X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fptlrpc%2Fclient.c;h=bac33017fec3a5f4d8e71b86e059ff20ecb3a0ab;hp=9f6d54a99fcbc5628bf1faac55ab52f54cc4f2ac;hb=3283170eeb05eb701197d981fa773b86f63c8a33;hpb=294e507d34de0d7990ec507334f71ff1d1a7db01 diff --git a/lustre/ptlrpc/client.c b/lustre/ptlrpc/client.c index 9f6d54a..bac3301 100644 --- a/lustre/ptlrpc/client.c +++ b/lustre/ptlrpc/client.c @@ -15,11 +15,7 @@ * * You should have received a copy of the GNU General Public License * version 2 along with this program; If not, see - * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf - * - * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, - * CA 95054 USA or visit www.sun.com if you need additional information or - * have any questions. + * http://www.gnu.org/licenses/gpl-2.0.html * * GPL HEADER END */ @@ -27,7 +23,7 @@ * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2011, 2013, Intel Corporation. + * Copyright (c) 2011, 2016, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -37,11 +33,6 @@ /** Implementation of client-side PortalRPC interfaces */ #define DEBUG_SUBSYSTEM S_RPC -#ifndef __KERNEL__ -#include -#include -#include -#endif #include #include @@ -52,8 +43,26 @@ #include "ptlrpc_internal.h" +const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_pin_ops = { + .add_kiov_frag = ptlrpc_prep_bulk_page_pin, + .release_frags = ptlrpc_release_bulk_page_pin, +}; +EXPORT_SYMBOL(ptlrpc_bulk_kiov_pin_ops); + +const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_nopin_ops = { + .add_kiov_frag = ptlrpc_prep_bulk_page_nopin, + .release_frags = ptlrpc_release_bulk_noop, +}; +EXPORT_SYMBOL(ptlrpc_bulk_kiov_nopin_ops); + +const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kvec_ops = { + .add_iov_frag = ptlrpc_prep_bulk_frag, +}; +EXPORT_SYMBOL(ptlrpc_bulk_kvec_ops); + static int ptlrpc_send_new_req(struct ptlrpc_request *req); static int ptlrpcd_check_work(struct ptlrpc_request *req); +static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async); /** * Initialize passed in client structure \a cl. @@ -70,82 +79,110 @@ EXPORT_SYMBOL(ptlrpc_init_client); /** * Return PortalRPC connection for remore uud \a uuid */ -struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid) +struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid, + lnet_nid_t nid4refnet) { - struct ptlrpc_connection *c; - lnet_nid_t self; - lnet_process_id_t peer; - int err; + struct ptlrpc_connection *c; + lnet_nid_t self; + struct lnet_process_id peer; + int err; /* ptlrpc_uuid_to_peer() initializes its 2nd parameter * before accessing its values. */ /* coverity[uninit_use_in_call] */ - err = ptlrpc_uuid_to_peer(uuid, &peer, &self); - if (err != 0) { - CNETERR("cannot find peer %s!\n", uuid->uuid); - return NULL; - } + peer.nid = nid4refnet; + err = ptlrpc_uuid_to_peer(uuid, &peer, &self); + if (err != 0) { + CNETERR("cannot find peer %s!\n", uuid->uuid); + return NULL; + } - c = ptlrpc_connection_get(peer, self, uuid); - if (c) { - memcpy(c->c_remote_uuid.uuid, - uuid->uuid, sizeof(c->c_remote_uuid.uuid)); - } + c = ptlrpc_connection_get(peer, self, uuid); + if (c) { + memcpy(c->c_remote_uuid.uuid, + uuid->uuid, sizeof(c->c_remote_uuid.uuid)); + } - CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c); + CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c); - return c; + return c; } -EXPORT_SYMBOL(ptlrpc_uuid_to_connection); /** * Allocate and initialize new bulk descriptor on the sender. * Returns pointer to the descriptor or NULL on error. */ -struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned npages, unsigned max_brw, - unsigned type, unsigned portal) +struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned nfrags, unsigned max_brw, + enum ptlrpc_bulk_op_type type, + unsigned portal, + const struct ptlrpc_bulk_frag_ops *ops) { struct ptlrpc_bulk_desc *desc; int i; - OBD_ALLOC(desc, offsetof(struct ptlrpc_bulk_desc, bd_iov[npages])); - if (!desc) + /* ensure that only one of KIOV or IOVEC is set but not both */ + LASSERT((ptlrpc_is_bulk_desc_kiov(type) && + ops->add_kiov_frag != NULL) || + (ptlrpc_is_bulk_desc_kvec(type) && + ops->add_iov_frag != NULL)); + + OBD_ALLOC_PTR(desc); + if (desc == NULL) return NULL; + if (type & PTLRPC_BULK_BUF_KIOV) { + OBD_ALLOC_LARGE(GET_KIOV(desc), + nfrags * sizeof(*GET_KIOV(desc))); + if (GET_KIOV(desc) == NULL) + goto out; + } else { + OBD_ALLOC_LARGE(GET_KVEC(desc), + nfrags * sizeof(*GET_KVEC(desc))); + if (GET_KVEC(desc) == NULL) + goto out; + } spin_lock_init(&desc->bd_lock); init_waitqueue_head(&desc->bd_waitq); - desc->bd_max_iov = npages; + desc->bd_max_iov = nfrags; desc->bd_iov_count = 0; desc->bd_portal = portal; desc->bd_type = type; desc->bd_md_count = 0; + desc->bd_frag_ops = (struct ptlrpc_bulk_frag_ops *) ops; LASSERT(max_brw > 0); desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT); /* PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this * node. Negotiated ocd_brw_size will always be <= this number. */ for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++) - LNetInvalidateHandle(&desc->bd_mds[i]); + LNetInvalidateMDHandle(&desc->bd_mds[i]); return desc; +out: + OBD_FREE_PTR(desc); + return NULL; } /** * Prepare bulk descriptor for specified outgoing request \a req that - * can fit \a npages * pages. \a type is bulk type. \a portal is where + * can fit \a nfrags * pages. \a type is bulk type. \a portal is where * the bulk to be sent. Used on client-side. * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on * error. */ struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req, - unsigned npages, unsigned max_brw, - unsigned type, unsigned portal) + unsigned nfrags, unsigned max_brw, + unsigned int type, + unsigned portal, + const struct ptlrpc_bulk_frag_ops + *ops) { struct obd_import *imp = req->rq_import; struct ptlrpc_bulk_desc *desc; ENTRY; - LASSERT(type == BULK_PUT_SINK || type == BULK_GET_SOURCE); - desc = ptlrpc_new_bulk(npages, max_brw, type, portal); + LASSERT(ptlrpc_is_bulk_op_passive(type)); + + desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops); if (desc == NULL) RETURN(NULL); @@ -163,60 +200,89 @@ struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req, } EXPORT_SYMBOL(ptlrpc_prep_bulk_imp); -/* - * Add a page \a page to the bulk descriptor \a desc. - * Data to transfer in the page starts at offset \a pageoffset and - * amount of data to transfer from the page is \a len - */ void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc, - struct page *page, int pageoffset, int len, int pin) + struct page *page, int pageoffset, int len, + int pin) { + lnet_kiov_t *kiov; + LASSERT(desc->bd_iov_count < desc->bd_max_iov); LASSERT(page != NULL); LASSERT(pageoffset >= 0); LASSERT(len > 0); - LASSERT(pageoffset + len <= PAGE_CACHE_SIZE); + LASSERT(pageoffset + len <= PAGE_SIZE); + LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type)); + + kiov = &BD_GET_KIOV(desc, desc->bd_iov_count); desc->bd_nob += len; if (pin) - page_cache_get(page); + get_page(page); + + kiov->kiov_page = page; + kiov->kiov_offset = pageoffset; + kiov->kiov_len = len; - ptlrpc_add_bulk_page(desc, page, pageoffset, len); + desc->bd_iov_count++; } EXPORT_SYMBOL(__ptlrpc_prep_bulk_page); -/** - * Uninitialize and free bulk descriptor \a desc. - * Works on bulk descriptors both from server and client side. - */ -void __ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc, int unpin) +int ptlrpc_prep_bulk_frag(struct ptlrpc_bulk_desc *desc, + void *frag, int len) +{ + struct kvec *iovec; + ENTRY; + + LASSERT(desc->bd_iov_count < desc->bd_max_iov); + LASSERT(frag != NULL); + LASSERT(len > 0); + LASSERT(ptlrpc_is_bulk_desc_kvec(desc->bd_type)); + + iovec = &BD_GET_KVEC(desc, desc->bd_iov_count); + + desc->bd_nob += len; + + iovec->iov_base = frag; + iovec->iov_len = len; + + desc->bd_iov_count++; + + RETURN(desc->bd_nob); +} +EXPORT_SYMBOL(ptlrpc_prep_bulk_frag); + +void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc) { - int i; ENTRY; LASSERT(desc != NULL); LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */ LASSERT(desc->bd_md_count == 0); /* network hands off */ LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL)); + LASSERT(desc->bd_frag_ops != NULL); - sptlrpc_enc_pool_put_pages(desc); + if (ptlrpc_is_bulk_desc_kiov(desc->bd_type)) + sptlrpc_enc_pool_put_pages(desc); if (desc->bd_export) class_export_put(desc->bd_export); else class_import_put(desc->bd_import); - if (unpin) { - for (i = 0; i < desc->bd_iov_count ; i++) - page_cache_release(desc->bd_iov[i].kiov_page); - } + if (desc->bd_frag_ops->release_frags != NULL) + desc->bd_frag_ops->release_frags(desc); - OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc, - bd_iov[desc->bd_max_iov])); + if (ptlrpc_is_bulk_desc_kiov(desc->bd_type)) + OBD_FREE_LARGE(GET_KIOV(desc), + desc->bd_max_iov * sizeof(*GET_KIOV(desc))); + else + OBD_FREE_LARGE(GET_KVEC(desc), + desc->bd_max_iov * sizeof(*GET_KVEC(desc))); + OBD_FREE_PTR(desc); EXIT; } -EXPORT_SYMBOL(__ptlrpc_free_bulk); +EXPORT_SYMBOL(ptlrpc_free_bulk); /** * Set server timelimit for this req, i.e. how long are we willing to wait @@ -286,22 +352,34 @@ int ptlrpc_at_get_net_latency(struct ptlrpc_request *req) } /* Adjust expected network latency */ -static void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req, - unsigned int service_time) +void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req, + unsigned int service_time) { unsigned int nl, oldnl; struct imp_at *at; - time_t now = cfs_time_current_sec(); + time64_t now = ktime_get_real_seconds(); LASSERT(req->rq_import); - at = &req->rq_import->imp_at; + + if (service_time > now - req->rq_sent + 3) { + /* bz16408, however, this can also happen if early reply + * is lost and client RPC is expired and resent, early reply + * or reply of original RPC can still be fit in reply buffer + * of resent RPC, now client is measuring time from the + * resent time, but server sent back service time of original + * RPC. + */ + CDEBUG((lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ? + D_ADAPTTO : D_WARNING, + "Reported service time %u > total measured time %lld\n", + service_time, now - req->rq_sent); + return; + } /* Network latency is total time less server processing time */ - nl = max_t(int, now - req->rq_sent - service_time, 0) +1/*st rounding*/; - if (service_time > now - req->rq_sent + 3 /* bz16408 */) - CWARN("Reported service time %u > total measured time " - CFS_DURATION_T"\n", service_time, - cfs_time_sub(now, req->rq_sent)); + nl = max_t(int, now - req->rq_sent - + service_time, 0) + 1; /* st rounding */ + at = &req->rq_import->imp_at; oldnl = at_measured(&at->iat_net_latency, nl); if (oldnl != 0) @@ -338,59 +416,64 @@ static int unpack_reply(struct ptlrpc_request *req) * If anything goes wrong just ignore it - same as if it never happened */ static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req) +__must_hold(&req->rq_lock) { - struct ptlrpc_request *early_req; - time_t olddl; - int rc; - ENTRY; + struct