/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
* vim:expandtab:shiftwidth=8:tabstop=8:
*
- * Copyright (C) 2002 Cluster File Systems, Inc.
+ * Copyright (c) 2002, 2003 Cluster File Systems, Inc.
*
* This file is part of Lustre, http://www.lustre.org.
*
#define DEBUG_SUBSYSTEM S_RPC
+#ifdef __KERNEL__
#include <linux/module.h>
+#else
+#include <liblustre.h>
+#endif
+#include <linux/obd_class.h>
#include <linux/lustre_net.h>
-ptl_handle_eq_t request_out_eq, reply_in_eq, reply_out_eq, bulk_source_eq,
- bulk_sink_eq;
-static const ptl_handle_ni_t *socknal_nip = NULL, *qswnal_nip = NULL;
+struct ptlrpc_ni ptlrpc_interfaces[NAL_MAX_NR];
+int ptlrpc_ninterfaces;
/*
* Free the packet when it has gone out
*/
-static int request_out_callback(ptl_event_t *ev, void *data)
+static int request_out_callback(ptl_event_t *ev)
{
+ struct ptlrpc_request *req = ev->mem_desc.user_ptr;
ENTRY;
+ /* requests always contiguous */
+ LASSERT((ev->mem_desc.options & PTL_MD_IOV) == 0);
+
if (ev->type != PTL_EVENT_SENT) {
// XXX make sure we understand all events, including ACK's
CERROR("Unknown event %d\n", ev->type);
LBUG();
}
+ /* this balances the atomic_inc in ptl_send_rpc */
+ ptlrpc_req_finished(req);
RETURN(1);
}
/*
* Free the packet when it has gone out
*/
-static int reply_out_callback(ptl_event_t *ev, void *data)
+static int reply_out_callback(ptl_event_t *ev)
{
ENTRY;
+ /* replies always contiguous */
+ LASSERT((ev->mem_desc.options & PTL_MD_IOV) == 0);
+
if (ev->type == PTL_EVENT_SENT) {
OBD_FREE(ev->mem_desc.start, ev->mem_desc.length);
+ } else if (ev->type == PTL_EVENT_ACK) {
+ struct ptlrpc_request *req = ev->mem_desc.user_ptr;
+ if (req->rq_flags & PTL_RPC_FL_WANT_ACK) {
+ req->rq_flags &= ~PTL_RPC_FL_WANT_ACK;
+ wake_up(&req->rq_wait_for_rep);
+ } else {
+ DEBUG_REQ(D_ERROR, req,
+ "ack received for reply, not wanted");
+ }
} else {
- // XXX make sure we understand all events, including ACK's
+ // XXX make sure we understand all events
CERROR("Unknown event %d\n", ev->type);
LBUG();
}
/*
* Wake up the thread waiting for the reply once it comes in.
*/
-static int reply_in_callback(ptl_event_t *ev, void *data)
+int reply_in_callback(ptl_event_t *ev)
{
- struct ptlrpc_request *rpc = ev->mem_desc.user_ptr;
+ struct ptlrpc_request *req = ev->mem_desc.user_ptr;
ENTRY;
+ /* replies always contiguous */
+ LASSERT((ev->mem_desc.options & PTL_MD_IOV) == 0);
+
+ if (req->rq_xid == 0x5a5a5a5a5a5a5a5a) {
+ CERROR("Reply received for freed request! Probably a missing "
+ "ptlrpc_abort()\n");
+ LBUG();
+ }
+
+ if (req->rq_xid != ev->match_bits) {
+ CERROR("Reply packet for wrong request\n");
+ LBUG();
+ }
+
if (ev->type == PTL_EVENT_PUT) {
- rpc->rq_repmsg = ev->mem_desc.start + ev->offset;
+ req->rq_repmsg = ev->mem_desc.start + ev->offset;
barrier();
