On 2020-07-07 4:36 p.m., Kees Cook wrote:
On Tue, Jul 07, 2020 at 10:13:42AM -0700, Scott Branden wrote:
You and others are certainly more experts in the filesystem and security
infrastructure of the kernel.
What I am trying to accomplish is a simple operation:
request part of a file into a buffer rather than the whole file.
If someone could add such support I would be more than happy to use it.
Sure, and I totally understand that, but as it happens, no one has stepped
up with spare time to do that work. Since you're the person with the need
for the API, it falls to you to do it. And I understand what feature creep
feels like (I needed to fix one design problem[1] with timers, and I spent
months sending hundreds of patches). Some times you get lucky and it's
easy, and sometimes you end up touching something that needs a LOT of work
to refactor before you can make your desired change work well with the
rest of the kernel and be maintainable by other people into the future.
Quick tangent: I can't find in the many many threads where you explain
how large these firmware images are and why existing kernel memory
allocations are insufficient to load them. How large are these[2] files?
/lib/firmware/vk-boot1-bcm958401m2.ecdsa.bin
This is on the order of a few MB at most.
/lib/firmware/vk-boot2-bcm958401m2_a72.ecdsa.bin
Some of these images are current 250MB. At we anticipate them growing
to 512MB.
And, we do have systems with the driver loading 16 cards in parallel
with no requirement that they are the same images
(although loading 16 different images to 16 different cards would be
strange but possible).
For me, the requirements for partial read support are these things,
which are the characteristics of the existing API:
- the LSM must be able to validate the entire file contents before
any data is available to any reader. (Which was pointed out back in
August 2019[3].)
I thought this was addressed in patch v6 "ima: aad FIRMWARE_PARTIAL_READ
support"
https://lkml.org/lkml/2020/6/5/1126
(although implementation not to your liking in current review)?
And "any" reader includes having a DMA window open
on the memory range used for reading the contents (which was pointed
out at by Mimi[4] but went unanswered and remains broken still in this
v10, but I will comment separately on that.)
- the integrity of the file contents must be maintained between
validation and delivery (currently this is handled internally via
disallow_writes()).
I don't understand what you are staying here: I am request a partial
firmware read into a buf.
At the time the partial firmware is read into a buf it is validated by
the security module if such integrity checks are enabled.
If, at another time I wish to request another partial firmware into a
buffer (could be the same part of the file or a different part of a file
or from another file), the integrity check is performed again and the
portion of the file I request should be put into the buffer.
If a lock on a file is needed by somebody between these partial reads
that is a different API and out of the scope of my patch series.
This has now bubbled into many other designs issues in the existing
codebase.
Correct -- this is one of the many difficulties of contributing to a
large and complex code base with many maintainers. There can be a lot
of requirements for the code that have nothing to do with seemingly more
narrow areas of endeavour.
Thanks at least for helping with guidance, I see your review is thought
out and
hopefully we can come to a conclusion as I feel we are fairly close with
your changes.
I will need more details on your comments - see below.
On 2020-07-06 8:08 p.m., Kees Cook wrote:
On Mon, Jul 06, 2020 at 04:23:09PM -0700, Scott Branden wrote:
Add FIRMWARE_PARTIAL_READ support for integrity
measurement on partial reads of firmware files.
Hi,
Several versions ago I'd suggested that the LSM infrastructure handle
the "full read" semantics so that individual LSMs don't need to each
duplicate the same efforts. As it happens, only IMA is impacted (SELinux
ignores everything except modules, and LoadPin only cares about origin
not contents).
Does your patch series "Fix misused kernel_read_file() enums" handle this
because this suggestion is outside the scope of my change?
My proposed patch series cleans up a number of mistakes that were made
to the kernel_read_file() API, and helps clarify my point about the
enums being used for *what*, and not *how* or *where*, which needs to
be fixed in this series and shouldn't be a big deal (I will reply to
individual patches).
Next is the problem that enum kernel_read_file_id is an object
TYPE enum, not a HOW enum. (And it seems I missed the addition of
READING_FIRMWARE_PREALLOC_BUFFER, which may share a similar problem.)
That it's a partial read doesn't change _what_ you're reading: that's an
internal API detail. What happens when I attempt to do a partial read of
a kexec image?
It does not appear there is any user of partial reads of kexec images?