ptlrpc_request *early_req; + time64_t olddl; + int rc; + ENTRY; req->rq_early = 0; spin_unlock(&req->rq_lock); rc = sptlrpc_cli_unwrap_early_reply(req, &early_req); if (rc) { spin_lock(&req->rq_lock); - RETURN(rc); - } - - rc = unpack_reply(early_req); - if (rc == 0) { - /* Expecting to increase the service time estimate here */ - ptlrpc_at_adj_service(req, - lustre_msg_get_timeout(early_req->rq_repmsg)); - ptlrpc_at_adj_net_latency(req, - lustre_msg_get_service_time(early_req->rq_repmsg)); - } - - sptlrpc_cli_finish_early_reply(early_req); + RETURN(rc); + } + rc = unpack_reply(early_req); if (rc != 0) { + sptlrpc_cli_finish_early_reply(early_req); spin_lock(&req->rq_lock); RETURN(rc); } - /* Adjust the local timeout for this req */ - ptlrpc_at_set_req_timeout(req); + /* Use new timeout value just to adjust the local value for this + * request, don't include it into at_history. It is unclear yet why + * service time increased and should it be counted or skipped, e.g. + * that can be recovery case or some error or server, the real reply + * will add all new data if it is worth to add. */ + req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg); + lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout); + + /* Network latency can be adjusted, it is pure network delays */ + ptlrpc_at_adj_net_latency(req, + lustre_msg_get_service_time(early_req->rq_repmsg)); + + sptlrpc_cli_finish_early_reply(early_req); spin_lock(&req->rq_lock); olddl = req->rq_deadline; - /* server assumes it now has rq_timeout from when it sent the - * early reply, so client should give it at least that long. */ - req->rq_deadline = cfs_time_current_sec() + req->rq_timeout + + /* server assumes it now has rq_timeout from when the request + * arrived, so the client should give it at least that long. + * since we don't know the arrival time we'll use the original + * sent time */ + req->rq_deadline = req->rq_sent + req->rq_timeout + ptlrpc_at_get_net_latency(req); DEBUG_REQ(D_ADAPTTO, req, - "Early reply #%d, new deadline in "CFS_DURATION_T"s " - "("CFS_DURATION_T"s)", req->rq_early_count, - cfs_time_sub(req->rq_deadline, cfs_time_current_sec()), - cfs_time_sub(req->rq_deadline, olddl)); + "Early reply #%d, new deadline in %llds (%llds)", + req->rq_early_count, + req->rq_deadline - ktime_get_real_seconds(), + req->rq_deadline - olddl); RETURN(rc); } -struct kmem_cache *request_cache; +static struct kmem_cache *request_cache; -int ptlrpc_request_cache_init() +int ptlrpc_request_cache_init(void) { request_cache = kmem_cache_create("ptlrpc_cache", sizeof(struct ptlrpc_request), @@ -398,12 +481,12 @@ int ptlrpc_request_cache_init() return request_cache == NULL ? -ENOMEM : 0; } -void ptlrpc_request_cache_fini() +void ptlrpc_request_cache_fini(void) { kmem_cache_destroy(request_cache); } -struct ptlrpc_request *ptlrpc_request_cache_alloc(int flags) +struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags) { struct ptlrpc_request *req; @@ -422,15 +505,15 @@ void ptlrpc_request_cache_free(struct ptlrpc_request *req) */ void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool) { - cfs_list_t *l, *tmp; + struct list_head *l, *tmp; struct ptlrpc_request *req; LASSERT(pool != NULL); spin_lock(&pool->prp_lock); - cfs_list_for_each_safe(l, tmp, &pool->prp_req_list) { - req = cfs_list_entry(l, struct ptlrpc_request, rq_list); - cfs_list_del(&req->rq_list); + list_for_each_safe(l, tmp, &pool->prp_req_list) { + req = list_entry(l, struct ptlrpc_request, rq_list); + list_del(&req->rq_list); LASSERT(req->rq_reqbuf); LASSERT(req->rq_reqbuf_len == pool->prp_rq_size); OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size); @@ -444,7 +527,7 @@ EXPORT_SYMBOL(ptlrpc_free_rq_pool); /** * Allocates, initializes and adds \a num_rq requests to the pool \a pool */ -void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq) +int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq) { int i; int size = 1; @@ -452,7 +535,7 @@ void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq) while (size < pool->prp_rq_size) size <<= 1; - LASSERTF(cfs_list_empty(&pool->prp_req_list) || + LASSERTF(list_empty(&pool->prp_req_list) || size == pool->prp_rq_size, "Trying to change pool size with nonempty pool " "from %d to %d bytes\n", pool->prp_rq_size, size); @@ -466,20 +549,20 @@ void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq) spin_unlock(&pool->prp_lock); req = ptlrpc_request_cache_alloc(GFP_NOFS); if (!req) - return; + return i; OBD_ALLOC_LARGE(msg, size); if (!msg) { ptlrpc_request_cache_free(req); - return; - } - req->rq_reqbuf = msg; - req->rq_reqbuf_len = size; - req->rq_pool = pool; + return i; + } + req->rq_reqbuf = msg; + req->rq_reqbuf_len = size; + req->rq_pool = pool; spin_lock(&pool->prp_lock); - cfs_list_add_tail(&req->rq_list, &pool->prp_req_list); + list_add_tail(&req->rq_list, &pool->prp_req_list); } spin_unlock(&pool->prp_lock); - return; + return num_rq; } EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool); @@ -493,30 +576,25 @@ EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool); */ struct ptlrpc_request_pool * ptlrpc_init_rq_pool(int num_rq, int msgsize, - void (*populate_pool)(struct ptlrpc_request_pool *, int)) + int (*populate_pool)(struct ptlrpc_request_pool *, int)) { - struct ptlrpc_request_pool *pool; + struct ptlrpc_request_pool *pool; - OBD_ALLOC(pool, sizeof (struct ptlrpc_request_pool)); - if (!pool) - return NULL; + OBD_ALLOC(pool, sizeof(struct ptlrpc_request_pool)); + if (!pool) + return NULL; - /* Request next power of two for the allocation, because internally - kernel would do exactly this */ + /* Request next power of two for the allocation, because internally + kernel would do exactly this */ spin_lock_init(&pool->prp_lock); - CFS_INIT_LIST_HEAD(&pool->prp_req_list); - pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD; - pool->prp_populate = populate_pool; + INIT_LIST_HEAD(&pool->prp_req_list); + pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD; + pool->prp_populate = populate_pool; - populate_pool(pool, num_rq); + populate_pool(pool, num_rq); - if (cfs_list_empty(&pool->prp_req_list)) { - /* have not allocated a single request for the pool */ - OBD_FREE(pool, sizeof (struct ptlrpc_request_pool)); - pool = NULL; - } - return pool; + return pool; } EXPORT_SYMBOL(ptlrpc_init_rq_pool); @@ -538,14 +616,14 @@ ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool) * in writeout path, where this matters, this is safe to do, because * nothing is lost in this case, and when some in-flight requests * complete, this code will be called again. */ - if (unlikely(cfs_list_empty(&pool->prp_req_list))) { + if (unlikely(list_empty(&pool->prp_req_list))) { spin_unlock(&pool->prp_lock); return NULL; } - request = cfs_list_entry(pool->prp_req_list.next, struct ptlrpc_request, - rq_list); - cfs_list_del_init(&request->rq_list); + request = list_entry(pool->prp_req_list.next, struct ptlrpc_request, + rq_list); + list_del_init(&request->rq_list); spin_unlock(&pool->prp_lock); LASSERT(request->rq_reqbuf); @@ -568,91 +646,141 @@ static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request) struct ptlrpc_request_pool *pool = request->rq_pool; spin_lock(&pool->prp_lock); - LASSERT(cfs_list_empty(&request->rq_list)); + LASSERT(list_empty(&request->rq_list)); LASSERT(!request->rq_receiving_reply); - cfs_list_add_tail(&request->rq_list, &pool->prp_req_list); + list_add_tail(&request->rq_list, &pool->prp_req_list); spin_unlock(&pool->prp_lock); } -static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request, - __u32 version, int opcode, - int count, __u32 *lengths, char **bufs, - struct ptlrpc_cli_ctx *ctx) +void ptlrpc_add_unreplied(struct ptlrpc_request *req) { - struct obd_import *imp = request->rq_import; - int rc; - ENTRY; + struct obd_import *imp = req->rq_import; + struct list_head *tmp; + struct ptlrpc_request *iter; - if (unlikely(ctx)) - request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx); - else { - rc = sptlrpc_req_get_ctx(request); - if (rc) - GOTO(out_free, rc); - } + assert_spin_locked(&imp->imp_lock); + LASSERT(list_empty(&req->rq_unreplied_list)); - sptlrpc_req_set_flavor(request, opcode); + /* unreplied list is sorted by xid in ascending order */ + list_for_each_prev(tmp, &imp->imp_unreplied_list) { + iter = list_entry(tmp, struct ptlrpc_request, + rq_unreplied_list); - rc = lustre_pack_request(request, imp->imp_msg_magic, count, - lengths, bufs); - if (rc) { - LASSERT(!request->rq_pool); - GOTO(out_ctx, rc); - } + LASSERT(req->rq_xid != iter->rq_xid); + if (req->rq_xid < iter->rq_xid) + continue; + list_add(&req->rq_unreplied_list, &iter->rq_unreplied_list); + return; + } + list_add(&req->rq_unreplied_list, &imp->imp_unreplied_list); +} + +void ptlrpc_assign_next_xid_nolock(struct ptlrpc_request *req) +{ + req->rq_xid = ptlrpc_next_xid(); + ptlrpc_add_unreplied(req); +} + +static inline void ptlrpc_assign_next_xid(struct ptlrpc_request *req) +{ + spin_lock(&req->rq_import->imp_lock); + ptlrpc_assign_next_xid_nolock(req); + spin_unlock(&req->rq_import->imp_lock); +} + +int ptlrpc_request_bufs_pack(struct ptlrpc_request *request, + __u32 version, int opcode, char **bufs, + struct ptlrpc_cli_ctx *ctx) +{ + int count; + struct obd_import *imp; + __u32 *lengths; + int rc; + + ENTRY; + + count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT); + imp = request->rq_import; + lengths = request->rq_pill.rc_area[RCL_CLIENT]; + + if (ctx != NULL) { + request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx); + } else { + rc = sptlrpc_req_get_ctx(request); + if (rc) + GOTO(out_free, rc); + } + sptlrpc_req_set_flavor(request, opcode); - lustre_msg_add_version(request->rq_reqmsg, version); - request->rq_send_state = LUSTRE_IMP_FULL; - request->rq_type = PTL_RPC_MSG_REQUEST; - request->rq_export = NULL; + rc = lustre_pack_request(request, imp->imp_msg_magic, count, + lengths, bufs); + if (rc) + GOTO(out_ctx, rc); - request->rq_req_cbid.cbid_fn = request_out_callback; - request->rq_req_cbid.cbid_arg = request; + lustre_msg_add_version(request->rq_reqmsg, version); + request->rq_send_state = LUSTRE_IMP_FULL; + request->rq_type = PTL_RPC_MSG_REQUEST; - request->rq_reply_cbid.cbid_fn = reply_in_callback; - request->rq_reply_cbid.cbid_arg = request; + request->rq_req_cbid.cbid_fn = request_out_callback; + request->rq_req_cbid.cbid_arg = request; - request->rq_reply_deadline = 0; - request->rq_phase = RQ_PHASE_NEW; - request->rq_next_phase = RQ_PHASE_UNDEFINED; + request->rq_reply_cbid.cbid_fn = reply_in_callback; + request->rq_reply_cbid.cbid_arg = request; - request->rq_request_portal = imp->imp_client->cli_request_portal; - request->rq_reply_portal = imp->imp_client->cli_reply_portal; + request->rq_reply_deadline = 0; + request->rq_bulk_deadline = 0; + request->rq_req_deadline = 0; + request->rq_phase = RQ_PHASE_NEW; + request->rq_next_phase = RQ_PHASE_UNDEFINED; - ptlrpc_at_set_req_timeout(request); + request->rq_request_portal = imp->imp_client->cli_request_portal; + request->rq_reply_portal = imp->imp_client->cli_reply_portal; - spin_lock_init(&request->rq_lock); - CFS_INIT_LIST_HEAD(&request->rq_list); - CFS_INIT_LIST_HEAD(&request->rq_timed_list); - CFS_INIT_LIST_HEAD(&request->rq_replay_list); - CFS_INIT_LIST_HEAD(&request->rq_ctx_chain); - CFS_INIT_LIST_HEAD(&request->rq_set_chain); - CFS_INIT_LIST_HEAD(&request->rq_history_list); - CFS_INIT_LIST_HEAD(&request->rq_exp_list); - init_waitqueue_head(&request->rq_reply_waitq); - init_waitqueue_head(&request->rq_set_waitq); - request->rq_xid = ptlrpc_next_xid(); - atomic_set(&request->rq_refcount, 1); + ptlrpc_at_set_req_timeout(request); lustre_msg_set_opc(request->rq_reqmsg, opcode); + ptlrpc_assign_next_xid(request); + + /* Let's setup deadline for req/reply/bulk unlink for opcode. */ + if (cfs_fail_val == opcode) { + time64_t *fail_t = NULL, *fail2_t = NULL; + + if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) + fail_t = &request->rq_bulk_deadline; + else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) + fail_t = &request->rq_reply_deadline; + else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK)) + fail_t = &request->rq_req_deadline; + else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BOTH_UNLINK)) { + fail_t = &request->rq_reply_deadline; + fail2_t = &request->rq_bulk_deadline; + } + + if (fail_t) { + *fail_t = ktime_get_real_seconds() + LONG_UNLINK; + + if (fail2_t) + *fail2_t = ktime_get_real_seconds() + + LONG_UNLINK; + + /* + * The RPC is infected, let the test to change the + * fail_loc + */ + msleep(4 * MSEC_PER_SEC); + } + } RETURN(0); + out_ctx: + LASSERT(!request->rq_pool); sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1); out_free: class_import_put(imp); - return rc; -} -int ptlrpc_request_bufs_pack(struct ptlrpc_request *request, - __u32 version, int opcode, char **bufs, - struct ptlrpc_cli_ctx *ctx) -{ - int count; + return rc; - count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT); - return __ptlrpc_request_bufs_pack(request, version, opcode, count, - request->rq_pill.rc_area[RCL_CLIENT], - bufs, ctx); } EXPORT_SYMBOL(ptlrpc_request_bufs_pack); @@ -702,16 +830,17 @@ struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp, { struct ptlrpc_request *request = NULL; - if (pool) - request = ptlrpc_prep_req_from_pool(pool); + request = ptlrpc_request_cache_alloc(GFP_NOFS); - if (!request) - request = ptlrpc_request_cache_alloc(GFP_NOFS); + if (!request && pool) + request = ptlrpc_prep_req_from_pool(pool); if (request) { + ptlrpc_cli_req_init(request); + LASSERTF((unsigned long)imp > 0x1000, "%p", imp); LASSERT(imp != LP_POISON); - LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p", + LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n", imp->imp_client); LASSERT(imp->imp_client != LP_POISON); @@ -807,67 +936,27 @@ struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp, EXPORT_SYMBOL(ptlrpc_request_alloc_pack); /** - * Prepare request (fetched from pool \a poolif not NULL) on import \a imp - * for operation \a opcode. Request would contain \a count buffers. - * Sizes of buffers are described in array \a lengths and buffers themselves - * are provided by a pointer \a bufs. - * Returns prepared request structure pointer or NULL on error. - */ -struct ptlrpc_request * -ptlrpc_prep_req_pool(struct obd_import *imp, - __u32 version, int opcode, - int count, __u32 *lengths, char **bufs, - struct ptlrpc_request_pool *pool) -{ - struct ptlrpc_request *request; - int rc; - - request = __ptlrpc_request_alloc(imp, pool); - if (!request) - return NULL; - - rc = __ptlrpc_request_bufs_pack(request, version, opcode, count, - lengths, bufs, NULL); - if (rc) { - ptlrpc_request_free(request); - request = NULL; - } - return request; -} -EXPORT_SYMBOL(ptlrpc_prep_req_pool); - -/** - * Same as ptlrpc_prep_req_pool, but without pool - */ -struct ptlrpc_request * -ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count, - __u32 *lengths, char **bufs) -{ - return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs, - NULL); -} -EXPORT_SYMBOL(ptlrpc_prep_req); - -/** - * Allocate and initialize new request set structure. + * Allocate and initialize new request set structure on the current CPT. * Returns a pointer to the newly allocated set structure or NULL on error. */ struct ptlrpc_request_set *ptlrpc_prep_set(void) { - struct ptlrpc_request_set *set; + struct ptlrpc_request_set *set; + int cpt; ENTRY; - OBD_ALLOC(set, sizeof *set); + cpt = cfs_cpt_current(cfs_cpt_table, 0); + OBD_CPT_ALLOC(set, cfs_cpt_table, cpt, sizeof *set); if (!set) RETURN(NULL); atomic_set(&set->set_refcount, 1); - CFS_INIT_LIST_HEAD(&set->set_requests); + INIT_LIST_HEAD(&set->set_requests); init_waitqueue_head(&set->set_waitq); atomic_set(&set->set_new_count, 0); atomic_set(&set->set_remaining, 0); spin_lock_init(&set->set_new_req_lock); - CFS_INIT_LIST_HEAD(&set->set_new_requests); - CFS_INIT_LIST_HEAD(&set->set_cblist); + INIT_LIST_HEAD(&set->set_new_requests); + INIT_LIST_HEAD(&set->set_cblist); set->set_max_inflight = UINT_MAX; set->set_producer = NULL; set->set_producer_arg = NULL; @@ -901,7 +990,6 @@ struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func, RETURN(set); } -EXPORT_SYMBOL(ptlrpc_prep_fcset); /** * Wind down and free request set structure previously allocated with @@ -913,33 +1001,33 @@ EXPORT_SYMBOL(ptlrpc_prep_fcset); */ void ptlrpc_set_destroy(struct ptlrpc_request_set *set) { - cfs_list_t *tmp; - cfs_list_t *next; - int expected_phase; - int n = 0; - ENTRY; + struct list_head *tmp; + struct list_head *next; + int expected_phase; + int n = 0; + ENTRY; - /* Requests on the set should either all be completed, or all be new */ + /* Requests on the set should either all be completed, or all be new */ expected_phase = (atomic_read(&set->set_remaining) == 0) ? - RQ_PHASE_COMPLETE : RQ_PHASE_NEW; - cfs_list_for_each (tmp, &set->set_requests) { - struct ptlrpc_request *req = - cfs_list_entry(tmp, struct ptlrpc_request, - rq_set_chain); - - LASSERT(req->rq_phase == expected_phase); - n++; - } + RQ_PHASE_COMPLETE : RQ_PHASE_NEW; + list_for_each(tmp, &set->set_requests) { + struct ptlrpc_request *req = + list_entry(tmp, struct ptlrpc_request, + rq_set_chain); + + LASSERT(req->rq_phase == expected_phase); + n++; + } LASSERTF(atomic_read(&set->set_remaining) == 0 || atomic_read(&set->set_remaining) == n, "%d / %d\n", atomic_read(&set->set_remaining), n); - cfs_list_for_each_safe(tmp, next, &set->set_requests) { - struct ptlrpc_request *req = - cfs_list_entry(tmp, struct ptlrpc_request, - rq_set_chain); - cfs_list_del_init(&req->rq_set_chain); + list_for_each_safe(tmp, next, &set->set_requests) { + struct ptlrpc_request *req = + list_entry(tmp, struct ptlrpc_request, + rq_set_chain); + list_del_init(&req->rq_set_chain); LASSERT(req->rq_phase == expected_phase); @@ -971,19 +1059,18 @@ EXPORT_SYMBOL(ptlrpc_set_destroy); int ptlrpc_set_add_cb(struct ptlrpc_request_set *set, set_interpreter_func fn, void *data) { - struct ptlrpc_set_cbdata *cbdata; + struct ptlrpc_set_cbdata *cbdata; - OBD_ALLOC_PTR(cbdata); - if (cbdata == NULL) - RETURN(-ENOMEM); + OBD_ALLOC_PTR(cbdata); + if (cbdata == NULL) + RETURN(-ENOMEM); - cbdata->psc_interpret = fn; - cbdata->psc_data = data; - cfs_list_add_tail(&cbdata->psc_item, &set->set_cblist); + cbdata->psc_interpret = fn; + cbdata->psc_data = data; + list_add_tail(&cbdata->psc_item, &set->set_cblist); - RETURN(0); + RETURN(0); } -EXPORT_SYMBOL(ptlrpc_set_add_cb); /** * Add a new request to the general purpose request set. @@ -992,10 +1079,13 @@ EXPORT_SYMBOL(ptlrpc_set_add_cb); void ptlrpc_set_add_req(struct ptlrpc_request_set *set, struct ptlrpc_request *req) { - LASSERT(cfs_list_empty(&req->rq_set_chain)); + LASSERT(list_empty(&req->rq_set_chain)); + + if (req->rq_allow_intr) + set->set_allow_intr = 1; /* The set takes over the caller's request reference */ - cfs_list_add_tail(&req->rq_set_chain, &set->set_requests); + list_add_tail(&req->rq_set_chain, &set->set_requests); req->rq_set = set; atomic_inc(&set->set_remaining); req->rq_queued_time = cfs_time_current(); @@ -1030,7 +1120,7 @@ void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc, */ req->rq_set = set; req->rq_queued_time = cfs_time_current(); - cfs_list_add_tail(&req->rq_set_chain, &set->set_new_requests); + list_add_tail(&req->rq_set_chain, &set->set_new_requests); count = atomic_inc_return(&set->set_new_count); spin_unlock(&set->set_new_req_lock); @@ -1045,7 +1135,6 @@ void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc, wake_up(&pc->pc_partners[i]->pc_set->set_waitq); } } -EXPORT_SYMBOL(ptlrpc_set_add_new_req); /** * Based on the current state of the import, determine if the request @@ -1077,9 +1166,9 @@ static int ptlrpc_import_delay_req(struct obd_import *imp, D_HA : D_ERROR, req, "IMP_CLOSED "); *status = -EIO; } else if (ptlrpc_send_limit_expired(req)) { - /* probably doesn't need to be a D_ERROR after initial testing */ - DEBUG_REQ(D_ERROR, req, "send limit expired "); - *status = -EIO; + /* probably doesn't need to be a D_ERROR after initial testing*/ + DEBUG_REQ(D_HA, req, "send limit expired "); + *status = -ETIMEDOUT; } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING && imp->imp_state == LUSTRE_IMP_CONNECTING) { /* allow CONNECT even if import is invalid */ ; @@ -1099,7 +1188,7 @@ static int ptlrpc_import_delay_req(struct obd_import *imp, if (atomic_read(&imp->imp_inval_count) != 0) { DEBUG_REQ(D_ERROR, req, "invalidate in flight"); *status = -EIO; - } else if (imp->imp_dlm_fake || req->rq_no_delay) { + } else if (req->rq_no_delay) { *status = -EWOULDBLOCK; } else if (req->rq_allow_replay && (imp->imp_state == LUSTRE_IMP_REPLAY || @@ -1116,35 +1205,40 @@ static int ptlrpc_import_delay_req(struct obd_import *imp, } /** - * Decide if the eror message regarding provided request \a req - * should be printed to the console or not. - * Makes it's decision on request status and other properties. - * Returns 1 to print error on the system console or 0 if not. + * Decide if the error message should be printed to the console or not. + * Makes its decision based on request type, status, and failure frequency. + * + * \param[in] req request that failed and may need a console message + * + * \retval false if no message should be printed + * \retval true if console message should be printed */ -static int ptlrpc_console_allow(struct ptlrpc_request *req) +static bool ptlrpc_console_allow(struct ptlrpc_request *req) { - __u32 opc; - int err; + __u32 opc; - LASSERT(req->rq_reqmsg != NULL); - opc = lustre_msg_get_opc(req->rq_reqmsg); + LASSERT(req->rq_reqmsg != NULL); + opc = lustre_msg_get_opc(req->rq_reqmsg); - /* Suppress particular reconnect errors which are to be expected. No - * errors are suppressed for the initial connection on an import */ - if ((lustre_handle_is_used(&req->rq_import->imp_remote_handle)) && - (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT)) { + /* Suppress particular reconnect errors which are to be expected. */ + if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) { + int err; - /* Suppress timed out reconnect requests */ - if (req->rq_timedout) - return 0; + /* Suppress timed out reconnect requests */ + if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) || + req->rq_timedout) + return false; - /* Suppress unavailable/again reconnect requests */ - err = lustre_msg_get_status(req->rq_repmsg); - if (err == -ENODEV || err == -EAGAIN) - return 0; - } + /* Suppress most unavailable/again reconnect requests, but + * print occasionally so it is clear client is trying to + * connect to a server where no target is running. */ + err = lustre_msg_get_status(req->rq_repmsg); + if ((err == -ENODEV || err == -EAGAIN) && + req->rq_import->imp_conn_cnt % 30 != 20) + return false; + } - return 1; + return true; } /** @@ -1159,14 +1253,17 @@ static int ptlrpc_check_status(struct ptlrpc_request *req) err = lustre_msg_get_status(req->rq_repmsg); if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) { struct obd_import *imp = req->rq_import; + lnet_nid_t nid = imp->imp_connection->c_peer.nid; __u32 opc = lustre_msg_get_opc(req->rq_reqmsg); - if (ptlrpc_console_allow(req)) - LCONSOLE_ERROR_MSG(0x011, "%s: Communicating with %s," - " operation %s failed with %d.