locking/refcounts, x86/asm: Use unique .text section for refcount exceptions
Subsystems (3)
PATCH 1/1 locking/refcounts, x86/asm: Use unique .text section for refcount exceptions
Using .text.unlikely for refcount exceptions isn't safe because gcc may move entire functions into .text.unlikely (e.g. in6_dev_get()), which would cause any uses of a protected refcount_t function to stay inline with the function, triggering the protection unconditionally:
.section .text.unlikely,"ax",@progbits
.type in6_dev_get, @functionin6_dev_getx:
.LFB4673:
.loc 2 4128 0
.cfi_startproc
...
lock; incl 480(%rbx)
js 111f
.pushsection .text.unlikely111: lea 480(%rbx), %rcx
112: .byte 0x0f, 0xff
.popsection
113:
This creates a unique .text section and adds an additional test to the exception handler to WARN in the case of having none of OF, SF, nor ZF set so we can see things like this more easily in the future.
--- arch/x86/Kconfig | 2 +- arch/x86/include/asm/refcount.h | 2 +- arch/x86/mm/extable.c | 7 ++++++- include/asm-generic/vmlinux.lds.h | 1 + 4 files changed, 9 insertions(+), 3 deletions(-)
arch/x86/Kconfig +1 -1
@@ -56,7 +56,7 @@ config X86 select ARCH_HAS_MMIO_FLUSH select ARCH_HAS_PMEM_API if X86_64 # Causing hangs/crashes, see the commit that added this change for details. - select ARCH_HAS_REFCOUNT if BROKEN + select ARCH_HAS_REFCOUNT select ARCH_HAS_UACCESS_FLUSHCACHE if X86_64 select ARCH_HAS_SET_MEMORY select ARCH_HAS_SG_CHAIN
arch/x86/include/asm/refcount.h +1 -1
@@ -15,7 +15,7 @@* back to the regular execution flow in .text. */ #define _REFCOUNT_EXCEPTION \ - ".pushsection .text.unlikely\n" \ + ".pushsection .text..refcount\n" \ "111:\tlea %[counter], %%" _ASM_CX "\n" \ "112:\t" ASM_UD0 "\n" \ ASM_UNREACHABLE \
arch/x86/mm/extable.c +6 -1
@@ -66,12 +66,17 @@ bool ex_handler_refcount(const struct exception_table_entry *fixup,* wrapped around) will be set. Additionally, seeing the refcount * reach 0 will set ZF (Zero Flag: result was zero). In each of * these cases we want a report, since it's a boundary condition. -* +* The SF case is not reported since it indicates post-boundary +* manipulations below zero or above INT_MAX. And if none of the +* flags are set, something has gone very wrong, so report it. */ if (regs->flags & (X86_EFLAGS_OF |X86_EFLAGS_ZF )) {bool zero = regs->flags &X86_EFLAGS_ZF ;refcount_error_report (regs, zero ?"hit zero" :"overflow" ); + }else if ((regs->flags &X86_EFLAGS_SF ) ==0 ) { +/* Report if none of OF, ZF, nor SF are set. */ +refcount_error_report (regs,"unexpected saturation" ); }return true;
include/asm-generic/vmlinux.lds.h +1 -0
@@ -459,6 +459,7 @@#define TEXT_TEXT \ ALIGN_FUNCTION(); \ *(.text.hot TEXT_MAIN .text.fixup .text.unlikely) \ + *(.text..refcount) \ *(.ref.text) \ MEM_KEEP(init.text) \ MEM_KEEP(exit.text) \
Signature · 6 lines
2.7.4 -- Kees Cook Pixel Security
Quoted 52 lines · Kees Cook
Using .text.unlikely for refcount exceptions isn't safe because gcc may move entire functions into .text.unlikely (e.g.
in6_dev_get()), which would cause any uses of a protectedrefcount_tfunction to stay inline with the function, triggering the protection unconditionally:.section .text.unlikely,"ax",@progbits
.typein6_dev_get, @functionin6_dev_getx:
.LFB4673:.loc 2 4128 0 .cfi_startproc ... lock; incl 480(%rbx) js 111f .pushsection .text.unlikely111: lea 480(%rbx), %rcx
112: .byte0x0f,0xff
.popsection
113:This creates a unique .text section and adds an additional test to the exception handler to WARN in the case of having none of OF, SF, nor ZF set so we can see things like this more easily in the future.
