Add `consensus_decode_from_finite_reader` optimization by dpc · Pull Request #1023 · rust-bitcoin/rust-bitcoin

dpc · 2022-05-29T04:10:57Z

As things are right now, memory exhaustion protection in Decodable
is based on checking input-decoded lengths against arbitrary limits,
and ad-hoc wrapping collection deserialization in Take.

The problem with that are two-fold:

Potential consensus bugs due to incorrect limits.
Performance degradation when decoding nested structured,
due to recursive Take<Take<..>> readers.

This change introduces a systematic approach to the problem.

A concept of a "size-limited-reader" is introduced to rely on
the input data to finish at enforced limit and fail deserialization.

Memory exhaustion protection is now achived by capping allocations
to reasonable values, yet allowing the underlying collections
to grow to accomodate rare yet legitmately oversized data (with tiny
performance cost), and reliance on input data size limit.

A set of simple rules allow avoiding recursive Take wrappers.

Fix #997

dpc · 2022-05-29T04:17:13Z

src/consensus/encode.rs

Not true, because network message is using it to decode payload.

MAX_VEC_SIZE it is then.

src/consensus/encode.rs

dpc · 2022-05-29T05:43:50Z

OK. I have to benchmark this on a laptop, with a lot of noise, so I've ran the bench multiple times and picked some best results for both master and this PR:

old:
test blockdata::block::benches::bench_block_deserialize                 ... bench:   1,381,111 ns/iter (+/- 18,765)
test blockdata::transaction::benches::bench_transaction_deserialize     ... bench:         133 ns/iter (+/- 5)

test blockdata::block::benches::bench_block_deserialize                 ... bench:   1,431,350 ns/iter (+/- 187,241)
test blockdata::transaction::benches::bench_transaction_deserialize     ... bench:         177 ns/iter (+/- 4)

test blockdata::block::benches::bench_block_deserialize                 ... bench:   1,364,857 ns/iter (+/- 22,960)
test blockdata::transaction::benches::bench_transaction_deserialize     ... bench:         127 ns/iter (+/- 6)


new:
test blockdata::block::benches::bench_block_deserialize                 ... bench:   1,112,535 ns/iter (+/- 45,702)
test blockdata::transaction::benches::bench_transaction_deserialize     ... bench:         145 ns/iter (+/- 3)

test blockdata::block::benches::bench_block_deserialize                 ... bench:   1,093,602 ns/iter (+/- 21,832)
test blockdata::transaction::benches::bench_transaction_deserialize     ... bench:          97 ns/iter (+/- 2)

test blockdata::block::benches::bench_block_deserialize                 ... bench:   1,133,440 ns/iter (+/- 36,818)
test blockdata::transaction::benches::bench_transaction_deserialize     ... bench:         146 ns/iter (+/- 2)

apoelstra · 2022-05-30T15:32:04Z

src/blockdata/script.rs

In ae28a501e0e41a78afb6b01dcf1fba8a689262dd:

Curious why this isn't just a default impl on the trait? This code looks like it's repeated for every individual type.

Ah, you already have a default impl going the other way.

FWIW I think it would be ok to have both methods default-impl'd to call each other, with doccomments instructing trait implementors to implement one. I believe that std does this somewhere, though I don't remember where..

I think I like it. I'll give it a try.

src/network/message.rs

apoelstra · 2022-05-30T15:42:13Z

src/consensus/encode.rs

In 3a87775a22f2957f8b71ca8a68dfaadd5615a3f0:

I'm a liiitle tempted to change the above line to Cursor::new(&data[..MAX_VEC_SIZE]) but maybe that's a little too weird/surprising.

I was thinking about it. It would give a little bit of protection if someone somehow passed a huge vector in here. But then - if the caller allows malicious input to create huge vectors, they kind of already lost anyway. And the ability of decoding/encoding huge data might be useful for someone that can trust it. Eg. Vec<Block> stored in the file system or something. And arbitrary limits might get in the way.

apoelstra

ACK 888a1ec5c4e830f905288b6a824cb23fc1375989

tcharding · 2022-05-31T00:52:33Z

I did a bunch of docs fixes (mainly grammar), if you'd like to use them I pushed a patch to branch dpc-issue-997-2 on my fork.

I had one question, just out of interest, is there a reason you use such short line length in the comments?

While cleaning up I've been favouring line length 100 for comments (I try to not touch them if they are at least 80 though if changing them does not make them noticeably nicer.)

dpc · 2022-05-31T00:57:30Z

I had one question, just out of interest, is there a reason you use such short line length in the comments?

