5 ms·
We have today 32+ cores CPUs with minimal or no branch prediction and speculation. They work spectacularly well on some specific workloads but are terrible for
by gpderetta 8y ago
We have today 32+ cores CPUs with minimal or no branch prediction and speculation. They work spectacularly well on some specific workloads but are terrible for general purpose computation. They are called GPUs.
Speculation and branch prediction is not going away. On the contrary is going to get more and more sophisticated.
What needs to die is the belief that you can rely on just software (i.e. memory safety) for isolation between trusted and untrusted code[1].
[1] Yes, meltdown bypasses hardware protection, but that's an intel specific fuck up, not inherent issue with speculation.
- twtw 8y ago> What needs to die is the belief that you can rely on just software (i.e. memory safety) for isolation between trusted and untrusted code[1]. I don't understand how you came to this conclusion. How does spectre indicate that you can't rely on software for isolation? > meltdown bypasses hardware protection, but that's Intel specific It's not Intel specific. Certain ARM and POWER architectures are also vulnerable. Spectre bypasses hardware protections too, just in a more subtle way.
- gizmo686 8y agoThe idea behind spectre is that speculative execution bypasses software protections. Given if f(A) B else C, one cannot safely assume that B only gets executed when A resolves to true. It is not reasonable to expect programmers to produce safe code in an environment where they have to doubt if statements. This could be solved without any major overhauls in hardware, by just updating it to make speculative execution actually speculative and not commit any side effects until it knows it went down the correct branch.
- gpderetta 8y agoUnfortunately fetching a new cache line from memory is a globally visible side effect. One of the major benefits of speculation and is that it enables additional memory level parallelism. Even otherwise strictly in order designs, when aiming for high performance incorporate some form of memory speculation (scout threading, run ahead execution).
- gizmo686 8y agoRight, and CPUs could, in principle, fix this. Eg. they could speculatively fetch the cache line, and only commit it to the cache once they know what branch they took. This still lets you do the slow memory access before you know if you will need it, without causing globally visible side effects.
- gpderetta 8y agoAs soon as you fetch the cache line, every other core will know you have done that via the cache coherency protocol. You have just disclosed an address potentially derived from some sensitive information. Edit: to be precise they can, for example, see the transition from exclusive to shared.
- gizmo686 8y agoI suppose this gets into what is consider a "major" overhaul of the architecture. The idea is that you don't actually "fetch" anything until you are no longer speculating. The core that is doing the speculating can pre-fetch the memory, and then commit it to the cache only once it knows that is the correct thing to do. There are still some cache coherence concerns, but nothing that strikes me as insurmountable.
- gpderetta 8y agoI do not understand your distinction between prefect and fetch. You either have the data, and you need to notify other cores that you do or you don't. Not notifying other CPUs seems pointless because before committing you would have to do a round-trip to the other core to verify whether the data is still valid making the prefetching pointless.
- twtw 8y agoWhat you are describing here is high cost, because you have to take care of coherence when the line is committed to the cache, which in your scenario would not be speculative. That would require notifying the other cores and potentially refetching if another core has written to the line between your prefetch and commit. A more practical solution is probably to have protection boundaries within the cache, like http://people.csail.mit.edu/vlk/dawg-micro18.pdf http://people.csail.mit.edu/vlk/dawg-micro18.pdf.
- twtw 8y agoSure, I get it. Perhaps I phrased my response poorly. This comment: > What needs to die is the belief that you can rely on just software (i.e. memory safety) for isolation between trusted and untrusted code is an argument that doubting if statements should be acceptable, and that the assumption that only B has architecturally visible side effects when A is true is not sound. I disagree strongly with that. You should be able to rely on software checks for some things. Going forward, the solution is not to rewrite software to not depend on software mechanisms (e.g. bounds checks) for protection (unclear what would be used instead...) but to fix the hardware so these software mechanisms work. What needs to "die" is not "belief that you can rely on just software," but hardware that violates fundamental guarantees. It sounds like we agree on this, I just wanted to make my point clear since I see now that my original comment was not clear. FWIW, this is essentially the argument Torvalds made when Intel tried to add feature flags for non-broken speculation.
- gpderetta 8y agoI don't think Linus has any expectation that spectre v1 will ever be fixed in hardware.
- loup-vaillant 8y agoThere's a difference between what you expect, and how you think things should be. Linus was probably making a normative statement, not a prediction.
- pmoriarty 8y ago"We have today 32+ cores CPUs ... They are called GPUs." There are also chips like the GA144, which has 144 cores: https://www.rs-online.com/designspark/hands-on-with-a-144-core-processor https://www.rs-online.com/designspark/hands-on-with-a-144-co...
- pdimitar 8y ago> What needs to die is the belief that you can rely on just software (i.e. memory safety) for isolation between trusted and untrusted code[1]. That is definitely true. We need a bit more security in the hardware. I am not saying we should go back to 300MHz CPUs just for the sake of security but surely there's a spectrum where Spectre / Meltdown / Rowhammer attacks are impossible and we still have very fast processors and RAM.
- gpderetta 8y agoMeltdown for sure. Spectre not so much. Note that 2+ decade old 200mhz pentium pro is very likely as vulnerable to spectre as the current generation. Frequency doesn't really factor in it.
- pdimitar 8y agoI wasn't clear. I meant that we shouldn't take huge performance hits for having some security at the hardware level.
- gpderetta 8y agoMeltdown shouldn't hopefully be an issue, as there are high performance CPUs which are immune. A proper Spectre V2 would probably require partitioning the jump prediction tables by process; in the meantime flushing seem to work (although at a non trivial performance cost). The issue is the basic spectre v1 vulnerability, the bound check bypass, is so firmly rooted in the basic mechanism of speculative execution that it is very unlikely to be fixable. But it only affect,as a first approximation, processes that try to execute code with different trust domains in the same address space. The existing software mitigations appear to help, address spac separation is probably a safer long term solution. There are simple in order CPUs that have a small speculation window and are, in practice not affected, but it is not a given that an high performance in order core is safe. Denver (an inorder VLIW design) is affected; I can't find anything about power6 (the last inorder CPU competitive with OoO).
- 8y ago