]> CyberLeo.Net >> Repos - FreeBSD/FreeBSD.git/blob - 6/lib/libc_r/uthread/uthread_writev.c
merge fix for boot-time hang on centos' xen
[FreeBSD/FreeBSD.git] / 6 / lib / libc_r / uthread / uthread_writev.c
1 /*
2  * Copyright (c) 1995-1998 John Birrell <jb@cimlogic.com.au>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  * 3. All advertising materials mentioning features or use of this software
14  *    must display the following acknowledgement:
15  *      This product includes software developed by John Birrell.
16  * 4. Neither the name of the author nor the names of any co-contributors
17  *    may be used to endorse or promote products derived from this software
18  *    without specific prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY JOHN BIRRELL AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  *
32  * $FreeBSD$
33  *
34  */
35 #include <sys/types.h>
36 #include <sys/fcntl.h>
37 #include <sys/uio.h>
38 #include <errno.h>
39 #include <stdlib.h>
40 #include <string.h>
41 #include <unistd.h>
42 #include <pthread.h>
43 #include "pthread_private.h"
44
45 __weak_reference(__writev, writev);
46
47 ssize_t
48 _writev(int fd, const struct iovec * iov, int iovcnt)
49 {
50         struct pthread  *curthread = _get_curthread();
51         int     blocking;
52         int     idx = 0;
53         int     type;
54         ssize_t cnt;
55         ssize_t n;
56         ssize_t num = 0;
57         ssize_t ret;
58         struct iovec liov[20];
59         struct iovec *p_iov = liov;
60
61         /* Check if the array size exceeds to compiled in size: */
62         if (iovcnt > (sizeof(liov) / sizeof(struct iovec))) {
63                 /* Allocate memory for the local array: */
64                 if ((p_iov = (struct iovec *)
65                     malloc(iovcnt * sizeof(struct iovec))) == NULL) {
66                         /* Insufficient memory: */
67                         errno = ENOMEM;
68                         return (-1);
69                 }
70         }
71
72         /* Copy the caller's array so that it can be modified locally: */
73         memcpy(p_iov,iov,iovcnt * sizeof(struct iovec));
74
75         /* Lock the file descriptor for write: */
76         if ((ret = _FD_LOCK(fd, FD_WRITE, NULL)) == 0) {
77                 /* Get the read/write mode type: */
78                 type = _thread_fd_getflags(fd) & O_ACCMODE;
79
80                 /* Check if the file is not open for write: */
81                 if (type != O_WRONLY && type != O_RDWR) {
82                         /* File is not open for write: */
83                         errno = EBADF;
84                         _FD_UNLOCK(fd, FD_WRITE);
85                         return (-1);
86                 }
87
88                 /* Check if file operations are to block */
89                 blocking = ((_thread_fd_getflags(fd) & O_NONBLOCK) == 0);
90
91                 /*
92                  * Loop while no error occurs and until the expected number
93                  * of bytes are written if performing a blocking write:
94                  */
95                 while (ret == 0) {
96                         /* Perform a non-blocking write syscall: */
97                         n = __sys_writev(fd, &p_iov[idx], iovcnt - idx);
98
99                         /* Check if one or more bytes were written: */
100                         if (n > 0) {
101                                 /*
102                                  * Keep a count of the number of bytes
103                                  * written:
104                                  */
105                                 num += n;
106
107                                 /*
108                                  * Enter a loop to check if a short write
109                                  * occurred and move the index to the
110                                  * array entry where the short write
111                                  * ended:
112                                  */
113                                 cnt = n;
114                                 while (cnt > 0 && idx < iovcnt) {
115                                         /*
116                                          * If the residual count exceeds
117                                          * the size of this vector, then
118                                          * it was completely written:
119                                          */
120                                         if (cnt >= p_iov[idx].iov_len)
121                                                 /*
122                                                  * Decrement the residual
123                                                  * count and increment the
124                                                  * index to the next array
125                                                  * entry:
126                                                  */
127                                                 cnt -= p_iov[idx++].iov_len;
128                                         else {
129                                                 /*
130                                                  * This entry was only
131                                                  * partially written, so
132                                                  * adjust it's length
133                                                  * and base pointer ready
134                                                  * for the next write:
135                                                  */
136                                                 p_iov[idx].iov_len -= cnt;
137                                                 p_iov[idx].iov_base =
138                                                     (char *)p_iov[idx].iov_base
139                                                     + cnt;
140                                                 cnt = 0;
141                                         }
142                                 }
143                         } else if (n == 0) {
144                                 /*
145                                  * Avoid an infinite loop if the last iov_len is
146                                  * 0.
147                                  */
148                                 while (idx < iovcnt && p_iov[idx].iov_len == 0)
149                                         idx++;
150
151                                 if (idx == iovcnt) {
152                                         ret = num;
153                                         break;
154                                 }
155                         }
156                        
157                         /*
158                          * If performing a blocking write, check if the
159                          * write would have blocked or if some bytes
160                          * were written but there are still more to
161                          * write:
162                          */
163                         if (blocking && ((n < 0 && (errno == EWOULDBLOCK ||
164                             errno == EAGAIN)) || (n >= 0 && idx < iovcnt))) {
165                                 curthread->data.fd.fd = fd;
166                                 _thread_kern_set_timeout(NULL);
167
168                                 /* Reset the interrupted operation flag: */
169                                 curthread->interrupted = 0;
170
171                                 _thread_kern_sched_state(PS_FDW_WAIT,
172                                     __FILE__, __LINE__);
173
174                                 /*
175                                  * Check if the operation was
176                                  * interrupted by a signal
177                                  */
178                                 if (curthread->interrupted) {
179                                         if (num > 0) {
180                                                 /* Return partial success: */
181                                                 ret = num;
182                                         } else {
183                                                 /* Return an error: */
184                                                 errno = EINTR;
185                                                 ret = -1;
186                                         }
187                                 }
188
189                         /*
190                          * If performing a non-blocking write,
191                          * just return whatever the write syscall did:
192                          */
193                         } else if (!blocking) {
194                                 /* A non-blocking call might return zero: */
195                                 ret = n;
196                                 break;
197
198                         /*
199                          * If there was an error, return partial success
200                          * (if any bytes were written) or else the error:
201                          */
202                         } else if (n < 0) {
203                                 if (num > 0)
204                                         ret = num;
205                                 else
206                                         ret = n;
207
208                         /* Check if the write has completed: */
209                         } else if (idx == iovcnt)
210                                 /* Return the number of bytes written: */
211                                 ret = num;
212                 }
213                 _FD_UNLOCK(fd, FD_RDWR);
214         }
215
216         /* If memory was allocated for the array, free it: */
217         if (p_iov != liov)
218                 free(p_iov);
219
220         return (ret);
221 }
222
223 ssize_t
224 __writev(int fd, const struct iovec *iov, int iovcnt)
225 {
226         ssize_t ret;
227
228         _thread_enter_cancellation_point();
229         ret = _writev(fd, iov, iovcnt);
230         _thread_leave_cancellation_point();
231
232         return ret;
233 }