Re: [PATCH] x86: add SEV hashing to fw_cfg for kernel/initrd/cmdline

[Dov Murik]

Hi Eduardo,

On 15/06/2021 18:20, Eduardo Habkost wrote:
> On Tue, May 25, 2021 at 06:59:31AM +0000, Dov Murik wrote:
>> From: James Bottomley <jejb@linux.ibm.com>
>>
>> If the VM is using memory encryption and also specifies a kernel/initrd
>> or appended command line, calculate the hashes and add them to the
>> encrypted data.  For this to work, OVMF must support an encrypted area
>> to place the data which is advertised via a special GUID in the OVMF
>> reset table (if the GUID doesn't exist, the user isn't allowed to pass
>> in the kernel/initrd/cmdline via the fw_cfg interface).
>>
>> The hashes of each of the files is calculated (or the string in the case
>> of the cmdline with trailing '\0' included).  Each entry in the hashes
>> table is GUID identified and since they're passed through the memcrypt
>> interface, the hash of the encrypted data will be accumulated by the
>> PSP.
>>
>> Signed-off-by: James Bottomley <jejb@linux.ibm.com>
>> Signed-off-by: Dov Murik <dovmurik@linux.ibm.com>
>> [dovmurik@linux.ibm.com: use machine->cgs, remove parsing of GUID
>> strings, remove GCC pragma, fix checkpatch errors]
>> ---
>>
>> OVMF support for handling the table of hashes (verifying that the
>> kernel/initrd/cmdline passed via the fw_cfg interface indeed correspond
>> to the measured hashes in the table) will be posted soon to edk2-devel.
>>
>> ---
>>  hw/i386/x86.c | 120 +++++++++++++++++++++++++++++++++++++++++++++++++-
>>  1 file changed, 119 insertions(+), 1 deletion(-)
>>
> 
> This is not an objection to the patch itself, but: can we do
> something to move all sev-related code to sev.c?  It would make
> the process of assigning a maintainer and reviewing/merging
> future patches much simpler.
> 

I'll look into this following Philippe's suggestions.

> I am not familiar with SEV internals, so my only question is
> about configurations where SEV is disabled:
> 
> [...]
>> diff --git a/hw/i386/x86.c b/hw/i386/x86.c
>> @@ -778,6 +818,11 @@ void x86_load_linux(X86MachineState *x86ms,
>>      const char *initrd_filename = machine->initrd_filename;
>>      const char *dtb_filename = machine->dtb;
>>      const char *kernel_cmdline = machine->kernel_cmdline;
>> +    uint8_t buf[HASH_SIZE];
>> +    uint8_t *hash = buf;
>> +    size_t hash_len = sizeof(buf);
>> +    struct sev_hash_table *sev_ht = NULL;
>> +    int sev_ht_index = 0;
>>  
>>      /* Align to 16 bytes as a paranoia measure */
>>      cmdline_size = (strlen(kernel_cmdline) + 16) & ~15;
>> @@ -799,6 +844,22 @@ void x86_load_linux(X86MachineState *x86ms,
>>          exit(1);
>>      }
>>  
>> +    if (machine->cgs && machine->cgs->ready) {
> 
> machine->cgs doesn't seem to be a SEV-specific field.
> What if machine->cgs->ready is set but SEV is disabled?
> 

You're right; I'll change this to sev_enabled() like in
hw/i386/pc_sysfw.c .

-Dov


>> +        uint8_t *data;
>> +        struct sev_hash_table_descriptor *area;
>> +
>> +        if (!pc_system_ovmf_table_find(SEV_HASH_TABLE_RV_GUID, &data, 
>> NULL)) {
>> +            fprintf(stderr, "qemu: kernel command line specified but OVMF 
>> has "
>> +                    "no hash table guid\n");
>> +            exit(1);
>> +        }
>> +        area = (struct sev_hash_table_descriptor *)data;
>> +
>> +        sev_ht = qemu_map_ram_ptr(NULL, area->base);
>> +        memcpy(sev_ht->guid, sev_hash_table_header_guid, 
>> sizeof(sev_ht->guid));
>> +        sev_ht->len = sizeof(*sev_ht);
>> +    }
>> +
>>      /* kernel protocol version */
>>      if (ldl_p(header + 0x202) == 0x53726448) {
>>          protocol = lduw_p(header + 0x206);
> [...]
>