Sadeh int err; 1360d29adb34SJosh Durgin bool was_paused; 13613d14c5d2SYehuda Sadeh 13626f6c7006SSage Weil dout("map_request %p tid %lld\n", req, req->r_tid); 136317a13e40SIlya Dryomov 136417a13e40SIlya Dryomov err = __calc_request_pg(osdc->osdmap, req, &pgid); 13656f6c7006SSage Weil if (err) { 13666f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 13673d14c5d2SYehuda Sadeh return err; 13686f6c7006SSage Weil } 13693d14c5d2SYehuda Sadeh req->r_pgid = pgid; 13703d14c5d2SYehuda Sadeh 13718008ab10SIlya Dryomov num = ceph_calc_pg_acting(osdc->osdmap, pgid, acting, &o); 13728008ab10SIlya Dryomov if (num < 0) 13738008ab10SIlya Dryomov num = 0; 13743d14c5d2SYehuda Sadeh 1375d29adb34SJosh Durgin was_paused = req->r_paused; 1376d29adb34SJosh Durgin req->r_paused = __req_should_be_paused(osdc, req); 1377d29adb34SJosh Durgin if (was_paused && !req->r_paused) 1378d29adb34SJosh Durgin force_resend = 1; 1379d29adb34SJosh Durgin 138038d6453cSSage Weil if ((!force_resend && 138138d6453cSSage Weil req->r_osd && req->r_osd->o_osd == o && 13823d14c5d2SYehuda Sadeh req->r_sent >= req->r_osd->o_incarnation && 13833d14c5d2SYehuda Sadeh req->r_num_pg_osds == num && 13843d14c5d2SYehuda Sadeh memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) || 1385d29adb34SJosh Durgin (req->r_osd == NULL && o == -1) || 1386d29adb34SJosh Durgin req->r_paused) 13873d14c5d2SYehuda Sadeh return 0; /* no change */ 13883d14c5d2SYehuda Sadeh 13895b191d99SSage Weil dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n", 13905b191d99SSage Weil req->r_tid, pgid.pool, pgid.seed, o, 13913d14c5d2SYehuda Sadeh req->r_osd ? req->r_osd->o_osd : -1); 13923d14c5d2SYehuda Sadeh 13933d14c5d2SYehuda Sadeh /* record full pg acting set */ 13943d14c5d2SYehuda Sadeh memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num); 13953d14c5d2SYehuda Sadeh req->r_num_pg_osds = num; 13963d14c5d2SYehuda Sadeh 13973d14c5d2SYehuda Sadeh if (req->r_osd) { 13983d14c5d2SYehuda Sadeh __cancel_request(req); 13993d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 1400a390de02SIlya Dryomov list_del_init(&req->r_linger_osd_item); 14013d14c5d2SYehuda Sadeh req->r_osd = NULL; 14023d14c5d2SYehuda Sadeh } 14033d14c5d2SYehuda Sadeh 14043d14c5d2SYehuda Sadeh req->r_osd = __lookup_osd(osdc, o); 14053d14c5d2SYehuda Sadeh if (!req->r_osd && o >= 0) { 14063d14c5d2SYehuda Sadeh err = -ENOMEM; 1407e10006f8SAlex Elder req->r_osd = create_osd(osdc, o); 14086f6c7006SSage Weil if (!req->r_osd) { 14096f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 14103d14c5d2SYehuda Sadeh goto out; 14116f6c7006SSage Weil } 14123d14c5d2SYehuda Sadeh 14136f6c7006SSage Weil dout("map_request osd %p is osd%d\n", req->r_osd, o); 14143d14c5d2SYehuda Sadeh __insert_osd(osdc, req->r_osd); 14153d14c5d2SYehuda Sadeh 1416b7a9e5ddSSage Weil ceph_con_open(&req->r_osd->o_con, 1417b7a9e5ddSSage Weil CEPH_ENTITY_TYPE_OSD, o, 1418b7a9e5ddSSage Weil &osdc->osdmap->osd_addr[o]); 14193d14c5d2SYehuda Sadeh } 14203d14c5d2SYehuda Sadeh 1421f671b581SIlya Dryomov __enqueue_request(req); 14223d14c5d2SYehuda Sadeh err = 1; /* osd or pg changed */ 14233d14c5d2SYehuda Sadeh 14243d14c5d2SYehuda Sadeh out: 14253d14c5d2SYehuda Sadeh return err; 14263d14c5d2SYehuda Sadeh } 14273d14c5d2SYehuda Sadeh 14283d14c5d2SYehuda Sadeh /* 14293d14c5d2SYehuda Sadeh * caller should hold map_sem (for read) and request_mutex 14303d14c5d2SYehuda Sadeh */ 143156e925b6SSage Weil static void __send_request(struct ceph_osd_client *osdc, 14323d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 14333d14c5d2SYehuda Sadeh { 14341b83bef2SSage Weil void *p; 14353d14c5d2SYehuda Sadeh 14361b83bef2SSage Weil dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n", 14371b83bef2SSage Weil req, req->r_tid, req->r_osd->o_osd, req->r_flags, 14381b83bef2SSage Weil (unsigned long long)req->r_pgid.pool, req->r_pgid.seed); 14393d14c5d2SYehuda Sadeh 14401b83bef2SSage Weil /* fill in message content that changes each time we send it */ 14411b83bef2SSage Weil put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch); 14421b83bef2SSage Weil put_unaligned_le32(req->r_flags, req->r_request_flags); 1443205ee118SIlya Dryomov put_unaligned_le64(req->r_target_oloc.pool, req->r_request_pool); 14441b83bef2SSage Weil p = req->r_request_pgid; 14451b83bef2SSage Weil ceph_encode_64(&p, req->r_pgid.pool); 14461b83bef2SSage Weil ceph_encode_32(&p, req->r_pgid.seed); 14471b83bef2SSage Weil put_unaligned_le64(1, req->r_request_attempts); /* FIXME */ 14481b83bef2SSage Weil memcpy(req->r_request_reassert_version, &req->r_reassert_version, 14491b83bef2SSage Weil sizeof(req->r_reassert_version)); 14502169aea6SSage Weil 14513d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 14523d14c5d2SYehuda Sadeh list_move_tail(&req->r_req_lru_item, &osdc->req_lru); 14533d14c5d2SYehuda Sadeh 14543d14c5d2SYehuda Sadeh ceph_msg_get(req->r_request); /* send consumes a ref */ 145526be8808SAlex Elder 14563d14c5d2SYehuda Sadeh req->r_sent = req->r_osd->o_incarnation; 145726be8808SAlex Elder 145826be8808SAlex Elder ceph_con_send(&req->r_osd->o_con, req->r_request); 14593d14c5d2SYehuda Sadeh } 14603d14c5d2SYehuda Sadeh 14613d14c5d2SYehuda Sadeh /* 14626f6c7006SSage Weil * Send any requests in the queue (req_unsent). 14636f6c7006SSage Weil */ 1464f9d25199SAlex Elder static void __send_queued(struct ceph_osd_client *osdc) 14656f6c7006SSage Weil { 14666f6c7006SSage Weil struct ceph_osd_request *req, *tmp; 14676f6c7006SSage Weil 1468f9d25199SAlex Elder dout("__send_queued\n"); 1469f9d25199SAlex Elder list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) 14706f6c7006SSage Weil __send_request(osdc, req); 14716f6c7006SSage Weil } 14726f6c7006SSage Weil 14736f6c7006SSage Weil /* 14740bbfdfe8SIlya Dryomov * Caller should hold map_sem for read and request_mutex. 14750bbfdfe8SIlya Dryomov */ 14760bbfdfe8SIlya Dryomov static int __ceph_osdc_start_request(struct ceph_osd_client *osdc, 14770bbfdfe8SIlya Dryomov struct ceph_osd_request *req, 14780bbfdfe8SIlya Dryomov bool nofail) 14790bbfdfe8SIlya Dryomov { 14800bbfdfe8SIlya Dryomov int rc; 14810bbfdfe8SIlya Dryomov 14820bbfdfe8SIlya Dryomov __register_request(osdc, req); 14830bbfdfe8SIlya Dryomov req->r_sent = 0; 14840bbfdfe8SIlya Dryomov req->r_got_reply = 0; 14850bbfdfe8SIlya Dryomov rc = __map_request(osdc, req, 0); 14860bbfdfe8SIlya Dryomov if (rc < 0) { 14870bbfdfe8SIlya Dryomov if (nofail) { 14880bbfdfe8SIlya Dryomov dout("osdc_start_request failed map, " 14890bbfdfe8SIlya Dryomov " will retry %lld\n", req->r_tid); 14900bbfdfe8SIlya Dryomov rc = 0; 14910bbfdfe8SIlya Dryomov } else { 14920bbfdfe8SIlya Dryomov __unregister_request(osdc, req); 14930bbfdfe8SIlya Dryomov } 14940bbfdfe8SIlya Dryomov return rc; 14950bbfdfe8SIlya Dryomov } 14960bbfdfe8SIlya Dryomov 14970bbfdfe8SIlya Dryomov if (req->r_osd == NULL) { 14980bbfdfe8SIlya Dryomov dout("send_request %p no up osds in pg\n", req); 14990bbfdfe8SIlya Dryomov ceph_monc_request_next_osdmap(&osdc->client->monc); 15000bbfdfe8SIlya Dryomov } else { 15010bbfdfe8SIlya Dryomov __send_queued(osdc); 15020bbfdfe8SIlya Dryomov } 15030bbfdfe8SIlya Dryomov 15040bbfdfe8SIlya Dryomov return 0; 15050bbfdfe8SIlya Dryomov } 15060bbfdfe8SIlya Dryomov 15070bbfdfe8SIlya Dryomov /* 15083d14c5d2SYehuda Sadeh * Timeout callback, called every N seconds when 1 or more osd 15093d14c5d2SYehuda Sadeh * requests has been active for more than N seconds. When this 15103d14c5d2SYehuda Sadeh * happens, we ping all OSDs with requests who have timed out to 15113d14c5d2SYehuda Sadeh * ensure any communications channel reset is detected. Reset the 15123d14c5d2SYehuda Sadeh * request timeouts another N seconds in the future as we go. 15133d14c5d2SYehuda Sadeh * Reschedule the timeout event another N seconds in future (unless 15143d14c5d2SYehuda Sadeh * there are no open requests). 15153d14c5d2SYehuda Sadeh */ 15163d14c5d2SYehuda Sadeh static void handle_timeout(struct work_struct *work) 15173d14c5d2SYehuda Sadeh { 15183d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 15193d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, timeout_work.work); 152083aff95eSSage Weil struct ceph_osd_request *req; 15213d14c5d2SYehuda Sadeh struct ceph_osd *osd; 15223d14c5d2SYehuda Sadeh unsigned long keepalive = 15233d14c5d2SYehuda Sadeh osdc->client->options->osd_keepalive_timeout * HZ; 15243d14c5d2SYehuda Sadeh struct list_head slow_osds; 15253d14c5d2SYehuda Sadeh dout("timeout\n"); 15263d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 15273d14c5d2SYehuda Sadeh 15283d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 15293d14c5d2SYehuda Sadeh 15303d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 15313d14c5d2SYehuda Sadeh 15323d14c5d2SYehuda Sadeh /* 15333d14c5d2SYehuda Sadeh * ping osds that are a bit slow. this ensures that if there 15343d14c5d2SYehuda Sadeh * is a break in the TCP connection we will notice, and reopen 15353d14c5d2SYehuda Sadeh * a connection with that osd (from the fault callback). 