[deliverable/linux.git] / arch / um / os-Linux / aio.c

/*
 * Copyright (C) 2004 - 2007 Jeff Dike (jdike@{addtoit,linux.intel}.com)
 * Licensed under the GPL
 */

#include <unistd.h>
#include <sched.h>
#include <signal.h>
#include <errno.h>
#include <sys/time.h>
#include <asm/unistd.h>
#include <aio.h>
#include <init.h>
#include <kern_util.h>
#include <os.h>

struct aio_thread_req {
	enum aio_type type;
	int io_fd;
	unsigned long long offset;
	char *buf;
	int len;
	struct aio_context *aio;
};

#if defined(HAVE_AIO_ABI)
#include <linux/aio_abi.h>

/*
 * If we have the headers, we are going to build with AIO enabled.
 * If we don't have aio in libc, we define the necessary stubs here.
 */

#if !defined(HAVE_AIO_LIBC)

static long io_setup(int n, aio_context_t *ctxp)
{
	return syscall(__NR_io_setup, n, ctxp);
}

static long io_submit(aio_context_t ctx, long nr, struct iocb **iocbpp)
{
	return syscall(__NR_io_submit, ctx, nr, iocbpp);
}

static long io_getevents(aio_context_t ctx_id, long min_nr, long nr,
			 struct io_event *events, struct timespec *timeout)
{
	return syscall(__NR_io_getevents, ctx_id, min_nr, nr, events, timeout);
}

#endif

/*
 * The AIO_MMAP cases force the mmapped page into memory here
 * rather than in whatever place first touches the data.  I used
 * to do this by touching the page, but that's delicate because
 * gcc is prone to optimizing that away.  So, what's done here
 * is we read from the descriptor from which the page was
 * mapped.  The caller is required to pass an offset which is
 * inside the page that was mapped.  Thus, when the read
 * returns, we know that the page is in the page cache, and
 * that it now backs the mmapped area.
 */

static int do_aio(aio_context_t ctx, enum aio_type type, int fd, char *buf,
		  int len, unsigned long long offset, struct aio_context *aio)
{
	struct iocb *iocbp = & ((struct iocb) {
				    .aio_data       = (unsigned long) aio,
				    .aio_fildes     = fd,
				    .aio_buf        = (unsigned long) buf,
				    .aio_nbytes     = len,
				    .aio_offset     = offset
			     });
	char c;

	switch (type) {
	case AIO_READ:
		iocbp->aio_lio_opcode = IOCB_CMD_PREAD;
		break;
	case AIO_WRITE:
		iocbp->aio_lio_opcode = IOCB_CMD_PWRITE;
		break;
	case AIO_MMAP:
		iocbp->aio_lio_opcode = IOCB_CMD_PREAD;
		iocbp->aio_buf = (unsigned long) &c;
		iocbp->aio_nbytes = sizeof(c);
		break;
	default:
		printk(UM_KERN_ERR "Bogus op in do_aio - %d\n", type);
		return -EINVAL;
	}

	return (io_submit(ctx, 1, &iocbp) > 0) ? 0 : -errno;
}

/* Initialized in an initcall and unchanged thereafter */
static aio_context_t ctx = 0;

static int aio_thread(void *arg)
{
	struct aio_thread_reply reply;
	struct io_event event;
	int err, n, reply_fd;

	signal(SIGWINCH, SIG_IGN);

	while (1) {
		n = io_getevents(ctx, 1, 1, &event, NULL);
		if (n < 0) {
			if (errno == EINTR)
				continue;
			printk(UM_KERN_ERR "aio_thread - io_getevents failed, "
			       "errno = %d\n", errno);
		}
		else {
			reply = ((struct aio_thread_reply)
				{ .data = (void *) (long) event.data,
						.err	= event.res });
			reply_fd = ((struct aio_context *) reply.data)->reply_fd;
			err = write(reply_fd, &reply, sizeof(reply));
			if (err != sizeof(reply))
				printk(UM_KERN_ERR "aio_thread - write failed, "
				       "fd = %d, err = %d\n", reply_fd, errno);
		}
	}
	return 0;
}

#endif

static int do_not_aio(struct aio_thread_req *req)
{
	char c;
	unsigned long long actual;
	int n;

	actual = lseek64(req->io_fd, req->offset, SEEK_SET);
	if (actual != req->offset)
		return -errno;

	switch (req->type) {
	case AIO_READ:
		n = read(req->io_fd, req->buf, req->len);
		break;
	case AIO_WRITE:
		n = write(req->io_fd, req->buf, req->len);
		break;
	case AIO_MMAP:
		n = read(req->io_fd, &c, sizeof(c));
		break;
	default:
		printk(UM_KERN_ERR "do_not_aio - bad request type : %d\n",
		       req->type);
		return -EINVAL;
	}

	if (n < 0)
		return -errno;
	return 0;
}

/* These are initialized in initcalls and not changed */
static int aio_req_fd_r = -1;
static int aio_req_fd_w = -1;
static int aio_pid = -1;
static unsigned long aio_stack;

static int not_aio_thread(void *arg)
{
	struct aio_thread_req req;
	struct aio_thread_reply reply;
	int err;

	signal(SIGWINCH, SIG_IGN);
	while (1) {
		err = read(aio_req_fd_r, &req, sizeof(req));
		if (err != sizeof(req)) {
			if (err < 0)
				printk(UM_KERN_ERR "not_aio_thread - "
				       "read failed, fd = %d, err = %d\n",
				       aio_req_fd_r,
				       errno);
			else {
				printk(UM_KERN_ERR "not_aio_thread - short "
				       "read, fd = %d, length = %d\n",
				       aio_req_fd_r, err);
			}
			continue;
		}
		err = do_not_aio(&req);
		reply = ((struct aio_thread_reply) { .data 	= req.aio,
						     .err	= err });
		err = write(req.aio->reply_fd, &reply, sizeof(reply));
		if (err != sizeof(reply))
			printk(UM_KERN_ERR "not_aio_thread - write failed, "
			       "fd = %d, err = %d\n", req.aio->reply_fd, errno);
	}

	return 0;
}

static int init_aio_24(void)
{
	int fds[2], err;

	err = os_pipe(fds, 1, 1);
	if (err)
		goto out;

	aio_req_fd_w = fds[0];
	aio_req_fd_r = fds[1];

	err = os_set_fd_block(aio_req_fd_w, 0);
	if (err)
		goto out_close_pipe;

	err = run_helper_thread(not_aio_thread, NULL,
				CLONE_FILES | CLONE_VM, &aio_stack);
	if (err < 0)
		goto out_close_pipe;

	aio_pid = err;
	goto out;

out_close_pipe:
	close(fds[0]);
	close(fds[1]);
	aio_req_fd_w = -1;
	aio_req_fd_r = -1;
out:
#ifndef HAVE_AIO_ABI
	printk(UM_KERN_INFO "/usr/include/linux/aio_abi.h not present during "
	       "build\n");
#endif
	printk(UM_KERN_INFO "2.6 host AIO support not used - falling back to "
	       "I/O thread\n");
	return 0;
}

#ifdef HAVE_AIO_ABI
#define DEFAULT_24_AIO 0
static int init_aio_26(void)
{
	int err;

	if (io_setup(256, &ctx)) {
		err = -errno;
		printk(UM_KERN_ERR "aio_thread failed to initialize context, "
		       "err = %d\n", errno);
		return err;
	}

	err = run_helper_thread(aio_thread, NULL,
				CLONE_FILES | CLONE_VM, &aio_stack);
	if (err < 0)
		return err;

	aio_pid = err;