- wake_up_interruptible(&rpc->rq_wait_for_rep);
+ wake_up(&req->rq_wait_for_rep);
} else {
// XXX make sure we understand all events, including ACK's
CERROR("Unknown event %d\n", ev->type);
RETURN(1);
}
-int request_in_callback(ptl_event_t *ev, void *data)
+int request_in_callback(ptl_event_t *ev)
{
- struct ptlrpc_service *service = data;
- int index;
+ struct ptlrpc_request_buffer_desc *rqbd = ev->mem_desc.user_ptr;
+ struct ptlrpc_srv_ni *srv_ni = rqbd->rqbd_srv_ni;
+ struct ptlrpc_service *service = srv_ni->sni_service;
+
+ /* requests always contiguous */
+ LASSERT((ev->mem_desc.options & PTL_MD_IOV) == 0);
+ /* we only enable puts */
+ LASSERT(ev->type == PTL_EVENT_PUT);
+ LASSERT(atomic_read(&srv_ni->sni_nrqbds_receiving) > 0);
+ LASSERT(atomic_read(&rqbd->rqbd_refcount) > 0);
if (ev->rlength != ev->mlength)
CERROR("Warning: Possibly truncated rpc (%d/%d)\n",
ev->mlength, ev->rlength);
- spin_lock(&service->srv_lock);
- for (index = 0; index < service->srv_ring_length; index++)
- if ( service->srv_buf[index] == ev->mem_desc.start)
- break;
+ if (ptl_is_valid_handle(&ev->unlinked_me)) {
+ /* This is the last request to be received into this
+ * request buffer. We don't bump the refcount, since the
+ * thread servicing this event is effectively taking over
+ * portals' reference.
+ */
+#warning ev->unlinked_me.nal_idx is not set properly in a callback
+ LASSERT(ev->unlinked_me.handle_idx==rqbd->rqbd_me_h.handle_idx);
+
+ /* we're off the air */
+ /* we'll probably start dropping packets in portals soon */
+ if (atomic_dec_and_test(&srv_ni->sni_nrqbds_receiving))
+ CERROR("All request buffers busy\n");
+ } else {
+ /* +1 ref for service thread */
+ atomic_inc(&rqbd->rqbd_refcount);
+ }
- if (index == service->srv_ring_length)
- LBUG();
+ wake_up(&service->srv_waitq);
- service->srv_ref_count[index]++;
+ return 0;
+}
- if (ptl_is_valid_handle(&ev->unlinked_me)) {
- int idx;
+static int bulk_put_source_callback(ptl_event_t *ev)
+{
+ struct ptlrpc_bulk_desc *desc = ev->mem_desc.user_ptr;
+ struct ptlrpc_bulk_page *bulk;
+ struct list_head *tmp;
+ struct list_head *next;
+ ENTRY;
- for (idx = 0; idx < service->srv_ring_length; idx++)
- if (service->srv_me_h[idx].handle_idx ==
- ev->unlinked_me.handle_idx)
- break;
- if (idx == service->srv_ring_length)
- LBUG();
+ CDEBUG(D_NET, "got %s event %d\n",
+ (ev->type == PTL_EVENT_SENT) ? "SENT" :
+ (ev->type == PTL_EVENT_ACK) ? "ACK" : "UNEXPECTED", ev->type);
- CDEBUG(D_NET, "unlinked %d\n", idx);
- ptl_set_inv_handle(&(service->srv_me_h[idx]));
+ LASSERT(ev->type == PTL_EVENT_SENT || ev->type == PTL_EVENT_ACK);
- if (service->srv_ref_count[idx] == 0)
- ptlrpc_link_svc_me(service, idx);
- }
+ LASSERT(atomic_read(&desc->bd_source_callback_count) > 0 &&
+ atomic_read(&desc->bd_source_callback_count) <= 2);