15363d14c5d2SYehuda Sadeh */ 15373d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&slow_osds); 15383d14c5d2SYehuda Sadeh list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) { 15393d14c5d2SYehuda Sadeh if (time_before(jiffies, req->r_stamp + keepalive)) 15403d14c5d2SYehuda Sadeh break; 15413d14c5d2SYehuda Sadeh 15423d14c5d2SYehuda Sadeh osd = req->r_osd; 15433d14c5d2SYehuda Sadeh BUG_ON(!osd); 15443d14c5d2SYehuda Sadeh dout(" tid %llu is slow, will send keepalive on osd%d\n", 15453d14c5d2SYehuda Sadeh req->r_tid, osd->o_osd); 15463d14c5d2SYehuda Sadeh list_move_tail(&osd->o_keepalive_item, &slow_osds); 15473d14c5d2SYehuda Sadeh } 15483d14c5d2SYehuda Sadeh while (!list_empty(&slow_osds)) { 15493d14c5d2SYehuda Sadeh osd = list_entry(slow_osds.next, struct ceph_osd, 15503d14c5d2SYehuda Sadeh o_keepalive_item); 15513d14c5d2SYehuda Sadeh list_del_init(&osd->o_keepalive_item); 15523d14c5d2SYehuda Sadeh ceph_con_keepalive(&osd->o_con); 15533d14c5d2SYehuda Sadeh } 15543d14c5d2SYehuda Sadeh 15553d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 1556f9d25199SAlex Elder __send_queued(osdc); 15573d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 15583d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 15593d14c5d2SYehuda Sadeh } 15603d14c5d2SYehuda Sadeh 15613d14c5d2SYehuda Sadeh static void handle_osds_timeout(struct work_struct *work) 15623d14c5d2SYehuda Sadeh { 15633d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 15643d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, 15653d14c5d2SYehuda Sadeh osds_timeout_work.work); 15663d14c5d2SYehuda Sadeh unsigned long delay = 15673d14c5d2SYehuda Sadeh osdc->client->options->osd_idle_ttl * HZ >> 2; 15683d14c5d2SYehuda Sadeh 15693d14c5d2SYehuda Sadeh dout("osds timeout\n"); 15703d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 1571aca420bcSSage Weil remove_old_osds(osdc); 15723d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 15733d14c5d2SYehuda Sadeh 15743d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 15753d14c5d2SYehuda Sadeh round_jiffies_relative(delay)); 15763d14c5d2SYehuda Sadeh } 15773d14c5d2SYehuda Sadeh 1578205ee118SIlya Dryomov static int ceph_oloc_decode(void **p, void *end, 1579205ee118SIlya Dryomov struct ceph_object_locator *oloc) 1580205ee118SIlya Dryomov { 1581205ee118SIlya Dryomov u8 struct_v, struct_cv; 1582205ee118SIlya Dryomov u32 len; 1583205ee118SIlya Dryomov void *struct_end; 1584205ee118SIlya Dryomov int ret = 0; 1585205ee118SIlya Dryomov 1586205ee118SIlya Dryomov ceph_decode_need(p, end, 1 + 1 + 4, e_inval); 1587205ee118SIlya Dryomov struct_v = ceph_decode_8(p); 1588205ee118SIlya Dryomov struct_cv = ceph_decode_8(p); 1589205ee118SIlya Dryomov if (struct_v < 3) { 1590205ee118SIlya Dryomov pr_warn("got v %d < 3 cv %d of ceph_object_locator\n", 1591205ee118SIlya Dryomov struct_v, struct_cv); 1592205ee118SIlya Dryomov goto e_inval; 1593205ee118SIlya Dryomov } 1594205ee118SIlya Dryomov if (struct_cv > 6) { 1595205ee118SIlya Dryomov pr_warn("got v %d cv %d > 6 of ceph_object_locator\n", 1596205ee118SIlya Dryomov struct_v, struct_cv); 1597205ee118SIlya Dryomov goto e_inval; 1598205ee118SIlya Dryomov } 1599205ee118SIlya Dryomov len = ceph_decode_32(p); 1600205ee118SIlya Dryomov ceph_decode_need(p, end, len, e_inval); 1601205ee118SIlya Dryomov struct_end = *p + len; 1602205ee118SIlya Dryomov 1603205ee118SIlya Dryomov oloc->pool = ceph_decode_64(p); 1604205ee118SIlya Dryomov *p += 4; /* skip preferred */ 1605205ee118SIlya Dryomov 1606205ee118SIlya Dryomov len = ceph_decode_32(p); 1607205ee118SIlya Dryomov if (len > 0) { 1608205ee118SIlya Dryomov pr_warn("ceph_object_locator::key is set\n"); 1609205ee118SIlya Dryomov goto e_inval; 1610205ee118SIlya Dryomov } 1611205ee118SIlya Dryomov 1612205ee118SIlya Dryomov if (struct_v >= 5) { 1613205ee118SIlya Dryomov len = ceph_decode_32(p); 1614205ee118SIlya Dryomov if (len > 0) { 1615205ee118SIlya Dryomov pr_warn("ceph_object_locator::nspace is set\n"); 1616205ee118SIlya Dryomov goto e_inval; 1617205ee118SIlya Dryomov } 1618205ee118SIlya Dryomov } 1619205ee118SIlya Dryomov 1620205ee118SIlya Dryomov if (struct_v >= 6) { 1621205ee118SIlya Dryomov s64 hash = ceph_decode_64(p); 1622205ee118SIlya Dryomov if (hash != -1) { 1623205ee118SIlya Dryomov pr_warn("ceph_object_locator::hash is set\n"); 1624205ee118SIlya Dryomov goto e_inval; 1625205ee118SIlya Dryomov } 1626205ee118SIlya Dryomov } 1627205ee118SIlya Dryomov 1628205ee118SIlya Dryomov /* skip the rest */ 1629205ee118SIlya Dryomov *p = struct_end; 1630205ee118SIlya Dryomov out: 1631205ee118SIlya Dryomov return ret; 1632205ee118SIlya Dryomov 1633205ee118SIlya Dryomov e_inval: 1634205ee118SIlya Dryomov ret = -EINVAL; 1635205ee118SIlya Dryomov goto out; 1636205ee118SIlya Dryomov } 1637205ee118SIlya Dryomov 1638205ee118SIlya Dryomov static int ceph_redirect_decode(void **p, void *end, 1639205ee118SIlya Dryomov struct ceph_request_redirect *redir) 1640205ee118SIlya Dryomov { 1641205ee118SIlya Dryomov u8 struct_v, struct_cv; 1642205ee118SIlya Dryomov u32 len; 1643205ee118SIlya Dryomov void *struct_end; 1644205ee118SIlya Dryomov int ret; 1645205ee118SIlya Dryomov 1646205ee118SIlya Dryomov ceph_decode_need(p, end, 1 + 1 + 4, e_inval); 1647205ee118SIlya Dryomov struct_v = ceph_decode_8(p); 1648205ee118SIlya Dryomov struct_cv = ceph_decode_8(p); 1649205ee118SIlya Dryomov if (struct_cv > 1) { 1650205ee118SIlya Dryomov pr_warn("got v %d cv %d > 1 of ceph_request_redirect\n", 1651205ee118SIlya Dryomov struct_v, struct_cv); 1652205ee118SIlya Dryomov goto e_inval; 1653205ee118SIlya Dryomov } 1654205ee118SIlya Dryomov len = ceph_decode_32(p); 1655205ee118SIlya Dryomov ceph_decode_need(p, end, len, e_inval); 1656205ee118SIlya Dryomov struct_end = *p + len; 1657205ee118SIlya Dryomov 1658205ee118SIlya Dryomov ret = ceph_oloc_decode(p, end, &redir->oloc); 1659205ee118SIlya Dryomov if (ret) 1660205ee118SIlya Dryomov goto out; 1661205ee118SIlya Dryomov 1662205ee118SIlya Dryomov len = ceph_decode_32(p); 1663205ee118SIlya Dryomov if (len > 0) { 1664205ee118SIlya Dryomov pr_warn("ceph_request_redirect::object_name is set\n"); 1665205ee118SIlya Dryomov goto e_inval; 1666205ee118SIlya Dryomov } 1667205ee118SIlya Dryomov 1668205ee118SIlya Dryomov len = ceph_decode_32(p); 1669205ee118SIlya Dryomov *p += len; /* skip osd_instructions */ 1670205ee118SIlya Dryomov 1671205ee118SIlya Dryomov /* skip the rest */ 1672205ee118SIlya Dryomov *p = struct_end; 1673205ee118SIlya Dryomov out: 1674205ee118SIlya Dryomov return ret; 1675205ee118SIlya Dryomov 1676205ee118SIlya Dryomov e_inval: 1677205ee118SIlya Dryomov ret = -EINVAL; 1678205ee118SIlya Dryomov goto out; 1679205ee118SIlya Dryomov } 1680205ee118SIlya Dryomov 168125845472SSage Weil static void complete_request(struct ceph_osd_request *req) 168225845472SSage Weil { 168325845472SSage Weil complete_all(&req->r_safe_completion); /* fsync waiter */ 168425845472SSage Weil } 168525845472SSage Weil 16863d14c5d2SYehuda Sadeh /* 16873d14c5d2SYehuda Sadeh * handle osd op reply. either call the callback if it is specified, 16883d14c5d2SYehuda Sadeh * or do the completion to wake up the waiting thread. 16893d14c5d2SYehuda Sadeh */ 16903d14c5d2SYehuda Sadeh static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg, 16913d14c5d2SYehuda Sadeh struct ceph_connection *con) 16923d14c5d2SYehuda Sadeh { 16931b83bef2SSage Weil void *p, *end; 16943d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 1695205ee118SIlya Dryomov struct ceph_request_redirect redir; 16963d14c5d2SYehuda Sadeh u64 tid; 16971b83bef2SSage Weil int object_len; 169879528734SAlex Elder unsigned int numops; 169979528734SAlex Elder int payload_len, flags; 17003d14c5d2SYehuda Sadeh s32 result; 17011b83bef2SSage Weil s32 retry_attempt; 17021b83bef2SSage Weil struct ceph_pg pg; 17031b83bef2SSage Weil int err; 17041b83bef2SSage Weil u32 reassert_epoch; 17051b83bef2SSage Weil u64 reassert_version; 17061b83bef2SSage Weil u32 osdmap_epoch; 17070d5af164SAlex Elder int already_completed; 17089fc6e064SAlex Elder u32 bytes; 170979528734SAlex Elder unsigned int i; 17103d14c5d2SYehuda Sadeh 17113d14c5d2SYehuda Sadeh tid = le64_to_cpu(msg->hdr.tid); 17121b83bef2SSage Weil dout("handle_reply %p tid %llu\n", msg, tid); 17131b83bef2SSage Weil 17141b83bef2SSage Weil p = msg->front.iov_base; 17151b83bef2SSage Weil end = p + msg->front.iov_len; 17161b83bef2SSage Weil 17171b83bef2SSage Weil ceph_decode_need(&p, end, 4, bad); 17181b83bef2SSage Weil object_len = ceph_decode_32(&p); 17191b83bef2SSage Weil ceph_decode_need(&p, end, object_len, bad); 17201b83bef2SSage Weil p += object_len; 17211b83bef2SSage Weil 1722ef4859d6SAlex Elder err = ceph_decode_pgid(&p, end, &pg); 17231b83bef2SSage Weil if (err) 17243d14c5d2SYehuda Sadeh goto bad; 17251b83bef2SSage Weil 17261b83bef2SSage Weil ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad); 17271b83bef2SSage Weil flags = ceph_decode_64(&p); 17281b83bef2SSage Weil result = ceph_decode_32(&p); 17291b83bef2SSage Weil reassert_epoch = ceph_decode_32(&p); 17301b83bef2SSage Weil reassert_version = ceph_decode_64(&p); 17311b83bef2SSage Weil osdmap_epoch = ceph_decode_32(&p); 17321b83bef2SSage Weil 17333d14c5d2SYehuda Sadeh /* lookup */ 1734ff513aceSIlya Dryomov down_read(&osdc->map_sem); 17353d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 17363d14c5d2SYehuda Sadeh req = __lookup_request(osdc, tid); 17373d14c5d2SYehuda Sadeh if (req == NULL) { 17383d14c5d2SYehuda Sadeh dout("handle_reply tid %llu dne\n", tid); 17398058fd45SAlex Elder goto bad_mutex; 17403d14c5d2SYehuda Sadeh } 17413d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 17421b83bef2SSage Weil 17431b83bef2SSage Weil dout("handle_reply %p tid %llu req %p result %d\n", msg, tid, 17441b83bef2SSage Weil req, result); 17451b83bef2SSage Weil 174618741196SDan Carpenter ceph_decode_need(&p, end, 4, bad_put); 17471b83bef2SSage Weil numops = ceph_decode_32(&p); 17481b83bef2SSage Weil if (numops > CEPH_OSD_MAX_OP) 17491b83bef2SSage Weil goto bad_put; 17501b83bef2SSage Weil if (numops != req->r_num_ops) 17511b83bef2SSage Weil goto bad_put; 17521b83bef2SSage Weil payload_len = 0; 175318741196SDan Carpenter ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad_put); 17541b83bef2SSage Weil for (i = 0; i < numops; i++) { 17551b83bef2SSage Weil struct ceph_osd_op *op = p; 17561b83bef2SSage Weil int len; 17571b83bef2SSage Weil 17581b83bef2SSage Weil len = le32_to_cpu(op->payload_len); 17591b83bef2SSage Weil req->r_reply_op_len[i] = len; 17601b83bef2SSage Weil dout(" op %d has %d bytes\n", i, len); 17611b83bef2SSage Weil payload_len += len; 17621b83bef2SSage Weil p += sizeof(*op); 17631b83bef2SSage Weil } 17649fc6e064SAlex Elder bytes = le32_to_cpu(msg->hdr.data_len); 17659fc6e064SAlex Elder if (payload_len != bytes) { 1766b9a67899SJoe Perches pr_warn("sum of op payload lens %d != data_len %d\n", 17679fc6e064SAlex Elder payload_len, bytes); 17681b83bef2SSage Weil goto bad_put; 17691b83bef2SSage Weil } 17701b83bef2SSage Weil 177118741196SDan Carpenter ceph_decode_need(&p, end, 4 + numops * 4, bad_put); 17721b83bef2SSage Weil retry_attempt = ceph_decode_32(&p); 17731b83bef2SSage Weil for (i = 0; i < numops; i++) 17741b83bef2SSage Weil req->r_reply_op_result[i] = ceph_decode_32(&p); 17753d14c5d2SYehuda Sadeh 1776205ee118SIlya Dryomov if (le16_to_cpu(msg->hdr.version) >= 6) { 1777205ee118SIlya Dryomov p += 8 + 4; /* skip replay_version */ 1778205ee118SIlya Dryomov p += 8; /* skip user_version */ 1779205ee118SIlya Dryomov 1780205ee118SIlya Dryomov err = ceph_redirect_decode(&p, end, &redir); 1781205ee118SIlya Dryomov if (err) 1782205ee118SIlya Dryomov goto bad_put; 1783205ee118SIlya Dryomov } else { 1784205ee118SIlya Dryomov redir.oloc.pool = -1; 1785205ee118SIlya Dryomov } 1786205ee118SIlya Dryomov 1787205ee118SIlya Dryomov if (redir.oloc.pool != -1) { 1788205ee118SIlya Dryomov dout("redirect pool %lld\n", redir.oloc.pool); 1789205ee118SIlya Dryomov 1790205ee118SIlya Dryomov __unregister_request(osdc, req); 1791205ee118SIlya Dryomov 1792205ee118SIlya Dryomov req->r_target_oloc = redir.oloc; /* struct */ 1793205ee118SIlya Dryomov 1794205ee118SIlya Dryomov /* 1795205ee118SIlya Dryomov * Start redirect requests with nofail=true. If 1796205ee118SIlya Dryomov * mapping fails, request will end up on the notarget 1797205ee118SIlya Dryomov * list, waiting for the new osdmap (which can take 1798205ee118SIlya Dryomov * a while), even though the original request mapped 1799205ee118SIlya Dryomov * successfully. In the future we might want to follow 1800205ee118SIlya Dryomov * original request's nofail setting here. 1801205ee118SIlya Dryomov */ 1802ff513aceSIlya Dryomov err = __ceph_osdc_start_request(osdc, req, true); 1803205ee118SIlya Dryomov BUG_ON(err); 1804205ee118SIlya Dryomov 1805ff513aceSIlya Dryomov goto out_unlock; 1806205ee118SIlya Dryomov } 1807205ee118SIlya Dryomov 1808eb845ff1SYan, Zheng already_completed = req->r_got_reply; 18093d14c5d2SYehuda Sadeh if (!req->r_got_reply) { 18101b83bef2SSage Weil req->r_result = result; 18113d14c5d2SYehuda Sadeh dout("handle_reply result %d bytes %d\n", req->r_result, 18123d14c5d2SYehuda Sadeh bytes); 18133d14c5d2SYehuda Sadeh if (req->r_result == 0) 18143d14c5d2SYehuda Sadeh req->r_result = bytes; 18153d14c5d2SYehuda Sadeh 18163d14c5d2SYehuda Sadeh /* in case this is a write and we need to replay, */ 18171b83bef2SSage Weil req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch); 18181b83bef2SSage Weil req->r_reassert_version.version = cpu_to_le64(reassert_version); 18193d14c5d2SYehuda Sadeh 18203d14c5d2SYehuda Sadeh req->r_got_reply = 1; 18213d14c5d2SYehuda Sadeh } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) { 18223d14c5d2SYehuda Sadeh dout("handle_reply tid %llu dup ack\n", tid); 1823ff513aceSIlya Dryomov goto out_unlock; 18243d14c5d2SYehuda Sadeh } 18253d14c5d2SYehuda Sadeh 18263d14c5d2SYehuda Sadeh dout("handle_reply tid %llu flags %d\n", tid, flags); 18273d14c5d2SYehuda Sadeh 1828a40c4f10SYehuda Sadeh if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK)) 1829a40c4f10SYehuda Sadeh __register_linger_request(osdc, req); 1830a40c4f10SYehuda Sadeh 18313d14c5d2SYehuda Sadeh /* either this is a read, or we got the safe response */ 18323d14c5d2SYehuda Sadeh if (result < 0 || 18333d14c5d2SYehuda Sadeh (flags & CEPH_OSD_FLAG_ONDISK) || 18343d14c5d2SYehuda Sadeh ((flags & CEPH_OSD_FLAG_WRITE) == 0)) 18353d14c5d2SYehuda Sadeh __unregister_request(osdc, req); 18363d14c5d2SYehuda Sadeh 18373d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 1838ff513aceSIlya Dryomov up_read(&osdc->map_sem); 18393d14c5d2SYehuda Sadeh 1840eb845ff1SYan, Zheng if (!already_completed) { 184161c5d6bfSYan, Zheng if (req->r_unsafe_callback && 184261c5d6bfSYan, Zheng result >= 0 && !(flags & CEPH_OSD_FLAG_ONDISK)) 184361c5d6bfSYan, Zheng req->r_unsafe_callback(req, true); 18443d14c5d2SYehuda Sadeh if (req->r_callback) 18453d14c5d2SYehuda Sadeh req->r_callback(req, msg); 18463d14c5d2SYehuda Sadeh else 18473d14c5d2SYehuda Sadeh complete_all(&req->r_completion); 1848eb845ff1SYan, Zheng } 18493d14c5d2SYehuda Sadeh 185061c5d6bfSYan, Zheng if (flags & CEPH_OSD_FLAG_ONDISK) { 185161c5d6bfSYan, Zheng if (req->r_unsafe_callback && already_completed) 185261c5d6bfSYan, Zheng req->r_unsafe_callback(req, false); 185325845472SSage Weil complete_request(req); 185461c5d6bfSYan, Zheng } 18553d14c5d2SYehuda Sadeh 1856ff513aceSIlya Dryomov out: 1857a40c4f10SYehuda Sadeh dout("req=%p req->r_linger=%d\n", req, req->r_linger); 18583d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 18593d14c5d2SYehuda Sadeh return; 1860ff513aceSIlya Dryomov out_unlock: 1861ff513aceSIlya Dryomov mutex_unlock(&osdc->request_mutex); 1862ff513aceSIlya Dryomov up_read(&osdc->map_sem); 1863ff513aceSIlya Dryomov goto out; 18643d14c5d2SYehuda Sadeh 18651b83bef2SSage Weil bad_put: 186637c89bdeSLi Wang req->r_result = -EIO; 186737c89bdeSLi Wang __unregister_request(osdc, req); 186837c89bdeSLi Wang if (req->r_callback) 186937c89bdeSLi Wang req->r_callback(req, msg); 187037c89bdeSLi Wang else 187137c89bdeSLi Wang complete_all(&req->r_completion); 187237c89bdeSLi Wang complete_request(req); 18731b83bef2SSage Weil ceph_osdc_put_request(req); 18748058fd45SAlex Elder bad_mutex: 18758058fd45SAlex Elder mutex_unlock(&osdc->request_mutex); 1876ff513aceSIlya Dryomov up_read(&osdc->map_sem); 18773d14c5d2SYehuda Sadeh bad: 18781b83bef2SSage Weil pr_err("corrupt osd_op_reply got %d %d\n", 18791b83bef2SSage Weil (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len)); 18803d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 18813d14c5d2SYehuda Sadeh } 18823d14c5d2SYehuda Sadeh 18836f6c7006SSage Weil static void reset_changed_osds(struct ceph_osd_client *osdc) 18843d14c5d2SYehuda Sadeh { 18853d14c5d2SYehuda Sadeh struct rb_node *p, *n; 18863d14c5d2SYehuda Sadeh 18873d14c5d2SYehuda Sadeh for (p = rb_first(&osdc->osds); p; p = n) { 18886f6c7006SSage Weil struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node); 18893d14c5d2SYehuda Sadeh 18903d14c5d2SYehuda Sadeh n = rb_next(p); 18913d14c5d2SYehuda Sadeh if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) || 18923d14c5d2SYehuda Sadeh memcmp(&osd->o_con.peer_addr, 18933d14c5d2SYehuda Sadeh ceph_osd_addr(osdc->osdmap, 18943d14c5d2SYehuda Sadeh osd->o_osd), 18953d14c5d2SYehuda Sadeh sizeof(struct ceph_entity_addr)) != 0) 18963d14c5d2SYehuda Sadeh __reset_osd(osdc, osd); 18973d14c5d2SYehuda Sadeh } 18983d14c5d2SYehuda Sadeh } 18993d14c5d2SYehuda Sadeh 19003d14c5d2SYehuda Sadeh /* 19016f6c7006SSage Weil * Requeue requests whose mapping to an OSD has changed. If requests map to 19026f6c7006SSage Weil * no osd, request a new map. 19033d14c5d2SYehuda Sadeh * 1904e6d50f67SAlex Elder * Caller should hold map_sem for read. 19053d14c5d2SYehuda Sadeh */ 19069a1ea2dbSJosh Durgin static void kick_requests(struct ceph_osd_client *osdc, bool force_resend, 19079a1ea2dbSJosh Durgin bool force_resend_writes) 19083d14c5d2SYehuda Sadeh { 1909a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 19106f6c7006SSage Weil struct rb_node *p; 19116f6c7006SSage Weil int needmap = 0; 19126f6c7006SSage Weil int err; 19139a1ea2dbSJosh Durgin bool force_resend_req; 19143d14c5d2SYehuda Sadeh 19159a1ea2dbSJosh Durgin dout("kick_requests %s %s\n", force_resend ? " (force resend)" : "", 19169a1ea2dbSJosh Durgin force_resend_writes ? " (force resend writes)" : ""); 19173d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 19186194ea89SSage Weil for (p = rb_first(&osdc->requests); p; ) { 19196f6c7006SSage Weil req = rb_entry(p, struct ceph_osd_request, r_node); 19206194ea89SSage Weil p = rb_next(p); 1921ab60b16dSAlex Elder 1922ab60b16dSAlex Elder /* 1923ab60b16dSAlex Elder * For linger requests that have not yet been 1924ab60b16dSAlex Elder * registered, move them to the linger list; they'll 1925ab60b16dSAlex Elder * be sent to the osd in the loop below. Unregister 1926ab60b16dSAlex Elder * the request before re-registering it as a linger 1927ab60b16dSAlex Elder * request to ensure the __map_request() below 1928ab60b16dSAlex Elder * will decide it needs to be sent. 1929ab60b16dSAlex Elder */ 1930ab60b16dSAlex Elder if (req->r_linger && list_empty(&req->r_linger_item)) { 1931ab60b16dSAlex Elder dout("%p tid %llu restart on osd%d\n", 1932ab60b16dSAlex Elder req, req->r_tid, 1933ab60b16dSAlex Elder req->r_osd ? req->r_osd->o_osd : -1); 193496e4dac6SAlex Elder ceph_osdc_get_request(req); 1935ab60b16dSAlex Elder __unregister_request(osdc, req); 1936ab60b16dSAlex Elder __register_linger_request(osdc, req); 193796e4dac6SAlex Elder ceph_osdc_put_request(req); 1938ab60b16dSAlex Elder continue; 1939ab60b16dSAlex Elder } 1940ab60b16dSAlex Elder 19419a1ea2dbSJosh Durgin force_resend_req = force_resend || 19429a1ea2dbSJosh Durgin (force_resend_writes && 19439a1ea2dbSJosh Durgin req->r_flags & CEPH_OSD_FLAG_WRITE); 19449a1ea2dbSJosh Durgin err = __map_request(osdc, req, force_resend_req); 19456f6c7006SSage Weil if (err < 0) 19466f6c7006SSage Weil continue; /* error */ 19476f6c7006SSage Weil if (req->r_osd == NULL) { 19486f6c7006SSage Weil dout("%p tid %llu maps to no osd\n", req, req->r_tid); 19496f6c7006SSage Weil needmap++; /* request a newer map */ 19506f6c7006SSage Weil } else if (err > 0) { 19516194ea89SSage Weil if (!req->r_linger) { 19526194ea89SSage Weil dout("%p tid %llu requeued on osd%d\n", req, 19536194ea89SSage Weil req->r_tid, 19546f6c7006SSage Weil req->r_osd ? req->r_osd->o_osd : -1); 19556f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 19566f6c7006SSage Weil } 19576f6c7006SSage Weil } 19586194ea89SSage Weil } 1959a40c4f10SYehuda Sadeh 