Skip to contentSkip to conversation

iommu: Add RCU-protected page free support

RFC PATCH In review Lu Baolu 4 participants 12 messages Jun 9, 2022 – Jun 20, 2022

Subsystems (4)
lore
Share & tools
lore

Lu Baolu

author maintainer
#0 lore raw

RFC PATCH 1/2 iommu: Add RCU-protected page free support

The IOMMU page tables are updated using iommu_map/unmap() interfaces. Currently, there is no mandatory requirement for drivers to use locks to ensure concurrent updates to page tables, because it's assumed that overlapping IOVA ranges do not have concurrent updates. Therefore the IOMMU drivers only need to take care of concurrent updates to level page table entries.

But enabling new features challenges this assumption. For example, the hardware assisted dirty page tracking feature requires scanning page tables in interfaces other than mapping and unmapping. This might result in a use-after-free scenario in which a level page table has been freed by the unmap() interface, while another thread is scanning the next level page table.

This adds RCU-protected page free support so that the pages are really freed and reused after a RCU grace period. Hence, the page tables are safe for scanning within a rcu_read_lock critical region. Considering that scanning the page table is a rare case, this also adds a domain flag and the RCU-protected page free is only used when this flat is set.

---
 include/linux/iommu.h |  9 +++++++++
 drivers/iommu/iommu.c | 23 +++++++++++++++++++++++
 2 files changed, 32 insertions(+)
include/linux/iommu.h +9 -0
index 5e1afe169549..6f68eabb8567 100644
@@ -95,6 +95,7 @@ struct iommu_domain {
 	void *handler_token;
 	struct iommu_domain_geometry geometry;
 	struct iommu_dma_cookie *iova_cookie;
+	unsigned long concurrent_traversal:1;
 };
 
 static inline bool iommu_is_dma_domain(struct iommu_domain *domain)
@@ -657,6 +658,12 @@ static inline void dev_iommu_priv_set(struct device *dev, void *priv)
 	dev->iommu->priv = priv;
 }
 
+static inline void domain_set_concurrent_traversal(struct iommu_domain *domain,
+						   bool value)
+{
+	domain->concurrent_traversal = value;
+}
+
 int iommu_probe_device(struct device *dev);
 void iommu_release_device(struct device *dev);
 
@@ -677,6 +684,8 @@ int iommu_group_claim_dma_owner(struct iommu_group *group, void *owner);
 void iommu_group_release_dma_owner(struct iommu_group *group);
 bool iommu_group_dma_owner_claimed(struct iommu_group *group);
 
+void iommu_free_pgtbl_pages(struct iommu_domain *domain,
+			    struct list_head *pages);
 #else /* CONFIG_IOMMU_API */
 
 struct iommu_ops {};
drivers/iommu/iommu.c +23 -0
index 847ad47a2dfd..ceeb97ebe3e2 100644
@@ -3252,3 +3252,26 @@ bool iommu_group_dma_owner_claimed(struct iommu_group *group)
 	return user;
 }
 EXPORT_SYMBOL_GPL(iommu_group_dma_owner_claimed);
+
+static void pgtble_page_free_rcu(struct rcu_head *rcu)
+{
+	struct page *page = container_of(rcu, struct page, rcu_head);
+
+	__free_pages(page, 0);
+}
+
+void iommu_free_pgtbl_pages(struct iommu_domain *domain,
+			    struct list_head *pages)
+{
+	struct page *page, *next;
+
+	if (!domain->concurrent_traversal) {
+		put_pages_list(pages);
+		return;
+	}
+
+	list_for_each_entry_safe(page, next, pages, lru) {
+		list_del(&page->lru);
+		call_rcu(&page->rcu_head, pgtble_page_free_rcu);
+	}
+}
Signature · 6 lines
2.25.1

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu
author maintainer
#1 lore raw

RFC PATCH 2/2 iommu: Replace put_pages_list() with iommu_free_pgtbl_pages()

Therefore, RCU protected page free will take effect if necessary.

