[PATCH 0/8] Passthrough support for AMD-IOMMU (AMD-Vi)

[PATCH 0/8] Passthrough support for AMD-IOMMU (AMD-Vi)

[Joerg Roedel]

Hi,

this series of patches implement iommu=pt support for AMD IOMMU (AMD-Vi). If
this option is passed on the kernel command line, the AMD-Vi driver will not
register itself as the dma_ops backend and allows all devices unlimited access
to main memory as long as no other kernel part (currently only KVM will do so)
assigns the device to another domain using the IOMMU-API.

Joerg

Diffstat:


 arch/x86/include/asm/amd_iommu.h       |    1 +
 arch/x86/include/asm/amd_iommu_types.h |    4 +
 arch/x86/kernel/amd_iommu.c            |  140 +++++++++++++++++++++++++++----
 arch/x86/kernel/amd_iommu_init.c       |    8 ++-
 arch/x86/kernel/pci-dma.c              |    9 ++-
 5 files changed, 142 insertions(+), 20 deletions(-)

Shortlog:

Joerg Roedel (8):
      x86/dma: Mark iommu_pass_through as __read_mostly
      x86/amd-iommu: Add core functions for pd allocation/freeing
      x86/amd-iommu: Add passthrough mode initialization functions
      x86/amd-iommu: Fix device table write order
      x86/amd-iommu: Align locking between attach_device and detach_device
      x86/amd-iommu: Make sure a device is assigned in passthrough mode
      x86/amd-iommu: Don't detach device from pt domain on driver unbind
      x86/amd-iommu: Initialize passthrough mode when requested

[PATCH 1/8] x86/dma: Mark iommu_pass_through as __read_mostly

[Joerg Roedel]

This variable is read most of the time. This patch marks it
as such. It also documents the meaning the this variable
while at it.

Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
---
 arch/x86/kernel/pci-dma.c |    9 ++++++++-
 1 files changed, 8 insertions(+), 1 deletions(-)

diff --git a/arch/x86/kernel/pci-dma.c b/arch/x86/kernel/pci-dma.c
index 1a041bc..873aa07 100644
--- a/arch/x86/kernel/pci-dma.c
+++ b/arch/x86/kernel/pci-dma.c
@@ -32,7 +32,14 @@ int no_iommu __read_mostly;
 /* Set this to 1 if there is a HW IOMMU in the system */
 int iommu_detected __read_mostly = 0;
 
-int iommu_pass_through;
+/*
+ * This variable becomes 1 if iommu=pt is passed on the kernel command line.
+ * If this variable is 1, IOMMU implementations do no DMA ranslation for
+ * devices and allow every device to access to whole physical memory. This is
+ * useful if a user want to use an IOMMU only for KVM device assignment to
+ * guests and not for driver dma translation.
+ */
+int iommu_pass_through __read_mostly;
 
 dma_addr_t bad_dma_address __read_mostly = 0;
 EXPORT_SYMBOL(bad_dma_address);
-- 
1.6.3.3

[PATCH 2/8] x86/amd-iommu: Add core functions for pd allocation/freeing

[Joerg Roedel]

This patch factors some code of protection domain allocation
into seperate functions. This way the logic can be used to
allocate the passthrough domain later. As a side effect this
patch fixes an unlikely domain id leakage bug.

Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
---
 arch/x86/kernel/amd_iommu.c |   36 ++++++++++++++++++++++++++++++++----
 1 files changed, 32 insertions(+), 4 deletions(-)

diff --git a/arch/x86/kernel/amd_iommu.c b/arch/x86/kernel/amd_iommu.c
index 6c99f50..0934348 100644
--- a/arch/x86/kernel/amd_iommu.c
+++ b/arch/x86/kernel/amd_iommu.c
@@ -1988,19 +1988,47 @@ static void cleanup_domain(struct protection_domain *domain)
 	write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
 }
 