1960a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osdc->req_linger, 1961a40c4f10SYehuda Sadeh r_linger_item) { 1962a40c4f10SYehuda Sadeh dout("linger req=%p req->r_osd=%p\n", req, req->r_osd); 1963a40c4f10SYehuda Sadeh 19649a1ea2dbSJosh Durgin err = __map_request(osdc, req, 19659a1ea2dbSJosh Durgin force_resend || force_resend_writes); 1966ab60b16dSAlex Elder dout("__map_request returned %d\n", err); 1967a40c4f10SYehuda Sadeh if (err == 0) 1968a40c4f10SYehuda Sadeh continue; /* no change and no osd was specified */ 1969a40c4f10SYehuda Sadeh if (err < 0) 1970a40c4f10SYehuda Sadeh continue; /* hrm! */ 1971a40c4f10SYehuda Sadeh if (req->r_osd == NULL) { 1972a40c4f10SYehuda Sadeh dout("tid %llu maps to no valid osd\n", req->r_tid); 1973a40c4f10SYehuda Sadeh needmap++; /* request a newer map */ 1974a40c4f10SYehuda Sadeh continue; 1975a40c4f10SYehuda Sadeh } 1976a40c4f10SYehuda Sadeh 1977a40c4f10SYehuda Sadeh dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid, 1978a40c4f10SYehuda Sadeh req->r_osd ? req->r_osd->o_osd : -1); 1979a40c4f10SYehuda Sadeh __register_request(osdc, req); 1980c89ce05eSAlex Elder __unregister_linger_request(osdc, req); 1981a40c4f10SYehuda Sadeh } 198214d2f38dSAlex Elder reset_changed_osds(osdc); 19833d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 19843d14c5d2SYehuda Sadeh 19853d14c5d2SYehuda Sadeh if (needmap) { 19863d14c5d2SYehuda Sadeh dout("%d requests for down osds, need new map\n", needmap); 19873d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 19883d14c5d2SYehuda Sadeh } 19893d14c5d2SYehuda Sadeh } 19906f6c7006SSage Weil 19916f6c7006SSage Weil 19923d14c5d2SYehuda Sadeh /* 19933d14c5d2SYehuda Sadeh * Process updated osd map. 19943d14c5d2SYehuda Sadeh * 19953d14c5d2SYehuda Sadeh * The message contains any number of incremental and full maps, normally 19963d14c5d2SYehuda Sadeh * indicating some sort of topology change in the cluster. Kick requests 19973d14c5d2SYehuda Sadeh * off to different OSDs as needed. 19983d14c5d2SYehuda Sadeh */ 19993d14c5d2SYehuda Sadeh void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg) 20003d14c5d2SYehuda Sadeh { 20013d14c5d2SYehuda Sadeh void *p, *end, *next; 20023d14c5d2SYehuda Sadeh u32 nr_maps, maplen; 20033d14c5d2SYehuda Sadeh u32 epoch; 20043d14c5d2SYehuda Sadeh struct ceph_osdmap *newmap = NULL, *oldmap; 20053d14c5d2SYehuda Sadeh int err; 20063d14c5d2SYehuda Sadeh struct ceph_fsid fsid; 20079a1ea2dbSJosh Durgin bool was_full; 20083d14c5d2SYehuda Sadeh 20093d14c5d2SYehuda Sadeh dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0); 20103d14c5d2SYehuda Sadeh p = msg->front.iov_base; 20113d14c5d2SYehuda Sadeh end = p + msg->front.iov_len; 20123d14c5d2SYehuda Sadeh 20133d14c5d2SYehuda Sadeh /* verify fsid */ 20143d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, sizeof(fsid), bad); 20153d14c5d2SYehuda Sadeh ceph_decode_copy(&p, &fsid, sizeof(fsid)); 20163d14c5d2SYehuda Sadeh if (ceph_check_fsid(osdc->client, &fsid) < 0) 20173d14c5d2SYehuda Sadeh return; 20183d14c5d2SYehuda Sadeh 20193d14c5d2SYehuda Sadeh down_write(&osdc->map_sem); 20203d14c5d2SYehuda Sadeh 20219a1ea2dbSJosh Durgin was_full = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL); 20229a1ea2dbSJosh Durgin 20233d14c5d2SYehuda Sadeh /* incremental maps */ 20243d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 20253d14c5d2SYehuda Sadeh dout(" %d inc maps\n", nr_maps); 20263d14c5d2SYehuda Sadeh while (nr_maps > 0) { 20273d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 20283d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 20293d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 20303d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 20313d14c5d2SYehuda Sadeh next = p + maplen; 20323d14c5d2SYehuda Sadeh if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) { 20333d14c5d2SYehuda Sadeh dout("applying incremental map %u len %d\n", 20343d14c5d2SYehuda Sadeh epoch, maplen); 20353d14c5d2SYehuda Sadeh newmap = osdmap_apply_incremental(&p, next, 20363d14c5d2SYehuda Sadeh osdc->osdmap, 203715d9882cSAlex Elder &osdc->client->msgr); 20383d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 20393d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 20403d14c5d2SYehuda Sadeh goto bad; 20413d14c5d2SYehuda Sadeh } 20423d14c5d2SYehuda Sadeh BUG_ON(!newmap); 20433d14c5d2SYehuda Sadeh if (newmap != osdc->osdmap) { 20443d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 20453d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 20463d14c5d2SYehuda Sadeh } 20479a1ea2dbSJosh Durgin was_full = was_full || 20489a1ea2dbSJosh Durgin ceph_osdmap_flag(osdc->osdmap, 20499a1ea2dbSJosh Durgin CEPH_OSDMAP_FULL); 20509a1ea2dbSJosh Durgin kick_requests(osdc, 0, was_full); 20513d14c5d2SYehuda Sadeh } else { 20523d14c5d2SYehuda Sadeh dout("ignoring incremental map %u len %d\n", 20533d14c5d2SYehuda Sadeh epoch, maplen); 20543d14c5d2SYehuda Sadeh } 20553d14c5d2SYehuda Sadeh p = next; 20563d14c5d2SYehuda Sadeh nr_maps--; 20573d14c5d2SYehuda Sadeh } 20583d14c5d2SYehuda Sadeh if (newmap) 20593d14c5d2SYehuda Sadeh goto done; 20603d14c5d2SYehuda Sadeh 20613d14c5d2SYehuda Sadeh /* full maps */ 20623d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 20633d14c5d2SYehuda Sadeh dout(" %d full maps\n", nr_maps); 20643d14c5d2SYehuda Sadeh while (nr_maps) { 20653d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 20663d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 20673d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 20683d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 20693d14c5d2SYehuda Sadeh if (nr_maps > 1) { 20703d14c5d2SYehuda Sadeh dout("skipping non-latest full map %u len %d\n", 20713d14c5d2SYehuda Sadeh epoch, maplen); 20723d14c5d2SYehuda Sadeh } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) { 20733d14c5d2SYehuda Sadeh dout("skipping full map %u len %d, " 20743d14c5d2SYehuda Sadeh "older than our %u\n", epoch, maplen, 20753d14c5d2SYehuda Sadeh osdc->osdmap->epoch); 20763d14c5d2SYehuda Sadeh } else { 207738d6453cSSage Weil int skipped_map = 0; 207838d6453cSSage Weil 20793d14c5d2SYehuda Sadeh dout("taking full map %u len %d\n", epoch, maplen); 2080a2505d63SIlya Dryomov newmap = ceph_osdmap_decode(&p, p+maplen); 20813d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 20823d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 20833d14c5d2SYehuda Sadeh goto bad; 20843d14c5d2SYehuda Sadeh } 20853d14c5d2SYehuda Sadeh BUG_ON(!newmap); 20863d14c5d2SYehuda Sadeh oldmap = osdc->osdmap; 20873d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 208838d6453cSSage Weil if (oldmap) { 208938d6453cSSage Weil if (oldmap->epoch + 1 < newmap->epoch) 209038d6453cSSage Weil skipped_map = 1; 20913d14c5d2SYehuda Sadeh ceph_osdmap_destroy(oldmap); 209238d6453cSSage Weil } 20939a1ea2dbSJosh Durgin was_full = was_full || 20949a1ea2dbSJosh Durgin ceph_osdmap_flag(osdc->osdmap, 20959a1ea2dbSJosh Durgin CEPH_OSDMAP_FULL); 20969a1ea2dbSJosh Durgin kick_requests(osdc, skipped_map, was_full); 20973d14c5d2SYehuda Sadeh } 20983d14c5d2SYehuda Sadeh p += maplen; 20993d14c5d2SYehuda Sadeh nr_maps--; 21003d14c5d2SYehuda Sadeh } 21013d14c5d2SYehuda Sadeh 2102b72e19b9SDan Carpenter if (!osdc->osdmap) 2103b72e19b9SDan Carpenter goto bad; 21043d14c5d2SYehuda Sadeh done: 21053d14c5d2SYehuda Sadeh downgrade_write(&osdc->map_sem); 21063d14c5d2SYehuda Sadeh ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch); 2107cd634fb6SSage Weil 2108cd634fb6SSage Weil /* 2109cd634fb6SSage Weil * subscribe to subsequent osdmap updates if full to ensure 2110cd634fb6SSage Weil * we find out when we are no longer full and stop returning 2111cd634fb6SSage Weil * ENOSPC. 2112cd634fb6SSage Weil */ 2113d29adb34SJosh Durgin if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) || 2114d29adb34SJosh Durgin ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD) || 2115d29adb34SJosh Durgin ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR)) 2116cd634fb6SSage Weil ceph_monc_request_next_osdmap(&osdc->client->monc); 2117cd634fb6SSage Weil 2118f9d25199SAlex Elder mutex_lock(&osdc->request_mutex); 2119f9d25199SAlex Elder __send_queued(osdc); 2120f9d25199SAlex Elder mutex_unlock(&osdc->request_mutex); 21213d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 21223d14c5d2SYehuda Sadeh wake_up_all(&osdc->client->auth_wq); 21233d14c5d2SYehuda Sadeh return; 21243d14c5d2SYehuda Sadeh 21253d14c5d2SYehuda Sadeh bad: 21263d14c5d2SYehuda Sadeh pr_err("osdc handle_map corrupt msg\n"); 21273d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 21283d14c5d2SYehuda Sadeh up_write(&osdc->map_sem); 21293d14c5d2SYehuda Sadeh } 21303d14c5d2SYehuda Sadeh 21313d14c5d2SYehuda Sadeh /* 2132a40c4f10SYehuda Sadeh * watch/notify callback event infrastructure 2133a40c4f10SYehuda Sadeh * 2134a40c4f10SYehuda Sadeh * These callbacks are used both for watch and notify operations. 