\n", - imp->imp_obd->obd_name, - libcfs_nid2str( - imp->imp_connection->c_peer.nid), - ll_opcode2str(opc), err); + + /* -EAGAIN is normal when using POSIX flocks */ + if (ptlrpc_console_allow(req) && + !(opc == LDLM_ENQUEUE && err == -EAGAIN)) + LCONSOLE_ERROR_MSG(0x11, "%s: operation %s to node %s " + "failed: rc = %d\n", + imp->imp_obd->obd_name, + ll_opcode2str(opc), + libcfs_nid2str(nid), err); RETURN(err < 0 ? err : -EINVAL); } @@ -1197,12 +1294,30 @@ static void ptlrpc_save_versions(struct ptlrpc_request *req) LASSERT(versions); lustre_msg_set_versions(reqmsg, versions); - CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n", + CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n", versions[0], versions[1]); EXIT; } +__u64 ptlrpc_known_replied_xid(struct obd_import *imp) +{ + struct ptlrpc_request *req; + + assert_spin_locked(&imp->imp_lock); + if (list_empty(&imp->imp_unreplied_list)) + return 0; + + req = list_entry(imp->imp_unreplied_list.next, struct ptlrpc_request, + rq_unreplied_list); + LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid); + + if (imp->imp_known_replied_xid < req->rq_xid - 1) + imp->imp_known_replied_xid = req->rq_xid - 1; + + return req->rq_xid - 1; +} + /** * Callback function called when client receives RPC reply for \a req. * Returns 0 on success or error code. @@ -1212,18 +1327,19 @@ static void ptlrpc_save_versions(struct ptlrpc_request *req) */ static int after_reply(struct ptlrpc_request *req) { - struct obd_import *imp = req->rq_import; - struct obd_device *obd = req->rq_import->imp_obd; - int rc; - struct timeval work_start; - long timediff; - ENTRY; + struct obd_import *imp = req->rq_import; + struct obd_device *obd = req->rq_import->imp_obd; + ktime_t work_start; + u64 committed; + s64 timediff; + int rc; + ENTRY; LASSERT(obd != NULL); /* repbuf must be unlinked */ - LASSERT(!req->rq_receiving_reply && !req->rq_must_unlink); + LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked); - if (req->rq_reply_truncate) { + if (req->rq_reply_truncated) { if (ptlrpc_no_resend(req)) { DEBUG_REQ(D_ERROR, req, "reply buffer overflow," " expected: %d, actual size: %d", @@ -1244,6 +1360,9 @@ static int after_reply(struct ptlrpc_request *req) RETURN(0); } + work_start = ktime_get_real(); + timediff = ktime_us_delta(req->rq_sent_ns, work_start); + /* * NB Until this point, the whole of the incoming message, * including buflens, status etc is in the sender's byte order. @@ -1267,21 +1386,14 @@ static int after_reply(struct ptlrpc_request *req) /* retry indefinitely on EINPROGRESS */ if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS && ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) { - time_t now = cfs_time_current_sec(); + time64_t now = ktime_get_real_seconds(); DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS"); + spin_lock(&req->rq_lock); req->rq_resend = 1; + spin_unlock(&req->rq_lock); req->rq_nr_resend++; - /* allocate new xid to avoid reply reconstruction */ - if (!req->rq_bulk) { - /* new xid is already allocated for bulk in - * ptlrpc_check_set() */ - req->rq_xid = ptlrpc_next_xid(); - DEBUG_REQ(D_RPCTRACE, req, "Allocating new xid for " - "resend on EINPROGRESS"); - } - /* Readjust the timeout for current conditions */ ptlrpc_at_set_req_timeout(req); /* delay resend to give a chance to the server to get ready. @@ -1293,11 +1405,14 @@ static int after_reply(struct ptlrpc_request *req) else req->rq_sent = now + req->rq_nr_resend; + /* Resend for EINPROGRESS will use a new XID */ + spin_lock(&imp->imp_lock); + list_del_init(&req->rq_unreplied_list); + spin_unlock(&imp->imp_lock); + RETURN(0); } - do_gettimeofday(&work_start); - timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL); if (obd->obd_svc_stats != NULL) { lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR, timediff); @@ -1320,20 +1435,20 @@ static int after_reply(struct ptlrpc_request *req) rc = ptlrpc_check_status(req); imp->imp_connect_error = rc; - if (rc) { - /* - * Either we've been evicted, or the server has failed for - * some reason. Try to reconnect, and if that fails, punt to - * the upcall. - */ - if (ll_rpc_recoverable_error(rc)) { - if (req->rq_send_state != LUSTRE_IMP_FULL || - imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) { - RETURN(rc); - } - ptlrpc_request_handle_notconn(req); - RETURN(rc); - } + if (rc) { + /* + * Either we've been evicted, or the server has failed for + * some reason. Try to reconnect, and if that fails, punt to + * the upcall. + */ + if (ptlrpc_recoverable_error(rc)) { + if (req->rq_send_state != LUSTRE_IMP_FULL || + imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) { + RETURN(rc); + } + ptlrpc_request_handle_notconn(req); + RETURN(rc); + } } else { /* * Let's look if server sent slv. Do it only for RPC with @@ -1376,19 +1491,18 @@ static int after_reply(struct ptlrpc_request *req) /* * Replay-enabled imports return commit-status information. */ - if (lustre_msg_get_last_committed(req->rq_repmsg)) { - imp->imp_peer_committed_transno = - lustre_msg_get_last_committed(req->rq_repmsg); - } + committed = lustre_msg_get_last_committed(req->rq_repmsg); + if (likely(committed > imp->imp_peer_committed_transno)) + imp->imp_peer_committed_transno = committed; ptlrpc_free_committed(imp); - if (!cfs_list_empty(&imp->imp_replay_list)) { + if (!list_empty(&imp->imp_replay_list)) { struct ptlrpc_request *last; - last = cfs_list_entry(imp->imp_replay_list.prev, - struct ptlrpc_request, - rq_replay_list); + last = list_entry(imp->imp_replay_list.prev, + struct ptlrpc_request, + rq_replay_list); /* * Requests with rq_replay stay on the list even if no * commit is expected. @@ -1411,11 +1525,19 @@ static int after_reply(struct ptlrpc_request *req) static int ptlrpc_send_new_req(struct ptlrpc_request *req) { struct obd_import *imp = req->rq_import; + __u64 min_xid = 0; int rc; ENTRY; LASSERT(req->rq_phase == RQ_PHASE_NEW); - if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()) && + + /* do not try to go further if there is not enough memory in enc_pool */ + if (req->rq_sent && req->rq_bulk != NULL) + if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() && + pool_is_at_full_capacity()) + RETURN(-ENOMEM); + + if (req->rq_sent && (req->rq_sent > ktime_get_real_seconds()) && (!req->rq_generation_set || req->rq_import_generation == imp->imp_generation)) RETURN (0); @@ -1424,6 +1546,9 @@ static int ptlrpc_send_new_req(struct ptlrpc_request *req) spin_lock(&imp->imp_lock); + LASSERT(req->rq_xid != 0); + LASSERT(!list_empty(&req->rq_unreplied_list)); + if (!req->rq_generation_set) req->rq_import_generation = imp->imp_generation; @@ -1432,12 +1557,11 @@ static int ptlrpc_send_new_req(struct ptlrpc_request *req) req->rq_waiting = 1; spin_unlock(&req->rq_lock); - DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: " - "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg), + DEBUG_REQ(D_HA, req, "req waiting for recovery: (%s != %s)", ptlrpc_import_state_name(req->rq_send_state), ptlrpc_import_state_name(imp->imp_state)); - LASSERT(cfs_list_empty(&req->rq_list)); - cfs_list_add_tail(&req->rq_list, &imp->imp_delayed_list); + LASSERT(list_empty(&req->rq_list)); + list_add_tail(&req->rq_list, &imp->imp_delayed_list); atomic_inc(&req->rq_import->imp_inflight); spin_unlock(&imp->imp_lock); RETURN(0); @@ -1450,11 +1574,28 @@ static int ptlrpc_send_new_req(struct ptlrpc_request *req) RETURN(rc); } - LASSERT(cfs_list_empty(&req->rq_list)); - cfs_list_add_tail(&req->rq_list, &imp->imp_sending_list); + LASSERT(list_empty(&req->rq_list)); + list_add_tail(&req->rq_list, &imp->imp_sending_list); atomic_inc(&req->rq_import->imp_inflight); + + /* find the known replied XID from the unreplied list, CONNECT + * and DISCONNECT requests are skipped to make the sanity check + * on server side happy. see process_req_last_xid(). + * + * For CONNECT: Because replay requests have lower XID, it'll + * break the sanity check if CONNECT bump the exp_last_xid on + * server. + * + * For DISCONNECT: Since client will abort inflight RPC before + * sending DISCONNECT, DISCONNECT may carry an XID which higher + * than the inflight RPC. + */ + if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req)) + min_xid = ptlrpc_known_replied_xid(imp); spin_unlock(&imp->imp_lock); + lustre_msg_set_last_xid(req->rq_reqmsg, min_xid); + lustre_msg_set_status(req->rq_reqmsg, current_pid()); rc = sptlrpc_req_refresh_ctx(req, -1); @@ -1471,13 +1612,23 @@ static int ptlrpc_send_new_req(struct ptlrpc_request *req) } CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc" - " %s:%s:%d:"LPU64":%s:%d\n", current_comm(), + " %s:%s:%d:%llu:%s:%d\n", current_comm(), imp->imp_obd->obd_uuid.uuid, lustre_msg_get_status(req->rq_reqmsg), req->rq_xid, libcfs_nid2str(imp->imp_connection->c_peer.nid), lustre_msg_get_opc(req->rq_reqmsg)); rc = ptl_send_rpc(req, 0); + if (rc == -ENOMEM) { + spin_lock(&imp->imp_lock); + if (!list_empty(&req->rq_list)) { + list_del_init(&req->rq_list); + atomic_dec(&req->rq_import->imp_inflight); + } + spin_unlock(&imp->imp_lock); + ptlrpc_rqphase_move(req, RQ_PHASE_NEW); + RETURN(rc); + } if (rc) { DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc); spin_lock(&req->rq_lock); @@ -1517,61 +1668,106 @@ static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set) * and no more replies are expected. * (it is possible to get less replies than requests sent e.g. due to timed out * requests or requests that we had trouble to send out) + * + * NOTE: This function contains a potential schedule point (cond_resched()). */ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set) { - cfs_list_t *tmp, *next; - int force_timer_recalc = 0; - ENTRY; + struct list_head *tmp, *next; + struct list_head comp_reqs; + int