intel_iommu: misc fixes for error exposed after error_report_once()
cover intel_iommu: misc fixes for error exposed after error_report_once()
Recently we have switched quite a few VT-d trace points into error_report_once()s and this does expose some errors that we didn't detect before (previously tracepoints won't trigger as long as tracepoints weren't enabled). These errors are not severe in that all of them will not affect functionality of the guest otherwise we'll notice them even earlier. However it still worths to fix all of them. This patchset tries to fix quite a few errors like this (except the last patch, which should also workaround an error but has nothing to do with the newly introduced error_report_once).
All the errors could easily be triggered by rebooting a guest with both vfio-pci device and vIOMMU, and we can see errors dumped in stderr like:
qemu-kvm: vtd_iova_to_slpte: detected slpte permission error (iova=0xffd9ce00, level=0x2, slpte=0x0, write=1)
qemu-kvm: vtd_iommu_translate: detected translation failure (dev=02:00:00, iova=0x0)
qemu-kvm: vtd_interrupt_remap_msi: MSI address low 32 bit invalid: 0x0
Regarding to the patchset itself:
Patch 1: fixes slpte permission error warning Patch 2: fixes intr_enabled flag reset missing Patch 3-4: fixes MSI translation warning Patch 5: workaround of a kernel bug that could cause UNMAP fail error
It was tested that this series can fix all the error messages observed in below bugs:
https://bugzilla.redhat.com/show_bug.cgi?id=1662270
https://bugzilla.redhat.com/show_bug.cgi?id=1662291
Please have a look, thanks.
Peter Xu (5):
intel_iommu: fix operator in vtd_switch_address_space
intel_iommu: reset intr_enabled when system reset pci/msi: export msi_is_masked()
i386/kvm: ignore masked irqs when update msi routes
vfio: retry one more time conditionally for type1 unmap
hw/i386/intel_iommu.c | 3 ++- hw/pci/msi.c | 2 +- hw/vfio/common.c | 16 ++++++++++++++++ hw/vfio/trace-events | 1 + include/hw/pci/msi.h | 1 + target/i386/kvm.c | 14 +++++++++++--- 6 files changed, 32 insertions(+), 5 deletions(-)
Signature · 1 line
2.17.1
PATCH 1/5 intel_iommu: fix operator in vtd_switch_address_space
When calculating use_iommu, we wanted to first detect whether DMAR is enabled, then check whether PT is enabled if DMAR is enabled. However in the current code we used "&" rather than "&&" so the ordering requirement is lost (instead it'll be an "AND" operation). This could introduce errors dumped in QEMU console when rebooting a guest with both assigned device and vIOMMU, like:
qemu-system-x86_64: vtd_dev_to_context_entry: invalid root entry:
rsvd=0xf000ff53f000e2c3, val=0xf000ff53f000ff53 (reserved nonzero)
--- hw/i386/intel_iommu.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
hw/i386/intel_iommu.c +1 -1
@@ -1153,7 +1153,7 @@ static bool vtd_switch_address_space(VTDAddressSpace *as)assert (as); - use_iommu = as->iommu_state->dmar_enabled & !vtd_dev_pt_enabled (as); + use_iommu = as->iommu_state->dmar_enabled && !vtd_dev_pt_enabled (as); trace_vtd_switch_address_space(pci_bus_num(as->bus), VTD_PCI_SLOT(as->devfn),
Signature · 1 line
2.17.1
Re: [Qemu-devel] [PATCH 1/5] intel_iommu: fix operator in vtd_switch_address_space
Quoted 20 lines · Peter Xu
When calculating
use_iommu, we wanted to first detect whether DMAR is enabled, then check whether PT is enabled if DMAR is enabled. However in the current code we used "&" rather than "&&" so the ordering requirement is lost (instead it'll be an "AND" operation). This could introduce errors dumped in QEMU console when rebooting a guest with both assigned device and vIOMMU, like:qemu-system-x86_64:
vtd_dev_to_context_entry: invalid root entry:rsvd=0xf000ff53f000e2c3, val=0xf000ff53f000ff53 (reserved nonzero)Signed-off-by Peter Xu <[email protected]>--- hw/i386/intel_iommu.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)hw/i386/intel_iommu.c +0
-0
hw/i386/intel_iommu.c static bool vtd_switch_address_space(VTDAddressSpace *as) · line 1159
Quoted 8 lines
hw/i386/intel_iommu.c +1
-1@@ -1153,7 +1153,7 @@ static bool vtd_switch_address_space(VTDAddressSpace *as)assert (as); - use_iommu = as->iommu_state->dmar_enabled & !vtd_dev_pt_enabled (as); + use_iommu = as->iommu_state->dmar_enabled && !vtd_dev_pt_enabled (as); trace_vtd_switch_address_space(pci_bus_num(as->bus), VTD_PCI_SLOT(as->devfn),
PATCH 2/5 intel_iommu: reset intr_enabled when system reset
In reply to Peter XuThis is found when I was debugging another problem. Until now no bug is reported with this but we'd better reset the IR status correctly after a system reset.
