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_OPREPLY_FRONT_LEN 512 233d14c5d2SYehuda Sadeh 245522ae0bSAlex Elder static struct kmem_cache *ceph_osd_request_cache; 255522ae0bSAlex Elder 263d14c5d2SYehuda Sadeh static const struct ceph_connection_operations osd_con_ops; 273d14c5d2SYehuda Sadeh 28f9d25199SAlex Elder static void __send_queued(struct ceph_osd_client *osdc); 296f6c7006SSage Weil static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd); 30a40c4f10SYehuda Sadeh static void __register_request(struct ceph_osd_client *osdc, 31a40c4f10SYehuda Sadeh struct ceph_osd_request *req); 322cc6128aSIlya Dryomov static void __unregister_request(struct ceph_osd_client *osdc, 332cc6128aSIlya Dryomov struct ceph_osd_request *req); 34a40c4f10SYehuda Sadeh static void __unregister_linger_request(struct ceph_osd_client *osdc, 35a40c4f10SYehuda Sadeh struct ceph_osd_request *req); 362cc6128aSIlya Dryomov static void __enqueue_request(struct ceph_osd_request *req); 373d14c5d2SYehuda Sadeh 383d14c5d2SYehuda Sadeh /* 393d14c5d2SYehuda Sadeh * Implement client access to distributed object storage cluster. 403d14c5d2SYehuda Sadeh * 413d14c5d2SYehuda Sadeh * All data objects are stored within a cluster/cloud of OSDs, or 423d14c5d2SYehuda Sadeh * "object storage devices." (Note that Ceph OSDs have _nothing_ to 433d14c5d2SYehuda Sadeh * do with the T10 OSD extensions to SCSI.) Ceph OSDs are simply 443d14c5d2SYehuda Sadeh * remote daemons serving up and coordinating consistent and safe 453d14c5d2SYehuda Sadeh * access to storage. 463d14c5d2SYehuda Sadeh * 473d14c5d2SYehuda Sadeh * Cluster membership and the mapping of data objects onto storage devices 483d14c5d2SYehuda Sadeh * are described by the osd map. 493d14c5d2SYehuda Sadeh * 503d14c5d2SYehuda Sadeh * We keep track of pending OSD requests (read, write), resubmit 513d14c5d2SYehuda Sadeh * requests to different OSDs when the cluster topology/data layout 523d14c5d2SYehuda Sadeh * change, or retry the affected requests when the communications 533d14c5d2SYehuda Sadeh * channel with an OSD is reset. 543d14c5d2SYehuda Sadeh */ 553d14c5d2SYehuda Sadeh 563d14c5d2SYehuda Sadeh /* 573d14c5d2SYehuda Sadeh * calculate the mapping of a file extent onto an object, and fill out the 583d14c5d2SYehuda Sadeh * request accordingly. shorten extent as necessary if it crosses an 593d14c5d2SYehuda Sadeh * object boundary. 603d14c5d2SYehuda Sadeh * 613d14c5d2SYehuda Sadeh * fill osd op in request message. 623d14c5d2SYehuda Sadeh */ 63dbe0fc41SAlex Elder static int calc_layout(struct ceph_file_layout *layout, u64 off, u64 *plen, 64a19dadfbSAlex Elder u64 *objnum, u64 *objoff, u64 *objlen) 653d14c5d2SYehuda Sadeh { 6660e56f13SAlex Elder u64 orig_len = *plen; 67d63b77f4SSage Weil int r; 683d14c5d2SYehuda Sadeh 6960e56f13SAlex Elder /* object extent? */ 7075d1c941SAlex Elder r = ceph_calc_file_object_mapping(layout, off, orig_len, objnum, 7175d1c941SAlex Elder objoff, objlen); 72d63b77f4SSage Weil if (r < 0) 73d63b77f4SSage Weil return r; 7475d1c941SAlex Elder if (*objlen < orig_len) { 7575d1c941SAlex Elder *plen = *objlen; 7660e56f13SAlex Elder dout(" skipping last %llu, final file extent %llu~%llu\n", 7760e56f13SAlex Elder orig_len - *plen, off, *plen); 7860e56f13SAlex Elder } 7960e56f13SAlex Elder 8075d1c941SAlex Elder dout("calc_layout objnum=%llx %llu~%llu\n", *objnum, *objoff, *objlen); 813d14c5d2SYehuda Sadeh 823ff5f385SAlex Elder return 0; 833d14c5d2SYehuda Sadeh } 843d14c5d2SYehuda Sadeh 85c54d47bfSAlex Elder static void ceph_osd_data_init(struct ceph_osd_data *osd_data) 86c54d47bfSAlex Elder { 87c54d47bfSAlex Elder memset(osd_data, 0, sizeof (*osd_data)); 88c54d47bfSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_NONE; 89c54d47bfSAlex Elder } 90c54d47bfSAlex Elder 91a4ce40a9SAlex Elder static void ceph_osd_data_pages_init(struct ceph_osd_data *osd_data, 9243bfe5deSAlex Elder struct page **pages, u64 length, u32 alignment, 9343bfe5deSAlex Elder bool pages_from_pool, bool own_pages) 9443bfe5deSAlex Elder { 9543bfe5deSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_PAGES; 9643bfe5deSAlex Elder osd_data->pages = pages; 9743bfe5deSAlex Elder osd_data->length = length; 9843bfe5deSAlex Elder osd_data->alignment = alignment; 9943bfe5deSAlex Elder osd_data->pages_from_pool = pages_from_pool; 10043bfe5deSAlex Elder osd_data->own_pages = own_pages; 10143bfe5deSAlex Elder } 10243bfe5deSAlex Elder 103a4ce40a9SAlex Elder static void ceph_osd_data_pagelist_init(struct ceph_osd_data *osd_data, 10443bfe5deSAlex Elder struct ceph_pagelist *pagelist) 10543bfe5deSAlex Elder { 10643bfe5deSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_PAGELIST; 10743bfe5deSAlex Elder osd_data->pagelist = pagelist; 10843bfe5deSAlex Elder } 10943bfe5deSAlex Elder 11043bfe5deSAlex Elder #ifdef CONFIG_BLOCK 111a4ce40a9SAlex Elder static void ceph_osd_data_bio_init(struct ceph_osd_data *osd_data, 11243bfe5deSAlex Elder struct bio *bio, size_t bio_length) 11343bfe5deSAlex Elder { 11443bfe5deSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_BIO; 11543bfe5deSAlex Elder osd_data->bio = bio; 11643bfe5deSAlex Elder osd_data->bio_length = bio_length; 11743bfe5deSAlex Elder } 11843bfe5deSAlex Elder #endif /* CONFIG_BLOCK */ 11943bfe5deSAlex Elder 120863c7eb5SAlex Elder #define osd_req_op_data(oreq, whch, typ, fld) \ 121863c7eb5SAlex Elder ({ \ 1228a703a38SIoana Ciornei struct ceph_osd_request *__oreq = (oreq); \ 1238a703a38SIoana Ciornei unsigned int __whch = (whch); \ 1248a703a38SIoana Ciornei BUG_ON(__whch >= __oreq->r_num_ops); \ 1258a703a38SIoana Ciornei &__oreq->r_ops[__whch].typ.fld; \ 126863c7eb5SAlex Elder }) 127863c7eb5SAlex Elder 12849719778SAlex Elder static struct ceph_osd_data * 12949719778SAlex Elder osd_req_op_raw_data_in(struct ceph_osd_request *osd_req, unsigned int which) 13049719778SAlex Elder { 13149719778SAlex Elder BUG_ON(which >= osd_req->r_num_ops); 13249719778SAlex Elder 13349719778SAlex Elder return &osd_req->r_ops[which].raw_data_in; 13449719778SAlex Elder } 13549719778SAlex Elder 136a4ce40a9SAlex Elder struct ceph_osd_data * 137a4ce40a9SAlex Elder osd_req_op_extent_osd_data(struct ceph_osd_request *osd_req, 138406e2c9fSAlex Elder unsigned int which) 139a4ce40a9SAlex Elder { 140863c7eb5SAlex Elder return osd_req_op_data(osd_req, which, extent, osd_data); 141a4ce40a9SAlex Elder } 142a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data); 143a4ce40a9SAlex Elder 14449719778SAlex Elder void osd_req_op_raw_data_in_pages(struct ceph_osd_request *osd_req, 14549719778SAlex Elder unsigned int which, struct page **pages, 14649719778SAlex Elder u64 length, u32 alignment, 14749719778SAlex Elder bool pages_from_pool, bool own_pages) 14849719778SAlex Elder { 14949719778SAlex Elder struct ceph_osd_data *osd_data; 15049719778SAlex Elder 15149719778SAlex Elder osd_data = osd_req_op_raw_data_in(osd_req, which); 15249719778SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 15349719778SAlex Elder pages_from_pool, own_pages); 15449719778SAlex Elder } 15549719778SAlex Elder EXPORT_SYMBOL(osd_req_op_raw_data_in_pages); 15649719778SAlex Elder 157a4ce40a9SAlex Elder void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *osd_req, 158406e2c9fSAlex Elder unsigned int which, struct page **pages, 159406e2c9fSAlex Elder u64 length, u32 alignment, 160a4ce40a9SAlex Elder bool pages_from_pool, bool own_pages) 161a4ce40a9SAlex Elder { 162a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 163a4ce40a9SAlex Elder 164863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, extent, osd_data); 165a4ce40a9SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 166a4ce40a9SAlex Elder pages_from_pool, own_pages); 167a4ce40a9SAlex Elder } 168a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data_pages); 169a4ce40a9SAlex Elder 170a4ce40a9SAlex Elder void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *osd_req, 171406e2c9fSAlex Elder unsigned int which, struct ceph_pagelist *pagelist) 172a4ce40a9SAlex Elder { 173a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 174a4ce40a9SAlex Elder 175863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, extent, osd_data); 176a4ce40a9SAlex Elder ceph_osd_data_pagelist_init(osd_data, pagelist); 177a4ce40a9SAlex Elder } 178a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data_pagelist); 179a4ce40a9SAlex Elder 180a4ce40a9SAlex Elder #ifdef CONFIG_BLOCK 181a4ce40a9SAlex Elder void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *osd_req, 182406e2c9fSAlex Elder unsigned int which, struct bio *bio, size_t bio_length) 183a4ce40a9SAlex Elder { 184a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 185863c7eb5SAlex Elder 186863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, extent, osd_data); 187a4ce40a9SAlex Elder ceph_osd_data_bio_init(osd_data, bio, bio_length); 188a4ce40a9SAlex Elder } 189a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data_bio); 190a4ce40a9SAlex Elder #endif /* CONFIG_BLOCK */ 191a4ce40a9SAlex Elder 192a4ce40a9SAlex Elder static void osd_req_op_cls_request_info_pagelist( 193a4ce40a9SAlex Elder struct ceph_osd_request *osd_req, 194a4ce40a9SAlex Elder unsigned int which, struct ceph_pagelist *pagelist) 195a4ce40a9SAlex Elder { 196a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 197a4ce40a9SAlex Elder 198863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, request_info); 199a4ce40a9SAlex Elder ceph_osd_data_pagelist_init(osd_data, pagelist); 200a4ce40a9SAlex Elder } 201a4ce40a9SAlex Elder 20204017e29SAlex Elder void osd_req_op_cls_request_data_pagelist( 20304017e29SAlex Elder struct ceph_osd_request *osd_req, 20404017e29SAlex Elder unsigned int which, struct ceph_pagelist *pagelist) 20504017e29SAlex Elder { 20604017e29SAlex Elder struct ceph_osd_data *osd_data; 20704017e29SAlex Elder 208863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, request_data); 20904017e29SAlex Elder ceph_osd_data_pagelist_init(osd_data, pagelist); 210bb873b53SIlya Dryomov osd_req->r_ops[which].cls.indata_len += pagelist->length; 211bb873b53SIlya Dryomov osd_req->r_ops[which].indata_len += pagelist->length; 21204017e29SAlex Elder } 21304017e29SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_request_data_pagelist); 21404017e29SAlex Elder 2156c57b554SAlex Elder void osd_req_op_cls_request_data_pages(struct ceph_osd_request *osd_req, 2166c57b554SAlex Elder unsigned int which, struct page **pages, u64 length, 2176c57b554SAlex Elder u32 alignment, bool pages_from_pool, bool own_pages) 2186c57b554SAlex Elder { 2196c57b554SAlex Elder struct ceph_osd_data *osd_data; 2206c57b554SAlex Elder 2216c57b554SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, request_data); 2226c57b554SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 2236c57b554SAlex Elder pages_from_pool, own_pages); 224bb873b53SIlya Dryomov osd_req->r_ops[which].cls.indata_len += length; 225bb873b53SIlya Dryomov osd_req->r_ops[which].indata_len += length; 2266c57b554SAlex Elder } 2276c57b554SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_request_data_pages); 2286c57b554SAlex Elder 229a4ce40a9SAlex Elder void osd_req_op_cls_response_data_pages(struct ceph_osd_request *osd_req, 230a4ce40a9SAlex Elder unsigned int which, struct page **pages, u64 length, 231a4ce40a9SAlex Elder u32 alignment, bool pages_from_pool, bool own_pages) 232a4ce40a9SAlex Elder { 233a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 234a4ce40a9SAlex Elder 235863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, response_data); 236a4ce40a9SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 237a4ce40a9SAlex Elder pages_from_pool, own_pages); 238a4ce40a9SAlex Elder } 239a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_response_data_pages); 240a4ce40a9SAlex Elder 24123c08a9cSAlex Elder static u64 ceph_osd_data_length(struct ceph_osd_data *osd_data) 24223c08a9cSAlex Elder { 24323c08a9cSAlex Elder switch (osd_data->type) { 24423c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_NONE: 24523c08a9cSAlex Elder return 0; 24623c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_PAGES: 24723c08a9cSAlex Elder return osd_data->length; 24823c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_PAGELIST: 24923c08a9cSAlex Elder return (u64)osd_data->pagelist->length; 25023c08a9cSAlex Elder #ifdef CONFIG_BLOCK 25123c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_BIO: 25223c08a9cSAlex Elder return (u64)osd_data->bio_length; 25323c08a9cSAlex Elder #endif /* CONFIG_BLOCK */ 25423c08a9cSAlex Elder default: 25523c08a9cSAlex Elder WARN(true, "unrecognized data type %d\n", (int)osd_data->type); 25623c08a9cSAlex Elder return 0; 25723c08a9cSAlex Elder } 25823c08a9cSAlex Elder } 25923c08a9cSAlex Elder 260c54d47bfSAlex Elder static void ceph_osd_data_release(struct ceph_osd_data *osd_data) 261c54d47bfSAlex Elder { 2625476492fSAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES && osd_data->own_pages) { 263c54d47bfSAlex Elder int num_pages; 264c54d47bfSAlex Elder 265c54d47bfSAlex Elder num_pages = calc_pages_for((u64)osd_data->alignment, 266c54d47bfSAlex Elder (u64)osd_data->length); 267c54d47bfSAlex Elder ceph_release_page_vector(osd_data->pages, num_pages); 268c54d47bfSAlex Elder } 2695476492fSAlex Elder ceph_osd_data_init(osd_data); 2705476492fSAlex Elder } 2715476492fSAlex Elder 2725476492fSAlex Elder static void osd_req_op_data_release(struct ceph_osd_request *osd_req, 2735476492fSAlex Elder unsigned int which) 2745476492fSAlex Elder { 2755476492fSAlex Elder struct ceph_osd_req_op *op; 2765476492fSAlex Elder 2775476492fSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 2785476492fSAlex Elder op = &osd_req->r_ops[which]; 2795476492fSAlex Elder 2805476492fSAlex Elder switch (op->op) { 2815476492fSAlex Elder case CEPH_OSD_OP_READ: 2825476492fSAlex Elder case CEPH_OSD_OP_WRITE: 283e30b7577SIlya Dryomov case CEPH_OSD_OP_WRITEFULL: 2845476492fSAlex Elder ceph_osd_data_release(&op->extent.osd_data); 2855476492fSAlex Elder break; 2865476492fSAlex Elder case CEPH_OSD_OP_CALL: 2875476492fSAlex Elder ceph_osd_data_release(&op->cls.request_info); 28804017e29SAlex Elder ceph_osd_data_release(&op->cls.request_data); 2895476492fSAlex Elder ceph_osd_data_release(&op->cls.response_data); 2905476492fSAlex Elder break; 291d74b50beSYan, Zheng case CEPH_OSD_OP_SETXATTR: 292d74b50beSYan, Zheng case CEPH_OSD_OP_CMPXATTR: 293d74b50beSYan, Zheng ceph_osd_data_release(&op->xattr.osd_data); 294d74b50beSYan, Zheng break; 29566ba609fSYan, Zheng case CEPH_OSD_OP_STAT: 29666ba609fSYan, Zheng ceph_osd_data_release(&op->raw_data_in); 29766ba609fSYan, Zheng break; 2985476492fSAlex Elder default: 2995476492fSAlex Elder break; 3005476492fSAlex Elder } 301c54d47bfSAlex Elder } 302c54d47bfSAlex Elder 3033d14c5d2SYehuda Sadeh /* 30463244fa1SIlya Dryomov * Assumes @t is zero-initialized. 30563244fa1SIlya Dryomov */ 30663244fa1SIlya Dryomov static void target_init(struct ceph_osd_request_target *t) 30763244fa1SIlya Dryomov { 30863244fa1SIlya Dryomov ceph_oid_init(&t->base_oid); 30963244fa1SIlya Dryomov ceph_oloc_init(&t->base_oloc); 31063244fa1SIlya Dryomov ceph_oid_init(&t->target_oid); 31163244fa1SIlya Dryomov ceph_oloc_init(&t->target_oloc); 31263244fa1SIlya Dryomov 31363244fa1SIlya Dryomov ceph_osds_init(&t->acting); 31463244fa1SIlya Dryomov ceph_osds_init(&t->up); 31563244fa1SIlya Dryomov t->size = -1; 31663244fa1SIlya Dryomov t->min_size = -1; 31763244fa1SIlya Dryomov 31863244fa1SIlya Dryomov t->osd = CEPH_HOMELESS_OSD; 31963244fa1SIlya Dryomov } 32063244fa1SIlya Dryomov 32163244fa1SIlya Dryomov static void target_destroy(struct ceph_osd_request_target *t) 32263244fa1SIlya Dryomov { 32363244fa1SIlya Dryomov ceph_oid_destroy(&t->base_oid); 32463244fa1SIlya Dryomov ceph_oid_destroy(&t->target_oid); 32563244fa1SIlya Dryomov } 32663244fa1SIlya Dryomov 32763244fa1SIlya Dryomov /* 3283d14c5d2SYehuda Sadeh * requests 3293d14c5d2SYehuda Sadeh */ 3309e94af20SIlya Dryomov static void ceph_osdc_release_request(struct kref *kref) 3313d14c5d2SYehuda Sadeh { 3329e94af20SIlya Dryomov struct ceph_osd_request *req = container_of(kref, 3339e94af20SIlya Dryomov struct ceph_osd_request, r_kref); 3345476492fSAlex Elder unsigned int which; 3353d14c5d2SYehuda Sadeh 3369e94af20SIlya Dryomov dout("%s %p (r_request %p r_reply %p)\n", __func__, req, 3379e94af20SIlya Dryomov req->r_request, req->r_reply); 3386562d661SIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&req->r_node)); 3396562d661SIlya Dryomov WARN_ON(!list_empty(&req->r_req_lru_item)); 3406562d661SIlya Dryomov WARN_ON(!list_empty(&req->r_osd_item)); 3416562d661SIlya Dryomov WARN_ON(!list_empty(&req->r_linger_item)); 3426562d661SIlya Dryomov WARN_ON(!list_empty(&req->r_linger_osd_item)); 3436562d661SIlya Dryomov WARN_ON(req->r_osd); 3449e94af20SIlya Dryomov 3453d14c5d2SYehuda Sadeh if (req->r_request) 3463d14c5d2SYehuda Sadeh ceph_msg_put(req->r_request); 347ace6d3a9SAlex Elder if (req->r_reply) { 3488921d114SAlex Elder ceph_msg_revoke_incoming(req->r_reply); 349ab8cb34aSAlex Elder ceph_msg_put(req->r_reply); 350ace6d3a9SAlex Elder } 3510fff87ecSAlex Elder 3525476492fSAlex Elder for (which = 0; which < req->r_num_ops; which++) 3535476492fSAlex Elder osd_req_op_data_release(req, which); 3540fff87ecSAlex Elder 355a66dd383SIlya Dryomov target_destroy(&req->r_t); 3563d14c5d2SYehuda Sadeh ceph_put_snap_context(req->r_snapc); 357d30291b9SIlya Dryomov 3583d14c5d2SYehuda Sadeh if (req->r_mempool) 3593d14c5d2SYehuda Sadeh mempool_free(req, req->r_osdc->req_mempool); 3603f1af42aSIlya Dryomov else if (req->r_num_ops <= CEPH_OSD_SLAB_OPS) 3615522ae0bSAlex Elder kmem_cache_free(ceph_osd_request_cache, req); 3623f1af42aSIlya Dryomov else 3633f1af42aSIlya Dryomov kfree(req); 3643d14c5d2SYehuda Sadeh } 3659e94af20SIlya Dryomov 3669e94af20SIlya Dryomov void ceph_osdc_get_request(struct ceph_osd_request *req) 3679e94af20SIlya Dryomov { 3689e94af20SIlya Dryomov dout("%s %p (was %d)\n", __func__, req, 3699e94af20SIlya Dryomov atomic_read(&req->r_kref.refcount)); 3709e94af20SIlya Dryomov kref_get(&req->r_kref); 3719e94af20SIlya Dryomov } 3729e94af20SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_get_request); 3739e94af20SIlya Dryomov 3749e94af20SIlya Dryomov void ceph_osdc_put_request(struct ceph_osd_request *req) 3759e94af20SIlya Dryomov { 3763ed97d63SIlya Dryomov if (req) { 3779e94af20SIlya Dryomov dout("%s %p (was %d)\n", __func__, req, 3789e94af20SIlya Dryomov atomic_read(&req->r_kref.refcount)); 3799e94af20SIlya Dryomov kref_put(&req->r_kref, ceph_osdc_release_request); 3809e94af20SIlya Dryomov } 3813ed97d63SIlya Dryomov } 3829e94af20SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_put_request); 3833d14c5d2SYehuda Sadeh 3843d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc, 3853d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 3861b83bef2SSage Weil unsigned int num_ops, 3873d14c5d2SYehuda Sadeh bool use_mempool, 38854a54007SAlex Elder gfp_t gfp_flags) 3893d14c5d2SYehuda Sadeh { 3903d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 3913d14c5d2SYehuda Sadeh 3923d14c5d2SYehuda Sadeh if (use_mempool) { 3933f1af42aSIlya Dryomov BUG_ON(num_ops > CEPH_OSD_SLAB_OPS); 3943d14c5d2SYehuda Sadeh req = mempool_alloc(osdc->req_mempool, gfp_flags); 3953f1af42aSIlya Dryomov } else if (num_ops <= CEPH_OSD_SLAB_OPS) { 3963f1af42aSIlya Dryomov req = kmem_cache_alloc(ceph_osd_request_cache, gfp_flags); 3973d14c5d2SYehuda Sadeh } else { 3983f1af42aSIlya Dryomov BUG_ON(num_ops > CEPH_OSD_MAX_OPS); 3993f1af42aSIlya Dryomov req = kmalloc(sizeof(*req) + num_ops * sizeof(req->r_ops[0]), 4003f1af42aSIlya Dryomov gfp_flags); 4013d14c5d2SYehuda Sadeh } 4023f1af42aSIlya Dryomov if (unlikely(!req)) 4033d14c5d2SYehuda Sadeh return NULL; 4043d14c5d2SYehuda Sadeh 4053f1af42aSIlya Dryomov /* req only, each op is zeroed in _osd_req_op_init() */ 4063f1af42aSIlya Dryomov memset(req, 0, sizeof(*req)); 4073f1af42aSIlya Dryomov 4083d14c5d2SYehuda Sadeh req->r_osdc = osdc; 4093d14c5d2SYehuda Sadeh req->r_mempool = use_mempool; 41079528734SAlex Elder req->r_num_ops = num_ops; 41184127282SIlya Dryomov req->r_snapid = CEPH_NOSNAP; 41284127282SIlya Dryomov req->r_snapc = ceph_get_snap_context(snapc); 4133d14c5d2SYehuda Sadeh 4143d14c5d2SYehuda Sadeh kref_init(&req->r_kref); 4153d14c5d2SYehuda Sadeh init_completion(&req->r_completion); 4163d14c5d2SYehuda Sadeh init_completion(&req->r_safe_completion); 417a978fa20SAlex Elder RB_CLEAR_NODE(&req->r_node); 4183d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&req->r_unsafe_item); 419a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&req->r_linger_item); 4201d0326b1SIlya Dryomov INIT_LIST_HEAD(&req->r_linger_osd_item); 421935b639aSSage Weil INIT_LIST_HEAD(&req->r_req_lru_item); 422cd43045cSSage Weil INIT_LIST_HEAD(&req->r_osd_item); 423cd43045cSSage Weil 424a66dd383SIlya Dryomov target_init(&req->r_t); 42522116525SIlya Dryomov 42613d1ad16SIlya Dryomov dout("%s req %p\n", __func__, req); 42713d1ad16SIlya Dryomov return req; 4283f1af42aSIlya Dryomov } 42913d1ad16SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_alloc_request); 4303f1af42aSIlya Dryomov 43113d1ad16SIlya Dryomov int ceph_osdc_alloc_messages(struct ceph_osd_request *req, gfp_t gfp) 43213d1ad16SIlya Dryomov { 43313d1ad16SIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 43413d1ad16SIlya Dryomov struct ceph_msg *msg; 43513d1ad16SIlya Dryomov int msg_size; 4363d14c5d2SYehuda Sadeh 437d30291b9SIlya Dryomov WARN_ON(ceph_oid_empty(&req->r_base_oid)); 438d30291b9SIlya Dryomov 43913d1ad16SIlya Dryomov /* create request message */ 440ae458f5aSIlya Dryomov msg_size = 4 + 4 + 4; /* client_inc, osdmap_epoch, flags */ 441ae458f5aSIlya Dryomov msg_size += 4 + 4 + 4 + 8; /* mtime, reassert_version */ 442ae458f5aSIlya Dryomov msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */ 443ae458f5aSIlya Dryomov msg_size += 1 + 8 + 4 + 4; /* pgid */ 44413d1ad16SIlya Dryomov msg_size += 4 + req->r_base_oid.name_len; /* oid */ 44513d1ad16SIlya Dryomov msg_size += 2 + req->r_num_ops * sizeof(struct ceph_osd_op); 446ae458f5aSIlya Dryomov msg_size += 8; /* snapid */ 447ae458f5aSIlya Dryomov msg_size += 8; /* snap_seq */ 44813d1ad16SIlya Dryomov msg_size += 4 + 8 * (req->r_snapc ? req->r_snapc->num_snaps : 0); 449ae458f5aSIlya Dryomov msg_size += 4; /* retry_attempt */ 450ae458f5aSIlya Dryomov 45113d1ad16SIlya Dryomov if (req->r_mempool) 4523d14c5d2SYehuda Sadeh msg = ceph_msgpool_get(&osdc->msgpool_op, 0); 4533d14c5d2SYehuda Sadeh else 45413d1ad16SIlya Dryomov msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp, true); 45513d1ad16SIlya Dryomov if (!msg) 45613d1ad16SIlya Dryomov return -ENOMEM; 4573d14c5d2SYehuda Sadeh 4583d14c5d2SYehuda Sadeh memset(msg->front.iov_base, 0, msg->front.iov_len); 4593d14c5d2SYehuda Sadeh req->r_request = msg; 4603d14c5d2SYehuda Sadeh 46113d1ad16SIlya Dryomov /* create reply message */ 46213d1ad16SIlya Dryomov msg_size = OSD_OPREPLY_FRONT_LEN; 463711da55dSIlya Dryomov msg_size += req->r_base_oid.name_len; 464711da55dSIlya Dryomov msg_size += req->r_num_ops * sizeof(struct ceph_osd_op); 46513d1ad16SIlya Dryomov 46613d1ad16SIlya Dryomov if (req->r_mempool) 46713d1ad16SIlya Dryomov msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0); 46813d1ad16SIlya Dryomov else 46913d1ad16SIlya Dryomov msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, msg_size, gfp, true); 47013d1ad16SIlya Dryomov if (!msg) 47113d1ad16SIlya Dryomov return -ENOMEM; 47213d1ad16SIlya Dryomov 47313d1ad16SIlya Dryomov req->r_reply = msg; 47413d1ad16SIlya Dryomov 47513d1ad16SIlya Dryomov return 0; 47613d1ad16SIlya Dryomov } 47713d1ad16SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_alloc_messages); 4783d14c5d2SYehuda Sadeh 479a8dd0a37SAlex Elder static bool osd_req_opcode_valid(u16 opcode) 480a8dd0a37SAlex Elder { 481a8dd0a37SAlex Elder switch (opcode) { 48270b5bfa3SIlya Dryomov #define GENERATE_CASE(op, opcode, str) case CEPH_OSD_OP_##op: return true; 48370b5bfa3SIlya Dryomov __CEPH_FORALL_OSD_OPS(GENERATE_CASE) 48470b5bfa3SIlya Dryomov #undef GENERATE_CASE 485a8dd0a37SAlex Elder default: 486a8dd0a37SAlex Elder return false; 487a8dd0a37SAlex Elder } 488a8dd0a37SAlex Elder } 489a8dd0a37SAlex Elder 49033803f33SAlex Elder /* 49133803f33SAlex Elder * This is an osd op init function for opcodes that have no data or 49233803f33SAlex Elder * other information associated with them. It also serves as a 49333803f33SAlex Elder * common init routine for all the other init functions, below. 49433803f33SAlex Elder */ 495c99d2d4aSAlex Elder static struct ceph_osd_req_op * 49649719778SAlex Elder _osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which, 497144cba14SYan, Zheng u16 opcode, u32 flags) 49833803f33SAlex Elder { 499c99d2d4aSAlex Elder struct ceph_osd_req_op *op; 500c99d2d4aSAlex Elder 501c99d2d4aSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 50233803f33SAlex Elder BUG_ON(!osd_req_opcode_valid(opcode)); 50333803f33SAlex Elder 504c99d2d4aSAlex Elder op = &osd_req->r_ops[which]; 50533803f33SAlex Elder memset(op, 0, sizeof (*op)); 50633803f33SAlex Elder op->op = opcode; 507144cba14SYan, Zheng op->flags = flags; 508c99d2d4aSAlex Elder 509c99d2d4aSAlex Elder return op; 51033803f33SAlex Elder } 51133803f33SAlex Elder 51249719778SAlex Elder void osd_req_op_init(struct ceph_osd_request *osd_req, 513144cba14SYan, Zheng unsigned int which, u16 opcode, u32 flags) 51449719778SAlex Elder { 515144cba14SYan, Zheng (void)_osd_req_op_init(osd_req, which, opcode, flags); 51649719778SAlex Elder } 51749719778SAlex Elder EXPORT_SYMBOL(osd_req_op_init); 51849719778SAlex Elder 519c99d2d4aSAlex Elder void osd_req_op_extent_init(struct ceph_osd_request *osd_req, 520c99d2d4aSAlex Elder unsigned int which, u16 opcode, 52133803f33SAlex Elder u64 offset, u64 length, 52233803f33SAlex Elder u64 truncate_size, u32 truncate_seq) 52333803f33SAlex Elder { 524144cba14SYan, Zheng struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, 525144cba14SYan, Zheng opcode, 0); 52633803f33SAlex Elder size_t payload_len = 0; 52733803f33SAlex Elder 528ad7a60deSLi Wang BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE && 529e30b7577SIlya Dryomov opcode != CEPH_OSD_OP_WRITEFULL && opcode != CEPH_OSD_OP_ZERO && 530e30b7577SIlya Dryomov opcode != CEPH_OSD_OP_TRUNCATE); 53133803f33SAlex Elder 53233803f33SAlex Elder op->extent.offset = offset; 53333803f33SAlex Elder op->extent.length = length; 53433803f33SAlex Elder op->extent.truncate_size = truncate_size; 53533803f33SAlex Elder op->extent.truncate_seq = truncate_seq; 536e30b7577SIlya Dryomov if (opcode == CEPH_OSD_OP_WRITE || opcode == CEPH_OSD_OP_WRITEFULL) 53733803f33SAlex Elder payload_len += length; 53833803f33SAlex Elder 539de2aa102SIlya Dryomov op->indata_len = payload_len; 54033803f33SAlex Elder } 54133803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_init); 54233803f33SAlex Elder 543c99d2d4aSAlex Elder void osd_req_op_extent_update(struct ceph_osd_request *osd_req, 544c99d2d4aSAlex Elder unsigned int which, u64 length) 545e5975c7cSAlex Elder { 546c99d2d4aSAlex Elder struct ceph_osd_req_op *op; 547c99d2d4aSAlex Elder u64 previous; 548c99d2d4aSAlex Elder 549c99d2d4aSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 550c99d2d4aSAlex Elder op = &osd_req->r_ops[which]; 551c99d2d4aSAlex Elder previous = op->extent.length; 552e5975c7cSAlex Elder 553e5975c7cSAlex Elder if (length == previous) 554e5975c7cSAlex Elder return; /* Nothing to do */ 555e5975c7cSAlex Elder BUG_ON(length > previous); 556e5975c7cSAlex Elder 557e5975c7cSAlex Elder op->extent.length = length; 558de2aa102SIlya Dryomov op->indata_len -= previous - length; 559e5975c7cSAlex Elder } 560e5975c7cSAlex Elder EXPORT_SYMBOL(osd_req_op_extent_update); 561e5975c7cSAlex Elder 5622c63f49aSYan, Zheng void osd_req_op_extent_dup_last(struct ceph_osd_request *osd_req, 5632c63f49aSYan, Zheng unsigned int which, u64 offset_inc) 5642c63f49aSYan, Zheng { 5652c63f49aSYan, Zheng struct ceph_osd_req_op *op, *prev_op; 5662c63f49aSYan, Zheng 5672c63f49aSYan, Zheng BUG_ON(which + 1 >= osd_req->r_num_ops); 5682c63f49aSYan, Zheng 5692c63f49aSYan, Zheng prev_op = &osd_req->r_ops[which]; 5702c63f49aSYan, Zheng op = _osd_req_op_init(osd_req, which + 1, prev_op->op, prev_op->flags); 5712c63f49aSYan, Zheng /* dup previous one */ 5722c63f49aSYan, Zheng op->indata_len = prev_op->indata_len; 5732c63f49aSYan, Zheng op->outdata_len = prev_op->outdata_len; 5742c63f49aSYan, Zheng op->extent = prev_op->extent; 5752c63f49aSYan, Zheng /* adjust offset */ 5762c63f49aSYan, Zheng op->extent.offset += offset_inc; 5772c63f49aSYan, Zheng op->extent.length -= offset_inc; 5782c63f49aSYan, Zheng 5792c63f49aSYan, Zheng if (op->op == CEPH_OSD_OP_WRITE || op->op == CEPH_OSD_OP_WRITEFULL) 5802c63f49aSYan, Zheng op->indata_len -= offset_inc; 5812c63f49aSYan, Zheng } 5822c63f49aSYan, Zheng EXPORT_SYMBOL(osd_req_op_extent_dup_last); 5832c63f49aSYan, Zheng 584c99d2d4aSAlex Elder void osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which, 58504017e29SAlex Elder u16 opcode, const char *class, const char *method) 58633803f33SAlex Elder { 587144cba14SYan, Zheng struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, 588144cba14SYan, Zheng opcode, 0); 5895f562df5SAlex Elder struct ceph_pagelist *pagelist; 59033803f33SAlex Elder size_t payload_len = 0; 59133803f33SAlex Elder size_t size; 59233803f33SAlex Elder 59333803f33SAlex Elder BUG_ON(opcode != CEPH_OSD_OP_CALL); 59433803f33SAlex Elder 5955f562df5SAlex Elder pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS); 5965f562df5SAlex Elder BUG_ON(!pagelist); 5975f562df5SAlex Elder ceph_pagelist_init(pagelist); 5985f562df5SAlex Elder 59933803f33SAlex Elder op->cls.class_name = class; 60033803f33SAlex Elder size = strlen(class); 60133803f33SAlex Elder BUG_ON(size > (size_t) U8_MAX); 60233803f33SAlex Elder op->cls.class_len = size; 6035f562df5SAlex Elder ceph_pagelist_append(pagelist, class, size); 60433803f33SAlex Elder payload_len += size; 60533803f33SAlex Elder 60633803f33SAlex Elder op->cls.method_name = method; 60733803f33SAlex Elder size = strlen(method); 60833803f33SAlex Elder BUG_ON(size > (size_t) U8_MAX); 60933803f33SAlex Elder op->cls.method_len = size; 6105f562df5SAlex Elder ceph_pagelist_append(pagelist, method, size); 61133803f33SAlex Elder payload_len += size; 61233803f33SAlex Elder 613a4ce40a9SAlex Elder osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist); 6145f562df5SAlex Elder 615de2aa102SIlya Dryomov op->indata_len = payload_len; 61633803f33SAlex Elder } 61733803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_init); 6188c042b0dSAlex Elder 619d74b50beSYan, Zheng int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which, 620d74b50beSYan, Zheng u16 opcode, const char *name, const void *value, 621d74b50beSYan, Zheng size_t size, u8 cmp_op, u8 cmp_mode) 622d74b50beSYan, Zheng { 623144cba14SYan, Zheng struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, 624144cba14SYan, Zheng opcode, 0); 625d74b50beSYan, Zheng struct ceph_pagelist *pagelist; 626d74b50beSYan, Zheng size_t payload_len; 627d74b50beSYan, Zheng 628d74b50beSYan, Zheng BUG_ON(opcode != CEPH_OSD_OP_SETXATTR && opcode != CEPH_OSD_OP_CMPXATTR); 629d74b50beSYan, Zheng 630d74b50beSYan, Zheng pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS); 631d74b50beSYan, Zheng if (!pagelist) 632d74b50beSYan, Zheng return -ENOMEM; 633d74b50beSYan, Zheng 634d74b50beSYan, Zheng ceph_pagelist_init(pagelist); 635d74b50beSYan, Zheng 636d74b50beSYan, Zheng payload_len = strlen(name); 637d74b50beSYan, Zheng op->xattr.name_len = payload_len; 638d74b50beSYan, Zheng ceph_pagelist_append(pagelist, name, payload_len); 639d74b50beSYan, Zheng 640d74b50beSYan, Zheng op->xattr.value_len = size; 641d74b50beSYan, Zheng ceph_pagelist_append(pagelist, value, size); 642d74b50beSYan, Zheng payload_len += size; 643d74b50beSYan, Zheng 644d74b50beSYan, Zheng op->xattr.cmp_op = cmp_op; 645d74b50beSYan, Zheng op->xattr.cmp_mode = cmp_mode; 646d74b50beSYan, Zheng 647d74b50beSYan, Zheng ceph_osd_data_pagelist_init(&op->xattr.osd_data, pagelist); 648de2aa102SIlya Dryomov op->indata_len = payload_len; 649d74b50beSYan, Zheng return 0; 650d74b50beSYan, Zheng } 651d74b50beSYan, Zheng EXPORT_SYMBOL(osd_req_op_xattr_init); 652d74b50beSYan, Zheng 653c99d2d4aSAlex Elder void osd_req_op_watch_init(struct ceph_osd_request *osd_req, 654c99d2d4aSAlex Elder unsigned int which, u16 opcode, 65533803f33SAlex Elder u64 cookie, u64 version, int flag) 65633803f33SAlex Elder { 657144cba14SYan, Zheng struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, 658144cba14SYan, Zheng opcode, 0); 65933803f33SAlex Elder 660c99d2d4aSAlex Elder BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH); 66133803f33SAlex Elder 66233803f33SAlex Elder op->watch.cookie = cookie; 6639ef1ee5aSAlex Elder op->watch.ver = version; 66433803f33SAlex Elder if (opcode == CEPH_OSD_OP_WATCH && flag) 66533803f33SAlex Elder op->watch.flag = (u8)1; 66633803f33SAlex Elder } 66733803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_watch_init); 66833803f33SAlex Elder 669c647b8a8SIlya Dryomov void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req, 670c647b8a8SIlya Dryomov unsigned int which, 671c647b8a8SIlya Dryomov u64 expected_object_size, 672c647b8a8SIlya Dryomov u64 expected_write_size) 673c647b8a8SIlya Dryomov { 674c647b8a8SIlya Dryomov struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, 675144cba14SYan, Zheng CEPH_OSD_OP_SETALLOCHINT, 676144cba14SYan, Zheng 0); 677c647b8a8SIlya Dryomov 678c647b8a8SIlya Dryomov op->alloc_hint.expected_object_size = expected_object_size; 679c647b8a8SIlya Dryomov op->alloc_hint.expected_write_size = expected_write_size; 680c647b8a8SIlya Dryomov 681c647b8a8SIlya Dryomov /* 682c647b8a8SIlya Dryomov * CEPH_OSD_OP_SETALLOCHINT op is advisory and therefore deemed 683c647b8a8SIlya Dryomov * not worth a feature bit. Set FAILOK per-op flag to make 684c647b8a8SIlya Dryomov * sure older osds don't trip over an unsupported opcode. 685c647b8a8SIlya Dryomov */ 686c647b8a8SIlya Dryomov op->flags |= CEPH_OSD_OP_FLAG_FAILOK; 687c647b8a8SIlya Dryomov } 688c647b8a8SIlya Dryomov EXPORT_SYMBOL(osd_req_op_alloc_hint_init); 689c647b8a8SIlya Dryomov 69090af3602SAlex Elder static void ceph_osdc_msg_data_add(struct ceph_msg *msg, 691ec9123c5SAlex Elder struct ceph_osd_data *osd_data) 692ec9123c5SAlex Elder { 693ec9123c5SAlex Elder u64 length = ceph_osd_data_length(osd_data); 694ec9123c5SAlex Elder 695ec9123c5SAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 696ec9123c5SAlex Elder BUG_ON(length > (u64) SIZE_MAX); 697ec9123c5SAlex Elder if (length) 69890af3602SAlex Elder ceph_msg_data_add_pages(msg, osd_data->pages, 699ec9123c5SAlex Elder length, osd_data->alignment); 700ec9123c5SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) { 701ec9123c5SAlex Elder BUG_ON(!length); 70290af3602SAlex Elder ceph_msg_data_add_pagelist(msg, osd_data->pagelist); 703ec9123c5SAlex Elder #ifdef CONFIG_BLOCK 704ec9123c5SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) { 70590af3602SAlex Elder ceph_msg_data_add_bio(msg, osd_data->bio, length); 706ec9123c5SAlex Elder #endif 707ec9123c5SAlex Elder } else { 708ec9123c5SAlex Elder BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE); 709ec9123c5SAlex Elder } 710ec9123c5SAlex Elder } 711ec9123c5SAlex Elder 712bb873b53SIlya Dryomov static u32 osd_req_encode_op(struct ceph_osd_op *dst, 713bb873b53SIlya Dryomov const struct ceph_osd_req_op *src) 7143d14c5d2SYehuda Sadeh { 715a8dd0a37SAlex Elder if (WARN_ON(!osd_req_opcode_valid(src->op))) { 716a8dd0a37SAlex Elder pr_err("unrecognized osd opcode %d\n", src->op); 717a8dd0a37SAlex Elder 718a8dd0a37SAlex Elder return 0; 719a8dd0a37SAlex Elder } 7203d14c5d2SYehuda Sadeh 721065a68f9SAlex Elder switch (src->op) { 722fbfab539SAlex Elder case CEPH_OSD_OP_STAT: 723fbfab539SAlex Elder break; 7243d14c5d2SYehuda Sadeh case CEPH_OSD_OP_READ: 7253d14c5d2SYehuda Sadeh case CEPH_OSD_OP_WRITE: 726e30b7577SIlya Dryomov case CEPH_OSD_OP_WRITEFULL: 727ad7a60deSLi Wang case CEPH_OSD_OP_ZERO: 728ad7a60deSLi Wang case CEPH_OSD_OP_TRUNCATE: 729175face2SAlex Elder dst->extent.offset = cpu_to_le64(src->extent.offset); 730175face2SAlex Elder dst->extent.length = cpu_to_le64(src->extent.length); 7313d14c5d2SYehuda Sadeh dst->extent.truncate_size = 7323d14c5d2SYehuda Sadeh cpu_to_le64(src->extent.truncate_size); 7333d14c5d2SYehuda Sadeh dst->extent.truncate_seq = 7343d14c5d2SYehuda Sadeh cpu_to_le32(src->extent.truncate_seq); 7353d14c5d2SYehuda Sadeh break; 7363d14c5d2SYehuda Sadeh case CEPH_OSD_OP_CALL: 7373d14c5d2SYehuda Sadeh dst->cls.class_len = src->cls.class_len; 7383d14c5d2SYehuda Sadeh dst->cls.method_len = src->cls.method_len; 739bb873b53SIlya Dryomov dst->cls.indata_len = cpu_to_le32(src->cls.indata_len); 7403d14c5d2SYehuda Sadeh break; 7413d14c5d2SYehuda Sadeh case CEPH_OSD_OP_STARTSYNC: 7423d14c5d2SYehuda Sadeh break; 743a40c4f10SYehuda Sadeh case CEPH_OSD_OP_NOTIFY_ACK: 744a40c4f10SYehuda Sadeh case CEPH_OSD_OP_WATCH: 745a40c4f10SYehuda Sadeh dst->watch.cookie = cpu_to_le64(src->watch.cookie); 746a40c4f10SYehuda Sadeh dst->watch.ver = cpu_to_le64(src->watch.ver); 747a40c4f10SYehuda Sadeh dst->watch.flag = src->watch.flag; 748a40c4f10SYehuda Sadeh break; 749c647b8a8SIlya Dryomov case CEPH_OSD_OP_SETALLOCHINT: 750c647b8a8SIlya Dryomov dst->alloc_hint.expected_object_size = 751c647b8a8SIlya Dryomov cpu_to_le64(src->alloc_hint.expected_object_size); 752c647b8a8SIlya Dryomov dst->alloc_hint.expected_write_size = 753c647b8a8SIlya Dryomov cpu_to_le64(src->alloc_hint.expected_write_size); 754c647b8a8SIlya Dryomov break; 755d74b50beSYan, Zheng case CEPH_OSD_OP_SETXATTR: 756d74b50beSYan, Zheng case CEPH_OSD_OP_CMPXATTR: 757d74b50beSYan, Zheng dst->xattr.name_len = cpu_to_le32(src->xattr.name_len); 758d74b50beSYan, Zheng dst->xattr.value_len = cpu_to_le32(src->xattr.value_len); 759d74b50beSYan, Zheng dst->xattr.cmp_op = src->xattr.cmp_op; 760d74b50beSYan, Zheng dst->xattr.cmp_mode = src->xattr.cmp_mode; 761d74b50beSYan, Zheng break; 762864e9197SYan, Zheng case CEPH_OSD_OP_CREATE: 763864e9197SYan, Zheng case CEPH_OSD_OP_DELETE: 764864e9197SYan, Zheng break; 7653d14c5d2SYehuda Sadeh default: 7664c46459cSAlex Elder pr_err("unsupported osd opcode %s\n", 7678f63ca2dSAlex Elder ceph_osd_op_name(src->op)); 7684c46459cSAlex Elder WARN_ON(1); 769a8dd0a37SAlex Elder 770a8dd0a37SAlex Elder return 0; 7713d14c5d2SYehuda Sadeh } 7727b25bf5fSIlya Dryomov 773a8dd0a37SAlex Elder dst->op = cpu_to_le16(src->op); 7747b25bf5fSIlya Dryomov dst->flags = cpu_to_le32(src->flags); 775de2aa102SIlya Dryomov dst->payload_len = cpu_to_le32(src->indata_len); 776175face2SAlex Elder 777bb873b53SIlya Dryomov return src->indata_len; 7783d14c5d2SYehuda Sadeh } 7793d14c5d2SYehuda Sadeh 7803d14c5d2SYehuda Sadeh /* 7813d14c5d2SYehuda Sadeh * build new request AND message, calculate layout, and adjust file 7823d14c5d2SYehuda Sadeh * extent as needed. 