Message ID | cover.1410302383.git.geoff@infradead.org |
---|---|
State | New |
Headers | show |
On 10 September 2014 00:49, Geoff Levand <geoff@infradead.org> wrote: > Some of the macros defined in kvm_arm.h are useful in the exception vector > routines, but they are not compatible with the assembler. Change the > definition of ESR_EL2_ISS to be compatible. > > Fixes build errors like these when using kvm_arm.h in assembly > source files: > > Error: unexpected characters following instruction at operand 3 -- `add x0,x1,#((1U<<25)-1)' > > Signed-off-by: Geoff Levand <geoff@infradead.org> > --- > arch/arm64/include/asm/kvm_arm.h | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/arch/arm64/include/asm/kvm_arm.h b/arch/arm64/include/asm/kvm_arm.h > index cc83520..e0e7e64 100644 > --- a/arch/arm64/include/asm/kvm_arm.h > +++ b/arch/arm64/include/asm/kvm_arm.h > @@ -176,7 +176,7 @@ > #define ESR_EL2_EC_SHIFT (26) > #define ESR_EL2_EC (0x3fU << ESR_EL2_EC_SHIFT) > #define ESR_EL2_IL (1U << 25) > -#define ESR_EL2_ISS (ESR_EL2_IL - 1) > +#define ESR_EL2_ISS (0xffff) Don't you mean 0x1ffffff? And, there is a macro UL() for this purpose, so I suppose you could redefine ESR_EL2_IL as (UL(1) << 25) as well. I know it is not strictly the same thing, but it should be good enough as this is arm64 only > #define ESR_EL2_ISV_SHIFT (24) > #define ESR_EL2_ISV (1U << ESR_EL2_ISV_SHIFT) > #define ESR_EL2_SAS_SHIFT (22) > -- > 1.9.1 > > > > _______________________________________________ > linux-arm-kernel mailing list > linux-arm-kernel@lists.infradead.org > http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
On Wed, 2014-09-10 at 10:40 +0200, Ard Biesheuvel wrote: > On 10 September 2014 00:49, Geoff Levand <geoff@infradead.org> wrote: > > Some of the macros defined in kvm_arm.h are useful in the exception vector > > routines, but they are not compatible with the assembler. Change the > > definition of ESR_EL2_ISS to be compatible. > > > > Fixes build errors like these when using kvm_arm.h in assembly > > source files: > > > > Error: unexpected characters following instruction at operand 3 -- `add x0,x1,#((1U<<25)-1)' > > > > Signed-off-by: Geoff Levand <geoff@infradead.org> > > --- > > arch/arm64/include/asm/kvm_arm.h | 2 +- > > 1 file changed, 1 insertion(+), 1 deletion(-) > > > > diff --git a/arch/arm64/include/asm/kvm_arm.h b/arch/arm64/include/asm/kvm_arm.h > > index cc83520..e0e7e64 100644 > > --- a/arch/arm64/include/asm/kvm_arm.h > > +++ b/arch/arm64/include/asm/kvm_arm.h > > @@ -176,7 +176,7 @@ > > #define ESR_EL2_EC_SHIFT (26) > > #define ESR_EL2_EC (0x3fU << ESR_EL2_EC_SHIFT) > > #define ESR_EL2_IL (1U << 25) > > -#define ESR_EL2_ISS (ESR_EL2_IL - 1) > > +#define ESR_EL2_ISS (0xffff) > > Don't you mean 0x1ffffff? Hcalls have a 16 bit 'payload', the upper bits of the ISS field are specified as zero by the architecture so 0xffff is the same as 0x1ffffff. > And, there is a macro UL() for this purpose, so I suppose you could > redefine ESR_EL2_IL as (UL(1) << 25) as well. I know it is not > strictly the same thing, but it should be good enough as this is arm64 > only Sure that will be OK. The one other use of ESR_EL2_IL will promote the operation to unsigned long without ill effect. I'll prepare an updated patch. -Geoff
On Tue, Sep 09, 2014 at 11:49:04PM +0100, Geoff Levand wrote: > Add the new hcall HVC_CALL_FUNC that allows execution of a function at EL2. > During CPU reset the CPU must be brought to the exception level it had on > entry to the kernel. The HVC_CALL_FUNC hcall will provide the mechanism > needed for this exception level switch. > > Signed-off-by: Geoff Levand <geoff@infradead.org> > --- > arch/arm64/include/asm/virt.h | 11 +++++++++++ > arch/arm64/kernel/hyp-stub.S | 10 ++++++++++ > 2 files changed, 21 insertions(+) [...] > diff --git a/arch/arm64/kernel/hyp-stub.S b/arch/arm64/kernel/hyp-stub.S > index 9ab5f70..a21cf51 100644 > --- a/arch/arm64/kernel/hyp-stub.S > +++ b/arch/arm64/kernel/hyp-stub.S > @@ -75,7 +75,17 @@ el1_sync: > 1: cmp x10, #HVC_SET_VECTORS > b.ne 1f > msr vbar_el2, x0 > + b 2f > > +1: cmp x10, #HVC_CALL_FUNC > + b.ne 1f > + mov x29, lr > + mov lr, x0 > + mov x0, x1 > + mov x1, x2 > + mov x2, x3 > + blr lr > + mov lr, x29 Why are you clobbering x29? Will
On 10 September 2014 18:35, Geoff Levand <geoff@infradead.org> wrote: > On Wed, 2014-09-10 at 10:40 +0200, Ard Biesheuvel wrote: >> On 10 September 2014 00:49, Geoff Levand <geoff@infradead.org> wrote: >> > Some of the macros defined in kvm_arm.h are useful in the exception vector >> > routines, but they are not compatible with the assembler. Change the >> > definition of ESR_EL2_ISS to be compatible. >> > >> > Fixes build errors like these when using kvm_arm.h in assembly >> > source files: >> > >> > Error: unexpected characters following instruction at operand 3 -- `add x0,x1,#((1U<<25)-1)' >> > >> > Signed-off-by: Geoff Levand <geoff@infradead.org> >> > --- >> > arch/arm64/include/asm/kvm_arm.h | 2 +- >> > 1 file changed, 1 insertion(+), 1 deletion(-) >> > >> > diff --git a/arch/arm64/include/asm/kvm_arm.h b/arch/arm64/include/asm/kvm_arm.h >> > index cc83520..e0e7e64 100644 >> > --- a/arch/arm64/include/asm/kvm_arm.h >> > +++ b/arch/arm64/include/asm/kvm_arm.h >> > @@ -176,7 +176,7 @@ >> > #define ESR_EL2_EC_SHIFT (26) >> > #define ESR_EL2_EC (0x3fU << ESR_EL2_EC_SHIFT) >> > #define ESR_EL2_IL (1U << 25) >> > -#define ESR_EL2_ISS (ESR_EL2_IL - 1) >> > +#define ESR_EL2_ISS (0xffff) >> >> Don't you mean 0x1ffffff? > > Hcalls have a 16 bit 'payload', the upper bits of the ISS field > are specified as zero by the architecture so 0xffff is the same > as 0x1ffffff. > Even if HVC is currently the only exception we are taking in EL2 (is that the case btw?), it seems wrong to define this field in such a way that it (a) deviates from how the architecture specifies ESR_ELx.ISS and (b) may cause surprises once someone unsuspectingly starts and'ing his ESR values produced by another exception class with it, expecting the macro's value to reflect its name >> And, there is a macro UL() for this purpose, so I suppose you could >> redefine ESR_EL2_IL as (UL(1) << 25) as well. I know it is not >> strictly the same thing, but it should be good enough as this is arm64 >> only > > Sure that will be OK. The one other use of ESR_EL2_IL will promote > the operation to unsigned long without ill effect. I'll prepare an > updated patch. > > -Geoff > >
Hi Will, On Wed, 2014-09-10 at 18:07 +0100, Will Deacon wrote: > On Tue, Sep 09, 2014 at 11:49:04PM +0100, Geoff Levand wrote: > > Add the new hcall HVC_CALL_FUNC that allows execution of a function at EL2. > > During CPU reset the CPU must be brought to the exception level it had on > > entry to the kernel. The HVC_CALL_FUNC hcall will provide the mechanism > > needed for this exception level switch. > > > > Signed-off-by: Geoff Levand <geoff@infradead.org> > > --- > > arch/arm64/include/asm/virt.h | 11 +++++++++++ > > arch/arm64/kernel/hyp-stub.S | 10 ++++++++++ > > 2 files changed, 21 insertions(+) > > [...] > > > diff --git a/arch/arm64/kernel/hyp-stub.S b/arch/arm64/kernel/hyp-stub.S > > index 9ab5f70..a21cf51 100644 > > --- a/arch/arm64/kernel/hyp-stub.S > > +++ b/arch/arm64/kernel/hyp-stub.S > > @@ -75,7 +75,17 @@ el1_sync: > > 1: cmp x10, #HVC_SET_VECTORS > > b.ne 1f > > msr vbar_el2, x0 > > + b 2f > > > > +1: cmp x10, #HVC_CALL_FUNC > > + b.ne 1f > > + mov x29, lr > > + mov lr, x0 > > + mov x0, x1 > > + mov x1, x2 > > + mov x2, x3 > > + blr lr > > + mov lr, x29 > > Why are you clobbering x29? I can change this to x28, unless you can recommend another? -Geoff
On Wed, Sep 10, 2014 at 06:23:57PM +0100, Geoff Levand wrote: > On Wed, 2014-09-10 at 18:07 +0100, Will Deacon wrote: > > On Tue, Sep 09, 2014 at 11:49:04PM +0100, Geoff Levand wrote: > > > diff --git a/arch/arm64/kernel/hyp-stub.S b/arch/arm64/kernel/hyp-stub.S > > > index 9ab5f70..a21cf51 100644 > > > --- a/arch/arm64/kernel/hyp-stub.S > > > +++ b/arch/arm64/kernel/hyp-stub.S > > > @@ -75,7 +75,17 @@ el1_sync: > > > 1: cmp x10, #HVC_SET_VECTORS > > > b.ne 1f > > > msr vbar_el2, x0 > > > + b 2f > > > > > > +1: cmp x10, #HVC_CALL_FUNC > > > + b.ne 1f > > > + mov x29, lr > > > + mov lr, x0 > > > + mov x0, x1 > > > + mov x1, x2 > > > + mov x2, x3 > > > + blr lr > > > + mov lr, x29 > > > > Why are you clobbering x29? > > I can change this to x28, unless you can recommend another? How about something that's not callee saved? Will
On Tue, Sep 09, 2014 at 11:49:04PM +0100, Geoff Levand wrote: > To improve the readability of the el1_sync routine in hyp-stub.S replace the > numeric immediate values with preprocessor macros ESR_EL2_EC_SHIFT and > ESR_EL2_EC_HVC64. > > Signed-off-by: Geoff Levand <geoff@infradead.org> > --- > arch/arm64/kernel/hyp-stub.S | 7 +++++-- > 1 file changed, 5 insertions(+), 2 deletions(-) > > diff --git a/arch/arm64/kernel/hyp-stub.S b/arch/arm64/kernel/hyp-stub.S > index a272f33..2d960a9 100644 > --- a/arch/arm64/kernel/hyp-stub.S > +++ b/arch/arm64/kernel/hyp-stub.S > @@ -52,10 +52,13 @@ ENDPROC(__hyp_stub_vectors) > > .align 11 > > +#define ESR_EL2_EC_SHIFT 26 > +#define ESR_EL2_EC_HVC64 0x16 These exist in arch/arm64/include/asm/kvm_arm.h, no? If anything that should be folded into arch/arm64/include/asm/esr.h... Mark. > + > el1_sync: > mrs x1, esr_el2 > - lsr x1, x1, #26 > - cmp x1, #0x16 > + lsr x1, x1, #ESR_EL2_EC_SHIFT > + cmp x1, #ESR_EL2_EC_HVC64 > b.ne 2f // Not an HVC trap > cbz x0, 1f > msr vbar_el2, x0 // Set vbar_el2 > -- > 1.9.1 > > > > _______________________________________________ > linux-arm-kernel mailing list > linux-arm-kernel@lists.infradead.org > http://lists.infradead.org/mailman/listinfo/linux-arm-kernel >
On Wed, Sep 10, 2014 at 06:09:07PM +0100, Ard Biesheuvel wrote: > On 10 September 2014 18:35, Geoff Levand <geoff@infradead.org> wrote: > > On Wed, 2014-09-10 at 10:40 +0200, Ard Biesheuvel wrote: > >> On 10 September 2014 00:49, Geoff Levand <geoff@infradead.org> wrote: > >> > Some of the macros defined in kvm_arm.h are useful in the exception vector > >> > routines, but they are not compatible with the assembler. Change the > >> > definition of ESR_EL2_ISS to be compatible. > >> > > >> > Fixes build errors like these when using kvm_arm.h in assembly > >> > source files: > >> > > >> > Error: unexpected characters following instruction at operand 3 -- `add x0,x1,#((1U<<25)-1)' > >> > > >> > Signed-off-by: Geoff Levand <geoff@infradead.org> > >> > --- > >> > arch/arm64/include/asm/kvm_arm.h | 2 +- > >> > 1 file changed, 1 insertion(+), 1 deletion(-) > >> > > >> > diff --git a/arch/arm64/include/asm/kvm_arm.h b/arch/arm64/include/asm/kvm_arm.h > >> > index cc83520..e0e7e64 100644 > >> > --- a/arch/arm64/include/asm/kvm_arm.h > >> > +++ b/arch/arm64/include/asm/kvm_arm.h > >> > @@ -176,7 +176,7 @@ > >> > #define ESR_EL2_EC_SHIFT (26) > >> > #define ESR_EL2_EC (0x3fU << ESR_EL2_EC_SHIFT) > >> > #define ESR_EL2_IL (1U << 25) > >> > -#define ESR_EL2_ISS (ESR_EL2_IL - 1) > >> > +#define ESR_EL2_ISS (0xffff) > >> > >> Don't you mean 0x1ffffff? > > > > Hcalls have a 16 bit 'payload', the upper bits of the ISS field > > are specified as zero by the architecture so 0xffff is the same > > as 0x1ffffff. > > > > Even if HVC is currently the only exception we are taking in EL2 (is > that the case btw?), it seems wrong to define this field in such a way > that it > (a) deviates from how the architecture specifies ESR_ELx.ISS and > (b) may cause surprises once someone unsuspectingly starts and'ing his > ESR values produced by another exception class with it, expecting the > macro's value to reflect its name Agreed. A macro called ESR_EL2_ISS should return the ISS field, and nothing less. > >> And, there is a macro UL() for this purpose, so I suppose you could > >> redefine ESR_EL2_IL as (UL(1) << 25) as well. I know it is not > >> strictly the same thing, but it should be good enough as this is arm64 > >> only This sounds good to me. Mark.
