[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_constants.h"
#include "kern_util.h"
#include "os.h"
#include "user.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>
75e5584c JD	12	#include "aio.h"
75e5584c JD	13	#include "init.h"
da3e30e7	14	#include "kern_constants.h"
edea1385	15	#include "kern_util.h"
ba180fd4 JD	16	#include "os.h"
ba180fd4 JD	17	#include "user.h"
75e5584c	18
91acb21f	19	struct aio_thread_req {
d50084a2 JD	20	enum aio_type type;
	21	int io_fd;
	22	unsigned long long offset;
	23	char *buf;
	24	int len;
	25	struct aio_context *aio;
91acb21f JD	26	};
91acb21f JD	27
75e5584c JD	28	#if defined(HAVE_AIO_ABI)
	29	#include <linux/aio_abi.h>
	30
ba180fd4 JD	31	/*
ba180fd4 JD	32	* If we have the headers, we are going to build with AIO enabled.
75e5584c JD	33	* If we don't have aio in libc, we define the necessary stubs here.
	34	*/
	35
	36	#if !defined(HAVE_AIO_LIBC)
	37
	38	static long io_setup(int n, aio_context_t *ctxp)
	39	{
d50084a2	40	return syscall(__NR_io_setup, n, ctxp);
75e5584c JD	41	}
	42
	43	static long io_submit(aio_context_t ctx, long nr, struct iocb **iocbpp)
	44	{
d50084a2	45	return syscall(__NR_io_submit, ctx, nr, iocbpp);
75e5584c JD	46	}
	47
	48	static long io_getevents(aio_context_t ctx_id, long min_nr, long nr,
d50084a2	49	struct io_event events, struct timespec timeout)
75e5584c	50	{
d50084a2	51	return syscall(__NR_io_getevents, ctx_id, min_nr, nr, events, timeout);
75e5584c JD	52	}
	53
	54	#endif
	55
ba180fd4 JD	56	/*
ba180fd4 JD	57	* The AIO_MMAP cases force the mmapped page into memory here
75e5584c JD	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.
	66	*/
	67
91acb21f	68	static int do_aio(aio_context_t ctx, enum aio_type type, int fd, char *buf,
d50084a2	69	int len, unsigned long long offset, struct aio_context *aio)
75e5584c	70	{
da3e30e7 JD	71	struct iocb *iocbp = & ((struct iocb) {
	72	.aio_data = (unsigned long) aio,
	73	.aio_fildes = fd,
	74	.aio_buf = (unsigned long) buf,
	75	.aio_nbytes = len,
	76	.aio_offset = offset
	77	});
d50084a2	78	char c;
da3e30e7 JD	79
da3e30e7 JD	80	switch (type) {
d50084a2	81	case AIO_READ:
da3e30e7	82	iocbp->aio_lio_opcode = IOCB_CMD_PREAD;
d50084a2 JD	83	break;
d50084a2 JD	84	case AIO_WRITE:
da3e30e7	85	iocbp->aio_lio_opcode = IOCB_CMD_PWRITE;
d50084a2 JD	86	break;
d50084a2 JD	87	case AIO_MMAP:
da3e30e7 JD	88	iocbp->aio_lio_opcode = IOCB_CMD_PREAD;
	89	iocbp->aio_buf = (unsigned long) &c;
	90	iocbp->aio_nbytes = sizeof(c);
d50084a2 JD	91	break;
d50084a2 JD	92	default:
da3e30e7 JD	93	printk(UM_KERN_ERR "Bogus op in do_aio - %d\n", type);
da3e30e7 JD	94	return -EINVAL;
d50084a2 JD	95	}
d50084a2 JD	96
da3e30e7	97	return (io_submit(ctx, 1, &iocbp) > 0) ? 0 : -errno;
75e5584c JD	98	}
75e5584c JD	99
9683da91	100	/* Initialized in an initcall and unchanged thereafter */
75e5584c JD	101	static aio_context_t ctx = 0;
	102
	103	static int aio_thread(void *arg)
	104	{
d50084a2 JD	105	struct aio_thread_reply reply;
	106	struct io_event event;
	107	int err, n, reply_fd;
	108
	109	signal(SIGWINCH, SIG_IGN);
	110
ba180fd4	111	while (1) {
d50084a2	112	n = io_getevents(ctx, 1, 1, &event, NULL);
ba180fd4 JD	113	if (n < 0) {
ba180fd4 JD	114	if (errno == EINTR)
d50084a2	115	continue;
ba180fd4	116	printk(UM_KERN_ERR "aio_thread - io_getevents failed, "
d50084a2 JD	117	"errno = %d\n", errno);
	118	}
	119	else {
	120	reply = ((struct aio_thread_reply)
	121	{ .data = (void *) (long) event.data,
	122	.err = event.res });
91acb21f	123	reply_fd = ((struct aio_context *) reply.data)->reply_fd;
a61f334f	124	err = write(reply_fd, &reply, sizeof(reply));
ba180fd4 JD	125	if (err != sizeof(reply))
	126	printk(UM_KERN_ERR "aio_thread - write failed, "
