Wrong layer.

SSDs know which blocks have been written to a lot, have been giving a lot of read errors before etc., and often even have heterogeneous storages (such as a bit of SLC for burst writing next to a bunch of MLC for density).

They can spend ECC bits much more efficiently with that information than a file system ever could, which usually sees the storage as a flat, linear array of blocks.

This is true, but nevertheless you cannot place your trust only in the manufacturer of the SSD/HDD, as I have seen enough cases when the SSD/HDD reports no errors, but nonetheless it returns corrupted data.

For any important data you should have your own file hashes, for corruption detection, and you should add some form of redundancy for file repair, either with a specialized tool or simply by duplicating the file on separate storage media.

A database with file hashes can also serve other purposes than corruption detection, e.g. it can be used to find duplicate data without physically accessing the archival storage media.

Verifying at higher layers can be ok (it's still not ideal!), but trying to actively fix things below that are broken usually quickly becomes a nightmare.

IMO it's exactly the right layer, just like for ECC memory.

There's a lot of potential for errors when the storage controller processes and turns the data into analog magic to transmit it.

In practice, this is a solved problem, but only until someone makes a mistake, then there will be a lot of trouble debugging it between the manufacturer certainly denying their mistake and people getting caught up on the usual suspects.

Doing all the ECC stuff right on the CPU gives you all the benefits against bitrot and resilience against all errors in transmission for free.

And if all things go just right we might even be getting better instruction support for ECC stuff. That'd be a nice bonus

> There's a lot of potential for errors when the storage controller processes and turns the data into analog magic to transmit it.

That's a physical layer, and as such should obviously have end-to-end ECC appropriate to the task. But the error distribution shape is probably very different from that of bytes in NAND data at rest, which is different from that of DRAM and PCI again.

For the same reason, IP does not do error correction, but rather relies on lower layers to present error-free datagram semantics to it: Ethernet, Wi-Fi, and (managed-spectrum) 5G all have dramatically different properties that higher layers have no business worrying about. And sticking with that example, once it becomes TCP's job to handle packet loss due to transmission errors (instead of just congestion), things go south pretty quickly.

> And sticking with that example, once it becomes TCP's job to handle packet loss due to transmission errors (instead of just congestion), things go south pretty quickly.

Outside of wireless links (where FEC of some degree is necessary regardless) this is mostly because TCP’s checksum is so weak. QUIC for example handles this much better, since the packet’s authenticated encryption doubles as a robust error detecting code. And unlike TLS over TCP, the connection is resilient to these failures: a TCP packet that is corrupted but passes the TCP checksum will kill the TLS connection on top of it instead of retransmitting.

Ah, I meant go south in terms of performance, not correctness. Most TCP congestion control algorithms interpret loss exclusively as a congestion signal, since that's what most lower layers have historically presented to it.

This is why newer TCP variants that use different congestion signals can deal with networks that violate that assumption better, such as e.g. Starlink: https://blog.apnic.net/2024/05/17/a-transport-protocols-view...

Other than that, I didn't realize that TLS has no way of just retransmitting broken data without breaking the entire connection (and a potentially expensive request or response with it)! Makes sense at that layer, but I never thought about it in detail. Good to know, thank you.

ECC memory modules don’t do their own very complicated remapping from linear addresses to physical blocks like SSDs do. ECC memory is also oriented toward fixing transient errors, not persistently bad physical blocks.

The filesystem doesn't have access to the right existing ECC data to be able to add a few bytes to do the job. It would need to store a whole extra copy.

There are potentially ways a filesystem could use heirarchical ECC to just store a small percentage extra, but it would be far from theoretically optimal and rely on the fact just a few logical blocks of the drive become unreadable, and those logical blocks aren't correlated in write time (which I imagine isn't true for most ssd firmware).

Reed Solomon codes, or forward error correction is what you’re discussing. All modern drives do it at low levels anyway.

It would not be hard for a COW file system to use them, but it can easily get out of control paranoia wise. Ideally you’d need them for every bit of data, including metadata.

That said, I did have a computer that randomly bit flipped when writing to storage sometimes (eventually traced it to an iffy power supply), and PAR (a type of reed solomon coding forward error correction library) worked great for getting a working backup off the machine. Every other thing I tried would end up with at least a couple bit flip errors per GB, which make it impossible.

CD storage has an interesting take, the available sector size varies by use, i.e. audio or MPEG1 video (VideoCD) at 2352 data octets per sector (with two media level ECCs), actual data at 2048 octets per sector where the extra EDC/ECC can be exposed by reading "raw". I learned this the hard way with VideoPack's malformed VCD images, I wrote a tool to post-process the images to recreate the correct EDC/ECC per sector. Fun fact, ISO9660 stores file metadata simultaneously in big-endian and little form (AFAIR VP used to fluff that up too).

Octets? Don't you mean "bytes"? Or is that word problematic now?

Personally, I prefer the word "bytes", but "octets" is technically more accurate as there are systems that use differently sized bytes. A lot of these are obsolete but there are also current examples, for example in most FPGA that provide SRAM blocks, it's actually arranged as 9, 18 or 36-bit wide with the expectation that you'll use the extra bits for parity or flags of some kind.

I wonder if OP used "octets" because physical pattern in the CD used to represent a byte is a sequence of 17 pits and lands.

BTW, byte size during the history varied from 4 to 24 bit! Even now, based on interpretation, you can say 16 bit bytes do exist.

Char type can be 16 bit on some DSP systems.

I was curious, so I checked. Before this comment, I only knew about 7 bit bytes.

The term octets is pretty common in network protocol RFCs, maybe their vocabulary is biased in the direction of that writing.

Octets is the term used in most international standards instead of the American "byte".

"Octet" has the advantage that it is not ambiguous. In old computer documentation, from the fifties to the late sixties, a "byte" could have meant any size between 6 bits and 16 bits, the same like "word", which could have meant anything between 8 bits and 64 bits, including values like 12 bits, 18 bits, 36 bits, 60 bits, or even 43 bits.

Traditionally, computer memory is divided in pages, which are divided in lines, which are divided in words, which are divided in bytes. However the sizes of any of those "units" has varied in very wide ranges in the early computers.

IBM System/360 has chosen the 8-bit byte, and the dominance of IBM has then forced this now ubiquitous meaning of "byte", but there were many computers before System/360 and many coexisting for some years with the IBM 360 and later mainframes, where byte meant something else.

Not problematic, minor pedantry. With much time spent reading (and occasionally writing) technical documentation it's octets, binary prefixes, and other wanton pedantry where likely to be understood/appreciated or precision is required.

FTR, ECMA-130 (the CD "yellow book" equivalent standard) is littered with the term "8-bit bytes", so it was certainly a thing then. Precision when simultaneously discussing eight-to-fourteen modulation, and the 17 encoding "bits" that hit the media for each octet as noted in a sibling comment.

Now, woktets on the other hand...

You can still do this for boot code if the error isn't significant enough to make all of the boot fail. The "fixing it by plugging it in somewhere else" could then also be simple enough to the point of being fully automated.

ZFS has "copies=2", but iirc there are no filesystems with support for single disk erasure codes, which is a huge shame because these can be several orders of magnitude more robust compared to a simple copy for the same space.

zfs can run with a single disk stripe. pfsense gladly runs this way. See https://docs.netgate.com/pfsense/en/latest/install/install-z...

You can, but only if your CPU is directly connected to a flash chip with no controller in the way. Linux calls it the mtd subsystem (memory technology device).