Brad Fitzpatrick [Sat, 15 Oct 2011 00:34:07 +0000 (17:34 -0700)]
http: DoS protection: cap non-Handler Request.Body reads
Previously, if an http.Handler didn't fully consume a
Request.Body before returning and the request and the response
from the handler indicated no reason to close the connection,
the server would read an unbounded amount of the request's
unread body to advance past the request message to find the
next request's header. That was a potential DoS.
With this CL there's a threshold under which we read
(currently 256KB) in order to keep the connection in
keep-alive mode, but once we hit that, we instead
switch into a "Connection: close" response and don't
read the request body.
Chris Farmiloe [Fri, 14 Oct 2011 21:29:54 +0000 (17:29 -0400)]
xml: match Marshal's XMLName behavior in Unmarshal
When xml.Marshal is called on a struct it will happily
reflect the information in the "tag" of an XMLName member
regardless of the type to give the struct a tag-name in
it's XML form. This is backed up by the documentation which
says:
However xml.Unmarshal *does* care about the XMLName field
being of type xml.Name, and currently returns the error
"field XMLName does not have type xml.Name" if you have it
set to something else.
This is firstly inconsistant with xml.Marshal but it also
makes it impossible to use xml.Marshal alongside other
Marshallers (like json/bson) without poluting the state's
namespace with XMLName fields. Inorder to exclude fields
from other Marshallers the convention has been started to
tag fields as "omitempty"; which will cause the field not
to display if it is at it's "zero" state, XMLName cannot
have such as zero-state since it is a struct, so it is nicer
to use a pointer/bool value for XMLName so it can be easily
excluded when I want to Marshal my struct by some other
wire format.
Attached is the proposed minor change, that simply stops
erring if it can't set the name on the XMLName field, which
is just optional metadata anyway.
Fixes #2265.
Adam Langley [Fri, 14 Oct 2011 19:11:21 +0000 (15:11 -0400)]
crypto/x509: add code for dealing with PKIX public keys.
We also have functions for dealing with PKCS#1 private keys. This
change adds functions for PKIX /public/ keys. Most of the time one
won't be parsing them because they usually come in certificates, but
marshaling them happens and I've previously copied the code from
x509.go for this.
Adam Langley [Fri, 14 Oct 2011 19:06:54 +0000 (15:06 -0400)]
crypto/x509: keep the raw Subject and Issuer.
X509 names, like everything else X509, are ludicrously general. This
change keeps the raw version of the subject and issuer around for
matching. Since certificates use a distinguished encoding, comparing
the encoding is the same as comparing the values directly. This came
up recently when parsing the NSS built-in certificates which use the
raw subject and issuer for matching trust records to certificates.
Luca Greco [Fri, 14 Oct 2011 18:27:45 +0000 (14:27 -0400)]
websocket: tweak hybi ReadHandshake to supports Firefox implementation
Firefox Websocket implementation send a "Connection: keep-alive, upgrade"
header during the handshake (and as descripted on the last hybi draft
the "Connection" header must include, but doesn't need to be equal to,
"upgrade":
'4. A "Connection" header field that includes the token "Upgrade",
treated as an ASCII case-insensitive value.'
Nigel Tao [Fri, 14 Oct 2011 04:22:02 +0000 (15:22 +1100)]
html: fix some tokenizer bugs with attribute key/values.
The relevant spec sections are 13.2.4.38-13.2.4.40.
http://www.whatwg.org/specs/web-apps/current-work/multipage/tokenization.html#attribute-value-(double-quoted)-state
Nigel Tao [Thu, 13 Oct 2011 22:58:39 +0000 (09:58 +1100)]
html: rewrite the tokenizer to be more consistent.
Previously, the tokenizer made two passes per token. The first pass
established the token boundary. The second pass picked out the tag name
and attributes inside that boundary. This was problematic when the two
passes disagreed. For example, "<p id=can't><p id=won't>" caused an
infinite loop because the first pass skipped everything inside the
single quotes, and recognized only one token, but the second pass never
got past the first '>'.
