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_constants.h"
15 #include "kern_util.h"
19 struct aio_thread_req {
22 unsigned long long offset;
25 struct aio_context *aio;
28 #if defined(HAVE_AIO_ABI)
29 #include <linux/aio_abi.h>
32 * If we have the headers, we are going to build with AIO enabled.
33 * If we don't have aio in libc, we define the necessary stubs here.
36 #if !defined(HAVE_AIO_LIBC)
38 static long io_setup(int n, aio_context_t *ctxp)
40 return syscall(__NR_io_setup, n, ctxp);
43 static long io_submit(aio_context_t ctx, long nr, struct iocb **iocbpp)
45 return syscall(__NR_io_submit, ctx, nr, iocbpp);
48 static long io_getevents(aio_context_t ctx_id, long min_nr, long nr,
49 struct io_event *events, struct timespec *timeout)
51 return syscall(__NR_io_getevents, ctx_id, min_nr, nr, events, timeout);
57 * The AIO_MMAP cases force the mmapped page into memory here
58 * rather than in whatever place first touches the data. I used
59 * to do this by touching the page, but that's delicate because
60 * gcc is prone to optimizing that away. So, what's done here
61 * is we read from the descriptor from which the page was
62 * mapped. The caller is required to pass an offset which is
63 * inside the page that was mapped. Thus, when the read
64 * returns, we know that the page is in the page cache, and
65 * that it now backs the mmapped area.
68 static int do_aio(aio_context_t ctx, enum aio_type type, int fd, char *buf,
69 int len, unsigned long long offset, struct aio_context *aio)
71 struct iocb *iocbp = & ((struct iocb) {
72 .aio_data = (unsigned long) aio,
74 .aio_buf = (unsigned long) buf,
82 iocbp->aio_lio_opcode = IOCB_CMD_PREAD;
85 iocbp->aio_lio_opcode = IOCB_CMD_PWRITE;
88 iocbp->aio_lio_opcode = IOCB_CMD_PREAD;
89 iocbp->aio_buf = (unsigned long) &c;
90 iocbp->aio_nbytes = sizeof(c);
93 printk(UM_KERN_ERR "Bogus op in do_aio - %d\n", type);
97 return (io_submit(ctx, 1, &iocbp) > 0) ? 0 : -errno;
100 /* Initialized in an initcall and unchanged thereafter */
101 static aio_context_t ctx = 0;
103 static int aio_thread(void *arg)
105 struct aio_thread_reply reply;
106 struct io_event event;
107 int err, n, reply_fd;
109 signal(SIGWINCH, SIG_IGN);
112 n = io_getevents(ctx, 1, 1, &event, NULL);
116 printk(UM_KERN_ERR "aio_thread - io_getevents failed, "
117 "errno = %d\n", errno);
120 reply = ((struct aio_thread_reply)
121 { .data = (void *) (long) event.data,
123 reply_fd = ((struct aio_context *) reply.data)->reply_fd;
124 err = write(reply_fd, &reply, sizeof(reply));
125 if (err != sizeof(reply))
126 printk(UM_KERN_ERR "aio_thread - write failed, "
127 "fd = %d, err = %d\n", reply_fd, errno);
135 static int do_not_aio(struct aio_thread_req *req)
138 unsigned long long actual;
141 actual = lseek64(req->io_fd, req->offset, SEEK_SET);
142 if (actual != req->offset)
147 n = read(req->io_fd, req->buf, req->len);
150 n = write(req->io_fd, req->buf, req->len);
153 n = read(req->io_fd, &c, sizeof(c));
156 printk(UM_KERN_ERR "do_not_aio - bad request type : %d\n",
166 /* These are initialized in initcalls and not changed */
167 static int aio_req_fd_r = -1;
168 static int aio_req_fd_w = -1;
169 static int aio_pid = -1;
170 static unsigned long aio_stack;
172 static int not_aio_thread(void *arg)
174 struct aio_thread_req req;
175 struct aio_thread_reply reply;
178 signal(SIGWINCH, SIG_IGN);
180 err = read(aio_req_fd_r, &req, sizeof(req));
181 if (err != sizeof(req)) {
183 printk(UM_KERN_ERR "not_aio_thread - "
184 "read failed, fd = %d, err = %d\n",
188 printk(UM_KERN_ERR "not_aio_thread - short "
189 "read, fd = %d, length = %d\n",
194 err = do_not_aio(&req);
195 reply = ((struct aio_thread_reply) { .data = req.aio,
