[PATCH 3/4] iommu/arm-smmu-v3: Use NUMA memory allocations for stream tables and comamnd queues
Robin Murphy
robin.murphy at arm.com
Wed Oct 18 06:36:03 PDT 2017
On 04/10/17 14:53, Ganapatrao Kulkarni wrote:
> Hi Robin,
>
>
> On Thu, Sep 21, 2017 at 5:28 PM, Robin Murphy <robin.murphy at arm.com> wrote:
>> [+Christoph and Marek]
>>
>> On 21/09/17 09:59, Ganapatrao Kulkarni wrote:
>>> Introduce smmu_alloc_coherent and smmu_free_coherent functions to
>>> allocate/free dma coherent memory from NUMA node associated with SMMU.
>>> Replace all calls of dmam_alloc_coherent with smmu_alloc_coherent
>>> for SMMU stream tables and command queues.
>>
>> This doesn't work - not only do you lose the 'managed' aspect and risk
>> leaking various tables on probe failure or device removal, but more
>> importantly, unless you add DMA syncs around all the CPU accesses to the
>> tables, you lose the critical 'coherent' aspect, and that's a horribly
>> invasive change that I really don't want to make.
>
> this implementation is similar to function used to allocate memory for
> translation tables.
The concept is similar, yes, and would work if implemented *correctly*
with the aforementioned comprehensive and hugely invasive changes. The
implementation as presented in this patch, however, is incomplete and
badly broken.
By way of comparison, the io-pgtable implementations contain all the
necessary dma_sync_* calls, never relied on devres, and only have one
DMA direction to worry about (hint: the queues don't all work
identically). There are also a couple of practical reasons for using
streaming mappings with the DMA == phys restriction there - tracking
both the CPU and DMA addresses for each table would significantly
increase the memory overhead, and using the cacheable linear map address
in all cases sidesteps any potential problems with the atomic PTE
updates. Neither of those concerns apply to the SMMUv3 data structures,
which are textbook coherent DMA allocations (being tied to the lifetime
of the device, rather than transient).
> why do you see it affects to stream tables and not to page tables.
> at runtime, both tables are accessed by SMMU only.
>
> As said in cover letter, having stream table from respective NUMA node
> is yielding
> around 30% performance!
> please suggest, if there is any better way to address this issue?
I fully agree that NUMA-aware allocations are a worthwhile thing that we
want. I just don't like the idea of going around individual drivers
replacing coherent API usage with bodged-up streaming mappings - I
really think it's worth making the effort to to tackle it once, in the
proper place, in a way that benefits all users together.
Robin.
>>
>> Christoph, Marek; how reasonable do you think it is to expect
>> dma_alloc_coherent() to be inherently NUMA-aware on NUMA-capable
>> systems? SWIOTLB looks fairly straightforward to fix up (for the simple
>> allocation case; I'm not sure it's even worth it for bounce-buffering),
>> but the likes of CMA might be a little trickier...
>>
>> Robin.
>>
>>> Signed-off-by: Ganapatrao Kulkarni <ganapatrao.kulkarni at cavium.com>
>>> ---
>>> drivers/iommu/arm-smmu-v3.c | 57 ++++++++++++++++++++++++++++++++++++++++-----
>>> 1 file changed, 51 insertions(+), 6 deletions(-)
>>>
>>> diff --git a/drivers/iommu/arm-smmu-v3.c b/drivers/iommu/arm-smmu-v3.c
>>> index e67ba6c..bc4ba1f 100644
>>> --- a/drivers/iommu/arm-smmu-v3.c
>>> +++ b/drivers/iommu/arm-smmu-v3.c
>>> @@ -1158,6 +1158,50 @@ static void arm_smmu_init_bypass_stes(u64 *strtab, unsigned int nent)
>>> }
>>> }
>>>
>>> +static void *smmu_alloc_coherent(struct arm_smmu_device *smmu, size_t size,
>>> + dma_addr_t *dma_handle, gfp_t gfp)
>>> +{
>>> + struct device *dev = smmu->dev;
>>> + void *pages;
>>> + dma_addr_t dma;
>>> + int numa_node = dev_to_node(dev);
>>> +
>>> + pages = alloc_pages_exact_nid(numa_node, size, gfp | __GFP_ZERO);
>>> + if (!pages)
>>> + return NULL;
>>> +
>>> + if (!(smmu->features & ARM_SMMU_FEAT_COHERENCY)) {
>>> + dma = dma_map_single(dev, pages, size, DMA_TO_DEVICE);
>>> + if (dma_mapping_error(dev, dma))
>>> + goto out_free;
>>> + /*
>>> + * We depend on the SMMU being able to work with any physical
>>> + * address directly, so if the DMA layer suggests otherwise by
>>> + * translating or truncating them, that bodes very badly...