--- hw/i386/intel_iommu.c | 1 + 1 file changed, 1 insertion(+)
hw/i386/intel_iommu.c +1 -0
@@ -3138,6 +3138,7 @@ static void vtd_init(IntelIOMMUState *s) s->root =0 ; s->root_extended = false; s->dmar_enabled = false; + s->intr_enabled = false; s->iq_head =0 ; s->iq_tail =0 ; s->iq =0 ;
Signature · 1 line
2.17.1
Re: [Qemu-devel] [PATCH 2/5] intel_iommu: reset intr_enabled when system reset
Quoted 14 lines · Peter Xu
This is found when I was debugging another problem. Until now no bug is reported with this but we'd better reset the IR status correctly after a system reset.
Signed-off-by Peter Xu <[email protected]>--- hw/i386/intel_iommu.c | 1 + 1 file changed, 1 insertion(+)hw/i386/intel_iommu.c +0
-0
hw/i386/intel_iommu.c static void vtd_init(IntelIOMMUState *s) · line 3144
Quoted 7 lines
hw/i386/intel_iommu.c +1
-0@@ -3138,6 +3138,7 @@ static void vtd_init(IntelIOMMUState *s) s->root =0 ; s->root_extended = false; s->dmar_enabled = false; + s->intr_enabled = false; s->iq_head =0 ; s->iq_tail =0 ; s->iq =0 ;
PATCH 3/5 pci/msi: export msi_is_masked()
In reply to Peter XuIt is going to be used later on outside MSI code to detect whether one MSI vector is masked out.
--- hw/pci/msi.c | 2 +- include/hw/pci/msi.h | 1 + 2 files changed, 2 insertions(+), 1 deletion(-)
hw/pci/msi.c +1 -1
@@ -286,7 +286,7 @@ void msi_reset(PCIDevice *dev)MSI_DEV_PRINTF (dev,"reset\n" ); } -static bool msi_is_masked (const PCIDevice *dev,unsigned int vector) +bool msi_is_masked (const PCIDevice *dev,unsigned int vector) {uint16_t flags =pci_get_word (dev->config +msi_flags_off (dev));uint32_t mask, data;
include/hw/pci/msi.h +1 -0
@@ -39,6 +39,7 @@ int msi_init(struct PCIDevice *dev, uint8_t offset, bool msi_per_vector_mask, Error **errp);void msi_uninit (struct PCIDevice *dev);void msi_reset (PCIDevice *dev); +bool msi_is_masked (const PCIDevice *dev,unsigned int vector);void msi_notify (PCIDevice *dev,unsigned int vector);void msi_send_message (PCIDevice *dev,MSIMessage msg);void msi_write_config (PCIDevice *dev,uint32_t addr,uint32_t val,int len);
Signature · 1 line
2.17.1
PATCH 4/5 i386/kvm: ignore masked irqs when update msi routes
In reply to Peter XuWhen we are with intel-iommu device and with IR on, KVM will register an IEC notifier to detect interrupt updates from the guest and we'll kick off kvm_update_msi_routes_all() when it happens to make sure kernel IRQ cache is matching the latest.
