| Message ID | 20250323140911.226137-2-nstange@suse.de (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | ima: get rid of hard dependency on SHA-1 | expand |
On Sun, 2025-03-23 at 15:08 +0100, Nicolai Stange wrote: > IMA creates one runtime_measurements_<hash-algo> sysfs file for every TPM > bank + for SHA1 if not covered by any such. These differ only in that the > template hash value for each record is of the file's associated algorithm > each. > > The kernel does not necessarily support each hash algorithm associated > with some TPM bank though -- the most common case probably being that the > algorithm is not built-in, but provided as a module, if at all, and thus > not available at IMA init time yet. At least from an IMA perspective, the IMA Kconfig selects "CRYPTO_SHA1", making it the only crypto hash algorithm required to be built into the kernel. > > If that happens to be the case, the behavior is a bit counter-intuitive: > probably for historic reasons and to still extend the TPM bank with > something, a record's template hash is filled with the padded SHA1 value. > That is, it is perfectly possible that runtime_measurements_sha256 contains > padded SHA1 template hashes if SHA-256 was unavailable at IMA init. As you explained in "[RFC PATCH v2 09/13] ima: invalidate unsupported PCR banks only once", not extending the TPM bank: b.) is a security risk, because the bank would validate an empty measurement list. The solution was to extend a padded SHA1 digest, which could still be verified. The question boils down to whether extending the TPM bank with a valid, maybe deprecated hash algo, is better than not extending it at all. Was that the right solution? I believe it was at the time. SHA1 was being deprecated for specific use cases, not all usecases. So the question is what to do going forward. > > I would argue that it's likely that no existing userspace tool is relying > on this fallback logic -- they either wouldn't consume the hash value from > the measurement list directly but recreate it by themselves, as is required > for verification against PCRs, or, if they did, they would somehow assume a > hash algorithm and expect the hashes in the measurement list to be of that > type. If of the latter kind, this could even lead to hard to debug > verification failures. For example, from looking at keylime's current > code, the verifier logic seems to assume that the template hashes found > in the provided measurement list are of the configured 'ima_log_hash_alg' > type. In particular, it does not check against padded SHA1 upon > mismatch. The downside, if none of the TPM bank hash algorithms are configured as builtin in the kernel, is the lack of a measurement list. True, not implementing the fallback is simpler, but it is their choice. > > That being said, there's also another dimension: currently IMA has a > hard requirement on SHA-1 and subsequent patches in this series will > attempt to get rid of that. If SHA-1 is not available at IMA init though, > it would also mean that padded SHA-1 values cannot get filled in as a > fallback for other unsupported algorithms. Substituting something like > hard coded all-zeroes or all-ones would be dangerous, because some > application or user scripts could perhaps (ab)use the template hashes from > the exported measurement lists for some kind of fingerprinting scheme or > so. Agreed, using the open-writer/ToMToU integrity violation to indicate an unsupported TPM bank would not be a good idea. > > In conclusion, I think it's best to not create the > runtime_measurements_<hash-algo> sysfs files for hash algorithms not > supported by the kernel. That way, applications expecting a certain > hash algorithm for the measurement list and which are not able to handle > the padded-SHA1 fallback scheme would fail with a clear indication on what > the proem is. Furthermore, as digests for unsupported banks are not > getting exposed to userspace anymore, we'll have all flexibility to > set it to any value internally, including all-ones as will be needed in > a subsequent patch when addressing PCR extend for unsupported banks. > > So, do not create runtime_measurements_<hash-algo> sysfs files for > unsupported hash algorithms. Likewise for their ascii