	127	"fd = %d, err = %d\n", reply_fd, errno);
d50084a2 JD	128	}
	129	}
	130	return 0;
75e5584c JD	131	}
	132
	133	#endif
	134
91acb21f	135	static int do_not_aio(struct aio_thread_req *req)
75e5584c	136	{
d50084a2	137	char c;
ef0470c0	138	unsigned long long actual;
a61f334f	139	int n;
d50084a2	140
ef0470c0	141	actual = lseek64(req->io_fd, req->offset, SEEK_SET);
ba180fd4	142	if (actual != req->offset)
ef0470c0 JD	143	return -errno;
ef0470c0 JD	144
5134d8fe	145	switch (req->type) {
d50084a2	146	case AIO_READ:
a61f334f	147	n = read(req->io_fd, req->buf, req->len);
d50084a2 JD	148	break;
d50084a2 JD	149	case AIO_WRITE:
a61f334f	150	n = write(req->io_fd, req->buf, req->len);
d50084a2 JD	151	break;
d50084a2 JD	152	case AIO_MMAP:
a61f334f	153	n = read(req->io_fd, &c, sizeof(c));
d50084a2 JD	154	break;
d50084a2 JD	155	default:
ba180fd4 JD	156	printk(UM_KERN_ERR "do_not_aio - bad request type : %d\n",
ba180fd4 JD	157	req->type);
a61f334f	158	return -EINVAL;
d50084a2 JD	159	}
d50084a2 JD	160
ba180fd4	161	if (n < 0)
a61f334f JD	162	return -errno;
a61f334f JD	163	return 0;
75e5584c JD	164	}
75e5584c JD	165
9683da91 JD	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;
c4399016	170	static unsigned long aio_stack;
9683da91	171
75e5584c JD	172	static int not_aio_thread(void *arg)
75e5584c JD	173	{
d50084a2 JD	174	struct aio_thread_req req;
	175	struct aio_thread_reply reply;
	176	int err;
	177
	178	signal(SIGWINCH, SIG_IGN);
ba180fd4	179	while (1) {
a61f334f	180	err = read(aio_req_fd_r, &req, sizeof(req));
ba180fd4 JD	181	if (err != sizeof(req)) {
	182	if (err < 0)
	183	printk(UM_KERN_ERR "not_aio_thread - "
	184	"read failed, fd = %d, err = %d\n",
	185	aio_req_fd_r,
a61f334f	186	errno);
d50084a2	187	else {
ba180fd4 JD	188	printk(UM_KERN_ERR "not_aio_thread - short "
	189	"read, fd = %d, length = %d\n",
	190	aio_req_fd_r, err);
d50084a2 JD	191	}
	192	continue;
	193	}
	194	err = do_not_aio(&req);
	195	reply = ((struct aio_thread_reply) { .data = req.aio,
ef0470c0	196	.err = err });
a61f334f	197	err = write(req.aio->reply_fd, &reply, sizeof(reply));
ba180fd4 JD	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);
d50084a2	201	}
1b57e9c2 JD	202
1b57e9c2 JD	203	return 0;
75e5584c JD	204	}
75e5584c JD	205
75e5584c JD	206	static int init_aio_24(void)
75e5584c JD	207	{
d50084a2 JD	208	int fds[2], err;
	209
	210	err = os_pipe(fds, 1, 1);
ba180fd4	211	if (err)
d50084a2 JD	212	goto out;
	213
	214	aio_req_fd_w = fds[0];
	215	aio_req_fd_r = fds[1];
8603ec81 JD	216
8603ec81 JD	217	err = os_set_fd_block(aio_req_fd_w, 0);
ba180fd4	218	if (err)
8603ec81 JD	219	goto out_close_pipe;
8603ec81 JD	220
d50084a2	221	err = run_helper_thread(not_aio_thread, NULL,
4dbed85a	222	CLONE_FILES \| CLONE_VM, &aio_stack);
ba180fd4	223	if (err < 0)
d50084a2 JD	224	goto out_close_pipe;
	225
	226	aio_pid = err;
	227	goto out;
	228
	229	out_close_pipe:
512b6fb1 JD	230	close(fds[0]);
512b6fb1 JD	231	close(fds[1]);
d50084a2 JD	232	aio_req_fd_w = -1;
	233	aio_req_fd_r = -1;
	234	out:
75e5584c	235	#ifndef HAVE_AIO_ABI
ba180fd4 JD	236	printk(UM_KERN_INFO "/usr/include/linux/aio_abi.h not present during "
ba180fd4 JD	237	"build\n");
75e5584c	238	#endif
ba180fd4 JD	239	printk(UM_KERN_INFO "2.6 host AIO support not used - falling back to "
ba180fd4 JD	240	"I/O thread\n");
d50084a2	241	return 0;
75e5584c JD	242	}
	243
	244	#ifdef HAVE_AIO_ABI
	245	#define DEFAULT_24_AIO 0
	246	static int init_aio_26(void)
	247	{
d50084a2	248	int err;
75e5584c	249
ba180fd4	250	if (io_setup(256, &ctx)) {
b4fd310e	251	err = -errno;
ba180fd4 JD	252	printk(UM_KERN_ERR "aio_thread failed to initialize context, "
ba180fd4 JD	253	"err = %d\n", errno);
