[kernel] x86/swiotlb/amd: Half the size if allocation failed

Message ID 20221027052607.260234-1-aik@amd.com
State New
Headers
Series [kernel] x86/swiotlb/amd: Half the size if allocation failed |

Commit Message

Alexey Kardashevskiy Oct. 27, 2022, 5:26 a.m. UTC
  At the moment the AMD encrypted platform reserves 6% of RAM for SWIOTLB
or 1GB, whichever is less. However it is possible that there is no block
big enough in the low memory which make SWIOTLB allocation fail and
the kernel continues without DMA. In such case a VM hangs on DMA.

This divides the size in half and tries again reusing the existing
remapping logic.

This updates default_nslabs on successful allocation which looks like
an oversight as not doing so should have broken callers of
swiotlb_size_or_default().

Signed-off-by: Alexey Kardashevskiy <aik@amd.com>
--

I hit the problem with
QEMU's "-m 16819M" where SWIOTLB was adjusted to
0x7e200 == 1,058,013,184 (slightly less than 1GB) while
0x7e180 still worked.

With guest errors enabled, there are many unassigned accesses from
virtio.

---
 kernel/dma/swiotlb.c | 20 +++++++++++++-------
 1 file changed, 13 insertions(+), 7 deletions(-)
  

Comments

Thadeu Lima de Souza Cascardo Oct. 27, 2022, 4:46 p.m. UTC | #1
On Thu, Oct 27, 2022 at 04:26:07PM +1100, Alexey Kardashevskiy wrote:
> At the moment the AMD encrypted platform reserves 6% of RAM for SWIOTLB
> or 1GB, whichever is less. However it is possible that there is no block
> big enough in the low memory which make SWIOTLB allocation fail and
> the kernel continues without DMA. In such case a VM hangs on DMA.
> 
> This divides the size in half and tries again reusing the existing
> remapping logic.
> 
> This updates default_nslabs on successful allocation which looks like
> an oversight as not doing so should have broken callers of
> swiotlb_size_or_default().
> 
> Signed-off-by: Alexey Kardashevskiy <aik@amd.com>

Should this have a
Fixes: e998879d4fb7 ("x86,swiotlb: Adjust SWIOTLB bounce buffer size for SEV guests")
?

Cascardo.

> --
> 
> I hit the problem with
> QEMU's "-m 16819M" where SWIOTLB was adjusted to
> 0x7e200 == 1,058,013,184 (slightly less than 1GB) while
> 0x7e180 still worked.
> 
> With guest errors enabled, there are many unassigned accesses from
> virtio.
> 
> ---
>  kernel/dma/swiotlb.c | 20 +++++++++++++-------
>  1 file changed, 13 insertions(+), 7 deletions(-)
> 
> diff --git a/kernel/dma/swiotlb.c b/kernel/dma/swiotlb.c
> index 339a990554e7..d28c294320fd 100644
> --- a/kernel/dma/swiotlb.c
> +++ b/kernel/dma/swiotlb.c
> @@ -338,21 +338,27 @@ void __init swiotlb_init_remap(bool addressing_limit, unsigned int flags,
>  	else
>  		tlb = memblock_alloc_low(bytes, PAGE_SIZE);
>  	if (!tlb) {
> -		pr_warn("%s: failed to allocate tlb structure\n", __func__);
> -		return;
> -	}
> -
> -	if (remap && remap(tlb, nslabs) < 0) {
> +		pr_warn("%s: Failed to allocate %zu bytes tlb structure\n",
> +			__func__, bytes);
> +	} else if (remap && remap(tlb, nslabs) < 0) {
>  		memblock_free(tlb, PAGE_ALIGN(bytes));
> +		tlb = NULL;
> +		pr_warn("%s: Failed to remap %zu bytes\n", __func__, bytes);
> +	}
>  
> +	if (!tlb) {
>  		nslabs = ALIGN(nslabs >> 1, IO_TLB_SEGSIZE);
>  		if (nslabs >= IO_TLB_MIN_SLABS)
>  			goto retry;
> -
> -		pr_warn("%s: Failed to remap %zu bytes\n", __func__, bytes);
>  		return;
>  	}
>  
> +	if (default_nslabs != nslabs) {
> +		pr_info("SWIOTLB bounce buffer size adjusted %lu -> %lu slabs",
> +			default_nslabs, nslabs);
> +		default_nslabs = nslabs;
> +	}
> +
>  	alloc_size = PAGE_ALIGN(array_size(sizeof(*mem->slots), nslabs));
>  	mem->slots = memblock_alloc(alloc_size, PAGE_SIZE);
>  	if (!mem->slots) {
> -- 
> 2.37.3
>
  
