[PATCH 0/5] AMD IOMMUv2 Cleanups and Fixes

[PATCH 0/5] AMD IOMMUv2 Cleanups and Fixes

[Joerg Roedel]

Hi,

here is a small patch-set with some clean-ups for the AMD IOMMUv2
driver. The most important change is the conversion to use the
mmu_notifier release call-back instead of the task_exit notifier to get
notified when a MM dies. Please review.

Thanks,

	Joerg

Diffstat:

 drivers/iommu/amd_iommu_v2.c |  184 +++++++++++++++++-------------------------

Shortlog:

Joerg Roedel (5):
      iommu/amd: Don't access IOMMUv2 state_table directly
      iommu/amd: Convert IOMMUv2 state_table into state_list
      iommu/amd: Implement mmu_notifier_release call-back
      iommu/amd: Remove IOMMUv2 pasid_state_list
      iommu/amd: Handle parallel invalidate_range_start/end calls correctly

 1 file changed, 74 insertions(+), 110 deletions(-)

[PATCH 1/5] iommu/amd: Don't access IOMMUv2 state_table directly

[Joerg Roedel]

From: Joerg Roedel <jroedel-l3A5Bk7waGM@public.gmane.org>

This is a preparation for converting the state_table into a
state_list.

Signed-off-by: Joerg Roedel <jroedel-l3A5Bk7waGM@public.gmane.org>
Tested-by: Jay Cornwall <Jay.Cornwall-5C7GfCeVMHo@public.gmane.org>
---
 drivers/iommu/amd_iommu_v2.c |   13 +++++++++----
 1 file changed, 9 insertions(+), 4 deletions(-)

diff --git a/drivers/iommu/amd_iommu_v2.c b/drivers/iommu/amd_iommu_v2.c
index 5208828..8c3086a 100644
--- a/drivers/iommu/amd_iommu_v2.c
+++ b/drivers/iommu/amd_iommu_v2.c
@@ -111,13 +111,18 @@ static u16 device_id(struct pci_dev *pdev)
 	return devid;
 }
 
+static struct device_state *__get_device_state(u16 devid)
+{
+	return state_table[devid];
+}
+
 static struct device_state *get_device_state(u16 devid)
 {
 	struct device_state *dev_state;
 	unsigned long flags;
 
 	spin_lock_irqsave(&state_lock, flags);
-	dev_state = state_table[devid];
+	dev_state = __get_device_state(devid);
 	if (dev_state != NULL)
 		atomic_inc(&dev_state->count);
 	spin_unlock_irqrestore(&state_lock, flags);
@@ -839,7 +844,7 @@ void amd_iommu_free_device(struct pci_dev *pdev)
 
 	spin_lock_irqsave(&state_lock, flags);
 
-	dev_state = state_table[devid];
+	dev_state = __get_device_state(devid);
 	if (dev_state == NULL) {
 		spin_unlock_irqrestore(&state_lock, flags);
 		return;
@@ -872,7 +877,7 @@ int amd_iommu_set_invalid_ppr_cb(struct pci_dev *pdev,
 	spin_lock_irqsave(&state_lock, flags);
 
 	ret = -EINVAL;
-	dev_state = state_table[devid];
+	dev_state = __get_device_state(devid);
 	if (dev_state == NULL)
 		goto out_unlock;
 
@@ -903,7 +908,7 @@ int amd_iommu_set_invalidate_ctx_cb(struct pci_dev *pdev,
 	spin_lock_irqsave(&state_lock, flags);
 
 	ret = -EINVAL;
-	dev_state = state_table[devid];
+	dev_state = __get_device_state(devid);
 	if (dev_state == NULL)
 		goto out_unlock;
 
-- 
1.7.9.5

[PATCH 2/5] iommu/amd: Convert IOMMUv2 state_table into state_list

[Joerg Roedel]

From: Joerg Roedel <jroedel-l3A5Bk7waGM@public.gmane.org>

The state_table consumes 512kb of memory and is only sparsly
populated. Convert it into a list to save memory. There
should be no measurable performance impact.

