[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-Id: <fcc94e7f347c3759a1698444239a7250b22cde7d.camel@au1.ibm.com>
Date: Thu, 22 Aug 2019 10:27:07 +1000
From: "Alastair D'Silva" <alastair@....ibm.com>
To: Christophe Leroy <christophe.leroy@....fr>,
Segher Boessenkool <segher@...nel.crashing.org>,
Michael Ellerman <mpe@...erman.id.au>
Cc: linux-kernel@...r.kernel.org, linuxppc-dev@...ts.ozlabs.org
Subject: RE: [RFC PATCH] powerpc: Convert ____flush_dcache_icache_phys() to C
On Wed, 2019-08-21 at 22:27 +0200, Christophe Leroy wrote:
>
> Le 20/08/2019 à 06:36, Alastair D'Silva a écrit :
> > On Fri, 2019-08-16 at 15:52 +0000, Christophe Leroy wrote:
>
> [...]
>
> >
> > Thanks Christophe,
> >
> > I'm trying a somewhat different approach that requires less
> > knowledge
> > of assembler. Handling of CPU_FTR_COHERENT_ICACHE is outside this
> > function. The code below is not a patch as my tree is a bit messy,
> > sorry:
>
> Can we be 100% sure that GCC won't add any code accessing some
> global
> data or stack while the Data MMU is OFF ?
>
> Christophe
>
+mpe
I'm not sure how we would go about making such a guarantee, but I've
tied every variable used to a register and addr is passed in a
register, so there is no stack usage, and every call in there only
operates on it's operands.
The calls to the inline cache helpers (for the PPC32 case) are all
constants, so I can't see a reasonable scenario where there would be a
function call and reordered to after the DR bit is turned off, but I
guess if we want to be paranoid, we could always add an mb() call
before the DR bit is manipulated to prevent the compiler from
reordering across the section where the data MMU is disabled.
>
> > /**
> > * flush_dcache_icache_phys() - Flush a page by it's physical
> > address
> > * @addr: the physical address of the page
> > */
> > static void flush_dcache_icache_phys(unsigned long addr)
> > {
> > register unsigned long msr;
> > register unsigned long dlines = PAGE_SIZE >> l1_dcache_shift();
> > register unsigned long dbytes = l1_dcache_bytes();
> > register unsigned long ilines = PAGE_SIZE >> l1_icache_shift();
> > register unsigned long ibytes = l1_icache_bytes();
> > register unsigned long i;
> > register unsigned long address = addr;
> >
> > /*
> > * Clear the DR bit so that we operate on physical
> > * rather than virtual addresses
> > */
> > msr = mfmsr();
> > mtmsr(msr & ~(MSR_DR));
> >
> > /* Write out the data cache */
> > for (i = 0; i < dlines; i++, address += dbytes)
> > dcbst((void *)address);
> >
> > /* Invalidate the instruction cache */
> > address = addr;
> > for (i = 0; i < ilines; i++, address += ibytes)
> > icbi((void *)address);
> >
> > mtmsr(msr);
> > }
> >
> > void test_flush_phys(unsigned long addr)
> > {
> > flush_dcache_icache_phys(addr);
> > }
> >
> >
> > This gives the following assembler (using pmac32_defconfig):
> > 000003cc <test_flush_phys>:
> > 3cc: 94 21 ff f0 stwu r1,-16(r1)
> > 3d0: 7d 00 00 a6 mfmsr r8
> > 3d4: 55 09 07 34 rlwinm r9,r8,0,28,26
> > 3d8: 7d 20 01 24 mtmsr r9
> > 3dc: 39 20 00 80 li r9,128
> > 3e0: 7d 29 03 a6 mtctr r9
> > 3e4: 39 43 10 00 addi r10,r3,4096
> > 3e8: 7c 69 1b 78 mr r9,r3
> > 3ec: 7c 00 48 6c dcbst 0,r9
> > 3f0: 39 29 00 20 addi r9,r9,32
> > 3f4: 42 00 ff f8 bdnz 3ec <test_flush_phys+0x20>
> > 3f8: 7c 00 1f ac icbi 0,r3
> > 3fc: 38 63 00 20 addi r3,r3,32
> > 400: 7f 8a 18 40 cmplw cr7,r10,r3
> > 404: 40 9e ff f4 bne cr7,3f8 <test_flush_phys+0x2c>
> > 408: 7d 00 01 24 mtmsr r8
> > 40c: 38 21 00 10 addi r1,r1,16
> > 410: 4e 80 00 20 blr
> >
> >
--
Alastair D'Silva
Open Source Developer
Linux Technology Centre, IBM Australia
mob: 0423 762 819
Powered by blists - more mailing lists