-static int amd_iommu_domain_init(struct iommu_domain *dom)
+static void protection_domain_free(struct protection_domain *domain)
+{
+	if (!domain)
+		return;
+
+	if (domain->id)
+		domain_id_free(domain->id);
+
+	kfree(domain);
+}
+
+static struct protection_domain *protection_domain_alloc(void)
 {
 	struct protection_domain *domain;
 
 	domain = kzalloc(sizeof(*domain), GFP_KERNEL);
 	if (!domain)
-		return -ENOMEM;
+		return NULL;
 
 	spin_lock_init(&domain->lock);
-	domain->mode = PAGE_MODE_3_LEVEL;
 	domain->id = domain_id_alloc();
 	if (!domain->id)
+		goto out_err;
+
+	return domain;
+
+out_err:
+	kfree(domain);
+
+	return NULL;
+}
+
+static int amd_iommu_domain_init(struct iommu_domain *dom)
+{
+	struct protection_domain *domain;
+
+	domain = protection_domain_alloc();
+	if (!domain)
 		goto out_free;
+
+	domain->mode    = PAGE_MODE_3_LEVEL;
 	domain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
 	if (!domain->pt_root)
 		goto out_free;
@@ -2010,7 +2038,7 @@ static int amd_iommu_domain_init(struct iommu_domain *dom)
 	return 0;
 
 out_free:
-	kfree(domain);
+	protection_domain_free(domain);
 
 	return -ENOMEM;
 }
-- 
1.6.3.3

[PATCH 3/8] x86/amd-iommu: Add passthrough mode initialization functions

[Joerg Roedel]

When iommu=pt is passed on kernel command line the devices
should run untranslated. This requires the allocation of a
special domain for that purpose. This patch implements the
allocation and initialization path for iommu=pt.

Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
---
 arch/x86/include/asm/amd_iommu.h       |    1 +
 arch/x86/include/asm/amd_iommu_types.h |    4 ++
 arch/x86/kernel/amd_iommu.c            |   72 ++++++++++++++++++++++++++++----
 3 files changed, 69 insertions(+), 8 deletions(-)

diff --git a/arch/x86/include/asm/amd_iommu.h b/arch/x86/include/asm/amd_iommu.h
index bdf96f1..ac95995 100644
--- a/arch/x86/include/asm/amd_iommu.h
+++ b/arch/x86/include/asm/amd_iommu.h
@@ -25,6 +25,7 @@
 #ifdef CONFIG_AMD_IOMMU
 extern int amd_iommu_init(void);
 extern int amd_iommu_init_dma_ops(void);
+extern int amd_iommu_init_passthrough(void);
 extern void amd_iommu_detect(void);
 extern irqreturn_t amd_iommu_int_handler(int irq, void *data);
 extern void amd_iommu_flush_all_domains(void);
diff --git a/arch/x86/include/asm/amd_iommu_types.h b/arch/x86/include/asm/amd_iommu_types.h
index 0c878ca..49f7453 100644
--- a/arch/x86/include/asm/amd_iommu_types.h
+++ b/arch/x86/include/asm/amd_iommu_types.h
@@ -143,6 +143,7 @@
 #define EVT_BUFFER_SIZE		8192 /* 512 entries */
 #define EVT_LEN_MASK		(0x9ULL << 56)
 
+#define PAGE_MODE_NONE    0x00
 #define PAGE_MODE_1_LEVEL 0x01
 #define PAGE_MODE_2_LEVEL 0x02
 #define PAGE_MODE_3_LEVEL 0x03
@@ -194,6 +195,9 @@
 #define PD_DMA_OPS_MASK		(1UL << 0) /* domain used for dma_ops */
 #define PD_DEFAULT_MASK		(1UL << 1) /* domain is a default dma_ops
 					      domain for an IOMMU */
+#define PD_PASSTHROUGH_MASK	(1UL << 2) /* domain has no page
+					      translation */
+
 extern bool amd_iommu_dump;
 #define DUMP_printk(format, arg...)					\
 	do {								\
diff --git a/arch/x86/kernel/amd_iommu.c b/arch/x86/kernel/amd_iommu.c
index 0934348..7987f20 100644
--- a/arch/x86/kernel/amd_iommu.c
+++ b/arch/x86/kernel/amd_iommu.c
@@ -41,6 +41,12 @@ static DEFINE_RWLOCK(amd_iommu_devtable_lock);
 static LIST_HEAD(iommu_pd_list);
 static DEFINE_SPINLOCK(iommu_pd_list_lock);
 
