2 * Copyright (C) 2004 - 2007 Jeff Dike (jdike@{addtoit,linux.intel}.com)
3 * Licensed under the GPL
11 #include <asm/unistd.h>
14 #include <kern_util.h>
17 struct aio_thread_req {
20 unsigned long long offset;
23 struct aio_context *aio;
26 #if defined(HAVE_AIO_ABI)
27 #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);
55 * The AIO_MMAP cases force the mmapped page into memory here
56 * rather than in whatever place first touches the data. I used
57 * to do this by touching the page, but that's delicate because
58 * gcc is prone to optimizing that away. So, what's done here
59 * is we read from the descriptor from which the page was
60 * mapped. The caller is required to pass an offset which is
61 * inside the page that was mapped. Thus, when the read
62 * returns, we know that the page is in the page cache, and
63 * that it now backs the mmapped area.
66 static int do_aio(aio_context_t ctx, enum aio_type type, int fd, char *buf,
67 int len, unsigned long long offset, struct aio_context *aio)
69 struct iocb *iocbp = & ((struct iocb) {
70 .aio_data = (unsigned long) aio,
72 .aio_buf = (unsigned long) buf,
80 iocbp->aio_lio_opcode = IOCB_CMD_PREAD;
83 iocbp->aio_lio_opcode = IOCB_CMD_PWRITE;
86 iocbp->aio_lio_opcode = IOCB_CMD_PREAD;
87 iocbp->aio_buf = (unsigned long) &c;
88 iocbp->aio_nbytes = sizeof(c);
91 printk(UM_KERN_ERR "Bogus op in do_aio - %d\n", type);
95 return (io_submit(ctx, 1, &iocbp) > 0) ? 0 : -errno;
98 /* Initialized in an initcall and unchanged thereafter */
99 static aio_context_t ctx = 0;
101 static int aio_thread(void *arg)
103 struct aio_thread_reply reply;
104 struct io_event event;
105 int err, n, reply_fd;
107 os_fix_helper_signals();
109 n = io_getevents(ctx, 1, 1, &event, NULL);
113 printk(UM_KERN_ERR "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(UM_KERN_ERR "aio_thread - write failed, "
124 "fd = %d, 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(UM_KERN_ERR "do_not_aio - bad request type : %d\n",
163 /* These are initialized in initcalls and not changed */
164 static int aio_req_fd_r = -1;
165 static int aio_req_fd_w = -1;
166 static int aio_pid = -1;
167 static unsigned long aio_stack;
169 static int not_aio_thread(void *arg)
171 struct aio_thread_req req;
172 struct aio_thread_reply reply;
175 os_fix_helper_signals();
177 err = read(aio_req_fd_r, &req, sizeof(req));
178 if (err != sizeof(req)) {
180 printk(UM_KERN_ERR "not_aio_thread - "
181 "read failed, fd = %d, err = %d\n",
185 printk(UM_KERN_ERR "not_aio_thread - short "
186 "read, fd = %d, length = %d\n",
191 err = do_not_aio(&req);
192 reply = ((struct aio_thread_reply) { .data = req.aio,
194 err = write(req.aio->reply_fd, &reply, sizeof(reply));
195 if (err != sizeof(reply))
196 printk(UM_KERN_ERR "not_aio_thread - write failed, "
197 "fd = %d, err = %d\n", req.aio->reply_fd, errno);
203 static int init_aio_24(void)
207 err = os_pipe(fds, 1, 1);
211 aio_req_fd_w = fds[0];
212 aio_req_fd_r = fds[1];
214 err = os_set_fd_block(aio_req_fd_w, 0);
218 err = run_helper_thread(not_aio_thread, NULL,
219 CLONE_FILES | CLONE_VM, &aio_stack);
233 printk(UM_KERN_INFO "/usr/include/linux/aio_abi.h not present during "
236 printk(UM_KERN_INFO "2.6 host AIO support not used - falling back to "
242 #define DEFAULT_24_AIO 0
243 static int init_aio_26(void)
247 if (io_setup(256, &ctx)) {
249 printk(UM_KERN_ERR "aio_thread failed to initialize context, "
250 "err = %d\n", errno);
254 err = run_helper_thread(aio_thread, NULL,
255 CLONE_FILES | CLONE_VM, &aio_stack);
261 printk(UM_KERN_INFO "Using 2.6 host AIO\n");
265 static int submit_aio_26(enum aio_type type, int io_fd, char *buf, int len,
266 unsigned long long offset, struct aio_context *aio)
268 struct aio_thread_reply reply;
271 err = do_aio(ctx, type, io_fd, buf, len, offset, aio);
273 reply = ((struct aio_thread_reply) { .data = aio,
275 err = write(aio->reply_fd, &reply, sizeof(reply));
276 if (err != sizeof(reply)) {
278 printk(UM_KERN_ERR "submit_aio_26 - write failed, "
279 "fd = %d, err = %d\n", aio->reply_fd, -err);
288 #define DEFAULT_24_AIO 1
289 static int init_aio_26(void)
294 static int submit_aio_26(enum aio_type type, int io_fd, char *buf, int len,
295 unsigned long long offset, struct aio_context *aio)
301 /* Initialized in an initcall and unchanged thereafter */
302 static int aio_24 = DEFAULT_24_AIO;
304 static int __init set_aio_24(char *name, int *add)
310 __uml_setup("aio=2.4", set_aio_24,
312 " This is used to force UML to use 2.4-style AIO even when 2.6 AIO is\n"
313 " available. 2.4 AIO is a single thread that handles one request at a\n"
314 " time, synchronously. 2.6 AIO is a thread which uses the 2.6 AIO \n"
315 " interface to handle an arbitrary number of pending requests. 2.6 AIO \n"
316 " is not available in tt mode, on 2.4 hosts, or when UML is built with\n"
317 " /usr/include/linux/aio_abi.h not available. Many distributions don't\n"
318 " include aio_abi.h, so you will need to copy it from a kernel tree to\n"
319 " your /usr/include/linux in order to build an AIO-capable UML\n\n"
322 static int init_aio(void)
328 if (err && (errno == ENOSYS)) {
329 printk(UM_KERN_INFO "2.6 AIO not supported on the "
330 "host - reverting to 2.4 AIO\n");
337 return init_aio_24();
343 * The reason for the __initcall/__uml_exitcall asymmetry is that init_aio
344 * needs to be called when the kernel is running because it calls run_helper,
345 * which needs get_free_page. exit_aio is a __uml_exitcall because the generic
346 * kernel does not run __exitcalls on shutdown, and can't because many of them
347 * break when called outside of module unloading.
349 __initcall(init_aio);
351 static void exit_aio(void)
354 os_kill_process(aio_pid, 1);
355 free_stack(aio_stack, 0);
359 __uml_exitcall(exit_aio);
361 static int submit_aio_24(enum aio_type type, int io_fd, char *buf, int len,
362 unsigned long long offset, struct aio_context *aio)
364 struct aio_thread_req req = { .type = type,
373 err = write(aio_req_fd_w, &req, sizeof(req));
374 if (err == sizeof(req))
381 int submit_aio(enum aio_type type, int io_fd, char *buf, int len,
382 unsigned long long offset, int reply_fd,
383 struct aio_context *aio)
385 aio->reply_fd = reply_fd;
387 return submit_aio_24(type, io_fd, buf, len, offset, aio);
389 return submit_aio_26(type, io_fd, buf, len, offset, aio);