| Message ID | 20240814205303.2619373-1-kevinloughlin@google.com |
|---|---|
| State | New |
| Headers | show |
| Series | device-dax: map dax memory as decrypted in CoCo guests | expand |
> Confidential Computing (CoCo) guests encrypt private memory by default. > DAX memory regions allow a guest to bypass its own (private) page cache > and instead use host memory, which is not private to the guest. > > Commit 867400af90f1 ("mm/memremap.c: map FS_DAX device memory as > decrypted") only ensures that FS_DAX memory is appropriately marked as > decrypted. As such, also mark device-dax memory as decrypted. Overall patch seems OK to me. How can I test this? Can I test it with virtio-pmem device? Thanks, Pankaj > > Signed-off-by: Kevin Loughlin <kevinloughlin@google.co > --- > drivers/dax/device.c | 3 +++ > 1 file changed, 3 insertions(+) > > diff --git a/drivers/dax/device.c b/drivers/dax/device.c > index 2051e4f73c8a..a284442d7ecc 100644 > --- a/drivers/dax/device.c > +++ b/drivers/dax/device.c > @@ -11,6 +11,7 @@ > #include <linux/fs.h> > #include <linux/mm.h> > #include <linux/mman.h> > +#include <linux/cc_platform.h> > #include "dax-private.h" > #include "bus.h" > > @@ -303,6 +304,8 @@ static int dax_mmap(struct file *filp, struct vm_area_struct *vma) > > vma->vm_ops = &dax_vm_ops; > vm_flags_set(vma, VM_HUGEPAGE); > + if (cc_platform_has(CC_ATTR_MEM_ENCRYPT)) > + vma->vm_page_prot = pgprot_decrypted(vma->vm_page_prot); > return 0; > } >
> How can I test this? Can I test it with virtio-pmem device? Correct. Assuming the CoCo guest accesses some virtio-pmem device in devdax mode, mmapping() this virtio-pmem device's memory region results in the guest and host reading the same (plaintext) values from the region. > > Thanks, > Pankaj > > > > Signed-off-by: Kevin Loughlin <kevinloughlin@google.co > --- > > drivers/dax/device.c | 3 +++ > > 1 file changed, 3 insertions(+) > > > > diff --git a/drivers/dax/device.c b/drivers/dax/device.c > > index 2051e4f73c8a..a284442d7ecc 100644 > > --- a/drivers/dax/device.c > > +++ b/drivers/dax/device.c > > @@ -11,6 +11,7 @@ > > #include <linux/fs.h> > > #include <linux/mm.h> > > #include <linux/mman.h> > > +#include <linux/cc_platform.h> > > #include "dax-private.h" > > #include "bus.h" > > > > @@ -303,6 +304,8 @@ static int dax_mmap(struct file *filp, struct vm_area_struct *vma) > > > > vma->vm_ops = &dax_vm_ops; > > vm_flags_set(vma, VM_HUGEPAGE); > > + if (cc_platform_has(CC_ATTR_MEM_ENCRYPT)) > > + vma->vm_page_prot = pgprot_decrypted(vma->vm_page_prot); > > return 0; > > } > > >
On 8/26/2024 11:35 PM, Kevin Loughlin wrote: >> How can I test this? Can I test it with virtio-pmem device? > > Correct. Assuming the CoCo guest accesses some virtio-pmem device in > devdax mode, mmapping() this virtio-pmem device's memory region > results in the guest and host reading the same (plaintext) values from > the region. I tried to test the daxdev with virtio-pmem but getting the below error (just tried without this patch) root@ubuntu:/home/amd# ndctl list -N { "dev":"namespace0.0", "mode":"devdax", "map":"dev", "size":1054867456, "uuid":"c8b15ce6-0c8f-4a1a-ada6-b19a90bdf1bb", "numa_node":0 } root@ubuntu:cat /dev/zero | daxio --output=/dev/namespace0.0 daxio: neither input or output is device dax Could you please share the instructions (to test with virtio-pmem or Qemu). If not, still okay. Thanks, Pankaj Best regards, Pankaj
On 9/10/2024 3:00 PM, Gupta, Pankaj wrote: > On 8/26/2024 11:35 PM, Kevin Loughlin wrote: >>> How can I test this? Can I test it with virtio-pmem device? >> >> Correct. Assuming the CoCo guest accesses some virtio-pmem device in >> devdax mode, mmapping() this virtio-pmem device's memory region >> results in the guest and host reading the same (plaintext) values from >> the region. > > I tried to test the daxdev with virtio-pmem but getting the below error > (just tried without this patch) > > root@ubuntu:/home/amd# ndctl list -N > { > "dev":"namespace0.0", > "mode":"devdax", > "map":"dev", > "size":1054867456, > "uuid":"c8b15ce6-0c8f-4a1a-ada6-b19a90bdf1bb", > "numa_node":0 > } > > root@ubuntu:cat /dev/zero | daxio --output=/dev/namespace0.0 > daxio: neither input or output is device dax > > Could you please share the instructions (to test with virtio-pmem or > Qemu). If not, still okay. Ah by mistake pressed entered early. Just want to see the behavior on a SEV system with daxdev (maybe above virtio-pmem or other emulated device) and see how this patch helps. Best regards, Pankaj
diff --git a/drivers/dax/device.c b/drivers/dax/device.c index 2051e4f73c8a..a284442d7ecc 100644 --- a/drivers/dax/device.c +++ b/drivers/dax/device.c @@ -11,6 +11,7 @@ #include <linux/fs.h> #include <linux/mm.h> #include <linux/mman.h> +#include <linux/cc_platform.h> #include "dax-private.h" #include "bus.h" @@ -303,6 +304,8 @@ static int dax_mmap(struct file *filp, struct vm_area_struct *vma) vma->vm_ops = &dax_vm_ops; vm_flags_set(vma, VM_HUGEPAGE); + if (cc_platform_has(CC_ATTR_MEM_ENCRYPT)) + vma->vm_page_prot = pgprot_decrypted(vma->vm_page_prot); return 0; }
Confidential Computing (CoCo) guests encrypt private memory by default. DAX memory regions allow a guest to bypass its own (private) page cache and instead use host memory, which is not private to the guest. Commit 867400af90f1 ("mm/memremap.c: map FS_DAX device memory as decrypted") only ensures that FS_DAX memory is appropriately marked as decrypted. As such, also mark device-dax memory as decrypted. Signed-off-by: Kevin Loughlin <kevinloughlin@google.com> --- drivers/dax/device.c | 3 +++ 1 file changed, 3 insertions(+)