---
 drivers/iommu/amd/io_pgtable.c | 5 ++---
 drivers/iommu/dma-iommu.c      | 6 ++++--
 drivers/iommu/intel/iommu.c    | 4 ++--
 3 files changed, 8 insertions(+), 7 deletions(-)
drivers/iommu/amd/io_pgtable.c +2 -3
index 6608d1717574..a62d5dafd7f2 100644
@@ -423,7 +423,7 @@ static int iommu_v1_map_page(struct io_pgtable_ops *ops, unsigned long iova,
 	}
 
 	/* Everything flushed out, free pages now */
-	put_pages_list(&freelist);
+	iommu_free_pgtbl_pages(&dom->domain, &freelist);
 
 	return ret;
 }
@@ -503,8 +503,7 @@ static void v1_free_pgtable(struct io_pgtable *iop)
 
 	/* Make changes visible to IOMMUs */
 	amd_iommu_domain_update(dom);
-
-	put_pages_list(&freelist);
+	iommu_free_pgtbl_pages(&dom->domain, &freelist);
 }
 
 static struct io_pgtable *v1_alloc_pgtable(struct io_pgtable_cfg *cfg, void *cookie)
drivers/iommu/dma-iommu.c +4 -2
index f90251572a5d..a948358c3e51 100644
@@ -132,7 +132,8 @@ static void fq_ring_free(struct iommu_dma_cookie *cookie, struct iova_fq *fq)
 		if (fq->entries[idx].counter >= counter)
 			break;
 
-		put_pages_list(&fq->entries[idx].freelist);
+		iommu_free_pgtbl_pages(cookie->fq_domain,
+				       &fq->entries[idx].freelist);
 		free_iova_fast(&cookie->iovad,
 			       fq->entries[idx].iova_pfn,
 			       fq->entries[idx].pages);
@@ -228,7 +229,8 @@ static void iommu_dma_free_fq(struct iommu_dma_cookie *cookie)
 		struct iova_fq *fq = per_cpu_ptr(cookie->fq, cpu);
 
 		fq_ring_for_each(idx, fq)
-			put_pages_list(&fq->entries[idx].freelist);
+			iommu_free_pgtbl_pages(cookie->fq_domain,
+					       &fq->entries[idx].freelist);
 	}
 
 	free_percpu(cookie->fq);
drivers/iommu/intel/iommu.c +2 -2
index 19024dc52735..f429671e837f 100644
@@ -1891,7 +1891,7 @@ static void domain_exit(struct dmar_domain *domain)
 		LIST_HEAD(freelist);
 
 		domain_unmap(domain, 0, DOMAIN_MAX_PFN(domain->gaw), &freelist);
-		put_pages_list(&freelist);
+		iommu_free_pgtbl_pages(&domain->domain, &freelist);
 	}
 
 	kfree(domain);
@@ -4510,7 +4510,7 @@ static void intel_iommu_tlb_sync(struct iommu_domain *domain,
 				      start_pfn, nrpages,
 				      list_empty(&gather->freelist), 0);
 
-	put_pages_list(&gather->freelist);
+	iommu_free_pgtbl_pages(domain, &gather->freelist);
 }
 
 static phys_addr_t intel_iommu_iova_to_phys(struct iommu_domain *domain,
Signature · 6 lines
2.25.1

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu
#2 lore raw
In reply to Lu Baolu
Quoted 39 lines · Lu Baolu

The IOMMU page tables are updated using iommu_map/unmap() interfaces. Currently, there is no mandatory requirement for drivers to use locks to ensure concurrent updates to page tables, because it's assumed that overlapping IOVA ranges do not have concurrent updates. Therefore the IOMMU drivers only need to take care of concurrent updates to level page table entries.

But enabling new features challenges this assumption. For example, the hardware assisted dirty page tracking feature requires scanning page tables in interfaces other than mapping and unmapping. This might result in a use-after-free scenario in which a level page table has been freed by the unmap() interface, while another thread is scanning the next level page table.

This adds RCU-protected page free support so that the pages are really freed and reused after a RCU grace period. Hence, the page tables are safe for scanning within a rcu_read_lock critical region. Considering that scanning the page table is a rare case, this also adds a domain flag and the RCU-protected page free is only used when this flat is set.

---
 include/linux/iommu.h |  9 +++++++++
 drivers/iommu/iommu.c | 23 +++++++++++++++++++++++
 2 files changed, 32 insertions(+)
include/linux/iommu.h +1 -0
index 5e1afe169549..6f68eabb8567 100644
@@ -95,6 +95,7 @@ struct iommu_domain {
 	void *handler_token;
 	struct iommu_domain_geometry geometry;
 	struct iommu_dma_cookie *iova_cookie;
+	unsigned long concurrent_traversal:1;
 };
 
 static inline bool iommu_is_dma_domain(struct iommu_domain *domain)

include/linux/iommu.h static inline void dev_iommu_priv_set(struct device *dev, void *priv) · line 665

Quoted 8 lines
include/linux/iommu.h +5 -0
@@ -657,6 +658,12 @@ static inline void dev_iommu_priv_set(struct device *dev, void *priv)
 	dev->iommu->priv = priv;
 }
 
+static inline void domain_set_concurrent_traversal(struct iommu_domain *domain,
+						   bool value)
+{
+	domain->concurrent_traversal = value;
+}

?? If you want it to be a driver opt in I would just add a flags to the domain ops. "DOMAIN_FLAG_RCU_FREE_PAGES"

Quoted 4 lines
include/linux/iommu.h +4 -0
+void iommu_free_pgtbl_pages(struct iommu_domain *domain,
+			    struct list_head *pages)
+{
+	struct page *page, *next;

include/linux/iommu.h static inline void dev_iommu_priv_set(struct device *dev, void *priv)

Quoted 10 lines
include/linux/iommu.h +10 -0
+
+	if (!domain->concurrent_traversal) {
+		put_pages_list(pages);
+		return;
+	}
+
+	list_for_each_entry_safe(page, next, pages, lru) {
+		list_del(&page->lru);
+		call_rcu(&page->rcu_head, pgtble_page_free_rcu);
+	}

It seems OK, but I wonder if there is benifit to using put_pages_list() from the rcu callback

Jason
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

maintainer
#3 lore raw
Quoted 43 lines · Jason Gunthorpe
Quoted 42 lines · Lu Baolu

The IOMMU page tables are updated using iommu_map/unmap() interfaces. Currently, there is no mandatory requirement for drivers to use locks to ensure concurrent updates to page tables, because it's assumed that overlapping IOVA ranges do not have concurrent updates. Therefore the IOMMU drivers only need to take care of concurrent updates to level page table entries.

But enabling new features challenges this assumption. For example, the hardware assisted dirty page tracking feature requires scanning page tables in interfaces other than mapping and unmapping. This might result in a use-after-free scenario in which a level page table has been freed by the unmap() interface, while another thread is scanning the next level page table.

This adds RCU-protected page free support so that the pages are really freed and reused after a RCU grace period. Hence, the page tables are safe for scanning within a rcu_read_lock critical region. Considering that scanning the page table is a rare case, this also adds a domain flag and the RCU-protected page free is only used when this flat is set.

---
  include/linux/iommu.h |  9 +++++++++
  drivers/iommu/iommu.c | 23 +++++++++++++++++++++++
  2 files changed, 32 insertions(+)
include/linux/iommu.h +1 -0
index 5e1afe169549..6f68eabb8567 100644
@@ -95,6 +95,7 @@ struct iommu_domain {
  	void *handler_token;
  	struct iommu_domain_geometry geometry;
  	struct iommu_dma_cookie *iova_cookie;
+	unsigned long concurrent_traversal:1;
  };
 
  static inline bool iommu_is_dma_domain(struct iommu_domain *domain)
@@ -657,6 +658,12 @@ static inline void dev_iommu_priv_set(struct device *dev, void *priv)
  	dev->iommu->priv = priv;
  }
 
Quoted 8 lines

include/linux/iommu.h static inline void dev_iommu_priv_set(struct device *dev, void *priv) · line 665

Quoted 5 lines
include/linux/iommu.h +5 -0
+static inline void domain_set_concurrent_traversal(struct iommu_domain *domain,
+						   bool value)
+{
+	domain->concurrent_traversal = value;
+}

?? If you want it to be a driver opt in I would just add a flags to the domain ops. "DOMAIN_FLAG_RCU_FREE_PAGES"

Is there a significant benefit to keeping both paths, or could we get away with just always using RCU? Realistically, pagetable pages aren't likely to be freed all that frequently, except perhaps at domain teardown, but that shouldn't really be performance-critical, and I guess we could stick an RCU sync point in iommu_domain_free() if we're really worried about releasing larger quantities of pages back to the allocator ASAP?

It's already a driver opt-in to use the iommu_iotlb_gather freelist in the first place, and right now the ones that do are also the ones that do lock-free table walks so will ultimately all want this as well.

