1a4ce40a9SAlex Elder 23d14c5d2SYehuda Sadeh #include <linux/ceph/ceph_debug.h> 33d14c5d2SYehuda Sadeh 43d14c5d2SYehuda Sadeh #include <linux/module.h> 53d14c5d2SYehuda Sadeh #include <linux/err.h> 63d14c5d2SYehuda Sadeh #include <linux/highmem.h> 73d14c5d2SYehuda Sadeh #include <linux/mm.h> 83d14c5d2SYehuda Sadeh #include <linux/pagemap.h> 93d14c5d2SYehuda Sadeh #include <linux/slab.h> 103d14c5d2SYehuda Sadeh #include <linux/uaccess.h> 113d14c5d2SYehuda Sadeh #ifdef CONFIG_BLOCK 123d14c5d2SYehuda Sadeh #include <linux/bio.h> 133d14c5d2SYehuda Sadeh #endif 143d14c5d2SYehuda Sadeh 153d14c5d2SYehuda Sadeh #include <linux/ceph/libceph.h> 163d14c5d2SYehuda Sadeh #include <linux/ceph/osd_client.h> 173d14c5d2SYehuda Sadeh #include <linux/ceph/messenger.h> 183d14c5d2SYehuda Sadeh #include <linux/ceph/decode.h> 193d14c5d2SYehuda Sadeh #include <linux/ceph/auth.h> 203d14c5d2SYehuda Sadeh #include <linux/ceph/pagelist.h> 213d14c5d2SYehuda Sadeh 223d14c5d2SYehuda Sadeh #define OSD_OP_FRONT_LEN 4096 233d14c5d2SYehuda Sadeh #define OSD_OPREPLY_FRONT_LEN 512 243d14c5d2SYehuda Sadeh 255522ae0bSAlex Elder static struct kmem_cache *ceph_osd_request_cache; 265522ae0bSAlex Elder 273d14c5d2SYehuda Sadeh static const struct ceph_connection_operations osd_con_ops; 283d14c5d2SYehuda Sadeh 29f9d25199SAlex Elder static void __send_queued(struct ceph_osd_client *osdc); 306f6c7006SSage Weil static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd); 31a40c4f10SYehuda Sadeh static void __register_request(struct ceph_osd_client *osdc, 32a40c4f10SYehuda Sadeh struct ceph_osd_request *req); 332cc6128aSIlya Dryomov static void __unregister_request(struct ceph_osd_client *osdc, 342cc6128aSIlya Dryomov struct ceph_osd_request *req); 35a40c4f10SYehuda Sadeh static void __unregister_linger_request(struct ceph_osd_client *osdc, 36a40c4f10SYehuda Sadeh struct ceph_osd_request *req); 372cc6128aSIlya Dryomov static void __enqueue_request(struct ceph_osd_request *req); 3856e925b6SSage Weil static void __send_request(struct ceph_osd_client *osdc, 39a40c4f10SYehuda Sadeh struct ceph_osd_request *req); 403d14c5d2SYehuda Sadeh 413d14c5d2SYehuda Sadeh /* 423d14c5d2SYehuda Sadeh * Implement client access to distributed object storage cluster. 433d14c5d2SYehuda Sadeh * 443d14c5d2SYehuda Sadeh * All data objects are stored within a cluster/cloud of OSDs, or 453d14c5d2SYehuda Sadeh * "object storage devices." (Note that Ceph OSDs have _nothing_ to 463d14c5d2SYehuda Sadeh * do with the T10 OSD extensions to SCSI.) Ceph OSDs are simply 473d14c5d2SYehuda Sadeh * remote daemons serving up and coordinating consistent and safe 483d14c5d2SYehuda Sadeh * access to storage. 493d14c5d2SYehuda Sadeh * 503d14c5d2SYehuda Sadeh * Cluster membership and the mapping of data objects onto storage devices 513d14c5d2SYehuda Sadeh * are described by the osd map. 523d14c5d2SYehuda Sadeh * 533d14c5d2SYehuda Sadeh * We keep track of pending OSD requests (read, write), resubmit 543d14c5d2SYehuda Sadeh * requests to different OSDs when the cluster topology/data layout 553d14c5d2SYehuda Sadeh * change, or retry the affected requests when the communications 563d14c5d2SYehuda Sadeh * channel with an OSD is reset. 573d14c5d2SYehuda Sadeh */ 583d14c5d2SYehuda Sadeh 593d14c5d2SYehuda Sadeh /* 603d14c5d2SYehuda Sadeh * calculate the mapping of a file extent onto an object, and fill out the 613d14c5d2SYehuda Sadeh * request accordingly. shorten extent as necessary if it crosses an 623d14c5d2SYehuda Sadeh * object boundary. 633d14c5d2SYehuda Sadeh * 643d14c5d2SYehuda Sadeh * fill osd op in request message. 653d14c5d2SYehuda Sadeh */ 66dbe0fc41SAlex Elder static int calc_layout(struct ceph_file_layout *layout, u64 off, u64 *plen, 67a19dadfbSAlex Elder u64 *objnum, u64 *objoff, u64 *objlen) 683d14c5d2SYehuda Sadeh { 6960e56f13SAlex Elder u64 orig_len = *plen; 70d63b77f4SSage Weil int r; 713d14c5d2SYehuda Sadeh 7260e56f13SAlex Elder /* object extent? */ 7375d1c941SAlex Elder r = ceph_calc_file_object_mapping(layout, off, orig_len, objnum, 7475d1c941SAlex Elder objoff, objlen); 75d63b77f4SSage Weil if (r < 0) 76d63b77f4SSage Weil return r; 7775d1c941SAlex Elder if (*objlen < orig_len) { 7875d1c941SAlex Elder *plen = *objlen; 7960e56f13SAlex Elder dout(" skipping last %llu, final file extent %llu~%llu\n", 8060e56f13SAlex Elder orig_len - *plen, off, *plen); 8160e56f13SAlex Elder } 8260e56f13SAlex Elder 8375d1c941SAlex Elder dout("calc_layout objnum=%llx %llu~%llu\n", *objnum, *objoff, *objlen); 843d14c5d2SYehuda Sadeh 853ff5f385SAlex Elder return 0; 863d14c5d2SYehuda Sadeh } 873d14c5d2SYehuda Sadeh 88c54d47bfSAlex Elder static void ceph_osd_data_init(struct ceph_osd_data *osd_data) 89c54d47bfSAlex Elder { 90c54d47bfSAlex Elder memset(osd_data, 0, sizeof (*osd_data)); 91c54d47bfSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_NONE; 92c54d47bfSAlex Elder } 93c54d47bfSAlex Elder 94a4ce40a9SAlex Elder static void ceph_osd_data_pages_init(struct ceph_osd_data *osd_data, 9543bfe5deSAlex Elder struct page **pages, u64 length, u32 alignment, 9643bfe5deSAlex Elder bool pages_from_pool, bool own_pages) 9743bfe5deSAlex Elder { 9843bfe5deSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_PAGES; 9943bfe5deSAlex Elder osd_data->pages = pages; 10043bfe5deSAlex Elder osd_data->length = length; 10143bfe5deSAlex Elder osd_data->alignment = alignment; 10243bfe5deSAlex Elder osd_data->pages_from_pool = pages_from_pool; 10343bfe5deSAlex Elder osd_data->own_pages = own_pages; 10443bfe5deSAlex Elder } 10543bfe5deSAlex Elder 106a4ce40a9SAlex Elder static void ceph_osd_data_pagelist_init(struct ceph_osd_data *osd_data, 10743bfe5deSAlex Elder struct ceph_pagelist *pagelist) 10843bfe5deSAlex Elder { 10943bfe5deSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_PAGELIST; 11043bfe5deSAlex Elder osd_data->pagelist = pagelist; 11143bfe5deSAlex Elder } 11243bfe5deSAlex Elder 11343bfe5deSAlex Elder #ifdef CONFIG_BLOCK 114a4ce40a9SAlex Elder static void ceph_osd_data_bio_init(struct ceph_osd_data *osd_data, 11543bfe5deSAlex Elder struct bio *bio, size_t bio_length) 11643bfe5deSAlex Elder { 11743bfe5deSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_BIO; 11843bfe5deSAlex Elder osd_data->bio = bio; 11943bfe5deSAlex Elder osd_data->bio_length = bio_length; 12043bfe5deSAlex Elder } 12143bfe5deSAlex Elder #endif /* CONFIG_BLOCK */ 12243bfe5deSAlex Elder 123863c7eb5SAlex Elder #define osd_req_op_data(oreq, whch, typ, fld) \ 124863c7eb5SAlex Elder ({ \ 1258a703a38SIoana Ciornei struct ceph_osd_request *__oreq = (oreq); \ 1268a703a38SIoana Ciornei unsigned int __whch = (whch); \ 1278a703a38SIoana Ciornei BUG_ON(__whch >= __oreq->r_num_ops); \ 1288a703a38SIoana Ciornei &__oreq->r_ops[__whch].typ.fld; \ 129863c7eb5SAlex Elder }) 130863c7eb5SAlex Elder 13149719778SAlex Elder static struct ceph_osd_data * 13249719778SAlex Elder osd_req_op_raw_data_in(struct ceph_osd_request *osd_req, unsigned int which) 13349719778SAlex Elder { 13449719778SAlex Elder BUG_ON(which >= osd_req->r_num_ops); 13549719778SAlex Elder 13649719778SAlex Elder return &osd_req->r_ops[which].raw_data_in; 13749719778SAlex Elder } 13849719778SAlex Elder 139a4ce40a9SAlex Elder struct ceph_osd_data * 140a4ce40a9SAlex Elder osd_req_op_extent_osd_data(struct ceph_osd_request *osd_req, 141406e2c9fSAlex Elder unsigned int which) 142a4ce40a9SAlex Elder { 143863c7eb5SAlex Elder return osd_req_op_data(osd_req, which, extent, osd_data); 144a4ce40a9SAlex Elder } 145a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data); 146a4ce40a9SAlex Elder 147a4ce40a9SAlex Elder struct ceph_osd_data * 148a4ce40a9SAlex Elder osd_req_op_cls_response_data(struct ceph_osd_request *osd_req, 149a4ce40a9SAlex Elder unsigned int which) 150a4ce40a9SAlex Elder { 151863c7eb5SAlex Elder return osd_req_op_data(osd_req, which, cls, response_data); 152a4ce40a9SAlex Elder } 153a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_response_data); /* ??? */ 154a4ce40a9SAlex Elder 15549719778SAlex Elder void osd_req_op_raw_data_in_pages(struct ceph_osd_request *osd_req, 15649719778SAlex Elder unsigned int which, struct page **pages, 15749719778SAlex Elder u64 length, u32 alignment, 15849719778SAlex Elder bool pages_from_pool, bool own_pages) 15949719778SAlex Elder { 16049719778SAlex Elder struct ceph_osd_data *osd_data; 16149719778SAlex Elder 16249719778SAlex Elder osd_data = osd_req_op_raw_data_in(osd_req, which); 16349719778SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 16449719778SAlex Elder pages_from_pool, own_pages); 16549719778SAlex Elder } 16649719778SAlex Elder EXPORT_SYMBOL(osd_req_op_raw_data_in_pages); 16749719778SAlex Elder 168a4ce40a9SAlex Elder void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *osd_req, 169406e2c9fSAlex Elder unsigned int which, struct page **pages, 170406e2c9fSAlex Elder u64 length, u32 alignment, 171a4ce40a9SAlex Elder bool pages_from_pool, bool own_pages) 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_pages_init(osd_data, pages, length, alignment, 177a4ce40a9SAlex Elder pages_from_pool, own_pages); 178a4ce40a9SAlex Elder } 179a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data_pages); 180a4ce40a9SAlex Elder 181a4ce40a9SAlex Elder void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *osd_req, 182406e2c9fSAlex Elder unsigned int which, struct ceph_pagelist *pagelist) 183a4ce40a9SAlex Elder { 184a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 185a4ce40a9SAlex Elder 186863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, extent, osd_data); 187a4ce40a9SAlex Elder ceph_osd_data_pagelist_init(osd_data, pagelist); 188a4ce40a9SAlex Elder } 189a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data_pagelist); 190a4ce40a9SAlex Elder 191a4ce40a9SAlex Elder #ifdef CONFIG_BLOCK 192a4ce40a9SAlex Elder void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *osd_req, 193406e2c9fSAlex Elder unsigned int which, struct bio *bio, size_t bio_length) 194a4ce40a9SAlex Elder { 195a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 196863c7eb5SAlex Elder 197863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, extent, osd_data); 198a4ce40a9SAlex Elder ceph_osd_data_bio_init(osd_data, bio, bio_length); 199a4ce40a9SAlex Elder } 200a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data_bio); 201a4ce40a9SAlex Elder #endif /* CONFIG_BLOCK */ 202a4ce40a9SAlex Elder 203a4ce40a9SAlex Elder static void osd_req_op_cls_request_info_pagelist( 204a4ce40a9SAlex Elder struct ceph_osd_request *osd_req, 205a4ce40a9SAlex Elder unsigned int which, struct ceph_pagelist *pagelist) 206a4ce40a9SAlex Elder { 207a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 208a4ce40a9SAlex Elder 209863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, request_info); 210a4ce40a9SAlex Elder ceph_osd_data_pagelist_init(osd_data, pagelist); 211a4ce40a9SAlex Elder } 212a4ce40a9SAlex Elder 21304017e29SAlex Elder void osd_req_op_cls_request_data_pagelist( 21404017e29SAlex Elder struct ceph_osd_request *osd_req, 21504017e29SAlex Elder unsigned int which, struct ceph_pagelist *pagelist) 21604017e29SAlex Elder { 21704017e29SAlex Elder struct ceph_osd_data *osd_data; 21804017e29SAlex Elder 219863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, request_data); 22004017e29SAlex Elder ceph_osd_data_pagelist_init(osd_data, pagelist); 22104017e29SAlex Elder } 22204017e29SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_request_data_pagelist); 22304017e29SAlex Elder 2246c57b554SAlex Elder void osd_req_op_cls_request_data_pages(struct ceph_osd_request *osd_req, 2256c57b554SAlex Elder unsigned int which, struct page **pages, u64 length, 2266c57b554SAlex Elder u32 alignment, bool pages_from_pool, bool own_pages) 2276c57b554SAlex Elder { 2286c57b554SAlex Elder struct ceph_osd_data *osd_data; 2296c57b554SAlex Elder 2306c57b554SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, request_data); 2316c57b554SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 2326c57b554SAlex Elder pages_from_pool, own_pages); 2336c57b554SAlex Elder } 2346c57b554SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_request_data_pages); 2356c57b554SAlex Elder 236a4ce40a9SAlex Elder void osd_req_op_cls_response_data_pages(struct ceph_osd_request *osd_req, 237a4ce40a9SAlex Elder unsigned int which, struct page **pages, u64 length, 238a4ce40a9SAlex Elder u32 alignment, bool pages_from_pool, bool own_pages) 239a4ce40a9SAlex Elder { 240a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 241a4ce40a9SAlex Elder 242863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, response_data); 243a4ce40a9SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 244a4ce40a9SAlex Elder pages_from_pool, own_pages); 245a4ce40a9SAlex Elder } 246a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_response_data_pages); 247a4ce40a9SAlex Elder 24823c08a9cSAlex Elder static u64 ceph_osd_data_length(struct ceph_osd_data *osd_data) 24923c08a9cSAlex Elder { 25023c08a9cSAlex Elder switch (osd_data->type) { 25123c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_NONE: 25223c08a9cSAlex Elder return 0; 25323c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_PAGES: 25423c08a9cSAlex Elder return osd_data->length; 25523c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_PAGELIST: 25623c08a9cSAlex Elder return (u64)osd_data->pagelist->length; 25723c08a9cSAlex Elder #ifdef CONFIG_BLOCK 25823c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_BIO: 25923c08a9cSAlex Elder return (u64)osd_data->bio_length; 26023c08a9cSAlex Elder #endif /* CONFIG_BLOCK */ 26123c08a9cSAlex Elder default: 26223c08a9cSAlex Elder WARN(true, "unrecognized data type %d\n", (int)osd_data->type); 26323c08a9cSAlex Elder return 0; 26423c08a9cSAlex Elder } 26523c08a9cSAlex Elder } 26623c08a9cSAlex Elder 267c54d47bfSAlex Elder static void ceph_osd_data_release(struct ceph_osd_data *osd_data) 268c54d47bfSAlex Elder { 2695476492fSAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES && osd_data->own_pages) { 270c54d47bfSAlex Elder int num_pages; 271c54d47bfSAlex Elder 272c54d47bfSAlex Elder num_pages = calc_pages_for((u64)osd_data->alignment, 273c54d47bfSAlex Elder (u64)osd_data->length); 274c54d47bfSAlex Elder ceph_release_page_vector(osd_data->pages, num_pages); 275c54d47bfSAlex Elder } 2765476492fSAlex Elder ceph_osd_data_init(osd_data); 2775476492fSAlex Elder } 2785476492fSAlex Elder 2795476492fSAlex Elder static void osd_req_op_data_release(struct ceph_osd_request *osd_req, 2805476492fSAlex Elder unsigned int which) 2815476492fSAlex Elder { 2825476492fSAlex Elder struct ceph_osd_req_op *op; 2835476492fSAlex Elder 2845476492fSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 2855476492fSAlex Elder op = &osd_req->r_ops[which]; 2865476492fSAlex Elder 2875476492fSAlex Elder switch (op->op) { 2885476492fSAlex Elder case CEPH_OSD_OP_READ: 2895476492fSAlex Elder case CEPH_OSD_OP_WRITE: 290e30b7577SIlya Dryomov case CEPH_OSD_OP_WRITEFULL: 2915476492fSAlex Elder ceph_osd_data_release(&op->extent.osd_data); 2925476492fSAlex Elder break; 2935476492fSAlex Elder case CEPH_OSD_OP_CALL: 2945476492fSAlex Elder ceph_osd_data_release(&op->cls.request_info); 29504017e29SAlex Elder ceph_osd_data_release(&op->cls.request_data); 2965476492fSAlex Elder ceph_osd_data_release(&op->cls.response_data); 2975476492fSAlex Elder break; 298d74b50beSYan, Zheng case CEPH_OSD_OP_SETXATTR: 299d74b50beSYan, Zheng case CEPH_OSD_OP_CMPXATTR: 300d74b50beSYan, Zheng ceph_osd_data_release(&op->xattr.osd_data); 301d74b50beSYan, Zheng break; 30266ba609fSYan, Zheng case CEPH_OSD_OP_STAT: 30366ba609fSYan, Zheng ceph_osd_data_release(&op->raw_data_in); 30466ba609fSYan, Zheng break; 3055476492fSAlex Elder default: 3065476492fSAlex Elder break; 3075476492fSAlex Elder } 308c54d47bfSAlex Elder } 309c54d47bfSAlex Elder 3103d14c5d2SYehuda Sadeh /* 3113d14c5d2SYehuda Sadeh * requests 3123d14c5d2SYehuda Sadeh */ 3139e94af20SIlya Dryomov static void ceph_osdc_release_request(struct kref *kref) 3143d14c5d2SYehuda Sadeh { 3159e94af20SIlya Dryomov struct ceph_osd_request *req = container_of(kref, 3169e94af20SIlya Dryomov struct ceph_osd_request, r_kref); 3175476492fSAlex Elder unsigned int which; 3183d14c5d2SYehuda Sadeh 3199e94af20SIlya Dryomov dout("%s %p (r_request %p r_reply %p)\n", __func__, req, 3209e94af20SIlya Dryomov req->r_request, req->r_reply); 3216562d661SIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&req->r_node)); 3226562d661SIlya Dryomov WARN_ON(!list_empty(&req->r_req_lru_item)); 3236562d661SIlya Dryomov WARN_ON(!list_empty(&req->r_osd_item)); 3246562d661SIlya Dryomov WARN_ON(!list_empty(&req->r_linger_item)); 3256562d661SIlya Dryomov WARN_ON(!list_empty(&req->r_linger_osd_item)); 3266562d661SIlya Dryomov WARN_ON(req->r_osd); 3279e94af20SIlya Dryomov 3283d14c5d2SYehuda Sadeh if (req->r_request) 3293d14c5d2SYehuda Sadeh ceph_msg_put(req->r_request); 330ace6d3a9SAlex Elder if (req->r_reply) { 3318921d114SAlex Elder ceph_msg_revoke_incoming(req->r_reply); 332ab8cb34aSAlex Elder ceph_msg_put(req->r_reply); 333ace6d3a9SAlex Elder } 3340fff87ecSAlex Elder 3355476492fSAlex Elder for (which = 0; which < req->r_num_ops; which++) 3365476492fSAlex Elder osd_req_op_data_release(req, which); 3370fff87ecSAlex Elder 3383d14c5d2SYehuda Sadeh ceph_put_snap_context(req->r_snapc); 3393d14c5d2SYehuda Sadeh if (req->r_mempool) 3403d14c5d2SYehuda Sadeh mempool_free(req, req->r_osdc->req_mempool); 3413f1af42aSIlya Dryomov else if (req->r_num_ops <= CEPH_OSD_SLAB_OPS) 3425522ae0bSAlex Elder kmem_cache_free(ceph_osd_request_cache, req); 3433f1af42aSIlya Dryomov else 3443f1af42aSIlya Dryomov kfree(req); 3453d14c5d2SYehuda Sadeh } 3469e94af20SIlya Dryomov 3479e94af20SIlya Dryomov void ceph_osdc_get_request(struct ceph_osd_request *req) 3489e94af20SIlya Dryomov { 3499e94af20SIlya Dryomov dout("%s %p (was %d)\n", __func__, req, 3509e94af20SIlya Dryomov atomic_read(&req->r_kref.refcount)); 3519e94af20SIlya Dryomov kref_get(&req->r_kref); 3529e94af20SIlya Dryomov } 3539e94af20SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_get_request); 3549e94af20SIlya Dryomov 3559e94af20SIlya Dryomov void ceph_osdc_put_request(struct ceph_osd_request *req) 3569e94af20SIlya Dryomov { 3573ed97d63SIlya Dryomov if (req) { 3589e94af20SIlya Dryomov dout("%s %p (was %d)\n", __func__, req, 3599e94af20SIlya Dryomov atomic_read(&req->r_kref.refcount)); 3609e94af20SIlya Dryomov kref_put(&req->r_kref, ceph_osdc_release_request); 3619e94af20SIlya Dryomov } 3623ed97d63SIlya Dryomov } 3639e94af20SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_put_request); 3643d14c5d2SYehuda Sadeh 3653d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc, 3663d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 3671b83bef2SSage Weil unsigned int num_ops, 3683d14c5d2SYehuda Sadeh bool use_mempool, 36954a54007SAlex Elder gfp_t gfp_flags) 3703d14c5d2SYehuda Sadeh { 3713d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 3723d14c5d2SYehuda Sadeh struct ceph_msg *msg; 3731b83bef2SSage Weil size_t msg_size; 3743d14c5d2SYehuda Sadeh 3753d14c5d2SYehuda Sadeh if (use_mempool) { 3763f1af42aSIlya Dryomov BUG_ON(num_ops > CEPH_OSD_SLAB_OPS); 3773d14c5d2SYehuda