Re: inappropriate use of in_atomic()

From: Stephen Rothwell
Date: Fri Mar 11 2005 - 01:27:15 EST


Hi Andrew,

On Thu, 10 Mar 2005 20:40:06 -0800 Andrew Morton <akpm@xxxxxxxx> wrote:
>
> in_atomic() is not a reliable indication of whether it is currently safe
> to call schedule().
>
> arch/ppc64/kernel/viopath.c

in_atomic() in viopath.c was just used to determine if we had initialised
enough to be able to wait in a semaphore (i.e. schedule). Thus it can be
replaced now with checking system_state for SYSTEM_RUNNING.

Signed-off-by: Stephen Rothwell <sfr@xxxxxxxxxxxxxxxx>

Test booted on iSeries (which is the only place it is used).
--
Cheers,
Stephen Rothwell sfr@xxxxxxxxxxxxxxxx
http://www.canb.auug.org.au/~sfr/

diff -ruNp linus/arch/ppc64/kernel/viopath.c linus-in_atomic/arch/ppc64/kernel/viopath.c
--- linus/arch/ppc64/kernel/viopath.c 2005-01-22 06:09:01.000000000 +1100
+++ linus-in_atomic/arch/ppc64/kernel/viopath.c 2005-03-11 17:19:45.000000000 +1100
@@ -79,7 +79,7 @@ static void handleMonitorEvent(struct Hv
/*
* We use this structure to handle asynchronous responses. The caller
* blocks on the semaphore and the handler posts the semaphore. However,
- * if in_atomic() is true in the caller, then wait_atomic is used ...
+ * if system_state is not SYSTEM_RUNNING, then wait_atomic is used ...
*/
struct doneAllocParms_t {
struct semaphore *sem;
@@ -465,7 +465,7 @@ static int allocateEvents(HvLpIndex remo
DECLARE_MUTEX_LOCKED(Semaphore);
atomic_t wait_atomic;

- if (in_atomic()) {
+ if (system_state != SYSTEM_RUNNING) {
parms.used_wait_atomic = 1;
atomic_set(&wait_atomic, 1);
parms.wait_atomic = &wait_atomic;
@@ -475,7 +475,7 @@ static int allocateEvents(HvLpIndex remo
}
mf_allocate_lp_events(remoteLp, HvLpEvent_Type_VirtualIo, 250, /* It would be nice to put a real number here! */
numEvents, &viopath_donealloc, &parms);
- if (in_atomic()) {
+ if (system_state != SYSTEM_RUNNING) {
while (atomic_read(&wait_atomic))
mb();
} else

Attachment: pgp00000.pgp
Description: PGP signature