Robin.

Quoted 18 lines
Quoted 4 lines
include/linux/iommu.h +4 -0
+void iommu_free_pgtbl_pages(struct iommu_domain *domain,
+			    struct list_head *pages)
+{
+	struct page *page, *next;

include/linux/iommu.h static inline void dev_iommu_priv_set(struct device *dev, void *priv)

Quoted 10 lines
include/linux/iommu.h +10 -0
+
+	if (!domain->concurrent_traversal) {
+		put_pages_list(pages);
+		return;
+	}
+
+	list_for_each_entry_safe(page, next, pages, lru) {
+		list_del(&page->lru);
+		call_rcu(&page->rcu_head, pgtble_page_free_rcu);
+	}

It seems OK, but I wonder if there is benifit to using put_pages_list() from the rcu callback

Quoted 1 line

Jason

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

#4 lore raw
Quoted 6 lines · Robin Murphy

Is there a significant benefit to keeping both paths, or could we get away with just always using RCU? Realistically, pagetable pages aren't likely to be freed all that frequently, except perhaps at domain teardown, but that shouldn't really be performance-critical, and I guess we could stick an RCU sync point in iommu_domain_free() if we're really worried about releasing larger quantities of pages back to the allocator ASAP?

I think you are right, anything that uses the iommu_iotlb_gather may as well use RCU too.

IIRC the allocators already know that RCU is often sitting on freed-memory and have some contigency to flush it out before OOMing, so nothing special should be needed.

Jason
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

#10 lore raw
In reply to Lu Baolu
Quoted 6 lines · Lu Baolu

The IOMMU page tables are updated using iommu_map/unmap() interfaces. Currently, there is no mandatory requirement for drivers to use locks to ensure concurrent updates to page tables, because it's assumed that overlapping IOVA ranges do not have concurrent updates. Therefore the IOMMU drivers only need to take care of concurrent updates to level page table entries.

The last part doesn't read well..
s/updates to level page table entries/ updates to page-table entries at the same level

Quoted 8 lines

But enabling new features challenges this assumption. For example, the hardware assisted dirty page tracking feature requires scanning page tables in interfaces other than mapping and unmapping. This might result in a use-after-free scenario in which a level page table has been freed by the unmap() interface, while another thread is scanning the next level page table.

Quoted 5 lines

This adds RCU-protected page free support so that the pages are really freed and reused after a RCU grace period. Hence, the page tables are safe for scanning within a rcu_read_lock critical region. Considering that scanning the page table is a rare case, this also adds a domain flag and the RCU-protected page free is only used when this flat is set.

s/flat/flag

Quoted 12 lines
---
 include/linux/iommu.h |  9 +++++++++
 drivers/iommu/iommu.c | 23 +++++++++++++++++++++++
 2 files changed, 32 insertions(+)
include/linux/iommu.h +0 -0
index 5e1afe169549..6f68eabb8567 100644

include/linux/iommu.h struct iommu_domain { · line 98

Quoted 4 lines
include/linux/iommu.h +1 -0
@@ -95,6 +95,7 @@ struct iommu_domain {
 	void *handler_token;
 	struct iommu_domain_geometry geometry;
 	struct iommu_dma_cookie *iova_cookie;
+	unsigned long concurrent_traversal:1;

Does this need to be a bitfield? Even though you are needing just one bit now, you can probably make have maskbits?

Quoted 30 lines
include/linux/iommu.h +8 -0
 };
 
 static inline bool iommu_is_dma_domain(struct iommu_domain *domain)
@@ -657,6 +658,12 @@ static inline void dev_iommu_priv_set(struct device *dev, void *priv)
 	dev->iommu->priv = priv;
 }
 
+static inline void domain_set_concurrent_traversal(struct iommu_domain *domain,
+						   bool value)
+{
+	domain->concurrent_traversal = value;
+}
+
 int iommu_probe_device(struct device *dev);
 void iommu_release_device(struct device *dev);
 
@@ -677,6 +684,8 @@ int iommu_group_claim_dma_owner(struct iommu_group *group, void *owner);
 void iommu_group_release_dma_owner(struct iommu_group *group);
 bool iommu_group_dma_owner_claimed(struct iommu_group *group);
 
