iommu/vt-d: Keep RMRR mappings around on driver unbind
Subsystems (2)
Joerg Roedel
maintainercover iommu/vt-d: Keep RMRR mappings around on driver unbind
Hi,
here is a patch-set to fix an issue recently discovered when the Intel IOMMU is in use with devices that need RMRR mappings.
The problem is that the RMRR mappings are destroyed when the device driver is unbound from the device, causing DMAR faults.
To solve this problem a device driver core change is necessary to catch the right point in time for the IOMMU code to destroy any mappings for a device.
With this patch-set the RMRR mappings are only destroyed when the device is actually removed from the system.
Please review.
Thanks,
Joerg
Joerg Roedel (2):
driver core: Add BUS_NOTIFY_REMOVED_DEVICE event iommu/vt-d: Only remove domain when device is removed
drivers/base/core.c | 3 +++ drivers/iommu/intel-iommu.c | 11 +---------- include/linux/device.h | 11 ++++++----- 3 files changed, 10 insertions(+), 15 deletions(-)
Signature · 1 line
1.9.1
Joerg Roedel
maintainerPATCH 1/2 driver core: Add BUS_NOTIFY_REMOVED_DEVICE event
From: Joerg Roedel <[email protected]>
This event closes an important gap in the bus notifiers. There is already the BUS_NOTIFY_DEL_DEVICE event, but that is sent when the device is still bound to its device driver.
This is too early for the IOMMU code to destroy any mappings for the device, as they might still be in use by the driver.
The new BUS_NOTIFY_REMOVED_DEVICE event introduced with this patch closes this gap as it is sent when the device is already unbound from its device driver and almost completly removed from the driver core.
With this event the IOMMU code can safely destroy any mappings and other data structures when a device is removed.
--- drivers/base/core.c | 3 +++ include/linux/device.h | 11 ++++++----- 2 files changed, 9 insertions(+), 5 deletions(-)
drivers/base/core.c +3 -0
@@ -1211,6 +1211,9 @@ void device_del(struct device *dev)*/ if (platform_notify_remove)platform_notify_remove (dev); +if (dev->bus) +blocking_notifier_call_chain (&dev->bus->p->bus_notifier, +BUS_NOTIFY_REMOVED_DEVICE , dev);kobject_uevent (&dev->kobj,KOBJ_REMOVE );cleanup_device_parent (dev);kobject_del (&dev->kobj);
include/linux/device.h +6 -5
@@ -181,13 +181,14 @@ extern int bus_unregister_notifier(struct bus_type *bus,* with the device lock held in the core, so be careful. */ #define BUS_NOTIFY_ADD_DEVICE 0x00000001/* device added */ -#define BUS_NOTIFY_DEL_DEVICE 0x00000002/* device removed */ -#define BUS_NOTIFY_BIND_DRIVER 0x00000003/* driver about to be +#define BUS_NOTIFY_DEL_DEVICE 0x00000002/* device to be removed */ +#define BUS_NOTIFY_REMOVED_DEVICE 0x00000003/* device removed */ +#define BUS_NOTIFY_BIND_DRIVER 0x00000004/* driver about to be bound */ -#define BUS_NOTIFY_BOUND_DRIVER 0x00000004/* driver bound to device */ -#define BUS_NOTIFY_UNBIND_DRIVER 0x00000005/* driver about to be +#define BUS_NOTIFY_BOUND_DRIVER 0x00000005/* driver bound to device */ +#define BUS_NOTIFY_UNBIND_DRIVER 0x00000006/* driver about to be unbound */ -#define BUS_NOTIFY_UNBOUND_DRIVER 0x00000006/* driver is unbound +#define BUS_NOTIFY_UNBOUND_DRIVER 0x00000007/* driver is unbound from the device */ extern struct kset *bus_get_kset (struct bus_type *bus);
Signature · 1 line
1.9.1
Joerg Roedel
maintainerPATCH 2/2 iommu/vt-d: Only remove domain when device is removed
In reply to Joerg RoedelFrom: Joerg Roedel <[email protected]>
This makes sure any RMRR mappings stay in place when the driver is unbound from the device.
