▲ 64 points
back
47 comments
It's my long standing opinion[1] that we shouldn't accept any RAM that can be subject to random bit flips. Most of the mitigations are just security through obscurity.
My computer with a lot of ECC DDR5 sometimes catches several bitflips in a day. I know this because they are reported to dmesg and sometimes I look at dmesg.
Intel decided long ago that you would need to pay more to not be subject to random bitflips. So it's an AMD system.
My computer with a lot of ECC DDR5 sometimes catches several bitflips in a day.
Replace your RAM or find what's causing this disturbance in the environment. Sooner or later you'll get a pattern the ECC won't be able to correct.
Ironically if you didn't have ECC you would instantly RMA such RAM as defective because even a simple memtest would quickly discover the problem, and "several bitflips in a day" would cause very obvious crashes all the time.
Agreed, my DIY NAS with ECC DDR4 has ran for over 5 years and I have 0 ECC errors in that time.
Something must be very wrong with OP's setup. Perhaps the RAM has been overclocked too far.
Well that's DDR4. Apparently DDR5 was made with tighter tolerances and less safety margin, which is why many people argue ECC should be mandatory in DDR5 - it actually relies on it.
There are plenty of DDR5 machines without ECC[1], and yet they're not constantly crashing in the way that frequent bitflips would suggest.
[1] DDR5's built-in ECC is not the same as traditional "end to end" ECC, and is meant to correct bitflips on the die, not between the RAM and the CPU, so ECC errors that you see are actually suggestive of a bad connection between the RAM and CPU.
Most bitflips don't cause immediate crashes, or crashes at all. They just silently corrupt your data.
For those who don't know, Intel only offers ECC in server CPUs. AMD supports it on all of their current CPUs including client aka consumer ones like in laptops and desktops.
I was shocked to learn that around 10% of Firefox crashes are caused by bit flips.
Most software is like this. Factorio had a weird one where some routers dropped UDP packets with zero checksum.
> It's my long standing opinion[1] that we shouldn't accept any RAM that can be subject to random bit flips.
Every system fails eventually.
That is why every subsystem must have means to detect the failures, to allow the replacement of the bad components.
ECC in DIMMs achieves this.
For myself, this has been the greatest advantage of always using ECC, that I have been warned early about the great increase in the frequency of errors in some DIMMs, after many years (over 5 years) of continuous usage, which has allowed me to replace the offending DIMMs and continue to use those systems for some more years.
Yes things fail, but currently random bit flips are not considered a failure.
The big companies consider them a failure in their own computers, so they use appropriate memory, but they consider that "consumers" either must accept them as normal and not a failure, or pay a great overprice in comparison with what a big company pays, in order to have memory devices with error detection.
They are, that’s why ECC exists, and those who are willing to pay for it, can get it.
100% agree. This is a pure case of defective product. Memory not behaving like memory.
It's very telling that when Rowhammer first appeared, authors of popular memory testing utilities added tests for it, and then quickly hid and disabled them by default because "too much RAM would test defective". I have no idea how they were convinced to do so.
Maybe they don't want to be seen as the boy who cried wolf? Or people would think it's a lie because it would still work for web browsing?
> Most of the mitigations are just security through obscurity.
When manufacturers are sufficiently lazy.
If you put a refresh counter into each memory row at like <1% overhead, you can actually guarantee that a thousand or a hundred accesses forces nearby cells to refresh and stay well inside their safety margin.
Intel MEE addressed this (https://eprint.iacr.org/2016/204.pdf), by treating DRAM as entirely untrusted (assuming an attacker with arbitrary physical read/write capabilities), but then they discontinued it.
Regarding your linked comment:
> If we had properly tested and validated RAM, RowHammer wouldn't work, ever.
While it's exacerbated by physical defects and tight timings, it's really a fundamental problem with how DRAM works. It's frankly a miracle it works in the first place.
