[PATCH -mm 1/5][AIO] - Rework compat_sys_io_submit

[PATCH -mm 1/5][AIO] - Rework compat_sys_io_submit

[Sébastien Dugué]

                 compat_sys_io_submit() cleanup


  Cleanup compat_sys_io_submit by duplicating some of the native syscall
logic in the compat layer and directly calling io_submit_one() instead
of fooling the syscall into thinking it is called from a native 64-bit
caller.

  This eliminates:

  - the overhead of copying the nr iocb pointers on the userspace stack

  - the PAGE_SIZE/(sizeof(void *)) limit on the number of iocbs that
    can be submitted.

  This is also needed for the completion notification patch to avoid having
to rewrite each iocb on the caller stack for io_submit_one() to find the
sigevents.


 compat.c |   61 ++++++++++++++++++++++++++++++++++---------------------------
 1 file changed, 34 insertions(+), 27 deletions(-)

Signed-off-by: Sébastien Dugué <sebastien.dugue@bull.net>

Index: linux-2.6.20-rc4-mm1/fs/compat.c
===================================================================
--- linux-2.6.20-rc4-mm1.orig/fs/compat.c	2007-01-12 11:40:28.000000000 +0100
+++ linux-2.6.20-rc4-mm1/fs/compat.c	2007-01-12 12:14:51.000000000 +0100
@@ -644,40 +644,47 @@ out:
 	return ret;
 }
 
-static inline long
-copy_iocb(long nr, u32 __user *ptr32, struct iocb __user * __user *ptr64)
-{
-	compat_uptr_t uptr;
-	int i;
-
-	for (i = 0; i < nr; ++i) {
-		if (get_user(uptr, ptr32 + i))
-			return -EFAULT;
-		if (put_user(compat_ptr(uptr), ptr64 + i))
-			return -EFAULT;
-	}
-	return 0;
-}
-
-#define MAX_AIO_SUBMITS 	(PAGE_SIZE/sizeof(struct iocb *))
-
 asmlinkage long
 compat_sys_io_submit(aio_context_t ctx_id, int nr, u32 __user *iocb)
 {
-	struct iocb __user * __user *iocb64; 
-	long ret;
+	struct kioctx *ctx;
+	long ret = 0;
+	int i;
 
 	if (unlikely(nr < 0))
 		return -EINVAL;
 
-	if (nr > MAX_AIO_SUBMITS)
-		nr = MAX_AIO_SUBMITS;
-	
-	iocb64 = compat_alloc_user_space(nr * sizeof(*iocb64));
-	ret = copy_iocb(nr, iocb, iocb64);
-	if (!ret)
-		ret = sys_io_submit(ctx_id, nr, iocb64);
-	return ret;
+	if (unlikely(!access_ok(VERIFY_READ, iocb, (nr * sizeof(u32)))))
+		return -EFAULT;
+
+	ctx = lookup_ioctx(ctx_id);
+	if (unlikely(!ctx))
+		return -EINVAL;
+
+	for (i=0; i<nr; i++) {
+		compat_uptr_t uptr;
+		struct iocb __user *user_iocb;
+		struct iocb tmp;
+
+		if (unlikely(get_user(uptr, iocb + i))) {
+			ret = -EFAULT;
+			break;
+		}
+
+		user_iocb = compat_ptr(uptr);
+
+		if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
+			ret = -EFAULT;
+			break;
+		}
+
+		ret = io_submit_one(ctx, user_iocb, &tmp);
+		if (ret)
+			break;
+	}
+
+	put_ioctx(ctx);
+	return i ? i: ret;
 }
 
 struct compat_ncp_mount_data {

[PATCH -mm 2/5][AIO] - fix aio.h includes

[Sébastien Dugué]

  Fix the double inclusion of linux/uio.h in linux/aio.h



 aio.h |    1 -
 1 file changed, 1 deletion(-)

Signed-off-by: Sébastien Dugué <sebastien.dugue@bull.net>

Index: linux-2.6.20-rc4-mm1/include/linux/aio.h
===================================================================
--- linux-2.6.20-rc4-mm1.orig/include/linux/aio.h	2007-01-12 11:40:39.000000000 +0100
+++ linux-2.6.20-rc4-mm1/include/linux/aio.h	2007-01-12 12:15:01.000000000 +0100
@@ -7,7 +7,6 @@
 #include <linux/uio.h>
 
 #include <asm/atomic.h>
-#include <linux/uio.h>
 
 #define AIO_MAXSEGS		4
 #define AIO_KIOGRP_NR_ATOMIC	8

[PATCH -mm 3/5][AIO] - Make good_sigevent non-static

[Sébastien Dugué]

                      Make good_sigevent() non-static


  Move good_sigevent() from posix-timers.c to signal.c where it belongs,
and make it non-static so that it can be used by other subsystems.


 include/linux/signal.h |    1 +
 kernel/posix-timers.c  |   17 -----------------
 kernel/signal.c        |   24 ++++++++++++++++++++++++
 3 files changed, 25 insertions(+), 17 deletions(-)

Signed-off-by: Sébastien Dugué <sebastien.dugue@bull.net>


Index: linux-2.6.20-rc4-mm1/include/linux/signal.h
===================================================================
--- linux-2.6.20-rc4-mm1.orig/include/linux/signal.h	2007-01-12 11:40:30.000000000 +0100
+++ linux-2.6.20-rc4-mm1/include/linux/signal.h	2007-01-12 12:15:06.000000000 +0100
@@ -240,6 +240,7 @@ extern int sigprocmask(int, sigset_t *, 
 
 struct pt_regs;
 extern int get_signal_to_deliver(siginfo_t *info, struct k_sigaction *return_ka, struct pt_regs *regs, void *cookie);
+extern struct task_struct * good_sigevent(sigevent_t *);
 
 extern struct kmem_cache *sighand_cachep;
 
Index: linux-2.6.20-rc4-mm1/kernel/posix-timers.c
===================================================================
--- linux-2.6.20-rc4-mm1.orig/kernel/posix-timers.c	2007-01-12 11:40:39.000000000 +0100
+++ linux-2.6.20-rc4-mm1/kernel/posix-timers.c	2007-01-12 12:15:06.000000000 +0100
@@ -367,23 +367,6 @@ static enum hrtimer_restart posix_timer_
 	return ret;
 }
 
-static struct task_struct * good_sigevent(sigevent_t * event)
-{
-	struct task_struct *rtn = current->group_leader;
-
-	if ((event->sigev_notify & SIGEV_THREAD_ID ) &&
-		(!(rtn = find_task_by_pid(event->sigev_notify_thread_id)) ||
-		 rtn->tgid != current->tgid ||
-		 (event->sigev_notify & ~SIGEV_THREAD_ID) != SIGEV_SIGNAL))
-		return NULL;
-
-	if (((event->sigev_notify & ~SIGEV_THREAD_ID) != SIGEV_NONE) &&
-	    ((event->sigev_signo <= 0) || (event->sigev_signo > SIGRTMAX)))
-		return NULL;
-
-	return rtn;
-}
-
 void register_posix_clock(const clockid_t clock_id, struct k_clock *new_clock)
 {
 	if ((unsigned) clock_id >= MAX_CLOCKS) {
Index: linux-2.6.20-rc4-mm1/kernel/signal.c
===================================================================
--- linux-2.6.20-rc4-mm1.orig/kernel/signal.c	2007-01-12 11:40:39.000000000 +0100
+++ linux-2.6.20-rc4-mm1/kernel/signal.c	2007-01-12 12:15:06.000000000 +0100
@@ -1213,6 +1213,30 @@ int group_send_sig_info(int sig, struct 
 	return ret;
 }
 
+/***
+ * good_sigevent - check and get target task from a sigevent.
+ * @event: the sigevent to be checked
+ *
+ * This function must be called with the tasklist_lock held for reading.
+ */
+struct task_struct * good_sigevent(sigevent_t * event)
+{
+	struct task_struct *task = current->group_leader;
+
+	if ((event->sigev_notify & SIGEV_THREAD_ID ) == SIGEV_THREAD_ID) {
+		task = find_task_by_pid(event->sigev_notify_thread_id);
+
+		if (!task || task->tgid != current->tgid)
+			return NULL;
+	} else if (event->sigev_notify == SIGEV_SIGNAL) {
+		if ((event->sigev_signo <= 0) || (event->sigev_signo > SIGRTMAX))
+			return NULL;
+	} else
+		return NULL;
+
+	return task;
+}
+
 /*
  * kill_pgrp_info() sends a signal to a process group: this is what the tty
  * control characters do (^C, ^Z etc)

[PATCH -mm 4/5][AIO] - AIO completion signal notification

[Sébastien Dugué]

                      AIO completion signal notification

  The current 2.6 kernel does not support notification of user space via
an RT signal upon an asynchronous IO completion. The POSIX specification
states that when an AIO request completes, a signal can be delivered to
the application as notification.

  This patch adds a struct sigevent *aio_sigeventp to the iocb.
The relevant fields (pid, signal number and value) are stored in the kiocb
for use when the request completes.

  That sigevent structure is filled by the application as part of the AIO
request preparation. Upon request completion, the kernel notifies the
application using those sigevent parameters. If SIGEV_NONE has been specified,
then the old behaviour is retained and the application must rely on polling
the completion queue using io_getevents().


  A struct sigevent *aio_sigeventp field is added to struct iocb in
include/linux/aio_abi.h

  A struct aio_notify containing the sigevent parameters is defined in aio.h:

  struct aio_notify {
	struct task_struct	*target;
	__u16			signo;
	__u16			notify;
	sigval_t		value;
  };

  A struct aio_notify ki_notify is added to struct kiocb in include/linux/aio.h

  In io_submit_one(), if the application provided a sigevent then
setup_sigevent() is called which does the following:

	- check access to the user sigevent and make a local copy

	- if the requested notification is SIGEV_NONE, then nothing to do

	- fill in the kiocb->ki_notify fields (notify, signo, value)

	- check sigevent consistency, get the signal target task and
	  save it in kiocb->ki_notify.target

	- preallocate a sigqueue for this event using sigqueue_alloc()

  Upon request completion, in aio_complete(), if notification is needed for
this request (iocb->ki_notify.notify != SIGEV_NONE), then aio_send_signal()
is called to signal the target task as follows:

	- fill in the siginfo struct to be sent to the task

	- if notify is SIGEV_THREAD_ID then send signal to specific task
	  using send_sigqueue()

	- else send signal to task group using send_5group_sigqueue()



Notes concerning sigqueue preallocation:

 To ensure reliable delivery of completion notification, the sigqueue is
preallocated in the submission path so that there is no chance it can fail
in the completion path.

 Unlike the posix-timers case (currently the single other user of sigqueue
preallocation), where the sigqueue is allocated for the lifetime of the timer
and freed at timer destruction time, the aio case is a bit more tricky due to
the async nature of the whole thing.

  In the aio case, the sigqueue exists for the lifetime of the request, therefore
it must be freed only once the signal for the request completion has been
delivered. This involves changing __sigqueue_free() to free the sigqueue when the
signal is collected if si_code is SI_ASYNCIO even if it was preallocated as well
as explicitly calling sigqueue_free() in submission and completion error paths.


 fs/aio.c                |  115 ++++++++++++++++++++++++++++++++++++++++++++++--
 fs/compat.c             |   18 +++++++
 include/linux/aio.h     |   12 +++++
 include/linux/aio_abi.h |    3 -
 kernel/signal.c         |    2
 5 files changed, 144 insertions(+), 6 deletions(-)

Signed-off-by: Sébastien Dugué <sebastien.dugue@bull.net>
Signed-off-by: Laurent Vivier <laurent.vivier@bull.net>

Index: linux-2.6.20-rc4-mm1/fs/aio.c
===================================================================
--- linux-2.6.20-rc4-mm1.orig/fs/aio.c	2007-01-12 11:40:38.000000000 +0100
+++ linux-2.6.20-rc4-mm1/fs/aio.c	2007-01-12 12:32:55.000000000 +0100
@@ -419,6 +419,7 @@ static struct kiocb fastcall *__aio_get_
 	req->ki_dtor = NULL;
 	req->private = NULL;
 	req->ki_iovec = NULL;
+	req->ki_notify.sigq = NULL;
 	INIT_LIST_HEAD(&req->ki_run_list);
 
 	/* Check if the completion queue has enough free space to
@@ -465,6 +466,12 @@ static inline void really_put_req(struct
 		req->ki_dtor(req);
 	if (req->ki_iovec != &req->ki_inline_vec)
 		kfree(req->ki_iovec);
+
+	/* Release task ref */
+	if (req->ki_notify.notify == SIGEV_THREAD_ID ||
+	    req->ki_notify.notify == SIGEV_SIGNAL)
+		put_task_struct(req->ki_notify.target);
+
 	kmem_cache_free(kiocb_cachep, req);
 	ctx->reqs_active--;
 
@@ -916,6 +923,79 @@ void fastcall kick_iocb(struct kiocb *io
 }
 EXPORT_SYMBOL(kick_iocb);
 
