Skip to contentSkip to conversation

aio: fix oops because of extra IO control block freeing.

PATCH Alan Stern 2 participants 3 messages Mar 5, 2007 1 message not in the archive

lore
Share & tools
lore

Alan Stern

author
#1 lore raw
In reply to a message not in the archive
Quoted 81 lines · Leonid Ananiev

From Leonid Ananiev

This patch finishes moving from using errno EIOCBRETRY to using flag in IO control block for aio retrying. After this change the process will be kicked for direct aio as it was for sync aio.

The patch is applied to 2.6.20 or 2.6.21-rc2 diff -uprN -X linux-2.6.20-aio21/Documentation/dontdiff linux-2.6.20-aio21/drivers/usb/gadget/inode.c linux-2.6.20-aio22/drivers/usb/gadget/inode.c
--- linux-2.6.20-aio21/drivers/usb/gadget/inode.c 2007-03-04
21:45:52.000000000 +0300
+++ linux-2.6.20-aio22/drivers/usb/gadget/inode.c 2007-03-05
18:19:35.000000000 +0300

(quoted diff) +4 -1
@@ -692,7 +692,10 @@ fail:
  		kfree(priv);
  		put_ep(epdata);
  	} else
-		value = (iv ? -EIOCBRETRY : -EIOCBQUEUED);
+		if (iv)
+			kiocbSetPgBusy(iocb);
+		else
+			value = -EIOCBQUEUED;
  	return value;
  }
 

diff -uprN -X linux-2.6.20-aio21/Documentation/dontdiff linux-2.6.20-aio21/fs/ocfs2/dlmglue.c linux-2.6.20-aio22/fs/ocfs2/dlmglue.c
--- linux-2.6.20-aio21/fs/ocfs2/dlmglue.c 2007-03-04 21:45:52.000000000
+0300
+++ linux-2.6.20-aio22/fs/ocfs2/dlmglue.c 2007-03-04 22:57:50.000000000
+0300

(quoted diff) +1 -1
@@ -1639,7 +1639,7 @@ int ocfs2_meta_lock_full(struct inode *i
 
  	status = ocfs2_cluster_lock(osb, lockres, level, dlm_flags, arg_flags);
  	if (status < 0) {
-		if (status != -EAGAIN && status != -EIOCBRETRY)
+		if (status != -EAGAIN)
  			mlog_errno(status);
  		goto bail;
  	}

diff -uprN -X linux-2.6.20-aio21/Documentation/dontdiff linux-2.6.20-aio21/include/linux/aio.h linux-2.6.20-aio22/include/linux/aio.h
--- linux-2.6.20-aio21/include/linux/aio.h 2007-03-04 21:46:45.000000000
+0300
+++ linux-2.6.20-aio22/include/linux/aio.h 2007-03-04 22:57:50.000000000
+0300

(quoted diff) +0 -0
@@ -79,15 +79,6 @@ struct kioctx;
   * not ask the method again -- ki_retry must ensure forward progress.
   * aio_complete() must be called once and only once in the future, 

multiple
* calls may result in undefined behaviour.
- *
- * If ki_retry returns -EIOCBRETRY it has made a promise that kick_iocb()
- * will be called on the kiocb pointer in the future. This may happen
- * through generic helpers that associate kiocb->ki_wait with a wait
- * queue head that ki_retry uses via current->io_wait. It can also happen
- * with custom tracking and manual calls to kick_iocb(), though that is
- * discouraged. In either case, kick_iocb() must be called once and only
- * once. ki_retry must ensure forward progress, the AIO core will wait
- * indefinitely for kick_iocb() to be called.
*/
struct kiocb {
struct list_head ki_run_list;
diff -uprN -X linux-2.6.20-aio21/Documentation/dontdiff linux-2.6.20-aio21/include/linux/errno.h linux-2.6.20-aio22/include/linux/errno.h
--- linux-2.6.20-aio21/include/linux/errno.h 2007-03-04
21:45:51.000000000 +0300
+++ linux-2.6.20-aio22/include/linux/errno.h 2007-03-04
22:57:50.000000000 +0300

(quoted diff) +0 -0
@@ -22,7 +22,6 @@
  #define EBADTYPE	527	/* Type not supported by server */
  #define EJUKEBOX	528	/* Request initiated, but will not complete 

before timeout */
#define EIOCBQUEUED 529 /* iocb queued, will get completion event */
-#define EIOCBRETRY 530 /* iocb queued, will trigger a retry */

Quoted 1 line
  #endif

Where is kiocbSetPgBusy() defined, and where in the documentation is it explained? For that matter, what is the reason for changing the return value at all?