+void iommu_free_pgtbl_pages(struct iommu_domain *domain,
+			    struct list_head *pages);
 #else /* CONFIG_IOMMU_API */
 
 struct iommu_ops {};
drivers/iommu/iommu.c +0 -0
index 847ad47a2dfd..ceeb97ebe3e2 100644

drivers/iommu/iommu.c bool iommu_group_dma_owner_claimed(struct iommu_group *group) · line 3256

Quoted 5 lines
drivers/iommu/iommu.c +2 -0
@@ -3252,3 +3252,26 @@ bool iommu_group_dma_owner_claimed(struct iommu_group *group)
 	return user;
 }
 EXPORT_SYMBOL_GPL(iommu_group_dma_owner_claimed);
+
+static void pgtble_page_free_rcu(struct rcu_head *rcu)

maybe the names can be consistent? pgtble_ vs pgtbl below.

vote to drop the 'e' :-)

Quoted 24 lines
drivers/iommu/iommu.c +21 -0
+{
+	struct page *page = container_of(rcu, struct page, rcu_head);
+
+	__free_pages(page, 0);
+}
+
+void iommu_free_pgtbl_pages(struct iommu_domain *domain,
+			    struct list_head *pages)
+{
+	struct page *page, *next;
+
+	if (!domain->concurrent_traversal) {
+		put_pages_list(pages);
+		return;
+	}
+
+	list_for_each_entry_safe(page, next, pages, lru) {
+		list_del(&page->lru);
+		call_rcu(&page->rcu_head, pgtble_page_free_rcu);
+	}
+}
Signature · 1 line
2.25.1

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

author maintainer
#6 lore raw
In reply to Jason Gunthorpe
Quoted 8 lines · Jason Gunthorpe
Quoted 8 lines
drivers/iommu/iommu.c +8 -0
+void iommu_free_pgtbl_pages(struct iommu_domain *domain,
+			    struct list_head *pages)
+{
+	struct page *page, *next;
+
+	if (!domain->concurrent_traversal) {
+		put_pages_list(pages);
+		return;
Quoted 8 lines

drivers/iommu/iommu.c bool iommu_group_dma_owner_claimed(struct iommu_group *group) · line 3276

Quoted 6 lines
drivers/iommu/iommu.c +6 -0
+	}
+
+	list_for_each_entry_safe(page, next, pages, lru) {
+		list_del(&page->lru);
+		call_rcu(&page->rcu_head, pgtble_page_free_rcu);
+	}

It seems OK, but I wonder if there is benifit to using put_pages_list() from the rcu callback

The price is that we need to allocate a "struct list_head" and free it in the rcu callback as well. Currently the list_head is sitting in the stack.

Best regards,
baolu
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

author maintainer
#5 lore raw
In reply to Jason Gunthorpe
Quoted 9 lines · Jason Gunthorpe
Quoted 6 lines · Robin Murphy

Is there a significant benefit to keeping both paths, or could we get away with just always using RCU? Realistically, pagetable pages aren't likely to be freed all that frequently, except perhaps at domain teardown, but that shouldn't really be performance-critical, and I guess we could stick an RCU sync point in iommu_domain_free() if we're really worried about releasing larger quantities of pages back to the allocator ASAP?

Quoted 6 lines

I think you are right, anything that uses the iommu_iotlb_gather may as well use RCU too.

IIRC the allocators already know that RCU is often sitting on freed-memory and have some contigency to flush it out before OOMing, so nothing special should be needed.

Fair enough. How about below code?

static void pgtble_page_free_rcu(struct rcu_head *rcu)

{
        struct page *page = container_of(rcu, struct page, rcu_head);

        __free_pages(page, 0);
}

/*
* Free pages gathered in the freelist of iommu_iotlb_gather. Use RCU free
* way so that it's safe for lock-free page table walk.
*/
void iommu_free_iotlb_gather_pages(struct iommu_iotlb_gather *iotlb_gather)

{
        struct page *page, *next;

        list_for_each_entry_safe(page, next, &iotlb_gather->freelist, 
lru) {
                list_del(&page->lru);
                call_rcu(&page->rcu_head, pgtble_page_free_rcu);
        }
}

Best regards,
baolu

author maintainer
#11 lore raw
In reply to Raj, Ashok
Quoted 20 lines · Ashok
Quoted 6 lines · Lu Baolu

The IOMMU page tables are updated using iommu_map/unmap() interfaces. Currently, there is no mandatory requirement for drivers to use locks to ensure concurrent updates to page tables, because it's assumed that overlapping IOVA ranges do not have concurrent updates. Therefore the IOMMU drivers only need to take care of concurrent updates to level page table entries.

