[alsa-devel] [PATCH v2 4/5] ALSA: emu10k1: make sure synth DMA pages are allocated with DMA functions
Commit a5003fc04113 ("[ALSA] emu10k1 - simplify page allocation for synth") switched from using the DMA allocator for synth DMA pages to manually calling alloc_page(). However, this usage has an implicit assumption that the DMA address space for the emu10k1-family chip is the same as the CPU physical address space which is not true for a system with a IOMMU.
Since this made the synth part of the driver non-functional on such systems let's effectively revert that commit.
Signed-off-by: Maciej S. Szmigiero mail@maciej.szmigiero.name --- Changes from v1: None in this patch.
sound/pci/emu10k1/memory.c | 70 ++++++++++++++++++++++++++-------------------- 1 file changed, 39 insertions(+), 31 deletions(-)
diff --git a/sound/pci/emu10k1/memory.c b/sound/pci/emu10k1/memory.c index fcb04cbbc9ab..5cdffe2d31e1 100644 --- a/sound/pci/emu10k1/memory.c +++ b/sound/pci/emu10k1/memory.c @@ -457,49 +457,44 @@ static void get_single_page_range(struct snd_util_memhdr *hdr, *last_page_ret = last_page; }
-/* release allocated pages */ -static void __synth_free_pages(struct snd_emu10k1 *emu, int first_page, - int last_page) -{ - int page; - - for (page = first_page; page <= last_page; page++) { - free_page((unsigned long)emu->page_ptr_table[page]); - emu->page_addr_table[page] = 0; - emu->page_ptr_table[page] = NULL; - } -} - /* * allocate kernel pages */ static int synth_alloc_pages(struct snd_emu10k1 *emu, struct snd_emu10k1_memblk *blk) { int page, first_page, last_page; + struct snd_dma_buffer dmab;
emu10k1_memblk_init(blk); get_single_page_range(emu->memhdr, blk, &first_page, &last_page); /* allocate kernel pages */ for (page = first_page; page <= last_page; page++) { - /* first try to allocate from <4GB zone */ - struct page *p = alloc_page(GFP_KERNEL | GFP_DMA32 | - __GFP_NOWARN); - if (!p || (page_to_pfn(p) & ~(emu->dma_mask >> PAGE_SHIFT))) { - if (p) - __free_page(p); - /* try to allocate from <16MB zone */ - p = alloc_page(GFP_ATOMIC | GFP_DMA | - __GFP_NORETRY | /* no OOM-killer */ - __GFP_NOWARN); - } - if (!p) { - __synth_free_pages(emu, first_page, page - 1); - return -ENOMEM; + if (snd_dma_alloc_pages(SNDRV_DMA_TYPE_DEV, + snd_dma_pci_data(emu->pci), + PAGE_SIZE, &dmab) < 0) + goto __fail; + if (!is_valid_page(emu, dmab.addr)) { + snd_dma_free_pages(&dmab); + goto __fail; } - emu->page_addr_table[page] = page_to_phys(p); - emu->page_ptr_table[page] = page_address(p); + emu->page_addr_table[page] = dmab.addr; + emu->page_ptr_table[page] = dmab.area; } return 0; + +__fail: + /* release allocated pages */ + last_page = page - 1; + for (page = first_page; page <= last_page; page++) { + dmab.area = emu->page_ptr_table[page]; + dmab.addr = emu->page_addr_table[page]; + dmab.bytes = PAGE_SIZE; + snd_dma_free_pages(&dmab); + emu->page_addr_table[page] = 0; + emu->page_ptr_table[page] = NULL; + } + + return -ENOMEM; }
/* @@ -507,10 +502,23 @@ static int synth_alloc_pages(struct snd_emu10k1 *emu, struct snd_emu10k1_memblk */ static int synth_free_pages(struct snd_emu10k1 *emu, struct snd_emu10k1_memblk *blk) { - int first_page, last_page; + int page, first_page, last_page; + struct snd_dma_buffer dmab;
get_single_page_range(emu->memhdr, blk, &first_page, &last_page); - __synth_free_pages(emu, first_page, last_page); + dmab.dev.type = SNDRV_DMA_TYPE_DEV; + dmab.dev.dev = snd_dma_pci_data(emu->pci); + for (page = first_page; page <= last_page; page++) { + if (emu->page_ptr_table[page] == NULL) + continue; + dmab.area = emu->page_ptr_table[page]; + dmab.addr = emu->page_addr_table[page]; + dmab.bytes = PAGE_SIZE; + snd_dma_free_pages(&dmab); + emu->page_addr_table[page] = 0; + emu->page_ptr_table[page] = NULL; + } + return 0; }
On Sat, 27 Jan 2018 15:42:47 +0100, Maciej S. Szmigiero wrote:
Commit a5003fc04113 ("[ALSA] emu10k1 - simplify page allocation for synth") switched from using the DMA allocator for synth DMA pages to manually calling alloc_page(). However, this usage has an implicit assumption that the DMA address space for the emu10k1-family chip is the same as the CPU physical address space which is not true for a system with a IOMMU.
Since this made the synth part of the driver non-functional on such systems let's effectively revert that commit.