force_timer_recalc = 0; + ENTRY; if (atomic_read(&set->set_remaining) == 0) - RETURN(1); + RETURN(1); - cfs_list_for_each_safe(tmp, next, &set->set_requests) { - struct ptlrpc_request *req = - cfs_list_entry(tmp, struct ptlrpc_request, - rq_set_chain); - struct obd_import *imp = req->rq_import; - int unregistered = 0; - int rc = 0; - - if (req->rq_phase == RQ_PHASE_NEW && - ptlrpc_send_new_req(req)) { - force_timer_recalc = 1; - } + INIT_LIST_HEAD(&comp_reqs); + list_for_each_safe(tmp, next, &set->set_requests) { + struct ptlrpc_request *req = + list_entry(tmp, struct ptlrpc_request, + rq_set_chain); + struct obd_import *imp = req->rq_import; + int unregistered = 0; + int async = 1; + int rc = 0; - /* delayed send - skip */ - if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent) + if (req->rq_phase == RQ_PHASE_COMPLETE) { + list_move_tail(&req->rq_set_chain, &comp_reqs); + continue; + } + + /* This schedule point is mainly for the ptlrpcd caller of this + * function. Most ptlrpc sets are not long-lived and unbounded + * in length, but at the least the set used by the ptlrpcd is. + * Since the processing time is unbounded, we need to insert an + * explicit schedule point to make the thread well-behaved. + */ + cond_resched(); + + /* If the caller requires to allow to be interpreted by force + * and it has really been interpreted, then move the request + * to RQ_PHASE_INTERPRET phase in spite of what the current + * phase is. */ + if (unlikely(req->rq_allow_intr && req->rq_intr)) { + req->rq_status = -EINTR; + ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET); + + /* Since it is interpreted and we have to wait for + * the reply to be unlinked, then use sync mode. */ + async = 0; + + GOTO(interpret, req->rq_status); + } + + if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req)) + force_timer_recalc = 1; + + /* delayed send - skip */ + if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent) continue; /* delayed resend - skip */ if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend && - req->rq_sent > cfs_time_current_sec()) + req->rq_sent > ktime_get_real_seconds()) continue; - if (!(req->rq_phase == RQ_PHASE_RPC || - req->rq_phase == RQ_PHASE_BULK || - req->rq_phase == RQ_PHASE_INTERPRET || - req->rq_phase == RQ_PHASE_UNREGISTERING || - req->rq_phase == RQ_PHASE_COMPLETE)) { - DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase); - LBUG(); - } + if (!(req->rq_phase == RQ_PHASE_RPC || + req->rq_phase == RQ_PHASE_BULK || + req->rq_phase == RQ_PHASE_INTERPRET || + req->rq_phase == RQ_PHASE_UNREG_RPC || + req->rq_phase == RQ_PHASE_UNREG_BULK)) { + DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase); + LBUG(); + } - if (req->rq_phase == RQ_PHASE_UNREGISTERING) { - LASSERT(req->rq_next_phase != req->rq_phase); - LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED); + if (req->rq_phase == RQ_PHASE_UNREG_RPC || + req->rq_phase == RQ_PHASE_UNREG_BULK) { + LASSERT(req->rq_next_phase != req->rq_phase); + LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED); + + if (req->rq_req_deadline && + !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK)) + req->rq_req_deadline = 0; + if (req->rq_reply_deadline && + !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) + req->rq_reply_deadline = 0; + if (req->rq_bulk_deadline && + !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) + req->rq_bulk_deadline = 0; - /* - * Skip processing until reply is unlinked. We - * can't return to pool before that and we can't - * call interpret before that. We need to make - * sure that all rdma transfers finished and will - * not corrupt any data. - */ - if (ptlrpc_client_recv_or_unlink(req) || - ptlrpc_client_bulk_active(req)) - continue; + /* + * Skip processing until reply is unlinked. We + * can't return to pool before that and we can't + * call interpret before that. We need to make + * sure that all rdma transfers finished and will + * not corrupt any data. + */ + if (req->rq_phase == RQ_PHASE_UNREG_RPC && + ptlrpc_client_recv_or_unlink(req)) + continue; + if (req->rq_phase == RQ_PHASE_UNREG_BULK && + ptlrpc_client_bulk_active(req)) + continue; /* * Turn fail_loc off to prevent it from looping @@ -1593,9 +1789,6 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set) ptlrpc_rqphase_move(req, req->rq_next_phase); } - if (req->rq_phase == RQ_PHASE_COMPLETE) - continue; - if (req->rq_phase == RQ_PHASE_INTERPRET) GOTO(interpret, req->rq_status); @@ -1634,7 +1827,7 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set) /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr * so it sets rq_intr regardless of individual rpc - * timeouts. The synchronous IO waiting path sets + * timeouts. The synchronous IO waiting path sets * rq_intr irrespective of whether ptlrpcd * has seen a timeout. Our policy is to only interpret * interrupted rpcs after they have timed out, so we @@ -1653,15 +1846,17 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set) req->rq_waiting || req->rq_wait_ctx) { int status; - if (!ptlrpc_unregister_reply(req, 1)) - continue; + if (!ptlrpc_unregister_reply(req, 1)) { + ptlrpc_unregister_bulk(req, 1); + continue; + } spin_lock(&imp->imp_lock); if (ptlrpc_import_delay_req(imp, req, &status)){ /* put on delay list - only if we wait * recovery finished - before send */ - cfs_list_del_init(&req->rq_list); - cfs_list_add_tail(&req->rq_list, + list_del_init(&req->rq_list); + list_add_tail(&req->rq_list, &imp-> imp_delayed_list); spin_unlock(&imp->imp_lock); @@ -1684,8 +1879,8 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set) GOTO(interpret, req->rq_status); } - cfs_list_del_init(&req->rq_list); - cfs_list_add_tail(&req->rq_list, + list_del_init(&req->rq_list); + list_add_tail(&req->rq_list, &imp->imp_sending_list); spin_unlock(&imp->imp_lock); @@ -1700,20 +1895,10 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set) spin_lock(&req->rq_lock); req->rq_resend = 1; spin_unlock(&req->rq_lock); - if (req->rq_bulk) { - __u64 old_xid; - - if (!ptlrpc_unregister_bulk(req, 1)) - continue; - - /* ensure previous bulk fails */ - old_xid = req->rq_xid; - req->rq_xid = ptlrpc_next_xid(); - CDEBUG(D_HA, "resend bulk " - "old x"LPU64 - " new x"LPU64"\n", - old_xid, req->rq_xid); - } + + if (req->rq_bulk != NULL && + !ptlrpc_unregister_bulk(req, 1)) + continue; } /* * rq_wait_ctx is only touched by ptlrpcd, @@ -1741,6 +1926,14 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set) } rc = ptl_send_rpc(req, 0); + if (rc == -ENOMEM) { + spin_lock(&imp->imp_lock); + if (!list_empty(&req->rq_list)) + list_del_init(&req->rq_list); + spin_unlock(&imp->imp_lock); + ptlrpc_rqphase_move(req, RQ_PHASE_NEW); + continue; + } if (rc) { DEBUG_REQ(D_HA, req, "send failed: rc = %d", rc); @@ -1812,27 +2005,27 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set) req->rq_status = -EIO; } - ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET); + ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET); - interpret: - LASSERT(req->rq_phase == RQ_PHASE_INTERPRET); + interpret: + LASSERT(req->rq_phase == RQ_PHASE_INTERPRET); - /* This moves to "unregistering" phase we need to wait for - * reply unlink. */ - if (!unregistered && !ptlrpc_unregister_reply(req, 1)) { - /* start async bulk unlink too */ - ptlrpc_unregister_bulk(req, 1); - continue; - } + /* This moves to "unregistering" phase we need to wait for + * reply unlink. */ + if (!unregistered && !ptlrpc_unregister_reply(req, async)) { + /* start async bulk unlink too */ + ptlrpc_unregister_bulk(req, 1); + continue; + } - if (!ptlrpc_unregister_bulk(req, 1)) - continue; + if (!ptlrpc_unregister_bulk(req, async)) + continue; - /* When calling interpret receiving already should be - * finished. */ - LASSERT(!req->rq_receiving_reply); + /* When calling interpret receiving already should be + * finished. */ + LASSERT(!req->rq_receiving_reply); - ptlrpc_req_interpret(env, req, req->rq_status); + ptlrpc_req_interpret(env, req, req->rq_status); if (ptlrpcd_check_work(req)) { atomic_dec(&set->set_remaining); @@ -1840,23 +2033,26 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set) } ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE); - CDEBUG(req->rq_reqmsg != NULL ? D_RPCTRACE : 0, - "Completed RPC pname:cluuid:pid:xid:nid:" - "opc %s:%s:%d:"LPU64":%s:%d\n", - current_comm(), imp->imp_obd->obd_uuid.uuid, - lustre_msg_get_status(req->rq_reqmsg), req->rq_xid, - libcfs_nid2str(imp->imp_connection->c_peer.nid), - lustre_msg_get_opc(req->rq_reqmsg)); + if (req->rq_reqmsg != NULL) + CDEBUG(D_RPCTRACE, + "Completed RPC pname:cluuid:pid:xid:nid:" + "opc %s:%s:%d:%llu:%s:%d\n", current_comm(), + imp->imp_obd->obd_uuid.uuid, + lustre_msg_get_status(req->rq_reqmsg), + req->rq_xid, + libcfs_nid2str(imp->imp_connection->c_peer.nid), + lustre_msg_get_opc(req->rq_reqmsg)); spin_lock(&imp->imp_lock); /* Request already may be not on sending or delaying list. This * may happen in the case of marking it erroneous for the case * ptlrpc_import_delay_req(req, status) find it impossible to * allow sending this rpc and returns *status != 0. */ - if (!cfs_list_empty(&req->rq_list)) { - cfs_list_del_init(&req->rq_list); + if (!list_empty(&req->rq_list)) { + list_del_init(&req->rq_list); atomic_dec(&imp->imp_inflight); } + list_del_init(&req->rq_unreplied_list); spin_unlock(&imp->imp_lock); atomic_dec(&set->set_remaining); @@ -1869,7 +2065,7 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set) /* free the request that has just been completed * in order not to pollute set->set_requests */ - cfs_list_del_init(&req->rq_set_chain); + list_del_init(&req->rq_set_chain); spin_lock(&req->rq_lock); req->rq_set = NULL; req->rq_invalid_rqset = 0; @@ -1879,9 +2075,15 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set) if (req->rq_status != 0) set->set_rc = req->rq_status; ptlrpc_req_finished(req); + } else { + list_move_tail(&req->rq_set_chain, &comp_reqs); } } + /* move completed request at the head of list so it's easier for + * caller to find them */ + list_splice(&comp_reqs, &set->set_requests); + /* If we hit an error, we want to recover promptly. */ RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc); } @@ -1902,17 +2104,16 @@ int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink) req->rq_timedout = 1; spin_unlock(&req->rq_lock); - DEBUG_REQ(D_WARNING, req, "Request sent has %s: [sent "CFS_DURATION_T - "/real "CFS_DURATION_T"]", + DEBUG_REQ(D_WARNING, req, "Request sent has %s: [sent %lld/real %lld]", req->rq_net_err ? "failed due to network error" : ((req->rq_real_sent == 0 || - cfs_time_before(req->rq_real_sent, req->rq_sent) || - cfs_time_aftereq(req->rq_real_sent, req->rq_deadline)) ? + req->rq_real_sent < req->rq_sent || + req->rq_real_sent >= req->rq_deadline) ? "timed out for sent delay" : "timed out for slow reply"), - req->rq_sent, req->rq_real_sent); + (s64)req->rq_sent, (s64)req->rq_real_sent); - if (imp != NULL && obd_debug_peer_on_timeout) - LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer); + if (imp != NULL && obd_debug_peer_on_timeout) + LNetDebugPeer(imp->imp_connection->c_peer); ptlrpc_unregister_reply(req, async_unlink); ptlrpc_unregister_bulk(req, async_unlink); @@ -1965,20 +2166,20 @@ int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink) */ int ptlrpc_expired_set(void *data) { - struct ptlrpc_request_set *set = data; - cfs_list_t *tmp; - time_t now = cfs_time_current_sec(); - ENTRY; + struct ptlrpc_request_set *set = data; + struct list_head *tmp; + time64_t now = ktime_get_real_seconds(); - LASSERT(set != NULL); + ENTRY; + LASSERT(set != NULL); - /* - * A timeout expired. See which reqs it applies to... - */ - cfs_list_for_each (tmp, &set->set_requests) { - struct ptlrpc_request *req = - cfs_list_entry(tmp, struct ptlrpc_request, - rq_set_chain); + /* + * A timeout expired. See which reqs it applies to... + */ + list_for_each(tmp, &set->set_requests) { + struct ptlrpc_request *req = + list_entry(tmp, struct ptlrpc_request, + rq_set_chain); /* don't expire request waiting for context */ if (req->rq_wait_ctx) @@ -2006,7 +2207,6 @@ int ptlrpc_expired_set(void *data) */ RETURN(1); } -EXPORT_SYMBOL(ptlrpc_expired_set); /** * Sets rq_intr flag in \a req under spinlock. @@ -2023,42 +2223,44 @@ EXPORT_SYMBOL(ptlrpc_mark_interrupted); * Interrupts (sets interrupted flag) all uncompleted requests in * a set \a data. Callback for l_wait_event for interruptible waits. */ -void ptlrpc_interrupted_set(void *data) +static void ptlrpc_interrupted_set(void *data) { - struct ptlrpc_request_set *set = data; - cfs_list_t *tmp; + struct ptlrpc_request_set *set = data; + struct list_head *tmp; - LASSERT(set != NULL); - CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set); + LASSERT(set != NULL); + CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set); - cfs_list_for_each(tmp, &set->set_requests) { - struct ptlrpc_request *req = - cfs_list_entry(tmp, struct ptlrpc_request, - rq_set_chain); + list_for_each(tmp, &set->set_requests) { + struct ptlrpc_request *req = + list_entry(tmp, struct ptlrpc_request, rq_set_chain); - if (req->rq_phase != RQ_PHASE_RPC && - req->rq_phase != RQ_PHASE_UNREGISTERING) - continue; + if (req->rq_intr) + continue; - ptlrpc_mark_interrupted(req); - } + if (req->rq_phase != RQ_PHASE_RPC && + req->rq_phase != RQ_PHASE_UNREG_RPC && + !req->rq_allow_intr) + continue; + + ptlrpc_mark_interrupted(req); + } } -EXPORT_SYMBOL(ptlrpc_interrupted_set); /** * Get the smallest timeout in the set; this does NOT set a timeout. */ int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set) { - cfs_list_t *tmp; - time_t now = cfs_time_current_sec(); - int timeout = 0; - struct ptlrpc_request *req; - int deadline; - ENTRY; + struct list_head *tmp; + time64_t now = ktime_get_real_seconds(); + int timeout = 0; + struct ptlrpc_request *req; + time64_t deadline; - cfs_list_for_each(tmp, &set->set_requests) { - req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain); + ENTRY; + list_for_each(tmp, &set->set_requests) { + req = list_entry(tmp, struct ptlrpc_request, rq_set_chain); /* * Request in-flight? @@ -2094,7 +2296,6 @@ int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set) } RETURN(timeout); } -EXPORT_SYMBOL(ptlrpc_set_next_timeout); /** * Send all unset request from the set and then wait untill all @@ -2104,7 +2305,7 @@ EXPORT_SYMBOL(ptlrpc_set_next_timeout); */ int ptlrpc_set_wait(struct ptlrpc_request_set *set) { - cfs_list_t *tmp; + struct list_head *tmp; struct ptlrpc_request *req; struct l_wait_info lwi; int rc, timeout; @@ -2113,14 +2314,14 @@ int ptlrpc_set_wait(struct ptlrpc_request_set *set) if (set->set_producer) (void)ptlrpc_set_producer(set); else - cfs_list_for_each(tmp, &set->set_requests) { - req = cfs_list_entry(tmp, struct ptlrpc_request, - rq_set_chain); + list_for_each(tmp, &set->set_requests) { + req = list_entry(tmp, struct ptlrpc_request, + rq_set_chain); if (req->rq_phase == RQ_PHASE_NEW) (void)ptlrpc_send_new_req(req); } - if (cfs_list_empty(&set->set_requests)) + if (list_empty(&set->set_requests)) RETURN(0); do { @@ -2131,21 +2332,21 @@ int ptlrpc_set_wait(struct ptlrpc_request_set *set) CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n", set, timeout); - if (timeout == 0 && !cfs_signal_pending()) - /* - * No requests are in-flight (ether timed out - * or delayed), so we can allow interrupts. - * We still want to block for a limited time, - * so we allow interrupts during the timeout. - */ - lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1), - ptlrpc_expired_set, - ptlrpc_interrupted_set, set); + if ((timeout == 0 && !signal_pending(current)) || + set->set_allow_intr) + /* No requests are in-flight (ether timed out + * or delayed), so we can allow interrupts. + * We still want to block for a limited time, + * so we allow interrupts during the timeout. */ + lwi = LWI_TIMEOUT_INTR_ALL( + cfs_time_seconds(timeout ? timeout : 1), + ptlrpc_expired_set, + ptlrpc_interrupted_set, set); else /* * At least one request is in flight, so no * interrupts are allowed. Wait until all - * complete, or an in-flight req times out. + * complete, or an in-flight req times out. */ lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1), ptlrpc_expired_set, set); @@ -2155,8 +2356,9 @@ int ptlrpc_set_wait(struct ptlrpc_request_set *set) /* LU-769 - if we ignored the signal because it was already * pending when we started, we need to handle it now or we risk * it being ignored forever */ - if (rc == -ETIMEDOUT && !lwi.lwi_allow_intr && - cfs_signal_pending()) { + if (rc == -ETIMEDOUT && + (!lwi.lwi_allow_intr || set->set_allow_intr) && + signal_pending(current)) { sigset_t blocked_sigs = cfs_block_sigsinv(LUSTRE_FATAL_SIGS); @@ -2164,7 +2366,7 @@ int ptlrpc_set_wait(struct ptlrpc_request_set *set) * like SIGINT or SIGKILL. We still ignore less * important signals since ptlrpc set is not easily * reentrant from userspace again */ - if (cfs_signal_pending()) + if (signal_pending(current)) ptlrpc_interrupted_set(set); cfs_restore_sigs(blocked_sigs); } @@ -2179,9 +2381,9 @@ int ptlrpc_set_wait(struct ptlrpc_request_set *set) * I don't really care if we go once more round the loop in * the error cases -eeb. */ if (rc == 0 && atomic_read(&set->set_remaining) == 0) { - cfs_list_for_each(tmp, &set->set_requests) { - req = cfs_list_entry(tmp, struct ptlrpc_request, - rq_set_chain); + list_for_each(tmp, &set->set_requests) { + req = list_entry(tmp, struct ptlrpc_request, + rq_set_chain); spin_lock(&req->rq_lock); req->rq_invalid_rqset = 1; spin_unlock(&req->rq_lock); @@ -2192,8 +2394,8 @@ int ptlrpc_set_wait(struct ptlrpc_request_set *set) LASSERT(atomic_read(&set->set_remaining) == 0); rc = set->set_rc; /* rq_status of already freed requests if any */ - cfs_list_for_each(tmp, &set->set_requests) { - req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain); + list_for_each(tmp, &set->set_requests) { + req = list_entry(tmp, struct ptlrpc_request, rq_set_chain); LASSERT(req->rq_phase == RQ_PHASE_COMPLETE); if (req->rq_status != 0) @@ -2208,9 +2410,9 @@ int ptlrpc_set_wait(struct ptlrpc_request_set *set) struct ptlrpc_set_cbdata *cbdata, *n; int err; - cfs_list_for_each_entry_safe(cbdata, n, + list_for_each_entry_safe(cbdata, n, &set->set_cblist, psc_item) { - cfs_list_del_init(&cbdata->psc_item); + list_del_init(&cbdata->psc_item); err = cbdata->psc_interpret(set, cbdata->psc_data, rc); if (err && !rc) rc = err; @@ -2232,31 +2434,31 @@ EXPORT_SYMBOL(ptlrpc_set_wait); */ static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked) { - ENTRY; - if (request == NULL) { - EXIT; - return; - } + ENTRY; - LASSERTF(!request->rq_receiving_reply, "req %p\n", request); - LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */ - LASSERTF(cfs_list_empty(&request->rq_list), "req %p\n", request); - LASSERTF(cfs_list_empty(&request->rq_set_chain), "req %p\n", request); - LASSERTF(cfs_list_empty(&request->rq_exp_list), "req %p\n", request); - LASSERTF(!request->rq_replay, "req %p\n", request); + if (request == NULL) + RETURN_EXIT; - req_capsule_fini(&request->rq_pill); + LASSERT(!request->rq_srv_req); + LASSERT(request->rq_export == NULL); + LASSERTF(!request->rq_receiving_reply, "req %p\n", request); + LASSERTF(list_empty(&request->rq_list), "req %p\n", request); + LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request); + LASSERTF(!request->rq_replay, "req %p\n", request); - /* We must take it off the imp_replay_list first. Otherwise, we'll set - * request->rq_reqmsg to NULL while osc_close is dereferencing it. */ - if (request->rq_import != NULL) { + req_capsule_fini(&request->rq_pill); + + /* We must take it off the imp_replay_list first. Otherwise, we'll set + * request->rq_reqmsg to NULL while osc_close is dereferencing it. */ + if (request->rq_import != NULL) { if (!locked) spin_lock(&request->rq_import->imp_lock); - cfs_list_del_init(&request->rq_replay_list); + list_del_init(&request->rq_replay_list); + list_del_init(&request->rq_unreplied_list); if (!locked) spin_unlock(&request->rq_import->imp_lock); } - LASSERTF(cfs_list_empty(&request->rq_replay_list), "req %p\n", request); + LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request); if (atomic_read(&request->rq_refcount) != 0) { DEBUG_REQ(D_ERROR, request, @@ -2266,16 +2468,13 @@ static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked) if (request->rq_repbuf != NULL) sptlrpc_cli_free_repbuf(request); - if (request->rq_export != NULL) { - class_export_put(request->rq_export); - request->rq_export = NULL; - } + if (request->rq_import != NULL) { class_import_put(request->rq_import); request->rq_import = NULL; } if (request->rq_bulk != NULL) - ptlrpc_free_bulk_pin(request->rq_bulk); + ptlrpc_free_bulk(request->rq_bulk); if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL) sptlrpc_cli_free_reqbuf(request); @@ -2293,43 +2492,67 @@ static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked) static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked); /** * Drop one request reference. Must be called with import imp_lock held. - * When reference count drops to zero, reuqest is freed. + * When reference count drops to zero, request is freed. */ void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request) { - LASSERT(spin_is_locked(&request->rq_import->imp_lock)); + assert_spin_locked(&request->rq_import->imp_lock); (void)__ptlrpc_req_finished(request, 1); } -EXPORT_SYMBOL(ptlrpc_req_finished_with_imp_lock); /** * Helper function * Drops one reference count for request \a request. * \a locked set indicates that caller holds import imp_lock. * Frees the request whe reference count reaches zero. + * + * \retval 1 the request is freed + * \retval 0 some others still hold references on the request */ static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked) { - ENTRY; - if (request == NULL) - RETURN(1); + int count; + ENTRY; - if (request == LP_POISON || - request->rq_reqmsg == LP_POISON) { - CERROR("dereferencing freed request (bug 575)\n"); - LBUG(); - RETURN(1); - } + if (!request) + RETURN(1); + + LASSERT(request != LP_POISON); + LASSERT(request->rq_reqmsg != LP_POISON); - DEBUG_REQ(D_INFO, request, "refcount now %u", + DEBUG_REQ(D_INFO, request, "refcount now %u", atomic_read(&request->rq_refcount) - 1); - if (atomic_dec_and_test(&request->rq_refcount)) { - __ptlrpc_free_req(request, locked); - RETURN(1); - } + spin_lock(&request->rq_lock); + count = atomic_dec_return(&request->rq_refcount); + LASSERTF(count >= 0, "Invalid ref count %d\n", count); - RETURN(0); + /* For open RPC, the client does not know the EA size (LOV, ACL, and + * so on) before replied, then the client has to reserve very large + * reply buffer. Such buffer will not be released until the RPC freed. + * Since The open RPC is replayable, we need to keep it in the replay + * list until close. If there are a lot of files opened concurrently, + * then the client may be OOM. + * + * If fact, it is unnecessary to keep reply buffer for open replay, + * related EAs have already been saved via mdc_save_lovea() before + * coming here. So it is safe to free the reply buffer some earlier + * before releasing the RPC to avoid client OOM. LU-9514 */ + if (count == 1 && request->rq_early_free_repbuf && request->rq_repbuf) { + spin_lock(&request->rq_early_free_lock); + sptlrpc_cli_free_repbuf(request); + request->rq_repbuf = NULL; + request->rq_repbuf_len = 0; + request->rq_repdata = NULL; + request->rq_reqdata_len = 0; + spin_unlock(&request->rq_early_free_lock); + } + spin_unlock(&request->rq_lock); + + if (!count) + __ptlrpc_free_req(request, locked); + + RETURN(!count); } /** @@ -2357,7 +2580,7 @@ EXPORT_SYMBOL(ptlrpc_req_xid); * The request owner (i.e. the thread doing the I/O) must call... * Returns 0 on success or 1 if unregistering cannot be made. */ -int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async) +static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async) { int rc; struct l_wait_info lwi; @@ -2367,12 +2590,11 @@ int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async) */ LASSERT(!in_interrupt()); - /* - * Let's setup deadline for reply unlink. - */ - if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) && - async && request->rq_reply_deadline == 0) - request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK; + /* Let's setup deadline for reply unlink. */ + if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) && + async && request->rq_reply_deadline == 0 && cfs_fail_val == 0) + request->rq_reply_deadline = ktime_get_real_seconds() + + LONG_UNLINK; /* * Nothing left to do. @@ -2388,10 +2610,8 @@ int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async) if (!ptlrpc_client_recv_or_unlink(request)) RETURN(1); - /* - * Move to "Unregistering" phase as reply was not unlinked yet. - */ - ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING); + /* Move to "Unregistering" phase as reply was not unlinked yet. */ + ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC); /* * Do not wait for unlink to finish. @@ -2405,12 +2625,10 @@ int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async) * unlinked before returning a req to the pool. */ for (;;) { -#ifdef __KERNEL__ /* The wq argument is ignored by user-space wait_event macros */ wait_queue_head_t *wq = (request->rq_set != NULL) ? &request->rq_set->set_waitq : &request->rq_reply_waitq; -#endif /* Network access will complete in finite time but the HUGE * timeout lets us CWARN for visibility of sluggish NALs */ lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK), @@ -2423,13 +2641,14 @@ int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async) } LASSERT(rc == -ETIMEDOUT); - DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout " - "rvcng=%d unlnk=%d", request->rq_receiving_reply, - request->rq_must_unlink); + DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout " + "receiving_reply=%d req_ulinked=%d reply_unlinked=%d", + request->rq_receiving_reply, + request->rq_req_unlinked, + request->rq_reply_unlinked); } RETURN(0); } -EXPORT_SYMBOL(ptlrpc_unregister_reply); static void ptlrpc_free_request(struct ptlrpc_request *req) { @@ -2439,7 +2658,7 @@ static void ptlrpc_free_request(struct ptlrpc_request *req) if (req->rq_commit_cb != NULL) req->rq_commit_cb(req); - cfs_list_del_init(&req->rq_replay_list); + list_del_init(&req->rq_replay_list); __ptlrpc_req_finished(req, 1); } @@ -2452,7 +2671,7 @@ void ptlrpc_request_committed(struct ptlrpc_request *req, int force) struct obd_import *imp = req->rq_import; spin_lock(&imp->imp_lock); - if (cfs_list_empty(&req->rq_replay_list)) { + if (list_empty(&req->rq_replay_list)) { spin_unlock(&imp->imp_lock); return; } @@ -2480,26 +2699,26 @@ void ptlrpc_free_committed(struct obd_import *imp) ENTRY; LASSERT(imp != NULL); - LASSERT(spin_is_locked(&imp->imp_lock)); - + assert_spin_locked(&imp->imp_lock); if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked && imp->imp_generation == imp->imp_last_generation_checked) { - CDEBUG(D_INFO, "%s: skip recheck: last_committed "LPU64"\n", + CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n", imp->imp_obd->obd_name, imp->imp_peer_committed_transno); RETURN_EXIT; } - CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n", + CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n", imp->imp_obd->obd_name, imp->imp_peer_committed_transno, imp->imp_generation); - if (imp->imp_generation != imp->imp_last_generation_checked) + if (imp->imp_generation != imp->imp_last_generation_checked || + imp->imp_last_transno_checked == 0) skip_committed_list = false; imp->imp_last_transno_checked = imp->imp_peer_committed_transno; imp->imp_last_generation_checked = imp->imp_generation; - cfs_list_for_each_entry_safe(req, saved, &imp->imp_replay_list, + list_for_each_entry_safe(req, saved, &imp->imp_replay_list, rq_replay_list) { /* XXX ok to remove when 1357 resolved - rread 05/29/03 */ LASSERT(req != last_req); @@ -2522,12 +2741,12 @@ void ptlrpc_free_committed(struct obd_import *imp) if (req->rq_replay) { DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)"); - cfs_list_move_tail(&req->rq_replay_list, + list_move_tail(&req->rq_replay_list, &imp->imp_committed_list); continue; } - DEBUG_REQ(D_INFO, req, "commit (last_committed "LPU64")", + DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)", imp->imp_peer_committed_transno); free_req: ptlrpc_free_request(req); @@ -2536,11 +2755,19 @@ free_req: if (skip_committed_list) GOTO(out, 0); - cfs_list_for_each_entry_safe(req, saved, &imp->imp_committed_list, - rq_replay_list) { + list_for_each_entry_safe(req, saved, &imp->imp_committed_list, + rq_replay_list) { LASSERT(req->rq_transno != 0); - if (req->rq_import_generation < imp->imp_generation) { - DEBUG_REQ(D_RPCTRACE, req, "free stale open request"); + if (req->rq_import_generation < imp->imp_generation || + !req->rq_replay) { + DEBUG_REQ(D_RPCTRACE, req, "free %s open request", + req->rq_import_generation < + imp->imp_generation ? "stale" : "closed"); + + if (imp->imp_replay_cursor == &req->rq_replay_list) + imp->imp_replay_cursor = + req->rq_replay_list.next; + ptlrpc_free_request(req); } } @@ -2553,7 +2780,6 @@ void ptlrpc_cleanup_client(struct obd_import *imp) ENTRY; EXIT; } -EXPORT_SYMBOL(ptlrpc_cleanup_client); /** * Schedule previously sent request for resend. @@ -2564,25 +2790,26 @@ EXPORT_SYMBOL(ptlrpc_cleanup_client); void ptlrpc_resend_req(struct ptlrpc_request *req) { DEBUG_REQ(D_HA, req, "going to resend"); + spin_lock(&req->rq_lock); + + /* Request got reply but linked to the import list still. + Let ptlrpc_check_set() to process it. */ + if (ptlrpc_client_replied(req)) { + spin_unlock(&req->rq_lock); + DEBUG_REQ(D_HA, req, "it has reply, so skip it"); + return; + } + lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 }); req->rq_status = -EAGAIN; - spin_lock(&req->rq_lock); req->rq_resend = 1; req->rq_net_err = 0; req->rq_timedout = 0; - if (req->rq_bulk) { - __u64 old_xid = req->rq_xid; - /* ensure previous bulk fails */ - req->rq_xid = ptlrpc_next_xid(); - CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n", - old_xid, req->rq_xid); - } ptlrpc_client_wake_req(req); spin_unlock(&req->rq_lock); } -EXPORT_SYMBOL(ptlrpc_resend_req); /* XXX: this function and rq_status are currently unused */ void ptlrpc_restart_req(struct ptlrpc_request *req) @@ -2596,7 +2823,6 @@ void ptlrpc_restart_req(struct ptlrpc_request *req) ptlrpc_client_wake_req(req); spin_unlock(&req->rq_lock); } -EXPORT_SYMBOL(ptlrpc_restart_req); /** * Grab additional reference on a request \a req @@ -2616,9 +2842,9 @@ EXPORT_SYMBOL(ptlrpc_request_addref); void ptlrpc_retain_replayable_request(struct ptlrpc_request *req, struct obd_import *imp) { - cfs_list_t *tmp; + struct list_head *tmp; - LASSERT(spin_is_locked(&imp->imp_lock)); + assert_spin_locked(&imp->imp_lock); if (req->rq_transno == 0) { DEBUG_REQ(D_EMERG, req, "saving request with zero transno"); @@ -2629,19 +2855,23 @@ void ptlrpc_retain_replayable_request(struct ptlrpc_request *req, as resent replayed requests. */ lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT); - /* don't re-add requests that have been replayed */ - if (!cfs_list_empty(&req->rq_replay_list)) - return; + /* don't re-add requests that have been replayed */ + if (!list_empty(&req->rq_replay_list)) + return; - lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY); + lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY); - LASSERT(imp->imp_replayable); - /* Balanced in ptlrpc_free_committed, usually. */ - ptlrpc_request_addref(req); - cfs_list_for_each_prev(tmp, &imp->imp_replay_list) { - struct ptlrpc_request *iter = - cfs_list_entry(tmp, struct ptlrpc_request, - rq_replay_list); + spin_lock(&req->rq_lock); + req->rq_resend = 0; + spin_unlock(&req->rq_lock); + + LASSERT(imp->imp_replayable); + /* Balanced in ptlrpc_free_committed, usually. */ + ptlrpc_request_addref(req); + list_for_each_prev(tmp, &imp->imp_replay_list) { + struct ptlrpc_request *iter = list_entry(tmp, + struct ptlrpc_request, + rq_replay_list); /* We may have duplicate transnos if we create and then * open a file, or for closes retained if to match creating @@ -2658,13 +2888,12 @@ void ptlrpc_retain_replayable_request(struct ptlrpc_request *req, continue; } - cfs_list_add(&req->rq_replay_list, &iter->rq_replay_list); - return; - } + list_add(&req->rq_replay_list, &iter->rq_replay_list); + return; + } - cfs_list_add(&req->rq_replay_list, &imp->imp_replay_list); + list_add(&req->rq_replay_list, &imp->imp_replay_list); } -EXPORT_SYMBOL(ptlrpc_retain_replayable_request); /** * Send request and wait until it completes. @@ -2679,11 +2908,11 @@ int ptlrpc_queue_wait(struct ptlrpc_request *req) LASSERT(req->rq_set == NULL); LASSERT(!req->rq_receiving_reply); - set = ptlrpc_prep_set(); - if (set == NULL) { - CERROR("Unable to allocate ptlrpc set."); - RETURN(-ENOMEM); - } + set = ptlrpc_prep_set(); + if (set == NULL) { + CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM); + RETURN(-ENOMEM); + } /* for distributed debugging */ lustre_msg_set_status(req->rq_reqmsg, current_pid()); @@ -2698,14 +2927,9 @@ int ptlrpc_queue_wait(struct ptlrpc_request *req) } EXPORT_SYMBOL(ptlrpc_queue_wait); -struct ptlrpc_replay_async_args { - int praa_old_state; - int praa_old_status; -}; - /** * Callback used for replayed requests reply processing. - * In case of succesful reply calls registeresd request replay callback. + * In case of successful reply calls registered request replay callback. * In case of error restart replay process. */ static int ptlrpc_replay_interpret(const struct lu_env *env, @@ -2718,10 +2942,15 @@ static int ptlrpc_replay_interpret(const struct lu_env *env, ENTRY; atomic_dec(&imp->imp_replay_inflight); - if (!ptlrpc_client_replied(req)) { - CERROR("request replay timed out, restarting recovery\n"); - GOTO(out, rc = -ETIMEDOUT); - } + /* Note: if it is bulk replay (MDS-MDS replay), then even if + * server got the request, but bulk transfer timeout, let's + * replay the bulk req again */ + if (!ptlrpc_client_replied(req) || + (req->rq_bulk != NULL && + lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) { + DEBUG_REQ(D_ERROR, req, "request replay timed out.