2135a40c4f10SYehuda Sadeh */ 2136a40c4f10SYehuda Sadeh static void __release_event(struct kref *kref) 2137a40c4f10SYehuda Sadeh { 2138a40c4f10SYehuda Sadeh struct ceph_osd_event *event = 2139a40c4f10SYehuda Sadeh container_of(kref, struct ceph_osd_event, kref); 2140a40c4f10SYehuda Sadeh 2141a40c4f10SYehuda Sadeh dout("__release_event %p\n", event); 2142a40c4f10SYehuda Sadeh kfree(event); 2143a40c4f10SYehuda Sadeh } 2144a40c4f10SYehuda Sadeh 2145a40c4f10SYehuda Sadeh static void get_event(struct ceph_osd_event *event) 2146a40c4f10SYehuda Sadeh { 2147a40c4f10SYehuda Sadeh kref_get(&event->kref); 2148a40c4f10SYehuda Sadeh } 2149a40c4f10SYehuda Sadeh 2150a40c4f10SYehuda Sadeh void ceph_osdc_put_event(struct ceph_osd_event *event) 2151a40c4f10SYehuda Sadeh { 2152a40c4f10SYehuda Sadeh kref_put(&event->kref, __release_event); 2153a40c4f10SYehuda Sadeh } 2154a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_put_event); 2155a40c4f10SYehuda Sadeh 2156a40c4f10SYehuda Sadeh static void __insert_event(struct ceph_osd_client *osdc, 2157a40c4f10SYehuda Sadeh struct ceph_osd_event *new) 2158a40c4f10SYehuda Sadeh { 2159a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 2160a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 2161a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 2162a40c4f10SYehuda Sadeh 2163a40c4f10SYehuda Sadeh while (*p) { 2164a40c4f10SYehuda Sadeh parent = *p; 2165a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 2166a40c4f10SYehuda Sadeh if (new->cookie < event->cookie) 2167a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 2168a40c4f10SYehuda Sadeh else if (new->cookie > event->cookie) 2169a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 2170a40c4f10SYehuda Sadeh else 2171a40c4f10SYehuda Sadeh BUG(); 2172a40c4f10SYehuda Sadeh } 2173a40c4f10SYehuda Sadeh 2174a40c4f10SYehuda Sadeh rb_link_node(&new->node, parent, p); 2175a40c4f10SYehuda Sadeh rb_insert_color(&new->node, &osdc->event_tree); 2176a40c4f10SYehuda Sadeh } 2177a40c4f10SYehuda Sadeh 2178a40c4f10SYehuda Sadeh static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc, 2179a40c4f10SYehuda Sadeh u64 cookie) 2180a40c4f10SYehuda Sadeh { 2181a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 2182a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 2183a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 2184a40c4f10SYehuda Sadeh 2185a40c4f10SYehuda Sadeh while (*p) { 2186a40c4f10SYehuda Sadeh parent = *p; 2187a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 2188a40c4f10SYehuda Sadeh if (cookie < event->cookie) 2189a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 2190a40c4f10SYehuda Sadeh else if (cookie > event->cookie) 2191a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 2192a40c4f10SYehuda Sadeh else 2193a40c4f10SYehuda Sadeh return event; 2194a40c4f10SYehuda Sadeh } 2195a40c4f10SYehuda Sadeh return NULL; 2196a40c4f10SYehuda Sadeh } 2197a40c4f10SYehuda Sadeh 2198a40c4f10SYehuda Sadeh static void __remove_event(struct ceph_osd_event *event) 2199a40c4f10SYehuda Sadeh { 2200a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 2201a40c4f10SYehuda Sadeh 2202a40c4f10SYehuda Sadeh if (!RB_EMPTY_NODE(&event->node)) { 2203a40c4f10SYehuda Sadeh dout("__remove_event removed %p\n", event); 2204a40c4f10SYehuda Sadeh rb_erase(&event->node, &osdc->event_tree); 2205a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 2206a40c4f10SYehuda Sadeh } else { 2207a40c4f10SYehuda Sadeh dout("__remove_event didn't remove %p\n", event); 2208a40c4f10SYehuda Sadeh } 2209a40c4f10SYehuda Sadeh } 2210a40c4f10SYehuda Sadeh 2211a40c4f10SYehuda Sadeh int ceph_osdc_create_event(struct ceph_osd_client *osdc, 2212a40c4f10SYehuda Sadeh void (*event_cb)(u64, u64, u8, void *), 22133c663bbdSAlex Elder void *data, struct ceph_osd_event **pevent) 2214a40c4f10SYehuda Sadeh { 2215a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 2216a40c4f10SYehuda Sadeh 2217a40c4f10SYehuda Sadeh event = kmalloc(sizeof(*event), GFP_NOIO); 2218a40c4f10SYehuda Sadeh if (!event) 2219a40c4f10SYehuda Sadeh return -ENOMEM; 2220a40c4f10SYehuda Sadeh 2221a40c4f10SYehuda Sadeh dout("create_event %p\n", event); 2222a40c4f10SYehuda Sadeh event->cb = event_cb; 22233c663bbdSAlex Elder event->one_shot = 0; 2224a40c4f10SYehuda Sadeh event->data = data; 2225a40c4f10SYehuda Sadeh event->osdc = osdc; 2226a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&event->osd_node); 22273ee5234dSAlex Elder RB_CLEAR_NODE(&event->node); 2228a40c4f10SYehuda Sadeh kref_init(&event->kref); /* one ref for us */ 2229a40c4f10SYehuda Sadeh kref_get(&event->kref); /* one ref for the caller */ 2230a40c4f10SYehuda Sadeh 2231a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 2232a40c4f10SYehuda Sadeh event->cookie = ++osdc->event_count; 2233a40c4f10SYehuda Sadeh __insert_event(osdc, event); 2234a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 2235a40c4f10SYehuda Sadeh 2236a40c4f10SYehuda Sadeh *pevent = event; 2237a40c4f10SYehuda Sadeh return 0; 2238a40c4f10SYehuda Sadeh } 2239a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_create_event); 2240a40c4f10SYehuda Sadeh 2241a40c4f10SYehuda Sadeh void ceph_osdc_cancel_event(struct ceph_osd_event *event) 2242a40c4f10SYehuda Sadeh { 2243a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 2244a40c4f10SYehuda Sadeh 2245a40c4f10SYehuda Sadeh dout("cancel_event %p\n", event); 2246a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 2247a40c4f10SYehuda Sadeh __remove_event(event); 2248a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 2249a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); /* caller's */ 2250a40c4f10SYehuda Sadeh } 2251a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_cancel_event); 2252a40c4f10SYehuda Sadeh 2253a40c4f10SYehuda Sadeh 2254a40c4f10SYehuda Sadeh static void do_event_work(struct work_struct *work) 2255a40c4f10SYehuda Sadeh { 2256a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work = 2257a40c4f10SYehuda Sadeh container_of(work, struct ceph_osd_event_work, work); 2258a40c4f10SYehuda Sadeh struct ceph_osd_event *event = event_work->event; 2259a40c4f10SYehuda Sadeh u64 ver = event_work->ver; 2260a40c4f10SYehuda Sadeh u64 notify_id = event_work->notify_id; 2261a40c4f10SYehuda Sadeh u8 opcode = event_work->opcode; 2262a40c4f10SYehuda Sadeh 2263a40c4f10SYehuda Sadeh dout("do_event_work completing %p\n", event); 2264a40c4f10SYehuda Sadeh event->cb(ver, notify_id, opcode, event->data); 2265a40c4f10SYehuda Sadeh dout("do_event_work completed %p\n", event); 2266a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 2267a40c4f10SYehuda Sadeh kfree(event_work); 2268a40c4f10SYehuda Sadeh } 2269a40c4f10SYehuda Sadeh 2270a40c4f10SYehuda Sadeh 2271a40c4f10SYehuda Sadeh /* 2272a40c4f10SYehuda Sadeh * Process osd watch notifications 2273a40c4f10SYehuda Sadeh */ 22743c663bbdSAlex Elder static void handle_watch_notify(struct ceph_osd_client *osdc, 22753c663bbdSAlex Elder struct ceph_msg *msg) 2276a40c4f10SYehuda Sadeh { 2277a40c4f10SYehuda Sadeh void *p, *end; 2278a40c4f10SYehuda Sadeh u8 proto_ver; 2279a40c4f10SYehuda Sadeh u64 cookie, ver, notify_id; 2280a40c4f10SYehuda Sadeh u8 opcode; 2281a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 2282a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work; 2283a40c4f10SYehuda Sadeh 2284a40c4f10SYehuda Sadeh p = msg->front.iov_base; 2285a40c4f10SYehuda Sadeh end = p + msg->front.iov_len; 2286a40c4f10SYehuda Sadeh 2287a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, proto_ver, bad); 2288a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, opcode, bad); 2289a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, cookie, bad); 2290a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, ver, bad); 2291a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, notify_id, bad); 2292a40c4f10SYehuda Sadeh 2293a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 2294a40c4f10SYehuda Sadeh event = __find_event(osdc, cookie); 2295a40c4f10SYehuda Sadeh if (event) { 22963c663bbdSAlex Elder BUG_ON(event->one_shot); 2297a40c4f10SYehuda Sadeh get_event(event); 2298a40c4f10SYehuda Sadeh } 2299a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 2300a40c4f10SYehuda Sadeh dout("handle_watch_notify cookie %lld ver %lld event %p\n", 2301a40c4f10SYehuda Sadeh cookie, ver, event); 2302a40c4f10SYehuda Sadeh if (event) { 2303a40c4f10SYehuda Sadeh event_work = kmalloc(sizeof(*event_work), GFP_NOIO); 2304a40c4f10SYehuda Sadeh if (!event_work) { 230591883cd2SIlya Dryomov pr_err("couldn't allocate event_work\n"); 230691883cd2SIlya Dryomov ceph_osdc_put_event(event); 230791883cd2SIlya Dryomov return; 2308a40c4f10SYehuda Sadeh } 23096b0ae409SMariusz Kozlowski INIT_WORK(&event_work->work, do_event_work); 2310a40c4f10SYehuda Sadeh event_work->event = event; 2311a40c4f10SYehuda Sadeh event_work->ver = ver; 2312a40c4f10SYehuda Sadeh event_work->notify_id = notify_id; 2313a40c4f10SYehuda Sadeh event_work->opcode = opcode; 231491883cd2SIlya Dryomov 231591883cd2SIlya Dryomov queue_work(osdc->notify_wq, &event_work->work); 2316a40c4f10SYehuda Sadeh } 2317a40c4f10SYehuda Sadeh 2318a40c4f10SYehuda Sadeh return; 2319a40c4f10SYehuda Sadeh 2320a40c4f10SYehuda Sadeh bad: 2321a40c4f10SYehuda Sadeh pr_err("osdc handle_watch_notify corrupt msg\n"); 2322a40c4f10SYehuda Sadeh } 2323a40c4f10SYehuda Sadeh 2324a40c4f10SYehuda Sadeh /* 2325e65550fdSAlex Elder * build new request AND message 2326e65550fdSAlex Elder * 2327e65550fdSAlex Elder */ 2328e65550fdSAlex Elder void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off, 2329e65550fdSAlex Elder struct ceph_snap_context *snapc, u64 snap_id, 2330e65550fdSAlex Elder struct timespec *mtime) 2331e65550fdSAlex Elder { 2332e65550fdSAlex Elder struct ceph_msg *msg = req->r_request; 2333e65550fdSAlex Elder void *p; 2334e65550fdSAlex Elder size_t msg_size; 2335e65550fdSAlex Elder int flags = req->r_flags; 2336e65550fdSAlex Elder u64 data_len; 2337e65550fdSAlex Elder unsigned int i; 2338e65550fdSAlex Elder 2339e65550fdSAlex Elder req->r_snapid = snap_id; 2340e65550fdSAlex Elder req->r_snapc = ceph_get_snap_context(snapc); 2341e65550fdSAlex Elder 2342e65550fdSAlex Elder /* encode request */ 2343e65550fdSAlex Elder msg->hdr.version = cpu_to_le16(4); 2344e65550fdSAlex Elder 2345e65550fdSAlex Elder p = msg->front.iov_base; 2346e65550fdSAlex Elder ceph_encode_32(&p, 1); /* client_inc is always 1 */ 2347e65550fdSAlex Elder req->r_request_osdmap_epoch = p; 2348e65550fdSAlex Elder p += 4; 2349e65550fdSAlex Elder req->r_request_flags = p; 2350e65550fdSAlex Elder p += 4; 2351e65550fdSAlex Elder if (req->r_flags & CEPH_OSD_FLAG_WRITE) 2352e65550fdSAlex Elder ceph_encode_timespec(p, mtime); 2353e65550fdSAlex Elder p += sizeof(struct ceph_timespec); 2354e65550fdSAlex Elder req->r_request_reassert_version = p; 2355e65550fdSAlex Elder p += sizeof(struct ceph_eversion); /* will get filled in */ 2356e65550fdSAlex Elder 2357e65550fdSAlex Elder /* oloc */ 2358e65550fdSAlex Elder ceph_encode_8(&p, 4); 2359e65550fdSAlex Elder ceph_encode_8(&p, 4); 2360e65550fdSAlex Elder ceph_encode_32(&p, 8 + 4 + 4); 2361e65550fdSAlex Elder req->r_request_pool = p; 2362e65550fdSAlex Elder p += 8; 2363e65550fdSAlex Elder ceph_encode_32(&p, -1); /* preferred */ 2364e65550fdSAlex Elder ceph_encode_32(&p, 0); /* key len */ 2365e65550fdSAlex Elder 2366e65550fdSAlex Elder ceph_encode_8(&p, 1); 2367e65550fdSAlex Elder req->r_request_pgid = p; 2368e65550fdSAlex Elder p += 8 + 4; 2369e65550fdSAlex Elder ceph_encode_32(&p, -1); /* preferred */ 2370e65550fdSAlex Elder 2371e65550fdSAlex Elder /* oid */ 23723c972c95SIlya Dryomov ceph_encode_32(&p, req->r_base_oid.name_len); 23733c972c95SIlya Dryomov memcpy(p, req->r_base_oid.name, req->r_base_oid.name_len); 23743c972c95SIlya Dryomov dout("oid '%.