Re: [PATCH] release quicklist before free_page

[Peter Zijlstra <peterz@infradead.org>]

On Mon, 2007-07-23 at 08:21 -0700, Daniel Walker wrote:
> Resolves,
> 
> BUG: sleeping function called from invalid context cc1(29651) at kernel/rtmutex.c:636
> in_atomic():1 [00000001], irqs_disabled():0
>  [<c0119f50>] __might_sleep+0xf3/0xf9
>  [<c031600e>] __rt_spin_lock+0x21/0x3c
>  [<c014102c>] get_zone_pcp+0x20/0x29
>  [<c0141a40>] free_hot_cold_page+0xdc/0x167
>  [<c013a3f4>] add_preempt_count+0x12/0xcc
>  [<c0110d92>] pgd_dtor+0x0/0x1
>  [<c015d865>] quicklist_trim+0xb7/0xe3
>  [<c0111025>] check_pgt_cache+0x19/0x1c
>  [<c0148df5>] free_pgtables+0x54/0x12c
>  [<c013a3f4>] add_preempt_count+0x12/0xcc
>  [<c014e5be>] unmap_region+0xeb/0x13b
> 
> 
> It looks like the quicklist isn't used after a few variables are evaluated.
> So no need to keep preemption disabled over the whole function.

Not quite, it uses preempt_disable() to avoid migration and stick to a
cpu. Without that it might end up freeing pages from another quicklist.

How about this - compile tested only
---

We cannot call the page allocator with preemption-disabled, use the
per_cpu_locked construct to allow preemption while guarding the per cpu
data.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
---
 include/linux/quicklist.h |   19 +++++++++++++++----
 mm/quicklist.c            |    9 +++++----
 2 files changed, 20 insertions(+), 8 deletions(-)

Index: linux-2.6/include/linux/quicklist.h
===================================================================
--- linux-2.6.orig/include/linux/quicklist.h
+++ linux-2.6/include/linux/quicklist.h
@@ -18,7 +18,7 @@ struct quicklist {
 	int nr_pages;
 };
 
-DECLARE_PER_CPU(struct quicklist, quicklist)[CONFIG_NR_QUICK];
+DECLARE_PER_CPU_LOCKED(struct quicklist, quicklist)[CONFIG_NR_QUICK];
 
 /*
  * The two key functions quicklist_alloc and quicklist_free are inline so
@@ -30,19 +30,30 @@ DECLARE_PER_CPU(struct quicklist, quickl
  * The fast patch in quicklist_alloc touched only a per cpu cacheline and
  * the first cacheline of the page itself. There is minmal overhead involved.
  */
-static inline void *quicklist_alloc(int nr, gfp_t flags, void (*ctor)(void *))
+static inline void *__quicklist_alloc(int cpu, int nr, gfp_t flags, void (*ctor)(void *))
 {
 	struct quicklist *q;
 	void **p = NULL;
 
-	q =&get_cpu_var(quicklist)[nr];
+	q = &__get_cpu_var_locked(quicklist, cpu)[nr];
 	p = q->page;
 	if (likely(p)) {
 		q->page = p[0];
 		p[0] = NULL;
 		q->nr_pages--;
 	}
-	put_cpu_var(quicklist);
+	return p;
+}
+
+static inline void *quicklist_alloc(int nr, gfp_t flags, void (*ctor)(void *))
+{
+	struct quicklist *q;
+	void **p = NULL;
+	int cpu;
+
+	(void)get_cpu_var_locked(quicklist, &cpu)[nr];
+	p = __quicklist_alloc(cpu, nr, flags, ctor);
+	put_cpu_var_locked(quicklist, cpu);
 	if (likely(p))
 		return p;
 
Index: linux-2.6/mm/quicklist.c
===================================================================
--- linux-2.6.orig/mm/quicklist.c
+++ linux-2.6/mm/quicklist.c
@@ -19,7 +19,7 @@
 #include <linux/module.h>
 #include <linux/quicklist.h>
 
-DEFINE_PER_CPU(struct quicklist, quicklist)[CONFIG_NR_QUICK];
+DEFINE_PER_CPU_LOCKED(struct quicklist, quicklist)[CONFIG_NR_QUICK];
 
 #define FRACTION_OF_NODE_MEM	16
 
@@ -51,8 +51,9 @@ void quicklist_trim(int nr, void (*dtor)
 {
 	long pages_to_free;
 	struct quicklist *q;
+	int cpu;
 
-	q = &get_cpu_var(quicklist)[nr];
+	q = &get_cpu_var_locked(quicklist, &cpu)[nr];
 	if (q->nr_pages > min_pages) {
 		pages_to_free = min_pages_to_free(q, min_pages, max_free);
 
@@ -61,7 +62,7 @@ void quicklist_trim(int nr, void (*dtor)
 			 * We pass a gfp_t of 0 to quicklist_alloc here
 			 * because we will never call into the page allocator.
 			 */
-			void *p = quicklist_alloc(nr, 0, NULL);
+			void *p = __quicklist_alloc(cpu, nr, 0, NULL);
 
 			if (dtor)
 				dtor(p);
@@ -69,7 +70,7 @@ void quicklist_trim(int nr, void (*dtor)
 			pages_to_free--;
 		}
 	}
-	put_cpu_var(quicklist);
+	put_cpu_var_locked(quicklist, cpu);
 }
 
 unsigned long quicklist_total_size(void)


-
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/