7833d14c5d2SYehuda Sadeh * 7843d14c5d2SYehuda Sadeh * if the file was recently truncated, we include information about its 7853d14c5d2SYehuda Sadeh * old and new size so that the object can be updated appropriately. (we 7863d14c5d2SYehuda Sadeh * avoid synchronously deleting truncated objects because it's slow.) 7873d14c5d2SYehuda Sadeh * 7883d14c5d2SYehuda Sadeh * if @do_sync, include a 'startsync' command so that the osd will flush 7893d14c5d2SYehuda Sadeh * data quickly. 7903d14c5d2SYehuda Sadeh */ 7913d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc, 7923d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 7933d14c5d2SYehuda Sadeh struct ceph_vino vino, 794715e4cd4SYan, Zheng u64 off, u64 *plen, 795715e4cd4SYan, Zheng unsigned int which, int num_ops, 7963d14c5d2SYehuda Sadeh int opcode, int flags, 7973d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 7983d14c5d2SYehuda Sadeh u32 truncate_seq, 7993d14c5d2SYehuda Sadeh u64 truncate_size, 800153e5167SAlex Elder bool use_mempool) 8013d14c5d2SYehuda Sadeh { 8023d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 80375d1c941SAlex Elder u64 objnum = 0; 80475d1c941SAlex Elder u64 objoff = 0; 80575d1c941SAlex Elder u64 objlen = 0; 8066816282dSSage Weil int r; 8073d14c5d2SYehuda Sadeh 808ad7a60deSLi Wang BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE && 809864e9197SYan, Zheng opcode != CEPH_OSD_OP_ZERO && opcode != CEPH_OSD_OP_TRUNCATE && 810864e9197SYan, Zheng opcode != CEPH_OSD_OP_CREATE && opcode != CEPH_OSD_OP_DELETE); 8113d14c5d2SYehuda Sadeh 812acead002SAlex Elder req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool, 813ae7ca4a3SAlex Elder GFP_NOFS); 81413d1ad16SIlya Dryomov if (!req) { 81513d1ad16SIlya Dryomov r = -ENOMEM; 81613d1ad16SIlya Dryomov goto fail; 81713d1ad16SIlya Dryomov } 81879528734SAlex Elder 8193d14c5d2SYehuda Sadeh /* calculate max write size */ 820a19dadfbSAlex Elder r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen); 82113d1ad16SIlya Dryomov if (r) 82213d1ad16SIlya Dryomov goto fail; 823a19dadfbSAlex Elder 824864e9197SYan, Zheng if (opcode == CEPH_OSD_OP_CREATE || opcode == CEPH_OSD_OP_DELETE) { 825144cba14SYan, Zheng osd_req_op_init(req, which, opcode, 0); 826864e9197SYan, Zheng } else { 827864e9197SYan, Zheng u32 object_size = le32_to_cpu(layout->fl_object_size); 828864e9197SYan, Zheng u32 object_base = off - objoff; 829ccca4e37SYan, Zheng if (!(truncate_seq == 1 && truncate_size == -1ULL)) { 830d18d1e28SAlex Elder if (truncate_size <= object_base) { 831d18d1e28SAlex Elder truncate_size = 0; 832d18d1e28SAlex Elder } else { 833d18d1e28SAlex Elder truncate_size -= object_base; 834d18d1e28SAlex Elder if (truncate_size > object_size) 835d18d1e28SAlex Elder truncate_size = object_size; 836d18d1e28SAlex Elder } 837ccca4e37SYan, Zheng } 838715e4cd4SYan, Zheng osd_req_op_extent_init(req, which, opcode, objoff, objlen, 839b0270324SAlex Elder truncate_size, truncate_seq); 840864e9197SYan, Zheng } 841d18d1e28SAlex Elder 842bb873b53SIlya Dryomov req->r_flags = flags; 8433c972c95SIlya Dryomov req->r_base_oloc.pool = ceph_file_layout_pg_pool(*layout); 844d30291b9SIlya Dryomov ceph_oid_printf(&req->r_base_oid, "%llx.%08llx", vino.ino, objnum); 845dbe0fc41SAlex Elder 846bb873b53SIlya Dryomov req->r_snapid = vino.snap; 847bb873b53SIlya Dryomov if (flags & CEPH_OSD_FLAG_WRITE) 848bb873b53SIlya Dryomov req->r_data_offset = off; 849bb873b53SIlya Dryomov 85013d1ad16SIlya Dryomov r = ceph_osdc_alloc_messages(req, GFP_NOFS); 85113d1ad16SIlya Dryomov if (r) 85213d1ad16SIlya Dryomov goto fail; 85313d1ad16SIlya Dryomov 8543d14c5d2SYehuda Sadeh return req; 85513d1ad16SIlya Dryomov 85613d1ad16SIlya Dryomov fail: 85713d1ad16SIlya Dryomov ceph_osdc_put_request(req); 85813d1ad16SIlya Dryomov return ERR_PTR(r); 8593d14c5d2SYehuda Sadeh } 8603d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_new_request); 8613d14c5d2SYehuda Sadeh 8623d14c5d2SYehuda Sadeh /* 8633d14c5d2SYehuda Sadeh * We keep osd requests in an rbtree, sorted by ->r_tid. 8643d14c5d2SYehuda Sadeh */ 865fcd00b68SIlya Dryomov DEFINE_RB_FUNCS(request, struct ceph_osd_request, r_tid, r_node) 8663d14c5d2SYehuda Sadeh 8673d14c5d2SYehuda Sadeh static struct ceph_osd_request * 8683d14c5d2SYehuda Sadeh __lookup_request_ge(struct ceph_osd_client *osdc, 8693d14c5d2SYehuda Sadeh u64 tid) 8703d14c5d2SYehuda Sadeh { 8713d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 8723d14c5d2SYehuda Sadeh struct rb_node *n = osdc->requests.rb_node; 8733d14c5d2SYehuda Sadeh 8743d14c5d2SYehuda Sadeh while (n) { 8753d14c5d2SYehuda Sadeh req = rb_entry(n, struct ceph_osd_request, r_node); 8763d14c5d2SYehuda Sadeh if (tid < req->r_tid) { 8773d14c5d2SYehuda Sadeh if (!n->rb_left) 8783d14c5d2SYehuda Sadeh return req; 8793d14c5d2SYehuda Sadeh n = n->rb_left; 8803d14c5d2SYehuda Sadeh } else if (tid > req->r_tid) { 8813d14c5d2SYehuda Sadeh n = n->rb_right; 8823d14c5d2SYehuda Sadeh } else { 8833d14c5d2SYehuda Sadeh return req; 8843d14c5d2SYehuda Sadeh } 8853d14c5d2SYehuda Sadeh } 8863d14c5d2SYehuda Sadeh return NULL; 8873d14c5d2SYehuda Sadeh } 8883d14c5d2SYehuda Sadeh 8892cc6128aSIlya Dryomov static void __kick_linger_request(struct ceph_osd_request *req) 8902cc6128aSIlya Dryomov { 8912cc6128aSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 8922cc6128aSIlya Dryomov struct ceph_osd *osd = req->r_osd; 8932cc6128aSIlya Dryomov 8942cc6128aSIlya Dryomov /* 8952cc6128aSIlya Dryomov * Linger requests need to be resent with a new tid to avoid 8962cc6128aSIlya Dryomov * the dup op detection logic on the OSDs. Achieve this with 8972cc6128aSIlya Dryomov * a re-register dance instead of open-coding. 8982cc6128aSIlya Dryomov */ 8992cc6128aSIlya Dryomov ceph_osdc_get_request(req); 9002cc6128aSIlya Dryomov if (!list_empty(&req->r_linger_item)) 9012cc6128aSIlya Dryomov __unregister_linger_request(osdc, req); 9022cc6128aSIlya Dryomov else 9032cc6128aSIlya Dryomov __unregister_request(osdc, req); 9042cc6128aSIlya Dryomov __register_request(osdc, req); 9052cc6128aSIlya Dryomov ceph_osdc_put_request(req); 9062cc6128aSIlya Dryomov 9072cc6128aSIlya Dryomov /* 9082cc6128aSIlya Dryomov * Unless request has been registered as both normal and 9092cc6128aSIlya Dryomov * lingering, __unregister{,_linger}_request clears r_osd. 9102cc6128aSIlya Dryomov * However, here we need to preserve r_osd to make sure we 9112cc6128aSIlya Dryomov * requeue on the same OSD. 9122cc6128aSIlya Dryomov */ 9132cc6128aSIlya Dryomov WARN_ON(req->r_osd || !osd); 9142cc6128aSIlya Dryomov req->r_osd = osd; 9152cc6128aSIlya Dryomov 9162cc6128aSIlya Dryomov dout("%s requeueing %p tid %llu\n", __func__, req, req->r_tid); 9172cc6128aSIlya Dryomov __enqueue_request(req); 9182cc6128aSIlya Dryomov } 9192cc6128aSIlya Dryomov 9206f6c7006SSage Weil /* 9216f6c7006SSage Weil * Resubmit requests pending on the given osd. 9226f6c7006SSage Weil */ 9236f6c7006SSage Weil static void __kick_osd_requests(struct ceph_osd_client *osdc, 9246f6c7006SSage Weil struct ceph_osd *osd) 9256f6c7006SSage Weil { 926a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 927e02493c0SAlex Elder LIST_HEAD(resend); 9282cc6128aSIlya Dryomov LIST_HEAD(resend_linger); 9296f6c7006SSage Weil int err; 9306f6c7006SSage Weil 9312cc6128aSIlya Dryomov dout("%s osd%d\n", __func__, osd->o_osd); 9326f6c7006SSage Weil err = __reset_osd(osdc, osd); 933685a7555SAlex Elder if (err) 9346f6c7006SSage Weil return; 9352cc6128aSIlya Dryomov 936e02493c0SAlex Elder /* 937e02493c0SAlex Elder * Build up a list of requests to resend by traversing the 938e02493c0SAlex Elder * osd's list of requests. Requests for a given object are 939e02493c0SAlex Elder * sent in tid order, and that is also the order they're 940e02493c0SAlex Elder * kept on this list. Therefore all requests that are in 941e02493c0SAlex Elder * flight will be found first, followed by all requests that 942e02493c0SAlex Elder * have not yet been sent. And to resend requests while 943e02493c0SAlex Elder * preserving this order we will want to put any sent 944e02493c0SAlex Elder * requests back on the front of the osd client's unsent 945e02493c0SAlex Elder * list. 946e02493c0SAlex Elder * 947e02493c0SAlex Elder * So we build a separate ordered list of already-sent 948e02493c0SAlex Elder * requests for the affected osd and splice it onto the 949e02493c0SAlex Elder * front of the osd client's unsent list. Once we've seen a 950e02493c0SAlex Elder * request that has not yet been sent we're done. Those 951e02493c0SAlex Elder * requests are already sitting right where they belong. 952e02493c0SAlex Elder */ 9536f6c7006SSage Weil list_for_each_entry(req, &osd->o_requests, r_osd_item) { 954e02493c0SAlex Elder if (!req->r_sent) 955e02493c0SAlex Elder break; 9562cc6128aSIlya Dryomov 9572cc6128aSIlya Dryomov if (!req->r_linger) { 9582cc6128aSIlya Dryomov dout("%s requeueing %p tid %llu\n", __func__, req, 9592cc6128aSIlya Dryomov req->r_tid); 960e02493c0SAlex Elder list_move_tail(&req->r_req_lru_item, &resend); 9616f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 9622cc6128aSIlya Dryomov } else { 9632cc6128aSIlya Dryomov list_move_tail(&req->r_req_lru_item, &resend_linger); 9642cc6128aSIlya Dryomov } 9656f6c7006SSage Weil } 966e02493c0SAlex Elder list_splice(&resend, &osdc->req_unsent); 967a40c4f10SYehuda Sadeh 968e02493c0SAlex Elder /* 9692cc6128aSIlya Dryomov * Both registered and not yet registered linger requests are 9702cc6128aSIlya Dryomov * enqueued with a new tid on the same OSD. We add/move them 9712cc6128aSIlya Dryomov * to req_unsent/o_requests at the end to keep things in tid 9722cc6128aSIlya Dryomov * order. 973e02493c0SAlex Elder */ 974a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osd->o_linger_requests, 9751d0326b1SIlya Dryomov r_linger_osd_item) { 9762cc6128aSIlya Dryomov WARN_ON(!list_empty(&req->r_req_lru_item)); 9772cc6128aSIlya Dryomov __kick_linger_request(req); 978a40c4f10SYehuda Sadeh } 9792cc6128aSIlya Dryomov 9802cc6128aSIlya Dryomov list_for_each_entry_safe(req, nreq, &resend_linger, r_req_lru_item) 9812cc6128aSIlya Dryomov __kick_linger_request(req); 9826f6c7006SSage Weil } 9836f6c7006SSage Weil 9843d14c5d2SYehuda Sadeh /* 9853d14c5d2SYehuda Sadeh * If the osd connection drops, we need to resubmit all requests. 9863d14c5d2SYehuda Sadeh */ 9873d14c5d2SYehuda Sadeh static void osd_reset(struct ceph_connection *con) 9883d14c5d2SYehuda Sadeh { 9893d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 9903d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 9913d14c5d2SYehuda Sadeh 9923d14c5d2SYehuda Sadeh if (!osd) 9933d14c5d2SYehuda Sadeh return; 9943d14c5d2SYehuda Sadeh dout("osd_reset osd%d\n", osd->o_osd); 9953d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 9963d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 99783aff95eSSage Weil mutex_lock(&osdc->request_mutex); 99883aff95eSSage Weil __kick_osd_requests(osdc, osd); 999f9d25199SAlex Elder __send_queued(osdc); 100083aff95eSSage Weil mutex_unlock(&osdc->request_mutex); 10013d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 10023d14c5d2SYehuda Sadeh } 10033d14c5d2SYehuda Sadeh 10043d14c5d2SYehuda Sadeh /* 10053d14c5d2SYehuda Sadeh * Track open sessions with osds. 10063d14c5d2SYehuda Sadeh */ 1007e10006f8SAlex Elder static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum) 10083d14c5d2SYehuda Sadeh { 10093d14c5d2SYehuda Sadeh struct ceph_osd *osd; 10103d14c5d2SYehuda Sadeh 10113d14c5d2SYehuda Sadeh osd = kzalloc(sizeof(*osd), GFP_NOFS); 10123d14c5d2SYehuda Sadeh if (!osd) 10133d14c5d2SYehuda Sadeh return NULL; 10143d14c5d2SYehuda Sadeh 10153d14c5d2SYehuda Sadeh atomic_set(&osd->o_ref, 1); 10163d14c5d2SYehuda Sadeh osd->o_osdc = osdc; 1017e10006f8SAlex Elder osd->o_osd = onum; 1018f407731dSAlex Elder RB_CLEAR_NODE(&osd->o_node); 10193d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_requests); 1020a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osd->o_linger_requests); 10213d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_osd_lru); 10223d14c5d2SYehuda Sadeh osd->o_incarnation = 1; 10233d14c5d2SYehuda Sadeh 1024b7a9e5ddSSage Weil ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr); 10253d14c5d2SYehuda Sadeh 10263d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_keepalive_item); 10273d14c5d2SYehuda Sadeh return osd; 10283d14c5d2SYehuda Sadeh } 10293d14c5d2SYehuda Sadeh 10303d14c5d2SYehuda Sadeh static struct ceph_osd *get_osd(struct ceph_osd *osd) 10313d14c5d2SYehuda Sadeh { 10323d14c5d2SYehuda Sadeh if (atomic_inc_not_zero(&osd->o_ref)) { 10333d14c5d2SYehuda Sadeh dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1, 10343d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref)); 10353d14c5d2SYehuda Sadeh return osd; 10363d14c5d2SYehuda Sadeh } else { 10373d14c5d2SYehuda Sadeh dout("get_osd %p FAIL\n", osd); 10383d14c5d2SYehuda Sadeh return NULL; 10393d14c5d2SYehuda Sadeh } 10403d14c5d2SYehuda Sadeh } 10413d14c5d2SYehuda Sadeh 10423d14c5d2SYehuda Sadeh static void put_osd(struct ceph_osd *osd) 10433d14c5d2SYehuda Sadeh { 10443d14c5d2SYehuda Sadeh dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref), 10453d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref) - 1); 1046b28ec2f3SIlya Dryomov if (atomic_dec_and_test(&osd->o_ref)) { 1047b28ec2f3SIlya Dryomov if (osd->o_auth.authorizer) 10486c1ea260SIlya Dryomov ceph_auth_destroy_authorizer(osd->o_auth.authorizer); 10493d14c5d2SYehuda Sadeh kfree(osd); 10503d14c5d2SYehuda Sadeh } 10513d14c5d2SYehuda Sadeh } 10523d14c5d2SYehuda Sadeh 1053fcd00b68SIlya Dryomov DEFINE_RB_FUNCS(osd, struct ceph_osd, o_osd, o_node) 1054fcd00b68SIlya Dryomov 10553d14c5d2SYehuda Sadeh /* 10563d14c5d2SYehuda Sadeh * remove an osd from our map 10573d14c5d2SYehuda Sadeh */ 10583d14c5d2SYehuda Sadeh static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 10593d14c5d2SYehuda Sadeh { 10607eb71e03SIlya Dryomov dout("%s %p osd%d\n", __func__, osd, osd->o_osd); 1061cc9f1f51SIlya Dryomov WARN_ON(!list_empty(&osd->o_requests)); 1062cc9f1f51SIlya Dryomov WARN_ON(!list_empty(&osd->o_linger_requests)); 10637c6e6fc5SIlya Dryomov 10643d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 1065fcd00b68SIlya Dryomov erase_osd(&osdc->osds, osd); 10667eb71e03SIlya Dryomov } 10677eb71e03SIlya Dryomov 10687eb71e03SIlya Dryomov static void remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 10697eb71e03SIlya Dryomov { 10707eb71e03SIlya Dryomov dout("%s %p osd%d\n", __func__, osd, osd->o_osd); 10717eb71e03SIlya Dryomov 10727eb71e03SIlya Dryomov if (!RB_EMPTY_NODE(&osd->o_node)) { 10733d14c5d2SYehuda Sadeh ceph_con_close(&osd->o_con); 10747eb71e03SIlya Dryomov __remove_osd(osdc, osd); 10753d14c5d2SYehuda Sadeh put_osd(osd); 10763d14c5d2SYehuda Sadeh } 10777eb71e03SIlya Dryomov } 10783d14c5d2SYehuda Sadeh 10793d14c5d2SYehuda Sadeh static void __move_osd_to_lru(struct ceph_osd_client *osdc, 10803d14c5d2SYehuda Sadeh struct ceph_osd *osd) 10813d14c5d2SYehuda Sadeh { 1082bbf37ec3SIlya Dryomov dout("%s %p\n", __func__, osd); 10833d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_osd_lru)); 1084bbf37ec3SIlya Dryomov 10853d14c5d2SYehuda Sadeh list_add_tail(&osd->o_osd_lru, &osdc->osd_lru); 1086a319bf56SIlya Dryomov osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl; 10873d14c5d2SYehuda Sadeh } 10883d14c5d2SYehuda Sadeh 1089bbf37ec3SIlya Dryomov static void maybe_move_osd_to_lru(struct ceph_osd_client *osdc, 1090bbf37ec3SIlya Dryomov struct ceph_osd *osd) 1091bbf37ec3SIlya Dryomov { 1092bbf37ec3SIlya Dryomov dout("%s %p\n", __func__, osd); 1093bbf37ec3SIlya Dryomov 1094bbf37ec3SIlya Dryomov if (list_empty(&osd->o_requests) && 1095bbf37ec3SIlya Dryomov list_empty(&osd->o_linger_requests)) 1096bbf37ec3SIlya Dryomov __move_osd_to_lru(osdc, osd); 1097bbf37ec3SIlya Dryomov } 1098bbf37ec3SIlya Dryomov 10993d14c5d2SYehuda Sadeh static void __remove_osd_from_lru(struct ceph_osd *osd) 11003d14c5d2SYehuda Sadeh { 11013d14c5d2SYehuda Sadeh dout("__remove_osd_from_lru %p\n", osd); 11023d14c5d2SYehuda Sadeh if (!list_empty(&osd->o_osd_lru)) 11033d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 11043d14c5d2SYehuda Sadeh } 11053d14c5d2SYehuda Sadeh 11063d14c5d2SYehuda Sadeh /* 11073d14c5d2SYehuda Sadeh * reset osd connect 11083d14c5d2SYehuda Sadeh */ 11093d14c5d2SYehuda Sadeh static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 11103d14c5d2SYehuda Sadeh { 1111c3acb181SAlex Elder struct ceph_entity_addr *peer_addr; 11123d14c5d2SYehuda Sadeh 11133d14c5d2SYehuda Sadeh dout("__reset_osd %p osd%d\n", osd, osd->o_osd); 1114a40c4f10SYehuda Sadeh if (list_empty(&osd->o_requests) && 1115a40c4f10SYehuda Sadeh list_empty(&osd->o_linger_requests)) { 11167eb71e03SIlya Dryomov remove_osd(osdc, osd); 1117c3acb181SAlex Elder return -ENODEV; 1118c3acb181SAlex Elder } 1119c3acb181SAlex Elder 1120c3acb181SAlex Elder peer_addr = &osdc->osdmap->osd_addr[osd->o_osd]; 1121c3acb181SAlex Elder if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) && 11223d14c5d2SYehuda Sadeh !ceph_con_opened(&osd->o_con)) { 1123c3acb181SAlex Elder struct ceph_osd_request *req; 1124c3acb181SAlex Elder 11253d14c5d2SYehuda Sadeh dout("osd addr hasn't changed and connection never opened, " 11260b4af2e8SIlya Dryomov "letting msgr retry\n"); 11273d14c5d2SYehuda Sadeh /* touch each r_stamp for handle_timeout()'s benfit */ 11283d14c5d2SYehuda Sadeh list_for_each_entry(req, &osd->o_requests, r_osd_item) 11293d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 1130c3acb181SAlex Elder 1131c3acb181SAlex Elder return -EAGAIN; 11323d14c5d2SYehuda Sadeh } 1133c3acb181SAlex Elder 1134c3acb181SAlex Elder ceph_con_close(&osd->o_con); 1135c3acb181SAlex Elder ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr); 1136c3acb181SAlex Elder osd->o_incarnation++; 1137c3acb181SAlex Elder 1138c3acb181SAlex Elder return 0; 11393d14c5d2SYehuda Sadeh } 11403d14c5d2SYehuda Sadeh 11413d14c5d2SYehuda Sadeh static void __schedule_osd_timeout(struct ceph_osd_client *osdc) 11423d14c5d2SYehuda Sadeh { 11433d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->timeout_work, 1144a319bf56SIlya Dryomov osdc->client->options->osd_keepalive_timeout); 11453d14c5d2SYehuda Sadeh } 11463d14c5d2SYehuda Sadeh 11473d14c5d2SYehuda Sadeh static void __cancel_osd_timeout(struct ceph_osd_client *osdc) 11483d14c5d2SYehuda Sadeh { 11493d14c5d2SYehuda Sadeh cancel_delayed_work(&osdc->timeout_work); 11503d14c5d2SYehuda Sadeh } 11513d14c5d2SYehuda Sadeh 11523d14c5d2SYehuda Sadeh /* 11533d14c5d2SYehuda Sadeh * Register request, assign tid. If this is the first request, set up 11543d14c5d2SYehuda Sadeh * the timeout event. 