+static int aio_send_signal(struct aio_notify *notify)
+{
+	struct sigqueue *sigq = notify->sigq;
+	struct siginfo *info = &sigq->info;
+	int ret;
+
+	memset(info, 0, sizeof(struct siginfo));
+
+	info->si_signo = notify->signo;
+	info->si_errno = 0;
+	info->si_code = SI_ASYNCIO;
+	info->si_pid = 0;
+	info->si_uid = 0;
+	info->si_value = notify->value;
+
+	if (notify->notify & SIGEV_THREAD_ID)
+		ret = send_sigqueue(notify->signo, sigq, notify->target);
+	else
+		ret = send_group_sigqueue(notify->signo, sigq, notify->target);
+
+	return ret;
+}
+
+static long aio_setup_sigevent(struct aio_notify *notify,
+			       struct sigevent __user *user_event)
+{
+	sigevent_t event;
+	struct task_struct *target;
+
+	if (copy_from_user(&event, user_event, sizeof (event)))
+		return -EFAULT;
+
+	if (event.sigev_notify == SIGEV_NONE)
+		return 0;
+
+	notify->notify = event.sigev_notify;
+	notify->signo = event.sigev_signo;
+	notify->value = event.sigev_value;
+
+	read_lock(&tasklist_lock);
+	target = good_sigevent(&event);
+
+	if (unlikely(!target || (target->flags & PF_EXITING)))
+		goto out_unlock;
+
+	/*
+	 * At this point, we know that notify is either SIGEV_SIGNAL or
+	 * SIGEV_THREAD_ID and the target task is valid. So get a reference
+	 * on the task, it will be dropped in really_put_req() when
+	 * we're done with the request.
+	 */
+	get_task_struct(target);
+	notify->target = target;
+	read_unlock(&tasklist_lock);
+
+	/*
+	 * NOTE: we cannot free the sigqueue in the completion path as
+	 * the signal may not have been delivered to the target task.
+	 * Therefore it has to be freed in __sigqueue_free() when the
+	 * signal is collected if si_code is SI_ASYNCIO.
+	 */
+	notify->sigq = sigqueue_alloc();
+
+	if (unlikely(!notify->sigq))
+		return -EAGAIN;
+
+	return 0;
+
+out_unlock:
+	read_unlock(&tasklist_lock);
+	return -EINVAL;
+}
+
 /* aio_complete
  *	Called when the io request on the given iocb is complete.
  *	Returns true if this is the last user of the request.  The 
@@ -963,8 +1043,11 @@ int fastcall aio_complete(struct kiocb *
 	 * cancelled requests don't get events, userland was given one
 	 * when the event got cancelled.
 	 */
-	if (kiocbIsCancelled(iocb))
+	if (kiocbIsCancelled(iocb)) {
+		if (iocb->ki_notify.sigq)
+			sigqueue_free(iocb->ki_notify.sigq);
 		goto put_rq;
+	}
 
 	ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
 
@@ -994,6 +1077,15 @@ int fastcall aio_complete(struct kiocb *
 	kunmap_atomic(ring, KM_IRQ1);
 
 	pr_debug("added to ring %p at [%lu]\n", iocb, tail);
+
+	if (iocb->ki_notify.notify != SIGEV_NONE) {
+		ret = aio_send_signal(&iocb->ki_notify);
+
+		/* If signal generation failed, release the sigqueue */
+		if (ret)
+			sigqueue_free(iocb->ki_notify.sigq);
+	}
+
 put_rq:
 	/* everything turned out well, dispose of the aiocb. */
 	ret = __aio_put_req(ctx, iocb);
@@ -1545,8 +1637,7 @@ int fastcall io_submit_one(struct kioctx
 	ssize_t ret;
 
 	/* enforce forwards compatibility on users */
-	if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2 ||
-		     iocb->aio_reserved3)) {
+	if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved3)) {
 		pr_debug("EINVAL: io_submit: reserve field set\n");
 		return -EINVAL;
 	}
@@ -1555,6 +1646,7 @@ int fastcall io_submit_one(struct kioctx
 	if (unlikely(
 	    (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
 	    (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
+	    (iocb->aio_sigeventp != (unsigned long)iocb->aio_sigeventp) ||
 	    ((ssize_t)iocb->aio_nbytes < 0)
 	   )) {
 		pr_debug("EINVAL: io_submit: overflow check\n");
@@ -1588,11 +1680,21 @@ int fastcall io_submit_one(struct kioctx
 	init_waitqueue_func_entry(&req->ki_wait.wait, aio_wake_function);
 	INIT_LIST_HEAD(&req->ki_wait.wait.task_list);
  	req->ki_run_list.next = req->ki_run_list.prev = NULL;
+	/* handle setting up the sigevent for POSIX AIO signals */
+	req->ki_notify.notify = SIGEV_NONE;
+
+	if (iocb->aio_sigeventp) {
+		ret = aio_setup_sigevent(&req->ki_notify,
+					 (struct sigevent __user *)(unsigned long)
+					 iocb->aio_sigeventp);
+		if (ret)
+			goto out_put_req;
+	}
 
 	ret = aio_setup_iocb(req);
 
 	if (ret)
-		goto out_put_req;
+		goto out_sigqfree;
 
 	spin_lock_irq(&ctx->ctx_lock);
 	aio_run_iocb(req);
@@ -1605,6 +1707,11 @@ int fastcall io_submit_one(struct kioctx
 	aio_put_req(req);	/* drop extra ref to req */
 	return 0;
 
+out_sigqfree:
+	/* Undo the sigqueue alloc if someting went bad */
+	if (req->ki_notify.sigq)
+		sigqueue_free(req->ki_notify.sigq);
+
 out_put_req:
 	aio_put_req(req);	/* drop extra ref to req */
 	aio_put_req(req);	/* drop i/o ref to req */
Index: linux-2.6.20-rc4-mm1/include/linux/aio_abi.h
===================================================================
--- linux-2.6.20-rc4-mm1.orig/include/linux/aio_abi.h	2006-11-29 22:57:37.000000000 +0100
+++ linux-2.6.20-rc4-mm1/include/linux/aio_abi.h	2007-01-12 12:30:09.000000000 +0100
@@ -82,8 +82,9 @@ struct iocb {
 	__u64	aio_nbytes;
 	__s64	aio_offset;
 
+	__u64	aio_sigeventp;	/* pointer to struct sigevent */
+
 	/* extra parameters */
-	__u64	aio_reserved2;	/* TODO: use this for a (struct sigevent *) */
 	__u64	aio_reserved3;
 }; /* 64 bytes */
 
Index: linux-2.6.20-rc4-mm1/include/linux/aio.h
===================================================================
--- linux-2.6.20-rc4-mm1.orig/include/linux/aio.h	2007-01-12 12:15:01.000000000 +0100
+++ linux-2.6.20-rc4-mm1/include/linux/aio.h	2007-01-12 12:30:09.000000000 +0100
@@ -7,6 +7,7 @@
 #include <linux/uio.h>
 
 #include <asm/atomic.h>
+#include <asm/siginfo.h>
 
 #define AIO_MAXSEGS		4
 #define AIO_KIOGRP_NR_ATOMIC	8
@@ -54,6 +55,14 @@ struct kioctx;
 #define kiocbIsCancelled(iocb)	test_bit(KIF_CANCELLED, &(iocb)->ki_flags)
 #define kiocbIsRestarted(iocb)	test_bit(KIF_RESTARTED, &(iocb)->ki_flags)
 
+struct aio_notify {
+	struct task_struct	*target;
+	__u16			signo;
+	__u16			notify;
+	sigval_t		value;
+	struct sigqueue		*sigq;
+};
+
 /* is there a better place to document function pointer methods? */
 /**
  * ki_retry	-	iocb forward progress callback
@@ -123,6 +132,9 @@ struct kiocb {
 
 	struct list_head	ki_list;	/* the aio core uses this
 						 * for cancellation */
+
+	/* to notify a process on I/O event */
+	struct aio_notify	ki_notify;
 };
 
 #define is_sync_kiocb(iocb)	((iocb)->ki_key == KIOCB_SYNC_KEY)
Index: linux-2.6.20-rc4-mm1/kernel/signal.c
===================================================================
--- linux-2.6.20-rc4-mm1.orig/kernel/signal.c	2007-01-12 12:15:06.000000000 +0100
+++ linux-2.6.20-rc4-mm1/kernel/signal.c	2007-01-12 12:30:09.000000000 +0100
@@ -297,7 +297,7 @@ static struct sigqueue *__sigqueue_alloc
 
 static void __sigqueue_free(struct sigqueue *q)
 {
-	if (q->flags & SIGQUEUE_PREALLOC)
+	if (q->flags & SIGQUEUE_PREALLOC && q->info.si_code != SI_ASYNCIO)
 		return;
 	atomic_dec(&q->user->sigpending);
 	free_uid(q->user);
Index: linux-2.6.20-rc4-mm1/fs/compat.c
===================================================================
--- linux-2.6.20-rc4-mm1.orig/fs/compat.c	2007-01-12 12:14:51.000000000 +0100
+++ linux-2.6.20-rc4-mm1/fs/compat.c	2007-01-12 12:30:09.000000000 +0100
@@ -665,6 +665,7 @@ compat_sys_io_submit(aio_context_t ctx_i
 		compat_uptr_t uptr;
 		struct iocb __user *user_iocb;
 		struct iocb tmp;
+		struct compat_sigevent __user *uevent;
 
 		if (unlikely(get_user(uptr, iocb + i))) {
 			ret = -EFAULT;
@@ -678,6 +679,23 @@ compat_sys_io_submit(aio_context_t ctx_i
 			break;
 		}
 
+		uevent = (struct compat_sigevent __user *)tmp.aio_sigeventp;
+
+		if (uevent) {
+			struct sigevent __user *event = NULL;
+			struct sigevent kevent;
+
+			event = compat_alloc_user_space(sizeof(*event));
+
+			if (get_compat_sigevent(&kevent, uevent) ||
+			    copy_to_user(event, &kevent, sizeof(*event))) {
+				ret = -EFAULT;
+				break;
+			}
+
+			tmp.aio_sigeventp = (__u64)event;
+		}
+
 		ret = io_submit_one(ctx, user_iocb, &tmp);
 		if (ret)
 			break;

[PATCH -mm 5/5][AIO] - Add listio syscall support

[Sébastien Dugué]

From: Bharata B Rao <bharata@in.ibm.com>

This patch provides POSIX listio support by means of a new system call.

long lio_submit(aio_context_t ctx_id, int mode, long nr,
	struct iocb __user * __user *iocbpp, struct sigevent __user *event)

This system call is similar to the io_submit() system call, but takes
two more arguments.

'mode' argument can be LIO_WAIT or LIO_NOWAIT.

'event' argument specifies the signal notification mechanism.

This patch is built on support provided by the aio signal notification
patch by Sebastien. The following two structures together provide
the support for grouping iocbs belonging to a list (lio).

struct aio_notify {
	struct task_struct	*target;
	__u16			signo;
	__u16			notify;
	sigval_t		value;
	struct sigqueue		*sigq;
};

struct lio_event {
	atomic_t		lio_users;
	struct aio_notify	lio_notify;
};

A single lio_event struct is maintained for the list of iocbs.
lio_users holds the number of requests attached to this lio and lio_notify
has the necessary information for generating completion notification
signal.

If the mode is LIO_WAIT, the event argument is ignored and the system
call waits until all the requests of the lio are completed.

If the mode is LIO_NOWAIT, the system call returns immediately after
submitting the io requests and may optionally notify the process on
list io completion depending on the event argument.

Signed-off-by: Sébastien Dugué <sebastien.dugue@bull.net>
Signed-off-by: Laurent Vivier <laurent.vivier@bull.net>
Signed-off-by: Bharata B Rao <bharata@in.ibm.com>
---

 arch/i386/kernel/syscall_table.S |    1 
 arch/x86_64/ia32/ia32entry.S     |    4 
 fs/aio.c                         |  175 ++++++++++++++++++++++++++++++++++-----
 fs/compat.c                      |  117 +++++++++++++++++++++-----
 include/asm-i386/unistd.h        |    3 
 include/asm-x86_64/unistd.h      |    4 
 include/linux/aio.h              |   14 ++-
 include/linux/aio_abi.h          |    5 +
 include/linux/syscalls.h         |    2 
 9 files changed, 280 insertions(+), 45 deletions(-)

Index: linux-2.6.20-rc4-mm1/arch/i386/kernel/syscall_table.S
===================================================================
--- linux-2.6.20-rc4-mm1.orig/arch/i386/kernel/syscall_table.S	2007-01-17 08:58:02.000000000 +0100
+++ linux-2.6.20-rc4-mm1/arch/i386/kernel/syscall_table.S	2007-01-17 08:59:16.000000000 +0100
@@ -319,3 +319,4 @@ ENTRY(sys_call_table)
 	.long sys_move_pages
 	.long sys_getcpu
 	.long sys_epoll_pwait
+	.long sys_lio_submit		/* 320 */
Index: linux-2.6.20-rc4-mm1/fs/aio.c
===================================================================
--- linux-2.6.20-rc4-mm1.orig/fs/aio.c	2007-01-17 08:58:02.000000000 +0100
+++ linux-2.6.20-rc4-mm1/fs/aio.c	2007-01-17 08:59:16.000000000 +0100
@@ -416,6 +416,7 @@ static struct kiocb fastcall *__aio_get_
 	req->ki_ctx = ctx;
 	req->ki_cancel = NULL;
 	req->ki_retry = NULL;
+	req->ki_lio = NULL;
 	req->ki_dtor = NULL;
 	req->private = NULL;
 	req->ki_iovec = NULL;
@@ -996,6 +997,59 @@ out_unlock:
 	return -EINVAL;
 }
 
