Message ID | 20210527150526.271941-4-pasha.tatashin@soleen.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | arm64: MMU enabled kexec relocation | expand |
On Thu, 27 May 2021 16:05:11 +0100, Pavel Tatashin <pasha.tatashin@soleen.com> wrote: > > The hyp-stub's elx_sync code fits in the vector. > > With this, all of the hyp-stubs behaviour is contained in its vectors. > This lets kexec and hibernate copy the hyp-stub when they need its > behaviour, instead of re-implementing it. > > Co-developed-by: James Morse <james.morse@arm.com> > Signed-off-by: Pavel Tatashin <pasha.tatashin@soleen.com> > --- > arch/arm64/kernel/hyp-stub.S | 64 +++++++++++++++++++----------------- > 1 file changed, 33 insertions(+), 31 deletions(-) > > diff --git a/arch/arm64/kernel/hyp-stub.S b/arch/arm64/kernel/hyp-stub.S > index 18a97bee3779..86af6c4e52b9 100644 > --- a/arch/arm64/kernel/hyp-stub.S > +++ b/arch/arm64/kernel/hyp-stub.S > @@ -21,6 +21,37 @@ SYM_CODE_START_LOCAL(\label) > .align 7 > b \label > SYM_CODE_END(\label) > +.endm > + > +.macro elx_sync_vector label > +SYM_CODE_START_LOCAL(\label) > + .align 7 > + cmp x0, #HVC_SET_VECTORS > + b.ne 1f > + msr vbar_el2, x1 > + b 9f > + > +1: cmp x0, #HVC_VHE_RESTART > + b.eq mutate_to_vhe Now that this has turned into a macro, what are the guarantees that mutate_to_vhe will be within reach of the site where this macro is expanded? It does work here, but what about the other expansion sites that will show up later in the series? What was wrong with directly branching to the original call site? Nothing in the commit message explains it. Thanks, M.
On Thu, May 27, 2021 at 11:54 AM Marc Zyngier <maz@kernel.org> wrote: > > On Thu, 27 May 2021 16:05:11 +0100, > Pavel Tatashin <pasha.tatashin@soleen.com> wrote: > > > > The hyp-stub's elx_sync code fits in the vector. > > > > With this, all of the hyp-stubs behaviour is contained in its vectors. > > This lets kexec and hibernate copy the hyp-stub when they need its > > behaviour, instead of re-implementing it. > > > > Co-developed-by: James Morse <james.morse@arm.com> > > Signed-off-by: Pavel Tatashin <pasha.tatashin@soleen.com> > > --- > > arch/arm64/kernel/hyp-stub.S | 64 +++++++++++++++++++----------------- > > 1 file changed, 33 insertions(+), 31 deletions(-) > > > > diff --git a/arch/arm64/kernel/hyp-stub.S b/arch/arm64/kernel/hyp-stub.S > > index 18a97bee3779..86af6c4e52b9 100644 > > --- a/arch/arm64/kernel/hyp-stub.S > > +++ b/arch/arm64/kernel/hyp-stub.S > > @@ -21,6 +21,37 @@ SYM_CODE_START_LOCAL(\label) > > .align 7 > > b \label > > SYM_CODE_END(\label) > > +.endm > > + > > +.macro elx_sync_vector label > > +SYM_CODE_START_LOCAL(\label) > > + .align 7 > > + cmp x0, #HVC_SET_VECTORS > > + b.ne 1f > > + msr vbar_el2, x1 > > + b 9f > > + > > +1: cmp x0, #HVC_VHE_RESTART > > + b.eq mutate_to_vhe > > Now that this has turned into a macro, what are the guarantees that > mutate_to_vhe will be within reach of the site where this macro is > expanded? It does work here, but what about the other expansion sites > that will show up later in the series? > > What was wrong with directly branching to the original call site? > Nothing in the commit message explains it. Hi Marc, I need to explain this better in the commit log. Later in the series we create our own vector copy that is outside of the old and new kernel so it does not get overwritten during kexec relocation. When VHE is enabled, the vector is passed so we can switch to el2 before jumping to the new kernel. arm64_relocate_new_kernel() which performs the relocation runs with MMU being enabled until relocation is done, after that disables MMU and if vectors are passed performs: mov x0, #HVC_SOFT_RESTART hvc #0 /* Jumps from el2 */ It cannot call mutate_to_vhe because #HVC_VHE_RESTART is not used here. But, if it had to it would not work as we cannot return to the old kernel text after relocation. Thanks, Pasha > > Thanks, > > M. > > -- > Without deviation from the norm, progress is not possible.