Hi Geoff, The general approach looks good to me, using the HVC immediate makes this look far nicer to me. Hopefully Marc and Christoffer agree on that. That said, I have some comments on the mechanics below. On Tue, Sep 09, 2014 at 11:49:04PM +0100, Geoff Levand wrote: > To allow for additional hcalls to be defined and to make the arm64 hcall API > more consistent across exception vector routines change the hcall implementations > to use the ISS field of the ESR_EL2 register to specify the hcall type. > > The existing arm64 hcall implementations are limited in that they only allow > for two distinct hcalls; with the x0 register either zero, or not zero. Also, > the API of the hyp-stub exception vector routines and the KVM exception vector > routines differ; hyp-stub uses a non-zero value in x0 to implement > __hyp_set_vectors, whereas KVM uses it to implement kvm_call_hyp. > > Define three new preprocessor macros HVC_GET_VECTORS, HVC_SET_VECTORS and > HVC_KVM_CALL_HYP and to be used as hcall type specifiers and convert the > existing __hyp_get_vectors(), __hyp_set_vectors() and kvm_call_hyp() routines > to use these new macros when executing and HVC call. Also change the > corresponding hyp-stub and KVM el1_sync exception vector routines to use these > new macros. > > Signed-off-by: Geoff Levand <geoff@infradead.org> > --- > arch/arm64/include/asm/virt.h | 20 ++++++++++++++++++++ > arch/arm64/kernel/hyp-stub.S | 38 ++++++++++++++++++++++++++------------ > arch/arm64/kvm/hyp.S | 19 ++++++++++++------- > 3 files changed, 58 insertions(+), 19 deletions(-) > > diff --git a/arch/arm64/include/asm/virt.h b/arch/arm64/include/asm/virt.h > index 7a5df52..894fe53 100644 > --- a/arch/arm64/include/asm/virt.h > +++ b/arch/arm64/include/asm/virt.h > @@ -21,6 +21,26 @@ > #define BOOT_CPU_MODE_EL1 (0xe11) > #define BOOT_CPU_MODE_EL2 (0xe12) > > +/* > + * HVC_GET_VECTORS - Return the value of the vbar_el2 register. > + */ > + > +#define HVC_GET_VECTORS 1 > + > +/* > + * HVC_SET_VECTORS - Set the value of the vbar_el2 register. > + * > + * @x0: Physical address of the new vector table. > + */ > + > +#define HVC_SET_VECTORS 2 > + > +/* > + * HVC_KVM_CALL_HYP - Execute kvm_call_hyp routine. > + */ > + > +#define HVC_KVM_CALL_HYP 3 If this can be used without KVM (e.g. in the hyp stub) I'd just call this HVC_CALL_HYP, or the name will be a little misleading. > + > #ifndef __ASSEMBLY__ > > /* > diff --git a/arch/arm64/kernel/hyp-stub.S b/arch/arm64/kernel/hyp-stub.S > index 2d960a9..9ab5f70 100644 > --- a/arch/arm64/kernel/hyp-stub.S > +++ b/arch/arm64/kernel/hyp-stub.S > @@ -54,16 +54,29 @@ ENDPROC(__hyp_stub_vectors) > > #define ESR_EL2_EC_SHIFT 26 > #define ESR_EL2_EC_HVC64 0x16 > +#define ESR_EL2_ISS 0xffff The last patch tried to add an identical macro to a header file. Can we use that header please? > > el1_sync: > - mrs x1, esr_el2 > - lsr x1, x1, #ESR_EL2_EC_SHIFT > - cmp x1, #ESR_EL2_EC_HVC64 > - b.ne 2f // Not an HVC trap > - cbz x0, 1f > - msr vbar_el2, x0 // Set vbar_el2 > + mrs x10, esr_el2 Any reason for using x10? If we want to preserve the lowest register numbers, start with the highest caller-saved register numbers (i.e. x18). At least for me it makes the code far easier to read; it doesn't make it look like x10 is special. > + lsr x9, x10, #ESR_EL2_EC_SHIFT // x9=EC > + and x10, x10, #ESR_EL2_ISS // x10=ISS The mnemonics make these comments redundant. > + cmp x9, #ESR_EL2_EC_HVC64 > + b.ne 2f // Not a host HVC trap Now that we have the nice mnemonic, we could get rid of the comment here. I'd drop the 'host' from the comment; it wasn't there orginally and it's somewhat meaningless for the stub (KVM isn't up yet, and the only the native OS can make a HVC). > + mrs x9, vttbr_el2 > + cbnz x9, 2f // Not a host HVC trap I don't understand this. When is vttbr_el2 non-zero, and why do we want to silently return from a HVC in that case? That didn't seem to be the case in the original code. > + > + cmp x10, #HVC_GET_VECTORS > + b.ne 1f > + mrs x0, vbar_el2 > b 2f > -1: mrs x0, vbar_el2 // Return vbar_el2 > + > +1: cmp x10, #HVC_SET_VECTORS > + b.ne 1f > + msr vbar_el2, x0 > + > +1: It feels like we should explode if we ever reach here from the host -- if we've made an unsupported HVC wereally want to know that we've done so. > 2: eret > ENDPROC(el1_sync) > > @@ -103,11 +116,12 @@ ENDPROC(\label) > * initialisation entry point. > */ > > -ENTRY(__hyp_get_vectors) > - mov x0, xzr > - // fall through > ENTRY(__hyp_set_vectors) > - hvc #0 > + hvc #HVC_SET_VECTORS > ret > -ENDPROC(__hyp_get_vectors) > ENDPROC(__hyp_set_vectors) > + > +ENTRY(__hyp_get_vectors) > + hvc #HVC_GET_VECTORS > + ret > +ENDPROC(__hyp_get_vectors) > diff --git a/arch/arm64/kvm/hyp.S b/arch/arm64/kvm/hyp.S > index b72aa9f..3972ee9 100644 > --- a/arch/arm64/kvm/hyp.S > +++ b/arch/arm64/kvm/hyp.S > @@ -26,6 +26,7 @@ > #include <asm/kvm_asm.h> > #include <asm/kvm_arm.h> > #include <asm/kvm_mmu.h> > +#include <asm/virt.h> > > #define CPU_GP_REG_OFFSET(x) (CPU_GP_REGS + x) > #define CPU_XREG_OFFSET(x) CPU_GP_REG_OFFSET(CPU_USER_PT_REGS + 8*x) > @@ -1105,12 +1106,9 @@ __hyp_panic_str: > * in Hyp mode (see init_hyp_mode in arch/arm/kvm/arm.c). Return values are > * passed in r0 and r1. > * > - * A function pointer with a value of 0 has a special meaning, and is > - * used to implement __hyp_get_vectors in the same way as in > - * arch/arm64/kernel/hyp_stub.S. > */ > ENTRY(kvm_call_hyp) > - hvc #0 > + hvc #HVC_KVM_CALL_HYP > ret > ENDPROC(kvm_call_hyp) > > @@ -1140,6 +1138,7 @@ el1_sync: // Guest trapped into EL2 > push x2, x3 > > mrs x1, esr_el2 > + and x0, x1, #ESR_EL2_ISS > lsr x2, x1, #ESR_EL2_EC_SHIFT > > cmp x2, #ESR_EL2_EC_HVC64 > @@ -1149,15 +1148,19 @@ el1_sync: // Guest trapped into EL2 > cbnz x3, el1_trap // called HVC > > /* Here, we're pretty sure the host called HVC. */ > + mov x10, x0 As above, please use the highest numbered caller-saved register so as to not make the register numbering look special. > pop x2, x3 > pop x0, x1 > > - /* Check for __hyp_get_vectors */ > - cbnz x0, 1f > + cmp x10, #HVC_GET_VECTORS > + b.ne 1f > mrs x0, vbar_el2 > b 2f > > -1: push lr, xzr > +1: cmp x10, #HVC_KVM_CALL_HYP > + b.ne 1f > + > + push lr, xzr > > /* > * Compute the function address in EL2, and shuffle the parameters. > @@ -1170,6 +1173,8 @@ el1_sync: // Guest trapped into EL2 > blr lr > > pop lr, xzr > + > +1: > 2: eret Any reason we need two labels here? If we've got here with an invalid HVC immediate, I think we should explode loudly. Mark.
On Tue, Sep 09, 2014 at 11:49:04PM +0100, Geoff Levand wrote: > Add the new hcall HVC_CALL_FUNC that allows execution of a function at EL2. > During CPU reset the CPU must be brought to the exception level it had on > entry to the kernel. The HVC_CALL_FUNC hcall will provide the mechanism > needed for this exception level switch. > > Signed-off-by: Geoff Levand <geoff@infradead.org> > --- > arch/arm64/include/asm/virt.h | 11 +++++++++++ > arch/arm64/kernel/hyp-stub.S | 10 ++++++++++ > 2 files changed, 21 insertions(+) > > diff --git a/arch/arm64/include/asm/virt.h b/arch/arm64/include/asm/virt.h > index 894fe53..b217fbc 100644 > --- a/arch/arm64/include/asm/virt.h > +++ b/arch/arm64/include/asm/virt.h > @@ -41,6 +41,17 @@ > > #define HVC_KVM_CALL_HYP 3 > > +/* > + * HVC_CALL_FUNC - Execute a function at EL2. > + * > + * @x0: Physical address of the funtion to be executed. > + * @x1: Passed as the first argument to @fn. > + * @x2: Passed as the second argument to @fn. > + * @x3: Passed as the third argument to @fn. > + */ > + > +#define HVC_CALL_FUNC 4 > + Can't we use the HVC_KVM_CALL_HYP for this as well? I thought we already added the code to the stub to do that in the last patch. Is there a difference between the two that I'm missing? > #ifndef __ASSEMBLY__ > > /* > diff --git a/arch/arm64/kernel/hyp-stub.S b/arch/arm64/kernel/hyp-stub.S > index 9ab5f70..a21cf51 100644 > --- a/arch/arm64/kernel/hyp-stub.S > +++ b/arch/arm64/kernel/hyp-stub.S > @@ -75,7 +75,17 @@ el1_sync: > 1: cmp x10, #HVC_SET_VECTORS > b.ne 1f > msr vbar_el2, x0 > + b 2f > > +1: cmp x10, #HVC_CALL_FUNC > + b.ne 1f > + mov x29, lr What's the contract for functions we call through the stub? If they can use all the caller-saved registers, then we need to stach the original LR before issuing the HVC. Otherwise we can stash it in x18 at EL2. Thanks, Mark.
