tahoe-lafs/trac-2024-07-25

downloader hangs when server returns empty string #2024

New issue

Open

opened 2013-07-16 16:17:53 +00:00 by warner · 2 comments

warner commented

2013-07-16 16:17:53 +00:00

Owner

While investigating the test_download.Corrupt.test_each_byte catalog_detection=True failure for #1382, after fixing the bitrot, we discovered that the downloader hangs when the server responds to a read request with zero bytes. In particular, when the test corrupts offset 8 (which is the MSB of the num_leases value), the storage server believes that the container includes a ridiculous number of leases, therefore the leases start (self._lease_offset) before the beginning of the file, therefore all reads are truncated down to nothing.

The actual bug is that the downloader doesn't handle this well. It looks like, when the read fails to satisfy any of the desired bytes, the downloader just issues a new read request, identical to the first. It then loops forever, trying to fetch the same range and always failing. This is an availability problem, since a corrupt/malicious server could prevent downloads from proceeding by mangling its responses in this way.

Instead, the downloader should never ask for a given range of bytes twice from the same storage server (at least without some intervening event like a reconnection). So the downloader will need to remember what it asked for, and if it doesn't get it, add those offsets to a list of "bytes that this server won't give us". Then, if we absolutely need any bytes that appear in that list, we declare the Share to be a loss and switch to a different one.

A simpler rule would probably work too: any zero-length reads are grounds to reject the share. We do some speculative reads (based upon assumptions about the segment size), so we'd need to look carefully at that code and make sure that the speculation cannot correctly yield zero bytes. In particular I'm thinking about the UEB fetch from the end of the share: its offset depends upon the size of the hash trees, so if our guessed segment size is too small, the UEB fetch might return zero bytes, but the correct thing to do is to re-fetch it from the correct place once we've grabbed the offset table.

The workaround I recommended for markb's work in #1382 is to just refrain from corrupting those four num_leases bytes. Once we fix this ticket, we should go back to test_download.py and remove that workaround.

While investigating the `test_download.Corrupt.test_each_byte catalog_detection=True` failure for #1382, after fixing the bitrot, we discovered that the downloader hangs when the server responds to a read request with zero bytes. In particular, when the test corrupts offset 8 (which is the MSB of the `num_leases` value), the storage server believes that the container includes a ridiculous number of leases, therefore the leases start (`self._lease_offset`) before the beginning of the file, therefore all reads are truncated down to nothing. The actual bug is that the downloader doesn't handle this well. It looks like, when the read fails to satisfy any of the desired bytes, the downloader just issues a new read request, identical to the first. It then loops forever, trying to fetch the same range and always failing. This is an availability problem, since a corrupt/malicious server could prevent downloads from proceeding by mangling its responses in this way. Instead, the downloader should never ask for a given range of bytes twice from the same storage server (at least without some intervening event like a reconnection). So the downloader will need to remember what it asked for, and if it doesn't get it, add those offsets to a list of "bytes that this server won't give us". Then, if we absolutely need any bytes that appear in that list, we declare the Share to be a loss and switch to a different one. A simpler rule would probably work too: any zero-length reads are grounds to reject the share. We do some speculative reads (based upon assumptions about the segment size), so we'd need to look carefully at that code and make sure that the speculation cannot correctly yield zero bytes. In particular I'm thinking about the UEB fetch from the end of the share: its offset depends upon the size of the hash trees, so if our guessed segment size is too small, the UEB fetch might return zero bytes, but the correct thing to do is to re-fetch it from the correct place once we've grabbed the offset table. The workaround I recommended for markb's work in #1382 is to just refrain from corrupting those four `num_leases` bytes. Once we fix this ticket, we should go back to `test_download.py` and remove that workaround.

tahoe-lafs added the

labels 2013-07-16 16:17:53 +00:00

tahoe-lafs added this to the eventually milestone 2013-07-16 16:17:53 +00:00

daira commented

2013-07-17 15:16:20 +00:00

Author

Owner

Replying to warner:

A simpler rule would probably work too: any zero-length reads are grounds to reject the share.

That wouldn't work against a malicious server that, say, returns 1 byte.

The workaround I recommended for markb's work in #1382 is to just refrain from corrupting those four num_leases bytes. Once we fix this ticket, we should go back to test_download.py and remove that workaround.

+1.

Replying to [warner](/tahoe-lafs/trac-2024-07-25/issues/9497): > A simpler rule would probably work too: any zero-length reads are grounds to reject the share. That wouldn't work against a malicious server that, say, returns 1 byte. > The workaround I recommended for markb's work in #1382 is to just refrain from corrupting those four `num_leases` bytes. Once we fix this ticket, we should go back to `test_download.py` and remove that workaround. +1.

daira commented

2013-07-17 15:18:02 +00:00

Author

Owner

Actually I'd like to have a test marked TODO for this bug first, to make sure it doesn't get lost.