The last part doesn't read well..
s/updates to level page table entries/ updates to page-table entries at the same level

Quoted 8 lines

But enabling new features challenges this assumption. For example, the hardware assisted dirty page tracking feature requires scanning page tables in interfaces other than mapping and unmapping. This might result in a use-after-free scenario in which a level page table has been freed by the unmap() interface, while another thread is scanning the next level page table.

Quoted 7 lines
Quoted 5 lines

This adds RCU-protected page free support so that the pages are really freed and reused after a RCU grace period. Hence, the page tables are safe for scanning within a rcu_read_lock critical region. Considering that scanning the page table is a rare case, this also adds a domain flag and the RCU-protected page free is only used when this flat is set.

s/flat/flag

Above updated. Thank you!

Quoted 7 lines
Quoted 7 lines
---
  include/linux/iommu.h |  9 +++++++++
  drivers/iommu/iommu.c | 23 +++++++++++++++++++++++
  2 files changed, 32 insertions(+)
Quoted 13 lines

include/linux/iommu.h struct iommu_domain { · line 98

Quoted 9 lines
include/linux/iommu.h +1 -0
index 5e1afe169549..6f68eabb8567 100644
@@ -95,6 +95,7 @@ struct iommu_domain {
  	void *handler_token;
  	struct iommu_domain_geometry geometry;
  	struct iommu_dma_cookie *iova_cookie;
+	unsigned long concurrent_traversal:1;

Does this need to be a bitfield? Even though you are needing just one bit now, you can probably make have maskbits?

As discussed in another reply, I am about to drop the driver opt-in and wrapper the helper around the iommu_iotlb_gather.

Quoted 39 lines
Quoted 31 lines
include/linux/iommu.h +8 -0
  };
 
  static inline bool iommu_is_dma_domain(struct iommu_domain *domain)
@@ -657,6 +658,12 @@ static inline void dev_iommu_priv_set(struct device *dev, void *priv)
  	dev->iommu->priv = priv;
  }
 
+static inline void domain_set_concurrent_traversal(struct iommu_domain *domain,
+						   bool value)
+{
+	domain->concurrent_traversal = value;
+}
+
  int iommu_probe_device(struct device *dev);
  void iommu_release_device(struct device *dev);
 
@@ -677,6 +684,8 @@ int iommu_group_claim_dma_owner(struct iommu_group *group, void *owner);
  void iommu_group_release_dma_owner(struct iommu_group *group);
  bool iommu_group_dma_owner_claimed(struct iommu_group *group);
 
+void iommu_free_pgtbl_pages(struct iommu_domain *domain,
+			    struct list_head *pages);
  #else /* CONFIG_IOMMU_API */
 
  struct iommu_ops {};
drivers/iommu/iommu.c +0 -0
index 847ad47a2dfd..ceeb97ebe3e2 100644

drivers/iommu/iommu.c bool iommu_group_dma_owner_claimed(struct iommu_group *group) · line 3256

Quoted 5 lines
drivers/iommu/iommu.c +2 -0
@@ -3252,3 +3252,26 @@ bool iommu_group_dma_owner_claimed(struct iommu_group *group)
  	return user;
  }
  EXPORT_SYMBOL_GPL(iommu_group_dma_owner_claimed);
+
+static void pgtble_page_free_rcu(struct rcu_head *rcu)

maybe the names can be consistent? pgtble_ vs pgtbl below.

Quoted 1 line

vote to drop the 'e' :-)

Updated.