Signed-off-by: Joerg Roedel <jroedel-l3A5Bk7waGM@public.gmane.org>
Tested-by: Jay Cornwall <Jay.Cornwall-5C7GfCeVMHo@public.gmane.org>
---
 drivers/iommu/amd_iommu_v2.c |   39 ++++++++++++++++++---------------------
 1 file changed, 18 insertions(+), 21 deletions(-)

diff --git a/drivers/iommu/amd_iommu_v2.c b/drivers/iommu/amd_iommu_v2.c
index 8c3086a..05e93f1 100644
--- a/drivers/iommu/amd_iommu_v2.c
+++ b/drivers/iommu/amd_iommu_v2.c
@@ -56,6 +56,8 @@ struct pasid_state {
 };
 
 struct device_state {
+	struct list_head list;
+	u16 devid;
 	atomic_t count;
 	struct pci_dev *pdev;
 	struct pasid_state **states;
@@ -81,7 +83,7 @@ struct fault {
 	u16 flags;
 };
 
-static struct device_state **state_table;
+static LIST_HEAD(state_list);
 static spinlock_t state_lock;
 
 /* List and lock for all pasid_states */
@@ -113,7 +115,14 @@ static u16 device_id(struct pci_dev *pdev)
 
 static struct device_state *__get_device_state(u16 devid)
 {
-	return state_table[devid];
+	struct device_state *dev_state;
+
+	list_for_each_entry(dev_state, &state_list, list) {
+		if (dev_state->devid == devid)
+			return dev_state;
+	}
+
+	return NULL;
 }
 
 static struct device_state *get_device_state(u16 devid)
@@ -776,7 +785,8 @@ int amd_iommu_init_device(struct pci_dev *pdev, int pasids)
 
 	spin_lock_init(&dev_state->lock);
 	init_waitqueue_head(&dev_state->wq);
-	dev_state->pdev = pdev;
+	dev_state->pdev  = pdev;
+	dev_state->devid = devid;
 
 	tmp = pasids;
 	for (dev_state->pasid_levels = 0; (tmp - 1) & ~0x1ff; tmp >>= 9)
@@ -806,13 +816,13 @@ int amd_iommu_init_device(struct pci_dev *pdev, int pasids)
 
 	spin_lock_irqsave(&state_lock, flags);
 
-	if (state_table[devid] != NULL) {
+	if (__get_device_state(devid) != NULL) {
 		spin_unlock_irqrestore(&state_lock, flags);
 		ret = -EBUSY;
 		goto out_free_domain;
 	}
 
-	state_table[devid] = dev_state;
+	list_add_tail(&dev_state->list, &state_list);
 
 	spin_unlock_irqrestore(&state_lock, flags);
 
@@ -850,7 +860,7 @@ void amd_iommu_free_device(struct pci_dev *pdev)
 		return;
 	}
 
-	state_table[devid] = NULL;
+	list_del(&dev_state->list);
 
 	spin_unlock_irqrestore(&state_lock, flags);
 
@@ -925,7 +935,6 @@ EXPORT_SYMBOL(amd_iommu_set_invalidate_ctx_cb);
 
 static int __init amd_iommu_v2_init(void)
 {
-	size_t state_table_size;
 	int ret;
 
 	pr_info("AMD IOMMUv2 driver by Joerg Roedel <joerg.roedel-5C7GfCeVMHo@public.gmane.org>\n");
@@ -941,16 +950,10 @@ static int __init amd_iommu_v2_init(void)
 
 	spin_lock_init(&state_lock);
 
-	state_table_size = MAX_DEVICES * sizeof(struct device_state *);
-	state_table = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
-					       get_order(state_table_size));
-	if (state_table == NULL)
-		return -ENOMEM;
-
 	ret = -ENOMEM;
 	iommu_wq = create_workqueue("amd_iommu_v2");
 	if (iommu_wq == NULL)
-		goto out_free;
+		goto out;
 
 	ret = -ENOMEM;
 	empty_page_table = (u64 *)get_zeroed_page(GFP_KERNEL);
@@ -965,16 +968,13 @@ static int __init amd_iommu_v2_init(void)
 out_destroy_wq:
 	destroy_workqueue(iommu_wq);
 
-out_free:
-	free_pages((unsigned long)state_table, get_order(state_table_size));
-
+out:
 	return ret;
 }
 
 static void __exit amd_iommu_v2_exit(void)
 {
 	struct device_state *dev_state;
-	size_t state_table_size;
 	int i;
 
 	if (!amd_iommu_v2_supported())
@@ -1003,9 +1003,6 @@ static void __exit amd_iommu_v2_exit(void)
 
 	destroy_workqueue(iommu_wq);
 
-	state_table_size = MAX_DEVICES * sizeof(struct device_state *);
-	free_pages((unsigned long)state_table, get_order(state_table_size));
-
 	free_page((unsigned long)empty_page_table);
 }
 
-- 
1.7.9.5

[PATCH 3/5] iommu/amd: Implement mmu_notifier_release call-back

[Joerg Roedel]

From: Joerg Roedel <jroedel-l3A5Bk7waGM@public.gmane.org>

Since mmu_notifier call-backs can sleep (because they use
SRCU now) we can use them to tear down PASID mappings. This
allows us to finally remove the hack to use the task_exit
notifier from oprofile to get notified when a process dies.

