[v2,2/2] dma-pool: Only allocate from CMA when in same memory zone
diff mbox series

Message ID 20200803160956.19235-3-nsaenzjulienne@suse.de
State Superseded
Headers show
Series
  • dma-pool fixes
Related show

Commit Message

Nicolas Saenz Julienne Aug. 3, 2020, 4:09 p.m. UTC
There is no guarantee to CMA's placement, so allocating a zone specific
atomic pool from CMA might return memory from a completely different
memory zone. To get around this double check CMA's placement before
allocating from it.

Signed-off-by: Nicolas Saenz Julienne <nsaenzjulienne@suse.de>
---

Changes since v1:
 - Make cma_in_zone() more strict, GFP_KERNEL doesn't default to true
   now

 kernel/dma/pool.c | 34 +++++++++++++++++++++++++++++++++-
 1 file changed, 33 insertions(+), 1 deletion(-)

Comments

Christoph Hellwig Aug. 4, 2020, 6:06 a.m. UTC | #1
On Mon, Aug 03, 2020 at 06:09:56PM +0200, Nicolas Saenz Julienne wrote:
> +	if (IS_ENABLED(CONFIG_ZONE_DMA) && (gfp & GFP_DMA))
> +		return end <= DMA_BIT_MASK(zone_dma_bits);
> +	if (IS_ENABLED(CONFIG_ZONE_DMA32) && (gfp & GFP_DMA32))
> +		return end <= DMA_BIT_MASK(32);
> +	if (gfp & GFP_KERNEL)
> +		return end > DMA_BIT_MASK(32);

So the GFP_KERNEL one here looks weird.  For one I don't think the if
line is needed at all, and it just confuses things.  Second I don't
see the need (and actually some harm) in preventing GFP_KERNEL
allocations from dipping into lower CMA areas - something that we did
support before 5.8 with the single pool.
Nicolas Saenz Julienne Aug. 4, 2020, 9:43 a.m. UTC | #2
On Tue, 2020-08-04 at 08:06 +0200, Christoph Hellwig wrote:
> On Mon, Aug 03, 2020 at 06:09:56PM +0200, Nicolas Saenz Julienne wrote:
> > +	if (IS_ENABLED(CONFIG_ZONE_DMA) && (gfp & GFP_DMA))
> > +		return end <= DMA_BIT_MASK(zone_dma_bits);
> > +	if (IS_ENABLED(CONFIG_ZONE_DMA32) && (gfp & GFP_DMA32))
> > +		return end <= DMA_BIT_MASK(32);
> > +	if (gfp & GFP_KERNEL)
> > +		return end > DMA_BIT_MASK(32);
> 
> So the GFP_KERNEL one here looks weird.  For one I don't think the if
> line is needed at all, and it just confuses things.

Yes, sorry, shoud've seen that.

> Second I don't see the need (and actually some harm) in preventing GFP_KERNEL
> allocations from dipping into lower CMA areas - something that we did support
> before 5.8 with the single pool.

My thinking is the least we pressure CMA the better, it's generally scarse, and
it'll not grow as the atomic pools grow. As far as harm is concerned, we now
check addresses for correctness, so we shouldn't run into problems.

There is a potential case for architectures defining a default CMA but not
defining DMA zones where this could be problematic. But isn't that just plain
abusing CMA? If you need low memory allocations, you should be defining DMA
zones.

Regards,
Nicolas
Christoph Hellwig Aug. 6, 2020, 5:18 a.m. UTC | #3
On Tue, Aug 04, 2020 at 11:43:15AM +0200, Nicolas Saenz Julienne wrote:
> > Second I don't see the need (and actually some harm) in preventing GFP_KERNEL
> > allocations from dipping into lower CMA areas - something that we did support
> > before 5.8 with the single pool.
> 
> My thinking is the least we pressure CMA the better, it's generally scarse, and
> it'll not grow as the atomic pools grow. As far as harm is concerned, we now
> check addresses for correctness, so we shouldn't run into problems.
> 
> There is a potential case for architectures defining a default CMA but not
> defining DMA zones where this could be problematic. But isn't that just plain
> abusing CMA? If you need low memory allocations, you should be defining DMA
> zones.