+void lio_check(struct lio_event *lio)
+{
+	int ret;
+
+	ret = atomic_dec_and_test(&lio->lio_users);
+
+	if (unlikely(ret) && lio->lio_notify.notify != SIGEV_NONE) {
+		/* last one -> notify process */
+		if (aio_send_signal(&lio->lio_notify))
+			sigqueue_free(lio->lio_notify.sigq);
+		kfree(lio);
+	}
+}
+
+struct lio_event *lio_create(struct sigevent __user *user_event,
+			int mode)
+{
+	int ret = 0;
+	struct lio_event *lio = NULL;
+
+	if (unlikely((mode == LIO_NOWAIT) && !user_event))
+		return lio;
+
+	lio = kzalloc(sizeof(*lio), GFP_KERNEL);
+
+	if (!lio)
+		return ERR_PTR(-EAGAIN);
+
+	/*
+	 * Grab an initial ref on the lio to avoid races between
+	 * submission and completion.
+	 */
+	atomic_set(&lio->lio_users, 1);
+
+	lio->lio_notify.notify = SIGEV_NONE;
+
+	/* sigevent argument is ignored with LIO_WAIT */
+	if (user_event && (mode == LIO_NOWAIT)) {
+		/*
+		 * User specified an event for this lio,
+		 * he wants to be notified upon lio completion.
+		 */
+		ret = aio_setup_sigevent(&lio->lio_notify, user_event);
+
+		if (ret) {
+			kfree(lio);
+			return ERR_PTR(ret);
+		}
+	}
+
+	return lio;
+}
+
 /* aio_complete
  *	Called when the io request on the given iocb is complete.
  *	Returns true if this is the last user of the request.  The 
@@ -1044,6 +1098,9 @@ int fastcall aio_complete(struct kiocb *
 	 * when the event got cancelled.
 	 */
 	if (kiocbIsCancelled(iocb)) {
+		if (iocb->ki_lio)
+			lio_check(iocb->ki_lio);
+
 		if (iocb->ki_notify.sigq)
 			sigqueue_free(iocb->ki_notify.sigq);
 		goto put_rq;
@@ -1086,6 +1143,9 @@ int fastcall aio_complete(struct kiocb *
 			sigqueue_free(iocb->ki_notify.sigq);
 	}
 
+	if (iocb->ki_lio)
+		lio_check(iocb->ki_lio);
+
 put_rq:
 	/* everything turned out well, dispose of the aiocb. */
 	ret = __aio_put_req(ctx, iocb);
@@ -1630,7 +1690,7 @@ static int aio_wake_function(wait_queue_
 }
 
 int fastcall io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
-			 struct iocb *iocb)
+			 struct iocb *iocb, struct lio_event *lio)
 {
 	struct kiocb *req;
 	struct file *file;
@@ -1691,6 +1751,9 @@ int fastcall io_submit_one(struct kioctx
 			goto out_put_req;
 	}
 
+	/* Attach this iocb to its lio */
+	req->ki_lio = lio;
+
 	ret = aio_setup_iocb(req);
 
 	if (ret)
@@ -1718,6 +1781,48 @@ out_put_req:
 	return ret;
 }
 
+static int io_submit_group(struct kioctx *ctx, long nr,
+	struct iocb __user * __user *iocbpp, struct lio_event *lio)
+{
+	int i;
+	long ret = 0;
+
+	/*
+	 * AKPM: should this return a partial result if some of the IOs were
+	 * successfully submitted?
+	 */
+	for (i = 0; i < nr; i++) {
+		struct iocb __user *user_iocb;
+		struct iocb tmp;
+
+		if (unlikely(__get_user(user_iocb, iocbpp + i))) {
+			ret = -EFAULT;
+			break;
+		}
+
+		if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
+			ret = -EFAULT;
+			break;
+		}
+
+		if (lio)
+			atomic_inc(&lio->lio_users);
+
+		ret = io_submit_one(ctx, user_iocb, &tmp, lio);
+		if (ret) {
+			if (lio) {
+				/*
+				 * In case of listio, continue with
+				 * the subsequent requests
+				 */
+				atomic_dec(&lio->lio_users);
+			} else
+				break;
+		}
+	}
+	return i ? i : ret;
+}
+
 /* sys_io_submit:
  *	Queue the nr iocbs pointed to by iocbpp for processing.  Returns
  *	the number of iocbs queued.  May return -EINVAL if the aio_context
@@ -1735,7 +1840,6 @@ asmlinkage long sys_io_submit(aio_contex
 {
 	struct kioctx *ctx;
 	long ret = 0;
-	int i;
 
 	if (unlikely(nr < 0))
 		return -EINVAL;
@@ -1749,31 +1853,60 @@ asmlinkage long sys_io_submit(aio_contex
 		return -EINVAL;
 	}
 
-	/*
-	 * AKPM: should this return a partial result if some of the IOs were
-	 * successfully submitted?
-	 */
-	for (i=0; i<nr; i++) {
-		struct iocb __user *user_iocb;
-		struct iocb tmp;
+	ret = io_submit_group(ctx, nr, iocbpp, NULL);
 
-		if (unlikely(__get_user(user_iocb, iocbpp + i))) {
-			ret = -EFAULT;
-			break;
-		}
+	put_ioctx(ctx);
+	return ret;
+}
 
-		if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
-			ret = -EFAULT;
-			break;
-		}
+asmlinkage long sys_lio_submit(aio_context_t ctx_id, int mode, long nr,
+	struct iocb __user * __user *iocbpp, struct sigevent __user *event)
+{
+	struct kioctx *ctx;
+	struct lio_event *lio = NULL;
+	long ret = 0;
 
-		ret = io_submit_one(ctx, user_iocb, &tmp);
-		if (ret)
-			break;
+	if (unlikely(nr < 0))
+		return -EINVAL;
+
+	if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
+		return -EFAULT;
+
+	ctx = lookup_ioctx(ctx_id);
+	if (unlikely(!ctx)) {
+		pr_debug("EINVAL: lio_submit: invalid context id\n");
+		return -EINVAL;
+	}
+
+	lio = lio_create(event, mode);
+
+	ret = PTR_ERR(lio);
+	if (IS_ERR(lio))
+		goto out_put_ctx;
+
+	ret = io_submit_group(ctx, nr, iocbpp, lio);
+
+	/* If we failed to submit even one request just return */
+	if (ret < 0 ) {
+		if (lio)
+			kfree(lio);
+		goto out_put_ctx;
+	}
+
+	/*
+	 * Drop extra ref on the lio now that we're done submitting requests.
+	 */
+	if (lio)
+		lio_check(lio);
+
+	if (mode == LIO_WAIT) {
+		wait_event(ctx->wait, atomic_read(&lio->lio_users) == 0);
+		kfree(lio);
 	}
 
+out_put_ctx:
 	put_ioctx(ctx);
-	return i ? i : ret;
+	return ret;
 }
 
 /* lookup_kiocb
Index: linux-2.6.20-rc4-mm1/fs/compat.c
===================================================================
--- linux-2.6.20-rc4-mm1.orig/fs/compat.c	2007-01-17 08:58:02.000000000 +0100
+++ linux-2.6.20-rc4-mm1/fs/compat.c	2007-01-17 08:59:16.000000000 +0100
@@ -644,24 +644,13 @@ out:
 	return ret;
 }
 
-asmlinkage long
-compat_sys_io_submit(aio_context_t ctx_id, int nr, u32 __user *iocb)
+static int compat_io_submit_group(struct kioctx *ctx, long nr,
+	u32 __user *iocb, struct lio_event *lio)
 {
-	struct kioctx *ctx;
-	long ret = 0;
 	int i;
+	long ret = 0;
 
-	if (unlikely(nr < 0))
-		return -EINVAL;
-
-	if (unlikely(!access_ok(VERIFY_READ, iocb, (nr * sizeof(u32)))))
-		return -EFAULT;
-
-	ctx = lookup_ioctx(ctx_id);
-	if (unlikely(!ctx))
-		return -EINVAL;
-
-	for (i=0; i<nr; i++) {
+	for (i = 0; i < nr; i++) {
 		compat_uptr_t uptr;
 		struct iocb __user *user_iocb;
 		struct iocb tmp;
@@ -696,13 +685,103 @@ compat_sys_io_submit(aio_context_t ctx_i
 			tmp.aio_sigeventp = (__u64)event;
 		}
 
-		ret = io_submit_one(ctx, user_iocb, &tmp);
-		if (ret)
-			break;
+		if (lio)
+			atomic_inc(&lio->lio_users);
+
+		ret = io_submit_one(ctx, user_iocb, &tmp, lio);
+		if (ret) {
+			if (lio) {
+				/*
+				 * In case of listio, continue with
+				 * the subsequent requests
+				 */
+				atomic_dec(&lio->lio_users);
+			} else
+				break;
+		}
+
+
 	}
+	return i ? i : ret;
+}
 
+asmlinkage long
+compat_sys_io_submit(aio_context_t ctx_id, int nr, u32 __user *iocb)
+{
+	struct kioctx *ctx;
+	long ret = 0;
+
+	if (unlikely(nr < 0))
+		return -EINVAL;
+
+	if (unlikely(!access_ok(VERIFY_READ, iocb, (nr * sizeof(u32)))))
+		return -EFAULT;
+
+	ctx = lookup_ioctx(ctx_id);
+	if (unlikely(!ctx))
+		return -EINVAL;
+
+	ret = compat_io_submit_group(ctx, nr, iocb, NULL);
+	put_ioctx(ctx);
+	return ret;
+}
+
+asmlinkage long
+compat_sys_lio_submit(aio_context_t ctx_id, int mode, int nr, u32 __user *iocb,
+		struct compat_sigevent __user *sig_user)
+{
+	struct kioctx *ctx;
+	struct lio_event *lio = NULL;
+	struct sigevent __user *event = NULL;
+	long ret = 0;
+
+	if (unlikely(nr < 0))
+		return -EINVAL;
+
+	if (unlikely(!access_ok(VERIFY_READ, iocb, (nr * sizeof(u32)))))
+		return -EFAULT;
+
+	ctx = lookup_ioctx(ctx_id);
+	if (unlikely(!ctx))
+		return -EINVAL;
+
+	if (sig_user) {
+		struct sigevent kevent;
+		event = compat_alloc_user_space(sizeof(struct sigevent));
+		if (get_compat_sigevent(&kevent, sig_user) ||
+			copy_to_user(event, &kevent, sizeof(struct sigevent)))
+			return -EFAULT;
+	}
+
+	lio = lio_create(event, mode);
+
+	ret = PTR_ERR(lio);
+	if (IS_ERR(lio))
+		goto out_put_ctx;
+
+	ret = compat_io_submit_group(ctx, nr, iocb, lio);
+
+	/* If we failed to submit even one request just return */
+	if (ret < 0) {
+		if (lio)
+			kfree(lio);
+		goto out_put_ctx;
+	}
+
+	/*
+	 * Drop extra ref on the lio now that we're done submitting requests.
+	 */
+	if (lio)
+		lio_check(lio);
+
+
+	if (mode == LIO_WAIT) {
+		wait_event(ctx->wait, atomic_read(&lio->lio_users) == 0);
+		kfree(lio);
+	}
+out_put_ctx:
 	put_ioctx(ctx);
-	return i ? i: ret;
+	return ret;
 }
 
 struct compat_ncp_mount_data {
Index: linux-2.6.20-rc4-mm1/include/asm-i386/unistd.h
===================================================================
--- linux-2.6.20-rc4-mm1.orig/include/asm-i386/unistd.h	2007-01-17 08:58:02.000000000 +0100
+++ linux-2.6.20-rc4-mm1/include/asm-i386/unistd.h	2007-01-17 08:59:16.000000000 +0100
@@ -325,10 +325,11 @@
 #define __NR_move_pages		317
 #define __NR_getcpu		318
 #define __NR_epoll_pwait	319
+#define __NR_lio_submit		320
 
 #ifdef __KERNEL__
 
-#define NR_syscalls 320
+#define NR_syscalls 321
 
 #define __ARCH_WANT_IPC_PARSE_VERSION
 #define __ARCH_WANT_OLD_READDIR
Index: linux-2.6.20-rc4-mm1/include/linux/aio_abi.h
===================================================================
--- linux-2.6.20-rc4-mm1.orig/include/linux/aio_abi.h	2007-01-17 08:58:02.000000000 +0100
+++ linux-2.6.20-rc4-mm1/include/linux/aio_abi.h	2007-01-17 08:59:16.000000000 +0100
@@ -45,6 +45,11 @@ enum {
 	IOCB_CMD_PWRITEV = 8,
 };
 
+enum {
+	LIO_WAIT = 0,
+	LIO_NOWAIT = 1,
+};
+
 /* read() from /dev/aio returns these structures. */
 struct io_event {
 	__u64		data;		/* the data field from the iocb */
Index: linux-2.6.20-rc4-mm1/include/linux/aio.h
===================================================================
--- linux-2.6.20-rc4-mm1.orig/include/linux/aio.h	2007-01-17 08:58:02.000000000 +0100
+++ linux-2.6.20-rc4-mm1/include/linux/aio.h	2007-01-17 08:59:16.000000000 +0100
@@ -63,6 +63,11 @@ struct aio_notify {
 	struct sigqueue		*sigq;
 };
 