11553d14c5d2SYehuda Sadeh */ 1156a40c4f10SYehuda Sadeh static void __register_request(struct ceph_osd_client *osdc, 11573d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 11583d14c5d2SYehuda Sadeh { 11593d14c5d2SYehuda Sadeh req->r_tid = ++osdc->last_tid; 11603d14c5d2SYehuda Sadeh req->r_request->hdr.tid = cpu_to_le64(req->r_tid); 116177f38e0eSSage Weil dout("__register_request %p tid %lld\n", req, req->r_tid); 1162fcd00b68SIlya Dryomov insert_request(&osdc->requests, req); 11633d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 11643d14c5d2SYehuda Sadeh osdc->num_requests++; 11653d14c5d2SYehuda Sadeh if (osdc->num_requests == 1) { 11663d14c5d2SYehuda Sadeh dout(" first request, scheduling timeout\n"); 11673d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 11683d14c5d2SYehuda Sadeh } 1169a40c4f10SYehuda Sadeh } 1170a40c4f10SYehuda Sadeh 11713d14c5d2SYehuda Sadeh /* 11723d14c5d2SYehuda Sadeh * called under osdc->request_mutex 11733d14c5d2SYehuda Sadeh */ 11743d14c5d2SYehuda Sadeh static void __unregister_request(struct ceph_osd_client *osdc, 11753d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 11763d14c5d2SYehuda Sadeh { 117735f9f8a0SSage Weil if (RB_EMPTY_NODE(&req->r_node)) { 117835f9f8a0SSage Weil dout("__unregister_request %p tid %lld not registered\n", 117935f9f8a0SSage Weil req, req->r_tid); 118035f9f8a0SSage Weil return; 118135f9f8a0SSage Weil } 118235f9f8a0SSage Weil 11833d14c5d2SYehuda Sadeh dout("__unregister_request %p tid %lld\n", req, req->r_tid); 1184fcd00b68SIlya Dryomov erase_request(&osdc->requests, req); 11853d14c5d2SYehuda Sadeh osdc->num_requests--; 11863d14c5d2SYehuda Sadeh 11873d14c5d2SYehuda Sadeh if (req->r_osd) { 11883d14c5d2SYehuda Sadeh /* make sure the original request isn't in flight. */ 11896740a845SAlex Elder ceph_msg_revoke(req->r_request); 11903d14c5d2SYehuda Sadeh 11913d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 1192bbf37ec3SIlya Dryomov maybe_move_osd_to_lru(osdc, req->r_osd); 11934f23409eSIlya Dryomov if (list_empty(&req->r_linger_osd_item)) 11943d14c5d2SYehuda Sadeh req->r_osd = NULL; 11953d14c5d2SYehuda Sadeh } 11963d14c5d2SYehuda Sadeh 11977d5f2481SAlex Elder list_del_init(&req->r_req_lru_item); 11983d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 11993d14c5d2SYehuda Sadeh 12003d14c5d2SYehuda Sadeh if (osdc->num_requests == 0) { 12013d14c5d2SYehuda Sadeh dout(" no requests, canceling timeout\n"); 12023d14c5d2SYehuda Sadeh __cancel_osd_timeout(osdc); 12033d14c5d2SYehuda Sadeh } 12043d14c5d2SYehuda Sadeh } 12053d14c5d2SYehuda Sadeh 12063d14c5d2SYehuda Sadeh /* 12073d14c5d2SYehuda Sadeh * Cancel a previously queued request message 12083d14c5d2SYehuda Sadeh */ 12093d14c5d2SYehuda Sadeh static void __cancel_request(struct ceph_osd_request *req) 12103d14c5d2SYehuda Sadeh { 12113d14c5d2SYehuda Sadeh if (req->r_sent && req->r_osd) { 12126740a845SAlex Elder ceph_msg_revoke(req->r_request); 12133d14c5d2SYehuda Sadeh req->r_sent = 0; 12143d14c5d2SYehuda Sadeh } 12153d14c5d2SYehuda Sadeh } 12163d14c5d2SYehuda Sadeh 1217a40c4f10SYehuda Sadeh static void __register_linger_request(struct ceph_osd_client *osdc, 1218a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1219a40c4f10SYehuda Sadeh { 1220af593064SIlya Dryomov dout("%s %p tid %llu\n", __func__, req, req->r_tid); 1221af593064SIlya Dryomov WARN_ON(!req->r_linger); 1222af593064SIlya Dryomov 122396e4dac6SAlex Elder ceph_osdc_get_request(req); 1224a40c4f10SYehuda Sadeh list_add_tail(&req->r_linger_item, &osdc->req_linger); 12256194ea89SSage Weil if (req->r_osd) 12261d0326b1SIlya Dryomov list_add_tail(&req->r_linger_osd_item, 12276194ea89SSage Weil &req->r_osd->o_linger_requests); 1228a40c4f10SYehuda Sadeh } 1229a40c4f10SYehuda Sadeh 1230a40c4f10SYehuda Sadeh static void __unregister_linger_request(struct ceph_osd_client *osdc, 1231a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1232a40c4f10SYehuda Sadeh { 1233af593064SIlya Dryomov WARN_ON(!req->r_linger); 1234af593064SIlya Dryomov 1235af593064SIlya Dryomov if (list_empty(&req->r_linger_item)) { 1236af593064SIlya Dryomov dout("%s %p tid %llu not registered\n", __func__, req, 1237af593064SIlya Dryomov req->r_tid); 1238af593064SIlya Dryomov return; 1239af593064SIlya Dryomov } 1240af593064SIlya Dryomov 1241af593064SIlya Dryomov dout("%s %p tid %llu\n", __func__, req, req->r_tid); 1242a40c4f10SYehuda Sadeh list_del_init(&req->r_linger_item); 1243af593064SIlya Dryomov 124461c74035SAlex Elder if (req->r_osd) { 12451d0326b1SIlya Dryomov list_del_init(&req->r_linger_osd_item); 1246bbf37ec3SIlya Dryomov maybe_move_osd_to_lru(osdc, req->r_osd); 1247fbdb9190SSage Weil if (list_empty(&req->r_osd_item)) 1248a40c4f10SYehuda Sadeh req->r_osd = NULL; 1249a40c4f10SYehuda Sadeh } 125096e4dac6SAlex Elder ceph_osdc_put_request(req); 1251a40c4f10SYehuda Sadeh } 1252a40c4f10SYehuda Sadeh 1253a40c4f10SYehuda Sadeh void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc, 1254a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1255a40c4f10SYehuda Sadeh { 1256a40c4f10SYehuda Sadeh if (!req->r_linger) { 1257a40c4f10SYehuda Sadeh dout("set_request_linger %p\n", req); 1258a40c4f10SYehuda Sadeh req->r_linger = 1; 1259a40c4f10SYehuda Sadeh } 1260a40c4f10SYehuda Sadeh } 1261a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_set_request_linger); 1262a40c4f10SYehuda Sadeh 126363244fa1SIlya Dryomov static bool __pool_full(struct ceph_pg_pool_info *pi) 126463244fa1SIlya Dryomov { 126563244fa1SIlya Dryomov return pi->flags & CEPH_POOL_FLAG_FULL; 126663244fa1SIlya Dryomov } 126763244fa1SIlya Dryomov 12683d14c5d2SYehuda Sadeh /* 1269d29adb34SJosh Durgin * Returns whether a request should be blocked from being sent 1270d29adb34SJosh Durgin * based on the current osdmap and osd_client settings. 1271d29adb34SJosh Durgin * 1272d29adb34SJosh Durgin * Caller should hold map_sem for read. 1273d29adb34SJosh Durgin */ 127463244fa1SIlya Dryomov static bool target_should_be_paused(struct ceph_osd_client *osdc, 127563244fa1SIlya Dryomov const struct ceph_osd_request_target *t, 127663244fa1SIlya Dryomov struct ceph_pg_pool_info *pi) 127763244fa1SIlya Dryomov { 127863244fa1SIlya Dryomov bool pauserd = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD); 127963244fa1SIlya Dryomov bool pausewr = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR) || 128063244fa1SIlya Dryomov ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) || 128163244fa1SIlya Dryomov __pool_full(pi); 128263244fa1SIlya Dryomov 128363244fa1SIlya Dryomov WARN_ON(pi->id != t->base_oloc.pool); 128463244fa1SIlya Dryomov return (t->flags & CEPH_OSD_FLAG_READ && pauserd) || 128563244fa1SIlya Dryomov (t->flags & CEPH_OSD_FLAG_WRITE && pausewr); 128663244fa1SIlya Dryomov } 128763244fa1SIlya Dryomov 128863244fa1SIlya Dryomov enum calc_target_result { 128963244fa1SIlya Dryomov CALC_TARGET_NO_ACTION = 0, 129063244fa1SIlya Dryomov CALC_TARGET_NEED_RESEND, 129163244fa1SIlya Dryomov CALC_TARGET_POOL_DNE, 129263244fa1SIlya Dryomov }; 129363244fa1SIlya Dryomov 129463244fa1SIlya Dryomov static enum calc_target_result calc_target(struct ceph_osd_client *osdc, 129563244fa1SIlya Dryomov struct ceph_osd_request_target *t, 129663244fa1SIlya Dryomov u32 *last_force_resend, 129763244fa1SIlya Dryomov bool any_change) 129863244fa1SIlya Dryomov { 129963244fa1SIlya Dryomov struct ceph_pg_pool_info *pi; 130063244fa1SIlya Dryomov struct ceph_pg pgid, last_pgid; 130163244fa1SIlya Dryomov struct ceph_osds up, acting; 130263244fa1SIlya Dryomov bool force_resend = false; 130363244fa1SIlya Dryomov bool need_check_tiering = false; 130463244fa1SIlya Dryomov bool need_resend = false; 130563244fa1SIlya Dryomov bool sort_bitwise = ceph_osdmap_flag(osdc->osdmap, 130663244fa1SIlya Dryomov CEPH_OSDMAP_SORTBITWISE); 130763244fa1SIlya Dryomov enum calc_target_result ct_res; 130863244fa1SIlya Dryomov int ret; 130963244fa1SIlya Dryomov 131063244fa1SIlya Dryomov pi = ceph_pg_pool_by_id(osdc->osdmap, t->base_oloc.pool); 131163244fa1SIlya Dryomov if (!pi) { 131263244fa1SIlya Dryomov t->osd = CEPH_HOMELESS_OSD; 131363244fa1SIlya Dryomov ct_res = CALC_TARGET_POOL_DNE; 131463244fa1SIlya Dryomov goto out; 131563244fa1SIlya Dryomov } 131663244fa1SIlya Dryomov 131763244fa1SIlya Dryomov if (osdc->osdmap->epoch == pi->last_force_request_resend) { 131863244fa1SIlya Dryomov if (last_force_resend && 131963244fa1SIlya Dryomov *last_force_resend < pi->last_force_request_resend) { 132063244fa1SIlya Dryomov *last_force_resend = pi->last_force_request_resend; 132163244fa1SIlya Dryomov force_resend = true; 132263244fa1SIlya Dryomov } else if (!last_force_resend) { 132363244fa1SIlya Dryomov force_resend = true; 132463244fa1SIlya Dryomov } 132563244fa1SIlya Dryomov } 132663244fa1SIlya Dryomov if (ceph_oid_empty(&t->target_oid) || force_resend) { 132763244fa1SIlya Dryomov ceph_oid_copy(&t->target_oid, &t->base_oid); 132863244fa1SIlya Dryomov need_check_tiering = true; 132963244fa1SIlya Dryomov } 133063244fa1SIlya Dryomov if (ceph_oloc_empty(&t->target_oloc) || force_resend) { 133163244fa1SIlya Dryomov ceph_oloc_copy(&t->target_oloc, &t->base_oloc); 133263244fa1SIlya Dryomov need_check_tiering = true; 133363244fa1SIlya Dryomov } 133463244fa1SIlya Dryomov 133563244fa1SIlya Dryomov if (need_check_tiering && 133663244fa1SIlya Dryomov (t->flags & CEPH_OSD_FLAG_IGNORE_OVERLAY) == 0) { 133763244fa1SIlya Dryomov if (t->flags & CEPH_OSD_FLAG_READ && pi->read_tier >= 0) 133863244fa1SIlya Dryomov t->target_oloc.pool = pi->read_tier; 133963244fa1SIlya Dryomov if (t->flags & CEPH_OSD_FLAG_WRITE && pi->write_tier >= 0) 134063244fa1SIlya Dryomov t->target_oloc.pool = pi->write_tier; 134163244fa1SIlya Dryomov } 134263244fa1SIlya Dryomov 134363244fa1SIlya Dryomov ret = ceph_object_locator_to_pg(osdc->osdmap, &t->target_oid, 134463244fa1SIlya Dryomov &t->target_oloc, &pgid); 134563244fa1SIlya Dryomov if (ret) { 134663244fa1SIlya Dryomov WARN_ON(ret != -ENOENT); 134763244fa1SIlya Dryomov t->osd = CEPH_HOMELESS_OSD; 134863244fa1SIlya Dryomov ct_res = CALC_TARGET_POOL_DNE; 134963244fa1SIlya Dryomov goto out; 135063244fa1SIlya Dryomov } 135163244fa1SIlya Dryomov last_pgid.pool = pgid.pool; 135263244fa1SIlya Dryomov last_pgid.seed = ceph_stable_mod(pgid.seed, t->pg_num, t->pg_num_mask); 135363244fa1SIlya Dryomov 135463244fa1SIlya Dryomov ceph_pg_to_up_acting_osds(osdc->osdmap, &pgid, &up, &acting); 135563244fa1SIlya Dryomov if (any_change && 135663244fa1SIlya Dryomov ceph_is_new_interval(&t->acting, 135763244fa1SIlya Dryomov &acting, 135863244fa1SIlya Dryomov &t->up, 135963244fa1SIlya Dryomov &up, 136063244fa1SIlya Dryomov t->size, 136163244fa1SIlya Dryomov pi->size, 136263244fa1SIlya Dryomov t->min_size, 136363244fa1SIlya Dryomov pi->min_size, 136463244fa1SIlya Dryomov t->pg_num, 136563244fa1SIlya Dryomov pi->pg_num, 136663244fa1SIlya Dryomov t->sort_bitwise, 136763244fa1SIlya Dryomov sort_bitwise, 136863244fa1SIlya Dryomov &last_pgid)) 136963244fa1SIlya Dryomov force_resend = true; 137063244fa1SIlya Dryomov 137163244fa1SIlya Dryomov if (t->paused && !target_should_be_paused(osdc, t, pi)) { 137263244fa1SIlya Dryomov t->paused = false; 137363244fa1SIlya Dryomov need_resend = true; 137463244fa1SIlya Dryomov } 137563244fa1SIlya Dryomov 137663244fa1SIlya Dryomov if (ceph_pg_compare(&t->pgid, &pgid) || 137763244fa1SIlya Dryomov ceph_osds_changed(&t->acting, &acting, any_change) || 137863244fa1SIlya Dryomov force_resend) { 137963244fa1SIlya Dryomov t->pgid = pgid; /* struct */ 138063244fa1SIlya Dryomov ceph_osds_copy(&t->acting, &acting); 138163244fa1SIlya Dryomov ceph_osds_copy(&t->up, &up); 138263244fa1SIlya Dryomov t->size = pi->size; 138363244fa1SIlya Dryomov t->min_size = pi->min_size; 138463244fa1SIlya Dryomov t->pg_num = pi->pg_num; 138563244fa1SIlya Dryomov t->pg_num_mask = pi->pg_num_mask; 138663244fa1SIlya Dryomov t->sort_bitwise = sort_bitwise; 138763244fa1SIlya Dryomov 138863244fa1SIlya Dryomov t->osd = acting.primary; 138963244fa1SIlya Dryomov need_resend = true; 139063244fa1SIlya Dryomov } 139163244fa1SIlya Dryomov 139263244fa1SIlya Dryomov ct_res = need_resend ? CALC_TARGET_NEED_RESEND : CALC_TARGET_NO_ACTION; 139363244fa1SIlya Dryomov out: 139463244fa1SIlya Dryomov dout("%s t %p -> ct_res %d osd %d\n", __func__, t, ct_res, t->osd); 139563244fa1SIlya Dryomov return ct_res; 139663244fa1SIlya Dryomov } 139763244fa1SIlya Dryomov 1398f671b581SIlya Dryomov static void __enqueue_request(struct ceph_osd_request *req) 1399f671b581SIlya Dryomov { 1400f671b581SIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 1401f671b581SIlya Dryomov 1402f671b581SIlya Dryomov dout("%s %p tid %llu to osd%d\n", __func__, req, req->r_tid, 1403f671b581SIlya Dryomov req->r_osd ? req->r_osd->o_osd : -1); 1404f671b581SIlya Dryomov 1405f671b581SIlya Dryomov if (req->r_osd) { 1406f671b581SIlya Dryomov __remove_osd_from_lru(req->r_osd); 1407f671b581SIlya Dryomov list_add_tail(&req->r_osd_item, &req->r_osd->o_requests); 1408f671b581SIlya Dryomov list_move_tail(&req->r_req_lru_item, &osdc->req_unsent); 1409f671b581SIlya Dryomov } else { 1410f671b581SIlya Dryomov list_move_tail(&req->r_req_lru_item, &osdc->req_notarget); 1411f671b581SIlya Dryomov } 1412f671b581SIlya Dryomov } 1413f671b581SIlya Dryomov 141417a13e40SIlya Dryomov /* 14153d14c5d2SYehuda Sadeh * Pick an osd (the first 'up' osd in the pg), allocate the osd struct 14163d14c5d2SYehuda Sadeh * (as needed), and set the request r_osd appropriately. If there is 141725985edcSLucas De Marchi * no up osd, set r_osd to NULL. Move the request to the appropriate list 14186f6c7006SSage Weil * (unsent, homeless) or leave on in-flight lru. 14193d14c5d2SYehuda Sadeh * 14203d14c5d2SYehuda Sadeh * Return 0 if unchanged, 1 if changed, or negative on error. 14213d14c5d2SYehuda Sadeh * 14223d14c5d2SYehuda Sadeh * Caller should hold map_sem for read and request_mutex. 14233d14c5d2SYehuda Sadeh */ 14246f6c7006SSage Weil static int __map_request(struct ceph_osd_client *osdc, 142538d6453cSSage Weil struct ceph_osd_request *req, int force_resend) 14263d14c5d2SYehuda Sadeh { 1427a66dd383SIlya Dryomov enum calc_target_result ct_res; 14283d14c5d2SYehuda Sadeh int err; 14293d14c5d2SYehuda Sadeh 14306f6c7006SSage Weil dout("map_request %p tid %lld\n", req, req->r_tid); 143117a13e40SIlya Dryomov 1432a66dd383SIlya Dryomov ct_res = calc_target(osdc, &req->r_t, NULL, force_resend); 1433a66dd383SIlya Dryomov switch (ct_res) { 1434a66dd383SIlya Dryomov case CALC_TARGET_POOL_DNE: 14356f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 1436a66dd383SIlya Dryomov return -EIO; 1437a66dd383SIlya Dryomov case CALC_TARGET_NO_ACTION: 14383d14c5d2SYehuda Sadeh return 0; /* no change */ 1439a66dd383SIlya Dryomov default: 1440a66dd383SIlya Dryomov BUG_ON(ct_res != CALC_TARGET_NEED_RESEND); 1441a66dd383SIlya Dryomov } 14423d14c5d2SYehuda Sadeh 14435b191d99SSage Weil dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n", 1444a66dd383SIlya Dryomov req->r_tid, req->r_t.pgid.pool, req->r_t.pgid.seed, req->r_t.osd, 14453d14c5d2SYehuda Sadeh req->r_osd ? req->r_osd->o_osd : -1); 14463d14c5d2SYehuda Sadeh 14473d14c5d2SYehuda Sadeh if (req->r_osd) { 14483d14c5d2SYehuda Sadeh __cancel_request(req); 14493d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 1450a390de02SIlya Dryomov list_del_init(&req->r_linger_osd_item); 14513d14c5d2SYehuda Sadeh req->r_osd = NULL; 14523d14c5d2SYehuda Sadeh } 14533d14c5d2SYehuda Sadeh 1454a66dd383SIlya Dryomov req->r_osd = lookup_osd(&osdc->osds, req->r_t.osd); 1455a66dd383SIlya Dryomov if (!req->r_osd && req->r_t.osd >= 0) { 14563d14c5d2SYehuda Sadeh err = -ENOMEM; 1457a66dd383SIlya Dryomov req->r_osd = create_osd(osdc, req->r_t.osd); 14586f6c7006SSage Weil if (!req->r_osd) { 14596f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 14603d14c5d2SYehuda Sadeh goto out; 14616f6c7006SSage Weil } 14623d14c5d2SYehuda Sadeh 14636f3bfd45SIlya Dryomov dout("map_request osd %p is osd%d\n", req->r_osd, 1464a66dd383SIlya Dryomov req->r_osd->o_osd); 1465fcd00b68SIlya Dryomov insert_osd(&osdc->osds, req->r_osd); 14663d14c5d2SYehuda Sadeh 1467b7a9e5ddSSage Weil ceph_con_open(&req->r_osd->o_con, 1468a66dd383SIlya Dryomov CEPH_ENTITY_TYPE_OSD, req->r_osd->o_osd, 1469a66dd383SIlya Dryomov &osdc->osdmap->osd_addr[req->r_osd->o_osd]); 14703d14c5d2SYehuda Sadeh } 14713d14c5d2SYehuda Sadeh 1472f671b581SIlya Dryomov __enqueue_request(req); 14733d14c5d2SYehuda Sadeh err = 1; /* osd or pg changed */ 14743d14c5d2SYehuda Sadeh 14753d14c5d2SYehuda Sadeh out: 14763d14c5d2SYehuda Sadeh return err; 14773d14c5d2SYehuda Sadeh } 14783d14c5d2SYehuda Sadeh 1479bb873b53SIlya Dryomov static void setup_request_data(struct ceph_osd_request *req, 1480bb873b53SIlya Dryomov struct ceph_msg *msg) 14813d14c5d2SYehuda Sadeh { 1482bb873b53SIlya Dryomov u32 data_len = 0; 1483bb873b53SIlya Dryomov int i; 14843d14c5d2SYehuda Sadeh 1485bb873b53SIlya Dryomov if (!list_empty(&msg->data)) 1486bb873b53SIlya Dryomov return; 14873d14c5d2SYehuda Sadeh 1488bb873b53SIlya Dryomov WARN_ON(msg->data_length); 1489bb873b53SIlya Dryomov for (i = 0; i < req->r_num_ops; i++) { 1490bb873b53SIlya Dryomov struct ceph_osd_req_op *op = &req->r_ops[i]; 1491bb873b53SIlya Dryomov 1492bb873b53SIlya Dryomov switch (op->op) { 1493bb873b53SIlya Dryomov /* request */ 1494bb873b53SIlya Dryomov case CEPH_OSD_OP_WRITE: 1495bb873b53SIlya Dryomov case CEPH_OSD_OP_WRITEFULL: 1496bb873b53SIlya Dryomov WARN_ON(op->indata_len != op->extent.length); 1497bb873b53SIlya Dryomov ceph_osdc_msg_data_add(msg, &op->extent.osd_data); 1498bb873b53SIlya Dryomov break; 1499bb873b53SIlya Dryomov case CEPH_OSD_OP_SETXATTR: 1500bb873b53SIlya Dryomov case CEPH_OSD_OP_CMPXATTR: 1501bb873b53SIlya Dryomov WARN_ON(op->indata_len != op->xattr.name_len + 1502bb873b53SIlya Dryomov op->xattr.value_len); 1503bb873b53SIlya Dryomov ceph_osdc_msg_data_add(msg, &op->xattr.osd_data); 1504bb873b53SIlya Dryomov break; 1505bb873b53SIlya Dryomov 1506bb873b53SIlya Dryomov /* reply */ 1507bb873b53SIlya Dryomov case CEPH_OSD_OP_STAT: 1508bb873b53SIlya Dryomov ceph_osdc_msg_data_add(req->r_reply, 1509bb873b53SIlya Dryomov &op->raw_data_in); 1510bb873b53SIlya Dryomov break; 1511bb873b53SIlya Dryomov case CEPH_OSD_OP_READ: 1512bb873b53SIlya Dryomov ceph_osdc_msg_data_add(req->r_reply, 1513bb873b53SIlya Dryomov &op->extent.osd_data); 1514bb873b53SIlya Dryomov break; 1515bb873b53SIlya Dryomov 1516bb873b53SIlya Dryomov /* both */ 1517bb873b53SIlya Dryomov case CEPH_OSD_OP_CALL: 1518bb873b53SIlya Dryomov WARN_ON(op->indata_len != op->cls.class_len + 1519bb873b53SIlya Dryomov op->cls.method_len + 1520bb873b53SIlya Dryomov op->cls.indata_len); 1521bb873b53SIlya Dryomov ceph_osdc_msg_data_add(msg, &op->cls.request_info); 1522bb873b53SIlya Dryomov /* optional, can be NONE */ 1523bb873b53SIlya Dryomov ceph_osdc_msg_data_add(msg, &op->cls.request_data); 1524bb873b53SIlya Dryomov /* optional, can be NONE */ 1525bb873b53SIlya Dryomov ceph_osdc_msg_data_add(req->r_reply, 1526bb873b53SIlya Dryomov &op->cls.response_data); 1527bb873b53SIlya Dryomov break; 1528bb873b53SIlya Dryomov } 1529bb873b53SIlya Dryomov 1530bb873b53SIlya Dryomov data_len += op->indata_len; 1531bb873b53SIlya Dryomov } 1532bb873b53SIlya Dryomov 1533bb873b53SIlya Dryomov WARN_ON(data_len != msg->data_length); 1534bb873b53SIlya Dryomov } 1535bb873b53SIlya Dryomov 1536bb873b53SIlya Dryomov static void encode_request(struct ceph_osd_request *req, struct ceph_msg *msg) 1537bb873b53SIlya Dryomov { 1538bb873b53SIlya Dryomov void *p = msg->front.iov_base; 1539bb873b53SIlya Dryomov void *const end = p + msg->front_alloc_len; 1540bb873b53SIlya Dryomov u32 data_len = 0; 1541bb873b53SIlya Dryomov int i; 1542bb873b53SIlya Dryomov 1543bb873b53SIlya Dryomov if (req->r_flags & CEPH_OSD_FLAG_WRITE) { 1544bb873b53SIlya Dryomov /* snapshots aren't writeable */ 1545bb873b53SIlya Dryomov WARN_ON(req->r_snapid != CEPH_NOSNAP); 1546bb873b53SIlya Dryomov } else { 1547bb873b53SIlya Dryomov WARN_ON(req->r_mtime.tv_sec || req->r_mtime.tv_nsec || 1548bb873b53SIlya Dryomov req->r_data_offset || req->r_snapc); 1549bb873b53SIlya Dryomov } 1550bb873b53SIlya Dryomov 1551bb873b53SIlya Dryomov setup_request_data(req, msg); 1552bb873b53SIlya Dryomov 1553bb873b53SIlya Dryomov ceph_encode_32(&p, 1); /* client_inc, always 1 */ 1554bb873b53SIlya Dryomov ceph_encode_32(&p, req->r_osdc->osdmap->epoch); 1555bb873b53SIlya Dryomov ceph_encode_32(&p, req->r_flags); 1556bb873b53SIlya Dryomov ceph_encode_timespec(p, &req->r_mtime); 1557bb873b53SIlya Dryomov p += sizeof(struct ceph_timespec); 1558bb873b53SIlya Dryomov /* aka reassert_version */ 1559bb873b53SIlya Dryomov memcpy(p, &req->r_replay_version, sizeof(req->r_replay_version)); 1560bb873b53SIlya Dryomov p += sizeof(req->r_replay_version); 1561bb873b53SIlya Dryomov 1562bb873b53SIlya Dryomov /* oloc */ 1563bb873b53SIlya Dryomov ceph_encode_8(&p, 4); 1564bb873b53SIlya Dryomov ceph_encode_8(&p, 4); 1565bb873b53SIlya Dryomov ceph_encode_32(&p, 8 + 4 + 4); 1566bb873b53SIlya Dryomov ceph_encode_64(&p, req->r_t.target_oloc.pool); 1567bb873b53SIlya Dryomov ceph_encode_32(&p, -1); /* preferred */ 1568bb873b53SIlya Dryomov ceph_encode_32(&p, 0); /* key len */ 1569bb873b53SIlya Dryomov 1570bb873b53SIlya Dryomov /* pgid */ 1571bb873b53SIlya Dryomov ceph_encode_8(&p, 1); 1572a66dd383SIlya Dryomov ceph_encode_64(&p, req->r_t.pgid.pool); 1573a66dd383SIlya Dryomov ceph_encode_32(&p, req->r_t.pgid.seed); 1574bb873b53SIlya Dryomov ceph_encode_32(&p, -1); /* preferred */ 15752169aea6SSage Weil 1576bb873b53SIlya Dryomov /* oid */ 1577bb873b53SIlya Dryomov ceph_encode_32(&p, req->r_t.target_oid.name_len); 1578bb873b53SIlya Dryomov memcpy(p, req->r_t.target_oid.name, req->r_t.target_oid.name_len); 1579bb873b53SIlya Dryomov p += req->r_t.target_oid.name_len; 1580bb873b53SIlya Dryomov 1581bb873b53SIlya Dryomov /* ops, can imply data */ 1582bb873b53SIlya Dryomov ceph_encode_16(&p, req->r_num_ops); 1583bb873b53SIlya Dryomov for (i = 0; i < req->r_num_ops; i++) { 1584bb873b53SIlya Dryomov data_len += osd_req_encode_op(p, &req->r_ops[i]); 1585bb873b53SIlya Dryomov p += sizeof(struct ceph_osd_op); 1586bb873b53SIlya Dryomov } 1587bb873b53SIlya Dryomov 1588bb873b53SIlya Dryomov ceph_encode_64(&p, req->r_snapid); /* snapid */ 1589bb873b53SIlya Dryomov if (req->r_snapc) { 1590bb873b53SIlya Dryomov ceph_encode_64(&p, req->r_snapc->seq); 1591bb873b53SIlya Dryomov ceph_encode_32(&p, req->r_snapc->num_snaps); 1592bb873b53SIlya Dryomov for (i = 0; i < req->r_snapc->num_snaps; i++) 1593bb873b53SIlya Dryomov ceph_encode_64(&p, req->r_snapc->snaps[i]); 1594bb873b53SIlya Dryomov } else { 1595bb873b53SIlya Dryomov ceph_encode_64(&p, 0); /* snap_seq */ 1596bb873b53SIlya Dryomov ceph_encode_32(&p, 0); /* snaps len */ 1597bb873b53SIlya Dryomov } 1598bb873b53SIlya Dryomov 1599bb873b53SIlya Dryomov ceph_encode_32(&p, req->r_attempts); /* retry_attempt */ 1600bb873b53SIlya Dryomov 1601bb873b53SIlya Dryomov BUG_ON(p > end); 1602bb873b53SIlya Dryomov msg->front.iov_len = p - msg->front.iov_base; 1603bb873b53SIlya Dryomov msg->hdr.version = cpu_to_le16(4); /* MOSDOp v4 */ 1604bb873b53SIlya Dryomov msg->hdr.front_len = cpu_to_le32(msg->front.iov_len); 1605bb873b53SIlya Dryomov msg->hdr.data_len = cpu_to_le32(data_len); 1606bb873b53SIlya Dryomov /* 1607bb873b53SIlya Dryomov * The header "data_off" is a hint to the receiver allowing it 1608bb873b53SIlya Dryomov * to align received data into its buffers such that there's no 1609bb873b53SIlya Dryomov * need to re-copy it before writing it to disk (direct I/O). 