On Thu, 27 May 2021 22:23:22 +0100, Pavel Tatashin <pasha.tatashin@soleen.com> wrote: > > On Thu, May 27, 2021 at 11:54 AM Marc Zyngier <maz@kernel.org> wrote: > > > > On Thu, 27 May 2021 16:05:11 +0100, > > Pavel Tatashin <pasha.tatashin@soleen.com> wrote: > > > > > > The hyp-stub's elx_sync code fits in the vector. > > > > > > With this, all of the hyp-stubs behaviour is contained in its vectors. > > > This lets kexec and hibernate copy the hyp-stub when they need its > > > behaviour, instead of re-implementing it. > > > > > > Co-developed-by: James Morse <james.morse@arm.com> > > > Signed-off-by: Pavel Tatashin <pasha.tatashin@soleen.com> > > > --- > > > arch/arm64/kernel/hyp-stub.S | 64 +++++++++++++++++++----------------- > > > 1 file changed, 33 insertions(+), 31 deletions(-) > > > > > > diff --git a/arch/arm64/kernel/hyp-stub.S b/arch/arm64/kernel/hyp-stub.S > > > index 18a97bee3779..86af6c4e52b9 100644 > > > --- a/arch/arm64/kernel/hyp-stub.S > > > +++ b/arch/arm64/kernel/hyp-stub.S > > > @@ -21,6 +21,37 @@ SYM_CODE_START_LOCAL(\label) > > > .align 7 > > > b \label > > > SYM_CODE_END(\label) > > > +.endm > > > + > > > +.macro elx_sync_vector label > > > +SYM_CODE_START_LOCAL(\label) > > > + .align 7 > > > + cmp x0, #HVC_SET_VECTORS > > > + b.ne 1f > > > + msr vbar_el2, x1 > > > + b 9f > > > + > > > +1: cmp x0, #HVC_VHE_RESTART > > > + b.eq mutate_to_vhe > > > > Now that this has turned into a macro, what are the guarantees that > > mutate_to_vhe will be within reach of the site where this macro is > > expanded? It does work here, but what about the other expansion sites > > that will show up later in the series? > > > > What was wrong with directly branching to the original call site? > > Nothing in the commit message explains it. > > Hi Marc, > > I need to explain this better in the commit log. Later in the series > we create our own vector copy that is outside of the old and new > kernel so it does not get overwritten during kexec relocation. When > VHE is enabled, the vector is passed so we can switch to el2 before > jumping to the new kernel. > > arm64_relocate_new_kernel() which performs the relocation runs with > MMU being enabled until relocation is done, after that disables MMU > and if vectors are passed performs: > > mov x0, #HVC_SOFT_RESTART > hvc #0 /* Jumps from el2 */ > > It cannot call mutate_to_vhe because #HVC_VHE_RESTART is not used > here. But, if it had to it would not work as we cannot return to the > old kernel text after relocation. OK, so you are happy with having a dangling branch pointing to nowhere? Something in me screams that it isn't a good idea, in general. If HVC_SOFT_RESTART is all you need, I'd rather you have a small stub that implements exactly that and nothing else. Feel free to extract it as a reusable macro if you want. Thanks, M.