counterparts. > > Note that at this point, SHA-1 is still mandatory, and thus, > runtime_measurements_sha1 as well as the "runtime_measurements" will > remain there, even though the code has provisions already to skip their > creation as well in case SHA-1 was unavailable. > > Signed-off-by: Nicolai Stange <nstange@suse.de> > If the purpose of this patch set is to actually remove IMA's dependency on a working SHA-1, at some point the Kconfig "select CRYPTO_SHA1" needs to be removed. Otherwise the kernel will be built with SHA1 builtin (CONFIG_CRYPTO_SHA1=y). If the purpose of this patch set is preparatory for eventually removing the SHA1 dependency, then the cover letter and the patch descriptions should indicate that. Assuming the latter, other than updating the patch description, the patch is fine. Before posting the non-RFC version of this patch, please trim the patch description. Reviewed-by: Mimi Zohar <zohar@linux.ibm.com> > --- > security/integrity/ima/ima_fs.c | 35 +++++++++++++++++++++------------ > 1 file changed, 22 insertions(+), 13 deletions(-) > > diff --git a/security/integrity/ima/ima_fs.c b/security/integrity/ima/ima_fs.c > index e4a79a9b2d58..a8df2fe5f4cb 100644 > --- a/security/integrity/ima/ima_fs.c > +++ b/security/integrity/ima/ima_fs.c > @@ -454,6 +454,9 @@ static int __init create_securityfs_measurement_lists(void) > return -ENOMEM; > > for (i = 0; i < securityfs_measurement_list_count; i++) { > + if (!ima_algo_array[i].tfm) > + continue; > + > algo = ima_algo_array[i].algo; > > sprintf(file_name, "ascii_runtime_measurements_%s", > @@ -573,20 +576,26 @@ int __init ima_fs_init(void) > if (ret != 0) > goto out; > > - binary_runtime_measurements = > - securityfs_create_symlink("binary_runtime_measurements", ima_dir, > - "binary_runtime_measurements_sha1", NULL); > - if (IS_ERR(binary_runtime_measurements)) { > - ret = PTR_ERR(binary_runtime_measurements); > - goto out; > - } > + if (ima_algo_array[ima_sha1_idx].tfm) { > + binary_runtime_measurements = > + securityfs_create_symlink("binary_runtime_measurements", > + ima_dir, > + "binary_runtime_measurements_sha1", > + NULL); > + if (IS_ERR(binary_runtime_measurements)) { > + ret = PTR_ERR(binary_runtime_measurements); > + goto out; > + } > > - ascii_runtime_measurements = > - securityfs_create_symlink("ascii_runtime_measurements", ima_dir, > - "ascii_runtime_measurements_sha1", NULL); > - if (IS_ERR(ascii_runtime_measurements)) { > - ret = PTR_ERR(ascii_runtime_measurements); > - goto out; > + ascii_runtime_measurements = > + securityfs_create_symlink("ascii_runtime_measurements", > + ima_dir, > + "ascii_runtime_measurements_sha1", > + NULL); > + if (IS_ERR(ascii_runtime_measurements)) { > + ret = PTR_ERR(ascii_runtime_measurements); > + goto out; > + } > } > > runtime_measurements_count =
Mimi Zohar <zohar@linux.ibm.com> writes: > On Sun, 2025-03-23 at 15:08 +0100, Nicolai Stange wrote: >> IMA creates one runtime_measurements_<hash-algo> sysfs file for every TPM >> bank + for SHA1 if not covered by any such. These differ only in that the >> template hash value for each record is of the file's associated algorithm >> each. >> >> The kernel does not necessarily support each hash algorithm associated >> with some TPM bank though -- the most common case probably being that the >> algorithm is not built-in, but provided as a module, if at all, and thus >> not available at IMA init time yet. > > At least from an IMA perspective, the IMA Kconfig selects "CRYPTO_SHA1", making > it the only crypto hash algorithm required to be built into the kernel. >> >> If that happens to be the case, the behavior is a bit counter-intuitive: >> probably for historic reasons and to still extend the TPM bank with >> something, a record's template hash is filled with the padded SHA1 value. >> That is, it is perfectly possible that runtime_measurements_sha256 contains >> padded SHA1 template hashes if SHA-256 was unavailable at IMA init. > > As you explained in "[RFC PATCH v2 09/13] ima: invalidate unsupported PCR banks > only once", not extending the TPM bank: > b.) is a security risk, because the bank would validate an empty measurement > list. > > The solution was to extend a padded SHA1 digest, which could still be verified. > The question boils down to whether