197 err = write(req.aio->reply_fd, &reply, sizeof(reply));
198 if (err != sizeof(reply))
199 printk(UM_KERN_ERR "not_aio_thread - write failed, "
200 "fd = %d, err = %d\n", req.aio->reply_fd, errno);
206 static int init_aio_24(void)
210 err = os_pipe(fds, 1, 1);
214 aio_req_fd_w = fds[0];
215 aio_req_fd_r = fds[1];
217 err = os_set_fd_block(aio_req_fd_w, 0);
221 err = run_helper_thread(not_aio_thread, NULL,
222 CLONE_FILES | CLONE_VM, &aio_stack);
236 printk(UM_KERN_INFO "/usr/include/linux/aio_abi.h not present during "
239 printk(UM_KERN_INFO "2.6 host AIO support not used - falling back to "
245 #define DEFAULT_24_AIO 0
246 static int init_aio_26(void)
250 if (io_setup(256, &ctx)) {
252 printk(UM_KERN_ERR "aio_thread failed to initialize context, "
253 "err = %d\n", errno);
257 err = run_helper_thread(aio_thread, NULL,
258 CLONE_FILES | CLONE_VM, &aio_stack);
264 printk(UM_KERN_INFO "Using 2.6 host AIO\n");
268 static int submit_aio_26(enum aio_type type, int io_fd, char *buf, int len,
269 unsigned long long offset, struct aio_context *aio)
271 struct aio_thread_reply reply;
274 err = do_aio(ctx, type, io_fd, buf, len, offset, aio);
276 reply = ((struct aio_thread_reply) { .data = aio,
278 err = write(aio->reply_fd, &reply, sizeof(reply));
279 if (err != sizeof(reply)) {
281 printk(UM_KERN_ERR "submit_aio_26 - write failed, "
282 "fd = %d, err = %d\n", aio->reply_fd, -err);
291 #define DEFAULT_24_AIO 1
292 static int init_aio_26(void)
297 static int submit_aio_26(enum aio_type type, int io_fd, char *buf, int len,
298 unsigned long long offset, struct aio_context *aio)
304 /* Initialized in an initcall and unchanged thereafter */
305 static int aio_24 = DEFAULT_24_AIO;
307 static int __init set_aio_24(char *name, int *add)
313 __uml_setup("aio=2.4", set_aio_24,
315 " This is used to force UML to use 2.4-style AIO even when 2.6 AIO is\n"
316 " available. 2.4 AIO is a single thread that handles one request at a\n"
317 " time, synchronously. 2.6 AIO is a thread which uses the 2.6 AIO \n"
318 " interface to handle an arbitrary number of pending requests. 2.6 AIO \n"
319 " is not available in tt mode, on 2.4 hosts, or when UML is built with\n"
320 " /usr/include/linux/aio_abi.h not available. Many distributions don't\n"
321 " include aio_abi.h, so you will need to copy it from a kernel tree to\n"
322 " your /usr/include/linux in order to build an AIO-capable UML\n\n"
325 static int init_aio(void)
331 if (err && (errno == ENOSYS)) {
332 printk(UM_KERN_INFO "2.6 AIO not supported on the "
333 "host - reverting to 2.4 AIO\n");
340 return init_aio_24();
346 * The reason for the __initcall/__uml_exitcall asymmetry is that init_aio
347 * needs to be called when the kernel is running because it calls run_helper,
348 * which needs get_free_page. exit_aio is a __uml_exitcall because the generic
349 * kernel does not run __exitcalls on shutdown, and can't because many of them
350 * break when called outside of module unloading.
352 __initcall(init_aio);
354 static void exit_aio(void)
357 os_kill_process(aio_pid, 1);
358 free_stack(aio_stack, 0);
362 __uml_exitcall(exit_aio);
364 static int submit_aio_24(enum aio_type type, int io_fd, char *buf, int len,
365 unsigned long long offset, struct aio_context *aio)
367 struct aio_thread_req req = { .type = type,
376 err = write(aio_req_fd_w, &req, sizeof(req));
377 if (err == sizeof(req))
384 int submit_aio(enum aio_type type, int io_fd, char *buf, int len,
385 unsigned long long offset, int reply_fd,
386 struct aio_context *aio)
388 aio->reply_fd = reply_fd;
390 return submit_aio_24(type, io_fd, buf, len, offset, aio);
392 return submit_aio_26(type, io_fd, buf, len, offset, aio);