Hi Simon,
I have some comments.
> 1. Make all calls to GET_THIS_PADDR before purging the DTR for PERCPU.
> This is because the kernel registers are saved in per_cpu data to allow
> the real kernel registers to be used by domains.
I did not notice that ia64_do_tlb_purge is called from ia64_jump_to_sal.
SET_PER_CPU_DATA() is called in ia64_os_mca_dispatch that calls
ia64_do_tlb_purge, because SET_PER_CPU_DATA() has to be called once.
> 2. Wrap the purging of the DTR for PERCPU in #ifdef XEN as Linux
> doesn't seem to do this (any more?)
The following lines are found in Linux tree.
arch/ia64/kernel/mca_asm.S:
104 // 2. Purge DTR for PERCPU data.
105 movl r16=PERCPU_ADDR
106 mov r18=PERCPU_PAGE_SHIFT<<2
107 ;;
108 ptr.d r16,r18
109 ;;
110 srlz.d
111 ;;
Isn't it such a meaning?
> 4. Fix VHPT purging code to treat what is at the IA64_DOMAIN_FLAGS_OFFSET
> offset as a litteral value (flag) rather than a pointer to a literal.
Suppose that,
struct domain *dom = 0xf000000007ccc080;
IA64_DOMAIN_FLAGS_OFFSET = 0x688
And I show how the value of r3 becomes since each instruction is executed.
@@ -289,25 +304,24 @@ ia64_do_tlb_purge:
ld8 r3=[r3];; // r3=0xf000000007ccc080 virtual address of dom
dep r3=0,r3,60,4;; // r3=0x0000000007ccc080 physical address of dom
add r3=IA64_DOMAIN_FLAGS_OFFSET,r3;; // r3=0x0000000007ccc708 physical
address of dom->arch.flags
ld8 r3=[r3];; // r3=2 load the value of domain->arch.flags
So 'r3=[r3]' is necessary.
Thanks,
KAZ
From: Simon Horman <horms@xxxxxxxxxxxx>
Subject: [Xen-ia64-devel] [patch 07/14] Kexec: Fix ia64_do_tlb_purge so that it
works with XEN
Date: Wed, 12 Sep 2007 17:08:52 +0900
> Fix ia64_do_tlb_purge, its broken in too many ways
>
> 1. Make all calls to GET_THIS_PADDR before purging the DTR for PERCPU.
> This is because the kernel registers are saved in per_cpu data to allow
> the real kernel registers to be used by domains.
>
> 2. Wrap the purging of the DTR for PERCPU in #ifdef XEN as Linux
> doesn't seem to do this (any more?)
>
> 3. Use the per_cpu variable to derive CURRENT_STACK_OFFSET rather
> than reading it from a kernel register. See 1) for explanation
> of why.
>
> 4. Fix VHPT purging code to treat what is at the IA64_DOMAIN_FLAGS_OFFSET
> offset as a litteral value (flag) rather than a pointer to a literal.
>
> 5. In the VHPT pruning code, don't use r25 as ia64_jump_to_sal,
> which branches to ia64_do_tlb_purge expects r25 to be preserved.
> There seems no reason not to use r2 as per the other purges
> done in ia64_do_tlb_purge. Furthermore use r16 and r18 instead
> of r20 and r24 for consistency reasons.
>
> 6. Move __va_ul(vcpu_vhpt_maddr(v)) comment outside of
> #if VHPT_ENABLED as it also applies to code further down that
> is outside the #if
>
> 7. There is no 7th change
>
> Cc: Tristan Gingold <tgingold@xxxxxxx>,
> Cc: Yutaka Ezaki <yutaka.ezaki@xxxxxxxxxxxxxx>,
> Cc: Masaki Kanno <kanno.masaki@xxxxxxxxxxxxxx>,
> Cc: Kazuhiro Suzuki <kaz@xxxxxxxxxxxxxx>,
> Signed-off-by: Simon Horman <horms@xxxxxxxxxxxx>
>
> Index: xen-unstable.hg/xen/arch/ia64/linux-xen/mca_asm.S
> ===================================================================
> --- xen-unstable.hg.orig/xen/arch/ia64/linux-xen/mca_asm.S 2007-07-11
> 12:05:33.000000000 +0900
> +++ xen-unstable.hg/xen/arch/ia64/linux-xen/mca_asm.S 2007-07-11
> 13:13:56.000000000 +0900
> @@ -195,6 +195,10 @@
> */
>
> ia64_do_tlb_purge:
> +#ifdef XEN
> + // This needs to be called in order for GET_THIS_PADDR to work
> + SET_PER_CPU_DATA();;
> +#endif
> #define O(member) IA64_CPUINFO_##member##_OFFSET
>
> GET_THIS_PADDR(r2, cpu_info) // load phys addr of cpu_info into r2
> @@ -244,6 +248,7 @@ ia64_do_tlb_purge:
> ;;
> srlz.d
> ;;
> +#ifdef XEN
> // 2. Purge DTR for PERCPU data.
