1a4ce40a9SAlex Elder 23d14c5d2SYehuda Sadeh #include <linux/ceph/ceph_debug.h> 33d14c5d2SYehuda Sadeh 43d14c5d2SYehuda Sadeh #include <linux/module.h> 53d14c5d2SYehuda Sadeh #include <linux/err.h> 63d14c5d2SYehuda Sadeh #include <linux/highmem.h> 73d14c5d2SYehuda Sadeh #include <linux/mm.h> 83d14c5d2SYehuda Sadeh #include <linux/pagemap.h> 93d14c5d2SYehuda Sadeh #include <linux/slab.h> 103d14c5d2SYehuda Sadeh #include <linux/uaccess.h> 113d14c5d2SYehuda Sadeh #ifdef CONFIG_BLOCK 123d14c5d2SYehuda Sadeh #include <linux/bio.h> 133d14c5d2SYehuda Sadeh #endif 143d14c5d2SYehuda Sadeh 153d14c5d2SYehuda Sadeh #include <linux/ceph/libceph.h> 163d14c5d2SYehuda Sadeh #include <linux/ceph/osd_client.h> 173d14c5d2SYehuda Sadeh #include <linux/ceph/messenger.h> 183d14c5d2SYehuda Sadeh #include <linux/ceph/decode.h> 193d14c5d2SYehuda Sadeh #include <linux/ceph/auth.h> 203d14c5d2SYehuda Sadeh #include <linux/ceph/pagelist.h> 213d14c5d2SYehuda Sadeh 223d14c5d2SYehuda Sadeh #define OSD_OP_FRONT_LEN 4096 233d14c5d2SYehuda Sadeh #define OSD_OPREPLY_FRONT_LEN 512 243d14c5d2SYehuda Sadeh 255522ae0bSAlex Elder static struct kmem_cache *ceph_osd_request_cache; 265522ae0bSAlex Elder 273d14c5d2SYehuda Sadeh static const struct ceph_connection_operations osd_con_ops; 283d14c5d2SYehuda Sadeh 29f9d25199SAlex Elder static void __send_queued(struct ceph_osd_client *osdc); 306f6c7006SSage Weil static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd); 31a40c4f10SYehuda Sadeh static void __register_request(struct ceph_osd_client *osdc, 32a40c4f10SYehuda Sadeh struct ceph_osd_request *req); 332cc6128aSIlya Dryomov static void __unregister_request(struct ceph_osd_client *osdc, 342cc6128aSIlya Dryomov struct ceph_osd_request *req); 35a40c4f10SYehuda Sadeh static void __unregister_linger_request(struct ceph_osd_client *osdc, 36a40c4f10SYehuda Sadeh struct ceph_osd_request *req); 372cc6128aSIlya Dryomov static void __enqueue_request(struct ceph_osd_request *req); 3856e925b6SSage Weil static void __send_request(struct ceph_osd_client *osdc, 39a40c4f10SYehuda Sadeh struct ceph_osd_request *req); 403d14c5d2SYehuda Sadeh 413d14c5d2SYehuda Sadeh /* 423d14c5d2SYehuda Sadeh * Implement client access to distributed object storage cluster. 433d14c5d2SYehuda Sadeh * 443d14c5d2SYehuda Sadeh * All data objects are stored within a cluster/cloud of OSDs, or 453d14c5d2SYehuda Sadeh * "object storage devices." (Note that Ceph OSDs have _nothing_ to 463d14c5d2SYehuda Sadeh * do with the T10 OSD extensions to SCSI.) Ceph OSDs are simply 473d14c5d2SYehuda Sadeh * remote daemons serving up and coordinating consistent and safe 483d14c5d2SYehuda Sadeh * access to storage. 493d14c5d2SYehuda Sadeh * 503d14c5d2SYehuda Sadeh * Cluster membership and the mapping of data objects onto storage devices 513d14c5d2SYehuda Sadeh * are described by the osd map. 523d14c5d2SYehuda Sadeh * 533d14c5d2SYehuda Sadeh * We keep track of pending OSD requests (read, write), resubmit 543d14c5d2SYehuda Sadeh * requests to different OSDs when the cluster topology/data layout 553d14c5d2SYehuda Sadeh * change, or retry the affected requests when the communications 563d14c5d2SYehuda Sadeh * channel with an OSD is reset. 573d14c5d2SYehuda Sadeh */ 583d14c5d2SYehuda Sadeh 593d14c5d2SYehuda Sadeh /* 603d14c5d2SYehuda Sadeh * calculate the mapping of a file extent onto an object, and fill out the 613d14c5d2SYehuda Sadeh * request accordingly. shorten extent as necessary if it crosses an 623d14c5d2SYehuda Sadeh * object boundary. 633d14c5d2SYehuda Sadeh * 643d14c5d2SYehuda Sadeh * fill osd op in request message. 653d14c5d2SYehuda Sadeh */ 66dbe0fc41SAlex Elder static int calc_layout(struct ceph_file_layout *layout, u64 off, u64 *plen, 67a19dadfbSAlex Elder u64 *objnum, u64 *objoff, u64 *objlen) 683d14c5d2SYehuda Sadeh { 6960e56f13SAlex Elder u64 orig_len = *plen; 70d63b77f4SSage Weil int r; 713d14c5d2SYehuda Sadeh 7260e56f13SAlex Elder /* object extent? */ 7375d1c941SAlex Elder r = ceph_calc_file_object_mapping(layout, off, orig_len, objnum, 7475d1c941SAlex Elder objoff, objlen); 75d63b77f4SSage Weil if (r < 0) 76d63b77f4SSage Weil return r; 7775d1c941SAlex Elder if (*objlen < orig_len) { 7875d1c941SAlex Elder *plen = *objlen; 7960e56f13SAlex Elder dout(" skipping last %llu, final file extent %llu~%llu\n", 8060e56f13SAlex Elder orig_len - *plen, off, *plen); 8160e56f13SAlex Elder } 8260e56f13SAlex Elder 8375d1c941SAlex Elder dout("calc_layout objnum=%llx %llu~%llu\n", *objnum, *objoff, *objlen); 843d14c5d2SYehuda Sadeh 853ff5f385SAlex Elder return 0; 863d14c5d2SYehuda Sadeh } 873d14c5d2SYehuda Sadeh 88c54d47bfSAlex Elder static void ceph_osd_data_init(struct ceph_osd_data *osd_data) 89c54d47bfSAlex Elder { 90c54d47bfSAlex Elder memset(osd_data, 0, sizeof (*osd_data)); 91c54d47bfSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_NONE; 92c54d47bfSAlex Elder } 93c54d47bfSAlex Elder 94a4ce40a9SAlex Elder static void ceph_osd_data_pages_init(struct ceph_osd_data *osd_data, 9543bfe5deSAlex Elder struct page **pages, u64 length, u32 alignment, 9643bfe5deSAlex Elder bool pages_from_pool, bool own_pages) 9743bfe5deSAlex Elder { 9843bfe5deSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_PAGES; 9943bfe5deSAlex Elder osd_data->pages = pages; 10043bfe5deSAlex Elder osd_data->length = length; 10143bfe5deSAlex Elder osd_data->alignment = alignment; 10243bfe5deSAlex Elder osd_data->pages_from_pool = pages_from_pool; 10343bfe5deSAlex Elder osd_data->own_pages = own_pages; 10443bfe5deSAlex Elder } 10543bfe5deSAlex Elder 106a4ce40a9SAlex Elder static void ceph_osd_data_pagelist_init(struct ceph_osd_data *osd_data, 10743bfe5deSAlex Elder struct ceph_pagelist *pagelist) 10843bfe5deSAlex Elder { 10943bfe5deSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_PAGELIST; 11043bfe5deSAlex Elder osd_data->pagelist = pagelist; 11143bfe5deSAlex Elder } 11243bfe5deSAlex Elder 11343bfe5deSAlex Elder #ifdef CONFIG_BLOCK 114a4ce40a9SAlex Elder static void ceph_osd_data_bio_init(struct ceph_osd_data *osd_data, 11543bfe5deSAlex Elder struct bio *bio, size_t bio_length) 11643bfe5deSAlex Elder { 11743bfe5deSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_BIO; 11843bfe5deSAlex Elder osd_data->bio = bio; 11943bfe5deSAlex Elder osd_data->bio_length = bio_length; 12043bfe5deSAlex Elder } 12143bfe5deSAlex Elder #endif /* CONFIG_BLOCK */ 12243bfe5deSAlex Elder 123863c7eb5SAlex Elder #define osd_req_op_data(oreq, whch, typ, fld) \ 124863c7eb5SAlex Elder ({ \ 125863c7eb5SAlex Elder BUG_ON(whch >= (oreq)->r_num_ops); \ 126863c7eb5SAlex Elder &(oreq)->r_ops[whch].typ.fld; \ 127863c7eb5SAlex Elder }) 128863c7eb5SAlex Elder 12949719778SAlex Elder static struct ceph_osd_data * 13049719778SAlex Elder osd_req_op_raw_data_in(struct ceph_osd_request *osd_req, unsigned int which) 13149719778SAlex Elder { 13249719778SAlex Elder BUG_ON(which >= osd_req->r_num_ops); 13349719778SAlex Elder 13449719778SAlex Elder return &osd_req->r_ops[which].raw_data_in; 13549719778SAlex Elder } 13649719778SAlex Elder 137a4ce40a9SAlex Elder struct ceph_osd_data * 138a4ce40a9SAlex Elder osd_req_op_extent_osd_data(struct ceph_osd_request *osd_req, 139406e2c9fSAlex Elder unsigned int which) 140a4ce40a9SAlex Elder { 141863c7eb5SAlex Elder return osd_req_op_data(osd_req, which, extent, osd_data); 142a4ce40a9SAlex Elder } 143a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data); 144a4ce40a9SAlex Elder 145a4ce40a9SAlex Elder struct ceph_osd_data * 146a4ce40a9SAlex Elder osd_req_op_cls_response_data(struct ceph_osd_request *osd_req, 147a4ce40a9SAlex Elder unsigned int which) 148a4ce40a9SAlex Elder { 149863c7eb5SAlex Elder return osd_req_op_data(osd_req, which, cls, response_data); 150a4ce40a9SAlex Elder } 151a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_response_data); /* ??? */ 152a4ce40a9SAlex Elder 15349719778SAlex Elder void osd_req_op_raw_data_in_pages(struct ceph_osd_request *osd_req, 15449719778SAlex Elder unsigned int which, struct page **pages, 15549719778SAlex Elder u64 length, u32 alignment, 15649719778SAlex Elder bool pages_from_pool, bool own_pages) 15749719778SAlex Elder { 15849719778SAlex Elder struct ceph_osd_data *osd_data; 15949719778SAlex Elder 16049719778SAlex Elder osd_data = osd_req_op_raw_data_in(osd_req, which); 16149719778SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 16249719778SAlex Elder pages_from_pool, own_pages); 16349719778SAlex Elder } 16449719778SAlex Elder EXPORT_SYMBOL(osd_req_op_raw_data_in_pages); 16549719778SAlex Elder 166a4ce40a9SAlex Elder void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *osd_req, 167406e2c9fSAlex Elder unsigned int which, struct page **pages, 168406e2c9fSAlex Elder u64 length, u32 alignment, 169a4ce40a9SAlex Elder bool pages_from_pool, bool own_pages) 170a4ce40a9SAlex Elder { 171a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 172a4ce40a9SAlex Elder 173863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, extent, osd_data); 174a4ce40a9SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 175a4ce40a9SAlex Elder pages_from_pool, own_pages); 176a4ce40a9SAlex Elder } 177a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data_pages); 178a4ce40a9SAlex Elder 179a4ce40a9SAlex Elder void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *osd_req, 180406e2c9fSAlex Elder unsigned int which, struct ceph_pagelist *pagelist) 181a4ce40a9SAlex Elder { 182a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 183a4ce40a9SAlex Elder 184863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, extent, osd_data); 185a4ce40a9SAlex Elder ceph_osd_data_pagelist_init(osd_data, pagelist); 186a4ce40a9SAlex Elder } 187a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data_pagelist); 188a4ce40a9SAlex Elder 189a4ce40a9SAlex Elder #ifdef CONFIG_BLOCK 190a4ce40a9SAlex Elder void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *osd_req, 191406e2c9fSAlex Elder unsigned int which, struct bio *bio, size_t bio_length) 192a4ce40a9SAlex Elder { 193a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 194863c7eb5SAlex Elder 195863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, extent, osd_data); 196a4ce40a9SAlex Elder ceph_osd_data_bio_init(osd_data, bio, bio_length); 197a4ce40a9SAlex Elder } 198a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data_bio); 199a4ce40a9SAlex Elder #endif /* CONFIG_BLOCK */ 200a4ce40a9SAlex Elder 201a4ce40a9SAlex Elder static void osd_req_op_cls_request_info_pagelist( 202a4ce40a9SAlex Elder struct ceph_osd_request *osd_req, 203a4ce40a9SAlex Elder unsigned int which, struct ceph_pagelist *pagelist) 204a4ce40a9SAlex Elder { 205a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 206a4ce40a9SAlex Elder 207863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, request_info); 208a4ce40a9SAlex Elder ceph_osd_data_pagelist_init(osd_data, pagelist); 209a4ce40a9SAlex Elder } 210a4ce40a9SAlex Elder 21104017e29SAlex Elder void osd_req_op_cls_request_data_pagelist( 21204017e29SAlex Elder struct ceph_osd_request *osd_req, 21304017e29SAlex Elder unsigned int which, struct ceph_pagelist *pagelist) 21404017e29SAlex Elder { 21504017e29SAlex Elder struct ceph_osd_data *osd_data; 21604017e29SAlex Elder 217863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, request_data); 21804017e29SAlex Elder ceph_osd_data_pagelist_init(osd_data, pagelist); 21904017e29SAlex Elder } 22004017e29SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_request_data_pagelist); 22104017e29SAlex Elder 2226c57b554SAlex Elder void osd_req_op_cls_request_data_pages(struct ceph_osd_request *osd_req, 2236c57b554SAlex Elder unsigned int which, struct page **pages, u64 length, 2246c57b554SAlex Elder u32 alignment, bool pages_from_pool, bool own_pages) 2256c57b554SAlex Elder { 2266c57b554SAlex Elder struct ceph_osd_data *osd_data; 2276c57b554SAlex Elder 2286c57b554SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, request_data); 2296c57b554SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 2306c57b554SAlex Elder pages_from_pool, own_pages); 2316c57b554SAlex Elder } 2326c57b554SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_request_data_pages); 2336c57b554SAlex Elder 234a4ce40a9SAlex Elder void osd_req_op_cls_response_data_pages(struct ceph_osd_request *osd_req, 235a4ce40a9SAlex Elder unsigned int which, struct page **pages, u64 length, 236a4ce40a9SAlex Elder u32 alignment, bool pages_from_pool, bool own_pages) 237a4ce40a9SAlex Elder { 238a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 239a4ce40a9SAlex Elder 240863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, response_data); 241a4ce40a9SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 242a4ce40a9SAlex Elder pages_from_pool, own_pages); 243a4ce40a9SAlex Elder } 244a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_response_data_pages); 245a4ce40a9SAlex Elder 24623c08a9cSAlex Elder static u64 ceph_osd_data_length(struct ceph_osd_data *osd_data) 24723c08a9cSAlex Elder { 24823c08a9cSAlex Elder switch (osd_data->type) { 24923c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_NONE: 25023c08a9cSAlex Elder return 0; 25123c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_PAGES: 25223c08a9cSAlex Elder return osd_data->length; 25323c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_PAGELIST: 25423c08a9cSAlex Elder return (u64)osd_data->pagelist->length; 25523c08a9cSAlex Elder #ifdef CONFIG_BLOCK 25623c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_BIO: 25723c08a9cSAlex Elder return (u64)osd_data->bio_length; 25823c08a9cSAlex Elder #endif /* CONFIG_BLOCK */ 25923c08a9cSAlex Elder default: 26023c08a9cSAlex Elder WARN(true, "unrecognized data type %d\n", (int)osd_data->type); 26123c08a9cSAlex Elder return 0; 26223c08a9cSAlex Elder } 26323c08a9cSAlex Elder } 26423c08a9cSAlex Elder 265c54d47bfSAlex Elder static void ceph_osd_data_release(struct ceph_osd_data *osd_data) 266c54d47bfSAlex Elder { 2675476492fSAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES && osd_data->own_pages) { 268c54d47bfSAlex Elder int num_pages; 269c54d47bfSAlex Elder 270c54d47bfSAlex Elder num_pages = calc_pages_for((u64)osd_data->alignment, 271c54d47bfSAlex Elder (u64)osd_data->length); 272c54d47bfSAlex Elder ceph_release_page_vector(osd_data->pages, num_pages); 273c54d47bfSAlex Elder } 2745476492fSAlex Elder ceph_osd_data_init(osd_data); 2755476492fSAlex Elder } 2765476492fSAlex Elder 2775476492fSAlex Elder static void osd_req_op_data_release(struct ceph_osd_request *osd_req, 2785476492fSAlex Elder unsigned int which) 2795476492fSAlex Elder { 2805476492fSAlex Elder struct ceph_osd_req_op *op; 2815476492fSAlex Elder 2825476492fSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 2835476492fSAlex Elder op = &osd_req->r_ops[which]; 2845476492fSAlex Elder 2855476492fSAlex Elder switch (op->op) { 2865476492fSAlex Elder case CEPH_OSD_OP_READ: 2875476492fSAlex Elder case CEPH_OSD_OP_WRITE: 2885476492fSAlex Elder ceph_osd_data_release(&op->extent.osd_data); 2895476492fSAlex Elder break; 2905476492fSAlex Elder case CEPH_OSD_OP_CALL: 2915476492fSAlex Elder ceph_osd_data_release(&op->cls.request_info); 29204017e29SAlex Elder ceph_osd_data_release(&op->cls.request_data); 2935476492fSAlex Elder ceph_osd_data_release(&op->cls.response_data); 2945476492fSAlex Elder break; 295d74b50beSYan, Zheng case CEPH_OSD_OP_SETXATTR: 296d74b50beSYan, Zheng case CEPH_OSD_OP_CMPXATTR: 297d74b50beSYan, Zheng ceph_osd_data_release(&op->xattr.osd_data); 298d74b50beSYan, Zheng break; 29966ba609fSYan, Zheng case CEPH_OSD_OP_STAT: 30066ba609fSYan, Zheng ceph_osd_data_release(&op->raw_data_in); 30166ba609fSYan, Zheng break; 3025476492fSAlex Elder default: 3035476492fSAlex Elder break; 3045476492fSAlex Elder } 305c54d47bfSAlex Elder } 306c54d47bfSAlex Elder 3073d14c5d2SYehuda Sadeh /* 3083d14c5d2SYehuda Sadeh * requests 3093d14c5d2SYehuda Sadeh */ 3109e94af20SIlya Dryomov static void ceph_osdc_release_request(struct kref *kref) 3113d14c5d2SYehuda Sadeh { 3129e94af20SIlya Dryomov struct ceph_osd_request *req = container_of(kref, 3139e94af20SIlya Dryomov struct ceph_osd_request, r_kref); 3145476492fSAlex Elder unsigned int which; 3153d14c5d2SYehuda Sadeh 3169e94af20SIlya Dryomov dout("%s %p (r_request %p r_reply %p)\n", __func__, req, 3179e94af20SIlya Dryomov req->r_request, req->r_reply); 3186562d661SIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&req->r_node)); 3196562d661SIlya Dryomov WARN_ON(!list_empty(&req->r_req_lru_item)); 3206562d661SIlya Dryomov WARN_ON(!list_empty(&req->r_osd_item)); 3216562d661SIlya Dryomov WARN_ON(!list_empty(&req->r_linger_item)); 3226562d661SIlya Dryomov WARN_ON(!list_empty(&req->r_linger_osd_item)); 3236562d661SIlya Dryomov WARN_ON(req->r_osd); 3249e94af20SIlya Dryomov 3253d14c5d2SYehuda Sadeh if (req->r_request) 3263d14c5d2SYehuda Sadeh ceph_msg_put(req->r_request); 327ace6d3a9SAlex Elder if (req->r_reply) { 3288921d114SAlex Elder ceph_msg_revoke_incoming(req->r_reply); 329ab8cb34aSAlex Elder ceph_msg_put(req->r_reply); 330ace6d3a9SAlex Elder } 3310fff87ecSAlex Elder 3325476492fSAlex Elder for (which = 0; which < req->r_num_ops; which++) 3335476492fSAlex Elder osd_req_op_data_release(req, which); 3340fff87ecSAlex Elder 3353d14c5d2SYehuda Sadeh ceph_put_snap_context(req->r_snapc); 3363d14c5d2SYehuda Sadeh if (req->r_mempool) 3373d14c5d2SYehuda Sadeh mempool_free(req, req->r_osdc->req_mempool); 3383d14c5d2SYehuda Sadeh else 3395522ae0bSAlex Elder kmem_cache_free(ceph_osd_request_cache, req); 3405522ae0bSAlex Elder 3413d14c5d2SYehuda Sadeh } 3429e94af20SIlya Dryomov 3439e94af20SIlya Dryomov void ceph_osdc_get_request(struct ceph_osd_request *req) 3449e94af20SIlya Dryomov { 3459e94af20SIlya Dryomov dout("%s %p (was %d)\n", __func__, req, 3469e94af20SIlya Dryomov atomic_read(&req->r_kref.refcount)); 3479e94af20SIlya Dryomov kref_get(&req->r_kref); 3489e94af20SIlya Dryomov } 3499e94af20SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_get_request); 3509e94af20SIlya Dryomov 3519e94af20SIlya Dryomov void ceph_osdc_put_request(struct ceph_osd_request *req) 3529e94af20SIlya Dryomov { 3539e94af20SIlya Dryomov dout("%s %p (was %d)\n", __func__, req, 3549e94af20SIlya Dryomov atomic_read(&req->r_kref.refcount)); 3559e94af20SIlya Dryomov kref_put(&req->r_kref, ceph_osdc_release_request); 3569e94af20SIlya Dryomov } 3579e94af20SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_put_request); 3583d14c5d2SYehuda Sadeh 3593d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc, 3603d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 3611b83bef2SSage Weil unsigned int num_ops, 3623d14c5d2SYehuda Sadeh bool use_mempool, 36354a54007SAlex Elder gfp_t gfp_flags) 3643d14c5d2SYehuda Sadeh { 3653d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 3663d14c5d2SYehuda Sadeh struct ceph_msg *msg; 3671b83bef2SSage