The 32-bit ARM kernel stops working when it exceeds a certain size
threshold.
This problem can occur, for instance, when enabling certain debug CONFIG
options or when linking numerous drivers as builtins, but reproducing
it primarily depends on the kernel size. To demonstrate this, I wrote a
debug Makefile code.
The following debug code generates and links drivers/dummy-*.S. The
NR_DUMMY_FILES variable specifies the number of dummy-*.S files to link.
The added code should be harmless except growing the kernel, as it
introduces unused symbols, dummy_0, dummy_1, and so on, each of which
occupies 64K code.
[Applying debug Makefile code]
Run the following code.
$ cat <<'EOF' >> drivers/Makefile
dummy-sources := $(shell i=0; while [ "$$i" -lt "$(NR_DUMMY_FILES)" ]; do echo dummy-$$i.S; i=$$(( i + 1 )); done)
targets += $(dummy-sources)
obj-y += $(patsubst %.S,%.o,$(dummy-sources))
quiet_cmd_gen_dummy = GEN $@
cmd_gen_dummy = { \
echo '\#include <linux/linkage.h>'; \
echo '.section ".text","ax"'; \
echo "ENTRY(dummy_$*)"; \
echo ".fill 65536, 1, 0"; \
echo "ENDPROC(dummy_$*)"; \
} > $@
$(obj)/dummy-%.S: FORCE; $(call if_changed,gen_dummy)
EOF
[Build result]
You can easily increase the .text section by giving NR_DUMMY_FILES from
the build command.
[1] ARM defconfig + NR_DUMMY_FILES=250
This inserts (64K * 250) dummy code on top of defconfig.
The kernel boots.
$ make -j$(nproc) ARCH=arm CROSS_COMPILE=arm-linux-gnueabihf- NR_DUMMY_FILES=250 defconfig all
$ arm-linux-gnueabihf-size vmlinux
text data bss dec hex filename
33361996 9278130 427088 43067214 291274e vmlinux
$ arm-linux-gnueabihf-nm vmlinux | grep ' [tT] ' | grep __get_user_4
c20e5dbc T __get_user_4
[2] ARM defconfig + NR_DUMMY_FILES=350
This inserts (64K * 350) dummy code on top of defconfig.
The kernel does not boot any more.
$ make -j$(nproc) ARCH=arm CROSS_COMPILE=arm-linux-gnueabihf- NR_DUMMY_FILES=350 defconfig all
$ arm-linux-gnueabihf-size vmlinux
text data bss dec hex filename
39946652 9282226 427088 49655966 2f5b09e vmlinux
$ arm-linux-gnueabihf-nm vmlinux | grep ' [tT] ' | grep __get_user_4
c272931c T __get_user_4
c06f6b10 t ____get_user_4_veneer
c0aef930 t ____get_user_4_veneer
I noticed ____get_user_4_veneer was causing the issue.
After commit 02e541db0540 ("ARM: 8323/1: force linker to use PIC veneers"),
veneers use the ip (= r12) register.
The disassembly of the veneer looks as follows:
c06f6b10 <____get_user_4_veneer>:
c06f6b10: e59fc000 ldr ip, [pc] @ c06f6b18 <____get_user_4_veneer+0x8>
c06f6b14: e08ff00c add pc, pc, ip
c06f6b18: 02032800 .word 0x02032800
When a routine is called via a veneer, the ip register is destroyed.
It is absolutely fine.
AAPCS32 [1] says:
"Register r12 (IP) may be used by a linker as a scratch register between
a routine and any subroutine it calls. It can also be used within a
routine to hold intermediate values between subroutine calls."
Of course, the compiler knows it, but the problem here is, the compiler
cannot notice the 'bl' instruction inside the inline asm.
In __get_user_check(), uaccess_save_and_enable() and uaccess_restore()
functions read and restore the value of the DACR register. There is no
obvious function call between them because __get_user_x() is a macro.
However, there is a branch to __get_user_4 buried within inline asm,
which the compiler is not aware of. Because only "lr" and "cc" are
listed in the clobber, the compiler considers it safe to use the ip
register to hold __ua_flags.