> > It cannot call mutate_to_vhe because #HVC_VHE_RESTART is not used > > here. But, if it had to it would not work as we cannot return to the > > old kernel text after relocation. > > OK, so you are happy with having a dangling branch pointing to > nowhere? Something in me screams that it isn't a good idea, in > general. > > If HVC_SOFT_RESTART is all you need, I'd rather you have a small stub > that implements exactly that and nothing else. Feel free to extract it > as a reusable macro if you want. Sure, I will add in the next revision. Thanks, Pasha
On Tue, Jun 1, 2021 at 9:18 PM Pavel Tatashin <pasha.tatashin@soleen.com> wrote: > > > > It cannot call mutate_to_vhe because #HVC_VHE_RESTART is not used > > > here. But, if it had to it would not work as we cannot return to the > > > old kernel text after relocation. > > > > OK, so you are happy with having a dangling branch pointing to > > nowhere? Something in me screams that it isn't a good idea, in > > general. > > > > If HVC_SOFT_RESTART is all you need, I'd rather you have a small stub > > that implements exactly that and nothing else. Feel free to extract it > > as a reusable macro if you want. Using macro won't help here to save kernel text. Optimally, we would want to use only one vector table to reduce kernel text memory usage. I could do that by overwriting sync entries in trans_pgd_copy_el2_vectors(). But, that would be ugly, as I would need to have some specific assumptions about what entries need to be overwritten. Therefore, I decided to move the vector table that we currently have in hibernate code, and make it common between kexec and hibernate; trans_pgd_copy_el2_vectors() will use that table's body to create copies. Thanks, Pasha
diff --git a/arch/arm64/kernel/hyp-stub.S b/arch/arm64/kernel/hyp-stub.S index 18a97bee3779..86af6c4e52b9 100644 --- a/arch/arm64/kernel/hyp-stub.S +++ b/arch/arm64/kernel/hyp-stub.S @@ -21,6 +21,37 @@ SYM_CODE_START_LOCAL(\label) .align 7 b \label SYM_CODE_END(\label) +.endm + +.macro elx_sync_vector label +SYM_CODE_START_LOCAL(\label) + .align 7 + cmp x0, #HVC_SET_VECTORS + b.ne 1f + msr vbar_el2, x1 + b 9f + +1: cmp x0, #HVC_VHE_RESTART + b.eq mutate_to_vhe + +2: cmp x0, #HVC_SOFT_RESTART + b.ne 3f + mov x0, x2 + mov x2, x4 + mov x4, x1 + mov x1, x3 + br x4 // no return + +3: cmp x0, #HVC_RESET_VECTORS + beq 9f // Nothing to reset! + + /* Someone called kvm_call_hyp() against the hyp-stub... */ + mov_q x0, HVC_STUB_ERR + eret + +9: mov x0, xzr + eret +SYM_CODE_END(\label) .endm .text @@ -34,12 +65,12 @@ SYM_CODE_START(__hyp_stub_vectors) invalid_vector hyp_stub_el2t_fiq_invalid // FIQ EL2t invalid_vector hyp_stub_el2t_error_invalid // Error EL2t - ventry elx_sync // Synchronous EL2h + elx_sync_vector el2h_sync // Synchronous EL2h invalid_vector hyp_stub_el2h_irq_invalid // IRQ EL2h invalid_vector hyp_stub_el2h_fiq_invalid // FIQ EL2h invalid_vector hyp_stub_el2h_error_invalid // Error EL2h - ventry elx_sync // Synchronous 64-bit EL1 + elx_sync_vector el1_sync // Synchronous 64-bit EL1 invalid_vector hyp_stub_el1_irq_invalid // IRQ 64-bit EL1 invalid_vector hyp_stub_el1_fiq_invalid // FIQ 64-bit EL1 invalid_vector hyp_stub_el1_error_invalid // Error 64-bit EL1 @@ -55,35 +86,6 @@ SYM_CODE_END(__hyp_stub_vectors) # Check the __hyp_stub_vectors didn't overflow .org . - (__hyp_stub_vectors_end - __hyp_stub_vectors) + SZ_2K - -SYM_CODE_START_LOCAL(elx_sync) - cmp x0, #HVC_SET_VECTORS - b.ne 1f - msr vbar_el2, x1 - b 9f - -1: cmp x0, #HVC_VHE_RESTART - b.eq mutate_to_vhe - -2: cmp x0, #HVC_SOFT_RESTART - b.ne 3f - mov x0, x2 - mov x2, x4 - mov x4, x1 - mov x1, x3 - br x4 // no return - -3: cmp x0, #HVC_RESET_VECTORS - beq 9f // Nothing to reset! - - /* Someone called kvm_call_hyp() against the hyp-stub... */ - mov_q x0, HVC_STUB_ERR - eret - -9: mov x0, xzr - eret -SYM_CODE_END(elx_sync) - // nVHE? No way! Give me the real thing! SYM_CODE_START_LOCAL(mutate_to_vhe) // Sanity check: MMU *must* be off
The hyp-stub's elx_sync code fits in the vector. With this, all of the hyp-stubs behaviour is contained in its vectors. This lets kexec and hibernate copy the hyp-stub when they need its behaviour, instead of re-implementing it. Co-developed-by: James Morse <james.morse@arm.com> Signed-off-by: Pavel Tatashin <pasha.tatashin@soleen.com> --- arch/arm64/kernel/hyp-stub.S | 64 +++++++++++++++++++----------------- 1 file changed, 33 insertions(+), 31 deletions(-)