Sadeh req = mempool_alloc(osdc->req_mempool, gfp_flags); 3783f1af42aSIlya Dryomov } else if (num_ops <= CEPH_OSD_SLAB_OPS) { 3793f1af42aSIlya Dryomov req = kmem_cache_alloc(ceph_osd_request_cache, gfp_flags); 3803d14c5d2SYehuda Sadeh } else { 3813f1af42aSIlya Dryomov BUG_ON(num_ops > CEPH_OSD_MAX_OPS); 3823f1af42aSIlya Dryomov req = kmalloc(sizeof(*req) + num_ops * sizeof(req->r_ops[0]), 3833f1af42aSIlya Dryomov gfp_flags); 3843d14c5d2SYehuda Sadeh } 3853f1af42aSIlya Dryomov if (unlikely(!req)) 3863d14c5d2SYehuda Sadeh return NULL; 3873d14c5d2SYehuda Sadeh 3883f1af42aSIlya Dryomov /* req only, each op is zeroed in _osd_req_op_init() */ 3893f1af42aSIlya Dryomov memset(req, 0, sizeof(*req)); 3903f1af42aSIlya Dryomov 3913d14c5d2SYehuda Sadeh req->r_osdc = osdc; 3923d14c5d2SYehuda Sadeh req->r_mempool = use_mempool; 39379528734SAlex Elder req->r_num_ops = num_ops; 39484127282SIlya Dryomov req->r_snapid = CEPH_NOSNAP; 39584127282SIlya Dryomov req->r_snapc = ceph_get_snap_context(snapc); 3963d14c5d2SYehuda Sadeh 3973d14c5d2SYehuda Sadeh kref_init(&req->r_kref); 3983d14c5d2SYehuda Sadeh init_completion(&req->r_completion); 3993d14c5d2SYehuda Sadeh init_completion(&req->r_safe_completion); 400a978fa20SAlex Elder RB_CLEAR_NODE(&req->r_node); 4013d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&req->r_unsafe_item); 402a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&req->r_linger_item); 4031d0326b1SIlya Dryomov INIT_LIST_HEAD(&req->r_linger_osd_item); 404935b639aSSage Weil INIT_LIST_HEAD(&req->r_req_lru_item); 405cd43045cSSage Weil INIT_LIST_HEAD(&req->r_osd_item); 406cd43045cSSage Weil 4073c972c95SIlya Dryomov req->r_base_oloc.pool = -1; 408205ee118SIlya Dryomov req->r_target_oloc.pool = -1; 40922116525SIlya Dryomov 4103f1af42aSIlya Dryomov msg_size = OSD_OPREPLY_FRONT_LEN; 4113f1af42aSIlya Dryomov if (num_ops > CEPH_OSD_SLAB_OPS) { 4123f1af42aSIlya Dryomov /* ceph_osd_op and rval */ 4133f1af42aSIlya Dryomov msg_size += (num_ops - CEPH_OSD_SLAB_OPS) * 4143f1af42aSIlya Dryomov (sizeof(struct ceph_osd_op) + 4); 4153f1af42aSIlya Dryomov } 4163f1af42aSIlya Dryomov 4173d14c5d2SYehuda Sadeh /* create reply message */ 4183d14c5d2SYehuda Sadeh if (use_mempool) 4193d14c5d2SYehuda Sadeh msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0); 4203d14c5d2SYehuda Sadeh else 4213f1af42aSIlya Dryomov msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, msg_size, 4223f1af42aSIlya Dryomov gfp_flags, true); 4233d14c5d2SYehuda Sadeh if (!msg) { 4243d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 4253d14c5d2SYehuda Sadeh return NULL; 4263d14c5d2SYehuda Sadeh } 4273d14c5d2SYehuda Sadeh req->r_reply = msg; 4283d14c5d2SYehuda Sadeh 429ae458f5aSIlya Dryomov msg_size = 4 + 4 + 4; /* client_inc, osdmap_epoch, flags */ 430ae458f5aSIlya Dryomov msg_size += 4 + 4 + 4 + 8; /* mtime, reassert_version */ 431ae458f5aSIlya Dryomov msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */ 432ae458f5aSIlya Dryomov msg_size += 1 + 8 + 4 + 4; /* pgid */ 433ae458f5aSIlya Dryomov msg_size += 4 + CEPH_MAX_OID_NAME_LEN; /* oid */ 434ae458f5aSIlya Dryomov msg_size += 2 + num_ops * sizeof(struct ceph_osd_op); 435ae458f5aSIlya Dryomov msg_size += 8; /* snapid */ 436ae458f5aSIlya Dryomov msg_size += 8; /* snap_seq */ 437ae458f5aSIlya Dryomov msg_size += 4 + 8 * (snapc ? snapc->num_snaps : 0); /* snaps */ 438ae458f5aSIlya Dryomov msg_size += 4; /* retry_attempt */ 439ae458f5aSIlya Dryomov 4403d14c5d2SYehuda Sadeh /* create request message; allow space for oid */ 4413d14c5d2SYehuda Sadeh if (use_mempool) 4423d14c5d2SYehuda Sadeh msg = ceph_msgpool_get(&osdc->msgpool_op, 0); 4433d14c5d2SYehuda Sadeh else 444b61c2763SSage Weil msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true); 4453d14c5d2SYehuda Sadeh if (!msg) { 4463d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 4473d14c5d2SYehuda Sadeh return NULL; 4483d14c5d2SYehuda Sadeh } 4493d14c5d2SYehuda Sadeh 4503d14c5d2SYehuda Sadeh memset(msg->front.iov_base, 0, msg->front.iov_len); 4513d14c5d2SYehuda Sadeh 4523d14c5d2SYehuda Sadeh req->r_request = msg; 4533d14c5d2SYehuda Sadeh 4543d14c5d2SYehuda Sadeh return req; 4553d14c5d2SYehuda Sadeh } 4563d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_alloc_request); 4573d14c5d2SYehuda Sadeh 458a8dd0a37SAlex Elder static bool osd_req_opcode_valid(u16 opcode) 459a8dd0a37SAlex Elder { 460a8dd0a37SAlex Elder switch (opcode) { 46170b5bfa3SIlya Dryomov #define GENERATE_CASE(op, opcode, str) case CEPH_OSD_OP_##op: return true; 46270b5bfa3SIlya Dryomov __CEPH_FORALL_OSD_OPS(GENERATE_CASE) 46370b5bfa3SIlya Dryomov #undef GENERATE_CASE 464a8dd0a37SAlex Elder default: 465a8dd0a37SAlex Elder return false; 466a8dd0a37SAlex Elder } 467a8dd0a37SAlex Elder } 468a8dd0a37SAlex Elder 46933803f33SAlex Elder /* 47033803f33SAlex Elder * This is an osd op init function for opcodes that have no data or 47133803f33SAlex Elder * other information associated with them. It also serves as a 47233803f33SAlex Elder * common init routine for all the other init functions, below. 47333803f33SAlex Elder */ 474c99d2d4aSAlex Elder static struct ceph_osd_req_op * 47549719778SAlex Elder _osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which, 476144cba14SYan, Zheng u16 opcode, u32 flags) 47733803f33SAlex Elder { 478c99d2d4aSAlex Elder struct ceph_osd_req_op *op; 479c99d2d4aSAlex Elder 480c99d2d4aSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 48133803f33SAlex Elder BUG_ON(!osd_req_opcode_valid(opcode)); 48233803f33SAlex Elder 483c99d2d4aSAlex Elder op = &osd_req->r_ops[which]; 48433803f33SAlex Elder memset(op, 0, sizeof (*op)); 48533803f33SAlex Elder op->op = opcode; 486144cba14SYan, Zheng op->flags = flags; 487c99d2d4aSAlex Elder 488c99d2d4aSAlex Elder return op; 48933803f33SAlex Elder } 49033803f33SAlex Elder 49149719778SAlex Elder void osd_req_op_init(struct ceph_osd_request *osd_req, 492144cba14SYan, Zheng unsigned int which, u16 opcode, u32 flags) 49349719778SAlex Elder { 494144cba14SYan, Zheng (void)_osd_req_op_init(osd_req, which, opcode, flags); 49549719778SAlex Elder } 49649719778SAlex Elder EXPORT_SYMBOL(osd_req_op_init); 49749719778SAlex Elder 498c99d2d4aSAlex Elder void osd_req_op_extent_init(struct ceph_osd_request *osd_req, 499c99d2d4aSAlex Elder unsigned int which, u16 opcode, 50033803f33SAlex Elder u64 offset, u64 length, 50133803f33SAlex Elder u64 truncate_size, u32 truncate_seq) 50233803f33SAlex Elder { 503144cba14SYan, Zheng struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, 504144cba14SYan, Zheng opcode, 0); 50533803f33SAlex Elder size_t payload_len = 0; 50633803f33SAlex Elder 507ad7a60deSLi Wang BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE && 508e30b7577SIlya Dryomov opcode != CEPH_OSD_OP_WRITEFULL && opcode != CEPH_OSD_OP_ZERO && 509e30b7577SIlya Dryomov opcode != CEPH_OSD_OP_TRUNCATE); 51033803f33SAlex Elder 51133803f33SAlex Elder op->extent.offset = offset; 51233803f33SAlex Elder op->extent.length = length; 51333803f33SAlex Elder op->extent.truncate_size = truncate_size; 51433803f33SAlex Elder op->extent.truncate_seq = truncate_seq; 515e30b7577SIlya Dryomov if (opcode == CEPH_OSD_OP_WRITE || opcode == CEPH_OSD_OP_WRITEFULL) 51633803f33SAlex Elder payload_len += length; 51733803f33SAlex Elder 518de2aa102SIlya Dryomov op->indata_len = payload_len; 51933803f33SAlex Elder } 52033803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_init); 52133803f33SAlex Elder 522c99d2d4aSAlex Elder void osd_req_op_extent_update(struct ceph_osd_request *osd_req, 523c99d2d4aSAlex Elder unsigned int which, u64 length) 524e5975c7cSAlex Elder { 525c99d2d4aSAlex Elder struct ceph_osd_req_op *op; 526c99d2d4aSAlex Elder u64 previous; 527c99d2d4aSAlex Elder 528c99d2d4aSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 529c99d2d4aSAlex Elder op = &osd_req->r_ops[which]; 530c99d2d4aSAlex Elder previous = op->extent.length; 531e5975c7cSAlex Elder 532e5975c7cSAlex Elder if (length == previous) 533e5975c7cSAlex Elder return; /* Nothing to do */ 534e5975c7cSAlex Elder BUG_ON(length > previous); 535e5975c7cSAlex Elder 536e5975c7cSAlex Elder op->extent.length = length; 537de2aa102SIlya Dryomov op->indata_len -= previous - length; 538e5975c7cSAlex Elder } 539e5975c7cSAlex Elder EXPORT_SYMBOL(osd_req_op_extent_update); 540e5975c7cSAlex Elder 5412c63f49aSYan, Zheng void osd_req_op_extent_dup_last(struct ceph_osd_request *osd_req, 5422c63f49aSYan, Zheng unsigned int which, u64 offset_inc) 5432c63f49aSYan, Zheng { 5442c63f49aSYan, Zheng struct ceph_osd_req_op *op, *prev_op; 5452c63f49aSYan, Zheng 5462c63f49aSYan, Zheng BUG_ON(which + 1 >= osd_req->r_num_ops); 5472c63f49aSYan, Zheng 5482c63f49aSYan, Zheng prev_op = &osd_req->r_ops[which]; 5492c63f49aSYan, Zheng op = _osd_req_op_init(osd_req, which + 1, prev_op->op, prev_op->flags); 5502c63f49aSYan, Zheng /* dup previous one */ 5512c63f49aSYan, Zheng op->indata_len = prev_op->indata_len; 5522c63f49aSYan, Zheng op->outdata_len = prev_op->outdata_len; 5532c63f49aSYan, Zheng op->extent = prev_op->extent; 5542c63f49aSYan, Zheng /* adjust offset */ 5552c63f49aSYan, Zheng op->extent.offset += offset_inc; 5562c63f49aSYan, Zheng op->extent.length -= offset_inc; 5572c63f49aSYan, Zheng 5582c63f49aSYan, Zheng if (op->op == CEPH_OSD_OP_WRITE || op->op == CEPH_OSD_OP_WRITEFULL) 5592c63f49aSYan, Zheng op->indata_len -= offset_inc; 5602c63f49aSYan, Zheng } 5612c63f49aSYan, Zheng EXPORT_SYMBOL(osd_req_op_extent_dup_last); 5622c63f49aSYan, Zheng 563c99d2d4aSAlex Elder void osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which, 56404017e29SAlex Elder u16 opcode, const char *class, const char *method) 56533803f33SAlex Elder { 566144cba14SYan, Zheng struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, 567144cba14SYan, Zheng opcode, 0); 5685f562df5SAlex Elder struct ceph_pagelist *pagelist; 56933803f33SAlex Elder size_t payload_len = 0; 57033803f33SAlex Elder size_t size; 57133803f33SAlex Elder 57233803f33SAlex Elder BUG_ON(opcode != CEPH_OSD_OP_CALL); 57333803f33SAlex Elder 5745f562df5SAlex Elder pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS); 5755f562df5SAlex Elder BUG_ON(!pagelist); 5765f562df5SAlex Elder ceph_pagelist_init(pagelist); 5775f562df5SAlex Elder 57833803f33SAlex Elder op->cls.class_name = class; 57933803f33SAlex Elder size = strlen(class); 58033803f33SAlex Elder BUG_ON(size > (size_t) U8_MAX); 58133803f33SAlex Elder op->cls.class_len = size; 5825f562df5SAlex Elder ceph_pagelist_append(pagelist, class, size); 58333803f33SAlex Elder payload_len += size; 58433803f33SAlex Elder 58533803f33SAlex Elder op->cls.method_name = method; 58633803f33SAlex Elder size = strlen(method); 58733803f33SAlex Elder BUG_ON(size > (size_t) U8_MAX); 58833803f33SAlex Elder op->cls.method_len = size; 5895f562df5SAlex Elder ceph_pagelist_append(pagelist, method, size); 59033803f33SAlex Elder payload_len += size; 59133803f33SAlex Elder 592a4ce40a9SAlex Elder osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist); 5935f562df5SAlex Elder 59433803f33SAlex Elder op->cls.argc = 0; /* currently unused */ 59533803f33SAlex Elder 596de2aa102SIlya Dryomov op->indata_len = payload_len; 59733803f33SAlex Elder } 59833803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_init); 5998c042b0dSAlex Elder 600d74b50beSYan, Zheng int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which, 601d74b50beSYan, Zheng u16 opcode, const char *name, const void *value, 602d74b50beSYan, Zheng size_t size, u8 cmp_op, u8 cmp_mode) 603d74b50beSYan, Zheng { 604144cba14SYan, Zheng struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, 605144cba14SYan, Zheng opcode, 0); 606d74b50beSYan, Zheng struct ceph_pagelist *pagelist; 607d74b50beSYan, Zheng size_t payload_len; 608d74b50beSYan, Zheng 609d74b50beSYan, Zheng BUG_ON(opcode != CEPH_OSD_OP_SETXATTR && opcode != CEPH_OSD_OP_CMPXATTR); 610d74b50beSYan, Zheng 611d74b50beSYan, Zheng pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS); 612d74b50beSYan, Zheng if (!pagelist) 613d74b50beSYan, Zheng return -ENOMEM; 614d74b50beSYan, Zheng 615d74b50beSYan, Zheng ceph_pagelist_init(pagelist); 616d74b50beSYan, Zheng 617d74b50beSYan, Zheng payload_len = strlen(name); 618d74b50beSYan, Zheng op->xattr.name_len = payload_len; 619d74b50beSYan, Zheng ceph_pagelist_append(pagelist, name, payload_len); 620d74b50beSYan, Zheng 621d74b50beSYan, Zheng op->xattr.value_len = size; 622d74b50beSYan, Zheng ceph_pagelist_append(pagelist, value, size); 623d74b50beSYan, Zheng payload_len += size; 624d74b50beSYan, Zheng 625d74b50beSYan, Zheng op->xattr.cmp_op = cmp_op; 626d74b50beSYan, Zheng op->xattr.cmp_mode = cmp_mode; 627d74b50beSYan, Zheng 628d74b50beSYan, Zheng ceph_osd_data_pagelist_init(&op->xattr.osd_data, pagelist); 629de2aa102SIlya Dryomov op->indata_len = payload_len; 630d74b50beSYan, Zheng return 0; 631d74b50beSYan, Zheng } 632d74b50beSYan, Zheng EXPORT_SYMBOL(osd_req_op_xattr_init); 633d74b50beSYan, Zheng 634c99d2d4aSAlex Elder void osd_req_op_watch_init(struct ceph_osd_request *osd_req, 635c99d2d4aSAlex Elder unsigned int which, u16 opcode, 63633803f33SAlex Elder u64 cookie, u64 version, int flag) 63733803f33SAlex Elder { 638144cba14SYan, Zheng struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, 639144cba14SYan, Zheng opcode, 0); 64033803f33SAlex Elder 641c99d2d4aSAlex Elder BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH); 64233803f33SAlex Elder 64333803f33SAlex Elder op->watch.cookie = cookie; 6449ef1ee5aSAlex Elder op->watch.ver = version; 64533803f33SAlex Elder if (opcode == CEPH_OSD_OP_WATCH && flag) 64633803f33SAlex Elder op->watch.flag = (u8)1; 64733803f33SAlex Elder } 64833803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_watch_init); 64933803f33SAlex Elder 650c647b8a8SIlya Dryomov void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req, 651c647b8a8SIlya Dryomov unsigned int which, 652c647b8a8SIlya Dryomov u64 expected_object_size, 653c647b8a8SIlya Dryomov u64 expected_write_size) 654c647b8a8SIlya Dryomov { 655c647b8a8SIlya Dryomov struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, 656144cba14SYan, Zheng CEPH_OSD_OP_SETALLOCHINT, 657144cba14SYan, Zheng 0); 658c647b8a8SIlya Dryomov 659c647b8a8SIlya Dryomov op->alloc_hint.expected_object_size = expected_object_size; 660c647b8a8SIlya Dryomov op->alloc_hint.expected_write_size = expected_write_size; 661c647b8a8SIlya Dryomov 662c647b8a8SIlya Dryomov /* 663c647b8a8SIlya Dryomov * CEPH_OSD_OP_SETALLOCHINT op is advisory and therefore deemed 664c647b8a8SIlya Dryomov * not worth a feature bit. Set FAILOK per-op flag to make 665c647b8a8SIlya Dryomov * sure older osds don't trip over an unsupported opcode. 666c647b8a8SIlya Dryomov */ 667c647b8a8SIlya Dryomov op->flags |= CEPH_OSD_OP_FLAG_FAILOK; 668c647b8a8SIlya Dryomov } 669c647b8a8SIlya Dryomov EXPORT_SYMBOL(osd_req_op_alloc_hint_init); 670c647b8a8SIlya Dryomov 67190af3602SAlex Elder static void ceph_osdc_msg_data_add(struct ceph_msg *msg, 672ec9123c5SAlex Elder struct ceph_osd_data *osd_data) 673ec9123c5SAlex Elder { 674ec9123c5SAlex Elder u64 length = ceph_osd_data_length(osd_data); 675ec9123c5SAlex Elder 676ec9123c5SAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 677ec9123c5SAlex Elder BUG_ON(length > (u64) SIZE_MAX); 678ec9123c5SAlex Elder if (length) 67990af3602SAlex Elder ceph_msg_data_add_pages(msg, osd_data->pages, 680ec9123c5SAlex Elder length, osd_data->alignment); 681ec9123c5SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) { 682ec9123c5SAlex Elder BUG_ON(!length); 68390af3602SAlex Elder ceph_msg_data_add_pagelist(msg, osd_data->pagelist); 684ec9123c5SAlex Elder #ifdef CONFIG_BLOCK 685ec9123c5SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) { 68690af3602SAlex Elder ceph_msg_data_add_bio(msg, osd_data->bio, length); 687ec9123c5SAlex Elder #endif 688ec9123c5SAlex Elder } else { 689ec9123c5SAlex Elder BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE); 690ec9123c5SAlex Elder } 691ec9123c5SAlex Elder } 692ec9123c5SAlex Elder 693175face2SAlex Elder static u64 osd_req_encode_op(struct ceph_osd_request *req, 69479528734SAlex Elder struct ceph_osd_op *dst, unsigned int which) 6953d14c5d2SYehuda Sadeh { 69679528734SAlex Elder struct ceph_osd_req_op *src; 69704017e29SAlex Elder struct ceph_osd_data *osd_data; 69854d50649SAlex Elder u64 request_data_len = 0; 69904017e29SAlex Elder u64 data_length; 700175face2SAlex Elder 70179528734SAlex Elder BUG_ON(which >= req->r_num_ops); 70279528734SAlex Elder src = &req->r_ops[which]; 703a8dd0a37SAlex Elder if (WARN_ON(!osd_req_opcode_valid(src->op))) { 704a8dd0a37SAlex Elder pr_err("unrecognized osd opcode %d\n", src->op); 705a8dd0a37SAlex Elder 706a8dd0a37SAlex Elder return 0; 707a8dd0a37SAlex Elder } 7083d14c5d2SYehuda Sadeh 709065a68f9SAlex Elder switch (src->op) { 710fbfab539SAlex Elder case CEPH_OSD_OP_STAT: 71149719778SAlex Elder osd_data = &src->raw_data_in; 71249719778SAlex Elder ceph_osdc_msg_data_add(req->r_reply, osd_data); 713fbfab539SAlex Elder break; 7143d14c5d2SYehuda Sadeh case CEPH_OSD_OP_READ: 7153d14c5d2SYehuda Sadeh case CEPH_OSD_OP_WRITE: 716e30b7577SIlya Dryomov case CEPH_OSD_OP_WRITEFULL: 717ad7a60deSLi Wang case CEPH_OSD_OP_ZERO: 718ad7a60deSLi Wang case CEPH_OSD_OP_TRUNCATE: 719e30b7577SIlya Dryomov if (src->op == CEPH_OSD_OP_WRITE || 720e30b7577SIlya Dryomov src->op == CEPH_OSD_OP_WRITEFULL) 72154d50649SAlex Elder request_data_len = src->extent.length; 722175face2SAlex Elder dst->extent.offset = cpu_to_le64(src->extent.offset); 723175face2SAlex Elder dst->extent.length = cpu_to_le64(src->extent.length); 7243d14c5d2SYehuda Sadeh dst->extent.truncate_size = 7253d14c5d2SYehuda Sadeh cpu_to_le64(src->extent.truncate_size); 7263d14c5d2SYehuda Sadeh dst->extent.truncate_seq = 7273d14c5d2SYehuda Sadeh cpu_to_le32(src->extent.truncate_seq); 72804017e29SAlex Elder osd_data = &src->extent.osd_data; 729e30b7577SIlya Dryomov if (src->op == CEPH_OSD_OP_WRITE || 730e30b7577SIlya Dryomov src->op == CEPH_OSD_OP_WRITEFULL) 73104017e29SAlex Elder ceph_osdc_msg_data_add(req->r_request, osd_data); 7325476492fSAlex Elder else 73304017e29SAlex Elder ceph_osdc_msg_data_add(req->r_reply, osd_data); 7343d14c5d2SYehuda Sadeh break; 7353d14c5d2SYehuda Sadeh case CEPH_OSD_OP_CALL: 7363d14c5d2SYehuda Sadeh dst->cls.class_len = src->cls.class_len; 7373d14c5d2SYehuda Sadeh dst->cls.method_len = src->cls.method_len; 73804017e29SAlex Elder osd_data = &src->cls.request_info; 73904017e29SAlex Elder ceph_osdc_msg_data_add(req->r_request, osd_data); 74004017e29SAlex Elder BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGELIST); 74104017e29SAlex Elder request_data_len = osd_data->pagelist->length; 74204017e29SAlex Elder 74304017e29SAlex Elder osd_data = &src->cls.request_data; 74404017e29SAlex Elder data_length = ceph_osd_data_length(osd_data); 74504017e29SAlex Elder if (data_length) { 74604017e29SAlex Elder BUG_ON(osd_data->type == CEPH_OSD_DATA_TYPE_NONE); 74704017e29SAlex Elder dst->cls.indata_len = cpu_to_le32(data_length); 74804017e29SAlex Elder ceph_osdc_msg_data_add(req->r_request, osd_data); 749de2aa102SIlya Dryomov src->indata_len += data_length; 75004017e29SAlex Elder request_data_len += data_length; 75104017e29SAlex Elder } 75204017e29SAlex Elder osd_data = &src->cls.response_data; 75304017e29SAlex Elder ceph_osdc_msg_data_add(req->r_reply, osd_data); 7543d14c5d2SYehuda Sadeh break; 7553d14c5d2SYehuda Sadeh case CEPH_OSD_OP_STARTSYNC: 7563d14c5d2SYehuda Sadeh break; 757a40c4f10SYehuda Sadeh case CEPH_OSD_OP_NOTIFY_ACK: 758a40c4f10SYehuda Sadeh case CEPH_OSD_OP_WATCH: 759a40c4f10SYehuda Sadeh dst->watch.cookie = cpu_to_le64(src->watch.cookie); 760a40c4f10SYehuda Sadeh dst->watch.ver = cpu_to_le64(src->watch.ver); 761a40c4f10SYehuda Sadeh dst->watch.flag = src->watch.flag; 762a40c4f10SYehuda Sadeh break; 763c647b8a8SIlya Dryomov case CEPH_OSD_OP_SETALLOCHINT: 764c647b8a8SIlya Dryomov dst->alloc_hint.expected_object_size = 765c647b8a8SIlya Dryomov cpu_to_le64(src->alloc_hint.expected_object_size); 766c647b8a8SIlya Dryomov dst->alloc_hint.expected_write_size = 767c647b8a8SIlya Dryomov cpu_to_le64(src->alloc_hint.expected_write_size); 768c647b8a8SIlya Dryomov break; 769d74b50beSYan, Zheng case CEPH_OSD_OP_SETXATTR: 770d74b50beSYan, Zheng case CEPH_OSD_OP_CMPXATTR: 771d74b50beSYan, Zheng dst->xattr.name_len = cpu_to_le32(src->xattr.name_len); 772d74b50beSYan, Zheng dst->xattr.value_len = cpu_to_le32(src->xattr.value_len); 773d74b50beSYan, Zheng dst->xattr.cmp_op = src->xattr.cmp_op; 774d74b50beSYan, Zheng dst->xattr.cmp_mode = src->xattr.cmp_mode; 775d74b50beSYan, Zheng osd_data = &src->xattr.osd_data; 776d74b50beSYan, Zheng ceph_osdc_msg_data_add(req->r_request, osd_data); 777d74b50beSYan, Zheng request_data_len = osd_data->pagelist->length; 778d74b50beSYan, Zheng break; 779864e9197SYan, Zheng case CEPH_OSD_OP_CREATE: 780864e9197SYan, Zheng case CEPH_OSD_OP_DELETE: 781864e9197SYan, Zheng break; 7823d14c5d2SYehuda Sadeh default: 7834c46459cSAlex Elder pr_err("unsupported osd opcode %s\n", 7848f63ca2dSAlex Elder ceph_osd_op_name(src->op)); 7854c46459cSAlex Elder WARN_ON(1); 786a8dd0a37SAlex Elder 787a8dd0a37SAlex Elder return 0; 7883d14c5d2SYehuda Sadeh } 7897b25bf5fSIlya Dryomov 790a8dd0a37SAlex Elder dst->op = cpu_to_le16(src->op); 7917b25bf5fSIlya Dryomov dst->flags = cpu_to_le32(src->flags); 792de2aa102SIlya Dryomov dst->payload_len = cpu_to_le32(src->indata_len); 793175face2SAlex Elder 79454d50649SAlex Elder return request_data_len; 7953d14c5d2SYehuda Sadeh } 7963d14c5d2SYehuda Sadeh 7973d14c5d2SYehuda Sadeh /* 7983d14c5d2SYehuda Sadeh * build new request AND message, calculate layout, and adjust file 7993d14c5d2SYehuda Sadeh * extent as needed. 