796aec4a5b
Sapphire Rapids and Emerald Rapids SOCs turn out to have a bug that has
security implications.
Both of these accelerators work by the application submitting a 64 byte
command to the device; this command contains an opcode as well as the
virtual address of the return value that the device will update on
completion... and a set of opcode specific values.
In a typical scenario a ring 3 application mmaps the device file and uses
the ENQCMD or MOVDIR64 instructions (which are variations of a 64 byte
atomic write) on this mmap'd memory region to directly submit commands to a
device hardware.
The return value as specified in the command, is supposed to be 32 (or 64)
bytes aligned in memory, and generally the hardware checks and enforces
this alignment.
However in testing it has been found that there are conditions (controlled
by the submitter) where this enforcement does not happen.... which makes it
possible for the return value to span a page boundary. And this is where it
goes wrong - the accelerators will perform the virtual to physical address
lookup on the first of the two pages, but end up continue writing to the
next consecutive physical (host) page rather than the consecutive virtual
page. In addition, the device will end up in a hung state on such unaligned
write of the return value.
This patch series has the proposed software side solution consisting of 3 parts
part 1: Don't allow these two PCI devices to be assigned to VM guests
(we cannot trust a VM guest to behave correctly and not cause this condition)
part 2: Don't allow ring 3 applications to set up the mmap unless they have
CAP_SYS_RAWIO permissions. This makes it no longer possible for non-root
applications to directly submit commands to the accelerator
part 3: Add a write() method to the device so that an application can submit
its commands to the kernel driver, which performs the needed sanity checks
before submitting it to the hardware.
This switch from mmap to write is an incompatible interface change to
non-root userspace, but we have not found a way to avoid this.
All software we know of uses a small set of accessor libraries for these
accelerators, for which libqpl and libdml (on github) are the most common.
As part of the security release, updated versions of these libraries will be
released that transparently fall back to write().
Intel has assigned CVE-2024-21823 to this hardware issue.
-----BEGIN PGP SIGNATURE-----
iHUEABYKAB0WIQR7kcW/9hKL3nKiNhEbbuU1gZlMiwUCZkJGUwAKCRAbbuU1gZlM
i9kzAQCzkirMANgOz8YEV3QqxPoE6pVWRL/mrsJw2yVzv3RcVwEA/suqCkBXO5zx
H9QzRXV0SGr5DJkra4ktSAwKUIfGYAQ=
=8m4L
-----END PGP SIGNATURE-----
Merge tag 'idxd-for-linus-may2024' of git bundle from Arjan
Pull DSA and IAA accelerator mis-alignment fix from Arjan van de Ven:
"The DSA (memory copy/zero/etc) and IAA (compression) accelerators in
the Sapphire Rapids and Emerald Rapids SOCs turn out to have a bug
that has security implications.
Both of these accelerators work by the application submitting a 64
byte command to the device; this command contains an opcode as well as
the virtual address of the return value that the device will update on
completion... and a set of opcode specific values.
In a typical scenario a ring 3 application mmaps the device file and
uses the ENQCMD or MOVDIR64 instructions (which are variations of a 64
byte atomic write) on this mmap'd memory region to directly submit
commands to a device hardware.
The return value as specified in the command, is supposed to be 32 (or
64) bytes aligned in memory, and generally the hardware checks and
enforces this alignment.
However in testing it has been found that there are conditions
(controlled by the submitter) where this enforcement does not
happen... which makes it possible for the return value to span a page
boundary. And this is where it goes wrong - the accelerators will
perform the virtual to physical address lookup on the first of the two
pages, but end up continue writing to the next consecutive physical
(host) page rather than the consecutive virtual page. In addition, the
device will end up in a hung state on such unaligned write of the
return value.
This patch series has the proposed software side solution consisting
of three parts:
- Don't allow these two PCI devices to be assigned to VM guests (we
cannot trust a VM guest to behave correctly and not cause this
condition)
- Don't allow ring 3 applications to set up the mmap unless they have
CAP_SYS_RAWIO permissions. This makes it no longer possible for
non-root applications to directly submit commands to the
accelerator
- Add a write() method to the device so that an application can
submit its commands to the kernel driver, which performs the needed
sanity checks before submitting it to the hardware.
This switch from mmap to write is an incompatible interface change to
non-root userspace, but we have not found a way to avoid this. All
software we know of uses a small set of accessor libraries for these
accelerators, for which libqpl and libdml (on github) are the most
common. As part of the security release, updated versions of these
libraries will be released that transparently fall back to write().
Intel has assigned CVE-2024-21823 to this hardware issue"
* tag 'idxd-for-linus-may2024' of git bundle from Arjan:
dmaengine: idxd: add a write() method for applications to submit work
dmaengine: idxd: add a new security check to deal with a hardware erratum
VFIO: Add the SPR_DSA and SPR_IAX devices to the denylist
Linux kernel
============
There are several guides for kernel developers and users. These guides can
be rendered in a number of formats, like HTML and PDF. Please read
Documentation/admin-guide/README.rst first.
In order to build the documentation, use ``make htmldocs`` or
``make pdfdocs``. The formatted documentation can also be read online at:
https://www.kernel.org/doc/html/latest/
There are various text files in the Documentation/ subdirectory,
several of them using the reStructuredText markup notation.
Please read the Documentation/process/changes.rst file, as it contains the
requirements for building and running the kernel, and information about
the problems which may result by upgrading your kernel.