>>> + */
>>> + if (dma != virt_to_phys(pages))
>>> + goto out_unmap;
>>> + }
>>> +
>>> + *dma_handle = (dma_addr_t)virt_to_phys(pages);
>>> + return pages;
>>> +
>>> +out_unmap:
>>> + dev_err(dev, "Cannot accommodate DMA translation for IOMMU page tables\n");
>>> + dma_unmap_single(dev, dma, size, DMA_TO_DEVICE);
>>> +out_free:
>>> + free_pages_exact(pages, size);
>>> + return NULL;
>>> +}
>>> +
>>> +static void smmu_free_coherent(struct arm_smmu_device *smmu, size_t size,
>>> + void *pages, dma_addr_t dma_handle)
>>> +{
>>> + if (!(smmu->features & ARM_SMMU_FEAT_COHERENCY))
>>> + dma_unmap_single(smmu->dev, dma_handle, size, DMA_TO_DEVICE);
>>> + free_pages_exact(pages, size);
>>> +}
>>> +
>>> static int arm_smmu_init_l2_strtab(struct arm_smmu_device *smmu, u32 sid)
>>> {
>>> size_t size;
>>> @@ -1172,7 +1216,7 @@ static int arm_smmu_init_l2_strtab(struct arm_smmu_device *smmu, u32 sid)
>>> strtab = &cfg->strtab[(sid >> STRTAB_SPLIT) * STRTAB_L1_DESC_DWORDS];
>>>
>>> desc->span = STRTAB_SPLIT + 1;
>>> - desc->l2ptr = dmam_alloc_coherent(smmu->dev, size, &desc->l2ptr_dma,
>>> + desc->l2ptr = smmu_alloc_coherent(smmu, size, &desc->l2ptr_dma,
>>> GFP_KERNEL | __GFP_ZERO);
>>> if (!desc->l2ptr) {
>>> dev_err(smmu->dev,
>>> @@ -1487,7 +1531,7 @@ static void arm_smmu_domain_free(struct iommu_domain *domain)
>>> struct arm_smmu_s1_cfg *cfg = &smmu_domain->s1_cfg;
>>>
>>> if (cfg->cdptr) {
>>> - dmam_free_coherent(smmu_domain->smmu->dev,
>>> + smmu_free_coherent(smmu,
>>> CTXDESC_CD_DWORDS << 3,
>>> cfg->cdptr,
>>> cfg->cdptr_dma);
>>> @@ -1515,7 +1559,7 @@ static int arm_smmu_domain_finalise_s1(struct arm_smmu_domain *smmu_domain,
>>> if (asid < 0)
>>> return asid;
>>>
>>> - cfg->cdptr = dmam_alloc_coherent(smmu->dev, CTXDESC_CD_DWORDS << 3,
>>> + cfg->cdptr = smmu_alloc_coherent(smmu, CTXDESC_CD_DWORDS << 3,
>>> &cfg->cdptr_dma,
>>> GFP_KERNEL | __GFP_ZERO);
>>> if (!cfg->cdptr) {
>>> @@ -1984,7 +2028,7 @@ static int arm_smmu_init_one_queue(struct arm_smmu_device *smmu,
>>> {
>>> size_t qsz = ((1 << q->max_n_shift) * dwords) << 3;
>>>
>>> - q->base = dmam_alloc_coherent(smmu->dev, qsz, &q->base_dma, GFP_KERNEL);
>>> + q->base = smmu_alloc_coherent(smmu, qsz, &q->base_dma, GFP_KERNEL);
>>> if (!q->base) {
>>> dev_err(smmu->dev, "failed to allocate queue (0x%zx bytes)\n",
>>> qsz);
>>> @@ -2069,7 +2113,7 @@ static int arm_smmu_init_strtab_2lvl(struct arm_smmu_device *smmu)
>>> size, smmu->sid_bits);
>>>
>>> l1size = cfg->num_l1_ents * (STRTAB_L1_DESC_DWORDS << 3);
>>> - strtab = dmam_alloc_coherent(smmu->dev, l1size, &cfg->strtab_dma,
>>> + strtab = smmu_alloc_coherent(smmu, l1size, &cfg->strtab_dma,
>>> GFP_KERNEL | __GFP_ZERO);
>>> if (!strtab) {
>>> dev_err(smmu->dev,
>>> @@ -2097,8 +2141,9 @@ static int arm_smmu_init_strtab_linear(struct arm_smmu_device *smmu)
>>> u32 size;
>>> struct arm_smmu_strtab_cfg *cfg = &smmu->strtab_cfg;
>>>
>>> +
>>> size = (1 << smmu->sid_bits) * (STRTAB_STE_DWORDS << 3);
>>> - strtab = dmam_alloc_coherent(smmu->dev, size, &cfg->strtab_dma,
>>> + strtab = smmu_alloc_coherent(smmu, size, &cfg->strtab_dma,
>>> GFP_KERNEL | __GFP_ZERO);
>>> if (!strtab) {
>>> dev_err(smmu->dev,
>>>
>>
>
> thanks
> Ganapat
>
More information about the linux-arm-kernel
mailing list