+/*
+ * Domain for untranslated devices - only allocated
+ * if iommu=pt passed on kernel cmd line.
+ */
+static struct protection_domain *pt_domain;
+
 #ifdef CONFIG_IOMMU_API
 static struct iommu_ops amd_iommu_ops;
 #endif
@@ -1067,9 +1073,9 @@ static struct protection_domain *domain_for_device(u16 devid)
  * If a device is not yet associated with a domain, this function does
  * assigns it visible for the hardware
  */
-static void attach_device(struct amd_iommu *iommu,
-			  struct protection_domain *domain,
-			  u16 devid)
+static void __attach_device(struct amd_iommu *iommu,
+			    struct protection_domain *domain,
+			    u16 devid)
 {
 	unsigned long flags;
 	u64 pte_root = virt_to_phys(domain->pt_root);
@@ -1087,12 +1093,19 @@ static void attach_device(struct amd_iommu *iommu,
 
 	amd_iommu_pd_table[devid] = domain;
 	write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
+}
 
-       /*
-        * We might boot into a crash-kernel here. The crashed kernel
-        * left the caches in the IOMMU dirty. So we have to flush
-        * here to evict all dirty stuff.
-        */
+static void attach_device(struct amd_iommu *iommu,
+			  struct protection_domain *domain,
+			  u16 devid)
+{
+	__attach_device(iommu, domain, devid);
+
+	/*
+	 * We might boot into a crash-kernel here. The crashed kernel
+	 * left the caches in the IOMMU dirty. So we have to flush
+	 * here to evict all dirty stuff.
+	 */
 	iommu_queue_inv_dev_entry(iommu, devid);
 	iommu_flush_tlb_pde(iommu, domain->id);
 }
@@ -2219,3 +2232,46 @@ static struct iommu_ops amd_iommu_ops = {
 	.domain_has_cap = amd_iommu_domain_has_cap,
 };
 
+/*****************************************************************************
+ *
+ * The next functions do a basic initialization of IOMMU for pass through
+ * mode
+ *
+ * In passthrough mode the IOMMU is initialized and enabled but not used for
+ * DMA-API translation.
+ *
+ *****************************************************************************/
+
+int __init amd_iommu_init_passthrough(void)
+{
+	struct pci_dev *dev = NULL;
+	u16 devid, devid2;
+
+	/* allocate passthroug domain */
+	pt_domain = protection_domain_alloc();
+	if (!pt_domain)
+		return -ENOMEM;
+
+	pt_domain->mode |= PAGE_MODE_NONE;
+
+	while ((dev = pci_get_device(PCI_ANY_ID, PCI_ANY_ID, dev)) != NULL) {
+		struct amd_iommu *iommu;
+
+		devid = calc_devid(dev->bus->number, dev->devfn);
+		if (devid > amd_iommu_last_bdf)
+			continue;
+
+		devid2 = amd_iommu_alias_table[devid];
+
+		iommu = amd_iommu_rlookup_table[devid2];
+		if (!iommu)
+			continue;
+
+		__attach_device(iommu, pt_domain, devid);
+		__attach_device(iommu, pt_domain, devid2);
+	}
+
+	pr_info("AMD-Vi: Initialized for Passthrough Mode\n");
+
+	return 0;
+}
-- 
1.6.3.3

[PATCH 4/8] x86/amd-iommu: Fix device table write order

[Joerg Roedel]

The V bit of the device table entry has to be set after the
rest of the entry is written to not confuse the hardware.

Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
---
 arch/x86/kernel/amd_iommu.c |    4 ++--
 1 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/arch/x86/kernel/amd_iommu.c b/arch/x86/kernel/amd_iommu.c
index 7987f20..2b1e77c 100644
--- a/arch/x86/kernel/amd_iommu.c
+++ b/arch/x86/kernel/amd_iommu.c
@@ -1087,9 +1087,9 @@ static void __attach_device(struct amd_iommu *iommu,
 	pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
 
 	write_lock_irqsave(&amd_iommu_devtable_lock, flags);
-	amd_iommu_dev_table[devid].data[0] = lower_32_bits(pte_root);
-	amd_iommu_dev_table[devid].data[1] = upper_32_bits(pte_root);
 	amd_iommu_dev_table[devid].data[2] = domain->id;
+	amd_iommu_dev_table[devid].data[1] = upper_32_bits(pte_root);
+	amd_iommu_dev_table[devid].data[0] = lower_32_bits(pte_root);
 
 	amd_iommu_pd_table[devid] = domain;
 	write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
-- 
1.6.3.3

[PATCH 5/8] x86/amd-iommu: Align locking between attach_device and detach_device

[Joerg Roedel]

This patch makes the locking behavior between the functions
attach_device and __attach_device consistent with the
locking behavior between detach_device and __detach_device.

Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
---
 arch/x86/kernel/amd_iommu.c |   19 ++++++++++++++-----
 1 files changed, 14 insertions(+), 5 deletions(-)

diff --git a/arch/x86/kernel/amd_iommu.c b/arch/x86/kernel/amd_iommu.c
index 2b1e77c..9aa135d 100644
--- a/arch/x86/kernel/amd_iommu.c
+++ b/arch/x86/kernel/amd_iommu.c
@@ -1077,29 +1077,38 @@ static void __attach_device(struct amd_iommu *iommu,
 			    struct protection_domain *domain,
 			    u16 devid)
 {
-	unsigned long flags;
-	u64 pte_root = virt_to_phys(domain->pt_root);
+	u64 pte_root;
 
-	domain->dev_cnt += 1;
+	/* lock domain */
+	spin_lock(&domain->lock);
+
+	pte_root = virt_to_phys(domain->pt_root);
 
 	pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
 		    << DEV_ENTRY_MODE_SHIFT;
 	pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
 
-	write_lock_irqsave(&amd_iommu_devtable_lock, flags);
 	amd_iommu_dev_table[devid].data[2] = domain->id;
 	amd_iommu_dev_table[devid].data[1] = upper_32_bits(pte_root);
 	amd_iommu_dev_table[devid].data[0] = lower_32_bits(pte_root);
 
 	amd_iommu_pd_table[devid] = domain;
-	write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
+
+	domain->dev_cnt += 1;
+
+	/* ready */
+	spin_unlock(&domain->lock);
 }
 
 static void attach_device(struct amd_iommu *iommu,
 			  struct protection_domain *domain,
 			  u16 devid)
 {
+	unsigned long flags;
+
+	write_lock_irqsave(&amd_iommu_devtable_lock, flags);
 	__attach_device(iommu, domain, devid);
+	write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
 
 	/*
 	 * We might boot into a crash-kernel here. The crashed kernel
-- 
1.6.3.3

[PATCH 6/8] x86/amd-iommu: Make sure a device is assigned in passthrough mode

[Joerg Roedel]

When the IOMMU driver runs in passthrough mode it has to
make sure that every device not assigned to an IOMMU-API
domain must be put into the passthrough domain instead of
keeping it unassigned.

Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
---
 arch/x86/kernel/amd_iommu.c |    9 +++++++++
 1 files changed, 9 insertions(+), 0 deletions(-)

diff --git a/arch/x86/kernel/amd_iommu.c b/arch/x86/kernel/amd_iommu.c
index 9aa135d..a8e74c3 100644
--- a/arch/x86/kernel/amd_iommu.c
+++ b/arch/x86/kernel/amd_iommu.c
@@ -1141,6 +1141,15 @@ static void __detach_device(struct protection_domain *domain, u16 devid)
 
 	/* ready */
 	spin_unlock(&domain->lock);
+
+	/*
+	 * If we run in passthrough mode the device must be assigned to the
+	 * passthrough domain if it is detached from any other domain
+	 */
+	if (iommu_pass_through) {
+		struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
+		__attach_device(iommu, pt_domain, devid);
+	}
 }
 
 /*
-- 
1.6.3.3

[PATCH 7/8] x86/amd-iommu: Don't detach device from pt domain on driver unbind

[Joerg Roedel]

This patch makes sure a device is not detached from the
passthrough domain when the device driver is unloaded or
does otherwise release the device.

Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
---
 arch/x86/kernel/amd_iommu.c |    2 ++
 1 files changed, 2 insertions(+), 0 deletions(-)

diff --git a/arch/x86/kernel/amd_iommu.c b/arch/x86/kernel/amd_iommu.c
index a8e74c3..12a541d 100644
--- a/arch/x86/kernel/amd_iommu.c
+++ b/arch/x86/kernel/amd_iommu.c
@@ -1195,6 +1195,8 @@ static int device_change_notifier(struct notifier_block *nb,
 	case BUS_NOTIFY_UNBOUND_DRIVER:
 		if (!domain)
 			goto out;
+		if (iommu_pass_through)
+			break;
 		detach_device(domain, devid);
 		break;
 	case BUS_NOTIFY_ADD_DEVICE:
-- 
1.6.3.3

[PATCH 8/8] x86/amd-iommu: Initialize passthrough mode when requested

[Joerg Roedel]

This patch enables the passthrough mode for AMD IOMMU by
running the initialization function when iommu=pt is passed
on the kernel command line.

Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
---
 arch/x86/kernel/amd_iommu_init.c |    8 +++++++-
 1 files changed, 7 insertions(+), 1 deletions(-)

diff --git a/arch/x86/kernel/amd_iommu_init.c b/arch/x86/kernel/amd_iommu_init.c
index c1b17e9..f00f489 100644
--- a/arch/x86/kernel/amd_iommu_init.c
+++ b/arch/x86/kernel/amd_iommu_init.c
@@ -1242,12 +1242,18 @@ int __init amd_iommu_init(void)
 	if (ret)
 		goto free;
 
-	ret = amd_iommu_init_dma_ops();
+	if (iommu_pass_through)
+		ret = amd_iommu_init_passthrough();
+	else
+		ret = amd_iommu_init_dma_ops();
 	if (ret)
 		goto free;
 
 	enable_iommus();
 
+	if (iommu_pass_through)
+		goto out;
+
 	printk(KERN_INFO "AMD IOMMU: device isolation ");
 	if (amd_iommu_isolate)
 		printk("enabled\n");
-- 
1.6.3.3

Re: [PATCH 0/8] Passthrough support for AMD-IOMMU (AMD-Vi)

[David Woodhouse]

On Fri, 2009-09-04 at 11:36 +0200, Joerg Roedel wrote:
> 
> this series of patches implement iommu=pt support for AMD IOMMU (AMD-Vi). If
> this option is passed on the kernel command line, the AMD-Vi driver will not
> register itself as the dma_ops backend and allows all devices unlimited access
> to main memory as long as no other kernel part (currently only KVM will do so)
> assigns the device to another domain using the IOMMU-API.

The Intel IOMMU used to behave like that in passthrough mode, but I've
now changed it. Now, it will still register itself as the dma_ops
backend, and it won't actually bypass the IOMMU for devices with a
limited dma_mask. So you don't end up having to use swiotlb for those
(which wasn't working anyway, iirc).

The devices for which you actually care about performance are still
going to end up with the passthrough mapping as you desire, anyway.

-- 
David Woodhouse                            Open Source Technology Centre
David.Woodhouse@intel.com                              Intel Corporation