Let's keep (or re-add after this revert) the simplification with __synth_free_pages(). Other than that I have no objection to this change.
thanks,
Takashi
Signed-off-by: Maciej S. Szmigiero mail@maciej.szmigiero.name
Changes from v1: None in this patch.
sound/pci/emu10k1/memory.c | 70 ++++++++++++++++++++++++++-------------------- 1 file changed, 39 insertions(+), 31 deletions(-)
diff --git a/sound/pci/emu10k1/memory.c b/sound/pci/emu10k1/memory.c index fcb04cbbc9ab..5cdffe2d31e1 100644 --- a/sound/pci/emu10k1/memory.c +++ b/sound/pci/emu10k1/memory.c @@ -457,49 +457,44 @@ static void get_single_page_range(struct snd_util_memhdr *hdr, *last_page_ret = last_page; }
-/* release allocated pages */ -static void __synth_free_pages(struct snd_emu10k1 *emu, int first_page,
int last_page)
-{
- int page;
- for (page = first_page; page <= last_page; page++) {
free_page((unsigned long)emu->page_ptr_table[page]);
emu->page_addr_table[page] = 0;
emu->page_ptr_table[page] = NULL;
- }
-}
/*
- allocate kernel pages
*/ static int synth_alloc_pages(struct snd_emu10k1 *emu, struct snd_emu10k1_memblk *blk) { int page, first_page, last_page;
struct snd_dma_buffer dmab;
emu10k1_memblk_init(blk); get_single_page_range(emu->memhdr, blk, &first_page, &last_page); /* allocate kernel pages */ for (page = first_page; page <= last_page; page++) {
/* first try to allocate from <4GB zone */
struct page *p = alloc_page(GFP_KERNEL | GFP_DMA32 |
__GFP_NOWARN);
if (!p || (page_to_pfn(p) & ~(emu->dma_mask >> PAGE_SHIFT))) {
if (p)
__free_page(p);
/* try to allocate from <16MB zone */
p = alloc_page(GFP_ATOMIC | GFP_DMA |
__GFP_NORETRY | /* no OOM-killer */
__GFP_NOWARN);
}
if (!p) {
__synth_free_pages(emu, first_page, page - 1);
return -ENOMEM;
if (snd_dma_alloc_pages(SNDRV_DMA_TYPE_DEV,
snd_dma_pci_data(emu->pci),
PAGE_SIZE, &dmab) < 0)
goto __fail;
if (!is_valid_page(emu, dmab.addr)) {
snd_dma_free_pages(&dmab);
}goto __fail;
emu->page_addr_table[page] = page_to_phys(p);
emu->page_ptr_table[page] = page_address(p);
emu->page_addr_table[page] = dmab.addr;
} return 0;emu->page_ptr_table[page] = dmab.area;
+__fail:
- /* release allocated pages */
- last_page = page - 1;
- for (page = first_page; page <= last_page; page++) {
dmab.area = emu->page_ptr_table[page];
dmab.addr = emu->page_addr_table[page];
dmab.bytes = PAGE_SIZE;
snd_dma_free_pages(&dmab);
emu->page_addr_table[page] = 0;
emu->page_ptr_table[page] = NULL;
- }
- return -ENOMEM;
}
/* @@ -507,10 +502,23 @@ static int synth_alloc_pages(struct snd_emu10k1 *emu, struct snd_emu10k1_memblk */ static int synth_free_pages(struct snd_emu10k1 *emu, struct snd_emu10k1_memblk *blk) {
- int first_page, last_page;
int page, first_page, last_page;
struct snd_dma_buffer dmab;
get_single_page_range(emu->memhdr, blk, &first_page, &last_page);
- __synth_free_pages(emu, first_page, last_page);
- dmab.dev.type = SNDRV_DMA_TYPE_DEV;
- dmab.dev.dev = snd_dma_pci_data(emu->pci);
- for (page = first_page; page <= last_page; page++) {
if (emu->page_ptr_table[page] == NULL)
continue;
dmab.area = emu->page_ptr_table[page];
dmab.addr = emu->page_addr_table[page];
dmab.bytes = PAGE_SIZE;
snd_dma_free_pages(&dmab);
emu->page_addr_table[page] = 0;
emu->page_ptr_table[page] = NULL;
- }
- return 0;
}
On 12.02.2018 13:44, Takashi Iwai wrote:
On Sat, 27 Jan 2018 15:42:47 +0100, Maciej S. Szmigiero wrote:
Commit a5003fc04113 ("[ALSA] emu10k1 - simplify page allocation for synth") switched from using the DMA allocator for synth DMA pages to manually calling alloc_page(). However, this usage has an implicit assumption that the DMA address space for the emu10k1-family chip is the same as the CPU physical address space which is not true for a system with a IOMMU.
Since this made the synth part of the driver non-functional on such systems let's effectively revert that commit.
Let's keep (or re-add after this revert) the simplification with __synth_free_pages(). Other than that I have no objection to this change.
Will keep the simplification via __synth_free_pages() function in a respin.
thanks,
Takashi
Thanks, Maciej
participants (2)
-
Maciej S. Szmigiero
-
Takashi Iwai