We should be careful about making the cache too big due to the memory overhead.
I'm not sure where we are with support for streaming parses (beyond the
--stream option in the CLI, which assumes multiple distinct documents that have
been concatted), so in most cases, a multiple of this overhead would be needed
by the infoset itself anyway.
The reason I bring up blobs is that they bypass the requirement to have a large
infoset. Imagine a 4GB document, consisting of a length prefix, 3.99GB worth of
binary blobs, and metadata. There is no good reason parsing this document
should require anywhere near 4GB memory overhead.
> The DFDL spec allows implementations to have limits on backtracking
> capability, it just requires us to document them.
The issue isn't so much backtracking, as fortracking. At the moment, we need to
read in all of dfdl:length before we begin parsing the first time, so we get a
problem even if we never attempt to backtrack.
________________________________
From: Beckerle, Mike <mbecke...@owlcyberdefense.com>
Sent: Friday, September 11, 2020 2:51 PM
To: dev@daffodil.apache.org <dev@daffodil.apache.org>
Subject: Re: Large dfdl:length values on complex types
We have tunable limits people have to sometimes enlarge. E.g. max size of a
regex match is limited.
The DFDL spec allows implementations to have limits on backtracking capability,
it just requires us to document them.
That said. We really don't want to have this sort of limit hard coded.
A non-streamable file format that just stores the total length in a header
record is pretty common, and people will use those for really big pieces of
data these days.
Can this 256MB just be a tunable constant that people can enlarge?
>From a security perspective, a stored length field is always something that
>should be sanity checked - if clobbered, it could contain -1 and that could be
>interpreted as maxInt or something, resulting in a denial-of-service attack.
>It would be better to get a processing error in that case.
Can we just check for complex type lengths against a tunable limit, and error
out before we even try to read it in?
________________________________
From: Sloane, Brandon <bslo...@owlcyberdefense.com>
Sent: Friday, September 11, 2020 1:53 PM
To: dev@daffodil.apache.org <dev@daffodil.apache.org>
Subject: Re: Large dfdl:length values on complex types
256MB doesn't strike me as that big. We haven't ad filesize limits measured in
GB since fat32 and ext1 (depending on settings; even ext4 can have a limit as
low as 16gb). All it takes is such a file to have a length prefix, and we can
very easily run into the limit. Combined with our support of large binary
blobs, it is not unreasonable that someone would want to use DFDL on such a
format.
________________________________
From: Beckerle, Mike <mbecke...@owlcyberdefense.com>
Sent: Friday, September 11, 2020 1:24 PM
To: dev@daffodil.apache.org <dev@daffodil.apache.org>
Subject: Re: Large dfdl:length values on complex types
Maybe a silly question, but why don't we just hit a tunable size limit
immediately before we "try to read" that data? 256MB is very big.
Is this a real format, or a test case designed to push the boundaries?
________________________________
From: Steve Lawrence <slawre...@apache.org>
Sent: Friday, September 11, 2020 1:14 PM
To: dev@daffodil.apache.org <dev@daffodil.apache.org>
Subject: Large dfdl:length values on complex types
I recently came across an issue where we have something like this:
<xs:element name="length" type="xs:int" ... />
<xs:element name="data"
dfdl:lengthKind="explicit" dfdl:length="{ ../length }">
<xs:complexType>
<xs:sequence>
<xs:element name="field1" ... />
<xs:element name="field2" ... />
...
<xs:element name="fieldN" ... />
</xs:sequence>
</xs:complexType>
</xs:element>
So we have a length element and a complex data field that uses this
length, and the data field is made up of a bunch of fields.
The issue I come across is related to how we cache bytes in buckets for
backtracking. As we fill up buckets, we currently limit the total amount
cache size of the buckets to 256MB. So if someone ever parses more than
256MB of data and then tries to backtrack past that, we error. The idea
being that we don't want to keep an infinite cache for potential
backtracking and people should have realized that they went down the
wrong branch much earlier.
Though, a problem occurs with the complex types with a large specified
length like above. When we have the complex type with expression
../length, before trying to parse any of the fields, we read that length
number of bytes into our cache buckets to confirm that that number of
bytes exists. The problem occurs if length is more than 256MB. In this
case, we read length number of bytes, and start removing elements from
the cache once we read more than 256MB.
But once that succeeds and we read length bytes, we then try to start
parsing the fields within the complex type, but we've removed those
early cached bytes, and so we fail with an unhelpful backtracking exception.
I'm not sure of the right solution here.
Perhaps we shouldn't be throwing away these bytes when dealing with
complex lengths?
Or perhaps we shouldn't even be trying to determine if that many bytes
are available when we have a specified length. Instead, maybe we should
just set the bit limit to make sure we don't parse more than than that?
And if eventually something tries to read a byte and there aren't enough
and we hit that limit, only then do we fail. This feels like the right
solution, but wanted to start a discussion to see if maybe there's a
reason we try to read the full length, or maybe there's another alternative?