- spin_unlock(&service->srv_lock);
- if (ev->type == PTL_EVENT_PUT)
- wake_up(&service->srv_waitq);
- else
- CERROR("Unexpected event type: %d\n", ev->type);
+ /* 1 fragment for each page always */
+ LASSERT(ev->mem_desc.niov == desc->bd_page_count);
- return 0;
+ if (atomic_dec_and_test(&desc->bd_source_callback_count)) {
+ void (*event_handler)(struct ptlrpc_bulk_desc *);
+
+ list_for_each_safe(tmp, next, &desc->bd_page_list) {
+ bulk = list_entry(tmp, struct ptlrpc_bulk_page,
+ bp_link);
+
+ if (bulk->bp_cb != NULL)
+ bulk->bp_cb(bulk);
+ }
+
+ /* We need to make a note of whether there's an event handler
+ * before we call wake_up, because if there is no event handler,
+ * 'desc' might be freed before we're scheduled again. */
+ event_handler = desc->bd_ptl_ev_hdlr;
+
+ desc->bd_flags |= PTL_BULK_FL_SENT;
+ wake_up(&desc->bd_waitq);
+ if (event_handler) {
+ LASSERT(desc->bd_ptl_ev_hdlr == event_handler);
+ event_handler(desc);
+ }
+ }
+
+ RETURN(0);
}
-static int bulk_source_callback(ptl_event_t *ev, void *data)
+static int bulk_put_sink_callback(ptl_event_t *ev)
{
- struct ptlrpc_bulk_page *bulk = ev->mem_desc.user_ptr;
- struct ptlrpc_bulk_desc *desc = bulk->b_desc;
+ struct ptlrpc_bulk_desc *desc = ev->mem_desc.user_ptr;
+ struct ptlrpc_bulk_page *bulk;
+ struct list_head *tmp;
+ struct list_head *next;
+ ptl_size_t total = 0;
+ void (*event_handler)(struct ptlrpc_bulk_desc *);
ENTRY;
- if (ev->type == PTL_EVENT_SENT) {
- CDEBUG(D_NET, "got SENT event\n");
- } else if (ev->type == PTL_EVENT_ACK) {
- CDEBUG(D_NET, "got ACK event\n");
- desc->b_flags |= PTL_BULK_FL_SENT;
- wake_up_interruptible(&desc->b_waitq);
- } else {
- CERROR("Unexpected event type!\n");
- LBUG();
+ LASSERT(ev->type == PTL_EVENT_PUT);
+
+ /* put with zero offset */
+ LASSERT(ev->offset == 0);
+ /* used iovs */
+ LASSERT((ev->mem_desc.options & PTL_MD_IOV) != 0);
+ /* 1 fragment for each page always */
+ LASSERT(ev->mem_desc.niov == desc->bd_page_count);
+
+ list_for_each_safe (tmp, next, &desc->bd_page_list) {
+ bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link);
+
+ total += bulk->bp_buflen;
+
+ if (bulk->bp_cb != NULL)
+ bulk->bp_cb(bulk);
+ }
+
+ LASSERT(ev->mem_desc.length == total);
+
+ /* We need to make a note of whether there's an event handler
+ * before we call wake_up, because if there is no event
+ * handler, 'desc' might be freed before we're scheduled again. */
+ event_handler = desc->bd_ptl_ev_hdlr;
+
+ desc->bd_flags |= PTL_BULK_FL_RCVD;
+ wake_up(&desc->bd_waitq);
+ if (event_handler) {
+ LASSERT(desc->bd_ptl_ev_hdlr == event_handler);
+ event_handler(desc);
}
RETURN(1);
}
-static int bulk_sink_callback(ptl_event_t *ev, void *data)
+static int bulk_get_source_callback(ptl_event_t *ev)
{
- struct ptlrpc_bulk_page *bulk = ev->mem_desc.user_ptr;