Reported-by Mike Galbraith <[email protected]>Fixes 7a46ec0e2f48 ("locking/refcounts, x86/asm: Implement fast refcount overflow protection")Signed-off-by Kees Cook <[email protected]>--- arch/x86/Kconfig | 2 +- arch/x86/include/asm/refcount.h | 2 +- arch/x86/mm/extable.c | 7 ++++++- include/asm-generic/vmlinux.lds.h | 1 + 4 files changed, 9 insertions(+), 3 deletions(-)arch/x86/Kconfig +1
-1@@ -56,7 +56,7 @@ config X86 select ARCH_HAS_MMIO_FLUSH select ARCH_HAS_PMEM_API if X86_64 # Causing hangs/crashes, see the commit that added this change for details. - select ARCH_HAS_REFCOUNT if BROKEN + select ARCH_HAS_REFCOUNT select ARCH_HAS_UACCESS_FLUSHCACHE if X86_64 select ARCH_HAS_SET_MEMORY select ARCH_HAS_SG_CHAINarch/x86/include/asm/refcount.h +0
-0
arch/x86/include/asm/refcount.h line 18
Quoted 5 lines
arch/x86/include/asm/refcount.h +1
-1@@ -15,7 +15,7 @@* back to the regular execution flow in .text. */ #define _REFCOUNT_EXCEPTION \ - ".pushsection .text.unlikely\n" \ + ".pushsection .text..refcount\n" \
You could also use a .subsection here: those are always emitted out of line, but into the same section.
Quoted 45 lines
arch/x86/include/asm/refcount.h +0
-0"111:\tlea %[counter], %%" _ASM_CX"\n" \"112:\t" ASM_UD0 "\n" \ASM_UNREACHABLE \arch/x86/mm/extable.c +6
-1@@ -66,12 +66,17 @@ bool ex_handler_refcount(const struct exception_table_entry *fixup,* wrapped around) will be set. Additionally, seeing the refcount * reach 0 will set ZF (Zero Flag: result was zero). In each of * these cases we want a report, since it's a boundary condition. -* +* The SF case is not reported since it indicates post-boundary +* manipulations below zero or above INT_MAX. And if none of the +* flags are set, something has gone very wrong, so report it. */ if (regs->flags & (X86_EFLAGS_OF |X86_EFLAGS_ZF )) {bool zero = regs->flags &X86_EFLAGS_ZF ;refcount_error_report (regs, zero ?"hit zero" :"overflow" ); + }else if ((regs->flags &X86_EFLAGS_SF ) ==0 ) { +/* Report if none of OF, ZF, nor SF are set. */ +refcount_error_report (regs,"unexpected saturation" ); }return true;include/asm-generic/vmlinux.lds.h +1
-1@@ -459,6 +459,7 @@#define TEXT_TEXT \ ALIGN_FUNCTION(); \ *(.text.hot TEXT_MAIN .text.fixup .text.unlikely) \ + *(.text..refcount) \ *(.ref.text) \ MEM_KEEP(init.text) \ MEM_KEEP(exit.text) \ --2.7.4
--
Kees Cook
Pixel Security
Quoted 53 lines · Ard Biesheuvel
Quoted 52 lines · Kees Cook
Using .text.unlikely for refcount exceptions isn't safe because gcc may move entire functions into .text.unlikely (e.g.
in6_dev_get()), which would cause any uses of a protectedrefcount_tfunction to stay inline with the function, triggering the protection unconditionally:.section .text.unlikely,"ax",@progbits
.typein6_dev_get, @functionin6_dev_getx:
.LFB4673:.loc 2 4128 0 .cfi_startproc ... lock; incl 480(%rbx) js 111f .pushsection .text.unlikely111: lea 480(%rbx), %rcx
112: .byte0x0f,0xff
.popsection
113:This creates a unique .text section and adds an additional test to the exception handler to WARN in the case of having none of OF, SF, nor ZF set so we can see things like this more easily in the future.
Reported-by Mike Galbraith <[email protected]>Fixes 7a46ec0e2f48 ("locking/refcounts, x86/asm: Implement fast refcount overflow protection")Signed-off-by Kees Cook <[email protected]>--- arch/x86/Kconfig | 2 +- arch/x86/include/asm/refcount.h | 2 +- arch/x86/mm/extable.c | 7 ++++++- include/asm-generic/vmlinux.lds.h | 1 + 4 files changed, 9 insertions(+), 3 deletions(-)arch/x86/Kconfig +1
-1@@ -56,7 +56,7 @@ config X86 select ARCH_HAS_MMIO_FLUSH select ARCH_HAS_PMEM_API if X86_64 # Causing hangs/crashes, see the commit that added this change for details. - select ARCH_HAS_REFCOUNT if BROKEN + select ARCH_HAS_REFCOUNT select ARCH_HAS_UACCESS_FLUSHCACHE if X86_64 select ARCH_HAS_SET_MEMORY select ARCH_HAS_SG_CHAINarch/x86/include/asm/refcount.h +0
-0@@ -15,7 +15,7 @@
Quoted 8 lines
arch/x86/include/asm/refcount.h line 18
Quoted 5 lines
arch/x86/include/asm/refcount.h +1
-1* back to the regular execution flow in .text. */ #define _REFCOUNT_EXCEPTION \ - ".pushsection .text.unlikely\n" \ + ".pushsection .text..refcount\n" \You could also use a .subsection here: those are always emitted out of line, but into the same section.
I considered it (especially after looking to see how you handled it in the arm64 port) but decided against it in favor of collecting them all at the end with .text.unlikely.
-Kees
Signature · 2 lines
Kees Cook Pixel Security
Quoted 64 lines · Kees Cook
Quoted 53 lines · Ard Biesheuvel
Quoted 52 lines · Kees Cook
Using .text.unlikely for refcount exceptions isn't safe because gcc may move entire functions into .text.unlikely (e.g.
in6_dev_get()), which would cause any uses of a protectedrefcount_tfunction to stay inline with the function, triggering the protection unconditionally:.section .text.unlikely,"ax",@progbits
.typein6_dev_get, @functionin6_dev_getx:
.LFB4673:.loc 2 4128 0 .cfi_startproc ... lock; incl 480(%rbx) js 111f .pushsection .text.unlikely111: lea 480(%rbx), %rcx
112: .byte0x0f,0xff
.popsection
113:This creates a unique .text section and adds an additional test to the exception handler to WARN in the case of having none of OF, SF, nor ZF set so we can see things like this more easily in the future.
Reported-by Mike Galbraith <[email protected]>Fixes 7a46ec0e2f48 ("locking/refcounts, x86/asm: Implement fast refcount overflow protection")Signed-off-by Kees Cook <[email protected]>--- arch/x86/Kconfig | 2 +- arch/x86/include/asm/refcount.h | 2 +- arch/x86/mm/extable.c | 7 ++++++- include/asm-generic/vmlinux.lds.h | 1 + 4 files changed, 9 insertions(+), 3 deletions(-)arch/x86/Kconfig +1
-1@@ -56,7 +56,7 @@ config X86 select ARCH_HAS_MMIO_FLUSH select ARCH_HAS_PMEM_API if X86_64 # Causing hangs/crashes, see the commit that added this change for details. - select ARCH_HAS_REFCOUNT if BROKEN + select ARCH_HAS_REFCOUNT select ARCH_HAS_UACCESS_FLUSHCACHE if X86_64 select ARCH_HAS_SET_MEMORY select ARCH_HAS_SG_CHAINarch/x86/include/asm/refcount.h +0
-0@@ -15,7 +15,7 @@Quoted 8 lines
arch/x86/include/asm/refcount.h line 18
Quoted 5 lines
arch/x86/include/asm/refcount.h +1
-1* back to the regular execution flow in .text. */ #define _REFCOUNT_EXCEPTION \ - ".pushsection .text.unlikely\n" \ + ".pushsection .text..refcount\n" \You could also use a .subsection here: those are always emitted out of line, but into the same section.
Quoted 4 lines
I considered it (especially after looking to see how you handled it in the arm64 port) but decided against it in favor of collecting them all at the end with .text.unlikely.
OK. On arm64, I needed this because the out of line sequence refers back to the inline code, which resulted in section mismatches if the refcount API is used from __init code. If that problem does not exist on x86, I guess it doesn't matter.
Quoted 15 lines · Kees Cook
Using .text.unlikely for refcount exceptions isn't safe because gcc may move entire functions into .text.unlikely (e.g.
in6_dev_get()), which would cause any uses of a protectedrefcount_tfunction to stay inline with the function, triggering the protection unconditionally:.section .text.unlikely,"ax",@progbits
.typein6_dev_get, @functionin6_dev_getx:
.LFB4673:.loc 2 4128 0 .cfi_startproc ... lock; incl 480(%rbx) js 111f .pushsection .text.unlikely
Quoted 8 lines
111: lea 480(%rbx), %rcx
112: .byte0x0f,0xff
.popsection
113:This creates a unique .text section and adds an additional test to the exception handler to WARN in the case of having none of OF, SF, nor ZF set so we can see things like this more easily in the future.
Closure: gcc-4.8.5 now builds a functional kernel as well, so that aspect of this bug was just a larger a dose of the same toxin.
Question below.
(quoted diff) +0 -0
b/arch/x86/include/asm/refcount.h
arch/x86/include/asm/refcount.h line 18
Quoted 9 lines
index
ff871210b9f2..4e44250e7d0d100644arch/x86/include/asm/refcount.h +1
-1@@ -15,7 +15,7 @@* back to the regular execution flow in .text. */ #define _REFCOUNT_EXCEPTION \ - ".pushsection .text.unlikely\n" \ + ".pushsection .text..refcount\n" \
Why two dots? (.text.refcount_ex?)
-Mike
Quoted 21 lines · Mike Galbraith
Quoted 20 lines · Kees Cook
Using .text.unlikely for refcount exceptions isn't safe because gcc may move entire functions into .text.unlikely (e.g.
in6_dev_get()), which would cause any uses of a protectedrefcount_tfunction to stay inline with the function, triggering the protection unconditionally:.section .text.unlikely,"ax",@progbits
.typein6_dev_get, @functionin6_dev_getx:
.LFB4673:.loc 2 4128 0 .cfi_startproc ... lock; incl 480(%rbx) js 111f .pushsection .text.unlikely111: lea 480(%rbx), %rcx
112: .byte0x0f,0xff
.popsection
113:
Quoted 6 lines
Quoted 3 lines
This creates a unique .text section and adds an additional test to the exception handler to WARN in the case of having none of OF, SF, nor ZF set so we can see things like this more easily in the future.
Closure: gcc-4.8.5 now builds a functional kernel as well, so that aspect of this bug was just a larger a dose of the same toxin.
Okay, excellent. Thanks for checking!
Quoted 8 lines
Question below.
arch/x86/include/asm/refcount.h +0
-0Quoted 3 lines
index
ff871210b9f2..4e44250e7d0d100644arch/x86/include/asm/refcount.h +0
-0
Quoted 8 lines
arch/x86/include/asm/refcount.h line 18
Quoted 6 lines
arch/x86/include/asm/refcount.h +1
-1@@ -15,7 +15,7 @@* back to the regular execution flow in .text. */ #define _REFCOUNT_EXCEPTION \ - ".pushsection .text.unlikely\n" \ + ".pushsection .text..refcount\n" \Why two dots? (.text.refcount_ex?)
A dot keeps it out of the TEXT_MAIN macro namespace (see cb87481ee89db in -next, which is function names: [a-zA-Z0-9_]) to avoid collisions and so it can be put at the end with text.unlikely to keep the cold code together.
-Kees
Signature · 2 lines
Kees Cook Pixel Security
Quoted 27 lines · Kees Cook
Using .text.unlikely for refcount exceptions isn't safe because gcc may move entire functions into .text.unlikely (e.g.
in6_dev_get()), which would cause any uses of a protectedrefcount_tfunction to stay inline with the function, triggering the protection unconditionally:.section .text.unlikely,"ax",@progbits
.typein6_dev_get, @functionin6_dev_getx:
.LFB4673:.loc 2 4128 0 .cfi_startproc ... lock; incl 480(%rbx) js 111f .pushsection .text.unlikely111: lea 480(%rbx), %rcx
112: .byte0x0f,0xff
.popsection
113:This creates a unique .text section and adds an additional test to the exception handler to WARN in the case of having none of OF, SF, nor ZF set so we can see things like this more easily in the future.
Reported-by Mike Galbraith <[email protected]>Fixes 7a46ec0e2f48 ("locking/refcounts, x86/asm: Implement fast refcount overflow protection")Signed-off-by Kees Cook <[email protected]>
Quoted 6 lines
--- arch/x86/Kconfig | 2 +- arch/x86/include/asm/refcount.h | 2 +- arch/x86/mm/extable.c | 7 ++++++- include/asm-generic/vmlinux.lds.h | 1 + 4 files changed, 9 insertions(+), 3 deletions(-)
Could you please split this into two patches: one that fixes the .unlikely bug, the other that re-enables the optimized version?
Should there be any other problem with refcounts this would make any bisection result more clear-cut.
Thanks,
Ingo
Quoted 37 lines · Ingo Molnar
Quoted 27 lines · Kees Cook
Using .text.unlikely for refcount exceptions isn't safe because gcc may move entire functions into .text.unlikely (e.g.
in6_dev_get()), which would cause any uses of a protectedrefcount_tfunction to stay inline with the function, triggering the protection unconditionally:.section .text.unlikely,"ax",@progbits
.typein6_dev_get, @functionin6_dev_getx:
.LFB4673:.loc 2 4128 0 .cfi_startproc ... lock; incl 480(%rbx) js 111f .pushsection .text.unlikely111: lea 480(%rbx), %rcx
112: .byte0x0f,0xff
.popsection
113:This creates a unique .text section and adds an additional test to the exception handler to WARN in the case of having none of OF, SF, nor ZF set so we can see things like this more easily in the future.
Reported-by Mike Galbraith <[email protected]>Fixes 7a46ec0e2f48 ("locking/refcounts, x86/asm: Implement fast refcount overflow protection")Signed-off-by Kees Cook <[email protected]>Quoted 6 lines
--- arch/x86/Kconfig | 2 +- arch/x86/include/asm/refcount.h | 2 +- arch/x86/mm/extable.c | 7 ++++++- include/asm-generic/vmlinux.lds.h | 1 + 4 files changed, 9 insertions(+), 3 deletions(-)
Quoted 5 lines
Could you please split this into two patches: one that fixes the .unlikely bug, the other that re-enables the optimized version?
Should there be any other problem with refcounts this would make any bisection result more clear-cut.
Certainly!
-Kees
Signature · 2 lines
Kees Cook Pixel Security