| Message ID | 20200709061911.954326-1-tyhicks@linux.microsoft.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | ima: Fix rule parsing bugs and extend KEXEC_CMDLINE rule support | expand |
On Thu, 2020-07-09 at 01:18 -0500, Tyler Hicks wrote: > This series ultimately extends the supported IMA rule conditionals for > the KEXEC_CMDLINE hook function. As of today, there's an imbalance in > IMA language conditional support for KEXEC_CMDLINE rules in comparison > to KEXEC_KERNEL_CHECK and KEXEC_INITRAMFS_CHECK rules. The KEXEC_CMDLINE > rules do not support *any* conditionals so you cannot have a sequence of > rules like this: > > dont_measure func=KEXEC_KERNEL_CHECK obj_type=foo_t > dont_measure func=KEXEC_INITRAMFS_CHECK obj_type=foo_t > dont_measure func=KEXEC_CMDLINE obj_type=foo_t > measure func=KEXEC_KERNEL_CHECK > measure func=KEXEC_INITRAMFS_CHECK > measure func=KEXEC_CMDLINE > > Instead, KEXEC_CMDLINE rules can only be measured or not measured and > there's no additional flexibility in today's implementation of the > KEXEC_CMDLINE hook function. > > With this series, the above sequence of rules becomes valid and any > calls to kexec_file_load() with a kernel and initramfs inode type of > foo_t will not be measured (that includes the kernel cmdline buffer) > while all other objects given to a kexec_file_load() syscall will be > measured. There's obviously not an inode directly associated with the > kernel cmdline buffer but this patch series ties the inode based > decision making for KEXEC_CMDLINE to the kernel's inode. I think this > will be intuitive to policy authors. > > While reading IMA code and preparing to make this change, I realized > that the buffer based hook functions (KEXEC_CMDLINE and KEY_CHECK) are > quite special in comparison to longer standing hook functions. These > buffer based hook functions can only support measure actions and there > are some restrictions on the conditionals that they support. However, > the rule parser isn't enforcing any of those restrictions and IMA policy > authors wouldn't have any immediate way of knowing that the policy that > they wrote is invalid. For example, the sequence of rules above parses > successfully in today's kernel but the > "dont_measure func=KEXEC_CMDLINE ..." rule is incorrectly handled in > ima_match_rules(). The dont_measure rule is *always* considered to be a > match so, surprisingly, no KEXEC_CMDLINE measurements are made. > > While making the rule parser more strict, I realized that the parser > does not correctly free all of the allocated memory associated with an > ima_rule_entry when going down some error paths. Invalid policy loaded > by the policy administrator could result in small memory leaks. > > I envision patches 1-7 going to stable. The series is ordered in a way > that has all the fixes up front, followed by cleanups, followed by the > feature patch. The breakdown of patches looks like so: > > Memory leak fixes: 1-3 > Parser strictness fixes: 4-7 > Code cleanups made possible by the fixes: 8-11 > Extend KEXEC_CMDLINE rule support: 12 Thanks, Tyler. This is a really nice patch set. The patches are now in the "next-integrity-testing" branch. Mimi
On 2020-07-17 00:31:33, Mimi Zohar wrote: > On Thu, 2020-07-09 at 01:18 -0500, Tyler Hicks wrote: > > This series ultimately extends the supported IMA rule conditionals for > > the KEXEC_CMDLINE hook function. As of today, there's an imbalance in > > IMA language conditional support for KEXEC_CMDLINE rules in comparison > > to KEXEC_KERNEL_CHECK and KEXEC_INITRAMFS_CHECK rules. The KEXEC_CMDLINE > > rules do not support *any* conditionals so you cannot have a sequence of > > rules like this: > > > > dont_measure func=KEXEC_KERNEL_CHECK obj_type=foo_t > > dont_measure func=KEXEC_INITRAMFS_CHECK obj_type=foo_t > > dont_measure func=KEXEC_CMDLINE obj_type=foo_t > > measure func=KEXEC_KERNEL_CHECK > > measure func=KEXEC_INITRAMFS_CHECK > > measure func=KEXEC_CMDLINE > > > > Instead, KEXEC_CMDLINE rules can only be measured or