Signed-off-by: Joerg Roedel <jroedel-l3A5Bk7waGM@public.gmane.org>
Tested-by: Jay Cornwall <Jay.Cornwall-5C7GfCeVMHo@public.gmane.org>
---
 drivers/iommu/amd_iommu_v2.c |   91 ++++++++++++++++--------------------------
 1 file changed, 35 insertions(+), 56 deletions(-)

diff --git a/drivers/iommu/amd_iommu_v2.c b/drivers/iommu/amd_iommu_v2.c
index 05e93f1..c06a29a 100644
--- a/drivers/iommu/amd_iommu_v2.c
+++ b/drivers/iommu/amd_iommu_v2.c
@@ -101,7 +101,6 @@ static u64 *empty_page_table;
 
 static void free_pasid_states(struct device_state *dev_state);
 static void unbind_pasid(struct device_state *dev_state, int pasid);
-static int task_exit(struct notifier_block *nb, unsigned long e, void *data);
 
 static u16 device_id(struct pci_dev *pdev)
 {
@@ -172,10 +171,6 @@ static void put_device_state_wait(struct device_state *dev_state)
 	free_device_state(dev_state);
 }
 
-static struct notifier_block profile_nb = {
-	.notifier_call = task_exit,
-};
-
 static void link_pasid_state(struct pasid_state *pasid_state)
 {
 	spin_lock(&ps_lock);
@@ -393,7 +388,12 @@ static void free_pasid_states(struct device_state *dev_state)
 			continue;
 
 		put_pasid_state(pasid_state);
-		unbind_pasid(dev_state, i);
+
+		/*
+		 * This will call the mn_release function and
+		 * unbind the PASID
+		 */
+		mmu_notifier_unregister(&pasid_state->mn, pasid_state->mm);
 	}
 
 	if (dev_state->pasid_levels == 2)
@@ -475,7 +475,24 @@ static void mn_invalidate_range_end(struct mmu_notifier *mn,
 				  __pa(pasid_state->mm->pgd));
 }
 
+static void mn_release(struct mmu_notifier *mn, struct mm_struct *mm)
+{
+	struct pasid_state *pasid_state;
+	struct device_state *dev_state;
+
+	might_sleep();
+
+	pasid_state = mn_to_state(mn);
+	dev_state   = pasid_state->device_state;
+
+	if (pasid_state->device_state->inv_ctx_cb)
+		dev_state->inv_ctx_cb(dev_state->pdev, pasid_state->pasid);
+
+	unbind_pasid(dev_state, pasid_state->pasid);
+}
+
 static struct mmu_notifier_ops iommu_mn = {
+	.release		= mn_release,
 	.clear_flush_young      = mn_clear_flush_young,
 	.change_pte             = mn_change_pte,
 	.invalidate_page        = mn_invalidate_page,
@@ -618,53 +635,6 @@ static struct notifier_block ppr_nb = {
 	.notifier_call = ppr_notifier,
 };
 
-static int task_exit(struct notifier_block *nb, unsigned long e, void *data)
-{
-	struct pasid_state *pasid_state;
-	struct task_struct *task;
-
-	task = data;
-
-	/*
-	 * Using this notifier is a hack - but there is no other choice
-	 * at the moment. What I really want is a sleeping notifier that
-	 * is called when an MM goes down. But such a notifier doesn't
-	 * exist yet. The notifier needs to sleep because it has to make
-	 * sure that the device does not use the PASID and the address
-	 * space anymore before it is destroyed. This includes waiting
-	 * for pending PRI requests to pass the workqueue. The
-	 * MMU-Notifiers would be a good fit, but they use RCU and so
-	 * they are not allowed to sleep. Lets see how we can solve this
-	 * in a more intelligent way in the future.
-	 */
-again:
-	spin_lock(&ps_lock);
-	list_for_each_entry(pasid_state, &pasid_state_list, list) {
-		struct device_state *dev_state;
-		int pasid;
-
-		if (pasid_state->task != task)
-			continue;
-
-		/* Drop Lock and unbind */
-		spin_unlock(&ps_lock);
-
-		dev_state = pasid_state->device_state;
-		pasid     = pasid_state->pasid;
-
-		if (pasid_state->device_state->inv_ctx_cb)
-			dev_state->inv_ctx_cb(dev_state->pdev, pasid);
-
-		unbind_pasid(dev_state, pasid);
-
-		/* Task may be in the list multiple times */
-		goto again;
-	}
-	spin_unlock(&ps_lock);
-
-	return NOTIFY_OK;
-}
-
 int amd_iommu_bind_pasid(struct pci_dev *pdev, int pasid,
 			 struct task_struct *task)
 {
@@ -739,6 +709,7 @@ EXPORT_SYMBOL(amd_iommu_bind_pasid);
 