> movl r16=PERCPU_ADDR
> mov r18=PERCPU_PAGE_SHIFT<<2
> @@ -252,6 +257,7 @@ ia64_do_tlb_purge:
> ;;
> srlz.d
> ;;
> +#endif
> // 3. Purge ITR for PAL code.
> GET_THIS_PADDR(r2, ia64_mca_pal_base)
> ;;
> @@ -263,7 +269,16 @@ ia64_do_tlb_purge:
> srlz.i
> ;;
> // 4. Purge DTR for stack.
> +#ifdef XEN
> + // Kernel registers are saved in a per_cpu cpu_kr_ia64_t
> + // to allow the kernel registers themselves to be used by domains.
> + GET_THIS_PADDR(r2, cpu_kr);;
> + add r2=IA64_KR_CURRENT_STACK_OFFSET,r2
> + ;;
> + ld8 r16=[r2]
> +#else
> mov r16=IA64_KR(CURRENT_STACK)
> +#endif
> ;;
> shl r16=r16,IA64_GRANULE_SHIFT
> movl r19=PAGE_OFFSET
> @@ -277,8 +292,8 @@ ia64_do_tlb_purge:
> ;;
> #ifdef XEN
> // 5. VHPT
> + // r2 = __va_ul(vcpu_vhpt_maddr(v));
> #if VHPT_ENABLED
> - // r25 = __va_ul(vcpu_vhpt_maddr(v));
> GET_THIS_PADDR(r2,cpu_kr);;
> add r2=IA64_KR_CURRENT_OFFSET,r2;;
> ld8 r2=[r2];;
> @@ -289,25 +304,24 @@ ia64_do_tlb_purge:
> ld8 r3=[r3];;
> dep r3=0,r3,60,4;; // virtual to physical
> add r3=IA64_DOMAIN_FLAGS_OFFSET,r3;;
> - ld8 r3=[r3];;
> and r3=HAS_PERVCPU_VHPT_MASK,r3;;
> cmp.eq p6,p0=r3,r0;;
> (p6) br.cond.sptk .not_pervcpu_vhpt
> add r2=IA64_VCPU_VHPT_MADDR_OFFSET,r2;;
> dep r2=0,r2,60,4;; // virtual to physical
> ld8 r2=[r2];;
> - dep r25=-1,r2,60,4;; // physical to virtual
> + dep r2=-1,r2,60,4;; // physical to virtual
> br.sptk .percpu_vhpt_done
> #endif
> .not_pervcpu_vhpt:
> GET_THIS_PADDR(r2, vhpt_paddr);;
> ld8 r2=[r2];;
> - dep r25=-1,r2,60,4;; // physical to virtual
> + dep r2=-1,r2,60,4;; // physical to virtual
> .percpu_vhpt_done:
> - dep r20=0,r25,0,IA64_GRANULE_SHIFT
> - mov r24=IA64_GRANULE_SHIFT<<2
> + dep r16=0,r2,0,IA64_GRANULE_SHIFT
> + mov r18=IA64_GRANULE_SHIFT<<2
> ;;
> - ptr.d r20,r24
> + ptr.d r16,r18
> ;;
> srlz.d
> ;;
>
> --
>
> --
> Horms
> H: http://www.vergenet.net/~horms/
> W: http://www.valinux.co.jp/en/
>
>
> _______________________________________________
> Xen-ia64-devel mailing list
> Xen-ia64-devel@xxxxxxxxxxxxxxxxxxx
> http://lists.xensource.com/xen-ia64-devel
_______________________________________________
Xen-ia64-devel mailing list
Xen-ia64-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-ia64-devel
|
Home