extending the TPM bank with a valid, maybe > deprecated hash algo, is better than not extending it at all. Was that the > right solution? I believe it was at the time. SHA1 was being deprecated for > specific use cases, not all usecases. No doubt about that, I'll drop the "counter-intuitive" part. Even now (or by 2030 to be more specific), it should be only FIPS users caring about SHA1 in this context at all. > So the question is what to do going forward. > >> >> I would argue that it's likely that no existing userspace tool is relying >> on this fallback logic -- they either wouldn't consume the hash value from >> the measurement list directly but recreate it by themselves, as is required >> for verification against PCRs, or, if they did, they would somehow assume a >> hash algorithm and expect the hashes in the measurement list to be of that >> type. If of the latter kind, this could even lead to hard to debug >> verification failures. For example, from looking at keylime's current >> code, the verifier logic seems to assume that the template hashes found >> in the provided measurement list are of the configured 'ima_log_hash_alg' >> type. In particular, it does not check against padded SHA1 upon >> mismatch. > > The downside, if none of the TPM bank hash algorithms are configured as builtin > in the kernel, is the lack of a measurement list. Yes. Just for the record, going forward SHA256 will be, with [v2 05/13]. So unless the SHA256 bank is disabled by firmware, it should be fine then. > True, not implementing the fallback is simpler, but it is their > choice. The point I was trying to make is that no tool currently exists that would successfully verify a /sys/*/*runtime_measurements_sha256 with SHA1s template hashes in it. keylime would not as outlined above, and ima-evm-utils ima_measurement command wouldn't either, because it works only on /sys/*/*runtime_measurements_sha1 to begin with. (I previously missed the latter fact, but its ima_verify_template_hash() does a hard-coded SHA1). That is, ima_measurement's fallback logic only applies to the PCR validations, not to the sysfs files. >> That being said, there's also another dimension: currently IMA has a >> hard requirement on SHA-1 and subsequent patches in this series will >> attempt to get rid of that. If SHA-1 is not available at IMA init though, >> it would also mean that padded SHA-1 values cannot get filled in as a >> fallback for other unsupported algorithms. Substituting something like >> hard coded all-zeroes or all-ones would be dangerous, because some >> application or user scripts could perhaps (ab)use the template hashes from >> the exported measurement lists for some kind of fingerprinting scheme or >> so. > > Agreed, using the open-writer/ToMToU integrity violation to indicate an > unsupported TPM bank would not be a good idea. > >> >> In conclusion, I think it's best to not create the >> runtime_measurements_<hash-algo> sysfs files for hash algorithms not >> supported by the kernel. That way, applications expecting a certain >> hash algorithm for the measurement list and which are not able to handle >> the padded-SHA1 fallback scheme would fail with a clear indication on what >> the proem is. Furthermore, as digests for unsupported banks are not >> getting exposed to userspace anymore, we'll have all flexibility to >> set it to any value internally, including all-ones as will be needed in >> a subsequent patch when addressing PCR extend for unsupported banks. >> >> So, do not create runtime_measurements_<hash-algo> sysfs files for >> unsupported hash algorithms. Likewise for their ascii counterparts. >> >> Note that at this point, SHA-1 is still mandatory, and thus, >> runtime_measurements_sha1 as well as the "runtime_measurements" will >> remain there, even though the code has provisions already to skip their >> creation as well in case SHA-1 was unavailable. >> >> Signed-off-by: Nicolai Stange <nstange@suse.de> >> > > If the purpose of this patch set is to actually remove IMA's dependency on a > working SHA-1, at some point the Kconfig "select CRYPTO_SHA1" needs to be > removed. Otherwise the kernel will be built with SHA1 builtin > (CONFIG_CRYPTO_SHA1=y). I should have described it better. In the first step at least, the