This change rewrites the tokenizer to use one pass, accumulating the
boundary points of token text, tag names, attribute keys and attribute
values as it looks for the token endpoint.
It should still be reasonably efficient: text, names, keys and values
are not lower-cased or unescaped (and converted from []byte to string)
until asked for.
One of the token_test test cases was fixed to be consistent with
html5lib. Three more test cases were temporarily disabled, and will be
re-enabled in a follow-up CL. All the parse_test test cases pass.
(more are possible but omitted for now as they are part of
specific tests where rather than changing what is there we
should probably expand the tests to cover the new case)
Dmitriy Vyukov [Wed, 12 Oct 2011 14:40:02 +0000 (17:40 +0300)]
runtime: faster strings
Use FlagNoPointers and do not zeroize memory when allocate strings.
test/garbage/parser.out old new
run #1 32.923s 32.065s
run #2 33.047s 31.931s
run #3 32.702s 31.841s
run #4 32.718s 31.838s
run #5 32.702s 31.868s
Robert Griesemer [Wed, 12 Oct 2011 05:28:56 +0000 (22:28 -0700)]
go/scanner: remove AllowIllegalChars mode
This mode was needed before for clients of
the go/scanner that were parsing non-Go code.
All those clients have been moved to scanner
or have been deleted from the library.
Robert Griesemer [Wed, 12 Oct 2011 04:49:53 +0000 (21:49 -0700)]
gofmt: fix a couple of crashes, disallow rewrites for incomplete programs
The current implementation of formatting for incomplete programs
cannot tolerate program rewrites; ignore -rewrite in that case
with a warning message (temporary solution).
Fix a couple of crashes that were introduced recently.
Louis Kruger [Tue, 11 Oct 2011 17:07:32 +0000 (13:07 -0400)]
crypto/tls: add 3DES ciphersuites
The following ciphersuites are added:
TLS_RSA_WITH_3DES_EDE_CBC_SHA
TLS_ECDHE_RSA_WITH_3DES_EDE_CBC_SHA
This change helps conform to the TLS1.1 standard because
the first ciphersuite is "mandatory" in RFC4346
Chris Farmiloe [Tue, 11 Oct 2011 16:53:16 +0000 (12:53 -0400)]
net: fix socket leak in case of Dial failure
Socket descriptors are not closed when fd.connect() fails during generic socket creation.
After a connection failure [ECONNREFUSED] descriptors are left in SYN_SENT state indefinitely (unless they get an explicit RST). Repeated failed connections will eventually cause your program to hit the user/system max-open-files limit.
Rob Pike [Mon, 10 Oct 2011 19:38:49 +0000 (12:38 -0700)]
gob: avoid one copy for every message written.
Plus the need for a second in-memory buffer.
Plays a bit fast and loose with the contents of a byte buffer,
but saves a potentially huge allocation. The gotest
run is about 10% faster overall after this change.
Adam Langley [Sat, 8 Oct 2011 14:06:53 +0000 (10:06 -0400)]
crypto/tls: add server side SNI support.
With this in place, a TLS server is capable of selecting the correct
certificate based on the client's ServerNameIndication extension.
The need to call Config.BuildNameToCertificate is unfortunate, but
adding a sync.Once to the Config structure made it uncopyable and I
felt that was too high a price to pay. Parsing the leaf certificates
in each handshake was too inefficient to consider.
Robert Griesemer [Fri, 7 Oct 2011 00:37:59 +0000 (17:37 -0700)]
go/token: document deserialization property
FileSet deserialization (Read) uses its own instance of a gob decoder.
If the FileSet data may be followed by other data on the reader, Read
may consume too much data that is lost unless the reader implements
ReadByte.
Robert Griesemer [Thu, 6 Oct 2011 23:07:56 +0000 (16:07 -0700)]
go/ast: don't remove function bodies when filtering exports
This is a semantic but no API change. It is a cleaner
implementation of pure filtering. Applications that
need function bodies stripped can easily do this them-
selves.