--- drivers/iommu/intel-iommu.c | 11 +---------- 1 file changed, 1 insertion(+), 10 deletions(-)
drivers/iommu/intel-iommu.c +1 -10
@@ -3865,16 +3865,7 @@ static int device_notifier(struct notifier_block *nb,if (iommu_dummy (dev))return 0 ; -if (action !=BUS_NOTIFY_UNBOUND_DRIVER && - action !=BUS_NOTIFY_DEL_DEVICE ) -return 0 ; - -/* -* If the device is still attached to a device driver we can't -* tear down the domain yet as DMA mappings may still be in use. -* Wait for the BUS_NOTIFY_UNBOUND_DRIVER event to do that. -*/ -if (action ==BUS_NOTIFY_DEL_DEVICE && dev->driver !=NULL ) +if (action !=BUS_NOTIFY_REMOVED_DEVICE )return 0 ; domain =find_domain (dev);
Signature · 1 line
1.9.1
Re: [PATCH 2/2] iommu/vt-d: Only remove domain when device is removed
Quoted 19 lines · Joerg Roedel
From: Joerg Roedel <[email protected]>
This makes sure any RMRR mappings stay in place when the driver is unbound from the device.
Signed-off-by Joerg Roedel <[email protected]>--- drivers/iommu/intel-iommu.c | 11 +---------- 1 file changed, 1 insertion(+), 10 deletions(-)drivers/iommu/intel-iommu.c +0
-1@@ -3865,16 +3865,7 @@ static int device_notifier(struct notifier_block *nb,if (iommu_dummy (dev))return 0 ; - if (action != BUS_NOTIFY_UNBOUND_DRIVER &&
drivers/iommu/intel-iommu.c static int device_notifier(struct notifier_block *nb, · line 3868
Quoted 10 lines
drivers/iommu/intel-iommu.c +1
-9- action != BUS_NOTIFY_DEL_DEVICE) -return 0 ; - -/* -* If the device is still attached to a device driver we can't -* tear down the domain yet as DMA mappings may still be in use. -* Wait for the BUS_NOTIFY_UNBOUND_DRIVER event to do that. -*/ -if (action ==BUS_NOTIFY_DEL_DEVICE && dev->driver !=NULL ) +if (action !=BUS_NOTIFY_REMOVED_DEVICE )
I haven't tested it, but I'm concerned whether this has introduced a domain leak. If we think about the case of unbinding a device from a host driver and attaching it to a domain through the IOMMU API, I think we used to count on this path to call domain_exit(), which made the domain_context_mapped() in intel_iommu_attach_device() "unlikely". With this change, isn't the test in intel_iommu_attach_device() now neither likely nor unlikely and we're only removing the dev_info from the domain and not destroying the domain itself? Thanks,
Alex
Re: [PATCH 2/2] iommu/vt-d: Only remove domain when device is removed
Hi Alex,
Quoted 8 lines · Alex Williamson
I haven't tested it, but I'm concerned whether this has introduced a domain leak. If we think about the case of unbinding a device from a host driver and attaching it to a domain through the IOMMU API, I think we used to count on this path to call
domain_exit(), which made thedomain_context_mapped()inintel_iommu_attach_device()"unlikely". With this change, isn't the test inintel_iommu_attach_device()now neither likely nor unlikely and we're only removing thedev_infofrom the domain and not destroying the domain itself? Thanks,
As I see it, there is no leak. The DMA-API domains are kept in the device_domain_list and re-used when the device driver re-attaches. But your are right that the unlikely in intel_iommu_attach_device() isn't true anymore. We could probably remove it.
Joerg
Re: [PATCH 2/2] iommu/vt-d: Only remove domain when device is removed
Quoted 12 lines · Joerg Roedel
Hi Alex,
Quoted 8 lines · Alex Williamson
I haven't tested it, but I'm concerned whether this has introduced a domain leak. If we think about the case of unbinding a device from a host driver and attaching it to a domain through the IOMMU API, I think we used to count on this path to call
domain_exit(), which made thedomain_context_mapped()inintel_iommu_attach_device()"unlikely". With this change, isn't the test inintel_iommu_attach_device()now neither likely nor unlikely and we're only removing thedev_infofrom the domain and not destroying the domain itself? Thanks,
Quoted 4 lines
As I see it, there is no leak. The DMA-API domains are kept in the
device_domain_listand re-used when the device driver re-attaches. But your are right that the unlikely inintel_iommu_attach_device()isn't true anymore. We could probably remove it.
