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 signal(SIGWINCH, SIG_IGN);
110 n = io_getevents(ctx, 1, 1, &event, NULL);
114 printk(UM_KERN_ERR "aio_thread - io_getevents failed, "
115 "errno = %d\n", errno);
118 reply = ((struct aio_thread_reply)
119 { .data = (void *) (long) event.data,
121 reply_fd = ((struct aio_context *) reply.data)->reply_fd;
122 err = write(reply_fd, &reply, sizeof(reply));
123 if (err != sizeof(reply))
124 printk(UM_KERN_ERR "aio_thread - write failed, "
125 "fd = %d, err = %d\n", reply_fd, errno);
133 static int do_not_aio(struct aio_thread_req *req)
136 unsigned long long actual;
139 actual = lseek64(req->io_fd, req->offset, SEEK_SET);
140 if (actual != req->offset)
145 n = read(req->io_fd, req->buf, req->len);
148 n = write(req->io_fd, req->buf, req->len);
151 n = read(req->io_fd, &c, sizeof(c));
154 printk(UM_KERN_ERR "do_not_aio - bad request type : %d\n",
164 /* These are initialized in initcalls and not changed */
165 static int aio_req_fd_r = -1;
166 static int aio_req_fd_w = -1;
167 static int aio_pid = -1;
168 static unsigned long aio_stack;
170 static int not_aio_thread(void *arg)
172 struct aio_thread_req req;
173 struct aio_thread_reply reply;
176 signal(SIGWINCH, SIG_IGN);
178 err = read(aio_req_fd_r, &req, sizeof(req));
179 if (err != sizeof(req)) {
181 printk(UM_KERN_ERR "not_aio_thread - "
182 "read failed, fd = %d, err = %d\n",
186 printk(UM_KERN_ERR "not_aio_thread - short "
187 "read, fd = %d, length = %d\n",
192 err = do_not_aio(&req);
193 reply = ((struct aio_thread_reply) { .data = req.aio,
195 err = write(req.aio->reply_fd, &reply, sizeof(reply));
196 if (err != sizeof(reply))
197 printk(UM_KERN_ERR "not_aio_thread - write failed, "
198 "fd = %d, err = %d\n", req.aio->reply_fd, errno);
204 static int init_aio_24(void)
208 err = os_pipe(fds, 1, 1);
212 aio_req_fd_w = fds[0];
213 aio_req_fd_r = fds[1];
215 err = os_set_fd_block(aio_req_fd_w, 0);
219 err = run_helper_thread(not_aio_thread, NULL,
220 CLONE_FILES | CLONE_VM, &aio_stack);
234 printk(UM_KERN_INFO "/usr/include/linux/aio_abi.h not present during "
237 printk(UM_KERN_INFO "2.6 host AIO support not used - falling back to "
243 #define DEFAULT_24_AIO 0
244 static int init_aio_26(void)
248 if (io_setup(256, &ctx)) {
250 printk(UM_KERN_ERR "aio_thread failed to initialize context, "
251 "err = %d\n", errno);
255 err = run_helper_thread(aio_thread, NULL,
256 CLONE_FILES | CLONE_VM, &aio_stack);
262 printk(UM_KERN_INFO "Using 2.6 host AIO\n");
266 static int submit_aio_26(enum aio_type type, int io_fd, char *buf, int len,
267 unsigned long long offset, struct aio_context *aio)
269 struct aio_thread_reply reply;
272 err = do_aio(ctx, type, io_fd, buf, len, offset, aio);
274 reply = ((struct aio_thread_reply) { .data = aio,
276 err = write(aio->reply_fd, &reply, sizeof(reply));
277 if (err != sizeof(reply)) {
279 printk(UM_KERN_ERR "submit_aio_26 - write failed, "
280 "fd = %d, err = %d\n", aio->reply_fd, -err);
289 #define DEFAULT_24_AIO 1
290 static int init_aio_26(void)
295 static int submit_aio_26(enum aio_type type, int io_fd, char *buf, int len,
296 unsigned long long offset, struct aio_context *aio)
302 /* Initialized in an initcall and unchanged thereafter */
303 static int aio_24 = DEFAULT_24_AIO;
305 static int __init set_aio_24(char *name, int *add)
311 __uml_setup("aio=2.4", set_aio_24,
313 " This is used to force UML to use 2.4-style AIO even when 2.6 AIO is\n"
314 " available. 2.4 AIO is a single thread that handles one request at a\n"
315 " time, synchronously. 2.6 AIO is a thread which uses the 2.6 AIO \n"
316 " interface to handle an arbitrary number of pending requests. 2.6 AIO \n"
317 " is not available in tt mode, on 2.4 hosts, or when UML is built with\n"
318 " /usr/include/linux/aio_abi.h not available. Many distributions don't\n"
319 " include aio_abi.h, so you will need to copy it from a kernel tree to\n"
320 " your /usr/include/linux in order to build an AIO-capable UML\n\n"
323 static int init_aio(void)
329 if (err && (errno == ENOSYS)) {
330 printk(UM_KERN_INFO "2.6 AIO not supported on the "
331 "host - reverting to 2.4 AIO\n");
338 return init_aio_24();
344 * The reason for the __initcall/__uml_exitcall asymmetry is that init_aio
345 * needs to be called when the kernel is running because it calls run_helper,
346 * which needs get_free_page. exit_aio is a __uml_exitcall because the generic
347 * kernel does not run __exitcalls on shutdown, and can't because many of them
348 * break when called outside of module unloading.
350 __initcall(init_aio);
352 static void exit_aio(void)
355 os_kill_process(aio_pid, 1);
356 free_stack(aio_stack, 0);
360 __uml_exitcall(exit_aio);
362 static int submit_aio_24(enum aio_type type, int io_fd, char *buf, int len,
363 unsigned long long offset, struct aio_context *aio)
365 struct aio_thread_req req = { .type = type,
374 err = write(aio_req_fd_w, &req, sizeof(req));
375 if (err == sizeof(req))
382 int submit_aio(enum aio_type type, int io_fd, char *buf, int len,
383 unsigned long long offset, int reply_fd,
384 struct aio_context *aio)
386 aio->reply_fd = reply_fd;
388 return submit_aio_24(type, io_fd, buf, len, offset, aio);
390 return submit_aio_26(type, io_fd, buf, len, offset, aio);