From: Miklos Szeredi <[email protected]>
This fixes a bug in reserve-slub.patch.
If cache_grow() was called with objp != NULL then the 'reserve' local
variable wasn't initialized. This resulted in ac->reserve being set to
a rubbish value. Due to this in some circumstances huge amounts of
slab pages were allocated (due to slab_force_alloc() returning true),
which caused atomic page allocation failures and slowdown of the
system.
Signed-off-by: Miklos Szeredi <[email protected]>
Signed-off-by: Suresh Jayaraman <[email protected]>
---
mm/slab.c | 5 +++--
1 file changed, 3 insertions(+), 2 deletions(-)
Index: mmotm/mm/slab.c
===================================================================
--- mmotm.orig/mm/slab.c
+++ mmotm/mm/slab.c
@@ -2760,7 +2760,7 @@ static int cache_grow(struct kmem_cache
size_t offset;
gfp_t local_flags;
struct kmem_list3 *l3;
- int reserve;
+ int reserve = -1;
/*
* Be lazy and only check for valid flags here, keeping it out of the
@@ -2816,7 +2816,8 @@ static int cache_grow(struct kmem_cache
if (local_flags & __GFP_WAIT)
local_irq_disable();
check_irq_off();
- slab_set_reserve(cachep, reserve);
+ if (reserve != -1)
+ slab_set_reserve(cachep, reserve);
spin_lock(&l3->list_lock);
/* Make slab active. */
On Thu, 1 Oct 2009, Suresh Jayaraman wrote:
> From: Miklos Szeredi <[email protected]>
>
> This fixes a bug in reserve-slub.patch.
>
> If cache_grow() was called with objp != NULL then the 'reserve' local
> variable wasn't initialized. This resulted in ac->reserve being set to
> a rubbish value. Due to this in some circumstances huge amounts of
> slab pages were allocated (due to slab_force_alloc() returning true),
> which caused atomic page allocation failures and slowdown of the
> system.
>
> Signed-off-by: Miklos Szeredi <[email protected]>
> Signed-off-by: Suresh Jayaraman <[email protected]>
> ---
> mm/slab.c | 5 +++--
> 1 file changed, 3 insertions(+), 2 deletions(-)
>
> Index: mmotm/mm/slab.c
> ===================================================================
> --- mmotm.orig/mm/slab.c
> +++ mmotm/mm/slab.c
> @@ -2760,7 +2760,7 @@ static int cache_grow(struct kmem_cache
> size_t offset;
> gfp_t local_flags;
> struct kmem_list3 *l3;
> - int reserve;
> + int reserve = -1;
>
> /*
> * Be lazy and only check for valid flags here, keeping it out of the
> @@ -2816,7 +2816,8 @@ static int cache_grow(struct kmem_cache
> if (local_flags & __GFP_WAIT)
> local_irq_disable();
> check_irq_off();
> - slab_set_reserve(cachep, reserve);
> + if (reserve != -1)
> + slab_set_reserve(cachep, reserve);
> spin_lock(&l3->list_lock);
>
> /* Make slab active. */
Given the patch description, shouldn't this be a test for objp != NULL
instead, then?
If so, it doesn't make sense because reserve will only be initialized when
objp == NULL in the call to kmem_getpages() from cache_grow().
The title of the patch suggests this is just dealing with an uninitialized
auto variable so the anticipated change would be from "int reserve" to
"int uninitialized_var(result)".
On Thursday October 1, [email protected] wrote:
> On Thu, 1 Oct 2009, Suresh Jayaraman wrote:
>
> > From: Miklos Szeredi <[email protected]>
> >
> > This fixes a bug in reserve-slub.patch.
> >
> > If cache_grow() was called with objp != NULL then the 'reserve' local
> > variable wasn't initialized. This resulted in ac->reserve being set to
> > a rubbish value. Due to this in some circumstances huge amounts of
> > slab pages were allocated (due to slab_force_alloc() returning true),
> > which caused atomic page allocation failures and slowdown of the
> > system.
