X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fptlrpc%2Fclient.c;h=d920faa031d17bc665d7232a2df5285c97cf3cac;hp=cec226ddad5a7fc65c84c01a4849d027f40c1116;hb=d1dcff3084e929f5768dc733cdc104cddf168c06;hpb=a46519b0971baeccc498594b2ec0e3f8c4304c82 diff --git a/lustre/ptlrpc/client.c b/lustre/ptlrpc/client.c index cec226d..d920faa 100644 --- a/lustre/ptlrpc/client.c +++ b/lustre/ptlrpc/client.c @@ -1,6 +1,4 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: - * +/* * GPL HEADER START * * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. @@ -26,14 +24,18 @@ * GPL HEADER END */ /* - * Copyright 2008 Sun Microsystems, Inc. All rights reserved + * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. + * + * Copyright (c) 2011, 2013, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ * Lustre is a trademark of Sun Microsystems, Inc. */ +/** Implementation of client-side PortalRPC interfaces */ + #define DEBUG_SUBSYSTEM S_RPC #ifndef __KERNEL__ #include @@ -50,6 +52,12 @@ #include "ptlrpc_internal.h" +static int ptlrpc_send_new_req(struct ptlrpc_request *req); +static int ptlrpcd_check_work(struct ptlrpc_request *req); + +/** + * Initialize passed in client structure \a cl. + */ void ptlrpc_init_client(int req_portal, int rep_portal, char *name, struct ptlrpc_client *cl) { @@ -57,7 +65,11 @@ void ptlrpc_init_client(int req_portal, int rep_portal, char *name, cl->cli_reply_portal = rep_portal; cl->cli_name = name; } +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 *c; @@ -65,9 +77,12 @@ struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid) lnet_process_id_t 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) { - CERROR("cannot find peer %s!\n", uuid->uuid); + CNETERR("cannot find peer %s!\n", uuid->uuid); return NULL; } @@ -81,37 +96,58 @@ struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid) return c; } +EXPORT_SYMBOL(ptlrpc_uuid_to_connection); -static inline struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal) +/** + * 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 *desc; + struct ptlrpc_bulk_desc *desc; + int i; - OBD_ALLOC(desc, offsetof (struct ptlrpc_bulk_desc, bd_iov[npages])); - if (!desc) - return NULL; + OBD_ALLOC(desc, offsetof(struct ptlrpc_bulk_desc, bd_iov[npages])); + if (!desc) + return NULL; - spin_lock_init(&desc->bd_lock); - cfs_waitq_init(&desc->bd_waitq); - desc->bd_max_iov = npages; - desc->bd_iov_count = 0; - LNetInvalidateHandle(&desc->bd_md_h); - desc->bd_portal = portal; - desc->bd_type = type; + spin_lock_init(&desc->bd_lock); + init_waitqueue_head(&desc->bd_waitq); + desc->bd_max_iov = npages; + desc->bd_iov_count = 0; + desc->bd_portal = portal; + desc->bd_type = type; + desc->bd_md_count = 0; + 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]); - return desc; + return desc; } +/** + * Prepare bulk descriptor for specified outgoing request \a req that + * can fit \a npages * 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, - int npages, int type, int portal) + unsigned npages, unsigned max_brw, + unsigned type, unsigned portal) { - struct obd_import *imp = req->rq_import; - struct ptlrpc_bulk_desc *desc; + struct obd_import *imp = req->rq_import; + struct ptlrpc_bulk_desc *desc; - ENTRY; - LASSERT(type == BULK_PUT_SINK || type == BULK_GET_SOURCE); - desc = new_bulk(npages, type, portal); - if (desc == NULL) - RETURN(NULL); + ENTRY; + LASSERT(type == BULK_PUT_SINK || type == BULK_GET_SOURCE); + desc = ptlrpc_new_bulk(npages, max_brw, type, portal); + if (desc == NULL) + RETURN(NULL); desc->bd_import_generation = req->rq_import_generation; desc->bd_import = class_import_get(imp); @@ -125,68 +161,67 @@ struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req, return desc; } +EXPORT_SYMBOL(ptlrpc_prep_bulk_imp); -struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req, - int npages, int type, int portal) +/* + * 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 obd_export *exp = req->rq_export; - struct ptlrpc_bulk_desc *desc; - - ENTRY; - LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK); - - desc = new_bulk(npages, type, portal); - if (desc == NULL) - RETURN(NULL); - - desc->bd_export = class_export_get(exp); - desc->bd_req = req; + LASSERT(desc->bd_iov_count < desc->bd_max_iov); + LASSERT(page != NULL); + LASSERT(pageoffset >= 0); + LASSERT(len > 0); + LASSERT(pageoffset + len <= PAGE_CACHE_SIZE); - desc->bd_cbid.cbid_fn = server_bulk_callback; - desc->bd_cbid.cbid_arg = desc; + desc->bd_nob += len; - /* NB we don't assign rq_bulk here; server-side requests are - * re-used, and the handler frees the bulk desc explicitly. */ + if (pin) + page_cache_get(page); - return desc; + ptlrpc_add_bulk_page(desc, page, pageoffset, len); } +EXPORT_SYMBOL(__ptlrpc_prep_bulk_page); -void ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc, - cfs_page_t *page, int pageoffset, int len) +/** + * 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) { - LASSERT(desc->bd_iov_count < desc->bd_max_iov); - LASSERT(page != NULL); - LASSERT(pageoffset >= 0); - LASSERT(len > 0); - LASSERT(pageoffset + len <= CFS_PAGE_SIZE); + int i; + ENTRY; - desc->bd_nob += len; - - ptlrpc_add_bulk_page(desc, page, pageoffset, len); -} - -void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc) -{ - 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 != NULL); - LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */ - LASSERT(!desc->bd_network_rw); /* network hands off or */ - LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL)); + sptlrpc_enc_pool_put_pages(desc); - sptlrpc_enc_pool_put_pages(desc); + if (desc->bd_export) + class_export_put(desc->bd_export); + else + class_import_put(desc->bd_import); - 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); + } - OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc, - bd_iov[desc->bd_max_iov])); - EXIT; + OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc, + bd_iov[desc->bd_max_iov])); + EXIT; } +EXPORT_SYMBOL(__ptlrpc_free_bulk); -/* Set server timelimit for this req */ +/** + * Set server timelimit for this req, i.e. how long are we willing to wait + * for reply before timing out this request. + */ void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req) { __u32 serv_est; @@ -197,17 +232,22 @@ void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req) if (AT_OFF) { /* non-AT settings */ + /** + * \a imp_server_timeout means this is reverse import and + * we send (currently only) ASTs to the client and cannot afford + * to wait too long for the reply, otherwise the other client + * (because of which we are sending this request) would + * timeout waiting for us + */ req->rq_timeout = req->rq_import->imp_server_timeout ? - obd_timeout / 2 : obd_timeout; - lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout); - return; + obd_timeout / 2 : obd_timeout; + } else { + at = &req->rq_import->imp_at; + idx = import_at_get_index(req->rq_import, + req->rq_request_portal); + serv_est = at_get(&at->iat_service_estimate[idx]); + req->rq_timeout = at_est2timeout(serv_est); } - - at = &req->rq_import->imp_at; - idx = import_at_get_index(req->rq_import, - req->rq_request_portal); - serv_est = at_get(&at->iat_service_estimate[idx]); - req->rq_timeout = at_est2timeout(serv_est); /* We could get even fancier here, using history to predict increased loading... */ @@ -215,6 +255,7 @@ void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req) reqmsg*/ lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout); } +EXPORT_SYMBOL(ptlrpc_at_set_req_timeout); /* Adjust max service estimate based on server value */ static void ptlrpc_at_adj_service(struct ptlrpc_request *req, @@ -224,18 +265,13 @@ static void ptlrpc_at_adj_service(struct ptlrpc_request *req, unsigned int oldse; struct imp_at *at; - /* do estimate only if is not in recovery */ - if ((req->rq_send_state != LUSTRE_IMP_FULL) && - (req->rq_send_state != LUSTRE_IMP_CONNECTING)) - return; - LASSERT(req->rq_import); at = &req->rq_import->imp_at; idx = import_at_get_index(req->rq_import, req->rq_request_portal); /* max service estimates are tracked on the server side, so just keep minimal history here */ - oldse = at_add(&at->iat_service_estimate[idx], serv_est); + oldse = at_measured(&at->iat_service_estimate[idx], serv_est); if (oldse != 0) CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d " "has changed from %d to %d\n", @@ -267,7 +303,7 @@ static void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req, CFS_DURATION_T"\n", service_time, cfs_time_sub(now, req->rq_sent)); - oldnl = at_add(&at->iat_net_latency, nl); + oldnl = at_measured(&at->iat_net_latency, nl); if (oldnl != 0) CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) " "has changed from %d to %d\n", @@ -281,13 +317,12 @@ static int unpack_reply(struct ptlrpc_request *req) { int rc; - /* Clear reply swab mask; we may have already swabbed an early reply */ - req->rq_rep_swab_mask = 0; - - rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen); - if (rc) { - DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc); - return(-EPROTO); + if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) { + rc = ptlrpc_unpack_rep_msg(req, req->rq_replen); + if (rc) { + DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc); + return(-EPROTO); + } } rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF); @@ -298,7 +333,7 @@ static int unpack_reply(struct ptlrpc_request *req) return 0; } -/* +/** * Handle an early reply message, called with the rq_lock held. * If anything goes wrong just ignore it - same as if it never happened */ @@ -310,11 +345,11 @@ static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req) ENTRY; req->rq_early = 0; - spin_unlock(&req->rq_lock); + spin_unlock(&req->rq_lock); - rc = sptlrpc_cli_unwrap_early_reply(req, &early_req); - if (rc) { - spin_lock(&req->rq_lock); + rc = sptlrpc_cli_unwrap_early_reply(req, &early_req); + if (rc) { + spin_lock(&req->rq_lock); RETURN(rc); } @@ -329,86 +364,133 @@ static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req) sptlrpc_cli_finish_early_reply(early_req); - spin_lock(&req->rq_lock); + if (rc != 0) { + spin_lock(&req->rq_lock); + RETURN(rc); + } - if (rc == 0) { - /* Adjust the local timeout for this req */ - ptlrpc_at_set_req_timeout(req); - - 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 + - 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)); - } + /* Adjust the local timeout for this req */ + ptlrpc_at_set_req_timeout(req); - RETURN(rc); + 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 + + 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)); + + RETURN(rc); +} + +struct kmem_cache *request_cache; + +int ptlrpc_request_cache_init() +{ + request_cache = kmem_cache_create("ptlrpc_cache", + sizeof(struct ptlrpc_request), + 0, SLAB_HWCACHE_ALIGN, NULL); + return request_cache == NULL ? -ENOMEM : 0; +} + +void ptlrpc_request_cache_fini() +{ + kmem_cache_destroy(request_cache); } +struct ptlrpc_request *ptlrpc_request_cache_alloc(int flags) +{ + struct ptlrpc_request *req; + + OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags); + return req; +} + +void ptlrpc_request_cache_free(struct ptlrpc_request *req) +{ + OBD_SLAB_FREE_PTR(req, request_cache); +} + +/** + * Wind down request pool \a pool. + * Frees all requests from the pool too + */ void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool) { - struct list_head *l, *tmp; - struct ptlrpc_request *req; + cfs_list_t *l, *tmp; + struct ptlrpc_request *req; - LASSERT(pool != NULL); + LASSERT(pool != NULL); - spin_lock(&pool->prp_lock); - 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(req->rq_reqbuf, pool->prp_rq_size); - OBD_FREE(req, sizeof(*req)); - } - spin_unlock(&pool->prp_lock); - OBD_FREE(pool, sizeof(*pool)); + 