goal is to remove the runtime dependency only. Because when SHA1's ->fips_allowed in crypto/testmgr.c gets flipped to false, SHA1 instantiation would fail with -ENOENT if the kernel was booted with a fips=1 on its command line. Users not interested in FIPS, i.e. the vast majority, might still want to use SHA1 and there's no real reason not to. But yes, it would definitely make sense to drop the CRYPTO_SHA1 dep, at least at some point. Perhaps by simply moving it to the new IMA_COMPAT_FALLBACK_TPM_EXTEND. I would personally not do that now though, just in case there'll be some unexpected fallout from this series. > If the purpose of this patch set is preparatory for eventually removing the SHA1 > dependency, then the cover letter and the patch descriptions should indicate > that. > > Assuming the latter, other than updating the patch description, the patch is > fine. > > Before posting the non-RFC version of this patch, please trim the patch > description. Point taken. > Reviewed-by: Mimi Zohar <zohar@linux.ibm.com> Thanks _a lot_! Nicolai >
> > > I would argue that it's likely that no existing userspace tool is relying > > > on this fallback logic -- they either wouldn't consume the hash value from > > > the measurement list directly but recreate it by themselves, as is required > > > for verification against PCRs, or, if they did, they would somehow assume a > > > hash algorithm and expect the hashes in the measurement list to be of that > > > type. If of the latter kind, this could even lead to hard to debug > > > verification failures. For example, from looking at keylime's current > > > code, the verifier logic seems to assume that the template hashes found > > > in the provided measurement list are of the configured 'ima_log_hash_alg' > > > type. In particular, it does not check against padded SHA1 upon > > > mismatch. > > > > The downside, if none of the TPM bank hash algorithms are configured as builtin > > in the kernel, is the lack of a measurement list. > > Yes. Just for the record, going forward SHA256 will be, with [v2 05/13]. So > unless the SHA256 bank is disabled by firmware, it should be fine then. When IMA goes into TPM-bypass mode, the measurement list still needs to exist. We cannot depend on the SHA256 TPM bank being enabled, nor SHA256 always being the default IMA file hash algorithm. For this reason sha256 [v2 05/13] needs to be added to the list of "extra" hashes, if the TPM sha256 bank is disabled. Refer to the comments on [02/13] for alternatives. > > If the purpose of this patch set is to actually remove IMA's dependency on a > > working SHA-1, at some point the Kconfig "select CRYPTO_SHA1" needs to be > > removed. Otherwise the kernel will be built with SHA1 builtin > > (CONFIG_CRYPTO_SHA1=y). > > I should have described it better. In the first step at least, the goal > is to remove the runtime dependency only. Because when SHA1's > ->fips_allowed in crypto/testmgr.c gets flipped to false, SHA1 > instantiation would fail with -ENOENT if the kernel was booted with a > fips=1 on its command line. Users not interested in FIPS, i.e. the vast > majority, might still want to use SHA1 and there's no real reason not > to. Thank you for the explanation. > > But yes, it would definitely make sense to drop the CRYPTO_SHA1 dep, at > least at some point. Perhaps by simply moving it to the new > IMA_COMPAT_FALLBACK_TPM_EXTEND. I would personally not do that now > though, just in case there'll be some unexpected fallout from this > series. As long as there is a way of testing the changes, I'm fine with not removing the select. Mimi
diff --git a/security/integrity/ima/ima_fs.c b/security/integrity/ima/ima_fs.c index e4a79a9b2d58..a8df2fe5f4cb 100644 --- a/security/integrity/ima/ima_fs.c +++ b/security/integrity/ima/ima_fs.c @@ -454,6 +454,9 @@ static int __init create_securityfs_measurement_lists(void) return -ENOMEM; for (i = 0; i < securityfs_measurement_list_count; i++) { + if (!ima_algo_array[i].tfm) + continue; + algo = ima_algo_array[i].algo; sprintf(file_name, "ascii_runtime_measurements_%s", @@ -573,20 +576,26 @@ int __init ima_fs_init(void) if (ret != 0) goto out; - binary_runtime_measurements = - securityfs_create_symlink("binary_runtime_measurements", ima_dir, - "binary_runtime_measurements_sha1", NULL); - if (IS_ERR(binary_runtime_measurements)) { - ret = PTR_ERR(binary_runtime_measurements); - goto out; - } + if (ima_algo_array[ima_sha1_idx].tfm) { + binary_runtime_measurements = + securityfs_create_symlink("binary_runtime_measurements", + ima_dir, + "binary_runtime_measurements_sha1", + NULL); + if (IS_ERR(binary_runtime_measurements)) { + ret = PTR_ERR(binary_runtime_measurements); + goto out; + } - ascii_runtime_measurements = - securityfs_create_symlink("ascii_runtime_measurements", ima_dir, - "ascii_runtime_measurements_sha1", NULL); - if (IS_ERR(ascii_runtime_measurements)) { - ret = PTR_ERR(ascii_runtime_measurements); - goto out; + ascii_runtime_measurements = + securityfs_create_symlink("ascii_runtime_measurements", + ima_dir, + "ascii_runtime_measurements_sha1", + NULL); + if (IS_ERR(ascii_runtime_measurements)) { + ret = PTR_ERR(ascii_runtime_measurements); + goto out; + } } runtime_measurements_count =
IMA creates one runtime_measurements_<hash-algo> sysfs file for every TPM bank + for SHA1 if not covered by any such. These differ only in that the template hash value for each record is of the file's associated algorithm each. The kernel does not necessarily support each hash algorithm associated with some TPM bank though -- the most common case probably being that the algorithm is not built-in, but provided as a module, if at all, and thus not available at IMA init time yet. If that happens to be the case, the behavior is a bit counter-intuitive: probably for historic reasons and to still extend the TPM bank with something, a record's template hash is filled with the padded SHA1 value. That is, it is perfectly possible that runtime_measurements_sha256 contains padded SHA1 template hashes if SHA-256 was unavailable at IMA init. I would argue that it's likely that no existing userspace tool is relying on this fallback logic -- they either wouldn't consume the hash value from the measurement list directly but recreate it by themselves, as is required for verification against PCRs, or, if they did, they would somehow assume a hash algorithm and expect the hashes in the measurement list to be of that type. If of the latter kind, this could even lead to hard to debug verification failures. For example, from looking at keylime's current code, the verifier logic seems to assume that the template hashes found in the provided measurement list are of the configured 'ima_log_hash_alg' type. In particular, it does not check against padded SHA1 upon mismatch. That being said, there's also another dimension: currently IMA has a hard requirement on SHA-1 and subsequent patches in this series will attempt to get rid of that. If SHA-1 is not available at IMA init though, it would also mean that padded SHA-1 values cannot get filled in as a fallback for other unsupported algorithms. Substituting something like hard coded all-zeroes or all-ones would be dangerous, because some application or user scripts could perhaps (ab)use the template hashes from the exported measurement lists for some kind of fingerprinting scheme or so. In conclusion, I think it's best to not create the runtime_measurements_<hash-algo> sysfs files for hash algorithms not supported by the kernel. That way, applications expecting a certain hash algorithm for the measurement list and which are not able to handle the padded-SHA1 fallback scheme would fail with a clear indication on what the problem is. Furthermore, as digests for unsupported banks are not getting exposed to userspace anymore, we'll have all flexibility to set it to any value internally, including all-ones as will be needed in a subsequent patch when addressing PCR extend for unsupported banks. So, do not create runtime_measurements_<hash-algo> sysfs files for unsupported hash algorithms. Likewise for their ascii counterparts. Note that at this point, SHA-1 is still mandatory, and thus, runtime_measurements_sha1 as well as the "runtime_measurements" will remain there, even though the code has provisions already to skip their creation as well in case SHA-1 was unavailable. Signed-off-by: Nicolai Stange <nstange@suse.de> --- security/integrity/ima/ima_fs.c | 35 +++++++++++++++++++++------------ 1 file changed, 22 insertions(+), 13 deletions(-)