- struct ptlrpc_bulk_desc *desc = bulk->b_desc;
+ struct ptlrpc_bulk_desc *desc = ev->mem_desc.user_ptr;
+ struct ptlrpc_bulk_page *bulk;
+ struct list_head *tmp;
+ struct list_head *next;
+ ptl_size_t total = 0;
+ void (*event_handler)(struct ptlrpc_bulk_desc *);
ENTRY;
- if (ev->type == PTL_EVENT_PUT) {
- if (bulk->b_buf != ev->mem_desc.start + ev->offset)
- CERROR("bulkbuf != mem_desc -- why?\n");
- desc->b_finished_count++;
- if (bulk->b_cb != NULL)
- bulk->b_cb(bulk);
- if (desc->b_finished_count == desc->b_page_count) {
- desc->b_flags |= PTL_BULK_FL_RCVD;
- wake_up_interruptible(&desc->b_waitq);
- if (desc->b_cb != NULL)
- desc->b_cb(desc);
- }
- } else {
- CERROR("Unexpected event type!\n");
- LBUG();
+ LASSERT(ev->type == PTL_EVENT_GET);
+
+ /* put with zero offset */
+ LASSERT(ev->offset == 0);
+ /* used iovs */
+ LASSERT((ev->mem_desc.options & PTL_MD_IOV) != 0);
+ /* 1 fragment for each page always */
+ LASSERT(ev->mem_desc.niov == desc->bd_page_count);
+
+ list_for_each_safe (tmp, next, &desc->bd_page_list) {
+ bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link);
+
+ total += bulk->bp_buflen;
+
+ if (bulk->bp_cb != NULL)
+ bulk->bp_cb(bulk);
+ }
+
+ LASSERT(ev->mem_desc.length == total);
+
+ /* We need to make a note of whether there's an event handler
+ * before we call wake_up, because if there is no event
+ * handler, 'desc' might be freed before we're scheduled again. */
+ event_handler = desc->bd_ptl_ev_hdlr;
+
+ desc->bd_flags |= PTL_BULK_FL_SENT;
+ wake_up(&desc->bd_waitq);
+ if (event_handler) {
+ LASSERT(desc->bd_ptl_ev_hdlr == event_handler);
+ event_handler(desc);
}
RETURN(1);
}
-int ptlrpc_init_portals(void)
+
+static int bulk_get_sink_callback(ptl_event_t *ev)
{
- int rc;
- ptl_handle_ni_t ni;
+ struct ptlrpc_bulk_desc *desc = ev->mem_desc.user_ptr;
+ struct ptlrpc_bulk_page *bulk;
+ struct list_head *tmp;
+ struct list_head *next;
+ ENTRY;
- socknal_nip = inter_module_get_request("ksocknal_ni", "ksocknal");
- qswnal_nip = inter_module_get_request("kqswnal_ni", "kqswnal");
- if (socknal_nip == NULL && qswnal_nip == NULL) {
- CERROR("get_ni failed: is a NAL module loaded?\n");
- return -EIO;
+ CDEBUG(D_NET, "got %s event %d\n",
+ (ev->type == PTL_EVENT_SENT) ? "SENT" :
+ (ev->type == PTL_EVENT_REPLY) ? "REPLY" : "UNEXPECTED",
+ ev->type);
+
+ LASSERT(ev->type == PTL_EVENT_SENT || ev->type == PTL_EVENT_REPLY);
+
+ LASSERT(atomic_read(&desc->bd_source_callback_count) > 0 &&
+ atomic_read(&desc->bd_source_callback_count) <= 2);
+
+ /* 1 fragment for each page always */
+ LASSERT(ev->mem_desc.niov == desc->bd_page_count);
+
+ if (atomic_dec_and_test(&desc->bd_source_callback_count)) {
+ void (*event_handler)(struct ptlrpc_bulk_desc *);
+
+ list_for_each_safe(tmp, next, &desc->bd_page_list) {