I typically let rustfmt deal with it, so I wasn't thinking much about. I'm so used to typing out garbage that rustfmt just fixes for me, that I have a hard enough time with the code, and wasn't even thinking about comments. :D

I did a bunch of docs fixes (mainly grammar), if you'd like to use them I pushed a patch to branch dpc-issue-997-2 on my fork.

Notes. Will do. Thanks a lot!

tcharding · 2022-05-31T01:20:55Z

I typically let rustfmt deal with it, so I wasn't thinking much about. I'm so used to typing out garbage that rustfmt just fixes for me, that I have a hard enough time with the code, and wasn't even thinking about comments. :D

ah, no sweat. We should have that fixed up soon hopefully :)

tcharding

All my comments are nits only.

ACK 888a1ec5c4e830f905288b6a824cb23fc1375989

tcharding · 2022-05-31T01:08:39Z

src/consensus/encode.rs

This comment is stale, there is no len argument now. This PR was a little confusing to review because the last patch modifies code in the first patch.

src/consensus/encode.rs

tcharding · 2022-05-31T01:11:33Z

src/consensus/encode.rs

A question please since we have been discussing debug asserts a bit but I'm still not totally understanding when to use what. This debug statement is defensive programming against insane input with chunck size 0 causing an infinite loop, right? What made you choose to use debug assert instead of either
A. a regular asset
B. returning an error

I wasn't really sure myself what to do. chunk_size should always be a constant, and this whole function is kind of internal, so I thought checking in debug_ is good enough to spot some silly mistake. I'm still not sure about it. I wish I could make chunk_size a const of some kind.

I just turned it into assert_ne. Compiler can probably optimize it anyway.

tcharding · 2022-05-31T01:17:56Z

src/consensus/encode.rs

Why references here? This works too

assert_eq!( read_bytes_from_finite_reader( io::Cursor::new(&data), ReadBytesFromFiniteReaderOpts { len: data.len(), chunk_size } ).unwrap(), data );

Natural possessiveness, I guess. 🤷

src/network/message.rs

tcharding · 2022-05-31T02:47:51Z

LGTM, can you squash this all down into a single commit please?

As things are right now, memory exhaustion protection in `Decodable` is based on checking input-decoded lengths against arbitrary limits, and ad-hoc wrapping collection deserialization in `Take`. The problem with that are two-fold: * Potential consensus bugs due to incorrect limits. * Performance degradation when decoding nested structured, due to recursive `Take<Take<..>>` readers. This change introduces a systematic approach to the problem. A concept of a "size-limited-reader" is introduced to rely on the input data to finish at enforced limit and fail deserialization. Memory exhaustion protection is now achived by capping allocations to reasonable values, yet allowing the underlying collections to grow to accomodate rare yet legitmately oversized data (with tiny performance cost), and reliance on input data size limit. A set of simple rules allow avoiding recursive `Take` wrappers. Fix #997

apoelstra

ACK 082e185

apoelstra · 2022-05-31T14:29:36Z

No comment on choice of assert ;)

tcharding

ACK 082e185

sanket1729

post merge ACK. Thanks for doing this cleanly

dpc · 2022-06-03T06:13:12Z

Bad news. Somehow, somewhere I've lost e6a3ec41bb9b3f072453665338059879d982eada . No idea when and why.

Oh, it doesn't even show up.

commit e6a3ec41bb9b3f072453665338059879d982eada (issue-997)
Reflog: HEAD@{87} (Dawid Ciężarkiewicz <dpc@dpc.pw>)
Reflog message: commit (amend): Forward default `consensus_decode` to `consensus_decode_from_finite_reader`
Author: Dawid Ciężarkiewicz <dpc@dpc.pw>
Date:   Mon May 30 18:14:09 2022 -0700

    Forward default `consensus_decode` to `consensus_decode_from_finite_reader`

I'll introduce it in another PR. You might want to take a second look if I didn't mess up something else, just in case. I'm really sorry - no idea RN how it happened.

*Edit: I push it again, should show up as 80d8cc2f6bc4e4ccbdc57ec5fff9d283c1e5e263 . I intended for it to be in there, would make this PR even smaller.

dpc · 2022-06-12T03:29:29Z

@RCasatta since I know you're using these traits in blocks_iterator and care about performance there you might want to be aware of this and #1035 .