+struct lio_event {
+	atomic_t		lio_users;
+	struct aio_notify	lio_notify;
+};
+
 /* is there a better place to document function pointer methods? */
 /**
  * ki_retry	-	iocb forward progress callback
@@ -117,6 +122,8 @@ struct kiocb {
 	__u64			ki_user_data;	/* user's data for completion */
 	struct wait_bit_queue	ki_wait;
 	loff_t			ki_pos;
+	/* lio this iocb might be attached to */
+	struct lio_event	*ki_lio;
 
 	atomic_t		ki_bio_count;	/* num bio used for this iocb */
 	void			*private;
@@ -225,12 +232,15 @@ struct mm_struct;
 extern void FASTCALL(exit_aio(struct mm_struct *mm));
 extern struct kioctx *lookup_ioctx(unsigned long ctx_id);
 extern int FASTCALL(io_submit_one(struct kioctx *ctx,
-			struct iocb __user *user_iocb, struct iocb *iocb));
+			struct iocb __user *user_iocb, struct iocb *iocb,
+			struct lio_event *lio));
+struct lio_event *lio_create(struct sigevent __user *user_event, int mode);
+void lio_check(struct lio_event *lio);
 
 /* semi private, but used by the 32bit emulations: */
 struct kioctx *lookup_ioctx(unsigned long ctx_id);
 int FASTCALL(io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
-				  struct iocb *iocb));
+				  struct iocb *iocb, struct lio_event *lio));
 
 #define get_ioctx(kioctx) do {						\
 	BUG_ON(atomic_read(&(kioctx)->users) <= 0);			\
Index: linux-2.6.20-rc4-mm1/include/linux/syscalls.h
===================================================================
--- linux-2.6.20-rc4-mm1.orig/include/linux/syscalls.h	2007-01-17 08:58:02.000000000 +0100
+++ linux-2.6.20-rc4-mm1/include/linux/syscalls.h	2007-01-17 08:59:16.000000000 +0100
@@ -317,6 +317,8 @@ asmlinkage long sys_io_getevents(aio_con
 				struct timespec __user *timeout);
 asmlinkage long sys_io_submit(aio_context_t, long,
 				struct iocb __user * __user *);
+asmlinkage long sys_lio_submit(aio_context_t, int, long,
+		struct iocb __user * __user *, struct sigevent __user *);
 asmlinkage long sys_io_cancel(aio_context_t ctx_id, struct iocb __user *iocb,
 			      struct io_event __user *result);
 asmlinkage ssize_t sys_sendfile(int out_fd, int in_fd,
Index: linux-2.6.20-rc4-mm1/include/asm-x86_64/unistd.h
===================================================================
--- linux-2.6.20-rc4-mm1.orig/include/asm-x86_64/unistd.h	2007-01-17 08:58:02.000000000 +0100
+++ linux-2.6.20-rc4-mm1/include/asm-x86_64/unistd.h	2007-01-17 08:59:16.000000000 +0100
@@ -619,8 +619,10 @@ __SYSCALL(__NR_sync_file_range, sys_sync
 __SYSCALL(__NR_vmsplice, sys_vmsplice)
 #define __NR_move_pages		279
 __SYSCALL(__NR_move_pages, sys_move_pages)
+#define __NR_lio_submit		280
+__SYSCALL(__NR_lio_submit, sys_lio_submit)
 
-#define __NR_syscall_max __NR_move_pages
+#define __NR_syscall_max __NR_lio_submit
 
 #ifndef __NO_STUBS
 #define __ARCH_WANT_OLD_READDIR
Index: linux-2.6.20-rc4-mm1/arch/x86_64/ia32/ia32entry.S
===================================================================
--- linux-2.6.20-rc4-mm1.orig/arch/x86_64/ia32/ia32entry.S	2007-01-17 08:58:02.000000000 +0100
+++ linux-2.6.20-rc4-mm1/arch/x86_64/ia32/ia32entry.S	2007-01-17 08:59:16.000000000 +0100
@@ -714,8 +714,10 @@ ia32_sys_call_table:
 	.quad compat_sys_get_robust_list
 	.quad sys_splice
 	.quad sys_sync_file_range
-	.quad sys_tee
+	.quad sys_tee			/* 315 */
 	.quad compat_sys_vmsplice
 	.quad compat_sys_move_pages
 	.quad sys_getcpu
+	.quad quiet_ni_syscall		/* sys_epoll_wait */
+	.quad compat_sys_lio_submit
 ia32_syscall_end:		

Re: [PATCH -mm 4/5][AIO] - AIO completion signal notification

[Andrew Morton]

On Wed, 17 Jan 2007 10:50:18 +0100
Sébastien Dugué <sebastien.dugue@bull.net> wrote:

> +static long aio_setup_sigevent(struct aio_notify *notify,
> +			       struct sigevent __user *user_event)
> +{
> +	sigevent_t event;
> +	struct task_struct *target;
> +
> +	if (copy_from_user(&event, user_event, sizeof (event)))
> +		return -EFAULT;
> +
> +	if (event.sigev_notify == SIGEV_NONE)
> +		return 0;
> +
> +	notify->notify = event.sigev_notify;
> +	notify->signo = event.sigev_signo;
> +	notify->value = event.sigev_value;
> +
> +	read_lock(&tasklist_lock);
> +	target = good_sigevent(&event);
> +
> +	if (unlikely(!target || (target->flags & PF_EXITING)))
> +		goto out_unlock;
> +
> +	/*
> +	 * At this point, we know that notify is either SIGEV_SIGNAL or
> +	 * SIGEV_THREAD_ID and the target task is valid. So get a reference
> +	 * on the task, it will be dropped in really_put_req() when
> +	 * we're done with the request.
> +	 */
> +	get_task_struct(target);
> +	notify->target = target;
> +	read_unlock(&tasklist_lock);
> +
> +	/*
> +	 * NOTE: we cannot free the sigqueue in the completion path as
> +	 * the signal may not have been delivered to the target task.
> +	 * Therefore it has to be freed in __sigqueue_free() when the
> +	 * signal is collected if si_code is SI_ASYNCIO.
> +	 */
> +	notify->sigq = sigqueue_alloc();
> +
> +	if (unlikely(!notify->sigq))
> +		return -EAGAIN;

Did this just leak a ref on the task_struct?

> +
> +	return 0;
> +
> +out_unlock:
> +	read_unlock(&tasklist_lock);
> +	return -EINVAL;
> +}

Re: [PATCH -mm 5/5][AIO] - Add listio syscall support

[Andrew Morton]

On Wed, 17 Jan 2007 10:55:54 +0100
Sébastien Dugué <sebastien.dugue@bull.net> wrote:

> +asmlinkage long
> +compat_sys_lio_submit(aio_context_t ctx_id, int mode, int nr, u32 __user *iocb,
> +		struct compat_sigevent __user *sig_user)
> +{
> +	struct kioctx *ctx;
> +	struct lio_event *lio = NULL;
> +	struct sigevent __user *event = NULL;
> +	long ret = 0;
> +
> +	if (unlikely(nr < 0))
> +		return -EINVAL;
> +
> +	if (unlikely(!access_ok(VERIFY_READ, iocb, (nr * sizeof(u32)))))
> +		return -EFAULT;
> +
> +	ctx = lookup_ioctx(ctx_id);
> +	if (unlikely(!ctx))
> +		return -EINVAL;
> +
> +	if (sig_user) {
> +		struct sigevent kevent;
> +		event = compat_alloc_user_space(sizeof(struct sigevent));
> +		if (get_compat_sigevent(&kevent, sig_user) ||
> +			copy_to_user(event, &kevent, sizeof(struct sigevent)))
> +			return -EFAULT;

I think we just leaked a ref against the ioctx.

That's two.  Please re-review the whole patchset for leaks like this.

Please also do not do returns from the middle of functions like this.  It's just
asking for resource leaks, either now or in the future.


> +	}
> +
> +	lio = lio_create(event, mode);
> +
> +	ret = PTR_ERR(lio);
> +	if (IS_ERR(lio))
> +		goto out_put_ctx;
> +
> +	ret = compat_io_submit_group(ctx, nr, iocb, lio);
> +
> +	/* If we failed to submit even one request just return */
> +	if (ret < 0) {
> +		if (lio)
> +			kfree(lio);
> +		goto out_put_ctx;
> +	}
> +
> +	/*
> +	 * Drop extra ref on the lio now that we're done submitting requests.
> +	 */
> +	if (lio)
> +		lio_check(lio);
> +
> +
> +	if (mode == LIO_WAIT) {
> +		wait_event(ctx->wait, atomic_read(&lio->lio_users) == 0);
> +		kfree(lio);
> +	}
> +out_put_ctx:
>  	put_ioctx(ctx);
> -	return i ? i: ret;
> +	return ret;
>  }

Re: [PATCH -mm 5/5][AIO] - Add listio syscall support

[Andrew Morton]

On Wed, 17 Jan 2007 10:55:54 +0100
Sébastien Dugué <sebastien.dugue@bull.net> wrote:

> +struct lio_event *lio_create(struct sigevent __user *user_event,
> +			int mode)
> +{
> +	int ret = 0;
> +	struct lio_event *lio = NULL;
> +
> +	if (unlikely((mode == LIO_NOWAIT) && !user_event))
> +		return lio;
> +
> +	lio = kzalloc(sizeof(*lio), GFP_KERNEL);
> +
> +	if (!lio)
> +		return ERR_PTR(-EAGAIN);
> +

Why EAGAIN and not ENOMEM?

Re: [PATCH -mm 5/5][AIO] - Add listio syscall support

[Andrew Morton]

On Wed, 17 Jan 2007 10:55:54 +0100
Sébastien Dugué <sebastien.dugue@bull.net> wrote:

> +void lio_check(struct lio_event *lio)
> +{
> +	int ret;
> +
> +	ret = atomic_dec_and_test(&lio->lio_users);
> +
> +	if (unlikely(ret) && lio->lio_notify.notify != SIGEV_NONE) {
> +		/* last one -> notify process */
> +		if (aio_send_signal(&lio->lio_notify))
> +			sigqueue_free(lio->lio_notify.sigq);
> +		kfree(lio);
> +	}
> +}

That's a scary function.  It may (or may not) free the memory at lio,
returning no indication to the caller whether or not that memory is still
allocated.  This is most peculiar - are you really sure there's no
potential for a use-after-free here?

The function is poorly named: I'd expect something called "foo_check" to
not have any side-effects.  This one has gross side-effects.  Want to think
up a better name, please?

And given that this function has global scope, perhaps a little explanatory
comment is in order?

> +struct lio_event *lio_create(struct sigevent __user *user_event,
> +			int mode)

Here too.

Re: [PATCH -mm 5/5][AIO] - Add listio syscall support

[Laurent Vivier]

[-- Attachment #1: Type: text/plain, Size: 987 bytes --]

Andrew Morton wrote:
> On Wed, 17 Jan 2007 10:55:54 +0100
> Sébastien Dugué <sebastien.dugue@bull.net> wrote:
> 
>> +struct lio_event *lio_create(struct sigevent __user *user_event,
>> +			int mode)
>> +{
>> +	int ret = 0;
>> +	struct lio_event *lio = NULL;
>> +
>> +	if (unlikely((mode == LIO_NOWAIT) && !user_event))
>> +		return lio;
>> +
>> +	lio = kzalloc(sizeof(*lio), GFP_KERNEL);
>> +
>> +	if (!lio)
>> +		return ERR_PTR(-EAGAIN);
>> +
> 
> Why EAGAIN and not ENOMEM?

According to the POSIX AIO specifications:

"ERRORS:
...
[EAGAIN]
    The resources necessary to queue all the I/O requests were not available. "

I think memory is "a resource necessary to queue all the I/O requests"...

http://www.opengroup.org/onlinepubs/009695399/functions/lio_listio.html

Regards,
Laurent
-- 
------------- Laurent.Vivier@bull.net  --------------
       "Any sufficiently advanced technology is
  indistinguishable from magic." - Arthur C. Clarke


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 189 bytes --]

Re: [PATCH -mm 4/5][AIO] - AIO completion signal notification

[Sébastien Dugué]

On Tue, 23 Jan 2007 21:35:13 -0800 Andrew Morton <akpm@osdl.org> wrote:

> On Wed, 17 Jan 2007 10:50:18 +0100
> Sébastien Dugué <sebastien.dugue@bull.net> wrote:
> 
> > +static long aio_setup_sigevent(struct aio_notify *notify,
> > +			       struct sigevent __user *user_event)
> > +{
> > +	sigevent_t event;
> > +	struct task_struct *target;
> > +
> > +	if (copy_from_user(&event, user_event, sizeof (event)))
> > +		return -EFAULT;
> > +
> > +	if (event.sigev_notify == SIGEV_NONE)
> > +		return 0;
> > +
> > +	notify->notify = event.sigev_notify;
> > +	notify->signo = event.sigev_signo;
> > +	notify->value = event.sigev_value;
> > +
> > +	read_lock(&tasklist_lock);
> > +	target = good_sigevent(&event);
> > +
> > +	if (unlikely(!target || (target->flags & PF_EXITING)))
> > +		goto out_unlock;
> > +
> > +	/*
> > +	 * At this point, we know that notify is either SIGEV_SIGNAL or
> > +	 * SIGEV_THREAD_ID and the target task is valid. So get a reference
> > +	 * on the task, it will be dropped in really_put_req() when
> > +	 * we're done with the request.
> > +	 */
> > +	get_task_struct(target);
> > +	notify->target = target;
> > +	read_unlock(&tasklist_lock);
> > +
> > +	/*
> > +	 * NOTE: we cannot free the sigqueue in the completion path as
> > +	 * the signal may not have been delivered to the target task.
> > +	 * Therefore it has to be freed in __sigqueue_free() when the
> > +	 * signal is collected if si_code is SI_ASYNCIO.
> > +	 */
> > +	notify->sigq = sigqueue_alloc();
> > +
> > +	if (unlikely(!notify->sigq))
> > +		return -EAGAIN;
> 
> Did this just leak a ref on the task_struct?
> 

  No, the ref is released in really_put_req() when we dispose of
the iocb.

