Re: [PATCH V3 09/30] arm/mm: Enable ARCH_HAS_VM_GET_PAGE_PROT

From: Geert Uytterhoeven
Date: Wed Mar 02 2022 - 05:05:53 EST

Hi Anshuman,

On Wed, Mar 2, 2022 at 10:51 AM Anshuman Khandual
<anshuman.khandual@xxxxxxx> wrote:
> On 3/2/22 12:35 PM, Christophe Leroy wrote:
> > Le 02/03/2022 à 04:22, Anshuman Khandual a écrit :
> >> On 3/1/22 1:46 PM, Christophe Leroy wrote:
> >>> Le 01/03/2022 à 01:31, Russell King (Oracle) a écrit :
> >>>> On Tue, Mar 01, 2022 at 05:30:41AM +0530, Anshuman Khandual wrote:
> >>>>> On 2/28/22 4:27 PM, Russell King (Oracle) wrote:
> >>>>>> On Mon, Feb 28, 2022 at 04:17:32PM +0530, Anshuman Khandual wrote:
> >>>>>>> This defines and exports a platform specific custom vm_get_page_prot() via
> >>>>>>> subscribing ARCH_HAS_VM_GET_PAGE_PROT. Subsequently all __SXXX and __PXXX
> >>>>>>> macros can be dropped which are no longer needed.
> >>>>>>
> >>>>>> What I would really like to know is why having to run _code_ to work out
> >>>>>> what the page protections need to be is better than looking it up in a
> >>>>>> table.
> >>>>>>
> >>>>>> Not only is this more expensive in terms of CPU cycles, it also brings
> >>>>>> additional code size with it.
> >>>>>>
> >>>>>> I'm struggling to see what the benefit is.
> >>>>>
> >>>>> Currently vm_get_page_prot() is also being _run_ to fetch required page
> >>>>> protection values. Although that is being run in the core MM and from a
> >>>>> platform perspective __SXXX, __PXXX are just being exported for a table.
> >>>>> Looking it up in a table (and applying more constructs there after) is
> >>>>> not much different than a clean switch case statement in terms of CPU
> >>>>> usage. So this is not more expensive in terms of CPU cycles.
> >>>>
> >>>> I disagree.
> >>>
> >>> So do I.
> >>>
> >>>>
> >>>> However, let's base this disagreement on some evidence. Here is the
> >>>> present 32-bit ARM implementation:
> >>>>
> >>>> 00000048 <vm_get_page_prot>:
> >>>> 48: e200000f and r0, r0, #15
> >>>> 4c: e3003000 movw r3, #0
> >>>> 4c: R_ARM_MOVW_ABS_NC .LANCHOR1
> >>>> 50: e3403000 movt r3, #0
> >>>> 50: R_ARM_MOVT_ABS .LANCHOR1
> >>>> 54: e7930100 ldr r0, [r3, r0, lsl #2]
> >>>> 58: e12fff1e bx lr
> >>>>
> >>>> That is five instructions long.
> >>>
> >>> On ppc32 I get:
> >>>
> >>> 00000094 <vm_get_page_prot>:
> >>> 94: 3d 20 00 00 lis r9,0
> >>> 96: R_PPC_ADDR16_HA .data..ro_after_init
> >>> 98: 54 84 16 ba rlwinm r4,r4,2,26,29
> >>> 9c: 39 29 00 00 addi r9,r9,0
> >>> 9e: R_PPC_ADDR16_LO .data..ro_after_init
> >>> a0: 7d 29 20 2e lwzx r9,r9,r4
> >>> a4: 91 23 00 00 stw r9,0(r3)
> >>> a8: 4e 80 00 20 blr
> >>>
> >>>
> >>>>
> >>>> Please show that your new implementation is not more expensive on
> >>>> 32-bit ARM. Please do so by building a 32-bit kernel, and providing
> >>>> the disassembly.
> >>>
> >>> With your series I get:
> >>>
> >>> 00000000 <vm_get_page_prot>:
> >>> 0: 3d 20 00 00 lis r9,0
> >>> 2: R_PPC_ADDR16_HA .rodata
> >>> 4: 39 29 00 00 addi r9,r9,0
> >>> 6: R_PPC_ADDR16_LO .rodata
> >>> 8: 54 84 16 ba rlwinm r4,r4,2,26,29
> >>> c: 7d 49 20 2e lwzx r10,r9,r4
> >>> 10: 7d 4a 4a 14 add r10,r10,r9
> >>> 14: 7d 49 03 a6 mtctr r10
> >>> 18: 4e 80 04 20 bctr
> >>> 1c: 39 20 03 15 li r9,789
> >>> 20: 91 23 00 00 stw r9,0(r3)
> >>> 24: 4e 80 00 20 blr
> >>> 28: 39 20 01 15 li r9,277
> >>> 2c: 91 23 00 00 stw r9,0(r3)
> >>> 30: 4e 80 00 20 blr
> >>> 34: 39 20 07 15 li r9,1813
> >>> 38: 91 23 00 00 stw r9,0(r3)
> >>> 3c: 4e 80 00 20 blr
> >>> 40: 39 20 05 15 li r9,1301
> >>> 44: 91 23 00 00 stw r9,0(r3)
> >>> 48: 4e 80 00 20 blr
> >>> 4c: 39 20 01 11 li r9,273
> >>> 50: 4b ff ff d0 b 20 <vm_get_page_prot+0x20>
> >>>
> >>>
> >>> That is definitely more expensive, it implements a table of branches.
> >>
> >> Okay, will split out the PPC32 implementation that retains existing
> >> table look up method. Also planning to keep that inside same file
> >> (arch/powerpc/mm/mmap.c), unless you have a difference preference.
> >
> > My point was not to get something specific for PPC32, but to amplify on
> > Russell's objection.
> >
> > As this is bad for ARM and bad for PPC32, do we have any evidence that
> > your change is good for any other architecture ?
> >
> > I checked PPC64 and there is exactly the same drawback. With the current
> > implementation it is a small function performing table read then a few
> > adjustment. After your change it is a bigger function implementing a
> > table of branches.
>
> I am wondering if this would not be the case for any other switch case
> statement on the platform ? Is there something specific/different just
> on vm_get_page_prot() implementation ? Are you suggesting that switch
> case statements should just be avoided instead ?
>
> >
> > So, as requested by Russell, could you look at the disassembly for other
> > architectures and show us that ARM and POWERPC are the only ones for
> > which your change is not optimal ?
>
> But the primary purpose of this series is not to guarantee optimized
> code on platform by platform basis, while migrating from a table based
> look up method into a switch case statement.
>
> But instead, the purposes is to remove current levels of unnecessary
> abstraction while converting a vm_flags access combination into page
> protection. The switch case statement for platform implementation of
> vm_get_page_prot() just seemed logical enough. Christoph's original
> suggestion patch for x86 had the same implementation as well.
>
> But if the table look up is still better/preferred method on certain
> platforms like arm or ppc32, will be happy to preserve that.

I doubt the switch() variant would give better code on any platform.

What about using tables everywhere, using designated initializers
to improve readability?

Gr{oetje,eeting}s,

Geert

--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@xxxxxxxxxxxxxx

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
-- Linus Torvalds