Hi Geoff, On Tue, Sep 09, 2014 at 11:49:04PM +0100, Geoff Levand wrote: > The kexec re-boot support that will be added in a subsequent patch in this > series will need to read the device tree CPU properties, and it is expected > that a rework of the SMP spin table code to handle cpu_die will also need this > functionality, so add two new common arm64 files cpu-properties.h and > cpu-properties.c that define a new structure cpu_properties that hold the > various CPU properties from a device tree, and the new routine > read_cpu_properties() that fills the structure from a device tree CPU node. I'm very much not keen on placing all this information in a single structure -- that adds a new tight coupling that we didn't have before, and it looks like it's going to be painful to maintain. If kexec uses the existing high-level hotplug infrastructure, it has no reason to go anywhere near this information. So I really don't like this patch. Why do you think we need this information to be centralized in this way? Thanks, Mark. > > Signed-off-by: Geoff Levand <geoff@infradead.org> > --- > arch/arm64/kernel/cpu-properties.c | 58 ++++++++++++++++++++++++++++++++++++++ > arch/arm64/kernel/cpu-properties.h | 39 +++++++++++++++++++++++++ > 2 files changed, 97 insertions(+) > create mode 100644 arch/arm64/kernel/cpu-properties.c > create mode 100644 arch/arm64/kernel/cpu-properties.h > > diff --git a/arch/arm64/kernel/cpu-properties.c b/arch/arm64/kernel/cpu-properties.c > new file mode 100644 > index 0000000..e64b34b > --- /dev/null > +++ b/arch/arm64/kernel/cpu-properties.c > @@ -0,0 +1,58 @@ > +/* > + * Copyright (C) Linaro. > + * > + * This program is free software; you can redistribute it and/or modify > + * it under the terms of the GNU General Public License version 2 as > + * published by the Free Software Foundation. > + */ > + > +#include "cpu-properties.h" > + > +int read_cpu_properties(struct cpu_properties *p, const struct device_node *dn) > +{ > + const u32 *cell; > + > + memset(p, 0, sizeof(*p)); > + p->hwid = INVALID_HWID; > + p->cpu_release_addr = INVALID_ADDR; > + > + cell = of_get_property(dn, "reg", NULL); > + > + if (!cell) { > + pr_err("%s: Error: %s: invalid reg property\n", > + __func__, dn->full_name); > + return -1; > + } > + > + p->hwid = of_read_number(cell, > + of_n_addr_cells((struct device_node *)dn)) & MPIDR_HWID_BITMASK; > + > + p->enable_method = of_get_property(dn, "enable-method", NULL); > + > + if (!p->enable_method) { > + pr_err("%s: Error: %s: invalid enable-method\n", > + __func__, dn->full_name); > + return -1; > + } > + > + if (!strcmp(p->enable_method, "psci")) { > + p->type = cpu_enable_method_psci; > + return 0; > + } > + > + if (strcmp(p->enable_method, "spin-table")) { > + p->type = cpu_enable_method_unknown; > + return -1; > + } > + > + p->type = cpu_enable_method_spin_table; > + > + if (of_property_read_u64(dn, "cpu-release-addr", > + &p->cpu_release_addr)) { > + pr_err("%s: Error: %s: invalid cpu-return-addr property\n", > + __func__, dn->full_name); > + return -1; > + } > + > + return 0; > +} > diff --git a/arch/arm64/kernel/cpu-properties.h b/arch/arm64/kernel/cpu-properties.h > new file mode 100644 > index 0000000..b4218ef > --- /dev/null > +++ b/arch/arm64/kernel/cpu-properties.h > @@ -0,0 +1,39 @@ > +/* > + * Copyright (C) Linaro. > + * > + * This program is free software; you can redistribute it and/or modify > + * it under the terms of the GNU General Public License version 2 as > + * published by the Free Software Foundation. > + */ > + > +#if !defined(__ARM64_CPU_PROPERTIES_H) > +#define __ARM64_CPU_PROPERTIES_H > + > +#include <asm/memory.h> > +#include <asm/cputype.h> > + > +#define INVALID_ADDR UL(~0) > + > +#if !defined(__ASSEMBLY__) > + > +#include <linux/kernel.h> > +#include <linux/of.h> > + > +enum cpu_enable_method { > + cpu_enable_method_unknown, > + cpu_enable_method_psci, > + cpu_enable_method_spin_table, > +}; > + > +struct cpu_properties { > + u64 hwid; > + u64 cpu_release_addr; > + const char *enable_method; > + enum cpu_enable_method type; > +}; > + > +int read_cpu_properties(struct cpu_properties *p, const struct device_node *dn); > + > +#endif /* !defined(__ASSEMBLY__) */ > + > +#endif > -- > 1.9.1 > > > > _______________________________________________ > linux-arm-kernel mailing list > linux-arm-kernel@lists.infradead.org > http://lists.infradead.org/mailman/listinfo/linux-arm-kernel >
Hi Geoff, On Tue, Sep 09, 2014 at 11:49:04PM +0100, Geoff Levand wrote: > Add the new arm64 routine local_disable() to allow the masking of several DAIF > flags in one operation. Currently, we only have routines to mask individual > flags, and to mask several flags multiple calls to daifset are required. > > Signed-off-by: Geoff Levand <geoff@infradead.org> > --- > arch/arm64/include/asm/irqflags.h | 13 +++++++++++++ > 1 file changed, 13 insertions(+) > > diff --git a/arch/arm64/include/asm/irqflags.h b/arch/arm64/include/asm/irqflags.h > index 11cc941..28521d4 100644 > --- a/arch/arm64/include/asm/irqflags.h > +++ b/arch/arm64/include/asm/irqflags.h > @@ -113,5 +113,18 @@ static inline int arch_irqs_disabled_flags(unsigned long flags) > #define local_dbg_enable() asm("msr daifclr, #8" : : : "memory") > #define local_dbg_disable() asm("msr daifset, #8" : : : "memory") > > +enum daif_flag { Is there any reason for this to be an enum rather than a set of #defines? It would be nice to be able to use these in asm. > + DAIF_FIQ = (1UL << 6), > + DAIF_IRQ = (1UL << 7), > + DAIF_ASYNC = (1UL << 8), > + DAIF_DBG = (1UL << 9), > + DAIF_ALL = (0xffUL << 6), Not 0xf? It would be nicer to OR the other flags. > +}; > + > +static inline void local_disable(unsigned long daif_flags) > +{ > + arch_local_irq_restore(daif_flags | arch_local_save_flags()); If we knew the value was a constant (which we could check with __builting_constant_p) we could use daifset here, rather than having a RMW sequence. With that, the other local_*_{enable,disable} calls could be rewritten in terms of this, without affecting the generated code. That would require some shifting to account for the difference between pstate and daif{clr,set} layout, but for constants that shouldn't be a problem. Thanks, Mark.
Hi Geoff, On Tue, Sep 09, 2014 at 11:49:05PM +0100, Geoff Levand wrote: > The current implementation of ipi_cpu_stop() is just a tight infinite loop > around cpu_relax(). This infinite loop implementation is OK if the machine > will soon do a poweroff, but it doesn't have any mechanism to allow a CPU > to be brought back on-line, nor is it compatible with kexec re-boot. I don't see why we should use this when we have disable_nonboot_cpus. If the kernel is alive and well, disable_nonboot_cpus will correctly shut down all but one CPU, returning an error if that fails, whereupon we can respect the error code and halt the kexec. If the kernel is not alive and well, we have no idea what CPUs are executing anyway, so all we can expect to do is to boot a (UP) crash kernel in some previously reserved memory. Trying to actually kill the CPUs is nice, but possibly not necessary. > Add a check for a valid cpu_die method of the appropriate cpu_ops structure, > and if a valid method is found, transfer control to that method. It is > expected that the cpu_die method puts the CPU into a state such that they can > be brought back on-line or progress through a kexec re-boot. > > Signed-off-by: Geoff Levand <geoff@infradead.org> > --- > arch/arm64/kernel/smp.c | 9 +++++++++ > 1 file changed, 9 insertions(+) > > diff --git a/arch/arm64/kernel/smp.c b/arch/arm64/kernel/smp.c > index 4743397..002aa8a 100644 > --- a/arch/arm64/kernel/smp.c > +++ b/arch/arm64/kernel/smp.c > @@ -555,6 +555,15 @@ static void ipi_cpu_stop(unsigned int cpu) > > local_irq_disable(); > > + /* If we have the cpu ops use them. */ > + > + if (cpu_ops[cpu]->cpu_disable && > + cpu_ops[cpu]->cpu_die && > + !cpu_ops[cpu]->cpu_disable(cpu)) > + cpu_ops[cpu]->cpu_die(cpu); I don't think kexec should handle this. The hotplug code already does this, better (calling cpu_kill and returning an error code), and having two callers of these functions is only going to lead to hard-to-debug drift between the two. > while (1) > cpu_relax(); Any CPUs left here are a major problem. We absolutely must fail kexec if a CPU is still in the kernel (in the pen or in the kernel proper), or they can do arbitrarily bad things when the kernel image gets clobbered. SO this is insufficient. As I mention above, a crash kernel might be an exception to that rule, but we shouldn't treat that as the usual case. Thanks, Mark.
On Tue, Sep 09, 2014 at 11:49:05PM +0100, Geoff Levand wrote: > The arm64 implementation of kexec expects an operational cpu_die method of > struct cpu_operations, so add defined(CONFIG_KEXEC) to the preprocessor > conditional that enables cpu_die. I very much do not like this, as I explained in previous postings and my comments on other patches in this series. Please use the hotplug infrastructure, and have kexec depend on CPU_HOTPLUG || !SMP. There is no reason for these files to have any knowledge of kexec whatsoever. Thanks, Mark > > Signed-off-by: Geoff Levand <geoff@infradead.org> > --- > arch/arm64/include/asm/cpu_ops.h | 2 +- > arch/arm64/kernel/psci.c | 4 ++-- > 2 files changed, 3 insertions(+), 3 deletions(-) > > diff --git a/arch/arm64/include/asm/cpu_ops.h b/arch/arm64/include/asm/cpu_ops.h > index d7b4b38..9fd7281 100644 > --- a/arch/arm64/include/asm/cpu_ops.h > +++ b/arch/arm64/include/asm/cpu_ops.h > @@ -50,7 +50,7 @@ struct cpu_operations { > int (*cpu_prepare)(unsigned int); > int (*cpu_boot)(unsigned int); > void (*cpu_postboot)(void); > -#ifdef CONFIG_HOTPLUG_CPU > +#if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_KEXEC) > int (*cpu_disable)(unsigned int cpu); > void (*cpu_die)(unsigned int cpu); > int (*cpu_kill)(unsigned int cpu); > diff --git a/arch/arm64/kernel/psci.c b/arch/arm64/kernel/psci.c > index 5539547..81dbbc9 100644 > --- a/arch/arm64/kernel/psci.c > +++ b/arch/arm64/kernel/psci.c > @@ -380,7 +380,7 @@ static int cpu_psci_cpu_boot(unsigned int cpu) > return err; > } > > -#ifdef CONFIG_HOTPLUG_CPU > +#if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_KEXEC) > static int cpu_psci_cpu_disable(unsigned int cpu) > { > /* Fail early if we don't have CPU_OFF support */ > @@ -442,7 +442,7 @@ const struct cpu_operations cpu_psci_ops = { > .cpu_init = cpu_psci_cpu_init, > .cpu_prepare = cpu_psci_cpu_prepare, > .cpu_boot = cpu_psci_cpu_boot, > -#ifdef CONFIG_HOTPLUG_CPU > +#if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_KEXEC) > .cpu_disable = cpu_psci_cpu_disable, > .cpu_die = cpu_psci_cpu_die, > .cpu_kill = cpu_psci_cpu_kill, > -- > 1.9.1 > > > > _______________________________________________ > linux-arm-kernel mailing list > linux-arm-kernel@lists.infradead.org > http://lists.infradead.org/mailman/listinfo/linux-arm-kernel >
Hi Geoff, On Tue, Sep 09, 2014 at 11:49:05PM +0100, Geoff Levand wrote: > Commit 90f51a09ef83 ("arm64: Fix machine_shutdown() definition") changed the > implementation of machine_shutdown() from calling smp_send_stop() to > calling disable_nonboot_cpus(). Revert that part of the commit so that > machine_shutdown() once again calls smp_send_stop(). I don't think this is the right way to go. I think we should be using the hotplug infrastructure (i.e. disable_nonboot_cpus()) rather than rolling our own, and we should be paying attention to the return value. So as far as I can tell, disable_nonboot_cpus is the right thing to call, but machine_shutdown is the wrong place to call it (due to not returning an error code). We need to call it when we can still abort the kexec. Mark. > With the application of another patch in this series, (arm64: Use cpu_ops for > smp_stop), smp_send_stop() will do the correct thing for a kexec reboot. This > change also corrects the source code comment for the machine_shutdown() routine. > > Signed-off-by: Geoff Levand <geoff@infradead.org> > --- > arch/arm64/kernel/process.c | 9 +++------ > 1 file changed, 3 insertions(+), 6 deletions(-) > > diff --git a/arch/arm64/kernel/process.c b/arch/arm64/kernel/process.c > index 0a3414b..cd0ae9d 100644 > --- a/arch/arm64/kernel/process.c > +++ b/arch/arm64/kernel/process.c > @@ -101,15 +101,12 @@ void arch_cpu_idle_dead(void) > /* > * Called by kexec, immediately prior to machine_kexec(). > * > - * This must completely disable all secondary CPUs; simply causing those CPUs > - * to execute e.g. a RAM-based pin loop is not sufficient. This allows the > - * kexec'd kernel to use any and all RAM as it sees fit, without having to > - * avoid any code or data used by any SW CPU pin loop. The CPU hotplug > - * functionality embodied in disable_nonboot_cpus() to achieve this. > + * This must shutdown all secondary CPUs. The functionality > + * embodied in smp_send_stop() will achieve this. > */ > void machine_shutdown(void) > { > - disable_nonboot_cpus(); > + smp_send_stop(); > } > > /* > -- > 1.9.1 > > > > _______________________________________________ > linux-arm-kernel mailing list > linux-arm-kernel@lists.infradead.org > http://lists.infradead.org/mailman/listinfo/linux-arm-kernel >
