2002-10-10 17:18:58

by Sylvain Pasche

Subject: 2.5.41 isofs patch to avoid "bad: scheduling while atomic!"

Without the patch:
bad: scheduling while atomic!
Call Trace:
[<c01161b1>] schedule+0x3d/0x4d4
[<c01123fd>] smp_apic_timer_interrupt+0x111/0x124
[<c0107595>] need_resched+0x1f/0x2a
[<c01a0068>] sysvipc_sem_read_proc+0x1b0/0x227
[<c01acba5>] serial_in+0x45/0x4c
[<c01aeadd>] serial8250_console_write+0x5d/0x1cc
[<c011ba26>] __call_console_drivers+0x3e/0x50
[<c011ba8b>] _call_console_drivers+0x53/0x58
[<c011bb45>] call_console_drivers+0xb5/0xe0
[<c011bed7>] release_console_sem+0xc3/0x164
[<c011bd7f>] printk+0x1a7/0x1f4
[<c011890f>] __might_sleep+0x4f/0x58
[<c013f2f4>] __alloc_pages+0x24/0x24c
[<c013f544>] __get_free_pages+0x28/0x60
[<f8998797>] isofs_readdir+0x6f/0xf7 [isofs]
[<c015dfd5>] vfs_readdir+0x75/0x88
[<c015e260>] filldir64+0x0/0x114
[<c015e3c3>] sys_getdents64+0x4f/0xb3
[<c015e260>] filldir64+0x0/0x114
[<c015d52d>] sys_fcntl64+0x85/0x98
[<c015d539>] sys_fcntl64+0x91/0x98
[<c01075d3>] syscall_call+0x7/0xb

the simple fix:

--- linux-2.5.41/fs/isofs/dir.c_old 2002-10-10 19:12:19.000000000 +0200
+++ linux-2.5.41/fs/isofs/dir.c 2002-10-10 19:13:26.000000000 +0200
@@ -256,7 +256,7 @@

lock_kernel();

- tmpname = (char *) __get_free_page(GFP_KERNEL);
+ tmpname = (char *) __get_free_page(GFP_KERNEL | GFP_ATOMIC);
if (!tmpname)
return -ENOMEM;
tmpde = (struct iso_directory_record *) (tmpname+1024);



Sylvain Pasche

2002-10-10 21:19:31

by Robert Love

Subject: Re: 2.5.41 isofs patch to avoid "bad: scheduling while atomic!"

On Thu, 2002-10-10 at 13:24, Sylvain Pasche wrote:


> - tmpname = (char *) __get_free_page(GFP_KERNEL);
> + tmpname = (char *) __get_free_page(GFP_KERNEL | GFP_ATOMIC);
> if (!tmpname)
> return -ENOMEM;
> tmpde = (struct iso_directory_record *) (tmpname+1024);

You just want GFP_ATOMIC, not the OR of both.

I do not see where the lock is in the call path, though. The
lock_kernel() is not counted.

I also wonder why the __might_sleep was just triggered but the
schedule() check was? Are you sure this is the culprit? By the looks
of it, I would think it is somewhere north of __might_sleep in the stack
trace.

Robert Love

2002-10-10 21:49:46

by Andrew Morton

Subject: Re: 2.5.41 isofs patch to avoid "bad: scheduling while atomic!"

Sylvain Pasche wrote:
>
> --- linux-2.5.41/fs/isofs/dir.c_old 2002-10-10 19:12:19.000000000 +0200
> +++ linux-2.5.41/fs/isofs/dir.c 2002-10-10 19:13:26.000000000 +0200
> @@ -256,7 +256,7 @@
>
> lock_kernel();
>
> - tmpname = (char *) __get_free_page(GFP_KERNEL);
> + tmpname = (char *) __get_free_page(GFP_KERNEL | GFP_ATOMIC);
> if (!tmpname)
> return -ENOMEM;

Not sure about the scheduling while atomic thing, but it is returning
with the lock held. I'll fix that up.