8003d14c5d2SYehuda Sadeh * 8013d14c5d2SYehuda Sadeh * if the file was recently truncated, we include information about its 8023d14c5d2SYehuda Sadeh * old and new size so that the object can be updated appropriately. (we 8033d14c5d2SYehuda Sadeh * avoid synchronously deleting truncated objects because it's slow.) 8043d14c5d2SYehuda Sadeh * 8053d14c5d2SYehuda Sadeh * if @do_sync, include a 'startsync' command so that the osd will flush 8063d14c5d2SYehuda Sadeh * data quickly. 8073d14c5d2SYehuda Sadeh */ 8083d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc, 8093d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 8103d14c5d2SYehuda Sadeh struct ceph_vino vino, 811715e4cd4SYan, Zheng u64 off, u64 *plen, 812715e4cd4SYan, Zheng unsigned int which, int num_ops, 8133d14c5d2SYehuda Sadeh int opcode, int flags, 8143d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 8153d14c5d2SYehuda Sadeh u32 truncate_seq, 8163d14c5d2SYehuda Sadeh u64 truncate_size, 817153e5167SAlex Elder bool use_mempool) 8183d14c5d2SYehuda Sadeh { 8193d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 82075d1c941SAlex Elder u64 objnum = 0; 82175d1c941SAlex Elder u64 objoff = 0; 82275d1c941SAlex Elder u64 objlen = 0; 8236816282dSSage Weil int r; 8243d14c5d2SYehuda Sadeh 825ad7a60deSLi Wang BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE && 826864e9197SYan, Zheng opcode != CEPH_OSD_OP_ZERO && opcode != CEPH_OSD_OP_TRUNCATE && 827864e9197SYan, Zheng opcode != CEPH_OSD_OP_CREATE && opcode != CEPH_OSD_OP_DELETE); 8283d14c5d2SYehuda Sadeh 829acead002SAlex Elder req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool, 830ae7ca4a3SAlex Elder GFP_NOFS); 8314ad12621SSage Weil if (!req) 8326816282dSSage Weil return ERR_PTR(-ENOMEM); 83379528734SAlex Elder 834d178a9e7SAlex Elder req->r_flags = flags; 8353d14c5d2SYehuda Sadeh 8363d14c5d2SYehuda Sadeh /* calculate max write size */ 837a19dadfbSAlex Elder r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen); 8383ff5f385SAlex Elder if (r < 0) { 8393ff5f385SAlex Elder ceph_osdc_put_request(req); 8406816282dSSage Weil return ERR_PTR(r); 8413ff5f385SAlex Elder } 842a19dadfbSAlex Elder 843864e9197SYan, Zheng if (opcode == CEPH_OSD_OP_CREATE || opcode == CEPH_OSD_OP_DELETE) { 844144cba14SYan, Zheng osd_req_op_init(req, which, opcode, 0); 845864e9197SYan, Zheng } else { 846864e9197SYan, Zheng u32 object_size = le32_to_cpu(layout->fl_object_size); 847864e9197SYan, Zheng u32 object_base = off - objoff; 848ccca4e37SYan, Zheng if (!(truncate_seq == 1 && truncate_size == -1ULL)) { 849d18d1e28SAlex Elder if (truncate_size <= object_base) { 850d18d1e28SAlex Elder truncate_size = 0; 851d18d1e28SAlex Elder } else { 852d18d1e28SAlex Elder truncate_size -= object_base; 853d18d1e28SAlex Elder if (truncate_size > object_size) 854d18d1e28SAlex Elder truncate_size = object_size; 855d18d1e28SAlex Elder } 856ccca4e37SYan, Zheng } 857715e4cd4SYan, Zheng osd_req_op_extent_init(req, which, opcode, objoff, objlen, 858b0270324SAlex Elder truncate_size, truncate_seq); 859864e9197SYan, Zheng } 860d18d1e28SAlex Elder 8613c972c95SIlya Dryomov req->r_base_oloc.pool = ceph_file_layout_pg_pool(*layout); 8623d14c5d2SYehuda Sadeh 8633c972c95SIlya Dryomov snprintf(req->r_base_oid.name, sizeof(req->r_base_oid.name), 8644295f221SIlya Dryomov "%llx.%08llx", vino.ino, objnum); 8653c972c95SIlya Dryomov req->r_base_oid.name_len = strlen(req->r_base_oid.name); 866dbe0fc41SAlex Elder 8673d14c5d2SYehuda Sadeh return req; 8683d14c5d2SYehuda Sadeh } 8693d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_new_request); 8703d14c5d2SYehuda Sadeh 8713d14c5d2SYehuda Sadeh /* 8723d14c5d2SYehuda Sadeh * We keep osd requests in an rbtree, sorted by ->r_tid. 8733d14c5d2SYehuda Sadeh */ 8743d14c5d2SYehuda Sadeh static void __insert_request(struct ceph_osd_client *osdc, 8753d14c5d2SYehuda Sadeh struct ceph_osd_request *new) 8763d14c5d2SYehuda Sadeh { 8773d14c5d2SYehuda Sadeh struct rb_node **p = &osdc->requests.rb_node; 8783d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 8793d14c5d2SYehuda Sadeh struct ceph_osd_request *req = NULL; 8803d14c5d2SYehuda Sadeh 8813d14c5d2SYehuda Sadeh while (*p) { 8823d14c5d2SYehuda Sadeh parent = *p; 8833d14c5d2SYehuda Sadeh req = rb_entry(parent, struct ceph_osd_request, r_node); 8843d14c5d2SYehuda Sadeh if (new->r_tid < req->r_tid) 8853d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 8863d14c5d2SYehuda Sadeh else if (new->r_tid > req->r_tid) 8873d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 8883d14c5d2SYehuda Sadeh else 8893d14c5d2SYehuda Sadeh BUG(); 8903d14c5d2SYehuda Sadeh } 8913d14c5d2SYehuda Sadeh 8923d14c5d2SYehuda Sadeh rb_link_node(&new->r_node, parent, p); 8933d14c5d2SYehuda Sadeh rb_insert_color(&new->r_node, &osdc->requests); 8943d14c5d2SYehuda Sadeh } 8953d14c5d2SYehuda Sadeh 8963d14c5d2SYehuda Sadeh static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc, 8973d14c5d2SYehuda Sadeh u64 tid) 8983d14c5d2SYehuda Sadeh { 8993d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 9003d14c5d2SYehuda Sadeh struct rb_node *n = osdc->requests.rb_node; 9013d14c5d2SYehuda Sadeh 9023d14c5d2SYehuda Sadeh while (n) { 9033d14c5d2SYehuda Sadeh req = rb_entry(n, struct ceph_osd_request, r_node); 9043d14c5d2SYehuda Sadeh if (tid < req->r_tid) 9053d14c5d2SYehuda Sadeh n = n->rb_left; 9063d14c5d2SYehuda Sadeh else if (tid > req->r_tid) 9073d14c5d2SYehuda Sadeh n = n->rb_right; 9083d14c5d2SYehuda Sadeh else 9093d14c5d2SYehuda Sadeh return req; 9103d14c5d2SYehuda Sadeh } 9113d14c5d2SYehuda Sadeh return NULL; 9123d14c5d2SYehuda Sadeh } 9133d14c5d2SYehuda Sadeh 9143d14c5d2SYehuda Sadeh static struct ceph_osd_request * 9153d14c5d2SYehuda Sadeh __lookup_request_ge(struct ceph_osd_client *osdc, 9163d14c5d2SYehuda Sadeh u64 tid) 9173d14c5d2SYehuda Sadeh { 9183d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 9193d14c5d2SYehuda Sadeh struct rb_node *n = osdc->requests.rb_node; 9203d14c5d2SYehuda Sadeh 9213d14c5d2SYehuda Sadeh while (n) { 9223d14c5d2SYehuda Sadeh req = rb_entry(n, struct ceph_osd_request, r_node); 9233d14c5d2SYehuda Sadeh if (tid < req->r_tid) { 9243d14c5d2SYehuda Sadeh if (!n->rb_left) 9253d14c5d2SYehuda Sadeh return req; 9263d14c5d2SYehuda Sadeh n = n->rb_left; 9273d14c5d2SYehuda Sadeh } else if (tid > req->r_tid) { 9283d14c5d2SYehuda Sadeh n = n->rb_right; 9293d14c5d2SYehuda Sadeh } else { 9303d14c5d2SYehuda Sadeh return req; 9313d14c5d2SYehuda Sadeh } 9323d14c5d2SYehuda Sadeh } 9333d14c5d2SYehuda Sadeh return NULL; 9343d14c5d2SYehuda Sadeh } 9353d14c5d2SYehuda Sadeh 9362cc6128aSIlya Dryomov static void __kick_linger_request(struct ceph_osd_request *req) 9372cc6128aSIlya Dryomov { 9382cc6128aSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 9392cc6128aSIlya Dryomov struct ceph_osd *osd = req->r_osd; 9402cc6128aSIlya Dryomov 9412cc6128aSIlya Dryomov /* 9422cc6128aSIlya Dryomov * Linger requests need to be resent with a new tid to avoid 9432cc6128aSIlya Dryomov * the dup op detection logic on the OSDs. Achieve this with 9442cc6128aSIlya Dryomov * a re-register dance instead of open-coding. 9452cc6128aSIlya Dryomov */ 9462cc6128aSIlya Dryomov ceph_osdc_get_request(req); 9472cc6128aSIlya Dryomov if (!list_empty(&req->r_linger_item)) 9482cc6128aSIlya Dryomov __unregister_linger_request(osdc, req); 9492cc6128aSIlya Dryomov else 9502cc6128aSIlya Dryomov __unregister_request(osdc, req); 9512cc6128aSIlya Dryomov __register_request(osdc, req); 9522cc6128aSIlya Dryomov ceph_osdc_put_request(req); 9532cc6128aSIlya Dryomov 9542cc6128aSIlya Dryomov /* 9552cc6128aSIlya Dryomov * Unless request has been registered as both normal and 9562cc6128aSIlya Dryomov * lingering, __unregister{,_linger}_request clears r_osd. 9572cc6128aSIlya Dryomov * However, here we need to preserve r_osd to make sure we 9582cc6128aSIlya Dryomov * requeue on the same OSD. 9592cc6128aSIlya Dryomov */ 9602cc6128aSIlya Dryomov WARN_ON(req->r_osd || !osd); 9612cc6128aSIlya Dryomov req->r_osd = osd; 9622cc6128aSIlya Dryomov 9632cc6128aSIlya Dryomov dout("%s requeueing %p tid %llu\n", __func__, req, req->r_tid); 9642cc6128aSIlya Dryomov __enqueue_request(req); 9652cc6128aSIlya Dryomov } 9662cc6128aSIlya Dryomov 9676f6c7006SSage Weil /* 9686f6c7006SSage Weil * Resubmit requests pending on the given osd. 9696f6c7006SSage Weil */ 9706f6c7006SSage Weil static void __kick_osd_requests(struct ceph_osd_client *osdc, 9716f6c7006SSage Weil struct ceph_osd *osd) 9726f6c7006SSage Weil { 973a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 974e02493c0SAlex Elder LIST_HEAD(resend); 9752cc6128aSIlya Dryomov LIST_HEAD(resend_linger); 9766f6c7006SSage Weil int err; 9776f6c7006SSage Weil 9782cc6128aSIlya Dryomov dout("%s osd%d\n", __func__, osd->o_osd); 9796f6c7006SSage Weil err = __reset_osd(osdc, osd); 980685a7555SAlex Elder if (err) 9816f6c7006SSage Weil return; 9822cc6128aSIlya Dryomov 983e02493c0SAlex Elder /* 984e02493c0SAlex Elder * Build up a list of requests to resend by traversing the 985e02493c0SAlex Elder * osd's list of requests. Requests for a given object are 986e02493c0SAlex Elder * sent in tid order, and that is also the order they're 987e02493c0SAlex Elder * kept on this list. Therefore all requests that are in 988e02493c0SAlex Elder * flight will be found first, followed by all requests that 989e02493c0SAlex Elder * have not yet been sent. And to resend requests while 990e02493c0SAlex Elder * preserving this order we will want to put any sent 991e02493c0SAlex Elder * requests back on the front of the osd client's unsent 992e02493c0SAlex Elder * list. 993e02493c0SAlex Elder * 994e02493c0SAlex Elder * So we build a separate ordered list of already-sent 995e02493c0SAlex Elder * requests for the affected osd and splice it onto the 996e02493c0SAlex Elder * front of the osd client's unsent list. Once we've seen a 997e02493c0SAlex Elder * request that has not yet been sent we're done. Those 998e02493c0SAlex Elder * requests are already sitting right where they belong. 999e02493c0SAlex Elder */ 10006f6c7006SSage Weil list_for_each_entry(req, &osd->o_requests, r_osd_item) { 1001e02493c0SAlex Elder if (!req->r_sent) 1002e02493c0SAlex Elder break; 10032cc6128aSIlya Dryomov 10042cc6128aSIlya Dryomov if (!req->r_linger) { 10052cc6128aSIlya Dryomov dout("%s requeueing %p tid %llu\n", __func__, req, 10062cc6128aSIlya Dryomov req->r_tid); 1007e02493c0SAlex Elder list_move_tail(&req->r_req_lru_item, &resend); 10086f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 10092cc6128aSIlya Dryomov } else { 10102cc6128aSIlya Dryomov list_move_tail(&req->r_req_lru_item, &resend_linger); 10112cc6128aSIlya Dryomov } 10126f6c7006SSage Weil } 1013e02493c0SAlex Elder list_splice(&resend, &osdc->req_unsent); 1014a40c4f10SYehuda Sadeh 1015e02493c0SAlex Elder /* 10162cc6128aSIlya Dryomov * Both registered and not yet registered linger requests are 10172cc6128aSIlya Dryomov * enqueued with a new tid on the same OSD. We add/move them 10182cc6128aSIlya Dryomov * to req_unsent/o_requests at the end to keep things in tid 10192cc6128aSIlya Dryomov * order. 1020e02493c0SAlex Elder */ 1021a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osd->o_linger_requests, 10221d0326b1SIlya Dryomov r_linger_osd_item) { 10232cc6128aSIlya Dryomov WARN_ON(!list_empty(&req->r_req_lru_item)); 10242cc6128aSIlya Dryomov __kick_linger_request(req); 1025a40c4f10SYehuda Sadeh } 10262cc6128aSIlya Dryomov 10272cc6128aSIlya Dryomov list_for_each_entry_safe(req, nreq, &resend_linger, r_req_lru_item) 10282cc6128aSIlya Dryomov __kick_linger_request(req); 10296f6c7006SSage Weil } 10306f6c7006SSage Weil 10313d14c5d2SYehuda Sadeh /* 10323d14c5d2SYehuda Sadeh * If the osd connection drops, we need to resubmit all requests. 10333d14c5d2SYehuda Sadeh */ 10343d14c5d2SYehuda Sadeh static void osd_reset(struct ceph_connection *con) 10353d14c5d2SYehuda Sadeh { 10363d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 10373d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 10383d14c5d2SYehuda Sadeh 10393d14c5d2SYehuda Sadeh if (!osd) 10403d14c5d2SYehuda Sadeh return; 10413d14c5d2SYehuda Sadeh dout("osd_reset osd%d\n", osd->o_osd); 10423d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 10433d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 104483aff95eSSage Weil mutex_lock(&osdc->request_mutex); 104583aff95eSSage Weil __kick_osd_requests(osdc, osd); 1046f9d25199SAlex Elder __send_queued(osdc); 104783aff95eSSage Weil mutex_unlock(&osdc->request_mutex); 10483d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 10493d14c5d2SYehuda Sadeh } 10503d14c5d2SYehuda Sadeh 10513d14c5d2SYehuda Sadeh /* 10523d14c5d2SYehuda Sadeh * Track open sessions with osds. 10533d14c5d2SYehuda Sadeh */ 1054e10006f8SAlex Elder static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum) 10553d14c5d2SYehuda Sadeh { 10563d14c5d2SYehuda Sadeh struct ceph_osd *osd; 10573d14c5d2SYehuda Sadeh 10583d14c5d2SYehuda Sadeh osd = kzalloc(sizeof(*osd), GFP_NOFS); 10593d14c5d2SYehuda Sadeh if (!osd) 10603d14c5d2SYehuda Sadeh return NULL; 10613d14c5d2SYehuda Sadeh 10623d14c5d2SYehuda Sadeh atomic_set(&osd->o_ref, 1); 10633d14c5d2SYehuda Sadeh osd->o_osdc = osdc; 1064e10006f8SAlex Elder osd->o_osd = onum; 1065f407731dSAlex Elder RB_CLEAR_NODE(&osd->o_node); 10663d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_requests); 1067a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osd->o_linger_requests); 10683d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_osd_lru); 10693d14c5d2SYehuda Sadeh osd->o_incarnation = 1; 10703d14c5d2SYehuda Sadeh 1071b7a9e5ddSSage Weil ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr); 10723d14c5d2SYehuda Sadeh 10733d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_keepalive_item); 10743d14c5d2SYehuda Sadeh return osd; 10753d14c5d2SYehuda Sadeh } 10763d14c5d2SYehuda Sadeh 10773d14c5d2SYehuda Sadeh static struct ceph_osd *get_osd(struct ceph_osd *osd) 10783d14c5d2SYehuda Sadeh { 10793d14c5d2SYehuda Sadeh if (atomic_inc_not_zero(&osd->o_ref)) { 10803d14c5d2SYehuda Sadeh dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1, 10813d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref)); 10823d14c5d2SYehuda Sadeh return osd; 10833d14c5d2SYehuda Sadeh } else { 10843d14c5d2SYehuda Sadeh dout("get_osd %p FAIL\n", osd); 10853d14c5d2SYehuda Sadeh return NULL; 10863d14c5d2SYehuda Sadeh } 10873d14c5d2SYehuda Sadeh } 10883d14c5d2SYehuda Sadeh 10893d14c5d2SYehuda Sadeh static void put_osd(struct ceph_osd *osd) 10903d14c5d2SYehuda Sadeh { 10913d14c5d2SYehuda Sadeh dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref), 10923d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref) - 1); 1093b28ec2f3SIlya Dryomov if (atomic_dec_and_test(&osd->o_ref)) { 1094b28ec2f3SIlya Dryomov if (osd->o_auth.authorizer) 10956c1ea260SIlya Dryomov ceph_auth_destroy_authorizer(osd->o_auth.authorizer); 10963d14c5d2SYehuda Sadeh kfree(osd); 10973d14c5d2SYehuda Sadeh } 10983d14c5d2SYehuda Sadeh } 10993d14c5d2SYehuda Sadeh 11003d14c5d2SYehuda Sadeh /* 11013d14c5d2SYehuda Sadeh * remove an osd from our map 11023d14c5d2SYehuda Sadeh */ 11033d14c5d2SYehuda Sadeh static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 11043d14c5d2SYehuda Sadeh { 11057eb71e03SIlya Dryomov dout("%s %p osd%d\n", __func__, osd, osd->o_osd); 1106cc9f1f51SIlya Dryomov WARN_ON(!list_empty(&osd->o_requests)); 1107cc9f1f51SIlya Dryomov WARN_ON(!list_empty(&osd->o_linger_requests)); 11087c6e6fc5SIlya Dryomov 11093d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 11107eb71e03SIlya Dryomov rb_erase(&osd->o_node, &osdc->osds); 11117eb71e03SIlya Dryomov RB_CLEAR_NODE(&osd->o_node); 11127eb71e03SIlya Dryomov } 11137eb71e03SIlya Dryomov 11147eb71e03SIlya Dryomov static void remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 11157eb71e03SIlya Dryomov { 11167eb71e03SIlya Dryomov dout("%s %p osd%d\n", __func__, osd, osd->o_osd); 11177eb71e03SIlya Dryomov 11187eb71e03SIlya Dryomov if (!RB_EMPTY_NODE(&osd->o_node)) { 11193d14c5d2SYehuda Sadeh ceph_con_close(&osd->o_con); 11207eb71e03SIlya Dryomov __remove_osd(osdc, osd); 11213d14c5d2SYehuda Sadeh put_osd(osd); 11223d14c5d2SYehuda Sadeh } 11237eb71e03SIlya Dryomov } 11243d14c5d2SYehuda Sadeh 1125aca420bcSSage Weil static void remove_all_osds(struct ceph_osd_client *osdc) 1126aca420bcSSage Weil { 1127048a9d2dSJiaju Zhang dout("%s %p\n", __func__, osdc); 1128aca420bcSSage Weil mutex_lock(&osdc->request_mutex); 1129aca420bcSSage Weil while (!RB_EMPTY_ROOT(&osdc->osds)) { 1130aca420bcSSage Weil struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds), 1131aca420bcSSage Weil struct ceph_osd, o_node); 11327eb71e03SIlya Dryomov remove_osd(osdc, osd); 1133aca420bcSSage Weil } 