1610bb873b53SIlya Dryomov */ 1611bb873b53SIlya Dryomov msg->hdr.data_off = cpu_to_le16(req->r_data_offset); 1612bb873b53SIlya Dryomov 1613bb873b53SIlya Dryomov dout("%s req %p oid %*pE oid_len %d front %zu data %u\n", __func__, 1614bb873b53SIlya Dryomov req, req->r_t.target_oid.name_len, req->r_t.target_oid.name, 1615bb873b53SIlya Dryomov req->r_t.target_oid.name_len, msg->front.iov_len, data_len); 1616bb873b53SIlya Dryomov } 1617bb873b53SIlya Dryomov 1618bb873b53SIlya Dryomov /* 1619bb873b53SIlya Dryomov * @req has to be assigned a tid and registered. 1620bb873b53SIlya Dryomov */ 1621bb873b53SIlya Dryomov static void send_request(struct ceph_osd_request *req) 1622bb873b53SIlya Dryomov { 1623bb873b53SIlya Dryomov struct ceph_osd *osd = req->r_osd; 1624bb873b53SIlya Dryomov 1625bb873b53SIlya Dryomov WARN_ON(osd->o_osd != req->r_t.osd); 1626bb873b53SIlya Dryomov 1627bb873b53SIlya Dryomov req->r_flags |= CEPH_OSD_FLAG_KNOWN_REDIR; 1628bb873b53SIlya Dryomov if (req->r_attempts) 1629bb873b53SIlya Dryomov req->r_flags |= CEPH_OSD_FLAG_RETRY; 1630bb873b53SIlya Dryomov else 1631bb873b53SIlya Dryomov WARN_ON(req->r_flags & CEPH_OSD_FLAG_RETRY); 1632bb873b53SIlya Dryomov 1633bb873b53SIlya Dryomov encode_request(req, req->r_request); 1634bb873b53SIlya Dryomov 1635bb873b53SIlya Dryomov dout("%s req %p tid %llu to pg %llu.%x osd%d flags 0x%x attempt %d\n", 1636bb873b53SIlya Dryomov __func__, req, req->r_tid, req->r_t.pgid.pool, req->r_t.pgid.seed, 1637bb873b53SIlya Dryomov req->r_t.osd, req->r_flags, req->r_attempts); 1638bb873b53SIlya Dryomov 1639bb873b53SIlya Dryomov req->r_t.paused = false; 16403d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 1641bb873b53SIlya Dryomov req->r_attempts++; 16423d14c5d2SYehuda Sadeh 1643bb873b53SIlya Dryomov req->r_sent = osd->o_incarnation; 1644bb873b53SIlya Dryomov req->r_request->hdr.tid = cpu_to_le64(req->r_tid); 1645bb873b53SIlya Dryomov ceph_con_send(&osd->o_con, ceph_msg_get(req->r_request)); 16463d14c5d2SYehuda Sadeh } 16473d14c5d2SYehuda Sadeh 16483d14c5d2SYehuda Sadeh /* 16496f6c7006SSage Weil * Send any requests in the queue (req_unsent). 16506f6c7006SSage Weil */ 1651f9d25199SAlex Elder static void __send_queued(struct ceph_osd_client *osdc) 16526f6c7006SSage Weil { 16536f6c7006SSage Weil struct ceph_osd_request *req, *tmp; 16546f6c7006SSage Weil 1655f9d25199SAlex Elder dout("__send_queued\n"); 1656bb873b53SIlya Dryomov list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) { 1657bb873b53SIlya Dryomov list_move_tail(&req->r_req_lru_item, &osdc->req_lru); 1658bb873b53SIlya Dryomov send_request(req); 1659bb873b53SIlya Dryomov } 16606f6c7006SSage Weil } 16616f6c7006SSage Weil 16626f6c7006SSage Weil /* 16630bbfdfe8SIlya Dryomov * Caller should hold map_sem for read and request_mutex. 16640bbfdfe8SIlya Dryomov */ 16650bbfdfe8SIlya Dryomov static int __ceph_osdc_start_request(struct ceph_osd_client *osdc, 16660bbfdfe8SIlya Dryomov struct ceph_osd_request *req, 16670bbfdfe8SIlya Dryomov bool nofail) 16680bbfdfe8SIlya Dryomov { 16690bbfdfe8SIlya Dryomov int rc; 16700bbfdfe8SIlya Dryomov 16710bbfdfe8SIlya Dryomov __register_request(osdc, req); 16720bbfdfe8SIlya Dryomov req->r_sent = 0; 16730bbfdfe8SIlya Dryomov req->r_got_reply = 0; 16740bbfdfe8SIlya Dryomov rc = __map_request(osdc, req, 0); 16750bbfdfe8SIlya Dryomov if (rc < 0) { 16760bbfdfe8SIlya Dryomov if (nofail) { 16770bbfdfe8SIlya Dryomov dout("osdc_start_request failed map, " 16780bbfdfe8SIlya Dryomov " will retry %lld\n", req->r_tid); 16790bbfdfe8SIlya Dryomov rc = 0; 16800bbfdfe8SIlya Dryomov } else { 16810bbfdfe8SIlya Dryomov __unregister_request(osdc, req); 16820bbfdfe8SIlya Dryomov } 16830bbfdfe8SIlya Dryomov return rc; 16840bbfdfe8SIlya Dryomov } 16850bbfdfe8SIlya Dryomov 16860bbfdfe8SIlya Dryomov if (req->r_osd == NULL) { 16870bbfdfe8SIlya Dryomov dout("send_request %p no up osds in pg\n", req); 16880bbfdfe8SIlya Dryomov ceph_monc_request_next_osdmap(&osdc->client->monc); 16890bbfdfe8SIlya Dryomov } else { 16900bbfdfe8SIlya Dryomov __send_queued(osdc); 16910bbfdfe8SIlya Dryomov } 16920bbfdfe8SIlya Dryomov 16930bbfdfe8SIlya Dryomov return 0; 16940bbfdfe8SIlya Dryomov } 16950bbfdfe8SIlya Dryomov 1696fe5da05eSIlya Dryomov static void __complete_request(struct ceph_osd_request *req) 1697fe5da05eSIlya Dryomov { 1698fe5da05eSIlya Dryomov if (req->r_callback) 1699fe5da05eSIlya Dryomov req->r_callback(req); 1700fe5da05eSIlya Dryomov else 1701fe5da05eSIlya Dryomov complete_all(&req->r_completion); 1702fe5da05eSIlya Dryomov } 1703fe5da05eSIlya Dryomov 17040bbfdfe8SIlya Dryomov /* 17053d14c5d2SYehuda Sadeh * Timeout callback, called every N seconds when 1 or more osd 17063d14c5d2SYehuda Sadeh * requests has been active for more than N seconds. When this 17073d14c5d2SYehuda Sadeh * happens, we ping all OSDs with requests who have timed out to 17083d14c5d2SYehuda Sadeh * ensure any communications channel reset is detected. Reset the 17093d14c5d2SYehuda Sadeh * request timeouts another N seconds in the future as we go. 17103d14c5d2SYehuda Sadeh * Reschedule the timeout event another N seconds in future (unless 17113d14c5d2SYehuda Sadeh * there are no open requests). 17123d14c5d2SYehuda Sadeh */ 17133d14c5d2SYehuda Sadeh static void handle_timeout(struct work_struct *work) 17143d14c5d2SYehuda Sadeh { 17153d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 17163d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, timeout_work.work); 1717a319bf56SIlya Dryomov struct ceph_options *opts = osdc->client->options; 171883aff95eSSage Weil struct ceph_osd_request *req; 17193d14c5d2SYehuda Sadeh struct ceph_osd *osd; 17203d14c5d2SYehuda Sadeh struct list_head slow_osds; 17213d14c5d2SYehuda Sadeh dout("timeout\n"); 17223d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 17233d14c5d2SYehuda Sadeh 17243d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 17253d14c5d2SYehuda Sadeh 17263d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 17273d14c5d2SYehuda Sadeh 17283d14c5d2SYehuda Sadeh /* 17293d14c5d2SYehuda Sadeh * ping osds that are a bit slow. this ensures that if there 17303d14c5d2SYehuda Sadeh * is a break in the TCP connection we will notice, and reopen 17313d14c5d2SYehuda Sadeh * a connection with that osd (from the fault callback). 17323d14c5d2SYehuda Sadeh */ 17333d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&slow_osds); 17343d14c5d2SYehuda Sadeh list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) { 1735a319bf56SIlya Dryomov if (time_before(jiffies, 1736a319bf56SIlya Dryomov req->r_stamp + opts->osd_keepalive_timeout)) 17373d14c5d2SYehuda Sadeh break; 17383d14c5d2SYehuda Sadeh 17393d14c5d2SYehuda Sadeh osd = req->r_osd; 17403d14c5d2SYehuda Sadeh BUG_ON(!osd); 17413d14c5d2SYehuda Sadeh dout(" tid %llu is slow, will send keepalive on osd%d\n", 17423d14c5d2SYehuda Sadeh req->r_tid, osd->o_osd); 17433d14c5d2SYehuda Sadeh list_move_tail(&osd->o_keepalive_item, &slow_osds); 17443d14c5d2SYehuda Sadeh } 17453d14c5d2SYehuda Sadeh while (!list_empty(&slow_osds)) { 17463d14c5d2SYehuda Sadeh osd = list_entry(slow_osds.next, struct ceph_osd, 17473d14c5d2SYehuda Sadeh o_keepalive_item); 17483d14c5d2SYehuda Sadeh list_del_init(&osd->o_keepalive_item); 17493d14c5d2SYehuda Sadeh ceph_con_keepalive(&osd->o_con); 17503d14c5d2SYehuda Sadeh } 17513d14c5d2SYehuda Sadeh 17523d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 1753f9d25199SAlex Elder __send_queued(osdc); 17543d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 17553d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 17563d14c5d2SYehuda Sadeh } 17573d14c5d2SYehuda Sadeh 17583d14c5d2SYehuda Sadeh static void handle_osds_timeout(struct work_struct *work) 17593d14c5d2SYehuda Sadeh { 17603d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 17613d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, 17623d14c5d2SYehuda Sadeh osds_timeout_work.work); 1763a319bf56SIlya Dryomov unsigned long delay = osdc->client->options->osd_idle_ttl / 4; 176442a2c09fSIlya Dryomov struct ceph_osd *osd, *nosd; 17653d14c5d2SYehuda Sadeh 176642a2c09fSIlya Dryomov dout("%s osdc %p\n", __func__, osdc); 17673d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 176842a2c09fSIlya Dryomov mutex_lock(&osdc->request_mutex); 17693d14c5d2SYehuda Sadeh 177042a2c09fSIlya Dryomov list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) { 177142a2c09fSIlya Dryomov if (time_before(jiffies, osd->lru_ttl)) 177242a2c09fSIlya Dryomov break; 177342a2c09fSIlya Dryomov 177442a2c09fSIlya Dryomov remove_osd(osdc, osd); 177542a2c09fSIlya Dryomov } 177642a2c09fSIlya Dryomov 177742a2c09fSIlya Dryomov mutex_unlock(&osdc->request_mutex); 177842a2c09fSIlya Dryomov up_read(&osdc->map_sem); 17793d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 17803d14c5d2SYehuda Sadeh round_jiffies_relative(delay)); 17813d14c5d2SYehuda Sadeh } 17823d14c5d2SYehuda Sadeh 1783205ee118SIlya Dryomov static int ceph_oloc_decode(void **p, void *end, 1784205ee118SIlya Dryomov struct ceph_object_locator *oloc) 1785205ee118SIlya Dryomov { 1786205ee118SIlya Dryomov u8 struct_v, struct_cv; 1787205ee118SIlya Dryomov u32 len; 1788205ee118SIlya Dryomov void *struct_end; 1789205ee118SIlya Dryomov int ret = 0; 1790205ee118SIlya Dryomov 1791205ee118SIlya Dryomov ceph_decode_need(p, end, 1 + 1 + 4, e_inval); 1792205ee118SIlya Dryomov struct_v = ceph_decode_8(p); 1793205ee118SIlya Dryomov struct_cv = ceph_decode_8(p); 1794205ee118SIlya Dryomov if (struct_v < 3) { 1795205ee118SIlya Dryomov pr_warn("got v %d < 3 cv %d of ceph_object_locator\n", 1796205ee118SIlya Dryomov struct_v, struct_cv); 1797205ee118SIlya Dryomov goto e_inval; 1798205ee118SIlya Dryomov } 1799205ee118SIlya Dryomov if (struct_cv > 6) { 1800205ee118SIlya Dryomov pr_warn("got v %d cv %d > 6 of ceph_object_locator\n", 1801205ee118SIlya Dryomov struct_v, struct_cv); 1802205ee118SIlya Dryomov goto e_inval; 1803205ee118SIlya Dryomov } 1804205ee118SIlya Dryomov len = ceph_decode_32(p); 1805205ee118SIlya Dryomov ceph_decode_need(p, end, len, e_inval); 1806205ee118SIlya Dryomov struct_end = *p + len; 1807205ee118SIlya Dryomov 1808205ee118SIlya Dryomov oloc->pool = ceph_decode_64(p); 1809205ee118SIlya Dryomov *p += 4; /* skip preferred */ 1810205ee118SIlya Dryomov 1811205ee118SIlya Dryomov len = ceph_decode_32(p); 1812205ee118SIlya Dryomov if (len > 0) { 1813205ee118SIlya Dryomov pr_warn("ceph_object_locator::key is set\n"); 1814205ee118SIlya Dryomov goto e_inval; 1815205ee118SIlya Dryomov } 1816205ee118SIlya Dryomov 1817205ee118SIlya Dryomov if (struct_v >= 5) { 1818205ee118SIlya Dryomov len = ceph_decode_32(p); 1819205ee118SIlya Dryomov if (len > 0) { 1820205ee118SIlya Dryomov pr_warn("ceph_object_locator::nspace is set\n"); 1821205ee118SIlya Dryomov goto e_inval; 1822205ee118SIlya Dryomov } 1823205ee118SIlya Dryomov } 1824205ee118SIlya Dryomov 1825205ee118SIlya Dryomov if (struct_v >= 6) { 1826205ee118SIlya Dryomov s64 hash = ceph_decode_64(p); 1827205ee118SIlya Dryomov if (hash != -1) { 1828205ee118SIlya Dryomov pr_warn("ceph_object_locator::hash is set\n"); 1829205ee118SIlya Dryomov goto e_inval; 1830205ee118SIlya Dryomov } 1831205ee118SIlya Dryomov } 1832205ee118SIlya Dryomov 1833205ee118SIlya Dryomov /* skip the rest */ 1834205ee118SIlya Dryomov *p = struct_end; 1835205ee118SIlya Dryomov out: 1836205ee118SIlya Dryomov return ret; 1837205ee118SIlya Dryomov 1838205ee118SIlya Dryomov e_inval: 1839205ee118SIlya Dryomov ret = -EINVAL; 1840205ee118SIlya Dryomov goto out; 1841205ee118SIlya Dryomov } 1842205ee118SIlya Dryomov 1843205ee118SIlya Dryomov static int ceph_redirect_decode(void **p, void *end, 1844205ee118SIlya Dryomov struct ceph_request_redirect *redir) 1845205ee118SIlya Dryomov { 1846205ee118SIlya Dryomov u8 struct_v, struct_cv; 1847205ee118SIlya Dryomov u32 len; 1848205ee118SIlya Dryomov void *struct_end; 1849205ee118SIlya Dryomov int ret; 1850205ee118SIlya Dryomov 1851205ee118SIlya Dryomov ceph_decode_need(p, end, 1 + 1 + 4, e_inval); 1852205ee118SIlya Dryomov struct_v = ceph_decode_8(p); 1853205ee118SIlya Dryomov struct_cv = ceph_decode_8(p); 1854205ee118SIlya Dryomov if (struct_cv > 1) { 1855205ee118SIlya Dryomov pr_warn("got v %d cv %d > 1 of ceph_request_redirect\n", 1856205ee118SIlya Dryomov struct_v, struct_cv); 1857205ee118SIlya Dryomov goto e_inval; 1858205ee118SIlya Dryomov } 1859205ee118SIlya Dryomov len = ceph_decode_32(p); 1860205ee118SIlya Dryomov ceph_decode_need(p, end, len, e_inval); 1861205ee118SIlya Dryomov struct_end = *p + len; 1862205ee118SIlya Dryomov 1863205ee118SIlya Dryomov ret = ceph_oloc_decode(p, end, &redir->oloc); 1864205ee118SIlya Dryomov if (ret) 1865205ee118SIlya Dryomov goto out; 1866205ee118SIlya Dryomov 1867205ee118SIlya Dryomov len = ceph_decode_32(p); 1868205ee118SIlya Dryomov if (len > 0) { 1869205ee118SIlya Dryomov pr_warn("ceph_request_redirect::object_name is set\n"); 1870205ee118SIlya Dryomov goto e_inval; 1871205ee118SIlya Dryomov } 1872205ee118SIlya Dryomov 1873205ee118SIlya Dryomov len = ceph_decode_32(p); 1874205ee118SIlya Dryomov *p += len; /* skip osd_instructions */ 1875205ee118SIlya Dryomov 1876205ee118SIlya Dryomov /* skip the rest */ 1877205ee118SIlya Dryomov *p = struct_end; 1878205ee118SIlya Dryomov out: 1879205ee118SIlya Dryomov return ret; 1880205ee118SIlya Dryomov 1881205ee118SIlya Dryomov e_inval: 1882205ee118SIlya Dryomov ret = -EINVAL; 1883205ee118SIlya Dryomov goto out; 1884205ee118SIlya Dryomov } 1885205ee118SIlya Dryomov 1886fe5da05eSIlya Dryomov struct MOSDOpReply { 1887fe5da05eSIlya Dryomov struct ceph_pg pgid; 1888fe5da05eSIlya Dryomov u64 flags; 1889fe5da05eSIlya Dryomov int result; 1890fe5da05eSIlya Dryomov u32 epoch; 1891fe5da05eSIlya Dryomov int num_ops; 1892fe5da05eSIlya Dryomov u32 outdata_len[CEPH_OSD_MAX_OPS]; 1893fe5da05eSIlya Dryomov s32 rval[CEPH_OSD_MAX_OPS]; 1894fe5da05eSIlya Dryomov int retry_attempt; 1895fe5da05eSIlya Dryomov struct ceph_eversion replay_version; 1896fe5da05eSIlya Dryomov u64 user_version; 1897fe5da05eSIlya Dryomov struct ceph_request_redirect redirect; 1898fe5da05eSIlya Dryomov }; 189925845472SSage Weil 1900fe5da05eSIlya Dryomov static int decode_MOSDOpReply(const struct ceph_msg *msg, struct MOSDOpReply *m) 19013d14c5d2SYehuda Sadeh { 1902fe5da05eSIlya Dryomov void *p = msg->front.iov_base; 1903fe5da05eSIlya Dryomov void *const end = p + msg->front.iov_len; 1904fe5da05eSIlya Dryomov u16 version = le16_to_cpu(msg->hdr.version); 1905fe5da05eSIlya Dryomov struct ceph_eversion bad_replay_version; 1906b0b31a8fSIlya Dryomov u8 decode_redir; 1907fe5da05eSIlya Dryomov u32 len; 1908fe5da05eSIlya Dryomov int ret; 1909fe5da05eSIlya Dryomov int i; 19103d14c5d2SYehuda Sadeh 1911fe5da05eSIlya Dryomov ceph_decode_32_safe(&p, end, len, e_inval); 1912fe5da05eSIlya Dryomov ceph_decode_need(&p, end, len, e_inval); 1913fe5da05eSIlya Dryomov p += len; /* skip oid */ 19141b83bef2SSage Weil 1915fe5da05eSIlya Dryomov ret = ceph_decode_pgid(&p, end, &m->pgid); 1916fe5da05eSIlya Dryomov if (ret) 1917fe5da05eSIlya Dryomov return ret; 19181b83bef2SSage Weil 1919fe5da05eSIlya Dryomov ceph_decode_64_safe(&p, end, m->flags, e_inval); 1920fe5da05eSIlya Dryomov ceph_decode_32_safe(&p, end, m->result, e_inval); 1921fe5da05eSIlya Dryomov ceph_decode_need(&p, end, sizeof(bad_replay_version), e_inval); 1922fe5da05eSIlya Dryomov memcpy(&bad_replay_version, p, sizeof(bad_replay_version)); 1923fe5da05eSIlya Dryomov p += sizeof(bad_replay_version); 1924fe5da05eSIlya Dryomov ceph_decode_32_safe(&p, end, m->epoch, e_inval); 19251b83bef2SSage Weil 1926fe5da05eSIlya Dryomov ceph_decode_32_safe(&p, end, m->num_ops, e_inval); 1927fe5da05eSIlya Dryomov if (m->num_ops > ARRAY_SIZE(m->outdata_len)) 1928fe5da05eSIlya Dryomov goto e_inval; 19291b83bef2SSage Weil 1930fe5da05eSIlya Dryomov ceph_decode_need(&p, end, m->num_ops * sizeof(struct ceph_osd_op), 1931fe5da05eSIlya Dryomov e_inval); 1932fe5da05eSIlya Dryomov for (i = 0; i < m->num_ops; i++) { 19331b83bef2SSage Weil struct ceph_osd_op *op = p; 19341b83bef2SSage Weil 1935fe5da05eSIlya Dryomov m->outdata_len[i] = le32_to_cpu(op->payload_len); 19361b83bef2SSage Weil p += sizeof(*op); 19371b83bef2SSage Weil } 1938fe5da05eSIlya Dryomov 1939fe5da05eSIlya Dryomov ceph_decode_32_safe(&p, end, m->retry_attempt, e_inval); 1940fe5da05eSIlya Dryomov for (i = 0; i < m->num_ops; i++) 1941fe5da05eSIlya Dryomov ceph_decode_32_safe(&p, end, m->rval[i], e_inval); 1942fe5da05eSIlya Dryomov 1943fe5da05eSIlya Dryomov if (version >= 5) { 1944fe5da05eSIlya Dryomov ceph_decode_need(&p, end, sizeof(m->replay_version), e_inval); 1945fe5da05eSIlya Dryomov memcpy(&m->replay_version, p, sizeof(m->replay_version)); 1946fe5da05eSIlya Dryomov p += sizeof(m->replay_version); 1947fe5da05eSIlya Dryomov ceph_decode_64_safe(&p, end, m->user_version, e_inval); 1948fe5da05eSIlya Dryomov } else { 1949fe5da05eSIlya Dryomov m->replay_version = bad_replay_version; /* struct */ 1950fe5da05eSIlya Dryomov m->user_version = le64_to_cpu(m->replay_version.version); 19511b83bef2SSage Weil } 19521b83bef2SSage Weil 1953fe5da05eSIlya Dryomov if (version >= 6) { 1954fe5da05eSIlya Dryomov if (version >= 7) 1955fe5da05eSIlya Dryomov ceph_decode_8_safe(&p, end, decode_redir, e_inval); 1956b0b31a8fSIlya Dryomov else 1957b0b31a8fSIlya Dryomov decode_redir = 1; 1958b0b31a8fSIlya Dryomov } else { 1959b0b31a8fSIlya Dryomov decode_redir = 0; 1960b0b31a8fSIlya Dryomov } 1961b0b31a8fSIlya Dryomov 1962b0b31a8fSIlya Dryomov if (decode_redir) { 1963fe5da05eSIlya Dryomov ret = ceph_redirect_decode(&p, end, &m->redirect); 1964fe5da05eSIlya Dryomov if (ret) 1965fe5da05eSIlya Dryomov return ret; 1966205ee118SIlya Dryomov } else { 1967fe5da05eSIlya Dryomov ceph_oloc_init(&m->redirect.oloc); 1968205ee118SIlya Dryomov } 1969205ee118SIlya Dryomov 1970fe5da05eSIlya Dryomov return 0; 1971205ee118SIlya Dryomov 1972fe5da05eSIlya Dryomov e_inval: 1973fe5da05eSIlya Dryomov return -EINVAL; 1974fe5da05eSIlya Dryomov } 1975fe5da05eSIlya Dryomov 1976fe5da05eSIlya Dryomov /* 1977fe5da05eSIlya Dryomov * We are done with @req if 1978fe5da05eSIlya Dryomov * - @m is a safe reply, or 1979fe5da05eSIlya Dryomov * - @m is an unsafe reply and we didn't want a safe one 1980fe5da05eSIlya Dryomov */ 1981fe5da05eSIlya Dryomov static bool done_request(const struct ceph_osd_request *req, 1982fe5da05eSIlya Dryomov const struct MOSDOpReply *m) 1983fe5da05eSIlya Dryomov { 1984fe5da05eSIlya Dryomov return (m->result < 0 || 1985fe5da05eSIlya Dryomov (m->flags & CEPH_OSD_FLAG_ONDISK) || 1986fe5da05eSIlya Dryomov !