Passes toolstash -cmp.

Change-Id: I38eb507de2e9dc2cf01822e420bf31a91fb1b720
Reviewed-on: https://go-review.googlesource.com/20766
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

The parser code was not reachable ever since some of the lexer cleanups.
We could recognize '~' in the lexer, complain, and return a '^' instead,
but it's been a few years since Go was new and this may have been a use-
ful error. The lexer complains with "illegal character U+007E '~'" which
is good enough.

For #13244.

Change-Id: Ie3283738486eb6f8462d594f2728ac98333c0520
Reviewed-on: https://go-review.googlesource.com/20768Reviewed-by: Matthew Dempsky <mdempsky@google.com>

Shrinks cmd/go by 30KB.

Change-Id: Ied31192e85af76ebac743f8cc12bd9ef6ec5048f
Reviewed-on: https://go-review.googlesource.com/20765Reviewed-by: Matthew Dempsky <mdempsky@google.com>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>

Change-Id: I0516d49ee8381c5e022d77c2fb41515c01c8a631
Reviewed-on: https://go-review.googlesource.com/20764Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: David Crawshaw <crawshaw@golang.org>

Use a local variable instead.

Passes toolstash -cmp.

Change-Id: I9623a40ff0d568f11afd1279b6aaa1c33eda644c
Reviewed-on: https://go-review.googlesource.com/20730Reviewed-by: Ian Lance Taylor <iant@golang.org>

Instead, use a slice.

Passes toolstash -cmp.

Change-Id: I889fdb4ae997416f907522f549b96506be13bec7
Reviewed-on: https://go-review.googlesource.com/20699Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

It is unused.

Passes toolstash -cmp.

Change-Id: I22ae2bb432ce6be377dea43cf018ffccb6e95f37
Reviewed-on: https://go-review.googlesource.com/20698Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

Currently we shrink stacks during STW mark termination because it used
to be unsafe to shrink them concurrently. For some programs, this
significantly increases pause time: stack shrinking costs ~5ms/MB
copied plus 2µs/shrink.

Now that we've made it safe to shrink a stack without the world being
stopped, shrink them during the concurrent mark phase.

This reduces the STW time in the program from issue #12967 by an order
of magnitude and brings it from over the 10ms goal to well under:

name           old 95%ile-markTerm-time  new 95%ile-markTerm-time  delta
Stackshrink-4               23.8ms ±60%               1.80ms ±39%  -92.44%  (p=0.008 n=5+5)

Fixes #12967.

This slows down the go1 and garbage benchmarks overall by < 0.5%.

name              old time/op  new time/op  delta
XBenchGarbage-12  2.48ms ± 1%  2.49ms ± 1%  +0.45%  (p=0.005 n=25+21)