Weil size_t msg_size; 3683d14c5d2SYehuda Sadeh 36979528734SAlex Elder BUILD_BUG_ON(CEPH_OSD_MAX_OP > U16_MAX); 37079528734SAlex Elder BUG_ON(num_ops > CEPH_OSD_MAX_OP); 37179528734SAlex Elder 3721b83bef2SSage Weil msg_size = 4 + 4 + 8 + 8 + 4+8; 3731b83bef2SSage Weil msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */ 3741b83bef2SSage Weil msg_size += 1 + 8 + 4 + 4; /* pg_t */ 3752d0ebc5dSIlya Dryomov msg_size += 4 + CEPH_MAX_OID_NAME_LEN; /* oid */ 3761b83bef2SSage Weil msg_size += 2 + num_ops*sizeof(struct ceph_osd_op); 3771b83bef2SSage Weil msg_size += 8; /* snapid */ 3781b83bef2SSage Weil msg_size += 8; /* snap_seq */ 3791b83bef2SSage Weil msg_size += 8 * (snapc ? snapc->num_snaps : 0); /* snaps */ 3801b83bef2SSage Weil msg_size += 4; 3813d14c5d2SYehuda Sadeh 3823d14c5d2SYehuda Sadeh if (use_mempool) { 3833d14c5d2SYehuda Sadeh req = mempool_alloc(osdc->req_mempool, gfp_flags); 3843d14c5d2SYehuda Sadeh memset(req, 0, sizeof(*req)); 3853d14c5d2SYehuda Sadeh } else { 3865522ae0bSAlex Elder req = kmem_cache_zalloc(ceph_osd_request_cache, gfp_flags); 3873d14c5d2SYehuda Sadeh } 3883d14c5d2SYehuda Sadeh if (req == NULL) 3893d14c5d2SYehuda Sadeh return NULL; 3903d14c5d2SYehuda Sadeh 3913d14c5d2SYehuda Sadeh req->r_osdc = osdc; 3923d14c5d2SYehuda Sadeh req->r_mempool = use_mempool; 39379528734SAlex Elder req->r_num_ops = num_ops; 3943d14c5d2SYehuda Sadeh 3953d14c5d2SYehuda Sadeh kref_init(&req->r_kref); 3963d14c5d2SYehuda Sadeh init_completion(&req->r_completion); 3973d14c5d2SYehuda Sadeh init_completion(&req->r_safe_completion); 398a978fa20SAlex Elder RB_CLEAR_NODE(&req->r_node); 3993d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&req->r_unsafe_item); 400a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&req->r_linger_item); 4011d0326b1SIlya Dryomov INIT_LIST_HEAD(&req->r_linger_osd_item); 402935b639aSSage Weil INIT_LIST_HEAD(&req->r_req_lru_item); 403cd43045cSSage Weil INIT_LIST_HEAD(&req->r_osd_item); 404cd43045cSSage Weil 4053c972c95SIlya Dryomov req->r_base_oloc.pool = -1; 406205ee118SIlya Dryomov req->r_target_oloc.pool = -1; 40722116525SIlya Dryomov 4083d14c5d2SYehuda Sadeh /* create reply message */ 4093d14c5d2SYehuda Sadeh if (use_mempool) 4103d14c5d2SYehuda Sadeh msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0); 4113d14c5d2SYehuda Sadeh else 4123d14c5d2SYehuda Sadeh msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, 413b61c2763SSage Weil OSD_OPREPLY_FRONT_LEN, gfp_flags, true); 4143d14c5d2SYehuda Sadeh if (!msg) { 4153d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 4163d14c5d2SYehuda Sadeh return NULL; 4173d14c5d2SYehuda Sadeh } 4183d14c5d2SYehuda Sadeh req->r_reply = msg; 4193d14c5d2SYehuda Sadeh 4203d14c5d2SYehuda Sadeh /* create request message; allow space for oid */ 4213d14c5d2SYehuda Sadeh if (use_mempool) 4223d14c5d2SYehuda Sadeh msg = ceph_msgpool_get(&osdc->msgpool_op, 0); 4233d14c5d2SYehuda Sadeh else 424b61c2763SSage Weil msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true); 4253d14c5d2SYehuda Sadeh if (!msg) { 4263d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 4273d14c5d2SYehuda Sadeh return NULL; 4283d14c5d2SYehuda Sadeh } 4293d14c5d2SYehuda Sadeh 4303d14c5d2SYehuda Sadeh memset(msg->front.iov_base, 0, msg->front.iov_len); 4313d14c5d2SYehuda Sadeh 4323d14c5d2SYehuda Sadeh req->r_request = msg; 4333d14c5d2SYehuda Sadeh 4343d14c5d2SYehuda Sadeh return req; 4353d14c5d2SYehuda Sadeh } 4363d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_alloc_request); 4373d14c5d2SYehuda Sadeh 438a8dd0a37SAlex Elder static bool osd_req_opcode_valid(u16 opcode) 439a8dd0a37SAlex Elder { 440a8dd0a37SAlex Elder switch (opcode) { 44170b5bfa3SIlya Dryomov #define GENERATE_CASE(op, opcode, str) case CEPH_OSD_OP_##op: return true; 44270b5bfa3SIlya Dryomov __CEPH_FORALL_OSD_OPS(GENERATE_CASE) 44370b5bfa3SIlya Dryomov #undef GENERATE_CASE 444a8dd0a37SAlex Elder default: 445a8dd0a37SAlex Elder return false; 446a8dd0a37SAlex Elder } 447a8dd0a37SAlex Elder } 448a8dd0a37SAlex Elder 44933803f33SAlex Elder /* 45033803f33SAlex Elder * This is an osd op init function for opcodes that have no data or 45133803f33SAlex Elder * other information associated with them. It also serves as a 45233803f33SAlex Elder * common init routine for all the other init functions, below. 45333803f33SAlex Elder */ 454c99d2d4aSAlex Elder static struct ceph_osd_req_op * 45549719778SAlex Elder _osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which, 456144cba14SYan, Zheng u16 opcode, u32 flags) 45733803f33SAlex Elder { 458c99d2d4aSAlex Elder struct ceph_osd_req_op *op; 459c99d2d4aSAlex Elder 460c99d2d4aSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 46133803f33SAlex Elder BUG_ON(!osd_req_opcode_valid(opcode)); 46233803f33SAlex Elder 463c99d2d4aSAlex Elder op = &osd_req->r_ops[which]; 46433803f33SAlex Elder memset(op, 0, sizeof (*op)); 46533803f33SAlex Elder op->op = opcode; 466144cba14SYan, Zheng op->flags = flags; 467c99d2d4aSAlex Elder 468c99d2d4aSAlex Elder return op; 46933803f33SAlex Elder } 47033803f33SAlex Elder 47149719778SAlex Elder void osd_req_op_init(struct ceph_osd_request *osd_req, 472144cba14SYan, Zheng unsigned int which, u16 opcode, u32 flags) 47349719778SAlex Elder { 474144cba14SYan, Zheng (void)_osd_req_op_init(osd_req, which, opcode, flags); 47549719778SAlex Elder } 47649719778SAlex Elder EXPORT_SYMBOL(osd_req_op_init); 47749719778SAlex Elder 478c99d2d4aSAlex Elder void osd_req_op_extent_init(struct ceph_osd_request *osd_req, 479c99d2d4aSAlex Elder unsigned int which, u16 opcode, 48033803f33SAlex Elder u64 offset, u64 length, 48133803f33SAlex Elder u64 truncate_size, u32 truncate_seq) 48233803f33SAlex Elder { 483144cba14SYan, Zheng struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, 484144cba14SYan, Zheng opcode, 0); 48533803f33SAlex Elder size_t payload_len = 0; 48633803f33SAlex Elder 487ad7a60deSLi Wang BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE && 488864e9197SYan, Zheng opcode != CEPH_OSD_OP_ZERO && opcode != CEPH_OSD_OP_TRUNCATE); 48933803f33SAlex Elder 49033803f33SAlex Elder op->extent.offset = offset; 49133803f33SAlex Elder op->extent.length = length; 49233803f33SAlex Elder op->extent.truncate_size = truncate_size; 49333803f33SAlex Elder op->extent.truncate_seq = truncate_seq; 49433803f33SAlex Elder if (opcode == CEPH_OSD_OP_WRITE) 49533803f33SAlex Elder payload_len += length; 49633803f33SAlex Elder 49733803f33SAlex Elder op->payload_len = payload_len; 49833803f33SAlex Elder } 49933803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_init); 50033803f33SAlex Elder 501c99d2d4aSAlex Elder void osd_req_op_extent_update(struct ceph_osd_request *osd_req, 502c99d2d4aSAlex Elder unsigned int which, u64 length) 503e5975c7cSAlex Elder { 504c99d2d4aSAlex Elder struct ceph_osd_req_op *op; 505c99d2d4aSAlex Elder u64 previous; 506c99d2d4aSAlex Elder 507c99d2d4aSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 508c99d2d4aSAlex Elder op = &osd_req->r_ops[which]; 509c99d2d4aSAlex Elder previous = op->extent.length; 510e5975c7cSAlex Elder 511e5975c7cSAlex Elder if (length == previous) 512e5975c7cSAlex Elder return; /* Nothing to do */ 513e5975c7cSAlex Elder BUG_ON(length > previous); 514e5975c7cSAlex Elder 515e5975c7cSAlex Elder op->extent.length = length; 516e5975c7cSAlex Elder op->payload_len -= previous - length; 517e5975c7cSAlex Elder } 518e5975c7cSAlex Elder EXPORT_SYMBOL(osd_req_op_extent_update); 519e5975c7cSAlex Elder 520c99d2d4aSAlex Elder void osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which, 52104017e29SAlex Elder u16 opcode, const char *class, const char *method) 52233803f33SAlex Elder { 523144cba14SYan, Zheng struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, 524144cba14SYan, Zheng opcode, 0); 5255f562df5SAlex Elder struct ceph_pagelist *pagelist; 52633803f33SAlex Elder size_t payload_len = 0; 52733803f33SAlex Elder size_t size; 52833803f33SAlex Elder 52933803f33SAlex Elder BUG_ON(opcode != CEPH_OSD_OP_CALL); 53033803f33SAlex Elder 5315f562df5SAlex Elder pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS); 5325f562df5SAlex Elder BUG_ON(!pagelist); 5335f562df5SAlex Elder ceph_pagelist_init(pagelist); 5345f562df5SAlex Elder 53533803f33SAlex Elder op->cls.class_name = class; 53633803f33SAlex Elder size = strlen(class); 53733803f33SAlex Elder BUG_ON(size > (size_t) U8_MAX); 53833803f33SAlex Elder op->cls.class_len = size; 5395f562df5SAlex Elder ceph_pagelist_append(pagelist, class, size); 54033803f33SAlex Elder payload_len += size; 54133803f33SAlex Elder 54233803f33SAlex Elder op->cls.method_name = method; 54333803f33SAlex Elder size = strlen(method); 54433803f33SAlex Elder BUG_ON(size > (size_t) U8_MAX); 54533803f33SAlex Elder op->cls.method_len = size; 5465f562df5SAlex Elder ceph_pagelist_append(pagelist, method, size); 54733803f33SAlex Elder payload_len += size; 54833803f33SAlex Elder 549a4ce40a9SAlex Elder osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist); 5505f562df5SAlex Elder 55133803f33SAlex Elder op->cls.argc = 0; /* currently unused */ 55233803f33SAlex Elder 55333803f33SAlex Elder op->payload_len = payload_len; 55433803f33SAlex Elder } 55533803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_init); 5568c042b0dSAlex Elder 557d74b50beSYan, Zheng int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which, 558d74b50beSYan, Zheng u16 opcode, const char *name, const void *value, 559d74b50beSYan, Zheng size_t size, u8 cmp_op, u8 cmp_mode) 560d74b50beSYan, Zheng { 561144cba14SYan, Zheng struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, 562144cba14SYan, Zheng opcode, 0); 563d74b50beSYan, Zheng struct ceph_pagelist *pagelist; 564d74b50beSYan, Zheng size_t payload_len; 565d74b50beSYan, Zheng 566d74b50beSYan, Zheng BUG_ON(opcode != CEPH_OSD_OP_SETXATTR && opcode != CEPH_OSD_OP_CMPXATTR); 567d74b50beSYan, Zheng 568d74b50beSYan, Zheng pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS); 569d74b50beSYan, Zheng if (!pagelist) 570d74b50beSYan, Zheng return -ENOMEM; 571d74b50beSYan, Zheng 572d74b50beSYan, Zheng ceph_pagelist_init(pagelist); 573d74b50beSYan, Zheng 574d74b50beSYan, Zheng payload_len = strlen(name); 575d74b50beSYan, Zheng op->xattr.name_len = payload_len; 576d74b50beSYan, Zheng ceph_pagelist_append(pagelist, name, payload_len); 577d74b50beSYan, Zheng 578d74b50beSYan, Zheng op->xattr.value_len = size; 579d74b50beSYan, Zheng ceph_pagelist_append(pagelist, value, size); 580d74b50beSYan, Zheng payload_len += size; 581d74b50beSYan, Zheng 582d74b50beSYan, Zheng op->xattr.cmp_op = cmp_op; 583d74b50beSYan, Zheng op->xattr.cmp_mode = cmp_mode; 584d74b50beSYan, Zheng 585d74b50beSYan, Zheng ceph_osd_data_pagelist_init(&op->xattr.osd_data, pagelist); 586d74b50beSYan, Zheng op->payload_len = payload_len; 587d74b50beSYan, Zheng return 0; 588d74b50beSYan, Zheng } 589d74b50beSYan, Zheng EXPORT_SYMBOL(osd_req_op_xattr_init); 590d74b50beSYan, Zheng 591c99d2d4aSAlex Elder void osd_req_op_watch_init(struct ceph_osd_request *osd_req, 592c99d2d4aSAlex Elder unsigned int which, u16 opcode, 59333803f33SAlex Elder u64 cookie, u64 version, int flag) 59433803f33SAlex Elder { 595144cba14SYan, Zheng struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, 596144cba14SYan, Zheng opcode, 0); 59733803f33SAlex Elder 598c99d2d4aSAlex Elder BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH); 59933803f33SAlex Elder 60033803f33SAlex Elder op->watch.cookie = cookie; 6019ef1ee5aSAlex Elder op->watch.ver = version; 60233803f33SAlex Elder if (opcode == CEPH_OSD_OP_WATCH && flag) 60333803f33SAlex Elder op->watch.flag = (u8)1; 60433803f33SAlex Elder } 60533803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_watch_init); 60633803f33SAlex Elder 607c647b8a8SIlya Dryomov void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req, 608c647b8a8SIlya Dryomov unsigned int which, 609c647b8a8SIlya Dryomov u64 expected_object_size, 610c647b8a8SIlya Dryomov u64 expected_write_size) 611c647b8a8SIlya Dryomov { 612c647b8a8SIlya Dryomov struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, 613144cba14SYan, Zheng CEPH_OSD_OP_SETALLOCHINT, 614144cba14SYan, Zheng 0); 615c647b8a8SIlya Dryomov 616c647b8a8SIlya Dryomov op->alloc_hint.expected_object_size = expected_object_size; 617c647b8a8SIlya Dryomov op->alloc_hint.expected_write_size = expected_write_size; 618c647b8a8SIlya Dryomov 619c647b8a8SIlya Dryomov /* 620c647b8a8SIlya Dryomov * CEPH_OSD_OP_SETALLOCHINT op is advisory and therefore deemed 621c647b8a8SIlya Dryomov * not worth a feature bit. Set FAILOK per-op flag to make 622c647b8a8SIlya Dryomov * sure older osds don't trip over an unsupported opcode. 623c647b8a8SIlya Dryomov */ 624c647b8a8SIlya Dryomov op->flags |= CEPH_OSD_OP_FLAG_FAILOK; 625c647b8a8SIlya Dryomov } 626c647b8a8SIlya Dryomov EXPORT_SYMBOL(osd_req_op_alloc_hint_init); 627c647b8a8SIlya Dryomov 62890af3602SAlex Elder static void ceph_osdc_msg_data_add(struct ceph_msg *msg, 629ec9123c5SAlex Elder struct ceph_osd_data *osd_data) 630ec9123c5SAlex Elder { 631ec9123c5SAlex Elder u64 length = ceph_osd_data_length(osd_data); 632ec9123c5SAlex Elder 633ec9123c5SAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 634ec9123c5SAlex Elder BUG_ON(length > (u64) SIZE_MAX); 635ec9123c5SAlex Elder if (length) 63690af3602SAlex Elder ceph_msg_data_add_pages(msg, osd_data->pages, 637ec9123c5SAlex Elder length, osd_data->alignment); 638ec9123c5SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) { 639ec9123c5SAlex Elder BUG_ON(!length); 64090af3602SAlex Elder ceph_msg_data_add_pagelist(msg, osd_data->pagelist); 641ec9123c5SAlex Elder #ifdef CONFIG_BLOCK 642ec9123c5SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) { 64390af3602SAlex Elder ceph_msg_data_add_bio(msg, osd_data->bio, length); 644ec9123c5SAlex Elder #endif 645ec9123c5SAlex Elder } else { 646ec9123c5SAlex Elder BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE); 647ec9123c5SAlex Elder } 648ec9123c5SAlex Elder } 649ec9123c5SAlex Elder 650175face2SAlex Elder static u64 osd_req_encode_op(struct ceph_osd_request *req, 65179528734SAlex Elder struct ceph_osd_op *dst, unsigned int which) 6523d14c5d2SYehuda Sadeh { 65379528734SAlex Elder struct ceph_osd_req_op *src; 65404017e29SAlex Elder struct ceph_osd_data *osd_data; 65554d50649SAlex Elder u64 request_data_len = 0; 65604017e29SAlex Elder u64 data_length; 657175face2SAlex Elder 65879528734SAlex Elder BUG_ON(which >= req->r_num_ops); 65979528734SAlex Elder src = &req->r_ops[which]; 660a8dd0a37SAlex Elder if (WARN_ON(!osd_req_opcode_valid(src->op))) { 661a8dd0a37SAlex Elder pr_err("unrecognized osd opcode %d\n", src->op); 662a8dd0a37SAlex Elder 663a8dd0a37SAlex Elder return 0; 664a8dd0a37SAlex Elder } 6653d14c5d2SYehuda Sadeh 666065a68f9SAlex Elder switch (src->op) { 667fbfab539SAlex Elder case CEPH_OSD_OP_STAT: 66849719778SAlex Elder osd_data = &src->raw_data_in; 66949719778SAlex Elder ceph_osdc_msg_data_add(req->r_reply, osd_data); 670fbfab539SAlex Elder break; 6713d14c5d2SYehuda Sadeh case CEPH_OSD_OP_READ: 6723d14c5d2SYehuda Sadeh case CEPH_OSD_OP_WRITE: 673ad7a60deSLi Wang case CEPH_OSD_OP_ZERO: 674ad7a60deSLi Wang case CEPH_OSD_OP_TRUNCATE: 675175face2SAlex Elder if (src->op == CEPH_OSD_OP_WRITE) 67654d50649SAlex Elder request_data_len = src->extent.length; 677175face2SAlex Elder dst->extent.offset = cpu_to_le64(src->extent.offset); 678175face2SAlex Elder dst->extent.length = cpu_to_le64(src->extent.length); 6793d14c5d2SYehuda Sadeh dst->extent.truncate_size = 6803d14c5d2SYehuda Sadeh cpu_to_le64(src->extent.truncate_size); 6813d14c5d2SYehuda Sadeh dst->extent.truncate_seq = 6823d14c5d2SYehuda Sadeh cpu_to_le32(src->extent.truncate_seq); 68304017e29SAlex Elder osd_data = &src->extent.osd_data; 6845476492fSAlex Elder if (src->op == CEPH_OSD_OP_WRITE) 68504017e29SAlex Elder ceph_osdc_msg_data_add(req->r_request, osd_data); 6865476492fSAlex Elder else 68704017e29SAlex Elder ceph_osdc_msg_data_add(req->r_reply, osd_data); 6883d14c5d2SYehuda Sadeh break; 6893d14c5d2SYehuda Sadeh case CEPH_OSD_OP_CALL: 6903d14c5d2SYehuda Sadeh dst->cls.class_len = src->cls.class_len; 6913d14c5d2SYehuda Sadeh dst->cls.method_len = src->cls.method_len; 69204017e29SAlex Elder osd_data = &src->cls.request_info; 69304017e29SAlex Elder ceph_osdc_msg_data_add(req->r_request, osd_data); 69404017e29SAlex Elder BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGELIST); 69504017e29SAlex Elder request_data_len = osd_data->pagelist->length; 69604017e29SAlex Elder 69704017e29SAlex Elder osd_data = &src->cls.request_data; 69804017e29SAlex Elder data_length = ceph_osd_data_length(osd_data); 69904017e29SAlex Elder if (data_length) { 70004017e29SAlex Elder BUG_ON(osd_data->type == CEPH_OSD_DATA_TYPE_NONE); 70104017e29SAlex Elder dst->cls.indata_len = cpu_to_le32(data_length); 70204017e29SAlex Elder ceph_osdc_msg_data_add(req->r_request, osd_data); 70304017e29SAlex Elder src->payload_len += data_length; 70404017e29SAlex Elder request_data_len += data_length; 70504017e29SAlex Elder } 70604017e29SAlex Elder osd_data = &src->cls.response_data; 70704017e29SAlex Elder ceph_osdc_msg_data_add(req->r_reply, osd_data); 7083d14c5d2SYehuda Sadeh break; 7093d14c5d2SYehuda Sadeh case CEPH_OSD_OP_STARTSYNC: 7103d14c5d2SYehuda Sadeh break; 711a40c4f10SYehuda Sadeh case CEPH_OSD_OP_NOTIFY_ACK: 712a40c4f10SYehuda Sadeh case CEPH_OSD_OP_WATCH: 713a40c4f10SYehuda Sadeh dst->watch.cookie = cpu_to_le64(src->watch.cookie); 714a40c4f10SYehuda Sadeh dst->watch.ver = cpu_to_le64(src->watch.ver); 715a40c4f10SYehuda Sadeh dst->watch.flag = src->watch.flag; 716a40c4f10SYehuda Sadeh break; 717c647b8a8SIlya Dryomov case CEPH_OSD_OP_SETALLOCHINT: 718c647b8a8SIlya Dryomov dst->alloc_hint.expected_object_size = 719c647b8a8SIlya Dryomov cpu_to_le64(src->alloc_hint.expected_object_size); 720c647b8a8SIlya Dryomov dst->alloc_hint.expected_write_size = 721c647b8a8SIlya Dryomov cpu_to_le64(src->alloc_hint.expected_write_size); 722c647b8a8SIlya Dryomov break; 723d74b50beSYan, Zheng case CEPH_OSD_OP_SETXATTR: 724d74b50beSYan, Zheng case CEPH_OSD_OP_CMPXATTR: 725d74b50beSYan, Zheng dst->xattr.name_len = cpu_to_le32(src->xattr.name_len); 726d74b50beSYan, Zheng dst->xattr.value_len = cpu_to_le32(src->xattr.value_len); 727d74b50beSYan, Zheng dst->xattr.cmp_op = src->xattr.cmp_op; 728d74b50beSYan, Zheng dst->xattr.cmp_mode = src->xattr.cmp_mode; 729d74b50beSYan, Zheng osd_data = &src->xattr.osd_data; 730d74b50beSYan, Zheng ceph_osdc_msg_data_add(req->r_request, osd_data); 731d74b50beSYan, Zheng request_data_len = osd_data->pagelist->length; 732d74b50beSYan, Zheng break; 733864e9197SYan, Zheng case CEPH_OSD_OP_CREATE: 734864e9197SYan, Zheng case CEPH_OSD_OP_DELETE: 735864e9197SYan, Zheng break; 7363d14c5d2SYehuda Sadeh default: 7374c46459cSAlex Elder pr_err("unsupported osd opcode %s\n", 7388f63ca2dSAlex Elder ceph_osd_op_name(src->op)); 7394c46459cSAlex Elder WARN_ON(1); 740a8dd0a37SAlex Elder 741a8dd0a37SAlex Elder return 0; 7423d14c5d2SYehuda Sadeh } 7437b25bf5fSIlya Dryomov 744a8dd0a37SAlex Elder dst->op = cpu_to_le16(src->op); 7457b25bf5fSIlya Dryomov dst->flags = cpu_to_le32(src->flags); 7463d14c5d2SYehuda Sadeh dst->payload_len = cpu_to_le32(src->payload_len); 747175face2SAlex Elder 74854d50649SAlex Elder return request_data_len; 7493d14c5d2SYehuda Sadeh } 7503d14c5d2SYehuda Sadeh 7513d14c5d2SYehuda Sadeh /* 7523d14c5d2SYehuda Sadeh * build new request AND message, calculate layout, and adjust file 7533d14c5d2SYehuda Sadeh * extent as needed. 