* Export InvalidPCRs field in ReplayError
In order to retrieve the Invalid PCRs which couldn't be replayed against the Event log, we need this field to be exported as this gives the exact and true information. Replay error events will give all the events, but doesn't give the exact PCR index which doesn't get replayed.
Following is the test to extend PCR 7 and verify the PCRs 7,8,9 against the Event log. Output:
```
event log failed to verify: the following registers failed to replay: [7]
ReplayError Events:=[107]
Replay Error Events PCR indexes=[0 7 2 3 6 9 8 1 4 5]
```
* Add Comment to the exported field
Systems with TXT enabled may issue the TPM2_Startup() command from a
locality other than 0. In this case, the initial value of PCR0 will
represent the locality that the call was made from. This is exposed to
higher layers by an EV_NO_ACTION event that has data containing the
NULL-terminated string "StartupLocality" followed by a single byte
representing the state of the locality. As this event is EV_NO_ACTION,
it does not represent an extension in itself.
So:
1) Ignore events that are EV_NO_ACTION when replaying the log, except:
2) For PCR0, if an event is EV_NO_ACTION and contains the string
"StartupLocality", use the final byte of the event data as the initial
value of PCR0 for the replay.
Using the length of a digest to infer the hash algorithm is somewhat
fragile - if we end up with multiple hash algorithms that share the same
digest length, things will break. Instead, pass more complete digest
information through to relevant functions and figure things out by
mapping the TPM hash algorithm to the appropriate Golang type.
Crypto agile logs may contain digest types that we don't currently
handle. However, we still need to know how long each digest is in order
to read over the appropriate amount of the buffer. This information is
provided to us as part of the spec header - make use of it rather than
hardcoding the set of digests and lengths we know about.
@brandonweeks detected another case of the "make([]T, untrustedValue)"
pattern, which would allow an attacker to cause the parser to allocate
an unbounded amount of memory.
Fix this by reading one algorithm at a time instead of pre-allocating a
slice of algorithms.
A go-fuzz target for the ParseEventLog function. It has been tested
with go-fuzz and go-fuzz + libFuzzer.
oss-fuzz requires a statically built fuzzer binary, so `gofuzz` build
tags are added to avoid building files that depend on go-tspi. A mock
tpm_other.go file is also included to satisfy the `platformTPM`
interface.
It's best practice to define as much code, especially exported API, in
files that can build on any platform. With as little code as possible in
OS specific files.
Ensure files with build tags don't contain any exported APIs. This helps
us not accidentally define API that only works on one platform, or have
incompatible method defintions between OSes.
TODO: follow up with an "unsupported" implementation so this builds on
Mac or without CGO (e.g. for servers)?
Ensure an attacker can't alter the value we interpret by appending an entry of
the same type to the eventlog. Don't worry about events that come before the
EV_SEPARATOR for now.
This is being prototyped in an internal package as we start to open
source. This code will either live in attest, or in a separate eventlog
package in the future.