not measured and > > there's no additional flexibility in today's implementation of the > > KEXEC_CMDLINE hook function. > > > > With this series, the above sequence of rules becomes valid and any > > calls to kexec_file_load() with a kernel and initramfs inode type of > > foo_t will not be measured (that includes the kernel cmdline buffer) > > while all other objects given to a kexec_file_load() syscall will be > > measured. There's obviously not an inode directly associated with the > > kernel cmdline buffer but this patch series ties the inode based > > decision making for KEXEC_CMDLINE to the kernel's inode. I think this > > will be intuitive to policy authors. > > > > While reading IMA code and preparing to make this change, I realized > > that the buffer based hook functions (KEXEC_CMDLINE and KEY_CHECK) are > > quite special in comparison to longer standing hook functions. These > > buffer based hook functions can only support measure actions and there > > are some restrictions on the conditionals that they support. However, > > the rule parser isn't enforcing any of those restrictions and IMA policy > > authors wouldn't have any immediate way of knowing that the policy that > > they wrote is invalid. For example, the sequence of rules above parses > > successfully in today's kernel but the > > "dont_measure func=KEXEC_CMDLINE ..." rule is incorrectly handled in > > ima_match_rules(). The dont_measure rule is *always* considered to be a > > match so, surprisingly, no KEXEC_CMDLINE measurements are made. > > > > While making the rule parser more strict, I realized that the parser > > does not correctly free all of the allocated memory associated with an > > ima_rule_entry when going down some error paths. Invalid policy loaded > > by the policy administrator could result in small memory leaks. > > > > I envision patches 1-7 going to stable. The series is ordered in a way > > that has all the fixes up front, followed by cleanups, followed by the > > feature patch. The breakdown of patches looks like so: > > > > Memory leak fixes: 1-3 > > Parser strictness fixes: 4-7 > > Code cleanups made possible by the fixes: 8-11 > > Extend KEXEC_CMDLINE rule support: 12 > > Thanks, Tyler. This is a really nice patch set. The patches are now > in the "next-integrity-testing" branch. Thank you for all the helpful review comments. You know where to find me if any bugs pop up during testing. :) Tyler > > Mimi
[Cc'ing Sasha] On Thu, 2020-07-09 at 01:18 -0500, Tyler Hicks wrote: > I envision patches 1-7 going to stable. The series is ordered in a way > that has all the fixes up front, followed by cleanups, followed by the > feature patch. The breakdown of patches looks like so: > > Memory leak fixes: 1-3 > Parser strictness fixes: 4-7 > Code cleanups made possible by the fixes: 8-11 > Extend KEXEC_CMDLINE rule support: 12 I agree they should be backported, but they don't apply cleanly before linux-5.6. The changes aren't that major. Some patch hunks apply cleanly, but won't compile, while others patch hunks need to be dropped based on when the feature was upstreamed. For these reasons, I'm not Cc'ing stable. Feature upstreamed: - LSM policy update: linux 5.3 - key command line: linux 5.3 - blacklist: linux 5.5 - keyrings: linux 5.6 For Linux 5.3: - Dependency on backporting commit 483ec26eed42 ("ima: ima/lsm policy rule loading logic bug fixes") to apply " ima: Free the entire rule if it fails to parse". Mimi
