Re: [PATCH v7u1 26/31] x86: Don't enable swiotlb if there is notenough ram for it

[Shuah Khan]

On Mon, Jan 7, 2013 at 8:26 AM, Konrad Rzeszutek Wilk
<konrad.wilk@xxxxxxxxxx> wrote:
> On Fri, Jan 04, 2013 at 02:10:25PM -0800, Yinghai Lu wrote:
>> On Fri, Jan 4, 2013 at 1:02 PM, Shuah Khan <shuahkhan@xxxxxxxxx> wrote:
>> > Pani'cing the system doesn't sound like a good option to me in this
>> > case. This change to disable swiotlb is made for kdump. However, with
>> > this change several system fail to boot, unless crashkernel_low=72M is
>> > specified.
>>
>> this patchset is new feature to put second kdump kernel above 4G.
>>
>> >
>> > I would the say the right approach to solve this would be to not
>> > change the current pci_swiotlb_detect_override() behavior and treat
>> > swiotlb =1 upon entry equivalent to swiotlb_force set.
>>
>> that will make intel system have to take crashkernel_low=72M too.
>> otherwise intel system will get panic during swiotlb allocation.
>
> Two things:
>
>  1). You need to wrap the 'is_enough_..' in CONFIG_KEXEC, which means
>     that the function needs to go in a header file.
>  2). The check for 1MB is suspect. Why only 1MB? You mentioned it is
>      b/c of crashkernel_low=72M (which I am not seeing in v3.8 kernel-parameters.txt?
>      Is that part of your mega-patchset?). Anyhow, there seems to be a disconnect -
>      what if the user supplied crashkernel_low=27M? Perhaps the 'is_enough'
>      should also parse the bootparams to double-check that there is enough
>      low-mem space? But then if the kernel grows then 72M might not be enough -
>      you might need 82M with 3.9.
>
>      Perhaps a better way for this is to do:
>         1). Change 'is_enough' to check only for 4MB.
>         2). When booting as kexec, the SWIOTLB would only use 4MB instead of 64MB?
>
>      Or, we could also use the post-late SWIOTLB initialization similiary to how it was
>      done on ia64. This would mean that the AMD VI code would just call the
>      .. something like this - NOT tested or even compile tested:
>
> diff --git a/drivers/iommu/amd_iommu.c b/drivers/iommu/amd_iommu.c
> index c1c74e0..e7fa8f7 100644
> --- a/drivers/iommu/amd_iommu.c
> +++ b/drivers/iommu/amd_iommu.c
> @@ -3173,6 +3173,24 @@ int __init amd_iommu_init_dma_ops(void)
>         if (unhandled && max_pfn > MAX_DMA32_PFN) {
>                 /* There are unhandled devices - initialize swiotlb for them */
>                 swiotlb = 1;
> +               /* Late (so no bootmem allocator) usage and only if the early SWIOTLB
> +                * hadn't been allocated (which can happen on kexec kernels booted
> +                * above 4GB). */
> +               if (!swiotlb_nr_tbl()) {
> +                       int retry = 3;
> +                       int mb_size = 64;
> +                       int rc = 0;
> +retry_me:
> +                       if (retry < 0)
> +                               panic("We tried setting %dMB for SWIOTLB but got -ENOMEM", mb_size << 1);
> +                       rc = swiotlb_late_init_with_default_size(mb_size * (1<<20));
> +                       if (rc) {
> +                               retry --;
> +                               mb_size >> 1;
> +                               goto retry_me;
> +                       }
> +                       dma_ops = &swiotlb_dma_ops;
> +               }
>         }
>
>         amd_iommu_stats_init();
>
> And then the early SWIOTLB initialization for 64MB can fail and we are still OK.
>>

Yinghai/Konrad,

Did more testing. btw this patch depends on your [v7u1,25/31]
memblock: add memblock_mem_size(). Here are the test results:

1. When there is not enough memory: (enough_mem_for_swiotlb() returns false)
system will panic in amd_iommu_init_dma_ops().

2. When there is enough memory: (enough_mem_for_swiotlb() returns true):
swiotlb is reserved
pci_swiotlb_late_init() leaves the buffer allocated since swiotlb=1
with that getting changed in amd_iommu_init_dma_ops().

I agree with Konrad that the logic should be wrapped in CONFIG_KEXEC.

Also, since IOMMU drivers can no longer assume swiotlb is allocated
enough_mem_for_swiotlb() check fails, AMD IOMMU or another other iommu
driver can't simply rely on changing swiotlb=1 and assuming the buffer
is there.

As Konrad suggested,  a hook is needed, however, I think the logic to
ensure switolb buffer belongs in swiotlb modules. How about changing
pci_swiolb_late_init() logic to ensure swioltb late init is done
instead of leaving it up to AMD IOMMU driver or some other driver.

The logic to update dma_ops doesn't really belong in
amd_iommu_init_dma_ops() anyways.

-- Shuah
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/