 void amd_iommu_unbind_pasid(struct pci_dev *pdev, int pasid)
 {
+	struct pasid_state *pasid_state;
 	struct device_state *dev_state;
 	u16 devid;
 
@@ -755,7 +726,17 @@ void amd_iommu_unbind_pasid(struct pci_dev *pdev, int pasid)
 	if (pasid < 0 || pasid >= dev_state->max_pasids)
 		goto out;
 
-	unbind_pasid(dev_state, pasid);
+	pasid_state = get_pasid_state(dev_state, pasid);
+	if (pasid_state == NULL)
+		goto out;
+	/*
+	 * Drop reference taken here. We are safe because we still hold
+	 * the reference taken in the amd_iommu_bind_pasid function.
+	 */
+	put_pasid_state(pasid_state);
+
+	/* This will call the mn_release function and unbind the PASID */
+	mmu_notifier_unregister(&pasid_state->mn, pasid_state->mm);
 
 out:
 	put_device_state(dev_state);
@@ -961,7 +942,6 @@ static int __init amd_iommu_v2_init(void)
 		goto out_destroy_wq;
 
 	amd_iommu_register_ppr_notifier(&ppr_nb);
-	profile_event_register(PROFILE_TASK_EXIT, &profile_nb);
 
 	return 0;
 
@@ -980,7 +960,6 @@ static void __exit amd_iommu_v2_exit(void)
 	if (!amd_iommu_v2_supported())
 		return;
 
-	profile_event_unregister(PROFILE_TASK_EXIT, &profile_nb);
 	amd_iommu_unregister_ppr_notifier(&ppr_nb);
 
 	flush_workqueue(iommu_wq);
-- 
1.7.9.5

[PATCH 4/5] iommu/amd: Remove IOMMUv2 pasid_state_list

[Joerg Roedel]

From: Joerg Roedel <jroedel-l3A5Bk7waGM@public.gmane.org>

This list was only used for the task_exit notifier function.
Now that it is gone we can remove it.

Signed-off-by: Joerg Roedel <jroedel-l3A5Bk7waGM@public.gmane.org>
Tested-by: Jay Cornwall <Jay.Cornwall-5C7GfCeVMHo@public.gmane.org>
---
 drivers/iommu/amd_iommu_v2.c |   26 --------------------------
 1 file changed, 26 deletions(-)

diff --git a/drivers/iommu/amd_iommu_v2.c b/drivers/iommu/amd_iommu_v2.c
index c06a29a..d9309a0 100644
--- a/drivers/iommu/amd_iommu_v2.c
+++ b/drivers/iommu/amd_iommu_v2.c
@@ -86,10 +86,6 @@ struct fault {
 static LIST_HEAD(state_list);
 static spinlock_t state_lock;
 
-/* List and lock for all pasid_states */
-static LIST_HEAD(pasid_state_list);
-static DEFINE_SPINLOCK(ps_lock);
-
 static struct workqueue_struct *iommu_wq;
 
 /*
@@ -171,25 +167,6 @@ static void put_device_state_wait(struct device_state *dev_state)
 	free_device_state(dev_state);
 }
 
-static void link_pasid_state(struct pasid_state *pasid_state)
-{
-	spin_lock(&ps_lock);
-	list_add_tail(&pasid_state->list, &pasid_state_list);
-	spin_unlock(&ps_lock);
-}
-
-static void __unlink_pasid_state(struct pasid_state *pasid_state)
-{
-	list_del(&pasid_state->list);
-}
-
-static void unlink_pasid_state(struct pasid_state *pasid_state)
-{
-	spin_lock(&ps_lock);
-	__unlink_pasid_state(pasid_state);
-	spin_unlock(&ps_lock);
-}
-
 /* Must be called under dev_state->lock */
 static struct pasid_state **__get_pasid_state_ptr(struct device_state *dev_state,
 						  int pasid, bool alloc)
@@ -346,7 +323,6 @@ static void unbind_pasid(struct device_state *dev_state, int pasid)
 	if (pasid_state == NULL)
 		return;
 