name                      old time/op    new time/op    delta
BinaryTree17-12              2.93s ± 2%     2.97s ± 2%  +1.34%  (p=0.002 n=19+20)
Fannkuch11-12                2.51s ± 1%     2.59s ± 0%  +3.09%  (p=0.000 n=18+18)
FmtFprintfEmpty-12          51.1ns ± 2%    51.5ns ± 1%    ~     (p=0.280 n=20+17)
FmtFprintfString-12          175ns ± 1%     169ns ± 1%  -3.01%  (p=0.000 n=20+20)
FmtFprintfInt-12             160ns ± 1%     160ns ± 0%  +0.53%  (p=0.000 n=20+20)
FmtFprintfIntInt-12          265ns ± 0%     266ns ± 1%  +0.59%  (p=0.000 n=20+20)
FmtFprintfPrefixedInt-12     237ns ± 1%     238ns ± 1%  +0.44%  (p=0.000 n=20+20)
FmtFprintfFloat-12           326ns ± 1%     341ns ± 1%  +4.55%  (p=0.000 n=20+19)
FmtManyArgs-12              1.01µs ± 0%    1.02µs ± 0%  +0.43%  (p=0.000 n=20+19)
GobDecode-12                8.41ms ± 1%    8.30ms ± 2%  -1.22%  (p=0.000 n=20+19)
GobEncode-12                6.66ms ± 1%    6.68ms ± 0%  +0.30%  (p=0.000 n=18+19)
Gzip-12                      322ms ± 1%     322ms ± 1%    ~     (p=1.000 n=20+20)
Gunzip-12                   42.8ms ± 0%    42.9ms ± 0%    ~     (p=0.174 n=20+20)
HTTPClientServer-12         69.7µs ± 1%    70.6µs ± 1%  +1.20%  (p=0.000 n=20+20)
JSONEncode-12               16.8ms ± 0%    16.8ms ± 1%    ~     (p=0.154 n=19+19)
JSONDecode-12               65.1ms ± 0%    65.3ms ± 1%  +0.34%  (p=0.003 n=20+20)
Mandelbrot200-12            3.93ms ± 0%    3.92ms ± 0%    ~     (p=0.396 n=19+20)
GoParse-12                  3.66ms ± 1%    3.65ms ± 1%    ~     (p=0.117 n=16+18)
RegexpMatchEasy0_32-12      85.0ns ± 2%    85.5ns ± 2%    ~     (p=0.143 n=20+20)
RegexpMatchEasy0_1K-12       267ns ± 1%     267ns ± 1%    ~     (p=0.867 n=20+17)
RegexpMatchEasy1_32-12      83.3ns ± 2%    83.8ns ± 1%    ~     (p=0.068 n=20+20)
RegexpMatchEasy1_1K-12       432ns ± 1%     432ns ± 1%    ~     (p=0.804 n=20+19)
RegexpMatchMedium_32-12      133ns ± 0%     133ns ± 0%    ~     (p=1.000 n=20+20)
RegexpMatchMedium_1K-12     40.3µs ± 1%    40.4µs ± 1%    ~     (p=0.319 n=20+19)
RegexpMatchHard_32-12       2.10µs ± 1%    2.10µs ± 1%    ~     (p=0.723 n=20+18)
RegexpMatchHard_1K-12       63.0µs ± 0%    63.0µs ± 0%    ~     (p=0.158 n=19+17)
Revcomp-12                   461ms ± 1%     476ms ± 8%  +3.29%  (p=0.002 n=20+20)
Template-12                 80.1ms ± 1%    79.3ms ± 1%  -1.00%  (p=0.000 n=20+20)
TimeParse-12                 360ns ± 0%     360ns ± 0%    ~     (p=0.802 n=18+19)
TimeFormat-12                374ns ± 1%     372ns ± 0%  -0.77%  (p=0.000 n=20+19)
[Geo mean]                  61.8µs         62.0µs       +0.40%

Change-Id: Ib60cd46b7a4987e07670eb271d22f6cee5802842
Reviewed-on: https://go-review.googlesource.com/20044Reviewed-by: Keith Randall <khr@golang.org>

We're about to add another root marking job that needs to happen only
during the first markroot pass (whether that's concurrent or STW),
just like finalizer scanning. Rather than introducing another flag
that has the same value as finalizersDone, just rename finalizersDone
to markrootDone.

Change-Id: I535356c6ea1f3734cb5b6add264cb7bf48de95e8
Reviewed-on: https://go-review.googlesource.com/20043Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

Currently shinkstack is only safe during STW because it adjusts
channel-related stack pointers and moves send/receive stack slots
without synchronizing with the channel code. Make it safe to use when
the world isn't stopped by:

1) Locking all channels the G is blocked on while adjusting the sudogs
   and copying the area of the stack that may contain send/receive
   slots.

2) For any stack frames that may contain send/receive slot, using an
   atomic CAS to adjust pointers to prevent races between adjusting a
   pointer in a receive slot and a concurrent send writing to that
   receive slot.

In principle, the synchronization could be finer-grained. For example,
we considered synchronizing around the sudogs, which would allow
channel operations involving other Gs to continue if the G being
shrunk was far enough down the send/receive queue. However, using the
channel lock means no additional locks are necessary in the channel
code. Furthermore, the stack shrinking code holds the channel lock for
a very short time (much less than the time required to shrink the
stack).

This does not yet make stack shrinking concurrent; it merely makes
doing so safe.

This has negligible effect on the go1 and garbage benchmarks.

For #12967.