7543d14c5d2SYehuda Sadeh * 7553d14c5d2SYehuda Sadeh * if the file was recently truncated, we include information about its 7563d14c5d2SYehuda Sadeh * old and new size so that the object can be updated appropriately. (we 7573d14c5d2SYehuda Sadeh * avoid synchronously deleting truncated objects because it's slow.) 7583d14c5d2SYehuda Sadeh * 7593d14c5d2SYehuda Sadeh * if @do_sync, include a 'startsync' command so that the osd will flush 7603d14c5d2SYehuda Sadeh * data quickly. 7613d14c5d2SYehuda Sadeh */ 7623d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc, 7633d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 7643d14c5d2SYehuda Sadeh struct ceph_vino vino, 765715e4cd4SYan, Zheng u64 off, u64 *plen, 766715e4cd4SYan, Zheng unsigned int which, int num_ops, 7673d14c5d2SYehuda Sadeh int opcode, int flags, 7683d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 7693d14c5d2SYehuda Sadeh u32 truncate_seq, 7703d14c5d2SYehuda Sadeh u64 truncate_size, 771153e5167SAlex Elder bool use_mempool) 7723d14c5d2SYehuda Sadeh { 7733d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 77475d1c941SAlex Elder u64 objnum = 0; 77575d1c941SAlex Elder u64 objoff = 0; 77675d1c941SAlex Elder u64 objlen = 0; 7776816282dSSage Weil int r; 7783d14c5d2SYehuda Sadeh 779ad7a60deSLi Wang BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE && 780864e9197SYan, Zheng opcode != CEPH_OSD_OP_ZERO && opcode != CEPH_OSD_OP_TRUNCATE && 781864e9197SYan, Zheng opcode != CEPH_OSD_OP_CREATE && opcode != CEPH_OSD_OP_DELETE); 7823d14c5d2SYehuda Sadeh 783acead002SAlex Elder req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool, 784ae7ca4a3SAlex Elder GFP_NOFS); 7854ad12621SSage Weil if (!req) 7866816282dSSage Weil return ERR_PTR(-ENOMEM); 78779528734SAlex Elder 788d178a9e7SAlex Elder req->r_flags = flags; 7893d14c5d2SYehuda Sadeh 7903d14c5d2SYehuda Sadeh /* calculate max write size */ 791a19dadfbSAlex Elder r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen); 7923ff5f385SAlex Elder if (r < 0) { 7933ff5f385SAlex Elder ceph_osdc_put_request(req); 7946816282dSSage Weil return ERR_PTR(r); 7953ff5f385SAlex Elder } 796a19dadfbSAlex Elder 797864e9197SYan, Zheng if (opcode == CEPH_OSD_OP_CREATE || opcode == CEPH_OSD_OP_DELETE) { 798144cba14SYan, Zheng osd_req_op_init(req, which, opcode, 0); 799864e9197SYan, Zheng } else { 800864e9197SYan, Zheng u32 object_size = le32_to_cpu(layout->fl_object_size); 801864e9197SYan, Zheng u32 object_base = off - objoff; 802ccca4e37SYan, Zheng if (!(truncate_seq == 1 && truncate_size == -1ULL)) { 803d18d1e28SAlex Elder if (truncate_size <= object_base) { 804d18d1e28SAlex Elder truncate_size = 0; 805d18d1e28SAlex Elder } else { 806d18d1e28SAlex Elder truncate_size -= object_base; 807d18d1e28SAlex Elder if (truncate_size > object_size) 808d18d1e28SAlex Elder truncate_size = object_size; 809d18d1e28SAlex Elder } 810ccca4e37SYan, Zheng } 811715e4cd4SYan, Zheng osd_req_op_extent_init(req, which, opcode, objoff, objlen, 812b0270324SAlex Elder truncate_size, truncate_seq); 813864e9197SYan, Zheng } 814d18d1e28SAlex Elder 8153c972c95SIlya Dryomov req->r_base_oloc.pool = ceph_file_layout_pg_pool(*layout); 8163d14c5d2SYehuda Sadeh 8173c972c95SIlya Dryomov snprintf(req->r_base_oid.name, sizeof(req->r_base_oid.name), 8184295f221SIlya Dryomov "%llx.%08llx", vino.ino, objnum); 8193c972c95SIlya Dryomov req->r_base_oid.name_len = strlen(req->r_base_oid.name); 820dbe0fc41SAlex Elder 8213d14c5d2SYehuda Sadeh return req; 8223d14c5d2SYehuda Sadeh } 8233d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_new_request); 8243d14c5d2SYehuda Sadeh 8253d14c5d2SYehuda Sadeh /* 8263d14c5d2SYehuda Sadeh * We keep osd requests in an rbtree, sorted by ->r_tid. 8273d14c5d2SYehuda Sadeh */ 8283d14c5d2SYehuda Sadeh static void __insert_request(struct ceph_osd_client *osdc, 8293d14c5d2SYehuda Sadeh struct ceph_osd_request *new) 8303d14c5d2SYehuda Sadeh { 8313d14c5d2SYehuda Sadeh struct rb_node **p = &osdc->requests.rb_node; 8323d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 8333d14c5d2SYehuda Sadeh struct ceph_osd_request *req = NULL; 8343d14c5d2SYehuda Sadeh 8353d14c5d2SYehuda Sadeh while (*p) { 8363d14c5d2SYehuda Sadeh parent = *p; 8373d14c5d2SYehuda Sadeh req = rb_entry(parent, struct ceph_osd_request, r_node); 8383d14c5d2SYehuda Sadeh if (new->r_tid < req->r_tid) 8393d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 8403d14c5d2SYehuda Sadeh else if (new->r_tid > req->r_tid) 8413d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 8423d14c5d2SYehuda Sadeh else 8433d14c5d2SYehuda Sadeh BUG(); 8443d14c5d2SYehuda Sadeh } 8453d14c5d2SYehuda Sadeh 8463d14c5d2SYehuda Sadeh rb_link_node(&new->r_node, parent, p); 8473d14c5d2SYehuda Sadeh rb_insert_color(&new->r_node, &osdc->requests); 8483d14c5d2SYehuda Sadeh } 8493d14c5d2SYehuda Sadeh 8503d14c5d2SYehuda Sadeh static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc, 8513d14c5d2SYehuda Sadeh u64 tid) 8523d14c5d2SYehuda Sadeh { 8533d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 8543d14c5d2SYehuda Sadeh struct rb_node *n = osdc->requests.rb_node; 8553d14c5d2SYehuda Sadeh 8563d14c5d2SYehuda Sadeh while (n) { 8573d14c5d2SYehuda Sadeh req = rb_entry(n, struct ceph_osd_request, r_node); 8583d14c5d2SYehuda Sadeh if (tid < req->r_tid) 8593d14c5d2SYehuda Sadeh n = n->rb_left; 8603d14c5d2SYehuda Sadeh else if (tid > req->r_tid) 8613d14c5d2SYehuda Sadeh n = n->rb_right; 8623d14c5d2SYehuda Sadeh else 8633d14c5d2SYehuda Sadeh return req; 8643d14c5d2SYehuda Sadeh } 8653d14c5d2SYehuda Sadeh return NULL; 8663d14c5d2SYehuda Sadeh } 8673d14c5d2SYehuda Sadeh 8683d14c5d2SYehuda Sadeh static struct ceph_osd_request * 8693d14c5d2SYehuda Sadeh __lookup_request_ge(struct ceph_osd_client *osdc, 8703d14c5d2SYehuda Sadeh u64 tid) 8713d14c5d2SYehuda Sadeh { 8723d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 8733d14c5d2SYehuda Sadeh struct rb_node *n = osdc->requests.rb_node; 8743d14c5d2SYehuda Sadeh 8753d14c5d2SYehuda Sadeh while (n) { 8763d14c5d2SYehuda Sadeh req = rb_entry(n, struct ceph_osd_request, r_node); 8773d14c5d2SYehuda Sadeh if (tid < req->r_tid) { 8783d14c5d2SYehuda Sadeh if (!n->rb_left) 8793d14c5d2SYehuda Sadeh return req; 8803d14c5d2SYehuda Sadeh n = n->rb_left; 8813d14c5d2SYehuda Sadeh } else if (tid > req->r_tid) { 8823d14c5d2SYehuda Sadeh n = n->rb_right; 8833d14c5d2SYehuda Sadeh } else { 8843d14c5d2SYehuda Sadeh return req; 8853d14c5d2SYehuda Sadeh } 8863d14c5d2SYehuda Sadeh } 8873d14c5d2SYehuda Sadeh return NULL; 8883d14c5d2SYehuda Sadeh } 8893d14c5d2SYehuda Sadeh 8902cc6128aSIlya Dryomov static void __kick_linger_request(struct ceph_osd_request *req) 8912cc6128aSIlya Dryomov { 8922cc6128aSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 8932cc6128aSIlya Dryomov struct ceph_osd *osd = req->r_osd; 8942cc6128aSIlya Dryomov 8952cc6128aSIlya Dryomov /* 8962cc6128aSIlya Dryomov * Linger requests need to be resent with a new tid to avoid 8972cc6128aSIlya Dryomov * the dup op detection logic on the OSDs. Achieve this with 8982cc6128aSIlya Dryomov * a re-register dance instead of open-coding. 8992cc6128aSIlya Dryomov */ 9002cc6128aSIlya Dryomov ceph_osdc_get_request(req); 9012cc6128aSIlya Dryomov if (!list_empty(&req->r_linger_item)) 9022cc6128aSIlya Dryomov __unregister_linger_request(osdc, req); 9032cc6128aSIlya Dryomov else 9042cc6128aSIlya Dryomov __unregister_request(osdc, req); 9052cc6128aSIlya Dryomov __register_request(osdc, req); 9062cc6128aSIlya Dryomov ceph_osdc_put_request(req); 9072cc6128aSIlya Dryomov 9082cc6128aSIlya Dryomov /* 9092cc6128aSIlya Dryomov * Unless request has been registered as both normal and 9102cc6128aSIlya Dryomov * lingering, __unregister{,_linger}_request clears r_osd. 9112cc6128aSIlya Dryomov * However, here we need to preserve r_osd to make sure we 9122cc6128aSIlya Dryomov * requeue on the same OSD. 9132cc6128aSIlya Dryomov */ 9142cc6128aSIlya Dryomov WARN_ON(req->r_osd || !osd); 9152cc6128aSIlya Dryomov req->r_osd = osd; 9162cc6128aSIlya Dryomov 9172cc6128aSIlya Dryomov dout("%s requeueing %p tid %llu\n", __func__, req, req->r_tid); 9182cc6128aSIlya Dryomov __enqueue_request(req); 9192cc6128aSIlya Dryomov } 9202cc6128aSIlya Dryomov 9216f6c7006SSage Weil /* 9226f6c7006SSage Weil * Resubmit requests pending on the given osd. 9236f6c7006SSage Weil */ 9246f6c7006SSage Weil static void __kick_osd_requests(struct ceph_osd_client *osdc, 9256f6c7006SSage Weil struct ceph_osd *osd) 9266f6c7006SSage Weil { 927a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 928e02493c0SAlex Elder LIST_HEAD(resend); 9292cc6128aSIlya Dryomov LIST_HEAD(resend_linger); 9306f6c7006SSage Weil int err; 9316f6c7006SSage Weil 9322cc6128aSIlya Dryomov dout("%s osd%d\n", __func__, osd->o_osd); 9336f6c7006SSage Weil err = __reset_osd(osdc, osd); 934685a7555SAlex Elder if (err) 9356f6c7006SSage Weil return; 9362cc6128aSIlya Dryomov 937e02493c0SAlex Elder /* 938e02493c0SAlex Elder * Build up a list of requests to resend by traversing the 939e02493c0SAlex Elder * osd's list of requests. Requests for a given object are 940e02493c0SAlex Elder * sent in tid order, and that is also the order they're 941e02493c0SAlex Elder * kept on this list. Therefore all requests that are in 942e02493c0SAlex Elder * flight will be found first, followed by all requests that 943e02493c0SAlex Elder * have not yet been sent. And to resend requests while 944e02493c0SAlex Elder * preserving this order we will want to put any sent 945e02493c0SAlex Elder * requests back on the front of the osd client's unsent 946e02493c0SAlex Elder * list. 947e02493c0SAlex Elder * 948e02493c0SAlex Elder * So we build a separate ordered list of already-sent 949e02493c0SAlex Elder * requests for the affected osd and splice it onto the 950e02493c0SAlex Elder * front of the osd client's unsent list. Once we've seen a 951e02493c0SAlex Elder * request that has not yet been sent we're done. Those 952e02493c0SAlex Elder * requests are already sitting right where they belong. 953e02493c0SAlex Elder */ 9546f6c7006SSage Weil list_for_each_entry(req, &osd->o_requests, r_osd_item) { 955e02493c0SAlex Elder if (!req->r_sent) 956e02493c0SAlex Elder break; 9572cc6128aSIlya Dryomov 9582cc6128aSIlya Dryomov if (!req->r_linger) { 9592cc6128aSIlya Dryomov dout("%s requeueing %p tid %llu\n", __func__, req, 9602cc6128aSIlya Dryomov req->r_tid); 961e02493c0SAlex Elder list_move_tail(&req->r_req_lru_item, &resend); 9626f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 9632cc6128aSIlya Dryomov } else { 9642cc6128aSIlya Dryomov list_move_tail(&req->r_req_lru_item, &resend_linger); 9652cc6128aSIlya Dryomov } 9666f6c7006SSage Weil } 967e02493c0SAlex Elder list_splice(&resend, &osdc->req_unsent); 968a40c4f10SYehuda Sadeh 969e02493c0SAlex Elder /* 9702cc6128aSIlya Dryomov * Both registered and not yet registered linger requests are 9712cc6128aSIlya Dryomov * enqueued with a new tid on the same OSD. We add/move them 9722cc6128aSIlya Dryomov * to req_unsent/o_requests at the end to keep things in tid 9732cc6128aSIlya Dryomov * order. 974e02493c0SAlex Elder */ 975a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osd->o_linger_requests, 9761d0326b1SIlya Dryomov r_linger_osd_item) { 9772cc6128aSIlya Dryomov WARN_ON(!list_empty(&req->r_req_lru_item)); 9782cc6128aSIlya Dryomov __kick_linger_request(req); 979a40c4f10SYehuda Sadeh } 9802cc6128aSIlya Dryomov 9812cc6128aSIlya Dryomov list_for_each_entry_safe(req, nreq, &resend_linger, r_req_lru_item) 9822cc6128aSIlya Dryomov __kick_linger_request(req); 9836f6c7006SSage Weil } 9846f6c7006SSage Weil 9853d14c5d2SYehuda Sadeh /* 9863d14c5d2SYehuda Sadeh * If the osd connection drops, we need to resubmit all requests. 9873d14c5d2SYehuda Sadeh */ 9883d14c5d2SYehuda Sadeh static void osd_reset(struct ceph_connection *con) 9893d14c5d2SYehuda Sadeh { 9903d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 9913d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 9923d14c5d2SYehuda Sadeh 9933d14c5d2SYehuda Sadeh if (!osd) 9943d14c5d2SYehuda Sadeh return; 9953d14c5d2SYehuda Sadeh dout("osd_reset osd%d\n", osd->o_osd); 9963d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 9973d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 99883aff95eSSage Weil mutex_lock(&osdc->request_mutex); 99983aff95eSSage Weil __kick_osd_requests(osdc, osd); 1000f9d25199SAlex Elder __send_queued(osdc); 100183aff95eSSage Weil mutex_unlock(&osdc->request_mutex); 10023d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 10033d14c5d2SYehuda Sadeh } 10043d14c5d2SYehuda Sadeh 10053d14c5d2SYehuda Sadeh /* 10063d14c5d2SYehuda Sadeh * Track open sessions with osds. 10073d14c5d2SYehuda Sadeh */ 1008e10006f8SAlex Elder static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum) 10093d14c5d2SYehuda Sadeh { 10103d14c5d2SYehuda Sadeh struct ceph_osd *osd; 10113d14c5d2SYehuda Sadeh 10123d14c5d2SYehuda Sadeh osd = kzalloc(sizeof(*osd), GFP_NOFS); 10133d14c5d2SYehuda Sadeh if (!osd) 10143d14c5d2SYehuda Sadeh return NULL; 10153d14c5d2SYehuda Sadeh 10163d14c5d2SYehuda Sadeh atomic_set(&osd->o_ref, 1); 10173d14c5d2SYehuda Sadeh osd->o_osdc = osdc; 1018e10006f8SAlex Elder osd->o_osd = onum; 1019f407731dSAlex Elder RB_CLEAR_NODE(&osd->o_node); 10203d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_requests); 1021a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osd->o_linger_requests); 10223d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_osd_lru); 10233d14c5d2SYehuda Sadeh osd->o_incarnation = 1; 10243d14c5d2SYehuda Sadeh 1025b7a9e5ddSSage Weil ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr); 10263d14c5d2SYehuda Sadeh 10273d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osd->o_keepalive_item); 10283d14c5d2SYehuda Sadeh return osd; 10293d14c5d2SYehuda Sadeh } 10303d14c5d2SYehuda Sadeh 10313d14c5d2SYehuda Sadeh static struct ceph_osd *get_osd(struct ceph_osd *osd) 10323d14c5d2SYehuda Sadeh { 10333d14c5d2SYehuda Sadeh if (atomic_inc_not_zero(&osd->o_ref)) { 10343d14c5d2SYehuda Sadeh dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1, 10353d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref)); 10363d14c5d2SYehuda Sadeh return osd; 10373d14c5d2SYehuda Sadeh } else { 10383d14c5d2SYehuda Sadeh dout("get_osd %p FAIL\n", osd); 10393d14c5d2SYehuda Sadeh return NULL; 10403d14c5d2SYehuda Sadeh } 10413d14c5d2SYehuda Sadeh } 10423d14c5d2SYehuda Sadeh 10433d14c5d2SYehuda Sadeh static void put_osd(struct ceph_osd *osd) 10443d14c5d2SYehuda Sadeh { 10453d14c5d2SYehuda Sadeh dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref), 10463d14c5d2SYehuda Sadeh atomic_read(&osd->o_ref) - 1); 1047b28ec2f3SIlya Dryomov if (atomic_dec_and_test(&osd->o_ref)) { 10483d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth; 10493d14c5d2SYehuda Sadeh 1050b28ec2f3SIlya Dryomov if (osd->o_auth.authorizer) 105127859f97SSage Weil ceph_auth_destroy_authorizer(ac, osd->o_auth.authorizer); 10523d14c5d2SYehuda Sadeh kfree(osd); 10533d14c5d2SYehuda Sadeh } 10543d14c5d2SYehuda Sadeh } 10553d14c5d2SYehuda Sadeh 10563d14c5d2SYehuda Sadeh /* 10573d14c5d2SYehuda Sadeh * remove an osd from our map 10583d14c5d2SYehuda Sadeh */ 10593d14c5d2SYehuda Sadeh static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 10603d14c5d2SYehuda Sadeh { 10617eb71e03SIlya Dryomov dout("%s %p osd%d\n", __func__, osd, osd->o_osd); 1062cc9f1f51SIlya Dryomov WARN_ON(!list_empty(&osd->o_requests)); 1063cc9f1f51SIlya Dryomov WARN_ON(!list_empty(&osd->o_linger_requests)); 10647c6e6fc5SIlya Dryomov 10653d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 10667eb71e03SIlya Dryomov rb_erase(&osd->o_node, &osdc->osds); 10677eb71e03SIlya Dryomov RB_CLEAR_NODE(&osd->o_node); 10687eb71e03SIlya Dryomov } 10697eb71e03SIlya Dryomov 10707eb71e03SIlya Dryomov static void remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 10717eb71e03SIlya Dryomov { 10727eb71e03SIlya Dryomov dout("%s %p osd%d\n", __func__, osd, osd->o_osd); 10737eb71e03SIlya Dryomov 10747eb71e03SIlya Dryomov if (!RB_EMPTY_NODE(&osd->o_node)) { 10753d14c5d2SYehuda Sadeh ceph_con_close(&osd->o_con); 10767eb71e03SIlya Dryomov __remove_osd(osdc, osd); 10773d14c5d2SYehuda Sadeh put_osd(osd); 10783d14c5d2SYehuda Sadeh } 10797eb71e03SIlya Dryomov } 10803d14c5d2SYehuda Sadeh 1081aca420bcSSage Weil static void remove_all_osds(struct ceph_osd_client *osdc) 1082aca420bcSSage Weil { 1083048a9d2dSJiaju Zhang dout("%s %p\n", __func__, osdc); 1084aca420bcSSage Weil mutex_lock(&osdc->request_mutex); 1085aca420bcSSage Weil while (!RB_EMPTY_ROOT(&osdc->osds)) { 1086aca420bcSSage Weil struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds), 1087aca420bcSSage Weil struct ceph_osd, o_node); 10887eb71e03SIlya Dryomov remove_osd(osdc, osd); 1089aca420bcSSage Weil } 1090aca420bcSSage Weil mutex_unlock(&osdc->request_mutex); 1091aca420bcSSage Weil } 1092aca420bcSSage Weil 10933d14c5d2SYehuda Sadeh static void __move_osd_to_lru(struct ceph_osd_client *osdc, 10943d14c5d2SYehuda Sadeh struct ceph_osd *osd) 10953d14c5d2SYehuda Sadeh { 1096bbf37ec3SIlya Dryomov dout("%s %p\n", __func__, osd); 10973d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_osd_lru)); 1098bbf37ec3SIlya Dryomov 10993d14c5d2SYehuda Sadeh list_add_tail(&osd->o_osd_lru, &osdc->osd_lru); 1100a319bf56SIlya Dryomov osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl; 11013d14c5d2SYehuda Sadeh } 11023d14c5d2SYehuda Sadeh 1103bbf37ec3SIlya Dryomov static void maybe_move_osd_to_lru(struct ceph_osd_client *osdc, 1104bbf37ec3SIlya Dryomov struct ceph_osd *osd) 1105bbf37ec3SIlya Dryomov { 1106bbf37ec3SIlya Dryomov dout("%s %p\n", __func__, osd); 1107bbf37ec3SIlya Dryomov 1108bbf37ec3SIlya Dryomov if (list_empty(&osd->o_requests) && 1109bbf37ec3SIlya Dryomov list_empty(&osd->o_linger_requests)) 1110bbf37ec3SIlya Dryomov __move_osd_to_lru(osdc, osd); 1111bbf37ec3SIlya Dryomov } 1112bbf37ec3SIlya Dryomov 11133d14c5d2SYehuda Sadeh static void __remove_osd_from_lru(struct ceph_osd *osd) 11143d14c5d2SYehuda Sadeh { 11153d14c5d2SYehuda Sadeh dout("__remove_osd_from_lru %p\n", osd); 11163d14c5d2SYehuda Sadeh if (!list_empty(&osd->o_osd_lru)) 11173d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 11183d14c5d2SYehuda