*s' len %d\n", req->r_base_oid.name_len, 23753c972c95SIlya Dryomov req->r_base_oid.name, req->r_base_oid.name_len); 23763c972c95SIlya Dryomov p += req->r_base_oid.name_len; 2377e65550fdSAlex Elder 2378e65550fdSAlex Elder /* ops--can imply data */ 2379e65550fdSAlex Elder ceph_encode_16(&p, (u16)req->r_num_ops); 2380e65550fdSAlex Elder data_len = 0; 2381e65550fdSAlex Elder for (i = 0; i < req->r_num_ops; i++) { 2382e65550fdSAlex Elder data_len += osd_req_encode_op(req, p, i); 2383e65550fdSAlex Elder p += sizeof(struct ceph_osd_op); 2384e65550fdSAlex Elder } 2385e65550fdSAlex Elder 2386e65550fdSAlex Elder /* snaps */ 2387e65550fdSAlex Elder ceph_encode_64(&p, req->r_snapid); 2388e65550fdSAlex Elder ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0); 2389e65550fdSAlex Elder ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0); 2390e65550fdSAlex Elder if (req->r_snapc) { 2391e65550fdSAlex Elder for (i = 0; i < snapc->num_snaps; i++) { 2392e65550fdSAlex Elder ceph_encode_64(&p, req->r_snapc->snaps[i]); 2393e65550fdSAlex Elder } 2394e65550fdSAlex Elder } 2395e65550fdSAlex Elder 2396e65550fdSAlex Elder req->r_request_attempts = p; 2397e65550fdSAlex Elder p += 4; 2398e65550fdSAlex Elder 2399e65550fdSAlex Elder /* data */ 2400e65550fdSAlex Elder if (flags & CEPH_OSD_FLAG_WRITE) { 2401e65550fdSAlex Elder u16 data_off; 2402e65550fdSAlex Elder 2403e65550fdSAlex Elder /* 2404e65550fdSAlex Elder * The header "data_off" is a hint to the receiver 2405e65550fdSAlex Elder * allowing it to align received data into its 2406e65550fdSAlex Elder * buffers such that there's no need to re-copy 2407e65550fdSAlex Elder * it before writing it to disk (direct I/O). 2408e65550fdSAlex Elder */ 2409e65550fdSAlex Elder data_off = (u16) (off & 0xffff); 2410e65550fdSAlex Elder req->r_request->hdr.data_off = cpu_to_le16(data_off); 2411e65550fdSAlex Elder } 2412e65550fdSAlex Elder req->r_request->hdr.data_len = cpu_to_le32(data_len); 2413e65550fdSAlex Elder 2414e65550fdSAlex Elder BUG_ON(p > msg->front.iov_base + msg->front.iov_len); 2415e65550fdSAlex Elder msg_size = p - msg->front.iov_base; 2416e65550fdSAlex Elder msg->front.iov_len = msg_size; 2417e65550fdSAlex Elder msg->hdr.front_len = cpu_to_le32(msg_size); 2418e65550fdSAlex Elder 2419e65550fdSAlex Elder dout("build_request msg_size was %d\n", (int)msg_size); 2420e65550fdSAlex Elder } 2421e65550fdSAlex Elder EXPORT_SYMBOL(ceph_osdc_build_request); 2422e65550fdSAlex Elder 2423e65550fdSAlex Elder /* 24243d14c5d2SYehuda Sadeh * Register request, send initial attempt. 24253d14c5d2SYehuda Sadeh */ 24263d14c5d2SYehuda Sadeh int ceph_osdc_start_request(struct ceph_osd_client *osdc, 24273d14c5d2SYehuda Sadeh struct ceph_osd_request *req, 24283d14c5d2SYehuda Sadeh bool nofail) 24293d14c5d2SYehuda Sadeh { 24300bbfdfe8SIlya Dryomov int rc; 24313d14c5d2SYehuda Sadeh 24323d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 24333d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 24340bbfdfe8SIlya Dryomov 24350bbfdfe8SIlya Dryomov rc = __ceph_osdc_start_request(osdc, req, nofail); 24360bbfdfe8SIlya Dryomov 24373d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 24383d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 24390bbfdfe8SIlya Dryomov 24403d14c5d2SYehuda Sadeh return rc; 24413d14c5d2SYehuda Sadeh } 24423d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_start_request); 24433d14c5d2SYehuda Sadeh 24443d14c5d2SYehuda Sadeh /* 2445c9f9b93dSIlya Dryomov * Unregister a registered request. The request is not completed (i.e. 2446c9f9b93dSIlya Dryomov * no callbacks or wakeups) - higher layers are supposed to know what 2447c9f9b93dSIlya Dryomov * they are canceling. 2448c9f9b93dSIlya Dryomov */ 2449c9f9b93dSIlya Dryomov void ceph_osdc_cancel_request(struct ceph_osd_request *req) 2450c9f9b93dSIlya Dryomov { 2451c9f9b93dSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 2452c9f9b93dSIlya Dryomov 2453c9f9b93dSIlya Dryomov mutex_lock(&osdc->request_mutex); 2454c9f9b93dSIlya Dryomov if (req->r_linger) 2455c9f9b93dSIlya Dryomov __unregister_linger_request(osdc, req); 2456c9f9b93dSIlya Dryomov __unregister_request(osdc, req); 2457c9f9b93dSIlya Dryomov mutex_unlock(&osdc->request_mutex); 2458c9f9b93dSIlya Dryomov 2459c9f9b93dSIlya Dryomov dout("%s %p tid %llu canceled\n", __func__, req, req->r_tid); 2460c9f9b93dSIlya Dryomov } 2461c9f9b93dSIlya Dryomov EXPORT_SYMBOL(ceph_osdc_cancel_request); 2462c9f9b93dSIlya Dryomov 2463c9f9b93dSIlya Dryomov /* 24643d14c5d2SYehuda Sadeh * wait for a request to complete 24653d14c5d2SYehuda Sadeh */ 24663d14c5d2SYehuda Sadeh int ceph_osdc_wait_request(struct ceph_osd_client *osdc, 24673d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 24683d14c5d2SYehuda Sadeh { 24693d14c5d2SYehuda Sadeh int rc; 24703d14c5d2SYehuda Sadeh 2471c9f9b93dSIlya Dryomov dout("%s %p tid %llu\n", __func__, req, req->r_tid); 2472c9f9b93dSIlya Dryomov 24733d14c5d2SYehuda Sadeh rc = wait_for_completion_interruptible(&req->r_completion); 24743d14c5d2SYehuda Sadeh if (rc < 0) { 2475c9f9b93dSIlya Dryomov dout("%s %p tid %llu interrupted\n", __func__, req, req->r_tid); 2476c9f9b93dSIlya Dryomov ceph_osdc_cancel_request(req); 247725845472SSage Weil complete_request(req); 24783d14c5d2SYehuda Sadeh return rc; 24793d14c5d2SYehuda Sadeh } 24803d14c5d2SYehuda Sadeh 2481c9f9b93dSIlya Dryomov dout("%s %p tid %llu result %d\n", __func__, req, req->r_tid, 2482c9f9b93dSIlya Dryomov req->r_result); 24833d14c5d2SYehuda Sadeh return req->r_result; 24843d14c5d2SYehuda Sadeh } 24853d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_wait_request); 24863d14c5d2SYehuda Sadeh 24873d14c5d2SYehuda Sadeh /* 24883d14c5d2SYehuda Sadeh * sync - wait for all in-flight requests to flush. avoid starvation. 24893d14c5d2SYehuda Sadeh */ 24903d14c5d2SYehuda Sadeh void ceph_osdc_sync(struct ceph_osd_client *osdc) 24913d14c5d2SYehuda Sadeh { 24923d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 24933d14c5d2SYehuda Sadeh u64 last_tid, next_tid = 0; 24943d14c5d2SYehuda Sadeh 24953d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 24963d14c5d2SYehuda Sadeh last_tid = osdc->last_tid; 24973d14c5d2SYehuda Sadeh while (1) { 24983d14c5d2SYehuda Sadeh req = __lookup_request_ge(osdc, next_tid); 24993d14c5d2SYehuda Sadeh if (!req) 25003d14c5d2SYehuda Sadeh break; 25013d14c5d2SYehuda Sadeh if (req->r_tid > last_tid) 25023d14c5d2SYehuda Sadeh break; 25033d14c5d2SYehuda Sadeh 25043d14c5d2SYehuda Sadeh next_tid = req->r_tid + 1; 25053d14c5d2SYehuda Sadeh if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0) 25063d14c5d2SYehuda Sadeh continue; 25073d14c5d2SYehuda Sadeh 25083d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 25093d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 25103d14c5d2SYehuda Sadeh dout("sync waiting on tid %llu (last is %llu)\n", 25113d14c5d2SYehuda Sadeh req->r_tid, last_tid); 25123d14c5d2SYehuda Sadeh wait_for_completion(&req->r_safe_completion); 25133d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 25143d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 25153d14c5d2SYehuda Sadeh } 25163d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 25173d14c5d2SYehuda Sadeh dout("sync done (thru tid %llu)\n", last_tid); 25183d14c5d2SYehuda Sadeh } 25193d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_sync); 25203d14c5d2SYehuda Sadeh 25213d14c5d2SYehuda Sadeh /* 2522dd935f44SJosh Durgin * Call all pending notify callbacks - for use after a watch is 2523dd935f44SJosh Durgin * unregistered, to make sure no more callbacks for it will be invoked 2524dd935f44SJosh Durgin */ 2525f6479449Sstephen hemminger void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc) 2526dd935f44SJosh Durgin { 2527dd935f44SJosh Durgin flush_workqueue(osdc->notify_wq); 2528dd935f44SJosh Durgin } 2529dd935f44SJosh Durgin EXPORT_SYMBOL(ceph_osdc_flush_notifies); 2530dd935f44SJosh Durgin 2531dd935f44SJosh Durgin 2532dd935f44SJosh Durgin /* 25333d14c5d2SYehuda Sadeh * init, shutdown 25343d14c5d2SYehuda Sadeh */ 25353d14c5d2SYehuda Sadeh int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client) 25363d14c5d2SYehuda Sadeh { 25373d14c5d2SYehuda Sadeh int err; 25383d14c5d2SYehuda Sadeh 25393d14c5d2SYehuda Sadeh dout("init\n"); 25403d14c5d2SYehuda Sadeh osdc->client = client; 25413d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 25423d14c5d2SYehuda Sadeh init_rwsem(&osdc->map_sem); 25433d14c5d2SYehuda Sadeh init_completion(&osdc->map_waiters); 25443d14c5d2SYehuda Sadeh osdc->last_requested_map = 0; 25453d14c5d2SYehuda Sadeh mutex_init(&osdc->request_mutex); 25463d14c5d2SYehuda Sadeh osdc->last_tid = 0; 25473d14c5d2SYehuda Sadeh osdc->osds = RB_ROOT; 25483d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->osd_lru); 25493d14c5d2SYehuda Sadeh osdc->requests = RB_ROOT; 25503d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_lru); 25516f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_unsent); 25526f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_notarget); 