3 * Licensed under the GPL
11 #include <sys/syscall.h>
16 #include "kern_constants.h"
18 struct aio_thread_req {
21 unsigned long long offset;
24 struct aio_context *aio;
27 #if defined(HAVE_AIO_ABI)
28 #include <linux/aio_abi.h>
30 /* If we have the headers, we are going to build with AIO enabled.
31 * If we don't have aio in libc, we define the necessary stubs here.
34 #if !defined(HAVE_AIO_LIBC)
36 static long io_setup(int n, aio_context_t *ctxp)
38 return syscall(__NR_io_setup, n, ctxp);
41 static long io_submit(aio_context_t ctx, long nr, struct iocb **iocbpp)
43 return syscall(__NR_io_submit, ctx, nr, iocbpp);
46 static long io_getevents(aio_context_t ctx_id, long min_nr, long nr,
47 struct io_event *events, struct timespec *timeout)
49 return syscall(__NR_io_getevents, ctx_id, min_nr, nr, events, timeout);
54 /* The AIO_MMAP cases force the mmapped page into memory here
55 * rather than in whatever place first touches the data. I used
56 * to do this by touching the page, but that's delicate because
57 * gcc is prone to optimizing that away. So, what's done here
58 * is we read from the descriptor from which the page was
59 * mapped. The caller is required to pass an offset which is
60 * inside the page that was mapped. Thus, when the read
61 * returns, we know that the page is in the page cache, and
62 * that it now backs the mmapped area.
65 static int do_aio(aio_context_t ctx, enum aio_type type, int fd, char *buf,
66 int len, unsigned long long offset, struct aio_context *aio)
68 struct iocb *iocbp = & ((struct iocb) {
69 .aio_data = (unsigned long) aio,
71 .aio_buf = (unsigned long) buf,
79 iocbp->aio_lio_opcode = IOCB_CMD_PREAD;
82 iocbp->aio_lio_opcode = IOCB_CMD_PWRITE;
85 iocbp->aio_lio_opcode = IOCB_CMD_PREAD;
86 iocbp->aio_buf = (unsigned long) &c;
87 iocbp->aio_nbytes = sizeof(c);
90 printk(UM_KERN_ERR "Bogus op in do_aio - %d\n", type);
94 return (io_submit(ctx, 1, &iocbp) > 0) ? 0 : -errno;
97 /* Initialized in an initcall and unchanged thereafter */
98 static aio_context_t ctx = 0;
100 static int aio_thread(void *arg)
102 struct aio_thread_reply reply;
103 struct io_event event;
104 int err, n, reply_fd;
106 signal(SIGWINCH, SIG_IGN);
109 n = io_getevents(ctx, 1, 1, &event, NULL);
113 printk("aio_thread - io_getevents failed, "
114 "errno = %d\n", errno);
117 reply = ((struct aio_thread_reply)
118 { .data = (void *) (long) event.data,
120 reply_fd = ((struct aio_context *) reply.data)->reply_fd;
121 err = write(reply_fd, &reply, sizeof(reply));
122 if(err != sizeof(reply))
123 printk("aio_thread - write failed, fd = %d, "
124 "err = %d\n", reply_fd, errno);
132 static int do_not_aio(struct aio_thread_req *req)
135 unsigned long long actual;
138 actual = lseek64(req->io_fd, req->offset, SEEK_SET);
139 if(actual != req->offset)
144 n = read(req->io_fd, req->buf, req->len);
147 n = write(req->io_fd, req->buf, req->len);
150 n = read(req->io_fd, &c, sizeof(c));
153 printk("do_not_aio - bad request type : %d\n", req->type);
162 /* These are initialized in initcalls and not changed */
163 static int aio_req_fd_r = -1;
164 static int aio_req_fd_w = -1;
165 static int aio_pid = -1;
166 static unsigned long aio_stack;
168 static int not_aio_thread(void *arg)
170 struct aio_thread_req req;
171 struct aio_thread_reply reply;
174 signal(SIGWINCH, SIG_IGN);
176 err = read(aio_req_fd_r, &req, sizeof(req));
177 if(err != sizeof(req)){
179 printk("not_aio_thread - read failed, "
180 "fd = %d, err = %d\n", aio_req_fd_r,
183 printk("not_aio_thread - short read, fd = %d, "
184 "length = %d\n", aio_req_fd_r, err);
188 err = do_not_aio(&req);
189 reply = ((struct aio_thread_reply) { .data = req.aio,