Sadeh } 11193d14c5d2SYehuda Sadeh 1120aca420bcSSage Weil static void remove_old_osds(struct ceph_osd_client *osdc) 11213d14c5d2SYehuda Sadeh { 11223d14c5d2SYehuda Sadeh struct ceph_osd *osd, *nosd; 11233d14c5d2SYehuda Sadeh 11243d14c5d2SYehuda Sadeh dout("__remove_old_osds %p\n", osdc); 11253d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 11263d14c5d2SYehuda Sadeh list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) { 1127aca420bcSSage Weil if (time_before(jiffies, osd->lru_ttl)) 11283d14c5d2SYehuda Sadeh break; 11297eb71e03SIlya Dryomov remove_osd(osdc, osd); 11303d14c5d2SYehuda Sadeh } 11313d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 11323d14c5d2SYehuda Sadeh } 11333d14c5d2SYehuda Sadeh 11343d14c5d2SYehuda Sadeh /* 11353d14c5d2SYehuda Sadeh * reset osd connect 11363d14c5d2SYehuda Sadeh */ 11373d14c5d2SYehuda Sadeh static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd) 11383d14c5d2SYehuda Sadeh { 1139c3acb181SAlex Elder struct ceph_entity_addr *peer_addr; 11403d14c5d2SYehuda Sadeh 11413d14c5d2SYehuda Sadeh dout("__reset_osd %p osd%d\n", osd, osd->o_osd); 1142a40c4f10SYehuda Sadeh if (list_empty(&osd->o_requests) && 1143a40c4f10SYehuda Sadeh list_empty(&osd->o_linger_requests)) { 11447eb71e03SIlya Dryomov remove_osd(osdc, osd); 1145c3acb181SAlex Elder return -ENODEV; 1146c3acb181SAlex Elder } 1147c3acb181SAlex Elder 1148c3acb181SAlex Elder peer_addr = &osdc->osdmap->osd_addr[osd->o_osd]; 1149c3acb181SAlex Elder if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) && 11503d14c5d2SYehuda Sadeh !ceph_con_opened(&osd->o_con)) { 1151c3acb181SAlex Elder struct ceph_osd_request *req; 1152c3acb181SAlex Elder 11533d14c5d2SYehuda Sadeh dout("osd addr hasn't changed and connection never opened, " 11540b4af2e8SIlya Dryomov "letting msgr retry\n"); 11553d14c5d2SYehuda Sadeh /* touch each r_stamp for handle_timeout()'s benfit */ 11563d14c5d2SYehuda Sadeh list_for_each_entry(req, &osd->o_requests, r_osd_item) 11573d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 1158c3acb181SAlex Elder 1159c3acb181SAlex Elder return -EAGAIN; 11603d14c5d2SYehuda Sadeh } 1161c3acb181SAlex Elder 1162c3acb181SAlex Elder ceph_con_close(&osd->o_con); 1163c3acb181SAlex Elder ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr); 1164c3acb181SAlex Elder osd->o_incarnation++; 1165c3acb181SAlex Elder 1166c3acb181SAlex Elder return 0; 11673d14c5d2SYehuda Sadeh } 11683d14c5d2SYehuda Sadeh 11693d14c5d2SYehuda Sadeh static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new) 11703d14c5d2SYehuda Sadeh { 11713d14c5d2SYehuda Sadeh struct rb_node **p = &osdc->osds.rb_node; 11723d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 11733d14c5d2SYehuda Sadeh struct ceph_osd *osd = NULL; 11743d14c5d2SYehuda Sadeh 1175aca420bcSSage Weil dout("__insert_osd %p osd%d\n", new, new->o_osd); 11763d14c5d2SYehuda Sadeh while (*p) { 11773d14c5d2SYehuda Sadeh parent = *p; 11783d14c5d2SYehuda Sadeh osd = rb_entry(parent, struct ceph_osd, o_node); 11793d14c5d2SYehuda Sadeh if (new->o_osd < osd->o_osd) 11803d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 11813d14c5d2SYehuda Sadeh else if (new->o_osd > osd->o_osd) 11823d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 11833d14c5d2SYehuda Sadeh else 11843d14c5d2SYehuda Sadeh BUG(); 11853d14c5d2SYehuda Sadeh } 11863d14c5d2SYehuda Sadeh 11873d14c5d2SYehuda Sadeh rb_link_node(&new->o_node, parent, p); 11883d14c5d2SYehuda Sadeh rb_insert_color(&new->o_node, &osdc->osds); 11893d14c5d2SYehuda Sadeh } 11903d14c5d2SYehuda Sadeh 11913d14c5d2SYehuda Sadeh static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o) 11923d14c5d2SYehuda Sadeh { 11933d14c5d2SYehuda Sadeh struct ceph_osd *osd; 11943d14c5d2SYehuda Sadeh struct rb_node *n = osdc->osds.rb_node; 11953d14c5d2SYehuda Sadeh 11963d14c5d2SYehuda Sadeh while (n) { 11973d14c5d2SYehuda Sadeh osd = rb_entry(n, struct ceph_osd, o_node); 11983d14c5d2SYehuda Sadeh if (o < osd->o_osd) 11993d14c5d2SYehuda Sadeh n = n->rb_left; 12003d14c5d2SYehuda Sadeh else if (o > osd->o_osd) 12013d14c5d2SYehuda Sadeh n = n->rb_right; 12023d14c5d2SYehuda Sadeh else 12033d14c5d2SYehuda Sadeh return osd; 12043d14c5d2SYehuda Sadeh } 12053d14c5d2SYehuda Sadeh return NULL; 12063d14c5d2SYehuda Sadeh } 12073d14c5d2SYehuda Sadeh 12083d14c5d2SYehuda Sadeh static void __schedule_osd_timeout(struct ceph_osd_client *osdc) 12093d14c5d2SYehuda Sadeh { 12103d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->timeout_work, 1211a319bf56SIlya Dryomov osdc->client->options->osd_keepalive_timeout); 12123d14c5d2SYehuda Sadeh } 12133d14c5d2SYehuda Sadeh 12143d14c5d2SYehuda Sadeh static void __cancel_osd_timeout(struct ceph_osd_client *osdc) 12153d14c5d2SYehuda Sadeh { 12163d14c5d2SYehuda Sadeh cancel_delayed_work(&osdc->timeout_work); 12173d14c5d2SYehuda Sadeh } 12183d14c5d2SYehuda Sadeh 12193d14c5d2SYehuda Sadeh /* 12203d14c5d2SYehuda Sadeh * Register request, assign tid. If this is the first request, set up 12213d14c5d2SYehuda Sadeh * the timeout event. 12223d14c5d2SYehuda Sadeh */ 1223a40c4f10SYehuda Sadeh static void __register_request(struct ceph_osd_client *osdc, 12243d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 12253d14c5d2SYehuda Sadeh { 12263d14c5d2SYehuda Sadeh req->r_tid = ++osdc->last_tid; 12273d14c5d2SYehuda Sadeh req->r_request->hdr.tid = cpu_to_le64(req->r_tid); 122877f38e0eSSage Weil dout("__register_request %p tid %lld\n", req, req->r_tid); 12293d14c5d2SYehuda Sadeh __insert_request(osdc, req); 12303d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 12313d14c5d2SYehuda Sadeh osdc->num_requests++; 12323d14c5d2SYehuda Sadeh if (osdc->num_requests == 1) { 12333d14c5d2SYehuda Sadeh dout(" first request, scheduling timeout\n"); 12343d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 12353d14c5d2SYehuda Sadeh } 1236a40c4f10SYehuda Sadeh } 1237a40c4f10SYehuda Sadeh 12383d14c5d2SYehuda Sadeh /* 12393d14c5d2SYehuda Sadeh * called under osdc->request_mutex 12403d14c5d2SYehuda Sadeh */ 12413d14c5d2SYehuda Sadeh static void __unregister_request(struct ceph_osd_client *osdc, 12423d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 12433d14c5d2SYehuda Sadeh { 124435f9f8a0SSage Weil if (RB_EMPTY_NODE(&req->r_node)) { 124535f9f8a0SSage Weil dout("__unregister_request %p tid %lld not registered\n", 124635f9f8a0SSage Weil req, req->r_tid); 124735f9f8a0SSage Weil return; 124835f9f8a0SSage Weil } 124935f9f8a0SSage Weil 12503d14c5d2SYehuda Sadeh dout("__unregister_request %p tid %lld\n", req, req->r_tid); 12513d14c5d2SYehuda Sadeh rb_erase(&req->r_node, &osdc->requests); 12526562d661SIlya Dryomov RB_CLEAR_NODE(&req->r_node); 12533d14c5d2SYehuda Sadeh osdc->num_requests--; 12543d14c5d2SYehuda Sadeh 12553d14c5d2SYehuda Sadeh if (req->r_osd) { 12563d14c5d2SYehuda Sadeh /* make sure the original request isn't in flight. */ 12576740a845SAlex Elder ceph_msg_revoke(req->r_request); 12583d14c5d2SYehuda Sadeh 12593d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 1260bbf37ec3SIlya Dryomov maybe_move_osd_to_lru(osdc, req->r_osd); 12614f23409eSIlya Dryomov if (list_empty(&req->r_linger_osd_item)) 12623d14c5d2SYehuda Sadeh req->r_osd = NULL; 12633d14c5d2SYehuda Sadeh } 12643d14c5d2SYehuda Sadeh 12657d5f2481SAlex Elder list_del_init(&req->r_req_lru_item); 12663d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 12673d14c5d2SYehuda Sadeh 12683d14c5d2SYehuda Sadeh if (osdc->num_requests == 0) { 12693d14c5d2SYehuda Sadeh dout(" no requests, canceling timeout\n"); 12703d14c5d2SYehuda Sadeh __cancel_osd_timeout(osdc); 12713d14c5d2SYehuda Sadeh } 12723d14c5d2SYehuda Sadeh } 12733d14c5d2SYehuda Sadeh 12743d14c5d2SYehuda Sadeh /* 12753d14c5d2SYehuda Sadeh * Cancel a previously queued request message 12763d14c5d2SYehuda Sadeh */ 12773d14c5d2SYehuda Sadeh static void __cancel_request(struct ceph_osd_request *req) 12783d14c5d2SYehuda Sadeh { 12793d14c5d2SYehuda Sadeh if (req->r_sent && req->r_osd) { 12806740a845SAlex Elder ceph_msg_revoke(req->r_request); 12813d14c5d2SYehuda Sadeh req->r_sent = 0; 12823d14c5d2SYehuda Sadeh } 12833d14c5d2SYehuda Sadeh } 12843d14c5d2SYehuda Sadeh 1285a40c4f10SYehuda Sadeh static void __register_linger_request(struct ceph_osd_client *osdc, 1286a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1287a40c4f10SYehuda Sadeh { 1288af593064SIlya Dryomov dout("%s %p tid %llu\n", __func__, req, req->r_tid); 1289af593064SIlya Dryomov WARN_ON(!req->r_linger); 1290af593064SIlya Dryomov 129196e4dac6SAlex Elder ceph_osdc_get_request(req); 1292a40c4f10SYehuda Sadeh list_add_tail(&req->r_linger_item, &osdc->req_linger); 12936194ea89SSage Weil if (req->r_osd) 12941d0326b1SIlya Dryomov list_add_tail(&req->r_linger_osd_item, 12956194ea89SSage Weil &req->r_osd->o_linger_requests); 1296a40c4f10SYehuda Sadeh } 1297a40c4f10SYehuda Sadeh 1298a40c4f10SYehuda Sadeh static void __unregister_linger_request(struct ceph_osd_client *osdc, 1299a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1300a40c4f10SYehuda Sadeh { 1301af593064SIlya Dryomov WARN_ON(!req->r_linger); 1302af593064SIlya Dryomov 1303af593064SIlya Dryomov if (list_empty(&req->r_linger_item)) { 1304af593064SIlya Dryomov dout("%s %p tid %llu not registered\n", __func__, req, 1305af593064SIlya Dryomov req->r_tid); 1306af593064SIlya Dryomov return; 1307af593064SIlya Dryomov } 1308af593064SIlya Dryomov 1309af593064SIlya Dryomov dout("%s %p tid %llu\n", __func__, req, req->r_tid); 1310a40c4f10SYehuda Sadeh list_del_init(&req->r_linger_item); 1311af593064SIlya Dryomov 131261c74035SAlex Elder if (req->r_osd) { 13131d0326b1SIlya Dryomov list_del_init(&req->r_linger_osd_item); 1314bbf37ec3SIlya Dryomov maybe_move_osd_to_lru(osdc, req->r_osd); 1315fbdb9190SSage Weil if (list_empty(&req->r_osd_item)) 1316a40c4f10SYehuda Sadeh req->r_osd = NULL; 1317a40c4f10SYehuda Sadeh } 131896e4dac6SAlex Elder ceph_osdc_put_request(req); 1319a40c4f10SYehuda Sadeh } 1320a40c4f10SYehuda Sadeh 1321a40c4f10SYehuda Sadeh void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc, 1322a40c4f10SYehuda Sadeh struct ceph_osd_request *req) 1323a40c4f10SYehuda Sadeh { 1324a40c4f10SYehuda Sadeh if (!req->r_linger) { 1325a40c4f10SYehuda Sadeh dout("set_request_linger %p\n", req); 1326a40c4f10SYehuda Sadeh req->r_linger = 1; 1327a40c4f10SYehuda Sadeh } 1328a40c4f10SYehuda Sadeh } 1329a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_set_request_linger); 1330a40c4f10SYehuda Sadeh 13313d14c5d2SYehuda Sadeh /* 1332d29adb34SJosh Durgin * Returns whether a request should be blocked from being sent 1333d29adb34SJosh Durgin * based on the current osdmap and osd_client settings. 1334d29adb34SJosh Durgin * 1335d29adb34SJosh Durgin * Caller should hold map_sem for read. 1336d29adb34SJosh Durgin */ 1337d29adb34SJosh Durgin static bool __req_should_be_paused(struct ceph_osd_client *osdc, 1338d29adb34SJosh Durgin struct ceph_osd_request *req) 1339d29adb34SJosh Durgin { 1340d29adb34SJosh Durgin bool pauserd = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD); 1341d29adb34SJosh Durgin bool pausewr = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR) || 1342d29adb34SJosh Durgin ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL); 1343d29adb34SJosh Durgin return (req->r_flags & CEPH_OSD_FLAG_READ && pauserd) || 1344d29adb34SJosh Durgin (req->r_flags & CEPH_OSD_FLAG_WRITE && pausewr); 1345d29adb34SJosh Durgin } 1346d29adb34SJosh Durgin 1347d29adb34SJosh Durgin /* 134817a13e40SIlya Dryomov * Calculate mapping of a request to a PG. Takes tiering into account. 134917a13e40SIlya Dryomov */ 135017a13e40SIlya Dryomov static int __calc_request_pg(struct ceph_osdmap *osdmap, 135117a13e40SIlya Dryomov struct ceph_osd_request *req, 135217a13e40SIlya Dryomov struct ceph_pg *pg_out) 135317a13e40SIlya Dryomov { 1354205ee118SIlya Dryomov bool need_check_tiering; 1355205ee118SIlya Dryomov 1356205ee118SIlya Dryomov need_check_tiering = false; 1357205ee118SIlya Dryomov if (req->r_target_oloc.pool == -1) { 1358205ee118SIlya Dryomov req->r_target_oloc = req->r_base_oloc; /* struct */ 1359205ee118SIlya Dryomov need_check_tiering = true; 1360205ee118SIlya Dryomov } 1361205ee118SIlya Dryomov if (req->r_target_oid.name_len == 0) { 1362205ee118SIlya Dryomov ceph_oid_copy(&req->r_target_oid, &req->r_base_oid); 1363205ee118SIlya Dryomov need_check_tiering = true; 1364205ee118SIlya Dryomov } 1365205ee118SIlya Dryomov 1366205ee118SIlya Dryomov if (need_check_tiering && 1367205ee118SIlya Dryomov (req->r_flags & CEPH_OSD_FLAG_IGNORE_OVERLAY) == 0) { 136817a13e40SIlya Dryomov struct ceph_pg_pool_info *pi; 136917a13e40SIlya Dryomov 1370205ee118SIlya Dryomov pi = ceph_pg_pool_by_id(osdmap, req->r_target_oloc.pool); 137117a13e40SIlya Dryomov if (pi) { 137217a13e40SIlya Dryomov if ((req->r_flags & CEPH_OSD_FLAG_READ) && 137317a13e40SIlya Dryomov pi->read_tier >= 0) 1374205ee118SIlya Dryomov req->r_target_oloc.pool = pi->read_tier; 137517a13e40SIlya Dryomov if ((req->r_flags & CEPH_OSD_FLAG_WRITE) && 137617a13e40SIlya Dryomov pi->write_tier >= 0) 1377205ee118SIlya Dryomov req->r_target_oloc.pool = pi->write_tier; 137817a13e40SIlya Dryomov } 137917a13e40SIlya Dryomov /* !pi is caught in ceph_oloc_oid_to_pg() */ 138017a13e40SIlya Dryomov } 138117a13e40SIlya Dryomov 1382205ee118SIlya Dryomov return ceph_oloc_oid_to_pg(osdmap, &req->r_target_oloc, 1383205ee118SIlya Dryomov &req->r_target_oid, pg_out); 138417a13e40SIlya Dryomov } 138517a13e40SIlya Dryomov 1386f671b581SIlya Dryomov static void __enqueue_request(struct ceph_osd_request *req) 1387f671b581SIlya Dryomov { 1388f671b581SIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 1389f671b581SIlya Dryomov 1390f671b581SIlya Dryomov dout("%s %p tid %llu to osd%d\n", __func__, req, req->r_tid, 1391f671b581SIlya Dryomov req->r_osd ? req->r_osd->o_osd : -1); 1392f671b581SIlya Dryomov 1393f671b581SIlya Dryomov if (req->r_osd) { 1394f671b581SIlya Dryomov __remove_osd_from_lru(req->r_osd); 1395f671b581SIlya Dryomov list_add_tail(&req->r_osd_item, &req->r_osd->o_requests); 1396f671b581SIlya Dryomov list_move_tail(&req->r_req_lru_item, &osdc->req_unsent); 1397f671b581SIlya Dryomov } else { 1398f671b581SIlya Dryomov list_move_tail(&req->r_req_lru_item, &osdc->req_notarget); 1399f671b581SIlya Dryomov } 1400f671b581SIlya Dryomov } 1401f671b581SIlya Dryomov 140217a13e40SIlya Dryomov /* 14033d14c5d2SYehuda Sadeh * Pick an osd (the first 'up' osd in the pg), allocate the osd struct 14043d14c5d2SYehuda Sadeh * (as needed), and set the request r_osd appropriately. If there is 140525985edcSLucas De Marchi * no up osd, set r_osd to NULL. Move the request to the appropriate list 14066f6c7006SSage Weil * (unsent, homeless) or leave on in-flight lru. 14073d14c5d2SYehuda Sadeh * 14083d14c5d2SYehuda Sadeh * Return 0 if unchanged, 1 if changed, or negative on error. 14093d14c5d2SYehuda Sadeh * 14103d14c5d2SYehuda Sadeh * Caller should hold map_sem for read and request_mutex. 14113d14c5d2SYehuda Sadeh */ 14126f6c7006SSage Weil static int __map_request(struct ceph_osd_client *osdc, 141338d6453cSSage Weil struct ceph_osd_request *req, int force_resend) 14143d14c5d2SYehuda Sadeh { 14155b191d99SSage Weil struct ceph_pg pgid; 14163d14c5d2SYehuda Sadeh int acting[CEPH_PG_MAX_SIZE]; 14178008ab10SIlya Dryomov int num, o; 14183d14c5d2SYehuda Sadeh int err; 1419d29adb34SJosh Durgin bool was_paused; 14203d14c5d2SYehuda Sadeh 14216f6c7006SSage Weil dout("map_request %p tid %lld\n", req, req->r_tid); 142217a13e40SIlya Dryomov 142317a13e40SIlya Dryomov err = __calc_request_pg(osdc->osdmap, req, &pgid); 14246f6c7006SSage Weil if (err) { 14256f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 14263d14c5d2SYehuda Sadeh return err; 14276f6c7006SSage Weil } 14283d14c5d2SYehuda Sadeh req->r_pgid = pgid; 14293d14c5d2SYehuda Sadeh 14308008ab10SIlya Dryomov num = ceph_calc_pg_acting(osdc->osdmap, pgid, acting, &o); 14318008ab10SIlya Dryomov if (num < 0) 14328008ab10SIlya Dryomov num = 0; 14333d14c5d2SYehuda Sadeh 1434d29adb34SJosh Durgin was_paused = req->r_paused; 1435d29adb34SJosh Durgin req->r_paused = __req_should_be_paused(osdc, req); 1436d29adb34SJosh Durgin if (was_paused && !req->r_paused) 1437d29adb34SJosh Durgin force_resend = 1; 1438d29adb34SJosh Durgin 143938d6453cSSage Weil if ((!force_resend && 144038d6453cSSage Weil req->r_osd && req->r_osd->o_osd == o && 14413d14c5d2SYehuda Sadeh req->r_sent >= req->r_osd->o_incarnation && 14423d14c5d2SYehuda Sadeh req->r_num_pg_osds == num && 14433d14c5d2SYehuda Sadeh memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) || 1444d29adb34SJosh Durgin (req->r_osd == NULL && o == -1) || 1445d29adb34SJosh Durgin req->r_paused) 14463d14c5d2SYehuda Sadeh return 0; /* no change */ 14473d14c5d2SYehuda Sadeh 14485b191d99SSage Weil dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n", 14495b191d99SSage Weil req->r_tid, pgid.pool, pgid.seed, o, 14503d14c5d2SYehuda Sadeh req->r_osd ? req->r_osd->o_osd : -1); 14513d14c5d2SYehuda Sadeh 14523d14c5d2SYehuda Sadeh /* record full pg acting set */ 14533d14c5d2SYehuda Sadeh memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num); 14543d14c5d2SYehuda Sadeh req->r_num_pg_osds = num; 14553d14c5d2SYehuda Sadeh 14563d14c5d2SYehuda Sadeh if (req->r_osd) { 14573d14c5d2SYehuda Sadeh __cancel_request(req); 14583d14c5d2SYehuda Sadeh list_del_init(&req->r_osd_item); 1459a390de02SIlya Dryomov list_del_init(&req->r_linger_osd_item); 14603d14c5d2SYehuda Sadeh req->r_osd = NULL; 14613d14c5d2SYehuda Sadeh } 14623d14c5d2SYehuda Sadeh 14633d14c5d2SYehuda Sadeh req->r_osd = __lookup_osd(osdc, o); 14643d14c5d2SYehuda Sadeh if (!req->r_osd && o >= 0) { 14653d14c5d2SYehuda Sadeh err = -ENOMEM; 1466e10006f8SAlex Elder req->r_osd = create_osd(osdc, o); 14676f6c7006SSage Weil if (!req->r_osd) { 14686f6c7006SSage Weil list_move(&req->r_req_lru_item, &osdc->req_notarget); 14693d14c5d2SYehuda Sadeh goto out; 14706f6c7006SSage Weil } 14713d14c5d2SYehuda Sadeh 14726f6c7006SSage Weil dout("map_request osd %p is osd%d\n", req->r_osd, o); 14733d14c5d2SYehuda Sadeh __insert_osd(osdc, req->r_osd); 14743d14c5d2SYehuda Sadeh 1475b7a9e5ddSSage Weil ceph_con_open(&req->r_osd->o_con, 1476b7a9e5ddSSage Weil CEPH_ENTITY_TYPE_OSD, o, 1477b7a9e5ddSSage Weil &osdc->osdmap->osd_addr[o]); 14783d14c5d2SYehuda Sadeh } 14793d14c5d2SYehuda Sadeh 1480f671b581SIlya Dryomov __enqueue_request(req); 14813d14c5d2SYehuda Sadeh err = 1; /* osd or pg changed */ 14823d14c5d2SYehuda Sadeh 14833d14c5d2SYehuda Sadeh out: 14843d14c5d2SYehuda Sadeh return err; 14853d14c5d2SYehuda Sadeh } 14863d14c5d2SYehuda Sadeh 14873d14c5d2SYehuda Sadeh /* 14883d14c5d2SYehuda Sadeh * caller should hold map_sem (for read) and request_mutex 14893d14c5d2SYehuda Sadeh */ 149056e925b6SSage Weil static void __send_request(struct ceph_osd_client *osdc, 14913d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 14923d14c5d2SYehuda Sadeh { 14931b83bef2SSage Weil void *p; 14943d14c5d2SYehuda Sadeh 14951b83bef2SSage Weil dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n", 14961b83bef2SSage Weil req, req->r_tid, req->r_osd->o_osd, req->r_flags, 14971b83bef2SSage Weil (unsigned long long)req->r_pgid.pool, req->r_pgid.seed); 14983d14c5d2SYehuda Sadeh 14991b83bef2SSage Weil /* fill in message content that changes each time we send it */ 15001b83bef2SSage Weil put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch); 15011b83bef2SSage Weil put_unaligned_le32(req->r_flags, req->r_request_flags); 1502205ee118SIlya Dryomov put_unaligned_le64(req->r_target_oloc.pool, req->r_request_pool); 15031b83bef2SSage Weil p = req->r_request_pgid; 15041b83bef2SSage Weil ceph_encode_64(&p, req->r_pgid.pool); 15051b83bef2SSage Weil ceph_encode_32(&p, req->r_pgid.seed); 15061b83bef2SSage Weil put_unaligned_le64(1, req->r_request_attempts); /* FIXME */ 15071b83bef2SSage Weil memcpy(req->r_request_reassert_version, &req->r_reassert_version, 15081b83bef2SSage Weil sizeof(req->r_reassert_version)); 15092169aea6SSage Weil 15103d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 15113d14c5d2SYehuda Sadeh list_move_tail(&req->r_req_lru_item, &osdc->req_lru); 15123d14c5d2SYehuda Sadeh 15133d14c5d2SYehuda Sadeh ceph_msg_get(req->r_request); /* send consumes a ref */ 151426be8808SAlex Elder 15153d14c5d2SYehuda Sadeh req->r_sent = req->r_osd->o_incarnation; 151626be8808SAlex Elder 151726be8808SAlex Elder ceph_con_send(&req->r_osd->o_con, req->r_request); 15183d14c5d2SYehuda Sadeh } 15193d14c5d2SYehuda Sadeh 15203d14c5d2SYehuda Sadeh /* 15216f6c7006SSage Weil * Send any requests in the queue (req_unsent). 