2553a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_linger); 25543d14c5d2SYehuda Sadeh osdc->num_requests = 0; 25553d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout); 25563d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout); 2557a40c4f10SYehuda Sadeh spin_lock_init(&osdc->event_lock); 2558a40c4f10SYehuda Sadeh osdc->event_tree = RB_ROOT; 2559a40c4f10SYehuda Sadeh osdc->event_count = 0; 25603d14c5d2SYehuda Sadeh 25613d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 25623d14c5d2SYehuda Sadeh round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ)); 25633d14c5d2SYehuda Sadeh 25643d14c5d2SYehuda Sadeh err = -ENOMEM; 25653d14c5d2SYehuda Sadeh osdc->req_mempool = mempool_create_kmalloc_pool(10, 25663d14c5d2SYehuda Sadeh sizeof(struct ceph_osd_request)); 25673d14c5d2SYehuda Sadeh if (!osdc->req_mempool) 25683d14c5d2SYehuda Sadeh goto out; 25693d14c5d2SYehuda Sadeh 2570d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP, 2571d50b409fSSage Weil OSD_OP_FRONT_LEN, 10, true, 25723d14c5d2SYehuda Sadeh "osd_op"); 25733d14c5d2SYehuda Sadeh if (err < 0) 25743d14c5d2SYehuda Sadeh goto out_mempool; 2575d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY, 25763d14c5d2SYehuda Sadeh OSD_OPREPLY_FRONT_LEN, 10, true, 25773d14c5d2SYehuda Sadeh "osd_op_reply"); 25783d14c5d2SYehuda Sadeh if (err < 0) 25793d14c5d2SYehuda Sadeh goto out_msgpool; 2580a40c4f10SYehuda Sadeh 2581dbcae088SDan Carpenter err = -ENOMEM; 2582a40c4f10SYehuda Sadeh osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify"); 2583dbcae088SDan Carpenter if (!osdc->notify_wq) 2584c172ec5cSIlya Dryomov goto out_msgpool_reply; 2585c172ec5cSIlya Dryomov 25863d14c5d2SYehuda Sadeh return 0; 25873d14c5d2SYehuda Sadeh 2588c172ec5cSIlya Dryomov out_msgpool_reply: 2589c172ec5cSIlya Dryomov ceph_msgpool_destroy(&osdc->msgpool_op_reply); 25903d14c5d2SYehuda Sadeh out_msgpool: 25913d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 25923d14c5d2SYehuda Sadeh out_mempool: 25933d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 25943d14c5d2SYehuda Sadeh out: 25953d14c5d2SYehuda Sadeh return err; 25963d14c5d2SYehuda Sadeh } 25973d14c5d2SYehuda Sadeh 25983d14c5d2SYehuda Sadeh void ceph_osdc_stop(struct ceph_osd_client *osdc) 25993d14c5d2SYehuda Sadeh { 2600a40c4f10SYehuda Sadeh flush_workqueue(osdc->notify_wq); 2601a40c4f10SYehuda Sadeh destroy_workqueue(osdc->notify_wq); 26023d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->timeout_work); 26033d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->osds_timeout_work); 26043d14c5d2SYehuda Sadeh if (osdc->osdmap) { 26053d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 26063d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 26073d14c5d2SYehuda Sadeh } 2608aca420bcSSage Weil remove_all_osds(osdc); 26093d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 26103d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 26113d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op_reply); 26123d14c5d2SYehuda Sadeh } 26133d14c5d2SYehuda Sadeh 26143d14c5d2SYehuda Sadeh /* 26153d14c5d2SYehuda Sadeh * Read some contiguous pages. If we cross a stripe boundary, shorten 26163d14c5d2SYehuda Sadeh * *plen. Return number of bytes read, or error. 26173d14c5d2SYehuda Sadeh */ 26183d14c5d2SYehuda Sadeh int ceph_osdc_readpages(struct ceph_osd_client *osdc, 26193d14c5d2SYehuda Sadeh struct ceph_vino vino, struct ceph_file_layout *layout, 26203d14c5d2SYehuda Sadeh u64 off, u64 *plen, 26213d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 2622b7495fc2SSage Weil struct page **pages, int num_pages, int page_align) 26233d14c5d2SYehuda Sadeh { 26243d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 26253d14c5d2SYehuda Sadeh int rc = 0; 26263d14c5d2SYehuda Sadeh 26273d14c5d2SYehuda Sadeh dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino, 26283d14c5d2SYehuda Sadeh vino.snap, off, *plen); 262979528734SAlex Elder req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 1, 26303d14c5d2SYehuda Sadeh CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ, 2631acead002SAlex Elder NULL, truncate_seq, truncate_size, 2632153e5167SAlex Elder false); 26336816282dSSage Weil if (IS_ERR(req)) 26346816282dSSage Weil return PTR_ERR(req); 26353d14c5d2SYehuda Sadeh 26363d14c5d2SYehuda Sadeh /* it may be a short read due to an object boundary */ 26370fff87ecSAlex Elder 2638406e2c9fSAlex Elder osd_req_op_extent_osd_data_pages(req, 0, 2639a4ce40a9SAlex Elder pages, *plen, page_align, false, false); 26403d14c5d2SYehuda Sadeh 2641e0c59487SAlex Elder dout("readpages final extent is %llu~%llu (%llu bytes align %d)\n", 264243bfe5deSAlex Elder off, *plen, *plen, page_align); 26433d14c5d2SYehuda Sadeh 264479528734SAlex Elder ceph_osdc_build_request(req, off, NULL, vino.snap, NULL); 264502ee07d3SAlex Elder 26463d14c5d2SYehuda Sadeh rc = ceph_osdc_start_request(osdc, req, false); 26473d14c5d2SYehuda Sadeh if (!rc) 26483d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 26493d14c5d2SYehuda Sadeh 26503d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 26513d14c5d2SYehuda Sadeh dout("readpages result %d\n", rc); 26523d14c5d2SYehuda Sadeh return rc; 26533d14c5d2SYehuda Sadeh } 26543d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_readpages); 26553d14c5d2SYehuda Sadeh 26563d14c5d2SYehuda Sadeh /* 26573d14c5d2SYehuda Sadeh * do a synchronous write on N pages 26583d14c5d2SYehuda Sadeh */ 26593d14c5d2SYehuda Sadeh int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino, 26603d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 26613d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 26623d14c5d2SYehuda Sadeh u64 off, u64 len, 26633d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 26643d14c5d2SYehuda Sadeh struct timespec *mtime, 266524808826SAlex Elder struct page **pages, int num_pages) 26663d14c5d2SYehuda Sadeh { 26673d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 26683d14c5d2SYehuda Sadeh int rc = 0; 2669b7495fc2SSage Weil int page_align = off & ~PAGE_MASK; 26703d14c5d2SYehuda Sadeh 2671acead002SAlex Elder BUG_ON(vino.snap != CEPH_NOSNAP); /* snapshots aren't writeable */ 267279528734SAlex Elder req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 1, 26733d14c5d2SYehuda Sadeh CEPH_OSD_OP_WRITE, 267424808826SAlex Elder CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE, 2675acead002SAlex Elder snapc, truncate_seq, truncate_size, 2676153e5167SAlex Elder true); 26776816282dSSage Weil if (IS_ERR(req)) 26786816282dSSage Weil return PTR_ERR(req); 26793d14c5d2SYehuda Sadeh 26803d14c5d2SYehuda Sadeh /* it may be a short write due to an object boundary */ 2681406e2c9fSAlex Elder osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align, 268243bfe5deSAlex Elder false, false); 268343bfe5deSAlex Elder dout("writepages %llu~%llu (%llu bytes)\n", off, len, len); 26843d14c5d2SYehuda Sadeh 268579528734SAlex Elder ceph_osdc_build_request(req, off, snapc, CEPH_NOSNAP, mtime); 268602ee07d3SAlex Elder 268787f979d3SAlex Elder rc = ceph_osdc_start_request(osdc, req, true); 26883d14c5d2SYehuda Sadeh if (!rc) 26893d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 26903d14c5d2SYehuda Sadeh 26913d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 26923d14c5d2SYehuda Sadeh if (rc == 0) 26933d14c5d2SYehuda Sadeh rc = len; 26943d14c5d2SYehuda Sadeh dout("writepages result %d\n", rc); 26953d14c5d2SYehuda Sadeh return rc; 26963d14c5d2SYehuda Sadeh } 26973d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_writepages); 26983d14c5d2SYehuda Sadeh 26995522ae0bSAlex Elder int ceph_osdc_setup(void) 27005522ae0bSAlex Elder { 27015522ae0bSAlex Elder BUG_ON(ceph_osd_request_cache); 27025522ae0bSAlex Elder ceph_osd_request_cache = kmem_cache_create("ceph_osd_request", 27035522ae0bSAlex Elder sizeof (struct ceph_osd_request), 27045522ae0bSAlex Elder __alignof__(struct ceph_osd_request), 27055522ae0bSAlex Elder 0, NULL); 27065522ae0bSAlex Elder 27075522ae0bSAlex Elder return ceph_osd_request_cache ? 0 : -ENOMEM; 27085522ae0bSAlex Elder } 27095522ae0bSAlex Elder EXPORT_SYMBOL(ceph_osdc_setup); 27105522ae0bSAlex Elder 27115522ae0bSAlex Elder void ceph_osdc_cleanup(void) 27125522ae0bSAlex Elder { 27135522ae0bSAlex Elder BUG_ON(!ceph_osd_request_cache); 27145522ae0bSAlex Elder kmem_cache_destroy(ceph_osd_request_cache); 27155522ae0bSAlex Elder ceph_osd_request_cache = NULL; 27165522ae0bSAlex Elder } 27175522ae0bSAlex Elder EXPORT_SYMBOL(ceph_osdc_cleanup); 27185522ae0bSAlex Elder 27193d14c5d2SYehuda Sadeh /* 27203d14c5d2SYehuda Sadeh * handle incoming message 27213d14c5d2SYehuda Sadeh */ 27223d14c5d2SYehuda Sadeh static void dispatch(struct ceph_connection *con, struct ceph_msg *msg) 27233d14c5d2SYehuda Sadeh { 27243d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 27253d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 27263d14c5d2SYehuda Sadeh int type = le16_to_cpu(msg->hdr.type); 27273d14c5d2SYehuda Sadeh 27283d14c5d2SYehuda Sadeh if (!osd) 27293d14c5d2SYehuda Sadeh goto out; 27303d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 27313d14c5d2SYehuda Sadeh 27323d14c5d2SYehuda Sadeh switch (type) { 27333d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 27343d14c5d2SYehuda Sadeh ceph_osdc_handle_map(osdc, msg); 27353d14c5d2SYehuda Sadeh break; 27363d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 27373d14c5d2SYehuda Sadeh handle_reply(osdc, msg, con); 27383d14c5d2SYehuda Sadeh break; 2739a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 2740a40c4f10SYehuda Sadeh handle_watch_notify(osdc, msg); 2741a40c4f10SYehuda Sadeh break; 27423d14c5d2SYehuda Sadeh 27433d14c5d2SYehuda Sadeh default: 27443d14c5d2SYehuda Sadeh pr_err("received unknown message type %d %s\n", type, 27453d14c5d2SYehuda Sadeh ceph_msg_type_name(type)); 27463d14c5d2SYehuda Sadeh } 27473d14c5d2SYehuda Sadeh out: 27483d14c5d2SYehuda Sadeh ceph_msg_put(msg); 27493d14c5d2SYehuda Sadeh } 27503d14c5d2SYehuda Sadeh 27513d14c5d2SYehuda Sadeh /* 27523d14c5d2SYehuda Sadeh * lookup and return message for incoming reply. set up reply message 27533d14c5d2SYehuda Sadeh * pages. 