1134aca420bcSSage Weil mutex_unlock(&osdc->request_mutex); 1135aca420bcSSage Weil } 1136aca420bcSSage Weil 11373d14c5d2SYehuda Sadeh static void __move_osd_to_lru(struct ceph_osd_client *osdc, 11383d14c5d2SYehuda Sadeh struct ceph_osd *osd) 11393d14c5d2SYehuda Sadeh { 1140bbf37ec3SIlya Dryomov dout("%s %p\n", __func__, osd); 11413d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_osd_lru)); 1142bbf37ec3SIlya Dryomov 11433d14c5d2SYehuda Sadeh list_add_tail(&osd->o_osd_lru, &osdc->osd_lru); 1144a319bf56SIlya Dryomov osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl; 11453d14c5d2SYehuda Sadeh } 11463d14c5d2SYehuda Sadeh 1147bbf37ec3SIlya Dryomov static void maybe_move_osd_to_lru(struct ceph_osd_client *osdc, 1148bbf37ec3SIlya Dryomov struct ceph_osd *osd) 1149bbf37ec3SIlya Dryomov { 1150bbf37ec3SIlya Dryomov dout("%s %p\n", __func__, osd); 1151bbf37ec3SIlya Dryomov 1152bbf37ec3SIlya Dryomov if (list_empty(&osd->o_requests) && 1153bbf37ec3SIlya Dryomov list_empty(&osd->o_linger_requests)) 1154bbf37ec3SIlya Dryomov __move_osd_to_lru(osdc, osd); 1155bbf37ec3SIlya Dryomov } 1156bbf37ec3SIlya Dryomov 11573d14c5d2SYehuda Sadeh static void __remove_osd_from_lru(struct ceph_osd *osd) 11583d14c5d2SYehuda Sadeh { 11593d14c5d2SYehuda Sadeh dout("__remove_osd_from_lru %p\n", osd); 11603d14c5d2SYehuda Sadeh if (!list_empty(&osd->o_osd_lru)) 11613d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 11623d14c5d2SYehuda Sadeh } 11633d14c5d2SYehuda Sadeh 1164aca420bcSSage Weil static void remove_old_osds(struct ceph_osd_client *osdc) 11653d14c5d2SYehuda Sadeh { 11663d14c5d2SYehuda Sadeh struct ceph_osd *osd, *nosd; 11673d14c5d2SYehuda Sadeh 11683d14c5d2SYehuda Sadeh dout("__remove_old_osds %p\n", osdc); 11693d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 11703d14c5d2SYehuda Sadeh list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) { 1171aca420bcSSage Weil if (time_before(jiffies, osd->lru_ttl)) 11723d14c5d2SYehuda Sadeh break; 11737eb71e03SIlya Dryomov remove_osd(osdc, osd); 11743d14c5d2SYehuda Sadeh } 11753d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 11763d14c5d2SYehuda Sadeh } 11773d14c5d2SYehuda Sadeh 11783d14c5d2SYehuda Sadeh /* 11793d14c5d2SYehuda Sadeh * reset osd connect 11803d14c5d2SYehuda Sadeh */ 11813d14c5d2SYehuda Sadeh static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 11823d14c5d2SYehuda Sadeh { 1183c3acb181SAlex Elder struct ceph_entity_addr *peer_addr; 11843d14c5d2SYehuda Sadeh 11853d14c5d2SYehuda Sadeh dout("__reset_osd %p osd%d\n", osd, osd->o_osd); 1186a40c4f10SYehuda Sadeh if (list_empty(&osd->o_requests) && 1187a40c4f10SYehuda Sadeh list_empty(&osd->o_linger_requests)) { 11887eb71e03SIlya Dryomov remove_osd(osdc, osd); 1189c3acb181SAlex Elder return -ENODEV; 1190c3acb181SAlex Elder } 1191c3acb181SAlex Elder 1192c3acb181SAlex Elder peer_addr = &osdc->osdmap->osd_addr[osd->o_osd]; 1193c3acb181SAlex Elder if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) && 11943d14c5d2SYehuda Sadeh !ceph_con_opened(&osd->o_con)) { 1195c3acb181SAlex Elder struct ceph_osd_request *req; 1196c3acb181SAlex Elder 11973d14c5d2SYehuda Sadeh dout("osd addr hasn't changed and connection never opened, " 11980b4af2e8SIlya Dryomov "letting msgr retry\n"); 11993d14c5d2SYehuda Sadeh /* touch each r_stamp for handle_timeout()'s benfit */ 12003d14c5d2SYehuda Sadeh list_for_each_entry(req, &osd->o_requests, r_osd_item) 12013d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 1202c3acb181SAlex Elder 1203c3acb181SAlex Elder return -EAGAIN; 12043d14c5d2SYehuda Sadeh } 1205c3acb181SAlex Elder 1206c3acb181SAlex Elder ceph_con_close(&osd->o_con); 1207c3acb181SAlex Elder ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr); 1208c3acb181SAlex Elder osd->o_incarnation++; 1209c3acb181SAlex Elder 1210c3acb181SAlex Elder return 0; 12113d14c5d2SYehuda Sadeh } 12123d14c5d2SYehuda Sadeh 12133d14c5d2SYehuda Sadeh static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new) 12143d14c5d2SYehuda Sadeh { 12153d14c5d2SYehuda Sadeh struct rb_node **p = &osdc->osds.rb_node; 12163d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 12173d14c5d2SYehuda Sadeh struct ceph_osd *osd = NULL; 12183d14c5d2SYehuda Sadeh 1219aca420bcSSage Weil dout("__insert_osd %p osd%d\n", new, new->o_osd); 12203d14c5d2SYehuda Sadeh while (*p) { 12213d14c5d2SYehuda Sadeh parent = *p; 12223d14c5d2SYehuda Sadeh osd = rb_entry(parent, struct ceph_osd, o_node); 12233d14c5d2SYehuda Sadeh if (new->o_osd < osd->o_osd) 12243d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 12253d14c5d2SYehuda Sadeh else if (new->o_osd > osd->o_osd) 12263d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 12273d14c5d2SYehuda Sadeh else 12283d14c5d2SYehuda Sadeh BUG(); 12293d14c5d2SYehuda Sadeh } 12303d14c5d2SYehuda Sadeh 12313d14c5d2SYehuda Sadeh rb_link_node(&new->o_node, parent, p); 12323d14c5d2SYehuda Sadeh rb_insert_color(&new->o_node, &osdc->osds); 12333d14c5d2SYehuda Sadeh } 12343d14c5d2SYehuda Sadeh 12353d14c5d2SYehuda Sadeh static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o) 12363d14c5d2SYehuda Sadeh { 12373d14c5d2SYehuda Sadeh struct ceph_osd *osd; 12383d14c5d2SYehuda Sadeh struct rb_node *n = osdc->osds.rb_node; 12393d14c5d2SYehuda Sadeh 12403d14c5d2SYehuda Sadeh while (n) { 12413d14c5d2SYehuda Sadeh osd = rb_entry(n, struct ceph_osd, o_node); 12423d14c5d2SYehuda Sadeh if (o < osd->o_osd) 12433d14c5d2SYehuda Sadeh n = n->rb_left; 12443d14c5d2SYehuda Sadeh else if (o > osd->o_osd) 12453d14c5d2SYehuda Sadeh n = n->rb_right; 12463d14c5d2SYehuda Sadeh else 12473d14c5d2SYehuda Sadeh return osd; 12483d14c5d2SYehuda Sadeh } 12493d14c5d2SYehuda Sadeh return NULL; 12503d14c5d2SYehuda Sadeh } 12513d14c5d2SYehuda Sadeh 12523d14c5d2SYehuda Sadeh static void __schedule_osd_timeout(struct ceph_osd_client *osdc) 12533d14c5d2SYehuda Sadeh { 12543d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->timeout_work, 1255a319bf56SIlya Dryomov osdc->client->options->osd_keepalive_timeout); 12563d14c5d2SYehuda Sadeh } 12573d14c5d2SYehuda Sadeh 12583d14c5d2SYehuda Sadeh static void __cancel_osd_timeout(struct ceph_osd_client *osdc) 12593d14c5d2SYehuda Sadeh { 12603d14c5d2SYehuda Sadeh cancel_delayed_work(&osdc->timeout_work); 12613d14c5d2SYehuda Sadeh } 12623d14c5d2SYehuda Sadeh 12633d14c5d2SYehuda Sadeh /* 12643d14c5d2SYehuda Sadeh * Register request, assign tid. If this is the first request, set up 12653d14c5d2SYehuda Sadeh * the timeout event. 12663d14c5d2SYehuda Sadeh */ 1267a40c4f10SYehuda Sadeh static void __register_request(struct ceph_osd_client *osdc, 12683d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 12693d14c5d2SYehuda Sadeh { 12703d14c5d2SYehuda Sadeh req->r_tid = ++osdc->last_tid; 12713d14c5d2SYehuda Sadeh req->r_request->hdr.tid = cpu_to_le64(req->r_tid); 127277f38e0eSSage Weil dout("__register_request %p tid %lld\n", req, req->r_tid); 12733d14c5d2SYehuda Sadeh __insert_request(osdc, req); 12743d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 12753d14c5d2SYehuda Sadeh osdc->num_requests++; 12763d14c5d2SYehuda Sadeh if (osdc->num_requests == 1) { 12773d14c5d2SYehuda Sadeh dout(" first request, scheduling timeout\n"); 12783d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 12793d14c5d2SYehuda Sadeh } 1280a40c4f10SYehuda Sadeh } 1281a40c4f10SYehuda Sadeh 12823d14c5d2SYehuda Sadeh /* 12833d14c5d2SYehuda Sadeh * called under osdc->request_mutex 12843d14c5d2SYehuda Sadeh */ 12853d14c5d2SYehuda Sadeh static void __unregister_request(struct ceph_osd_client *osdc, 12863d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 12873d14c5d2SYehuda Sadeh { 128835f9f8a0SSage Weil if (RB_EMPTY_NODE(&req->r_node)) { 128935f9f8a0SSage Weil dout("__unregister_request %p tid %lld not registered\n", 129035f9f8a0SSage Weil req, req->r_tid); 129135f9f8a0SSage Weil return; 129235f9f8a0SSage Weil } 129335f9f8a0SSage Weil 12943d14c5d2SYehuda Sadeh dout("__unregister_request %p tid %lld\n", req, req->r_tid); 12953d14c5d2SYehuda Sadeh rb_erase(&req->r_node, &osdc->requests); 12966562d661SIlya Dryomov RB_CLEAR_NODE(&req->r_node); 12973d14c5d2SYehuda Sadeh osdc->num_requests--; 12983d14c5d2SYehuda Sadeh 12993d14c5d2SYehuda Sadeh if (req->r_osd) { 13003d14c5d2SYehuda Sadeh /* make sure the original request isn't in flight. */ 13016740a845SAlex Elder ceph_msg_revoke(req->r_request); 13023d14c5d2SYehuda Sadeh 13033d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 1304bbf37ec3SIlya Dryomov maybe_move_osd_to_lru(osdc, req->r_osd); 13054f23409eSIlya Dryomov if (list_empty(&req->r_linger_osd_item)) 13063d14c5d2SYehuda Sadeh req->r_osd = NULL; 13073d14c5d2SYehuda Sadeh } 13083d14c5d2SYehuda Sadeh 13097d5f2481SAlex Elder list_del_init(&req->r_req_lru_item); 13103d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 13113d14c5d2SYehuda Sadeh 13123d14c5d2SYehuda Sadeh if (osdc->num_requests == 0) { 13133d14c5d2SYehuda Sadeh dout(" no requests, canceling timeout\n"); 13143d14c5d2SYehuda Sadeh __cancel_osd_timeout(osdc); 13153d14c5d2SYehuda Sadeh } 13163d14c5d2SYehuda Sadeh } 13173d14c5d2SYehuda Sadeh 13183d14c5d2SYehuda Sadeh /* 13193d14c5d2SYehuda Sadeh * Cancel a previously queued request message 13203d14c5d2SYehuda Sadeh */ 13213d14c5d2SYehuda Sadeh static void __cancel_request(struct ceph_osd_request *req) 13223d14c5d2SYehuda Sadeh { 13233d14c5d2SYehuda Sadeh if (req->r_sent && req->r_osd) { 13246740a845SAlex Elder ceph_msg_revoke(req->r_request); 13253d14c5d2SYehuda Sadeh req->r_sent = 0; 13263d14c5d2SYehuda Sadeh } 13273d14c5d2SYehuda Sadeh } 13283d14c5d2SYehuda Sadeh 1329a40c4f10SYehuda Sadeh static void __register_linger_request(struct ceph_osd_client *osdc, 1330a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1331a40c4f10SYehuda Sadeh { 1332af593064SIlya Dryomov dout("%s %p tid %llu\n", __func__, req, req->r_tid); 1333af593064SIlya Dryomov WARN_ON(!req->r_linger); 1334af593064SIlya Dryomov 133596e4dac6SAlex Elder ceph_osdc_get_request(req); 1336a40c4f10SYehuda Sadeh list_add_tail(&req->r_linger_item, &osdc->req_linger); 13376194ea89SSage Weil if (req->r_osd) 13381d0326b1SIlya Dryomov list_add_tail(&req->r_linger_osd_item, 13396194ea89SSage Weil &req->r_osd->o_linger_requests); 1340a40c4f10SYehuda Sadeh } 1341a40c4f10SYehuda Sadeh 1342a40c4f10SYehuda Sadeh static void __unregister_linger_request(struct ceph_osd_client *osdc, 1343a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1344a40c4f10SYehuda Sadeh { 1345af593064SIlya Dryomov WARN_ON(!req->r_linger); 1346af593064SIlya Dryomov 1347af593064SIlya Dryomov if (list_empty(&req->r_linger_item)) { 1348af593064SIlya Dryomov dout("%s %p tid %llu not registered\n", __func__, req, 1349af593064SIlya Dryomov req->r_tid); 1350af593064SIlya Dryomov return; 1351af593064SIlya Dryomov } 1352af593064SIlya Dryomov 1353af593064SIlya Dryomov dout("%s %p tid %llu\n", __func__, req, req->r_tid); 1354a40c4f10SYehuda Sadeh list_del_init(&req->r_linger_item); 1355af593064SIlya Dryomov 135661c74035SAlex Elder if (req->r_osd) { 13571d0326b1SIlya Dryomov list_del_init(&req->r_linger_osd_item); 1358bbf37ec3SIlya Dryomov maybe_move_osd_to_lru(osdc, req->r_osd); 1359fbdb9190SSage Weil if (list_empty(&req->r_osd_item)) 1360a40c4f10SYehuda Sadeh req->r_osd = NULL; 1361a40c4f10SYehuda Sadeh } 136296e4dac6SAlex Elder ceph_osdc_put_request(req); 1363a40c4f10SYehuda Sadeh } 1364a40c4f10SYehuda Sadeh 1365a40c4f10SYehuda Sadeh void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc, 1366a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1367a40c4f10SYehuda Sadeh { 1368a40c4f10SYehuda Sadeh if (!req->r_linger) { 1369a40c4f10SYehuda Sadeh dout("set_request_linger %p\n", req); 1370a40c4f10SYehuda Sadeh req->r_linger = 1; 1371a40c4f10SYehuda Sadeh } 1372a40c4f10SYehuda Sadeh } 1373a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_set_request_linger); 1374a40c4f10SYehuda Sadeh 13753d14c5d2SYehuda Sadeh /* 1376d29adb34SJosh Durgin * Returns whether a request should be blocked from being sent 1377d29adb34SJosh Durgin * based on the current osdmap and osd_client settings. 1378d29adb34SJosh Durgin * 1379d29adb34SJosh Durgin * Caller should hold map_sem for read. 1380d29adb34SJosh Durgin */ 1381d29adb34SJosh Durgin static bool __req_should_be_paused(struct ceph_osd_client *osdc, 1382d29adb34SJosh Durgin struct ceph_osd_request *req) 1383d29adb34SJosh Durgin { 1384d29adb34SJosh Durgin bool pauserd = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD); 1385d29adb34SJosh Durgin bool pausewr = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR) || 1386d29adb34SJosh Durgin ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL); 1387d29adb34SJosh Durgin return (req->r_flags & CEPH_OSD_FLAG_READ && pauserd) || 1388d29adb34SJosh Durgin (req->r_flags & CEPH_OSD_FLAG_WRITE && pausewr); 1389d29adb34SJosh Durgin } 1390d29adb34SJosh Durgin 1391d29adb34SJosh Durgin /* 139217a13e40SIlya Dryomov * Calculate mapping of a request to a PG. Takes tiering into account. 139317a13e40SIlya Dryomov */ 139417a13e40SIlya Dryomov static int __calc_request_pg(struct ceph_osdmap *osdmap, 139517a13e40SIlya Dryomov struct ceph_osd_request *req, 139617a13e40SIlya Dryomov struct ceph_pg *pg_out) 139717a13e40SIlya Dryomov { 1398205ee118SIlya Dryomov bool need_check_tiering; 1399205ee118SIlya Dryomov 1400205ee118SIlya Dryomov need_check_tiering = false; 1401205ee118SIlya Dryomov if (req->r_target_oloc.pool == -1) { 1402205ee118SIlya Dryomov req->r_target_oloc = req->r_base_oloc; /* struct */ 1403205ee118SIlya Dryomov need_check_tiering = true; 1404205ee118SIlya Dryomov } 1405205ee118SIlya Dryomov if (req->r_target_oid.name_len == 0) { 1406205ee118SIlya Dryomov ceph_oid_copy(&req->r_target_oid, &req->r_base_oid); 1407205ee118SIlya Dryomov need_check_tiering = true; 1408205ee118SIlya Dryomov } 1409205ee118SIlya Dryomov 1410205ee118SIlya Dryomov if (need_check_tiering && 1411205ee118SIlya Dryomov (req->r_flags & CEPH_OSD_FLAG_IGNORE_OVERLAY) == 0) { 141217a13e40SIlya Dryomov struct ceph_pg_pool_info *pi; 141317a13e40SIlya Dryomov 1414205ee118SIlya Dryomov pi = ceph_pg_pool_by_id(osdmap, req->r_target_oloc.pool); 141517a13e40SIlya Dryomov if (pi) { 141617a13e40SIlya Dryomov if ((req->r_flags & CEPH_OSD_FLAG_READ) && 141717a13e40SIlya Dryomov pi->read_tier >= 0) 1418205ee118SIlya Dryomov req->r_target_oloc.pool = pi->read_tier; 141917a13e40SIlya Dryomov if ((req->r_flags & CEPH_OSD_FLAG_WRITE) && 142017a13e40SIlya Dryomov pi->write_tier >= 0) 1421205ee118SIlya Dryomov req->r_target_oloc.pool = pi->write_tier; 142217a13e40SIlya Dryomov } 142317a13e40SIlya Dryomov /* !pi is caught in ceph_oloc_oid_to_pg() */ 142417a13e40SIlya Dryomov } 142517a13e40SIlya Dryomov 1426205ee118SIlya Dryomov return ceph_oloc_oid_to_pg(osdmap, &req->r_target_oloc, 1427205ee118SIlya Dryomov &req->r_target_oid, pg_out); 142817a13e40SIlya Dryomov } 142917a13e40SIlya Dryomov 1430f671b581SIlya Dryomov static void __enqueue_request(struct ceph_osd_request *req) 1431f671b581SIlya Dryomov { 1432f671b581SIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 1433f671b581SIlya Dryomov 1434f671b581SIlya Dryomov dout("%s %p tid %llu to osd%d\n", __func__, req, req->r_tid, 1435f671b581SIlya Dryomov req->r_osd ? req->r_osd->o_osd : -1); 1436f671b581SIlya Dryomov 1437f671b581SIlya Dryomov if (req->r_osd) { 1438f671b581SIlya Dryomov __remove_osd_from_lru(req->r_osd); 1439f671b581SIlya Dryomov list_add_tail(&req->r_osd_item, &req->r_osd->o_requests); 1440f671b581SIlya Dryomov list_move_tail(&req->r_req_lru_item, &osdc->req_unsent); 1441f671b581SIlya Dryomov } else { 1442f671b581SIlya Dryomov list_move_tail(&req->r_req_lru_item, &osdc->req_notarget); 1443f671b581SIlya Dryomov } 1444f671b581SIlya Dryomov } 1445f671b581SIlya Dryomov 144617a13e40SIlya Dryomov /* 14473d14c5d2SYehuda Sadeh * Pick an osd (the first 'up' osd in the pg), allocate the osd struct 14483d14c5d2SYehuda Sadeh * (as needed), and set the request r_osd appropriately. If there is 144925985edcSLucas De Marchi * no up osd, set r_osd to NULL. Move the request to the appropriate list 14506f6c7006SSage Weil * (unsent, homeless) or leave on in-flight lru. 14513d14c5d2SYehuda Sadeh * 14523d14c5d2SYehuda Sadeh * Return 0 if unchanged, 1 if changed, or negative on error. 14533d14c5d2SYehuda Sadeh * 14543d14c5d2SYehuda Sadeh * Caller should hold map_sem for read and request_mutex. 14553d14c5d2SYehuda Sadeh */ 14566f6c7006SSage Weil static int __map_request(struct ceph_osd_client *osdc, 145738d6453cSSage Weil struct ceph_osd_request *req, int force_resend) 14583d14c5d2SYehuda Sadeh { 14595b191d99SSage Weil struct ceph_pg pgid; 14603d14c5d2SYehuda Sadeh int acting[CEPH_PG_MAX_SIZE]; 14618008ab10SIlya Dryomov int num, o; 14623d14c5d2SYehuda Sadeh int err; 1463d29adb34SJosh Durgin bool was_paused; 14643d14c5d2SYehuda Sadeh 14656f6c7006SSage Weil dout("map_request %p tid %lld\n", req, req->r_tid); 146617a13e40SIlya Dryomov 146717a13e40SIlya Dryomov err = __calc_request_pg(osdc->osdmap, req, &pgid); 14686f6c7006SSage Weil if (err) { 14696f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 14703d14c5d2SYehuda Sadeh return err; 14716f6c7006SSage Weil } 14723d14c5d2SYehuda Sadeh req->r_pgid = pgid; 14733d14c5d2SYehuda Sadeh 14748008ab10SIlya Dryomov num = ceph_calc_pg_acting(osdc->osdmap, pgid, acting, &o); 14758008ab10SIlya Dryomov if (num < 0) 14768008ab10SIlya Dryomov num = 0; 14773d14c5d2SYehuda Sadeh 1478d29adb34SJosh Durgin was_paused = req->r_paused; 1479d29adb34SJosh Durgin req->r_paused = __req_should_be_paused(osdc, req); 1480d29adb34SJosh Durgin if (was_paused && !req->r_paused) 1481d29adb34SJosh Durgin force_resend = 1; 1482d29adb34SJosh Durgin 148338d6453cSSage Weil if ((!force_resend && 148438d6453cSSage Weil req->r_osd && req->r_osd->o_osd == o && 14853d14c5d2SYehuda Sadeh req->r_sent >= req->r_osd->o_incarnation && 14863d14c5d2SYehuda Sadeh req->r_num_pg_osds == num && 14873d14c5d2SYehuda Sadeh memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) || 1488d29adb34SJosh Durgin (req->r_osd == NULL && o == -1) || 1489d29adb34SJosh Durgin req->r_paused) 14903d14c5d2SYehuda Sadeh return 0; /* no change */ 14913d14c5d2SYehuda Sadeh 14925b191d99SSage Weil dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n", 14935b191d99SSage Weil req->r_tid, pgid.pool, pgid.seed, o, 14943d14c5d2SYehuda Sadeh req->r_osd ? req->r_osd->o_osd : -1); 14953d14c5d2SYehuda Sadeh 14963d14c5d2SYehuda Sadeh /* record full pg acting set */ 14973d14c5d2SYehuda Sadeh memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num); 14983d14c5d2SYehuda Sadeh req->r_num_pg_osds = num; 14993d14c5d2SYehuda Sadeh 15003d14c5d2SYehuda Sadeh if (req->r_osd) { 15013d14c5d2SYehuda Sadeh __cancel_request(req); 15023d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 1503a390de02SIlya Dryomov list_del_init(&req->r_linger_osd_item); 15043d14c5d2SYehuda Sadeh req->r_osd = NULL; 15053d14c5d2SYehuda Sadeh } 15063d14c5d2SYehuda Sadeh 15073d14c5d2SYehuda Sadeh req->r_osd = __lookup_osd(osdc, o); 15083d14c5d2SYehuda Sadeh if (!req->r_osd && o >= 0) { 15093d14c5d2SYehuda Sadeh err = -ENOMEM; 1510e10006f8SAlex Elder req->r_osd = create_osd(osdc, o); 15116f6c7006SSage Weil if (!req->r_osd) { 15126f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 15133d14c5d2SYehuda Sadeh goto out; 15146f6c7006SSage Weil } 15153d14c5d2SYehuda Sadeh 15166f6c7006SSage Weil dout("map_request osd %p is osd%d\n", req->r_osd, o); 15173d14c5d2SYehuda Sadeh __insert_osd(osdc, req->r_osd); 15183d14c5d2SYehuda Sadeh 1519b7a9e5ddSSage Weil ceph_con_open(&req->r_osd->o_con, 1520b7a9e5ddSSage Weil CEPH_ENTITY_TYPE_OSD, o, 1521b7a9e5ddSSage Weil &osdc->osdmap->osd_addr[o]); 15223d14c5d2SYehuda Sadeh } 15233d14c5d2SYehuda Sadeh 1524f671b581SIlya Dryomov __enqueue_request(req); 15253d14c5d2SYehuda Sadeh err = 1; /* osd or pg changed */ 15263d14c5d2SYehuda Sadeh 15273d14c5d2SYehuda Sadeh out: 15283d14c5d2SYehuda Sadeh return err; 15293d14c5d2SYehuda Sadeh } 15303d14c5d2SYehuda Sadeh 15313d14c5d2SYehuda Sadeh /* 15323d14c5d2SYehuda Sadeh * caller should hold map_sem (for read) and request_mutex 15333d14c5d2SYehuda Sadeh */ 153456e925b6SSage Weil static void __send_request(struct ceph_osd_client *osdc, 15353d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 15363d14c5d2SYehuda Sadeh { 15371b83bef2SSage Weil void *p; 15383d14c5d2SYehuda Sadeh 15391b83bef2SSage Weil dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n", 15401b83bef2SSage Weil req, req->r_tid, req->r_osd->o_osd, req->r_flags, 15411b83bef2SSage Weil (unsigned long long)req->r_pgid.pool, req->r_pgid.seed); 15423d14c5d2SYehuda Sadeh 15431b83bef2SSage Weil /* fill in message content that changes each time we send it */ 15441b83bef2SSage Weil put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch); 15451b83bef2SSage Weil put_unaligned_le32(req->r_flags, req->r_request_flags); 1546205ee118SIlya Dryomov put_unaligned_le64(req->r_target_oloc.pool, req->r_request_pool); 15471b83bef2SSage Weil p = req->r_request_pgid; 15481b83bef2SSage Weil ceph_encode_64(&p, req->r_pgid.pool); 15491b83bef2SSage Weil ceph_encode_32(&p, req->r_pgid.seed); 15501b83bef2SSage Weil put_unaligned_le64(1, req->r_request_attempts); /* FIXME */ 15511b83bef2SSage Weil memcpy(req->r_request_reassert_version, &req->r_reassert_version, 15521b83bef2SSage Weil sizeof(req->r_reassert_version)); 15532169aea6SSage Weil 15543d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 15553d14c5d2SYehuda Sadeh list_move_tail(&req->r_req_lru_item, &osdc->req_lru); 15563d14c5d2SYehuda Sadeh 15573d14c5d2SYehuda Sadeh ceph_msg_get(req->r_request); /* send consumes a ref */ 155826be8808SAlex Elder 15593d14c5d2SYehuda Sadeh req->r_sent = req->r_osd->o_incarnation; 156026be8808SAlex Elder 156126be8808SAlex Elder ceph_con_send(&req->r_osd->o_con, req->r_request); 15623d14c5d2SYehuda Sadeh } 15633d14c5d2SYehuda Sadeh 15643d14c5d2SYehuda Sadeh /* 15656f6c7006SSage Weil * Send any requests in the queue (req_unsent). 