15226f6c7006SSage Weil */ 1523f9d25199SAlex Elder static void __send_queued(struct ceph_osd_client *osdc) 15246f6c7006SSage Weil { 15256f6c7006SSage Weil struct ceph_osd_request *req, *tmp; 15266f6c7006SSage Weil 1527f9d25199SAlex Elder dout("__send_queued\n"); 1528f9d25199SAlex Elder list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) 15296f6c7006SSage Weil __send_request(osdc, req); 15306f6c7006SSage Weil } 15316f6c7006SSage Weil 15326f6c7006SSage Weil /* 15330bbfdfe8SIlya Dryomov * Caller should hold map_sem for read and request_mutex. 15340bbfdfe8SIlya Dryomov */ 15350bbfdfe8SIlya Dryomov static int __ceph_osdc_start_request(struct ceph_osd_client *osdc, 15360bbfdfe8SIlya Dryomov struct ceph_osd_request *req, 15370bbfdfe8SIlya Dryomov bool nofail) 15380bbfdfe8SIlya Dryomov { 15390bbfdfe8SIlya Dryomov int rc; 15400bbfdfe8SIlya Dryomov 15410bbfdfe8SIlya Dryomov __register_request(osdc, req); 15420bbfdfe8SIlya Dryomov req->r_sent = 0; 15430bbfdfe8SIlya Dryomov req->r_got_reply = 0; 15440bbfdfe8SIlya Dryomov rc = __map_request(osdc, req, 0); 15450bbfdfe8SIlya Dryomov if (rc < 0) { 15460bbfdfe8SIlya Dryomov if (nofail) { 15470bbfdfe8SIlya Dryomov dout("osdc_start_request failed map, " 15480bbfdfe8SIlya Dryomov " will retry %lld\n", req->r_tid); 15490bbfdfe8SIlya Dryomov rc = 0; 15500bbfdfe8SIlya Dryomov } else { 15510bbfdfe8SIlya Dryomov __unregister_request(osdc, req); 15520bbfdfe8SIlya Dryomov } 15530bbfdfe8SIlya Dryomov return rc; 15540bbfdfe8SIlya Dryomov } 15550bbfdfe8SIlya Dryomov 15560bbfdfe8SIlya Dryomov if (req->r_osd == NULL) { 15570bbfdfe8SIlya Dryomov dout("send_request %p no up osds in pg\n", req); 15580bbfdfe8SIlya Dryomov ceph_monc_request_next_osdmap(&osdc->client->monc); 15590bbfdfe8SIlya Dryomov } else { 15600bbfdfe8SIlya Dryomov __send_queued(osdc); 15610bbfdfe8SIlya Dryomov } 15620bbfdfe8SIlya Dryomov 15630bbfdfe8SIlya Dryomov return 0; 15640bbfdfe8SIlya Dryomov } 15650bbfdfe8SIlya Dryomov 15660bbfdfe8SIlya Dryomov /* 15673d14c5d2SYehuda Sadeh * Timeout callback, called every N seconds when 1 or more osd 15683d14c5d2SYehuda Sadeh * requests has been active for more than N seconds. When this 15693d14c5d2SYehuda Sadeh * happens, we ping all OSDs with requests who have timed out to 15703d14c5d2SYehuda Sadeh * ensure any communications channel reset is detected. Reset the 15713d14c5d2SYehuda Sadeh * request timeouts another N seconds in the future as we go. 15723d14c5d2SYehuda Sadeh * Reschedule the timeout event another N seconds in future (unless 15733d14c5d2SYehuda Sadeh * there are no open requests). 15743d14c5d2SYehuda Sadeh */ 15753d14c5d2SYehuda Sadeh static void handle_timeout(struct work_struct *work) 15763d14c5d2SYehuda Sadeh { 15773d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 15783d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, timeout_work.work); 1579a319bf56SIlya Dryomov struct ceph_options *opts = osdc->client->options; 158083aff95eSSage Weil struct ceph_osd_request *req; 15813d14c5d2SYehuda Sadeh struct ceph_osd *osd; 15823d14c5d2SYehuda Sadeh struct list_head slow_osds; 15833d14c5d2SYehuda Sadeh dout("timeout\n"); 15843d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 15853d14c5d2SYehuda Sadeh 15863d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 15873d14c5d2SYehuda Sadeh 15883d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 15893d14c5d2SYehuda Sadeh 15903d14c5d2SYehuda Sadeh /* 15913d14c5d2SYehuda Sadeh * ping osds that are a bit slow. this ensures that if there 15923d14c5d2SYehuda Sadeh * is a break in the TCP connection we will notice, and reopen 15933d14c5d2SYehuda Sadeh * a connection with that osd (from the fault callback). 15943d14c5d2SYehuda Sadeh */ 15953d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&slow_osds); 15963d14c5d2SYehuda Sadeh list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) { 1597a319bf56SIlya Dryomov if (time_before(jiffies, 1598a319bf56SIlya Dryomov req->r_stamp + opts->osd_keepalive_timeout)) 15993d14c5d2SYehuda Sadeh break; 16003d14c5d2SYehuda Sadeh 16013d14c5d2SYehuda Sadeh osd = req->r_osd; 16023d14c5d2SYehuda Sadeh BUG_ON(!osd); 16033d14c5d2SYehuda Sadeh dout(" tid %llu is slow, will send keepalive on osd%d\n", 16043d14c5d2SYehuda Sadeh req->r_tid, osd->o_osd); 16053d14c5d2SYehuda Sadeh list_move_tail(&osd->o_keepalive_item, &slow_osds); 16063d14c5d2SYehuda Sadeh } 16073d14c5d2SYehuda Sadeh while (!list_empty(&slow_osds)) { 16083d14c5d2SYehuda Sadeh osd = list_entry(slow_osds.next, struct ceph_osd, 16093d14c5d2SYehuda Sadeh o_keepalive_item); 16103d14c5d2SYehuda Sadeh list_del_init(&osd->o_keepalive_item); 16113d14c5d2SYehuda Sadeh ceph_con_keepalive(&osd->o_con); 16123d14c5d2SYehuda Sadeh } 16133d14c5d2SYehuda Sadeh 16143d14c5d2SYehuda Sadeh __schedule_osd_timeout(osdc); 1615f9d25199SAlex Elder __send_queued(osdc); 16163d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 16173d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 16183d14c5d2SYehuda Sadeh } 16193d14c5d2SYehuda Sadeh 16203d14c5d2SYehuda Sadeh static void handle_osds_timeout(struct work_struct *work) 16213d14c5d2SYehuda Sadeh { 16223d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 16233d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, 16243d14c5d2SYehuda Sadeh osds_timeout_work.work); 1625a319bf56SIlya Dryomov unsigned long delay = osdc->client->options->osd_idle_ttl / 4; 16263d14c5d2SYehuda Sadeh 16273d14c5d2SYehuda Sadeh dout("osds timeout\n"); 16283d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 1629aca420bcSSage Weil remove_old_osds(osdc); 16303d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 16313d14c5d2SYehuda Sadeh 16323d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 16333d14c5d2SYehuda Sadeh round_jiffies_relative(delay)); 16343d14c5d2SYehuda Sadeh } 16353d14c5d2SYehuda Sadeh 1636205ee118SIlya Dryomov static int ceph_oloc_decode(void **p, void *end, 1637205ee118SIlya Dryomov struct ceph_object_locator *oloc) 1638205ee118SIlya Dryomov { 1639205ee118SIlya Dryomov u8 struct_v, struct_cv; 1640205ee118SIlya Dryomov u32 len; 1641205ee118SIlya Dryomov void *struct_end; 1642205ee118SIlya Dryomov int ret = 0; 1643205ee118SIlya Dryomov 1644205ee118SIlya Dryomov ceph_decode_need(p, end, 1 + 1 + 4, e_inval); 1645205ee118SIlya Dryomov struct_v = ceph_decode_8(p); 1646205ee118SIlya Dryomov struct_cv = ceph_decode_8(p); 1647205ee118SIlya Dryomov if (struct_v < 3) { 1648205ee118SIlya Dryomov pr_warn("got v %d < 3 cv %d of ceph_object_locator\n", 1649205ee118SIlya Dryomov struct_v, struct_cv); 1650205ee118SIlya Dryomov goto e_inval; 1651205ee118SIlya Dryomov } 1652205ee118SIlya Dryomov if (struct_cv > 6) { 1653205ee118SIlya Dryomov pr_warn("got v %d cv %d > 6 of ceph_object_locator\n", 1654205ee118SIlya Dryomov struct_v, struct_cv); 1655205ee118SIlya Dryomov goto e_inval; 1656205ee118SIlya Dryomov } 1657205ee118SIlya Dryomov len = ceph_decode_32(p); 1658205ee118SIlya Dryomov ceph_decode_need(p, end, len, e_inval); 1659205ee118SIlya Dryomov struct_end = *p + len; 1660205ee118SIlya Dryomov 1661205ee118SIlya Dryomov oloc->pool = ceph_decode_64(p); 1662205ee118SIlya Dryomov *p += 4; /* skip preferred */ 1663205ee118SIlya Dryomov 1664205ee118SIlya Dryomov len = ceph_decode_32(p); 1665205ee118SIlya Dryomov if (len > 0) { 1666205ee118SIlya Dryomov pr_warn("ceph_object_locator::key is set\n"); 1667205ee118SIlya Dryomov goto e_inval; 1668205ee118SIlya Dryomov } 1669205ee118SIlya Dryomov 1670205ee118SIlya Dryomov if (struct_v >= 5) { 1671205ee118SIlya Dryomov len = ceph_decode_32(p); 1672205ee118SIlya Dryomov if (len > 0) { 1673205ee118SIlya Dryomov pr_warn("ceph_object_locator::nspace is set\n"); 1674205ee118SIlya Dryomov goto e_inval; 1675205ee118SIlya Dryomov } 1676205ee118SIlya Dryomov } 1677205ee118SIlya Dryomov 1678205ee118SIlya Dryomov if (struct_v >= 6) { 1679205ee118SIlya Dryomov s64 hash = ceph_decode_64(p); 1680205ee118SIlya Dryomov if (hash != -1) { 1681205ee118SIlya Dryomov pr_warn("ceph_object_locator::hash is set\n"); 1682205ee118SIlya Dryomov goto e_inval; 1683205ee118SIlya Dryomov } 1684205ee118SIlya Dryomov } 1685205ee118SIlya Dryomov 1686205ee118SIlya Dryomov /* skip the rest */ 1687205ee118SIlya Dryomov *p = struct_end; 1688205ee118SIlya Dryomov out: 1689205ee118SIlya Dryomov return ret; 1690205ee118SIlya Dryomov 1691205ee118SIlya Dryomov e_inval: 1692205ee118SIlya Dryomov ret = -EINVAL; 1693205ee118SIlya Dryomov goto out; 1694205ee118SIlya Dryomov } 1695205ee118SIlya Dryomov 1696205ee118SIlya Dryomov static int ceph_redirect_decode(void **p, void *end, 1697205ee118SIlya Dryomov struct ceph_request_redirect *redir) 1698205ee118SIlya Dryomov { 1699205ee118SIlya Dryomov u8 struct_v, struct_cv; 1700205ee118SIlya Dryomov u32 len; 1701205ee118SIlya Dryomov void *struct_end; 1702205ee118SIlya Dryomov int ret; 1703205ee118SIlya Dryomov 1704205ee118SIlya Dryomov ceph_decode_need(p, end, 1 + 1 + 4, e_inval); 1705205ee118SIlya Dryomov struct_v = ceph_decode_8(p); 1706205ee118SIlya Dryomov struct_cv = ceph_decode_8(p); 1707205ee118SIlya Dryomov if (struct_cv > 1) { 1708205ee118SIlya Dryomov pr_warn("got v %d cv %d > 1 of ceph_request_redirect\n", 1709205ee118SIlya Dryomov struct_v, struct_cv); 1710205ee118SIlya Dryomov goto e_inval; 1711205ee118SIlya Dryomov } 1712205ee118SIlya Dryomov len = ceph_decode_32(p); 1713205ee118SIlya Dryomov ceph_decode_need(p, end, len, e_inval); 1714205ee118SIlya Dryomov struct_end = *p + len; 1715205ee118SIlya Dryomov 1716205ee118SIlya Dryomov ret = ceph_oloc_decode(p, end, &redir->oloc); 1717205ee118SIlya Dryomov if (ret) 1718205ee118SIlya Dryomov goto out; 1719205ee118SIlya Dryomov 1720205ee118SIlya Dryomov len = ceph_decode_32(p); 1721205ee118SIlya Dryomov if (len > 0) { 1722205ee118SIlya Dryomov pr_warn("ceph_request_redirect::object_name is set\n"); 1723205ee118SIlya Dryomov goto e_inval; 1724205ee118SIlya Dryomov } 1725205ee118SIlya Dryomov 1726205ee118SIlya Dryomov len = ceph_decode_32(p); 1727205ee118SIlya Dryomov *p += len; /* skip osd_instructions */ 1728205ee118SIlya Dryomov 1729205ee118SIlya Dryomov /* skip the rest */ 1730205ee118SIlya Dryomov *p = struct_end; 1731205ee118SIlya Dryomov out: 1732205ee118SIlya Dryomov return ret; 1733205ee118SIlya Dryomov 1734205ee118SIlya Dryomov e_inval: 1735205ee118SIlya Dryomov ret = -EINVAL; 1736205ee118SIlya Dryomov goto out; 1737205ee118SIlya Dryomov } 1738205ee118SIlya Dryomov 173925845472SSage Weil static void complete_request(struct ceph_osd_request *req) 174025845472SSage Weil { 174125845472SSage Weil complete_all(&req->r_safe_completion); /* fsync waiter */ 174225845472SSage Weil } 174325845472SSage Weil 17443d14c5d2SYehuda Sadeh /* 17453d14c5d2SYehuda Sadeh * handle osd op reply. either call the callback if it is specified, 17463d14c5d2SYehuda Sadeh * or do the completion to wake up the waiting thread. 17473d14c5d2SYehuda Sadeh */ 17483d14c5d2SYehuda Sadeh static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg, 17493d14c5d2SYehuda Sadeh struct ceph_connection *con) 17503d14c5d2SYehuda Sadeh { 17511b83bef2SSage Weil void *p, *end; 17523d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 1753205ee118SIlya Dryomov struct ceph_request_redirect redir; 17543d14c5d2SYehuda Sadeh u64 tid; 17551b83bef2SSage Weil int object_len; 175679528734SAlex Elder unsigned int numops; 175779528734SAlex Elder int payload_len, flags; 17583d14c5d2SYehuda Sadeh s32 result; 17591b83bef2SSage Weil s32 retry_attempt; 17601b83bef2SSage Weil struct ceph_pg pg; 17611b83bef2SSage Weil int err; 17621b83bef2SSage Weil u32 reassert_epoch; 17631b83bef2SSage Weil u64 reassert_version; 17641b83bef2SSage Weil u32 osdmap_epoch; 17650d5af164SAlex Elder int already_completed; 17669fc6e064SAlex Elder u32 bytes; 176779528734SAlex Elder unsigned int i; 17683d14c5d2SYehuda Sadeh 17693d14c5d2SYehuda Sadeh tid = le64_to_cpu(msg->hdr.tid); 17701b83bef2SSage Weil dout("handle_reply %p tid %llu\n", msg, tid); 17711b83bef2SSage Weil 17721b83bef2SSage Weil p = msg->front.iov_base; 17731b83bef2SSage Weil end = p + msg->front.iov_len; 17741b83bef2SSage Weil 17751b83bef2SSage Weil ceph_decode_need(&p, end, 4, bad); 17761b83bef2SSage Weil object_len = ceph_decode_32(&p); 17771b83bef2SSage Weil ceph_decode_need(&p, end, object_len, bad); 17781b83bef2SSage Weil p += object_len; 17791b83bef2SSage Weil 1780ef4859d6SAlex Elder err = ceph_decode_pgid(&p, end, &pg); 17811b83bef2SSage Weil if (err) 17823d14c5d2SYehuda Sadeh goto bad; 17831b83bef2SSage Weil 17841b83bef2SSage Weil ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad); 17851b83bef2SSage Weil flags = ceph_decode_64(&p); 17861b83bef2SSage Weil result = ceph_decode_32(&p); 17871b83bef2SSage Weil reassert_epoch = ceph_decode_32(&p); 17881b83bef2SSage Weil reassert_version = ceph_decode_64(&p); 17891b83bef2SSage Weil osdmap_epoch = ceph_decode_32(&p); 17901b83bef2SSage Weil 17913d14c5d2SYehuda Sadeh /* lookup */ 1792ff513aceSIlya Dryomov down_read(&osdc->map_sem); 17933d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 17943d14c5d2SYehuda Sadeh req = __lookup_request(osdc, tid); 17953d14c5d2SYehuda Sadeh if (req == NULL) { 17963d14c5d2SYehuda Sadeh dout("handle_reply tid %llu dne\n", tid); 17978058fd45SAlex Elder goto bad_mutex; 17983d14c5d2SYehuda Sadeh } 17993d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 18001b83bef2SSage Weil 18011b83bef2SSage Weil dout("handle_reply %p tid %llu req %p result %d\n", msg, tid, 18021b83bef2SSage Weil req, result); 18031b83bef2SSage Weil 180418741196SDan Carpenter ceph_decode_need(&p, end, 4, bad_put); 18051b83bef2SSage Weil numops = ceph_decode_32(&p); 18061b83bef2SSage Weil if (numops > CEPH_OSD_MAX_OP) 18071b83bef2SSage Weil goto bad_put; 18081b83bef2SSage Weil if (numops != req->r_num_ops) 18091b83bef2SSage Weil goto bad_put; 18101b83bef2SSage Weil payload_len = 0; 181118741196SDan Carpenter ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad_put); 18121b83bef2SSage Weil for (i = 0; i < numops; i++) { 18131b83bef2SSage Weil struct ceph_osd_op *op = p; 18141b83bef2SSage Weil int len; 18151b83bef2SSage Weil 18161b83bef2SSage Weil len = le32_to_cpu(op->payload_len); 18171b83bef2SSage Weil req->r_reply_op_len[i] = len; 18181b83bef2SSage Weil dout(" op %d has %d bytes\n", i, len); 18191b83bef2SSage Weil payload_len += len; 18201b83bef2SSage Weil p += sizeof(*op); 18211b83bef2SSage Weil } 18229fc6e064SAlex Elder bytes = le32_to_cpu(msg->hdr.data_len); 18239fc6e064SAlex Elder if (payload_len != bytes) { 1824b9a67899SJoe Perches pr_warn("sum of op payload lens %d != data_len %d\n", 18259fc6e064SAlex Elder payload_len, bytes); 18261b83bef2SSage Weil goto bad_put; 18271b83bef2SSage Weil } 18281b83bef2SSage Weil 182918741196SDan Carpenter ceph_decode_need(&p, end, 4 + numops * 4, bad_put); 18301b83bef2SSage Weil retry_attempt = ceph_decode_32(&p); 18311b83bef2SSage Weil for (i = 0; i < numops; i++) 18321b83bef2SSage Weil req->r_reply_op_result[i] = ceph_decode_32(&p); 18333d14c5d2SYehuda Sadeh 1834205ee118SIlya Dryomov if (le16_to_cpu(msg->hdr.version) >= 6) { 1835205ee118SIlya Dryomov p += 8 + 4; /* skip replay_version */ 1836205ee118SIlya Dryomov p += 8; /* skip user_version */ 1837205ee118SIlya Dryomov 1838205ee118SIlya Dryomov err = ceph_redirect_decode(&p, end, &redir); 1839205ee118SIlya Dryomov if (err) 1840205ee118SIlya Dryomov goto bad_put; 1841205ee118SIlya Dryomov } else { 1842205ee118SIlya Dryomov redir.oloc.pool = -1; 1843205ee118SIlya Dryomov } 1844205ee118SIlya Dryomov 1845205ee118SIlya Dryomov if (redir.oloc.pool != -1) { 1846205ee118SIlya Dryomov dout("redirect pool %lld\n", redir.oloc.pool); 1847205ee118SIlya Dryomov 1848205ee118SIlya Dryomov __unregister_request(osdc, req); 1849205ee118SIlya Dryomov 1850205ee118SIlya Dryomov req->r_target_oloc = redir.oloc; /* struct */ 1851205ee118SIlya Dryomov 1852205ee118SIlya Dryomov /* 1853205ee118SIlya Dryomov * Start redirect requests with nofail=true. If 1854205ee118SIlya Dryomov * mapping fails, request will end up on the notarget 1855205ee118SIlya Dryomov * list, waiting for the new osdmap (which can take 1856205ee118SIlya Dryomov * a while), even though the original request mapped 1857205ee118SIlya Dryomov * successfully. In the future we might want to follow 1858205ee118SIlya Dryomov * original request's nofail setting here. 1859205ee118SIlya Dryomov */ 1860ff513aceSIlya Dryomov err = __ceph_osdc_start_request(osdc, req, true); 1861205ee118SIlya Dryomov BUG_ON(err); 1862205ee118SIlya Dryomov 1863ff513aceSIlya Dryomov goto out_unlock; 1864205ee118SIlya Dryomov } 1865205ee118SIlya Dryomov 1866eb845ff1SYan, Zheng already_completed = req->r_got_reply; 18673d14c5d2SYehuda Sadeh if (!req->r_got_reply) { 18681b83bef2SSage Weil req->r_result = result; 18693d14c5d2SYehuda Sadeh dout("handle_reply result %d bytes %d\n", req->r_result, 18703d14c5d2SYehuda Sadeh bytes); 18713d14c5d2SYehuda Sadeh if (req->r_result == 0) 18723d14c5d2SYehuda Sadeh req->r_result = bytes; 18733d14c5d2SYehuda Sadeh 18743d14c5d2SYehuda Sadeh /* in case this is a write and we need to replay, */ 18751b83bef2SSage Weil req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch); 18761b83bef2SSage Weil req->r_reassert_version.version = cpu_to_le64(reassert_version); 18773d14c5d2SYehuda Sadeh 18783d14c5d2SYehuda Sadeh req->r_got_reply = 1; 18793d14c5d2SYehuda Sadeh } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) { 18803d14c5d2SYehuda Sadeh dout("handle_reply tid %llu dup ack\n", tid); 1881ff513aceSIlya Dryomov goto out_unlock; 18823d14c5d2SYehuda Sadeh } 18833d14c5d2SYehuda Sadeh 18843d14c5d2SYehuda Sadeh dout("handle_reply tid %llu flags %d\n", tid, flags); 18853d14c5d2SYehuda Sadeh 1886a40c4f10SYehuda Sadeh if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK)) 1887a40c4f10SYehuda Sadeh __register_linger_request(osdc, req); 1888a40c4f10SYehuda Sadeh 18893d14c5d2SYehuda Sadeh /* either this is a read, or we got the safe response */ 18903d14c5d2SYehuda Sadeh if (result < 0 || 18913d14c5d2SYehuda Sadeh (flags & CEPH_OSD_FLAG_ONDISK) || 18923d14c5d2SYehuda Sadeh ((flags & CEPH_OSD_FLAG_WRITE) == 0)) 18933d14c5d2SYehuda Sadeh __unregister_request(osdc, req); 18943d14c5d2SYehuda Sadeh 18953d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 1896ff513aceSIlya Dryomov up_read(&osdc->map_sem); 18973d14c5d2SYehuda Sadeh 1898eb845ff1SYan, Zheng if (!already_completed) { 189961c5d6bfSYan, Zheng if (req->r_unsafe_callback && 190061c5d6bfSYan, Zheng result >= 0 && !