27543d14c5d2SYehuda Sadeh */ 27553d14c5d2SYehuda Sadeh static struct ceph_msg *get_reply(struct ceph_connection *con, 27563d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 27573d14c5d2SYehuda Sadeh int *skip) 27583d14c5d2SYehuda Sadeh { 27593d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 27603d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = osd->o_osdc; 27613d14c5d2SYehuda Sadeh struct ceph_msg *m; 27623d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 27633f0a4ac5SIlya Dryomov int front_len = le32_to_cpu(hdr->front_len); 27643d14c5d2SYehuda Sadeh int data_len = le32_to_cpu(hdr->data_len); 27653d14c5d2SYehuda Sadeh u64 tid; 27663d14c5d2SYehuda Sadeh 27673d14c5d2SYehuda Sadeh tid = le64_to_cpu(hdr->tid); 27683d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 27693d14c5d2SYehuda Sadeh req = __lookup_request(osdc, tid); 27703d14c5d2SYehuda Sadeh if (!req) { 27713d14c5d2SYehuda Sadeh *skip = 1; 27723d14c5d2SYehuda Sadeh m = NULL; 2773756a16a5SSage Weil dout("get_reply unknown tid %llu from osd%d\n", tid, 27743d14c5d2SYehuda Sadeh osd->o_osd); 27753d14c5d2SYehuda Sadeh goto out; 27763d14c5d2SYehuda Sadeh } 27773d14c5d2SYehuda Sadeh 2778ace6d3a9SAlex Elder if (req->r_reply->con) 27798921d114SAlex Elder dout("%s revoking msg %p from old con %p\n", __func__, 2780ace6d3a9SAlex Elder req->r_reply, req->r_reply->con); 27818921d114SAlex Elder ceph_msg_revoke_incoming(req->r_reply); 27823d14c5d2SYehuda Sadeh 2783f2be82b0SIlya Dryomov if (front_len > req->r_reply->front_alloc_len) { 2784b9a67899SJoe Perches pr_warn("get_reply front %d > preallocated %d (%u#%llu)\n", 2785f2be82b0SIlya Dryomov front_len, req->r_reply->front_alloc_len, 27864974341eSAlex Elder (unsigned int)con->peer_name.type, 27874974341eSAlex Elder le64_to_cpu(con->peer_name.num)); 27883f0a4ac5SIlya Dryomov m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front_len, GFP_NOFS, 27893f0a4ac5SIlya Dryomov false); 27903d14c5d2SYehuda Sadeh if (!m) 27913d14c5d2SYehuda Sadeh goto out; 27923d14c5d2SYehuda Sadeh ceph_msg_put(req->r_reply); 27933d14c5d2SYehuda Sadeh req->r_reply = m; 27943d14c5d2SYehuda Sadeh } 27953d14c5d2SYehuda Sadeh m = ceph_msg_get(req->r_reply); 27963d14c5d2SYehuda Sadeh 27973d14c5d2SYehuda Sadeh if (data_len > 0) { 2798a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 27990fff87ecSAlex Elder 2800a4ce40a9SAlex Elder /* 2801a4ce40a9SAlex Elder * XXX This is assuming there is only one op containing 2802a4ce40a9SAlex Elder * XXX page data. Probably OK for reads, but this 2803a4ce40a9SAlex Elder * XXX ought to be done more generally. 2804a4ce40a9SAlex Elder */ 2805406e2c9fSAlex Elder osd_data = osd_req_op_extent_osd_data(req, 0); 28060fff87ecSAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 28070fff87ecSAlex Elder if (osd_data->pages && 2808e0c59487SAlex Elder unlikely(osd_data->length < data_len)) { 28092ac2b7a6SAlex Elder 2810b9a67899SJoe Perches pr_warn("tid %lld reply has %d bytes we had only %llu bytes ready\n", 2811e0c59487SAlex Elder tid, data_len, osd_data->length); 28123d14c5d2SYehuda Sadeh *skip = 1; 28133d14c5d2SYehuda Sadeh ceph_msg_put(m); 28143d14c5d2SYehuda Sadeh m = NULL; 28153d14c5d2SYehuda Sadeh goto out; 28163d14c5d2SYehuda Sadeh } 28173d14c5d2SYehuda Sadeh } 28182ac2b7a6SAlex Elder } 28193d14c5d2SYehuda Sadeh *skip = 0; 28203d14c5d2SYehuda Sadeh dout("get_reply tid %lld %p\n", tid, m); 28213d14c5d2SYehuda Sadeh 28223d14c5d2SYehuda Sadeh out: 28233d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 28243d14c5d2SYehuda Sadeh return m; 28253d14c5d2SYehuda Sadeh 28263d14c5d2SYehuda Sadeh } 28273d14c5d2SYehuda Sadeh 28283d14c5d2SYehuda Sadeh static struct ceph_msg *alloc_msg(struct ceph_connection *con, 28293d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 28303d14c5d2SYehuda Sadeh int *skip) 28313d14c5d2SYehuda Sadeh { 28323d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 28333d14c5d2SYehuda Sadeh int type = le16_to_cpu(hdr->type); 28343d14c5d2SYehuda Sadeh int front = le32_to_cpu(hdr->front_len); 28353d14c5d2SYehuda Sadeh 28361c20f2d2SAlex Elder *skip = 0; 28373d14c5d2SYehuda Sadeh switch (type) { 28383d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 2839a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 2840b61c2763SSage Weil return ceph_msg_new(type, front, GFP_NOFS, false); 28413d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 28423d14c5d2SYehuda Sadeh return get_reply(con, hdr, skip); 28433d14c5d2SYehuda Sadeh default: 28443d14c5d2SYehuda Sadeh pr_info("alloc_msg unexpected msg type %d from osd%d\n", type, 28453d14c5d2SYehuda Sadeh osd->o_osd); 28463d14c5d2SYehuda Sadeh *skip = 1; 28473d14c5d2SYehuda Sadeh return NULL; 28483d14c5d2SYehuda Sadeh } 28493d14c5d2SYehuda Sadeh } 28503d14c5d2SYehuda Sadeh 28513d14c5d2SYehuda Sadeh /* 28523d14c5d2SYehuda Sadeh * Wrappers to refcount containing ceph_osd struct 28533d14c5d2SYehuda Sadeh */ 28543d14c5d2SYehuda Sadeh static struct ceph_connection *get_osd_con(struct ceph_connection *con) 28553d14c5d2SYehuda Sadeh { 28563d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 28573d14c5d2SYehuda Sadeh if (get_osd(osd)) 28583d14c5d2SYehuda Sadeh return con; 28593d14c5d2SYehuda Sadeh return NULL; 28603d14c5d2SYehuda Sadeh } 28613d14c5d2SYehuda Sadeh 28623d14c5d2SYehuda Sadeh static void put_osd_con(struct ceph_connection *con) 28633d14c5d2SYehuda Sadeh { 28643d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 28653d14c5d2SYehuda Sadeh put_osd(osd); 28663d14c5d2SYehuda Sadeh } 28673d14c5d2SYehuda Sadeh 28683d14c5d2SYehuda Sadeh /* 28693d14c5d2SYehuda Sadeh * authentication 28703d14c5d2SYehuda Sadeh */ 2871a3530df3SAlex Elder /* 2872a3530df3SAlex Elder * Note: returned pointer is the address of a structure that's 2873a3530df3SAlex Elder * managed separately. Caller must *not* attempt to free it. 2874a3530df3SAlex Elder */ 2875a3530df3SAlex Elder static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con, 28768f43fb53SAlex Elder int *proto, int force_new) 28773d14c5d2SYehuda Sadeh { 28783d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 28793d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 28803d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 288174f1869fSAlex Elder struct ceph_auth_handshake *auth = &o->o_auth; 28823d14c5d2SYehuda Sadeh 288374f1869fSAlex Elder if (force_new && auth->authorizer) { 288427859f97SSage Weil ceph_auth_destroy_authorizer(ac, auth->authorizer); 288574f1869fSAlex Elder auth->authorizer = NULL; 28863d14c5d2SYehuda Sadeh } 288727859f97SSage Weil if (!auth->authorizer) { 288827859f97SSage Weil int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 2889a3530df3SAlex Elder auth); 28903d14c5d2SYehuda Sadeh if (ret) 2891a3530df3SAlex Elder return ERR_PTR(ret); 289227859f97SSage Weil } else { 289327859f97SSage Weil int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 28940bed9b5cSSage Weil auth); 28950bed9b5cSSage Weil if (ret) 28960bed9b5cSSage Weil return ERR_PTR(ret); 28973d14c5d2SYehuda Sadeh } 28983d14c5d2SYehuda Sadeh *proto = ac->protocol; 289974f1869fSAlex Elder 2900a3530df3SAlex Elder return auth; 29013d14c5d2SYehuda Sadeh } 29023d14c5d2SYehuda Sadeh 29033d14c5d2SYehuda Sadeh 29043d14c5d2SYehuda Sadeh static int verify_authorizer_reply(struct ceph_connection *con, int len) 29053d14c5d2SYehuda Sadeh { 29063d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 29073d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 29083d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 29093d14c5d2SYehuda Sadeh 291027859f97SSage Weil return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len); 29113d14c5d2SYehuda Sadeh } 29123d14c5d2SYehuda Sadeh 29133d14c5d2SYehuda Sadeh static int invalidate_authorizer(struct ceph_connection *con) 29143d14c5d2SYehuda Sadeh { 29153d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 29163d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 29173d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 29183d14c5d2SYehuda Sadeh 291927859f97SSage Weil ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD); 29203d14c5d2SYehuda Sadeh return ceph_monc_validate_auth(&osdc->client->monc); 29213d14c5d2SYehuda Sadeh } 29223d14c5d2SYehuda Sadeh 292333d07337SYan, Zheng static int sign_message(struct ceph_connection *con, struct ceph_msg *msg) 292433d07337SYan, Zheng { 292533d07337SYan, Zheng struct ceph_osd *o = con->private; 292633d07337SYan, Zheng struct ceph_auth_handshake *auth = &o->o_auth; 292733d07337SYan, Zheng return ceph_auth_sign_message(auth, msg); 292833d07337SYan, Zheng } 292933d07337SYan, Zheng 293033d07337SYan, Zheng static int check_message_signature(struct ceph_connection *con, struct ceph_msg *msg) 293133d07337SYan, Zheng { 293233d07337SYan, Zheng struct ceph_osd *o = con->private; 293333d07337SYan, Zheng struct ceph_auth_handshake *auth = &o->o_auth; 293433d07337SYan, Zheng return ceph_auth_check_message_signature(auth, msg); 293533d07337SYan, Zheng } 293633d07337SYan, Zheng 29373d14c5d2SYehuda Sadeh static const struct ceph_connection_operations osd_con_ops = { 29383d14c5d2SYehuda Sadeh .get = get_osd_con, 29393d14c5d2SYehuda Sadeh .put = put_osd_con, 29403d14c5d2SYehuda Sadeh .dispatch = dispatch, 29413d14c5d2SYehuda Sadeh .get_authorizer = get_authorizer, 29423d14c5d2SYehuda Sadeh .verify_authorizer_reply = verify_authorizer_reply, 29433d14c5d2SYehuda Sadeh .invalidate_authorizer = invalidate_authorizer, 29443d14c5d2SYehuda Sadeh .alloc_msg = alloc_msg, 294533d07337SYan, Zheng .sign_message = sign_message, 294633d07337SYan, Zheng .check_message_signature = check_message_signature, 29473d14c5d2SYehuda Sadeh .fault = osd_reset, 29483d14c5d2SYehuda Sadeh }; 2949