[mirror_ubuntu-eoan-kernel.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 "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 | SIGCHLD, &aio_stack);
	if (err < 0)
		goto out_close_pipe;

	aio_pid = err;
	goto out;

out_close_pipe:
	os_close_file(fds[0]);
	os_close_file(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 | SIGCHLD, &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"
ba180fd4 JD	15	#include "os.h"
ba180fd4 JD	16	#include "user.h"
75e5584c	17
91acb21f	18	struct aio_thread_req {
d50084a2 JD	19	enum aio_type type;
	20	int io_fd;
	21	unsigned long long offset;
	22	char *buf;
	23	int len;
	24	struct aio_context *aio;
91acb21f JD	25	};
91acb21f JD	26
75e5584c JD	27	#if defined(HAVE_AIO_ABI)
	28	#include <linux/aio_abi.h>
	29
ba180fd4 JD	30	/*
ba180fd4 JD	31	* If we have the headers, we are going to build with AIO enabled.
75e5584c JD	32	* If we don't have aio in libc, we define the necessary stubs here.
	33	*/
	34
	35	#if !defined(HAVE_AIO_LIBC)
	36
	37	static long io_setup(int n, aio_context_t *ctxp)
	38	{
d50084a2	39	return syscall(__NR_io_setup, n, ctxp);
75e5584c JD	40	}
	41
	42	static long io_submit(aio_context_t ctx, long nr, struct iocb **iocbpp)
	43	{
d50084a2	44	return syscall(__NR_io_submit, ctx, nr, iocbpp);
75e5584c JD	45	}
	46
	47	static long io_getevents(aio_context_t ctx_id, long min_nr, long nr,
d50084a2	48	struct io_event events, struct timespec timeout)
75e5584c	49	{
d50084a2	50	return syscall(__NR_io_getevents, ctx_id, min_nr, nr, events, timeout);
75e5584c JD	51	}
	52
	53	#endif
	54
ba180fd4 JD	55	/*
ba180fd4 JD	56	* The AIO_MMAP cases force the mmapped page into memory here
75e5584c JD	57	* rather than in whatever place first touches the data. I used
	58	* to do this by touching the page, but that's delicate because
	59	* gcc is prone to optimizing that away. So, what's done here
	60	* is we read from the descriptor from which the page was
	61	* mapped. The caller is required to pass an offset which is
	62	* inside the page that was mapped. Thus, when the read
	63	* returns, we know that the page is in the page cache, and
	64	* that it now backs the mmapped area.
	65	*/
	66
91acb21f	67	static int do_aio(aio_context_t ctx, enum aio_type type, int fd, char *buf,
d50084a2	68	int len, unsigned long long offset, struct aio_context *aio)
75e5584c	69	{
da3e30e7 JD	70	struct iocb *iocbp = & ((struct iocb) {
	71	.aio_data = (unsigned long) aio,
	72	.aio_fildes = fd,
	73	.aio_buf = (unsigned long) buf,
	74	.aio_nbytes = len,
	75	.aio_offset = offset
	76	});
d50084a2	77	char c;
da3e30e7 JD	78
da3e30e7 JD	79	switch (type) {
d50084a2	80	case AIO_READ:
da3e30e7	81	iocbp->aio_lio_opcode = IOCB_CMD_PREAD;
d50084a2 JD	82	break;
d50084a2 JD	83	case AIO_WRITE:
da3e30e7	84	iocbp->aio_lio_opcode = IOCB_CMD_PWRITE;
d50084a2 JD	85	break;
d50084a2 JD	86	case AIO_MMAP:
da3e30e7 JD	87	iocbp->aio_lio_opcode = IOCB_CMD_PREAD;
	88	iocbp->aio_buf = (unsigned long) &c;
	89	iocbp->aio_nbytes = sizeof(c);
d50084a2 JD	90	break;
d50084a2 JD	91	default:
da3e30e7 JD	92	printk(UM_KERN_ERR "Bogus op in do_aio - %d\n", type);
da3e30e7 JD	93	return -EINVAL;
d50084a2 JD	94	}
d50084a2 JD	95