This series ultimately extends the supported IMA rule conditionals for the KEXEC_CMDLINE hook function. As of today, there's an imbalance in IMA language conditional support for KEXEC_CMDLINE rules in comparison to KEXEC_KERNEL_CHECK and KEXEC_INITRAMFS_CHECK rules. The KEXEC_CMDLINE rules do not support *any* conditionals so you cannot have a sequence of rules like this: dont_measure func=KEXEC_KERNEL_CHECK obj_type=foo_t dont_measure func=KEXEC_INITRAMFS_CHECK obj_type=foo_t dont_measure func=KEXEC_CMDLINE obj_type=foo_t measure func=KEXEC_KERNEL_CHECK measure func=KEXEC_INITRAMFS_CHECK measure func=KEXEC_CMDLINE Instead, KEXEC_CMDLINE rules can only be measured or not measured and there's no additional flexibility in today's implementation of the KEXEC_CMDLINE hook function. With this series, the above sequence of rules becomes valid and any calls to kexec_file_load() with a kernel and initramfs inode type of foo_t will not be measured (that includes the kernel cmdline buffer) while all other objects given to a kexec_file_load() syscall will be measured. There's obviously not an inode directly associated with the kernel cmdline buffer but this patch series ties the inode based decision making for KEXEC_CMDLINE to the kernel's inode. I think this will be intuitive to policy authors. While reading IMA code and preparing to make this change, I realized that the buffer based hook functions (KEXEC_CMDLINE and KEY_CHECK) are quite special in comparison to longer standing hook functions. These buffer based hook functions can only support measure actions and there are some restrictions on the conditionals that they support. However, the rule parser isn't enforcing any of those restrictions and IMA policy authors wouldn't have any immediate way of knowing that the policy that they wrote is invalid. For example, the sequence of rules above parses successfully in today's kernel but the "dont_measure func=KEXEC_CMDLINE ..." rule is incorrectly handled in ima_match_rules(). The dont_measure rule is *always* considered to be a match so, surprisingly, no KEXEC_CMDLINE measurements are made. While making the rule parser more strict, I realized that the parser does not correctly free all of the allocated memory associated with an ima_rule_entry when going down some error paths. Invalid policy loaded by the policy administrator could result in small memory leaks. I envision patches 1-7 going to stable. The series is ordered in a way that has all the fixes up front, followed by cleanups, followed by the feature patch. The breakdown of patches looks like so: Memory leak fixes: 1-3 Parser strictness fixes: 4-7 Code cleanups made possible by the fixes: 8-11 Extend KEXEC_CMDLINE rule support: 12 Perhaps the most logical ordering for code review is: 1, 2, 3, 8, 9, 4, 5, 6, 7, 10, 11, 12 If you'd like me to re-order or split up the series, just let me know. Thanks for considering these patches! * Series-wide v3 changes - Indentation changes in patch #4 which caused some churn - Added patch #7 - Significant changes to patch #10 to address Mimi's requests * Series-wide v2 changes - Rebased onto next-integrity-testing - Squashed patches 2 and 3 from v1 + Updated this cover letter to account for changes to patch index changes + See patch 2 for specific code changes Tyler Tyler Hicks (12): ima: Have the LSM free its audit rule ima: Free the entire rule when deleting a list of rules ima: Free the entire rule if it fails to parse ima: Fail rule parsing when buffer hook functions have an invalid action ima: Fail rule parsing when the KEXEC_CMDLINE hook is combined with an invalid cond ima: Fail rule parsing when the KEY_CHECK hook is combined with an invalid cond ima: Fail rule parsing when appraise_flag=blacklist is unsupportable ima: Shallow copy the args_p member of ima_rule_entry.lsm elements ima: Use correct type for the args_p member of ima_rule_entry.lsm elements ima: Move comprehensive rule validation checks out of the token parser ima: Use the common function to detect LSM conditionals in a rule ima: Support additional conditionals in the KEXEC_CMDLINE hook function include/linux/ima.h | 4 +- kernel/kexec_file.c | 2 +- security/integrity/ima/ima.h | 13 +- security/integrity/ima/ima_api.c | 2 +- security/integrity/ima/ima_appraise.c | 2 +- security/integrity/ima/ima_asymmetric_keys.c | 2 +- security/integrity/ima/ima_main.c | 23 ++- security/integrity/ima/ima_modsig.c | 20 -- security/integrity/ima/ima_policy.c | 206 ++++++++++++++----- security/integrity/ima/ima_queue_keys.c | 2 +- 10 files changed, 182 insertions(+), 94 deletions(-)