…te_reader` optimization 082e185 Add `consensus_decode_from_finite_reader` optimization (Dawid Ciężarkiewicz) Pull request description: As things are right now, memory exhaustion protection in `Decodable` is based on checking input-decoded lengths against arbitrary limits, and ad-hoc wrapping collection deserialization in `Take`. The problem with that are two-fold: * Potential consensus bugs due to incorrect limits. * Performance degradation when decoding nested structured, due to recursive `Take<Take<..>>` readers. This change introduces a systematic approach to the problem. A concept of a "size-limited-reader" is introduced to rely on the input data to finish at enforced limit and fail deserialization. Memory exhaustion protection is now achived by capping allocations to reasonable values, yet allowing the underlying collections to grow to accomodate rare yet legitmately oversized data (with tiny performance cost), and reliance on input data size limit. A set of simple rules allow avoiding recursive `Take` wrappers. Fix #997 ACKs for top commit: apoelstra: ACK 082e185 tcharding: ACK 082e185 Tree-SHA512: fa04b62a4799c9a11c5f85ec78a18fa9c2cd4819c83a0d6148fbb203c6fa15c2689cb0847e612b35b8c285a756d81690b31a9bede4486b845f0c16b9fcc6d097

a0489d4 fuzz: use travis-fuzz.sh in CI (Andrew Poelstra) 4c6f9b3 fuzz: remove mysteriously-not-necessary quotes from gh action script (Andrew Poelstra) 7baa21c fuzz: disable features in honggfuzz (Andrew Poelstra) e003e57 Add `consensus_decode_from_finite_reader` optimization (Dawid Ciężarkiewicz) Pull request description: Backport for #1023. Required for #997. Addresses issues like #1359 ACKs for top commit: tcharding: ACK a0489d4 TheBlueMatt: ACK a0489d4. Tree-SHA512: 9145d9666e35ae77598aaecf89222c7234637b57ded39b69fbb93ee9ce01c6d7c938b36a2d86319ba84155f2424e524386593d6c0d7af449be6bd118f729fd64

dpc mentioned this pull request May 29, 2022

Refactor trait Decodable (sketch) #1019

Closed

dpc commented May 29, 2022

View reviewed changes

src/consensus/encode.rs Outdated Show resolved Hide resolved

dpc marked this pull request as ready for review May 29, 2022 05:44

apoelstra reviewed May 30, 2022

View reviewed changes

src/network/message.rs Outdated Show resolved Hide resolved

apoelstra reviewed May 30, 2022

View reviewed changes

apoelstra previously approved these changes May 30, 2022

View reviewed changes

tcharding previously approved these changes May 31, 2022

View reviewed changes

dpc dismissed stale reviews from tcharding and apoelstra via af2c62f May 31, 2022 01:51

tcharding reviewed May 31, 2022

View reviewed changes

src/network/message.rs Outdated Show resolved Hide resolved

apoelstra approved these changes May 31, 2022

View reviewed changes

tcharding approved these changes May 31, 2022

View reviewed changes

apoelstra merged commit 58a62c0 into rust-bitcoin:master Jun 1, 2022

sanket1729 reviewed Jun 1, 2022

View reviewed changes

TheBlueMatt mentioned this pull request Nov 1, 2022

Add a test parsing transaction with a huge witness #1359

Merged

This was referenced Nov 1, 2022

Bug: "oversized vector allocation" due to large transaction (lots of witness inputs) romanz/electrs#783

Closed

[backport] consensus decode from finite decoder. #1360

Merged

apoelstra mentioned this pull request Aug 10, 2025

Implement Encodable/Decodable for list of u8 #4762

Merged

Conversation

dpc commented May 29, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

dpc commented May 29, 2022

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

apoelstra left a comment

Choose a reason for hiding this comment

Uh oh!

tcharding commented May 31, 2022

Uh oh!

dpc commented May 31, 2022

Uh oh!

tcharding commented May 31, 2022

Uh oh!

tcharding left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

tcharding commented May 31, 2022

Uh oh!

apoelstra left a comment

Choose a reason for hiding this comment

Uh oh!

apoelstra commented May 31, 2022

Uh oh!

tcharding left a comment

Choose a reason for hiding this comment

Uh oh!

sanket1729 left a comment

Choose a reason for hiding this comment

Uh oh!

dpc commented Jun 3, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

dpc commented Jun 12, 2022

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

dpc commented May 29, 2022 •

edited

Loading

dpc commented Jun 3, 2022 •

edited

Loading