	printk(UM_KERN_INFO "Using 2.6 host AIO\n");
	return 0;
}

static int submit_aio_26(enum aio_type type, int io_fd, char *buf, int len,
			 unsigned long long offset, struct aio_context *aio)
{
	struct aio_thread_reply reply;
	int err;

	err = do_aio(ctx, type, io_fd, buf, len, offset, aio);
	if (err) {
		reply = ((struct aio_thread_reply) { .data = aio,
					 .err  = err });
		err = write(aio->reply_fd, &reply, sizeof(reply));
		if (err != sizeof(reply)) {
			err = -errno;
			printk(UM_KERN_ERR "submit_aio_26 - write failed, "
			       "fd = %d, err = %d\n", aio->reply_fd, -err);
		}
		else err = 0;
	}

	return err;
}

#else
#define DEFAULT_24_AIO 1
static int init_aio_26(void)
{
	return -ENOSYS;
}

static int submit_aio_26(enum aio_type type, int io_fd, char *buf, int len,
			 unsigned long long offset, struct aio_context *aio)
{
	return -ENOSYS;
}
#endif

/* Initialized in an initcall and unchanged thereafter */
static int aio_24 = DEFAULT_24_AIO;

static int __init set_aio_24(char *name, int *add)
{
	aio_24 = 1;
	return 0;
}

__uml_setup("aio=2.4", set_aio_24,
"aio=2.4\n"
"    This is used to force UML to use 2.4-style AIO even when 2.6 AIO is\n"
"    available.  2.4 AIO is a single thread that handles one request at a\n"
"    time, synchronously.  2.6 AIO is a thread which uses the 2.6 AIO \n"
"    interface to handle an arbitrary number of pending requests.  2.6 AIO \n"
"    is not available in tt mode, on 2.4 hosts, or when UML is built with\n"
"    /usr/include/linux/aio_abi.h not available.  Many distributions don't\n"
"    include aio_abi.h, so you will need to copy it from a kernel tree to\n"
"    your /usr/include/linux in order to build an AIO-capable UML\n\n"
);

static int init_aio(void)
{
	int err;

	if (!aio_24) {
		err = init_aio_26();
		if (err && (errno == ENOSYS)) {
			printk(UM_KERN_INFO "2.6 AIO not supported on the "
			       "host - reverting to 2.4 AIO\n");
			aio_24 = 1;
		}
		else return err;
	}

	if (aio_24)
		return init_aio_24();

	return 0;
}

/*
 * The reason for the __initcall/__uml_exitcall asymmetry is that init_aio
 * needs to be called when the kernel is running because it calls run_helper,
 * which needs get_free_page.  exit_aio is a __uml_exitcall because the generic
 * kernel does not run __exitcalls on shutdown, and can't because many of them
 * break when called outside of module unloading.
 */
__initcall(init_aio);

static void exit_aio(void)
{
	if (aio_pid != -1) {
		os_kill_process(aio_pid, 1);
		free_stack(aio_stack, 0);
	}
}

__uml_exitcall(exit_aio);

static int submit_aio_24(enum aio_type type, int io_fd, char *buf, int len,
			 unsigned long long offset, struct aio_context *aio)
{
	struct aio_thread_req req = { .type 		= type,
				      .io_fd		= io_fd,
				      .offset		= offset,
				      .buf		= buf,
				      .len		= len,
				      .aio		= aio,
	};
	int err;

	err = write(aio_req_fd_w, &req, sizeof(req));
	if (err == sizeof(req))
		err = 0;
	else err = -errno;