15666f6c7006SSage Weil */ 1567f9d25199SAlex Elder static void __send_queued(struct ceph_osd_client *osdc) 15686f6c7006SSage Weil { 15696f6c7006SSage Weil struct ceph_osd_request *req, *tmp; 15706f6c7006SSage Weil 1571f9d25199SAlex Elder dout("__send_queued\n"); 1572f9d25199SAlex Elder list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) 15736f6c7006SSage Weil __send_request(osdc, req); 15746f6c7006SSage Weil } 15756f6c7006SSage Weil 15766f6c7006SSage Weil /* 15770bbfdfe8SIlya Dryomov * Caller should hold map_sem for read and request_mutex. 15780bbfdfe8SIlya Dryomov */ 15790bbfdfe8SIlya Dryomov static int __ceph_osdc_start_request(struct ceph_osd_client *osdc, 15800bbfdfe8SIlya Dryomov struct ceph_osd_request *req, 15810bbfdfe8SIlya Dryomov bool nofail) 15820bbfdfe8SIlya Dryomov { 15830bbfdfe8SIlya Dryomov int rc; 15840bbfdfe8SIlya Dryomov 15850bbfdfe8SIlya Dryomov __register_request(osdc, req); 15860bbfdfe8SIlya Dryomov req->r_sent = 0; 15870bbfdfe8SIlya Dryomov req->r_got_reply = 0; 15880bbfdfe8SIlya Dryomov rc = __map_request(osdc, req, 0); 15890bbfdfe8SIlya Dryomov if (rc < 0) { 15900bbfdfe8SIlya Dryomov if (nofail) { 15910bbfdfe8SIlya Dryomov dout("osdc_start_request failed map, " 15920bbfdfe8SIlya Dryomov " will retry %lld\n", req->r_tid); 15930bbfdfe8SIlya Dryomov rc = 0; 15940bbfdfe8SIlya Dryomov } else { 15950bbfdfe8SIlya Dryomov __unregister_request(osdc, req); 15960bbfdfe8SIlya Dryomov } 15970bbfdfe8SIlya Dryomov return rc; 15980bbfdfe8SIlya Dryomov } 15990bbfdfe8SIlya Dryomov 16000bbfdfe8SIlya Dryomov if (req->r_osd == NULL) { 16010bbfdfe8SIlya Dryomov dout("send_request %p no up osds in pg\n", req); 16020bbfdfe8SIlya Dryomov ceph_monc_request_next_osdmap(&osdc->client->monc); 16030bbfdfe8SIlya Dryomov } else { 16040bbfdfe8SIlya Dryomov __send_queued(osdc); 16050bbfdfe8SIlya Dryomov } 16060bbfdfe8SIlya Dryomov 16070bbfdfe8SIlya Dryomov return 0; 16080bbfdfe8SIlya Dryomov } 16090bbfdfe8SIlya Dryomov 16100bbfdfe8SIlya Dryomov /* 16113d14c5d2SYehuda Sadeh * Timeout callback, called every N seconds when 1 or more osd 16123d14c5d2SYehuda Sadeh * requests has been active for more than N seconds. When this 16133d14c5d2SYehuda Sadeh * happens, we ping all OSDs with requests who have timed out to 16143d14c5d2SYehuda Sadeh * ensure any communications channel reset is detected. Reset the 16153d14c5d2SYehuda Sadeh * request timeouts another N seconds in the future as we go. 16163d14c5d2SYehuda Sadeh * Reschedule the timeout event another N seconds in future (unless 16173d14c5d2SYehuda Sadeh * there are no open requests). 16183d14c5d2SYehuda Sadeh */ 16193d14c5d2SYehuda Sadeh static void handle_timeout(struct work_struct *work) 16203d14c5d2SYehuda Sadeh { 16213d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 16223d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, timeout_work.work); 1623a319bf56SIlya Dryomov struct ceph_options *opts = osdc->client->options; 162483aff95eSSage Weil struct ceph_osd_request *req; 16253d14c5d2SYehuda Sadeh struct ceph_osd *osd; 16263d14c5d2SYehuda Sadeh struct list_head slow_osds; 16273d14c5d2SYehuda Sadeh dout("timeout\n"); 16283d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 16293d14c5d2SYehuda Sadeh 16303d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 16313d14c5d2SYehuda Sadeh 16323d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 16333d14c5d2SYehuda Sadeh 16343d14c5d2SYehuda Sadeh /* 16353d14c5d2SYehuda Sadeh * ping osds that are a bit slow. this ensures that if there 16363d14c5d2SYehuda Sadeh * is a break in the TCP connection we will notice, and reopen 16373d14c5d2SYehuda Sadeh * a connection with that osd (from the fault callback). 16383d14c5d2SYehuda Sadeh */ 16393d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&slow_osds); 16403d14c5d2SYehuda Sadeh list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) { 1641a319bf56SIlya Dryomov if (time_before(jiffies, 1642a319bf56SIlya Dryomov req->r_stamp + opts->osd_keepalive_timeout)) 16433d14c5d2SYehuda Sadeh break; 16443d14c5d2SYehuda Sadeh 16453d14c5d2SYehuda Sadeh osd = req->r_osd; 16463d14c5d2SYehuda Sadeh BUG_ON(!osd); 16473d14c5d2SYehuda Sadeh dout(" tid %llu is slow, will send keepalive on osd%d\n", 16483d14c5d2SYehuda Sadeh req->r_tid, osd->o_osd); 16493d14c5d2SYehuda Sadeh list_move_tail(&osd->o_keepalive_item, &slow_osds); 16503d14c5d2SYehuda Sadeh } 16513d14c5d2SYehuda Sadeh while (!list_empty(&slow_osds)) { 16523d14c5d2SYehuda Sadeh osd = list_entry(slow_osds.next, struct ceph_osd, 16533d14c5d2SYehuda Sadeh o_keepalive_item); 16543d14c5d2SYehuda Sadeh list_del_init(&osd->o_keepalive_item); 16553d14c5d2SYehuda Sadeh ceph_con_keepalive(&osd->o_con); 16563d14c5d2SYehuda Sadeh } 16573d14c5d2SYehuda Sadeh 16583d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 1659f9d25199SAlex Elder __send_queued(osdc); 16603d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 16613d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 16623d14c5d2SYehuda Sadeh } 16633d14c5d2SYehuda Sadeh 16643d14c5d2SYehuda Sadeh static void handle_osds_timeout(struct work_struct *work) 16653d14c5d2SYehuda Sadeh { 16663d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 16673d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, 16683d14c5d2SYehuda Sadeh osds_timeout_work.work); 1669a319bf56SIlya Dryomov unsigned long delay = osdc->client->options->osd_idle_ttl / 4; 16703d14c5d2SYehuda Sadeh 16713d14c5d2SYehuda Sadeh dout("osds timeout\n"); 16723d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 1673aca420bcSSage Weil remove_old_osds(osdc); 16743d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 16753d14c5d2SYehuda Sadeh 16763d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 16773d14c5d2SYehuda Sadeh round_jiffies_relative(delay)); 16783d14c5d2SYehuda Sadeh } 16793d14c5d2SYehuda Sadeh 1680205ee118SIlya Dryomov static int ceph_oloc_decode(void **p, void *end, 1681205ee118SIlya Dryomov struct ceph_object_locator *oloc) 1682205ee118SIlya Dryomov { 1683205ee118SIlya Dryomov u8 struct_v, struct_cv; 1684205ee118SIlya Dryomov u32 len; 1685205ee118SIlya Dryomov void *struct_end; 1686205ee118SIlya Dryomov int ret = 0; 1687205ee118SIlya Dryomov 1688205ee118SIlya Dryomov ceph_decode_need(p, end, 1 + 1 + 4, e_inval); 1689205ee118SIlya Dryomov struct_v = ceph_decode_8(p); 1690205ee118SIlya Dryomov struct_cv = ceph_decode_8(p); 1691205ee118SIlya Dryomov if (struct_v < 3) { 1692205ee118SIlya Dryomov pr_warn("got v %d < 3 cv %d of ceph_object_locator\n", 1693205ee118SIlya Dryomov struct_v, struct_cv); 1694205ee118SIlya Dryomov goto e_inval; 1695205ee118SIlya Dryomov } 1696205ee118SIlya Dryomov if (struct_cv > 6) { 1697205ee118SIlya Dryomov pr_warn("got v %d cv %d > 6 of ceph_object_locator\n", 1698205ee118SIlya Dryomov struct_v, struct_cv); 1699205ee118SIlya Dryomov goto e_inval; 1700205ee118SIlya Dryomov } 1701205ee118SIlya Dryomov len = ceph_decode_32(p); 1702205ee118SIlya Dryomov ceph_decode_need(p, end, len, e_inval); 1703205ee118SIlya Dryomov struct_end = *p + len; 1704205ee118SIlya Dryomov 1705205ee118SIlya Dryomov oloc->pool = ceph_decode_64(p); 1706205ee118SIlya Dryomov *p += 4; /* skip preferred */ 1707205ee118SIlya Dryomov 1708205ee118SIlya Dryomov len = ceph_decode_32(p); 1709205ee118SIlya Dryomov if (len > 0) { 1710205ee118SIlya Dryomov pr_warn("ceph_object_locator::key is set\n"); 1711205ee118SIlya Dryomov goto e_inval; 1712205ee118SIlya Dryomov } 1713205ee118SIlya Dryomov 1714205ee118SIlya Dryomov if (struct_v >= 5) { 1715205ee118SIlya Dryomov len = ceph_decode_32(p); 1716205ee118SIlya Dryomov if (len > 0) { 1717205ee118SIlya Dryomov pr_warn("ceph_object_locator::nspace is set\n"); 1718205ee118SIlya Dryomov goto e_inval; 1719205ee118SIlya Dryomov } 1720205ee118SIlya Dryomov } 1721205ee118SIlya Dryomov 1722205ee118SIlya Dryomov if (struct_v >= 6) { 1723205ee118SIlya Dryomov s64 hash = ceph_decode_64(p); 1724205ee118SIlya Dryomov if (hash != -1) { 1725205ee118SIlya Dryomov pr_warn("ceph_object_locator::hash is set\n"); 1726205ee118SIlya Dryomov goto e_inval; 1727205ee118SIlya Dryomov } 1728205ee118SIlya Dryomov } 1729205ee118SIlya Dryomov 1730205ee118SIlya Dryomov /* skip the rest */ 1731205ee118SIlya Dryomov *p = struct_end; 1732205ee118SIlya Dryomov out: 1733205ee118SIlya Dryomov return ret; 1734205ee118SIlya Dryomov 1735205ee118SIlya Dryomov e_inval: 1736205ee118SIlya Dryomov ret = -EINVAL; 1737205ee118SIlya Dryomov goto out; 1738205ee118SIlya Dryomov } 1739205ee118SIlya Dryomov 1740205ee118SIlya Dryomov static int ceph_redirect_decode(void **p, void *end, 1741205ee118SIlya Dryomov struct ceph_request_redirect *redir) 1742205ee118SIlya Dryomov { 1743205ee118SIlya Dryomov u8 struct_v, struct_cv; 1744205ee118SIlya Dryomov u32 len; 1745205ee118SIlya Dryomov void *struct_end; 1746205ee118SIlya Dryomov int ret; 1747205ee118SIlya Dryomov 1748205ee118SIlya Dryomov ceph_decode_need(p, end, 1 + 1 + 4, e_inval); 1749205ee118SIlya Dryomov struct_v = ceph_decode_8(p); 1750205ee118SIlya Dryomov struct_cv = ceph_decode_8(p); 1751205ee118SIlya Dryomov if (struct_cv > 1) { 1752205ee118SIlya Dryomov pr_warn("got v %d cv %d > 1 of ceph_request_redirect\n", 1753205ee118SIlya Dryomov struct_v, struct_cv); 1754205ee118SIlya Dryomov goto e_inval; 1755205ee118SIlya Dryomov } 1756205ee118SIlya Dryomov len = ceph_decode_32(p); 1757205ee118SIlya Dryomov ceph_decode_need(p, end, len, e_inval); 1758205ee118SIlya Dryomov struct_end = *p + len; 1759205ee118SIlya Dryomov 1760205ee118SIlya Dryomov ret = ceph_oloc_decode(p, end, &redir->oloc); 1761205ee118SIlya Dryomov if (ret) 1762205ee118SIlya Dryomov goto out; 1763205ee118SIlya Dryomov 1764205ee118SIlya Dryomov len = ceph_decode_32(p); 1765205ee118SIlya Dryomov if (len > 0) { 1766205ee118SIlya Dryomov pr_warn("ceph_request_redirect::object_name is set\n"); 1767205ee118SIlya Dryomov goto e_inval; 1768205ee118SIlya Dryomov } 1769205ee118SIlya Dryomov 1770205ee118SIlya Dryomov len = ceph_decode_32(p); 1771205ee118SIlya Dryomov *p += len; /* skip osd_instructions */ 1772205ee118SIlya Dryomov 1773205ee118SIlya Dryomov /* skip the rest */ 1774205ee118SIlya Dryomov *p = struct_end; 1775205ee118SIlya Dryomov out: 1776205ee118SIlya Dryomov return ret; 1777205ee118SIlya Dryomov 1778205ee118SIlya Dryomov e_inval: 1779205ee118SIlya Dryomov ret = -EINVAL; 1780205ee118SIlya Dryomov goto out; 1781205ee118SIlya Dryomov } 1782205ee118SIlya Dryomov 178325845472SSage Weil static void complete_request(struct ceph_osd_request *req) 178425845472SSage Weil { 178525845472SSage Weil complete_all(&req->r_safe_completion); /* fsync waiter */ 178625845472SSage Weil } 178725845472SSage Weil 17883d14c5d2SYehuda Sadeh /* 17893d14c5d2SYehuda Sadeh * handle osd op reply. either call the callback if it is specified, 17903d14c5d2SYehuda Sadeh * or do the completion to wake up the waiting thread. 17913d14c5d2SYehuda Sadeh */ 179270cf052dSShraddha Barke static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg) 17933d14c5d2SYehuda Sadeh { 17941b83bef2SSage Weil void *p, *end; 17953d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 1796205ee118SIlya Dryomov struct ceph_request_redirect redir; 17973d14c5d2SYehuda Sadeh u64 tid; 17981b83bef2SSage Weil int object_len; 179979528734SAlex Elder unsigned int numops; 180079528734SAlex Elder int payload_len, flags; 18013d14c5d2SYehuda Sadeh s32 result; 18021b83bef2SSage Weil s32 retry_attempt; 18031b83bef2SSage Weil struct ceph_pg pg; 18041b83bef2SSage Weil int err; 18051b83bef2SSage Weil u32 reassert_epoch; 18061b83bef2SSage Weil u64 reassert_version; 18071b83bef2SSage Weil u32 osdmap_epoch; 18080d5af164SAlex Elder int already_completed; 18099fc6e064SAlex Elder u32 bytes; 1810b0b31a8fSIlya Dryomov u8 decode_redir; 181179528734SAlex Elder unsigned int i; 18123d14c5d2SYehuda Sadeh 18133d14c5d2SYehuda Sadeh tid = le64_to_cpu(msg->hdr.tid); 18141b83bef2SSage Weil dout("handle_reply %p tid %llu\n", msg, tid); 18151b83bef2SSage Weil 18161b83bef2SSage Weil p = msg->front.iov_base; 18171b83bef2SSage Weil end = p + msg->front.iov_len; 18181b83bef2SSage Weil 18191b83bef2SSage Weil ceph_decode_need(&p, end, 4, bad); 18201b83bef2SSage Weil object_len = ceph_decode_32(&p); 18211b83bef2SSage Weil ceph_decode_need(&p, end, object_len, bad); 18221b83bef2SSage Weil p += object_len; 18231b83bef2SSage Weil 1824ef4859d6SAlex Elder err = ceph_decode_pgid(&p, end, &pg); 18251b83bef2SSage Weil if (err) 18263d14c5d2SYehuda Sadeh goto bad; 18271b83bef2SSage Weil 18281b83bef2SSage Weil ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad); 18291b83bef2SSage Weil flags = ceph_decode_64(&p); 18301b83bef2SSage Weil result = ceph_decode_32(&p); 18311b83bef2SSage Weil reassert_epoch = ceph_decode_32(&p); 18321b83bef2SSage Weil reassert_version = ceph_decode_64(&p); 18331b83bef2SSage Weil osdmap_epoch = ceph_decode_32(&p); 18341b83bef2SSage Weil 18353d14c5d2SYehuda Sadeh /* lookup */ 1836ff513aceSIlya Dryomov down_read(&osdc->map_sem); 18373d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 18383d14c5d2SYehuda Sadeh req = __lookup_request(osdc, tid); 18393d14c5d2SYehuda Sadeh if (req == NULL) { 18403d14c5d2SYehuda Sadeh dout("handle_reply tid %llu dne\n", tid); 18418058fd45SAlex Elder goto bad_mutex; 18423d14c5d2SYehuda Sadeh } 18433d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 18441b83bef2SSage Weil 18451b83bef2SSage Weil dout("handle_reply %p tid %llu req %p result %d\n", msg, tid, 18461b83bef2SSage Weil req, result); 18471b83bef2SSage Weil 184818741196SDan Carpenter ceph_decode_need(&p, end, 4, bad_put); 18491b83bef2SSage Weil numops = ceph_decode_32(&p); 18503f1af42aSIlya Dryomov if (numops > CEPH_OSD_MAX_OPS) 18511b83bef2SSage Weil goto bad_put; 18521b83bef2SSage Weil if (numops != req->r_num_ops) 18531b83bef2SSage Weil goto bad_put; 18541b83bef2SSage Weil payload_len = 0; 185518741196SDan Carpenter ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad_put); 18561b83bef2SSage Weil for (i = 0; i < numops; i++) { 18571b83bef2SSage Weil struct ceph_osd_op *op = p; 18581b83bef2SSage Weil int len; 18591b83bef2SSage Weil 18601b83bef2SSage Weil len = le32_to_cpu(op->payload_len); 18617665d85bSYan, Zheng req->r_ops[i].outdata_len = len; 18621b83bef2SSage Weil dout(" op %d has %d bytes\n", i, len); 18631b83bef2SSage Weil payload_len += len; 18641b83bef2SSage Weil p += sizeof(*op); 18651b83bef2SSage Weil } 18669fc6e064SAlex Elder bytes = le32_to_cpu(msg->hdr.data_len); 18679fc6e064SAlex Elder if (payload_len != bytes) { 1868b9a67899SJoe Perches pr_warn("sum of op payload lens %d != data_len %d\n", 18699fc6e064SAlex Elder payload_len, bytes); 18701b83bef2SSage Weil goto bad_put; 18711b83bef2SSage Weil } 18721b83bef2SSage Weil 187318741196SDan Carpenter ceph_decode_need(&p, end, 4 + numops * 4, bad_put); 18741b83bef2SSage Weil retry_attempt = ceph_decode_32(&p); 18751b83bef2SSage Weil for (i = 0; i < numops; i++) 18767665d85bSYan, Zheng req->r_ops[i].rval = ceph_decode_32(&p); 18773d14c5d2SYehuda Sadeh 1878205ee118SIlya Dryomov if (le16_to_cpu(msg->hdr.version) >= 6) { 1879205ee118SIlya Dryomov p += 8 + 4; /* skip replay_version */ 1880205ee118SIlya Dryomov p += 8; /* skip user_version */ 1881205ee118SIlya Dryomov 1882b0b31a8fSIlya Dryomov if (le16_to_cpu(msg->hdr.version) >= 7) 1883b0b31a8fSIlya Dryomov ceph_decode_8_safe(&p, end, decode_redir, bad_put); 1884b0b31a8fSIlya Dryomov else 1885b0b31a8fSIlya Dryomov decode_redir = 1; 1886b0b31a8fSIlya Dryomov } else { 1887b0b31a8fSIlya Dryomov decode_redir = 0; 1888b0b31a8fSIlya Dryomov } 1889b0b31a8fSIlya Dryomov 1890b0b31a8fSIlya Dryomov if (decode_redir) { 1891205ee118SIlya Dryomov err = ceph_redirect_decode(&p, end, &redir); 1892205ee118SIlya Dryomov if (err) 1893205ee118SIlya Dryomov goto bad_put; 1894205ee118SIlya Dryomov } else { 1895205ee118SIlya Dryomov redir.oloc.pool = -1; 1896205ee118SIlya Dryomov } 1897205ee118SIlya Dryomov 1898205ee118SIlya Dryomov if (redir.oloc.pool != -1) { 1899205ee118SIlya Dryomov dout("redirect pool %lld\n", redir.oloc.pool); 1900205ee118SIlya Dryomov 1901205ee118SIlya Dryomov __unregister_request(osdc, req); 1902205ee118SIlya Dryomov 1903205ee118SIlya Dryomov req->r_target_oloc = redir.oloc; /* struct */ 1904205ee118SIlya Dryomov 1905205ee118SIlya Dryomov /* 1906205ee118SIlya Dryomov * Start redirect requests with nofail=true. If 1907205ee118SIlya Dryomov * mapping fails, request will end up on the notarget 1908205ee118SIlya Dryomov * list, waiting for the new osdmap (which can take 1909205ee118SIlya Dryomov * a while), even though the original request mapped 1910205ee118SIlya Dryomov * successfully. In the future we might want to follow 1911205ee118SIlya Dryomov * original request's nofail setting here. 1912205ee118SIlya Dryomov */ 1913ff513aceSIlya Dryomov err = __ceph_osdc_start_request(osdc, req, true); 1914205ee118SIlya Dryomov BUG_ON(err); 1915205ee118SIlya Dryomov 1916ff513aceSIlya Dryomov goto out_unlock; 1917205ee118SIlya Dryomov } 1918205ee118SIlya Dryomov 1919eb845ff1SYan, Zheng already_completed = req->r_got_reply; 19203d14c5d2SYehuda Sadeh if (!req->r_got_reply) { 19211b83bef2SSage Weil req->r_result = result; 19223d14c5d2SYehuda Sadeh dout("handle_reply result %d bytes %d\n", req->r_result, 19233d14c5d2SYehuda Sadeh bytes); 19243d14c5d2SYehuda Sadeh if (req->r_result == 0) 19253d14c5d2SYehuda Sadeh req->r_result = bytes; 19263d14c5d2SYehuda Sadeh 19273d14c5d2SYehuda Sadeh /* in case this is a write and we need to replay, */ 19281b83bef2SSage Weil req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch); 19291b83bef2SSage Weil req->r_reassert_version.version = cpu_to_le64(reassert_version); 19303d14c5d2SYehuda Sadeh 19313d14c5d2SYehuda Sadeh req->r_got_reply = 1; 19323d14c5d2SYehuda Sadeh } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) { 19333d14c5d2SYehuda Sadeh dout("handle_reply tid %llu dup ack\n", tid); 1934ff513aceSIlya Dryomov goto out_unlock; 19353d14c5d2SYehuda Sadeh } 19363d14c5d2SYehuda Sadeh 19373d14c5d2SYehuda Sadeh dout("handle_reply tid %llu flags %d\n", tid, flags); 19383d14c5d2SYehuda Sadeh 1939a40c4f10SYehuda Sadeh if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK)) 1940a40c4f10SYehuda Sadeh __register_linger_request(osdc, req); 1941a40c4f10SYehuda Sadeh 19423d14c5d2SYehuda Sadeh /* either this is a read, or we got the safe response */ 19433d14c5d2SYehuda Sadeh if (result < 0 || 19443d14c5d2SYehuda Sadeh (flags & CEPH_OSD_FLAG_ONDISK) || 19453d14c5d2SYehuda Sadeh ((flags & CEPH_OSD_FLAG_WRITE) == 0)) 19463d14c5d2SYehuda Sadeh __unregister_request(osdc, req); 19473d14c5d2SYehuda Sadeh 19483d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 1949ff513aceSIlya Dryomov up_read(&osdc->map_sem); 19503d14c5d2SYehuda Sadeh 1951eb845ff1SYan, Zheng if (!already_completed) { 195261c5d6bfSYan, Zheng if (req->r_unsafe_callback && 195361c5d6bfSYan, Zheng result >= 0 && !