Skip to contentSkip to conversation

amd-iommu fixes for tip/core/iommu

GIT PULL In review Joerg Roedel 3 participants 7 messages Jun 9, 2009 x86

lore
Share & tools
lore
#0 lore raw

Hi Ingo,

The following changes since commit 62a6f465f6572e1f28765c583c12753bb3e23715:
Ingo Molnar (1):

        Merge branch 'dma-debug/2.6.31' of git://git.kernel.org/.../joro/linux-2.6-iommu into core/iommu

are available in the git repository at:

git://git.kernel.org/pub/scm/linux/kernel/git/joro/linux-2.6-iommu.git amd-iommu/fixes

Jiri Slaby (1):
amd_iommu: fix lock imbalance

Joerg Roedel (4):
Merge commit 'tip/core/iommu' into amd-iommu/fixes
amd-iommu: remove BUS_NOTIFY_BOUND_DRIVER handling
amd-iommu: detach device explicitly before attaching it to a new domain
amd-iommu: remove unnecessary "AMD IOMMU: " prefix

 arch/x86/kernel/amd_iommu.c |   22 ++++++----------------
 1 files changed, 6 insertions(+), 16 deletions(-)

These patches fix issues with device assignment to KVM, a possible deadlock due to an unreleased lock and a minor printk style issue. Please pull.

Thanks,

Joerg

#1 lore raw

PATCH 1/4 amd_iommu: fix lock imbalance

From: Jiri Slaby <[email protected]>

In alloc_coherent there is an omitted unlock on the path where mapping fails. Add the unlock.

[ Impact: fix lock imbalance in alloc_coherent ]

---
 arch/x86/kernel/amd_iommu.c |    4 +++-
 1 files changed, 3 insertions(+), 1 deletions(-)
arch/x86/kernel/amd_iommu.c +3 -1
index d689883..9f89bb6 100644
@@ -1541,8 +1541,10 @@ static void *alloc_coherent(struct device *dev, size_t size,
 	*dma_addr = __map_single(dev, iommu, domain->priv, paddr,
 				 size, DMA_BIDIRECTIONAL, true, dma_mask);
 
-	if (*dma_addr == bad_dma_address)
+	if (*dma_addr == bad_dma_address) {
+		spin_unlock_irqrestore(&domain->lock, flags);
 		goto out_free;
+	}
 
 	iommu_completion_wait(iommu);
 
Signature · 1 line
1.6.3.1
#2 lore raw

PATCH 2/4 amd-iommu: remove BUS_NOTIFY_BOUND_DRIVER handling

In reply to Joerg Roedel

Handling this event causes device assignment in KVM to fail because the device gets re-attached as soon as the pci-stub registers as the driver for the device.

---
 arch/x86/kernel/amd_iommu.c |   11 -----------
 1 files changed, 0 insertions(+), 11 deletions(-)
arch/x86/kernel/amd_iommu.c +0 -11
index 8510e90..8187260 100644
@@ -1145,17 +1145,6 @@ static int device_change_notifier(struct notifier_block *nb,
 			  "to a non-dma-ops domain\n", dev_name(dev));
 
 	switch (action) {
-	case BUS_NOTIFY_BOUND_DRIVER:
-		if (domain)
-			goto out;
-		dma_domain = find_protection_domain(devid);
-		if (!dma_domain)
-			dma_domain = iommu->default_dom;
-		attach_device(iommu, &dma_domain->domain, devid);
-		DUMP_printk(KERN_INFO "AMD IOMMU: Using protection domain "
-			    "%d for device %s\n",
-			    dma_domain->domain.id, dev_name(dev));
-		break;
 	case BUS_NOTIFY_UNBOUND_DRIVER:
 		if (!domain)
 			goto out;
Signature · 1 line
1.6.3.1
Acked-by on v1 2/4
#3 lore raw

Re: [PATCH 2/4] amd-iommu: remove BUS_NOTIFY_BOUND_DRIVER handling

Quoted 5 lines · * Joerg Roedel ([email protected]) wrote:

Handling this event causes device assignment in KVM to fail because the device gets re-attached as soon as the pci-stub registers as the driver for the device.

Yep, makes good sense, will get picked up from get_device_resources.

thanks,
-chris

#4 lore raw

PATCH 3/4 amd-iommu: detach device explicitly before attaching it to a new domain

In reply to Joerg Roedel

This fixes a bug with a device that could not be assigned to a KVM guest because it is still assigned to a dma_ops protection domain.

[chrisw: simply remove WARN_ON(), will always fire since dev->driver will be pci-sub]

---
 arch/x86/kernel/amd_iommu.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)
arch/x86/kernel/amd_iommu.c +1 -1
index 8187260..772e910 100644
@@ -2073,7 +2073,7 @@ static int amd_iommu_attach_device(struct iommu_domain *dom,
 
 	old_domain = domain_for_device(devid);
 	if (old_domain)
-		return -EBUSY;
+		detach_device(old_domain, devid);
 
 	attach_device(iommu, domain, devid);
 
Signature · 1 line
1.6.3.1
#5 lore raw

PATCH 4/4 amd-iommu: remove unnecessary "AMD IOMMU: " prefix

In reply to Joerg Roedel

That prefix is already included in the DUMP_printk macro. So there is no need to repeat it in the format string.

---
 arch/x86/kernel/amd_iommu.c |    5 ++---
 1 files changed, 2 insertions(+), 3 deletions(-)
arch/x86/kernel/amd_iommu.c +2 -3
index 772e910..1c60554 100644
@@ -1266,9 +1266,8 @@ static int get_device_resources(struct device *dev,
 			dma_dom = (*iommu)->default_dom;
 		*domain = &dma_dom->domain;
 		attach_device(*iommu, *domain, *bdf);
-		DUMP_printk(KERN_INFO "AMD IOMMU: Using protection domain "
-				"%d for device %s\n",
-				(*domain)->id, dev_name(dev));
+		DUMP_printk("Using protection domain %d for device %s\n",
+			    (*domain)->id, dev_name(dev));
 	}
 
 	if (domain_for_device(_bdf) == NULL)
Signature · 1 line
1.6.3.1
#6 lore raw
In reply to Joerg Roedel
Quoted 19 lines · Joerg Roedel

Hi Ingo,

The following changes since commit 62a6f465f6572e1f28765c583c12753bb3e23715:
Ingo Molnar (1):

        Merge branch 'dma-debug/2.6.31' of git://git.kernel.org/.../joro/linux-2.6-iommu into core/iommu

are available in the git repository at:

git://git.kernel.org/pub/scm/linux/kernel/git/joro/linux-2.6-iommu.git amd-iommu/fixes

Jiri Slaby (1):
amd_iommu: fix lock imbalance

Joerg Roedel (4):
Merge commit 'tip/core/iommu' into amd-iommu/fixes
amd-iommu: remove BUS_NOTIFY_BOUND_DRIVER handling
amd-iommu: detach device explicitly before attaching it to a new domain
amd-iommu: remove unnecessary "AMD IOMMU: " prefix

Quoted 6 lines
 arch/x86/kernel/amd_iommu.c |   22 ++++++----------------
 1 files changed, 6 insertions(+), 16 deletions(-)

These patches fix issues with device assignment to KVM, a possible deadlock due to an unreleased lock and a minor printk style issue. Please pull.

Pulled into tip:core/iommu, thanks a lot Joerg!

Ingo