Fix 2 bugs in non-raw send with encryption

[gamanakis]

Motivation and Context

Closes #12014

Description

Bisecting identified the redacted send/receive as the source of the bug
for issue #12014. Specifically the call to
dsl_dataset_hold_obj(&fromds) has been replaced by
dsl_dataset_hold_obj_flags() which passes a DECRYPT flag and creates
a key mapping. A subsequent dsl_dataset_rele_flag(&fromds) is missing
and the key mapping is not cleared. This may be inadvertedly used, which
results in arc_untransform failing with ECKSUM in:
arc_untransform+0x96/0xb0 [zfs]
dbuf_read_verify_dnode_crypt+0x196/0x350 [zfs]
dbuf_read+0x56/0x770 [zfs]
dmu_buf_hold_by_dnode+0x4a/0x80 [zfs]
zap_lockdir+0x87/0xf0 [zfs]
zap_lookup_norm+0x5c/0xd0 [zfs]
zap_lookup+0x16/0x20 [zfs]
zfs_get_zplprop+0x8d/0x1d0 [zfs]
setup_featureflags+0x267/0x2e0 [zfs]
dmu_send_impl+0xe7/0xcb0 [zfs]
dmu_send_obj+0x265/0x360 [zfs]
zfs_ioc_send+0x10c/0x280 [zfs]

Fix this by restoring the call to dsl_dataset_hold_obj().

The same applies for to_ds: here replace dsl_dataset_rele(&to_ds) with
dsl_dataset_rele_flags().

Both leaked key mappings will cause a panic when exporting the
sending pool or unloading the zfs module after a non-raw send from
an encrypted filesystem.

How Has This Been Tested?

Manually running the scripts https://github.com/HankB/provoke_ZFS_corruption

Types of changes

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Performance enhancement (non-breaking change which improves efficiency)
• Code cleanup (non-breaking change which makes code smaller or more readable)
• Breaking change (fix or feature that would cause existing functionality to change)
• Library ABI change (libzfs, libzfs_core, libnvpair, libuutil and libzfsbootenv)
• Documentation (a change to man pages or other documentation)

Checklist:

• My code follows the OpenZFS code style requirements.
• I have updated the documentation accordingly.
• I have read the contributing document.
• I have added tests to cover my changes.
• I have run the ZFS Test Suite with this change applied.
• All commit messages are properly formatted and contain Signed-off-by.

[gamanakis]

How can I acknowledge @HankB and @pcd1193182 for the significant effort they put into this?

[amotin]

Makes sense to me. Though after searching around this area for couple hours I still have no idea why it helps with the corruptions. I think it is only a trigger, but for what?

[tonyhutter]

How can I acknowledge @HankB and @pcd1193182 for the significant effort they put into this?

Contributions-by: would be my recommendation. I see some of those in the git logs.

[behlendorf]

The fixes here make sense. But I agree, it's not at all clear to me how this would lead to a corruption. Was the final reproducer for this something which could be reasonably adapted and included in the test suite?

[ryao]

The fixes here make sense. But I agree, it's not at all clear to me how this would lead to a corruption. Was the final reproducer for this something which could be reasonably adapted and included in the test suite?

I have a WIP CodeQL check for instances of this problem in my git repository:

https://github.com/ryao/zfs/tree/issue-12014

I am waiting to see if it works. If it does, we should think about adding checks for other functions that could be accidentally mismatched.

[ryao]

ryao/zfs@7c921e3 works and detected the two issues fixed in this patch. The current iteration is limited to cases where the mismatched hold and release functions are called from the same function. I might try to handle more cases before I open a PR with the check.

[ryao]

The current iteration is limited to cases where the mismatched hold and release functions are called from the same function. I might try to handle more cases before I open a PR with the check.

I just did a manual search to see if this is even necessary. It turns out that it is. dmu_objset_hold_flags() calls dsl_dataset_hold_flags(), which calls dsl_dataset_hold_obj_flags(). The returned dsl_dataset_t pointer is then passed to dsl_dataset_rele(), which is wrong, on error from dmu_objset_from_ds() in dmu_objset_hold_flags(). If I expand the power of the check to catch mismatched hold and release functions across function calls, the check should be able to identify this and possibly other bugs.

[gamanakis]

@ryao thank you for taking this to the next level.
@behlendorf I think expanding the ZTS would be great, but the scripts take about 2-3h to reproduce this.

[gamanakis]

I believe the crucial issue in #12014 is the failure to clear key mappings, which is not currently handled in the code base. My suspicion is that the apparent corruption in ARC stems from a scenario where an object (snapshot) is freed, but its key mapping persists. In the current code base this happens for both fromds and to_ds. This suggests the stale key mappings are inadvertently reused.

In this context, it would likely be safe to use dsl_dataset_hold_obj_flags(&fromds) as long as we call dsl_dataset_rele_flags(&fromds) afterwards.

[HankB]

Was the final reproducer for this something which could be reasonably adapted and included in the test suite?

I'll comment because I wrote the reproducer. (Unfortunately I remain blissfully ignorant of the test suite and how it is run.)

I produced a series of scripts that:

• Populate a pool with compressible and random files in multiple nested datasets.
• Modify files, create and delete snapshots and send the encrypted pool to an unencrypted pool. (IOW, thrash on the pool.)

The setup of the tests is manual (copying and pasting commands from previous test notes.) This could all be scripted, I just didn't do that.

On my H/W - ten year old server motherboard, two 500GB SATA SSDs - it takes about 2 hours typically to reproduce corruption. One problem with automating it is that the results are indeterminate. It's "whale on it to see if it breaks." If it typically breaks at 2 hours, how long insures with reasonable confidence that the problem does not exist? Some earlier testing took on the order of a day to provoke corruption but I was able to bring that down by increasing the activity on the pool. It's conceivable that other changes to the stack could result in reducing the chance of corruption, delaying onset past the typical two hours.

Perhaps the next thing would be to fully automate the setup and bring the existing scripts up to "production quality." Others have used the scripts and been kind not comment on the code quality (beyond submitting PRs for bugs that I had overlooked.) Whether the scripts could be included in a test suite is left to those more familiar with the test suite.

[gamanakis]

62f125b: updated the commit message

[AndrewJDR]

I’m just an observer here, but i’m curious: Is the context and timing of the uncleared key mapping reusage understood well enough to design a synthetic test for the crash?

[owlshrimp]

I’m just an observer here, but i’m curious: Is the context and timing of the uncleared key mapping reusage understood well enough to design a synthetic test for the crash?

Or alternatively, can the theory be tested with some printf's or the like?

[gamanakis]

40567d0: updated the commit message, it's not really a panic but arc_untransform failing with ECKSUM.

The theory that the stale key mapping is reused is the only possible explanation I see. I have been trying to see if this is true but it's tricky with cmn_err(). In the current codebase all key mappings are leaked.

[behlendorf]

@HankB thanks for the clarification on the reproducer. I figured it wouldn't be easy to add to the CI or George would have done it in this PR already, but I still wanted to ask!

@ryao the codeQL check would be a nice addition to help us catch future issues, even if it is limited to within the scope of a single calling function (for now). Plus having a working example should make it pretty easy to extend to other similar interfaces.