Hi Geoff, On Tue, Sep 09, 2014 at 11:49:05PM +0100, Geoff Levand wrote: > Add three new files, kexec.h, machine_kexec.c and relocate_kernel.S to the > arm64 architecture that add support for the kexec re-boot mechanism > (CONFIG_KEXEC) on arm64 platforms. > > Signed-off-by: Geoff Levand <geoff@infradead.org> > --- > arch/arm64/Kconfig | 8 + > arch/arm64/include/asm/kexec.h | 52 +++ > arch/arm64/kernel/Makefile | 2 + > arch/arm64/kernel/machine_kexec.c | 612 ++++++++++++++++++++++++++++++++++++ > arch/arm64/kernel/relocate_kernel.S | 185 +++++++++++ > include/uapi/linux/kexec.h | 1 + > 6 files changed, 860 insertions(+) > create mode 100644 arch/arm64/include/asm/kexec.h > create mode 100644 arch/arm64/kernel/machine_kexec.c > create mode 100644 arch/arm64/kernel/relocate_kernel.S > > diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig > index f0d3a2d..6f0e1f1 100644 > --- a/arch/arm64/Kconfig > +++ b/arch/arm64/Kconfig > @@ -313,6 +313,14 @@ config ARCH_HAS_CACHE_LINE_SIZE > > source "mm/Kconfig" > > +config KEXEC > + bool "kexec system call" > + ---help--- > + kexec is a system call that implements the ability to shutdown your > + current kernel, and to start another kernel. It is like a reboot > + but it is independent of the system firmware. And like a reboot > + you can start any kernel with it, not just Linux. > + > config XEN_DOM0 > def_bool y > depends on XEN > diff --git a/arch/arm64/include/asm/kexec.h b/arch/arm64/include/asm/kexec.h > new file mode 100644 > index 0000000..9a3932c > --- /dev/null > +++ b/arch/arm64/include/asm/kexec.h > @@ -0,0 +1,52 @@ > +/* > + * kexec for arm64 > + * > + * Copyright (C) Linaro. > + * > + * This program is free software; you can redistribute it and/or modify > + * it under the terms of the GNU General Public License version 2 as > + * published by the Free Software Foundation. > + */ > + > +#if !defined(_ARM64_KEXEC_H) > +#define _ARM64_KEXEC_H > + > +/* Maximum physical address we can use pages from */ > + > +#define KEXEC_SOURCE_MEMORY_LIMIT (-1UL) > + > +/* Maximum address we can reach in physical address mode */ > + > +#define KEXEC_DESTINATION_MEMORY_LIMIT (-1UL) > + > +/* Maximum address we can use for the control code buffer */ > + > +#define KEXEC_CONTROL_MEMORY_LIMIT (-1UL) > + What are these used for? I see that other architectures seem to do the same thing, but they look odd. > +#define KEXEC_CONTROL_PAGE_SIZE 4096 What's this used for? Does this work with 64k pages, and is there any reason we can't figure out the actual size of the code (so we don't get bitten if it grows)? > + > +#define KEXEC_ARCH KEXEC_ARCH_ARM64 > + > +#define ARCH_HAS_KIMAGE_ARCH > + > +#if !defined(__ASSEMBLY__) > + > +struct kimage_arch { > + void *ctx; > +}; > + > +/** > + * crash_setup_regs() - save registers for the panic kernel > + * > + * @newregs: registers are saved here > + * @oldregs: registers to be saved (may be %NULL) > + */ > + > +static inline void crash_setup_regs(struct pt_regs *newregs, > + struct pt_regs *oldregs) > +{ > +} It would be nice to know what we're going to do for this. Is this a required function, or can we get away without crash kernel support for the moment? > + > +#endif /* !defined(__ASSEMBLY__) */ > + > +#endif > diff --git a/arch/arm64/kernel/Makefile b/arch/arm64/kernel/Makefile > index df7ef87..8b7c029 100644 > --- a/arch/arm64/kernel/Makefile > +++ b/arch/arm64/kernel/Makefile > @@ -29,6 +29,8 @@ arm64-obj-$(CONFIG_ARM64_CPU_SUSPEND) += sleep.o suspend.o > arm64-obj-$(CONFIG_JUMP_LABEL) += jump_label.o > arm64-obj-$(CONFIG_KGDB) += kgdb.o > arm64-obj-$(CONFIG_EFI) += efi.o efi-stub.o efi-entry.o > +arm64-obj-$(CONFIG_KEXEC) += machine_kexec.o relocate_kernel.o \ > + cpu-properties.o > > obj-y += $(arm64-obj-y) vdso/ > obj-m += $(arm64-obj-m) > diff --git a/arch/arm64/kernel/machine_kexec.c b/arch/arm64/kernel/machine_kexec.c > new file mode 100644 > index 0000000..043a3bc > --- /dev/null > +++ b/arch/arm64/kernel/machine_kexec.c > @@ -0,0 +1,612 @@ > +/* > + * kexec for arm64 > + * > + * Copyright (C) Linaro. > + * > + * This program is free software; you can redistribute it and/or modify > + * it under the terms of the GNU General Public License version 2 as > + * published by the Free Software Foundation. > + */ > + > +#include <linux/kernel.h> > +#include <linux/kexec.h> > +#include <linux/of_fdt.h> > +#include <linux/slab.h> > +#include <linux/uaccess.h> > + > +#include <asm/cacheflush.h> > +#include <asm/cpu_ops.h> > +#include <asm/system_misc.h> > + > +#include "cpu-properties.h" > + > +#if defined(DEBUG) > +static const int debug = 1; > +#else > +static const int debug; > +#endif I don't think we need this. > + > +typedef struct dtb_buffer {char b[0]; } dtb_t; It would be nice for this to be consistent with other dtb uses; if we need a dtb type then it shouldn't be specific to kexec. [...] > +static struct kexec_ctx *current_ctx; > + > +static int kexec_ctx_alloc(struct kimage *image) > +{ > + BUG_ON(image->arch.ctx); > + > + image->arch.ctx = kmalloc(sizeof(struct kexec_ctx), GFP_KERNEL); > + > + if (!image->arch.ctx) > + return -ENOMEM; > + > + current_ctx = (struct kexec_ctx *)image->arch.ctx; This seems to be the only use of current_ctx. I take it this is a leftover from debugging? [...] > +/** > + * kexec_list_walk - Helper to walk the kimage page list. > + */ Please keep this associated with the function it refers to (nothing should be between this comment and the function prototype). > + > +#define IND_FLAGS (IND_DESTINATION | IND_INDIRECTION | IND_DONE | IND_SOURCE) Can't this live in include/linux/kexec.h, where these flags are defined. The meaning of these doesn't seem to be documented anywhere. Would you be able to explain what each of these means? > +static void kexec_list_walk(void *ctx, unsigned long kimage_head, > + void (*cb)(void *ctx, unsigned int flag, void *addr, void *dest)) > +{ > + void *dest; > + unsigned long *entry; > + > + for (entry = &kimage_head, dest = NULL; ; entry++) { > + unsigned int flag = *entry & IND_FLAGS; > + void *addr = phys_to_virt(*entry & PAGE_MASK); > + > + switch (flag) { > + case IND_INDIRECTION: > + entry = (unsigned long *)addr - 1; > + cb(ctx, flag, addr, NULL); > + break; > + case IND_DESTINATION: > + dest = addr; > + cb(ctx, flag, addr, NULL); > + break; > + case IND_SOURCE: > + cb(ctx, flag, addr, dest); > + dest += PAGE_SIZE; I really don't understand what's going on with dest here, but that's probably because I don't understand the meaning of the flags. > + break; > + case IND_DONE: > + cb(ctx, flag , NULL, NULL); > + return; > + default: > + pr_devel("%s:%d unknown flag %xh\n", __func__, __LINE__, > + flag); Wouldn't pr_warn would be more appropriate here? > + cb(ctx, flag, addr, NULL); > + break; > + } > + } > +} > + > +/** > + * kexec_image_info - For debugging output. > + */ > + > +#define kexec_image_info(_i) _kexec_image_info(__func__, __LINE__, _i) > +static void _kexec_image_info(const char *func, int line, > + const struct kimage *image) > +{ > + if (debug) { > + unsigned long i; > + > + pr_devel("%s:%d:\n", func, line); > + pr_devel(" kexec image info:\n"); > + pr_devel(" type: %d\n", image->type); > + pr_devel(" start: %lx\n", image->start); > + pr_devel(" head: %lx\n", image->head); > + pr_devel(" nr_segments: %lu\n", image->nr_segments); > + > + for (i = 0; i < image->nr_segments; i++) { > + pr_devel(" segment[%lu]: %016lx - %016lx, " > + "%lxh bytes, %lu pages\n", > + i, > + image->segment[i].mem, > + image->segment[i].mem + image->segment[i].memsz, > + image->segment[i].memsz, > + image->segment[i].memsz / PAGE_SIZE); > + > + if (kexec_is_dtb_user(image->segment[i].buf)) > + pr_devel(" dtb segment\n"); > + } > + } > +} pr_devel is already dependent on DEBUG, so surely we don't need to check the debug variable? > + > +/** > + * kexec_find_dtb_seg - Helper routine to find the dtb segment. > + */ > + > +static const struct kexec_segment *kexec_find_dtb_seg( > + const struct kimage *image) > +{ > + int i; > + > + for (i = 0; i < image->nr_segments; i++) { > + if (kexec_is_dtb_user(image->segment[i].buf)) > + return &image->segment[i]; > + } > + > + return NULL; > +} I'm really not keen on having the kernel guess which blobs need special treatment, though we seem to do that for arm. It would be far nicer if we could pass flags for each segment to describe what it is (e.g. kernel image, dtb, other binary blob), so we can do things like pass multiple DTBs (so we load two kernels at once and pass each a unique DTB if we want to boot a new kernel + crashkernel pair). Unfortunately that would require some fairly invasive rework of the kexec core. For secureboot we can't trust a dtb from userspace, and will have to use kexec_file_load. To work with that we can either: * Reuse the original DTB, patched with the new command line. This may have statefulness issues (for things like simplefb). * Build a new DTB by flattening the current live tree. This would rely on drivers that modify state to patch the tree appropriately. Both of those are somewhat ugly. :( [...] > +/** > + * kexec_cpu_info_init - Initialize an array of kexec_cpu_info structures. > + * > + * Allocates a cpu info array and fills it with info for all cpus found in > + * the device tree passed. > + */ > + > +static int kexec_cpu_info_init(const struct device_node *dn, > + struct kexec_boot_info *info) > +{ > + int result; > + unsigned int cpu; > + > + info->cp = kmalloc( > + info->cpu_count * sizeof(*info->cp), GFP_KERNEL); > + > + if (!info->cp) { > + pr_err("%s: Error: Out of memory.", __func__); > + return -ENOMEM; > + } > + > + for (cpu = 0; cpu < info->cpu_count; cpu++) { > + struct cpu_properties *cp = &info->cp[cpu]; > + > + dn = of_find_node_by_type((struct device_node *)dn, "cpu"); > + > + if (!dn) { > + pr_devel("%s:%d: bad node\n", __func__, __LINE__); > + goto on_error; > + } > + > + result = read_cpu_properties(cp, dn); > + > + if (result) { > + pr_devel("%s:%d: bad node\n", __func__, __LINE__); > + goto on_error; > + } > + > + if (cp->type == cpu_enable_method_psci) > + pr_devel("%s:%d: cpu-%u: hwid-%llx, '%s'\n", > + __func__, __LINE__, cpu, cp->hwid, > + cp->enable_method); > + else > + pr_devel("%s:%d: cpu-%u: hwid-%llx, '%s', " > + "cpu-release-addr %llx\n", > + __func__, __LINE__, cpu, cp->hwid, > + cp->enable_method, > + cp->cpu_release_addr); > + } > + > + return 0; > + > +on_error: > + kfree(info->cp); > + info->cp = NULL; > + return -EINVAL; > +} I don't see why we should need this at all. If we use the hotplug infrastructure, we don't need access to the enable-method and related properties, and the kexec code need only deal with a single CPU. The only case where kexec needs to deal with other CPUs is when some are sat in the holding pen, but this code doesn't seem to handle that. as I believe I mentioned before, we should be able to extend the holding pen code to get those CPUs to increment a sat-in-pen counter and if that's non-zero after SMP bringup we print a warning (and disallow kexec). [...] > +/** > +* kexec_compat_check - Iterator for kexec_cpu_check. > +*/ > + > +static int kexec_compat_check(const struct kexec_ctx *ctx) > +{ > + unsigned int cpu_1; > + unsigned int to_process; > + > + to_process = min(ctx->first.cpu_count, ctx->second.cpu_count); > + > + if (ctx->first.cpu_count != ctx->second.cpu_count) > + pr_warn("%s: Warning: CPU count mismatch %u != %u.