RE: [RFD] x86/split_lock: Request to Intel

[hpa]

On October 18, 2019 3:45:14 AM PDT, David Laight <David.Laight@xxxxxxxxxx> wrote:
>From: Luck, Tony
>> Sent: 18 October 2019 00:28
>...
>> 2) Fix set_bit() et. al. to not issue atomic operations that cross
>boundaries.
>> 
>> Fenghua had been pursuing option #1 in previous iterations. He found
>a few
>> more places with the help of the "grep" patterns suggested by David
>Laight.
>> So that path is up to ~8 patches now that do one of:
>> 	+ Change from u32 to u64
>> 	+ Force alignment with a union with a u64
>> 	+ Change to non-atomic (places that didn't need atomic)
>> 
>> Downside of strategy #1 is that people will add new misaligned cases
>in the
>> future. So this process has no defined end point.
>> 
>> Strategy #2 begun when I looked at the split-lock issue I saw that
>with a
>> constant bit argument set_bit() just does a "ORB" on the affected
>byte (i.e.
>> no split lock). Similar for clear_bit() and change_bit(). Changing
>code to also
>> do that for the variable bit case is easy.
>> 
>> test_and_clr_bit() needs more care, but luckily, we had Peter Anvin
>nearby
>> to give us a neat solution.
>
>Changing the x86-64 bitops to use 32bit memory cycles is trivial
>(provided you are willing to accept a limit of 2G bits).
>
>OTOH this only works because x86 is LE.
>On any BE systems passing an 'int []' to any of the bit-functions is so
>terribly
>wrong it is unbelievable.
>
>So changing the x86-64 bitops is largely papering over a crack.
>
>In essence any code that casts the argument to any of the bitops
>functions
>is almost certainly badly broken on BE systems.
>
>The x86 cpu features code is always LE.
>It probably ought to have a typedef for a union of long [] and int [].
>
>	David
>
>-
>Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes,
>MK1 1PT, UK
>Registration No: 1397386 (Wales)

One thing I suggested is that we should actually expose the violations at committee time either by wrapping them in macros using __alignof__ and/or make the kernel compile with -Wcast-align.

On x86 the btsl/btcl/btrl instructions can be used without limiting to 2Gbit of the address is computed, the way one does for plain and, or, etc. However, if the real toes for the arguments are exposed then or is possible to do better.

Finally, as far as bigendian is concerned: the problem Linux has on bigendian machines is that it tries to use littleendian bitmaps on bigendian machines: on bigendian machines, bit 0 is naturally the MSB. If your reaction is "but that is absurd", then you have just grokked why bigendian is fundamentally broken.
-- 
Sent from my Android device with K-9 Mail. Please excuse my brevity.