Quoted 17 lines
Quoted 17 lines
drivers/iommu/iommu.c +16 -0
+{
+	struct page *page = container_of(rcu, struct page, rcu_head);
+
+	__free_pages(page, 0);
+}
+
+void iommu_free_pgtbl_pages(struct iommu_domain *domain,
+			    struct list_head *pages)
+{
+	struct page *page, *next;
+
+	if (!domain->concurrent_traversal) {
+		put_pages_list(pages);
+		return;
+	}
+
Quoted 8 lines
Quoted 8 lines
drivers/iommu/iommu.c +5 -0
+	list_for_each_entry_safe(page, next, pages, lru) {
+		list_del(&page->lru);
+		call_rcu(&page->rcu_head, pgtble_page_free_rcu);
+	}
+}
Signature · 1 line
2.25.1

Best regards,
baolu
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

#7 lore raw
In reply to Baolu Lu
Quoted 18 lines · Baolu Lu
Quoted 8 lines · Jason Gunthorpe
Quoted 8 lines
drivers/iommu/iommu.c +8 -0
+void iommu_free_pgtbl_pages(struct iommu_domain *domain,
+			    struct list_head *pages)
+{
+	struct page *page, *next;
+
+	if (!domain->concurrent_traversal) {
+		put_pages_list(pages);
+		return;
Quoted 8 lines

drivers/iommu/iommu.c bool iommu_group_dma_owner_claimed(struct iommu_group *group) · line 3276

Quoted 6 lines
drivers/iommu/iommu.c +6 -0
+	}
+
+	list_for_each_entry_safe(page, next, pages, lru) {
+		list_del(&page->lru);
+		call_rcu(&page->rcu_head, pgtble_page_free_rcu);
+	}

It seems OK, but I wonder if there is benifit to using put_pages_list() from the rcu callback

Quoted 3 lines

The price is that we need to allocate a "struct list_head" and free it in the rcu callback as well. Currently the list_head is sitting in the stack.

You'd have to use a different struct page layout so that the list_head was in the struct page and didn't overlap with the rcu_head

Jason

author maintainer
#8 lore raw
Quoted 19 lines · Jason Gunthorpe
Quoted 18 lines · Baolu Lu
Quoted 16 lines · Jason Gunthorpe
Quoted 4 lines
drivers/iommu/iommu.c +4 -0
+void iommu_free_pgtbl_pages(struct iommu_domain *domain,
+			    struct list_head *pages)
+{
+	struct page *page, *next;

drivers/iommu/iommu.c bool iommu_group_dma_owner_claimed(struct iommu_group *group) · line 3276

Quoted 10 lines
drivers/iommu/iommu.c +10 -0
+
+	if (!domain->concurrent_traversal) {
+		put_pages_list(pages);
+		return;
+	}
+
+	list_for_each_entry_safe(page, next, pages, lru) {
+		list_del(&page->lru);
+		call_rcu(&page->rcu_head, pgtble_page_free_rcu);
+	}

It seems OK, but I wonder if there is benifit to using put_pages_list() from the rcu callback

Quoted 6 lines
Quoted 3 lines

The price is that we need to allocate a "struct list_head" and free it in the rcu callback as well. Currently the list_head is sitting in the stack.

You'd have to use a different struct page layout so that the list_head was in the struct page and didn't overlap with the rcu_head

Okay, let me head this direction in the next version.

Best regards,
baolu

#9 lore raw
Quoted 20 lines · Baolu Lu
Quoted 19 lines · Jason Gunthorpe
Quoted 18 lines · Baolu Lu
Quoted 16 lines · Jason Gunthorpe
Quoted 4 lines
drivers/iommu/iommu.c +4 -0
+void iommu_free_pgtbl_pages(struct iommu_domain *domain,
+			    struct list_head *pages)
+{
+	struct page *page, *next;

drivers/iommu/iommu.c bool iommu_group_dma_owner_claimed(struct iommu_group *group) · line 3276

Quoted 10 lines
drivers/iommu/iommu.c +10 -0
+
+	if (!domain->concurrent_traversal) {
+		put_pages_list(pages);
+		return;
+	}
+
+	list_for_each_entry_safe(page, next, pages, lru) {
+		list_del(&page->lru);
+		call_rcu(&page->rcu_head, pgtble_page_free_rcu);
+	}

It seems OK, but I wonder if there is benifit to using put_pages_list() from the rcu callback

Quoted 8 lines
Quoted 6 lines
Quoted 3 lines

The price is that we need to allocate a "struct list_head" and free it in the rcu callback as well. Currently the list_head is sitting in the stack.

You'd have to use a different struct page layout so that the list_head was in the struct page and didn't overlap with the rcu_head

Okay, let me head this direction in the next version.

I'm not sure it is worth it performance wise, would be interesting to know perhaps. But see my prior email about how slab has a custom struct page.

Jason
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu