Skip to contentSkip to conversation

iommu/vt-d: Handle hotplug devices' default identity mapping setting

PATCH In review James Dong 4 participants 11 messages Feb 22, 2019 – Feb 24, 2019 2 messages not in the archive IOMMU

lore
Share & tools
lore

James Dong

author
#0 lore raw

PATCH 1/1 iommu/vt-d: Handle hotplug devices' default identity mapping setting

With specific Linux command line option such as "iommu=pt", some types of hotplug devices will be assigned to the static 1:1 iommu domain by default during kernel init.

If such a hotplug device is rescanned after being removed, for example, with following sample commands:
echo 1 > /sys/bus/pci/devices/0000\:03\:00.1/remove
echo 1 > /sys/bus/pci/rescan
, an iommu group for this hotplug device will be added without attaching it to the default static 1:1 domain. As a result, this device's I/O operation won't work.

Keep hotplug devices with default static 1:1 iommu mapping in their default domain after such hotplug devices are rescanned.

---
 drivers/iommu/intel-iommu.c | 14 +++++++++++++-
 1 file changed, 13 insertions(+), 1 deletion(-)
drivers/iommu/intel-iommu.c +13 -1
index 78188bf7e90d..4b02949e58ca 100644
@@ -5230,6 +5230,7 @@ static int intel_iommu_add_device(struct device *dev)
 	struct intel_iommu *iommu;
 	struct iommu_group *group;
 	u8 bus, devfn;
+	int ret = 0;
 
 	iommu = device_to_iommu(dev, &bus, &devfn);
 	if (!iommu)
@@ -5242,8 +5243,19 @@ static int intel_iommu_add_device(struct device *dev)
 	if (IS_ERR(group))
 		return PTR_ERR(group);
 
+	if (!iommu_group_default_domain(group) &&
+	    !find_domain(dev) && iommu_should_identity_map(dev, 0)) {
+		ret = domain_add_dev_info(si_domain, dev);
+		if (!ret)
+			pr_info("identity mapping for device %s\n",
+				dev_name(dev));
+		else
+			pr_info("identity mapping failed (%d) for device %s\n",
+				ret, dev_name(dev));
+	}
+
 	iommu_group_put(group);
-	return 0;
+	return ret;
 }
 
 static void intel_iommu_remove_device(struct device *dev)
Signature · 1 line
2.21.0.rc0.258.g878e2cd30e-goog
#1 lore raw

Hi,

Quoted 11 lines · James Dong

With specific Linux command line option such as "iommu=pt", some types of hotplug devices will be assigned to the static 1:1 iommu domain by default during kernel init.

If such a hotplug device is rescanned after being removed, for example, with following sample commands:
echo 1 > /sys/bus/pci/devices/0000\:03\:00.1/remove
echo 1 > /sys/bus/pci/rescan
, an iommu group for this hotplug device will be added without attaching it to the default static 1:1 domain. As a result, this device's I/O operation won't work.

What does "I/O operation won't work" exactly mean here? Do you see any IOMMU fault message? Or, something doesn't work as expected?

Quoted 27 lines

Keep hotplug devices with default static 1:1 iommu mapping in their default domain after such hotplug devices are rescanned.

---
  drivers/iommu/intel-iommu.c | 14 +++++++++++++-
  1 file changed, 13 insertions(+), 1 deletion(-)
drivers/iommu/intel-iommu.c +2 -0
index 78188bf7e90d..4b02949e58ca 100644
@@ -5230,6 +5230,7 @@ static int intel_iommu_add_device(struct device *dev)
  	struct intel_iommu *iommu;
  	struct iommu_group *group;
  	u8 bus, devfn;
+	int ret = 0;
 
  	iommu = device_to_iommu(dev, &bus, &devfn);
  	if (!iommu)
@@ -5242,8 +5243,19 @@ static int intel_iommu_add_device(struct device *dev)
  	if (IS_ERR(group))
  		return PTR_ERR(group);
 
+	if (!iommu_group_default_domain(group) &&

drivers/iommu/intel-iommu.c static int intel_iommu_add_device(struct device *dev) · line 5256

Quoted 10 lines
drivers/iommu/intel-iommu.c +10 -0
+	    !find_domain(dev) && iommu_should_identity_map(dev, 0)) {
+		ret = domain_add_dev_info(si_domain, dev);
+		if (!ret)
+			pr_info("identity mapping for device %s\n",
+				dev_name(dev));
+		else
+			pr_info("identity mapping failed (%d) for device %s\n",
+				ret, dev_name(dev));
+	}
+

I am not sure about whether this is a good place to fix this problem before Intel IOMMU driver switches to default domain.

Do you mind checking this?

drivers/iommu/intel-iommu.c +3 -0
index 6ecdcf8fc8c0..f62f30bc1339 100644
@@ -2632,6 +2632,9 @@ static struct dmar_domain *find_or_alloc_domain(struct device *dev, int gaw)
                        goto out;
        }
 
+       if (!iommu_should_identity_map(dev, 0))
+               return si_domain;
+
        /* Allocate and initialize new domain for the device */
        domain = alloc_domain(0);
        if (!domain)

Best regards,
Lu Baolu

#2 lore raw

Hi

Quoted 16 lines · Lu Baolu

Do you mind checking this?

drivers/iommu/intel-iommu.c +3 -0
index 6ecdcf8fc8c0..f62f30bc1339 100644
@@ -2632,6 +2632,9 @@ static struct dmar_domain 
 *find_or_alloc_domain(struct device *dev, int gaw)
                         goto out;
         }
 
+       if (!iommu_should_identity_map(dev, 0))
+               return si_domain;
+
         /* Allocate and initialize new domain for the device */
         domain = alloc_domain(0);
Quoted 1 line

        if (!domain)

Oops! This can't be compiled. Please try below instead if you'd like to. Sorry about it.

drivers/iommu/intel-iommu.c +5 -0
index 6ecdcf8fc8c0..b89f5ba6a3c8 100644
@@ -174,6 +174,8 @@ static inline unsigned long virt_to_dma_pfn(void *p)
        return page_to_dma_pfn(virt_to_page(p));
 }
 
+static int iommu_should_identity_map(struct device *dev, int startup);
+
 /* global iommu list, set NULL for ignored DMAR units */
 static struct intel_iommu **g_iommus;
 
@@ -2632,6 +2634,9 @@ static struct dmar_domain *find_or_alloc_domain(struct device *dev, int gaw)
                        goto out;
        }
 
+       if (iommu_should_identity_map(dev, 0))
+               return si_domain;
+
        /* Allocate and initialize new domain for the device */
        domain = alloc_domain(0);
        if (!domain)

Best regards,
Lu Baolu

A message that is not in the archive (sent off-list or not yet received): fe6dcb4e-03c4-0bbf-c51f-5964b20593dd-VuQAYsv1563Yd54FQh9/[email protected]

#4 lore raw

Baolu:

Thanks for your comments and your patch. Please find below our responses to each of your comments:

Quoted 2 lines

What does "I/O operation won't work" exactly mean here? Do you see any IOMMU fault message? Or, something doesn't work as expected?

Yes, DMAR fault messages as following came out:

[  354.939896] DMAR: DMAR:[DMA Read] Request device [03:00.1]fault addr
1fdfe80000
[  354.939896] DMAR:[fault reason 02] Present bit in context entry is clear
Quoted 15 lines

Do you mind checking this?

index 6ecdcf8fc8c0..f62f30bc1339 100644

drivers/iommu/intel-iommu.c +3 -0
@@ -2632,6 +2632,9 @@ static struct dmar_domain
 *find_or_alloc_domain(struct device *dev, int gaw)
                         goto out;
         }
 
+       if (!iommu_should_identity_map(dev, 0))
+               return si_domain;
+
         /* Allocate and initialize new domain for the device */
         domain = alloc_domain(0);
Quoted 1 line

if (!domain)

Tried this patch, and the same DMAR fault message came out.