(req->r_flags & CEPH_OSD_FLAG_ONDISK)); 1987fe5da05eSIlya Dryomov } 1988fe5da05eSIlya Dryomov 1989fe5da05eSIlya Dryomov /* 1990fe5da05eSIlya Dryomov * handle osd op reply. either call the callback if it is specified, 1991fe5da05eSIlya Dryomov * or do the completion to wake up the waiting thread. 1992fe5da05eSIlya Dryomov * 1993fe5da05eSIlya Dryomov * ->r_unsafe_callback is set? yes no 1994fe5da05eSIlya Dryomov * 1995fe5da05eSIlya Dryomov * first reply is OK (needed r_cb/r_completion, r_cb/r_completion, 1996fe5da05eSIlya Dryomov * any or needed/got safe) r_safe_completion r_safe_completion 1997fe5da05eSIlya Dryomov * 1998fe5da05eSIlya Dryomov * first reply is unsafe r_unsafe_cb(true) (nothing) 1999fe5da05eSIlya Dryomov * 2000fe5da05eSIlya Dryomov * when we get the safe reply r_unsafe_cb(false), r_cb/r_completion, 2001fe5da05eSIlya Dryomov * r_safe_completion r_safe_completion 2002fe5da05eSIlya Dryomov */ 2003fe5da05eSIlya Dryomov static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg) 2004fe5da05eSIlya Dryomov { 2005fe5da05eSIlya Dryomov struct ceph_osd_request *req; 2006fe5da05eSIlya Dryomov struct MOSDOpReply m; 2007fe5da05eSIlya Dryomov u64 tid = le64_to_cpu(msg->hdr.tid); 2008fe5da05eSIlya Dryomov u32 data_len = 0; 2009fe5da05eSIlya Dryomov bool already_acked; 2010fe5da05eSIlya Dryomov int ret; 2011fe5da05eSIlya Dryomov int i; 2012fe5da05eSIlya Dryomov 2013fe5da05eSIlya Dryomov dout("%s msg %p tid %llu\n", __func__, msg, tid); 2014fe5da05eSIlya Dryomov 2015fe5da05eSIlya Dryomov down_read(&osdc->map_sem); 2016fe5da05eSIlya Dryomov mutex_lock(&osdc->request_mutex); 2017fe5da05eSIlya Dryomov req = lookup_request(&osdc->requests, tid); 2018fe5da05eSIlya Dryomov if (!req) { 2019fe5da05eSIlya Dryomov dout("%s no tid %llu\n", __func__, tid); 2020fe5da05eSIlya Dryomov goto out_unlock; 2021fe5da05eSIlya Dryomov } 2022fe5da05eSIlya Dryomov ceph_osdc_get_request(req); 2023fe5da05eSIlya Dryomov 2024fe5da05eSIlya Dryomov ret = decode_MOSDOpReply(msg, &m); 2025fe5da05eSIlya Dryomov if (ret) { 2026fe5da05eSIlya Dryomov pr_err("failed to decode MOSDOpReply for tid %llu: %d\n", 2027fe5da05eSIlya Dryomov req->r_tid, ret); 2028fe5da05eSIlya Dryomov ceph_msg_dump(msg); 2029fe5da05eSIlya Dryomov goto fail_request; 2030fe5da05eSIlya Dryomov } 2031fe5da05eSIlya Dryomov dout("%s req %p tid %llu flags 0x%llx pgid %llu.%x epoch %u attempt %d v %u'%llu uv %llu\n", 2032fe5da05eSIlya Dryomov __func__, req, req->r_tid, m.flags, m.pgid.pool, m.pgid.seed, 2033fe5da05eSIlya Dryomov m.epoch, m.retry_attempt, le32_to_cpu(m.replay_version.epoch), 2034fe5da05eSIlya Dryomov le64_to_cpu(m.replay_version.version), m.user_version); 2035fe5da05eSIlya Dryomov 2036fe5da05eSIlya Dryomov if (m.retry_attempt >= 0) { 2037fe5da05eSIlya Dryomov if (m.retry_attempt != req->r_attempts - 1) { 2038fe5da05eSIlya Dryomov dout("req %p tid %llu retry_attempt %d != %d, ignoring\n", 2039fe5da05eSIlya Dryomov req, req->r_tid, m.retry_attempt, 2040fe5da05eSIlya Dryomov req->r_attempts - 1); 2041fe5da05eSIlya Dryomov goto out_put; 2042fe5da05eSIlya Dryomov } 2043fe5da05eSIlya Dryomov } else { 2044fe5da05eSIlya Dryomov WARN_ON(1); /* MOSDOpReply v4 is assumed */ 2045fe5da05eSIlya Dryomov } 2046fe5da05eSIlya Dryomov 2047fe5da05eSIlya Dryomov if (!ceph_oloc_empty(&m.redirect.oloc)) { 2048fe5da05eSIlya Dryomov dout("req %p tid %llu redirect pool %lld\n", req, req->r_tid, 2049fe5da05eSIlya Dryomov m.redirect.oloc.pool); 2050205ee118SIlya Dryomov __unregister_request(osdc, req); 2051205ee118SIlya Dryomov 2052fe5da05eSIlya Dryomov ceph_oloc_copy(&req->r_t.target_oloc, &m.redirect.oloc); 2053205ee118SIlya Dryomov 2054205ee118SIlya Dryomov /* 2055205ee118SIlya Dryomov * Start redirect requests with nofail=true. If 2056205ee118SIlya Dryomov * mapping fails, request will end up on the notarget 2057205ee118SIlya Dryomov * list, waiting for the new osdmap (which can take 2058205ee118SIlya Dryomov * a while), even though the original request mapped 2059205ee118SIlya Dryomov * successfully. In the future we might want to follow 2060205ee118SIlya Dryomov * original request's nofail setting here. 2061205ee118SIlya Dryomov */ 2062fe5da05eSIlya Dryomov ret = __ceph_osdc_start_request(osdc, req, true); 2063fe5da05eSIlya Dryomov BUG_ON(ret); 2064205ee118SIlya Dryomov 2065fe5da05eSIlya Dryomov goto out_put; 2066205ee118SIlya Dryomov } 2067205ee118SIlya Dryomov 2068fe5da05eSIlya Dryomov if (m.num_ops != req->r_num_ops) { 2069fe5da05eSIlya Dryomov pr_err("num_ops %d != %d for tid %llu\n", m.num_ops, 2070fe5da05eSIlya Dryomov req->r_num_ops, req->r_tid); 2071fe5da05eSIlya Dryomov goto fail_request; 2072fe5da05eSIlya Dryomov } 2073fe5da05eSIlya Dryomov for (i = 0; i < req->r_num_ops; i++) { 2074fe5da05eSIlya Dryomov dout(" req %p tid %llu op %d rval %d len %u\n", req, 2075fe5da05eSIlya Dryomov req->r_tid, i, m.rval[i], m.outdata_len[i]); 2076fe5da05eSIlya Dryomov req->r_ops[i].rval = m.rval[i]; 2077fe5da05eSIlya Dryomov req->r_ops[i].outdata_len = m.outdata_len[i]; 2078fe5da05eSIlya Dryomov data_len += m.outdata_len[i]; 2079fe5da05eSIlya Dryomov } 2080fe5da05eSIlya Dryomov if (data_len != le32_to_cpu(msg->hdr.data_len)) { 2081fe5da05eSIlya Dryomov pr_err("sum of lens %u != %u for tid %llu\n", data_len, 2082fe5da05eSIlya Dryomov le32_to_cpu(msg->hdr.data_len), req->r_tid); 2083fe5da05eSIlya Dryomov goto fail_request; 2084fe5da05eSIlya Dryomov } 2085fe5da05eSIlya Dryomov dout("%s req %p tid %llu acked %d result %d data_len %u\n", __func__, 2086fe5da05eSIlya Dryomov req, req->r_tid, req->r_got_reply, m.result, data_len); 20873d14c5d2SYehuda Sadeh 2088fe5da05eSIlya Dryomov already_acked = req->r_got_reply; 2089fe5da05eSIlya Dryomov if (!already_acked) { 2090fe5da05eSIlya Dryomov req->r_result = m.result ?: data_len; 2091fe5da05eSIlya Dryomov req->r_replay_version = m.replay_version; /* struct */ 2092fe5da05eSIlya Dryomov req->r_got_reply = true; 2093fe5da05eSIlya Dryomov } else if (!(m.flags & CEPH_OSD_FLAG_ONDISK)) { 2094fe5da05eSIlya Dryomov dout("req %p tid %llu dup ack\n", req, req->r_tid); 2095fe5da05eSIlya Dryomov goto out_put; 20963d14c5d2SYehuda Sadeh } 20973d14c5d2SYehuda Sadeh 2098fe5da05eSIlya Dryomov if (done_request(req, &m)) { 20993d14c5d2SYehuda Sadeh __unregister_request(osdc, req); 2100fe5da05eSIlya Dryomov if (req->r_linger) { 2101fe5da05eSIlya Dryomov WARN_ON(req->r_unsafe_callback); 2102fe5da05eSIlya Dryomov __register_linger_request(osdc, req); 2103fe5da05eSIlya Dryomov } 2104fe5da05eSIlya Dryomov } 21053d14c5d2SYehuda Sadeh 21063d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 2107ff513aceSIlya Dryomov up_read(&osdc->map_sem); 21083d14c5d2SYehuda Sadeh 2109fe5da05eSIlya Dryomov if (done_request(req, &m)) { 2110fe5da05eSIlya Dryomov if (already_acked && req->r_unsafe_callback) { 2111fe5da05eSIlya Dryomov dout("req %p tid %llu safe-cb\n", req, req->r_tid); 211261c5d6bfSYan, Zheng req->r_unsafe_callback(req, false); 2113fe5da05eSIlya Dryomov } else { 2114fe5da05eSIlya Dryomov dout("req %p tid %llu cb\n", req, req->r_tid); 2115fe5da05eSIlya Dryomov __complete_request(req); 211661c5d6bfSYan, Zheng } 2117fe5da05eSIlya Dryomov } else { 2118fe5da05eSIlya Dryomov if (req->r_unsafe_callback) { 2119fe5da05eSIlya Dryomov dout("req %p tid %llu unsafe-cb\n", req, req->r_tid); 2120fe5da05eSIlya Dryomov req->r_unsafe_callback(req, true); 2121fe5da05eSIlya Dryomov } else { 2122fe5da05eSIlya Dryomov WARN_ON(1); 2123fe5da05eSIlya Dryomov } 2124fe5da05eSIlya Dryomov } 2125fe5da05eSIlya Dryomov if (m.flags & CEPH_OSD_FLAG_ONDISK) 2126fe5da05eSIlya Dryomov complete_all(&req->r_safe_completion); 21273d14c5d2SYehuda Sadeh 21283d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 21293d14c5d2SYehuda Sadeh return; 2130fe5da05eSIlya Dryomov 2131fe5da05eSIlya Dryomov fail_request: 2132fe5da05eSIlya Dryomov req->r_result = -EIO; 2133fe5da05eSIlya Dryomov __unregister_request(osdc, req); 2134fe5da05eSIlya Dryomov __complete_request(req); 2135fe5da05eSIlya Dryomov complete_all(&req->r_safe_completion); 2136fe5da05eSIlya Dryomov out_put: 2137fe5da05eSIlya Dryomov ceph_osdc_put_request(req); 2138ff513aceSIlya Dryomov out_unlock: 2139ff513aceSIlya Dryomov mutex_unlock(&osdc->request_mutex); 2140ff513aceSIlya Dryomov up_read(&osdc->map_sem); 21413d14c5d2SYehuda Sadeh } 21423d14c5d2SYehuda Sadeh 21436f6c7006SSage Weil static void reset_changed_osds(struct ceph_osd_client *osdc) 21443d14c5d2SYehuda Sadeh { 21453d14c5d2SYehuda Sadeh struct rb_node *p, *n; 21463d14c5d2SYehuda Sadeh 21477eb71e03SIlya Dryomov dout("%s %p\n", __func__, osdc); 21483d14c5d2SYehuda Sadeh for (p = rb_first(&osdc->osds); p; p = n) { 21496f6c7006SSage Weil struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node); 21503d14c5d2SYehuda Sadeh 21513d14c5d2SYehuda Sadeh n = rb_next(p); 21523d14c5d2SYehuda Sadeh if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) || 21533d14c5d2SYehuda Sadeh memcmp(&osd->o_con.peer_addr, 21543d14c5d2SYehuda Sadeh ceph_osd_addr(osdc->osdmap, 21553d14c5d2SYehuda Sadeh osd->o_osd), 21563d14c5d2SYehuda Sadeh sizeof(struct ceph_entity_addr)) != 0) 21573d14c5d2SYehuda Sadeh __reset_osd(osdc, osd); 21583d14c5d2SYehuda Sadeh } 21593d14c5d2SYehuda Sadeh } 21603d14c5d2SYehuda Sadeh 21613d14c5d2SYehuda Sadeh /* 21626f6c7006SSage Weil * Requeue requests whose mapping to an OSD has changed. If requests map to 21636f6c7006SSage Weil * no osd, request a new map. 21643d14c5d2SYehuda Sadeh * 2165e6d50f67SAlex Elder * Caller should hold map_sem for read. 21663d14c5d2SYehuda Sadeh */ 21679a1ea2dbSJosh Durgin static void kick_requests(struct ceph_osd_client *osdc, bool force_resend, 21689a1ea2dbSJosh Durgin bool force_resend_writes) 21693d14c5d2SYehuda Sadeh { 2170a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 21716f6c7006SSage Weil struct rb_node *p; 21726f6c7006SSage Weil int needmap = 0; 21736f6c7006SSage Weil int err; 21749a1ea2dbSJosh Durgin bool force_resend_req; 21753d14c5d2SYehuda Sadeh 21769a1ea2dbSJosh Durgin dout("kick_requests %s %s\n", force_resend ? " (force resend)" : "", 21779a1ea2dbSJosh Durgin force_resend_writes ? " (force resend writes)" : ""); 21783d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 21796194ea89SSage Weil for (p = rb_first(&osdc->requests); p; ) { 21806f6c7006SSage Weil req = rb_entry(p, struct ceph_osd_request, r_node); 21816194ea89SSage Weil p = rb_next(p); 2182ab60b16dSAlex Elder 2183ab60b16dSAlex Elder /* 2184ab60b16dSAlex Elder * For linger requests that have not yet been 2185ab60b16dSAlex Elder * registered, move them to the linger list; they'll 2186ab60b16dSAlex Elder * be sent to the osd in the loop below. Unregister 2187ab60b16dSAlex Elder * the request before re-registering it as a linger 2188ab60b16dSAlex Elder * request to ensure the __map_request() below 2189ab60b16dSAlex Elder * will decide it needs to be sent. 2190ab60b16dSAlex Elder */ 2191ab60b16dSAlex Elder if (req->r_linger && list_empty(&req->r_linger_item)) { 2192ab60b16dSAlex Elder dout("%p tid %llu restart on osd%d\n", 2193ab60b16dSAlex Elder req, req->r_tid, 2194ab60b16dSAlex Elder req->r_osd ? req->r_osd->o_osd : -1); 219596e4dac6SAlex Elder ceph_osdc_get_request(req); 2196ab60b16dSAlex Elder __unregister_request(osdc, req); 2197ab60b16dSAlex Elder __register_linger_request(osdc, req); 219896e4dac6SAlex Elder ceph_osdc_put_request(req); 2199ab60b16dSAlex Elder continue; 2200ab60b16dSAlex Elder } 2201ab60b16dSAlex Elder 22029a1ea2dbSJosh Durgin force_resend_req = force_resend || 22039a1ea2dbSJosh Durgin (force_resend_writes && 22049a1ea2dbSJosh Durgin req->r_flags & CEPH_OSD_FLAG_WRITE); 22059a1ea2dbSJosh Durgin err = __map_request(osdc, req, force_resend_req); 22066f6c7006SSage Weil if (err < 0) 22076f6c7006SSage Weil continue; /* error */ 22086f6c7006SSage Weil if (req->r_osd == NULL) { 22096f6c7006SSage Weil dout("%p tid %llu maps to no osd\n", req, req->r_tid); 22106f6c7006SSage Weil needmap++; /* request a newer map */ 22116f6c7006SSage Weil } else if (err > 0) { 22126194ea89SSage Weil if (!req->r_linger) { 22136194ea89SSage Weil dout("%p tid %llu requeued on osd%d\n", req, 22146194ea89SSage Weil req->r_tid, 22156f6c7006SSage Weil req->r_osd ? req->r_osd->o_osd : -1); 22166f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 22176f6c7006SSage Weil } 22186f6c7006SSage Weil } 22196194ea89SSage Weil } 2220a40c4f10SYehuda Sadeh 2221a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osdc->req_linger, 2222a40c4f10SYehuda Sadeh r_linger_item) { 2223a40c4f10SYehuda Sadeh dout("linger req=%p req->r_osd=%p\n", req, req->r_osd); 2224a40c4f10SYehuda Sadeh 22259a1ea2dbSJosh Durgin err = __map_request(osdc, req, 22269a1ea2dbSJosh Durgin force_resend || force_resend_writes); 2227ab60b16dSAlex Elder dout("__map_request returned %d\n", err); 2228a40c4f10SYehuda Sadeh if (err < 0) 2229a40c4f10SYehuda Sadeh continue; /* hrm! */ 2230b0494532SIlya Dryomov if (req->r_osd == NULL || err > 0) { 2231a40c4f10SYehuda Sadeh if (req->r_osd == NULL) { 2232b0494532SIlya Dryomov dout("lingering %p tid %llu maps to no osd\n", 2233b0494532SIlya Dryomov req, req->r_tid); 2234b0494532SIlya Dryomov /* 2235b0494532SIlya Dryomov * A homeless lingering request makes 2236b0494532SIlya Dryomov * no sense, as it's job is to keep 2237b0494532SIlya Dryomov * a particular OSD connection open. 2238b0494532SIlya Dryomov * Request a newer map and kick the 2239b0494532SIlya Dryomov * request, knowing that it won't be 2240b0494532SIlya Dryomov * resent until we actually get a map 2241b0494532SIlya Dryomov * that can tell us where to send it. 2242b0494532SIlya Dryomov */ 2243b0494532SIlya Dryomov needmap++; 2244a40c4f10SYehuda Sadeh } 2245a40c4f10SYehuda Sadeh 2246b0494532SIlya Dryomov dout("kicking lingering %p tid %llu osd%d\n", req, 2247b0494532SIlya Dryomov req->r_tid, req->r_osd ? req->r_osd->o_osd : -1); 2248a40c4f10SYehuda Sadeh __register_request(osdc, req); 2249c89ce05eSAlex Elder __unregister_linger_request(osdc, req); 2250a40c4f10SYehuda Sadeh } 2251b0494532SIlya Dryomov } 225214d2f38dSAlex Elder reset_changed_osds(osdc); 22533d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 22543d14c5d2SYehuda Sadeh 22553d14c5d2SYehuda Sadeh if (needmap) { 22563d14c5d2SYehuda Sadeh dout("%d requests for down osds, need new map\n", needmap); 22573d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 22583d14c5d2SYehuda Sadeh } 22593d14c5d2SYehuda Sadeh } 22606f6c7006SSage Weil 22616f6c7006SSage Weil 22623d14c5d2SYehuda Sadeh /* 22633d14c5d2SYehuda Sadeh * Process updated osd map. 22643d14c5d2SYehuda Sadeh * 22653d14c5d2SYehuda Sadeh * The message contains any number of incremental and full maps, normally 22663d14c5d2SYehuda Sadeh * indicating some sort of topology change in the cluster. Kick requests 22673d14c5d2SYehuda Sadeh * off to different OSDs as needed. 22683d14c5d2SYehuda Sadeh */ 22693d14c5d2SYehuda Sadeh void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg) 22703d14c5d2SYehuda Sadeh { 22713d14c5d2SYehuda Sadeh void *p, *end, *next; 22723d14c5d2SYehuda Sadeh u32 nr_maps, maplen; 22733d14c5d2SYehuda Sadeh u32 epoch; 22743d14c5d2SYehuda Sadeh struct ceph_osdmap *newmap = NULL, *oldmap; 22753d14c5d2SYehuda Sadeh int err; 22763d14c5d2SYehuda Sadeh struct ceph_fsid fsid; 22779a1ea2dbSJosh Durgin bool was_full; 22783d14c5d2SYehuda Sadeh 22793d14c5d2SYehuda Sadeh dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0); 22803d14c5d2SYehuda Sadeh p = msg->front.iov_base; 22813d14c5d2SYehuda Sadeh end = p + msg->front.iov_len; 22823d14c5d2SYehuda Sadeh 22833d14c5d2SYehuda Sadeh /* verify fsid */ 22843d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, sizeof(fsid), bad); 22853d14c5d2SYehuda Sadeh ceph_decode_copy(&p, &fsid, sizeof(fsid)); 22863d14c5d2SYehuda Sadeh if (ceph_check_fsid(osdc->client, &fsid) < 0) 22873d14c5d2SYehuda Sadeh return; 22883d14c5d2SYehuda Sadeh 22893d14c5d2SYehuda Sadeh down_write(&osdc->map_sem); 22903d14c5d2SYehuda Sadeh 22919a1ea2dbSJosh Durgin was_full = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL); 22929a1ea2dbSJosh Durgin 22933d14c5d2SYehuda Sadeh /* incremental maps */ 22943d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 22953d14c5d2SYehuda Sadeh dout(" %d inc maps\n", nr_maps); 22963d14c5d2SYehuda Sadeh while (nr_maps > 0) { 22973d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 22983d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 22993d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 23003d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 23013d14c5d2SYehuda Sadeh next = p + maplen; 23023d14c5d2SYehuda Sadeh if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) { 23033d14c5d2SYehuda Sadeh dout("applying incremental map %u len %d\n", 23043d14c5d2SYehuda Sadeh epoch, maplen); 23053d14c5d2SYehuda Sadeh newmap = osdmap_apply_incremental(&p, next, 23060c0a8de1SIlya Dryomov osdc->osdmap); 23073d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 23083d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 23093d14c5d2SYehuda Sadeh goto bad; 23103d14c5d2SYehuda Sadeh } 23113d14c5d2SYehuda Sadeh BUG_ON(!newmap); 23123d14c5d2SYehuda Sadeh if (newmap != osdc->osdmap) { 23133d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 23143d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 23153d14c5d2SYehuda Sadeh } 23169a1ea2dbSJosh Durgin was_full = was_full || 23179a1ea2dbSJosh Durgin ceph_osdmap_flag(osdc->osdmap, 23189a1ea2dbSJosh Durgin CEPH_OSDMAP_FULL); 23199a1ea2dbSJosh Durgin kick_requests(osdc, 0, was_full); 23203d14c5d2SYehuda Sadeh } else { 23213d14c5d2SYehuda Sadeh dout("ignoring incremental map %u len %d\n", 23223d14c5d2SYehuda Sadeh epoch, maplen); 23233d14c5d2SYehuda Sadeh } 23243d14c5d2SYehuda Sadeh p = next; 23253d14c5d2SYehuda Sadeh nr_maps--; 23263d14c5d2SYehuda Sadeh } 23273d14c5d2SYehuda Sadeh if (newmap) 23283d14c5d2SYehuda Sadeh goto done; 23293d14c5d2SYehuda Sadeh 23303d14c5d2SYehuda Sadeh /* full maps */ 23313d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 23323d14c5d2SYehuda Sadeh dout(" %d full maps\n", nr_maps); 23333d14c5d2SYehuda Sadeh while (nr_maps) { 23343d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 23353d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 23363d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 23373d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 23383d14c5d2SYehuda Sadeh if (nr_maps > 1) { 23393d14c5d2SYehuda Sadeh dout("skipping non-latest full map %u len %d\n", 23403d14c5d2SYehuda Sadeh epoch, maplen); 23413d14c5d2SYehuda Sadeh } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) { 23423d14c5d2SYehuda Sadeh dout("skipping full map %u len %d, " 23433d14c5d2SYehuda Sadeh "older than our %u\n", epoch, maplen, 23443d14c5d2SYehuda Sadeh osdc->osdmap->epoch); 23453d14c5d2SYehuda Sadeh } else { 234638d6453cSSage Weil int skipped_map = 0; 234738d6453cSSage Weil 23483d14c5d2SYehuda Sadeh dout("taking full map %u len %d\n", epoch, maplen); 2349a2505d63SIlya Dryomov newmap = ceph_osdmap_decode(&p, p+maplen); 23503d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 23513d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 23523d14c5d2SYehuda Sadeh goto bad; 23533d14c5d2SYehuda Sadeh } 23543d14c5d2SYehuda Sadeh BUG_ON(!newmap); 23553d14c5d2SYehuda Sadeh oldmap = osdc->osdmap; 23563d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 235738d6453cSSage Weil if (oldmap) { 235838d6453cSSage Weil if (oldmap->epoch + 1 < newmap->epoch) 235938d6453cSSage Weil skipped_map = 1; 23603d14c5d2SYehuda Sadeh ceph_osdmap_destroy(oldmap); 236138d6453cSSage Weil } 23629a1ea2dbSJosh Durgin was_full = was_full || 23639a1ea2dbSJosh Durgin ceph_osdmap_flag(osdc->osdmap, 23649a1ea2dbSJosh Durgin CEPH_OSDMAP_FULL); 23659a1ea2dbSJosh Durgin kick_requests(osdc, skipped_map, was_full); 23663d14c5d2SYehuda Sadeh } 23673d14c5d2SYehuda Sadeh p += maplen; 23683d14c5d2SYehuda Sadeh nr_maps--; 23693d14c5d2SYehuda Sadeh } 23703d14c5d2SYehuda Sadeh 2371b72e19b9SDan Carpenter if (!osdc->osdmap) 2372b72e19b9SDan Carpenter goto bad; 23733d14c5d2SYehuda Sadeh done: 23743d14c5d2SYehuda Sadeh downgrade_write(&osdc->map_sem); 237582dcabadSIlya Dryomov ceph_monc_got_map(&osdc->client->monc, CEPH_SUB_OSDMAP, 237682dcabadSIlya Dryomov osdc->osdmap->epoch); 2377cd634fb6SSage Weil 2378cd634fb6SSage Weil /* 2379cd634fb6SSage Weil * subscribe to subsequent osdmap updates if full to ensure 2380cd634fb6SSage Weil * we find out when we are no longer full and stop returning 2381cd634fb6SSage Weil * ENOSPC. 