\n", > + __func__, ctx->first.cpu_count, ctx->second.cpu_count); > + > + for (cpu_1 = 0; cpu_1 < ctx->first.cpu_count; cpu_1++) { > + unsigned int cpu_2; > + struct cpu_properties *cp_1 = &ctx->first.cp[cpu_1]; > + > + for (cpu_2 = 0; cpu_2 < ctx->second.cpu_count; cpu_2++) { > + struct cpu_properties *cp_2 = &ctx->second.cp[cpu_2]; > + > + if (cp_1->hwid != cp_2->hwid) > + continue; > + > + if (!kexec_cpu_check(cp_1, cp_2)) > + return -EINVAL; > + > + to_process--; > + } > + } > + > + if (to_process) { > + pr_warn("%s: Warning: Failed to process %u CPUs.\n", __func__, > + to_process); > + return -EINVAL; > + } > + > + return 0; > +} I don't see the point in checking this in the kernel. If I pass the second stage kernel a new dtb where my enable methods are different, that was my choice as a user. If that doesn't work, that's my fault. There are plenty of other things that might be completely different that we don't sanity check, so I don't see why enable methods should be any different. [...] > +/** > + * kexec_check_cpu_die - Check if cpu_die() will work on secondary processors. > + */ > + > +static int kexec_check_cpu_die(void) > +{ > + unsigned int cpu; > + unsigned int sum = 0; > + > + /* For simplicity this also checks the primary CPU. */ > + > + for_each_cpu(cpu, cpu_all_mask) { > + if (cpu && (!cpu_ops[cpu] || !cpu_ops[cpu]->cpu_disable || > + cpu_ops[cpu]->cpu_disable(cpu))) { > + sum++; > + pr_err("%s: Error: " > + "CPU %u does not support hot un-plug.\n", > + __func__, cpu); > + } > + } > + > + return sum ? -EOPNOTSUPP : 0; > +} We should really use disable_nonboot_cpus() for this. That way we don't end up with a slightly different hotplug implementation for kexec. The above is missing cpu_kill calls, for example, and I'm worried by the possibility of further drift over time. I understand from our face-to-face discussion that you didn't want to require the PM infrastructure that disable_nonboot_cpus currently pulls in due to the being dependent on CONFIG_PM_SLEEP_SMP which selects CONFIG_PM_SLEEP and so on. The solution to that is to refactor the Kconfig so we can have disable_nonboot_cpus without all the other PM infrastructure. > + > +/** > + * machine_kexec_prepare - Prepare for a kexec reboot. > + * > + * Called from the core kexec code when a kernel image is loaded. > + */ > + > +int machine_kexec_prepare(struct kimage *image) > +{ > + int result; This seems to always be an error code. Call it 'err'. > + dtb_t *dtb = NULL; > + struct kexec_ctx *ctx; > + const struct kexec_segment *dtb_seg; > + > + kexec_image_info(image); > + > + result = kexec_check_cpu_die(); > + > + if (result) > + goto on_error; > + > + result = kexec_ctx_alloc(image); > + > + if (result) > + goto on_error; > + > + ctx = kexec_image_to_ctx(image); > + > + result = kexec_boot_info_init(&ctx->first, NULL); > + > + if (result) > + goto on_error; > + > + dtb_seg = kexec_find_dtb_seg(image); > + > + if (!dtb_seg) { > + result = -EINVAL; > + goto on_error; > + } > + > + result = kexec_copy_dtb(dtb_seg, &dtb); > + > + if (result) > + goto on_error; > + > + result = kexec_boot_info_init(&ctx->second, dtb); > + > + if (result) > + goto on_error; > + > + result = kexec_compat_check(ctx); > + > + if (result) > + goto on_error; > + > + kexec_dtb_addr = dtb_seg->mem; > + kexec_kimage_start = image->start; > + > + goto on_exit; > + > +on_error: > + kexec_ctx_clean(image); > +on_exit: on_* looks weird, and doesn't match the style of other labels in arch/arm64. Could we call these 'out_clean' and 'out' instead? > + kfree(dtb); > + return result; > +} > + > +/** > + * kexec_list_flush_cb - Callback to flush the kimage list to PoC. > + */ > + > +static void kexec_list_flush_cb(void *ctx , unsigned int flag, > + void *addr, void *dest) > +{ > + switch (flag) { > + case IND_INDIRECTION: > + case IND_SOURCE: > + __flush_dcache_area(addr, PAGE_SIZE); Is PAGE_SIZE always big enough? Do we not have a more accurate size? Perhaps I've misunderstood what's going on here. > + break; > + default: > + break; > + } > +} > + > +/** > + * machine_kexec - Do the kexec reboot. > + * > + * Called from the core kexec code for a sys_reboot with LINUX_REBOOT_CMD_KEXEC. > + */ > + > +void machine_kexec(struct kimage *image) > +{ > + phys_addr_t reboot_code_buffer_phys; > + void *reboot_code_buffer; > + struct kexec_ctx *ctx = kexec_image_to_ctx(image); > + > + BUG_ON(relocate_new_kernel_size > KEXEC_CONTROL_PAGE_SIZE); It looks like relocate_new_kernel_size is a build-time constant. If we need that to be less than KEXEC_CONTROL_PAGE_SIZE, then we should make that a build-time check. > + BUG_ON(num_online_cpus() > 1); > + BUG_ON(!ctx); > + > + kexec_image_info(image); > + > + kexec_kimage_head = image->head; > + > + reboot_code_buffer_phys = page_to_phys(image->control_code_page); > + reboot_code_buffer = phys_to_virt(reboot_code_buffer_phys); > + > + pr_devel("%s:%d: control_code_page: %p\n", __func__, __LINE__, > + (void *)image->control_code_page); This is already a pointer. Is the cast to void necessary? > + pr_devel("%s:%d: reboot_code_buffer_phys: %p\n", __func__, __LINE__, > + (void *)reboot_code_buffer_phys); Use %pa and pass &reboot_code_buffer_phys, no cast necessary. > + pr_devel("%s:%d: reboot_code_buffer: %p\n", __func__, __LINE__, > + reboot_code_buffer); > + pr_devel("%s:%d: relocate_new_kernel: %p\n", __func__, __LINE__, > + relocate_new_kernel); > + pr_devel("%s:%d: relocate_new_kernel_size: %lxh(%lu) bytes\n", __func__, > + __LINE__, relocate_new_kernel_size, relocate_new_kernel_size); Please use an '0x' prefix rather than a 'h' suffix. Do we need in print in both hex and decimal? > + > + pr_devel("%s:%d: kexec_dtb_addr: %p\n", __func__, __LINE__, > + (void *)kexec_dtb_addr); > + pr_devel("%s:%d: kexec_kimage_head: %p\n", __func__, __LINE__, > + (void *)kexec_kimage_head); > + pr_devel("%s:%d: kexec_kimage_start: %p\n", __func__, __LINE__, > + (void *)kexec_kimage_start); These are all unsigned long, so why not use the existing mechanism for printing unsigned long? > + > + /* > + * Copy relocate_new_kernel to the reboot_code_buffer for use > + * after the kernel is shut down. > + */ > + > + memcpy(reboot_code_buffer, relocate_new_kernel, > + relocate_new_kernel_size); > + > + /* Assure reboot_code_buffer is copied. */ > + > + mb(); I don't think we need the mb if this is only to guarantee completion before the cache flush -- cacheable memory accesses should hazard against cache flushes by VA. > + > + pr_info("Bye!\n"); > + > + local_disable(DAIF_ALL); We can move these two right before the soft_restart, after the cache maintenance. That way the print is closer to the exit of the current kernel. > + > + /* Flush the reboot_code_buffer in preparation for its execution. */ > + > + __flush_dcache_area(reboot_code_buffer, relocate_new_kernel_size); > + > + /* Flush the kimage list. */ > + > + kexec_list_walk(NULL, image->head, kexec_list_flush_cb); > + > + soft_restart(reboot_code_buffer_phys); > +} > + > +void machine_crash_shutdown(struct pt_regs *regs) > +{ > + /* Empty routine needed to avoid build errors. */ > +} > diff --git a/arch/arm64/kernel/relocate_kernel.S b/arch/arm64/kernel/relocate_kernel.S > new file mode 100644 > index 0000000..92aba9d > --- /dev/null > +++ b/arch/arm64/kernel/relocate_kernel.S > @@ -0,0 +1,185 @@ > +/* > + * kexec for arm64 > + * > + * Copyright (C) Linaro. > + * > + * This program is free software; you can redistribute it and/or modify > + * it under the terms of the GNU General Public License version 2 as > + * published by the Free Software Foundation. > + */ > + > +#include <asm/assembler.h> > +#include <asm/memory.h> > +#include <asm/page.h> > + > +/* The list entry flags. */ > + > +#define IND_DESTINATION_BIT 0 > +#define IND_INDIRECTION_BIT 1 > +#define IND_DONE_BIT 2 > +#define IND_SOURCE_BIT 3 Given these ned to match the existing IND_* flags in include/linux/kexec.h, and they aren't in any way specific to arm64, please put these ina an asm-generic header and redefine the existing IND_* flags in terms of them. > + > +/* > + * relocate_new_kernel - Put the 2nd stage kernel image in place and boot it. > + * > + * The memory that the old kernel occupies may be overwritten when coping the > + * new kernel to its final location. To assure that the relocate_new_kernel > + * routine which does that copy is not overwritten all code and data needed > + * by relocate_new_kernel must be between the symbols relocate_new_kernel and > + * relocate_new_kernel_end. The machine_kexec() routine will copy > + * relocate_new_kernel to the kexec control_code_page, a special page which > + * has been set up to be preserved during the kernel copy operation. > + */ > + > +.align 3 Surely this isn't necessary? > + > +.globl relocate_new_kernel > +relocate_new_kernel: > + > + /* Setup the list loop variables. */ > + > + ldr x10, kexec_kimage_head /* x10 = list entry */ Any reason for using x10 rather than starting with x0? Or x18, if you need to preserve the low registers? > + > + mrs x0, ctr_el0 > + ubfm x0, x0, #16, #19 > + mov x11, #4 > + lsl x11, x11, x0 /* x11 = dcache line size */ Any reason we can't use dcache_line_size, given it's a macro? > + > + mov x12, xzr /* x12 = segment start */ > + mov x13, xzr /* x13 = entry ptr */ > + mov x14, xzr /* x14 = copy dest */ > + > + /* Check if the new kernel needs relocation. */ > + > + cbz x10, .Ldone > + tbnz x10, IND_DONE_BIT, .Ldone > + > +.Lloop: Is there any reason for the '.L' on all of these? We only seem to do that in the lib code that was imported from elsewhere, and it doesn't match the rest of the arm64 asm. > + and x15, x10, PAGE_MASK /* x15 = addr */ > + > + /* Test the entry flags. */ > + > +.Ltest_source: > + tbz x10, IND_SOURCE_BIT, .Ltest_indirection > + > + /* copy_page(x20 = dest, x21 = src) */ > + > + mov x20, x14 > + mov x21, x15 > + > +1: ldp x22, x23, [x21] > + ldp x24, x25, [x21, #16] > + ldp x26, x27, [x21, #32] > + ldp x28, x29, [x21, #48] > + add x21, x21, #64 > + stnp x22, x23, [x20] > + stnp x24, x25, [x20, #16] > + stnp x26, x27, [x20, #32] > + stnp x28, x29, [x20, #48] > + add x20, x20, #64 > + tst x21, #(PAGE_SIZE - 1) > + b.ne 1b It's a shame we can't reuse copy_page directly. Could we not move the body to a macro we can reuse here? > + > + /* dest += PAGE_SIZE */ > + > + add x14, x14, PAGE_SIZE > + b .Lnext > + > +.Ltest_indirection: > + tbz x10, IND_INDIRECTION_BIT, .Ltest_destination > + > + /* ptr = addr */ > + > + mov x13, x15 > + b .Lnext > + > +.Ltest_destination: > + tbz x10, IND_DESTINATION_BIT, .Lnext > + > + /* flush segment */ > + > + bl .Lflush > + mov x12, x15 > + > + /* dest = addr */ > + > + mov x14, x15 > + > +.Lnext: > + /* entry = *ptr++ */ > + > + ldr x10, [x13] > + add x13, x13, 8 This can be: ldr x10, [x13], #8 > + > + /* while (!(entry & DONE)) */ > + > + tbz x10, IND_DONE_BIT, .Lloop > + > +.Ldone: > + /* flush last segment */ > + > + bl .Lflush > + > + dsb sy > + isb > + ic ialluis This doesn't look right; we need a dsb and an isb after the instruction cache maintenance (or the icache could still be flushing when we branch to the new kernel). > + > + /* start_new_kernel */ > + > + ldr x4, kexec_kimage_start > + ldr x0, kexec_dtb_addr > + mov x1, xzr > + mov x2, xzr > + mov x3, xzr > + br x4 > + > +/* flush - x11 = line size, x12 = start addr, x14 = end addr. */ > + > +.Lflush: > + cbz x12, 2f > + mov x0, x12 > + sub x1, x11, #1 > + bic x0, x0, x1 > +1: dc civac, x0 > + add x0, x0, x11 > + cmp x0, x14 > + b.lo 1b > +2: ret It would be nice if this were earlier in the file, before its callers. > + > +.align 3 We should have a comment as to why this is needed (to keep the 64-bit values below naturally aligned). Thanks, Mark.