The disassembly of do_undefinstr() looks as follows:
c030b2bc <do_undefinstr>:
[snip]
c030b3d8: e1a01006 mov r1, r6
c030b3dc: e306077c movw r0, #26492 @ 0x677c
c030b3e0: e34c02c5 movt r0, #49861 @ 0xc2c5
c030b3e4: e3a02000 mov r2, #0
c030b3e8: e8bd41f0 pop {r4, r5, r6, r7, r8, lr}
c030b3ec: eafffe2e b c030acac <die>
c030b3f0: ee13cf10 mrc 15, 0, ip, cr3, cr0, {0} <== DACR is loaded to ip
c030b3f4: e3cc300c bic r3, ip, #12
c030b3f8: e1a00007 mov r0, r7
c030b3fc: e3833004 orr r3, r3, #4
c030b400: e3e01441 mvn r1, #1090519040 @ 0x41000000
c030b404: ee033f10 mcr 15, 0, r3, cr3, cr0, {0}
c030b408: f57ff06f isb sy
c030b40c: eb0fadbf bl c06f6b10 <____get_user_4_veneer> <== ip is destroyed here
c030b410: ee03cf10 mcr 15, 0, ip, cr3, cr0, {0} <== ip is written back to DACR
c030b414: f57ff06f isb sy
c030b418: e3500000 cmp r0, #0
Add "ip" to the clobber list to inform the compiler that it may be
destroyed if we end up with a veneer.
[1]: https://github.com/ARM-software/abi-aa/blob/2023Q1/aapcs32/aapcs32.rst
Signed-off-by: Masahiro Yamada <[email protected]>
---
I did not check the entire ARM code to seek for a similar pattern,
but I confirmed __put_user_check() had "ip" "lr" "cc" in the clobber
list.
Perhaps, another way is to list out all registers not guaranteed
in the spec.
AAPCS32 says, "A subroutine must preserve the contents of the registers
r4-r8, r10, r11 and SP".
The others might be clobbered. Of course, clobbered registers are
predicable as we write the subroutines in assembly code (only ip is
conditionally clobbered if veneers are generated), but it will be
safe in case we change the assembly code.
arch/arm/include/asm/uaccess.h | 14 ++------------
1 file changed, 2 insertions(+), 12 deletions(-)
diff --git a/arch/arm/include/asm/uaccess.h b/arch/arm/include/asm/uaccess.h
index bb5c81823117..c28f5ec21e41 100644
--- a/arch/arm/include/asm/uaccess.h
+++ b/arch/arm/include/asm/uaccess.h
@@ -109,16 +109,6 @@ extern int __get_user_64t_1(void *);
extern int __get_user_64t_2(void *);
extern int __get_user_64t_4(void *);
-#define __GUP_CLOBBER_1 "lr", "cc"
-#ifdef CONFIG_CPU_USE_DOMAINS
-#define __GUP_CLOBBER_2 "ip", "lr", "cc"
-#else
-#define __GUP_CLOBBER_2 "lr", "cc"
-#endif
-#define __GUP_CLOBBER_4 "lr", "cc"
-#define __GUP_CLOBBER_32t_8 "lr", "cc"
-#define __GUP_CLOBBER_8 "lr", "cc"
-
#define __get_user_x(__r2, __p, __e, __l, __s) \
__asm__ __volatile__ ( \
__asmeq("%0", "r0") __asmeq("%1", "r2") \
@@ -126,7 +116,7 @@ extern int __get_user_64t_4(void *);
"bl __get_user_" #__s \
: "=&r" (__e), "=r" (__r2) \
: "0" (__p), "r" (__l) \
- : __GUP_CLOBBER_##__s)
+ : "ip", "lr", "cc")
/* narrowing a double-word get into a single 32bit word register: */
#ifdef __ARMEB__
@@ -148,7 +138,7 @@ extern int __get_user_64t_4(void *);
"bl __get_user_64t_" #__s \
: "=&r" (__e), "=r" (__r2) \
: "0" (__p), "r" (__l) \
- : __GUP_CLOBBER_##__s)
+ : "ip", "lr", "cc")
#else
#define __get_user_x_64t __get_user_x
#endif
--
2.39.2
Hello Masahiro,
Thanks for getting to the bottom of this.
