On Mon, Aug 25, 2025 at 11:40:48AM +0200, Christophe Leroy wrote:
> Hi Gabriel,
>
> Le 25/08/2025 à 11:04, Gabriel Paubert a écrit :
> > [Vous ne recevez pas souvent de courriers de paub...@iram.es. D?couvrez
> > pourquoi ceci est important ?
> > https://urldefense.com/v3/__https://aka.ms/LearnAboutSenderIdentification__;!!D9dNQwwGXtA!QUcSIXoDBBj9wAtcyQ-z3nPEAj-RnJpPgYwjOeb6LZWLejdLzq4uYsPMecQuK5Qy3147APjCNc-hcXGT71XuBh1AJI2M$
> > ]
> >
> > Hi Christophe,
> >
> > On Fri, Aug 22, 2025 at 11:58:06AM +0200, Christophe Leroy wrote:
> > > Masked user access avoids the address/size verification by access_ok().
> > > Allthough its main purpose is to skip the speculation in the
> > > verification of user address and size hence avoid the need of spec
> > > mitigation, it also has the advantage of reducing the amount of
> > > instructions required so it even benefits to platforms that don't
> > > need speculation mitigation, especially when the size of the copy is
> > > not know at build time.
> > >
> > > So implement masked user access on powerpc. The only requirement is
> > > to have memory gap that faults between the top user space and the
> > > real start of kernel area.
> > >
> > > On 64 bits platforms the address space is divided that way:
> > >
> > > 0xffffffffffffffff +------------------+
> > > | |
> > > | kernel space |
> > > | |
> > > 0xc000000000000000 +------------------+ <== PAGE_OFFSET
> > > |//////////////////|
> > > |//////////////////|
> > > 0x8000000000000000 |//////////////////|
> > > |//////////////////|
> > > |//////////////////|
> > > 0x0010000000000000 +------------------+ <== TASK_SIZE_MAX
> > > | |
> > > | user space |
> > > | |
> > > 0x0000000000000000 +------------------+
> > >
> > > Kernel is always above 0x8000000000000000 and user always
> > > below, with a gap in-between. It leads to a 4 instructions sequence:
> > >
> > > 80: 7c 69 1b 78 mr r9,r3
> > > 84: 7c 63 fe 76 sradi r3,r3,63
> > > 88: 7d 29 18 78 andc r9,r9,r3
> > > 8c: 79 23 00 4c rldimi r3,r9,0,1
> > >
> > > This sequence leaves r3 unmodified when it is below 0x8000000000000000
> > > and clamps it to 0x8000000000000000 if it is above.
> > >
> >
> > This comment looks wrong: the second instruction converts r3 to a
> > replicated sign bit of the address ((addr>0)?0:-1) if treating the
> > address as signed. After that the code only modifies the MSB of r3. So I
> > don't see how r3 could be unchanged from the original value...
>
> Unless I'm missing something, the above rldimi leaves the MSB of r3
> unmodified and replaces all other bits by the same in r9.
>
> This is the code generated by GCC for the following:
>
> unsigned long mask = (unsigned long)((long)addr >> 63);
>
> addr = ((addr & ~mask) & (~0UL >> 1)) | (mask & (1UL << 63));
>
>
> >
> > OTOH, I believe the following 3 instructions sequence would work,
> > input address (a) in r3, scratch value (tmp) in r9, both intptr_t:
> >
> > sradi r9,r3,63 ; tmp = (a >= 0) ? 0L : -1L;
> > andc r3,r3,r9 ; a = a & ~tmp; (equivalently a = (a >= 0) ? a : 0)
> > rldimi r3,r9,0,1 ; copy MSB of tmp to MSB of a
> >
> > But maybe I goofed...
> >
>
> From my understanding of rldimi, your proposed code would:
> - Keep r3 unmodified when it is above 0x8000000000000000
> - Set r3 to 0x7fffffffffffffff when it is below 0x8000000000000000
>
> Extract of ppc64 ABI:
>
> rldimi RA,RS,SH,MB
>
> The contents of register RS are rotated 64 left SH bits.
> A mask is generated having 1-bits from bit MB
> through bit 63− SH and 0-bits elsewhere. The rotated
> data are inserted into register RA under control of the
> generated mask.
Sorry, you are right, I got the polarity of the mask reversed in my
head.
Once again I may goof, but I believe that the following sequence
would work:
sradi r9,r3,63
andc r3,r3,r9
rldimi r3,r9,63,0 ; insert LSB of r9 into MSB of R3
Cheers,
Gabriel