And if you are going to change, why not make the return value something more logical? For instance, return 0 to indicate the iocb is complete and -EINPROGRESS to indicate another kick will occur.

Alan Stern

#2 lore raw
Quoted 3 lines

Where is kiocbSetPgBusy() defined,

It is defined in include/linux/aio.h (introduced in [PATCH 1/3])

Quoted 1 line

and where in the documentation is it explained?

I will add a comment for kiocbSetPgBusy()in aio.h.

Quoted 1 line

For that matter, what is the reason for changing the return value at

all?
EIORETRY is not IO return value at all. It is processing stage. It conflicts with IO return value. That is why a flag in iocb is used. It should be noted that loop but not EIOCBRETRY was used for vector IO (mm/filemap.c)

Leonid

Quoted 6 lines

-----Original Message-----
From: Alan Stern [mailto:[email protected]]
Sent: Monday, March 05, 2007 10:36 PM
To: Ananiev, Leonid I
Cc: Kernel development list
Subject: Re: [PATCH 2/3] aio: fix oops because of extra IO control

block freeing.

Quoted 6 lines · Leonid Ananiev
Quoted 4 lines

From Leonid Ananiev

This patch finishes moving from using errno EIOCBRETRY to using flag

in

Quoted 1 line
Quoted 1 line

IO control block for aio retrying. After this change the process will

be

Quoted 24 lines
Quoted 24 lines

kicked for direct aio as it was for sync aio.

The patch is applied to 2.6.20 or 2.6.21-rc2 diff -uprN -X linux-2.6.20-aio21/Documentation/dontdiff linux-2.6.20-aio21/drivers/usb/gadget/inode.c linux-2.6.20-aio22/drivers/usb/gadget/inode.c
--- linux-2.6.20-aio21/drivers/usb/gadget/inode.c 2007-03-04
21:45:52.000000000 +0300
+++ linux-2.6.20-aio22/drivers/usb/gadget/inode.c 2007-03-05
18:19:35.000000000 +0300

(quoted diff) +4 -1
@@ -692,7 +692,10 @@ fail:
  		kfree(priv);
  		put_ep(epdata);
  	} else
-		value = (iv ? -EIOCBRETRY : -EIOCBQUEUED);
+		if (iv)
+			kiocbSetPgBusy(iocb);
+		else
+			value = -EIOCBQUEUED;
  	return value;
  }
 
Quoted 2 lines
Quoted 2 lines

diff -uprN -X linux-2.6.20-aio21/Documentation/dontdiff linux-2.6.20-aio21/fs/ocfs2/dlmglue.c

linux-2.6.20-aio22/fs/ocfs2/dlmglue.c

Quoted 1 line
Quoted 1 line

--- linux-2.6.20-aio21/fs/ocfs2/dlmglue.c 2007-03-04

21:45:52.000000000

Quoted 2 lines
Quoted 2 lines

+0300
+++ linux-2.6.20-aio22/fs/ocfs2/dlmglue.c 2007-03-04

22:57:50.000000000

Quoted 4 lines

Review int ocfs2_meta_lock_full(struct inode *i · line 1640

Quoted 4 lines

+0300

(quoted diff) +0 -0
@@ -1639,7 +1639,7 @@ int ocfs2_meta_lock_full(struct inode *i
 
  	status = ocfs2_cluster_lock(osb, lockres, level, dlm_flags,

arg_flags);