The patch looks correct to me, but could we please drop this huge
commit log text? The AAPCS simply states that a BL instruction might
clobber IP, so any inline asm that contains a BL should have IP in its
clobber list. In our case, the use of PIC veneers is what might cause
this, so we should mention that as well. But your entire journey to
get to that conclusion is not something that deserves to be included
in the git commit log forever, imho.
So with the commit log trimmed:
Reviewed-by: Ard Biesheuvel <[email protected]>
Thanks,
On Mon, 25 Sept 2023 at 11:00, Masahiro Yamada <[email protected]> wrote:
>
> The 32-bit ARM kernel stops working when it exceeds a certain size
> threshold.
>
> This problem can occur, for instance, when enabling certain debug CONFIG
> options or when linking numerous drivers as builtins, but reproducing
> it primarily depends on the kernel size. To demonstrate this, I wrote a
> debug Makefile code.
>
> The following debug code generates and links drivers/dummy-*.S. The
> NR_DUMMY_FILES variable specifies the number of dummy-*.S files to link.
> The added code should be harmless except growing the kernel, as it
> introduces unused symbols, dummy_0, dummy_1, and so on, each of which
> occupies 64K code.
>
> [Applying debug Makefile code]
>
> Run the following code.
>
> $ cat <<'EOF' >> drivers/Makefile
>
> dummy-sources := $(shell i=0; while [ "$$i" -lt "$(NR_DUMMY_FILES)" ]; do echo dummy-$$i.S; i=$$(( i + 1 )); done)
> targets += $(dummy-sources)
> obj-y += $(patsubst %.S,%.o,$(dummy-sources))
>
> quiet_cmd_gen_dummy = GEN $@
> cmd_gen_dummy = { \
> echo '\#include <linux/linkage.h>'; \
> echo '.section ".text","ax"'; \
> echo "ENTRY(dummy_$*)"; \
> echo ".fill 65536, 1, 0"; \
> echo "ENDPROC(dummy_$*)"; \
> } > $@
>
> $(obj)/dummy-%.S: FORCE; $(call if_changed,gen_dummy)
> EOF
>
> [Build result]
>
> You can easily increase the .text section by giving NR_DUMMY_FILES from
> the build command.
>
> [1] ARM defconfig + NR_DUMMY_FILES=250
>
> This inserts (64K * 250) dummy code on top of defconfig.
> The kernel boots.
>
> $ make -j$(nproc) ARCH=arm CROSS_COMPILE=arm-linux-gnueabihf- NR_DUMMY_FILES=250 defconfig all
> $ arm-linux-gnueabihf-size vmlinux
> text data bss dec hex filename
> 33361996 9278130 427088 43067214 291274e vmlinux
> $ arm-linux-gnueabihf-nm vmlinux | grep ' [tT] ' | grep __get_user_4
> c20e5dbc T __get_user_4
>
> [2] ARM defconfig + NR_DUMMY_FILES=350
>
> This inserts (64K * 350) dummy code on top of defconfig.
> The kernel does not boot any more.
>
> $ make -j$(nproc) ARCH=arm CROSS_COMPILE=arm-linux-gnueabihf- NR_DUMMY_FILES=350 defconfig all
> $ arm-linux-gnueabihf-size vmlinux
> text data bss dec hex filename
> 39946652 9282226 427088 49655966 2f5b09e vmlinux
> $ arm-linux-gnueabihf-nm vmlinux | grep ' [tT] ' | grep __get_user_4
> c272931c T __get_user_4
> c06f6b10 t ____get_user_4_veneer
> c0aef930 t ____get_user_4_veneer
>
> I noticed ____get_user_4_veneer was causing the issue.
>
> After commit 02e541db0540 ("ARM: 8323/1: force linker to use PIC veneers"),
> veneers use the ip (= r12) register.
>
> The disassembly of the veneer looks as follows:
>
> c06f6b10 <____get_user_4_veneer>:
> c06f6b10: e59fc000 ldr ip, [pc] @ c06f6b18 <____get_user_4_veneer+0x8>
> c06f6b14: e08ff00c add pc, pc, ip
> c06f6b18: 02032800 .word 0x02032800
>
> When a routine is called via a veneer, the ip register is destroyed.
>
> It is absolutely fine.
>
> AAPCS32 [1] says:
>
> "Register r12 (IP) may be used by a linker as a scratch register between
> a routine and any subroutine it calls. It can also be used within a
> routine to hold intermediate values between subroutine calls."