+ bulk = list_entry(tmp, struct ptlrpc_bulk_page,
+ bp_link);
+
+ if (bulk->bp_cb != NULL)
+ bulk->bp_cb(bulk);
+ }
+
+ /* We need to make a note of whether there's an event handler
+ * before we call wake_up, because if there is no event handler,
+ * 'desc' might be freed before we're scheduled again. */
+ event_handler = desc->bd_ptl_ev_hdlr;
+
+ desc->bd_flags |= PTL_BULK_FL_RCVD;
+ wake_up(&desc->bd_waitq);
+ if (event_handler) {
+ LASSERT(desc->bd_ptl_ev_hdlr == event_handler);
+ event_handler(desc);
+ }
}
- /* Use the qswnal if it's there */
- if (qswnal_nip != NULL)
- ni = *qswnal_nip;
- else
- ni = *socknal_nip;
+ RETURN(0);
+}
- rc = PtlEQAlloc(ni, 128, request_out_callback, NULL, &request_out_eq);
- if (rc != PTL_OK)
- CERROR("PtlEQAlloc failed: %d\n", rc);
+int ptlrpc_uuid_to_peer (struct obd_uuid *uuid, struct ptlrpc_peer *peer)
+{
+ struct ptlrpc_ni *pni;
+ struct lustre_peer lpeer;
+ int i;
+ int rc = lustre_uuid_to_peer (uuid->uuid, &lpeer);
+
+ if (rc != 0)
+ RETURN (rc);
+
+ for (i = 0; i < ptlrpc_ninterfaces; i++) {
+ pni = &ptlrpc_interfaces[i];
+
+ if (!memcmp (&lpeer.peer_ni, &pni->pni_ni_h,
+ sizeof (lpeer.peer_ni))) {
+ peer->peer_nid = lpeer.peer_nid;
+ peer->peer_ni = pni;
+ return (0);
+ }
+ }
+
+ CERROR ("Can't find ptlrpc interface for "LPX64" ni handle %08lx %08lx\n",
+ lpeer.peer_nid, lpeer.peer_ni.nal_idx, lpeer.peer_ni.handle_idx);
+ return (-ENOENT);
+}
- rc = PtlEQAlloc(ni, 128, reply_out_callback, NULL, &reply_out_eq);
- if (rc != PTL_OK)
- CERROR("PtlEQAlloc failed: %d\n", rc);
+void ptlrpc_ni_fini (struct ptlrpc_ni *pni)
+{
+ PtlEQFree(pni->pni_request_out_eq_h);
+ PtlEQFree(pni->pni_reply_out_eq_h);
+ PtlEQFree(pni->pni_reply_in_eq_h);
+ PtlEQFree(pni->pni_bulk_put_source_eq_h);
+ PtlEQFree(pni->pni_bulk_put_sink_eq_h);
+ PtlEQFree(pni->pni_bulk_get_source_eq_h);
+ PtlEQFree(pni->pni_bulk_get_sink_eq_h);
+
+ inter_module_put(pni->pni_name);
+}
- rc = PtlEQAlloc(ni, 128, reply_in_callback, NULL, &reply_in_eq);
- if (rc != PTL_OK)
- CERROR("PtlEQAlloc failed: %d\n", rc);
+int ptlrpc_ni_init (char *name, struct ptlrpc_ni *pni)
+{
+ int rc;
+ ptl_handle_ni_t *nip;
- rc = PtlEQAlloc(ni, 128, bulk_source_callback, NULL, &bulk_source_eq);
+ nip = (ptl_handle_ni_t *)inter_module_get (name);
+ if (nip == NULL) {
+ CDEBUG (D_NET, "Network interface %s not loaded\n", name);
+ return (-ENOENT);
+ }
+
+ CDEBUG (D_NET, "init %s: nal_idx %ld\n", name, nip->nal_idx);
+
+ pni->pni_name = name;
+ pni->pni_ni_h = *nip;
+
+ ptl_set_inv_handle (&pni->pni_request_out_eq_h);
+ ptl_set_inv_handle (&pni->pni_reply_out_eq_h);
+ ptl_set_inv_handle (&pni->pni_reply_in_eq_h);
+ ptl_set_inv_handle (&pni->pni_bulk_put_source_eq_h);
+ ptl_set_inv_handle (&pni->pni_bulk_put_sink_eq_h);
+ ptl_set_inv_handle (&pni->pni_bulk_get_source_eq_h);