Hi Mark, On Mon, 2014-09-15 at 17:10 +0100, Mark Rutland wrote: > On Tue, Sep 09, 2014 at 11:49:04PM +0100, Geoff Levand wrote: > > +#define ESR_EL2_EC_SHIFT 26 > > +#define ESR_EL2_EC_HVC64 0x16 > > These exist in arch/arm64/include/asm/kvm_arm.h, no? > > If anything that should be folded into arch/arm64/include/asm/esr.h... OK, I'll remove this patch and add the esr.h fixup into my generic 'arm64 enhancements' series. -Geoff
Hi Mark, On Mon, 2014-09-15 at 18:57 +0100, Mark Rutland wrote: > On Tue, Sep 09, 2014 at 11:49:04PM +0100, Geoff Levand wrote: > > +/* > > + * HVC_GET_VECTORS - Return the value of the vbar_el2 register. > > + */ > > + > > +#define HVC_GET_VECTORS 1 > > + > > +/* > > + * HVC_SET_VECTORS - Set the value of the vbar_el2 register. > > + * > > + * @x0: Physical address of the new vector table. > > + */ > > + > > +#define HVC_SET_VECTORS 2 > > + > > +/* > > + * HVC_KVM_CALL_HYP - Execute kvm_call_hyp routine. > > + */ > > + > > +#define HVC_KVM_CALL_HYP 3 > > If this can be used without KVM (e.g. in the hyp stub) I'd just call > this HVC_CALL_HYP, or the name will be a little misleading. Yes, it is more or less generic, so we could have it as HVC_CALL_HYP. > > #ifndef __ASSEMBLY__ > > > > /* > > diff --git a/arch/arm64/kernel/hyp-stub.S b/arch/arm64/kernel/hyp-stub.S > > index 2d960a9..9ab5f70 100644 > > --- a/arch/arm64/kernel/hyp-stub.S > > +++ b/arch/arm64/kernel/hyp-stub.S > > @@ -54,16 +54,29 @@ ENDPROC(__hyp_stub_vectors) > > > > #define ESR_EL2_EC_SHIFT 26 > > #define ESR_EL2_EC_HVC64 0x16 > > +#define ESR_EL2_ISS 0xffff > > The last patch tried to add an identical macro to a header file. Can we > use that header please? As I mentioned in my reply to your comment on that patch, I'll prepare a separate header/macro cleanup patch. > > > > el1_sync: > > - mrs x1, esr_el2 > > - lsr x1, x1, #ESR_EL2_EC_SHIFT > > - cmp x1, #ESR_EL2_EC_HVC64 > > - b.ne 2f // Not an HVC trap > > - cbz x0, 1f > > - msr vbar_el2, x0 // Set vbar_el2 > > + mrs x10, esr_el2 > > Any reason for using x10? > > If we want to preserve the lowest register numbers, start with the > highest caller-saved register numbers (i.e. x18). At least for me it > makes the code far easier to read; it doesn't make it look like x10 is > special. OK, sure. > > + lsr x9, x10, #ESR_EL2_EC_SHIFT // x9=EC > > + and x10, x10, #ESR_EL2_ISS // x10=ISS > > The mnemonics make these comments redundant. > > > + cmp x9, #ESR_EL2_EC_HVC64 > > + b.ne 2f // Not a host HVC trap > > Now that we have the nice mnemonic, we could get rid of the comment > here. I'd drop the 'host' from the comment; it wasn't there orginally > and it's somewhat meaningless for the stub (KVM isn't up yet, and the > only the native OS can make a HVC). Sure, I copied this from the KVM vector so they would be more similar. > > > + mrs x9, vttbr_el2 > > + cbnz x9, 2f // Not a host HVC trap > > I don't understand this. When is vttbr_el2 non-zero, and why do we want > to silently return from a HVC in that case? That didn't seem to be the > case in the original code. No it is not in the original. I copied this from the KVM vector so they would be more similar. > > + > > + cmp x10, #HVC_GET_VECTORS > > + b.ne 1f > > + mrs x0, vbar_el2 > > b 2f > > -1: mrs x0, vbar_el2 // Return vbar_el2 > > + > > +1: cmp x10, #HVC_SET_VECTORS > > + b.ne 1f > > + msr vbar_el2, x0 > > + > > +1: > > It feels like we should explode if we ever reach here from the host -- > if we've made an unsupported HVC wereally want to know that we've done > so. Sure, I can put something in, but I would think that whoever tries a bad hcall is going to find out it doesn't work regardless. Do you have a good idea of how we can do this 'explode'? -Geoff
Hi Mark, On Mon, 2014-09-15 at 19:42 +0100, Mark Rutland wrote: > Hi Geoff, > > On Tue, Sep 09, 2014 at 11:49:04PM +0100, Geoff Levand wrote: > > The kexec re-boot support that will be added in a subsequent patch in this > > series will need to read the device tree CPU properties, and it is expected > > that a rework of the SMP spin table code to handle cpu_die will also need this > > functionality, so add two new common arm64 files cpu-properties.h and > > cpu-properties.c that define a new structure cpu_properties that hold the > > various CPU properties from a device tree, and the new routine > > read_cpu_properties() that fills the structure from a device tree CPU node. > > I'm very much not keen on placing all this information in a single > structure -- that adds a new tight coupling that we didn't have before, > and it looks like it's going to be painful to maintain. > > If kexec uses the existing high-level hotplug infrastructure, it has no > reason to go anywhere near this information. > > So I really don't like this patch. I decided to just strip all the checking out of the kernel with the expectation that the user space program will set things up correctly, so this patch is no longer needed. -Geoff
Hi Mark, On Mon, 2014-09-15 at 19:56 +0100, Mark Rutland wrote: > Hi Geoff, > > On Tue, Sep 09, 2014 at 11:49:04PM +0100, Geoff Levand wrote: > > Add the new arm64 routine local_disable() to allow the masking of several DAIF > > flags in one operation. Currently, we only have routines to mask individual > > flags, and to mask several flags multiple calls to daifset are required. > > > > Signed-off-by: Geoff Levand <geoff@infradead.org> > > --- > > arch/arm64/include/asm/irqflags.h | 13 +++++++++++++ > > 1 file changed, 13 insertions(+) > > > > diff --git a/arch/arm64/include/asm/irqflags.h b/arch/arm64/include/asm/irqflags.h > > index 11cc941..28521d4 100644 > > --- a/arch/arm64/include/asm/irqflags.h > > +++ b/arch/arm64/include/asm/irqflags.h > > @@ -113,5 +113,18 @@ static inline int arch_irqs_disabled_flags(unsigned long flags) > > #define local_dbg_enable() asm("msr daifclr, #8" : : : "memory") > > #define local_dbg_disable() asm("msr daifset, #8" : : : "memory") > > > > +enum daif_flag { > > Is there any reason for this to be an enum rather than a set of > #defines? > > It would be nice to be able to use these in asm. > > > + DAIF_FIQ = (1UL << 6), > > + DAIF_IRQ = (1UL << 7), > > + DAIF_ASYNC = (1UL << 8), > > + DAIF_DBG = (1UL << 9), > > + DAIF_ALL = (0xffUL << 6), > > Not 0xf? > > It would be nicer to OR the other flags. > > > +}; > > + > > +static inline void local_disable(unsigned long daif_flags) > > +{ > > + arch_local_irq_restore(daif_flags | arch_local_save_flags()); > > If we knew the value was a constant (which we could check with > __builting_constant_p) we could use daifset here, rather than having a > RMW sequence. > > With that, the other local_*_{enable,disable} calls could be rewritten > in terms of this, without affecting the generated code. That would > require some shifting to account for the difference between pstate and > daif{clr,set} layout, but for constants that shouldn't be a problem. Since kexec just needs to mask all the DAIF exceptions before the main CPU is reset I'll just do that inline, and put this rework of the exception mask routines on my todo list. -Geoff
Hi Mark, On Mon, 2014-09-15 at 19:11 +0100, Mark Rutland wrote: > On Tue, Sep 09, 2014 at 11:49:04PM +0100, Geoff Levand wrote: > > Add the new hcall HVC_CALL_FUNC that allows execution of a function at EL2. > > During CPU reset the CPU must be brought to the exception level it had on > > entry to the kernel. The HVC_CALL_FUNC hcall will provide the mechanism > > needed for this exception level switch. > > > > Signed-off-by: Geoff Levand <geoff@infradead.org> > > --- > > arch/arm64/include/asm/virt.h | 11 +++++++++++ > > arch/arm64/kernel/hyp-stub.S | 10 ++++++++++ > > 2 files changed, 21 insertions(+) > > > > diff --git a/arch/arm64/include/asm/virt.h b/arch/arm64/include/asm/virt.h > > index 894fe53..b217fbc 100644 > > --- a/arch/arm64/include/asm/virt.h > > +++ b/arch/arm64/include/asm/virt.h > > @@ -41,6 +41,17 @@ > > > > #define HVC_KVM_CALL_HYP 3 > > > > +/* > > + * HVC_CALL_FUNC - Execute a function at EL2. > > + * > > + * @x0: Physical address of the funtion to be executed. > > + * @x1: Passed as the first argument to @fn. > > + * @x2: Passed as the second argument to @fn. > > + * @x3: Passed as the third argument to @fn. > > + */ > > + > > +#define HVC_CALL_FUNC 4 > > + > > Can't we use the HVC_KVM_CALL_HYP for this as well? I thought we already > added the code to the stub to do that in the last patch. The last patch (arm64: Convert hcalls to use ISS field) only added function call support to the kvm vector. > Is there a difference between the two that I'm missing? The kvm and stub el1_sync vectors are different in that the kvm vector expects a stack to be set up whereas the stub vector does not. We have plenty of hcall numbers, so I thought it better to have two different numbers so improper usage would be easier to detect. > > #ifndef __ASSEMBLY__ > > > > /* > > diff --git a/arch/arm64/kernel/hyp-stub.S b/arch/arm64/kernel/hyp-stub.S > > index 9ab5f70..a21cf51 100644 > > --- a/arch/arm64/kernel/hyp-stub.S > > +++ b/arch/arm64/kernel/hyp-stub.S > > @@ -75,7 +75,17 @@ el1_sync: > > 1: cmp x10, #HVC_SET_VECTORS > > b.ne 1f > > msr vbar_el2, x0 > > + b 2f > > > > +1: cmp x10, #HVC_CALL_FUNC > > + b.ne 1f > > + mov x29, lr > > What's the contract for functions we call through the stub? My thinking was that this is just a mechanism to do the call. The caller and callee would need to set the conditions of the call. The only restriction is that the called function would need to preserve the register that lr is saved in (x29 here, but x18 in the current implementation). -Geoff
Hi Mark, On Mon, 2014-09-15 at 20:06 +0100, Mark Rutland wrote: > Hi Geoff, > > On Tue, Sep 09, 2014 at 11:49:05PM +0100, Geoff Levand wrote: > > The current implementation of ipi_cpu_stop() is just a tight infinite loop > > around cpu_relax(). This infinite loop implementation is OK if the machine > > will soon do a poweroff, but it doesn't have any mechanism to allow a CPU > > to be brought back on-line, nor is it compatible with kexec re-boot. > > I don't see why we should use this when we have disable_nonboot_cpus. I decided to use disable_nonboot_cpus, so this patch is no longer needed. I'll put the rework of disable_nonboot_cpus to not depend on CONFIG_PM_SLEEP_SMP on my todo list. -Geoff