(flags & CEPH_OSD_FLAG_ONDISK)) 195461c5d6bfSYan, Zheng req->r_unsafe_callback(req, true); 19553d14c5d2SYehuda Sadeh if (req->r_callback) 19563d14c5d2SYehuda Sadeh req->r_callback(req, msg); 19573d14c5d2SYehuda Sadeh else 19583d14c5d2SYehuda Sadeh complete_all(&req->r_completion); 1959eb845ff1SYan, Zheng } 19603d14c5d2SYehuda Sadeh 196161c5d6bfSYan, Zheng if (flags & CEPH_OSD_FLAG_ONDISK) { 196261c5d6bfSYan, Zheng if (req->r_unsafe_callback && already_completed) 196361c5d6bfSYan, Zheng req->r_unsafe_callback(req, false); 196425845472SSage Weil complete_request(req); 196561c5d6bfSYan, Zheng } 19663d14c5d2SYehuda Sadeh 1967ff513aceSIlya Dryomov out: 1968a40c4f10SYehuda Sadeh dout("req=%p req->r_linger=%d\n", req, req->r_linger); 19693d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 19703d14c5d2SYehuda Sadeh return; 1971ff513aceSIlya Dryomov out_unlock: 1972ff513aceSIlya Dryomov mutex_unlock(&osdc->request_mutex); 1973ff513aceSIlya Dryomov up_read(&osdc->map_sem); 1974ff513aceSIlya Dryomov goto out; 19753d14c5d2SYehuda Sadeh 19761b83bef2SSage Weil bad_put: 197737c89bdeSLi Wang req->r_result = -EIO; 197837c89bdeSLi Wang __unregister_request(osdc, req); 197937c89bdeSLi Wang if (req->r_callback) 198037c89bdeSLi Wang req->r_callback(req, msg); 198137c89bdeSLi Wang else 198237c89bdeSLi Wang complete_all(&req->r_completion); 198337c89bdeSLi Wang complete_request(req); 19841b83bef2SSage Weil ceph_osdc_put_request(req); 19858058fd45SAlex Elder bad_mutex: 19868058fd45SAlex Elder mutex_unlock(&osdc->request_mutex); 1987ff513aceSIlya Dryomov up_read(&osdc->map_sem); 19883d14c5d2SYehuda Sadeh bad: 19891b83bef2SSage Weil pr_err("corrupt osd_op_reply got %d %d\n", 19901b83bef2SSage Weil (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len)); 19913d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 19923d14c5d2SYehuda Sadeh } 19933d14c5d2SYehuda Sadeh 19946f6c7006SSage Weil static void reset_changed_osds(struct ceph_osd_client *osdc) 19953d14c5d2SYehuda Sadeh { 19963d14c5d2SYehuda Sadeh struct rb_node *p, *n; 19973d14c5d2SYehuda Sadeh 19987eb71e03SIlya Dryomov dout("%s %p\n", __func__, osdc); 19993d14c5d2SYehuda Sadeh for (p = rb_first(&osdc->osds); p; p = n) { 20006f6c7006SSage Weil struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node); 20013d14c5d2SYehuda Sadeh 20023d14c5d2SYehuda Sadeh n = rb_next(p); 20033d14c5d2SYehuda Sadeh if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) || 20043d14c5d2SYehuda Sadeh memcmp(&osd->o_con.peer_addr, 20053d14c5d2SYehuda Sadeh ceph_osd_addr(osdc->osdmap, 20063d14c5d2SYehuda Sadeh osd->o_osd), 20073d14c5d2SYehuda Sadeh sizeof(struct ceph_entity_addr)) != 0) 20083d14c5d2SYehuda Sadeh __reset_osd(osdc, osd); 20093d14c5d2SYehuda Sadeh } 20103d14c5d2SYehuda Sadeh } 20113d14c5d2SYehuda Sadeh 20123d14c5d2SYehuda Sadeh /* 20136f6c7006SSage Weil * Requeue requests whose mapping to an OSD has changed. If requests map to 20146f6c7006SSage Weil * no osd, request a new map. 20153d14c5d2SYehuda Sadeh * 2016e6d50f67SAlex Elder * Caller should hold map_sem for read. 20173d14c5d2SYehuda Sadeh */ 20189a1ea2dbSJosh Durgin static void kick_requests(struct ceph_osd_client *osdc, bool force_resend, 20199a1ea2dbSJosh Durgin bool force_resend_writes) 20203d14c5d2SYehuda Sadeh { 2021a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 20226f6c7006SSage Weil struct rb_node *p; 20236f6c7006SSage Weil int needmap = 0; 20246f6c7006SSage Weil int err; 20259a1ea2dbSJosh Durgin bool force_resend_req; 20263d14c5d2SYehuda Sadeh 20279a1ea2dbSJosh Durgin dout("kick_requests %s %s\n", force_resend ? " (force resend)" : "", 20289a1ea2dbSJosh Durgin force_resend_writes ? " (force resend writes)" : ""); 20293d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 20306194ea89SSage Weil for (p = rb_first(&osdc->requests); p; ) { 20316f6c7006SSage Weil req = rb_entry(p, struct ceph_osd_request, r_node); 20326194ea89SSage Weil p = rb_next(p); 2033ab60b16dSAlex Elder 2034ab60b16dSAlex Elder /* 2035ab60b16dSAlex Elder * For linger requests that have not yet been 2036ab60b16dSAlex Elder * registered, move them to the linger list; they'll 2037ab60b16dSAlex Elder * be sent to the osd in the loop below. Unregister 2038ab60b16dSAlex Elder * the request before re-registering it as a linger 2039ab60b16dSAlex Elder * request to ensure the __map_request() below 2040ab60b16dSAlex Elder * will decide it needs to be sent. 2041ab60b16dSAlex Elder */ 2042ab60b16dSAlex Elder if (req->r_linger && list_empty(&req->r_linger_item)) { 2043ab60b16dSAlex Elder dout("%p tid %llu restart on osd%d\n", 2044ab60b16dSAlex Elder req, req->r_tid, 2045ab60b16dSAlex Elder req->r_osd ? req->r_osd->o_osd : -1); 204696e4dac6SAlex Elder ceph_osdc_get_request(req); 2047ab60b16dSAlex Elder __unregister_request(osdc, req); 2048ab60b16dSAlex Elder __register_linger_request(osdc, req); 204996e4dac6SAlex Elder ceph_osdc_put_request(req); 2050ab60b16dSAlex Elder continue; 2051ab60b16dSAlex Elder } 2052ab60b16dSAlex Elder 20539a1ea2dbSJosh Durgin force_resend_req = force_resend || 20549a1ea2dbSJosh Durgin (force_resend_writes && 20559a1ea2dbSJosh Durgin req->r_flags & CEPH_OSD_FLAG_WRITE); 20569a1ea2dbSJosh Durgin err = __map_request(osdc, req, force_resend_req); 20576f6c7006SSage Weil if (err < 0) 20586f6c7006SSage Weil continue; /* error */ 20596f6c7006SSage Weil if (req->r_osd == NULL) { 20606f6c7006SSage Weil dout("%p tid %llu maps to no osd\n", req, req->r_tid); 20616f6c7006SSage Weil needmap++; /* request a newer map */ 20626f6c7006SSage Weil } else if (err > 0) { 20636194ea89SSage Weil if (!req->r_linger) { 20646194ea89SSage Weil dout("%p tid %llu requeued on osd%d\n", req, 20656194ea89SSage Weil req->r_tid, 20666f6c7006SSage Weil req->r_osd ? req->r_osd->o_osd : -1); 20676f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 20686f6c7006SSage Weil } 20696f6c7006SSage Weil } 20706194ea89SSage Weil } 2071a40c4f10SYehuda Sadeh 2072a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osdc->req_linger, 2073a40c4f10SYehuda Sadeh r_linger_item) { 2074a40c4f10SYehuda Sadeh dout("linger req=%p req->r_osd=%p\n", req, req->r_osd); 2075a40c4f10SYehuda Sadeh 20769a1ea2dbSJosh Durgin err = __map_request(osdc, req, 20779a1ea2dbSJosh Durgin force_resend || force_resend_writes); 2078ab60b16dSAlex Elder dout("__map_request returned %d\n", err); 2079a40c4f10SYehuda Sadeh if (err < 0) 2080a40c4f10SYehuda Sadeh continue; /* hrm! */ 2081b0494532SIlya Dryomov if (req->r_osd == NULL || err > 0) { 2082a40c4f10SYehuda Sadeh if (req->r_osd == NULL) { 2083b0494532SIlya Dryomov dout("lingering %p tid %llu maps to no osd\n", 2084b0494532SIlya Dryomov req, req->r_tid); 2085b0494532SIlya Dryomov /* 2086b0494532SIlya Dryomov * A homeless lingering request makes 2087b0494532SIlya Dryomov * no sense, as it's job is to keep 2088b0494532SIlya Dryomov * a particular OSD connection open. 2089b0494532SIlya Dryomov * Request a newer map and kick the 2090b0494532SIlya Dryomov * request, knowing that it won't be 2091b0494532SIlya Dryomov * resent until we actually get a map 2092b0494532SIlya Dryomov * that can tell us where to send it. 2093b0494532SIlya Dryomov */ 2094b0494532SIlya Dryomov needmap++; 2095a40c4f10SYehuda Sadeh } 2096a40c4f10SYehuda Sadeh 2097b0494532SIlya Dryomov dout("kicking lingering %p tid %llu osd%d\n", req, 2098b0494532SIlya Dryomov req->r_tid, req->r_osd ? req->r_osd->o_osd : -1); 2099a40c4f10SYehuda Sadeh __register_request(osdc, req); 2100c89ce05eSAlex Elder __unregister_linger_request(osdc, req); 2101a40c4f10SYehuda Sadeh } 2102b0494532SIlya Dryomov } 210314d2f38dSAlex Elder reset_changed_osds(osdc); 21043d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 21053d14c5d2SYehuda Sadeh 21063d14c5d2SYehuda Sadeh if (needmap) { 21073d14c5d2SYehuda Sadeh dout("%d requests for down osds, need new map\n", needmap); 21083d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 21093d14c5d2SYehuda Sadeh } 21103d14c5d2SYehuda Sadeh } 21116f6c7006SSage Weil 21126f6c7006SSage Weil 21133d14c5d2SYehuda Sadeh /* 21143d14c5d2SYehuda Sadeh * Process updated osd map. 21153d14c5d2SYehuda Sadeh * 21163d14c5d2SYehuda Sadeh * The message contains any number of incremental and full maps, normally 21173d14c5d2SYehuda Sadeh * indicating some sort of topology change in the cluster. Kick requests 21183d14c5d2SYehuda Sadeh * off to different OSDs as needed. 21193d14c5d2SYehuda Sadeh */ 21203d14c5d2SYehuda Sadeh void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg) 21213d14c5d2SYehuda Sadeh { 21223d14c5d2SYehuda Sadeh void *p, *end, *next; 21233d14c5d2SYehuda Sadeh u32 nr_maps, maplen; 21243d14c5d2SYehuda Sadeh u32 epoch; 21253d14c5d2SYehuda Sadeh struct ceph_osdmap *newmap = NULL, *oldmap; 21263d14c5d2SYehuda Sadeh int err; 21273d14c5d2SYehuda Sadeh struct ceph_fsid fsid; 21289a1ea2dbSJosh Durgin bool was_full; 21293d14c5d2SYehuda Sadeh 21303d14c5d2SYehuda Sadeh dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0); 21313d14c5d2SYehuda Sadeh p = msg->front.iov_base; 21323d14c5d2SYehuda Sadeh end = p + msg->front.iov_len; 21333d14c5d2SYehuda Sadeh 21343d14c5d2SYehuda Sadeh /* verify fsid */ 21353d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, sizeof(fsid), bad); 21363d14c5d2SYehuda Sadeh ceph_decode_copy(&p, &fsid, sizeof(fsid)); 21373d14c5d2SYehuda Sadeh if (ceph_check_fsid(osdc->client, &fsid) < 0) 21383d14c5d2SYehuda Sadeh return; 21393d14c5d2SYehuda Sadeh 21403d14c5d2SYehuda Sadeh down_write(&osdc->map_sem); 21413d14c5d2SYehuda Sadeh 21429a1ea2dbSJosh Durgin was_full = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL); 21439a1ea2dbSJosh Durgin 21443d14c5d2SYehuda Sadeh /* incremental maps */ 21453d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 21463d14c5d2SYehuda Sadeh dout(" %d inc maps\n", nr_maps); 21473d14c5d2SYehuda Sadeh while (nr_maps > 0) { 21483d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 21493d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 21503d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 21513d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 21523d14c5d2SYehuda Sadeh next = p + maplen; 21533d14c5d2SYehuda Sadeh if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) { 21543d14c5d2SYehuda Sadeh dout("applying incremental map %u len %d\n", 21553d14c5d2SYehuda Sadeh epoch, maplen); 21563d14c5d2SYehuda Sadeh newmap = osdmap_apply_incremental(&p, next, 21573d14c5d2SYehuda Sadeh osdc->osdmap, 215815d9882cSAlex Elder &osdc->client->msgr); 21593d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 21603d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 21613d14c5d2SYehuda Sadeh goto bad; 21623d14c5d2SYehuda Sadeh } 21633d14c5d2SYehuda Sadeh BUG_ON(!newmap); 21643d14c5d2SYehuda Sadeh if (newmap != osdc->osdmap) { 21653d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 21663d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 21673d14c5d2SYehuda Sadeh } 21689a1ea2dbSJosh Durgin was_full = was_full || 21699a1ea2dbSJosh Durgin ceph_osdmap_flag(osdc->osdmap, 21709a1ea2dbSJosh Durgin CEPH_OSDMAP_FULL); 21719a1ea2dbSJosh Durgin kick_requests(osdc, 0, was_full); 21723d14c5d2SYehuda Sadeh } else { 21733d14c5d2SYehuda Sadeh dout("ignoring incremental map %u len %d\n", 21743d14c5d2SYehuda Sadeh epoch, maplen); 21753d14c5d2SYehuda Sadeh } 21763d14c5d2SYehuda Sadeh p = next; 21773d14c5d2SYehuda Sadeh nr_maps--; 21783d14c5d2SYehuda Sadeh } 21793d14c5d2SYehuda Sadeh if (newmap) 21803d14c5d2SYehuda Sadeh goto done; 21813d14c5d2SYehuda Sadeh 21823d14c5d2SYehuda Sadeh /* full maps */ 21833d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 21843d14c5d2SYehuda Sadeh dout(" %d full maps\n", nr_maps); 21853d14c5d2SYehuda Sadeh while (nr_maps) { 21863d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 21873d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 21883d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 21893d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 21903d14c5d2SYehuda Sadeh if (nr_maps > 1) { 21913d14c5d2SYehuda Sadeh dout("skipping non-latest full map %u len %d\n", 21923d14c5d2SYehuda Sadeh epoch, maplen); 21933d14c5d2SYehuda Sadeh } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) { 21943d14c5d2SYehuda Sadeh dout("skipping full map %u len %d, " 21953d14c5d2SYehuda Sadeh "older than our %u\n", epoch, maplen, 21963d14c5d2SYehuda Sadeh osdc->osdmap->epoch); 21973d14c5d2SYehuda Sadeh } else { 219838d6453cSSage Weil int skipped_map = 0; 219938d6453cSSage Weil 22003d14c5d2SYehuda Sadeh dout("taking full map %u len %d\n", epoch, maplen); 2201a2505d63SIlya Dryomov newmap = ceph_osdmap_decode(&p, p+maplen); 22023d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 22033d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 22043d14c5d2SYehuda Sadeh goto bad; 22053d14c5d2SYehuda Sadeh } 22063d14c5d2SYehuda Sadeh BUG_ON(!newmap); 22073d14c5d2SYehuda Sadeh oldmap = osdc->osdmap; 22083d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 220938d6453cSSage Weil if (oldmap) { 221038d6453cSSage Weil if (oldmap->epoch + 1 < newmap->epoch) 221138d6453cSSage Weil skipped_map = 1; 22123d14c5d2SYehuda Sadeh ceph_osdmap_destroy(oldmap); 221338d6453cSSage Weil } 22149a1ea2dbSJosh Durgin was_full = was_full || 22159a1ea2dbSJosh Durgin ceph_osdmap_flag(osdc->osdmap, 22169a1ea2dbSJosh Durgin CEPH_OSDMAP_FULL); 22179a1ea2dbSJosh Durgin kick_requests(osdc, skipped_map, was_full); 22183d14c5d2SYehuda Sadeh } 22193d14c5d2SYehuda Sadeh p += maplen; 22203d14c5d2SYehuda Sadeh nr_maps--; 22213d14c5d2SYehuda Sadeh } 22223d14c5d2SYehuda Sadeh 2223b72e19b9SDan Carpenter if (!osdc->osdmap) 2224b72e19b9SDan Carpenter goto bad; 22253d14c5d2SYehuda Sadeh done: 22263d14c5d2SYehuda Sadeh downgrade_write(&osdc->map_sem); 222782dcabadSIlya Dryomov ceph_monc_got_map(&osdc->client->monc, CEPH_SUB_OSDMAP, 222882dcabadSIlya Dryomov osdc->osdmap->epoch); 2229cd634fb6SSage Weil 2230cd634fb6SSage Weil /* 2231cd634fb6SSage Weil * subscribe to subsequent osdmap updates if full to ensure 2232cd634fb6SSage Weil * we find out when we are no longer full and stop returning 2233cd634fb6SSage Weil * ENOSPC. 2234cd634fb6SSage Weil */ 2235d29adb34SJosh Durgin if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) || 2236d29adb34SJosh Durgin ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD) || 2237d29adb34SJosh Durgin ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR)) 2238cd634fb6SSage Weil ceph_monc_request_next_osdmap(&osdc->client->monc); 2239cd634fb6SSage Weil 2240f9d25199SAlex Elder mutex_lock(&osdc->request_mutex); 2241f9d25199SAlex Elder __send_queued(osdc); 2242f9d25199SAlex Elder mutex_unlock(&osdc->request_mutex); 22433d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 22443d14c5d2SYehuda Sadeh wake_up_all(&osdc->client->auth_wq); 22453d14c5d2SYehuda Sadeh return; 22463d14c5d2SYehuda Sadeh 22473d14c5d2SYehuda Sadeh bad: 22483d14c5d2SYehuda Sadeh pr_err("osdc handle_map corrupt msg\n"); 22493d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 22503d14c5d2SYehuda Sadeh up_write(&osdc->map_sem); 22513d14c5d2SYehuda Sadeh } 22523d14c5d2SYehuda Sadeh 22533d14c5d2SYehuda Sadeh /* 2254a40c4f10SYehuda Sadeh * watch/notify callback event infrastructure 2255a40c4f10SYehuda Sadeh * 2256a40c4f10SYehuda Sadeh * These callbacks are used both for watch and notify operations. 