Guess it is because of the iommu code path for hotplug devices. If a hotplug device is rescanned after removal, iommu_bus_notifier will be called as part of the notifier chains to handle BUS_NOTIFY_ADD_DEVICE event. Along the code path, intel_iommu_ops->add_device() created an iommu group for this hotplug device, but failed to create an iommu domain because of the default domain type IOMMU_DOMAIN_IDENTITY imposed by current IOMMU command line option got declined by intel_iommu_ops->domain_alloc().

In your patch, function find_or_alloc_domain() is not even in the code path of BUS_NOTIFY_ADD_DEVICE event notifier chain.

Please let us know if your have more concerns and suggestions.

Best Regards,
James

Quoted 55 lines · Lu Baolu

Hi

Quoted 16 lines · Lu Baolu

Do you mind checking this?

drivers/iommu/intel-iommu.c +3 -0
index 6ecdcf8fc8c0..f62f30bc1339 100644
@@ -2632,6 +2632,9 @@ static struct dmar_domain
 *find_or_alloc_domain(struct device *dev, int gaw)
                         goto out;
         }
 
+       if (!iommu_should_identity_map(dev, 0))
+               return si_domain;
+
         /* Allocate and initialize new domain for the device */
         domain = alloc_domain(0);
Quoted 1 line

if (!domain)

Oops! This can't be compiled. Please try below instead if you'd like to. Sorry about it.

drivers/iommu/intel-iommu.c +5 -0
index 6ecdcf8fc8c0..b89f5ba6a3c8 100644
@@ -174,6 +174,8 @@ static inline unsigned long virt_to_dma_pfn(void *p)
         return page_to_dma_pfn(virt_to_page(p));
  }
 
+static int iommu_should_identity_map(struct device *dev, int startup);
+
  /* global iommu list, set NULL for ignored DMAR units */
  static struct intel_iommu **g_iommus;
 
@@ -2632,6 +2634,9 @@ static struct dmar_domain
 *find_or_alloc_domain(struct device *dev, int gaw)
                         goto out;
         }
 
+       if (iommu_should_identity_map(dev, 0))
+               return si_domain;
+
         /* Allocate and initialize new domain for the device */
         domain = alloc_domain(0);

if (!domain)

Best regards,
Lu Baolu

#5 lore raw

Hi James,

Quoted 11 lines · James Dong

Tried this patch, and the same DMAR fault message came out.

Guess it is because of the iommu code path for hotplug devices. If a hotplug device is rescanned after removal, iommu_bus_notifier will be called as part of the notifier chains to handle BUS_NOTIFY_ADD_DEVICE event. Along the code path, intel_iommu_ops->add_device() created an iommu group for this hotplug device, but failed to create an iommu domain because of the default domain type IOMMU_DOMAIN_IDENTITY imposed by current IOMMU command line option got declined by intel_iommu_ops->domain_alloc().

Quoted 4 lines

In your patch, function find_or_alloc_domain() is not even in the code path of BUS_NOTIFY_ADD_DEVICE event notifier chain.

Please let us know if your have more concerns and suggestions.

Can I reproduce this with a local machine? If so, how should I do?

Quoted 3 lines

Best Regards,
James

Best regards,
Lu Baolu

author
#6 lore raw

Baolu:

The reproduction depends on devices. HW passthrough PCIe devices with default identity map could have the issue. Make sure that messages like following came out in dmesg and their mapping does not change after booting:

  [   10.167809] DMAR: Hardware identity mapping for device 0000:30:00.0
  [   10.167823] DMAR: Hardware identity mapping for device 0000:30:00.1

Devices which make following true could also be used for the experiment:

Quoted 6 lines

static int iommu_should_identity_map(struct device *dev, int startup)

{
		if ((iommu_identity_mapping & IDENTMAP_AZALIA) && IS_AZALIA(pdev))
			return 1;
		if ((iommu_identity_mapping & IDENTMAP_GFX) && IS_GFX_DEVICE(pdev))
			return 1;

Once they are up, remove them first by following command:
echo 1 > /sys/bus/pci/devices/0000\:03\:00.1/remove

Then trigger the hotplug device rescanning:
echo 1 > /sys/bus/pci/rescan

To provide an example of specific devices on the market, I need to try out. Or, if it is fine with you, forcing a PCIe NIC card to be default hardware passthrough by changing the intel-iommu.c is another easy way to reproduce this issue.

Best Regards,
James

A message that is not in the archive (sent off-list or not yet received): 20190222082829.75567-1-xmdong-hpIqsD4AKlfQT0dZR+AlfA@public.gmane.org

#8 lore raw

An NVMe device configured in static identity mapping should also cause this error when removed and rescanned. Essentially, a device that does a DMA when its driver inits, or one that you can force a DMA from.

Quoted 34 lines · James Dong

Baolu:

The reproduction depends on devices. HW passthrough PCIe devices with default
identity map could have the issue. Make sure that messages like following came
out in dmesg and their mapping does not change after booting:

  [   10.167809] DMAR: Hardware identity mapping for device 0000:30:00.0
  [   10.167823] DMAR: Hardware identity mapping for device 0000:30:00.1

Devices which make following true could also be used for the experiment:

Quoted 3 lines

static int iommu_should_identity_map(struct device *dev, int startup)

{
            if ((iommu_identity_mapping & IDENTMAP_AZALIA) &&

IS_AZALIA(pdev))

Quoted 2 lines
                    return 1;
            if ((iommu_identity_mapping & IDENTMAP_GFX) &&

IS_GFX_DEVICE(pdev))

Quoted 1 line
                    return 1;

Once they are up, remove them first by following command:
echo 1 > /sys/bus/pci/devices/0000\:03\:00.1/remove

Then trigger the hotplug device rescanning:
echo 1 > /sys/bus/pci/rescan

To provide an example of specific devices on the market, I need to try out. Or, if it is fine with you, forcing a PCIe NIC card to be default hardware passthrough by changing the intel-iommu.c is another easy way to reproduce this issue.

Best Regards,
James

#9 lore raw
In reply to James Dong

An NVMe device configured in static identity mapping should also cause this error when removed and rescanned. Essentially, a device that does a DMA when its driver inits, or one that you can force a DMA from.

-Jis

Quoted 30 lines · James Dong

Baolu:

The reproduction depends on devices. HW passthrough PCIe devices with default identity map could have the issue. Make sure that messages like following came out in dmesg and their mapping does not change after booting:

  [   10.167809] DMAR: Hardware identity mapping for device 0000:30:00.0
  [   10.167823] DMAR: Hardware identity mapping for device 0000:30:00.1

Devices which make following true could also be used for the experiment:

Quoted 6 lines

static int iommu_should_identity_map(struct device *dev, int startup)

{
            if ((iommu_identity_mapping & IDENTMAP_AZALIA) && IS_AZALIA(pdev))
                    return 1;
            if ((iommu_identity_mapping & IDENTMAP_GFX) && IS_GFX_DEVICE(pdev))
                    return 1;

Once they are up, remove them first by following command:
echo 1 > /sys/bus/pci/devices/0000\:03\:00.1/remove

Then trigger the hotplug device rescanning:
echo 1 > /sys/bus/pci/rescan

To provide an example of specific devices on the market, I need to try out. Or, if it is fine with you, forcing a PCIe NIC card to be default hardware passthrough by changing the intel-iommu.c is another easy way to reproduce this issue.

Best Regards,
James

author
#10 lore raw
In reply to Lu Baolu

Baolu:

Sorry that my last reply email seems not text format. Resend it now.

Thanks for your comments and your patch. Please find below our responses to each of your comments:

Quoted 2 lines

What does "I/O operation won't work" exactly mean here? Do you see any IOMMU fault message? Or, something doesn't work as expected?

Yes, DMAR fault messages as following came out:

[  354.939896] DMAR: DMAR:[DMA Read] Request device [03:00.1]fault addr 1fdfe80000 
[  354.939896] DMAR:[fault reason 02] Present bit in context entry is clear
Quoted 15 lines

Do you mind checking this?

index 6ecdcf8fc8c0..f62f30bc1339 100644

drivers/iommu/intel-iommu.c +3 -0
@@ -2632,6 +2632,9 @@ static struct dmar_domain 
 *find_or_alloc_domain(struct device *dev, int gaw)
                         goto out;
         }
 
+       if (!iommu_should_identity_map(dev, 0))
+               return si_domain;
+
         /* Allocate and initialize new domain for the device */
         domain = alloc_domain(0);
Quoted 1 line

if (!domain)

Tried this patch, and the same DMAR fault message came out.