Alexey Kardashevskiy Oct. 28, 2022, 1:14 a.m. UTC | #2
On 28/10/22 03:46, Thadeu Lima de Souza Cascardo wrote:
> On Thu, Oct 27, 2022 at 04:26:07PM +1100, Alexey Kardashevskiy wrote:
>> At the moment the AMD encrypted platform reserves 6% of RAM for SWIOTLB
>> or 1GB, whichever is less. However it is possible that there is no block
>> big enough in the low memory which make SWIOTLB allocation fail and
>> the kernel continues without DMA. In such case a VM hangs on DMA.
>>
>> This divides the size in half and tries again reusing the existing
>> remapping logic.
>>
>> This updates default_nslabs on successful allocation which looks like
>> an oversight as not doing so should have broken callers of
>> swiotlb_size_or_default().
>>
>> Signed-off-by: Alexey Kardashevskiy <aik@amd.com>
> 
> Should this have a
> Fixes: e998879d4fb7 ("x86,swiotlb: Adjust SWIOTLB bounce buffer size for SEV guests")
> ?


Well, the problem was there before that patch, the allocation failure 
was not handled while remap failure was. e998879d4fb7 just made it 
easier to see. But still worth mentioning I guess... Thanks,


> 
> Cascardo.
> 
>> --
>>
>> I hit the problem with
>> QEMU's "-m 16819M" where SWIOTLB was adjusted to
>> 0x7e200 == 1,058,013,184 (slightly less than 1GB) while
>> 0x7e180 still worked.
>>
>> With guest errors enabled, there are many unassigned accesses from
>> virtio.
>>
>> ---
>>   kernel/dma/swiotlb.c | 20 +++++++++++++-------
>>   1 file changed, 13 insertions(+), 7 deletions(-)
>>
>> diff --git a/kernel/dma/swiotlb.c b/kernel/dma/swiotlb.c
>> index 339a990554e7..d28c294320fd 100644
>> --- a/kernel/dma/swiotlb.c
>> +++ b/kernel/dma/swiotlb.c
>> @@ -338,21 +338,27 @@ void __init swiotlb_init_remap(bool addressing_limit, unsigned int flags,
>>   	else
>>   		tlb = memblock_alloc_low(bytes, PAGE_SIZE);
>>   	if (!tlb) {
>> -		pr_warn("%s: failed to allocate tlb structure\n", __func__);
>> -		return;
>> -	}
>> -
>> -	if (remap && remap(tlb, nslabs) < 0) {
>> +		pr_warn("%s: Failed to allocate %zu bytes tlb structure\n",
>> +			__func__, bytes);
>> +	} else if (remap && remap(tlb, nslabs) < 0) {
>>   		memblock_free(tlb, PAGE_ALIGN(bytes));
>> +		tlb = NULL;
>> +		pr_warn("%s: Failed to remap %zu bytes\n", __func__, bytes);
>> +	}
>>   
>> +	if (!tlb) {
>>   		nslabs = ALIGN(nslabs >> 1, IO_TLB_SEGSIZE);
>>   		if (nslabs >= IO_TLB_MIN_SLABS)
>>   			goto retry;
>> -
>> -		pr_warn("%s: Failed to remap %zu bytes\n", __func__, bytes);
>>   		return;
>>   	}
>>   
>> +	if (default_nslabs != nslabs) {
>> +		pr_info("SWIOTLB bounce buffer size adjusted %lu -> %lu slabs",
>> +			default_nslabs, nslabs);
>> +		default_nslabs = nslabs;
>> +	}
>> +
>>   	alloc_size = PAGE_ALIGN(array_size(sizeof(*mem->slots), nslabs));
>>   	mem->slots = memblock_alloc(alloc_size, PAGE_SIZE);
>>   	if (!mem->slots) {
>> -- 
>> 2.37.3
>>
  
Tom Lendacky Oct. 28, 2022, 2:13 p.m. UTC | #3
On 10/27/22 00:26, Alexey Kardashevskiy wrote:
> At the moment the AMD encrypted platform reserves 6% of RAM for SWIOTLB
> or 1GB, whichever is less. However it is possible that there is no block
> big enough in the low memory which make SWIOTLB allocation fail and
> the kernel continues without DMA. In such case a VM hangs on DMA.
> 
> This divides the size in half and tries again reusing the existing
> remapping logic.
> 
> This updates default_nslabs on successful allocation which looks like
> an oversight as not doing so should have broken callers of
> swiotlb_size_or_default().
> 
> Signed-off-by: Alexey Kardashevskiy <aik@amd.com>