da3e30e7	96	return (io_submit(ctx, 1, &iocbp) > 0) ? 0 : -errno;
75e5584c JD	97	}
75e5584c JD	98
9683da91	99	/* Initialized in an initcall and unchanged thereafter */
75e5584c JD	100	static aio_context_t ctx = 0;
	101
	102	static int aio_thread(void *arg)
	103	{
d50084a2 JD	104	struct aio_thread_reply reply;
	105	struct io_event event;
	106	int err, n, reply_fd;
	107
	108	signal(SIGWINCH, SIG_IGN);
	109
ba180fd4	110	while (1) {
d50084a2	111	n = io_getevents(ctx, 1, 1, &event, NULL);
ba180fd4 JD	112	if (n < 0) {
ba180fd4 JD	113	if (errno == EINTR)
d50084a2	114	continue;
ba180fd4	115	printk(UM_KERN_ERR "aio_thread - io_getevents failed, "
d50084a2 JD	116	"errno = %d\n", errno);
	117	}
	118	else {
	119	reply = ((struct aio_thread_reply)
	120	{ .data = (void *) (long) event.data,
	121	.err = event.res });
91acb21f	122	reply_fd = ((struct aio_context *) reply.data)->reply_fd;
a61f334f	123	err = write(reply_fd, &reply, sizeof(reply));
ba180fd4 JD	124	if (err != sizeof(reply))
	125	printk(UM_KERN_ERR "aio_thread - write failed, "
	126	"fd = %d, err = %d\n", reply_fd, errno);
d50084a2 JD	127	}
	128	}
	129	return 0;
75e5584c JD	130	}
	131
	132	#endif
	133
91acb21f	134	static int do_not_aio(struct aio_thread_req *req)
75e5584c	135	{
d50084a2	136	char c;
ef0470c0	137	unsigned long long actual;
a61f334f	138	int n;
d50084a2	139
ef0470c0	140	actual = lseek64(req->io_fd, req->offset, SEEK_SET);
ba180fd4	141	if (actual != req->offset)
ef0470c0 JD	142	return -errno;
ef0470c0 JD	143
ba180fd4	144	switch(req->type) {
d50084a2	145	case AIO_READ:
a61f334f	146	n = read(req->io_fd, req->buf, req->len);
d50084a2 JD	147	break;
d50084a2 JD	148	case AIO_WRITE:
a61f334f	149	n = write(req->io_fd, req->buf, req->len);
d50084a2 JD	150	break;
d50084a2 JD	151	case AIO_MMAP:
a61f334f	152	n = read(req->io_fd, &c, sizeof(c));
d50084a2 JD	153	break;
d50084a2 JD	154	default:
ba180fd4 JD	155	printk(UM_KERN_ERR "do_not_aio - bad request type : %d\n",
ba180fd4 JD	156	req->type);
a61f334f	157	return -EINVAL;
d50084a2 JD	158	}
d50084a2 JD	159
ba180fd4	160	if (n < 0)
a61f334f JD	161	return -errno;
a61f334f JD	162	return 0;
75e5584c JD	163	}
75e5584c JD	164
9683da91 JD	165	/* These are initialized in initcalls and not changed */
	166	static int aio_req_fd_r = -1;
	167	static int aio_req_fd_w = -1;
	168	static int aio_pid = -1;
c4399016	169	static unsigned long aio_stack;
9683da91	170
75e5584c JD	171	static int not_aio_thread(void *arg)
75e5584c JD	172	{
d50084a2 JD	173	struct aio_thread_req req;
	174	struct aio_thread_reply reply;
	175	int err;
	176
	177	signal(SIGWINCH, SIG_IGN);
ba180fd4	178	while (1) {
a61f334f	179	err = read(aio_req_fd_r, &req, sizeof(req));
ba180fd4 JD	180	if (err != sizeof(req)) {
	181	if (err < 0)
	182	printk(UM_KERN_ERR "not_aio_thread - "
	183	"read failed, fd = %d, err = %d\n",
	184	aio_req_fd_r,
a61f334f	185	errno);
d50084a2	186	else {
ba180fd4 JD	187	printk(UM_KERN_ERR "not_aio_thread - short "
	188	"read, fd = %d, length = %d\n",
	189	aio_req_fd_r, err);
d50084a2 JD	190	}
	191	continue;
	192	}
	193	err = do_not_aio(&req);
	194	reply = ((struct aio_thread_reply) { .data = req.aio,
ef0470c0	195	.err = err });
