ALSA: memalloc: Add fallback SG-buffer allocations for x86

The recent change for memory allocator replaced the SG-buffer handling
helper for x86 with the standard non-contiguous page handler.  This
works for most cases, but there is a corner case I obviously
overlooked, namely, the fallback of non-contiguous handler without
IOMMU.  When the system runs without IOMMU, the core handler tries to
use the continuous pages with a single SGL entry.  It works nicely for
most cases, but when the system memory gets fragmented, the large
allocation may fail frequently.

Ideally the non-contig handler could deal with the proper SG pages,
it's cumbersome to extend for now.  As a workaround, here we add new
types for (minimalistic) SG allocations, instead, so that the
allocator falls back to those types automatically when the allocation
with the standard API failed.

BTW, one better (but pretty minor) improvement from the previous
SG-buffer code is that this provides the proper mmap support without
the PCM's page fault handling.

Fixes: 2c95b92ecd ("ALSA: memalloc: Unify x86 SG-buffer handling (take#3)")
BugLink: https://gitlab.freedesktop.org/pipewire/pipewire/-/issues/2272
BugLink: https://bugzilla.suse.com/show_bug.cgi?id=1198248
Cc: <stable@vger.kernel.org>
Link: https://lore.kernel.org/r/20220413054808.7547-1-tiwai@suse.de
Signed-off-by: Takashi Iwai <tiwai@suse.de>

include/sound/memalloc.h

@@ -51,6 +51,11 @@ struct snd_dma_device {
 #define SNDRV_DMA_TYPE_DEV_SG	SNDRV_DMA_TYPE_DEV /* no SG-buf support */
 #define SNDRV_DMA_TYPE_DEV_WC_SG	SNDRV_DMA_TYPE_DEV_WC
 #endif
+/* fallback types, don't use those directly */
+#ifdef CONFIG_SND_DMA_SGBUF
+#define SNDRV_DMA_TYPE_DEV_SG_FALLBACK		10
+#define SNDRV_DMA_TYPE_DEV_WC_SG_FALLBACK	11
+#endif
 
 /*
  * info for buffer allocation

sound/core/memalloc.c

@@ -499,6 +499,10 @@ static const struct snd_malloc_ops snd_dma_wc_ops = {
 };
 #endif /* CONFIG_X86 */
 
+#ifdef CONFIG_SND_DMA_SGBUF
+static void *snd_dma_sg_fallback_alloc(struct snd_dma_buffer *dmab, size_t size);
+#endif
+
 /*
  * Non-contiguous pages allocator
  */
@@ -509,8 +513,18 @@ static void *snd_dma_noncontig_alloc(struct snd_dma_buffer *dmab, size_t size)
 
 	sgt = dma_alloc_noncontiguous(dmab->dev.dev, size, dmab->dev.dir,
 				      DEFAULT_GFP, 0);
-	if (!sgt)
+	if (!sgt) {
+#ifdef CONFIG_SND_DMA_SGBUF
+		if (dmab->dev.type == SNDRV_DMA_TYPE_DEV_WC_SG)
+			dmab->dev.type = SNDRV_DMA_TYPE_DEV_WC_SG_FALLBACK;
+		else
+			dmab->dev.type = SNDRV_DMA_TYPE_DEV_SG_FALLBACK;
+		return snd_dma_sg_fallback_alloc(dmab, size);
+#else
 		return NULL;
+#endif
+	}
+
 	dmab->dev.need_sync = dma_need_sync(dmab->dev.dev,
 					    sg_dma_address(sgt->sgl));
 	p = dma_vmap_noncontiguous(dmab->dev.dev, size, sgt);
@@ -633,6 +647,8 @@ static void *snd_dma_sg_wc_alloc(struct snd_dma_buffer *dmab, size_t size)
 