(flags & CEPH_OSD_FLAG_ONDISK)) 190161c5d6bfSYan, Zheng req->r_unsafe_callback(req, true); 19023d14c5d2SYehuda Sadeh if (req->r_callback) 19033d14c5d2SYehuda Sadeh req->r_callback(req, msg); 19043d14c5d2SYehuda Sadeh else 19053d14c5d2SYehuda Sadeh complete_all(&req->r_completion); 1906eb845ff1SYan, Zheng } 19073d14c5d2SYehuda Sadeh 190861c5d6bfSYan, Zheng if (flags & CEPH_OSD_FLAG_ONDISK) { 190961c5d6bfSYan, Zheng if (req->r_unsafe_callback && already_completed) 191061c5d6bfSYan, Zheng req->r_unsafe_callback(req, false); 191125845472SSage Weil complete_request(req); 191261c5d6bfSYan, Zheng } 19133d14c5d2SYehuda Sadeh 1914ff513aceSIlya Dryomov out: 1915a40c4f10SYehuda Sadeh dout("req=%p req->r_linger=%d\n", req, req->r_linger); 19163d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 19173d14c5d2SYehuda Sadeh return; 1918ff513aceSIlya Dryomov out_unlock: 1919ff513aceSIlya Dryomov mutex_unlock(&osdc->request_mutex); 1920ff513aceSIlya Dryomov up_read(&osdc->map_sem); 1921ff513aceSIlya Dryomov goto out; 19223d14c5d2SYehuda Sadeh 19231b83bef2SSage Weil bad_put: 192437c89bdeSLi Wang req->r_result = -EIO; 192537c89bdeSLi Wang __unregister_request(osdc, req); 192637c89bdeSLi Wang if (req->r_callback) 192737c89bdeSLi Wang req->r_callback(req, msg); 192837c89bdeSLi Wang else 192937c89bdeSLi Wang complete_all(&req->r_completion); 193037c89bdeSLi Wang complete_request(req); 19311b83bef2SSage Weil ceph_osdc_put_request(req); 19328058fd45SAlex Elder bad_mutex: 19338058fd45SAlex Elder mutex_unlock(&osdc->request_mutex); 1934ff513aceSIlya Dryomov up_read(&osdc->map_sem); 19353d14c5d2SYehuda Sadeh bad: 19361b83bef2SSage Weil pr_err("corrupt osd_op_reply got %d %d\n", 19371b83bef2SSage Weil (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len)); 19383d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 19393d14c5d2SYehuda Sadeh } 19403d14c5d2SYehuda Sadeh 19416f6c7006SSage Weil static void reset_changed_osds(struct ceph_osd_client *osdc) 19423d14c5d2SYehuda Sadeh { 19433d14c5d2SYehuda Sadeh struct rb_node *p, *n; 19443d14c5d2SYehuda Sadeh 19457eb71e03SIlya Dryomov dout("%s %p\n", __func__, osdc); 19463d14c5d2SYehuda Sadeh for (p = rb_first(&osdc->osds); p; p = n) { 19476f6c7006SSage Weil struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node); 19483d14c5d2SYehuda Sadeh 19493d14c5d2SYehuda Sadeh n = rb_next(p); 19503d14c5d2SYehuda Sadeh if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) || 19513d14c5d2SYehuda Sadeh memcmp(&osd->o_con.peer_addr, 19523d14c5d2SYehuda Sadeh ceph_osd_addr(osdc->osdmap, 19533d14c5d2SYehuda Sadeh osd->o_osd), 19543d14c5d2SYehuda Sadeh sizeof(struct ceph_entity_addr)) != 0) 19553d14c5d2SYehuda Sadeh __reset_osd(osdc, osd); 19563d14c5d2SYehuda Sadeh } 19573d14c5d2SYehuda Sadeh } 19583d14c5d2SYehuda Sadeh 19593d14c5d2SYehuda Sadeh /* 19606f6c7006SSage Weil * Requeue requests whose mapping to an OSD has changed. If requests map to 19616f6c7006SSage Weil * no osd, request a new map. 19623d14c5d2SYehuda Sadeh * 1963e6d50f67SAlex Elder * Caller should hold map_sem for read. 19643d14c5d2SYehuda Sadeh */ 19659a1ea2dbSJosh Durgin static void kick_requests(struct ceph_osd_client *osdc, bool force_resend, 19669a1ea2dbSJosh Durgin bool force_resend_writes) 19673d14c5d2SYehuda Sadeh { 1968a40c4f10SYehuda Sadeh struct ceph_osd_request *req, *nreq; 19696f6c7006SSage Weil struct rb_node *p; 19706f6c7006SSage Weil int needmap = 0; 19716f6c7006SSage Weil int err; 19729a1ea2dbSJosh Durgin bool force_resend_req; 19733d14c5d2SYehuda Sadeh 19749a1ea2dbSJosh Durgin dout("kick_requests %s %s\n", force_resend ? " (force resend)" : "", 19759a1ea2dbSJosh Durgin force_resend_writes ? " (force resend writes)" : ""); 19763d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 19776194ea89SSage Weil for (p = rb_first(&osdc->requests); p; ) { 19786f6c7006SSage Weil req = rb_entry(p, struct ceph_osd_request, r_node); 19796194ea89SSage Weil p = rb_next(p); 1980ab60b16dSAlex Elder 1981ab60b16dSAlex Elder /* 1982ab60b16dSAlex Elder * For linger requests that have not yet been 1983ab60b16dSAlex Elder * registered, move them to the linger list; they'll 1984ab60b16dSAlex Elder * be sent to the osd in the loop below. Unregister 1985ab60b16dSAlex Elder * the request before re-registering it as a linger 1986ab60b16dSAlex Elder * request to ensure the __map_request() below 1987ab60b16dSAlex Elder * will decide it needs to be sent. 1988ab60b16dSAlex Elder */ 1989ab60b16dSAlex Elder if (req->r_linger && list_empty(&req->r_linger_item)) { 1990ab60b16dSAlex Elder dout("%p tid %llu restart on osd%d\n", 1991ab60b16dSAlex Elder req, req->r_tid, 1992ab60b16dSAlex Elder req->r_osd ? req->r_osd->o_osd : -1); 199396e4dac6SAlex Elder ceph_osdc_get_request(req); 1994ab60b16dSAlex Elder __unregister_request(osdc, req); 1995ab60b16dSAlex Elder __register_linger_request(osdc, req); 199696e4dac6SAlex Elder ceph_osdc_put_request(req); 1997ab60b16dSAlex Elder continue; 1998ab60b16dSAlex Elder } 1999ab60b16dSAlex Elder 20009a1ea2dbSJosh Durgin force_resend_req = force_resend || 20019a1ea2dbSJosh Durgin (force_resend_writes && 20029a1ea2dbSJosh Durgin req->r_flags & CEPH_OSD_FLAG_WRITE); 20039a1ea2dbSJosh Durgin err = __map_request(osdc, req, force_resend_req); 20046f6c7006SSage Weil if (err < 0) 20056f6c7006SSage Weil continue; /* error */ 20066f6c7006SSage Weil if (req->r_osd == NULL) { 20076f6c7006SSage Weil dout("%p tid %llu maps to no osd\n", req, req->r_tid); 20086f6c7006SSage Weil needmap++; /* request a newer map */ 20096f6c7006SSage Weil } else if (err > 0) { 20106194ea89SSage Weil if (!req->r_linger) { 20116194ea89SSage Weil dout("%p tid %llu requeued on osd%d\n", req, 20126194ea89SSage Weil req->r_tid, 20136f6c7006SSage Weil req->r_osd ? req->r_osd->o_osd : -1); 20146f6c7006SSage Weil req->r_flags |= CEPH_OSD_FLAG_RETRY; 20156f6c7006SSage Weil } 20166f6c7006SSage Weil } 20176194ea89SSage Weil } 2018a40c4f10SYehuda Sadeh 2019a40c4f10SYehuda Sadeh list_for_each_entry_safe(req, nreq, &osdc->req_linger, 2020a40c4f10SYehuda Sadeh r_linger_item) { 2021a40c4f10SYehuda Sadeh dout("linger req=%p req->r_osd=%p\n", req, req->r_osd); 2022a40c4f10SYehuda Sadeh 20239a1ea2dbSJosh Durgin err = __map_request(osdc, req, 20249a1ea2dbSJosh Durgin force_resend || force_resend_writes); 2025ab60b16dSAlex Elder dout("__map_request returned %d\n", err); 2026a40c4f10SYehuda Sadeh if (err < 0) 2027a40c4f10SYehuda Sadeh continue; /* hrm! */ 2028b0494532SIlya Dryomov if (req->r_osd == NULL || err > 0) { 2029a40c4f10SYehuda Sadeh if (req->r_osd == NULL) { 2030b0494532SIlya Dryomov dout("lingering %p tid %llu maps to no osd\n", 2031b0494532SIlya Dryomov req, req->r_tid); 2032b0494532SIlya Dryomov /* 2033b0494532SIlya Dryomov * A homeless lingering request makes 2034b0494532SIlya Dryomov * no sense, as it's job is to keep 2035b0494532SIlya Dryomov * a particular OSD connection open. 2036b0494532SIlya Dryomov * Request a newer map and kick the 2037b0494532SIlya Dryomov * request, knowing that it won't be 2038b0494532SIlya Dryomov * resent until we actually get a map 2039b0494532SIlya Dryomov * that can tell us where to send it. 2040b0494532SIlya Dryomov */ 2041b0494532SIlya Dryomov needmap++; 2042a40c4f10SYehuda Sadeh } 2043a40c4f10SYehuda Sadeh 2044b0494532SIlya Dryomov dout("kicking lingering %p tid %llu osd%d\n", req, 2045b0494532SIlya Dryomov req->r_tid, req->r_osd ? req->r_osd->o_osd : -1); 2046a40c4f10SYehuda Sadeh __register_request(osdc, req); 2047c89ce05eSAlex Elder __unregister_linger_request(osdc, req); 2048a40c4f10SYehuda Sadeh } 2049b0494532SIlya Dryomov } 205014d2f38dSAlex Elder reset_changed_osds(osdc); 20513d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 20523d14c5d2SYehuda Sadeh 20533d14c5d2SYehuda Sadeh if (needmap) { 20543d14c5d2SYehuda Sadeh dout("%d requests for down osds, need new map\n", needmap); 20553d14c5d2SYehuda Sadeh ceph_monc_request_next_osdmap(&osdc->client->monc); 20563d14c5d2SYehuda Sadeh } 20573d14c5d2SYehuda Sadeh } 20586f6c7006SSage Weil 20596f6c7006SSage Weil 20603d14c5d2SYehuda Sadeh /* 20613d14c5d2SYehuda Sadeh * Process updated osd map. 20623d14c5d2SYehuda Sadeh * 20633d14c5d2SYehuda Sadeh * The message contains any number of incremental and full maps, normally 20643d14c5d2SYehuda Sadeh * indicating some sort of topology change in the cluster. Kick requests 20653d14c5d2SYehuda Sadeh * off to different OSDs as needed. 20663d14c5d2SYehuda Sadeh */ 20673d14c5d2SYehuda Sadeh void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg) 20683d14c5d2SYehuda Sadeh { 20693d14c5d2SYehuda Sadeh void *p, *end, *next; 20703d14c5d2SYehuda Sadeh u32 nr_maps, maplen; 20713d14c5d2SYehuda Sadeh u32 epoch; 20723d14c5d2SYehuda Sadeh struct ceph_osdmap *newmap = NULL, *oldmap; 20733d14c5d2SYehuda Sadeh int err; 20743d14c5d2SYehuda Sadeh struct ceph_fsid fsid; 20759a1ea2dbSJosh Durgin bool was_full; 20763d14c5d2SYehuda Sadeh 20773d14c5d2SYehuda Sadeh dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0); 20783d14c5d2SYehuda Sadeh p = msg->front.iov_base; 20793d14c5d2SYehuda Sadeh end = p + msg->front.iov_len; 20803d14c5d2SYehuda Sadeh 20813d14c5d2SYehuda Sadeh /* verify fsid */ 20823d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, sizeof(fsid), bad); 20833d14c5d2SYehuda Sadeh ceph_decode_copy(&p, &fsid, sizeof(fsid)); 20843d14c5d2SYehuda Sadeh if (ceph_check_fsid(osdc->client, &fsid) < 0) 20853d14c5d2SYehuda Sadeh return; 20863d14c5d2SYehuda Sadeh 20873d14c5d2SYehuda Sadeh down_write(&osdc->map_sem); 20883d14c5d2SYehuda Sadeh 20899a1ea2dbSJosh Durgin was_full = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL); 20909a1ea2dbSJosh Durgin 20913d14c5d2SYehuda Sadeh /* incremental maps */ 20923d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 20933d14c5d2SYehuda Sadeh dout(" %d inc maps\n", nr_maps); 20943d14c5d2SYehuda Sadeh while (nr_maps > 0) { 20953d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 20963d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 20973d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 20983d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 20993d14c5d2SYehuda Sadeh next = p + maplen; 21003d14c5d2SYehuda Sadeh if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) { 21013d14c5d2SYehuda Sadeh dout("applying incremental map %u len %d\n", 21023d14c5d2SYehuda Sadeh epoch, maplen); 21033d14c5d2SYehuda Sadeh newmap = osdmap_apply_incremental(&p, next, 21043d14c5d2SYehuda Sadeh osdc->osdmap, 210515d9882cSAlex Elder &osdc->client->msgr); 21063d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 21073d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 21083d14c5d2SYehuda Sadeh goto bad; 21093d14c5d2SYehuda Sadeh } 21103d14c5d2SYehuda Sadeh BUG_ON(!newmap); 21113d14c5d2SYehuda Sadeh if (newmap != osdc->osdmap) { 21123d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 21133d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 21143d14c5d2SYehuda Sadeh } 21159a1ea2dbSJosh Durgin was_full = was_full || 21169a1ea2dbSJosh Durgin ceph_osdmap_flag(osdc->osdmap, 21179a1ea2dbSJosh Durgin CEPH_OSDMAP_FULL); 21189a1ea2dbSJosh Durgin kick_requests(osdc, 0, was_full); 21193d14c5d2SYehuda Sadeh } else { 21203d14c5d2SYehuda Sadeh dout("ignoring incremental map %u len %d\n", 21213d14c5d2SYehuda Sadeh epoch, maplen); 21223d14c5d2SYehuda Sadeh } 21233d14c5d2SYehuda Sadeh p = next; 21243d14c5d2SYehuda Sadeh nr_maps--; 21253d14c5d2SYehuda Sadeh } 21263d14c5d2SYehuda Sadeh if (newmap) 21273d14c5d2SYehuda Sadeh goto done; 21283d14c5d2SYehuda Sadeh 21293d14c5d2SYehuda Sadeh /* full maps */ 21303d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 21313d14c5d2SYehuda Sadeh dout(" %d full maps\n", nr_maps); 21323d14c5d2SYehuda Sadeh while (nr_maps) { 21333d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 21343d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 21353d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 21363d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 21373d14c5d2SYehuda Sadeh if (nr_maps > 1) { 21383d14c5d2SYehuda Sadeh dout("skipping non-latest full map %u len %d\n", 21393d14c5d2SYehuda Sadeh epoch, maplen); 21403d14c5d2SYehuda Sadeh } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) { 21413d14c5d2SYehuda Sadeh dout("skipping full map %u len %d, " 21423d14c5d2SYehuda Sadeh "older than our %u\n", epoch, maplen, 21433d14c5d2SYehuda Sadeh osdc->osdmap->epoch); 21443d14c5d2SYehuda Sadeh } else { 214538d6453cSSage Weil int skipped_map = 0; 214638d6453cSSage Weil 21473d14c5d2SYehuda Sadeh dout("taking full map %u len %d\n", epoch, maplen); 2148a2505d63SIlya Dryomov newmap = ceph_osdmap_decode(&p, p+maplen); 21493d14c5d2SYehuda Sadeh if (IS_ERR(newmap)) { 21503d14c5d2SYehuda Sadeh err = PTR_ERR(newmap); 21513d14c5d2SYehuda Sadeh goto bad; 21523d14c5d2SYehuda Sadeh } 21533d14c5d2SYehuda Sadeh BUG_ON(!newmap); 21543d14c5d2SYehuda Sadeh oldmap = osdc->osdmap; 21553d14c5d2SYehuda Sadeh osdc->osdmap = newmap; 215638d6453cSSage Weil if (oldmap) { 215738d6453cSSage Weil if (oldmap->epoch + 1 < newmap->epoch) 215838d6453cSSage Weil skipped_map = 1; 21593d14c5d2SYehuda Sadeh ceph_osdmap_destroy(oldmap); 216038d6453cSSage Weil } 21619a1ea2dbSJosh Durgin was_full = was_full || 21629a1ea2dbSJosh Durgin ceph_osdmap_flag(osdc->osdmap, 21639a1ea2dbSJosh Durgin CEPH_OSDMAP_FULL); 21649a1ea2dbSJosh Durgin kick_requests(osdc, skipped_map, was_full); 21653d14c5d2SYehuda Sadeh } 21663d14c5d2SYehuda Sadeh p += maplen; 21673d14c5d2SYehuda Sadeh nr_maps--; 21683d14c5d2SYehuda Sadeh } 21693d14c5d2SYehuda Sadeh 2170b72e19b9SDan Carpenter if (!osdc->osdmap) 2171b72e19b9SDan Carpenter goto bad; 21723d14c5d2SYehuda Sadeh done: 21733d14c5d2SYehuda Sadeh downgrade_write(&osdc->map_sem); 21743d14c5d2SYehuda Sadeh ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch); 2175cd634fb6SSage Weil 2176cd634fb6SSage Weil /* 2177cd634fb6SSage Weil * subscribe to subsequent osdmap updates if full to ensure 2178cd634fb6SSage Weil * we find out when we are no longer full and stop returning 2179cd634fb6SSage Weil * ENOSPC. 2180cd634fb6SSage Weil */ 2181d29adb34SJosh Durgin if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) || 2182d29adb34SJosh Durgin ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD) || 2183d29adb34SJosh Durgin ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR)) 2184cd634fb6SSage Weil ceph_monc_request_next_osdmap(&osdc->client->monc); 2185cd634fb6SSage Weil 2186f9d25199SAlex Elder mutex_lock(&osdc->request_mutex); 2187f9d25199SAlex Elder __send_queued(osdc); 2188f9d25199SAlex Elder mutex_unlock(&osdc->request_mutex); 21893d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 21903d14c5d2SYehuda Sadeh wake_up_all(&osdc->client->auth_wq); 21913d14c5d2SYehuda Sadeh return; 21923d14c5d2SYehuda Sadeh 21933d14c5d2SYehuda Sadeh bad: 21943d14c5d2SYehuda Sadeh pr_err("osdc handle_map corrupt msg\n"); 21953d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 21963d14c5d2SYehuda Sadeh up_write(&osdc->map_sem); 21973d14c5d2SYehuda Sadeh } 21983d14c5d2SYehuda Sadeh 21993d14c5d2SYehuda Sadeh /* 2200a40c4f10SYehuda Sadeh * watch/notify callback event infrastructure 2201a40c4f10SYehuda Sadeh * 2202a40c4f10SYehuda Sadeh * These callbacks are used both for watch and notify operations. 