But the domains are unlinked from device_domain_list using unlink_domain_info() which is called from both domain_remove_dev_info() and domain_remove_one_dev_info() which are both part of that more likely, unlikely branch in intel_iommu_attach_device(). So it seems like any time we switch a device from the DMA-API to the IOMMU-API, we lose the reference to the domain. Is that incorrect? I'll try to test. Thanks,
Alex
Re: [PATCH 2/2] iommu/vt-d: Only remove domain when device is removed
Quoted 18 lines · Alex Williamson
Quoted 12 lines · Joerg Roedel
Hi Alex,
Quoted 8 lines · Alex Williamson
I haven't tested it, but I'm concerned whether this has introduced a domain leak. If we think about the case of unbinding a device from a host driver and attaching it to a domain through the IOMMU API, I think we used to count on this path to call
domain_exit(), which made thedomain_context_mapped()inintel_iommu_attach_device()"unlikely". With this change, isn't the test inintel_iommu_attach_device()now neither likely nor unlikely and we're only removing thedev_infofrom the domain and not destroying the domain itself? Thanks,Quoted 4 lines
As I see it, there is no leak. The DMA-API domains are kept in the
device_domain_listand re-used when the device driver re-attaches. But your are right that the unlikely inintel_iommu_attach_device()isn't true anymore. We could probably remove it.
Quoted 6 lines
But the domains are unlinked from
device_domain_listusingunlink_domain_info()which is called from bothdomain_remove_dev_info()anddomain_remove_one_dev_info()which are both part of that more likely, unlikely branch inintel_iommu_attach_device(). So it seems like any time we switch a device from the DMA-API to the IOMMU-API, we lose the reference to the domain. Is that incorrect? I'll try to test.
Trying the simple approach, a printk in each of alloc_domain() and free_domain_mem(), this is what I see when I start and stop a VM with an assigned device:
alloc_domain(): ffff8801e22ac000free_domain_mem(ffff8801e22ac000)alloc_domain(): ffff8801e3425c80
The IOMMU API domain is alloc'd and free'd, then a new DMA-API domain is alloc'd. There are no frees of the DMA-API domain. Thanks,
Alex
Joerg Roedel
maintainerRe: [PATCH 2/2] iommu/vt-d: Only remove domain when device is removed
In reply to Alex WilliamsonQuoted 6 lines · Alex Williamson
But the domains are unlinked from
device_domain_listusingunlink_domain_info()which is called from bothdomain_remove_dev_info()anddomain_remove_one_dev_info()which are both part of that more likely, unlikely branch inintel_iommu_attach_device(). So it seems like any time we switch a device from the DMA-API to the IOMMU-API, we lose the reference to the domain. Is that incorrect? I'll try to test.
Okay, I thought a while about that and it looks like a real fix needs a rewrite of the domain handling code in the VT-d driver to better handle domain lifetime. We'll get this for free when we add default domains and more domain handling logic to the iommu core, so I think we don't need to start rewriting the VT-d driver for this.
But for the time being, here is a simple fix for the leak in iommu_attach_domain:
Quoted 1 line
From
d65b236d0f27fe3ef7ac4d12cceb0da67aec86ceMon Sep 17 00:00:00 2001
From: Joerg Roedel <[email protected]>
Date: Tue, 9 Dec 2014 12:56:45 +0100
Subject: [PATCH] iommu/vt-d: Fix dmar_domain leak in iommu_attach_device
Since commit 1196c2f a domain is only destroyed in the notifier path if it is hot-unplugged. This caused a domain leakage in iommu_attach_device when a driver was unbound from the device and bound to VFIO. In this case the device is attached to a new domain and unlinked from the old domain. At this point nothing points to the old domain anymore and its memory is leaked.
Fix this by explicitly freeing the old domain in iommu_attach_domain.
--- drivers/iommu/intel-iommu.c | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-)
drivers/iommu/intel-iommu.c +5 -2
@@ -4424,10 +4424,13 @@ static int intel_iommu_attach_device(struct iommu_domain *domain, old_domain =find_domain (dev);if (old_domain) { -if (domain_type_is_vm_or_si (dmar_domain)) +if (domain_type_is_vm_or_si (dmar_domain)) {domain_remove_one_dev_info (old_domain, dev); -else + }else {domain_remove_dev_info (old_domain); +if (list_empty (&old_domain->devices)) +domain_exit (old_domain); + } } }
Signature · 1 line
1.8.4.5
Re: [PATCH 2/2] iommu/vt-d: Only remove domain when device is removed
Quoted 8 lines · Joerg Roedel
Quoted 6 lines · Alex Williamson
But the domains are unlinked from
device_domain_listusingunlink_domain_info()which is called from bothdomain_remove_dev_info()anddomain_remove_one_dev_info()which are both part of that more likely, unlikely branch inintel_iommu_attach_device(). So it seems like any time we switch a device from the DMA-API to the IOMMU-API, we lose the reference to the domain. Is that incorrect? I'll try to test.
