On January 8, 2025 2:19 AM, Patrick Steinhardt wrote: >On Tue, Jan 07, 2025 at 06:54:02PM -0500, rsbecker@xxxxxxxxxxxxx wrote: >> On January 7, 2025 6:22 PM, brian m. carlson wrote: >> >On 2025-01-07 at 15:26:58, Patrick Steinhardt wrote: >> >> Hi, >> >> >> >> this small patch series fixes the issue reported by Randall [1], >> >> where an exhausted entropy pool can cause us to die when writing a >> >> new table to the reftable stack. I _think_ that this is only an >> >> issue with the OpenSSL backend of `csprng_bytes()`: >> >> >> >> - `arc4random_buf()` never returns an error. >> >> >> >> - `getrandom()` pulls from "/dev/urandom" by default. >> >> >> >> - `getentropy()` seems to block when there is not enough randomness >> >> available. >> >> >> >> - `GtlGenRandom()` I cannot really tell. >> >> >> >> - The fallback reads from "/dev/urandom", which also returns bytes in >> >> case the entropy pool is drained. >> >> >> >> So OpenSSL's `RAND_bytes()` seems to be the only one that returns >> >> an error when the entropy pool is empty. I did wonder whether we >> >> even need to introduce the new flag in the first place, or whether >> >> we cannot just use `RAND_pseudo_bytes()` unconditionally. But I'm a >> >> bit uneasy about it given that OpenSSL has this doc: >> >> >> >> RAND_pseudo_bytes() puts num pseudo-random bytes into buf. >> >> Pseudo-random byte sequences generated by RAND_pseudo_bytes() will >> >> be unique if they are of sufficient length, but are not necessarily >> >> unpredictable. They can be used for non-cryptographic purposes and >> >> for certain purposes in cryptographic protocols, but usually not for >> >> key generation etc. >> >> >> >> It might be too easy to accidentally rely on `csprng_bytes()` where >> >> it actually requires strong cryptographic data, so I was erring on >> >> the side of caution. >> > >> >The reason I didn't use RAND_pseudo_bytes is because it's been >> >deprecated since OpenSSL 1.1.0 and RAND_bytes uses a CSPRNG just like >> >RAND_pseudo_bytes as of that version. Once it's seeded, it should be >> >able to generate plenty of bytes, because I believe it uses a >> >CTR-DRBG, which only needs to be reseeded after 2^48 bytes (which is far more >than we should be using). >> > >> >We can full well use RAND_pseudo_bytes, but all operating systems >> >should provide an appropriate entropy source that can provide 256 bits of >entropy on startup. >> >arc4random will just kill the process if it can't seed itself, so >> >your changes won't actually prevent dying on a lack of entropy. >> > >> >I don't want an option that chooses "insecure" bytes. My preference >> >is that we require people use a different backend or an up-to-date >> >OpenSSL version that shouldn't have this problem. We can use >> >RAND_pseudo_bytes if we really need to support older versions, but >> >there are also no major operating systems which require that old of a >> >version (CentOS 7, which is dead, used OpenSSL 1.0.2, and CentOS 8 uses >1.1.1k), so it's probably not within our support policy to do that. >> > >> >Note also that if OpenSSL is being used for TLS, a lack of entropy >> >will result in TLS not working, which means that Git will be randomly >> >broken on that system, which is not really an experience that we want >> >to encourage, so that should be taken into account. >> > >> >Can we get some more information about what version of OpenSSL is >> >being used and what the system entropy source is? >> >> In my situation, OpenSSL 3.0.11 on ia64. 3.4.1 and 3.0.13 on x86 (x86 >> works fine Because OpenSSL uses hardware. On ia64, we end up on PRNGD, >which does fail. > >You reported in [1] that a couple more tests are indeed failing, not only t0610. That >changes things in my opinion as it shows that this is not a localized issue in the >reftable library, but likely in multiple callsites where we use randomness. So my >current patch series is not sufficient as it only fixes up the reftable codebase. But in >the case where it's a general issue I tend to agree with brian, because I don't want to >play whack-a-mole with all the callsites of `git_rand()` where we can indeed use >insecure bytes. > >Honestly, this rather makes me want to remove the OpenSSL backend for our >CSRNG completely. NonStop is the only platform that uses it right now, and it seems >to be easy to misconfigure. All the other backends we have don't have the same >issue as explained further up in my message. So does NonStop support any of the >alternative backends that Git has, like `arc4random_buf()`, `getrandom()`, >`getentropy()` or reading from "/dev/urandom"? > >Might be I'm coming to conclusions too fast, so if I'm missing obvious usecases then >please stop me :) > >Randall, you mentioned that your platform had a maintenance window right during >the release of v2.48.0-rc2 [2]. You never mentioned issues with randomness before >that maintenance window, and after it you hit them in many tests without any >changes to the CSPRNG between v2.48.0-rc1 and -rc2. Could it be that something >broke on your end? Unfortunately, ia64 is not a great platform for randomness. There are no alternates available. We have a case open on PRNGD, but it is unlikely to be fixed any time soon. The ia64 platform goes off support at the end of 2025, so we will stop building git for that platform when that happens. If there is some stopgap solution we can use, even PRNGD, but warn about reducing randomness load, it might work. For x86, the hardware randomizer used in OpenSSL is fine.