d50084a2 JD	254	return err;
d50084a2 JD	255	}
75e5584c	256
d50084a2	257	err = run_helper_thread(aio_thread, NULL,
4dbed85a	258	CLONE_FILES \| CLONE_VM, &aio_stack);
ba180fd4	259	if (err < 0)
d50084a2	260	return err;
75e5584c	261
d50084a2	262	aio_pid = err;
75e5584c	263
ba180fd4	264	printk(UM_KERN_INFO "Using 2.6 host AIO\n");
d50084a2	265	return 0;
91acb21f JD	266	}
	267
	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)
	270	{
d50084a2 JD	271	struct aio_thread_reply reply;
	272	int err;
	273
	274	err = do_aio(ctx, type, io_fd, buf, len, offset, aio);
ba180fd4	275	if (err) {
d50084a2 JD	276	reply = ((struct aio_thread_reply) { .data = aio,
d50084a2 JD	277	.err = err });
a61f334f	278	err = write(aio->reply_fd, &reply, sizeof(reply));
ba180fd4	279	if (err != sizeof(reply)) {
a61f334f	280	err = -errno;
ba180fd4	281	printk(UM_KERN_ERR "submit_aio_26 - write failed, "
d50084a2	282	"fd = %d, err = %d\n", aio->reply_fd, -err);
a61f334f	283	}
d50084a2 JD	284	else err = 0;
	285	}
	286
	287	return err;
75e5584c JD	288	}
	289
	290	#else
	291	#define DEFAULT_24_AIO 1
91acb21f	292	static int init_aio_26(void)
75e5584c	293	{
d50084a2	294	return -ENOSYS;
75e5584c JD	295	}
75e5584c JD	296
91acb21f JD	297	static int submit_aio_26(enum aio_type type, int io_fd, char *buf, int len,
91acb21f JD	298	unsigned long long offset, struct aio_context *aio)
75e5584c	299	{
d50084a2	300	return -ENOSYS;
75e5584c JD	301	}
	302	#endif
	303
9683da91	304	/* Initialized in an initcall and unchanged thereafter */
75e5584c JD	305	static int aio_24 = DEFAULT_24_AIO;
	306
	307	static int __init set_aio_24(char name, int add)
	308	{
d50084a2 JD	309	aio_24 = 1;
d50084a2 JD	310	return 0;
75e5584c JD	311	}
	312
	313	__uml_setup("aio=2.4", set_aio_24,
	314	"aio=2.4\n"
	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"
	323	);
	324
	325	static int init_aio(void)
	326	{
d50084a2 JD	327	int err;
d50084a2 JD	328
ba180fd4	329	if (!aio_24) {
d50084a2	330	err = init_aio_26();
ba180fd4 JD	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");
d50084a2 JD	334	aio_24 = 1;
	335	}
	336	else return err;
	337	}
	338
ba180fd4	339	if (aio_24)
d50084a2 JD	340	return init_aio_24();
	341
	342	return 0;
75e5584c JD	343	}
75e5584c JD	344
ba180fd4 JD	345	/*
ba180fd4 JD	346	* The reason for the __initcall/__uml_exitcall asymmetry is that init_aio
75e5584c JD	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.
	351	*/
	352	__initcall(init_aio);
	353
	354	static void exit_aio(void)
	355	{
c4399016	356	if (aio_pid != -1) {
d50084a2	357	os_kill_process(aio_pid, 1);
c4399016 JD	358	free_stack(aio_stack, 0);
c4399016 JD	359	}
75e5584c JD	360	}
	361
	362	__uml_exitcall(exit_aio);
	363
91acb21f JD	364	static int submit_aio_24(enum aio_type type, int io_fd, char *buf, int len,
91acb21f JD	365	unsigned long long offset, struct aio_context *aio)
75e5584c	366	{
d50084a2 JD	367	struct aio_thread_req req = { .type = type,
	368	.io_fd = io_fd,
	369	.offset = offset,
	370	.buf = buf,
	371	.len = len,
	372	.aio = aio,
	373	};
	374	int err;
	375
a61f334f	376	err = write(aio_req_fd_w, &req, sizeof(req));
ba180fd4	377	if (err == sizeof(req))
d50084a2	378	err = 0;
a61f334f	379	else err = -errno;
d50084a2 JD	380
d50084a2 JD	381	return err;
91acb21f JD	382	}
	383
	384	int submit_aio(enum aio_type type, int io_fd, char *buf, int len,
d50084a2 JD	385	unsigned long long offset, int reply_fd,
d50084a2 JD	386	struct aio_context *aio)
91acb21f	387	{
d50084a2	388	aio->reply_fd = reply_fd;
ba180fd4	389	if (aio_24)
d50084a2	390	return submit_aio_24(type, io_fd, buf, len, offset, aio);
ba180fd4	391	else
d50084a2	392	return submit_aio_26(type, io_fd, buf, len, offset, aio);
75e5584c	393	}