* Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
*
- * Copyright (c) 2011 Whamcloud, Inc.
- *
+ * Copyright (c) 2011, 2012, Whamcloud, Inc.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
* Allocate and initialize new bulk descriptor
* Returns pointer to the descriptor or NULL on error.
*/
-static inline struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal)
+struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal)
{
struct ptlrpc_bulk_desc *desc;
}
/**
- * Prepare bulk descriptor for specified incoming 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 server-side after request was already
- * received.
- * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
- * error.
- */
-struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
- int npages, int type, int portal)
-{
- 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;
-
- desc->bd_cbid.cbid_fn = server_bulk_callback;
- desc->bd_cbid.cbid_arg = desc;
-
- /* NB we don't assign rq_bulk here; server-side requests are
- * re-used, and the handler frees the bulk desc explicitly. */
-
- return desc;
-}
-
-/**
* 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
*/
static int ptlrpc_console_allow(struct ptlrpc_request *req)
{
- __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
+ __u32 opc;
int err;
+ /* Fake requests include no rq_reqmsg */
+ if (req->rq_fake)
+ return 0;
+
+ 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)) &&
* 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;
cfs_spin_lock(&imp->imp_lock);
- req->rq_import_generation = imp->imp_generation;
+ if (!req->rq_generation_set)
+ req->rq_import_generation = imp->imp_generation;
if (ptlrpc_import_delay_req(imp, req, &rc)) {
cfs_spin_lock(&req->rq_lock);
* 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);
}
* was good after getting the REPLY for her GET or
* the ACK for her PUT. */
DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
- LBUG();
+ req->rq_status = -EIO;
}
ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
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);
/* 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");
cfs_atomic_inc(&req->rq_import->imp_replay_inflight);
#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 {
+ __u64 magic;
+ int (*cb)(const struct lu_env *, void *);
+ void *cbdata;
+};
+
+#define PTLRPC_WORK_MAGIC 0x6655436b676f4f44ULL /* magic code */
+
+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(arg->magic == PTLRPC_WORK_MAGIC);
+ LASSERT(arg->cb != NULL);
+
+ return arg->cb(env, arg->cbdata);
+}
+
+/**
+ * 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;
+
+ cfs_might_sleep();
+
+ if (cb == NULL)
+ RETURN(ERR_PTR(-EINVAL));
+
+ /* copy some code from deprecated fakereq. */
+ OBD_ALLOC_PTR(req);
+ 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;
+
+ cfs_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);
+ cfs_waitq_init(&req->rq_reply_waitq);
+ cfs_waitq_init(&req->rq_set_waitq);
+ cfs_atomic_set(&req->rq_refcount, 1);
+
+ CLASSERT (sizeof(*args) <= sizeof(req->rq_async_args));
+ args = ptlrpc_req_async_args(req);
+ args->magic = PTLRPC_WORK_MAGIC;
+ 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(cfs_atomic_read(&req->rq_refcount) > 0);
+ if (cfs_atomic_read(&req->rq_refcount) > 1)
+ return -EBUSY;
+
+ if (cfs_atomic_inc_return(&req->rq_refcount) > 2) { /* race */
+ cfs_atomic_dec(&req->rq_refcount);
+ return -EBUSY;
+ }
+
+ /* 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);
+ return 0;
+}
+EXPORT_SYMBOL(ptlrpcd_queue_work);
git://git.whamcloud.com / fs / lustre-release.git / blobdiff