vfio: keep interrupt source read only
[dpdk.git] / lib / librte_eal / linuxapp / eal / eal_alarm.c
1 /*-
2  *   BSD LICENSE
3  *
4  *   Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
5  *   All rights reserved.
6  *
7  *   Redistribution and use in source and binary forms, with or without
8  *   modification, are permitted provided that the following conditions
9  *   are met:
10  *
11  *     * Redistributions of source code must retain the above copyright
12  *       notice, this list of conditions and the following disclaimer.
13  *     * Redistributions in binary form must reproduce the above copyright
14  *       notice, this list of conditions and the following disclaimer in
15  *       the documentation and/or other materials provided with the
16  *       distribution.
17  *     * Neither the name of Intel Corporation nor the names of its
18  *       contributors may be used to endorse or promote products derived
19  *       from this software without specific prior written permission.
20  *
21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32  */
33 #include <stdio.h>
34 #include <stdint.h>
35 #include <signal.h>
36 #include <errno.h>
37 #include <string.h>
38 #include <sys/queue.h>
39 #include <sys/time.h>
40 #include <sys/timerfd.h>
41
42 #include <rte_memory.h>
43 #include <rte_memzone.h>
44 #include <rte_interrupts.h>
45 #include <rte_alarm.h>
46 #include <rte_common.h>
47 #include <rte_per_lcore.h>
48 #include <rte_eal.h>
49 #include <rte_launch.h>
50 #include <rte_lcore.h>
51 #include <rte_errno.h>
52 #include <rte_malloc.h>
53 #include <rte_spinlock.h>
54 #include <eal_private.h>
55
56 #ifndef TFD_NONBLOCK
57 #include <fcntl.h>
58 #define TFD_NONBLOCK    O_NONBLOCK
59 #endif
60
61 #define NS_PER_US 1000
62 #define US_PER_MS 1000
63 #define MS_PER_S 1000
64 #define US_PER_S (US_PER_MS * MS_PER_S)
65
66 #ifdef CLOCK_MONOTONIC_RAW /* Defined in glibc bits/time.h */
67 #define CLOCK_TYPE_ID CLOCK_MONOTONIC_RAW
68 #else
69 #define CLOCK_TYPE_ID CLOCK_MONOTONIC
70 #endif
71
72 struct alarm_entry {
73         LIST_ENTRY(alarm_entry) next;
74         struct timeval time;
75         rte_eal_alarm_callback cb_fn;
76         void *cb_arg;
77         volatile uint8_t executing;
78         volatile pthread_t executing_id;
79 };
80
81 static LIST_HEAD(alarm_list, alarm_entry) alarm_list = LIST_HEAD_INITIALIZER();
82 static rte_spinlock_t alarm_list_lk = RTE_SPINLOCK_INITIALIZER;
83
84 static struct rte_intr_handle intr_handle = {.fd = -1 };
85 static int handler_registered = 0;
86 static void eal_alarm_callback(struct rte_intr_handle *hdl, void *arg);
87
88 int
89 rte_eal_alarm_init(void)
90 {
91         intr_handle.type = RTE_INTR_HANDLE_ALARM;
92         /* create a timerfd file descriptor */
93         intr_handle.fd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK);
94         if (intr_handle.fd == -1)
95                 goto error;
96
97         return 0;
98
99 error:
100         rte_errno = errno;
101         return -1;
102 }
103
104 static void
105 eal_alarm_callback(struct rte_intr_handle *hdl __rte_unused,
106                 void *arg __rte_unused)
107 {
108         struct timespec now;
109         struct alarm_entry *ap;
110
111         rte_spinlock_lock(&alarm_list_lk);
112         while ((ap = LIST_FIRST(&alarm_list)) !=NULL &&
113                         clock_gettime(CLOCK_TYPE_ID, &now) == 0 &&
114                         (ap->time.tv_sec < now.tv_sec || (ap->time.tv_sec == now.tv_sec &&
115                                                 (ap->time.tv_usec * NS_PER_US) <= now.tv_nsec))) {
116                 ap->executing = 1;
117                 ap->executing_id = pthread_self();
118                 rte_spinlock_unlock(&alarm_list_lk);
119
120                 ap->cb_fn(ap->cb_arg);
121
122                 rte_spinlock_lock(&alarm_list_lk);
123
124                 LIST_REMOVE(ap, next);
125                 rte_free(ap);
126         }
127
128         if (!LIST_EMPTY(&alarm_list)) {
129                 struct itimerspec atime = { .it_interval = { 0, 0 } };
130
131                 ap = LIST_FIRST(&alarm_list);
132                 atime.it_value.tv_sec = ap->time.tv_sec;
133                 atime.it_value.tv_nsec = ap->time.tv_usec * NS_PER_US;
134                 /* perform borrow for subtraction if necessary */
135                 if (now.tv_nsec > (ap->time.tv_usec * NS_PER_US))
136                         atime.it_value.tv_sec--, atime.it_value.tv_nsec += US_PER_S * NS_PER_US;
137
138                 atime.it_value.tv_sec -= now.tv_sec;
139                 atime.it_value.tv_nsec -= now.tv_nsec;
140                 timerfd_settime(intr_handle.fd, 0, &atime, NULL);
141         }
142         rte_spinlock_unlock(&alarm_list_lk);
143 }
144
145 int
146 rte_eal_alarm_set(uint64_t us, rte_eal_alarm_callback cb_fn, void *cb_arg)
147 {
148         struct timespec now;
149         int ret = 0;
150         struct alarm_entry *ap, *new_alarm;
151
152         /* Check parameters, including that us won't cause a uint64_t overflow */
153         if (us < 1 || us > (UINT64_MAX - US_PER_S) || cb_fn == NULL)
154                 return -EINVAL;
155