2257a40c4f10SYehuda Sadeh */ 2258a40c4f10SYehuda Sadeh static void __release_event(struct kref *kref) 2259a40c4f10SYehuda Sadeh { 2260a40c4f10SYehuda Sadeh struct ceph_osd_event *event = 2261a40c4f10SYehuda Sadeh container_of(kref, struct ceph_osd_event, kref); 2262a40c4f10SYehuda Sadeh 2263a40c4f10SYehuda Sadeh dout("__release_event %p\n", event); 2264a40c4f10SYehuda Sadeh kfree(event); 2265a40c4f10SYehuda Sadeh } 2266a40c4f10SYehuda Sadeh 2267a40c4f10SYehuda Sadeh static void get_event(struct ceph_osd_event *event) 2268a40c4f10SYehuda Sadeh { 2269a40c4f10SYehuda Sadeh kref_get(&event->kref); 2270a40c4f10SYehuda Sadeh } 2271a40c4f10SYehuda Sadeh 2272a40c4f10SYehuda Sadeh void ceph_osdc_put_event(struct ceph_osd_event *event) 2273a40c4f10SYehuda Sadeh { 2274a40c4f10SYehuda Sadeh kref_put(&event->kref, __release_event); 2275a40c4f10SYehuda Sadeh } 2276a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_put_event); 2277a40c4f10SYehuda Sadeh 2278a40c4f10SYehuda Sadeh static void __insert_event(struct ceph_osd_client *osdc, 2279a40c4f10SYehuda Sadeh struct ceph_osd_event *new) 2280a40c4f10SYehuda Sadeh { 2281a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 2282a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 2283a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 2284a40c4f10SYehuda Sadeh 2285a40c4f10SYehuda Sadeh while (*p) { 2286a40c4f10SYehuda Sadeh parent = *p; 2287a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 2288a40c4f10SYehuda Sadeh if (new->cookie < event->cookie) 2289a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 2290a40c4f10SYehuda Sadeh else if (new->cookie > event->cookie) 2291a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 2292a40c4f10SYehuda Sadeh else 2293a40c4f10SYehuda Sadeh BUG(); 2294a40c4f10SYehuda Sadeh } 2295a40c4f10SYehuda Sadeh 2296a40c4f10SYehuda Sadeh rb_link_node(&new->node, parent, p); 2297a40c4f10SYehuda Sadeh rb_insert_color(&new->node, &osdc->event_tree); 2298a40c4f10SYehuda Sadeh } 2299a40c4f10SYehuda Sadeh 2300a40c4f10SYehuda Sadeh static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc, 2301a40c4f10SYehuda Sadeh u64 cookie) 2302a40c4f10SYehuda Sadeh { 2303a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 2304a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 2305a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 2306a40c4f10SYehuda Sadeh 2307a40c4f10SYehuda Sadeh while (*p) { 2308a40c4f10SYehuda Sadeh parent = *p; 2309a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 2310a40c4f10SYehuda Sadeh if (cookie < event->cookie) 2311a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 2312a40c4f10SYehuda Sadeh else if (cookie > event->cookie) 2313a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 2314a40c4f10SYehuda Sadeh else 2315a40c4f10SYehuda Sadeh return event; 2316a40c4f10SYehuda Sadeh } 2317a40c4f10SYehuda Sadeh return NULL; 2318a40c4f10SYehuda Sadeh } 2319a40c4f10SYehuda Sadeh 2320a40c4f10SYehuda Sadeh static void __remove_event(struct ceph_osd_event *event) 2321a40c4f10SYehuda Sadeh { 2322a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 2323a40c4f10SYehuda Sadeh 2324a40c4f10SYehuda Sadeh if (!RB_EMPTY_NODE(&event->node)) { 2325a40c4f10SYehuda Sadeh dout("__remove_event removed %p\n", event); 2326a40c4f10SYehuda Sadeh rb_erase(&event->node, &osdc->event_tree); 2327a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 2328a40c4f10SYehuda Sadeh } else { 2329a40c4f10SYehuda Sadeh dout("__remove_event didn't remove %p\n", event); 2330a40c4f10SYehuda Sadeh } 2331a40c4f10SYehuda Sadeh } 2332a40c4f10SYehuda Sadeh 2333a40c4f10SYehuda Sadeh int ceph_osdc_create_event(struct ceph_osd_client *osdc, 2334a40c4f10SYehuda Sadeh void (*event_cb)(u64, u64, u8, void *), 23353c663bbdSAlex Elder void *data, struct ceph_osd_event **pevent) 2336a40c4f10SYehuda Sadeh { 2337a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 2338a40c4f10SYehuda Sadeh 2339a40c4f10SYehuda Sadeh event = kmalloc(sizeof(*event), GFP_NOIO); 2340a40c4f10SYehuda Sadeh if (!event) 2341a40c4f10SYehuda Sadeh return -ENOMEM; 2342a40c4f10SYehuda Sadeh 2343a40c4f10SYehuda Sadeh dout("create_event %p\n", event); 2344a40c4f10SYehuda Sadeh event->cb = event_cb; 23453c663bbdSAlex Elder event->one_shot = 0; 2346a40c4f10SYehuda Sadeh event->data = data; 2347a40c4f10SYehuda Sadeh event->osdc = osdc; 2348a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&event->osd_node); 23493ee5234dSAlex Elder RB_CLEAR_NODE(&event->node); 2350a40c4f10SYehuda Sadeh kref_init(&event->kref); /* one ref for us */ 2351a40c4f10SYehuda Sadeh kref_get(&event->kref); /* one ref for the caller */ 2352a40c4f10SYehuda Sadeh 2353a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 2354a40c4f10SYehuda Sadeh event->cookie = ++osdc->event_count; 2355a40c4f10SYehuda Sadeh __insert_event(osdc, event); 2356a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 2357a40c4f10SYehuda Sadeh 2358a40c4f10SYehuda Sadeh *pevent = event; 2359a40c4f10SYehuda Sadeh return 0; 2360a40c4f10SYehuda Sadeh } 2361a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_create_event); 2362a40c4f10SYehuda Sadeh 2363a40c4f10SYehuda Sadeh void ceph_osdc_cancel_event(struct ceph_osd_event *event) 2364a40c4f10SYehuda Sadeh { 2365a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 2366a40c4f10SYehuda Sadeh 2367a40c4f10SYehuda Sadeh dout("cancel_event %p\n", event); 2368a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 2369a40c4f10SYehuda Sadeh __remove_event(event); 2370a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 2371a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); /* caller's */ 2372a40c4f10SYehuda Sadeh } 2373a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_cancel_event); 2374a40c4f10SYehuda Sadeh 2375a40c4f10SYehuda Sadeh 2376a40c4f10SYehuda Sadeh static void do_event_work(struct work_struct *work) 2377a40c4f10SYehuda Sadeh { 2378a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work = 2379a40c4f10SYehuda Sadeh container_of(work, struct ceph_osd_event_work, work); 2380a40c4f10SYehuda Sadeh struct ceph_osd_event *event = event_work->event; 2381a40c4f10SYehuda Sadeh u64 ver = event_work->ver; 2382a40c4f10SYehuda Sadeh u64 notify_id = event_work->notify_id; 2383a40c4f10SYehuda Sadeh u8 opcode = event_work->opcode; 2384a40c4f10SYehuda Sadeh 2385a40c4f10SYehuda Sadeh dout("do_event_work completing %p\n", event); 2386a40c4f10SYehuda Sadeh event->cb(ver, notify_id, opcode, event->data); 2387a40c4f10SYehuda Sadeh dout("do_event_work completed %p\n", event); 2388a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 2389a40c4f10SYehuda Sadeh kfree(event_work); 2390a40c4f10SYehuda Sadeh } 2391a40c4f10SYehuda Sadeh 2392a40c4f10SYehuda Sadeh 2393a40c4f10SYehuda Sadeh /* 2394a40c4f10SYehuda Sadeh * Process osd watch notifications 2395a40c4f10SYehuda Sadeh */ 23963c663bbdSAlex Elder static void handle_watch_notify(struct ceph_osd_client *osdc, 23973c663bbdSAlex Elder struct ceph_msg *msg) 2398a40c4f10SYehuda Sadeh { 2399a40c4f10SYehuda Sadeh void *p, *end; 2400a40c4f10SYehuda Sadeh u8 proto_ver; 2401a40c4f10SYehuda Sadeh u64 cookie, ver, notify_id; 2402a40c4f10SYehuda Sadeh u8 opcode; 2403a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 2404a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work; 2405a40c4f10SYehuda Sadeh 2406a40c4f10SYehuda Sadeh p = msg->front.iov_base; 2407a40c4f10SYehuda Sadeh end = p + msg->front.iov_len; 2408a40c4f10SYehuda Sadeh 2409a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, proto_ver, bad); 2410a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, opcode, bad); 2411a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, cookie, bad); 2412a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, ver, bad); 2413a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, notify_id, bad); 2414a40c4f10SYehuda Sadeh 2415a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 2416a40c4f10SYehuda Sadeh event = __find_event(osdc, cookie); 2417a40c4f10SYehuda Sadeh if (event) { 24183c663bbdSAlex Elder BUG_ON(event->one_shot); 2419a40c4f10SYehuda Sadeh get_event(event); 2420a40c4f10SYehuda Sadeh } 2421a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 2422a40c4f10SYehuda Sadeh dout("handle_watch_notify cookie %lld ver %lld event %p\n", 2423a40c4f10SYehuda Sadeh cookie, ver, event); 2424a40c4f10SYehuda Sadeh if (event) { 2425a40c4f10SYehuda Sadeh event_work = kmalloc(sizeof(*event_work), GFP_NOIO); 2426a40c4f10SYehuda Sadeh if (!event_work) { 242791883cd2SIlya Dryomov pr_err("couldn't allocate event_work\n"); 242891883cd2SIlya Dryomov ceph_osdc_put_event(event); 242991883cd2SIlya Dryomov return; 2430a40c4f10SYehuda Sadeh } 24316b0ae409SMariusz Kozlowski INIT_WORK(&event_work->work, do_event_work); 2432a40c4f10SYehuda Sadeh event_work->event = event; 2433a40c4f10SYehuda Sadeh event_work->ver = ver; 2434a40c4f10SYehuda Sadeh event_work->notify_id = notify_id; 2435a40c4f10SYehuda Sadeh event_work->opcode = opcode; 243691883cd2SIlya Dryomov 243791883cd2SIlya Dryomov queue_work(osdc->notify_wq, &event_work->work); 2438a40c4f10SYehuda Sadeh } 2439a40c4f10SYehuda Sadeh 2440a40c4f10SYehuda Sadeh return; 2441a40c4f10SYehuda Sadeh 2442a40c4f10SYehuda Sadeh bad: 2443a40c4f10SYehuda Sadeh pr_err("osdc handle_watch_notify corrupt msg\n"); 2444a40c4f10SYehuda Sadeh } 2445a40c4f10SYehuda Sadeh 2446a40c4f10SYehuda Sadeh /* 2447e65550fdSAlex Elder * build new request AND message 2448e65550fdSAlex Elder * 2449e65550fdSAlex Elder */ 2450e65550fdSAlex Elder void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off, 2451e65550fdSAlex Elder struct ceph_snap_context *snapc, u64 snap_id, 2452e65550fdSAlex Elder struct timespec *mtime) 2453e65550fdSAlex Elder { 2454e65550fdSAlex Elder struct ceph_msg *msg = req->r_request; 2455e65550fdSAlex Elder void *p; 2456e65550fdSAlex Elder size_t msg_size; 2457e65550fdSAlex Elder int flags = req->r_flags; 2458e65550fdSAlex Elder u64 data_len; 2459e65550fdSAlex Elder unsigned int i; 2460e65550fdSAlex Elder 2461e65550fdSAlex Elder req->r_snapid = snap_id; 246284127282SIlya Dryomov WARN_ON(snapc != req->r_snapc); 2463e65550fdSAlex Elder 2464e65550fdSAlex Elder /* encode request */ 2465e65550fdSAlex Elder msg->hdr.version = cpu_to_le16(4); 2466e65550fdSAlex Elder 2467e65550fdSAlex Elder p = msg->front.iov_base; 2468e65550fdSAlex Elder ceph_encode_32(&p, 1); /* client_inc is always 1 */ 2469e65550fdSAlex Elder req->r_request_osdmap_epoch = p; 2470e65550fdSAlex Elder p += 4; 2471e65550fdSAlex Elder req->r_request_flags = p; 2472e65550fdSAlex Elder p += 4; 2473e65550fdSAlex Elder if (req->r_flags & CEPH_OSD_FLAG_WRITE) 2474e65550fdSAlex Elder ceph_encode_timespec(p, mtime); 2475e65550fdSAlex Elder p += sizeof(struct ceph_timespec); 2476e65550fdSAlex Elder req->r_request_reassert_version = p; 2477e65550fdSAlex Elder p += sizeof(struct ceph_eversion); /* will get filled in */ 2478e65550fdSAlex Elder 2479e65550fdSAlex Elder /* oloc */ 2480e65550fdSAlex Elder ceph_encode_8(&p, 4); 2481e65550fdSAlex Elder ceph_encode_8(&p, 4); 2482e65550fdSAlex Elder ceph_encode_32(&p, 8 + 4 + 4); 2483e65550fdSAlex Elder req->r_request_pool = p; 2484e65550fdSAlex Elder p += 8; 2485e65550fdSAlex Elder ceph_encode_32(&p, -1); /* preferred */ 2486e65550fdSAlex Elder ceph_encode_32(&p, 0); /* key len */ 2487e65550fdSAlex Elder 2488e65550fdSAlex Elder ceph_encode_8(&p, 1); 2489e65550fdSAlex Elder req->r_request_pgid = p; 2490e65550fdSAlex Elder p += 8 + 4; 2491e65550fdSAlex Elder ceph_encode_32(&p, -1); /* preferred */ 2492e65550fdSAlex Elder 2493e65550fdSAlex Elder /* oid */ 24943c972c95SIlya Dryomov ceph_encode_32(&p, req->r_base_oid.name_len); 24953c972c95SIlya Dryomov memcpy(p, req->r_base_oid.name, req->r_base_oid.name_len); 24963c972c95SIlya Dryomov dout("oid '%.*s' len %d\n", req->r_base_oid.name_len, 24973c972c95SIlya Dryomov req->r_base_oid.name, req->r_base_oid.name_len); 24983c972c95SIlya Dryomov p += req->r_base_oid.name_len; 2499e65550fdSAlex Elder 2500e65550fdSAlex Elder /* ops--can imply data */ 2501e65550fdSAlex Elder ceph_encode_16(&p, (u16)req->r_num_ops); 2502e65550fdSAlex Elder data_len = 0; 2503e65550fdSAlex Elder for (i = 0; i < req->r_num_ops; i++) { 2504e65550fdSAlex Elder data_len += osd_req_encode_op(req, p, i); 2505e65550fdSAlex Elder p += sizeof(struct ceph_osd_op); 2506e65550fdSAlex Elder } 2507e65550fdSAlex Elder 2508e65550fdSAlex Elder /* snaps */ 2509e65550fdSAlex Elder ceph_encode_64(&p, req->r_snapid); 2510e65550fdSAlex Elder ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0); 2511e65550fdSAlex Elder ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0); 2512e65550fdSAlex Elder if (req->r_snapc) { 251384127282SIlya Dryomov for (i = 0; i < req->r_snapc->num_snaps; i++) { 2514e65550fdSAlex Elder ceph_encode_64(&p, req->r_snapc->snaps[i]); 2515e65550fdSAlex Elder } 2516e65550fdSAlex Elder } 2517e65550fdSAlex Elder 2518e65550fdSAlex Elder req->r_request_attempts = p; 2519e65550fdSAlex Elder p += 4; 2520e65550fdSAlex Elder 2521e65550fdSAlex Elder /* data */ 2522e65550fdSAlex Elder if (flags & CEPH_OSD_FLAG_WRITE) { 2523e65550fdSAlex Elder u16 data_off; 2524e65550fdSAlex Elder 2525e65550fdSAlex Elder /* 2526e65550fdSAlex Elder * The header "data_off" is a hint to the receiver 2527e65550fdSAlex Elder * allowing it to align received data into its 2528e65550fdSAlex Elder * buffers such that there's no need to re-copy 2529e65550fdSAlex Elder * it before writing it to disk (direct I/O). 2530e65550fdSAlex Elder */ 2531e65550fdSAlex Elder data_off = (u16) (off & 0xffff); 2532e65550fdSAlex Elder req->r_request->hdr.data_off = cpu_to_le16(data_off); 2533e65550fdSAlex Elder } 2534e65550fdSAlex Elder req->r_request->hdr.data_len = cpu_to_le32(data_len); 2535e65550fdSAlex Elder 2536e65550fdSAlex Elder BUG_ON(p > msg->front.iov_base + msg->front.iov_len); 2537e65550fdSAlex Elder msg_size = p - msg->front.iov_base; 2538e65550fdSAlex Elder msg->front.iov_len = msg_size; 2539e65550fdSAlex Elder msg->hdr.front_len = cpu_to_le32(msg_size); 2540e65550fdSAlex Elder 2541e65550fdSAlex Elder dout("build_request msg_size was %d\n", (int)msg_size); 2542e65550fdSAlex Elder } 2543e65550fdSAlex Elder EXPORT_SYMBOL(ceph_osdc_build_request); 2544e65550fdSAlex Elder 2545e65550fdSAlex Elder /* 25463d14c5d2SYehuda Sadeh * Register request, send initial attempt. 25473d14c5d2SYehuda Sadeh */ 25483d14c5d2SYehuda Sadeh int ceph_osdc_start_request(struct ceph_osd_client *osdc, 25493d14c5d2SYehuda Sadeh struct ceph_osd_request *req, 25503d14c5d2SYehuda Sadeh bool nofail) 25513d14c5d2SYehuda Sadeh { 25520bbfdfe8SIlya Dryomov int rc; 25533d14c5d2SYehuda Sadeh 25543d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 25553d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 25560bbfdfe8SIlya Dryomov 25570bbfdfe8SIlya Dryomov rc = __ceph_osdc_start_request(osdc, req, nofail); 25580bbfdfe8SIlya Dryomov 25593d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 25603d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 25610bbfdfe8SIlya Dryomov 25623d14c5d2SYehuda Sadeh return rc; 25633d14c5d2SYehuda Sadeh } 25643d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_start_request); 25653d14c5d2SYehuda Sadeh 25663d14c5d2SYehuda Sadeh /* 2567c9f9b93dSIlya Dryomov * Unregister a registered request. The request is not completed (i.e. 2568c9f9b93dSIlya Dryomov * no callbacks or wakeups) - higher layers are supposed to know what 2569c9f9b93dSIlya Dryomov * they are canceling. 2570c9f9b93dSIlya Dryomov */ 2571c9f9b93dSIlya Dryomov void ceph_osdc_cancel_request(struct ceph_osd_request *req) 2572c9f9b93dSIlya Dryomov { 2573c9f9b93dSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 2574c9f9b93dSIlya Dryomov 2575c9f9b93dSIlya Dryomov mutex_lock(&osdc->request_mutex); 2576c9f9b93dSIlya Dryomov if (req->r_linger) 2577c9f9b93dSIlya Dryomov __unregister_linger_request(osdc, req); 2578c9f9b93dSIlya Dryomov __unregister_request(osdc, req); 2579c9f9b93dSIlya Dryomov mutex_unlock(&osdc->request_mutex); 2580c9f9b93dSIlya Dryomov 2581c9f9b93dSIlya Dryomov dout("%s %p tid %llu canceled\n", __func__, req, req->r_tid); 2582c9f9b93dSIlya Dryomov } 2583c9f9b93dSIlya Dryomov EXPORT_SYMBOL(ceph_osdc_cancel_request); 2584c9f9b93dSIlya Dryomov 2585c9f9b93dSIlya Dryomov /* 25863d14c5d2SYehuda Sadeh * wait for a request to complete 25873d14c5d2SYehuda Sadeh */ 25883d14c5d2SYehuda Sadeh int ceph_osdc_wait_request(struct ceph_osd_client *osdc, 25893d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 25903d14c5d2SYehuda Sadeh { 25913d14c5d2SYehuda Sadeh int rc; 25923d14c5d2SYehuda Sadeh 2593c9f9b93dSIlya Dryomov dout("%s %p tid %llu\n", __func__, req, req->r_tid); 2594c9f9b93dSIlya Dryomov 25953d14c5d2SYehuda Sadeh rc = wait_for_completion_interruptible(&req->r_completion); 25963d14c5d2SYehuda Sadeh if (rc < 0) { 2597c9f9b93dSIlya Dryomov dout("%s %p tid %llu interrupted\n", __func__, req, req->r_tid); 2598c9f9b93dSIlya Dryomov ceph_osdc_cancel_request(req); 259925845472SSage Weil complete_request(req); 26003d14c5d2SYehuda Sadeh return rc; 26013d14c5d2SYehuda Sadeh } 26023d14c5d2SYehuda Sadeh 2603c9f9b93dSIlya Dryomov dout("%s %p tid %llu result %d\n", __func__, req, req->r_tid, 2604c9f9b93dSIlya Dryomov req->r_result); 26053d14c5d2SYehuda Sadeh return req->r_result; 26063d14c5d2SYehuda Sadeh } 26073d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_wait_request); 26083d14c5d2SYehuda Sadeh 26093d14c5d2SYehuda Sadeh /* 26103d14c5d2SYehuda Sadeh * sync - wait for all in-flight requests to flush. avoid starvation. 26113d14c5d2SYehuda Sadeh */ 26123d14c5d2SYehuda Sadeh void ceph_osdc_sync(struct ceph_osd_client *osdc) 26133d14c5d2SYehuda Sadeh { 26143d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 26153d14c5d2SYehuda Sadeh u64 last_tid, next_tid = 0; 26163d14c5d2SYehuda Sadeh 26173d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 26183d14c5d2SYehuda Sadeh last_tid = osdc->last_tid; 26193d14c5d2SYehuda Sadeh while (1) { 26203d14c5d2SYehuda Sadeh req = __lookup_request_ge(osdc, next_tid); 26213d14c5d2SYehuda Sadeh if (!req) 26223d14c5d2SYehuda Sadeh break; 26233d14c5d2SYehuda Sadeh if (req->r_tid > last_tid) 26243d14c5d2SYehuda Sadeh break; 26253d14c5d2SYehuda Sadeh 26263d14c5d2SYehuda Sadeh next_tid = req->r_tid + 1; 26273d14c5d2SYehuda Sadeh if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0) 26283d14c5d2SYehuda Sadeh continue; 26293d14c5d2SYehuda Sadeh 26303d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 26313d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 26323d14c5d2SYehuda Sadeh dout("sync waiting on tid %llu (last is %llu)\n", 26333d14c5d2SYehuda Sadeh req->r_tid, last_tid); 26343d14c5d2SYehuda Sadeh wait_for_completion(&req->r_safe_completion); 26353d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 26363d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 26373d14c5d2SYehuda Sadeh } 26383d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 26393d14c5d2SYehuda Sadeh dout("sync done (thru tid %llu)\n", last_tid); 26403d14c5d2SYehuda Sadeh } 26413d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_sync); 26423d14c5d2SYehuda Sadeh 26433d14c5d2SYehuda Sadeh /* 2644dd935f44SJosh Durgin * Call all pending notify callbacks - for use after a watch is 2645dd935f44SJosh Durgin * unregistered, to make sure no more callbacks for it will be invoked 2646dd935f44SJosh Durgin */ 2647f6479449Sstephen hemminger void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc) 2648dd935f44SJosh Durgin { 2649dd935f44SJosh Durgin flush_workqueue(osdc->notify_wq); 2650dd935f44SJosh Durgin } 2651dd935f44SJosh Durgin EXPORT_SYMBOL(ceph_osdc_flush_notifies); 2652dd935f44SJosh Durgin 2653dd935f44SJosh Durgin 2654dd935f44SJosh Durgin /* 26553d14c5d2SYehuda Sadeh * init, shutdown 26563d14c5d2SYehuda Sadeh */ 26573d14c5d2SYehuda Sadeh int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client) 26583d14c5d2SYehuda Sadeh { 26593d14c5d2SYehuda Sadeh int err; 26603d14c5d2SYehuda Sadeh 26613d14c5d2SYehuda Sadeh dout("init\n"); 26623d14c5d2SYehuda Sadeh osdc->client = client; 26633d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 26643d14c5d2SYehuda Sadeh init_rwsem(&osdc->map_sem); 26653d14c5d2SYehuda Sadeh init_completion(&osdc->map_waiters); 26663d14c5d2SYehuda Sadeh osdc->last_requested_map = 0; 26673d14c5d2SYehuda Sadeh mutex_init(&osdc->request_mutex); 26683d14c5d2SYehuda Sadeh osdc->last_tid = 0; 26693d14c5d2SYehuda Sadeh osdc->osds = RB_ROOT; 26703d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->osd_lru); 26713d14c5d2SYehuda Sadeh osdc->requests = RB_ROOT; 26723d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_lru); 26736f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_unsent); 