Quoted 7 lines
Okay, I thought a while about that and it looks like a real fix needs a rewrite of the domain handling code in the VT-d driver to better handle domain lifetime. We'll get this for free when we add default domains and more domain handling logic to the iommu core, so I think we don't need to start rewriting the VT-d driver for this.
But for the time being, here is a simple fix for the leak iniommu_attach_domain:
Joerg,
This patch doesn't seem to be fixing the memory leak.
I am testing with a 3.18.0-rc7 kernel applied to a RHEL 7.0 system.
I added printk in free_domain_mem and alloc_domain to first reproduce Alex's observation. I created a VM and assigned a PCI NIC w/ no associated RMRR to the VM.
The pattern I see is when starting the VM is:alloc_domain -> X
When powering off the VM:free_domain_mem(X)alloc_domain -> Y
I then applied the patch below and I still see the same pattern of two alloc_domain and one free_domain_mem when powering on/off the VM.
I added some additional instrumentation and I do not see the new call to domain_exit being executed. (See inline comments below.)
Jerry
Quoted 27 lines
Quoted 2 lines
From
d65b236d0f27fe3ef7ac4d12cceb0da67aec86ceMon Sep 17 00:00:00 2001From: Joerg Roedel <[email protected]>
Date: Tue, 9 Dec 2014 12:56:45 +0100
Subject: [PATCH] iommu/vt-d: Fixdmar_domainleak iniommu_attach_deviceSince commit 1196c2f a domain is only destroyed in the notifier path if it is hot-unplugged. This caused a domain leakage in
iommu_attach_devicewhen a driver was unbound from the device and bound to VFIO. In this case the device is attached to a new domain and unlinked from the old domain. At this point nothing points to the old domain anymore and its memory is leaked.
Fix this by explicitly freeing the old domain iniommu_attach_domain.Fixes 1196c2f 'iommu/vt-d: Only remove domain when device is removed'Signed-off-by Joerg Roedel <[email protected]>--- drivers/iommu/intel-iommu.c | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-)drivers/iommu/intel-iommu.c +0
-0
drivers/iommu/intel-iommu.c static int intel_iommu_attach_device(struct iommu_domain *domain, · line 4427
Quoted 5 lines
drivers/iommu/intel-iommu.c +1
-1@@ -4424,10 +4424,13 @@ static int intel_iommu_attach_device(struct iommu_domain *domain, old_domain =find_domain (dev);if (old_domain) { -if (domain_type_is_vm_or_si (dmar_domain)) +if (domain_type_is_vm_or_si (dmar_domain)) {
JAH> This path is executed when starting the VM.
drivers/iommu/intel-iommu.c static int intel_iommu_attach_device(struct iommu_domain *domain, · line 4429
Quoted 3 lines
drivers/iommu/intel-iommu.c +1
-1domain_remove_one_dev_info (old_domain, dev); - else + }else {
JAH> I don't see this path being executed.
Quoted 14 lines
drivers/iommu/intel-iommu.c +3
-0domain_remove_dev_info (old_domain); +if (list_empty (&old_domain->devices)) +domain_exit (old_domain); + } } }Signature · 6 lines
1.8.4.5 _______________________________________________ iommu mailing list iommu-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA@public.gmane.org https://lists.linuxfoundation.org/mailman/listinfo/iommu
Signature · 7 lines
----------------------------------------------------------------------------
Jerry Hoemann Software Engineer Hewlett-Packard
3404 E Harmony Rd. MS 36 phone: (970) 898-1022
Ft. Collins, CO 80528 FAX: (970) 898-0707
email: [email protected]
----------------------------------------------------------------------------Joerg Roedel
maintainerRe: [PATCH 2/2] iommu/vt-d: Only remove domain when device is removed
Hi Jerry,
Quoted 44 lines · Jerry Hoemann
Quoted 26 lines · Joerg Roedel
Quoted 1 line
From
d65b236d0f27fe3ef7ac4d12cceb0da67aec86ceMon Sep 17 00:00:00 2001From: Joerg Roedel <[email protected]>
Date: Tue, 9 Dec 2014 12:56:45 +0100
Subject: [PATCH] iommu/vt-d: Fixdmar_domainleak iniommu_attach_deviceSince commit 1196c2f a domain is only destroyed in the notifier path if it is hot-unplugged. This caused a domain leakage in
iommu_attach_devicewhen a driver was unbound from the device and bound to VFIO. In this case the device is attached to a new domain and unlinked from the old domain. At this point nothing points to the old domain anymore and its memory is leaked.