a61f334f	196	err = write(req.aio->reply_fd, &reply, sizeof(reply));
ba180fd4 JD	197	if (err != sizeof(reply))
	198	printk(UM_KERN_ERR "not_aio_thread - write failed, "
	199	"fd = %d, err = %d\n", req.aio->reply_fd, errno);
d50084a2	200	}
1b57e9c2 JD	201
1b57e9c2 JD	202	return 0;
75e5584c JD	203	}
75e5584c JD	204
75e5584c JD	205	static int init_aio_24(void)
75e5584c JD	206	{
d50084a2 JD	207	int fds[2], err;
	208
	209	err = os_pipe(fds, 1, 1);
ba180fd4	210	if (err)
d50084a2 JD	211	goto out;
	212
	213	aio_req_fd_w = fds[0];
	214	aio_req_fd_r = fds[1];
8603ec81 JD	215
8603ec81 JD	216	err = os_set_fd_block(aio_req_fd_w, 0);
ba180fd4	217	if (err)
8603ec81 JD	218	goto out_close_pipe;
8603ec81 JD	219
d50084a2	220	err = run_helper_thread(not_aio_thread, NULL,
c4399016	221	CLONE_FILES \| CLONE_VM \| SIGCHLD, &aio_stack);
ba180fd4	222	if (err < 0)
d50084a2 JD	223	goto out_close_pipe;
	224
	225	aio_pid = err;
	226	goto out;
	227
	228	out_close_pipe:
	229	os_close_file(fds[0]);
	230	os_close_file(fds[1]);
	231	aio_req_fd_w = -1;
	232	aio_req_fd_r = -1;
	233	out:
75e5584c	234	#ifndef HAVE_AIO_ABI
ba180fd4 JD	235	printk(UM_KERN_INFO "/usr/include/linux/aio_abi.h not present during "
ba180fd4 JD	236	"build\n");
75e5584c	237	#endif
ba180fd4 JD	238	printk(UM_KERN_INFO "2.6 host AIO support not used - falling back to "
ba180fd4 JD	239	"I/O thread\n");
d50084a2	240	return 0;
75e5584c JD	241	}
	242
	243	#ifdef HAVE_AIO_ABI
	244	#define DEFAULT_24_AIO 0
	245	static int init_aio_26(void)
	246	{
d50084a2	247	int err;
75e5584c	248
ba180fd4	249	if (io_setup(256, &ctx)) {
b4fd310e	250	err = -errno;
ba180fd4 JD	251	printk(UM_KERN_ERR "aio_thread failed to initialize context, "
ba180fd4 JD	252	"err = %d\n", errno);
d50084a2 JD	253	return err;
d50084a2 JD	254	}
75e5584c	255
d50084a2	256	err = run_helper_thread(aio_thread, NULL,
c4399016	257	CLONE_FILES \| CLONE_VM \| SIGCHLD, &aio_stack);
ba180fd4	258	if (err < 0)
d50084a2	259	return err;
75e5584c	260
d50084a2	261	aio_pid = err;
75e5584c	262
ba180fd4	263	printk(UM_KERN_INFO "Using 2.6 host AIO\n");
d50084a2	264	return 0;
91acb21f JD	265	}
	266
	267	static int submit_aio_26(enum aio_type type, int io_fd, char *buf, int len,
	268	unsigned long long offset, struct aio_context *aio)
	269	{
d50084a2 JD	270	struct aio_thread_reply reply;
	271	int err;
	272
	273	err = do_aio(ctx, type, io_fd, buf, len, offset, aio);
ba180fd4	274	if (err) {
d50084a2 JD	275	reply = ((struct aio_thread_reply) { .data = aio,
d50084a2 JD	276	.err = err });
a61f334f	277	err = write(aio->reply_fd, &reply, sizeof(reply));
ba180fd4	278	if (err != sizeof(reply)) {
a61f334f	279	err = -errno;
ba180fd4	280	printk(UM_KERN_ERR "submit_aio_26 - write failed, "
d50084a2	281	"fd = %d, err = %d\n", aio->reply_fd, -err);
a61f334f	282	}
d50084a2 JD	283	else err = 0;
	284	}
	285
	286	return err;
75e5584c JD	287	}
	288
	289	#else
	290	#define DEFAULT_24_AIO 1
91acb21f	291	static int init_aio_26(void)
75e5584c	292	{
d50084a2	293	return -ENOSYS;
75e5584c JD	294	}
75e5584c JD	295
91acb21f JD	296	static int submit_aio_26(enum aio_type type, int io_fd, char *buf, int len,
91acb21f JD	297	unsigned long long offset, struct aio_context *aio)
75e5584c	298	{
d50084a2	299	return -ENOSYS;
75e5584c JD	300	}
	301	#endif
	302
9683da91	303	/* Initialized in an initcall and unchanged thereafter */
75e5584c JD	304	static int aio_24 = DEFAULT_24_AIO;
	305
	306	static int __init set_aio_24(char name, int add)
	307	{
d50084a2 JD	308	aio_24 = 1;
d50084a2 JD	309	return 0;
75e5584c JD	310	}
	311
	312	__uml_setup("aio=2.4", set_aio_24,
	313	"aio=2.4\n"
	314	" This is used to force UML to use 2.4-style AIO even when 2.6 AIO is\n"
	315	" available. 2.4 AIO is a single thread that handles one request at a\n"
	316	" time, synchronously. 2.6 AIO is a thread which uses the 2.6 AIO \n"
	317	" interface to handle an arbitrary number of pending requests. 2.6 AIO \n"
	318	" is not available in tt mode, on 2.4 hosts, or when UML is built with\n"
	319	" /usr/include/linux/aio_abi.h not available. Many distributions don't\n"
	320	" include aio_abi.h, so you will need to copy it from a kernel tree to\n"
	321	" your /usr/include/linux in order to build an AIO-capable UML\n\n"
	322	);
	323
	324	static int init_aio(void)
	325	{
d50084a2 JD	326	int err;
d50084a2 JD	327
ba180fd4	328	if (!aio_24) {
d50084a2	329	err = init_aio_26();
ba180fd4 JD	330	if (err && (errno == ENOSYS)) {
	331	printk(UM_KERN_INFO "2.6 AIO not supported on the "
	332	"host - reverting to 2.4 AIO\n");
d50084a2 JD	333	aio_24 = 1;
	334	}
	335	else return err;
	336	}
	337
ba180fd4	338	if (aio_24)
d50084a2 JD	339	return init_aio_24();
	340
	341	return 0;
75e5584c JD	342	}
75e5584c JD	343
ba180fd4 JD	344	/*
ba180fd4 JD	345	* The reason for the __initcall/__uml_exitcall asymmetry is that init_aio
75e5584c JD	346	* needs to be called when the kernel is running because it calls run_helper,
	347	* which needs get_free_page. exit_aio is a __uml_exitcall because the generic
	348	* kernel does not run __exitcalls on shutdown, and can't because many of them
	349	* break when called outside of module unloading.
	350	*/
	351	__initcall(init_aio);
	352
	353	static void exit_aio(void)
	354	{
c4399016	355	if (aio_pid != -1) {
d50084a2	356	os_kill_process(aio_pid, 1);
c4399016 JD	357	free_stack(aio_stack, 0);
c4399016 JD	358	}
75e5584c JD	359	}
	360
	361	__uml_exitcall(exit_aio);
	362
91acb21f JD	363	static int submit_aio_24(enum aio_type type, int io_fd, char *buf, int len,
91acb21f JD	364	unsigned long long offset, struct aio_context *aio)
75e5584c	365	{
d50084a2 JD	366	struct aio_thread_req req = { .type = type,
	367	.io_fd = io_fd,
	368	.offset = offset,
	369	.buf = buf,
	370	.len = len,
	371	.aio = aio,
	372	};
	373	int err;
	374
a61f334f	375	err = write(aio_req_fd_w, &req, sizeof(req));
ba180fd4	376	if (err == sizeof(req))
d50084a2	377	err = 0;
a61f334f	378	else err = -errno;
d50084a2 JD	379
d50084a2 JD	380	return err;
91acb21f JD	381	}
	382
	383	int submit_aio(enum aio_type type, int io_fd, char *buf, int len,
d50084a2 JD	384	unsigned long long offset, int reply_fd,
d50084a2 JD	385	struct aio_context *aio)
91acb21f	386	{
d50084a2	387	aio->reply_fd = reply_fd;
ba180fd4	388	if (aio_24)
d50084a2	389	return submit_aio_24(type, io_fd, buf, len, offset, aio);
ba180fd4	390	else
d50084a2	391	return submit_aio_26(type, io_fd, buf, len, offset, aio);
75e5584c	392	}