26746f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_notarget); 2675a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_linger); 26763d14c5d2SYehuda Sadeh osdc->num_requests = 0; 26773d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout); 26783d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout); 2679a40c4f10SYehuda Sadeh spin_lock_init(&osdc->event_lock); 2680a40c4f10SYehuda Sadeh osdc->event_tree = RB_ROOT; 2681a40c4f10SYehuda Sadeh osdc->event_count = 0; 26823d14c5d2SYehuda Sadeh 26833d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 2684a319bf56SIlya Dryomov round_jiffies_relative(osdc->client->options->osd_idle_ttl)); 26853d14c5d2SYehuda Sadeh 26863d14c5d2SYehuda Sadeh err = -ENOMEM; 26879e767adbSIlya Dryomov osdc->req_mempool = mempool_create_slab_pool(10, 26889e767adbSIlya Dryomov ceph_osd_request_cache); 26893d14c5d2SYehuda Sadeh if (!osdc->req_mempool) 26903d14c5d2SYehuda Sadeh goto out; 26913d14c5d2SYehuda Sadeh 2692d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP, 2693d50b409fSSage Weil OSD_OP_FRONT_LEN, 10, true, 26943d14c5d2SYehuda Sadeh "osd_op"); 26953d14c5d2SYehuda Sadeh if (err < 0) 26963d14c5d2SYehuda Sadeh goto out_mempool; 2697d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY, 26983d14c5d2SYehuda Sadeh OSD_OPREPLY_FRONT_LEN, 10, true, 26993d14c5d2SYehuda Sadeh "osd_op_reply"); 27003d14c5d2SYehuda Sadeh if (err < 0) 27013d14c5d2SYehuda Sadeh goto out_msgpool; 2702a40c4f10SYehuda Sadeh 2703dbcae088SDan Carpenter err = -ENOMEM; 2704a40c4f10SYehuda Sadeh osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify"); 2705dbcae088SDan Carpenter if (!osdc->notify_wq) 2706c172ec5cSIlya Dryomov goto out_msgpool_reply; 2707c172ec5cSIlya Dryomov 27083d14c5d2SYehuda Sadeh return 0; 27093d14c5d2SYehuda Sadeh 2710c172ec5cSIlya Dryomov out_msgpool_reply: 2711c172ec5cSIlya Dryomov ceph_msgpool_destroy(&osdc->msgpool_op_reply); 27123d14c5d2SYehuda Sadeh out_msgpool: 27133d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 27143d14c5d2SYehuda Sadeh out_mempool: 27153d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 27163d14c5d2SYehuda Sadeh out: 27173d14c5d2SYehuda Sadeh return err; 27183d14c5d2SYehuda Sadeh } 27193d14c5d2SYehuda Sadeh 27203d14c5d2SYehuda Sadeh void ceph_osdc_stop(struct ceph_osd_client *osdc) 27213d14c5d2SYehuda Sadeh { 2722a40c4f10SYehuda Sadeh flush_workqueue(osdc->notify_wq); 2723a40c4f10SYehuda Sadeh destroy_workqueue(osdc->notify_wq); 27243d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->timeout_work); 27253d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->osds_timeout_work); 27263d14c5d2SYehuda Sadeh if (osdc->osdmap) { 27273d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 27283d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 27293d14c5d2SYehuda Sadeh } 2730aca420bcSSage Weil remove_all_osds(osdc); 27313d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 27323d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 27333d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op_reply); 27343d14c5d2SYehuda Sadeh } 27353d14c5d2SYehuda Sadeh 27363d14c5d2SYehuda Sadeh /* 27373d14c5d2SYehuda Sadeh * Read some contiguous pages. If we cross a stripe boundary, shorten 27383d14c5d2SYehuda Sadeh * *plen. Return number of bytes read, or error. 27393d14c5d2SYehuda Sadeh */ 27403d14c5d2SYehuda Sadeh int ceph_osdc_readpages(struct ceph_osd_client *osdc, 27413d14c5d2SYehuda Sadeh struct ceph_vino vino, struct ceph_file_layout *layout, 27423d14c5d2SYehuda Sadeh u64 off, u64 *plen, 27433d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 2744b7495fc2SSage Weil struct page **pages, int num_pages, int page_align) 27453d14c5d2SYehuda Sadeh { 27463d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 27473d14c5d2SYehuda Sadeh int rc = 0; 27483d14c5d2SYehuda Sadeh 27493d14c5d2SYehuda Sadeh dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino, 27503d14c5d2SYehuda Sadeh vino.snap, off, *plen); 2751715e4cd4SYan, Zheng req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 0, 1, 27523d14c5d2SYehuda Sadeh CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ, 2753acead002SAlex Elder NULL, truncate_seq, truncate_size, 2754153e5167SAlex Elder false); 27556816282dSSage Weil if (IS_ERR(req)) 27566816282dSSage Weil return PTR_ERR(req); 27573d14c5d2SYehuda Sadeh 27583d14c5d2SYehuda Sadeh /* it may be a short read due to an object boundary */ 27590fff87ecSAlex Elder 2760406e2c9fSAlex Elder osd_req_op_extent_osd_data_pages(req, 0, 2761a4ce40a9SAlex Elder pages, *plen, page_align, false, false); 27623d14c5d2SYehuda Sadeh 2763e0c59487SAlex Elder dout("readpages final extent is %llu~%llu (%llu bytes align %d)\n", 276443bfe5deSAlex Elder off, *plen, *plen, page_align); 27653d14c5d2SYehuda Sadeh 276679528734SAlex Elder ceph_osdc_build_request(req, off, NULL, vino.snap, NULL); 276702ee07d3SAlex Elder 27683d14c5d2SYehuda Sadeh rc = ceph_osdc_start_request(osdc, req, false); 27693d14c5d2SYehuda Sadeh if (!rc) 27703d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 27713d14c5d2SYehuda Sadeh 27723d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 27733d14c5d2SYehuda Sadeh dout("readpages result %d\n", rc); 27743d14c5d2SYehuda Sadeh return rc; 27753d14c5d2SYehuda Sadeh } 27763d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_readpages); 27773d14c5d2SYehuda Sadeh 27783d14c5d2SYehuda Sadeh /* 27793d14c5d2SYehuda Sadeh * do a synchronous write on N pages 27803d14c5d2SYehuda Sadeh */ 27813d14c5d2SYehuda Sadeh int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino, 27823d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 27833d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 27843d14c5d2SYehuda Sadeh u64 off, u64 len, 27853d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 27863d14c5d2SYehuda Sadeh struct timespec *mtime, 278724808826SAlex Elder struct page **pages, int num_pages) 27883d14c5d2SYehuda Sadeh { 27893d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 27903d14c5d2SYehuda Sadeh int rc = 0; 2791b7495fc2SSage Weil int page_align = off & ~PAGE_MASK; 27923d14c5d2SYehuda Sadeh 2793acead002SAlex Elder BUG_ON(vino.snap != CEPH_NOSNAP); /* snapshots aren't writeable */ 2794715e4cd4SYan, Zheng req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 0, 1, 27953d14c5d2SYehuda Sadeh CEPH_OSD_OP_WRITE, 279624808826SAlex Elder CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE, 2797acead002SAlex Elder snapc, truncate_seq, truncate_size, 2798153e5167SAlex Elder true); 27996816282dSSage Weil if (IS_ERR(req)) 28006816282dSSage Weil return PTR_ERR(req); 28013d14c5d2SYehuda Sadeh 28023d14c5d2SYehuda Sadeh /* it may be a short write due to an object boundary */ 2803406e2c9fSAlex Elder osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align, 280443bfe5deSAlex Elder false, false); 280543bfe5deSAlex Elder dout("writepages %llu~%llu (%llu bytes)\n", off, len, len); 28063d14c5d2SYehuda Sadeh 280779528734SAlex Elder ceph_osdc_build_request(req, off, snapc, CEPH_NOSNAP, mtime); 280802ee07d3SAlex Elder 280987f979d3SAlex Elder rc = ceph_osdc_start_request(osdc, req, true); 28103d14c5d2SYehuda Sadeh if (!rc) 28113d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 28123d14c5d2SYehuda Sadeh 28133d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 28143d14c5d2SYehuda Sadeh if (rc == 0) 28153d14c5d2SYehuda Sadeh rc = len; 28163d14c5d2SYehuda Sadeh dout("writepages result %d\n", rc); 28173d14c5d2SYehuda Sadeh return rc; 28183d14c5d2SYehuda Sadeh } 28193d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_writepages); 28203d14c5d2SYehuda Sadeh 28215522ae0bSAlex Elder int ceph_osdc_setup(void) 28225522ae0bSAlex Elder { 28233f1af42aSIlya Dryomov size_t size = sizeof(struct ceph_osd_request) + 28243f1af42aSIlya Dryomov CEPH_OSD_SLAB_OPS * sizeof(struct ceph_osd_req_op); 28253f1af42aSIlya Dryomov 28265522ae0bSAlex Elder BUG_ON(ceph_osd_request_cache); 28273f1af42aSIlya Dryomov ceph_osd_request_cache = kmem_cache_create("ceph_osd_request", size, 28283f1af42aSIlya Dryomov 0, 0, NULL); 28295522ae0bSAlex Elder 28305522ae0bSAlex Elder return ceph_osd_request_cache ? 0 : -ENOMEM; 28315522ae0bSAlex Elder } 28325522ae0bSAlex Elder EXPORT_SYMBOL(ceph_osdc_setup); 28335522ae0bSAlex Elder 28345522ae0bSAlex Elder void ceph_osdc_cleanup(void) 28355522ae0bSAlex Elder { 28365522ae0bSAlex Elder BUG_ON(!ceph_osd_request_cache); 28375522ae0bSAlex Elder kmem_cache_destroy(ceph_osd_request_cache); 28385522ae0bSAlex Elder ceph_osd_request_cache = NULL; 28395522ae0bSAlex Elder } 28405522ae0bSAlex Elder EXPORT_SYMBOL(ceph_osdc_cleanup); 28415522ae0bSAlex Elder 28423d14c5d2SYehuda Sadeh /* 28433d14c5d2SYehuda Sadeh * handle incoming message 28443d14c5d2SYehuda Sadeh */ 28453d14c5d2SYehuda Sadeh static void dispatch(struct ceph_connection *con, struct ceph_msg *msg) 28463d14c5d2SYehuda Sadeh { 28473d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 28483d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 28493d14c5d2SYehuda Sadeh int type = le16_to_cpu(msg->hdr.type); 28503d14c5d2SYehuda Sadeh 28513d14c5d2SYehuda Sadeh if (!osd) 28523d14c5d2SYehuda Sadeh goto out; 28533d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 28543d14c5d2SYehuda Sadeh 28553d14c5d2SYehuda Sadeh switch (type) { 28563d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 28573d14c5d2SYehuda Sadeh ceph_osdc_handle_map(osdc, msg); 28583d14c5d2SYehuda Sadeh break; 28593d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 286070cf052dSShraddha Barke handle_reply(osdc, msg); 28613d14c5d2SYehuda Sadeh break; 2862a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 2863a40c4f10SYehuda Sadeh handle_watch_notify(osdc, msg); 2864a40c4f10SYehuda Sadeh break; 28653d14c5d2SYehuda Sadeh 28663d14c5d2SYehuda Sadeh default: 28673d14c5d2SYehuda Sadeh pr_err("received unknown message type %d %s\n", type, 28683d14c5d2SYehuda Sadeh ceph_msg_type_name(type)); 28693d14c5d2SYehuda Sadeh } 28703d14c5d2SYehuda Sadeh out: 28713d14c5d2SYehuda Sadeh ceph_msg_put(msg); 28723d14c5d2SYehuda Sadeh } 28733d14c5d2SYehuda Sadeh 28743d14c5d2SYehuda Sadeh /* 2875d15f9d69SIlya Dryomov * Lookup and return message for incoming reply. Don't try to do 2876d15f9d69SIlya Dryomov * anything about a larger than preallocated data portion of the 2877d15f9d69SIlya Dryomov * message at the moment - for now, just skip the message. 28783d14c5d2SYehuda Sadeh */ 28793d14c5d2SYehuda Sadeh static struct ceph_msg *get_reply(struct ceph_connection *con, 28803d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 28813d14c5d2SYehuda Sadeh int *skip) 28823d14c5d2SYehuda Sadeh { 28833d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 28843d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = osd->o_osdc; 28853d14c5d2SYehuda Sadeh struct ceph_msg *m; 28863d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 28873f0a4ac5SIlya Dryomov int front_len = le32_to_cpu(hdr->front_len); 28883d14c5d2SYehuda Sadeh int data_len = le32_to_cpu(hdr->data_len); 28893d14c5d2SYehuda Sadeh u64 tid; 28903d14c5d2SYehuda Sadeh 28913d14c5d2SYehuda Sadeh tid = le64_to_cpu(hdr->tid); 28923d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 28933d14c5d2SYehuda Sadeh req = __lookup_request(osdc, tid); 28943d14c5d2SYehuda Sadeh if (!req) { 2895cd8140c6SIlya Dryomov dout("%s osd%d tid %llu unknown, skipping\n", __func__, 2896cd8140c6SIlya Dryomov osd->o_osd, tid); 28973d14c5d2SYehuda Sadeh m = NULL; 2898d15f9d69SIlya Dryomov *skip = 1; 28993d14c5d2SYehuda Sadeh goto out; 29003d14c5d2SYehuda Sadeh } 29013d14c5d2SYehuda Sadeh 29028921d114SAlex Elder ceph_msg_revoke_incoming(req->r_reply); 29033d14c5d2SYehuda Sadeh 2904f2be82b0SIlya Dryomov if (front_len > req->r_reply->front_alloc_len) { 2905d15f9d69SIlya Dryomov pr_warn("%s osd%d tid %llu front %d > preallocated %d\n", 2906d15f9d69SIlya Dryomov __func__, osd->o_osd, req->r_tid, front_len, 2907d15f9d69SIlya Dryomov req->r_reply->front_alloc_len); 29083f0a4ac5SIlya Dryomov m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front_len, GFP_NOFS, 29093f0a4ac5SIlya Dryomov false); 29103d14c5d2SYehuda Sadeh if (!m) 29113d14c5d2SYehuda Sadeh goto out; 29123d14c5d2SYehuda Sadeh ceph_msg_put(req->r_reply); 29133d14c5d2SYehuda Sadeh req->r_reply = m; 29143d14c5d2SYehuda Sadeh } 29153d14c5d2SYehuda Sadeh 2916d15f9d69SIlya Dryomov if (data_len > req->r_reply->data_length) { 2917d15f9d69SIlya Dryomov pr_warn("%s osd%d tid %llu data %d > preallocated %zu, skipping\n", 2918d15f9d69SIlya Dryomov __func__, osd->o_osd, req->r_tid, data_len, 2919d15f9d69SIlya Dryomov req->r_reply->data_length); 29203d14c5d2SYehuda Sadeh m = NULL; 2921d15f9d69SIlya Dryomov *skip = 1; 29223d14c5d2SYehuda Sadeh goto out; 29233d14c5d2SYehuda Sadeh } 2924d15f9d69SIlya Dryomov 2925d15f9d69SIlya Dryomov m = ceph_msg_get(req->r_reply); 29263d14c5d2SYehuda Sadeh dout("get_reply tid %lld %p\n", tid, m); 29273d14c5d2SYehuda Sadeh 29283d14c5d2SYehuda Sadeh out: 29293d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 29303d14c5d2SYehuda Sadeh return m; 29313d14c5d2SYehuda Sadeh } 29323d14c5d2SYehuda Sadeh 29333d14c5d2SYehuda Sadeh static struct ceph_msg *alloc_msg(struct ceph_connection *con, 29343d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 29353d14c5d2SYehuda Sadeh int *skip) 29363d14c5d2SYehuda Sadeh { 29373d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 29383d14c5d2SYehuda Sadeh int type = le16_to_cpu(hdr->type); 29393d14c5d2SYehuda Sadeh int front = le32_to_cpu(hdr->front_len); 29403d14c5d2SYehuda Sadeh 29411c20f2d2SAlex Elder *skip = 0; 29423d14c5d2SYehuda Sadeh switch (type) { 29433d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 2944a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 2945b61c2763SSage Weil return ceph_msg_new(type, front, GFP_NOFS, false); 29463d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 29473d14c5d2SYehuda Sadeh return get_reply(con, hdr, skip); 29483d14c5d2SYehuda Sadeh default: 29493d14c5d2SYehuda Sadeh pr_info("alloc_msg unexpected msg type %d from osd%d\n", type, 29503d14c5d2SYehuda Sadeh osd->o_osd); 29513d14c5d2SYehuda Sadeh *skip = 1; 29523d14c5d2SYehuda Sadeh return NULL; 29533d14c5d2SYehuda Sadeh } 29543d14c5d2SYehuda Sadeh } 29553d14c5d2SYehuda Sadeh 29563d14c5d2SYehuda Sadeh /* 29573d14c5d2SYehuda Sadeh * Wrappers to refcount containing ceph_osd struct 29583d14c5d2SYehuda Sadeh */ 29593d14c5d2SYehuda Sadeh static struct ceph_connection *get_osd_con(struct ceph_connection *con) 29603d14c5d2SYehuda Sadeh { 29613d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 29623d14c5d2SYehuda Sadeh if (get_osd(osd)) 29633d14c5d2SYehuda Sadeh return con; 29643d14c5d2SYehuda Sadeh return NULL; 29653d14c5d2SYehuda Sadeh } 29663d14c5d2SYehuda Sadeh 29673d14c5d2SYehuda Sadeh static void put_osd_con(struct ceph_connection *con) 29683d14c5d2SYehuda Sadeh { 29693d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 29703d14c5d2SYehuda Sadeh put_osd(osd); 29713d14c5d2SYehuda Sadeh } 29723d14c5d2SYehuda Sadeh 29733d14c5d2SYehuda Sadeh /* 29743d14c5d2SYehuda Sadeh * authentication 29753d14c5d2SYehuda Sadeh */ 2976a3530df3SAlex Elder /* 2977a3530df3SAlex Elder * Note: returned pointer is the address of a structure that's 2978a3530df3SAlex Elder * managed separately. Caller must *not* attempt to free it. 2979a3530df3SAlex Elder */ 2980a3530df3SAlex Elder static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con, 29818f43fb53SAlex Elder int *proto, int force_new) 29823d14c5d2SYehuda Sadeh { 29833d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 29843d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 29853d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 298674f1869fSAlex Elder struct ceph_auth_handshake *auth = &o->o_auth; 29873d14c5d2SYehuda Sadeh 298874f1869fSAlex Elder if (force_new && auth->authorizer) { 29896c1ea260SIlya Dryomov ceph_auth_destroy_authorizer(auth->authorizer); 299074f1869fSAlex Elder auth->authorizer = NULL; 29913d14c5d2SYehuda Sadeh } 299227859f97SSage Weil if (!auth->authorizer) { 299327859f97SSage Weil int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 2994a3530df3SAlex Elder auth); 29953d14c5d2SYehuda Sadeh if (ret) 2996a3530df3SAlex Elder return ERR_PTR(ret); 299727859f97SSage Weil } else { 299827859f97SSage Weil int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 29990bed9b5cSSage Weil auth); 30000bed9b5cSSage Weil if (ret) 30010bed9b5cSSage Weil return ERR_PTR(ret); 30023d14c5d2SYehuda Sadeh } 30033d14c5d2SYehuda Sadeh *proto = ac->protocol; 300474f1869fSAlex Elder 3005a3530df3SAlex Elder return auth; 30063d14c5d2SYehuda Sadeh } 30073d14c5d2SYehuda Sadeh 30083d14c5d2SYehuda Sadeh 30093d14c5d2SYehuda Sadeh static int verify_authorizer_reply(struct ceph_connection *con, int len) 30103d14c5d2SYehuda Sadeh { 30113d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 30123d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 30133d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 30143d14c5d2SYehuda Sadeh 301527859f97SSage Weil return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len); 30163d14c5d2SYehuda Sadeh } 30173d14c5d2SYehuda Sadeh 30183d14c5d2SYehuda Sadeh static int invalidate_authorizer(struct ceph_connection *con) 30193d14c5d2SYehuda Sadeh { 30203d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 30213d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 30223d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 30233d14c5d2SYehuda Sadeh 302427859f97SSage Weil ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD); 30253d14c5d2SYehuda Sadeh return ceph_monc_validate_auth(&osdc->client->monc); 30263d14c5d2SYehuda Sadeh } 30273d14c5d2SYehuda Sadeh 302879dbd1baSIlya Dryomov static int osd_sign_message(struct ceph_msg *msg) 302933d07337SYan, Zheng { 303079dbd1baSIlya Dryomov struct ceph_osd *o = msg->con->private; 303133d07337SYan, Zheng struct ceph_auth_handshake *auth = &o->o_auth; 303279dbd1baSIlya Dryomov 303333d07337SYan, Zheng return ceph_auth_sign_message(auth, msg); 303433d07337SYan, Zheng } 303533d07337SYan, Zheng 303679dbd1baSIlya Dryomov static int osd_check_message_signature(struct ceph_msg *msg) 303733d07337SYan, Zheng { 303879dbd1baSIlya Dryomov struct ceph_osd *o = msg->con->private; 303933d07337SYan, Zheng struct ceph_auth_handshake *auth = &o->o_auth; 304079dbd1baSIlya Dryomov 304133d07337SYan, Zheng return ceph_auth_check_message_signature(auth, msg); 304233d07337SYan, Zheng } 304333d07337SYan, Zheng 30443d14c5d2SYehuda Sadeh static const struct ceph_connection_operations osd_con_ops = { 30453d14c5d2SYehuda Sadeh .get = get_osd_con, 30463d14c5d2SYehuda Sadeh .put = put_osd_con, 30473d14c5d2SYehuda Sadeh .dispatch = dispatch, 30483d14c5d2SYehuda Sadeh .get_authorizer = get_authorizer, 30493d14c5d2SYehuda Sadeh .verify_authorizer_reply = verify_authorizer_reply, 30503d14c5d2SYehuda Sadeh .invalidate_authorizer = invalidate_authorizer, 30513d14c5d2SYehuda Sadeh .alloc_msg = alloc_msg, 305279dbd1baSIlya Dryomov .sign_message = osd_sign_message, 305379dbd1baSIlya Dryomov .check_message_signature = osd_check_message_signature, 30543d14c5d2SYehuda Sadeh .fault = osd_reset, 30553d14c5d2SYehuda Sadeh }; 3056