>
> Of course, the compiler knows it, but the problem here is, the compiler
> cannot notice the 'bl' instruction inside the inline asm.
>
> In __get_user_check(), uaccess_save_and_enable() and uaccess_restore()
> functions read and restore the value of the DACR register. There is no
> obvious function call between them because __get_user_x() is a macro.
> However, there is a branch to __get_user_4 buried within inline asm,
> which the compiler is not aware of. Because only "lr" and "cc" are
> listed in the clobber, the compiler considers it safe to use the ip
> register to hold __ua_flags.
>
> The disassembly of do_undefinstr() looks as follows:
>
> c030b2bc <do_undefinstr>:
> [snip]
> c030b3d8: e1a01006 mov r1, r6
> c030b3dc: e306077c movw r0, #26492 @ 0x677c
> c030b3e0: e34c02c5 movt r0, #49861 @ 0xc2c5
> c030b3e4: e3a02000 mov r2, #0
> c030b3e8: e8bd41f0 pop {r4, r5, r6, r7, r8, lr}
> c030b3ec: eafffe2e b c030acac <die>
> c030b3f0: ee13cf10 mrc 15, 0, ip, cr3, cr0, {0} <== DACR is loaded to ip
> c030b3f4: e3cc300c bic r3, ip, #12
> c030b3f8: e1a00007 mov r0, r7
> c030b3fc: e3833004 orr r3, r3, #4
> c030b400: e3e01441 mvn r1, #1090519040 @ 0x41000000
> c030b404: ee033f10 mcr 15, 0, r3, cr3, cr0, {0}
> c030b408: f57ff06f isb sy
> c030b40c: eb0fadbf bl c06f6b10 <____get_user_4_veneer> <== ip is destroyed here
> c030b410: ee03cf10 mcr 15, 0, ip, cr3, cr0, {0} <== ip is written back to DACR
> c030b414: f57ff06f isb sy
> c030b418: e3500000 cmp r0, #0
>
> Add "ip" to the clobber list to inform the compiler that it may be
> destroyed if we end up with a veneer.
>
> [1]: https://github.com/ARM-software/abi-aa/blob/2023Q1/aapcs32/aapcs32.rst
>
> Signed-off-by: Masahiro Yamada <[email protected]>
> ---
>
> I did not check the entire ARM code to seek for a similar pattern,
> but I confirmed __put_user_check() had "ip" "lr" "cc" in the clobber
> list.
>
> Perhaps, another way is to list out all registers not guaranteed
> in the spec.
>
> AAPCS32 says, "A subroutine must preserve the contents of the registers
> r4-r8, r10, r11 and SP".
>
> The others might be clobbered. Of course, clobbered registers are
> predicable as we write the subroutines in assembly code (only ip is
> conditionally clobbered if veneers are generated), but it will be
> safe in case we change the assembly code.
>
>
> arch/arm/include/asm/uaccess.h | 14 ++------------
> 1 file changed, 2 insertions(+), 12 deletions(-)
>
> diff --git a/arch/arm/include/asm/uaccess.h b/arch/arm/include/asm/uaccess.h
> index bb5c81823117..c28f5ec21e41 100644
> --- a/arch/arm/include/asm/uaccess.h
> +++ b/arch/arm/include/asm/uaccess.h
> @@ -109,16 +109,6 @@ extern int __get_user_64t_1(void *);
> extern int __get_user_64t_2(void *);
> extern int __get_user_64t_4(void *);
>
> -#define __GUP_CLOBBER_1 "lr", "cc"
> -#ifdef CONFIG_CPU_USE_DOMAINS
> -#define __GUP_CLOBBER_2 "ip", "lr", "cc"
> -#else
> -#define __GUP_CLOBBER_2 "lr", "cc"
> -#endif
> -#define __GUP_CLOBBER_4 "lr", "cc"
> -#define __GUP_CLOBBER_32t_8 "lr", "cc"
> -#define __GUP_CLOBBER_8 "lr", "cc"
> -
> #define __get_user_x(__r2, __p, __e, __l, __s) \
> __asm__ __volatile__ ( \
> __asmeq("%0", "r0") __asmeq("%1", "r2") \
> @@ -126,7 +116,7 @@ extern int __get_user_64t_4(void *);
> "bl __get_user_" #__s \
> : "=&r" (__e), "=r" (__r2) \
> : "0" (__p), "r" (__l) \
> - : __GUP_CLOBBER_##__s)
> + : "ip", "lr", "cc")
>
> /* narrowing a double-word get into a single 32bit word register: */
> #ifdef __ARMEB__
> @@ -148,7 +138,7 @@ extern int __get_user_64t_4(void *);
> "bl __get_user_64t_" #__s \
> : "=&r" (__e), "=r" (__r2) \
> : "0" (__p), "r" (__l) \
> - : __GUP_CLOBBER_##__s)
> + : "ip", "lr", "cc")
> #else
> #define __get_user_x_64t __get_user_x
> #endif
> --
> 2.39.2
>
On Tue, Sep 26, 2023 at 8:37 PM Ard Biesheuvel <[email protected]> wrote:
>
> Hello Masahiro,
>
> Thanks for getting to the bottom of this.