Quoted 10 lines
Quoted 10 lines
(quoted diff) +1 -1
  	if (status < 0) {
-		if (status != -EAGAIN && status != -EIOCBRETRY)
+		if (status != -EAGAIN)
  			mlog_errno(status);
  		goto bail;
  	}

diff -uprN -X linux-2.6.20-aio21/Documentation/dontdiff linux-2.6.20-aio21/include/linux/aio.h linux-2.6.20-aio22/include/linux/aio.h
--- linux-2.6.20-aio21/include/linux/aio.h 2007-03-04

21:46:45.000000000

Quoted 2 lines
Quoted 2 lines

+0300
+++ linux-2.6.20-aio22/include/linux/aio.h 2007-03-04

22:57:50.000000000

Quoted 3 lines

Review struct kioctx; · line 79

Quoted 3 lines

+0300

(quoted diff) +0 -0
@@ -79,15 +79,6 @@ struct kioctx;
   * not ask the method again -- ki_retry must ensure forward

progress.

Quoted 5 lines
Quoted 5 lines
(quoted diff) +0 -0
   * aio_complete() must be called once and only once in the future,

multiple
* calls may result in undefined behaviour.
- *
- * If ki_retry returns -EIOCBRETRY it has made a promise that

kick_iocb()

Quoted 1 line
Quoted 1 line

- * will be called on the kiocb pointer in the future. This may

happen

Quoted 2 lines
Quoted 2 lines

- * through generic helpers that associate kiocb->ki_wait with a wait
- * queue head that ki_retry uses via current->io_wait. It can also

happen

Quoted 1 line
Quoted 1 line

- * with custom tracking and manual calls to kick_iocb(), though that

is

Quoted 1 line
Quoted 1 line

- * discouraged. In either case, kick_iocb() must be called once and

only

Quoted 1 line
Quoted 1 line

- * once. ki_retry must ensure forward progress, the AIO core will

wait

Quoted 6 lines
Quoted 6 lines

- * indefinitely for kick_iocb() to be called. */
struct kiocb {
struct list_head ki_run_list;
diff -uprN -X linux-2.6.20-aio21/Documentation/dontdiff linux-2.6.20-aio21/include/linux/errno.h

Quoted 8 lines

Review line 23

Quoted 8 lines

linux-2.6.20-aio22/include/linux/errno.h
--- linux-2.6.20-aio21/include/linux/errno.h 2007-03-04
21:45:51.000000000 +0300
+++ linux-2.6.20-aio22/include/linux/errno.h 2007-03-04
22:57:50.000000000 +0300

(quoted diff) +0 -0
@@ -22,7 +22,6 @@
  #define EBADTYPE	527	/* Type not supported by server */
  #define EJUKEBOX	528	/* Request initiated, but will not

complete

Quoted 2 lines
Quoted 2 lines

before timeout */
#define EIOCBQUEUED 529 /* iocb queued, will get

completion event */

Quoted 10 lines
Quoted 3 lines

-#define EIOCBRETRY 530 /* iocb queued, will trigger a retry */

  #endif

Where is kiocbSetPgBusy() defined, and where in the documentation is it explained? For that matter, what is the reason for changing the return value at all?

And if you are going to change, why not make the return value something more logical? For instance, return 0 to indicate the iocb is complete

and

Quoted 3 lines

-EINPROGRESS to indicate another kick will occur.

Alan Stern

author
#3 lore raw
Quoted 5 lines · Leonid I
Quoted 3 lines

Where is kiocbSetPgBusy() defined,

It is defined in include/linux/aio.h (introduced in [PATCH 1/3])

Quoted 1 line

and where in the documentation is it explained?

Quoted 8 lines

I will add a comment for kiocbSetPgBusy()in aio.h.

Quoted 1 line

For that matter, what is the reason for changing the return value at

all?
EIORETRY is not IO return value at all. It is processing stage. It conflicts with IO return value. That is why a flag in iocb is used. It should be noted that loop but not EIOCBRETRY was used for vector IO (mm/filemap.c)

Okay. But I don't like the name. "kiocbSetPgBusy" might be meaningful to people who understand fs/aio.c, but it isn't meaningful to other people.

Could you add this to aio.h:

#define mark_iocb_not_complete(iocb) kiocbSetPgBusy(iocb)

That would be a lot easier to understand.

Alan Stern