On Fri, May 13, 2022, Peter Gonda wrote:
> On Thu, May 12, 2022 at 4:23 PM Ashish Kalra <[email protected]> wrote:
> >
> > From: Ashish Kalra <[email protected]>
> >
> > For some sev ioctl interfaces, the length parameter that is passed maybe
> > less than or equal to SEV_FW_BLOB_MAX_SIZE, but larger than the data
> > that PSP firmware returns. In this case, kmalloc will allocate memory
> > that is the size of the input rather than the size of the data.
> > Since PSP firmware doesn't fully overwrite the allocated buffer, these
> > sev ioctl interface may return uninitialized kernel slab memory.
> >
> > Reported-by: Andy Nguyen <[email protected]>
> > Suggested-by: David Rientjes <[email protected]>
> > Suggested-by: Peter Gonda <[email protected]>
> > Cc: [email protected]
> > Cc: [email protected]
> > Signed-off-by: Ashish Kalra <[email protected]>
> > ---
> > arch/x86/kvm/svm/sev.c | 6 +++---
> > 1 file changed, 3 insertions(+), 3 deletions(-)
> >
>
> Can we just update all the kmalloc()s that buffers get given to the
> PSP? For instance doesn't sev_send_update_data() have an issue?
> Reading the PSP spec it seems like a user can call this ioctl with a
> large hdr_len and the PSP will only fill out what's actually required
> like in these fixed up cases? This is assuming the PSP is written to
> spec (and just the current version). I'd rather have all of these
> instances updated.
Agreed, the kernel should explicitly initialize any copy_to_user() to source and
never rely on the PSP to fill the entire blob unless there's an ironclad guarantee
the entire struct/blob will be written. E.g. it's probably ok to skip zeroing
"data" in sev_ioctl_do_platform_status(), but even then it might be wortwhile as
defense-in-depth.
Looking through other copy_to_user() calls:
- "blob" in sev_ioctl_do_pek_csr()
- "id_blob" in sev_ioctl_do_get_id2()
- "pdh_blob" and "cert_blob" in sev_ioctl_do_pdh_export()
The last one is probably fine since the copy length comes from the PSP, but it's
not like these ioctls are performance critical...
/* If we query the length, FW responded with expected data. */
input.cert_chain_len = data.cert_chain_len;
input.pdh_cert_len = data.pdh_cert_len;
Hello Sean & Peter,
On 5/13/22 14:49, Sean Christopherson wrote:
> On Fri, May 13, 2022, Peter Gonda wrote:
>> On Thu, May 12, 2022 at 4:23 PM Ashish Kalra <[email protected]> wrote:
>>> From: Ashish Kalra <[email protected]>
>>>
>>> For some sev ioctl interfaces, the length parameter that is passed maybe
>>> less than or equal to SEV_FW_BLOB_MAX_SIZE, but larger than the data
>>> that PSP firmware returns. In this case, kmalloc will allocate memory
>>> that is the size of the input rather than the size of the data.
>>> Since PSP firmware doesn't fully overwrite the allocated buffer, these
>>> sev ioctl interface may return uninitialized kernel slab memory.
>>>
>>> Reported-by: Andy Nguyen <[email protected]>
>>> Suggested-by: David Rientjes <[email protected]>
>>> Suggested-by: Peter Gonda <[email protected]>
>>> Cc: [email protected]
>>> Cc: [email protected]
>>> Signed-off-by: Ashish Kalra <[email protected]>
>>> ---
>>> arch/x86/kvm/svm/sev.c | 6 +++---
>>> 1 file changed, 3 insertions(+), 3 deletions(-)
>>>
>> Can we just update all the kmalloc()s that buffers get given to the
>> PSP? For instance doesn't sev_send_update_data() have an issue?
>> Reading the PSP spec it seems like a user can call this ioctl with a
>> large hdr_len and the PSP will only fill out what's actually required
>> like in these fixed up cases? This is assuming the PSP is written to
>> spec (and just the current version). I'd rather have all of these
>> instances updated.
Yes, this function is also vulnerable as it allocates the return buffer
using kmalloc() and copies back to user the buffer sized as per the user
provided length (and not the FW returned length), so it surely needs fixup.
I will update all these instances to use kzalloc() instead of kmalloc().
> Agreed, the kernel should explicitly initialize any copy_to_user() to source and
> never rely on the PSP to fill the entire blob unless there's an ironclad guarantee
> the entire struct/blob will be written. E.g. it's probably ok to skip zeroing
> "data" in sev_ioctl_do_platform_status(), but even then it might be wortwhile as
> defense-in-depth.