>
> The patch looks correct to me, but could we please drop this huge
> commit log text? The AAPCS simply states that a BL instruction might
> clobber IP, so any inline asm that contains a BL should have IP in its
> clobber list. In our case, the use of PIC veneers is what might cause
> this, so we should mention that as well. But your entire journey to
> get to that conclusion is not something that deserves to be included
> in the git commit log forever, imho.
>
> So with the commit log trimmed:
>
> Reviewed-by: Ard Biesheuvel <[email protected]>
Is this commit log OK?
===================================>8============================
The 32-bit ARM kernel stops working if the kernel grows to the point
where veneers for __get_user_* are created.
AAPCS32 [1] states, "Register r12 (IP) may be used by a linker as a
scratch register between a routine and any subroutine it calls. It
can also be used within a routine to hold intermediate values between
subroutine calls."
However, the bl instruction buried within the inline asm is unpredictable
for compilers; hence, "ip" must be added to the clobber list.
This becomes critical when veneers for __get_user_* are created because
veneers use the ip register since commit 02e541db0540 ("ARM: 8323/1:
force linker to use PIC veneers").
[1]: https://github.com/ARM-software/abi-aa/blob/2023Q1/aapcs32/aapcs32.rst
===================================>8============================
--
Best Regards
Masahiro Yamada
On Tue, 26 Sept 2023 at 14:27, Masahiro Yamada <[email protected]> wrote:
>
> On Tue, Sep 26, 2023 at 8:37 PM Ard Biesheuvel <[email protected]> wrote:
> >
> > Hello Masahiro,
> >
> > Thanks for getting to the bottom of this.
> >
> > The patch looks correct to me, but could we please drop this huge
> > commit log text? The AAPCS simply states that a BL instruction might
> > clobber IP, so any inline asm that contains a BL should have IP in its
> > clobber list. In our case, the use of PIC veneers is what might cause
> > this, so we should mention that as well. But your entire journey to
> > get to that conclusion is not something that deserves to be included
> > in the git commit log forever, imho.
> >
> > So with the commit log trimmed:
> >
> > Reviewed-by: Ard Biesheuvel <[email protected]>
>
>
>
> Is this commit log OK?
>
Yes, thanks for rewriting it.
I don't think a resend is necessary - please drop this in Russell's
patch tracker.
Thanks,
>
> ===================================>8============================
> The 32-bit ARM kernel stops working if the kernel grows to the point
> where veneers for __get_user_* are created.
>
> AAPCS32 [1] states, "Register r12 (IP) may be used by a linker as a
> scratch register between a routine and any subroutine it calls. It
> can also be used within a routine to hold intermediate values between
> subroutine calls."
>
> However, the bl instruction buried within the inline asm is unpredictable
> for compilers; hence, "ip" must be added to the clobber list.
>
> This becomes critical when veneers for __get_user_* are created because
> veneers use the ip register since commit 02e541db0540 ("ARM: 8323/1:
> force linker to use PIC veneers").
>
> [1]: https://github.com/ARM-software/abi-aa/blob/2023Q1/aapcs32/aapcs32.rst
> ===================================>8============================
>
>
>
>
>
>
> --
> Best Regards
> Masahiro Yamada