2382cd634fb6SSage Weil */ 2383d29adb34SJosh Durgin if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) || 2384d29adb34SJosh Durgin ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD) || 2385d29adb34SJosh Durgin ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR)) 2386cd634fb6SSage Weil ceph_monc_request_next_osdmap(&osdc->client->monc); 2387cd634fb6SSage Weil 2388f9d25199SAlex Elder mutex_lock(&osdc->request_mutex); 2389f9d25199SAlex Elder __send_queued(osdc); 2390f9d25199SAlex Elder mutex_unlock(&osdc->request_mutex); 23913d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 23923d14c5d2SYehuda Sadeh wake_up_all(&osdc->client->auth_wq); 23933d14c5d2SYehuda Sadeh return; 23943d14c5d2SYehuda Sadeh 23953d14c5d2SYehuda Sadeh bad: 23963d14c5d2SYehuda Sadeh pr_err("osdc handle_map corrupt msg\n"); 23973d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 23983d14c5d2SYehuda Sadeh up_write(&osdc->map_sem); 23993d14c5d2SYehuda Sadeh } 24003d14c5d2SYehuda Sadeh 24013d14c5d2SYehuda Sadeh /* 2402a40c4f10SYehuda Sadeh * watch/notify callback event infrastructure 2403a40c4f10SYehuda Sadeh * 2404a40c4f10SYehuda Sadeh * These callbacks are used both for watch and notify operations. 2405a40c4f10SYehuda Sadeh */ 2406a40c4f10SYehuda Sadeh static void __release_event(struct kref *kref) 2407a40c4f10SYehuda Sadeh { 2408a40c4f10SYehuda Sadeh struct ceph_osd_event *event = 2409a40c4f10SYehuda Sadeh container_of(kref, struct ceph_osd_event, kref); 2410a40c4f10SYehuda Sadeh 2411a40c4f10SYehuda Sadeh dout("__release_event %p\n", event); 2412a40c4f10SYehuda Sadeh kfree(event); 2413a40c4f10SYehuda Sadeh } 2414a40c4f10SYehuda Sadeh 2415a40c4f10SYehuda Sadeh static void get_event(struct ceph_osd_event *event) 2416a40c4f10SYehuda Sadeh { 2417a40c4f10SYehuda Sadeh kref_get(&event->kref); 2418a40c4f10SYehuda Sadeh } 2419a40c4f10SYehuda Sadeh 2420a40c4f10SYehuda Sadeh void ceph_osdc_put_event(struct ceph_osd_event *event) 2421a40c4f10SYehuda Sadeh { 2422a40c4f10SYehuda Sadeh kref_put(&event->kref, __release_event); 2423a40c4f10SYehuda Sadeh } 2424a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_put_event); 2425a40c4f10SYehuda Sadeh 2426a40c4f10SYehuda Sadeh static void __insert_event(struct ceph_osd_client *osdc, 2427a40c4f10SYehuda Sadeh struct ceph_osd_event *new) 2428a40c4f10SYehuda Sadeh { 2429a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 2430a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 2431a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 2432a40c4f10SYehuda Sadeh 2433a40c4f10SYehuda Sadeh while (*p) { 2434a40c4f10SYehuda Sadeh parent = *p; 2435a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 2436a40c4f10SYehuda Sadeh if (new->cookie < event->cookie) 2437a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 2438a40c4f10SYehuda Sadeh else if (new->cookie > event->cookie) 2439a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 2440a40c4f10SYehuda Sadeh else 2441a40c4f10SYehuda Sadeh BUG(); 2442a40c4f10SYehuda Sadeh } 2443a40c4f10SYehuda Sadeh 2444a40c4f10SYehuda Sadeh rb_link_node(&new->node, parent, p); 2445a40c4f10SYehuda Sadeh rb_insert_color(&new->node, &osdc->event_tree); 2446a40c4f10SYehuda Sadeh } 2447a40c4f10SYehuda Sadeh 2448a40c4f10SYehuda Sadeh static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc, 2449a40c4f10SYehuda Sadeh u64 cookie) 2450a40c4f10SYehuda Sadeh { 2451a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 2452a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 2453a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 2454a40c4f10SYehuda Sadeh 2455a40c4f10SYehuda Sadeh while (*p) { 2456a40c4f10SYehuda Sadeh parent = *p; 2457a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 2458a40c4f10SYehuda Sadeh if (cookie < event->cookie) 2459a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 2460a40c4f10SYehuda Sadeh else if (cookie > event->cookie) 2461a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 2462a40c4f10SYehuda Sadeh else 2463a40c4f10SYehuda Sadeh return event; 2464a40c4f10SYehuda Sadeh } 2465a40c4f10SYehuda Sadeh return NULL; 2466a40c4f10SYehuda Sadeh } 2467a40c4f10SYehuda Sadeh 2468a40c4f10SYehuda Sadeh static void __remove_event(struct ceph_osd_event *event) 2469a40c4f10SYehuda Sadeh { 2470a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 2471a40c4f10SYehuda Sadeh 2472a40c4f10SYehuda Sadeh if (!RB_EMPTY_NODE(&event->node)) { 2473a40c4f10SYehuda Sadeh dout("__remove_event removed %p\n", event); 2474a40c4f10SYehuda Sadeh rb_erase(&event->node, &osdc->event_tree); 2475a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 2476a40c4f10SYehuda Sadeh } else { 2477a40c4f10SYehuda Sadeh dout("__remove_event didn't remove %p\n", event); 2478a40c4f10SYehuda Sadeh } 2479a40c4f10SYehuda Sadeh } 2480a40c4f10SYehuda Sadeh 2481a40c4f10SYehuda Sadeh int ceph_osdc_create_event(struct ceph_osd_client *osdc, 2482a40c4f10SYehuda Sadeh void (*event_cb)(u64, u64, u8, void *), 24833c663bbdSAlex Elder void *data, struct ceph_osd_event **pevent) 2484a40c4f10SYehuda Sadeh { 2485a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 2486a40c4f10SYehuda Sadeh 2487a40c4f10SYehuda Sadeh event = kmalloc(sizeof(*event), GFP_NOIO); 2488a40c4f10SYehuda Sadeh if (!event) 2489a40c4f10SYehuda Sadeh return -ENOMEM; 2490a40c4f10SYehuda Sadeh 2491a40c4f10SYehuda Sadeh dout("create_event %p\n", event); 2492a40c4f10SYehuda Sadeh event->cb = event_cb; 24933c663bbdSAlex Elder event->one_shot = 0; 2494a40c4f10SYehuda Sadeh event->data = data; 2495a40c4f10SYehuda Sadeh event->osdc = osdc; 2496a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&event->osd_node); 24973ee5234dSAlex Elder RB_CLEAR_NODE(&event->node); 2498a40c4f10SYehuda Sadeh kref_init(&event->kref); /* one ref for us */ 2499a40c4f10SYehuda Sadeh kref_get(&event->kref); /* one ref for the caller */ 2500a40c4f10SYehuda Sadeh 2501a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 2502a40c4f10SYehuda Sadeh event->cookie = ++osdc->event_count; 2503a40c4f10SYehuda Sadeh __insert_event(osdc, event); 2504a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 2505a40c4f10SYehuda Sadeh 2506a40c4f10SYehuda Sadeh *pevent = event; 2507a40c4f10SYehuda Sadeh return 0; 2508a40c4f10SYehuda Sadeh } 2509a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_create_event); 2510a40c4f10SYehuda Sadeh 2511a40c4f10SYehuda Sadeh void ceph_osdc_cancel_event(struct ceph_osd_event *event) 2512a40c4f10SYehuda Sadeh { 2513a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 2514a40c4f10SYehuda Sadeh 2515a40c4f10SYehuda Sadeh dout("cancel_event %p\n", event); 2516a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 2517a40c4f10SYehuda Sadeh __remove_event(event); 2518a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 2519a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); /* caller's */ 2520a40c4f10SYehuda Sadeh } 2521a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_cancel_event); 2522a40c4f10SYehuda Sadeh 2523a40c4f10SYehuda Sadeh 2524a40c4f10SYehuda Sadeh static void do_event_work(struct work_struct *work) 2525a40c4f10SYehuda Sadeh { 2526a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work = 2527a40c4f10SYehuda Sadeh container_of(work, struct ceph_osd_event_work, work); 2528a40c4f10SYehuda Sadeh struct ceph_osd_event *event = event_work->event; 2529a40c4f10SYehuda Sadeh u64 ver = event_work->ver; 2530a40c4f10SYehuda Sadeh u64 notify_id = event_work->notify_id; 2531a40c4f10SYehuda Sadeh u8 opcode = event_work->opcode; 2532a40c4f10SYehuda Sadeh 2533a40c4f10SYehuda Sadeh dout("do_event_work completing %p\n", event); 2534a40c4f10SYehuda Sadeh event->cb(ver, notify_id, opcode, event->data); 2535a40c4f10SYehuda Sadeh dout("do_event_work completed %p\n", event); 2536a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 2537a40c4f10SYehuda Sadeh kfree(event_work); 2538a40c4f10SYehuda Sadeh } 2539a40c4f10SYehuda Sadeh 2540a40c4f10SYehuda Sadeh 2541a40c4f10SYehuda Sadeh /* 2542a40c4f10SYehuda Sadeh * Process osd watch notifications 2543a40c4f10SYehuda Sadeh */ 25443c663bbdSAlex Elder static void handle_watch_notify(struct ceph_osd_client *osdc, 25453c663bbdSAlex Elder struct ceph_msg *msg) 2546a40c4f10SYehuda Sadeh { 2547a40c4f10SYehuda Sadeh void *p, *end; 2548a40c4f10SYehuda Sadeh u8 proto_ver; 2549a40c4f10SYehuda Sadeh u64 cookie, ver, notify_id; 2550a40c4f10SYehuda Sadeh u8 opcode; 2551a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 2552a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work; 2553a40c4f10SYehuda Sadeh 2554a40c4f10SYehuda Sadeh p = msg->front.iov_base; 2555a40c4f10SYehuda Sadeh end = p + msg->front.iov_len; 2556a40c4f10SYehuda Sadeh 2557a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, proto_ver, bad); 2558a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, opcode, bad); 2559a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, cookie, bad); 2560a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, ver, bad); 2561a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, notify_id, bad); 2562a40c4f10SYehuda Sadeh 2563a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 2564a40c4f10SYehuda Sadeh event = __find_event(osdc, cookie); 2565a40c4f10SYehuda Sadeh if (event) { 25663c663bbdSAlex Elder BUG_ON(event->one_shot); 2567a40c4f10SYehuda Sadeh get_event(event); 2568a40c4f10SYehuda Sadeh } 2569a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 2570a40c4f10SYehuda Sadeh dout("handle_watch_notify cookie %lld ver %lld event %p\n", 2571a40c4f10SYehuda Sadeh cookie, ver, event); 2572a40c4f10SYehuda Sadeh if (event) { 2573a40c4f10SYehuda Sadeh event_work = kmalloc(sizeof(*event_work), GFP_NOIO); 2574a40c4f10SYehuda Sadeh if (!event_work) { 257591883cd2SIlya Dryomov pr_err("couldn't allocate event_work\n"); 257691883cd2SIlya Dryomov ceph_osdc_put_event(event); 257791883cd2SIlya Dryomov return; 2578a40c4f10SYehuda Sadeh } 25796b0ae409SMariusz Kozlowski INIT_WORK(&event_work->work, do_event_work); 2580a40c4f10SYehuda Sadeh event_work->event = event; 2581a40c4f10SYehuda Sadeh event_work->ver = ver; 2582a40c4f10SYehuda Sadeh event_work->notify_id = notify_id; 2583a40c4f10SYehuda Sadeh event_work->opcode = opcode; 258491883cd2SIlya Dryomov 258591883cd2SIlya Dryomov queue_work(osdc->notify_wq, &event_work->work); 2586a40c4f10SYehuda Sadeh } 2587a40c4f10SYehuda Sadeh 2588a40c4f10SYehuda Sadeh return; 2589a40c4f10SYehuda Sadeh 2590a40c4f10SYehuda Sadeh bad: 2591a40c4f10SYehuda Sadeh pr_err("osdc handle_watch_notify corrupt msg\n"); 2592a40c4f10SYehuda Sadeh } 2593a40c4f10SYehuda Sadeh 2594a40c4f10SYehuda Sadeh /* 25953d14c5d2SYehuda Sadeh * Register request, send initial attempt. 25963d14c5d2SYehuda Sadeh */ 25973d14c5d2SYehuda Sadeh int ceph_osdc_start_request(struct ceph_osd_client *osdc, 25983d14c5d2SYehuda Sadeh struct ceph_osd_request *req, 25993d14c5d2SYehuda Sadeh bool nofail) 26003d14c5d2SYehuda Sadeh { 26010bbfdfe8SIlya Dryomov int rc; 26023d14c5d2SYehuda Sadeh 26033d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 26043d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 26050bbfdfe8SIlya Dryomov 26060bbfdfe8SIlya Dryomov rc = __ceph_osdc_start_request(osdc, req, nofail); 26070bbfdfe8SIlya Dryomov 26083d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 26093d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 26100bbfdfe8SIlya Dryomov 26113d14c5d2SYehuda Sadeh return rc; 26123d14c5d2SYehuda Sadeh } 26133d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_start_request); 26143d14c5d2SYehuda Sadeh 26153d14c5d2SYehuda Sadeh /* 2616c9f9b93dSIlya Dryomov * Unregister a registered request. The request is not completed (i.e. 2617c9f9b93dSIlya Dryomov * no callbacks or wakeups) - higher layers are supposed to know what 2618c9f9b93dSIlya Dryomov * they are canceling. 2619c9f9b93dSIlya Dryomov */ 2620c9f9b93dSIlya Dryomov void ceph_osdc_cancel_request(struct ceph_osd_request *req) 2621c9f9b93dSIlya Dryomov { 2622c9f9b93dSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 2623c9f9b93dSIlya Dryomov 2624c9f9b93dSIlya Dryomov mutex_lock(&osdc->request_mutex); 2625c9f9b93dSIlya Dryomov if (req->r_linger) 2626c9f9b93dSIlya Dryomov __unregister_linger_request(osdc, req); 2627c9f9b93dSIlya Dryomov __unregister_request(osdc, req); 2628c9f9b93dSIlya Dryomov mutex_unlock(&osdc->request_mutex); 2629c9f9b93dSIlya Dryomov 2630c9f9b93dSIlya Dryomov dout("%s %p tid %llu canceled\n", __func__, req, req->r_tid); 2631c9f9b93dSIlya Dryomov } 2632c9f9b93dSIlya Dryomov EXPORT_SYMBOL(ceph_osdc_cancel_request); 2633c9f9b93dSIlya Dryomov 2634c9f9b93dSIlya Dryomov /* 26353d14c5d2SYehuda Sadeh * wait for a request to complete 26363d14c5d2SYehuda Sadeh */ 26373d14c5d2SYehuda Sadeh int ceph_osdc_wait_request(struct ceph_osd_client *osdc, 26383d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 26393d14c5d2SYehuda Sadeh { 26403d14c5d2SYehuda Sadeh int rc; 26413d14c5d2SYehuda Sadeh 2642c9f9b93dSIlya Dryomov dout("%s %p tid %llu\n", __func__, req, req->r_tid); 2643c9f9b93dSIlya Dryomov 26443d14c5d2SYehuda Sadeh rc = wait_for_completion_interruptible(&req->r_completion); 26453d14c5d2SYehuda Sadeh if (rc < 0) { 2646c9f9b93dSIlya Dryomov dout("%s %p tid %llu interrupted\n", __func__, req, req->r_tid); 2647c9f9b93dSIlya Dryomov ceph_osdc_cancel_request(req); 2648fe5da05eSIlya Dryomov 2649fe5da05eSIlya Dryomov /* kludge - need to to wake ceph_osdc_sync() */ 2650fe5da05eSIlya Dryomov complete_all(&req->r_safe_completion); 26513d14c5d2SYehuda Sadeh return rc; 26523d14c5d2SYehuda Sadeh } 26533d14c5d2SYehuda Sadeh 2654c9f9b93dSIlya Dryomov dout("%s %p tid %llu result %d\n", __func__, req, req->r_tid, 2655c9f9b93dSIlya Dryomov req->r_result); 26563d14c5d2SYehuda Sadeh return req->r_result; 26573d14c5d2SYehuda Sadeh } 26583d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_wait_request); 26593d14c5d2SYehuda Sadeh 26603d14c5d2SYehuda Sadeh /* 26613d14c5d2SYehuda Sadeh * sync - wait for all in-flight requests to flush. avoid starvation. 