Though, kvm_update_msi_routes_all() is buggy in that it ignored the mask bit of either MSI/MSIX messages and it tries to translate the message even if the corresponding message was already masked by the guest driver (hence the MSI/MSIX message will be invalid).
Without this patch, we can receive an error message when we reboot a guest with both an assigned vfio-pci device and intel-iommu enabled:
qemu-system-x86_64: vtd_interrupt_remap_msi: MSI address low 32 bit invalid: 0x0
The error does not affect functionality of the guest since when we failed to translate we'll just silently continue (which makes sense since crashing the VM for this seems even worse), but still it's better to fix it up.
--- target/i386/kvm.c | 14 +++++++++++--- 1 file changed, 11 insertions(+), 3 deletions(-)
target/i386/kvm.c +11 -3
@@ -3889,7 +3889,7 @@ static QLIST_HEAD(, MSIRouteEntry) msi_route_list = \static void kvm_update_msi_routes_all (void *private,bool global,uint32_t index,uint32_t mask) { -int cnt =0 ; +int cnt =0 , vector;MSIRouteEntry *entry;MSIMessage msg;PCIDevice *dev;
@@ -3897,11 +3897,19 @@ static void kvm_update_msi_routes_all(void *private, bool global,/* TODO: explicit route update */ QLIST_FOREACH(entry, &msi_route_list, list) { cnt++; + vector = entry->vector; dev = entry->dev; -if (!msix_enabled (dev) && !msi_enabled (dev)) { +if (msix_enabled (dev) && !msix_is_masked (dev, vector)) { + msg =msix_get_message (dev, vector); + }else if (msi_enabled (dev) && !msi_is_masked (dev, vector)) { + msg =msi_get_message (dev, vector); + }else { +/* +* Either MSI/MSIX is disabled for the device, or the +* specific message was masked out. Skip this one. +*/ continue ; } - msg =pci_get_msi_message (dev, entry->vector);kvm_irqchip_update_msi_route (kvm_state, entry->virq, msg, dev); }kvm_irqchip_commit_routes (kvm_state);
Signature · 1 line
2.17.1
PATCH 5/5 vfio: retry one more time conditionally for type1 unmap
In reply to Peter XuIn Linux version v4.15+ there's a bug (introduced in 71a7d3d78e3c, "vfio/type1: silence integer overflow warning", 2017-10-20) that could potentially reject a valid unmap region that covers exactly the whole u64 address space (like iova=0xfef00000, size=2^64-0xfef00000). Besides a fix on the kernel side, QEMU also needs to live well even with the old kernels. When booting a guest with both vfio-pci and intel-iommu device, we can see error dumped:
qemu-kvm: VFIO_UNMAP_DMA: -22
qemu-kvm: vfio_dma_unmap(0x561f059948f0, 0xfef00000,
0xffffffff01100000) = -22 (Invalid argument)
This patch gives another shot of the UNMAP ioctl if the specific error was detected, while in the second UNMAP ioctl we skip the last page assuming that it's never used. In our case, currently only Intel VT-d is using this code and it should never use the iova address 2^64-4096 (so far largest supported GAW is 57 bits) so ignoring that page should be fine.
After this patch is applied, the errors go away.
