| Message ID | 20230905194849.v1.2.Id3ab7cf665bcead097654937233a645722a4cce3@changeid (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | Clean-up arm-smmu-v3-sva.c: remove arm_smmu_bond | expand |
On Tue, Sep 05, 2023 at 07:49:13PM +0800, Michael Shavit wrote: > Always allocate a new arm_smmu_bond in __arm_smmu_sva_bind and remove > the bond refcount since arm_smmu_bond can never be shared across calls > to __arm_smmu_sva_bind. > > The iommu framework will not allocate multiple SVA domains for the same > (device/mm) pair, nor will it call set_dev_pasid for a device if a > domain is already attached on the given pasid. There's also a one-to-one > mapping between MM and PASID. __arm_smmu_sva_bind is therefore never > called with the same (device/mm) pair, and so there's no reason to try > and normalize allocations of the arm_smmu_bond struct for a (device/mm) > pair across set_dev_pasid. > > Signed-off-by: Michael Shavit <mshavit@google.com> > --- > Note that this is true today because iommu_sva_bind_device calls > iommu_get_domain_for_dev_pasid to elude the iommu_attach_device_pasid if > a domain is already attached. > But even with Tina's patch series where iommu_get_domain_for_dev_pasid > is no longer used, iommu_attach_device_pasid also checks whether a > domain is already attached in the group's pasid_array. > > drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3-sva.c | 12 +----------- > 1 file changed, 1 insertion(+), 11 deletions(-) Reviewed-by: Jason Gunthorpe <jgg@nvidia.com> Jason
diff --git a/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3-sva.c b/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3-sva.c index 32784758ccce6..9fb6907c5e7d4 100644 --- a/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3-sva.c +++ b/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3-sva.c @@ -28,7 +28,6 @@ struct arm_smmu_bond { struct mm_struct *mm; struct arm_smmu_mmu_notifier *smmu_mn; struct list_head list; - refcount_t refs; }; #define sva_to_bond(handle) \ @@ -330,20 +329,11 @@ static int __arm_smmu_sva_bind(struct device *dev, struct mm_struct *mm) if (!master || !master->sva_enabled) return -ENODEV; - /* If bind() was already called for this {dev, mm} pair, reuse it. */ - list_for_each_entry(bond, &master->bonds, list) { - if (bond->mm == mm) { - refcount_inc(&bond->refs); - return &bond->sva; - } - } - bond = kzalloc(sizeof(*bond), GFP_KERNEL); if (!bond) return -ENOMEM; bond->mm = mm; - refcount_set(&bond->refs, 1); bond->smmu_mn = arm_smmu_mmu_notifier_get(smmu_domain, mm); if (IS_ERR(bond->smmu_mn)) { @@ -522,7 +512,7 @@ void arm_smmu_sva_remove_dev_pasid(struct iommu_domain *domain, } } - if (!WARN_ON(!bond) && refcount_dec_and_test(&bond->refs)) { + if (!WARN_ON(!bond)) { list_del(&bond->list); arm_smmu_mmu_notifier_put(bond->smmu_mn); kfree(bond);
Always allocate a new arm_smmu_bond in __arm_smmu_sva_bind and remove the bond refcount since arm_smmu_bond can never be shared across calls to __arm_smmu_sva_bind. The iommu framework will not allocate multiple SVA domains for the same (device/mm) pair, nor will it call set_dev_pasid for a device if a domain is already attached on the given pasid. There's also a one-to-one mapping between MM and PASID. __arm_smmu_sva_bind is therefore never called with the same (device/mm) pair, and so there's no reason to try and normalize allocations of the arm_smmu_bond struct for a (device/mm) pair across set_dev_pasid. Signed-off-by: Michael Shavit <mshavit@google.com> --- Note that this is true today because iommu_sva_bind_device calls iommu_get_domain_for_dev_pasid to elude the iommu_attach_device_pasid if a domain is already attached. But even with Tina's patch series where iommu_get_domain_for_dev_pasid is no longer used, iommu_attach_device_pasid also checks whether a domain is already attached in the group's pasid_array. drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3-sva.c | 12 +----------- 1 file changed, 1 insertion(+), 11 deletions(-)