[PATCH 0/3] fs: reduce export usage of kerne_read*() calls

[Scott Branden]

Hi Mimi,

On 2020-05-23 7:52 p.m., Mimi Zohar wrote:
> On Fri, 2020-05-22 at 16:25 -0700, Scott Branden wrote:
>> Hi Kees,
>>
>> On 2020-05-22 4:04 p.m., Kees Cook wrote:
>>> On Fri, May 22, 2020 at 03:24:32PM -0700, Scott Branden wrote:
>>>> On 2020-05-18 5:37 a.m., Mimi Zohar wrote:
>>>>> On Sun, 2020-05-17 at 23:22 -0700, Christoph Hellwig wrote:
>>>>>> On Fri, May 15, 2020 at 09:29:33PM +0000, Luis Chamberlain wrote:
>>>>>>> On Wed, May 13, 2020 at 11:17:36AM -0700, Christoph Hellwig wrote:
>>>>>>>> Can you also move kernel_read_* out of fs.h?  That header gets pulled
>>>>>>>> in just about everywhere and doesn't really need function not related
>>>>>>>> to the general fs interface.
>>>>>>> Sure, where should I dump these?
>>>>>> Maybe a new linux/kernel_read_file.h?  Bonus points for a small top
>>>>>> of the file comment explaining the point of the interface, which I
>>>>>> still don't get :)
>>>>> Instead of rolling your own method of having the kernel read a file,
>>>>> which requires call specific security hooks, this interface provides a
>>>>> single generic set of pre and post security hooks.  The
>>>>> kernel_read_file_id enumeration permits the security hook to
>>>>> differentiate between callers.
>>>>>
>>>>> To comply with secure and trusted boot concepts, a file cannot be
>>>>> accessible to the caller until after it has been measured and/or the
>>>>> integrity (hash/signature) appraised.
>>>>>
>>>>> In some cases, the file was previously read twice, first to measure
>>>>> and/or appraise the file and then read again into a buffer for
>>>>> use.  This interface reads the file into a buffer once, calls the
>>>>> generic post security hook, before providing the buffer to the caller.
>>>>>     (Note using firmware pre-allocated memory might be an issue.)
>>>>>
>>>>> Partial reading firmware will result in needing to pre-read the entire
>>>>> file, most likely on the security pre hook.
>>>> The entire file may be very large and not fit into a buffer.
>>>> Hence one of the reasons for a partial read of the file.
>>>> For security purposes, you need to change your code to limit the amount
>>>> of data it reads into a buffer at one time to not consume or run out of much
>>>> memory.
>>> Hm? That's not how whole-file hashing works. :)
>>> These hooks need to finish their hashing and policy checking before they
>>> can allow the rest of the code to move forward. (That's why it's a
>>> security hook.) If kernel memory utilization is the primary concern,
>>> then sure, things could be rearranged to do partial read and update the
>>> hash incrementally, but the entire file still needs to be locked,
>>> entirely hashed by hook, then read by the caller, then unlocked and
>>> released.
> Exactly.
>
>>> So, if you want to have partial file reads work, you'll need to
>>> rearchitect the way this works to avoid regressing the security coverage
>>> of these operations.
>> I am not familiar with how the security handling code works at all.
>> Is the same security check run on files opened from user space?
>> A file could be huge.
>>
>> If it assumes there is there is enough memory available to read the
>> entire file into kernel space then the improvement below can be left as
>> a memory optimization to be done in an independent (or future) patch series.
> There are two security hooks - security_kernel_read_file(),
> security_kernel_post_read_file - in kernel_read_file().  The first
> hook is called before the file is read into a buffer, while the second
> hook is called afterwards.
>
> For partial reads, measuring the firmware and verifying the firmware's
> signature will need to be done on the security_kernel_read_file()
> hook.
>
>>> So, probably, the code will look something like:
>>>
>>>
>>> file = kernel_open_file_for_reading(...)
>>> 	file = open...
>>> 	disallow_writes(file);
>>> 	while (processed < size-of-file) {
>>> 		buf = read(file, size...)
>>> 		security_file_read_partial(buf)
>>> 	}
>>> 	ret = security_file_read_finished(file);
>>> 	if (ret < 0) {
>>> 		allow_writes(file);
>>> 		return PTR_ERR(ret);
>>> 	}
>>> 	return file;
>>>
>>> while (processed < size-of-file) {
>>> 	buf = read(file, size...)
>>> 	firmware_send_partial(buf);
>>> }
>>>
>>> kernel_close_file_for_reading(file)
>>> 	allow_writes(file);
> Right, the ima_file_mmap(), ima_bprm_check(), and ima_file_check()
> hooks call process_measurement() to do this.  ima_post_read_file()
> passes a buffer to process_measurement() instead.
>
> Scott, the change should be straight forward.  The additional patch
> needs to:
> - define a new kernel_read_file_id enumeration, like
> FIRMWARE_PARTIAL_READ.
> - Currently ima_read_file() has a comment about pre-allocated firmware
> buffers.  Update ima_read_file() to call process_measurement() for the
> new enumeration FIRMWARE_PARTIAL_READ and update ima_post_read_file()
> to return immediately.
Should this be what is in ima_read_file?
{
     enum ima_hooks func;
     u32 secid;

     if (read_id != READING_FIRMWARE_PARTIAL_READ)
         return 0;

     if (!file) { /* should never happen */
         if (ima_appraise & IMA_APPRAISE_ENFORCE)
             return -EACCES;
         return 0;
     }

     security_task_getsecid(current, &secid);
     return process_measurement(file, current_cred(), secid, NULL,
                    0, MAY_READ, FILE_CHECK);
}
>
> The built-in IMA measurement policy contains a rule to measure
> firmware.  The policy can be specified on the boot command line by
> specifying "ima_policy=tcb".  After reading the firmware, the firmware
> measurement should be in <securityfs>/ima/ascii_runtime_measurements.
>
> thanks,
>
> Mimi