\n"); + GOTO(out, rc = -ETIMEDOUT); + } if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR && (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN || @@ -2742,7 +2971,7 @@ static int ptlrpc_replay_interpret(const struct lu_env *env, LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) == lustre_msg_get_transno(req->rq_repmsg) || lustre_msg_get_transno(req->rq_repmsg) == 0, - LPX64"/"LPX64"\n", + "%#llx/%#llx\n", lustre_msg_get_transno(req->rq_reqmsg), lustre_msg_get_transno(req->rq_repmsg)); } @@ -2758,8 +2987,8 @@ static int ptlrpc_replay_interpret(const struct lu_env *env, /* transaction number shouldn't be bigger than the latest replayed */ if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) { DEBUG_REQ(D_ERROR, req, - "Reported transno "LPU64" is bigger than the " - "replayed one: "LPU64, req->rq_transno, + "Reported transno %llu is bigger than the " + "replayed one: %llu", req->rq_transno, lustre_msg_get_transno(req->rq_reqmsg)); GOTO(out, rc = -EINVAL); } @@ -2775,6 +3004,49 @@ static int ptlrpc_replay_interpret(const struct lu_env *env, DEBUG_REQ(D_ERROR, req, "status %d, old was %d", lustre_msg_get_status(req->rq_repmsg), aa->praa_old_status); + + /* Note: If the replay fails for MDT-MDT recovery, let's + * abort all of the following requests in the replay + * and sending list, because MDT-MDT update requests + * are dependent on each other, see LU-7039 */ + if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) { + struct ptlrpc_request *free_req; + struct ptlrpc_request *tmp; + + spin_lock(&imp->imp_lock); + list_for_each_entry_safe(free_req, tmp, + &imp->imp_replay_list, + rq_replay_list) { + ptlrpc_free_request(free_req); + } + + list_for_each_entry_safe(free_req, tmp, + &imp->imp_committed_list, + rq_replay_list) { + ptlrpc_free_request(free_req); + } + + list_for_each_entry_safe(free_req, tmp, + &imp->imp_delayed_list, + rq_list) { + spin_lock(&free_req->rq_lock); + free_req->rq_err = 1; + free_req->rq_status = -EIO; + ptlrpc_client_wake_req(free_req); + spin_unlock(&free_req->rq_lock); + } + + list_for_each_entry_safe(free_req, tmp, + &imp->imp_sending_list, + rq_list) { + spin_lock(&free_req->rq_lock); + free_req->rq_err = 1; + free_req->rq_status = -EIO; + ptlrpc_client_wake_req(free_req); + spin_unlock(&free_req->rq_lock); + } + spin_unlock(&imp->imp_lock); + } } else { /* Put it back for re-replay. */ lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status); @@ -2834,34 +3106,37 @@ int ptlrpc_replay_req(struct ptlrpc_request *req) DEBUG_REQ(D_HA, req, "REPLAY"); atomic_inc(&req->rq_import->imp_replay_inflight); + spin_lock(&req->rq_lock); + req->rq_early_free_repbuf = 0; + spin_unlock(&req->rq_lock); ptlrpc_request_addref(req); /* ptlrpcd needs a ref */ - ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1); + ptlrpcd_add_req(req); RETURN(0); } -EXPORT_SYMBOL(ptlrpc_replay_req); /** * Aborts all in-flight request on import \a imp sending and delayed lists */ void ptlrpc_abort_inflight(struct obd_import *imp) { - cfs_list_t *tmp, *n; - ENTRY; + struct list_head *tmp, *n; + ENTRY; - /* Make sure that no new requests get processed for this import. - * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing - * this flag and then putting requests on sending_list or delayed_list. - */ + /* Make sure that no new requests get processed for this import. + * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing + * this flag and then putting requests on sending_list or delayed_list. + */ spin_lock(&imp->imp_lock); - /* XXX locking? Maybe we should remove each request with the list - * locked? Also, how do we know if the requests on the list are - * being freed at this time? - */ - cfs_list_for_each_safe(tmp, n, &imp->imp_sending_list) { - struct ptlrpc_request *req = - cfs_list_entry(tmp, struct ptlrpc_request, rq_list); + /* XXX locking? Maybe we should remove each request with the list + * locked? Also, how do we know if the requests on the list are + * being freed at this time? + */ + list_for_each_safe(tmp, n, &imp->imp_sending_list) { + struct ptlrpc_request *req = list_entry(tmp, + struct ptlrpc_request, + rq_list); DEBUG_REQ(D_RPCTRACE, req, "inflight"); @@ -2874,9 +3149,9 @@ void ptlrpc_abort_inflight(struct obd_import *imp) spin_unlock(&req->rq_lock); } - cfs_list_for_each_safe(tmp, n, &imp->imp_delayed_list) { + list_for_each_safe(tmp, n, &imp->imp_delayed_list) { struct ptlrpc_request *req = - cfs_list_entry(tmp, struct ptlrpc_request, rq_list); + list_entry(tmp, struct ptlrpc_request, rq_list); DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req"); @@ -2898,21 +3173,20 @@ void ptlrpc_abort_inflight(struct obd_import *imp) EXIT; } -EXPORT_SYMBOL(ptlrpc_abort_inflight); /** * Abort all uncompleted requests in request set \a set */ void ptlrpc_abort_set(struct ptlrpc_request_set *set) { - cfs_list_t *tmp, *pos; + struct list_head *tmp, *pos; - LASSERT(set != NULL); + LASSERT(set != NULL); - cfs_list_for_each_safe(pos, tmp, &set->set_requests) { - struct ptlrpc_request *req = - cfs_list_entry(pos, struct ptlrpc_request, - rq_set_chain); + list_for_each_safe(pos, tmp, &set->set_requests) { + struct ptlrpc_request *req = + list_entry(pos, struct ptlrpc_request, + rq_set_chain); spin_lock(&req->rq_lock); if (req->rq_phase != RQ_PHASE_RPC) { @@ -2948,7 +3222,7 @@ static spinlock_t ptlrpc_last_xid_lock; #define YEAR_2004 (1ULL << 30) void ptlrpc_init_xid(void) { - time_t now = cfs_time_current_sec(); + time64_t now = ktime_get_real_seconds(); spin_lock_init(&ptlrpc_last_xid_lock); if (now < YEAR_2004) { @@ -2987,7 +3261,66 @@ __u64 ptlrpc_next_xid(void) return next; } -EXPORT_SYMBOL(ptlrpc_next_xid); + +/** + * If request has a new allocated XID (new request or EINPROGRESS resend), + * use this XID as matchbits of bulk, otherwise allocate a new matchbits for + * request to ensure previous bulk fails and avoid problems with lost replies + * and therefore several transfers landing into the same buffer from different + * sending attempts. + */ +void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req) +{ + struct ptlrpc_bulk_desc *bd = req->rq_bulk; + + LASSERT(bd != NULL); + + /* Generate new matchbits for all resend requests, including + * resend replay. */ + if (req->rq_resend) { + __u64 old_mbits = req->rq_mbits; + + /* First time resend on -EINPROGRESS will generate new xid, + * so we can actually use the rq_xid as rq_mbits in such case, + * however, it's bit hard to distinguish such resend with a + * 'resend for the -EINPROGRESS resend'. To make it simple, + * we opt to generate mbits for all resend cases. */ + if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS)){ + req->rq_mbits = ptlrpc_next_xid(); + } else { + /* Old version transfers rq_xid to peer as + * matchbits. */ + spin_lock(&req->rq_import->imp_lock); + list_del_init(&req->rq_unreplied_list); + ptlrpc_assign_next_xid_nolock(req); + spin_unlock(&req->rq_import->imp_lock); + req->rq_mbits = req->rq_xid; + } + CDEBUG(D_HA, "resend bulk old x%llu new x%llu\n", + old_mbits, req->rq_mbits); + } else if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) { + /* Request being sent first time, use xid as matchbits. */ + req->rq_mbits = req->rq_xid; + } else { + /* Replay request, xid and matchbits have already been + * correctly assigned. */ + return; + } + + /* For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so + * that server can infer the number of bulks that were prepared, + * see LU-1431 */ + req->rq_mbits += ((bd->bd_iov_count + LNET_MAX_IOV - 1) / + LNET_MAX_IOV) - 1; + + /* Set rq_xid as rq_mbits to indicate the final bulk for the old + * server which does not support OBD_CONNECT_BULK_MBITS. LU-6808. + * + * It's ok to directly set the rq_xid here, since this xid bump + * won't affect the request position in unreplied list. */ + if (!OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS)) + req->rq_xid = req->rq_mbits; +} /** * Get a glimpse at what next xid value might have been. @@ -3037,15 +3370,14 @@ static void ptlrpcd_add_work_req(struct ptlrpc_request *req) { /* re-initialize the req */ req->rq_timeout = obd_timeout; - req->rq_sent = cfs_time_current_sec(); + req->rq_sent = ktime_get_real_seconds(); req->rq_deadline = req->rq_sent + req->rq_timeout; - req->rq_reply_deadline = req->rq_deadline; req->rq_phase = RQ_PHASE_INTERPRET; req->rq_next_phase = RQ_PHASE_COMPLETE; req->rq_xid = ptlrpc_next_xid(); req->rq_import_generation = req->rq_import->imp_generation; - ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1); + ptlrpcd_add_req(req); } static int work_interpreter(const struct lu_env *env, @@ -3081,7 +3413,7 @@ static int ptlrpcd_check_work(struct ptlrpc_request *req) void *ptlrpcd_alloc_work(struct obd_import *imp, int (*cb)(const struct lu_env *, void *), void *cbdata) { - struct ptlrpc_request *req = NULL; + struct ptlrpc_request *req = NULL; struct ptlrpc_work_async_args *args; ENTRY; @@ -3097,27 +3429,16 @@ void *ptlrpcd_alloc_work(struct obd_import *imp, RETURN(ERR_PTR(-ENOMEM)); } + ptlrpc_cli_req_init(req); + req->rq_send_state = LUSTRE_IMP_FULL; req->rq_type = PTL_RPC_MSG_REQUEST; req->rq_import = class_import_get(imp); - req->rq_export = NULL; req->rq_interpret_reply = work_interpreter; /* don't want reply */ - req->rq_receiving_reply = 0; - req->rq_must_unlink = 0; req->rq_no_delay = req->rq_no_resend = 1; req->rq_pill.rc_fmt = (void *)&worker_format; - spin_lock_init(&req->rq_lock); - CFS_INIT_LIST_HEAD(&req->rq_list); - CFS_INIT_LIST_HEAD(&req->rq_replay_list); - CFS_INIT_LIST_HEAD(&req->rq_set_chain); - CFS_INIT_LIST_HEAD(&req->rq_history_list); - CFS_INIT_LIST_HEAD(&req->rq_exp_list); - init_waitqueue_head(&req->rq_reply_waitq); - init_waitqueue_head(&req->rq_set_waitq); - atomic_set(&req->rq_refcount, 1); - CLASSERT (sizeof(*args) <= sizeof(req->rq_async_args)); args = ptlrpc_req_async_args(req); args->cb = cb;