Bugzilla: https://bugzilla.redhat.com/show_bug.cgi?id=1662291
--- hw/vfio/common.c | 16 ++++++++++++++++ hw/vfio/trace-events | 1 + 2 files changed, 17 insertions(+)
hw/vfio/common.c +16 -0
@@ -221,6 +221,22 @@ static int vfio_dma_unmap(VFIOContainer *container, };if (ioctl (container->fd,VFIO_IOMMU_UNMAP_DMA , &unmap)) { +/* +* Give it another shot due to a bug in kernel (v4.15-v4.20) +* that could potentially reject a region that exactly covers +* the whole u64 address space (71a7d3d78e3c, "vfio/type1: +* silence integer overflow warning", 2017-10-20). If that +* happens, we retry for one more time assuming that the last +* page of the address space (2^64-getpagesize()) has already +* been dropped. +*/ +if (errno ==EINVAL && unmap.size && unmap.iova + unmap.size ==0 ) { +trace_vfio_dma_unmap_workaround_overflow (); + unmap.size -=getpagesize (); +if (ioctl (container->fd,VFIO_IOMMU_UNMAP_DMA , &unmap) ==0 ) { +return 0 ; + } + }error_report ("VFIO_UNMAP_DMA: %d" , -errno);return -errno; }
hw/vfio/trace-events +1 -0
@@ -110,6 +110,7 @@ vfio_region_mmaps_set_enabled(const char *name, bool enabled) "Region %s mmaps e vfio_region_sparse_mmap_header(const char *name, int index, int nr_areas) "Device %s region %d: %d sparse mmap entries" vfio_region_sparse_mmap_entry(int i, unsigned long start, unsigned long end) "sparse entry %d [0x%lx - 0x%lx]" vfio_get_dev_region(const char *name, int index, uint32_t type, uint32_t subtype) "%s index %d, %08x/%0x8" +vfio_dma_unmap_workaround_overflow(void) "" # hw/vfio/platform.c vfio_platform_base_device_init(char *name, int groupid) "%s belongs to group #%d"
Signature · 1 line
2.17.1
Re: [Qemu-devel] [PATCH 5/5] vfio: retry one more time conditionally for type1 unmap
Quoted 62 lines · On Tue, 8 Jan 2019 19:47:20 +0800 Peter Xu <[email protected]> wrote:
In Linux version v4.15+ there's a bug (introduced in
71a7d3d78e3c, "vfio/type1: silence integer overflow warning", 2017-10-20) that could potentially reject a valid unmap region that covers exactly the whole u64 address space (like iova=0xfef00000, size=2^64-0xfef00000). Besides a fix on the kernel side, QEMU also needs to live well even with the old kernels. When booting a guest with both vfio-pci and intel-iommu device, we can see error dumped:qemu-kvm:
VFIO_UNMAP_DMA: -22
qemu-kvm: vfio_dma_unmap(0x561f059948f0,0xfef00000,0xffffffff01100000) = -22 (Invalid argument)This patch gives another shot of the UNMAP ioctl if the specific error was detected, while in the second UNMAP ioctl we skip the last page assuming that it's never used. In our case, currently only Intel VT-d is using this code and it should never use the iova address 2^64-4096 (so far largest supported GAW is 57 bits) so ignoring that page should be fine.
After this patch is applied, the errors go away.
Reported-by Pei Zhang <[email protected]>Bugzilla: https://bugzilla.redhat.com/show_bug.cgi?id=1662291
Suggested-by Alex Williamson <[email protected]>Signed-off-by Peter Xu <[email protected]>--- hw/vfio/common.c | 16 ++++++++++++++++ hw/vfio/trace-events | 1 + 2 files changed, 17 insertions(+)hw/vfio/common.c +16
-0@@ -221,6 +221,22 @@ static int vfio_dma_unmap(VFIOContainer *container, };if (ioctl (container->fd,VFIO_IOMMU_UNMAP_DMA , &unmap)) { +/* +* Give it another shot due to a bug in kernel (v4.15-v4.20) +* that could potentially reject a region that exactly covers +* the whole u64 address space (71a7d3d78e3c, "vfio/type1: +* silence integer overflow warning", 2017-10-20). If that +* happens, we retry for one more time assuming that the last +* page of the address space (2^64-getpagesize()) has already +* been dropped. +*/ +if (errno ==EINVAL && unmap.size && unmap.iova + unmap.size ==0 ) { +trace_vfio_dma_unmap_workaround_overflow (); + unmap.size -=getpagesize (); +if (ioctl (container->fd,VFIO_IOMMU_UNMAP_DMA , &unmap) ==0 ) { +return 0 ; + } + }error_report ("VFIO_UNMAP_DMA: %d" , -errno);return -errno; }hw/vfio/trace-events +0