191 err = write(req.aio->reply_fd, &reply, sizeof(reply));
192 if(err != sizeof(reply))
193 printk("not_aio_thread - write failed, fd = %d, "
194 "err = %d\n", req.aio->reply_fd, errno);
200 static int init_aio_24(void)
204 err = os_pipe(fds, 1, 1);
208 aio_req_fd_w = fds[0];
209 aio_req_fd_r = fds[1];
211 err = os_set_fd_block(aio_req_fd_w, 0);
215 err = run_helper_thread(not_aio_thread, NULL,
216 CLONE_FILES | CLONE_VM | SIGCHLD, &aio_stack);
224 os_close_file(fds[0]);
225 os_close_file(fds[1]);
230 printk("/usr/include/linux/aio_abi.h not present during build\n");
232 printk("2.6 host AIO support not used - falling back to I/O "
238 #define DEFAULT_24_AIO 0
239 static int init_aio_26(void)
243 if(io_setup(256, &ctx)){
245 printk("aio_thread failed to initialize context, err = %d\n",
250 err = run_helper_thread(aio_thread, NULL,
251 CLONE_FILES | CLONE_VM | SIGCHLD, &aio_stack);
257 printk("Using 2.6 host AIO\n");
261 static int submit_aio_26(enum aio_type type, int io_fd, char *buf, int len,
262 unsigned long long offset, struct aio_context *aio)
264 struct aio_thread_reply reply;
267 err = do_aio(ctx, type, io_fd, buf, len, offset, aio);
269 reply = ((struct aio_thread_reply) { .data = aio,
271 err = write(aio->reply_fd, &reply, sizeof(reply));
272 if(err != sizeof(reply)){
274 printk("submit_aio_26 - write failed, "
275 "fd = %d, err = %d\n", aio->reply_fd, -err);
284 #define DEFAULT_24_AIO 1
285 static int init_aio_26(void)
290 static int submit_aio_26(enum aio_type type, int io_fd, char *buf, int len,
291 unsigned long long offset, struct aio_context *aio)
297 /* Initialized in an initcall and unchanged thereafter */
298 static int aio_24 = DEFAULT_24_AIO;
300 static int __init set_aio_24(char *name, int *add)
306 __uml_setup("aio=2.4", set_aio_24,
308 " This is used to force UML to use 2.4-style AIO even when 2.6 AIO is\n"
309 " available. 2.4 AIO is a single thread that handles one request at a\n"
310 " time, synchronously. 2.6 AIO is a thread which uses the 2.6 AIO \n"
311 " interface to handle an arbitrary number of pending requests. 2.6 AIO \n"
312 " is not available in tt mode, on 2.4 hosts, or when UML is built with\n"
313 " /usr/include/linux/aio_abi.h not available. Many distributions don't\n"
314 " include aio_abi.h, so you will need to copy it from a kernel tree to\n"
315 " your /usr/include/linux in order to build an AIO-capable UML\n\n"
318 static int init_aio(void)
324 if(err && (errno == ENOSYS)){
325 printk("2.6 AIO not supported on the host - "
326 "reverting to 2.4 AIO\n");
333 return init_aio_24();
338 /* The reason for the __initcall/__uml_exitcall asymmetry is that init_aio
339 * needs to be called when the kernel is running because it calls run_helper,
340 * which needs get_free_page. exit_aio is a __uml_exitcall because the generic
341 * kernel does not run __exitcalls on shutdown, and can't because many of them
342 * break when called outside of module unloading.
344 __initcall(init_aio);
346 static void exit_aio(void)
349 os_kill_process(aio_pid, 1);
350 free_stack(aio_stack, 0);
354 __uml_exitcall(exit_aio);
356 static int submit_aio_24(enum aio_type type, int io_fd, char *buf, int len,
357 unsigned long long offset, struct aio_context *aio)
359 struct aio_thread_req req = { .type = type,
368 err = write(aio_req_fd_w, &req, sizeof(req));
369 if(err == sizeof(req))
376 int submit_aio(enum aio_type type, int io_fd, char *buf, int len,
377 unsigned long long offset, int reply_fd,
378 struct aio_context *aio)
380 aio->reply_fd = reply_fd;
382 return submit_aio_24(type, io_fd, buf, len, offset, aio);
384 return submit_aio_26(type, io_fd, buf, len, offset, aio);