	return err;
}

int submit_aio(enum aio_type type, int io_fd, char *buf, int len,
	       unsigned long long offset, int reply_fd,
	       struct aio_context *aio)
{
	aio->reply_fd = reply_fd;
	if (aio_24)
		return submit_aio_24(type, io_fd, buf, len, offset, aio);
	else
		return submit_aio_26(type, io_fd, buf, len, offset, aio);
}
Commit	Line	Data
75e5584c	1	/*
ba180fd4	2	* Copyright (C) 2004 - 2007 Jeff Dike (jdike@{addtoit,linux.intel}.com)
75e5584c JD	3	* Licensed under the GPL
	4	*/
	5
75e5584c	6	#include <unistd.h>
ba180fd4	7	#include <sched.h>
75e5584c JD	8	#include <signal.h>
75e5584c JD	9	#include <errno.h>
ba180fd4 JD	10	#include <sys/time.h>
ba180fd4 JD	11	#include <asm/unistd.h>
37185b33 AV	12	#include <aio.h>
	13	#include <init.h>
	14	#include <kern_util.h>
	15	#include <os.h>
75e5584c	16
91acb21f	17	struct aio_thread_req {
d50084a2 JD	18	enum aio_type type;
	19	int io_fd;
	20	unsigned long long offset;
	21	char *buf;
	22	int len;
	23	struct aio_context *aio;
91acb21f JD	24	};
91acb21f JD	25
75e5584c JD	26	#if defined(HAVE_AIO_ABI)
	27	#include <linux/aio_abi.h>
	28
ba180fd4 JD	29	/*
ba180fd4 JD	30	* If we have the headers, we are going to build with AIO enabled.
75e5584c JD	31	* If we don't have aio in libc, we define the necessary stubs here.
	32	*/
	33
	34	#if !defined(HAVE_AIO_LIBC)
	35
	36	static long io_setup(int n, aio_context_t *ctxp)
	37	{
d50084a2	38	return syscall(__NR_io_setup, n, ctxp);
75e5584c JD	39	}
	40
	41	static long io_submit(aio_context_t ctx, long nr, struct iocb **iocbpp)
	42	{
d50084a2	43	return syscall(__NR_io_submit, ctx, nr, iocbpp);
75e5584c JD	44	}
	45
	46	static long io_getevents(aio_context_t ctx_id, long min_nr, long nr,
d50084a2	47	struct io_event events, struct timespec timeout)
75e5584c	48	{
d50084a2	49	return syscall(__NR_io_getevents, ctx_id, min_nr, nr, events, timeout);
75e5584c JD	50	}
	51
	52	#endif
	53
ba180fd4 JD	54	/*
ba180fd4 JD	55	* The AIO_MMAP cases force the mmapped page into memory here
75e5584c JD	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.
	64	*/
	65
91acb21f	66	static int do_aio(aio_context_t ctx, enum aio_type type, int fd, char *buf,
d50084a2	67	int len, unsigned long long offset, struct aio_context *aio)
75e5584c	68	{
da3e30e7 JD	69	struct iocb *iocbp = & ((struct iocb) {
	70	.aio_data = (unsigned long) aio,
	71	.aio_fildes = fd,
	72	.aio_buf = (unsigned long) buf,
	73	.aio_nbytes = len,
	74	.aio_offset = offset
	75	});
d50084a2	76	char c;
da3e30e7 JD	77
da3e30e7 JD	78	switch (type) {
d50084a2	79	case AIO_READ:
da3e30e7	80	iocbp->aio_lio_opcode = IOCB_CMD_PREAD;
d50084a2 JD	81	break;
d50084a2 JD	82	case AIO_WRITE:
da3e30e7	83	iocbp->aio_lio_opcode = IOCB_CMD_PWRITE;
d50084a2 JD	84	break;
d50084a2 JD	85	case AIO_MMAP:
da3e30e7 JD	86	iocbp->aio_lio_opcode = IOCB_CMD_PREAD;
	87	iocbp->aio_buf = (unsigned long) &c;
