Re: [PATCH] regression: vmalloc easily fail.

[Nick Piggin <npiggin@suse.de>]

On Wed, Oct 29, 2008 at 08:07:37PM -0200, Glauber Costa wrote:
> On Wed, Oct 29, 2008 at 11:43:33AM +0100, Nick Piggin wrote:
> > On Wed, Oct 29, 2008 at 12:29:40PM +0200, Avi Kivity wrote:
> > > Nick Piggin wrote:
> > > >Hmm, spanning <30MB of memory... how much vmalloc space do you have?
> > > >
> > > >  
> > > 
> > > From the original report:
> > > 
> > > >VmallocTotal:     122880 kB
> > > >VmallocUsed:       15184 kB
> > > >VmallocChunk:      83764 kB
> > > 
> > > So it seems there's quite a bit of free space.
> > > 
> > > Chunk is the largest free contiguous region, right?  If so, it seems the 
> > 
> > Yes.
> > 
> > 
> > > problem is unrelated to guard pages, instead the search isn't finding a 
> > > 1-page area (with two guard pages) for some reason, even though lots of 
> > > free space is available.
> > 
> > Hmm. The free area search could be buggy...
> Do you want me to grab any specific info of it? Or should I just hack myself
> randomly into it? I'll probably have some time for that tomorrow.

I took a bit of a look. Does this help you at all?

I still think we should get rid of the guard pages in non-debug kernels
completely, but hopefully this will fix your problems?
--

- Fix off by one bug in the KVA allocator that can leave gaps 
- An initial vmalloc failure should start off a synchronous flush of lazy
  areas, in case someone is in progress flushing them already.
- Purge lock can be a mutex so we can sleep while that's going on.
 
Signed-off-by: Nick Piggin <npiggin@suse.de>
---
Index: linux-2.6/mm/vmalloc.c
===================================================================
--- linux-2.6.orig/mm/vmalloc.c
+++ linux-2.6/mm/vmalloc.c
@@ -14,6 +14,7 @@
 #include <linux/highmem.h>
 #include <linux/slab.h>
 #include <linux/spinlock.h>
+#include <linux/mutex.h>
 #include <linux/interrupt.h>
 #include <linux/proc_fs.h>
 #include <linux/seq_file.h>
@@ -362,7 +363,7 @@ retry:
 				goto found;
 		}
 
-		while (addr + size >= first->va_start && addr + size <= vend) {
+		while (addr + size > first->va_start && addr + size <= vend) {
 			addr = ALIGN(first->va_end + PAGE_SIZE, align);
 
 			n = rb_next(&first->rb_node);
@@ -472,7 +473,7 @@ static atomic_t vmap_lazy_nr = ATOMIC_IN
 static void __purge_vmap_area_lazy(unsigned long *start, unsigned long *end,
 					int sync, int force_flush)
 {
-	static DEFINE_SPINLOCK(purge_lock);
+	static DEFINE_MUTEX(purge_lock);
 	LIST_HEAD(valist);
 	struct vmap_area *va;
 	int nr = 0;
@@ -483,10 +484,10 @@ static void __purge_vmap_area_lazy(unsig
 	 * the case that isn't actually used at the moment anyway.
 	 */
 	if (!sync && !force_flush) {
-		if (!spin_trylock(&purge_lock))
+		if (!mutex_trylock(&purge_lock))
 			return;
 	} else
-		spin_lock(&purge_lock);
+		mutex_lock(&purge_lock);
 
 	rcu_read_lock();
 	list_for_each_entry_rcu(va, &vmap_area_list, list) {
@@ -518,7 +519,18 @@ static void __purge_vmap_area_lazy(unsig
 			__free_vmap_area(va);
 		spin_unlock(&vmap_area_lock);
 	}
-	spin_unlock(&purge_lock);
+	mutex_unlock(&purge_lock);
+}
+
+/*
+ * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
+ * is already purging.
+ */
+static void try_purge_vmap_area_lazy(void)
+{
+	unsigned long start = ULONG_MAX, end = 0;
+
+	__purge_vmap_area_lazy(&start, &end, 0, 0);
 }
 
 /*
@@ -528,7 +540,7 @@ static void purge_vmap_area_lazy(void)
 {
 	unsigned long start = ULONG_MAX, end = 0;
 
-	__purge_vmap_area_lazy(&start, &end, 0, 0);
+	__purge_vmap_area_lazy(&start, &end, 1, 0);
 }
 
 /*
@@ -539,7 +551,7 @@ static void free_unmap_vmap_area(struct 
 	va->flags |= VM_LAZY_FREE;
 	atomic_add((va->va_end - va->va_start) >> PAGE_SHIFT, &vmap_lazy_nr);
 	if (unlikely(atomic_read(&vmap_lazy_nr) > lazy_max_pages()))
-		purge_vmap_area_lazy();
+		try_purge_vmap_area_lazy();
 }
 
 static struct vmap_area *find_vmap_area(unsigned long addr)