Reviewed-by: Tom Lendacky <thomas.lendacky@amd.com>

> --
> 
> I hit the problem with
> QEMU's "-m 16819M" where SWIOTLB was adjusted to
> 0x7e200 == 1,058,013,184 (slightly less than 1GB) while
> 0x7e180 still worked.
> 
> With guest errors enabled, there are many unassigned accesses from
> virtio.
> 
> ---
>   kernel/dma/swiotlb.c | 20 +++++++++++++-------
>   1 file changed, 13 insertions(+), 7 deletions(-)
> 
> diff --git a/kernel/dma/swiotlb.c b/kernel/dma/swiotlb.c
> index 339a990554e7..d28c294320fd 100644
> --- a/kernel/dma/swiotlb.c
> +++ b/kernel/dma/swiotlb.c
> @@ -338,21 +338,27 @@ void __init swiotlb_init_remap(bool addressing_limit, unsigned int flags,
>   	else
>   		tlb = memblock_alloc_low(bytes, PAGE_SIZE);
>   	if (!tlb) {
> -		pr_warn("%s: failed to allocate tlb structure\n", __func__);
> -		return;
> -	}
> -
> -	if (remap && remap(tlb, nslabs) < 0) {
> +		pr_warn("%s: Failed to allocate %zu bytes tlb structure\n",
> +			__func__, bytes);
> +	} else if (remap && remap(tlb, nslabs) < 0) {
>   		memblock_free(tlb, PAGE_ALIGN(bytes));
> +		tlb = NULL;
> +		pr_warn("%s: Failed to remap %zu bytes\n", __func__, bytes);
> +	}
>   
> +	if (!tlb) {
>   		nslabs = ALIGN(nslabs >> 1, IO_TLB_SEGSIZE);
>   		if (nslabs >= IO_TLB_MIN_SLABS)
>   			goto retry;
> -
> -		pr_warn("%s: Failed to remap %zu bytes\n", __func__, bytes);
>   		return;
>   	}
>   
> +	if (default_nslabs != nslabs) {
> +		pr_info("SWIOTLB bounce buffer size adjusted %lu -> %lu slabs",
> +			default_nslabs, nslabs);
> +		default_nslabs = nslabs;
> +	}
> +
>   	alloc_size = PAGE_ALIGN(array_size(sizeof(*mem->slots), nslabs));
>   	mem->slots = memblock_alloc(alloc_size, PAGE_SIZE);
>   	if (!mem->slots) {
  
Christoph Hellwig Oct. 30, 2022, 8:05 a.m. UTC | #4
The subject looks wrong - this just touches the swiotlb code, and
nothing x86 or AMD specific.

The code flow now looks a little confusing.  Can yu just add a little
swiotlb_memblock_alloc helper that does the two variants of memblock
allocation and remap, so that we can keep a simply
loop around them in the caller?
  

Patch

diff --git a/kernel/dma/swiotlb.c b/kernel/dma/swiotlb.c
index 339a990554e7..d28c294320fd 100644
--- a/kernel/dma/swiotlb.c
+++ b/kernel/dma/swiotlb.c
@@ -338,21 +338,27 @@  void __init swiotlb_init_remap(bool addressing_limit, unsigned int flags,
 	else
 		tlb = memblock_alloc_low(bytes, PAGE_SIZE);
 	if (!tlb) {
-		pr_warn("%s: failed to allocate tlb structure\n", __func__);
-		return;
-	}
-
-	if (remap && remap(tlb, nslabs) < 0) {
+		pr_warn("%s: Failed to allocate %zu bytes tlb structure\n",
+			__func__, bytes);
+	} else if (remap && remap(tlb, nslabs) < 0) {
 		memblock_free(tlb, PAGE_ALIGN(bytes));
+		tlb = NULL;
+		pr_warn("%s: Failed to remap %zu bytes\n", __func__, bytes);
+	}
 
+	if (!tlb) {
 		nslabs = ALIGN(nslabs >> 1, IO_TLB_SEGSIZE);
 		if (nslabs >= IO_TLB_MIN_SLABS)
 			goto retry;
-
-		pr_warn("%s: Failed to remap %zu bytes\n", __func__, bytes);
 		return;
 	}
 
+	if (default_nslabs != nslabs) {
+		pr_info("SWIOTLB bounce buffer size adjusted %lu -> %lu slabs",
+			default_nslabs, nslabs);
+		default_nslabs = nslabs;
+	}
+
 	alloc_size = PAGE_ALIGN(array_size(sizeof(*mem->slots), nslabs));
 	mem->slots = memblock_alloc(alloc_size, PAGE_SIZE);
 	if (!mem->slots) {