>
> Looking through other copy_to_user() calls:
>
> - "blob" in sev_ioctl_do_pek_csr()
> - "id_blob" in sev_ioctl_do_get_id2()
> - "pdh_blob" and "cert_blob" in sev_ioctl_do_pdh_export()
These functions are part of the ccp driver and a fix for them has
already been sent upstream to [email protected] and
[email protected]:
[PATCH] crypto: ccp - Use kzalloc for sev ioctl interfaces to prevent
kernel memory leak
Thanks,
Ashish
>
> The last one is probably fine since the copy length comes from the PSP, but it's
> not like these ioctls are performance critical...
>
> /* If we query the length, FW responded with expected data. */
> input.cert_chain_len = data.cert_chain_len;
> input.pdh_cert_len = data.pdh_cert_len;
On Fri, May 13, 2022 at 2:11 PM Ashish Kalra <[email protected]> wrote:
>
> Hello Sean & Peter,
>
> On 5/13/22 14:49, Sean Christopherson wrote:
> > On Fri, May 13, 2022, Peter Gonda wrote:
> >> On Thu, May 12, 2022 at 4:23 PM Ashish Kalra <[email protected]> wrote:
> >>> From: Ashish Kalra <[email protected]>
> >>>
> >>> For some sev ioctl interfaces, the length parameter that is passed maybe
> >>> less than or equal to SEV_FW_BLOB_MAX_SIZE, but larger than the data
> >>> that PSP firmware returns. In this case, kmalloc will allocate memory
> >>> that is the size of the input rather than the size of the data.
> >>> Since PSP firmware doesn't fully overwrite the allocated buffer, these
> >>> sev ioctl interface may return uninitialized kernel slab memory.
> >>>
> >>> Reported-by: Andy Nguyen <[email protected]>
> >>> Suggested-by: David Rientjes <[email protected]>
> >>> Suggested-by: Peter Gonda <[email protected]>
> >>> Cc: [email protected]
> >>> Cc: [email protected]
> >>> Signed-off-by: Ashish Kalra <[email protected]>
> >>> ---
> >>> arch/x86/kvm/svm/sev.c | 6 +++---
> >>> 1 file changed, 3 insertions(+), 3 deletions(-)
> >>>
> >> Can we just update all the kmalloc()s that buffers get given to the
> >> PSP? For instance doesn't sev_send_update_data() have an issue?
> >> Reading the PSP spec it seems like a user can call this ioctl with a
> >> large hdr_len and the PSP will only fill out what's actually required
> >> like in these fixed up cases? This is assuming the PSP is written to
> >> spec (and just the current version). I'd rather have all of these
> >> instances updated.
>
> Yes, this function is also vulnerable as it allocates the return buffer
> using kmalloc() and copies back to user the buffer sized as per the user
> provided length (and not the FW returned length), so it surely needs fixup.
>
> I will update all these instances to use kzalloc() instead of kmalloc().
Do we need the alloc_page() in __sev_dbg_encrypt_user() to have __GFP_ZERO too?
>
> > Agreed, the kernel should explicitly initialize any copy_to_user() to source and
> > never rely on the PSP to fill the entire blob unless there's an ironclad guarantee
> > the entire struct/blob will be written. E.g. it's probably ok to skip zeroing
> > "data" in sev_ioctl_do_platform_status(), but even then it might be wortwhile as
> > defense-in-depth.
> >
> > Looking through other copy_to_user() calls:
> >
> > - "blob" in sev_ioctl_do_pek_csr()
> > - "id_blob" in sev_ioctl_do_get_id2()
> > - "pdh_blob" and "cert_blob" in sev_ioctl_do_pdh_export()
>
> These functions are part of the ccp driver and a fix for them has
> already been sent upstream to [email protected] and
> [email protected]:
>
> [PATCH] crypto: ccp - Use kzalloc for sev ioctl interfaces to prevent
> kernel memory leak
>
> Thanks,
>
> Ashish
>
> >
> > The last one is probably fine since the copy length comes from the PSP, but it's
> > not like these ioctls are performance critical...
