Hi Roberto, [cc'ing tpmdd-devel]
On Tue, 2017-07-25 at 17:44 +0200, Roberto Sassu wrote: > This patch set applies on top of kernel v4.13-rc2. > > IMA, for each file matching policy rules, calculates a digest, creates > a new entry in the measurement list and extends a TPM PCR with the digest > of entry data. The last step causes a noticeable performance reduction. > > Since systems likely access the same files, repeating the above tasks at > every boot can be avoided by replacing individual measurements of likely > accessed files with only one measurement of their digests: the advantage > is that the system performance significantly improves due to less PCR > extend operations; on the other hand, the information about which files > have exactly been accessed and in which sequence is lost. > > If this new measurement reports only good digests (e.g. those of > files included in a Linux distribution), and if verifiers only check > that a system executed good software and didn't access malicious data, > the disadvantages reported earlier would be acceptable. > > The Trusted Computing paradigm measure & load is still respected by IMA > with the proposed optimization. If a file being accessed is not in a > measured digest list, a measurement will be recorded as before. If it is, > the list has already been measured, and the verifier must assume that > files with digest in the list have been accessed. > > Measuring digest lists gives the following benefits: > > - boot time reduction > For a minimal Linux installation with 1400 measurements, the boot time > decreases from 1 minute 30 seconds to 15 seconds, after loading to IMA > the digest of all files packaged by the distribution (32000). The new > list contains 92 entries. Without IMA, the boot time is 8.5 seconds. Before we "fix" a TPM performance problem in IMA, we need to really understand the performance problem first. We've added a "TPM peformance" topic to the Linux Plumber Conference TPM microconference - http://wiki.linuxplumbersconf.org/2017:tpms. We've benchmarked a couple of different TPMs on different systems with TPMs on LPC, I2C, and STI. Originally we were seeing even worse performance than your 1 minute 30 seconds for 1400 measurements. Fortunately, we were able to bring it down to about 17 seconds for a 1000 TPM extends. Refer to commits a233a0289cf9 "tpm: msleep() delays - replace with usleep_range() in i2c nuvoton driver" and 0afb7118ae02 "tpm: add sleep only for retry in i2c_nuvoton_write_status()" for the details. Hamza Attak posted a similar patch to the tpmdd-devel mailing list replacing msleep() with usleep_range() calls. Unfortunately, we're seeing really poor performance with another TPM for other reasons. Mimi > > - lower network and CPU requirements for remote attestation > With the IMA optimization, both the measurement and digest lists > must be verified for a complete evaluation. However, since the lists > are fixed, they could be sent to and checked by the verifier only once. > Then, during a remote attestation, the only remaining task is to verify > the short measurement list. > > - signature-based remote attestation > Digest list signature can be used as a proof of the provenance for the > files whose digest is in the list. Then, if verifiers trust the signer > and only check provenance, remote attestation verification would simply > consist on checking digest lists signatures and that the measurement > list only contain list metadata digests (reference measurement databases > would be no longer required). An example of a signed digest list, > that can be parsed with this patch set, is the RPM package header. > > Digest lists are loaded in two stages by IMA through the new securityfs > interface called 'digest_lists'. Users supply metadata, for the digest > lists they want to load: path, format, digest, signature and algorithm > of the digest. > > Then, after the metadata digest is added to the measurement list, IMA > reads the digest lists at the path specified and loads the digests in > a hash table (digest lists are not measured, since their digest is already > included in the metadata). With metadata measurement instead of digest list > measurement, it is possible to avoid a performance reduction that would > occur by measuring many digest lists (e.g. RPM headers) individually. > If, alternatively, digest lists are loaded together, their signature > cannot be verified. > > Lastly, when a file is accessed, IMA searches the calculated digest in > the hash table. Only if the digest is not found a new entry is added > to the measurement list. > > > Roberto Sassu (12): > ima: generalize ima_read_policy() > ima: generalize ima_write_policy() > ima: generalize policy file operations > ima: use ima_show_htable_value to show hash table data > ima: add functions to manage digest lists > ima: added parser of digest lists metadata > ima: added parser for compact digest list > ima: added parser for RPM data type > ima: introduce securityfs interfaces for digest lists > ima: disable digest lookup if digest lists are not measured > ima: don't report measurements if digests are included in the loaded > lists > ima: added Documentation/security/IMA-digest-lists.txt > > Documentation/security/IMA-digest-lists.txt | 150 +++++++++++++++++ > include/linux/fs.h | 1 + > security/integrity/ima/Kconfig | 11 ++ > security/integrity/ima/Makefile | 1 + > security/integrity/ima/ima.h | 17 ++ > security/integrity/ima/ima_digest_list.c | 247 > ++++++++++++++++++++++++++++ > security/integrity/ima/ima_fs.c | 178 ++++++++++++-------- > security/integrity/ima/ima_main.c | 23 ++- > security/integrity/ima/ima_policy.c | 1 + > security/integrity/ima/ima_queue.c | 39 +++++ > 10 files changed, 602 insertions(+), 66 deletions(-) > create mode 100644 Documentation/security/IMA-digest-lists.txt > create mode 100644 security/integrity/ima/ima_digest_list.c >