  Thanks,

  Sébastien.

Re: [PATCH -mm 5/5][AIO] - Add listio syscall support

[Sébastien Dugué]

On Tue, 23 Jan 2007 21:50:47 -0800 Andrew Morton <akpm@osdl.org> wrote:

> On Wed, 17 Jan 2007 10:55:54 +0100
> Sébastien Dugué <sebastien.dugue@bull.net> wrote:
> 
> > +asmlinkage long
> > +compat_sys_lio_submit(aio_context_t ctx_id, int mode, int nr, u32 __user *iocb,
> > +		struct compat_sigevent __user *sig_user)
> > +{
> > +	struct kioctx *ctx;
> > +	struct lio_event *lio = NULL;
> > +	struct sigevent __user *event = NULL;
> > +	long ret = 0;
> > +
> > +	if (unlikely(nr < 0))
> > +		return -EINVAL;
> > +
> > +	if (unlikely(!access_ok(VERIFY_READ, iocb, (nr * sizeof(u32)))))
> > +		return -EFAULT;
> > +
> > +	ctx = lookup_ioctx(ctx_id);
> > +	if (unlikely(!ctx))
> > +		return -EINVAL;
> > +
> > +	if (sig_user) {
> > +		struct sigevent kevent;
> > +		event = compat_alloc_user_space(sizeof(struct sigevent));
> > +		if (get_compat_sigevent(&kevent, sig_user) ||
> > +			copy_to_user(event, &kevent, sizeof(struct sigevent)))
> > +			return -EFAULT;
> 
> I think we just leaked a ref against the ioctx.

  Right, will fix.

> 
> That's two.  Please re-review the whole patchset for leaks like this.

  Well, this particular patch did not have much review so far (contrary to the
the other patches in the patchset). That's why I posted it mainly
for comments and not expecting it to make it to -mm so soon.

> 
> Please also do not do returns from the middle of functions like this.  It's just
> asking for resource leaks, either now or in the future.
> 

  Argh, yes. Who wrote that code? ;-)

  Thanks,

  Sébastien.

Re: [PATCH -mm 5/5][AIO] - Add listio syscall support

[Sébastien Dugué]

On Wed, 24 Jan 2007 10:52:58 +0100 Laurent Vivier <Laurent.Vivier@bull.net> wrote:

> Andrew Morton wrote:
> > On Wed, 17 Jan 2007 10:55:54 +0100
> > Sébastien Dugué <sebastien.dugue@bull.net> wrote:
> > 
> >> +struct lio_event *lio_create(struct sigevent __user *user_event,
> >> +			int mode)
> >> +{
> >> +	int ret = 0;
> >> +	struct lio_event *lio = NULL;
> >> +
> >> +	if (unlikely((mode == LIO_NOWAIT) && !user_event))
> >> +		return lio;
> >> +
> >> +	lio = kzalloc(sizeof(*lio), GFP_KERNEL);
> >> +
> >> +	if (!lio)
> >> +		return ERR_PTR(-EAGAIN);
> >> +
> > 
> > Why EAGAIN and not ENOMEM?
> 
> According to the POSIX AIO specifications:
> 
> "ERRORS:
> ...
> [EAGAIN]
>     The resources necessary to queue all the I/O requests were not available. "
> 
> I think memory is "a resource necessary to queue all the I/O requests"...
> 
> http://www.opengroup.org/onlinepubs/009695399/functions/lio_listio.html
> 
> Regards,
> Laurent

  Thanks Laurent,

  just forgot the whys here.

  Sébastien.

Re: [PATCH -mm 5/5][AIO] - Add listio syscall support

[Sébastien Dugué]

On Tue, 23 Jan 2007 22:04:33 -0800 Andrew Morton <akpm@osdl.org> wrote:

> On Wed, 17 Jan 2007 10:55:54 +0100
> Sébastien Dugué <sebastien.dugue@bull.net> wrote:
> 
> > +void lio_check(struct lio_event *lio)
> > +{
> > +	int ret;
> > +
> > +	ret = atomic_dec_and_test(&lio->lio_users);
> > +
> > +	if (unlikely(ret) && lio->lio_notify.notify != SIGEV_NONE) {
> > +		/* last one -> notify process */
> > +		if (aio_send_signal(&lio->lio_notify))
> > +			sigqueue_free(lio->lio_notify.sigq);
> > +		kfree(lio);
> > +	}
> > +}
> 
> That's a scary function.  It may (or may not) free the memory at lio,
> returning no indication to the caller whether or not that memory is still
> allocated.  This is most peculiar - are you really sure there's no
> potential for a use-after-free here?

  Right again, this patch definitely needs more eyes peering over.

> 
> The function is poorly named: I'd expect something called "foo_check" to
> not have any side-effects.  This one has gross side-effects.  Want to think
> up a better name, please?
> 
> And given that this function has global scope, perhaps a little explanatory
> comment is in order?
> 
> > +struct lio_event *lio_create(struct sigevent __user *user_event,
> > +			int mode)
> 
> Here too.

  OK, will look into this. In the meantime, maybe you should drop this one
patch entirely.

  Thanks,

  Sébastien.

Re: [PATCH -mm 5/5][AIO] - Add listio syscall support

[Bharata B Rao]

On Tue, Jan 23, 2007 at 10:04:33PM -0800, Andrew Morton wrote:
> On Wed, 17 Jan 2007 10:55:54 +0100
> Sébastien Dugué <sebastien.dugue@bull.net> wrote:
> 
> > +void lio_check(struct lio_event *lio)
> > +{
> > +	int ret;
> > +
> > +	ret = atomic_dec_and_test(&lio->lio_users);
> > +
> > +	if (unlikely(ret) && lio->lio_notify.notify != SIGEV_NONE) {
> > +		/* last one -> notify process */
> > +		if (aio_send_signal(&lio->lio_notify))
> > +			sigqueue_free(lio->lio_notify.sigq);
> > +		kfree(lio);
> > +	}
> > +}
> 
> That's a scary function.  It may (or may not) free the memory at lio,
> returning no indication to the caller whether or not that memory is still
> allocated.  This is most peculiar - are you really sure there's no
> potential for a use-after-free here?

Yes, this function looks peculiar. Actually lio gets freed here only
for LIO_NOWAIT case. For LIO_WAIT case, it gets freed at the end
of sys_lio_submit() after it is done waiting for all io's.
But yes, all this is not very obvious.

> 
> The function is poorly named: I'd expect something called "foo_check" to
> not have any side-effects.  This one has gross side-effects.  Want to think
> up a better name, please?
> 
> And given that this function has global scope, perhaps a little explanatory
> comment is in order?
> 
> > +struct lio_event *lio_create(struct sigevent __user *user_event,
> > +			int mode)
> 
> Here too.

Ok, will try to take care of all these in the next iteration.

Thanks for your review.

Regards,
Bharata.

Re: [PATCH -mm 4/5][AIO] - AIO completion signal notification

[Christoph Hellwig]

On Wed, Jan 24, 2007 at 12:11:30PM +0100, S?bastien Dugu? wrote:
> > > +	if (unlikely(!notify->sigq))
> > > +		return -EAGAIN;
> > 
> > Did this just leak a ref on the task_struct?
> > 
> 
>   No, the ref is released in really_put_req() when we dispose of
> the iocb.

And the code really needs a comment explaining this.  I tripped over
this before, and I think it's even already the second time Andrew
stumbled over it.

Re: [PATCH -mm 4/5][AIO] - AIO completion signal notification

[Sébastien Dugué]

On Thu, 25 Jan 2007 05:42:42 +0000 Christoph Hellwig <hch@infradead.org> wrote:

> On Wed, Jan 24, 2007 at 12:11:30PM +0100, S?bastien Dugu? wrote:
> > > > +	if (unlikely(!notify->sigq))
> > > > +		return -EAGAIN;
> > > 
> > > Did this just leak a ref on the task_struct?
> > > 
> > 
> >   No, the ref is released in really_put_req() when we dispose of
> > the iocb.
> 
> And the code really needs a comment explaining this.  I tripped over
> this before, and I think it's even already the second time Andrew
> stumbled over it.
> 

  OK, will do.

  Thanks,

  Sébastien.

Re: [PATCH -mm 4/5][AIO] - AIO completion signal notification

[Sébastien Dugué]

                      AIO completion signal notification

  The current 2.6 kernel does not support notification of user space via
an RT signal upon an asynchronous IO completion. The POSIX specification
states that when an AIO request completes, a signal can be delivered to
the application as notification.

  This patch adds a struct sigevent *aio_sigeventp to the iocb.
The relevant fields (pid, signal number and value) are stored in the kiocb
for use when the request completes.

  That sigevent structure is filled by the application as part of the AIO
request preparation. Upon request completion, the kernel notifies the
application using those sigevent parameters. If SIGEV_NONE has been specified,
then the old behaviour is retained and the application must rely on polling
the completion queue using io_getevents().


  A struct sigevent *aio_sigeventp field is added to struct iocb in
include/linux/aio_abi.h

  A struct aio_notify containing the sigevent parameters is defined in aio.h:

  struct aio_notify {
	struct task_struct	*target;
	__u16			signo;
	__u16			notify;
	sigval_t		value;
  };

  A struct aio_notify ki_notify is added to struct kiocb in include/linux/aio.h

  In io_submit_one(), if the application provided a sigevent then
setup_sigevent() is called which does the following:

	- check access to the user sigevent and make a local copy

	- if the requested notification is SIGEV_NONE, then nothing to do

	- fill in the kiocb->ki_notify fields (notify, signo, value)

	- check sigevent consistency, get the signal target task and
	  save it in kiocb->ki_notify.target

	- preallocate a sigqueue for this event using sigqueue_alloc()

  Upon request completion, in aio_complete(), if notification is needed for
this request (iocb->ki_notify.notify != SIGEV_NONE), then aio_send_signal()
is called to signal the target task as follows:

	- fill in the siginfo struct to be sent to the task

	- if notify is SIGEV_THREAD_ID then send signal to specific task
	  using send_sigqueue()

	- else send signal to task group using send_5group_sigqueue()



Notes concerning sigqueue preallocation:

 To ensure reliable delivery of completion notification, the sigqueue is
preallocated in the submission path so that there is no chance it can fail
in the completion path.

 Unlike the posix-timers case (currently the single other user of sigqueue
preallocation), where the sigqueue is allocated for the lifetime of the timer
and freed at timer destruction time, the aio case is a bit more tricky due to
the async nature of the whole thing.

