Skip to contentSkip to conversation

fix GFP_COMP use in noncoherent dma allocators

PATCH In review Christoph Hellwig 4 participants 7 messages Dec 20, 2022 – Dec 21, 2022

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

Hi all,

this series fixe the regression where dma-iommu stopped clearing GFP_COMP from noncoherent dma allocations by never requesting those in the sound code and rejecting them in the core DMA API code.

Diffstat:

 kernel/dma/mapping.c  |    4 ++++
 sound/core/memalloc.c |    4 ++--
 2 files changed, 6 insertions(+), 2 deletions(-)
#1 lore raw

PATCH 1/2 ALSA: memalloc: don't use GFP_COMP for non-coherent dma allocations

While not quite as bogus as for the dma-coherent allocations that were fixed earlier, GFP_COMP for these allocations has no benefits for the dma-direct case, and can't be supported at all by dma dma-iommu backend which splits up allocations into smaller orders. Due to an oversight in ffcb75458460 that flag stopped being cleared for all dma allocations, but only got rejected for coherent ones.

Start fixing this by not requesting __GFP_COMP in the sound code, which is the only place that did this.

Fixes ffcb75458460 ("dma-mapping: reject __GFP_COMP in dma_alloc_attrs")
---
 sound/core/memalloc.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
sound/core/memalloc.c +2 -2
index 3cf5a87d69eaf3..81025f50a54229 100644
@@ -542,7 +542,7 @@ static void *snd_dma_noncontig_alloc(struct snd_dma_buffer *dmab, size_t size)
 	void *p;
 
 	sgt = dma_alloc_noncontiguous(dmab->dev.dev, size, dmab->dev.dir,
-				      DEFAULT_GFP | __GFP_COMP, 0);
+				      DEFAULT_GFP, 0);
 #ifdef CONFIG_SND_DMA_SGBUF
 	if (!sgt && !get_dma_ops(dmab->dev.dev)) {
 		if (dmab->dev.type == SNDRV_DMA_TYPE_DEV_WC_SG)
@@ -820,7 +820,7 @@ static void *snd_dma_noncoherent_alloc(struct snd_dma_buffer *dmab, size_t size)
 	void *p;
 
 	p = dma_alloc_noncoherent(dmab->dev.dev, size, &dmab->addr,
-				  dmab->dev.dir, DEFAULT_GFP | __GFP_COMP);
+				  dmab->dev.dir, DEFAULT_GFP);
 	if (p)
 		dmab->dev.need_sync = dma_need_sync(dmab->dev.dev, dmab->addr);
 	return p;
Signature · 1 line
2.35.1
#2 lore raw

PATCH 2/2 dma-mapping: reject GFP_COMP for noncohernt allocaions

In reply to Christoph Hellwig

While not quite as bogus as for the dma-coherent allocations that were fixed earlier, GFP_COMP for these allocations has no benefits for the dma-direct case, and can't be supported at all by dma dma-iommu backend which splits up allocations into smaller orders. Due to an oversight in ffcb75458460 that flag stopped being cleared for all dma allocations, but only got rejected for coherent ones, so fix up these callers to not allow __GFP_COMP as well after the sound code has been fixed to not ask for it.

Fixes ffcb75458460 ("dma-mapping: reject __GFP_COMP in dma_alloc_attrs")
---
 kernel/dma/mapping.c | 4 ++++
 1 file changed, 4 insertions(+)
kernel/dma/mapping.c +4 -0
index c026a5a5e0466e..68106e3791f6c3 100644
@@ -560,6 +560,8 @@ static struct page *__dma_alloc_pages(struct device *dev, size_t size,
 		return NULL;
 	if (WARN_ON_ONCE(gfp & (__GFP_DMA | __GFP_DMA32 | __GFP_HIGHMEM)))
 		return NULL;
+	if (WARN_ON_ONCE(gfp & __GFP_COMP))
+		return NULL;
 
 	size = PAGE_ALIGN(size);
 	if (dma_alloc_direct(dev, ops))
@@ -645,6 +647,8 @@ struct sg_table *dma_alloc_noncontiguous(struct device *dev, size_t size,
 
 	if (WARN_ON_ONCE(attrs & ~DMA_ATTR_ALLOC_SINGLE_PAGES))
 		return NULL;
+	if (WARN_ON_ONCE(gfp & __GFP_COMP))
+		return NULL;
 
 	if (ops && ops->alloc_noncontiguous)
 		sgt = ops->alloc_noncontiguous(dev, size, dir, gfp, attrs);
Signature · 1 line
2.35.1
Tested-by on v1 2/2
#3 lore raw

Re: [PATCH 2/2] dma-mapping: reject GFP_COMP for noncohernt allocaions

Hey,

Quoted 3 lines · Christoph Hellwig

While not quite as bogus as for the dma-coherent allocations that were fixed earlier, GFP_COMP for these allocations has no benefits for the dma-direct case, and can't be supported at all by dma dma-iommu

tested the series and this fixes the issue:

Minor nit, typo "noncohernt allocaions" in subject of this second patch.

Br, Kai

#4 lore raw

Re: [PATCH 2/2] dma-mapping: reject GFP_COMP for noncohernt allocaions

Quoted 9 lines · Kai Vehmanen

Hey,

Quoted 3 lines · Christoph Hellwig

While not quite as bogus as for the dma-coherent allocations that were fixed earlier, GFP_COMP for these allocations has no benefits for the dma-direct case, and can't be supported at all by dma dma-iommu

Quoted 6 lines

tested the series and this fixes the issue:

Minor nit, typo "noncohernt allocaions" in subject of this second patch.

Br, Kai

I also confirm that after applying this patch series issue was gone.

Thanks.

Signature · 2 lines
Best Regards,
Mike Gavrilov.
#5 lore raw

Re: [PATCH 2/2] dma-mapping: reject GFP_COMP for noncohernt allocaions

In reply to Kai Vehmanen
Quoted 1 line · Kai Vehmanen

Minor nit, typo "noncohernt allocaions" in subject of this second patch.

Thanks,

I've fixed this and queued up the series.

Acked-by
#6 lore raw
In reply to Christoph Hellwig
Quoted 11 lines · Christoph Hellwig

Hi all,

this series fixe the regression where dma-iommu stopped clearing GFP_COMP from noncoherent dma allocations by never requesting those in the sound code and rejecting them in the core DMA API code.

Diffstat:

 kernel/dma/mapping.c  |    4 ++++
 sound/core/memalloc.c |    4 ++--
 2 files changed, 6 insertions(+), 2 deletions(-)

Thank you Christoph for taking care of this mess, and apologies for my delayed reaction, as I've been (still) on vacation since the last week.

FWIW, feel free to take my ack for the sound part:

Takashi