Hi David,

On Wed, Dec 04, 2019 at 11:14:08AM +0000, David Laight wrote:

From: Paul Burton

...

Sent: 03 December 2019 20:50 Our FPU emulator currently uses __get_user() & __put_user() to perform emulated loads & stores. This is problematic because __get_user() & __put_user() are only suitable for naturally aligned memory accesses, and the address we're accessing is entirely under the control of userland.

This allows userland to cause a kernel panic by simply performing an unaligned floating point load or store - the kernel will handle the address error exception by attempting to emulate the instruction, and in the process it may generate another address error exception itself. This time the exception is taken with EPC pointing at the kernels FPU emulation code, and we hit a die_if_kernel() in emulate_load_store_insn().

Won't this be true of almost all code that uses get_user() and put_user() (with or without the leading __).

Only if the address being accessed is under the control of userland to the extent that it can create an unaligned address. You're right that may be more widespread though; it needs checking...

We used to have separate get_user_unaligned() & put_user_unaligned() which would suggest that it's expected that get_user() & put_user() require their accesses be aligned, but they were removed by commit 3170d8d226c2 ("kill {__,}{get,put}_user_unaligned()") in v4.13.

But perhaps we should just take the second AdEL exception & recover via the fixups table. We definitely don't right now... Needs further investigation...

...

Fix this up by using __copy_from_user() instead of __get_user() and __copy_to_user() instead of __put_user(). These replacements will handle arbitrary alignment without problems.

They'll also kill performance.....

Sure they're heavier, but if you're hitting the FPU emulator you're already slow - trapping to the kernel for instruction emulation is hardly a hot path. If you care about performance at all then this is already a code path to avoid at all costs.

Thanks, Paul

On Wed, 4 Dec 2019, David Laight wrote:

...

We used to have separate get_user_unaligned() & put_user_unaligned() which would suggest that it's expected that get_user() & put_user() require their accesses be aligned, but they were removed by commit 3170d8d226c2 ("kill {__,}{get,put}_user_unaligned()") in v4.13.

But perhaps we should just take the second AdEL exception & recover via the fixups table. We definitely don't right now... Needs further investigation...

get/put_user can fault because the user page is absent (etc). So there must be code to 'expect' a fault on those instructions.

As I recall we only emulate unaligned accesses with a subset of integer load/store instructions (and then only if TIF_FIXADE is set, which is the default), and never with FP load/store instructions. Consequently I see no point in doing this in the FP emulator either and I think these ought to just send SIGBUS instead. Otherwise you'll end up with user code that works differently depending on whether the FP hardware is real or emulated, which is really bad.

FWIW,

Maciej

Hi Paul,

Sorry to take so long; it took me a while to track down the discussion I had in mind, and I was quite busy too. Also greetings from linux.conf.au!

...

As I recall we only emulate unaligned accesses with a subset of integer load/store instructions (and then only if TIF_FIXADE is set, which is the default), and never with FP load/store instructions. Consequently I see no point in doing this in the FP emulator either and I think these ought to just send SIGBUS instead. Otherwise you'll end up with user code that works differently depending on whether the FP hardware is real or emulated, which is really bad.

That might simplify things here, but it's incorrect. I'm fairly certain the intent is that emulate_load_store_insn() handles all non-FP loads & stores (though looking at it we're missing some instructions added in r6). More importantly though we've been emulating FP loads & stores since v3.10 which introduced the change alongside microMIPS support in commit 102cedc32a6e ("MIPS: microMIPS: Floating point support."). The commit contains no description of why, and I'm not aware of any reason microMIPS specifically would need this so I suspect that commit bundled this change for no good reason...

See the thread of discussion starting from this submission:

https://www.linux-mips.org/cgi-bin/mesg.cgi?a=linux-mips&i=20120615234641.6938B58FE7C%40mail.viric.name

and in particular Ralf's response (not referred directly due to the monthly archive rollover):

https://www.linux-mips.org/cgi-bin/mesg.cgi?a=linux-mips&i=20120731134001.GA14151%40linux-mips.org

I think Ralf's argument still stands and I find it regrettable that an unwanted feature was sneaked in with a trick along with a submission supposed to only add a different, unrelated feature.

I can't even track down a public submission/review of the change you refer, which is not how things are supposed to work with Linux! And neither the `Signed-off-by' tags help figuring out what the route of the change was to get there upstream. At that time there was supposed to be Ralf's tag there, as it was him who was the sole port maintainer.

It's also worth noting that some hardware will handle unaligned FP loads/stores, which means having the emulator reject them will result in more of a visible difference to userland. ie. on some hardware they'll work just fine, but on some you'd get SIGBUS. So I do think emulating them makes some sense - just as for non-FP loads & stores it lets userland not care whether the hardware will handle them, so long as it's not performance critical code. If we knew that had never been used then perhaps we could enforce the alignment requirement (and maybe that's what you recall doing), but since we've been emulating them for the past 6 years it's too late for that now.

I don't think it's ever too late to remove a broken feature that everyone knows is not a part of the architecture and the emulation of which has never been advertised as a part of the Linux ABI either. You just don't make it a part of the ABI when you sneak in a feature without a proper review, we do not accept the fait accompli method in Linux development.

The presence of unaligned FP data is a sign of user code breakage and whoever caused that breakage will best know that ASAP by seeing their program trap (they can emulate the trap in their software by installing a suitable signal handler if they are so desperate to have unaligned FP data handled).

So I think that not only the new submission should be rejected, but also parts of commit 102cedc32a6e ("MIPS: microMIPS: Floating point support.") reverted that are not a part of actual microMIPS support. If someone relied on it by accident or ignorance, they'll simply have to adjust.

Maciej