Change-Id: Ia49df3a8a7be4b36e365aac4155a2416b94b988c
Reviewed-on: https://go-review.googlesource.com/20042Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Austin Clements <austin@google.com>

Currently, locking a G's stack by setting its status to _Gcopystack or
_Gscan is unordered with respect to channel locks. However, when we
make stack shrinking concurrent, stack shrinking will need to lock the
G and then acquire channel locks, which imposes an order on these.

Document this lock ordering and fix closechan to respect it.
Everything else already happens to respect it.

For #12967.

Change-Id: I4dd02675efffb3e7daa5285cf75bf24f987d90d4
Reviewed-on: https://go-review.googlesource.com/20041Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

Currently sudog.elem is never accessed concurrently, so in several
cases we drop the channel lock just before reading/writing the
sent/received value from/to sudog.elem. However, concurrent stack
shrinking is going to have to adjust sudog.elem to point to the new
stack, which means it needs a way to synchronize with accesses to
sudog.elem. Hence, add sudog.elem to the fields protected by
hchan.lock and scoot the unlocks down past the uses of sudog.elem.

While we're here, better document the channel synchronization rules.

For #12967.

Change-Id: I3ad0ca71f0a74b0716c261aef21b2f7f13f74917
Reviewed-on: https://go-review.googlesource.com/20040Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

With concurrent stack shrinking, the stack can move the instant after
a G enters _Gwaiting. There are only two places that put a G into
_Gwaiting: gopark and newstack. We fixed uses of gopark. This commit
fixes newstack by simplifying its G transitions and, in particular,
eliminating or narrowing the transient _Gwaiting states it passes
through so it's clear nothing in the G is accessed while in _Gwaiting.

For #12967.

Change-Id: I2440ead411d2bc61beb1e2ab020ebe3cb3481af9
Reviewed-on: https://go-review.googlesource.com/20039Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

gopark calls the unlock function after setting the G to _Gwaiting.
This means it's generally unsafe to access the G's stack from the
unlock function because the G may start running on another P. Once we
start shrinking stacks concurrently, a stack shrink could also move
the stack the moment after it enters _Gwaiting and before the unlock
function is called.

Document this restriction and fix the two places where we currently
violate it.

This is unlikely to be a problem in practice for these two places
right now, but they're already skating on thin ice. For example, the
following sequence could in principle cause corruption, deadlock, or a
panic in the select code:

On M1/P1:
1. G1 selects on channels A and B.
2. selectgoImpl calls gopark.
3. gopark puts G1 in _Gwaiting.
4. gopark calls selparkcommit.
5. selparkcommit releases the lock on channel A.

On M2/P2:
6. G2 sends to channel A.
7. The send puts G1 in _Grunnable and puts it on P2's run queue.
8. The scheduler runs, selects G1, puts it in _Grunning, and resumes G1.
9. On G1, the sellock immediately following the gopark gets called.
10. sellock grows and moves the stack.

On M1/P1:
11. selparkcommit continues to scan the lock order for the next
channel to unlock, but it's now reading from a freed (and possibly
reused) stack.

This shouldn't happen in practice because step 10 isn't the first call
to sellock, so the stack should already be big enough. However, once
we start shrinking stacks concurrently, this reasoning won't work any
more.

For #12967.

Change-Id: I3660c5be37e5be9f87433cb8141bdfdf37fadc4c
Reviewed-on: https://go-review.googlesource.com/20038Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

Currently the g.waiting list created by a select is in poll order.
However, nothing depends on this, and we're going to need access to
the channel lock order in other places shortly, so modify select to
put the waiting list in channel lock order.

For #12967.

Change-Id: If0d38816216ecbb37a36624d9b25dd96e0a775ec
Reviewed-on: https://go-review.googlesource.com/20037Reviewed-by: Rick Hudson <rlh@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Austin Clements <austin@google.com>

Currently the select lock order is a []*hchan. We're going to need to
refer to things other than the channel itself in lock order shortly,
so switch this to a []uint16 of indexes into the select cases. This
parallels the existing representation for the poll order.

Change-Id: I89262223fe20b4ddf5321592655ba9eac489cda1
Reviewed-on: https://go-review.googlesource.com/20036Reviewed-by: Rick Hudson <rlh@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

Given a G, there's currently no way to find the channel it's blocking
on. We'll need this information to fix a (probably theoretical) bug in
select and to implement concurrent stack shrinking, so record the
channel in the sudog.