  In the aio case, the sigqueue exists for the lifetime of the request, therefore
it must be freed only once the signal for the request completion has been
delivered. This involves changing __sigqueue_free() to free the sigqueue when the
signal is collected if si_code is SI_ASYNCIO even if it was preallocated as well
as explicitly calling sigqueue_free() in submission and completion error paths.


 fs/aio.c                |  115 ++++++++++++++++++++++++++++++++++++++++++++++--
 fs/compat.c             |   18 +++++++
 include/linux/aio.h     |   12 +++++
 include/linux/aio_abi.h |    3 -
 kernel/signal.c         |    2
 5 files changed, 144 insertions(+), 6 deletions(-)

Signed-off-by: Sébastien Dugué <sebastien.dugue@bull.net>
Signed-off-by: Laurent Vivier <laurent.vivier@bull.net>

Index: linux-2.6.19-rc6-mm2/fs/aio.c
===================================================================
--- linux-2.6.19-rc6-mm2.orig/fs/aio.c	2006-11-30 10:54:16.000000000 +0100
+++ linux-2.6.19-rc6-mm2/fs/aio.c	2006-11-30 15:18:52.000000000 +0100
@@ -416,6 +416,7 @@ static struct kiocb fastcall *__aio_get_
 	req->ki_dtor = NULL;
 	req->private = NULL;
 	req->ki_iovec = NULL;
+	req->ki_notify.sigq = NULL;
 	INIT_LIST_HEAD(&req->ki_run_list);
 
 	/* Check if the completion queue has enough free space to
@@ -463,6 +464,12 @@ static inline void really_put_req(struct
 		req->ki_dtor(req);
 	if (req->ki_iovec != &req->ki_inline_vec)
 		kfree(req->ki_iovec);
+
+	/* Release task ref */
+	if (req->ki_notify.notify == SIGEV_THREAD_ID ||
+	    req->ki_notify.notify == SIGEV_SIGNAL)
+		put_task_struct(req->ki_notify.target);
+
 	kmem_cache_free(kiocb_cachep, req);
 	ctx->reqs_active--;
 
@@ -929,6 +936,79 @@ void fastcall kick_iocb(struct kiocb *io
 }
 EXPORT_SYMBOL(kick_iocb);
 
+static int aio_send_signal(struct aio_notify *notify)
+{
+	struct sigqueue *sigq = notify->sigq;
+	struct siginfo *info = &sigq->info;
+	int ret;
+
+	memset(info, 0, sizeof(struct siginfo));
+
+	info->si_signo = notify->signo;
+	info->si_errno = 0;
+	info->si_code = SI_ASYNCIO;
+	info->si_pid = 0;
+	info->si_uid = 0;
+	info->si_value = notify->value;
+
+	if (notify->notify & SIGEV_THREAD_ID)
+		ret = send_sigqueue(notify->signo, sigq, notify->target);
+	else
+		ret = send_group_sigqueue(notify->signo, sigq, notify->target);
+
+	return ret;
+}
+
+static long aio_setup_sigevent(struct aio_notify *notify,
+			       struct sigevent __user *user_event)
+{
+	sigevent_t event;
+	struct task_struct *target;
+
+	if (copy_from_user(&event, user_event, sizeof (event)))
+		return -EFAULT;
+
+	if (event.sigev_notify == SIGEV_NONE)
+		return 0;
+
+	notify->notify = event.sigev_notify;
+	notify->signo = event.sigev_signo;
+	notify->value = event.sigev_value;
+
+	read_lock(&tasklist_lock);
+	target = good_sigevent(&event);
+
+	if (unlikely(!target || (target->flags & PF_EXITING)))
+		goto out_unlock;
+
+	/*
+	 * At this point, we know that notify is either SIGEV_SIGNAL or
+	 * SIGEV_THREAD_ID and the target task is valid. So get a reference
+	 * on the task, it will be dropped in really_put_req() when
+	 * we're done with the request.
+	 */
+	get_task_struct(target);
+	notify->target = target;
+	read_unlock(&tasklist_lock);
+
+	/*
+	 * NOTE: we cannot free the sigqueue in the completion path as
+	 * the signal may not have been delivered to the target task.
+	 * Therefore it has to be freed in __sigqueue_free() when the
+	 * signal is collected if si_code is SI_ASYNCIO.
+	 */
+	notify->sigq = sigqueue_alloc();
+
+	if (unlikely(!notify->sigq))
+		return -EAGAIN;
+
+	return 0;
+
+out_unlock:
+	read_unlock(&tasklist_lock);
+	return -EINVAL;
+}
+
 /* aio_complete
  *	Called when the io request on the given iocb is complete.
  *	Returns true if this is the last user of the request.  The 
@@ -976,8 +1056,11 @@ int fastcall aio_complete(struct kiocb *
 	 * cancelled requests don't get events, userland was given one
 	 * when the event got cancelled.
 	 */
-	if (kiocbIsCancelled(iocb))
+	if (kiocbIsCancelled(iocb)) {
+		if (iocb->ki_notify.sigq)
+			sigqueue_free(iocb->ki_notify.sigq);
 		goto put_rq;
+	}
 
 	ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
 
@@ -1008,6 +1091,14 @@ int fastcall aio_complete(struct kiocb *
 
 	pr_debug("added to ring %p at [%lu]\n", iocb, tail);
 
+	if (iocb->ki_notify.notify != SIGEV_NONE) {
+		ret = aio_send_signal(&iocb->ki_notify);
+
+		/* If signal generation failed, release the sigqueue */
+		if (ret)
+			sigqueue_free(iocb->ki_notify.sigq);
+	}
+
 	pr_debug("%ld retries: %zd of %zd\n", iocb->ki_retried,
 		iocb->ki_nbytes - iocb->ki_left, iocb->ki_nbytes);
 put_rq:
@@ -1549,8 +1640,7 @@ int fastcall io_submit_one(struct kioctx
 	ssize_t ret;
 
 	/* enforce forwards compatibility on users */
-	if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2 ||
-		     iocb->aio_reserved3)) {
+	if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved3)) {
 		pr_debug("EINVAL: io_submit: reserve field set\n");
 		return -EINVAL;
 	}
@@ -1559,6 +1649,7 @@ int fastcall io_submit_one(struct kioctx
 	if (unlikely(
 	    (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
 	    (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
+	    (iocb->aio_sigeventp != (unsigned long)iocb->aio_sigeventp) ||
 	    ((ssize_t)iocb->aio_nbytes < 0)
 	   )) {
 		pr_debug("EINVAL: io_submit: overflow check\n");
@@ -1593,10 +1684,21 @@ int fastcall io_submit_one(struct kioctx
 	INIT_LIST_HEAD(&req->ki_wait.task_list);
 	req->ki_retried = 0;
 
+	/* handle setting up the sigevent for POSIX AIO signals */
+	req->ki_notify.notify = SIGEV_NONE;
+
+	if (iocb->aio_sigeventp) {
+		ret = aio_setup_sigevent(&req->ki_notify,
+					 (struct sigevent __user *)(unsigned long)
+					 iocb->aio_sigeventp);
+		if (ret)
+			goto out_put_req;
+	}
+
 	ret = aio_setup_iocb(req);
 
 	if (ret)
-		goto out_put_req;
+		goto out_sigqfree;
 
 	spin_lock_irq(&ctx->ctx_lock);
 	aio_run_iocb(req);
@@ -1609,6 +1711,11 @@ int fastcall io_submit_one(struct kioctx
 	aio_put_req(req);	/* drop extra ref to req */
 	return 0;
 
+out_sigqfree:
+	/* Undo the sigqueue alloc if someting went bad */
+	if (req->ki_notify.sigq)
+		sigqueue_free(req->ki_notify.sigq);
+
 out_put_req:
 	aio_put_req(req);	/* drop extra ref to req */
 	aio_put_req(req);	/* drop i/o ref to req */
Index: linux-2.6.19-rc6-mm2/include/linux/aio_abi.h
===================================================================
--- linux-2.6.19-rc6-mm2.orig/include/linux/aio_abi.h	2006-11-30 10:54:16.000000000 +0100
+++ linux-2.6.19-rc6-mm2/include/linux/aio_abi.h	2006-11-30 15:06:05.000000000 +0100
@@ -82,8 +82,9 @@ struct iocb {
 	__u64	aio_nbytes;
 	__s64	aio_offset;
 
+	__u64	aio_sigeventp;	/* pointer to struct sigevent */
+
 	/* extra parameters */
-	__u64	aio_reserved2;	/* TODO: use this for a (struct sigevent *) */
 	__u64	aio_reserved3;
 }; /* 64 bytes */
 
Index: linux-2.6.19-rc6-mm2/include/linux/aio.h
===================================================================
--- linux-2.6.19-rc6-mm2.orig/include/linux/aio.h	2006-11-30 13:18:15.000000000 +0100
+++ linux-2.6.19-rc6-mm2/include/linux/aio.h	2006-11-30 15:06:05.000000000 +0100
@@ -7,6 +7,7 @@
 #include <linux/uio.h>
 
 #include <asm/atomic.h>
+#include <asm/siginfo.h>
 
 #define AIO_MAXSEGS		4
 #define AIO_KIOGRP_NR_ATOMIC	8
@@ -49,6 +50,14 @@ struct kioctx;
 #define kiocbIsKicked(iocb)	test_bit(KIF_KICKED, &(iocb)->ki_flags)
 #define kiocbIsCancelled(iocb)	test_bit(KIF_CANCELLED, &(iocb)->ki_flags)
 
+struct aio_notify {
+	struct task_struct	*target;
+	__u16			signo;
+	__u16			notify;
+	sigval_t		value;
+	struct sigqueue		*sigq;
+};
+
 /* is there a better place to document function pointer methods? */
 /**
  * ki_retry	-	iocb forward progress callback
@@ -118,6 +127,9 @@ struct kiocb {
 
 	struct list_head	ki_list;	/* the aio core uses this
 						 * for cancellation */
+
+	/* to notify a process on I/O event */
+	struct aio_notify	ki_notify;
 };
 
 #define is_sync_kiocb(iocb)	((iocb)->ki_key == KIOCB_SYNC_KEY)
Index: linux-2.6.19-rc6-mm2/kernel/signal.c
===================================================================
--- linux-2.6.19-rc6-mm2.orig/kernel/signal.c	2006-11-30 13:33:10.000000000 +0100
+++ linux-2.6.19-rc6-mm2/kernel/signal.c	2006-11-30 15:06:05.000000000 +0100
@@ -297,7 +297,7 @@ static struct sigqueue *__sigqueue_alloc
 
 static void __sigqueue_free(struct sigqueue *q)
 {
-	if (q->flags & SIGQUEUE_PREALLOC)
+	if (q->flags & SIGQUEUE_PREALLOC && q->info.si_code != SI_ASYNCIO)
 		return;
 	atomic_dec(&q->user->sigpending);
 	free_uid(q->user);
Index: linux-2.6.19-rc6-mm2/fs/compat.c
===================================================================
--- linux-2.6.19-rc6-mm2.orig/fs/compat.c	2006-11-30 13:12:32.000000000 +0100
+++ linux-2.6.19-rc6-mm2/fs/compat.c	2006-11-30 15:06:05.000000000 +0100
@@ -663,6 +663,7 @@ compat_sys_io_submit(aio_context_t ctx_i
 		compat_uptr_t uptr;
 		struct iocb __user *user_iocb;
 		struct iocb tmp;
+		struct compat_sigevent __user *uevent;
 
 		if (unlikely(get_user(uptr, iocb + i))) {
 			ret = -EFAULT;
@@ -676,6 +677,23 @@ compat_sys_io_submit(aio_context_t ctx_i
 			break;
 		}
 
+		uevent = (struct compat_sigevent __user *)tmp.aio_sigeventp;
+
+		if (uevent) {
+			struct sigevent __user *event = NULL;
+			struct sigevent kevent;
+
+			event = compat_alloc_user_space(sizeof(*event));
+
+			if (get_compat_sigevent(&kevent, uevent) ||
+			    copy_to_user(event, &kevent, sizeof(*event))) {
+				ret = -EFAULT;
+				break;
+			}
+
+			tmp.aio_sigeventp = (__u64)event;
+		}
+
 		ret = io_submit_one(ctx, user_iocb, &tmp);
 		if (ret)
 			break;

Re: [PATCH -mm 4/5][AIO] - AIO completion signal notification

[Sébastien Dugué]

On Wed, 29 Nov 2006 13:50:12 +0000, Christoph Hellwig <hch@infradead.org> wrote:

> On Wed, Nov 29, 2006 at 02:08:01PM +0100, S?bastien Dugu? wrote:
> > On Wed, 29 Nov 2006 10:51:50 +0000, Christoph Hellwig <hch@infradead.org> wrote:
> > 
> > > I'm a little bit unhappy about the usage of the notify flag.  The usage
> > > seems correct but very confusing:
> > 
> >   Well, I followed the logic from posix-timers.c, but it may be a poor
> > choice ;-)
> > 
> >   For a start, the SIGEV_* flags are quite confusing (for me at least).
> > SIGEV_SIGNAL is defined as 0, SIGEV_NONE as 1 and SIGEV_THREAD_ID as 4. I
> > would rather have seen SIGEV_NONE defined as 0 to avoid all this.
> > 
> >   I also wish I knew why those SIGEV_* constants were defined that way.
> 
> Ah, I missed that.  It explains some of the more wierd bits.  I suspect
> we should then use != SIGEV_NONE for the any kind of signal notification
> bit and == SIGEV_THREAD_ID for the case where we want to deliver to
> a particular thread.

  Right, that would make things much cleaner. Will try for it.

> 
> But this means we only get a thread reference for SIGEV_THREAD_ID
> here:
> 
> > > > +	if (notify->notify == (SIGEV_SIGNAL|SIGEV_THREAD_ID)) {
> > > > +		/*
> > > > +		 * This reference will be dropped in really_put_req() when
> > > > +		 * we're done with the request.
> > > > +		 */
> > > > +		get_task_struct(target);
> > > > +	}

  It's the way it is in posix-timers and I'm not sure I understand why. We take
a ref on the specific task if notify is SIGEV_THREAD_ID but not for
SIGEV_SIGNAL.

  I'm wondering what I'm missing here, shouldn't we also take a ref on the task
group leader in the SIGEV_SIGNAL case in posix-timers? 

> 
> But even use it for SIGEV_SIGNAL without SIGEV_THREAD_ID here:
> 
> > > > +	if (notify->notify & SIGEV_THREAD_ID)
> > > > +		ret = send_sigqueue(notify->signo, sigq, notify->target);
> > > > +	else
> > > > +		ret = send_group_sigqueue(notify->signo, sigq, notify->target);
> 
> Or do I miss something?

  I missing something too here ;-)

  If someone cared to explain why there is no ref taken on the task for the
SIGEV_SIGNAL case, it would be much appreciated. Is this a bug in posix-timers?


  Thanks,

  Sébastien.