Hi Mark, On Thu, 2014-09-18 at 02:13 +0100, Mark Rutland wrote: > On Tue, Sep 09, 2014 at 11:49:05PM +0100, Geoff Levand wrote: > > +++ b/arch/arm64/include/asm/kexec.h > > @@ -0,0 +1,52 @@ > > +/* > > + * kexec for arm64 > > + * > > + * Copyright (C) Linaro. > > + * > > + * This program is free software; you can redistribute it and/or modify > > + * it under the terms of the GNU General Public License version 2 as > > + * published by the Free Software Foundation. > > + */ > > + > > +#if !defined(_ARM64_KEXEC_H) > > +#define _ARM64_KEXEC_H > > + > > +/* Maximum physical address we can use pages from */ > > + > > +#define KEXEC_SOURCE_MEMORY_LIMIT (-1UL) > > + > > +/* Maximum address we can reach in physical address mode */ > > + > > +#define KEXEC_DESTINATION_MEMORY_LIMIT (-1UL) > > + > > +/* Maximum address we can use for the control code buffer */ > > + > > +#define KEXEC_CONTROL_MEMORY_LIMIT (-1UL) > > + > > What are these used for? I see that other architectures seem to do the > same thing, but they look odd. They need to be defined for the core kexec code, but arm64 doesn't use them. > > +#define KEXEC_CONTROL_PAGE_SIZE 4096 > > What's this used for? This is the size reserved for the reboot_code_buffer defined in kexec's core code. For arm64, we copy our relocate_new_kernel routine into the reboot_code_buffer. > Does this work with 64k pages, and is there any reason we can't figure > out the actual size of the code (so we don't get bitten if it grows)? Kexec will reserve pages to satisfy KEXEC_CONTROL_PAGE_SIZE, so for all arm64 page configs one page will be reserved for this value (4096). I have a check if relocate_new_kernel is too big '.org KEXEC_CONTROL_PAGE_SIZE' in the latest implementation. > > + > > +#define KEXEC_ARCH KEXEC_ARCH_ARM64 > > + > > +#define ARCH_HAS_KIMAGE_ARCH > > + > > +#if !defined(__ASSEMBLY__) > > + > > +struct kimage_arch { > > + void *ctx; > > +}; > > + > > +/** > > + * crash_setup_regs() - save registers for the panic kernel > > + * > > + * @newregs: registers are saved here > > + * @oldregs: registers to be saved (may be %NULL) > > + */ > > + > > +static inline void crash_setup_regs(struct pt_regs *newregs, > > + struct pt_regs *oldregs) > > +{ > > +} > > It would be nice to know what we're going to do for this. > > Is this a required function, or can we get away without crash kernel > support for the moment? This is just to avoid a build error. It is not used for kexec re-boot. > > + > > +#endif /* !defined(__ASSEMBLY__) */ > > + > > +#endif > > diff --git a/arch/arm64/kernel/Makefile b/arch/arm64/kernel/Makefile > > index df7ef87..8b7c029 100644 > > --- a/arch/arm64/kernel/Makefile > > +++ b/arch/arm64/kernel/Makefile > > @@ -29,6 +29,8 @@ arm64-obj-$(CONFIG_ARM64_CPU_SUSPEND) += sleep.o suspend.o > > arm64-obj-$(CONFIG_JUMP_LABEL) += jump_label.o > > arm64-obj-$(CONFIG_KGDB) += kgdb.o > > arm64-obj-$(CONFIG_EFI) += efi.o efi-stub.o efi-entry.o > > +arm64-obj-$(CONFIG_KEXEC) += machine_kexec.o relocate_kernel.o \ > > + cpu-properties.o > > > > obj-y += $(arm64-obj-y) vdso/ > > obj-m += $(arm64-obj-m) > > diff --git a/arch/arm64/kernel/machine_kexec.c b/arch/arm64/kernel/machine_kexec.c > > new file mode 100644 > > index 0000000..043a3bc > > --- /dev/null > > +++ b/arch/arm64/kernel/machine_kexec.c > > @@ -0,0 +1,612 @@ > > +/* > > + * kexec for arm64 > > + * > > + * Copyright (C) Linaro. > > + * > > + * This program is free software; you can redistribute it and/or modify > > + * it under the terms of the GNU General Public License version 2 as > > + * published by the Free Software Foundation. > > + */ > > + > > +#include <linux/kernel.h> > > +#include <linux/kexec.h> > > +#include <linux/of_fdt.h> > > +#include <linux/slab.h> > > +#include <linux/uaccess.h> > > + > > +#include <asm/cacheflush.h> > > +#include <asm/cpu_ops.h> > > +#include <asm/system_misc.h> > > + > > +#include "cpu-properties.h" > > + > > +#if defined(DEBUG) > > +static const int debug = 1; > > +#else > > +static const int debug; > > +#endif > > I don't think we need this. I put the debug output into another patch, which I'll decide to post or not later. > > + > > +typedef struct dtb_buffer {char b[0]; } dtb_t; > > It would be nice for this to be consistent with other dtb uses; if we > need a dtb type then it shouldn't be specific to kexec. This was to avoid errors due to the lack of type checking with void* types. I've reworked this in the latest patch. > > +static struct kexec_ctx *current_ctx; > > + > > +static int kexec_ctx_alloc(struct kimage *image) > > +{ > > + BUG_ON(image->arch.ctx); > > + > > + image->arch.ctx = kmalloc(sizeof(struct kexec_ctx), GFP_KERNEL); > > + > > + if (!image->arch.ctx) > > + return -ENOMEM; > > + > > + current_ctx = (struct kexec_ctx *)image->arch.ctx; > > This seems to be the only use of current_ctx. I take it this is a > leftover from debugging? > > [...] > > > +/** > > + * kexec_list_walk - Helper to walk the kimage page list. > > + */ > > Please keep this associated with the function it refers to (nothing > should be between this comment and the function prototype). > > > + > > +#define IND_FLAGS (IND_DESTINATION | IND_INDIRECTION | IND_DONE | IND_SOURCE) > > Can't this live in include/linux/kexec.h, where these flags are defined. I have a kexec patch submitted to clean this up. I'll re-factor this when that patch is upstream. https://lists.ozlabs.org/pipermail/linuxppc-dev/2014-August/120368.html > The meaning of these doesn't seem to be documented anywhere. Would you > be able to explain what each of these means? I think lack of comments/documentation is a general weakness of the core kexec code. > > +static void kexec_list_walk(void *ctx, unsigned long kimage_head, > > + void (*cb)(void *ctx, unsigned int flag, void *addr, void *dest)) > > +{ > > + void *dest; > > + unsigned long *entry; > > + > > + for (entry = &kimage_head, dest = NULL; ; entry++) { > > + unsigned int flag = *entry & IND_FLAGS; > > + void *addr = phys_to_virt(*entry & PAGE_MASK); > > + > > + switch (flag) { > > + case IND_INDIRECTION: > > + entry = (unsigned long *)addr - 1; > > + cb(ctx, flag, addr, NULL); > > + break; > > + case IND_DESTINATION: > > + dest = addr; > > + cb(ctx, flag, addr, NULL); > > + break; > > + case IND_SOURCE: > > + cb(ctx, flag, addr, dest); > > + dest += PAGE_SIZE; > > I really don't understand what's going on with dest here, but that's > probably because I don't understand the meaning of the flags. IND_SOURCE means the entry is a page of the current segment. dest is the address of that page. When you have a new source page the destination is post incremented. Think foo(src, dest++). > > + break; > > + case IND_DONE: > > + cb(ctx, flag , NULL, NULL); > > + return; > > + default: > > + pr_devel("%s:%d unknown flag %xh\n", __func__, __LINE__, > > + flag); > > Wouldn't pr_warn would be more appropriate here? We don't really don't need a message since the IND_ flags are well established. I'll remove this. > > > + cb(ctx, flag, addr, NULL); > > + break; > > + } > > + } > > +} > > + > > +/** > > + * kexec_image_info - For debugging output. > > + */ > > + > > +#define kexec_image_info(_i) _kexec_image_info(__func__, __LINE__, _i) > > +static void _kexec_image_info(const char *func, int line, > > + const struct kimage *image) > > +{ > > + if (debug) { > > + unsigned long i; > > + > > + pr_devel("%s:%d:\n", func, line); > > + pr_devel(" kexec image info:\n"); > > + pr_devel(" type: %d\n", image->type); > > + pr_devel(" start: %lx\n", image->start); > > + pr_devel(" head: %lx\n", image->head); > > + pr_devel(" nr_segments: %lu\n", image->nr_segments); > > + > > + for (i = 0; i < image->nr_segments; i++) { > > + pr_devel(" segment[%lu]: %016lx - %016lx, " > > + "%lxh bytes, %lu pages\n", > > + i, > > + image->segment[i].mem, > > + image->segment[i].mem + image->segment[i].memsz, > > + image->segment[i].memsz, > > + image->segment[i].memsz / PAGE_SIZE); > > + > > + if (kexec_is_dtb_user(image->segment[i].buf)) > > + pr_devel(" dtb segment\n"); > > + } > > + } > > +} > > pr_devel is already dependent on DEBUG, so surely we don't need to check > the debug variable? I'm not sure how much of this would be removed as dead code. If the compiler is cleaver enough it all should be. > > + > > +/** > > + * kexec_find_dtb_seg - Helper routine to find the dtb segment. > > + */ > > + > > +static const struct kexec_segment *kexec_find_dtb_seg( > > + const struct kimage *image) > > +{ > > + int i; > > + > > + for (i = 0; i < image->nr_segments; i++) { > > + if (kexec_is_dtb_user(image->segment[i].buf)) > > + return &image->segment[i]; > > + } > > + > > + return NULL; > > +} > > I'm really not keen on having the kernel guess which blobs need special > treatment, though we seem to do that for arm. Yes, to pass the dtb in r0 when th new kernel is entered. > It would be far nicer if we could pass flags for each segment to > describe what it is (e.g. kernel image, dtb, other binary blob), Well, we do pass a flag of sorts, the DT magic value. > so we > can do things like pass multiple DTBs (so we load two kernels at once > and pass each a unique DTB if we want to boot a new kernel + crashkernel > pair). Unfortunately that would require some fairly invasive rework of > the kexec core. I don't think I'll attempt that any time soon. Feel free to give it a try. > For secureboot we can't trust a dtb from userspace, and will have to use > kexec_file_load. To work with that we can either: > > * Reuse the original DTB, patched with the new command line. This may > have statefulness issues (for things like simplefb). > > * Build a new DTB by flattening the current live tree. This would rely > on drivers that modify state to patch the tree appropriately. I have not yet looked into how to do this yet. > [...] > > > +/** > > + * kexec_cpu_info_init - Initialize an array of kexec_cpu_info structures. > > + * > > + * Allocates a cpu info array and fills it with info for all cpus found in > > + * the device tree passed. > > + */ > > + > > +static int kexec_cpu_info_init(const struct device_node *dn, > > + struct kexec_boot_info *info) > > +{ > > + int result; > > + unsigned int cpu; > > + > > + info->cp = kmalloc( > > + info->cpu_count * sizeof(*info->cp), GFP_KERNEL); > > + > > + if (!info->cp) { > > + pr_err("%s: Error: Out of memory.", __func__); > > + return -ENOMEM; > > + } > > + > > + for (cpu = 0; cpu < info->cpu_count; cpu++) { > > + struct cpu_properties *cp = &info->cp[cpu]; > > + > > + dn = of_find_node_by_type((struct device_node *)dn, "cpu"); > > + > > + if (!dn) { > > + pr_devel("%s:%d: bad node\n", __func__, __LINE__); > > + goto on_error; > > + } > > + > > + result = read_cpu_properties(cp, dn); > > + > > + if (result) { > > + pr_devel("%s:%d: bad node\n", __func__, __LINE__); > > + goto on_error; > > + } > > + > > + if (cp->type == cpu_enable_method_psci) > > + pr_devel("%s:%d: cpu-%u: hwid-%llx, '%s'\n", > > + __func__, __LINE__, cpu, cp->hwid, > > + cp->enable_method); > > + else > > + pr_devel("%s:%d: cpu-%u: hwid-%llx, '%s', " > > + "cpu-release-addr %llx\n", > > + __func__, __LINE__, cpu, cp->hwid, > > + cp->enable_method, > > + cp->cpu_release_addr); > > + } > > + > > + return 0; > > + > > +on_error: > > + kfree(info->cp); > > + info->cp = NULL; > > + return -EINVAL; > > +} > > I don't see why we should need this at all. If we use the hotplug > infrastructure, we don't need access to the enable-method and related > properties, and the kexec code need only deal with a single CPU. I removed all the checking in the latest patch. > The only case where kexec needs to deal with other CPUs is when some are > sat in the holding pen, but this code doesn't seem to handle that. > > as I believe I mentioned before, we should be able to extend the holding > pen code to get those CPUs to increment a sat-in-pen counter and if > that's non-zero after SMP bringup we print a warning (and disallow > kexec). I have some work-in-progress patches that try to do this, but I will not include those in this series. See my spin-table branch: https://git.linaro.org/people/geoff.levand/linux-kexec.git > > +/** > > +* kexec_compat_check - Iterator for kexec_cpu_check. > > +*/ > > + > > +static int kexec_compat_check(const struct kexec_ctx *ctx) > > +{ > > + unsigned int cpu_1; > > + unsigned int to_process; > > + > > + to_process = min(ctx->first.cpu_count, ctx->second.cpu_count); > > + > > + if (ctx->first.cpu_count != ctx->second.cpu_count) > > + pr_warn("%s: Warning: CPU count mismatch %u != %u.\n", > > + __func__, ctx->first.cpu_count, ctx->second.cpu_count); > > + > > + for (cpu_1 = 0; cpu_1 < ctx->first.cpu_count; cpu_1++) { > > + unsigned int cpu_2; > > + struct cpu_properties *cp_1 = &ctx->first.cp[cpu_1]; > > + > > + for (cpu_2 = 0; cpu_2 < ctx->second.cpu_count; cpu_2++) { > > + struct cpu_properties *cp_2 = &ctx->second.cp[cpu_2]; > > + > > + if (cp_1->hwid != cp_2->hwid) > > + continue; > > + > > + if (!kexec_cpu_check(cp_1, cp_2)) > > + return -EINVAL; > > + > > + to_process--; > > + } > > + } > > + > > + if (to_process) { > > + pr_warn("%s: Warning: Failed to process %u CPUs.\n", __func__, > > + to_process); > > + return -EINVAL; > > + } > > + > > + return 0; > > +} > > I don't see the point in checking this in the kernel. If I pass the > second stage kernel a new dtb where my enable methods are different, > that was my choice as a user. If that doesn't work, that's my fault. > > There are plenty of other things that might be completely different that > we don't sanity check, so I don't see why enable methods should be any > different. > > [...] > > > +/** > > + * kexec_check_cpu_die - Check if cpu_die() will work on secondary processors. > > + */ > > + > > +static int kexec_check_cpu_die(void) > > +{ > > + unsigned int cpu; > > + unsigned int sum = 0; > > + > > + /* For simplicity this also checks the primary CPU. */ > > + > > + for_each_cpu(cpu, cpu_all_mask) { > > + if (cpu && (!cpu_ops[cpu] || !cpu_ops[cpu]->cpu_disable || > > + cpu_ops[cpu]->cpu_disable(cpu))) { > > + sum++; > > + pr_err("%s: Error: " > > + "CPU %u does not support hot un-plug.