Whamcloud - gitweb
fixed bulk source callback race
[fs/lustre-release.git] / lustre / ptlrpc / events.c
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  *  Copyright (C) 2002 Cluster File Systems, Inc.
5  *
6  *   This file is part of Lustre, http://www.lustre.org.
7  *
8  *   Lustre is free software; you can redistribute it and/or
9  *   modify it under the terms of version 2 of the GNU General Public
10  *   License as published by the Free Software Foundation.
11  *
12  *   Lustre is distributed in the hope that it will be useful,
13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
15  *   GNU General Public License for more details.
16  *
17  *   You should have received a copy of the GNU General Public License
18  *   along with Lustre; if not, write to the Free Software
19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
20  *
21  */
22
23 #define DEBUG_SUBSYSTEM S_RPC
24
25 #include <linux/module.h>
26 #include <linux/obd_support.h>
27 #include <linux/lustre_net.h>
28
29 ptl_handle_eq_t request_out_eq, reply_in_eq, reply_out_eq, bulk_source_eq,
30         bulk_sink_eq;
31 static const ptl_handle_ni_t *socknal_nip = NULL, *qswnal_nip = NULL;
32
33 /*
34  *  Free the packet when it has gone out
35  */
36 static int request_out_callback(ptl_event_t *ev)
37 {
38         ENTRY;
39
40         LASSERT ((ev->mem_desc.options & PTL_MD_IOV) == 0); /* requests always contiguous */
41
42         if (ev->type != PTL_EVENT_SENT) {
43                 // XXX make sure we understand all events, including ACK's
44                 CERROR("Unknown event %d\n", ev->type);
45                 LBUG();
46         }
47
48         RETURN(1);
49 }
50
51
52 /*
53  *  Free the packet when it has gone out
54  */
55 static int reply_out_callback(ptl_event_t *ev)
56 {
57         ENTRY;
58
59         LASSERT ((ev->mem_desc.options & PTL_MD_IOV) == 0); /* replies always contiguous */
60
61         if (ev->type == PTL_EVENT_SENT) {
62                 OBD_FREE(ev->mem_desc.start, ev->mem_desc.length);
63         } else {
64                 // XXX make sure we understand all events, including ACK's
65                 CERROR("Unknown event %d\n", ev->type);
66                 LBUG();
67         }
68
69         RETURN(1);
70 }
71
72 /*
73  * Wake up the thread waiting for the reply once it comes in.
74  */
75 static int reply_in_callback(ptl_event_t *ev)
76 {
77         struct ptlrpc_request *req = ev->mem_desc.user_ptr;
78         ENTRY;
79
80         LASSERT ((ev->mem_desc.options & PTL_MD_IOV) == 0); /* replies always contiguous */
81
82         if (req->rq_xid == 0x5a5a5a5a5a5a5a5a) {
83                 CERROR("Reply received for freed request!  Probably a missing "
84                        "ptlrpc_abort()\n");
85                 LBUG();
86         }
87
88         if (req->rq_xid != ev->match_bits) {
89                 CERROR("Reply packet for wrong request\n");
90                 LBUG();
91         }
92
93         if (ev->type == PTL_EVENT_PUT) {
94                 req->rq_repmsg = ev->mem_desc.start + ev->offset;
95                 barrier();
96                 wake_up(&req->rq_wait_for_rep);
97         } else {
98                 // XXX make sure we understand all events, including ACK's
99                 CERROR("Unknown event %d\n", ev->type);
100                 LBUG();
101         }
102
103         RETURN(1);
104 }
105
106 int request_in_callback(ptl_event_t *ev)
107 {
108         struct ptlrpc_request_buffer_desc *rqbd = ev->mem_desc.user_ptr;
109         struct ptlrpc_service *service = rqbd->rqbd_service;
110
111         LASSERT ((ev->mem_desc.options & PTL_MD_IOV) == 0); /* requests always contiguous */
112
113         if (ev->rlength != ev->mlength)
114                 CERROR("Warning: Possibly truncated rpc (%d/%d)\n",
115                        ev->mlength, ev->rlength);
116
117         if (ev->type == PTL_EVENT_PUT)
118                 wake_up(&service->srv_waitq);
119         else
120                 CERROR("Unexpected event type: %d\n", ev->type);
121
122         return 0;
123 }
124
125 static int bulk_source_callback(ptl_event_t *ev)
126 {
127         struct ptlrpc_bulk_desc *desc = ev->mem_desc.user_ptr;
128         struct ptlrpc_bulk_page *bulk;
129         struct list_head        *tmp;
130         struct list_head        *next;
131         ENTRY;
132
133         CDEBUG(D_NET, "got %s event %d\n",
134                (ev->type == PTL_EVENT_SENT) ? "SENT" :
135                (ev->type == PTL_EVENT_ACK)  ? "ACK"  : "UNEXPECTED", ev->type);
136
137         LASSERT (ev->type == PTL_EVENT_SENT ||
138                  ev->type == PTL_EVENT_ACK);
139
140         LASSERT (atomic_read (&desc->bd_source_callback_count) > 0 &&
141                  atomic_read (&desc->bd_source_callback_count) <= 2);