-0
hw/vfio/trace-events vfio_region_mmaps_set_enabled(const char *name, bool enabled) "Region %s mmaps e · line 116
Quoted 7 lines
hw/vfio/trace-events +1
-0@@ -110,6 +110,7 @@ vfio_region_mmaps_set_enabled(const char *name, bool enabled) "Region %s mmaps e vfio_region_sparse_mmap_header(const char *name, int index, int nr_areas) "Device %s region %d: %d sparse mmap entries" vfio_region_sparse_mmap_entry(int i, unsigned long start, unsigned long end) "sparse entry %d [0x%lx - 0x%lx]" vfio_get_dev_region(const char *name, int index, uint32_t type, uint32_t subtype) "%s index %d, %08x/%0x8" +vfio_dma_unmap_workaround_overflow(void) "" # hw/vfio/platform.c vfio_platform_base_device_init(char *name, int groupid) "%s belongs to group #%d"
Hi Peter,
I was working on a slightly different version:
hw/vfio/common.c +18 -1
@@ -220,7 +220,24 @@ static int vfio_dma_unmap(VFIOContainer *container, .size = size, }; -if (ioctl (container->fd,VFIO_IOMMU_UNMAP_DMA , &unmap)) { +while (ioctl (container->fd,VFIO_IOMMU_UNMAP_DMA , &unmap)) { +/* +* The type1 backend has an off-by-one bug in the kernel (71a7d3d78e3c +* v4.15) where its overflow check prevents us from unmapping the last +* page of the address space. Test for the error condition and re-try +* the unmap excluding the last page. The expectation is that we've +* never mapped the last page anyway and this unmap request comes via +* vIOMMU support which also makes it unlikely that this page is used. +* This bug was introduced well after type1 v2 support was introduced, +* so we shouldn't need to test for v1. A fix is proposed for kernel +* v5.0 so this workaround can be removed once affected kernels are +* sufficiently deprecated. +*/ +if (errno ==EINVAL && unmap.size && !(unmap.iova + unmap.size) && + container->iommu_type == VFIO_TYPE1v2_IOMMU) { + unmap.size -=1ULL <<ctz64 (container->pgsizes); +continue ; + }error_report ("VFIO_UNMAP_DMA: %d" , -errno);return -errno; }
I like your addition of tracing, but I prefer the type1v2 test (the bug is specific to the type1 backend) and using the iommu minimum page size rather than the cpu page size. Do you want to incorporate or would you prefer I post mine? Thanks,
Alex
Re: [Qemu-devel] [PATCH 5/5] vfio: retry one more time conditionally for type1 unmap
Quoted 108 lines · Alex Williamson
Quoted 62 lines · On Tue, 8 Jan 2019 19:47:20 +0800 Peter Xu <[email protected]> wrote:
In Linux version v4.15+ there's a bug (introduced in
71a7d3d78e3c, "vfio/type1: silence integer overflow warning", 2017-10-20) that could potentially reject a valid unmap region that covers exactly the whole u64 address space (like iova=0xfef00000, size=2^64-0xfef00000). Besides a fix on the kernel side, QEMU also needs to live well even with the old kernels. When booting a guest with both vfio-pci and intel-iommu device, we can see error dumped:qemu-kvm:
VFIO_UNMAP_DMA: -22
qemu-kvm: vfio_dma_unmap(0x561f059948f0,0xfef00000,0xffffffff01100000) = -22 (Invalid argument)This patch gives another shot of the UNMAP ioctl if the specific error was detected, while in the second UNMAP ioctl we skip the last page assuming that it's never used. In our case, currently only Intel VT-d is using this code and it should never use the iova address 2^64-4096 (so far largest supported GAW is 57 bits) so ignoring that page should be fine.
After this patch is applied, the errors go away.