For #12967.

Change-Id: If8fb63a140f1d07175818824d08c0ebeec2bdf66
Reviewed-on: https://go-review.googlesource.com/20035Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

gcMarkRootCheck is too expensive to do during mark termination.
However, since it's a useful check and it complements checkmark mode
nicely, enable it during mark termination is checkmark is enabled.

Change-Id: Icd9039e85e6e9d22747454441b50f1cdd1412202
Reviewed-on: https://go-review.googlesource.com/20663Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

I was wondering why cmd/go includes the HTTP server implementations.

Dumping the linker's deadcode dependency graph into a file and doing
some graph analysis, I found that the only reason cmd/go included an
HTTP server was because the maxBytesReader type (used by both the HTTP
transport & HTTP server) did a static type assertion to an HTTP server
type.

Changing it to a interface type assertion reduces the size of cmd/go
by 533KB (5.2%)

On linux/amd64, cmd/go goes from 10549200 to 10002624 bytes.

Add a test too so this doesn't regress. The test uses cmd/go as the
binary to test (a binary which needs the HTTP client but not the HTTP
server), but this change and test are equally applicable to any such
program.

Change-Id: I93865f43ec03b06d09241fbd9ea381817c2909c5
Reviewed-on: https://go-review.googlesource.com/20763Reviewed-by: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: David Crawshaw <crawshaw@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

Step 2 of stream-lining parameter parsing

- do parameter validity checks in parser
- two passes instead of multiple (and theoretically quadratic) passes
  when checking parameters
- removes the need for OKEY and some ONONAME nodes in those passes

This removes allocation of ~123K OKEY (incl. some ONONAME) nodes
out of a total of ~10M allocated nodes when running make.bash, or
a reduction of the number of alloacted nodes by ~1.2%.

Change-Id: I4a8ec578d0ee2a7b99892ac6b92e56f8e0415f03
Reviewed-on: https://go-review.googlesource.com/20748Reviewed-by: Matthew Dempsky <mdempsky@google.com>
Run-TryBot: Robert Griesemer <gri@golang.org>

Change-Id: I049d2596fe8ce0e93391599f5c224779fd8e316f
Reviewed-on: https://go-review.googlesource.com/20761Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>

Step 1 of streamlining parameter parsing.

Change-Id: If9fd38295ccc08aafc7f1d26188d0926dd73058b
Reviewed-on: https://go-review.googlesource.com/20747Reviewed-by: Matthew Dempsky <mdempsky@google.com>
Run-TryBot: Robert Griesemer <gri@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

Fixes #14825

Change-Id: Ib44d80579a55c15d75ea2ad1ef54efa6ca66a9a6
Reviewed-on: https://go-review.googlesource.com/20745
Run-TryBot: Todd Neal <todd@tneal.org>
Reviewed-by: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

Use The fmt internal buffer for character formatting instead of
the pp Printer rune decoding buffer.

Uses an uint64 instead of int64 argument to fmt_c and fmt_qc for easier
range checks since no valid runes are represented by negative numbers or
are above 0x10ffff.

Add range checks to fmt_c and fmt_qc to guarantee that a RuneError
character is returned by the functions for any invalid code point
in range uint64. For invalid code points in range utf8.MaxRune
the used utf8 and strconv functions already return a RuneError.

Change-Id: I9772f804dfcd79c3826fa7f6c5ebfbf4b5304a51
Reviewed-on: https://go-review.googlesource.com/20373
Run-TryBot: Rob Pike <r@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rob Pike <r@golang.org>

Remove check for %p and %T in printValue.
These verbs are not recursive and are handled already in
printArg which is called on any argument before printValue.

Format the type string for %T directly instead of invoking
the more complex printArg with %s on the type string.

Decouple the %T tests from variables declared in scan_test.go.

Change-Id: Ibd51566bd4cc1a260ce6d052f36382ed05020b48
Reviewed-on: https://go-review.googlesource.com/20622
Run-TryBot: Rob Pike <r@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rob Pike <r@golang.org>

Fixes #14664

Change-Id: I8bda2435857772f590859808904c48d768b87d46
Reviewed-on: https://go-review.googlesource.com/20254
Run-TryBot: Rob Pike <r@golang.org>
Reviewed-by: Rob Pike <r@golang.org>

The Join test was doing something remarkable and unnecessary instead of
just using ... on a slice. Maybe it was an editing relic.