142         
143         if (ev->mem_desc.niov != desc->bd_page_count)
144         {
145                 int mdniov = ev->mem_desc.niov;
146                 struct ptlrpc_bulk_desc desc_snapshot;
147
148                 desc_snapshot = *desc;
149                 
150                 printk ("ev(%p)->mem_desc.niov %d != desc(%p)->bd_page_count %d, snapshot %p\n", 
151                         ev, mdniov, desc, desc_snapshot.bd_page_count, &desc_snapshot);
152                 LBUG();
153         }
154         /* 1 fragment for each page always */
155         LASSERT (ev->mem_desc.niov == desc->bd_page_count);
156
157         if (atomic_dec_and_test (&desc->bd_source_callback_count))
158         {
159                 list_for_each_safe(tmp, next, &desc->bd_page_list) {
160                         bulk = list_entry(tmp, struct ptlrpc_bulk_page,
161                                           bp_link);
162                         
163                         if (bulk->bp_cb != NULL)
164                                 bulk->bp_cb(bulk);
165                 }
166                 desc->bd_flags |= PTL_BULK_FL_SENT;
167                 wake_up(&desc->bd_waitq);
168                 if (desc->bd_cb != NULL)
169                         desc->bd_cb(desc, desc->bd_cb_data);
170         }
171
172         RETURN(0);
173 }
174
175 static int bulk_sink_callback(ptl_event_t *ev)
176 {
177         struct ptlrpc_bulk_desc *desc = ev->mem_desc.user_ptr;
178         struct ptlrpc_bulk_page *bulk;
179         struct list_head        *tmp;
180         struct list_head        *next;
181         ptl_size_t               total = 0;
182         ENTRY;
183
184         if (ev->type == PTL_EVENT_PUT) {
185                 /* put with zero offset */
186                 LASSERT (ev->offset == 0);
187                 /* used iovs */
188                 LASSERT ((ev->mem_desc.options & PTL_MD_IOV) != 0);
189                 /* 1 fragment for each page always */
190                 LASSERT (ev->mem_desc.niov == desc->bd_page_count);
191
192                 list_for_each_safe (tmp, next, &desc->bd_page_list) {
193                         bulk = list_entry(tmp, struct ptlrpc_bulk_page,
194                                           bp_link);
195
196                         total += bulk->bp_buflen;
197
198                         if (bulk->bp_cb != NULL)
199                                 bulk->bp_cb(bulk);
200                 }
201
202                 LASSERT (ev->mem_desc.length == total);
203
204                 desc->bd_flags |= PTL_BULK_FL_RCVD;
205                 wake_up(&desc->bd_waitq);
206                 if (desc->bd_cb != NULL)
207                         desc->bd_cb(desc, desc->bd_cb_data);
208         } else {
209                 CERROR("Unexpected event type!\n");
210                 LBUG();
211         }
212
213         RETURN(1);
214 }
215
216 int ptlrpc_init_portals(void)
217 {
218         int rc;
219         ptl_handle_ni_t ni;
220
221         socknal_nip = inter_module_get_request("ksocknal_ni", "ksocknal");
222         qswnal_nip = inter_module_get_request("kqswnal_ni", "kqswnal");
223         if (socknal_nip == NULL && qswnal_nip == NULL) {
224                 CERROR("get_ni failed: is a NAL module loaded?\n");
225                 return -EIO;
226         }
227
228         /* Use the qswnal if it's there */
229         if (qswnal_nip != NULL)
230                 ni = *qswnal_nip;
231         else
232                 ni = *socknal_nip;
233
234         rc = PtlEQAlloc(ni, 1024, request_out_callback, &request_out_eq);
235         if (rc != PTL_OK)
236                 CERROR("PtlEQAlloc failed: %d\n", rc);
237
238         rc = PtlEQAlloc(ni, 1024, reply_out_callback, &reply_out_eq);
239         if (rc != PTL_OK)
240                 CERROR("PtlEQAlloc failed: %d\n", rc);
241
242         rc = PtlEQAlloc(ni, 1024, reply_in_callback, &reply_in_eq);
243         if (rc != PTL_OK)
244                 CERROR("PtlEQAlloc failed: %d\n", rc);
245
246         rc = PtlEQAlloc(ni, 1024, bulk_source_callback, &bulk_source_eq);
247         if (rc != PTL_OK)
248                 CERROR("PtlEQAlloc failed: %d\n", rc);
249
250         rc = PtlEQAlloc(ni, 1024, bulk_sink_callback, &bulk_sink_eq);
251         if (rc != PTL_OK)
252                 CERROR("PtlEQAlloc failed: %d\n", rc);
253
254         return rc;
255 }
256
257 void ptlrpc_exit_portals(void)
258 {
259         PtlEQFree(request_out_eq);
260         PtlEQFree(reply_out_eq);
261         PtlEQFree(reply_in_eq);
262         PtlEQFree(bulk_source_eq);
263         PtlEQFree(bulk_sink_eq);
264
265         if (qswnal_nip != NULL)
266                 inter_module_put("kqswnal_ni");
267         if (socknal_nip != NULL)
268                 inter_module_put("ksocknal_ni");
269 }