-	unlink_pasid_state(pasid_state);
 	__unbind_pasid(pasid_state);
 	put_pasid_state_wait(pasid_state); /* Reference taken in this function */
 }
@@ -687,8 +663,6 @@ int amd_iommu_bind_pasid(struct pci_dev *pdev, int pasid,
 	if (ret)
 		goto out_clear_state;
 
-	link_pasid_state(pasid_state);
-
 	return 0;
 
 out_clear_state:
-- 
1.7.9.5

[PATCH 5/5] iommu/amd: Handle parallel invalidate_range_start/end calls correctly

[Joerg Roedel]

From: Joerg Roedel <jroedel-l3A5Bk7waGM@public.gmane.org>

Add a counter to the pasid_state so that we do not restore
the original page-table before all invalidate_range_start
to invalidate_range_end sections have finished.

Signed-off-by: Joerg Roedel <jroedel-l3A5Bk7waGM@public.gmane.org>
---
 drivers/iommu/amd_iommu_v2.c |   17 +++++++++++++----
 1 file changed, 13 insertions(+), 4 deletions(-)

diff --git a/drivers/iommu/amd_iommu_v2.c b/drivers/iommu/amd_iommu_v2.c
index d9309a0..9ed9da2 100644
--- a/drivers/iommu/amd_iommu_v2.c
+++ b/drivers/iommu/amd_iommu_v2.c
@@ -45,6 +45,8 @@ struct pri_queue {
 struct pasid_state {
 	struct list_head list;			/* For global state-list */
 	atomic_t count;				/* Reference count */
+	atomic_t mmu_notifier_count;		/* Counting nested mmu_notifier
+						   calls */
 	struct task_struct *task;		/* Task bound to this PASID */
 	struct mm_struct *mm;			/* mm_struct for the faults */
 	struct mmu_notifier mn;                 /* mmu_otifier handle */
@@ -433,8 +435,11 @@ static void mn_invalidate_range_start(struct mmu_notifier *mn,
 	pasid_state = mn_to_state(mn);
 	dev_state   = pasid_state->device_state;
 
-	amd_iommu_domain_set_gcr3(dev_state->domain, pasid_state->pasid,
-				  __pa(empty_page_table));
+	if (atomic_add_return(1, &pasid_state->mmu_notifier_count) == 1) {
+		amd_iommu_domain_set_gcr3(dev_state->domain,
+					  pasid_state->pasid,
+					  __pa(empty_page_table));
+	}
 }
 
 static void mn_invalidate_range_end(struct mmu_notifier *mn,
@@ -447,8 +452,11 @@ static void mn_invalidate_range_end(struct mmu_notifier *mn,
 	pasid_state = mn_to_state(mn);
 	dev_state   = pasid_state->device_state;
 
-	amd_iommu_domain_set_gcr3(dev_state->domain, pasid_state->pasid,
-				  __pa(pasid_state->mm->pgd));
+	if (atomic_dec_and_test(&pasid_state->mmu_notifier_count)) {
+		amd_iommu_domain_set_gcr3(dev_state->domain,
+					  pasid_state->pasid,
+					  __pa(pasid_state->mm->pgd));
+	}
 }
 
 static void mn_release(struct mmu_notifier *mn, struct mm_struct *mm)
@@ -640,6 +648,7 @@ int amd_iommu_bind_pasid(struct pci_dev *pdev, int pasid,
 		goto out;
 
 	atomic_set(&pasid_state->count, 1);
+	atomic_set(&pasid_state->mmu_notifier_count, 0);
 	init_waitqueue_head(&pasid_state->wq);
 	spin_lock_init(&pasid_state->lock);
 
-- 
1.7.9.5

Re: [PATCH 0/5] AMD IOMMUv2 Cleanups and Fixes

[Joerg Roedel]

On Tue, May 20, 2014 at 11:18:21PM +0200, Joerg Roedel wrote:
> Joerg Roedel (5):
>       iommu/amd: Don't access IOMMUv2 state_table directly
>       iommu/amd: Convert IOMMUv2 state_table into state_list
>       iommu/amd: Implement mmu_notifier_release call-back
>       iommu/amd: Remove IOMMUv2 pasid_state_list
>       iommu/amd: Handle parallel invalidate_range_start/end calls correctly

Applied.