x86/virt/tdx: Further fix tdh_vp_enter() calls instrumentable code warning
Subsystems (2)
PATCH 1/1 x86/virt/tdx: Further fix tdh_vp_enter() calls instrumentable code warning
tdh_vp_enter() needs to be marked noinstr, which means it can't call any non-inlined noinstr functions. Commit e9f17038d814 ("x86/tdx: mark tdh_vp_enter() as __flatten") tried to address a build warning caused by tdx_tdvpr_pa() not getting inlined. Unfortunately that commit didn't fix the warning completely due to the inconsistent behavior of the __flatten annotation.
There are two problems that can come up depending on the compiler and config. One is that tdx_tdvpr_pa() doesn't get inlined, the other is that page_to_phys() doesn't get inlined.
The __flatten annotation makes the compiler inline all function calls that the annotated function makes, and the aforementioned commit assumed this is always honored, recursively. But it turns out it's not always true:
- Gcc may ignore __flatten when CONFIG_CC_OPTIMIZE_FOR_SIZE=y.
- Clang doesn't support recursive inlining for __flatten, which can trigger another similar warning when page_to_phys() calls pfn_valid() when CONFIG_DEBUG_VIRTUAL=y.
Therefore using __flatten is not the right fix.
To fix the first problem, remove the __flatten for tdh_vp_enter() and instead annotate tdx_tdvpr_pa() with __always_inline to make sure it is always inlined.
To fix the second problem, change tdx_tdvpr_pa() to use PFN_PHYS(page_to_pfn()) instead of page_to_phys() so that there will be no more function call inside tdx_tdvpr_pa()[*].
The TDVPR page is always an actual page out of page allocator, so the additional warning around pfn_valid() check in page_to_phys() doesn't help a lot anyway. It's not worth complicating the code for such warning when CONFIG_DEBUG_VIRTUAL=y.
[*] Since commit cba5d9b3e99d ("x86/mm/64: Make SPARSEMEM_VMEMMAP the
only memory model") page_to_pfn() has been a simple macro without any function call.
--- arch/x86/virt/vmx/tdx/tdx.c | 11 ++++++++--- 1 file changed, 8 insertions(+), 3 deletions(-)
arch/x86/virt/vmx/tdx/tdx.c +8 -3
@@ -1502,9 +1502,14 @@ static inline u64 tdx_tdr_pa(struct tdx_td *td)return page_to_phys (td->tdr_page); } -static inline u64 tdx_tdvpr_pa (struct tdx_vp *td) +static __always_inline u64 tdx_tdvpr_pa(struct tdx_vp *td) { -return page_to_phys (td->tdvpr_page); +/* +* Don't use page_to_phys() because tdh_vp_enter() calls this +* function from 'noinstr' code, and page_to_phys() can call +* uninlined functions on some compiler/configs. +*/ +return PFN_PHYS (page_to_pfn (td->tdvpr_page)); }/*
@@ -1518,7 +1523,7 @@ static void tdx_clflush_page(struct page *page)clflush_cache_range (page_to_virt (page),PAGE_SIZE ); } -noinstr __flatten u64 tdh_vp_enter(struct tdx_vp *td, struct tdx_module_args *args) +noinstr u64 tdh_vp_enter(struct tdx_vp *td, struct tdx_module_args *args) { args->rcx =tdx_tdvpr_pa (td);
base-commit: b7e21417e1f2c9b2d5c15b0a7d866e810de772aa
Signature · 1 line
2.49.0
Edgecombe, Rick P
maintainerQuoted 23 lines · Kai Huang
tdh_vp_enter()needs to be marked noinstr, which means it can't call any non-inlined noinstr functions. Commite9f17038d814("x86/tdx: marktdh_vp_enter()as__flatten") tried to address a build warning caused bytdx_tdvpr_pa()not getting inlined. Unfortunately that commit didn't fix the warning completely due to the inconsistent behavior of the__flattenannotation.There are two problems that can come up depending on the compiler and config. One is that
tdx_tdvpr_pa()doesn't get inlined, the other is thatpage_to_phys()doesn't get inlined.The
__flattenannotation makes the compiler inline all function calls that the annotated function makes, and the aforementioned commit assumed this is always honored, recursively. But it turns out it's not always true:- Gcc may ignore
__flattenwhenCONFIG_CC_OPTIMIZE_FOR_SIZE=y.
- Clang doesn't support recursive inlining for__flatten, which can trigger another similar warning whenpage_to_phys()callspfn_valid()whenCONFIG_DEBUG_VIRTUAL=y.Therefore using
__flattenis not the right fix.
Quoted 7 lines
To fix the first problem, remove the
__flattenfortdh_vp_enter()and instead annotatetdx_tdvpr_pa()with__always_inlineto make sure it is always inlined.To fix the second problem, change
tdx_tdvpr_pa()to use PFN_PHYS(page_to_pfn()) instead ofpage_to_phys()so that there will be no more function call inside tdx_tdvpr_pa()[*].
To check my understanding, page_to_pfn() on CONFIG_SPARSEMEM_VMEMMAP or CONFIG_FLATMEM has no function calls, but on CONFIG_SPARSEMEM, it does. We are counting on x86_64 to not use CONFIG_SPARSEMEM?
Quoted 6 lines
The TDVPR page is always an actual page out of page allocator, so the additional warning around
pfn_valid()check inpage_to_phys()doesn't help a lot anyway. It's not worth complicating the code for such warning whenCONFIG_DEBUG_VIRTUAL=y.
Quoted 6 lines
[*] Since commit
cba5d9b3e99d("x86/mm/64: MakeSPARSEMEM_VMEMMAPthe only memory model")page_to_pfn()has been a simple macro without any function call.Fixes e9f17038d814 ("x86/tdx: mark tdh_vp_enter() as __flatten")Signed-off-by Kai Huang <[email protected]>
Otherwise, LGTM.
Quoted 8 lines
Quoted 4 lines
To fix the second problem, change
tdx_tdvpr_pa()to use PFN_PHYS(page_to_pfn()) instead ofpage_to_phys()so that there will be no more function call inside tdx_tdvpr_pa()[*].To check my understanding,
page_to_pfn()onCONFIG_SPARSEMEM_VMEMMAPorCONFIG_FLATMEMhas no function calls, but onCONFIG_SPARSEMEM, it does. We are counting onx86_64to not useCONFIG_SPARSEMEM?
Yes. Please see include/asm-generic/memory_model.h.
Quoted 7 lines
Quoted 7 lines
The TDVPR page is always an actual page out of page allocator, so the additional warning around
pfn_valid()check inpage_to_phys()doesn't help a lot anyway. It's not worth complicating the code for such warning whenCONFIG_DEBUG_VIRTUAL=y.
Quoted 8 lines
Quoted 6 lines
[*] Since commit
cba5d9b3e99d("x86/mm/64: MakeSPARSEMEM_VMEMMAPthe only memory model")page_to_pfn()has been a simple macro without any function call.Fixes e9f17038d814 ("x86/tdx: mark tdh_vp_enter() as __flatten")Signed-off-by Kai Huang <[email protected]>Otherwise, LGTM.
Is this a RB? :-)
Quoted 11 lines · Kai
Quoted 9 lines
Quoted 4 lines
To fix the second problem, change
tdx_tdvpr_pa()to use PFN_PHYS(page_to_pfn()) instead ofpage_to_phys()so that there will be no more function call inside tdx_tdvpr_pa()[*].To check my understanding,
page_to_pfn()onCONFIG_SPARSEMEM_VMEMMAPorCONFIG_FLATMEMhas no function calls, but onCONFIG_SPARSEMEM, it does. We are
counting onx86_64to not useCONFIG_SPARSEMEM?Yes. Please see
include/asm-generic/memory_model.h.
It's a little brittle, but the compiler should warn at least.
Quoted 12 lines
Quoted 12 lines
Quoted 12 lines
The TDVPR page is always an actual page out of page allocator, so the additional warning around
pfn_valid()check inpage_to_phys()doesn't help a lot anyway. It's not worth complicating the code for such warning whenCONFIG_DEBUG_VIRTUAL=y.[*] Since commit
cba5d9b3e99d("x86/mm/64: MakeSPARSEMEM_VMEMMAPthe only memory model")page_to_pfn()has been a simple macro without any function call.
Quoted 6 lines
Quoted 4 lines
Quoted 2 lines
Fixes e9f17038d814 ("x86/tdx: mark tdh_vp_enter() as __flatten")Signed-off-by Kai Huang <[email protected]>Otherwise, LGTM.
Is this a RB? :-)
Yes, I think this problem has been annoying. This is a decent solution all things considered.