Guess it is because of the iommu code path for hotplug devices. If a hotplug device is rescanned after removal, iommu_bus_notifier will be called as part of the notifier chains to handle BUS_NOTIFY_ADD_DEVICE event. Along the code path, intel_iommu_ops->add_device() created an iommu group for this hotplug device, but failed to create an iommu domain because of the default domain type IOMMU_DOMAIN_IDENTITY imposed by current IOMMU command line option got declined by intel_iommu_ops->domain_alloc().

Since si_domain is type of "struct dmar_domain", which is platform dependent, it is hard to make this change in intel_iommu_ops->domain_alloc().

In your patch, function find_or_alloc_domain() is not in the code path of BUS_NOTIFY_ADD_DEVICE event notifier chain.

Please let us know if your have more concerns and suggestions.

Best Regards,
James

#11 lore raw

Hi James,

Quoted 25 lines · James Dong

Baolu:

Sorry that my last reply email seems not text format. Resend it now.

Thanks for your comments and your patch. Please find below our responses to each of your comments:

Quoted 2 lines

What does "I/O operation won't work" exactly mean here? Do you see any IOMMU fault message? Or, something doesn't work as expected?

Yes, DMAR fault messages as following came out:

[  354.939896] DMAR: DMAR:[DMA Read] Request device [03:00.1]fault addr 1fdfe80000
[  354.939896] DMAR:[fault reason 02] Present bit in context entry is clear
Quoted 10 lines

Do you mind checking this?

index 6ecdcf8fc8c0..f62f30bc1339 100644

drivers/iommu/intel-iommu.c +0 -0
@@ -2632,6 +2632,9 @@ static struct dmar_domain
 *find_or_alloc_domain(struct device *dev, int gaw)
                          goto out;
          }
 
Quoted 8 lines
Quoted 6 lines
drivers/iommu/intel-iommu.c +3 -0
+       if (!iommu_should_identity_map(dev, 0))
+               return si_domain;
+
          /* Allocate and initialize new domain for the device */
          domain = alloc_domain(0);

if (!domain)

Tried this patch, and the same DMAR fault message came out.

Thank you!

Quoted 8 lines

Guess it is because of the iommu code path for hotplug devices. If a hotplug device is rescanned after removal, iommu_bus_notifier will be called as part of the notifier chains to handle BUS_NOTIFY_ADD_DEVICE event. Along the code path, intel_iommu_ops->add_device() created an iommu group for this hotplug device, but failed to create an iommu domain because of the default domain type IOMMU_DOMAIN_IDENTITY imposed by current IOMMU command line option got declined by intel_iommu_ops->domain_alloc().

The Intel IOMMU driver hasn't switched to default domain yet although it's in the pipe line. So, there should be no domain allocated when a group is allocated for the device.

The problem is we need to check whether a hot-added device requires identity map instead of allocating a normal domain blindly.

Quoted 8 lines

Since si_domain is type of "struct dmar_domain", which is platform dependent, it is hard to make this change in intel_iommu_ops->domain_alloc().

In your patch, function find_or_alloc_domain() is not in the code path of BUS_NOTIFY_ADD_DEVICE event notifier chain.

Please let us know if your have more concerns and suggestions.

Can you please try the patch attached? I think this is a generic issue as I described in the commit message.

Best regards,
Lu Baolu

Quoted 1 line

From d942e60557fc7ea6fee535fb9a0a7d334d65b636 Mon Sep 17 00:00:00 2001

From: Lu Baolu <baolu.lu-VuQAYsv1563Yd54FQh9/[email protected]>
Date: Sun, 24 Feb 2019 10:01:03 +0800
Subject: [PATCH 1/1] iommu/vt-d: Check identity map for hot-added devices

The Intel IOMMU driver will put devices into a static identity mapped domain during boot if the kernel parameter "iommu=pt" is used. That means the IOMMU hardware will translate a DMA address into the same memory address.