2203a40c4f10SYehuda Sadeh */ 2204a40c4f10SYehuda Sadeh static void __release_event(struct kref *kref) 2205a40c4f10SYehuda Sadeh { 2206a40c4f10SYehuda Sadeh struct ceph_osd_event *event = 2207a40c4f10SYehuda Sadeh container_of(kref, struct ceph_osd_event, kref); 2208a40c4f10SYehuda Sadeh 2209a40c4f10SYehuda Sadeh dout("__release_event %p\n", event); 2210a40c4f10SYehuda Sadeh kfree(event); 2211a40c4f10SYehuda Sadeh } 2212a40c4f10SYehuda Sadeh 2213a40c4f10SYehuda Sadeh static void get_event(struct ceph_osd_event *event) 2214a40c4f10SYehuda Sadeh { 2215a40c4f10SYehuda Sadeh kref_get(&event->kref); 2216a40c4f10SYehuda Sadeh } 2217a40c4f10SYehuda Sadeh 2218a40c4f10SYehuda Sadeh void ceph_osdc_put_event(struct ceph_osd_event *event) 2219a40c4f10SYehuda Sadeh { 2220a40c4f10SYehuda Sadeh kref_put(&event->kref, __release_event); 2221a40c4f10SYehuda Sadeh } 2222a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_put_event); 2223a40c4f10SYehuda Sadeh 2224a40c4f10SYehuda Sadeh static void __insert_event(struct ceph_osd_client *osdc, 2225a40c4f10SYehuda Sadeh struct ceph_osd_event *new) 2226a40c4f10SYehuda Sadeh { 2227a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 2228a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 2229a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 2230a40c4f10SYehuda Sadeh 2231a40c4f10SYehuda Sadeh while (*p) { 2232a40c4f10SYehuda Sadeh parent = *p; 2233a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 2234a40c4f10SYehuda Sadeh if (new->cookie < event->cookie) 2235a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 2236a40c4f10SYehuda Sadeh else if (new->cookie > event->cookie) 2237a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 2238a40c4f10SYehuda Sadeh else 2239a40c4f10SYehuda Sadeh BUG(); 2240a40c4f10SYehuda Sadeh } 2241a40c4f10SYehuda Sadeh 2242a40c4f10SYehuda Sadeh rb_link_node(&new->node, parent, p); 2243a40c4f10SYehuda Sadeh rb_insert_color(&new->node, &osdc->event_tree); 2244a40c4f10SYehuda Sadeh } 2245a40c4f10SYehuda Sadeh 2246a40c4f10SYehuda Sadeh static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc, 2247a40c4f10SYehuda Sadeh u64 cookie) 2248a40c4f10SYehuda Sadeh { 2249a40c4f10SYehuda Sadeh struct rb_node **p = &osdc->event_tree.rb_node; 2250a40c4f10SYehuda Sadeh struct rb_node *parent = NULL; 2251a40c4f10SYehuda Sadeh struct ceph_osd_event *event = NULL; 2252a40c4f10SYehuda Sadeh 2253a40c4f10SYehuda Sadeh while (*p) { 2254a40c4f10SYehuda Sadeh parent = *p; 2255a40c4f10SYehuda Sadeh event = rb_entry(parent, struct ceph_osd_event, node); 2256a40c4f10SYehuda Sadeh if (cookie < event->cookie) 2257a40c4f10SYehuda Sadeh p = &(*p)->rb_left; 2258a40c4f10SYehuda Sadeh else if (cookie > event->cookie) 2259a40c4f10SYehuda Sadeh p = &(*p)->rb_right; 2260a40c4f10SYehuda Sadeh else 2261a40c4f10SYehuda Sadeh return event; 2262a40c4f10SYehuda Sadeh } 2263a40c4f10SYehuda Sadeh return NULL; 2264a40c4f10SYehuda Sadeh } 2265a40c4f10SYehuda Sadeh 2266a40c4f10SYehuda Sadeh static void __remove_event(struct ceph_osd_event *event) 2267a40c4f10SYehuda Sadeh { 2268a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 2269a40c4f10SYehuda Sadeh 2270a40c4f10SYehuda Sadeh if (!RB_EMPTY_NODE(&event->node)) { 2271a40c4f10SYehuda Sadeh dout("__remove_event removed %p\n", event); 2272a40c4f10SYehuda Sadeh rb_erase(&event->node, &osdc->event_tree); 2273a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 2274a40c4f10SYehuda Sadeh } else { 2275a40c4f10SYehuda Sadeh dout("__remove_event didn't remove %p\n", event); 2276a40c4f10SYehuda Sadeh } 2277a40c4f10SYehuda Sadeh } 2278a40c4f10SYehuda Sadeh 2279a40c4f10SYehuda Sadeh int ceph_osdc_create_event(struct ceph_osd_client *osdc, 2280a40c4f10SYehuda Sadeh void (*event_cb)(u64, u64, u8, void *), 22813c663bbdSAlex Elder void *data, struct ceph_osd_event **pevent) 2282a40c4f10SYehuda Sadeh { 2283a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 2284a40c4f10SYehuda Sadeh 2285a40c4f10SYehuda Sadeh event = kmalloc(sizeof(*event), GFP_NOIO); 2286a40c4f10SYehuda Sadeh if (!event) 2287a40c4f10SYehuda Sadeh return -ENOMEM; 2288a40c4f10SYehuda Sadeh 2289a40c4f10SYehuda Sadeh dout("create_event %p\n", event); 2290a40c4f10SYehuda Sadeh event->cb = event_cb; 22913c663bbdSAlex Elder event->one_shot = 0; 2292a40c4f10SYehuda Sadeh event->data = data; 2293a40c4f10SYehuda Sadeh event->osdc = osdc; 2294a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&event->osd_node); 22953ee5234dSAlex Elder RB_CLEAR_NODE(&event->node); 2296a40c4f10SYehuda Sadeh kref_init(&event->kref); /* one ref for us */ 2297a40c4f10SYehuda Sadeh kref_get(&event->kref); /* one ref for the caller */ 2298a40c4f10SYehuda Sadeh 2299a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 2300a40c4f10SYehuda Sadeh event->cookie = ++osdc->event_count; 2301a40c4f10SYehuda Sadeh __insert_event(osdc, event); 2302a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 2303a40c4f10SYehuda Sadeh 2304a40c4f10SYehuda Sadeh *pevent = event; 2305a40c4f10SYehuda Sadeh return 0; 2306a40c4f10SYehuda Sadeh } 2307a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_create_event); 2308a40c4f10SYehuda Sadeh 2309a40c4f10SYehuda Sadeh void ceph_osdc_cancel_event(struct ceph_osd_event *event) 2310a40c4f10SYehuda Sadeh { 2311a40c4f10SYehuda Sadeh struct ceph_osd_client *osdc = event->osdc; 2312a40c4f10SYehuda Sadeh 2313a40c4f10SYehuda Sadeh dout("cancel_event %p\n", event); 2314a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 2315a40c4f10SYehuda Sadeh __remove_event(event); 2316a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 2317a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); /* caller's */ 2318a40c4f10SYehuda Sadeh } 2319a40c4f10SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_cancel_event); 2320a40c4f10SYehuda Sadeh 2321a40c4f10SYehuda Sadeh 2322a40c4f10SYehuda Sadeh static void do_event_work(struct work_struct *work) 2323a40c4f10SYehuda Sadeh { 2324a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work = 2325a40c4f10SYehuda Sadeh container_of(work, struct ceph_osd_event_work, work); 2326a40c4f10SYehuda Sadeh struct ceph_osd_event *event = event_work->event; 2327a40c4f10SYehuda Sadeh u64 ver = event_work->ver; 2328a40c4f10SYehuda Sadeh u64 notify_id = event_work->notify_id; 2329a40c4f10SYehuda Sadeh u8 opcode = event_work->opcode; 2330a40c4f10SYehuda Sadeh 2331a40c4f10SYehuda Sadeh dout("do_event_work completing %p\n", event); 2332a40c4f10SYehuda Sadeh event->cb(ver, notify_id, opcode, event->data); 2333a40c4f10SYehuda Sadeh dout("do_event_work completed %p\n", event); 2334a40c4f10SYehuda Sadeh ceph_osdc_put_event(event); 2335a40c4f10SYehuda Sadeh kfree(event_work); 2336a40c4f10SYehuda Sadeh } 2337a40c4f10SYehuda Sadeh 2338a40c4f10SYehuda Sadeh 2339a40c4f10SYehuda Sadeh /* 2340a40c4f10SYehuda Sadeh * Process osd watch notifications 2341a40c4f10SYehuda Sadeh */ 23423c663bbdSAlex Elder static void handle_watch_notify(struct ceph_osd_client *osdc, 23433c663bbdSAlex Elder struct ceph_msg *msg) 2344a40c4f10SYehuda Sadeh { 2345a40c4f10SYehuda Sadeh void *p, *end; 2346a40c4f10SYehuda Sadeh u8 proto_ver; 2347a40c4f10SYehuda Sadeh u64 cookie, ver, notify_id; 2348a40c4f10SYehuda Sadeh u8 opcode; 2349a40c4f10SYehuda Sadeh struct ceph_osd_event *event; 2350a40c4f10SYehuda Sadeh struct ceph_osd_event_work *event_work; 2351a40c4f10SYehuda Sadeh 2352a40c4f10SYehuda Sadeh p = msg->front.iov_base; 2353a40c4f10SYehuda Sadeh end = p + msg->front.iov_len; 2354a40c4f10SYehuda Sadeh 2355a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, proto_ver, bad); 2356a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, opcode, bad); 2357a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, cookie, bad); 2358a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, ver, bad); 2359a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, notify_id, bad); 2360a40c4f10SYehuda Sadeh 2361a40c4f10SYehuda Sadeh spin_lock(&osdc->event_lock); 2362a40c4f10SYehuda Sadeh event = __find_event(osdc, cookie); 2363a40c4f10SYehuda Sadeh if (event) { 23643c663bbdSAlex Elder BUG_ON(event->one_shot); 2365a40c4f10SYehuda Sadeh get_event(event); 2366a40c4f10SYehuda Sadeh } 2367a40c4f10SYehuda Sadeh spin_unlock(&osdc->event_lock); 2368a40c4f10SYehuda Sadeh dout("handle_watch_notify cookie %lld ver %lld event %p\n", 2369a40c4f10SYehuda Sadeh cookie, ver, event); 2370a40c4f10SYehuda Sadeh if (event) { 2371a40c4f10SYehuda Sadeh event_work = kmalloc(sizeof(*event_work), GFP_NOIO); 2372a40c4f10SYehuda Sadeh if (!event_work) { 237391883cd2SIlya Dryomov pr_err("couldn't allocate event_work\n"); 237491883cd2SIlya Dryomov ceph_osdc_put_event(event); 237591883cd2SIlya Dryomov return; 2376a40c4f10SYehuda Sadeh } 23776b0ae409SMariusz Kozlowski INIT_WORK(&event_work->work, do_event_work); 2378a40c4f10SYehuda Sadeh event_work->event = event; 2379a40c4f10SYehuda Sadeh event_work->ver = ver; 2380a40c4f10SYehuda Sadeh event_work->notify_id = notify_id; 2381a40c4f10SYehuda Sadeh event_work->opcode = opcode; 238291883cd2SIlya Dryomov 238391883cd2SIlya Dryomov queue_work(osdc->notify_wq, &event_work->work); 2384a40c4f10SYehuda Sadeh } 2385a40c4f10SYehuda Sadeh 2386a40c4f10SYehuda Sadeh return; 2387a40c4f10SYehuda Sadeh 2388a40c4f10SYehuda Sadeh bad: 2389a40c4f10SYehuda Sadeh pr_err("osdc handle_watch_notify corrupt msg\n"); 2390a40c4f10SYehuda Sadeh } 2391a40c4f10SYehuda Sadeh 2392a40c4f10SYehuda Sadeh /* 2393e65550fdSAlex Elder * build new request AND message 2394e65550fdSAlex Elder * 2395e65550fdSAlex Elder */ 2396e65550fdSAlex Elder void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off, 2397e65550fdSAlex Elder struct ceph_snap_context *snapc, u64 snap_id, 2398e65550fdSAlex Elder struct timespec *mtime) 2399e65550fdSAlex Elder { 2400e65550fdSAlex Elder struct ceph_msg *msg = req->r_request; 2401e65550fdSAlex Elder void *p; 2402e65550fdSAlex Elder size_t msg_size; 2403e65550fdSAlex Elder int flags = req->r_flags; 2404e65550fdSAlex Elder u64 data_len; 2405e65550fdSAlex Elder unsigned int i; 2406e65550fdSAlex Elder 2407e65550fdSAlex Elder req->r_snapid = snap_id; 2408e65550fdSAlex Elder req->r_snapc = ceph_get_snap_context(snapc); 2409e65550fdSAlex Elder 2410e65550fdSAlex Elder /* encode request */ 2411e65550fdSAlex Elder msg->hdr.version = cpu_to_le16(4); 2412e65550fdSAlex Elder 2413e65550fdSAlex Elder p = msg->front.iov_base; 2414e65550fdSAlex Elder ceph_encode_32(&p, 1); /* client_inc is always 1 */ 2415e65550fdSAlex Elder req->r_request_osdmap_epoch = p; 2416e65550fdSAlex Elder p += 4; 2417e65550fdSAlex Elder req->r_request_flags = p; 2418e65550fdSAlex Elder p += 4; 2419e65550fdSAlex Elder if (req->r_flags & CEPH_OSD_FLAG_WRITE) 2420e65550fdSAlex Elder ceph_encode_timespec(p, mtime); 2421e65550fdSAlex Elder p += sizeof(struct ceph_timespec); 2422e65550fdSAlex Elder req->r_request_reassert_version = p; 2423e65550fdSAlex Elder p += sizeof(struct ceph_eversion); /* will get filled in */ 2424e65550fdSAlex Elder 2425e65550fdSAlex Elder /* oloc */ 2426e65550fdSAlex Elder ceph_encode_8(&p, 4); 2427e65550fdSAlex Elder ceph_encode_8(&p, 4); 2428e65550fdSAlex Elder ceph_encode_32(&p, 8 + 4 + 4); 2429e65550fdSAlex Elder req->r_request_pool = p; 2430e65550fdSAlex Elder p += 8; 2431e65550fdSAlex Elder ceph_encode_32(&p, -1); /* preferred */ 2432e65550fdSAlex Elder ceph_encode_32(&p, 0); /* key len */ 2433e65550fdSAlex Elder 2434e65550fdSAlex Elder ceph_encode_8(&p, 1); 2435e65550fdSAlex Elder req->r_request_pgid = p; 2436e65550fdSAlex Elder p += 8 + 4; 2437e65550fdSAlex Elder ceph_encode_32(&p, -1); /* preferred */ 2438e65550fdSAlex Elder 2439e65550fdSAlex Elder /* oid */ 24403c972c95SIlya Dryomov ceph_encode_32(&p, req->r_base_oid.name_len); 24413c972c95SIlya Dryomov memcpy(p, req->r_base_oid.name, req->r_base_oid.name_len); 24423c972c95SIlya Dryomov dout("oid '%.*s' len %d\n", req->r_base_oid.name_len, 24433c972c95SIlya Dryomov req->r_base_oid.name, req->r_base_oid.name_len); 24443c972c95SIlya Dryomov p += req->r_base_oid.name_len; 2445e65550fdSAlex Elder 2446e65550fdSAlex Elder /* ops--can imply data */ 2447e65550fdSAlex Elder ceph_encode_16(&p, (u16)req->r_num_ops); 2448e65550fdSAlex Elder data_len = 0; 2449e65550fdSAlex Elder for (i = 0; i < req->r_num_ops; i++) { 2450e65550fdSAlex Elder data_len += osd_req_encode_op(req, p, i); 2451e65550fdSAlex Elder p += sizeof(struct ceph_osd_op); 2452e65550fdSAlex Elder } 2453e65550fdSAlex Elder 2454e65550fdSAlex Elder /* snaps */ 2455e65550fdSAlex Elder ceph_encode_64(&p, req->r_snapid); 2456e65550fdSAlex Elder ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0); 2457e65550fdSAlex Elder ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0); 2458e65550fdSAlex Elder if (req->r_snapc) { 2459e65550fdSAlex Elder for (i = 0; i < snapc->num_snaps; i++) { 2460e65550fdSAlex Elder ceph_encode_64(&p, req->r_snapc->snaps[i]); 2461e65550fdSAlex Elder } 2462e65550fdSAlex Elder } 2463e65550fdSAlex Elder 2464e65550fdSAlex Elder req->r_request_attempts = p; 2465e65550fdSAlex Elder p += 4; 2466e65550fdSAlex Elder 2467e65550fdSAlex Elder /* data */ 2468e65550fdSAlex Elder if (flags & CEPH_OSD_FLAG_WRITE) { 2469e65550fdSAlex Elder u16 data_off; 2470e65550fdSAlex Elder 2471e65550fdSAlex Elder /* 2472e65550fdSAlex Elder * The header "data_off" is a hint to the receiver 2473e65550fdSAlex Elder * allowing it to align received data into its 2474e65550fdSAlex Elder * buffers such that there's no need to re-copy 2475e65550fdSAlex Elder * it before writing it to disk (direct I/O). 2476e65550fdSAlex Elder */ 2477e65550fdSAlex Elder data_off = (u16) (off & 0xffff); 2478e65550fdSAlex Elder req->r_request->hdr.data_off = cpu_to_le16(data_off); 2479e65550fdSAlex Elder } 2480e65550fdSAlex Elder req->r_request->hdr.data_len = cpu_to_le32(data_len); 2481e65550fdSAlex Elder 2482e65550fdSAlex Elder BUG_ON(p > msg->front.iov_base + msg->front.iov_len); 2483e65550fdSAlex Elder msg_size = p - msg->front.iov_base; 2484e65550fdSAlex Elder msg->front.iov_len = msg_size; 2485e65550fdSAlex Elder msg->hdr.front_len = cpu_to_le32(msg_size); 2486e65550fdSAlex Elder 2487e65550fdSAlex Elder dout("build_request msg_size was %d\n", (int)msg_size); 2488e65550fdSAlex Elder } 2489e65550fdSAlex Elder EXPORT_SYMBOL(ceph_osdc_build_request); 2490e65550fdSAlex Elder 2491e65550fdSAlex Elder /* 24923d14c5d2SYehuda Sadeh * Register request, send initial attempt. 24933d14c5d2SYehuda Sadeh */ 24943d14c5d2SYehuda Sadeh int ceph_osdc_start_request(struct ceph_osd_client *osdc, 24953d14c5d2SYehuda Sadeh struct ceph_osd_request *req, 24963d14c5d2SYehuda Sadeh bool nofail) 24973d14c5d2SYehuda Sadeh { 24980bbfdfe8SIlya Dryomov int rc; 24993d14c5d2SYehuda Sadeh 25003d14c5d2SYehuda Sadeh down_read(&osdc->map_sem); 25013d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 25020bbfdfe8SIlya Dryomov 25030bbfdfe8SIlya Dryomov rc = __ceph_osdc_start_request(osdc, req, nofail); 25040bbfdfe8SIlya Dryomov 25053d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 25063d14c5d2SYehuda Sadeh up_read(&osdc->map_sem); 25070bbfdfe8SIlya Dryomov 25083d14c5d2SYehuda Sadeh return rc; 25093d14c5d2SYehuda Sadeh } 25103d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_start_request); 25113d14c5d2SYehuda Sadeh 25123d14c5d2SYehuda Sadeh /* 2513c9f9b93dSIlya Dryomov * Unregister a registered request. The request is not completed (i.e. 2514c9f9b93dSIlya Dryomov * no callbacks or wakeups) - higher layers are supposed to know what 2515c9f9b93dSIlya Dryomov * they are canceling. 2516c9f9b93dSIlya Dryomov */ 2517c9f9b93dSIlya Dryomov void ceph_osdc_cancel_request(struct ceph_osd_request *req) 2518c9f9b93dSIlya Dryomov { 2519c9f9b93dSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 2520c9f9b93dSIlya Dryomov 2521c9f9b93dSIlya Dryomov mutex_lock(&osdc->request_mutex); 2522c9f9b93dSIlya Dryomov if (req->r_linger) 2523c9f9b93dSIlya Dryomov __unregister_linger_request(osdc, req); 2524c9f9b93dSIlya Dryomov __unregister_request(osdc, req); 2525c9f9b93dSIlya Dryomov mutex_unlock(&osdc->request_mutex); 2526c9f9b93dSIlya Dryomov 2527c9f9b93dSIlya Dryomov dout("%s %p tid %llu canceled\n", __func__, req, req->r_tid); 2528c9f9b93dSIlya Dryomov } 2529c9f9b93dSIlya Dryomov EXPORT_SYMBOL(ceph_osdc_cancel_request); 2530c9f9b93dSIlya Dryomov 2531c9f9b93dSIlya Dryomov /* 25323d14c5d2SYehuda Sadeh * wait for a request to complete 25333d14c5d2SYehuda Sadeh */ 25343d14c5d2SYehuda Sadeh int ceph_osdc_wait_request(struct ceph_osd_client *osdc, 25353d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 25363d14c5d2SYehuda Sadeh { 25373d14c5d2SYehuda Sadeh int rc; 25383d14c5d2SYehuda Sadeh 2539c9f9b93dSIlya Dryomov dout("%s %p tid %llu\n", __func__, req, req->r_tid); 2540c9f9b93dSIlya Dryomov 25413d14c5d2SYehuda Sadeh rc = wait_for_completion_interruptible(&req->r_completion); 25423d14c5d2SYehuda Sadeh if (rc < 0) { 2543c9f9b93dSIlya Dryomov dout("%s %p tid %llu interrupted\n", __func__, req, req->r_tid); 2544c9f9b93dSIlya Dryomov ceph_osdc_cancel_request(req); 254525845472SSage Weil complete_request(req); 25463d14c5d2SYehuda Sadeh return rc; 25473d14c5d2SYehuda Sadeh } 25483d14c5d2SYehuda Sadeh 2549c9f9b93dSIlya Dryomov dout("%s %p tid %llu result %d\n", __func__, req, req->r_tid, 2550c9f9b93dSIlya Dryomov req->r_result); 25513d14c5d2SYehuda Sadeh return req->r_result; 25523d14c5d2SYehuda Sadeh } 25533d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_wait_request); 25543d14c5d2SYehuda Sadeh 25553d14c5d2SYehuda Sadeh /* 25563d14c5d2SYehuda Sadeh * sync - wait for all in-flight requests to flush. avoid starvation. 25573d14c5d2SYehuda Sadeh */ 25583d14c5d2SYehuda Sadeh void ceph_osdc_sync(struct ceph_osd_client *osdc) 25593d14c5d2SYehuda Sadeh { 25603d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 25613d14c5d2SYehuda Sadeh u64 last_tid, next_tid = 0; 25623d14c5d2SYehuda Sadeh 25633d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 25643d14c5d2SYehuda Sadeh last_tid = osdc->last_tid; 25653d14c5d2SYehuda Sadeh while (1) { 25663d14c5d2SYehuda Sadeh req = __lookup_request_ge(osdc, next_tid); 25673d14c5d2SYehuda Sadeh if (!req) 25683d14c5d2SYehuda Sadeh break; 25693d14c5d2SYehuda Sadeh if (req->r_tid > last_tid) 25703d14c5d2SYehuda Sadeh break; 25713d14c5d2SYehuda Sadeh 25723d14c5d2SYehuda Sadeh next_tid = req->r_tid + 1; 25733d14c5d2SYehuda Sadeh if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0) 25743d14c5d2SYehuda Sadeh continue; 25753d14c5d2SYehuda Sadeh 25763d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 25773d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 25783d14c5d2SYehuda Sadeh dout("sync waiting on tid %llu (last is %llu)\n", 25793d14c5d2SYehuda Sadeh req->r_tid, last_tid); 25803d14c5d2SYehuda Sadeh wait_for_completion(&req->r_safe_completion); 25813d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 25823d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 25833d14c5d2SYehuda Sadeh } 25843d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 25853d14c5d2SYehuda Sadeh dout("sync done (thru tid %llu)\n", last_tid); 25863d14c5d2SYehuda Sadeh } 25873d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_sync); 25883d14c5d2SYehuda Sadeh 25893d14c5d2SYehuda Sadeh /* 2590dd935f44SJosh Durgin * Call all pending notify callbacks - for use after a watch is 2591dd935f44SJosh Durgin * unregistered, to make sure no more callbacks for it will be invoked 2592dd935f44SJosh Durgin */ 2593f6479449Sstephen hemminger void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc) 2594dd935f44SJosh Durgin { 2595dd935f44SJosh Durgin flush_workqueue(osdc->notify_wq); 2596dd935f44SJosh Durgin } 2597dd935f44SJosh Durgin EXPORT_SYMBOL(ceph_osdc_flush_notifies); 2598dd935f44SJosh Durgin 2599dd935f44SJosh Durgin 2600dd935f44SJosh Durgin /* 26013d14c5d2SYehuda Sadeh * init, shutdown 26023d14c5d2SYehuda Sadeh */ 26033d14c5d2SYehuda Sadeh int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client) 26043d14c5d2SYehuda Sadeh { 26053d14c5d2SYehuda Sadeh int err; 26063d14c5d2SYehuda Sadeh 26073d14c5d2SYehuda Sadeh dout("init\n"); 26083d14c5d2SYehuda Sadeh osdc->client = client; 26093d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 26103d14c5d2SYehuda Sadeh init_rwsem(&osdc->map_sem); 26113d14c5d2SYehuda Sadeh init_completion(&osdc->map_waiters); 26123d14c5d2SYehuda Sadeh osdc->last_requested_map = 0; 26133d14c5d2SYehuda Sadeh mutex_init(&osdc->request_mutex); 26143d14c5d2SYehuda Sadeh osdc->last_tid = 0; 26153d14c5d2SYehuda Sadeh osdc->osds = RB_ROOT; 26163d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->osd_lru); 26173d14c5d2SYehuda Sadeh osdc->requests = RB_ROOT; 26183d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_lru); 26196f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_unsent); 26206f6c7006SSage Weil INIT_LIST_HEAD(&osdc->req_notarget); 2621a40c4f10SYehuda Sadeh INIT_LIST_HEAD(&osdc->req_linger); 26223d14c5d2SYehuda Sadeh osdc->num_requests = 0; 26233d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout); 26243d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout); 2625a40c4f10SYehuda Sadeh spin_lock_init(&osdc->event_lock); 2626a40c4f10SYehuda