Fix this by explicitly freeing the old domain iniommu_attach_domain.Fixes 1196c2f 'iommu/vt-d: Only remove domain when device is removed'Signed-off-by Joerg Roedel <[email protected]>--- drivers/iommu/intel-iommu.c | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-)drivers/iommu/intel-iommu.c +0
-0drivers/iommu/intel-iommu.c static int intel_iommu_attach_device(struct iommu_domain *domain, · line 4427
Quoted 5 lines
drivers/iommu/intel-iommu.c +1
-1@@ -4424,10 +4424,13 @@ static int intel_iommu_attach_device(struct iommu_domain *domain, old_domain =find_domain (dev);if (old_domain) { -if (domain_type_is_vm_or_si (dmar_domain)) +if (domain_type_is_vm_or_si (dmar_domain)) {JAH> This path is executed when starting the VM.
drivers/iommu/intel-iommu.c static int intel_iommu_attach_device(struct iommu_domain *domain, · line 4429
Quoted 3 lines
drivers/iommu/intel-iommu.c +1
-1domain_remove_one_dev_info (old_domain, dev); - else + }else {JAH> I don't see this path being executed.
Quoted 4 lines
drivers/iommu/intel-iommu.c static int intel_iommu_attach_device(struct iommu_domain *domain, · line 4433
Quoted 4 lines
drivers/iommu/intel-iommu.c +3
-0domain_remove_dev_info (old_domain); +if (list_empty (&old_domain->devices)) +domain_exit (old_domain); + }
You are right, thanks for testing. The reason is that the check for domain_type_is_vm_or_si(dmar_domain) uses the new domain and not the old one. I'll post a new patch.
Joerg
Quoted 10 lines · Joerg Roedel
Hi,
here is a patch-set to fix an issue recently discovered when the Intel IOMMU is in use with devices that need RMRR mappings.
The problem is that the RMRR mappings are destroyed when the device driver is unbound from the device, causing DMAR faults.
Quoted 8 lines
To solve this problem a device driver core change is necessary to catch the right point in time for the IOMMU code to destroy any mappings for a device.
With this patch-set the RMRR mappings are only destroyed when the device is actually removed from the system.
Please review.
I have no objection to these, do you want me to take them through my tree? Or if they are going through some other one feel free to add:
To the first patch.
thanks,
greg k-h
Quoted 6 lines · Greg Kroah-Hartman
I have no objection to these, do you want me to take them through my tree? Or if they are going through some other one feel free to add:
To the first patch.
Thanks Greg! I added the patches to the IOMMU tree.
Joerg
Quoted 23 lines · Joerg Roedel
Hi,
here is a patch-set to fix an issue recently discovered when the Intel IOMMU is in use with devices that need RMRR mappings.
The problem is that the RMRR mappings are destroyed when the device driver is unbound from the device, causing DMAR faults.
To solve this problem a device driver core change is necessary to catch the right point in time for the IOMMU code to destroy any mappings for a device.
With this patch-set the RMRR mappings are only destroyed when the device is actually removed from the system.
Please review.
Thanks,
Joerg
Quoted 11 lines
Joerg Roedel (2):
driver core: AddBUS_NOTIFY_REMOVED_DEVICEevent iommu/vt-d: Only remove domain when device is removeddrivers/base/core.c | 3 +++ drivers/iommu/intel-iommu.c | 11 +---------- include/linux/device.h | 11 ++++++----- 3 files changed, 10 insertions(+), 15 deletions(-)Signature · 1 line
1.9.1
Joerg,
I tested on HP Gen7 and Gen9 systems for which we experience dmar faults when we rmmod a driver whose device had RMRR regions associated with it.
We don't see problem when patch set is applied.
Thanks,
Signature · 7 lines
----------------------------------------------------------------------------
Jerry Hoemann Software Engineer Hewlett-Packard
3404 E Harmony Rd. MS 36 phone: (970) 898-1022
Ft. Collins, CO 80528 FAX: (970) 898-0707
email: [email protected]
----------------------------------------------------------------------------