> >
> > /* If we query the length, FW responded with expected data. */
> > input.cert_chain_len = data.cert_chain_len;
> > input.pdh_cert_len = data.pdh_cert_len;
Hello Peter,
On 5/13/22 20:09, Peter Gonda wrote:
> On Fri, May 13, 2022 at 2:11 PM Ashish Kalra <[email protected]> wrote:
>> Hello Sean & Peter,
>>
>> On 5/13/22 14:49, Sean Christopherson wrote:
>>> On Fri, May 13, 2022, Peter Gonda wrote:
>>>> On Thu, May 12, 2022 at 4:23 PM Ashish Kalra <[email protected]> wrote:
>>>>> From: Ashish Kalra <[email protected]>
>>>>>
>>>>> For some sev ioctl interfaces, the length parameter that is passed maybe
>>>>> less than or equal to SEV_FW_BLOB_MAX_SIZE, but larger than the data
>>>>> that PSP firmware returns. In this case, kmalloc will allocate memory
>>>>> that is the size of the input rather than the size of the data.
>>>>> Since PSP firmware doesn't fully overwrite the allocated buffer, these
>>>>> sev ioctl interface may return uninitialized kernel slab memory.
>>>>>
>>>>> Reported-by: Andy Nguyen <[email protected]>
>>>>> Suggested-by: David Rientjes <[email protected]>
>>>>> Suggested-by: Peter Gonda <[email protected]>
>>>>> Cc: [email protected]
>>>>> Cc: [email protected]
>>>>> Signed-off-by: Ashish Kalra <[email protected]>
>>>>> ---
>>>>> arch/x86/kvm/svm/sev.c | 6 +++---
>>>>> 1 file changed, 3 insertions(+), 3 deletions(-)
>>>>>
>>>> Can we just update all the kmalloc()s that buffers get given to the
>>>> PSP? For instance doesn't sev_send_update_data() have an issue?
>>>> Reading the PSP spec it seems like a user can call this ioctl with a
>>>> large hdr_len and the PSP will only fill out what's actually required
>>>> like in these fixed up cases? This is assuming the PSP is written to
>>>> spec (and just the current version). I'd rather have all of these
>>>> instances updated.
>> Yes, this function is also vulnerable as it allocates the return buffer
>> using kmalloc() and copies back to user the buffer sized as per the user
>> provided length (and not the FW returned length), so it surely needs fixup.
>>
>> I will update all these instances to use kzalloc() instead of kmalloc().
> Do we need the alloc_page() in __sev_dbg_encrypt_user() to have __GFP_ZERO too?
Actually this is used to allocate intermediate buffers to do
copy_from_user, so it should be safe as size here is used for actual
guest memory size to be encrypted.
Thanks, Ashish
>
>
>>> Agreed, the kernel should explicitly initialize any copy_to_user() to source and
>>> never rely on the PSP to fill the entire blob unless there's an ironclad guarantee
>>> the entire struct/blob will be written. E.g. it's probably ok to skip zeroing
>>> "data" in sev_ioctl_do_platform_status(), but even then it might be wortwhile as
>>> defense-in-depth.
>>>
>>> Looking through other copy_to_user() calls:
>>>
>>> - "blob" in sev_ioctl_do_pek_csr()
>>> - "id_blob" in sev_ioctl_do_get_id2()
>>> - "pdh_blob" and "cert_blob" in sev_ioctl_do_pdh_export()
>> These functions are part of the ccp driver and a fix for them has
>> already been sent upstream to [email protected] and
>> [email protected]:
>>
>> [PATCH] crypto: ccp - Use kzalloc for sev ioctl interfaces to prevent
>> kernel memory leak
>>
>> Thanks,
>>
>> Ashish
>>
>>> The last one is probably fine since the copy length comes from the PSP, but it's
>>> not like these ioctls are performance critical...
>>>
>>> /* If we query the length, FW responded with expected data. */
>>> input.cert_chain_len = data.cert_chain_len;
>>> input.pdh_cert_len = data.pdh_cert_len;
On Fri, May 13, 2022, Ashish Kalra wrote:
> Hello Sean & Peter,
> > Looking through other copy_to_user() calls:
> >
> > - "blob" in sev_ioctl_do_pek_csr()
> > - "id_blob" in sev_ioctl_do_get_id2()
> > - "pdh_blob" and "cert_blob" in sev_ioctl_do_pdh_export()
>
> These functions are part of the ccp driver and a fix for them has already
> been sent upstream to [email protected] and
> [email protected]:
>
> [PATCH] crypto: ccp - Use kzalloc for sev ioctl interfaces to prevent kernel
> memory leak
Ha, that's why I was getting a bit of deja vu. I saw that fly by and then got it
confused with this patch.