26623d14c5d2SYehuda Sadeh */ 26633d14c5d2SYehuda Sadeh void ceph_osdc_sync(struct ceph_osd_client *osdc) 26643d14c5d2SYehuda Sadeh { 26653d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 26663d14c5d2SYehuda Sadeh u64 last_tid, next_tid = 0; 26673d14c5d2SYehuda Sadeh 26683d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 26693d14c5d2SYehuda Sadeh last_tid = osdc->last_tid; 26703d14c5d2SYehuda Sadeh while (1) { 26713d14c5d2SYehuda Sadeh req = __lookup_request_ge(osdc, next_tid); 26723d14c5d2SYehuda Sadeh if (!req) 26733d14c5d2SYehuda Sadeh break; 26743d14c5d2SYehuda Sadeh if (req->r_tid > last_tid) 26753d14c5d2SYehuda Sadeh break; 26763d14c5d2SYehuda Sadeh 26773d14c5d2SYehuda Sadeh next_tid = req->r_tid + 1; 26783d14c5d2SYehuda Sadeh if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0) 26793d14c5d2SYehuda Sadeh continue; 26803d14c5d2SYehuda Sadeh 26813d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 26823d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 26833d14c5d2SYehuda Sadeh dout("sync waiting on tid %llu (last is %llu)\n", 26843d14c5d2SYehuda Sadeh req->r_tid, last_tid); 26853d14c5d2SYehuda Sadeh wait_for_completion(&req->r_safe_completion); 26863d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 26873d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 26883d14c5d2SYehuda Sadeh } 26893d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 26903d14c5d2SYehuda Sadeh dout("sync done (thru tid %llu)\n", last_tid); 26913d14c5d2SYehuda Sadeh } 26923d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_sync); 26933d14c5d2SYehuda Sadeh 26943d14c5d2SYehuda Sadeh /* 2695dd935f44SJosh Durgin * Call all pending notify callbacks - for use after a watch is 2696dd935f44SJosh Durgin * unregistered, to make sure no more callbacks for it will be invoked 2697dd935f44SJosh Durgin */ 2698f6479449Sstephen hemminger void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc) 2699dd935f44SJosh Durgin { 2700dd935f44SJosh Durgin flush_workqueue(osdc->notify_wq); 2701dd935f44SJosh Durgin } 2702dd935f44SJosh Durgin EXPORT_SYMBOL(ceph_osdc_flush_notifies); 2703dd935f44SJosh Durgin 2704dd935f44SJosh Durgin 2705dd935f44SJosh Durgin /* 27063d14c5d2SYehuda Sadeh * init, shutdown 27073d14c5d2SYehuda Sadeh */ 27083d14c5d2SYehuda Sadeh int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client) 27093d14c5d2SYehuda Sadeh { 27103d14c5d2SYehuda Sadeh int err; 27113d14c5d2SYehuda Sadeh 27123d14c5d2SYehuda Sadeh dout("init\n"); 27133d14c5d2SYehuda Sadeh osdc->client = client; 27143d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 27153d14c5d2SYehuda Sadeh init_rwsem(&osdc->map_sem); 27163d14c5d2SYehuda Sadeh mutex_init(&osdc->request_mutex); 27173d14c5d2SYehuda Sadeh osdc->last_tid = 0; 27183d14c5d2SYehuda Sadeh osdc->osds = RB_ROOT; 27193d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->osd_lru); 27203d14c5d2SYehuda Sadeh osdc->requests = RB_ROOT; 27213d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_lru); 27226f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_unsent); 27236f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_notarget); 2724a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_linger); 27253d14c5d2SYehuda Sadeh osdc->num_requests = 0; 27263d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout); 27273d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout); 2728a40c4f10SYehuda Sadeh spin_lock_init(&osdc->event_lock); 2729a40c4f10SYehuda Sadeh osdc->event_tree = RB_ROOT; 2730a40c4f10SYehuda Sadeh osdc->event_count = 0; 27313d14c5d2SYehuda Sadeh 27323d14c5d2SYehuda Sadeh err = -ENOMEM; 27339e767adbSIlya Dryomov osdc->req_mempool = mempool_create_slab_pool(10, 27349e767adbSIlya Dryomov ceph_osd_request_cache); 27353d14c5d2SYehuda Sadeh if (!osdc->req_mempool) 27363d14c5d2SYehuda Sadeh goto out; 27373d14c5d2SYehuda Sadeh 2738d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP, 2739711da55dSIlya Dryomov PAGE_SIZE, 10, true, "osd_op"); 27403d14c5d2SYehuda Sadeh if (err < 0) 27413d14c5d2SYehuda Sadeh goto out_mempool; 2742d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY, 2743711da55dSIlya Dryomov PAGE_SIZE, 10, true, "osd_op_reply"); 27443d14c5d2SYehuda Sadeh if (err < 0) 27453d14c5d2SYehuda Sadeh goto out_msgpool; 2746a40c4f10SYehuda Sadeh 2747dbcae088SDan Carpenter err = -ENOMEM; 2748a40c4f10SYehuda Sadeh osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify"); 2749dbcae088SDan Carpenter if (!osdc->notify_wq) 2750c172ec5cSIlya Dryomov goto out_msgpool_reply; 2751c172ec5cSIlya Dryomov 2752b37ee1b9SIlya Dryomov schedule_delayed_work(&osdc->osds_timeout_work, 2753b37ee1b9SIlya Dryomov round_jiffies_relative(osdc->client->options->osd_idle_ttl)); 2754b37ee1b9SIlya Dryomov 27553d14c5d2SYehuda Sadeh return 0; 27563d14c5d2SYehuda Sadeh 2757c172ec5cSIlya Dryomov out_msgpool_reply: 2758c172ec5cSIlya Dryomov ceph_msgpool_destroy(&osdc->msgpool_op_reply); 27593d14c5d2SYehuda Sadeh out_msgpool: 27603d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 27613d14c5d2SYehuda Sadeh out_mempool: 27623d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 27633d14c5d2SYehuda Sadeh out: 27643d14c5d2SYehuda Sadeh return err; 27653d14c5d2SYehuda Sadeh } 27663d14c5d2SYehuda Sadeh 27673d14c5d2SYehuda Sadeh void ceph_osdc_stop(struct ceph_osd_client *osdc) 27683d14c5d2SYehuda Sadeh { 2769a40c4f10SYehuda Sadeh flush_workqueue(osdc->notify_wq); 2770a40c4f10SYehuda Sadeh destroy_workqueue(osdc->notify_wq); 27713d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->timeout_work); 27723d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->osds_timeout_work); 277342a2c09fSIlya Dryomov 277442a2c09fSIlya Dryomov mutex_lock(&osdc->request_mutex); 277542a2c09fSIlya Dryomov while (!RB_EMPTY_ROOT(&osdc->osds)) { 277642a2c09fSIlya Dryomov struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds), 277742a2c09fSIlya Dryomov struct ceph_osd, o_node); 277842a2c09fSIlya Dryomov remove_osd(osdc, osd); 277942a2c09fSIlya Dryomov } 278042a2c09fSIlya Dryomov mutex_unlock(&osdc->request_mutex); 278142a2c09fSIlya Dryomov 27823d14c5d2SYehuda Sadeh if (osdc->osdmap) { 27833d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 27843d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 27853d14c5d2SYehuda Sadeh } 27863d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 27873d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 27883d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op_reply); 27893d14c5d2SYehuda Sadeh } 27903d14c5d2SYehuda Sadeh 27913d14c5d2SYehuda Sadeh /* 27923d14c5d2SYehuda Sadeh * Read some contiguous pages. If we cross a stripe boundary, shorten 27933d14c5d2SYehuda Sadeh * *plen. Return number of bytes read, or error. 27943d14c5d2SYehuda Sadeh */ 27953d14c5d2SYehuda Sadeh int ceph_osdc_readpages(struct ceph_osd_client *osdc, 27963d14c5d2SYehuda Sadeh struct ceph_vino vino, struct ceph_file_layout *layout, 27973d14c5d2SYehuda Sadeh u64 off, u64 *plen, 27983d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 2799b7495fc2SSage Weil struct page **pages, int num_pages, int page_align) 28003d14c5d2SYehuda Sadeh { 28013d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 28023d14c5d2SYehuda Sadeh int rc = 0; 28033d14c5d2SYehuda Sadeh 28043d14c5d2SYehuda Sadeh dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino, 28053d14c5d2SYehuda Sadeh vino.snap, off, *plen); 2806715e4cd4SYan, Zheng req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 0, 1, 28073d14c5d2SYehuda Sadeh CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ, 2808acead002SAlex Elder NULL, truncate_seq, truncate_size, 2809153e5167SAlex Elder false); 28106816282dSSage Weil if (IS_ERR(req)) 28116816282dSSage Weil return PTR_ERR(req); 28123d14c5d2SYehuda Sadeh 28133d14c5d2SYehuda Sadeh /* it may be a short read due to an object boundary */ 2814406e2c9fSAlex Elder osd_req_op_extent_osd_data_pages(req, 0, 2815a4ce40a9SAlex Elder pages, *plen, page_align, false, false); 28163d14c5d2SYehuda Sadeh 2817e0c59487SAlex Elder dout("readpages final extent is %llu~%llu (%llu bytes align %d)\n", 281843bfe5deSAlex Elder off, *plen, *plen, page_align); 28193d14c5d2SYehuda Sadeh 28203d14c5d2SYehuda Sadeh rc = ceph_osdc_start_request(osdc, req, false); 28213d14c5d2SYehuda Sadeh if (!rc) 28223d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 28233d14c5d2SYehuda Sadeh 28243d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 28253d14c5d2SYehuda Sadeh dout("readpages result %d\n", rc); 28263d14c5d2SYehuda Sadeh return rc; 28273d14c5d2SYehuda Sadeh } 28283d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_readpages); 28293d14c5d2SYehuda Sadeh 28303d14c5d2SYehuda Sadeh /* 28313d14c5d2SYehuda Sadeh * do a synchronous write on N pages 28323d14c5d2SYehuda Sadeh */ 28333d14c5d2SYehuda Sadeh int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino, 28343d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 28353d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 28363d14c5d2SYehuda Sadeh u64 off, u64 len, 28373d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 28383d14c5d2SYehuda Sadeh struct timespec *mtime, 283924808826SAlex Elder struct page **pages, int num_pages) 28403d14c5d2SYehuda Sadeh { 28413d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 28423d14c5d2SYehuda Sadeh int rc = 0; 2843b7495fc2SSage Weil int page_align = off & ~PAGE_MASK; 28443d14c5d2SYehuda Sadeh 2845715e4cd4SYan, Zheng req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 0, 1, 28463d14c5d2SYehuda Sadeh CEPH_OSD_OP_WRITE, 284724808826SAlex Elder CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE, 2848acead002SAlex Elder snapc, truncate_seq, truncate_size, 2849153e5167SAlex Elder true); 28506816282dSSage Weil if (IS_ERR(req)) 28516816282dSSage Weil return PTR_ERR(req); 28523d14c5d2SYehuda Sadeh 28533d14c5d2SYehuda Sadeh /* it may be a short write due to an object boundary */ 2854406e2c9fSAlex Elder osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align, 285543bfe5deSAlex Elder false, false); 285643bfe5deSAlex Elder dout("writepages %llu~%llu (%llu bytes)\n", off, len, len); 28573d14c5d2SYehuda Sadeh 2858bb873b53SIlya Dryomov req->r_mtime = *mtime; 285987f979d3SAlex Elder rc = ceph_osdc_start_request(osdc, req, true); 28603d14c5d2SYehuda Sadeh if (!rc) 28613d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 28623d14c5d2SYehuda Sadeh 28633d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 28643d14c5d2SYehuda Sadeh if (rc == 0) 28653d14c5d2SYehuda Sadeh rc = len; 28663d14c5d2SYehuda Sadeh dout("writepages result %d\n", rc); 28673d14c5d2SYehuda Sadeh return rc; 28683d14c5d2SYehuda Sadeh } 28693d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_writepages); 28703d14c5d2SYehuda Sadeh 28715522ae0bSAlex Elder int ceph_osdc_setup(void) 28725522ae0bSAlex Elder { 28733f1af42aSIlya Dryomov size_t size = sizeof(struct ceph_osd_request) + 28743f1af42aSIlya Dryomov CEPH_OSD_SLAB_OPS * sizeof(struct ceph_osd_req_op); 28753f1af42aSIlya Dryomov 28765522ae0bSAlex Elder BUG_ON(ceph_osd_request_cache); 28773f1af42aSIlya Dryomov ceph_osd_request_cache = kmem_cache_create("ceph_osd_request", size, 28783f1af42aSIlya Dryomov 0, 0, NULL); 28795522ae0bSAlex Elder 28805522ae0bSAlex Elder return ceph_osd_request_cache ? 0 : -ENOMEM; 28815522ae0bSAlex Elder } 28825522ae0bSAlex Elder EXPORT_SYMBOL(ceph_osdc_setup); 28835522ae0bSAlex Elder 28845522ae0bSAlex Elder void ceph_osdc_cleanup(void) 28855522ae0bSAlex Elder { 28865522ae0bSAlex Elder BUG_ON(!ceph_osd_request_cache); 28875522ae0bSAlex Elder kmem_cache_destroy(ceph_osd_request_cache); 28885522ae0bSAlex Elder ceph_osd_request_cache = NULL; 28895522ae0bSAlex Elder } 28905522ae0bSAlex Elder EXPORT_SYMBOL(ceph_osdc_cleanup); 28915522ae0bSAlex Elder 28923d14c5d2SYehuda Sadeh /* 28933d14c5d2SYehuda Sadeh * handle incoming message 28943d14c5d2SYehuda Sadeh */ 28953d14c5d2SYehuda Sadeh static void dispatch(struct ceph_connection *con, struct ceph_msg *msg) 28963d14c5d2SYehuda Sadeh { 28973d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 28983d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 28993d14c5d2SYehuda Sadeh int type = le16_to_cpu(msg->hdr.type); 29003d14c5d2SYehuda Sadeh 29013d14c5d2SYehuda Sadeh if (!osd) 29023d14c5d2SYehuda Sadeh goto out; 29033d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 29043d14c5d2SYehuda Sadeh 29053d14c5d2SYehuda Sadeh switch (type) { 29063d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 29073d14c5d2SYehuda Sadeh ceph_osdc_handle_map(osdc, msg); 29083d14c5d2SYehuda Sadeh break; 29093d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 291070cf052dSShraddha Barke handle_reply(osdc, msg); 29113d14c5d2SYehuda Sadeh break; 2912a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 2913a40c4f10SYehuda Sadeh handle_watch_notify(osdc, msg); 2914a40c4f10SYehuda Sadeh break; 29153d14c5d2SYehuda Sadeh 29163d14c5d2SYehuda Sadeh default: 29173d14c5d2SYehuda Sadeh pr_err("received unknown message type %d %s\n", type, 29183d14c5d2SYehuda Sadeh ceph_msg_type_name(type)); 29193d14c5d2SYehuda Sadeh } 29203d14c5d2SYehuda Sadeh out: 29213d14c5d2SYehuda Sadeh ceph_msg_put(msg); 29223d14c5d2SYehuda Sadeh } 29233d14c5d2SYehuda Sadeh 29243d14c5d2SYehuda Sadeh /* 2925d15f9d69SIlya Dryomov * Lookup and return message for incoming reply. Don't try to do 2926d15f9d69SIlya Dryomov * anything about a larger than preallocated data portion of the 2927d15f9d69SIlya Dryomov * message at the moment - for now, just skip the message. 29283d14c5d2SYehuda Sadeh */ 29293d14c5d2SYehuda Sadeh static struct ceph_msg *get_reply(struct ceph_connection *con, 29303d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 29313d14c5d2SYehuda Sadeh int *skip) 29323d14c5d2SYehuda Sadeh { 29333d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 29343d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = osd->o_osdc; 29353d14c5d2SYehuda Sadeh struct ceph_msg *m; 29363d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 29373f0a4ac5SIlya Dryomov int front_len = le32_to_cpu(hdr->front_len); 29383d14c5d2SYehuda Sadeh int data_len = le32_to_cpu(hdr->data_len); 29393d14c5d2SYehuda Sadeh u64 tid; 29403d14c5d2SYehuda Sadeh 29413d14c5d2SYehuda Sadeh tid = le64_to_cpu(hdr->tid); 29423d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 2943fcd00b68SIlya Dryomov req = lookup_request(&osdc->requests, tid); 29443d14c5d2SYehuda Sadeh if (!req) { 2945cd8140c6SIlya Dryomov dout("%s osd%d tid %llu unknown, skipping\n", __func__, 2946cd8140c6SIlya Dryomov osd->o_osd, tid); 29473d14c5d2SYehuda Sadeh m = NULL; 2948d15f9d69SIlya Dryomov *skip = 1; 29493d14c5d2SYehuda Sadeh goto out; 29503d14c5d2SYehuda Sadeh } 29513d14c5d2SYehuda Sadeh 29528921d114SAlex Elder ceph_msg_revoke_incoming(req->r_reply); 29533d14c5d2SYehuda Sadeh 2954f2be82b0SIlya Dryomov if (front_len > req->r_reply->front_alloc_len) { 2955d15f9d69SIlya Dryomov pr_warn("%s osd%d tid %llu front %d > preallocated %d\n", 2956d15f9d69SIlya Dryomov __func__, osd->o_osd, req->r_tid, front_len, 2957d15f9d69SIlya Dryomov req->r_reply->front_alloc_len); 29583f0a4ac5SIlya Dryomov m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front_len, GFP_NOFS, 29593f0a4ac5SIlya Dryomov false); 29603d14c5d2SYehuda Sadeh if (!m) 29613d14c5d2SYehuda Sadeh goto out; 29623d14c5d2SYehuda Sadeh ceph_msg_put(req->r_reply); 29633d14c5d2SYehuda Sadeh req->r_reply = m; 29643d14c5d2SYehuda Sadeh } 29653d14c5d2SYehuda Sadeh 2966d15f9d69SIlya Dryomov if (data_len > req->r_reply->data_length) { 2967d15f9d69SIlya Dryomov pr_warn("%s osd%d tid %llu data %d > preallocated %zu, skipping\n", 2968d15f9d69SIlya Dryomov __func__, osd->o_osd, req->r_tid, data_len, 2969d15f9d69SIlya Dryomov req->r_reply->data_length); 29703d14c5d2SYehuda Sadeh m = NULL; 2971d15f9d69SIlya Dryomov *skip = 1; 29723d14c5d2SYehuda Sadeh goto out; 29733d14c5d2SYehuda Sadeh } 2974d15f9d69SIlya Dryomov 2975d15f9d69SIlya Dryomov m = ceph_msg_get(req->r_reply); 29763d14c5d2SYehuda Sadeh dout("get_reply tid %lld %p\n", tid, m); 29773d14c5d2SYehuda Sadeh 29783d14c5d2SYehuda Sadeh out: 29793d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 29803d14c5d2SYehuda Sadeh return m; 29813d14c5d2SYehuda Sadeh } 29823d14c5d2SYehuda Sadeh 29833d14c5d2SYehuda Sadeh static struct ceph_msg *alloc_msg(struct ceph_connection *con, 29843d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 29853d14c5d2SYehuda Sadeh int *skip) 29863d14c5d2SYehuda Sadeh { 29873d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 29883d14c5d2SYehuda Sadeh int type = le16_to_cpu(hdr->type); 29893d14c5d2SYehuda Sadeh int front = le32_to_cpu(hdr->front_len); 29903d14c5d2SYehuda Sadeh 29911c20f2d2SAlex Elder *skip = 0; 29923d14c5d2SYehuda Sadeh switch (type) { 29933d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 2994a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 2995b61c2763SSage Weil return ceph_msg_new(type, front, GFP_NOFS, false); 29963d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 29973d14c5d2SYehuda Sadeh return get_reply(con, hdr, skip); 29983d14c5d2SYehuda Sadeh default: 29993d14c5d2SYehuda Sadeh pr_info("alloc_msg unexpected msg type %d from osd%d\n", type, 30003d14c5d2SYehuda Sadeh osd->o_osd); 30013d14c5d2SYehuda Sadeh *skip = 1; 30023d14c5d2SYehuda Sadeh return NULL; 30033d14c5d2SYehuda Sadeh } 30043d14c5d2SYehuda Sadeh } 30053d14c5d2SYehuda Sadeh 30063d14c5d2SYehuda Sadeh /* 30073d14c5d2SYehuda Sadeh * Wrappers to refcount containing ceph_osd struct 30083d14c5d2SYehuda Sadeh */ 30093d14c5d2SYehuda Sadeh static struct ceph_connection *get_osd_con(struct ceph_connection *con) 30103d14c5d2SYehuda Sadeh { 30113d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 30123d14c5d2SYehuda Sadeh if (get_osd(osd)) 30133d14c5d2SYehuda Sadeh return con; 30143d14c5d2SYehuda Sadeh return NULL; 30153d14c5d2SYehuda Sadeh } 30163d14c5d2SYehuda Sadeh 30173d14c5d2SYehuda Sadeh static void put_osd_con(struct ceph_connection *con) 30183d14c5d2SYehuda Sadeh { 30193d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 30203d14c5d2SYehuda Sadeh put_osd(osd); 30213d14c5d2SYehuda Sadeh } 30223d14c5d2SYehuda Sadeh 30233d14c5d2SYehuda Sadeh /* 30243d14c5d2SYehuda Sadeh * authentication 30253d14c5d2SYehuda Sadeh */ 3026a3530df3SAlex Elder /* 3027a3530df3SAlex Elder * Note: returned pointer is the address of a structure that's 3028a3530df3SAlex Elder * managed separately. Caller must *not* attempt to free it. 3029a3530df3SAlex Elder */ 3030a3530df3SAlex Elder static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con, 30318f43fb53SAlex Elder int *proto, int force_new) 30323d14c5d2SYehuda Sadeh { 30333d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 30343d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 30353d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 303674f1869fSAlex Elder struct ceph_auth_handshake *auth = &o->o_auth; 30373d14c5d2SYehuda Sadeh 303874f1869fSAlex Elder if (force_new && auth->authorizer) { 30396c1ea260SIlya Dryomov ceph_auth_destroy_authorizer(auth->authorizer); 304074f1869fSAlex Elder auth->authorizer = NULL; 30413d14c5d2SYehuda Sadeh } 304227859f97SSage Weil if (!auth->authorizer) { 304327859f97SSage Weil int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 3044a3530df3SAlex Elder auth); 30453d14c5d2SYehuda Sadeh if (ret) 3046a3530df3SAlex Elder return ERR_PTR(ret); 304727859f97SSage Weil } else { 304827859f97SSage Weil int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 30490bed9b5cSSage Weil auth); 30500bed9b5cSSage Weil if (ret) 30510bed9b5cSSage Weil return ERR_PTR(ret); 30523d14c5d2SYehuda Sadeh } 30533d14c5d2SYehuda Sadeh *proto = ac->protocol; 305474f1869fSAlex Elder 3055a3530df3SAlex Elder return auth; 30563d14c5d2SYehuda Sadeh } 30573d14c5d2SYehuda Sadeh 30583d14c5d2SYehuda Sadeh 30593d14c5d2SYehuda Sadeh static int verify_authorizer_reply(struct ceph_connection *con, int len) 30603d14c5d2SYehuda Sadeh { 30613d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 30623d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 30633d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 30643d14c5d2SYehuda Sadeh 306527859f97SSage Weil return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len); 30663d14c5d2SYehuda Sadeh } 30673d14c5d2SYehuda Sadeh 30683d14c5d2SYehuda Sadeh static int invalidate_authorizer(struct ceph_connection *con) 30693d14c5d2SYehuda Sadeh { 30703d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 30713d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 30723d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 30733d14c5d2SYehuda Sadeh 307427859f97SSage Weil ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD); 30753d14c5d2SYehuda Sadeh return ceph_monc_validate_auth(&osdc->client->monc); 30763d14c5d2SYehuda Sadeh } 30773d14c5d2SYehuda Sadeh 307879dbd1baSIlya Dryomov static int osd_sign_message(struct ceph_msg *msg) 307933d07337SYan, Zheng { 308079dbd1baSIlya Dryomov struct ceph_osd *o = msg->con->private; 308133d07337SYan, Zheng struct ceph_auth_handshake *auth = &o->o_auth; 308279dbd1baSIlya Dryomov 308333d07337SYan, Zheng return ceph_auth_sign_message(auth, msg); 308433d07337SYan, Zheng } 308533d07337SYan, Zheng 308679dbd1baSIlya Dryomov static int osd_check_message_signature(struct ceph_msg *msg) 308733d07337SYan, Zheng { 308879dbd1baSIlya Dryomov struct ceph_osd *o = msg->con->private; 308933d07337SYan, Zheng struct ceph_auth_handshake *auth = &o->o_auth; 309079dbd1baSIlya Dryomov 309133d07337SYan, Zheng return ceph_auth_check_message_signature(auth, msg); 309233d07337SYan, Zheng } 309333d07337SYan, Zheng 30943d14c5d2SYehuda Sadeh static const struct ceph_connection_operations osd_con_ops = { 30953d14c5d2SYehuda Sadeh .get = get_osd_con, 30963d14c5d2SYehuda Sadeh .put = put_osd_con, 30973d14c5d2SYehuda Sadeh .dispatch = dispatch, 30983d14c5d2SYehuda Sadeh .get_authorizer = get_authorizer, 30993d14c5d2SYehuda Sadeh .verify_authorizer_reply = verify_authorizer_reply, 31003d14c5d2SYehuda Sadeh .invalidate_authorizer = invalidate_authorizer, 31013d14c5d2SYehuda Sadeh .alloc_msg = alloc_msg, 310279dbd1baSIlya Dryomov .sign_message = osd_sign_message, 310379dbd1baSIlya Dryomov .check_message_signature = osd_check_message_signature, 31043d14c5d2SYehuda Sadeh .fault = osd_reset, 31053d14c5d2SYehuda Sadeh }; 3106