> >
> > Signed-off-by: Miklos Szeredi <[email protected]>
> > Signed-off-by: Suresh Jayaraman <[email protected]>
> > ---
> > mm/slab.c | 5 +++--
> > 1 file changed, 3 insertions(+), 2 deletions(-)
> >
> > Index: mmotm/mm/slab.c
> > ===================================================================
> > --- mmotm.orig/mm/slab.c
> > +++ mmotm/mm/slab.c
> > @@ -2760,7 +2760,7 @@ static int cache_grow(struct kmem_cache
> > size_t offset;
> > gfp_t local_flags;
> > struct kmem_list3 *l3;
> > - int reserve;
> > + int reserve = -1;
> >
> > /*
> > * Be lazy and only check for valid flags here, keeping it out of the
> > @@ -2816,7 +2816,8 @@ static int cache_grow(struct kmem_cache
> > if (local_flags & __GFP_WAIT)
> > local_irq_disable();
> > check_irq_off();
> > - slab_set_reserve(cachep, reserve);
> > + if (reserve != -1)
> > + slab_set_reserve(cachep, reserve);
> > spin_lock(&l3->list_lock);
> >
> > /* Make slab active. */
>
> Given the patch description, shouldn't this be a test for objp != NULL
> instead, then?
In between those to patch hunks, cache_grow contains the code:
if (!objp)
objp = kmem_getpages(cachep, local_flags, nodeid, &reserve);
if (!objp)
goto failed;
We can no longer test if objp was NULL on entry to the function.
We could take a copy of objp on entry to the function, and test it
here. But initialising 'reserve' to an invalid value is easier.
>
> If so, it doesn't make sense because reserve will only be initialized when
> objp == NULL in the call to kmem_getpages() from cache_grow().
>
>
> The title of the patch suggests this is just dealing with an uninitialized
> auto variable so the anticipated change would be from "int reserve" to
> "int uninitialized_var(result)".
That change is only appropriate when the compiler is issuing a
warning that the variable is used before it is initialised, but we
know that not to be the case.
In this situation, we know it *is* being used before it is
initialised, and so we need to initialise it to something.
Thanks,
NeilBrown
On Fri, 2 Oct 2009, Neil Brown wrote:
> > > Index: mmotm/mm/slab.c
> > > ===================================================================
> > > --- mmotm.orig/mm/slab.c
> > > +++ mmotm/mm/slab.c
> > > @@ -2760,7 +2760,7 @@ static int cache_grow(struct kmem_cache
> > > size_t offset;
> > > gfp_t local_flags;
> > > struct kmem_list3 *l3;
> > > - int reserve;
> > > + int reserve = -1;
> > >
> > > /*
> > > * Be lazy and only check for valid flags here, keeping it out of the
> > > @@ -2816,7 +2816,8 @@ static int cache_grow(struct kmem_cache
> > > if (local_flags & __GFP_WAIT)
> > > local_irq_disable();
> > > check_irq_off();
> > > - slab_set_reserve(cachep, reserve);
> > > + if (reserve != -1)
> > > + slab_set_reserve(cachep, reserve);
> > > spin_lock(&l3->list_lock);
> > >
> > > /* Make slab active. */
> >
> > Given the patch description, shouldn't this be a test for objp != NULL
> > instead, then?
>
> In between those to patch hunks, cache_grow contains the code:
> if (!objp)
> objp = kmem_getpages(cachep, local_flags, nodeid, &reserve);
> if (!objp)
> goto failed;
>
> We can no longer test if objp was NULL on entry to the function.
> We could take a copy of objp on entry to the function, and test it
> here. But initialising 'reserve' to an invalid value is easier.
>
Seems like you could do all this in kmem_getpages(), then, by calling
slab_set_reserve(cachep, page->reserve) before returning the new page?
[ I'd also drop the branch in slab_set_reserve(), it's faster to just
assign it unconditionally. ]