The latter is pretty much what I expect, as we only support the default and
per-device DMA CMAs.
Nicolas Saenz Julienne Aug. 6, 2020, 11:50 a.m. UTC | #4
On Thu, 2020-08-06 at 07:18 +0200, Christoph Hellwig wrote:
> On Tue, Aug 04, 2020 at 11:43:15AM +0200, Nicolas Saenz Julienne wrote:
> > > Second I don't see the need (and actually some harm) in preventing GFP_KERNEL
> > > allocations from dipping into lower CMA areas - something that we did support
> > > before 5.8 with the single pool.
> > 
> > My thinking is the least we pressure CMA the better, it's generally scarse, and
> > it'll not grow as the atomic pools grow. As far as harm is concerned, we now
> > check addresses for correctness, so we shouldn't run into problems.
> > 
> > There is a potential case for architectures defining a default CMA but not
> > defining DMA zones where this could be problematic. But isn't that just plain
> > abusing CMA? If you need low memory allocations, you should be defining DMA
> > zones.
> 
> The latter is pretty much what I expect, as we only support the default and
> per-device DMA CMAs.

Fair enough, should I send a v3 with everything cleaned-up/rebased, or you'd
rather pick it up from your version?
Christoph Hellwig Aug. 6, 2020, 1:55 p.m. UTC | #5
On Thu, Aug 06, 2020 at 01:50:29PM +0200, Nicolas Saenz Julienne wrote:
> > The latter is pretty much what I expect, as we only support the default and
> > per-device DMA CMAs.
> 
> Fair enough, should I send a v3 with everything cleaned-up/rebased, or you'd
> rather pick it up from your version?

Please just resend the whole thing.

Patch
diff mbox series

diff --git a/kernel/dma/pool.c b/kernel/dma/pool.c
index 5d071d4a3cba..582523ccf4fe 100644
--- a/kernel/dma/pool.c
+++ b/kernel/dma/pool.c
@@ -3,7 +3,9 @@ 
  * Copyright (C) 2012 ARM Ltd.
  * Copyright (C) 2020 Google LLC
  */
+#include <linux/cma.h>
 #include <linux/debugfs.h>
+#include <linux/dma-contiguous.h>
 #include <linux/dma-direct.h>
 #include <linux/dma-noncoherent.h>
 #include <linux/init.h>
@@ -55,6 +57,32 @@  static void dma_atomic_pool_size_add(gfp_t gfp, size_t size)
 		pool_size_kernel += size;
 }
 
+static bool cma_in_zone(gfp_t gfp)
+{
+	unsigned long size;
+	phys_addr_t end;
+	struct cma *cma;
+
+	cma = dev_get_cma_area(NULL);
+	if (!cma)
+		return false;
+
+	size = cma_get_size(cma);
+	if (!size)
+		return false;
+
+	/* CMA can't cross zone boundaries, see cma_activate_area() */
+	end = cma_get_base(cma) - memblock_start_of_DRAM() + size - 1;
+	if (IS_ENABLED(CONFIG_ZONE_DMA) && (gfp & GFP_DMA))
+		return end <= DMA_BIT_MASK(zone_dma_bits);
+	if (IS_ENABLED(CONFIG_ZONE_DMA32) && (gfp & GFP_DMA32))
+		return end <= DMA_BIT_MASK(32);
+	if (gfp & GFP_KERNEL)
+		return end > DMA_BIT_MASK(32);
+
+	return false;
+}
+
 static int atomic_pool_expand(struct gen_pool *pool, size_t pool_size,
 			      gfp_t gfp)
 {
@@ -68,7 +96,11 @@  static int atomic_pool_expand(struct gen_pool *pool, size_t pool_size,
 
 	do {
 		pool_size = 1 << (PAGE_SHIFT + order);
-		page = alloc_pages(gfp, order);
+		if (cma_in_zone(gfp))
+ 			page = dma_alloc_from_contiguous(NULL, 1 << order,
+ 							 order, false);
+		if (!page)
+			page = alloc_pages(gfp, order);
 	} while (!page && order-- > 0);
 	if (!page)
 		goto out;