Potential DoS vector: blobs that are too big to be included in a square

[rootulp]

Problem

Originally reported via @tuxcanfly. See https://gist.github.com/tuxcanfly/207ab3dc1fac4bb9c38bedad2a053f18

Proposal

Is it possible for CheckTx to reject this type of transaction prior admittance into the mempool?

• A stateful check could be based on the GovSquareSize of 64 which would reject transactions larger than ~2MiB
• A stateless check could be based on the SquareSizeUpperBound of 128 which would reject transactions larger than ~8MiB

[rootulp]

Note we already have an ante handler that rejects txs where the total blob size is too large. It's possible that ante handler doesn't catch edge cases near the total blob size because it derives an upper bound based on a number of parameters. See

• What happens to transactions that contain a blobs that are too large? #2156
• feat: reject tx if total blob size too large #2202
• Pass app version to CheckTx #2214

[cmwaters]

I think we calculate the space taken by the blobs incorrectly

celestia-app/x/blob/ante/max_total_blob_size_ante.go

Lines 74 to 79 in 2d261a2

	func getTotal(sizes []uint32) (sum int) {
	for _, size := range sizes {
	sum += int(size)
	}
	return sum
	}

it should be in regards of the amount of shares * 512 bytes i.e.

for _, size := range sizes {
	sum += uint64(shares.SparseSharesNeeded(size)) * appconsts.ShareSize
}

Also I would consider of adding an extra share or two as buffer for padding

[rootulp]

getTotal is only used here where we want the number of total bytes in the blob sizes.

Also I would consider of adding an extra share or two as buffer for padding

Agreed. That ante handler doesn't account for padding shares so this is related to #3081

[cmwaters]

getTotal is only used here where we want the number of total bytes in the blob sizes.

yes but we want the total bytes based on the amount of shares that those blobs will occupy; not the raw bytes of the blobs

[rootulp]

🤦‍♂️ you're right, good catch!

[rootulp]

This issue may be partially mitigated by the fix in #3082 which will be released in the next major release.

[evan-forbes]

@rootulp since this is a bug, is it okay if we treat is as such? and is this still an open issue given #3082?

[rootulp]

TBH I still think this is open. #3082 improves the ante handler to reject PFBs that are guaranteed to not fit in a data square. But I think the DoS vector still exists and a malicious actor could craft PFBs that pass the ante handler but fail to fit in a data square.

Agreed it is a bug. What do you mean by "treat it as such"?

[evan-forbes]

Agreed it is a bug. What do you mean by "treat it as such"?

mainly that we just add a label lol and discuss what is needed to fix it ofc. that is weird wording tho, apologies

do we need more tests than what were added in #3082? what do we need to do to be confident that this isn't occuring? I could be missing something, but we should be able to deterministically test for this, correct?

[rootulp]

what do we need to do to be confident that this isn't occuring?

One more idea to mitigate the DoS: modify BroadcastTx from blocking to non-blocking

do we need more tests

We could use a fuzz test that submits blobs to the ante handler and verifies that if a valid PFB tx passes the ante handler then it also must be able to be included in an empty block (i.e. a block where it is the only tx).

[evan-forbes]

One more idea to mitigate the DoS: modify BroadcastTx from blocking to non-blocking

does the signer in the user packager using BroadcastMode Sync count for non-blocking? In general I agree that we should get rid of broadmost mode block, and perhaps a stupic question, but how does this reduce DoS? just that it reduces the resources required per connection? If so, we do have configurable limits on the time each node will wait and how many active connections it maintains help?

We could use a fuzz test that submits blobs to the ante handler and verifies that if a valid PFB tx passes the ante handler then it also must be able to be included in an empty block (i.e. a block where it is the only tx).

adding a fuzzer is a good idea! for this specific issue, perhaps we could limit to the scope to just the size of the PFB and the problematic antehandler?

[evan-forbes]

to close issue, we need to add a test case for the boundary of max blob size in this test

celestia-app/x/blob/ante/blob_share_decorator_test.go

Line 21 in a93bb62

func TestBlobShareDecorator(t *testing.T) {

[evan-forbes]

I think we can close this as completed @rootulp correct

[rootulp]

We can close this and open a new issue for

We could use a fuzz test that submits blobs to the ante handler and verifies that if a valid PFB tx passes the ante handler then it also must be able to be included in an empty block (i.e. a block where it is the only tx).

Created #4275