	88	iocbp->aio_nbytes = sizeof(c);
d50084a2 JD	89	break;
d50084a2 JD	90	default:
da3e30e7 JD	91	printk(UM_KERN_ERR "Bogus op in do_aio - %d\n", type);
da3e30e7 JD	92	return -EINVAL;
d50084a2 JD	93	}
d50084a2 JD	94
da3e30e7	95	return (io_submit(ctx, 1, &iocbp) > 0) ? 0 : -errno;
75e5584c JD	96	}
75e5584c JD	97
9683da91	98	/* Initialized in an initcall and unchanged thereafter */
75e5584c JD	99	static aio_context_t ctx = 0;
	100
	101	static int aio_thread(void *arg)
	102	{
d50084a2 JD	103	struct aio_thread_reply reply;
	104	struct io_event event;
	105	int err, n, reply_fd;
	106
	107	signal(SIGWINCH, SIG_IGN);
	108
ba180fd4	109	while (1) {
d50084a2	110	n = io_getevents(ctx, 1, 1, &event, NULL);
ba180fd4 JD	111	if (n < 0) {
ba180fd4 JD	112	if (errno == EINTR)
d50084a2	113	continue;
ba180fd4	114	printk(UM_KERN_ERR "aio_thread - io_getevents failed, "
d50084a2 JD	115	"errno = %d\n", errno);
	116	}
	117	else {
	118	reply = ((struct aio_thread_reply)
	119	{ .data = (void *) (long) event.data,
	120	.err = event.res });
91acb21f	121	reply_fd = ((struct aio_context *) reply.data)->reply_fd;
a61f334f	122	err = write(reply_fd, &reply, sizeof(reply));
ba180fd4 JD	123	if (err != sizeof(reply))
	124	printk(UM_KERN_ERR "aio_thread - write failed, "
	125	"fd = %d, err = %d\n", reply_fd, errno);
d50084a2 JD	126	}
	127	}
	128	return 0;
75e5584c JD	129	}
	130
	131	#endif
	132
91acb21f	133	static int do_not_aio(struct aio_thread_req *req)
75e5584c	134	{
d50084a2	135	char c;
ef0470c0	136	unsigned long long actual;
a61f334f	137	int n;
d50084a2	138
ef0470c0	139	actual = lseek64(req->io_fd, req->offset, SEEK_SET);
ba180fd4	140	if (actual != req->offset)
ef0470c0 JD	141	return -errno;
ef0470c0 JD	142
5134d8fe	143	switch (req->type) {
d50084a2	144	case AIO_READ:
a61f334f	145	n = read(req->io_fd, req->buf, req->len);
d50084a2 JD	146	break;
d50084a2 JD	147	case AIO_WRITE:
a61f334f	148	n = write(req->io_fd, req->buf, req->len);
d50084a2 JD	149	break;
d50084a2 JD	150	case AIO_MMAP:
a61f334f	151	n = read(req->io_fd, &c, sizeof(c));
d50084a2 JD	152	break;
d50084a2 JD	153	default:
ba180fd4 JD	154	printk(UM_KERN_ERR "do_not_aio - bad request type : %d\n",
ba180fd4 JD	155	req->type);
a61f334f	156	return -EINVAL;
d50084a2 JD	157	}
d50084a2 JD	158
ba180fd4	159	if (n < 0)
a61f334f JD	160	return -errno;
a61f334f JD	161	return 0;
75e5584c JD	162	}
75e5584c JD	163
9683da91 JD	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;
c4399016	168	static unsigned long aio_stack;
9683da91	169
75e5584c JD	170	static int not_aio_thread(void *arg)
75e5584c JD	171	{
d50084a2 JD	172	struct aio_thread_req req;
	173	struct aio_thread_reply reply;
	174	int err;
	175
	176	signal(SIGWINCH, SIG_IGN);
ba180fd4	177	while (1) {
a61f334f	178	err = read(aio_req_fd_r, &req, sizeof(req));
ba180fd4 JD	179	if (err != sizeof(req)) {
	180	if (err < 0)
	181	printk(UM_KERN_ERR "not_aio_thread - "