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); + LASSERT(req->rq_reqbuf); + LASSERT(req->rq_reqbuf_len == pool->prp_rq_size); + OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size); + ptlrpc_request_cache_free(req); + } + spin_unlock(&pool->prp_lock); + OBD_FREE(pool, sizeof(*pool)); } +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 i; int size = 1; - while (size < pool->prp_rq_size + SPTLRPC_MAX_PAYLOAD) + while (size < pool->prp_rq_size) size <<= 1; - LASSERTF(list_empty(&pool->prp_req_list) || size == pool->prp_rq_size, + LASSERTF(cfs_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); - spin_lock(&pool->prp_lock); - pool->prp_rq_size = size; - for (i = 0; i < num_rq; i++) { - struct ptlrpc_request *req; - struct lustre_msg *msg; - - spin_unlock(&pool->prp_lock); - OBD_ALLOC(req, sizeof(struct ptlrpc_request)); - if (!req) - return; - OBD_ALLOC_GFP(msg, size, CFS_ALLOC_STD); - if (!msg) { - OBD_FREE(req, sizeof(struct ptlrpc_request)); - return; + spin_lock(&pool->prp_lock); + pool->prp_rq_size = size; + for (i = 0; i < num_rq; i++) { + struct ptlrpc_request *req; + struct lustre_msg *msg; + + spin_unlock(&pool->prp_lock); + req = ptlrpc_request_cache_alloc(__GFP_IO); + if (!req) + return; + 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; - spin_lock(&pool->prp_lock); - list_add_tail(&req->rq_list, &pool->prp_req_list); - } - spin_unlock(&pool->prp_lock); - return; + spin_lock(&pool->prp_lock); + cfs_list_add_tail(&req->rq_list, &pool->prp_req_list); + } + spin_unlock(&pool->prp_lock); + return; } +EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool); +/** + * Create and initialize new request pool with given attributes: + * \a num_rq - initial number of requests to create for the pool + * \a msgsize - maximum message size possible for requests in thid pool + * \a populate_pool - function to be called when more requests need to be added + * to the pool + * Returns pointer to newly created pool or NULL on error. + */ struct ptlrpc_request_pool * ptlrpc_init_rq_pool(int num_rq, int msgsize, void (*populate_pool)(struct ptlrpc_request_pool *, int)) @@ -422,21 +504,25 @@ ptlrpc_init_rq_pool(int num_rq, int msgsize, /* Request next power of two for the allocation, because internally kernel would do exactly this */ - spin_lock_init(&pool->prp_lock); + spin_lock_init(&pool->prp_lock); CFS_INIT_LIST_HEAD(&pool->prp_req_list); - pool->prp_rq_size = msgsize; + pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD; pool->prp_populate = populate_pool; populate_pool(pool, num_rq); - if (list_empty(&pool->prp_req_list)) { + 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; } +EXPORT_SYMBOL(ptlrpc_init_rq_pool); +/** + * Fetches one request from pool \a pool + */ static struct ptlrpc_request * ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool) { @@ -446,21 +532,21 @@ ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool) if (!pool) return NULL; - spin_lock(&pool->prp_lock); + spin_lock(&pool->prp_lock); - /* See if we have anything in a pool, and bail out if nothing, - * 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(list_empty(&pool->prp_req_list))) { - spin_unlock(&pool->prp_lock); - return NULL; - } + /* See if we have anything in a pool, and bail out if nothing, + * 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))) { + spin_unlock(&pool->prp_lock); + return NULL; + } - request = list_entry(pool->prp_req_list.next, struct ptlrpc_request, - rq_list); - list_del_init(&request->rq_list); - spin_unlock(&pool->prp_lock); + request = cfs_list_entry(pool->prp_req_list.next, struct ptlrpc_request, + rq_list); + cfs_list_del_init(&request->rq_list); + spin_unlock(&pool->prp_lock); LASSERT(request->rq_reqbuf); LASSERT(request->rq_pool); @@ -474,15 +560,18 @@ ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool) return request; } +/** + * Returns freed \a request to pool. + */ static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request) { - struct ptlrpc_request_pool *pool = request->rq_pool; + struct ptlrpc_request_pool *pool = request->rq_pool; - spin_lock(&pool->prp_lock); - LASSERT(list_empty(&request->rq_list)); - LASSERT(!request->rq_receiving_reply); - list_add_tail(&request->rq_list, &pool->prp_req_list); - spin_unlock(&pool->prp_lock); + spin_lock(&pool->prp_lock); + LASSERT(cfs_list_empty(&request->rq_list)); + LASSERT(!request->rq_receiving_reply); + cfs_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, @@ -531,26 +620,27 @@ static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request, ptlrpc_at_set_req_timeout(request); - 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); - cfs_waitq_init(&request->rq_reply_waitq); - request->rq_xid = ptlrpc_next_xid(); - atomic_set(&request->rq_refcount, 1); - - lustre_msg_set_opc(request->rq_reqmsg, opcode); - - RETURN(0); + 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(); + cfs_atomic_set(&request->rq_refcount, 1); + + lustre_msg_set_opc(request->rq_reqmsg, opcode); + + RETURN(0); out_ctx: - sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1); + sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1); out_free: - class_import_put(imp); - return rc; + class_import_put(imp); + return rc; } int ptlrpc_request_bufs_pack(struct ptlrpc_request *request, @@ -566,23 +656,57 @@ int ptlrpc_request_bufs_pack(struct ptlrpc_request *request, } EXPORT_SYMBOL(ptlrpc_request_bufs_pack); +/** + * Pack request buffers for network transfer, performing necessary encryption + * steps if necessary. + */ int ptlrpc_request_pack(struct ptlrpc_request *request, __u32 version, int opcode) { - return ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL); -} + int rc; + rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL); + if (rc) + return rc; + + /* For some old 1.8 clients (< 1.8.7), they will LASSERT the size of + * ptlrpc_body sent from server equal to local ptlrpc_body size, so we + * have to send old ptlrpc_body to keep interoprability with these + * clients. + * + * Only three kinds of server->client RPCs so far: + * - LDLM_BL_CALLBACK + * - LDLM_CP_CALLBACK + * - LDLM_GL_CALLBACK + * + * XXX This should be removed whenever we drop the interoprability with + * the these old clients. + */ + if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK || + opcode == LDLM_GL_CALLBACK) + req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY, + sizeof(struct ptlrpc_body_v2), RCL_CLIENT); + + return rc; +} +EXPORT_SYMBOL(ptlrpc_request_pack); +/** + * Helper function to allocate new request on import \a imp + * and possibly using existing request from pool \a pool if provided. + * Returns allocated request structure with import field filled or + * NULL on error. + */ static inline struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp, struct ptlrpc_request_pool *pool) { - struct ptlrpc_request *request = NULL; + struct ptlrpc_request *request = NULL; - if (pool) - request = ptlrpc_prep_req_from_pool(pool); + if (pool) + request = ptlrpc_prep_req_from_pool(pool); - if (!request) - OBD_ALLOC_PTR(request); + if (!request) + request = ptlrpc_request_cache_alloc(__GFP_IO); if (request) { LASSERTF((unsigned long)imp > 0x1000, "%p", imp); @@ -599,6 +723,12 @@ struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp, return request; } +/** + * Helper function for creating a request. + * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits + * buffer structures according to capsule template \a format. + * Returns allocated request structure pointer or NULL on error. + */ static struct ptlrpc_request * ptlrpc_request_alloc_internal(struct obd_import *imp, struct ptlrpc_request_pool * pool, @@ -615,27 +745,49 @@ ptlrpc_request_alloc_internal(struct obd_import *imp, return request; } +/** + * Allocate new request structure for import \a imp and initialize its + * buffer structure according to capsule template \a format. + */ struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp, const struct req_format *format) { return ptlrpc_request_alloc_internal(imp, NULL, format); } +EXPORT_SYMBOL(ptlrpc_request_alloc); +/** + * Allocate new request structure for import \a imp from pool \a pool and + * initialize its buffer structure according to capsule template \a format. + */ struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp, struct ptlrpc_request_pool * pool, const struct req_format *format) { return ptlrpc_request_alloc_internal(imp, pool, format); } +EXPORT_SYMBOL(ptlrpc_request_alloc_pool); +/** + * For requests not from pool, free memory of the request structure. + * For requests obtained from a pool earlier, return request back to pool. + */ void ptlrpc_request_free(struct ptlrpc_request *request) { - if (request->rq_pool) - __ptlrpc_free_req_to_pool(request); - else - OBD_FREE_PTR(request); + if (request->rq_pool) + __ptlrpc_free_req_to_pool(request); + else + ptlrpc_request_cache_free(request); } +EXPORT_SYMBOL(ptlrpc_request_free); +/** + * Allocate new request for operatione \a opcode and immediatelly pack it for + * network transfer. + * Only used for simple requests like OBD_PING where the only important + * part of the request is operation itself. + * Returns allocated request or NULL on error. + */ struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp, const struct req_format *format, __u32 version, int opcode) @@ -652,7 +804,15 @@ struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp, } return req; } +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, @@ -674,7 +834,11 @@ ptlrpc_prep_req_pool(struct obd_import *imp, } 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) @@ -682,131 +846,128 @@ ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count, return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs, NULL); } +EXPORT_SYMBOL(ptlrpc_prep_req); -struct ptlrpc_request *ptlrpc_prep_fakereq(struct obd_import *imp, - unsigned int timeout, - ptlrpc_interpterer_t interpreter) +/** + * Allocate and initialize new request set structure. + * Returns a pointer to the newly allocated set structure or NULL on error. + */ +struct ptlrpc_request_set *ptlrpc_prep_set(void) { - struct ptlrpc_request *request = NULL; - ENTRY; - - OBD_ALLOC(request, sizeof(*request)); - if (!request) { - CERROR("request allocation out of memory\n"); - RETURN(NULL); - } - - request->rq_send_state = LUSTRE_IMP_FULL; - request->rq_type = PTL_RPC_MSG_REQUEST; - request->rq_import = class_import_get(imp); - request->rq_export = NULL; - request->rq_import_generation = imp->imp_generation; - - request->rq_timeout = timeout; - request->rq_sent = cfs_time_current_sec(); - request->rq_deadline = request->rq_sent + timeout; - request->rq_reply_deadline = request->rq_deadline; - request->rq_interpret_reply = interpreter; - request->rq_phase = RQ_PHASE_RPC; - request->rq_next_phase = RQ_PHASE_INTERPRET; - /* don't want reply */ - request->rq_receiving_reply = 0; - request->rq_must_unlink = 0; - request->rq_no_delay = request->rq_no_resend = 1; - request->rq_fake = 1; + struct ptlrpc_request_set *set; + + ENTRY; + OBD_ALLOC(set, sizeof *set); + if (!set) + RETURN(NULL); + cfs_atomic_set(&set->set_refcount, 1); + CFS_INIT_LIST_HEAD(&set->set_requests); + init_waitqueue_head(&set->set_waitq); + cfs_atomic_set(&set->set_new_count, 0); + cfs_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); + set->set_max_inflight = UINT_MAX; + set->set_producer = NULL; + set->set_producer_arg = NULL; + set->set_rc = 0; + + RETURN(set); +} +EXPORT_SYMBOL(ptlrpc_prep_set); - spin_lock_init(&request->rq_lock); - CFS_INIT_LIST_HEAD(&request->rq_list); - CFS_INIT_LIST_HEAD(&request->rq_replay_list); - CFS_INIT_LIST_HEAD(&request->rq_set_chain); - CFS_INIT_LIST_HEAD(&request->rq_history_list); - CFS_INIT_LIST_HEAD(&request->rq_exp_list); - cfs_waitq_init(&request->rq_reply_waitq); - - request->rq_xid = ptlrpc_next_xid(); - atomic_set(&request->rq_refcount, 1); - - RETURN(request); -} +/** + * Allocate and initialize new request set structure with flow control + * extension. This extension allows to control the number of requests in-flight + * for the whole set. A callback function to generate requests must be provided + * and the request set will keep the number of requests sent over the wire to + * @max_inflight. + * Returns a pointer to the newly allocated set structure or NULL on error. + */ +struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func, + void *arg) -void ptlrpc_fakereq_finished(struct ptlrpc_request *req) { - /* if we kill request before timeout - need adjust counter */ - if (req->rq_phase == RQ_PHASE_RPC) { - struct ptlrpc_request_set *set = req->rq_set; - - if (set) - set->set_remaining --; - } - - ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE); - list_del_init(&req->rq_list); -} - + struct ptlrpc_request_set *set; -struct ptlrpc_request_set *ptlrpc_prep_set(void) -{ - struct ptlrpc_request_set *set; + set = ptlrpc_prep_set(); + if (!set) + RETURN(NULL); - ENTRY; - OBD_ALLOC(set, sizeof *set); - if (!set) - RETURN(NULL); - CFS_INIT_LIST_HEAD(&set->set_requests); - cfs_waitq_init(&set->set_waitq); - 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); + set->set_max_inflight = max; + set->set_producer = func; + set->set_producer_arg = arg; - RETURN(set); + RETURN(set); } +EXPORT_SYMBOL(ptlrpc_prep_fcset); -/* Finish with this set; opposite of prep_set. */ +/** + * Wind down and free request set structure previously allocated with + * ptlrpc_prep_set. + * Ensures that all requests on the set have completed and removes + * all requests from the request list in a set. + * If any unsent request happen to be on the list, pretends that they got + * an error in flight and calls their completion handler. + */ void ptlrpc_set_destroy(struct ptlrpc_request_set *set) { - struct list_head *tmp; - struct list_head *next; + cfs_list_t *tmp; + cfs_list_t *next; int expected_phase; int n = 0; ENTRY; /* Requests on the set should either all be completed, or all be new */ - expected_phase = (set->set_remaining == 0) ? + expected_phase = (cfs_atomic_read(&set->set_remaining) == 0) ? RQ_PHASE_COMPLETE : RQ_PHASE_NEW; - list_for_each (tmp, &set->set_requests) { + cfs_list_for_each (tmp, &set->set_requests) { struct ptlrpc_request *req = - list_entry(tmp, struct ptlrpc_request, rq_set_chain); + cfs_list_entry(tmp, struct ptlrpc_request, + rq_set_chain); LASSERT(req->rq_phase == expected_phase); n++; } - LASSERTF(set->set_remaining == 0 || set->set_remaining == n, "%d / %d\n", - set->set_remaining, n); + LASSERTF(cfs_atomic_read(&set->set_remaining) == 0 || + cfs_atomic_read(&set->set_remaining) == n, "%d / %d\n", + cfs_atomic_read(&set->set_remaining), n); - list_for_each_safe(tmp, next, &set->set_requests) { + cfs_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); + cfs_list_entry(tmp, struct ptlrpc_request, + rq_set_chain); + cfs_list_del_init(&req->rq_set_chain); LASSERT(req->rq_phase == expected_phase); if (req->rq_phase == RQ_PHASE_NEW) { ptlrpc_req_interpret(NULL, req, -EBADR); - set->set_remaining--; + cfs_atomic_dec(&set->set_remaining); } - req->rq_set = NULL; + spin_lock(&req->rq_lock); + req->rq_set = NULL; + req->rq_invalid_rqset = 0; + spin_unlock(&req->rq_lock); + ptlrpc_req_finished (req); } - LASSERT(set->set_remaining == 0); + LASSERT(cfs_atomic_read(&set->set_remaining) == 0); - OBD_FREE(set, sizeof(*set)); + ptlrpc_reqset_put(set); EXIT; } +EXPORT_SYMBOL(ptlrpc_set_destroy); +/** + * Add a callback function \a fn to the set. + * This function would be called when all requests on this set are completed. + * The function will be passed \a data argument. + */ int ptlrpc_set_add_cb(struct ptlrpc_request_set *set, set_interpreter_func fn, void *data) { @@ -818,53 +979,75 @@ int ptlrpc_set_add_cb(struct ptlrpc_request_set *set, cbdata->psc_interpret = fn; cbdata->psc_data = data; - list_add_tail(&cbdata->psc_item, &set->set_cblist); + cfs_list_add_tail(&cbdata->psc_item, &set->set_cblist); RETURN(0); } +EXPORT_SYMBOL(ptlrpc_set_add_cb); +/** + * Add a new request to the general purpose request set. + * Assumes request reference from the caller. + */ void ptlrpc_set_add_req(struct ptlrpc_request_set *set, struct ptlrpc_request *req) { - /* The set takes over the caller's request reference */ - list_add_tail(&req->rq_set_chain, &set->set_requests); - req->rq_set = set; - set->set_remaining++; + LASSERT(cfs_list_empty(&req->rq_set_chain)); + + /* The set takes over the caller's request reference */ + cfs_list_add_tail(&req->rq_set_chain, &set->set_requests); + req->rq_set = set; + cfs_atomic_inc(&set->set_remaining); + req->rq_queued_time = cfs_time_current(); + + if (req->rq_reqmsg != NULL) + lustre_msg_set_jobid(req->rq_reqmsg, NULL); + + if (set->set_producer != NULL) + /* If the request set has a producer callback, the RPC must be + * sent straight away */ + ptlrpc_send_new_req(req); } +EXPORT_SYMBOL(ptlrpc_set_add_req); /** - * Lock so many callers can add things, the context that owns the set - * is supposed to notice these and move them into the set proper. + * Add a request to a request with dedicated server thread + * and wake the thread to make any necessary processing. + * Currently only used for ptlrpcd. */ -int ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc, +void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc, struct ptlrpc_request *req) { struct ptlrpc_request_set *set = pc->pc_set; + int count, i; - /* - * Let caller know that we stopped and will not handle this request. - * It needs to take care itself of request. - */ - if (test_bit(LIOD_STOP, &pc->pc_flags)) - return -EALREADY; - - spin_lock(&set->set_new_req_lock); - /* - * The set takes over the caller's request reference. - */ - list_add_tail(&req->rq_set_chain, &set->set_new_requests); - req->rq_set = set; - spin_unlock(&set->set_new_req_lock); - - /* - * Let thead know that we added something and better it to wake up - * and process. - */ - cfs_waitq_signal(&set->set_waitq); - return 0; -} + LASSERT(req->rq_set == NULL); + LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0); + + spin_lock(&set->set_new_req_lock); + /* + * The set takes over the caller's request reference. + */ + req->rq_set = set; + req->rq_queued_time = cfs_time_current(); + cfs_list_add_tail(&req->rq_set_chain, &set->set_new_requests); + count = cfs_atomic_inc_return(&set->set_new_count); + spin_unlock(&set->set_new_req_lock); + + /* Only need to call wakeup once for the first entry. */ + if (count == 1) { + wake_up(&set->set_waitq); + + /* XXX: It maybe unnecessary to wakeup all the partners. But to + * guarantee the async RPC can be processed ASAP, we have + * no other better choice. It maybe fixed in future. */ + for (i = 0; i < pc->pc_npartners; i++) + 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 * can be sent, is an error, or should be delayed. * @@ -888,90 +1071,64 @@ static int ptlrpc_import_delay_req(struct obd_import *imp, } else if (imp->imp_state == LUSTRE_IMP_NEW) { DEBUG_REQ(D_ERROR, req, "Uninitialized import."); *status = -EIO; - LBUG(); - } else if (imp->imp_state == LUSTRE_IMP_CLOSED) { - DEBUG_REQ(D_ERROR, req, "IMP_CLOSED "); + } else if (imp->imp_state == LUSTRE_IMP_CLOSED) { + /* pings may safely race with umount */ + DEBUG_REQ(lustre_msg_get_opc(req->rq_reqmsg) == OBD_PING ? + 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; } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING && imp->imp_state == LUSTRE_IMP_CONNECTING) { /* allow CONNECT even if import is invalid */ ; - if (atomic_read(&imp->imp_inval_count) != 0) { + if (cfs_atomic_read(&imp->imp_inval_count) != 0) { DEBUG_REQ(D_ERROR, req, "invalidate in flight"); *status = -EIO; } - } else if ((imp->imp_invalid && (!imp->imp_recon_bk)) || - imp->imp_obd->obd_no_recov) { - /* If the import has been invalidated (such as by an OST - * failure), and if the import(MGC) tried all of its connection - * list (Bug 13464), the request must fail with -ESHUTDOWN. - * This indicates the requests should be discarded; an -EIO - * may result in a resend of the request. */ - if (!imp->imp_deactive) - DEBUG_REQ(D_ERROR, req, "IMP_INVALID"); - *status = -ESHUTDOWN; /* bz 12940 */ - } else if (req->rq_import_generation != imp->imp_generation) { + } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) { + if (!imp->imp_deactive) + DEBUG_REQ(D_NET, req, "IMP_INVALID"); + *status = -ESHUTDOWN; /* bz 12940 */ + } else if (req->rq_import_generation != imp->imp_generation) { DEBUG_REQ(D_ERROR, req, "req wrong generation:"); *status = -EIO; } else if (req->rq_send_state != imp->imp_state) { /* invalidate in progress - any requests should be drop */ - if (atomic_read(&imp->imp_inval_count) != 0) { + if (cfs_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) { *status = -EWOULDBLOCK; - } else { - delay = 1; - } - } - - RETURN(delay); -} - -static int ptlrpc_check_reply(struct ptlrpc_request *req) -{ - int rc = 0; - ENTRY; - - /* serialise with network callback */ - spin_lock(&req->rq_lock); - - if (ptlrpc_client_replied(req)) - GOTO(out, rc = 1); - - if (req->rq_net_err && !req->rq_timedout) { - spin_unlock(&req->rq_lock); - rc = ptlrpc_expire_one_request(req, 0); - spin_lock(&req->rq_lock); - GOTO(out, rc); - } - - if (req->rq_err) - GOTO(out, rc = 1); - - if (req->rq_resend) - GOTO(out, rc = 1); + } else if (req->rq_allow_replay && + (imp->imp_state == LUSTRE_IMP_REPLAY || + imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS || + imp->imp_state == LUSTRE_IMP_REPLAY_WAIT || + imp->imp_state == LUSTRE_IMP_RECOVER)) { + DEBUG_REQ(D_HA, req, "allow during recovery.\n"); + } else { + delay = 1; + } + } - if (req->rq_restart) - GOTO(out, rc = 1); - - if (ptlrpc_client_early(req)) { - ptlrpc_at_recv_early_reply(req); - GOTO(out, rc = 0); /* keep waiting */ - } - - EXIT; - out: - spin_unlock(&req->rq_lock); - DEBUG_REQ(D_NET, req, "rc = %d for", rc); - return rc; + RETURN(delay); } -/* Conditionally suppress specific console messages */ +/** + * 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. + */ static int ptlrpc_console_allow(struct ptlrpc_request *req) { - __u32 opc = lustre_msg_get_opc(req->rq_reqmsg); + __u32 opc; int err; + 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)) && @@ -990,22 +1147,28 @@ static int ptlrpc_console_allow(struct ptlrpc_request *req) return 1; } - +/** + * Check request processing status. + * Returns the status. + */ static int ptlrpc_check_status(struct ptlrpc_request *req) { int err; ENTRY; 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; - __u32 opc = lustre_msg_get_opc(req->rq_reqmsg); - LCONSOLE_ERROR_MSG(0x011,"an error occurred while communicating" - " with %s. The %s operation failed with %d\n", - libcfs_nid2str(imp->imp_connection->c_peer.nid), - ll_opcode2str(opc), err); - RETURN(err < 0 ? err : -EINVAL); - } + if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) { + struct obd_import *imp = req->rq_import; + __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); + RETURN(err < 0 ? err : -EINVAL); + } if (err < 0) { DEBUG_REQ(D_INFO, req, "status is %d", err); @@ -1014,26 +1177,13 @@ static int ptlrpc_check_status(struct ptlrpc_request *req) DEBUG_REQ(D_INFO, req, "status is %d", err); } - if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) { - struct obd_import *imp = req->rq_import; - __u32 opc = lustre_msg_get_opc(req->rq_reqmsg); - - if (ptlrpc_console_allow(req)) - LCONSOLE_ERROR_MSG(0x011,"an error occurred while " - "communicating with %s. The %s " - "operation failed with %d\n", - libcfs_nid2str( - imp->imp_connection->c_peer.nid), - ll_opcode2str(opc), err); - - RETURN(err < 0 ? err : -EINVAL); - } - RETURN(err); } /** - * save pre-versions for replay + * save pre-versions of objects into request for replay. + * Versions are obtained from server reply. + * used for VBR. */ static void ptlrpc_save_versions(struct ptlrpc_request *req) { @@ -1055,6 +1205,10 @@ static void ptlrpc_save_versions(struct ptlrpc_request *req) /** * Callback function called when client receives RPC reply for \a req. + * Returns 0 on success or error code. + * The return alue would be assigned to req->rq_status by the caller + * as request processing status. + * This function also decides if the request needs to be saved for later replay. */ static int after_reply(struct ptlrpc_request *req) { @@ -1065,38 +1219,90 @@ static int after_reply(struct ptlrpc_request *req) long timediff; ENTRY; - LASSERT(!req->rq_receiving_reply); - LASSERT(obd); - LASSERT(req->rq_nob_received <= req->rq_repbuf_len); + LASSERT(obd != NULL); + /* repbuf must be unlinked */ + LASSERT(!req->rq_receiving_reply && !req->rq_must_unlink); + + if (req->rq_reply_truncate) { + if (ptlrpc_no_resend(req)) { + DEBUG_REQ(D_ERROR, req, "reply buffer overflow," + " expected: %d, actual size: %d", + req->rq_nob_received, req->rq_repbuf_len); + RETURN(-EOVERFLOW); + } + + sptlrpc_cli_free_repbuf(req); + /* Pass the required reply buffer size (include + * space for early reply). + * NB: no need to roundup because alloc_repbuf + * will roundup it */ + req->rq_replen = req->rq_nob_received; + req->rq_nob_received = 0; + spin_lock(&req->rq_lock); + req->rq_resend = 1; + spin_unlock(&req->rq_lock); + RETURN(0); + } /* * NB Until this point, the whole of the incoming message, * including buflens, status etc is in the sender's byte order. */ - rc = sptlrpc_cli_unwrap_reply(req); if (rc) { DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc); RETURN(rc); } - /* - * Security layer unwrap might ask resend this request. - */ - if (req->rq_resend) - RETURN(0); - - rc = unpack_reply(req); - if (rc) - RETURN(rc); - - 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); - ptlrpc_lprocfs_rpc_sent(req, timediff); - } + /* + * Security layer unwrap might ask resend this request. + */ + if (req->rq_resend) + RETURN(0); + + rc = unpack_reply(req); + if (rc) + RETURN(rc); + + /* 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(); + + DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS"); + req->rq_resend = 1; + 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. + * The delay is increased by 1s on every resend and is capped to + * the current request timeout (i.e. obd_timeout if AT is off, + * or AT service time x 125% + 5s, see at_est2timeout) */ + if (req->rq_nr_resend > req->rq_timeout) + req->rq_sent = now + req->rq_timeout; + else + req->rq_sent = now + req->rq_nr_resend; + + 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); + ptlrpc_lprocfs_rpc_sent(req, timediff); + } if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY && lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) { @@ -1105,7 +1311,8 @@ static int after_reply(struct ptlrpc_request *req) RETURN(-EPROTO); } - OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, obd_fail_val); + if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING) + CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val); ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg)); ptlrpc_at_adj_net_latency(req, lustre_msg_get_service_time(req->rq_repmsg)); @@ -1144,7 +1351,7 @@ static int after_reply(struct ptlrpc_request *req) } if (imp->imp_replayable) { - spin_lock(&imp->imp_lock); + spin_lock(&imp->imp_lock); /* * No point in adding already-committed requests to the replay * list, we will just remove them immediately. b=9829 @@ -1156,10 +1363,14 @@ static int after_reply(struct ptlrpc_request *req) /** version recovery */ ptlrpc_save_versions(req); ptlrpc_retain_replayable_request(req, imp); - } else if (req->rq_commit_cb != NULL) { - spin_unlock(&imp->imp_lock); - req->rq_commit_cb(req); - spin_lock(&imp->imp_lock); + } else if (req->rq_commit_cb != NULL && + list_empty(&req->rq_replay_list)) { + /* NB: don't call rq_commit_cb if it's already on + * rq_replay_list, ptlrpc_free_committed() will call + * it later, see LU-3618 for details */ + spin_unlock(&imp->imp_lock); + req->rq_commit_cb(req); + spin_lock(&imp->imp_lock); } /* @@ -1169,59 +1380,82 @@ static int after_reply(struct ptlrpc_request *req) imp->imp_peer_committed_transno = lustre_msg_get_last_committed(req->rq_repmsg); } - ptlrpc_free_committed(imp); - spin_unlock(&imp->imp_lock); - } - RETURN(rc); + ptlrpc_free_committed(imp); + + if (!cfs_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); + /* + * Requests with rq_replay stay on the list even if no + * commit is expected. + */ + if (last->rq_transno > imp->imp_peer_committed_transno) + ptlrpc_pinger_commit_expected(imp); + } + + spin_unlock(&imp->imp_lock); + } + + RETURN(rc); } +/** + * Helper function to send request \a req over the network for the first time + * Also adjusts request phase. + * Returns 0 on success or error code. + */ static int ptlrpc_send_new_req(struct ptlrpc_request *req) { - struct obd_import *imp; + struct obd_import *imp = req->rq_import; int rc; ENTRY; LASSERT(req->rq_phase == RQ_PHASE_NEW); - if (req->rq_sent && (req->rq_sent > cfs_time_current_sec())) + if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()) && + (!req->rq_generation_set || + req->rq_import_generation == imp->imp_generation)) RETURN (0); ptlrpc_rqphase_move(req, RQ_PHASE_RPC); - imp = req->rq_import; - spin_lock(&imp->imp_lock); - - req->rq_import_generation = imp->imp_generation; - - if (ptlrpc_import_delay_req(imp, req, &rc)) { - spin_lock(&req->rq_lock); - 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), - ptlrpc_import_state_name(req->rq_send_state), - ptlrpc_import_state_name(imp->imp_state)); - 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); - } - - if (rc != 0) { - spin_unlock(&imp->imp_lock); - req->rq_status = rc; - ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET); - RETURN(rc); - } - - LASSERT(list_empty(&req->rq_list)); - list_add_tail(&req->rq_list, &imp->imp_sending_list); - atomic_inc(&req->rq_import->imp_inflight); - spin_unlock(&imp->imp_lock); - - lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid()); + spin_lock(&imp->imp_lock); + + if (!req->rq_generation_set) + req->rq_import_generation = imp->imp_generation; + + if (ptlrpc_import_delay_req(imp, req, &rc)) { + spin_lock(&req->rq_lock); + 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), + 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); + cfs_atomic_inc(&req->rq_import->imp_inflight); + spin_unlock(&imp->imp_lock); + RETURN(0); + } + + if (rc != 0) { + spin_unlock(&imp->imp_lock); + req->rq_status = rc; + ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET); + RETURN(rc); + } + + LASSERT(cfs_list_empty(&req->rq_list)); + cfs_list_add_tail(&req->rq_list, &imp->imp_sending_list); + cfs_atomic_inc(&req->rq_import->imp_inflight); + spin_unlock(&imp->imp_lock); + + lustre_msg_set_status(req->rq_reqmsg, current_pid()); rc = sptlrpc_req_refresh_ctx(req, -1); if (rc) { @@ -1229,41 +1463,76 @@ static int ptlrpc_send_new_req(struct ptlrpc_request *req) req->rq_status = rc; RETURN(1); } else { - req->rq_wait_ctx = 1; + spin_lock(&req->rq_lock); + req->rq_wait_ctx = 1; + spin_unlock(&req->rq_lock); RETURN(0); } } - CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc" - " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_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)); + CDEBUG(D_RPCTRACE, "Sending 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)); rc = ptl_send_rpc(req, 0); if (rc) { DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc); - req->rq_net_err = 1; + spin_lock(&req->rq_lock); + req->rq_net_err = 1; + spin_unlock(&req->rq_lock); RETURN(rc); } RETURN(0); } -/* this sends any unsent RPCs in @set and returns TRUE if all are sent */ +static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set) +{ + int remaining, rc; + ENTRY; + + LASSERT(set->set_producer != NULL); + + remaining = cfs_atomic_read(&set->set_remaining); + + /* populate the ->set_requests list with requests until we + * reach the maximum number of RPCs in flight for this set */ + while (cfs_atomic_read(&set->set_remaining) < set->set_max_inflight) { + rc = set->set_producer(set, set->set_producer_arg); + if (rc == -ENOENT) { + /* no more RPC to produce */ + set->set_producer = NULL; + set->set_producer_arg = NULL; + RETURN(0); + } + } + + RETURN((cfs_atomic_read(&set->set_remaining) - remaining)); +} + +/** + * this sends any unsent RPCs in \a set and returns 1 if all are sent + * 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) + */ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set) { - struct list_head *tmp; + cfs_list_t *tmp, *next; int force_timer_recalc = 0; ENTRY; - if (set->set_remaining == 0) + if (cfs_atomic_read(&set->set_remaining) == 0) RETURN(1); - list_for_each(tmp, &set->set_requests) { + cfs_list_for_each_safe(tmp, next, &set->set_requests) { struct ptlrpc_request *req = - list_entry(tmp, struct ptlrpc_request, rq_set_chain); + 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 && @@ -1273,7 +1542,12 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set) /* delayed send - skip */ if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent) - continue; + continue; + + /* delayed resend - skip */ + if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend && + req->rq_sent > cfs_time_current_sec()) + continue; if (!(req->rq_phase == RQ_PHASE_RPC || req->rq_phase == RQ_PHASE_BULK || @@ -1337,21 +1611,36 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set) if (ptlrpc_client_recv_or_unlink(req) || ptlrpc_client_bulk_active(req)) continue; + /* If there is no need to resend, fail it now. */ + if (req->rq_no_resend) { + if (req->rq_status == 0) + req->rq_status = -EIO; + ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET); + GOTO(interpret, req->rq_status); + } else { + continue; + } } if (req->rq_err) { - req->rq_replied = 0; + spin_lock(&req->rq_lock); + req->rq_replied = 0; + spin_unlock(&req->rq_lock); if (req->rq_status == 0) req->rq_status = -EIO; ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET); GOTO(interpret, req->rq_status); } - /* ptlrpc_queue_wait->l_wait_event guarantees that rq_intr - * will only be set after rq_timedout, but 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 */ + /* 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 + * 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 + * need to enforce that here. + */ + if (req->rq_intr && (req->rq_timedout || req->rq_waiting || req->rq_wait_ctx)) { req->rq_status = -EINTR; @@ -1367,13 +1656,15 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set) if (!ptlrpc_unregister_reply(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 */ - list_del_init(&req->rq_list); - list_add_tail(&req->rq_list, &imp->imp_delayed_list); - spin_unlock(&imp->imp_lock); + 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, + &imp-> + imp_delayed_list); + spin_unlock(&imp->imp_lock); continue; } @@ -1381,29 +1672,34 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set) req->rq_status = status; ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET); - spin_unlock(&imp->imp_lock); - GOTO(interpret, req->rq_status); - } - if (req->rq_no_resend && !req->rq_wait_ctx) { - req->rq_status = -ENOTCONN; - ptlrpc_rqphase_move(req, - RQ_PHASE_INTERPRET); - spin_unlock(&imp->imp_lock); - GOTO(interpret, req->rq_status); - } - - list_del_init(&req->rq_list); - list_add_tail(&req->rq_list, - &imp->imp_sending_list); - - spin_unlock(&imp->imp_lock); - - req->rq_waiting = 0; - - if (req->rq_timedout||req->rq_resend) { - /* This is re-sending anyways, - * let's mark req as resend. */ - req->rq_resend = 1; + spin_unlock(&imp->imp_lock); + GOTO(interpret, req->rq_status); + } + if (ptlrpc_no_resend(req) && + !req->rq_wait_ctx) { + req->rq_status = -ENOTCONN; + ptlrpc_rqphase_move(req, + RQ_PHASE_INTERPRET); + spin_unlock(&imp->imp_lock); + GOTO(interpret, req->rq_status); + } + + cfs_list_del_init(&req->rq_list); + cfs_list_add_tail(&req->rq_list, + &imp->imp_sending_list); + + spin_unlock(&imp->imp_lock); + + spin_lock(&req->rq_lock); + req->rq_waiting = 0; + spin_unlock(&req->rq_lock); + + if (req->rq_timedout || req->rq_resend) { + /* This is re-sending anyways, + * let's mark req as resend. */ + spin_lock(&req->rq_lock); + req->rq_resend = 1; + spin_unlock(&req->rq_lock); if (req->rq_bulk) { __u64 old_xid; @@ -1427,49 +1723,64 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set) if (status) { if (req->rq_err) { req->rq_status = status; - req->rq_wait_ctx = 0; - force_timer_recalc = 1; - } else { - req->rq_wait_ctx = 1; - } - - continue; - } else { - req->rq_wait_ctx = 0; - } - - rc = ptl_send_rpc(req, 0); - if (rc) { - DEBUG_REQ(D_HA, req, "send failed (%d)", - rc); - force_timer_recalc = 1; - req->rq_net_err = 1; - } - /* need to reset the timeout */ - force_timer_recalc = 1; - } - - spin_lock(&req->rq_lock); - - if (ptlrpc_client_early(req)) { - ptlrpc_at_recv_early_reply(req); - spin_unlock(&req->rq_lock); + spin_lock(&req->rq_lock); + req->rq_wait_ctx = 0; + spin_unlock(&req->rq_lock); + force_timer_recalc = 1; + } else { + spin_lock(&req->rq_lock); + req->rq_wait_ctx = 1; + spin_unlock(&req->rq_lock); + } + + continue; + } else { + spin_lock(&req->rq_lock); + req->rq_wait_ctx = 0; + spin_unlock(&req->rq_lock); + } + + rc = ptl_send_rpc(req, 0); + if (rc) { + DEBUG_REQ(D_HA, req, + "send failed: rc = %d", rc); + force_timer_recalc = 1; + spin_lock(&req->rq_lock); + req->rq_net_err = 1; + spin_unlock(&req->rq_lock); + continue; + } + /* need to reset the timeout */ + force_timer_recalc = 1; + } + + spin_lock(&req->rq_lock); + + if (ptlrpc_client_early(req)) { + ptlrpc_at_recv_early_reply(req); + spin_unlock(&req->rq_lock); + continue; + } + + /* Still waiting for a reply? */ + if (ptlrpc_client_recv(req)) { + spin_unlock(&req->rq_lock); + continue; + } + + /* Did we actually receive a reply? */ + if (!ptlrpc_client_replied(req)) { + spin_unlock(&req->rq_lock); + continue; + } + + spin_unlock(&req->rq_lock); + + /* unlink from net because we are going to + * swab in-place of reply buffer */ + unregistered = ptlrpc_unregister_reply(req, 1); + if (!unregistered) continue; - } - - /* Still waiting for a reply? */ - if (ptlrpc_client_recv(req)) { - spin_unlock(&req->rq_lock); - continue; - } - - /* Did we actually receive a reply? */ - if (!ptlrpc_client_replied(req)) { - spin_unlock(&req->rq_lock); - continue; - } - - spin_unlock(&req->rq_lock); req->rq_status = after_reply(req); if (req->rq_resend) @@ -1480,7 +1791,7 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set) * process the reply. Similarly if the RPC returned * an error, and therefore the bulk will never arrive. */ - if (req->rq_bulk == NULL || req->rq_status != 0) { + if (req->rq_bulk == NULL || req->rq_status < 0) { ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET); GOTO(interpret, req->rq_status); } @@ -1492,14 +1803,14 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set) if (ptlrpc_client_bulk_active(req)) continue; - if (!req->rq_bulk->bd_success) { - /* The RPC reply arrived OK, but the bulk screwed - * up! Dead wierd since the server told us the RPC - * was good after getting the REPLY for her GET or - * the ACK for her PUT. */ - DEBUG_REQ(D_ERROR, req, "bulk transfer failed"); - LBUG(); - } + if (req->rq_bulk->bd_failure) { + /* The RPC reply arrived OK, but the bulk screwed + * up! Dead weird since the server told us the RPC + * was good after getting the REPLY for her GET or + * the ACK for her PUT. */ + DEBUG_REQ(D_ERROR, req, "bulk transfer failed"); + req->rq_status = -EIO; + } ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET); @@ -1508,8 +1819,11 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set) /* This moves to "unregistering" phase we need to wait for * reply unlink. */ - if (!ptlrpc_unregister_reply(req, 1)) + if (!unregistered && !ptlrpc_unregister_reply(req, 1)) { + /* start async bulk unlink too */ + ptlrpc_unregister_bulk(req, 1); continue; + } if (!ptlrpc_unregister_bulk(req, 1)) continue; @@ -1520,53 +1834,82 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set) ptlrpc_req_interpret(env, req, req->rq_status); - ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE); - - CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:" - "opc %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(), - imp->imp_obd->obd_uuid.uuid, - req->rq_reqmsg ? lustre_msg_get_status(req->rq_reqmsg):-1, - req->rq_xid, - libcfs_nid2str(imp->imp_connection->c_peer.nid), - req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : -1); - - spin_lock(&imp->imp_lock); - /* Request already may be not on sending or delaying list. This - * may happen in the case of marking it errorneous for the case - * ptlrpc_import_delay_req(req, status) find it impossible to - * allow sending this rpc and returns *status != 0. */ - if (!list_empty(&req->rq_list)) { - list_del_init(&req->rq_list); - atomic_dec(&imp->imp_inflight); - } - spin_unlock(&imp->imp_lock); - - set->set_remaining--; - cfs_waitq_broadcast(&imp->imp_recovery_waitq); + if (ptlrpcd_check_work(req)) { + atomic_dec(&set->set_remaining); + continue; + } + 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)); + + 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); + cfs_atomic_dec(&imp->imp_inflight); + } + spin_unlock(&imp->imp_lock); + + cfs_atomic_dec(&set->set_remaining); + wake_up_all(&imp->imp_recovery_waitq); + + if (set->set_producer) { + /* produce a new request if possible */ + if (ptlrpc_set_producer(set) > 0) + force_timer_recalc = 1; + + /* free the request that has just been completed + * in order not to pollute set->set_requests */ + cfs_list_del_init(&req->rq_set_chain); + spin_lock(&req->rq_lock); + req->rq_set = NULL; + req->rq_invalid_rqset = 0; + spin_unlock(&req->rq_lock); + + /* record rq_status to compute the final status later */ + if (req->rq_status != 0) + set->set_rc = req->rq_status; + ptlrpc_req_finished(req); + } } /* If we hit an error, we want to recover promptly. */ - RETURN(set->set_remaining == 0 || force_timer_recalc); + RETURN(cfs_atomic_read(&set->set_remaining) == 0 || force_timer_recalc); } +EXPORT_SYMBOL(ptlrpc_check_set); -/* Return 1 if we should give up, else 0 */ +/** + * Time out request \a req. is \a async_unlink is set, that means do not wait + * until LNet actually confirms network buffer unlinking. + * Return 1 if we should give up further retrying attempts or 0 otherwise. + */ int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink) { - struct obd_import *imp = req->rq_import; - int rc = 0; - ENTRY; + struct obd_import *imp = req->rq_import; + int rc = 0; + ENTRY; - spin_lock(&req->rq_lock); - req->rq_timedout = 1; - spin_unlock(&req->rq_lock); + spin_lock(&req->rq_lock); + req->rq_timedout = 1; + spin_unlock(&req->rq_lock); - DEBUG_REQ(D_WARNING, req, "Request x"LPU64" sent from %s to NID %s " - CFS_DURATION_T"s ago has %s (limit "CFS_DURATION_T"s).\n", - req->rq_xid, imp ? imp->imp_obd->obd_name : "", - imp ? libcfs_nid2str(imp->imp_connection->c_peer.nid) : "", - cfs_time_sub(cfs_time_current_sec(), req->rq_sent), - req->rq_net_err ? "failed due to network error" : "timed out", - cfs_time_sub(req->rq_deadline, req->rq_sent)); + DEBUG_REQ(D_WARNING, req, "Request sent has %s: [sent "CFS_DURATION_T + "/real "CFS_DURATION_T"]", + 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)) ? + "timed out for sent delay" : "timed out for slow reply"), + req->rq_sent, req->rq_real_sent); if (imp != NULL && obd_debug_peer_on_timeout) LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer); @@ -1582,10 +1925,7 @@ int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink) RETURN(1); } - if (req->rq_fake) - RETURN(1); - - atomic_inc(&imp->imp_timeouts); + cfs_atomic_inc(&imp->imp_timeouts); /* The DLM server doesn't want recovery run on its imports. */ if (imp->imp_dlm_fake) @@ -1599,16 +1939,16 @@ int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink) DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)", ptlrpc_import_state_name(req->rq_send_state), ptlrpc_import_state_name(imp->imp_state)); - spin_lock(&req->rq_lock); - req->rq_status = -ETIMEDOUT; - req->rq_err = 1; - spin_unlock(&req->rq_lock); - RETURN(1); + spin_lock(&req->rq_lock); + req->rq_status = -ETIMEDOUT; + req->rq_err = 1; + spin_unlock(&req->rq_lock); + RETURN(1); } /* if a request can't be resent we can't wait for an answer after the timeout */ - if (req->rq_no_resend) { + if (ptlrpc_no_resend(req)) { DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:"); rc = 1; } @@ -1618,10 +1958,15 @@ int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink) RETURN(rc); } +/** + * Time out all uncompleted requests in request set pointed by \a data + * Callback used when waiting on sets with l_wait_event. + * Always returns 1. + */ int ptlrpc_expired_set(void *data) { struct ptlrpc_request_set *set = data; - struct list_head *tmp; + cfs_list_t *tmp; time_t now = cfs_time_current_sec(); ENTRY; @@ -1630,9 +1975,10 @@ int ptlrpc_expired_set(void *data) /* * A timeout expired. See which reqs it applies to... */ - list_for_each (tmp, &set->set_requests) { + cfs_list_for_each (tmp, &set->set_requests) { struct ptlrpc_request *req = - list_entry(tmp, struct ptlrpc_request, rq_set_chain); + cfs_list_entry(tmp, struct ptlrpc_request, + rq_set_chain); /* don't expire request waiting for context */ if (req->rq_wait_ctx) @@ -1654,31 +2000,41 @@ int ptlrpc_expired_set(void *data) } /* - * When waiting for a whole set, we always to break out of the + * When waiting for a whole set, we always break out of the * sleep so we can recalculate the timeout, or enable interrupts * if everyone's timed out. */ RETURN(1); } +EXPORT_SYMBOL(ptlrpc_expired_set); +/** + * Sets rq_intr flag in \a req under spinlock. + */ void ptlrpc_mark_interrupted(struct ptlrpc_request *req) { - spin_lock(&req->rq_lock); - req->rq_intr = 1; - spin_unlock(&req->rq_lock); + spin_lock(&req->rq_lock); + req->rq_intr = 1; + spin_unlock(&req->rq_lock); } +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) { struct ptlrpc_request_set *set = data; - struct list_head *tmp; + cfs_list_t *tmp; LASSERT(set != NULL); - CERROR("INTERRUPTED SET %p\n", set); + CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set); - list_for_each(tmp, &set->set_requests) { + cfs_list_for_each(tmp, &set->set_requests) { struct ptlrpc_request *req = - list_entry(tmp, struct ptlrpc_request, rq_set_chain); + cfs_list_entry(tmp, struct ptlrpc_request, + rq_set_chain); if (req->rq_phase != RQ_PHASE_RPC && req->rq_phase != RQ_PHASE_UNREGISTERING) @@ -1687,23 +2043,22 @@ void ptlrpc_interrupted_set(void *data) 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) { - struct list_head *tmp; + cfs_list_t *tmp; time_t now = cfs_time_current_sec(); int timeout = 0; struct ptlrpc_request *req; int deadline; ENTRY; - SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */ - - list_for_each(tmp, &set->set_requests) { - req = list_entry(tmp, struct ptlrpc_request, rq_set_chain); + cfs_list_for_each(tmp, &set->set_requests) { + req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain); /* * Request in-flight? @@ -1727,6 +2082,8 @@ int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set) if (req->rq_phase == RQ_PHASE_NEW) deadline = req->rq_sent; + else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend) + deadline = req->rq_sent; else deadline = req->rq_sent + req->rq_timeout; @@ -1737,24 +2094,35 @@ 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 + * requests in the set complete (either get a reply, timeout, get an + * error or otherwise be interrupted). + * Returns 0 on success or error code otherwise. + */ int ptlrpc_set_wait(struct ptlrpc_request_set *set) { - struct list_head *tmp; + cfs_list_t *tmp; struct ptlrpc_request *req; struct l_wait_info lwi; int rc, timeout; ENTRY; - if (list_empty(&set->set_requests)) + 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); + if (req->rq_phase == RQ_PHASE_NEW) + (void)ptlrpc_send_new_req(req); + } + + if (cfs_list_empty(&set->set_requests)) RETURN(0); - 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); - } - do { timeout = ptlrpc_set_next_timeout(set); @@ -1762,28 +2130,70 @@ int ptlrpc_set_wait(struct ptlrpc_request_set *set) * req times out */ CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n", set, timeout); - lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout ? timeout : 1), - ptlrpc_expired_set, - ptlrpc_interrupted_set, set); - rc = l_wait_event(set->set_waitq, - ptlrpc_check_set(NULL, set), &lwi); + + 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); + else + /* + * At least one request is in flight, so no + * interrupts are allowed. Wait until all + * complete, or an in-flight req times out. + */ + lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1), + ptlrpc_expired_set, set); + + rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi); + + /* 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()) { + sigset_t blocked_sigs = + cfs_block_sigsinv(LUSTRE_FATAL_SIGS); + + /* In fact we only interrupt for the "fatal" signals + * like SIGINT or SIGKILL. We still ignore less + * important signals since ptlrpc set is not easily + * reentrant from userspace again */ + if (cfs_signal_pending()) + ptlrpc_interrupted_set(set); + cfs_restore_sigs(blocked_sigs); + } LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT); /* -EINTR => all requests have been flagged rq_intr so next * check completes. - * -ETIMEOUTD => someone timed out. When all reqs have + * -ETIMEDOUT => someone timed out. When all reqs have * timed out, signals are enabled allowing completion with * EINTR. * I don't really care if we go once more round the loop in * the error cases -eeb. */ - } while (rc != 0 || set->set_remaining != 0); + if (rc == 0 && cfs_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); + spin_lock(&req->rq_lock); + req->rq_invalid_rqset = 1; + spin_unlock(&req->rq_lock); + } + } + } while (rc != 0 || cfs_atomic_read(&set->set_remaining) != 0); - LASSERT(set->set_remaining == 0); + LASSERT(cfs_atomic_read(&set->set_remaining) == 0); - rc = 0; - list_for_each(tmp, &set->set_requests) { - req = list_entry(tmp, struct ptlrpc_request, rq_set_chain); + 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); LASSERT(req->rq_phase == RQ_PHASE_COMPLETE); if (req->rq_status != 0) @@ -1798,9 +2208,9 @@ int ptlrpc_set_wait(struct ptlrpc_request_set *set) struct ptlrpc_set_cbdata *cbdata, *n; int err; - list_for_each_entry_safe(cbdata, n, + cfs_list_for_each_entry_safe(cbdata, n, &set->set_cblist, psc_item) { - list_del_init(&cbdata->psc_item); + cfs_list_del_init(&cbdata->psc_item); err = cbdata->psc_interpret(set, cbdata->psc_data, rc); if (err && !rc) rc = err; @@ -1810,7 +2220,16 @@ int ptlrpc_set_wait(struct ptlrpc_request_set *set) RETURN(rc); } +EXPORT_SYMBOL(ptlrpc_set_wait); +/** + * Helper fuction for request freeing. + * Called when request count reached zero and request needs to be freed. + * Removes request from all sorts of sending/replay lists it might be on, + * frees network buffers if any are present. + * If \a locked is set, that means caller is already holding import imp_lock + * and so we no longer need to reobtain it (for certain lists manipulations) + */ static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked) { ENTRY; @@ -1821,26 +2240,25 @@ static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked) LASSERTF(!request->rq_receiving_reply, "req %p\n", request); LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */ - LASSERTF(list_empty(&request->rq_list), "req %p\n", request); - LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request); - LASSERTF(list_empty(&request->rq_exp_list), "req %p\n", request); + 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); - LASSERT(request->rq_cli_ctx || request->rq_fake); 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); - list_del_init(&request->rq_replay_list); - if (!locked) - spin_unlock(&request->rq_import->imp_lock); + if (!locked) + spin_lock(&request->rq_import->imp_lock); + cfs_list_del_init(&request->rq_replay_list); + if (!locked) + spin_unlock(&request->rq_import->imp_lock); } - LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request); + LASSERTF(cfs_list_empty(&request->rq_replay_list), "req %p\n", request); - if (atomic_read(&request->rq_refcount) != 0) { + if (cfs_atomic_read(&request->rq_refcount) != 0) { DEBUG_REQ(D_ERROR, request, "freeing request with nonzero refcount"); LBUG(); @@ -1856,8 +2274,8 @@ static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked) class_import_put(request->rq_import); request->rq_import = NULL; } - if (request->rq_bulk != NULL) - ptlrpc_free_bulk(request->rq_bulk); + if (request->rq_bulk != NULL) + ptlrpc_free_bulk_pin(request->rq_bulk); if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL) sptlrpc_cli_free_reqbuf(request); @@ -1868,17 +2286,28 @@ static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked) if (request->rq_pool) __ptlrpc_free_req_to_pool(request); else - OBD_FREE(request, sizeof(*request)); - EXIT; + ptlrpc_request_cache_free(request); + EXIT; } 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. + */ void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request) { - LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock); - (void)__ptlrpc_req_finished(request, 1); + LASSERT(spin_is_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. + */ static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked) { ENTRY; @@ -1893,9 +2322,9 @@ static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked) } DEBUG_REQ(D_INFO, request, "refcount now %u", - atomic_read(&request->rq_refcount) - 1); + cfs_atomic_read(&request->rq_refcount) - 1); - if (atomic_dec_and_test(&request->rq_refcount)) { + if (cfs_atomic_dec_and_test(&request->rq_refcount)) { __ptlrpc_free_req(request, locked); RETURN(1); } @@ -1903,36 +2332,44 @@ static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked) RETURN(0); } +/** + * Drops one reference count for a request. + */ void ptlrpc_req_finished(struct ptlrpc_request *request) { __ptlrpc_req_finished(request, 0); } +EXPORT_SYMBOL(ptlrpc_req_finished); +/** + * Returns xid of a \a request + */ __u64 ptlrpc_req_xid(struct ptlrpc_request *request) { return request->rq_xid; } EXPORT_SYMBOL(ptlrpc_req_xid); -/* Disengage the client's reply buffer from the network +/** + * Disengage the client's reply buffer from the network * NB does _NOT_ unregister any client-side bulk. * IDEMPOTENT, but _not_ safe against concurrent callers. * 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) { - int rc; - cfs_waitq_t *wq; - struct l_wait_info lwi; + int rc; + struct l_wait_info lwi; - /* - * Might sleep. - */ - LASSERT(!in_interrupt()); + /* + * Might sleep. + */ + LASSERT(!in_interrupt()); - /* - * Let's setup deadline for reply unlink. - */ + /* + * 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; @@ -1967,12 +2404,13 @@ int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async) * a chance to run reply_in_callback(), and to make sure we've * unlinked before returning a req to the pool. */ - if (request->rq_set != NULL) - wq = &request->rq_set->set_waitq; - else - wq = &request->rq_reply_waitq; - 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), @@ -1991,36 +2429,78 @@ int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async) } RETURN(0); } +EXPORT_SYMBOL(ptlrpc_unregister_reply); -/* caller must hold imp->imp_lock */ -void ptlrpc_free_committed(struct obd_import *imp) +static void ptlrpc_free_request(struct ptlrpc_request *req) { - struct list_head *tmp, *saved; - struct ptlrpc_request *req; - struct ptlrpc_request *last_req = NULL; /* temporary fire escape */ - ENTRY; + spin_lock(&req->rq_lock); + req->rq_replay = 0; + spin_unlock(&req->rq_lock); + + if (req->rq_commit_cb != NULL) + req->rq_commit_cb(req); + cfs_list_del_init(&req->rq_replay_list); + + __ptlrpc_req_finished(req, 1); +} + +/** + * the request is committed and dropped from the replay list of its import + */ +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)) { + spin_unlock(&imp->imp_lock); + return; + } + + if (force || req->rq_transno <= imp->imp_peer_committed_transno) + ptlrpc_free_request(req); + + spin_unlock(&imp->imp_lock); +} +EXPORT_SYMBOL(ptlrpc_request_committed); - LASSERT(imp != NULL); +/** + * Iterates through replay_list on import and prunes + * all requests have transno smaller than last_committed for the + * import and don't have rq_replay set. + * Since requests are sorted in transno order, stops when meetign first + * transno bigger than last_committed. + * caller must hold imp->imp_lock + */ +void ptlrpc_free_committed(struct obd_import *imp) +{ + struct ptlrpc_request *req, *saved; + struct ptlrpc_request *last_req = NULL; /* temporary fire escape */ + bool skip_committed_list = true; + ENTRY; - LASSERT_SPIN_LOCKED(&imp->imp_lock); + LASSERT(imp != NULL); + LASSERT(spin_is_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_RPCTRACE, "%s: skip recheck: last_committed "LPU64"\n", + CDEBUG(D_INFO, "%s: skip recheck: last_committed "LPU64"\n", imp->imp_obd->obd_name, imp->imp_peer_committed_transno); - EXIT; - return; + RETURN_EXIT; } CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" 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) + skip_committed_list = false; + imp->imp_last_transno_checked = imp->imp_peer_committed_transno; imp->imp_last_generation_checked = imp->imp_generation; - list_for_each_safe(tmp, saved, &imp->imp_replay_list) { - req = list_entry(tmp, struct ptlrpc_request, rq_replay_list); - + cfs_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); last_req = req; @@ -2034,47 +2514,60 @@ void ptlrpc_free_committed(struct obd_import *imp) GOTO(free_req, 0); } - if (req->rq_replay) { - DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)"); - continue; - } - /* not yet committed */ if (req->rq_transno > imp->imp_peer_committed_transno) { DEBUG_REQ(D_RPCTRACE, req, "stopping search"); break; } - DEBUG_REQ(D_RPCTRACE, req, "commit (last_committed "LPU64")", + if (req->rq_replay) { + DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)"); + cfs_list_move_tail(&req->rq_replay_list, + &imp->imp_committed_list); + continue; + } + + DEBUG_REQ(D_INFO, req, "commit (last_committed "LPU64")", imp->imp_peer_committed_transno); free_req: - spin_lock(&req->rq_lock); - req->rq_replay = 0; - spin_unlock(&req->rq_lock); - if (req->rq_commit_cb != NULL) - req->rq_commit_cb(req); - list_del_init(&req->rq_replay_list); - __ptlrpc_req_finished(req, 1); + ptlrpc_free_request(req); } + if (skip_committed_list) + GOTO(out, 0); + + cfs_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"); + ptlrpc_free_request(req); + } + } +out: EXIT; - return; } void ptlrpc_cleanup_client(struct obd_import *imp) { ENTRY; EXIT; - return; } +EXPORT_SYMBOL(ptlrpc_cleanup_client); +/** + * Schedule previously sent request