 	if (!p)
 		return NULL;
+	if (dmab->dev.type != SNDRV_DMA_TYPE_DEV_WC_SG)
+		return p;
 	for_each_sgtable_page(sgt, &iter, 0)
 		set_memory_wc(sg_wc_address(&iter), 1);
 	return p;
@@ -665,6 +681,95 @@ static const struct snd_malloc_ops snd_dma_sg_wc_ops = {
 	.get_page = snd_dma_noncontig_get_page,
 	.get_chunk_size = snd_dma_noncontig_get_chunk_size,
 };
+
+/* Fallback SG-buffer allocations for x86 */
+struct snd_dma_sg_fallback {
+	size_t count;
+	struct page **pages;
+	dma_addr_t *addrs;
+};
+
+static void __snd_dma_sg_fallback_free(struct snd_dma_buffer *dmab,
+				       struct snd_dma_sg_fallback *sgbuf)
+{
+	size_t i;
+
+	if (sgbuf->count && dmab->dev.type == SNDRV_DMA_TYPE_DEV_WC_SG_FALLBACK)
+		set_pages_array_wb(sgbuf->pages, sgbuf->count);
+	for (i = 0; i < sgbuf->count && sgbuf->pages[i]; i++)
+		dma_free_coherent(dmab->dev.dev, PAGE_SIZE,
+				  page_address(sgbuf->pages[i]),
+				  sgbuf->addrs[i]);
+	kvfree(sgbuf->pages);
+	kvfree(sgbuf->addrs);
+	kfree(sgbuf);
+}
+
+static void *snd_dma_sg_fallback_alloc(struct snd_dma_buffer *dmab, size_t size)
+{
+	struct snd_dma_sg_fallback *sgbuf;
+	struct page **pages;
+	size_t i, count;
+	void *p;
+
+	sgbuf = kzalloc(sizeof(*sgbuf), GFP_KERNEL);
+	if (!sgbuf)
+		return NULL;
+	count = PAGE_ALIGN(size) >> PAGE_SHIFT;
+	pages = kvcalloc(count, sizeof(*pages), GFP_KERNEL);
+	if (!pages)
+		goto error;
+	sgbuf->pages = pages;
+	sgbuf->addrs = kvcalloc(count, sizeof(*sgbuf->addrs), GFP_KERNEL);
+	if (!sgbuf->addrs)
+		goto error;
+
+	for (i = 0; i < count; sgbuf->count++, i++) {
+		p = dma_alloc_coherent(dmab->dev.dev, PAGE_SIZE,
+				       &sgbuf->addrs[i], DEFAULT_GFP);
+		if (!p)
+			goto error;
+		sgbuf->pages[i] = virt_to_page(p);
+	}
+
+	if (dmab->dev.type == SNDRV_DMA_TYPE_DEV_WC_SG_FALLBACK)
+		set_pages_array_wc(pages, count);
+	p = vmap(pages, count, VM_MAP, PAGE_KERNEL);
+	if (!p)
+		goto error;
+	dmab->private_data = sgbuf;
+	return p;
+
+ error:
+	__snd_dma_sg_fallback_free(dmab, sgbuf);
+	return NULL;
+}
+
+static void snd_dma_sg_fallback_free(struct snd_dma_buffer *dmab)
+{
+	vunmap(dmab->area);
+	__snd_dma_sg_fallback_free(dmab, dmab->private_data);
+}
+
+static int snd_dma_sg_fallback_mmap(struct snd_dma_buffer *dmab,
+				    struct vm_area_struct *area)
+{
+	struct snd_dma_sg_fallback *sgbuf = dmab->private_data;
+
+	if (dmab->dev.type == SNDRV_DMA_TYPE_DEV_WC_SG_FALLBACK)
+		area->vm_page_prot = pgprot_writecombine(area->vm_page_prot);
+	return vm_map_pages(area, sgbuf->pages, sgbuf->count);
+}
+
+static const struct snd_malloc_ops snd_dma_sg_fallback_ops = {
+	.alloc = snd_dma_sg_fallback_alloc,
+	.free = snd_dma_sg_fallback_free,
+	.mmap = snd_dma_sg_fallback_mmap,
+	/* reuse vmalloc helpers */
+	.get_addr = snd_dma_vmalloc_get_addr,
+	.get_page = snd_dma_vmalloc_get_page,
+	.get_chunk_size = snd_dma_vmalloc_get_chunk_size,
+};
 #endif /* CONFIG_SND_DMA_SGBUF */
 
 /*
@@ -736,6 +841,10 @@ static const struct snd_malloc_ops *dma_ops[] = {
 #ifdef CONFIG_GENERIC_ALLOCATOR
 	[SNDRV_DMA_TYPE_DEV_IRAM] = &snd_dma_iram_ops,
 #endif /* CONFIG_GENERIC_ALLOCATOR */
+#ifdef CONFIG_SND_DMA_SGBUF
+	[SNDRV_DMA_TYPE_DEV_SG_FALLBACK] = &snd_dma_sg_fallback_ops,
+	[SNDRV_DMA_TYPE_DEV_WC_SG_FALLBACK] = &snd_dma_sg_fallback_ops,
+#endif
 #endif /* CONFIG_HAS_DMA */
 };