(from patchset 1, c12c890c64dd6372b3893af1e6f5ab11802c9e81, of
https://go-review.googlesource.com/c/go/+/230025/1, with merges fixes
due to parent commit's differents from its ps1..ps3)
Instead of parsing the PEM files and then storing the *Certificate
values forever, still parse them to see if they're valid and pick out
some fields, but then only store the decoded pem.Block.Bytes until
that cert is first needed.
Saves about 500K of memory on my (Debian stable) machine after doing a
tls.Dial or calling x509.SystemCertPool.
A more aggressive version of this is still possible: we can not keep
the pem.Block.Bytes in memory either, and re-read them from disk when
necessary. But dealing with files disappearing and even large
multi-cert PEM files changing (with offsets sliding around) made this
conservative version attractive. It doesn't change the
slurp-roots-on-startup semantics. It just does so with less memory
retained.
Change-Id: I3aea333f4749ae3b0026042ec3ff7ac015c72204
(from patchset 1, 7cdc3c3e7427c9ef69e19224d6036c09c5ea1723, of
https://go-review.googlesource.com/c/go/+/229917/1)
This will allow building CertPools that consume less memory. (Most
certs are never accessed. Different users/programs access different
ones, but not many.)
This CL only adds the new internal mechanism (and uses it for the
old AddCert) but does not modify any existing root pool behavior.
(That is, the default Unix roots are still all slurped into memory as
of this CL)
Change-Id: Ib3a42e4050627b5e34413c595d8ced839c7bfa14
Snapshotted from Go commit 619c7a48a38b28b521591b490fd14ccb7ea5e821
(https://go-review.googlesource.com/c/go/+/229762,
"crypto/x509: add x509omitbundledroots build tag to not embed roots")
With 975c01342a25899962969833d8b2873dc8856a4f
(https://go-review.googlesource.com/c/go/+/220721) removed, because it
depends on other stuff in Go std that doesn't yet exist in a Go
release.
Also, add a subset fork of Go's internal/testenv, for use by x509's tests.
Makes parsing 4.6x faster.
name old time/op new time/op delta
ParseInt-12 32.1ns ± 1% 6.9ns ± 2% -78.55% (p=0.000 n=10+9)
Signed-off-by: David Anderson <danderson@tailscale.com>
When unregistering a replaced client connection, move the
still-connected peers to the current client connecition. Inform
the peers that we've gone only when unregistering the active
client connection.
Signed-off-by: Dmitry Adamushko <da@stablebits.net>
This was only done occasionally, but was extremely disruptive
when done and is no longer necessary.
It used to be that when switching links, we had to immediately
generate handshakes to everyone we were communicating with to
punch a hole in any NAT we were talking through. (This ended up
not really working, because in the process we got rid of our
session keys and ended up having a futile conversation for many
seconds.)
Now we have DERP, our link change propogates to the other side
as a new list of endpoints, so they start spraying packets.
We will definitely get one thanks to DERP, which will cause us
to spray, opening any NAT we are behind.
The result is that for good connections, we don't trash session
keys and cause an interruption.
Signed-off-by: David Crawshaw <crawshaw@tailscale.com>
It was one of the top garbage producers on my phone.
It's slated to be deleted and replaced anyway, but this helps in the
meantime.
The go.sum changes look scary, but the new dep only adds 240 bytes to
the binary. The go.sum noise is just cmd/go being aggressive in
including a lot of stuff (which is being fixed in Go 1.15, for what I
understand). And I ran a go mod tidy, which added some too. (I had to
write a custom wrapper around go mod tidy because this mod tidy
normally breaks on tailscale.io/control being missing but referenced
in tests)
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