	182	"read failed, fd = %d, err = %d\n",
	183	aio_req_fd_r,
a61f334f	184	errno);
d50084a2	185	else {
ba180fd4 JD	186	printk(UM_KERN_ERR "not_aio_thread - short "
	187	"read, fd = %d, length = %d\n",
	188	aio_req_fd_r, err);
d50084a2 JD	189	}
	190	continue;
	191	}
	192	err = do_not_aio(&req);
	193	reply = ((struct aio_thread_reply) { .data = req.aio,
ef0470c0	194	.err = err });
a61f334f	195	err = write(req.aio->reply_fd, &reply, sizeof(reply));
ba180fd4 JD	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);
d50084a2	199	}
1b57e9c2 JD	200
1b57e9c2 JD	201	return 0;
75e5584c JD	202	}
75e5584c JD	203
75e5584c JD	204	static int init_aio_24(void)
75e5584c JD	205	{
d50084a2 JD	206	int fds[2], err;
	207
	208	err = os_pipe(fds, 1, 1);
ba180fd4	209	if (err)
d50084a2 JD	210	goto out;
	211
	212	aio_req_fd_w = fds[0];
	213	aio_req_fd_r = fds[1];
8603ec81 JD	214
8603ec81 JD	215	err = os_set_fd_block(aio_req_fd_w, 0);
ba180fd4	216	if (err)
8603ec81 JD	217	goto out_close_pipe;
8603ec81 JD	218
d50084a2	219	err = run_helper_thread(not_aio_thread, NULL,
4dbed85a	220	CLONE_FILES \| CLONE_VM, &aio_stack);
ba180fd4	221	if (err < 0)
d50084a2 JD	222	goto out_close_pipe;
	223
	224	aio_pid = err;
	225	goto out;
	226
	227	out_close_pipe:
512b6fb1 JD	228	close(fds[0]);
512b6fb1 JD	229	close(fds[1]);
d50084a2 JD	230	aio_req_fd_w = -1;
	231	aio_req_fd_r = -1;
	232	out:
75e5584c	233	#ifndef HAVE_AIO_ABI
ba180fd4 JD	234	printk(UM_KERN_INFO "/usr/include/linux/aio_abi.h not present during "
ba180fd4 JD	235	"build\n");
75e5584c	236	#endif
ba180fd4 JD	237	printk(UM_KERN_INFO "2.6 host AIO support not used - falling back to "
ba180fd4 JD	238	"I/O thread\n");
d50084a2	239	return 0;
75e5584c JD	240	}
	241
	242	#ifdef HAVE_AIO_ABI
	243	#define DEFAULT_24_AIO 0
	244	static int init_aio_26(void)
	245	{
d50084a2	246	int err;
75e5584c	247
ba180fd4	248	if (io_setup(256, &ctx)) {
b4fd310e	249	err = -errno;
ba180fd4 JD	250	printk(UM_KERN_ERR "aio_thread failed to initialize context, "
ba180fd4 JD	251	"err = %d\n", errno);
d50084a2 JD	252	return err;
d50084a2 JD	253	}
75e5584c	254
d50084a2	255	err = run_helper_thread(aio_thread, NULL,
4dbed85a	256	CLONE_FILES \| CLONE_VM, &aio_stack);
ba180fd4	257	if (err < 0)
d50084a2	258	return err;
75e5584c	259
d50084a2	260	aio_pid = err;
75e5584c	261
ba180fd4	262	printk(UM_KERN_INFO "Using 2.6 host AIO\n");
d50084a2	263	return 0;
91acb21f JD	264	}
	265
	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)
	268	{
d50084a2 JD	269	struct aio_thread_reply reply;
	270	int err;
	271
	272	err = do_aio(ctx, type, io_fd, buf, len, offset, aio);
ba180fd4	273	if (err) {
d50084a2 JD	274	reply = ((struct aio_thread_reply) { .data = aio,
d50084a2 JD	275	.err = err });
a61f334f	276	err = write(aio->reply_fd, &reply, sizeof(reply));
ba180fd4	277	if (err != sizeof(reply)) {
a61f334f	278	err = -errno;
ba180fd4	279	printk(UM_KERN_ERR "submit_aio_26 - write failed, "
d50084a2	280	"fd = %d, err = %d\n", aio->reply_fd, -err);
a61f334f	281	}
d50084a2 JD	282	else err = 0;
	283	}
	284
	285	return err;
75e5584c JD	286	}
	287
	288	#else
	289	#define DEFAULT_24_AIO 1
91acb21f	290	static int init_aio_26(void)
75e5584c	291	{
d50084a2	292	return -ENOSYS;
75e5584c JD	293	}
75e5584c JD	294
91acb21f JD	295	static int submit_aio_26(enum aio_type type, int io_fd, char *buf, int len,
91acb21f JD	296	unsigned long long offset, struct aio_context *aio)
75e5584c	297	{
d50084a2	298	return -ENOSYS;
75e5584c JD	299	}
	300	#endif
	301
9683da91	302	/* Initialized in an initcall and unchanged thereafter */
75e5584c JD	303	static int aio_24 = DEFAULT_24_AIO;
	304
	305	static int __init set_aio_24(char name, int add)
	306	{
d50084a2 JD	307	aio_24 = 1;
d50084a2 JD	308	return 0;
75e5584c JD	309	}
	310
	311	__uml_setup("aio=2.4", set_aio_24,
	312	"aio=2.4\n"
	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"
	321	);
	322
	323	static int init_aio(void)
	324	{
d50084a2 JD	325	int err;
d50084a2 JD	326
ba180fd4	327	if (!aio_24) {
d50084a2	328	err = init_aio_26();
ba180fd4 JD	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");
d50084a2 JD	332	aio_24 = 1;
	333	}
	334	else return err;
	335	}
	336
ba180fd4	337	if (aio_24)
d50084a2 JD	338	return init_aio_24();
	339
	340	return 0;
75e5584c JD	341	}
75e5584c JD	342
ba180fd4 JD	343	/*
ba180fd4 JD	344	* The reason for the __initcall/__uml_exitcall asymmetry is that init_aio
75e5584c JD	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.
	349	*/
	350	__initcall(init_aio);
	351
	352	static void exit_aio(void)
	353	{
c4399016	354	if (aio_pid != -1) {
d50084a2	355	os_kill_process(aio_pid, 1);
c4399016 JD	356	free_stack(aio_stack, 0);
c4399016 JD	357	}
75e5584c JD	358	}
	359
	360	__uml_exitcall(exit_aio);
	361
91acb21f JD	362	static int submit_aio_24(enum aio_type type, int io_fd, char *buf, int len,
91acb21f JD	363	unsigned long long offset, struct aio_context *aio)
75e5584c	364	{
d50084a2 JD	365	struct aio_thread_req req = { .type = type,
	366	.io_fd = io_fd,
	367	.offset = offset,
	368	.buf = buf,
	369	.len = len,
	370	.aio = aio,
	371	};
	372	int err;
	373
a61f334f	374	err = write(aio_req_fd_w, &req, sizeof(req));
ba180fd4	375	if (err == sizeof(req))
d50084a2	376	err = 0;
a61f334f	377	else err = -errno;
d50084a2 JD	378
d50084a2 JD	379	return err;
91acb21f JD	380	}
	381
	382	int submit_aio(enum aio_type type, int io_fd, char *buf, int len,
d50084a2 JD	383	unsigned long long offset, int reply_fd,
d50084a2 JD	384	struct aio_context *aio)
91acb21f	385	{
d50084a2	386	aio->reply_fd = reply_fd;
ba180fd4	387	if (aio_24)
d50084a2	388	return submit_aio_24(type, io_fd, buf, len, offset, aio);
ba180fd4	389	else
d50084a2	390	return submit_aio_26(type, io_fd, buf, len, offset, aio);
75e5584c	391	}