\n", > > + __func__, cpu); > > + } > > + } > > + > > + return sum ? -EOPNOTSUPP : 0; > > +} > > We should really use disable_nonboot_cpus() for this. That way we don't > end up with a slightly different hotplug implementation for kexec. The > above is missing cpu_kill calls, for example, and I'm worried by the > possibility of further drift over time. > > I understand from our face-to-face discussion that you didn't want to > require the PM infrastructure that disable_nonboot_cpus currently pulls > in due to the being dependent on CONFIG_PM_SLEEP_SMP which selects > CONFIG_PM_SLEEP and so on. The solution to that is to refactor the > Kconfig so we can have disable_nonboot_cpus without all the other PM > infrastructure. I switch the current patch to use disable_nonboot_cpus(). > > + > > +/** > > + * machine_kexec_prepare - Prepare for a kexec reboot. > > + * > > + * Called from the core kexec code when a kernel image is loaded. > > + */ > > + > > +int machine_kexec_prepare(struct kimage *image) > > +{ > > + int result; > > This seems to always be an error code. Call it 'err'. > > > + dtb_t *dtb = NULL; > > + struct kexec_ctx *ctx; > > + const struct kexec_segment *dtb_seg; > > + > > + kexec_image_info(image); > > + > > + result = kexec_check_cpu_die(); > > + > > + if (result) > > + goto on_error; > > + > > + result = kexec_ctx_alloc(image); > > + > > + if (result) > > + goto on_error; > > + > > + ctx = kexec_image_to_ctx(image); > > + > > + result = kexec_boot_info_init(&ctx->first, NULL); > > + > > + if (result) > > + goto on_error; > > + > > + dtb_seg = kexec_find_dtb_seg(image); > > + > > + if (!dtb_seg) { > > + result = -EINVAL; > > + goto on_error; > > + } > > + > > + result = kexec_copy_dtb(dtb_seg, &dtb); > > + > > + if (result) > > + goto on_error; > > + > > + result = kexec_boot_info_init(&ctx->second, dtb); > > + > > + if (result) > > + goto on_error; > > + > > + result = kexec_compat_check(ctx); > > + > > + if (result) > > + goto on_error; > > + > > + kexec_dtb_addr = dtb_seg->mem; > > + kexec_kimage_start = image->start; > > + > > + goto on_exit; > > + > > +on_error: > > + kexec_ctx_clean(image); > > +on_exit: > > on_* looks weird, and doesn't match the style of other labels in > arch/arm64. Could we call these 'out_clean' and 'out' instead? > > > + kfree(dtb); > > + return result; > > +} > > + > > +/** > > + * kexec_list_flush_cb - Callback to flush the kimage list to PoC. > > + */ > > + > > +static void kexec_list_flush_cb(void *ctx , unsigned int flag, > > + void *addr, void *dest) > > +{ > > + switch (flag) { > > + case IND_INDIRECTION: > > + case IND_SOURCE: > > + __flush_dcache_area(addr, PAGE_SIZE); > > Is PAGE_SIZE always big enough? Do we not have a more accurate size? > Perhaps I've misunderstood what's going on here. The image list is a list of pages, so PAGE_SIZE should be OK. > > + break; > > + default: > > + break; > > + } > > +} > > + > > +/** > > + * machine_kexec - Do the kexec reboot. > > + * > > + * Called from the core kexec code for a sys_reboot with LINUX_REBOOT_CMD_KEXEC. > > + */ > > + > > +void machine_kexec(struct kimage *image) > > +{ > > + phys_addr_t reboot_code_buffer_phys; > > + void *reboot_code_buffer; > > + struct kexec_ctx *ctx = kexec_image_to_ctx(image); > > + > > + BUG_ON(relocate_new_kernel_size > KEXEC_CONTROL_PAGE_SIZE); > > It looks like relocate_new_kernel_size is a build-time constant. If we > need that to be less than KEXEC_CONTROL_PAGE_SIZE, then we should make > that a build-time check. I moved this check into relocate_new_kernel with a '.org KEXEC_CONTROL_PAGE_SIZE'. > > + BUG_ON(num_online_cpus() > 1); > > + BUG_ON(!ctx); > > + > > + kexec_image_info(image); > > + > > + kexec_kimage_head = image->head; > > + > > + reboot_code_buffer_phys = page_to_phys(image->control_code_page); > > + reboot_code_buffer = phys_to_virt(reboot_code_buffer_phys); > > + > > + pr_devel("%s:%d: control_code_page: %p\n", __func__, __LINE__, > > + (void *)image->control_code_page); > > This is already a pointer. Is the cast to void necessary? > > > + pr_devel("%s:%d: reboot_code_buffer_phys: %p\n", __func__, __LINE__, > > + (void *)reboot_code_buffer_phys); > > Use %pa and pass &reboot_code_buffer_phys, no cast necessary. > > > + pr_devel("%s:%d: reboot_code_buffer: %p\n", __func__, __LINE__, > > + reboot_code_buffer); > > + pr_devel("%s:%d: relocate_new_kernel: %p\n", __func__, __LINE__, > > + relocate_new_kernel); > > + pr_devel("%s:%d: relocate_new_kernel_size: %lxh(%lu) bytes\n", __func__, > > + __LINE__, relocate_new_kernel_size, relocate_new_kernel_size); > > Please use an '0x' prefix rather than a 'h' suffix. Do we need in print > in both hex and decimal? > > > + > > + pr_devel("%s:%d: kexec_dtb_addr: %p\n", __func__, __LINE__, > > + (void *)kexec_dtb_addr); > > + pr_devel("%s:%d: kexec_kimage_head: %p\n", __func__, __LINE__, > > + (void *)kexec_kimage_head); > > + pr_devel("%s:%d: kexec_kimage_start: %p\n", __func__, __LINE__, > > + (void *)kexec_kimage_start); > > These are all unsigned long, so why not use the existing mechanism for > printing unsigned long? > > > + > > + /* > > + * Copy relocate_new_kernel to the reboot_code_buffer for use > > + * after the kernel is shut down. > > + */ > > + > > + memcpy(reboot_code_buffer, relocate_new_kernel, > > + relocate_new_kernel_size); > > + > > + /* Assure reboot_code_buffer is copied. */ > > + > > + mb(); > > I don't think we need the mb if this is only to guarantee completion > before the cache flush -- cacheable memory accesses should hazard > against cache flushes by VA. OK. > > + > > + pr_info("Bye!\n"); > > + > > + local_disable(DAIF_ALL); > > We can move these two right before the soft_restart, after the cache > maintenance. That way the print is closer to the exit of the current > kernel. OK. > > + > > + /* Flush the reboot_code_buffer in preparation for its execution. */ > > + > > + __flush_dcache_area(reboot_code_buffer, relocate_new_kernel_size); > > + > > + /* Flush the kimage list. */ > > + > > + kexec_list_walk(NULL, image->head, kexec_list_flush_cb); > > + > > + soft_restart(reboot_code_buffer_phys); > > +} > > + > > +void machine_crash_shutdown(struct pt_regs *regs) > > +{ > > + /* Empty routine needed to avoid build errors. */ > > +} > > diff --git a/arch/arm64/kernel/relocate_kernel.S b/arch/arm64/kernel/relocate_kernel.S > > new file mode 100644 > > index 0000000..92aba9d > > --- /dev/null > > +++ b/arch/arm64/kernel/relocate_kernel.S > > @@ -0,0 +1,185 @@ > > +/* > > + * kexec for arm64 > > + * > > + * Copyright (C) Linaro. > > + * > > + * This program is free software; you can redistribute it and/or modify > > + * it under the terms of the GNU General Public License version 2 as > > + * published by the Free Software Foundation. > > + */ > > + > > +#include <asm/assembler.h> > > +#include <asm/memory.h> > > +#include <asm/page.h> > > + > > +/* The list entry flags. */ > > + > > +#define IND_DESTINATION_BIT 0 > > +#define IND_INDIRECTION_BIT 1 > > +#define IND_DONE_BIT 2 > > +#define IND_SOURCE_BIT 3 > > Given these ned to match the existing IND_* flags in > include/linux/kexec.h, and they aren't in any way specific to arm64, > please put these ina an asm-generic header and redefine the existing > IND_* flags in terms of them. See my patch that does that: https://lists.ozlabs.org/pipermail/linuxppc-dev/2014-August/120368.html > > > + > > +/* > > + * relocate_new_kernel - Put the 2nd stage kernel image in place and boot it. > > + * > > + * The memory that the old kernel occupies may be overwritten when coping the > > + * new kernel to its final location. To assure that the relocate_new_kernel > > + * routine which does that copy is not overwritten all code and data needed > > + * by relocate_new_kernel must be between the symbols relocate_new_kernel and > > + * relocate_new_kernel_end. The machine_kexec() routine will copy > > + * relocate_new_kernel to the kexec control_code_page, a special page which > > + * has been set up to be preserved during the kernel copy operation. > > + */ > > + > > +.align 3 > > Surely this isn't necessary? No, the code should be properly aligned. > > + > > +.globl relocate_new_kernel > > +relocate_new_kernel: > > + > > + /* Setup the list loop variables. */ > > + > > + ldr x10, kexec_kimage_head /* x10 = list entry */ > > Any reason for using x10 rather than starting with x0? Or x18, if you > need to preserve the low registers? > > > + > > + mrs x0, ctr_el0 > > + ubfm x0, x0, #16, #19 > > + mov x11, #4 > > + lsl x11, x11, x0 /* x11 = dcache line size */ > > Any reason we can't use dcache_line_size, given it's a macro? > > > + > > + mov x12, xzr /* x12 = segment start */ > > + mov x13, xzr /* x13 = entry ptr */ > > + mov x14, xzr /* x14 = copy dest */ > > + > > + /* Check if the new kernel needs relocation. */ > > + > > + cbz x10, .Ldone > > + tbnz x10, IND_DONE_BIT, .Ldone > > + > > +.Lloop: > > Is there any reason for the '.L' on all of these? We only seem to do > that in the lib code that was imported from elsewhere, and it doesn't > match the rest of the arm64 asm. .L is a local label prefix in gas. I don't think it would be good to have these with larger scope. > > + and x15, x10, PAGE_MASK /* x15 = addr */ > > + > > + /* Test the entry flags. */ > > + > > +.Ltest_source: > > + tbz x10, IND_SOURCE_BIT, .Ltest_indirection > > + > > + /* copy_page(x20 = dest, x21 = src) */ > > + > > + mov x20, x14 > > + mov x21, x15 > > + > > +1: ldp x22, x23, [x21] > > + ldp x24, x25, [x21, #16] > > + ldp x26, x27, [x21, #32] > > + ldp x28, x29, [x21, #48] > > + add x21, x21, #64 > > + stnp x22, x23, [x20] > > + stnp x24, x25, [x20, #16] > > + stnp x26, x27, [x20, #32] > > + stnp x28, x29, [x20, #48] > > + add x20, x20, #64 > > + tst x21, #(PAGE_SIZE - 1) > > + b.ne 1b > > It's a shame we can't reuse copy_page directly. Could we not move the > body to a macro we can reuse here? copy_page() also does some memory pre-fetch, which Arun said caused problems on the APM board. If that board were available to me for testing I could investigate, but at this time I will put this suggestion on my todo list. > > + > > + /* dest += PAGE_SIZE */ > > + > > + add x14, x14, PAGE_SIZE > > + b .Lnext > > + > > +.Ltest_indirection: > > + tbz x10, IND_INDIRECTION_BIT, .Ltest_destination > > + > > + /* ptr = addr */ > > + > > + mov x13, x15 > > + b .Lnext > > + > > +.Ltest_destination: > > + tbz x10, IND_DESTINATION_BIT, .Lnext > > + > > + /* flush segment */ > > + > > + bl .Lflush > > + mov x12, x15 > > + > > + /* dest = addr */ > > + > > + mov x14, x15 > > + > > +.Lnext: > > + /* entry = *ptr++ */ > > + > > + ldr x10, [x13] > > + add x13, x13, 8 > > This can be: > > ldr x10, [x13], #8 > > > + > > + /* while (!(entry & DONE)) */ > > + > > + tbz x10, IND_DONE_BIT, .Lloop > > + > > +.Ldone: > > + /* flush last segment */ > > + > > + bl .Lflush > > + > > + dsb sy > > + isb > > + ic ialluis > > This doesn't look right; we need a dsb and an isb after the instruction > cache maintenance (or the icache could still be flushing when we branch > to the new kernel). OK. > > + > > + /* start_new_kernel */ > > + > > + ldr x4, kexec_kimage_start > > + ldr x0, kexec_dtb_addr > > + mov x1, xzr > > + mov x2, xzr > > + mov x3, xzr > > + br x4 > > + > > +/* flush - x11 = line size, x12 = start addr, x14 = end addr. */ > > + > > +.Lflush: > > + cbz x12, 2f > > + mov x0, x12 > > + sub x1, x11, #1 > > + bic x0, x0, x1 > > +1: dc civac, x0 > > + add x0, x0, x11 > > + cmp x0, x14 > > + b.lo 1b > > +2: ret > > It would be nice if this were earlier in the file, before its callers. Then we would need to jump over it, which I don't think is very clean. > > > + > > +.align 3 > > We should have a comment as to why this is needed (to keep the 64-bit > values below naturally aligned). I haven't seen such an .align directive comment in any arm64 code yet. -Geoff