No. If it was designed really well, it would trigger extra refreshes as needed. It would know how much accessing a row disturbs nearby rows and how much disturbance a row can suffer before it needs a refresh to prevent data corruption. It would know how long a row can be held open for, too. There is no fundamental theorem that rowhammer must work. It was an engineering tradeoff.
I don't think we can blame DRAM designers for flying a teensy bit too close to the sun here, since this is no problem in normal operation and only appears under adversarial scenarios. We can blame them for not fixing it once it was discovered. And we can definitely blame Intel for making ECC RAM a market segmentation feature.
What you describe is the Target Row Refresh (TRR) mitigation, and it can be bypassed via "half-double" attack.
Where did I imply that they should use an incorrect underestimation of row disturbance? If they actually measured these things, they'd know how far away a disturbance can be created before it becomes so weak the normal refresh cycle fixes it.
It is not feasible for DRAM (or the memory controller) to maintain a fully accurate simulation of DRAM. If it was easy they'd have done it.
It doesn't need to be fully accurate, merely always conservative. It's a question of cost (how much performance does the conservative estimate waste) not feasibility.
> If it was easy they'd have done it.
Why? Something like 70->90% of the world's DRAM is produced by three companies that are very friendly to each other. It makes no business sense to fix problems that you know your peer companies will not fix... that's spending money that you absolutely do not have to.
If the business/economic theory doesn't sway you, look way back to what happened to ISP speeds and pricing when Google Fiber so much as credibly threatened to start providing service in an area served by a mono/duo/triopoly. Or -more recently- how SpaceX demonstrated that the defense-contractor-owned space launch companies had spend decades wasting enormous amounts of taxpayer money by refusing to do any significant amount of research into bringing the cost to launch down substantially. ISPs and the space launch companies had no peers that would spend the resources required to provide a better and/or cheaper service to their customers, so it made absolutely no sense for any one of them to spend resources to break the truce and make them all far less money in the long run.
Okay then refresh the whole thing and not just a single row whenever TRR is triggered.
Or refresh the whole thing after a certain number of row activations.
Or interleave activations with refreshes. Say, after every 5 activations, refresh the next row in the refresh cycle.
I'm not a DRAM expert. They can figure it out. I promise if you refresh the whole chip after every row activation you won't have rowhammer. It'll be too slow though. Somewhere in between is the fastest point where there isn't rowhammer.
It doesn't need to be a fully accurate simulation, any more than your car needs a fully accurate ICE simulation to decide "Change the engine oil".
Just count accesses to each row, and refresh the adjacent row(s) when some limit is reached.
You don't have to simulate it that hard. If you do TRR semi-aggressively and have it give a big boost to the neighboring rows' counter you can handle further away aggressor rows.
This problem has never existed with bigger DRAM cells.
At some point in time, a few generations of DRAM ago, they have reduced the dimensions so much and without discovering adequate mitigations for the problems introduced by this, that the DRAM reliability has become inadequate.
The reason why they did this was to reduce the fabrication costs. It is likely that the pressure to reduce the fabrication costs has been caused more by the desire to increase the profit margins than by the intention to enable any price reductions, because even before the recent price increases there have been around 15 years with only negligible reductions in memory prices.
By increasing the fabrication costs, it would be easy to eliminate the RowHammer problem, while still having memory prices several times lower than the current prices.
However, the vendors do not want this. They want to find some kind of mitigation that would not cause any measurable increase in the fabrication costs. Until now they have failed to do this, but it is not clear how hard they have tried.
It is very likely that their failure to find anything that works has been caused in a good part by the secrecy that is typical for nowadays.
In the earlier times of the semiconductor industry, every manufacturing problem was described in public research papers, with complete details, and usually the right solution was found by someone else and then it spread quickly in all the industry, with much less concerns about "IP" than today.
Only this openness has allowed the creation of the successful semiconductor industry and of the "Silicon Valley".
Little known fun fact: the MEE's checksum updates... were one of the most successful Rowhammer patterns... :-p
Interesting - but presumably used to attack rows outside of SGX?
both, inside and outside
Single event upsets are forced on you by physics. There is no possible way to stop them all.