Reported-by Pei Zhang <[email protected]>Bugzilla: https://bugzilla.redhat.com/show_bug.cgi?id=1662291
Suggested-by Alex Williamson <[email protected]>Signed-off-by Peter Xu <[email protected]>--- hw/vfio/common.c | 16 ++++++++++++++++ hw/vfio/trace-events | 1 + 2 files changed, 17 insertions(+)hw/vfio/common.c +16
-0@@ -221,6 +221,22 @@ static int vfio_dma_unmap(VFIOContainer *container, };if (ioctl (container->fd,VFIO_IOMMU_UNMAP_DMA , &unmap)) { +/* +* Give it another shot due to a bug in kernel (v4.15-v4.20) +* that could potentially reject a region that exactly covers +* the whole u64 address space (71a7d3d78e3c, "vfio/type1: +* silence integer overflow warning", 2017-10-20). If that +* happens, we retry for one more time assuming that the last +* page of the address space (2^64-getpagesize()) has already +* been dropped. +*/ +if (errno ==EINVAL && unmap.size && unmap.iova + unmap.size ==0 ) { +trace_vfio_dma_unmap_workaround_overflow (); + unmap.size -=getpagesize (); +if (ioctl (container->fd,VFIO_IOMMU_UNMAP_DMA , &unmap) ==0 ) { +return 0 ; + } + }error_report ("VFIO_UNMAP_DMA: %d" , -errno);return -errno; }hw/vfio/trace-events +0
-0hw/vfio/trace-events vfio_region_mmaps_set_enabled(const char *name, bool enabled) "Region %s mmaps e · line 116
Quoted 7 lines
hw/vfio/trace-events +1
-0@@ -110,6 +110,7 @@ vfio_region_mmaps_set_enabled(const char *name, bool enabled) "Region %s mmaps e vfio_region_sparse_mmap_header(const char *name, int index, int nr_areas) "Device %s region %d: %d sparse mmap entries" vfio_region_sparse_mmap_entry(int i, unsigned long start, unsigned long end) "sparse entry %d [0x%lx - 0x%lx]" vfio_get_dev_region(const char *name, int index, uint32_t type, uint32_t subtype) "%s index %d, %08x/%0x8" +vfio_dma_unmap_workaround_overflow(void) "" # hw/vfio/platform.c vfio_platform_base_device_init(char *name, int groupid) "%s belongs to group #%d"Hi Peter,
I was working on a slightly different version:
hw/vfio/common.c +18
-1@@ -220,7 +220,24 @@ static int vfio_dma_unmap(VFIOContainer *container, .size = size, }; -if (ioctl (container->fd,VFIO_IOMMU_UNMAP_DMA , &unmap)) { +while (ioctl (container->fd,VFIO_IOMMU_UNMAP_DMA , &unmap)) { +/* +* The type1 backend has an off-by-one bug in the kernel (71a7d3d78e3c +* v4.15) where its overflow check prevents us from unmapping the last +* page of the address space. Test for the error condition and re-try +* the unmap excluding the last page. The expectation is that we've +* never mapped the last page anyway and this unmap request comes via +* vIOMMU support which also makes it unlikely that this page is used. +* This bug was introduced well after type1 v2 support was introduced, +* so we shouldn't need to test for v1. A fix is proposed for kernel +* v5.0 so this workaround can be removed once affected kernels are +* sufficiently deprecated. +*/ +if (errno ==EINVAL && unmap.size && !(unmap.iova + unmap.size) && + container->iommu_type == VFIO_TYPE1v2_IOMMU) { + unmap.size -=1ULL <<ctz64 (container->pgsizes); +continue ; + }error_report ("VFIO_UNMAP_DMA: %d" , -errno);return -errno; }
Quoted 4 lines
I like your addition of tracing, but I prefer the type1v2 test (the bug is specific to the type1 backend) and using the iommu minimum page size rather than the cpu page size. Do you want to incorporate or would you prefer I post mine? Thanks,
Hi, Alex,
I think the type check and using the container->pgsizes are better! Please use your version, I'll simply drop mine from the series.
Thanks,
Signature · 1 line
Peter Xu