crypto/x509: Mac without CGO is very slow, times out with large numbers of certs #38215
Comments
|
Is this a duplicate? In this case, it isn't just about optimization, but about real issues that arise because of that. |
|
/cc @FiloSottile for triage / if it's a duplicate |
|
I actually am unsure what the logic is doing and why it is necessary. As far as I can tell, it does the following:
The part that surprise me is:
In the end, we just return the system certs and all of those from the two |
|
Yeah, the nocgo fallback is absolutely not great. The policy logic is extremely complex and we've gotten it wrong a few times before. The APIs don't help us. The two options here are:
I'd rather keep only one copy of this logic, so I want to try (2) before (1). So yeah, duplicate of #32604, but I appreciate the detailed user impact report, it helps with prioritization. |
I agree. Given the material impact it can have on end-users, and it doesn't look like option 2 is happening too soon (I am very happy to be wrong), can we consider some optimization to the nocgo fallback to carry us through? Also, is it worth documenting somewhere reasonably prominent, maybe in
No problem. :-) |
Let me see what I can do =) It really is a bad and inconsistent experience, sorry about that. |
|
I can try and give a hand. I am not sure if I captured the logic and options correctly here |
|
We'd basically drop the nocgo code and port the cgo one to hardcoded references to Security.framework. The hard part is in what the linker needs to do for that, so I reached out to the cmd/link owners. I'll keep #32604 up to date! |
|
Thanks. I subscribed to that one. If it really is hard, perhaps we can just do what you suggested and parse the xml, buy is some breathing room? |
|
We'll see. I would really regret depending on encoding/xml in crypto/x509 and rewriting all that logic. |
|
Change https://golang.org/cl/227037 mentions this issue: |
|
FWIW, I ran the gotip version that @FiloSottile put up in the link. It passed all tests, and the time is effectively identical to cgo. Also had someone whose cert store is quite large and it took up to 10s before with nocgo. With this patch, 1.3s. Much better! |
+----------------------------------------------------------------------+ | Hello, if you are reading this and run macOS, please test this code: | | | | $ GO111MODULE=on go get golang.org/dl/gotip@latest | | $ gotip download | | $ GODEBUG=x509roots=1 gotip test crypto/x509 -v -run TestSystemRoots | +----------------------------------------------------------------------+ We currently have two code paths to extract system roots on macOS: one uses cgo to invoke a maze of Security.framework APIs; the other is a horrible fallback that runs "/usr/bin/security verify-cert" on every root that has custom policies to check if it's trusted for SSL. The fallback is not only terrifying because it shells out to a binary, but also because it lets in certificates that are not trusted roots but are signed by trusted roots, and because it applies some filters (EKUs and expiration) only to roots with custom policies, as the others are not passed to verify-cert. The other code path, of course, requires cgo, so can't be used when cross-compiling and involves a large ball of C. It's all a mess, and it broke oh-so-many times (#14514, #16532, #19436, #20990, #21416, #24437, #24652, #25649, #26073, #27958, #28025, #28092, #29497, #30471, #30672, #30763, #30889, #32891, #38215, #38365, ...). Since macOS does not have a stable syscall ABI, we already dynamically link and invoke libSystem.dylib regardless of cgo availability (#17490). How that works is that functions in package syscall (like syscall.Open) take the address of assembly trampolines (like libc_open_trampoline) that jump to symbols imported with cgo_import_dynamic (like libc_open), and pass them along with arguments to syscall.syscall (which is implemented as runtime.syscall_syscall). syscall_syscall informs the scheduler and profiler, and then uses asmcgocall to switch to a system stack and invoke runtime.syscall. The latter is an assembly trampoline that unpacks the Go ABI arguments passed to syscall.syscall, finally calls the remote function, and puts the return value on the Go stack. (This last bit is the part that cgo compiles from a C wrapper.) We can do something similar to link and invoke Security.framework! The one difference is that runtime.syscall and friends check errors based on the errno convention, which Security doesn't follow, so I added runtime.syscallNoErr which just skips interpreting the return value. We only need a variant with six arguments because the calling convention is register-based, and extra arguments simply zero out some registers. That's plumbed through as crypto/x509/internal/macOS.syscall. The rest of that package is a set of wrappers for Security.framework and Core Foundation functions, like syscall is for libSystem. In theory, as long as macOS respects ABI backwards compatibility (a.k.a. as long as binaries built for a previous OS version keep running) this should be stable, as the final result is not different from what a C compiler would make. (One exception might be dictionary key strings, which we make our own copy of instead of using the dynamic symbol. If they change the value of those strings things might break. But why would they.) Finally, I rewrote the crypto/x509 cgo logic in Go using those wrappers. It works! I tried to make it match 1:1 the old logic, so that root_darwin_amd64.go can be reviewed by comparing it to root_cgo_darwin_amd64.go. The only difference is that we do proper error handling now, and assume that if there is no error the return values are there, while before we'd just check for nil pointers and move on. I kept the cgo logic to help with review and testing, but we should delete it once we are confident the new code works. The nocgo logic is gone and we shall never speak of it again. Fixes #32604 Fixes #19561 Fixes #38365 Awakens Cthulhu Change-Id: Id850962bad667f71e3af594bdfebbbb1edfbcbb4 Reviewed-on: https://go-review.googlesource.com/c/go/+/227037 Reviewed-by: Katie Hockman <katie@golang.org>
What version of Go are you using (
go version)?Does this issue reproduce with the latest release?
Yes
What operating system and processor architecture are you using (
go env)?go envOutputWhat did you do?
CGO_ENABLED=0Even without a timeout, the initial response can take many seconds. The default mac has ~22 certificates and it can take 3-4 seconds. Any corporate system with more certs can take many multiples.
What did you expect to see?
Reasonable return times
What did you see instead?
Long waits and timeouts.
Lots More Detail
I spent quite some time tracking it down. When compiling with
CGO_ENABLED=1, which is fine, although may require more tools installed and, critically, makes cross-compiling and some CI nearly impossible, the lookup time is roughly well below O(n). When compiling withCGO_ENABLED=0, which is desired when possible, every cert in the system increases the time to verify the server cert.The core code issue is in root_darwin.go, and specifically execSecurityRoots().
It does the following (simplified):
/usr/bin/securitywith appropriate args/usr/bin/securityon it with appropriate argsThe above is, essentially, O(n), and rises linearly with the number of certs. The comments here indicate that it should be optimized to reasonably close to with
CGO, but I haven't found that to be the case. I have had several runs where 160 certs and a timeout of 5 seconds would, well, time out.I created a repo with a program which recreates it. It also includes directions and output from my mac. In my case, it was 244ms for CGO and 1.7s for no CGO. I worked with someone who had ~160 certs (not unreasonable); it took well over 5 seconds.
repo here
The text was updated successfully, but these errors were encountered: