Slashdot Mirror
Google Says CPU Patches Cause 'Negligible Impact On Performance' With New 'Retpoline' Technique (theverge.com)
In a post on Google's Online Security Blog, two engineers described a novel chip-level patch that has been deployed across the company's entire infrastructure, resulting in only minor declines in performance in most cases. "The company has also posted details of the new technique, called Retpoline, in the hopes that other companies will be able to follow the same technique," reports The Verge. "If the claims hold, it would mean Intel and others have avoided the catastrophic slowdowns that many had predicted." From the report: "There has been speculation that the deployment of KPTI causes significant performance slowdowns," the post reads, referring to the company's "Kernel Page Table Isolation" technique. "Performance can vary, as the impact of the KPTI mitigations depends on the rate of system calls made by an application. On most of our workloads, including our cloud infrastructure, we see negligible impact on performance." "Of course, Google recommends thorough testing in your environment before deployment," the post continues. "We cannot guarantee any particular performance or operational impact."
Notably, the new technique only applies to one of the three variants involved in the new attacks. However, it's the variant that is arguably the most difficult to address. The other two vulnerabilities -- "bounds check bypass" and "rogue data cache load" -- would be addressed at the program and operating system level, respectively, and are unlikely to result in the same system-wide slowdowns.
Notably, the new technique only applies to one of the three variants involved in the new attacks. However, it's the variant that is arguably the most difficult to address. The other two vulnerabilities -- "bounds check bypass" and "rogue data cache load" -- would be addressed at the program and operating system level, respectively, and are unlikely to result in the same system-wide slowdowns.
Or just Buy AMD & get no slow down with more pci-e lanes.
Geez... You make it sound like this is the first ever time someone has had to write a software patch to bypass a hardware flaw. Driver developers have had to come up with clever workarounds to hardware defects since the the dawn of computing.
These Intel firmware fixes are just going to become part of yet another security update that will be required to keep systems secure.
Because it doesn't make sense: Intel has a KNOWN UNFIXABLE FLAW in Meltdown. It cannot be fixed. You are saying "don't switch to AMD because they might have a major flaw too at some point". Meltdown is a much larger problem than Spectre is.
More Intel spin. Spectre and Meltdown are different flaws. Meltdown is severe and unfixable and only affects Intel.
Google has created "retpoline", a technique which allows an indirect branch (e.g. a vtable call) to occur in a way that effectively disables speculative execution by isolating branch target prediction into a safe effectless loop. This addresses Variant 2 (aka Spectre).
Retpoline does not depend on or assist a CPU or an OS patch: it is done purely at the software level, per-app, by a compiler. There is no simple OS-wide patch.
Google says a retpoline call has performance "within cycles" of a regular old mispredicted branch. The zero-cost predictions we're used to are a thing of the past, because it effectively forces misprediction. I'd be curious to see a benchmark of an indirection-heavy platform like .NET.
This does not help address or optimize Variant 3, which is what the big kernel patches for Page Table Isolation are needed for. So, your I/O-dependent apps like databases are still going to take a big performance hit. Nor does it address Variant 1.
Google's technique ... has a small performance hit but much smaller than KPTI.
Keep in mind Google's technique (retpoline) is not an alternative to KPTI. Retpoline addresses Variant 2. KPTI addresses Variant 3. Both are required.
Correction, they speculated that they were able to get AMD chips to do that. Their toy attack (within process) succeeded showing AMD chips will do speculative ordering. No actual security risk there, beause processes can read their own memory.
BUT, they didn't know for a fact why they didn't succeed in attacking the kernel.
We've now had statements from AMD (after the paper was released) - namely, that permission bits are checked BEFORE issuing instructions so kernel memory isn't readable, even speculatively.
So.. .yeah, remember the paper is only what they think could be happening.