Re: [PATCH -mm 4/5][AIO] - AIO completion signal notification

[Christoph Hellwig]

On Wed, Nov 29, 2006 at 02:08:01PM +0100, S?bastien Dugu? wrote:
> On Wed, 29 Nov 2006 10:51:50 +0000, Christoph Hellwig <hch@infradead.org> wrote:
> 
> > I'm a little bit unhappy about the usage of the notify flag.  The usage
> > seems correct but very confusing:
> 
>   Well, I followed the logic from posix-timers.c, but it may be a poor
> choice ;-)
> 
>   For a start, the SIGEV_* flags are quite confusing (for me at least).
> SIGEV_SIGNAL is defined as 0, SIGEV_NONE as 1 and SIGEV_THREAD_ID as 4. I
> would rather have seen SIGEV_NONE defined as 0 to avoid all this.
> 
>   I also wish I knew why those SIGEV_* constants were defined that way.

Ah, I missed that.  It explains some of the more wierd bits.  I suspect
we should then use != SIGEV_NONE for the any kind of signal notification
bit and == SIGEV_THREAD_ID for the case where we want to deliver to
a particular thread.

But this means we only get a thread reference for SIGEV_THREAD_ID
here:

> > > +	if (notify->notify == (SIGEV_SIGNAL|SIGEV_THREAD_ID)) {
> > > +		/*
> > > +		 * This reference will be dropped in really_put_req() when
> > > +		 * we're done with the request.
> > > +		 */
> > > +		get_task_struct(target);
> > > +	}

But even use it for SIGEV_SIGNAL without SIGEV_THREAD_ID here:

> > > +	if (notify->notify & SIGEV_THREAD_ID)
> > > +		ret = send_sigqueue(notify->signo, sigq, notify->target);
> > > +	else
> > > +		ret = send_group_sigqueue(notify->signo, sigq, notify->target);

Or do I miss something?

Re: [PATCH -mm 4/5][AIO] - AIO completion signal notification

[Sébastien Dugué]

On Wed, 29 Nov 2006 06:33:35 -0500, Jakub Jelinek <jakub@redhat.com> wrote:

> On Wed, Nov 29, 2006 at 11:33:01AM +0100, S?bastien Dugu? wrote:
> >                       AIO completion signal notification
> > 
> >   The current 2.6 kernel does not support notification of user space via
> > an RT signal upon an asynchronous IO completion. The POSIX specification
> > states that when an AIO request completes, a signal can be delivered to
> > the application as notification.
> > 
> >   This patch adds a struct sigevent *aio_sigeventp to the iocb.
> > The relevant fields (pid, signal number and value) are stored in the kiocb
> > for use when the request completes.
> > 
> >   That sigevent structure is filled by the application as part of the AIO
> > request preparation. Upon request completion, the kernel notifies the
> > application using those sigevent parameters. If SIGEV_NONE has been specified,
> > then the old behaviour is retained and the application must rely on polling
> > the completion queue using io_getevents().
> 
> Well, from what I see applications must rely on polling the completion
> queue using io_getevents() in any case, isn't that the only way how to free
> the kernel resources associated with the AIO request, even if it uses
> SIGEV_SIGNAL or thread notification?

  Well, applications do not need to do any polling on the queue anymore.
io_getevents() needs to be called only once when the signal has been received,
either from a signal handler or from a thread blocking on the signal.

> aio_error/aio_return/aio_suspend
> will still need to io_getevents,

  Right, but only once.

> the only important difference with this
> patch is that a) the polling doesn't need to be asynchronous (i.e. have
> a special thread which just loops doing io_getevents)

  Yes, no more polling loop and I think this makes a big difference.

> b) it doesn't need to care about notification itself.
> 

  Uh! what do you mean here?

  Sébastien.

Re: [PATCH -mm 4/5][AIO] - AIO completion signal notification

[Sébastien Dugué]

On Wed, 29 Nov 2006 10:51:50 +0000, Christoph Hellwig <hch@infradead.org> wrote:

> I'm a little bit unhappy about the usage of the notify flag.  The usage
> seems correct but very confusing:

  Well, I followed the logic from posix-timers.c, but it may be a poor
choice ;-)

  For a start, the SIGEV_* flags are quite confusing (for me at least).
SIGEV_SIGNAL is defined as 0, SIGEV_NONE as 1 and SIGEV_THREAD_ID as 4. I
would rather have seen SIGEV_NONE defined as 0 to avoid all this.

  I also wish I knew why those SIGEV_* constants were defined that way.

> 
> In io_submit_one we set ki_notify.notify to SIGEV_NONE and possibly
> call aio_setup_sigevent:
> 
> > +	/* handle setting up the sigevent for POSIX AIO signals */
> > +	req->ki_notify.notify = SIGEV_NONE;
> > +
> > +	if (iocb->aio_sigeventp) {
> > +		ret = aio_setup_sigevent(&req->ki_notify,
> > +					 (struct sigevent __user *)(unsigned
> > long)
> > +					 iocb->aio_sigeventp);
> > +		if (ret)
> > +			goto out_put_req;
> > +	}
> > +
> 
> aio_setup_sigevent then checks the user passed even for which notify type
> we have, and returns if it's none or otherwise sets notify->notify to it.
> 
> > +	if (event.sigev_notify == SIGEV_NONE)
> > +		return 0;
> > +
> > +	notify->notify = event.sigev_notify;
> 
> Later aio_setup_sigevent gets a reference to the target task_structure
> if notify->notify is (SIGEV_SIGNAL|SIGEV_THREAD_ID) but _always_ stores
> the target pointer.

  Yep, as SIGEV_SIGNAL is 0, this in fact checks that notify is SIGEV_THREAD_ID.
It could have been written as:

	if (notify->notify == SIGEV_THREAD_ID)

  And the target pointer is always store because at this point notify is either
SIGEV_SIGNAL or SIGEV_THREAD_ID.

> 
> > +	if (notify->notify == (SIGEV_SIGNAL|SIGEV_THREAD_ID)) {
> > +		/*
> > +		 * This reference will be dropped in really_put_req() when
> > +		 * we're done with the request.
> > +		 */
> > +		get_task_struct(target);
> > +	}
> > +
> > +	notify->target = target;
> 
> 
> Once we're done with the iocb aio_complete aclls aio_send_signal if
> notify.notify is not SIGEV_NONE.

  Again, if it's not SIGEV_NONE, then it's either SIGEV_SIGNAL or
SIGEV_THREAD_ID.

> 
> > +	if (iocb->ki_notify.notify != SIGEV_NONE) {
> > +		ret = aio_send_signal(&iocb->ki_notify);
> > +
> > +		/* If signal generation failed, release the sigqueue */
> > +		if (ret)
> > +			sigqueue_free(iocb->ki_notify.sigq);
> > +	}
> > +
> 
> Which then uses notify->target to send the signal:
> > +	if (notify->notify & SIGEV_THREAD_ID)
> > +		ret = send_sigqueue(notify->signo, sigq, notify->target);
> > +	else
> > +		ret = send_group_sigqueue(notify->signo, sigq, notify->target);
> 
> And finally really_put_req puts the target if notify.notify contains
> either SIGEV_SIGNAL or SIGEV_THREAD_ID.
> 
> > +	/* Release task ref */
> > +	if (req->ki_notify.notify == (SIGEV_SIGNAL|SIGEV_THREAD_ID))
> > +		put_task_struct(req->ki_notify.target);

  Could have been if (req->ki_notify.notify == SIGEV_THREAD_ID)

> 
> Do you see the confusing?  I think all the notify.notify != SIGEV_NONE
> in the above code should be replaces by the much more descriptive
> notify.notify == (SIGEV_SIGNAL|SIGEV_THREAD_ID). In addition we should
> only store the target pointer inside the (SIGEV_SIGNAL|SIGEV_THREAD_ID)
> if block that gets a reference to it.

  No, I don't think so, notify == (SIGEV_SIGNAL|SIGEV_THREAD_ID) really means
notify == SIGEV_THREAD_ID which leaves out notify == SIGEV_SIGNAL. Or
am I grossly mistaken?

  Thanks,

  Sébastien.



> -
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/

Re: [PATCH -mm 4/5][AIO] - AIO completion signal notification

[Jakub Jelinek]

On Wed, Nov 29, 2006 at 11:33:01AM +0100, S?bastien Dugu? wrote:
>                       AIO completion signal notification
> 
>   The current 2.6 kernel does not support notification of user space via
> an RT signal upon an asynchronous IO completion. The POSIX specification
> states that when an AIO request completes, a signal can be delivered to
> the application as notification.
> 
>   This patch adds a struct sigevent *aio_sigeventp to the iocb.
> The relevant fields (pid, signal number and value) are stored in the kiocb
> for use when the request completes.
> 
>   That sigevent structure is filled by the application as part of the AIO
> request preparation. Upon request completion, the kernel notifies the
> application using those sigevent parameters. If SIGEV_NONE has been specified,
> then the old behaviour is retained and the application must rely on polling
> the completion queue using io_getevents().

Well, from what I see applications must rely on polling the completion
queue using io_getevents() in any case, isn't that the only way how to free
the kernel resources associated with the AIO request, even if it uses
SIGEV_SIGNAL or thread notification?  aio_error/aio_return/aio_suspend
will still need to io_getevents, the only important difference with this
patch is that a) the polling doesn't need to be asynchronous (i.e. have
a special thread which just loops doing io_getevents)
b) it doesn't need to care about notification itself.

	Jakub

Re: [PATCH -mm 4/5][AIO] - AIO completion signal notification

[Christoph Hellwig]

I'm a little bit unhappy about the usage of the notify flag.  The usage
seems correct but very confusing:

In io_submit_one we set ki_notify.notify to SIGEV_NONE and possibly
call aio_setup_sigevent:

> +	/* handle setting up the sigevent for POSIX AIO signals */
> +	req->ki_notify.notify = SIGEV_NONE;
> +
> +	if (iocb->aio_sigeventp) {
> +		ret = aio_setup_sigevent(&req->ki_notify,
> +					 (struct sigevent __user *)(unsigned
> long)
> +					 iocb->aio_sigeventp);
> +		if (ret)
> +			goto out_put_req;
> +	}
> +

aio_setup_sigevent then checks the user passed even for which notify type
we have, and returns if it's none or otherwise sets notify->notify to it.

> +	if (event.sigev_notify == SIGEV_NONE)
> +		return 0;
> +
> +	notify->notify = event.sigev_notify;

Later aio_setup_sigevent gets a reference to the target task_structure
if notify->notify is (SIGEV_SIGNAL|SIGEV_THREAD_ID) but _always_ stores
the target pointer.

> +	if (notify->notify == (SIGEV_SIGNAL|SIGEV_THREAD_ID)) {
> +		/*
> +		 * This reference will be dropped in really_put_req() when
> +		 * we're done with the request.
> +		 */
> +		get_task_struct(target);
> +	}
> +
> +	notify->target = target;


Once we're done with the iocb aio_complete aclls aio_send_signal if
notify.notify is not SIGEV_NONE.

> +	if (iocb->ki_notify.notify != SIGEV_NONE) {
> +		ret = aio_send_signal(&iocb->ki_notify);
> +
> +		/* If signal generation failed, release the sigqueue */
> +		if (ret)
> +			sigqueue_free(iocb->ki_notify.sigq);
> +	}
> +

Which then uses notify->target to send the signal:
> +	if (notify->notify & SIGEV_THREAD_ID)
> +		ret = send_sigqueue(notify->signo, sigq, notify->target);
> +	else
> +		ret = send_group_sigqueue(notify->signo, sigq, notify->target);

And finally really_put_req puts the target if notify.notify contains
either SIGEV_SIGNAL or SIGEV_THREAD_ID.

> +	/* Release task ref */
> +	if (req->ki_notify.notify == (SIGEV_SIGNAL|SIGEV_THREAD_ID))
> +		put_task_struct(req->ki_notify.target);

Do you see the confusing?  I think all the notify.notify != SIGEV_NONE
in the above code should be replaces by the much more descriptive
notify.notify == (SIGEV_SIGNAL|SIGEV_THREAD_ID). In addition we should
only store the target pointer inside the (SIGEV_SIGNAL|SIGEV_THREAD_ID)
if block that gets a reference to it.

[PATCH -mm 4/5][AIO] - AIO completion signal notification

[Sébastien Dugué]

                      AIO completion signal notification

  The current 2.6 kernel does not support notification of user space via
an RT signal upon an asynchronous IO completion. The POSIX specification
states that when an AIO request completes, a signal can be delivered to
the application as notification.

  This patch adds a struct sigevent *aio_sigeventp to the iocb.
The relevant fields (pid, signal number and value) are stored in the kiocb
for use when the request completes.

  That sigevent structure is filled by the application as part of the AIO
request preparation. Upon request completion, the kernel notifies the
application using those sigevent parameters. If SIGEV_NONE has been specified,
then the old behaviour is retained and the application must rely on polling
the completion queue using io_getevents().