Sadeh osdc->event_tree = RB_ROOT; 2627a40c4f10SYehuda Sadeh osdc->event_count = 0; 26283d14c5d2SYehuda Sadeh 26293d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 2630a319bf56SIlya Dryomov round_jiffies_relative(osdc->client->options->osd_idle_ttl)); 26313d14c5d2SYehuda Sadeh 26323d14c5d2SYehuda Sadeh err = -ENOMEM; 26333d14c5d2SYehuda Sadeh osdc->req_mempool = mempool_create_kmalloc_pool(10, 26343d14c5d2SYehuda Sadeh sizeof(struct ceph_osd_request)); 26353d14c5d2SYehuda Sadeh if (!osdc->req_mempool) 26363d14c5d2SYehuda Sadeh goto out; 26373d14c5d2SYehuda Sadeh 2638d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP, 2639d50b409fSSage Weil OSD_OP_FRONT_LEN, 10, true, 26403d14c5d2SYehuda Sadeh "osd_op"); 26413d14c5d2SYehuda Sadeh if (err < 0) 26423d14c5d2SYehuda Sadeh goto out_mempool; 2643d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY, 26443d14c5d2SYehuda Sadeh OSD_OPREPLY_FRONT_LEN, 10, true, 26453d14c5d2SYehuda Sadeh "osd_op_reply"); 26463d14c5d2SYehuda Sadeh if (err < 0) 26473d14c5d2SYehuda Sadeh goto out_msgpool; 2648a40c4f10SYehuda Sadeh 2649dbcae088SDan Carpenter err = -ENOMEM; 2650a40c4f10SYehuda Sadeh osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify"); 2651dbcae088SDan Carpenter if (!osdc->notify_wq) 2652c172ec5cSIlya Dryomov goto out_msgpool_reply; 2653c172ec5cSIlya Dryomov 26543d14c5d2SYehuda Sadeh return 0; 26553d14c5d2SYehuda Sadeh 2656c172ec5cSIlya Dryomov out_msgpool_reply: 2657c172ec5cSIlya Dryomov ceph_msgpool_destroy(&osdc->msgpool_op_reply); 26583d14c5d2SYehuda Sadeh out_msgpool: 26593d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 26603d14c5d2SYehuda Sadeh out_mempool: 26613d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 26623d14c5d2SYehuda Sadeh out: 26633d14c5d2SYehuda Sadeh return err; 26643d14c5d2SYehuda Sadeh } 26653d14c5d2SYehuda Sadeh 26663d14c5d2SYehuda Sadeh void ceph_osdc_stop(struct ceph_osd_client *osdc) 26673d14c5d2SYehuda Sadeh { 2668a40c4f10SYehuda Sadeh flush_workqueue(osdc->notify_wq); 2669a40c4f10SYehuda Sadeh destroy_workqueue(osdc->notify_wq); 26703d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->timeout_work); 26713d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->osds_timeout_work); 26723d14c5d2SYehuda Sadeh if (osdc->osdmap) { 26733d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 26743d14c5d2SYehuda Sadeh osdc->osdmap = NULL; 26753d14c5d2SYehuda Sadeh } 2676aca420bcSSage Weil remove_all_osds(osdc); 26773d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 26783d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 26793d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op_reply); 26803d14c5d2SYehuda Sadeh } 26813d14c5d2SYehuda Sadeh 26823d14c5d2SYehuda Sadeh /* 26833d14c5d2SYehuda Sadeh * Read some contiguous pages. If we cross a stripe boundary, shorten 26843d14c5d2SYehuda Sadeh * *plen. Return number of bytes read, or error. 26853d14c5d2SYehuda Sadeh */ 26863d14c5d2SYehuda Sadeh int ceph_osdc_readpages(struct ceph_osd_client *osdc, 26873d14c5d2SYehuda Sadeh struct ceph_vino vino, struct ceph_file_layout *layout, 26883d14c5d2SYehuda Sadeh u64 off, u64 *plen, 26893d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 2690b7495fc2SSage Weil struct page **pages, int num_pages, int page_align) 26913d14c5d2SYehuda Sadeh { 26923d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 26933d14c5d2SYehuda Sadeh int rc = 0; 26943d14c5d2SYehuda Sadeh 26953d14c5d2SYehuda Sadeh dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino, 26963d14c5d2SYehuda Sadeh vino.snap, off, *plen); 2697715e4cd4SYan, Zheng req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 0, 1, 26983d14c5d2SYehuda Sadeh CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ, 2699acead002SAlex Elder NULL, truncate_seq, truncate_size, 2700153e5167SAlex Elder false); 27016816282dSSage Weil if (IS_ERR(req)) 27026816282dSSage Weil return PTR_ERR(req); 27033d14c5d2SYehuda Sadeh 27043d14c5d2SYehuda Sadeh /* it may be a short read due to an object boundary */ 27050fff87ecSAlex Elder 2706406e2c9fSAlex Elder osd_req_op_extent_osd_data_pages(req, 0, 2707a4ce40a9SAlex Elder pages, *plen, page_align, false, false); 27083d14c5d2SYehuda Sadeh 2709e0c59487SAlex Elder dout("readpages final extent is %llu~%llu (%llu bytes align %d)\n", 271043bfe5deSAlex Elder off, *plen, *plen, page_align); 27113d14c5d2SYehuda Sadeh 271279528734SAlex Elder ceph_osdc_build_request(req, off, NULL, vino.snap, NULL); 271302ee07d3SAlex Elder 27143d14c5d2SYehuda Sadeh rc = ceph_osdc_start_request(osdc, req, false); 27153d14c5d2SYehuda Sadeh if (!rc) 27163d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 27173d14c5d2SYehuda Sadeh 27183d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 27193d14c5d2SYehuda Sadeh dout("readpages result %d\n", rc); 27203d14c5d2SYehuda Sadeh return rc; 27213d14c5d2SYehuda Sadeh } 27223d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_readpages); 27233d14c5d2SYehuda Sadeh 27243d14c5d2SYehuda Sadeh /* 27253d14c5d2SYehuda Sadeh * do a synchronous write on N pages 27263d14c5d2SYehuda Sadeh */ 27273d14c5d2SYehuda Sadeh int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino, 27283d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 27293d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 27303d14c5d2SYehuda Sadeh u64 off, u64 len, 27313d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 27323d14c5d2SYehuda Sadeh struct timespec *mtime, 273324808826SAlex Elder struct page **pages, int num_pages) 27343d14c5d2SYehuda Sadeh { 27353d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 27363d14c5d2SYehuda Sadeh int rc = 0; 2737b7495fc2SSage Weil int page_align = off & ~PAGE_MASK; 27383d14c5d2SYehuda Sadeh 2739acead002SAlex Elder BUG_ON(vino.snap != CEPH_NOSNAP); /* snapshots aren't writeable */ 2740715e4cd4SYan, Zheng req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 0, 1, 27413d14c5d2SYehuda Sadeh CEPH_OSD_OP_WRITE, 274224808826SAlex Elder CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE, 2743acead002SAlex Elder snapc, truncate_seq, truncate_size, 2744153e5167SAlex Elder true); 27456816282dSSage Weil if (IS_ERR(req)) 27466816282dSSage Weil return PTR_ERR(req); 27473d14c5d2SYehuda Sadeh 27483d14c5d2SYehuda Sadeh /* it may be a short write due to an object boundary */ 2749406e2c9fSAlex Elder osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align, 275043bfe5deSAlex Elder false, false); 275143bfe5deSAlex Elder dout("writepages %llu~%llu (%llu bytes)\n", off, len, len); 27523d14c5d2SYehuda Sadeh 275379528734SAlex Elder ceph_osdc_build_request(req, off, snapc, CEPH_NOSNAP, mtime); 275402ee07d3SAlex Elder 275587f979d3SAlex Elder rc = ceph_osdc_start_request(osdc, req, true); 27563d14c5d2SYehuda Sadeh if (!rc) 27573d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 27583d14c5d2SYehuda Sadeh 27593d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 27603d14c5d2SYehuda Sadeh if (rc == 0) 27613d14c5d2SYehuda Sadeh rc = len; 27623d14c5d2SYehuda Sadeh dout("writepages result %d\n", rc); 27633d14c5d2SYehuda Sadeh return rc; 27643d14c5d2SYehuda Sadeh } 27653d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_writepages); 27663d14c5d2SYehuda Sadeh 27675522ae0bSAlex Elder int ceph_osdc_setup(void) 27685522ae0bSAlex Elder { 27695522ae0bSAlex Elder BUG_ON(ceph_osd_request_cache); 27705522ae0bSAlex Elder ceph_osd_request_cache = kmem_cache_create("ceph_osd_request", 27715522ae0bSAlex Elder sizeof (struct ceph_osd_request), 27725522ae0bSAlex Elder __alignof__(struct ceph_osd_request), 27735522ae0bSAlex Elder 0, NULL); 27745522ae0bSAlex Elder 27755522ae0bSAlex Elder return ceph_osd_request_cache ? 0 : -ENOMEM; 27765522ae0bSAlex Elder } 27775522ae0bSAlex Elder EXPORT_SYMBOL(ceph_osdc_setup); 27785522ae0bSAlex Elder 27795522ae0bSAlex Elder void ceph_osdc_cleanup(void) 27805522ae0bSAlex Elder { 27815522ae0bSAlex Elder BUG_ON(!ceph_osd_request_cache); 27825522ae0bSAlex Elder kmem_cache_destroy(ceph_osd_request_cache); 27835522ae0bSAlex Elder ceph_osd_request_cache = NULL; 27845522ae0bSAlex Elder } 27855522ae0bSAlex Elder EXPORT_SYMBOL(ceph_osdc_cleanup); 27865522ae0bSAlex Elder 27873d14c5d2SYehuda Sadeh /* 27883d14c5d2SYehuda Sadeh * handle incoming message 27893d14c5d2SYehuda Sadeh */ 27903d14c5d2SYehuda Sadeh static void dispatch(struct ceph_connection *con, struct ceph_msg *msg) 27913d14c5d2SYehuda Sadeh { 27923d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 27933d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc; 27943d14c5d2SYehuda Sadeh int type = le16_to_cpu(msg->hdr.type); 27953d14c5d2SYehuda Sadeh 27963d14c5d2SYehuda Sadeh if (!osd) 27973d14c5d2SYehuda Sadeh goto out; 27983d14c5d2SYehuda Sadeh osdc = osd->o_osdc; 27993d14c5d2SYehuda Sadeh 28003d14c5d2SYehuda Sadeh switch (type) { 28013d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 28023d14c5d2SYehuda Sadeh ceph_osdc_handle_map(osdc, msg); 28033d14c5d2SYehuda Sadeh break; 28043d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 28053d14c5d2SYehuda Sadeh handle_reply(osdc, msg, con); 28063d14c5d2SYehuda Sadeh break; 2807a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 2808a40c4f10SYehuda Sadeh handle_watch_notify(osdc, msg); 2809a40c4f10SYehuda Sadeh break; 28103d14c5d2SYehuda Sadeh 28113d14c5d2SYehuda Sadeh default: 28123d14c5d2SYehuda Sadeh pr_err("received unknown message type %d %s\n", type, 28133d14c5d2SYehuda Sadeh ceph_msg_type_name(type)); 28143d14c5d2SYehuda Sadeh } 28153d14c5d2SYehuda Sadeh out: 28163d14c5d2SYehuda Sadeh ceph_msg_put(msg); 28173d14c5d2SYehuda Sadeh } 28183d14c5d2SYehuda Sadeh 28193d14c5d2SYehuda Sadeh /* 2820d15f9d69SIlya Dryomov * Lookup and return message for incoming reply. Don't try to do 2821d15f9d69SIlya Dryomov * anything about a larger than preallocated data portion of the 2822d15f9d69SIlya Dryomov * message at the moment - for now, just skip the message. 28233d14c5d2SYehuda Sadeh */ 28243d14c5d2SYehuda Sadeh static struct ceph_msg *get_reply(struct ceph_connection *con, 28253d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 28263d14c5d2SYehuda Sadeh int *skip) 28273d14c5d2SYehuda Sadeh { 28283d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 28293d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = osd->o_osdc; 28303d14c5d2SYehuda Sadeh struct ceph_msg *m; 28313d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 28323f0a4ac5SIlya Dryomov int front_len = le32_to_cpu(hdr->front_len); 28333d14c5d2SYehuda Sadeh int data_len = le32_to_cpu(hdr->data_len); 28343d14c5d2SYehuda Sadeh u64 tid; 28353d14c5d2SYehuda Sadeh 28363d14c5d2SYehuda Sadeh tid = le64_to_cpu(hdr->tid); 28373d14c5d2SYehuda Sadeh mutex_lock(&osdc->request_mutex); 28383d14c5d2SYehuda Sadeh req = __lookup_request(osdc, tid); 28393d14c5d2SYehuda Sadeh if (!req) { 2840d15f9d69SIlya Dryomov pr_warn("%s osd%d tid %llu unknown, skipping\n", 2841d15f9d69SIlya Dryomov __func__, osd->o_osd, tid); 28423d14c5d2SYehuda Sadeh m = NULL; 2843d15f9d69SIlya Dryomov *skip = 1; 28443d14c5d2SYehuda Sadeh goto out; 28453d14c5d2SYehuda Sadeh } 28463d14c5d2SYehuda Sadeh 2847ace6d3a9SAlex Elder if (req->r_reply->con) 28488921d114SAlex Elder dout("%s revoking msg %p from old con %p\n", __func__, 2849ace6d3a9SAlex Elder req->r_reply, req->r_reply->con); 28508921d114SAlex Elder ceph_msg_revoke_incoming(req->r_reply); 28513d14c5d2SYehuda Sadeh 2852f2be82b0SIlya Dryomov if (front_len > req->r_reply->front_alloc_len) { 2853d15f9d69SIlya Dryomov pr_warn("%s osd%d tid %llu front %d > preallocated %d\n", 2854d15f9d69SIlya Dryomov __func__, osd->o_osd, req->r_tid, front_len, 2855d15f9d69SIlya Dryomov req->r_reply->front_alloc_len); 28563f0a4ac5SIlya Dryomov m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front_len, GFP_NOFS, 28573f0a4ac5SIlya Dryomov false); 28583d14c5d2SYehuda Sadeh if (!m) 28593d14c5d2SYehuda Sadeh goto out; 28603d14c5d2SYehuda Sadeh ceph_msg_put(req->r_reply); 28613d14c5d2SYehuda Sadeh req->r_reply = m; 28623d14c5d2SYehuda Sadeh } 28633d14c5d2SYehuda Sadeh 2864d15f9d69SIlya Dryomov if (data_len > req->r_reply->data_length) { 2865d15f9d69SIlya Dryomov pr_warn("%s osd%d tid %llu data %d > preallocated %zu, skipping\n", 2866d15f9d69SIlya Dryomov __func__, osd->o_osd, req->r_tid, data_len, 2867d15f9d69SIlya Dryomov req->r_reply->data_length); 28683d14c5d2SYehuda Sadeh m = NULL; 2869d15f9d69SIlya Dryomov *skip = 1; 28703d14c5d2SYehuda Sadeh goto out; 28713d14c5d2SYehuda Sadeh } 2872d15f9d69SIlya Dryomov 2873d15f9d69SIlya Dryomov m = ceph_msg_get(req->r_reply); 28743d14c5d2SYehuda Sadeh dout("get_reply tid %lld %p\n", tid, m); 28753d14c5d2SYehuda Sadeh 28763d14c5d2SYehuda Sadeh out: 28773d14c5d2SYehuda Sadeh mutex_unlock(&osdc->request_mutex); 28783d14c5d2SYehuda Sadeh return m; 28793d14c5d2SYehuda Sadeh } 28803d14c5d2SYehuda Sadeh 28813d14c5d2SYehuda Sadeh static struct ceph_msg *alloc_msg(struct ceph_connection *con, 28823d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 28833d14c5d2SYehuda Sadeh int *skip) 28843d14c5d2SYehuda Sadeh { 28853d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 28863d14c5d2SYehuda Sadeh int type = le16_to_cpu(hdr->type); 28873d14c5d2SYehuda Sadeh int front = le32_to_cpu(hdr->front_len); 28883d14c5d2SYehuda Sadeh 28891c20f2d2SAlex Elder *skip = 0; 28903d14c5d2SYehuda Sadeh switch (type) { 28913d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 2892a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 2893b61c2763SSage Weil return ceph_msg_new(type, front, GFP_NOFS, false); 28943d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 28953d14c5d2SYehuda Sadeh return get_reply(con, hdr, skip); 28963d14c5d2SYehuda Sadeh default: 28973d14c5d2SYehuda Sadeh pr_info("alloc_msg unexpected msg type %d from osd%d\n", type, 28983d14c5d2SYehuda Sadeh osd->o_osd); 28993d14c5d2SYehuda Sadeh *skip = 1; 29003d14c5d2SYehuda Sadeh return NULL; 29013d14c5d2SYehuda Sadeh } 29023d14c5d2SYehuda Sadeh } 29033d14c5d2SYehuda Sadeh 29043d14c5d2SYehuda Sadeh /* 29053d14c5d2SYehuda Sadeh * Wrappers to refcount containing ceph_osd struct 29063d14c5d2SYehuda Sadeh */ 29073d14c5d2SYehuda Sadeh static struct ceph_connection *get_osd_con(struct ceph_connection *con) 29083d14c5d2SYehuda Sadeh { 29093d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 29103d14c5d2SYehuda Sadeh if (get_osd(osd)) 29113d14c5d2SYehuda Sadeh return con; 29123d14c5d2SYehuda Sadeh return NULL; 29133d14c5d2SYehuda Sadeh } 29143d14c5d2SYehuda Sadeh 29153d14c5d2SYehuda Sadeh static void put_osd_con(struct ceph_connection *con) 29163d14c5d2SYehuda Sadeh { 29173d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 29183d14c5d2SYehuda Sadeh put_osd(osd); 29193d14c5d2SYehuda Sadeh } 29203d14c5d2SYehuda Sadeh 29213d14c5d2SYehuda Sadeh /* 29223d14c5d2SYehuda Sadeh * authentication 29233d14c5d2SYehuda Sadeh */ 2924a3530df3SAlex Elder /* 2925a3530df3SAlex Elder * Note: returned pointer is the address of a structure that's 2926a3530df3SAlex Elder * managed separately. Caller must *not* attempt to free it. 2927a3530df3SAlex Elder */ 2928a3530df3SAlex Elder static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con, 29298f43fb53SAlex Elder int *proto, int force_new) 29303d14c5d2SYehuda Sadeh { 29313d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 29323d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 29333d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 293474f1869fSAlex Elder struct ceph_auth_handshake *auth = &o->o_auth; 29353d14c5d2SYehuda Sadeh 293674f1869fSAlex Elder if (force_new && auth->authorizer) { 293727859f97SSage Weil ceph_auth_destroy_authorizer(ac, auth->authorizer); 293874f1869fSAlex Elder auth->authorizer = NULL; 29393d14c5d2SYehuda Sadeh } 294027859f97SSage Weil if (!auth->authorizer) { 294127859f97SSage Weil int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 2942a3530df3SAlex Elder auth); 29433d14c5d2SYehuda Sadeh if (ret) 2944a3530df3SAlex Elder return ERR_PTR(ret); 294527859f97SSage Weil } else { 294627859f97SSage Weil int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 29470bed9b5cSSage Weil auth); 29480bed9b5cSSage Weil if (ret) 29490bed9b5cSSage Weil return ERR_PTR(ret); 29503d14c5d2SYehuda Sadeh } 29513d14c5d2SYehuda Sadeh *proto = ac->protocol; 295274f1869fSAlex Elder 2953a3530df3SAlex Elder return auth; 29543d14c5d2SYehuda Sadeh } 29553d14c5d2SYehuda Sadeh 29563d14c5d2SYehuda Sadeh 29573d14c5d2SYehuda Sadeh static int verify_authorizer_reply(struct ceph_connection *con, int len) 29583d14c5d2SYehuda Sadeh { 29593d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 29603d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 29613d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 29623d14c5d2SYehuda Sadeh 296327859f97SSage Weil return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len); 29643d14c5d2SYehuda Sadeh } 29653d14c5d2SYehuda Sadeh 29663d14c5d2SYehuda Sadeh static int invalidate_authorizer(struct ceph_connection *con) 29673d14c5d2SYehuda Sadeh { 29683d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 29693d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 29703d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 29713d14c5d2SYehuda Sadeh 297227859f97SSage Weil ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD); 29733d14c5d2SYehuda Sadeh return ceph_monc_validate_auth(&osdc->client->monc); 29743d14c5d2SYehuda Sadeh } 29753d14c5d2SYehuda Sadeh 297633d07337SYan, Zheng static int sign_message(struct ceph_connection *con, struct ceph_msg *msg) 297733d07337SYan, Zheng { 297833d07337SYan, Zheng struct ceph_osd *o = con->private; 297933d07337SYan, Zheng struct ceph_auth_handshake *auth = &o->o_auth; 298033d07337SYan, Zheng return ceph_auth_sign_message(auth, msg); 298133d07337SYan, Zheng } 298233d07337SYan, Zheng 298333d07337SYan, Zheng static int check_message_signature(struct ceph_connection *con, struct ceph_msg *msg) 298433d07337SYan, Zheng { 298533d07337SYan, Zheng struct ceph_osd *o = con->private; 298633d07337SYan, Zheng struct ceph_auth_handshake *auth = &o->o_auth; 298733d07337SYan, Zheng return ceph_auth_check_message_signature(auth, msg); 298833d07337SYan, Zheng } 298933d07337SYan, Zheng 29903d14c5d2SYehuda Sadeh static const struct ceph_connection_operations osd_con_ops = { 29913d14c5d2SYehuda Sadeh .get = get_osd_con, 29923d14c5d2SYehuda Sadeh .put = put_osd_con, 29933d14c5d2SYehuda Sadeh .dispatch = dispatch, 29943d14c5d2SYehuda Sadeh .get_authorizer = get_authorizer, 29953d14c5d2SYehuda Sadeh .verify_authorizer_reply = verify_authorizer_reply, 29963d14c5d2SYehuda Sadeh .invalidate_authorizer = invalidate_authorizer, 29973d14c5d2SYehuda Sadeh .alloc_msg = alloc_msg, 299833d07337SYan, Zheng .sign_message = sign_message, 299933d07337SYan, Zheng .check_message_signature = check_message_signature, 30003d14c5d2SYehuda Sadeh .fault = osd_reset, 30013d14c5d2SYehuda Sadeh }; 3002