I have been informed by Greg K-H to not add apis that are not used so such
support doesn't make sense to add at this time.
But you are proposing a generic API enhancement that any other user in
the kernel may end up using. Just because the bcm-vk driver is the only
user now, and IMA is the only LSM performing content analysis, it
doesn't mean that there won't be another driver added later, nor another
LSM. In fact, the new BPF LSM means that anything exposed by LSM hooks
is now available for analysis.
I'll use kernel_pread_file() and pass READING_KEXEC_IMAGE,
but the LSMs will have no idea it's a partial read.
The addition I am adding is for request_partial_firmware_into_buf.
In order to do so it adds internal support for partial reads of firmware
files,
not kexec image.
Yes, but you're changing kernel_read_file() APIs to do it. There are
plenty of users of that API. Maybe they would like to also use partial
reads?
$ git grep kernel_read_file | wc -l
77
The above seems outside the scope of my patch?
Unfortunately, it is not. Part of your responsibility as a kernel
developer making API changes/additions is that those changes need to
interact correctly with the rest of the kernel (and be maintainable).
Finally, what keeps the contents of the file from changing between the
first call (which IMA will read the entire file for) and the next reads
which will bypass IMA?
The request is for a partial read. IMA ensures the whole file integrity
even though I only do a partial read.
The next partial read will re-read and check integrity of file.
So, while terribly inefficient, I guess this approach is tenable. It
means that each partial read will trigger a full read for LSMs that care
about the hook.
So, to that end, I wonder why IMA doesn't do this for all file types?
It also means that we won't have a strict pairing of
security_kernel_read_file() to security_kernel_post_read_file() in the
LSMs, but it seems that only IMA currently explicitly cares about this
(or maybe not at all).
I'm not entirely happy about using this design, but it does appear
sufficient.
Yes, terribly inefficient, but if somebody wants to do some optimization
they are welcome to it.
In fact, I want to call a partial read of a file with NO security. Can I
add such a call instead?
If we did that now then the inefficient read of the file multiple times
to authenticate it each time wouldn't be introduced.
In my use case the linux kernel security on the file is quite
meaningless and a waste of time to even perform.
Whether the file has been compromised, corrupt or otherwise the image is
validated by the hardware after
the image is loaded to the card to ensure it passes authentication.
I'd suggested that the open file must have writes
disabled on it (as execve() does).
The file will be reopened and integrity checked on the next partial read (if
there is one).
So I don't think there is any change to be made here.
If writes aren't already disabled for a whole file read then that is
something that needs to be fixed in the existing code.
My suggestion quoted here was operating on the idea that whole-file
verification wasn't happening on every partial read, so this isn't a
problem in that case.
So, please redesign this:
- do not add an enum
I used existing infrastructure provided by Mimi but now looks like it will
have to fit with your patches from yesterday.
Right, this won't be hard. I will send a v2 of my patches to clarify the
purpose of the 3 file content hooks (load_data, read_file,
post_read_file), which might need renaming...
I see your cleanup and merged with it. Will need to test everything again.
- make the file unwritable for the life of having the handle open
It's no different than a full file read so no change to be made here.
Correct.
- make the "full read" happen as part of the first partial read so the
LSMs don't have to reimplement everything
Each partial read is an individual operation so I think a "full read" is
performed every time
if your security IMA is enabled. If someone wants to add a file lock and
then partial reads in the kernel
then that would be different than what is needed by the kernel driver.
So, given that Mimi is (I think?) satisfied with your approach here, I
can't realistically complain. I still don't like the idea of each LSM
needing to perform it's full read loop for the contents, but so be it:
IMA will have the code, SELinux doesn't care (yet), and LoadPin doesn't
care about contents.
I obviously am not aware of all of the security hooks and architecture
in place but I do see your cleanups as a simplification over what was there.
-Kees
[1] https://lore.kernel.org/lkml/20170808003345.GA107289@beast/
git log --oneline --grep 'Convert timer' --author "Kees Cook" | wc -l
271
[2] https://lore.kernel.org/lkml/824407ae-8ab8-0fe3-bd72-270fce960ac5@xxxxxxxxxxxx/
[3] https://lore.kernel.org/lkml/s5hsgpsqd49.wl-tiwai@xxxxxxx/
[4] https://lore.kernel.org/lkml/1591622862.4638.74.camel@xxxxxxxxxxxxx/