  A struct sigevent *aio_sigeventp field is added to struct iocb in
include/linux/aio_abi.h

  A struct aio_notify containing the sigevent parameters is defined in aio.h:

  struct aio_notify {
	struct task_struct	*target;
	__u16			signo;
	__u16			notify;
	sigval_t		value;
  };

  A struct aio_notify ki_notify is added to struct kiocb in include/linux/aio.h

  In io_submit_one(), if the application provided a sigevent then
setup_sigevent() is called which does the following:

	- check access to the user sigevent and make a local copy

	- if the requested notification is SIGEV_NONE, then nothing to do

	- fill in the kiocb->ki_notify fields (notify, signo, value)

	- check sigevent consistency, get the signal target task and
	  save it in kiocb->ki_notify.target

	- preallocate a sigqueue for this event using sigqueue_alloc()

  Upon request completion, in aio_complete(), if notification is needed for
this request (iocb->ki_notify.notify != SIGEV_NONE), then aio_send_signal()
is called to signal the target task as follows:

	- fill in the siginfo struct to be sent to the task

	- if notify is SIGEV_THREAD_ID then send signal to specific task
	  using send_sigqueue()

	- else send signal to task group using send_5group_sigqueue()



Notes concerning sigqueue preallocation:

 To ensure reliable delivery of completion notification, the sigqueue is
preallocated in the submission path so that there is no chance it can fail
in the completion path.

 Unlike the posix-timers case (currently the single other user of sigqueue
preallocation), where the sigqueue is allocated for the lifetime of the timer
and freed at timer destruction time, the aio case is a bit more tricky due to
the async nature of the whole thing.

  In the aio case, the sigqueue exists for the lifetime of the request,
therefore it must be freed only once the signal for the request completion has
been delivered. This involves changing __sigqueue_free() to free the sigqueue
when the signal is collected if si_code is SI_ASYNCIO even if it was
preallocated as well as explicitly calling sigqueue_free() in submission and
completion error paths.


 fs/aio.c                |  115 ++++++++++++++++++++++++++++++++++++++++++++++--
 fs/compat.c             |   18 +++++++
 include/linux/aio.h     |   12 +++++
 include/linux/aio_abi.h |    3 -
 kernel/signal.c         |    2
 5 files changed, 144 insertions(+), 6 deletions(-)

Signed-off-by: Sébastien Dugué <sebastien.dugue@bull.net>
Signed-off-by: Laurent Vivier <laurent.vivier@bull.net>

Index: linux-2.6.19-rc6-mm2/fs/aio.c
===================================================================
--- linux-2.6.19-rc6-mm2.orig/fs/aio.c	2006-11-28 12:49:40.000000000
+0100 +++ linux-2.6.19-rc6-mm2/fs/aio.c	2006-11-29 10:14:47.000000000
+0100 @@ -416,6 +416,7 @@ static struct kiocb fastcall *__aio_get_
 	req->ki_dtor = NULL;
 	req->private = NULL;
 	req->ki_iovec = NULL;
+	req->ki_notify.sigq = NULL;
 	INIT_LIST_HEAD(&req->ki_run_list);
 
 	/* Check if the completion queue has enough free space to
@@ -463,6 +464,11 @@ static inline void really_put_req(struct
 		req->ki_dtor(req);
 	if (req->ki_iovec != &req->ki_inline_vec)
 		kfree(req->ki_iovec);
+
+	/* Release task ref */
+	if (req->ki_notify.notify == (SIGEV_SIGNAL|SIGEV_THREAD_ID))
+		put_task_struct(req->ki_notify.target);
+
 	kmem_cache_free(kiocb_cachep, req);
 	ctx->reqs_active--;
 
@@ -929,6 +935,80 @@ void fastcall kick_iocb(struct kiocb *io
 }
 EXPORT_SYMBOL(kick_iocb);
 
+static int aio_send_signal(struct aio_notify *notify)
+{
+	struct sigqueue *sigq = notify->sigq;
+	struct siginfo *info = &sigq->info;
+	int ret;
+
+	memset(info, 0, sizeof(struct siginfo));
+
+	info->si_signo = notify->signo;
+	info->si_errno = 0;
+	info->si_code = SI_ASYNCIO;
+	info->si_pid = 0;
+	info->si_uid = 0;
+	info->si_value = notify->value;
+
+	if (notify->notify & SIGEV_THREAD_ID)
+		ret = send_sigqueue(notify->signo, sigq, notify->target);
+	else
+		ret = send_group_sigqueue(notify->signo, sigq, notify->target);
+
+	return ret;
+}
+
+static long aio_setup_sigevent(struct aio_notify *notify,
+			       struct sigevent __user *user_event)
+{
+	sigevent_t event;
+	struct task_struct *target;
+
+	if (copy_from_user(&event, user_event, sizeof (event)))
+		return -EFAULT;
+
+	if (event.sigev_notify == SIGEV_NONE)
+		return 0;
+
+	notify->notify = event.sigev_notify;
+	notify->signo = event.sigev_signo;
+	notify->value = event.sigev_value;
+
+	read_lock(&tasklist_lock);
+	target = good_sigevent(&event);
+
+	if (unlikely(!target || (target->flags & PF_EXITING)))
+		goto out_unlock;
+
+	if (notify->notify == (SIGEV_SIGNAL|SIGEV_THREAD_ID)) {
+		/*
+		 * This reference will be dropped in really_put_req() when
+		 * we're done with the request.
+		 */
+		get_task_struct(target);
+	}
+
+	notify->target = target;
+	read_unlock(&tasklist_lock);
+
+	/*
+	 * NOTE: we cannot free the sigqueue in the completion path as
+	 * the signal may not have been delivered to the target task.
+	 * Therefore it has to be freed in __sigqueue_free() when the
+	 * signal is collected if si_code is SI_ASYNCIO.
+	 */
+	notify->sigq = sigqueue_alloc();
+
+	if (unlikely(!notify->sigq))
+		return -EAGAIN;
+
+	return 0;
+
+out_unlock:
+	read_unlock(&tasklist_lock);
+	return -EINVAL;
+}
+
 /* aio_complete
  *	Called when the io request on the given iocb is complete.
  *	Returns true if this is the last user of the request.  The 
@@ -976,8 +1056,11 @@ int fastcall aio_complete(struct kiocb *
 	 * cancelled requests don't get events, userland was given one
 	 * when the event got cancelled.
 	 */
-	if (kiocbIsCancelled(iocb))
+	if (kiocbIsCancelled(iocb)) {
+		if (iocb->ki_notify.sigq)
+			sigqueue_free(iocb->ki_notify.sigq);
 		goto put_rq;
+	}
 
 	ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
 
@@ -1008,6 +1091,14 @@ int fastcall aio_complete(struct kiocb *
 
 	pr_debug("added to ring %p at [%lu]\n", iocb, tail);
 
+	if (iocb->ki_notify.notify != SIGEV_NONE) {
+		ret = aio_send_signal(&iocb->ki_notify);
+
+		/* If signal generation failed, release the sigqueue */
+		if (ret)
+			sigqueue_free(iocb->ki_notify.sigq);
+	}
+
 	pr_debug("%ld retries: %zd of %zd\n", iocb->ki_retried,
 		iocb->ki_nbytes - iocb->ki_left, iocb->ki_nbytes);
 put_rq:
@@ -1549,8 +1640,7 @@ int fastcall io_submit_one(struct kioctx
 	ssize_t ret;
 
 	/* enforce forwards compatibility on users */
-	if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2 ||
-		     iocb->aio_reserved3)) {
+	if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved3)) {
 		pr_debug("EINVAL: io_submit: reserve field set\n");
 		return -EINVAL;
 	}
@@ -1559,6 +1649,7 @@ int fastcall io_submit_one(struct kioctx
 	if (unlikely(
 	    (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
 	    (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
+	    (iocb->aio_sigeventp != (unsigned long)iocb->aio_sigeventp) ||
 	    ((ssize_t)iocb->aio_nbytes < 0)
 	   )) {
 		pr_debug("EINVAL: io_submit: overflow check\n");
@@ -1593,10 +1684,21 @@ int fastcall io_submit_one(struct kioctx
 	INIT_LIST_HEAD(&req->ki_wait.task_list);
 	req->ki_retried = 0;
 
+	/* handle setting up the sigevent for POSIX AIO signals */
+	req->ki_notify.notify = SIGEV_NONE;
+
+	if (iocb->aio_sigeventp) {
+		ret = aio_setup_sigevent(&req->ki_notify,
+					 (struct sigevent __user *)(unsigned
long)
+					 iocb->aio_sigeventp);
+		if (ret)
+			goto out_put_req;
+	}
+
 	ret = aio_setup_iocb(req);
 
 	if (ret)
-		goto out_put_req;
+		goto out_sigqfree;
 
 	spin_lock_irq(&ctx->ctx_lock);
 	aio_run_iocb(req);
@@ -1609,6 +1711,11 @@ int fastcall io_submit_one(struct kioctx
 	aio_put_req(req);	/* drop extra ref to req */
 	return 0;
 
+out_sigqfree:
+	/* Undo the sigqueue alloc if someting went bad */
+	if (req->ki_notify.sigq)
+		sigqueue_free(req->ki_notify.sigq);
+
 out_put_req:
 	aio_put_req(req);	/* drop extra ref to req */
 	aio_put_req(req);	/* drop i/o ref to req */
Index: linux-2.6.19-rc6-mm2/include/linux/aio_abi.h
===================================================================
--- linux-2.6.19-rc6-mm2.orig/include/linux/aio_abi.h	2006-11-16
05:03:40.000000000 +0100 +++
linux-2.6.19-rc6-mm2/include/linux/aio_abi.h	2006-11-29
10:14:29.000000000 +0100 @@ -82,8 +82,9 @@ struct iocb { __u64
aio_nbytes; __s64	aio_offset;
 
+	__u64	aio_sigeventp;	/* pointer to struct sigevent */
+
 	/* extra parameters */
-	__u64	aio_reserved2;	/* TODO: use this for a (struct
sigevent *) */ __u64	aio_reserved3;
 }; /* 64 bytes */
 
Index: linux-2.6.19-rc6-mm2/include/linux/aio.h
===================================================================
--- linux-2.6.19-rc6-mm2.orig/include/linux/aio.h	2006-11-28
12:51:41.000000000 +0100 +++ linux-2.6.19-rc6-mm2/include/linux/aio.h
2006-11-29 10:14:29.000000000 +0100 @@ -7,6 +7,7 @@
 #include <linux/uio.h>
 
 #include <asm/atomic.h>
+#include <asm/siginfo.h>
 
 #define AIO_MAXSEGS		4
 #define AIO_KIOGRP_NR_ATOMIC	8
@@ -49,6 +50,14 @@ struct kioctx;
 #define kiocbIsKicked(iocb)	test_bit(KIF_KICKED, &(iocb)->ki_flags)
 #define kiocbIsCancelled(iocb)	test_bit(KIF_CANCELLED,
&(iocb)->ki_flags) 
+struct aio_notify {
+	struct task_struct	*target;
+	__u16			signo;
+	__u16			notify;
+	sigval_t		value;
+	struct sigqueue		*sigq;
+};
+
 /* is there a better place to document function pointer methods? */
 /**
  * ki_retry	-	iocb forward progress callback
@@ -118,6 +127,9 @@ struct kiocb {
 
 	struct list_head	ki_list;	/* the aio core uses this
 						 * for cancellation */
+
+	/* to notify a process on I/O event */
+	struct aio_notify	ki_notify;
 };
 
 #define is_sync_kiocb(iocb)	((iocb)->ki_key == KIOCB_SYNC_KEY)
Index: linux-2.6.19-rc6-mm2/kernel/signal.c
===================================================================
--- linux-2.6.19-rc6-mm2.orig/kernel/signal.c	2006-11-28
12:51:43.000000000 +0100 +++ linux-2.6.19-rc6-mm2/kernel/signal.c
2006-11-28 12:51:45.000000000 +0100 @@ -297,7 +297,7 @@ static struct sigqueue
*__sigqueue_alloc 
 static void __sigqueue_free(struct sigqueue *q)
 {
-	if (q->flags & SIGQUEUE_PREALLOC)
+	if (q->flags & SIGQUEUE_PREALLOC && q->info.si_code != SI_ASYNCIO)
 		return;
 	atomic_dec(&q->user->sigpending);
 	free_uid(q->user);
Index: linux-2.6.19-rc6-mm2/fs/compat.c
===================================================================
--- linux-2.6.19-rc6-mm2.orig/fs/compat.c	2006-11-28 12:51:35.000000000
+0100 +++ linux-2.6.19-rc6-mm2/fs/compat.c	2006-11-29 10:14:29.000000000
+0100 @@ -663,6 +663,7 @@ compat_sys_io_submit(aio_context_t ctx_i
 		compat_uptr_t uptr;
 		struct iocb __user *user_iocb;
 		struct iocb tmp;
+		struct compat_sigevent __user *uevent;
 
 		if (get_user(uptr, iocb + i)) {
 			ret = -EFAULT;
@@ -676,6 +677,23 @@ compat_sys_io_submit(aio_context_t ctx_i
 			break;
 		}
 
+		uevent = (struct compat_sigevent __user *)tmp.aio_sigeventp;
+
+		if (uevent) {
+			struct sigevent __user *event = NULL;
+			struct sigevent kevent;
+
+			event = compat_alloc_user_space(sizeof(*event));
+
+			if (get_compat_sigevent(&kevent, uevent) ||
+			    copy_to_user(event, &kevent, sizeof(*event))) {
+				ret = -EFAULT;
+				break;
+			}
+
+			tmp.aio_sigeventp = (__u64)event;
+		}
+
 		ret = io_submit_one(ctx, user_iocb, &tmp);
 
 		if (ret)