+ ptl_set_inv_handle (&pni->pni_bulk_get_sink_eq_h);
+
+ /* NB We never actually PtlEQGet() out of these events queues since
+ * we're only interested in the event callback, so we can just let
+ * them wrap. Their sizes aren't a big deal, apart from providing
+ * a little history for debugging... */
+
+ rc = PtlEQAlloc(pni->pni_ni_h, 1024, request_out_callback,
+ &pni->pni_request_out_eq_h);
if (rc != PTL_OK)
- CERROR("PtlEQAlloc failed: %d\n", rc);
-
- rc = PtlEQAlloc(ni, 128, bulk_sink_callback, NULL, &bulk_sink_eq);
+ GOTO (fail, rc = -ENOMEM);
+
+ rc = PtlEQAlloc(pni->pni_ni_h, 1024, reply_out_callback,
+ &pni->pni_reply_out_eq_h);
+ if (rc != PTL_OK)
+ GOTO (fail, rc = -ENOMEM);
+
+ rc = PtlEQAlloc(pni->pni_ni_h, 1024, reply_in_callback,
+ &pni->pni_reply_in_eq_h);
if (rc != PTL_OK)
- CERROR("PtlEQAlloc failed: %d\n", rc);
+ GOTO (fail, rc = -ENOMEM);
+
+ rc = PtlEQAlloc(pni->pni_ni_h, 1024, bulk_put_source_callback,
+ &pni->pni_bulk_put_source_eq_h);
+ if (rc != PTL_OK)
+ GOTO (fail, rc = -ENOMEM);
+
+ rc = PtlEQAlloc(pni->pni_ni_h, 1024, bulk_put_sink_callback,
+ &pni->pni_bulk_put_sink_eq_h);
+ if (rc != PTL_OK)
+ GOTO (fail, rc = -ENOMEM);
+
+ rc = PtlEQAlloc(pni->pni_ni_h, 1024, bulk_get_source_callback,
+ &pni->pni_bulk_get_source_eq_h);
+ if (rc != PTL_OK)
+ GOTO (fail, rc = -ENOMEM);
+
+ rc = PtlEQAlloc(pni->pni_ni_h, 1024, bulk_get_sink_callback,
+ &pni->pni_bulk_get_sink_eq_h);
+ if (rc != PTL_OK)
+ GOTO (fail, rc = -ENOMEM);
+
+ return (0);
+ fail:
+ CERROR ("Failed to initialise network interface %s: %d\n",
+ name, rc);
+
+ /* OK to do complete teardown since we invalidated the handles above... */
+ ptlrpc_ni_fini (pni);
+ return (rc);
+}
- return rc;
+int ptlrpc_init_portals(void)
+{
+ /* Add new portals network interface names here.
+ * Order is irrelevent! */
+ char *ni_names[] = { "kqswnal_ni",
+ "kgmnal_ni",
+ "ksocknal_ni",
+ "ktoenal_ni",
+ "tcpnal_ni",
+ NULL };
+ int rc;
+ int i;
+
+ LASSERT (ptlrpc_ninterfaces == 0);
+
+ for (i = 0; ni_names[i] != NULL; i++) {
+ LASSERT (ptlrpc_ninterfaces <
+ sizeof (ptlrpc_interfaces)/sizeof (ptlrpc_interfaces[0]));
+
+ rc = ptlrpc_ni_init (ni_names[i],
+ &ptlrpc_interfaces[ptlrpc_ninterfaces]);
+ if (rc == 0)
+ ptlrpc_ninterfaces++;
+ }
+
+ if (ptlrpc_ninterfaces == 0) {
+ CERROR("network initialisation failed: is a NAL module loaded?\n");
+ return -EIO;
+ }
+ return 0;
}
void ptlrpc_exit_portals(void)
{
- PtlEQFree(request_out_eq);
- PtlEQFree(reply_out_eq);
- PtlEQFree(reply_in_eq);
- PtlEQFree(bulk_source_eq);
- PtlEQFree(bulk_sink_eq);
-
- if (qswnal_nip != NULL)
- inter_module_put("kqswnal_ni");
- if (socknal_nip != NULL)
- inter_module_put("ksocknal_ni");
+ while (ptlrpc_ninterfaces > 0)
+ ptlrpc_ni_fini (&ptlrpc_interfaces[--ptlrpc_ninterfaces]);
}