Fix it by deleting the monstrosity.

Change-Id: I5b90c6d539d334a9c27e57d26dacd831721cfcfe
Reviewed-on: https://go-review.googlesource.com/20727Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

Do a reset of the fmt flags before printing the extra argument
error message to prevent a malformed printing of extra arguments.

Regroup tests for extra argument error strings.

Change-Id: Ifd97f5ca36f6c97ed5a380d975cf154d17997d3f
Reviewed-on: https://go-review.googlesource.com/20571
Run-TryBot: Rob Pike <r@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rob Pike <r@golang.org>

This change filters out destination addresses by address family when
source address is specified to avoid running Dial operation with wrong
addressing scopes.

Fixes #11837.

Change-Id: I10b7a1fa325add2cd8ed58f105d527700a10d342
Reviewed-on: https://go-review.googlesource.com/20586Reviewed-by: Paul Marks <pmarks@google.com>

On the latest darwin kernels, kevent in runtime-integrated network
poller sometimes reports SYN-SENT state sockets as ESTABLISHED ones,
though it's still unclear what's the root cause.

This change prevents such spurious notifications by additional connect
system calls.

Fixes #14548.

Change-Id: Ie29788e38ca735ca77259befeba3229d6a30ac52
Reviewed-on: https://go-review.googlesource.com/20468
Run-TryBot: Mikio Hara <mikioh.mikioh@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>

This change consolidates functions and methods related to UnixAddr,
UnixConn and UnixListener for maintenance purpose, especially for
documentation.

The followup changes will update comments and examples.

Updates #10624.

Change-Id: I372d152099ac10956284e6b3863d7e4d9fe5c8e9
Reviewed-on: https://go-review.googlesource.com/20125
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>

This change consolidates functions and methods related to IPAddr and
IPConn for maintenance purpose, especially for documentation.

The followup changes will update comments and examples.

Updates #10624.

Change-Id: Ia5146f234225704a3c0b6459e1903e56a7b68134
Reviewed-on: https://go-review.googlesource.com/20124
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>

This change consolidates functions and methods related to UDPAddr and
UDPConn for maintenance purpose, especially for documentation.

The followup changes will update comments and examples.

Updates #10624.

Change-Id: Idfe9be8ea46ade1111b0ae176862b2048eafc7be
Reviewed-on: https://go-review.googlesource.com/20120
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>

Change-Id: I5dbcdf0ee0b46b760b2a7decb1d937aac2a6fa8d
Reviewed-on: https://go-review.googlesource.com/20585
Run-TryBot: Mikio Hara <mikioh.mikioh@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>

Change-Id: I01c000ff3f6dc6b0ed691e289eeef0fa61500337
Reviewed-on: https://go-review.googlesource.com/20744Reviewed-by: Keith Randall <khr@golang.org>

Use constants instead of dynamically computed values to determine
the bit sizes of types similar to how strconv and other packages
directly compute these sizes. Move these constants near the code
that uses them.

Change-Id: I78d113b7e697466097e32653975df5990380c2c1
Reviewed-on: https://go-review.googlesource.com/20514
Run-TryBot: Rob Pike <r@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rob Pike <r@golang.org>

This change also refactors SplitHostPort to avoid using gotos and
naked returns.

Fixes #14827

Change-Id: I4dca528936757fd06da76c23af8a0f6175bbedd1
Reviewed-on: https://go-review.googlesource.com/20726Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>

Change-Id: Ieefeceea40bd29657fd519368b0920dad8443844
Reviewed-on: https://go-review.googlesource.com/20712
Run-TryBot: Todd Neal <todd@tneal.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>

They've been on for a few weeks of general use and nothing
has tripped up on them yet.

Makes the compiler ~18% faster.

Change-Id: I42d7bbc0581597f9cf4fb28989847814c81b08a2
Reviewed-on: https://go-review.googlesource.com/20741Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>

The new link is https://golang.org/s/go15heapdump.

Change-Id: Ifcaf8572bfe815ffaa78442a1991f6e20e990a50
Reviewed-on: https://go-review.googlesource.com/20740Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>