for resend. + * For bulk requests we assign new xid (to avoid problems with + * lost replies and therefore several transfers landing into same buffer + * from different sending attempts). + */ void ptlrpc_resend_req(struct ptlrpc_request *req) { DEBUG_REQ(D_HA, req, "going to resend"); lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 }); req->rq_status = -EAGAIN; - spin_lock(&req->rq_lock); + spin_lock(&req->rq_lock); req->rq_resend = 1; req->rq_net_err = 0; req->rq_timedout = 0; @@ -2087,64 +2580,45 @@ void ptlrpc_resend_req(struct ptlrpc_request *req) old_xid, req->rq_xid); } ptlrpc_client_wake_req(req); - spin_unlock(&req->rq_lock); + 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) { - DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request"); - req->rq_status = -ERESTARTSYS; + DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request"); + req->rq_status = -ERESTARTSYS; - spin_lock(&req->rq_lock); - req->rq_restart = 1; - req->rq_timedout = 0; - ptlrpc_client_wake_req(req); - spin_unlock(&req->rq_lock); -} - -static int expired_request(void *data) -{ - struct ptlrpc_request *req = data; - ENTRY; - - /* - * Some failure can suspend regular timeouts. - */ - if (ptlrpc_check_suspend()) - RETURN(1); - - /* - * Deadline may have changed with an early reply. - */ - if (req->rq_deadline > cfs_time_current_sec()) - RETURN(1); - - RETURN(ptlrpc_expire_one_request(req, 0)); -} - -static void interrupted_request(void *data) -{ - struct ptlrpc_request *req = data; - DEBUG_REQ(D_HA, req, "request interrupted"); - spin_lock(&req->rq_lock); - req->rq_intr = 1; - spin_unlock(&req->rq_lock); + spin_lock(&req->rq_lock); + req->rq_restart = 1; + req->rq_timedout = 0; + ptlrpc_client_wake_req(req); + spin_unlock(&req->rq_lock); } +EXPORT_SYMBOL(ptlrpc_restart_req); +/** + * Grab additional reference on a request \a req + */ struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req) { ENTRY; - atomic_inc(&req->rq_refcount); + cfs_atomic_inc(&req->rq_refcount); RETURN(req); } +EXPORT_SYMBOL(ptlrpc_request_addref); +/** + * Add a request to import replay_list. + * Must be called under imp_lock + */ void ptlrpc_retain_replayable_request(struct ptlrpc_request *req, struct obd_import *imp) { - struct list_head *tmp; + cfs_list_t *tmp; - LASSERT_SPIN_LOCKED(&imp->imp_lock); + LASSERT(spin_is_locked(&imp->imp_lock)); if (req->rq_transno == 0) { DEBUG_REQ(D_EMERG, req, "saving request with zero transno"); @@ -2156,7 +2630,7 @@ void ptlrpc_retain_replayable_request(struct ptlrpc_request *req, lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT); /* don't re-add requests that have been replayed */ - if (!list_empty(&req->rq_replay_list)) + if (!cfs_list_empty(&req->rq_replay_list)) return; lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY); @@ -2164,9 +2638,10 @@ void ptlrpc_retain_replayable_request(struct ptlrpc_request *req, LASSERT(imp->imp_replayable); /* Balanced in ptlrpc_free_committed, usually. */ ptlrpc_request_addref(req); - list_for_each_prev(tmp, &imp->imp_replay_list) { + cfs_list_for_each_prev(tmp, &imp->imp_replay_list) { struct ptlrpc_request *iter = - list_entry(tmp, struct ptlrpc_request, rq_replay_list); + cfs_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 @@ -2183,243 +2658,56 @@ void ptlrpc_retain_replayable_request(struct ptlrpc_request *req, continue; } - list_add(&req->rq_replay_list, &iter->rq_replay_list); + cfs_list_add(&req->rq_replay_list, &iter->rq_replay_list); return; } - list_add_tail(&req->rq_replay_list, &imp->imp_replay_list); + cfs_list_add(&req->rq_replay_list, &imp->imp_replay_list); } +EXPORT_SYMBOL(ptlrpc_retain_replayable_request); +/** + * Send request and wait until it completes. + * Returns request processing status. + */ int ptlrpc_queue_wait(struct ptlrpc_request *req) { - int rc = 0; - int brc; - struct l_wait_info lwi; - struct obd_import *imp = req->rq_import; - cfs_duration_t timeout = CFS_TICK; - long timeoutl; + struct ptlrpc_request_set *set; + int rc; ENTRY; LASSERT(req->rq_set == NULL); LASSERT(!req->rq_receiving_reply); - /* for distributed debugging */ - lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid()); - LASSERT(imp->imp_obd != NULL); - CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc " - "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_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)); - - /* Mark phase here for a little debug help */ - ptlrpc_rqphase_move(req, RQ_PHASE_RPC); - - spin_lock(&imp->imp_lock); - req->rq_import_generation = imp->imp_generation; -restart: - if (ptlrpc_import_delay_req(imp, req, &rc)) { - list_del_init(&req->rq_list); - list_add_tail(&req->rq_list, &imp->imp_delayed_list); - atomic_inc(&imp->imp_inflight); - spin_unlock(&imp->imp_lock); - - DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%s != %s)", - cfs_curproc_comm(), - ptlrpc_import_state_name(req->rq_send_state), - ptlrpc_import_state_name(imp->imp_state)); - lwi = LWI_INTR(interrupted_request, req); - rc = l_wait_event(req->rq_reply_waitq, - (req->rq_send_state == imp->imp_state || - req->rq_err || req->rq_intr), - &lwi); - DEBUG_REQ(D_HA, req, "\"%s\" awake: (%s == %s or %d/%d == 1)", - cfs_curproc_comm(), - ptlrpc_import_state_name(imp->imp_state), - ptlrpc_import_state_name(req->rq_send_state), - req->rq_err, req->rq_intr); - - spin_lock(&imp->imp_lock); - list_del_init(&req->rq_list); - atomic_dec(&imp->imp_inflight); - - if (req->rq_err) { - /* rq_status was set locally */ - rc = -EIO; - } - else if (req->rq_intr) { - rc = -EINTR; - } - else if (req->rq_no_resend) { - spin_unlock(&imp->imp_lock); - GOTO(out, rc = -ETIMEDOUT); - } - else { - GOTO(restart, rc); - } - } - - if (rc != 0) { - spin_unlock(&imp->imp_lock); - req->rq_status = rc; // XXX this ok? - GOTO(out, rc); - } - - if (req->rq_resend) { - if (req->rq_bulk != NULL) { - ptlrpc_unregister_bulk(req, 0); - - /* bulk requests are supposed to be - * idempotent, so we are free to bump the xid - * here, which we need to do before - * registering the bulk again (bug 6371). - * print the old xid first for sanity. - */ - DEBUG_REQ(D_HA, req, "bumping xid for bulk: "); - req->rq_xid = ptlrpc_next_xid(); - } - - DEBUG_REQ(D_HA, req, "resending: "); - } - - /* XXX this is the same as ptlrpc_set_wait */ - LASSERT(list_empty(&req->rq_list)); - list_add_tail(&req->rq_list, &imp->imp_sending_list); - atomic_inc(&imp->imp_inflight); - spin_unlock(&imp->imp_lock); - - rc = sptlrpc_req_refresh_ctx(req, 0); - if (rc) { - if (req->rq_err) { - /* we got fatal ctx refresh error, directly jump out - * thus we can pass back the actual error code. - */ - spin_lock(&imp->imp_lock); - list_del_init(&req->rq_list); - atomic_dec(&imp->imp_inflight); - spin_unlock(&imp->imp_lock); - - CERROR("Failed to refresh ctx of req %p: %d\n", - req, rc); - GOTO(out, rc); - } - /* simulating we got error during send rpc */ - goto after_send; - } - - rc = ptl_send_rpc(req, 0); - if (rc) - DEBUG_REQ(D_HA, req, "send failed (%d); recovering", rc); -repeat: - timeoutl = req->rq_deadline - cfs_time_current_sec(); - timeout = (timeoutl <= 0 || rc) ? CFS_TICK : - cfs_time_seconds(timeoutl); - DEBUG_REQ(D_NET, req, - "-- sleeping for "CFS_DURATION_T" ticks", timeout); - lwi = LWI_TIMEOUT_INTR(timeout, expired_request, interrupted_request, - req); - brc = l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi); - if (brc == -ETIMEDOUT && ((req->rq_deadline > cfs_time_current_sec()) || - ptlrpc_check_and_wait_suspend(req))) - goto repeat; - -after_send: - CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:opc " - "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_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 the reply was received normally, this just grabs the spinlock - * (ensuring the reply callback has returned), sees that - * req->rq_receiving_reply is clear and returns. */ - ptlrpc_unregister_reply(req, 0); - - spin_lock(&imp->imp_lock); - list_del_init(&req->rq_list); - atomic_dec(&imp->imp_inflight); - spin_unlock(&imp->imp_lock); - - if (req->rq_err) { - DEBUG_REQ(D_RPCTRACE, req, "err rc=%d status=%d", - rc, req->rq_status); - GOTO(out, rc = rc ? rc : -EIO); - } - - if (req->rq_intr) { - /* Should only be interrupted if we timed out. */ - if (!req->rq_timedout) - DEBUG_REQ(D_ERROR, req, - "rq_intr set but rq_timedout not"); - GOTO(out, rc = -EINTR); - } - - /* Resend if we need to */ - if (req->rq_resend||req->rq_timedout) { - /* ...unless we were specifically told otherwise. */ - if (req->rq_no_resend) - GOTO(out, rc = -ETIMEDOUT); - spin_lock(&imp->imp_lock); - /* we can have rq_timeout on dlm fake import which not support - * recovery - but me need resend request on this import instead - * of return error */ - req->rq_resend = 1; - goto restart; - } - - if (!ptlrpc_client_replied(req)) { - /* How can this be? -eeb */ - DEBUG_REQ(D_ERROR, req, "!rq_replied: "); - LBUG(); - GOTO(out, rc = req->rq_status); + set = ptlrpc_prep_set(); + if (set == NULL) { + CERROR("Unable to allocate ptlrpc set."); + RETURN(-ENOMEM); } - rc = after_reply(req); - /* NB may return +ve success rc */ - if (req->rq_resend) { - spin_lock(&imp->imp_lock); - goto restart; - } + /* for distributed debugging */ + lustre_msg_set_status(req->rq_reqmsg, current_pid()); - out: - if (req->rq_bulk != NULL) { - if (rc >= 0) { - /* success so far. Note that anything going wrong - * with bulk now, is EXTREMELY strange, since the - * server must have believed that the bulk - * tranferred OK before she replied with success to - * me. */ - lwi = LWI_TIMEOUT(timeout, NULL, NULL); - brc = l_wait_event(req->rq_reply_waitq, - !ptlrpc_client_bulk_active(req), - &lwi); - LASSERT(brc == 0 || brc == -ETIMEDOUT); - if (brc != 0) { - LASSERT(brc == -ETIMEDOUT); - DEBUG_REQ(D_ERROR, req, "bulk timed out"); - rc = brc; - } else if (!req->rq_bulk->bd_success) { - DEBUG_REQ(D_ERROR, req, "bulk transfer failed"); - rc = -EIO; - } - } - if (rc < 0) - ptlrpc_unregister_bulk(req, 0); - } + /* add a ref for the set (see comment in ptlrpc_set_add_req) */ + ptlrpc_request_addref(req); + ptlrpc_set_add_req(set, req); + rc = ptlrpc_set_wait(set); + ptlrpc_set_destroy(set); - LASSERT(!req->rq_receiving_reply); - ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET); - cfs_waitq_broadcast(&imp->imp_recovery_waitq); RETURN(rc); } +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 error restart replay process. + */ static int ptlrpc_replay_interpret(const struct lu_env *env, struct ptlrpc_request *req, void * data, int rc) @@ -2428,7 +2716,7 @@ static int ptlrpc_replay_interpret(const struct lu_env *env, struct obd_import *imp = req->rq_import; ENTRY; - atomic_dec(&imp->imp_replay_inflight); + cfs_atomic_dec(&imp->imp_replay_inflight); if (!ptlrpc_client_replied(req)) { CERROR("request replay timed out, restarting recovery\n"); @@ -2442,12 +2730,13 @@ static int ptlrpc_replay_interpret(const struct lu_env *env, /** VBR: check version failure */ if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) { - /** replay was failed due to version mismatch */ - DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n"); - spin_lock(&imp->imp_lock); - imp->imp_vbr_failed = 1; - imp->imp_no_lock_replay = 1; - spin_unlock(&imp->imp_lock); + /** replay was failed due to version mismatch */ + DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n"); + spin_lock(&imp->imp_lock); + imp->imp_vbr_failed = 1; + imp->imp_no_lock_replay = 1; + spin_unlock(&imp->imp_lock); + lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status); } else { /** The transno had better not change over replay. */ LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) == @@ -2458,14 +2747,23 @@ static int ptlrpc_replay_interpret(const struct lu_env *env, lustre_msg_get_transno(req->rq_repmsg)); } - spin_lock(&imp->imp_lock); - /** if replays by version then gap was occur on server, no trust to locks */ - if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY) - imp->imp_no_lock_replay = 1; - imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg); - spin_unlock(&imp->imp_lock); + spin_lock(&imp->imp_lock); + /** if replays by version then gap occur on server, no trust to locks */ + if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY) + imp->imp_no_lock_replay = 1; + imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg); + spin_unlock(&imp->imp_lock); LASSERT(imp->imp_last_replay_transno); + /* 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, + lustre_msg_get_transno(req->rq_reqmsg)); + GOTO(out, rc = -EINVAL); + } + DEBUG_REQ(D_HA, req, "got rep"); /* let the callback do fixups, possibly including in the request */ @@ -2486,13 +2784,9 @@ static int ptlrpc_replay_interpret(const struct lu_env *env, * Errors while replay can set transno to 0, but * imp_last_replay_transno shouldn't be set to 0 anyway */ - if (req->rq_transno > 0) { - spin_lock(&imp->imp_lock); - LASSERT(req->rq_transno <= imp->imp_last_replay_transno); - imp->imp_last_replay_transno = req->rq_transno; - spin_unlock(&imp->imp_lock); - } else + if (req->rq_transno == 0) CERROR("Transno is 0 during replay!\n"); + /* continue with recovery */ rc = ptlrpc_import_recovery_state_machine(imp); out: @@ -2500,19 +2794,22 @@ static int ptlrpc_replay_interpret(const struct lu_env *env, if (rc != 0) /* this replay failed, so restart recovery */ - ptlrpc_connect_import(imp, NULL); + ptlrpc_connect_import(imp); RETURN(rc); } +/** + * Prepares and queues request for replay. + * Adds it to ptlrpcd queue for actual sending. + * Returns 0 on success. + */ int ptlrpc_replay_req(struct ptlrpc_request *req) { struct ptlrpc_replay_async_args *aa; ENTRY; LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY); - /* Not handling automatic bulk replay yet (or ever?) */ - LASSERT(req->rq_bulk == NULL); LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args)); aa = ptlrpc_req_async_args(req); @@ -2530,97 +2827,111 @@ int ptlrpc_replay_req(struct ptlrpc_request *req) /* Readjust the timeout for current conditions */ ptlrpc_at_set_req_timeout(req); + /* Tell server the net_latency, so the server can calculate how long + * it should wait for next replay */ + lustre_msg_set_service_time(req->rq_reqmsg, + ptlrpc_at_get_net_latency(req)); DEBUG_REQ(D_HA, req, "REPLAY"); - atomic_inc(&req->rq_import->imp_replay_inflight); + cfs_atomic_inc(&req->rq_import->imp_replay_inflight); ptlrpc_request_addref(req); /* ptlrpcd needs a ref */ - ptlrpcd_add_req(req, PSCOPE_OTHER); + ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1); 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) { - struct list_head *tmp, *n; + cfs_list_t *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. */ - spin_lock(&imp->imp_lock); + 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? */ - list_for_each_safe(tmp, n, &imp->imp_sending_list) { + cfs_list_for_each_safe(tmp, n, &imp->imp_sending_list) { struct ptlrpc_request *req = - list_entry(tmp, struct ptlrpc_request, rq_list); + cfs_list_entry(tmp, struct ptlrpc_request, rq_list); DEBUG_REQ(D_RPCTRACE, req, "inflight"); - spin_lock (&req->rq_lock); - if (req->rq_import_generation < imp->imp_generation) { - req->rq_err = 1; - req->rq_status = -EINTR; - ptlrpc_client_wake_req(req); - } - spin_unlock (&req->rq_lock); - } + spin_lock(&req->rq_lock); + if (req->rq_import_generation < imp->imp_generation) { + req->rq_err = 1; + req->rq_status = -EIO; + ptlrpc_client_wake_req(req); + } + spin_unlock(&req->rq_lock); + } - list_for_each_safe(tmp, n, &imp->imp_delayed_list) { - struct ptlrpc_request *req = - list_entry(tmp, struct ptlrpc_request, rq_list); + cfs_list_for_each_safe(tmp, n, &imp->imp_delayed_list) { + struct ptlrpc_request *req = + cfs_list_entry(tmp, struct ptlrpc_request, rq_list); - DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req"); + DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req"); - spin_lock (&req->rq_lock); - if (req->rq_import_generation < imp->imp_generation) { - req->rq_err = 1; - req->rq_status = -EINTR; - ptlrpc_client_wake_req(req); - } - spin_unlock (&req->rq_lock); - } + spin_lock(&req->rq_lock); + if (req->rq_import_generation < imp->imp_generation) { + req->rq_err = 1; + req->rq_status = -EIO; + ptlrpc_client_wake_req(req); + } + spin_unlock(&req->rq_lock); + } - /* Last chance to free reqs left on the replay list, but we - * will still leak reqs that haven't committed. */ - if (imp->imp_replayable) - ptlrpc_free_committed(imp); + /* Last chance to free reqs left on the replay list, but we + * will still leak reqs that haven't committed. */ + if (imp->imp_replayable) + ptlrpc_free_committed(imp); - spin_unlock(&imp->imp_lock); + spin_unlock(&imp->imp_lock); - EXIT; + EXIT; } +EXPORT_SYMBOL(ptlrpc_abort_inflight); +/** + * Abort all uncompleted requests in request set \a set + */ void ptlrpc_abort_set(struct ptlrpc_request_set *set) { - struct list_head *tmp, *pos; + cfs_list_t *tmp, *pos; LASSERT(set != NULL); - list_for_each_safe(pos, tmp, &set->set_requests) { + cfs_list_for_each_safe(pos, tmp, &set->set_requests) { struct ptlrpc_request *req = - list_entry(pos, struct ptlrpc_request, rq_set_chain); + cfs_list_entry(pos, struct ptlrpc_request, + rq_set_chain); - spin_lock(&req->rq_lock); - if (req->rq_phase != RQ_PHASE_RPC) { - spin_unlock(&req->rq_lock); - continue; - } + spin_lock(&req->rq_lock); + if (req->rq_phase != RQ_PHASE_RPC) { + spin_unlock(&req->rq_lock); + continue; + } - req->rq_err = 1; - req->rq_status = -EINTR; - ptlrpc_client_wake_req(req); - spin_unlock(&req->rq_lock); - } + req->rq_err = 1; + req->rq_status = -EINTR; + ptlrpc_client_wake_req(req); + spin_unlock(&req->rq_lock); + } } static __u64 ptlrpc_last_xid; static spinlock_t ptlrpc_last_xid_lock; -/* Initialize the XID for the node. This is common among all requests on +/** + * Initialize the XID for the node. This is common among all requests on * this node, and only requires the property that it is monotonically * increasing. It does not need to be sequential. Since this is also used * as the RDMA match bits, it is important that a single client NOT have @@ -2637,39 +2948,209 @@ static spinlock_t ptlrpc_last_xid_lock; #define YEAR_2004 (1ULL << 30) void ptlrpc_init_xid(void) { - time_t now = cfs_time_current_sec(); + time_t now = cfs_time_current_sec(); - spin_lock_init(&ptlrpc_last_xid_lock); - if (now < YEAR_2004) { - ll_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid)); - ptlrpc_last_xid >>= 2; - ptlrpc_last_xid |= (1ULL << 61); - } else { - ptlrpc_last_xid = (__u64)now << 20; - } + spin_lock_init(&ptlrpc_last_xid_lock); + if (now < YEAR_2004) { + cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid)); + ptlrpc_last_xid >>= 2; + ptlrpc_last_xid |= (1ULL << 61); + } else { + ptlrpc_last_xid = (__u64)now << 20; + } + + /* Need to always be aligned to a power-of-two for mutli-bulk BRW */ + CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0); + ptlrpc_last_xid &= PTLRPC_BULK_OPS_MASK; } +/** + * Increase xid and returns resulting new value to the caller. + * + * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting + * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC + * itself uses the last bulk xid needed, so the server can determine the + * the number of bulk transfers from the RPC XID and a bitmask. The starting + * xid must align to a power-of-two value. + * + * This is assumed to be true due to the initial ptlrpc_last_xid + * value also being initialized to a power-of-two value. LU-1431 + */ __u64 ptlrpc_next_xid(void) { - __u64 tmp; - spin_lock(&ptlrpc_last_xid_lock); - tmp = ++ptlrpc_last_xid; - spin_unlock(&ptlrpc_last_xid_lock); - return tmp; + __u64 next; + + spin_lock(&ptlrpc_last_xid_lock); + next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT; + ptlrpc_last_xid = next; + spin_unlock(&ptlrpc_last_xid_lock); + + return next; } +EXPORT_SYMBOL(ptlrpc_next_xid); +/** + * Get a glimpse at what next xid value might have been. + * Returns possible next xid. + */ __u64 ptlrpc_sample_next_xid(void) { #if BITS_PER_LONG == 32 - /* need to avoid possible word tearing on 32-bit systems */ - __u64 tmp; - spin_lock(&ptlrpc_last_xid_lock); - tmp = ptlrpc_last_xid + 1; - spin_unlock(&ptlrpc_last_xid_lock); - return tmp; + /* need to avoid possible word tearing on 32-bit systems */ + __u64 next; + + spin_lock(&ptlrpc_last_xid_lock); + next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT; + spin_unlock(&ptlrpc_last_xid_lock); + + return next; #else - /* No need to lock, since returned value is racy anyways */ - return ptlrpc_last_xid + 1; + /* No need to lock, since returned value is racy anyways */ + return ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT; #endif } EXPORT_SYMBOL(ptlrpc_sample_next_xid); + +/** + * Functions for operating ptlrpc workers. + * + * A ptlrpc work is a function which will be running inside ptlrpc context. + * The callback shouldn't sleep otherwise it will block that ptlrpcd thread. + * + * 1. after a work is created, it can be used many times, that is: + * handler = ptlrpcd_alloc_work(); + * ptlrpcd_queue_work(); + * + * queue it again when necessary: + * ptlrpcd_queue_work(); + * ptlrpcd_destroy_work(); + * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but + * it will only be queued once in any time. Also as its name implies, it may + * have delay before it really runs by ptlrpcd thread. + */ +struct ptlrpc_work_async_args { + int (*cb)(const struct lu_env *, void *); + void *cbdata; +}; + +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_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); +} + +static int work_interpreter(const struct lu_env *env, + struct ptlrpc_request *req, void *data, int rc) +{ + struct ptlrpc_work_async_args *arg = data; + + LASSERT(ptlrpcd_check_work(req)); + LASSERT(arg->cb != NULL); + + rc = arg->cb(env, arg->cbdata); + + list_del_init(&req->rq_set_chain); + req->rq_set = NULL; + + if (atomic_dec_return(&req->rq_refcount) > 1) { + atomic_set(&req->rq_refcount, 2); + ptlrpcd_add_work_req(req); + } + return rc; +} + +static int worker_format; + +static int ptlrpcd_check_work(struct ptlrpc_request *req) +{ + return req->rq_pill.rc_fmt == (void *)&worker_format; +} + +/** + * Create a work for ptlrpc. + */ +void *ptlrpcd_alloc_work(struct obd_import *imp, + int (*cb)(const struct lu_env *, void *), void *cbdata) +{ + struct ptlrpc_request *req = NULL; + struct ptlrpc_work_async_args *args; + ENTRY; + + might_sleep(); + + if (cb == NULL) + RETURN(ERR_PTR(-EINVAL)); + + /* copy some code from deprecated fakereq. */ + req = ptlrpc_request_cache_alloc(__GFP_IO); + if (req == NULL) { + CERROR("ptlrpc: run out of memory!\n"); + RETURN(ERR_PTR(-ENOMEM)); + } + + 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; + args->cbdata = cbdata; + + RETURN(req); +} +EXPORT_SYMBOL(ptlrpcd_alloc_work); + +void ptlrpcd_destroy_work(void *handler) +{ + struct ptlrpc_request *req = handler; + + if (req) + ptlrpc_req_finished(req); +} +EXPORT_SYMBOL(ptlrpcd_destroy_work); + +int ptlrpcd_queue_work(void *handler) +{ + struct ptlrpc_request *req = handler; + + /* + * Check if the req is already being queued. + * + * Here comes a trick: it lacks a way of checking if a req is being + * processed reliably in ptlrpc. Here I have to use refcount of req + * for this purpose. This is okay because the caller should use this + * req as opaque data. - Jinshan + */ + LASSERT(atomic_read(&req->rq_refcount) > 0); + if (atomic_inc_return(&req->rq_refcount) == 2) + ptlrpcd_add_work_req(req); + return 0; +} +EXPORT_SYMBOL(ptlrpcd_queue_work);