Unfortunately, a hot-added device doesn't subject to this. That results in some devices not working properly after hot added. A quick way to reproduce this issue is to boot a system with

iommu=pt

and, remove then readd the pci device with

echo 1 > /sys/bus/pci/devices/[pci_source_id]/remove echo 1 > /sys/bus/pci/rescan

You will find the identity mapped domain was replaced with a normal domain.

Cc Ashok Raj <[email protected]>
Cc Jacob Pan <jacob.jun.pan-VuQAYsv1563Yd54FQh9/[email protected]>
---
 drivers/iommu/intel-iommu.c | 27 +++++++++++++++------------
 1 file changed, 15 insertions(+), 12 deletions(-)
drivers/iommu/intel-iommu.c +15 -12
index 6ecdcf8fc8c0..730ee29d561b 100644
@@ -3001,9 +3001,9 @@ static int iommu_should_identity_map(struct device *dev, int startup)
 	}
 
 	/*
-	 * At boot time, we don't yet know if devices will be 64-bit capable.
-	 * Assume that they will — if they turn out not to be, then we can
-	 * take them out of the 1:1 domain later.
+	 * At boot time or hot added, we don't yet know if devices will be
+	 * 64-bit capable. Assume that they will — if they turn out not to
+	 * be, then we can take them out of the 1:1 domain later.
 	 */
 	if (!startup) {
 		/*
@@ -4807,16 +4807,19 @@ static int device_notifier(struct notifier_block *nb,
 	if (iommu_dummy(dev))
 		return 0;
 
-	if (action != BUS_NOTIFY_REMOVED_DEVICE)
-		return 0;
-
-	domain = find_domain(dev);
-	if (!domain)
-		return 0;
+	if (action == BUS_NOTIFY_REMOVED_DEVICE) {
+		domain = find_domain(dev);
+		if (!domain)
+			return 0;
 
-	dmar_remove_one_dev_info(dev);
-	if (!domain_type_is_vm_or_si(domain) && list_empty(&domain->devices))
-		domain_exit(domain);
+		dmar_remove_one_dev_info(dev);
+		if (!domain_type_is_vm_or_si(domain) &&
+		    list_empty(&domain->devices))
+			domain_exit(domain);
+	} else if (action == BUS_NOTIFY_ADD_DEVICE) {
+		if (iommu_should_identity_map(dev, 1))
+			domain_add_dev_info(si_domain, dev);
+	}
 
 	return 0;
 }
Signature · 1 line
2.17.1
author
#12 lore raw

Baolu:

Yes, it is a generic issue for hotplug devices with current Intel IOMMU driver, as reported in this thread as well.

The patch you provided does the job in our case. Please update this thread once your patch is merged. Thanks.

Best Regards,
James

On 2/23/19 12:56 AM, Lu Baolu wrote

Quoted 27 lines
drivers/iommu/intel-iommu.c +12 -9
@@ -4807,16 +4807,19 @@ static int device_notifier(struct notifier_block *nb,
 	if (iommu_dummy(dev))
 		return 0;
 
-	if (action != BUS_NOTIFY_REMOVED_DEVICE)
-		return 0;
-
-	domain = find_domain(dev);
-	if (!domain)
-		return 0;
+	if (action == BUS_NOTIFY_REMOVED_DEVICE) {
+		domain = find_domain(dev);
+		if (!domain)
+			return 0;
 
-	dmar_remove_one_dev_info(dev);
-	if (!domain_type_is_vm_or_si(domain) && list_empty(&domain->devices))
-		domain_exit(domain);
+		dmar_remove_one_dev_info(dev);
+		if (!domain_type_is_vm_or_si(domain) &&
+		    list_empty(&domain->devices))
+			domain_exit(domain);
+	} else if (action == BUS_NOTIFY_ADD_DEVICE) {
+		if (iommu_should_identity_map(dev, 1))
+			domain_add_dev_info(si_domain, dev);
+	}