156         new_alarm = rte_zmalloc(NULL, sizeof(*new_alarm), 0);
157         if (new_alarm == NULL)
158                 return -ENOMEM;
159
160         /* use current time to calculate absolute time of alarm */
161         clock_gettime(CLOCK_TYPE_ID, &now);
162
163         new_alarm->cb_fn = cb_fn;
164         new_alarm->cb_arg = cb_arg;
165         new_alarm->time.tv_usec = ((now.tv_nsec / NS_PER_US) + us) % US_PER_S;
166         new_alarm->time.tv_sec = now.tv_sec + (((now.tv_nsec / NS_PER_US) + us) / US_PER_S);
167
168         rte_spinlock_lock(&alarm_list_lk);
169         if (!handler_registered) {
170                 ret |= rte_intr_callback_register(&intr_handle,
171                                 eal_alarm_callback, NULL);
172                 handler_registered = (ret == 0) ? 1 : 0;
173         }
174
175         if (LIST_EMPTY(&alarm_list))
176                 LIST_INSERT_HEAD(&alarm_list, new_alarm, next);
177         else {
178                 LIST_FOREACH(ap, &alarm_list, next) {
179                         if (ap->time.tv_sec > new_alarm->time.tv_sec ||
180                                         (ap->time.tv_sec == new_alarm->time.tv_sec &&
181                                                         ap->time.tv_usec > new_alarm->time.tv_usec)){
182                                 LIST_INSERT_BEFORE(ap, new_alarm, next);
183                                 break;
184                         }
185                         if (LIST_NEXT(ap, next) == NULL) {
186                                 LIST_INSERT_AFTER(ap, new_alarm, next);
187                                 break;
188                         }
189                 }
190         }
191
192         if (LIST_FIRST(&alarm_list) == new_alarm) {
193                 struct itimerspec alarm_time = {
194                         .it_interval = {0, 0},
195                         .it_value = {
196                                 .tv_sec = us / US_PER_S,
197                                 .tv_nsec = (us % US_PER_S) * NS_PER_US,
198                         },
199                 };
200                 ret |= timerfd_settime(intr_handle.fd, 0, &alarm_time, NULL);
201         }
202         rte_spinlock_unlock(&alarm_list_lk);
203
204         return ret;
205 }
206
207 int
208 rte_eal_alarm_cancel(rte_eal_alarm_callback cb_fn, void *cb_arg)
209 {
210         struct alarm_entry *ap, *ap_prev;
211         int count = 0;
212         int err = 0;
213         int executing;
214
215         if (!cb_fn) {
216                 rte_errno = EINVAL;
217                 return -1;
218         }
219
220         do {
221                 executing = 0;
222                 rte_spinlock_lock(&alarm_list_lk);
223                 /* remove any matches at the start of the list */
224                 while ((ap = LIST_FIRST(&alarm_list)) != NULL &&
225                                 cb_fn == ap->cb_fn &&
226                                 (cb_arg == (void *)-1 || cb_arg == ap->cb_arg)) {
227
228                         if (ap->executing == 0) {
229                                 LIST_REMOVE(ap, next);
230                                 rte_free(ap);
231                                 count++;
232                         } else {
233                                 /* If calling from other context, mark that alarm is executing
234                                  * so loop can spin till it finish. Otherwise we are trying to
235                                  * cancel our self - mark it by EINPROGRESS */
236                                 if (pthread_equal(ap->executing_id, pthread_self()) == 0)
237                                         executing++;
238                                 else
239                                         err = EINPROGRESS;
240
241                                 break;
242                         }
243                 }
244                 ap_prev = ap;
245
246                 /* now go through list, removing entries not at start */
247                 LIST_FOREACH(ap, &alarm_list, next) {
248                         /* this won't be true first time through */
249                         if (cb_fn == ap->cb_fn &&
250                                         (cb_arg == (void *)-1 || cb_arg == ap->cb_arg)) {
251
252                                 if (ap->executing == 0) {
253                                         LIST_REMOVE(ap, next);
254                                         rte_free(ap);
255                                         count++;
256                                         ap = ap_prev;
257                                 } else if (pthread_equal(ap->executing_id, pthread_self()) == 0)
258                                         executing++;
259                                 else
260                                         err = EINPROGRESS;
261                         }
262                         ap_prev = ap;
263                 }
264                 rte_spinlock_unlock(&alarm_list_lk);
265         } while (executing != 0);
266
267         if (count == 0 && err == 0)
268                 rte_errno = ENOENT;
269         else if (err)
270                 rte_errno = err;
271
272         return count;
273 }