Increase COMMAND_LINE_SIZE as the current default value is too low
for syzbot kernel command line.
Reported-by: Dmitry Vyukov <[email protected]>
Signed-off-by: Alexandre Ghiti <[email protected]>
---
arch/riscv/include/uapi/asm/setup.h | 8 ++++++++
1 file changed, 8 insertions(+)
create mode 100644 arch/riscv/include/uapi/asm/setup.h
diff --git a/arch/riscv/include/uapi/asm/setup.h b/arch/riscv/include/uapi/asm/setup.h
new file mode 100644
index 000000000000..66b13a522880
--- /dev/null
+++ b/arch/riscv/include/uapi/asm/setup.h
@@ -0,0 +1,8 @@
+/* SPDX-License-Identifier: GPL-2.0-only WITH Linux-syscall-note */
+
+#ifndef _UAPI_ASM_RISCV_SETUP_H
+#define _UAPI_ASM_RISCV_SETUP_H
+
+#define COMMAND_LINE_SIZE 1024
+
+#endif /* _UAPI_ASM_RISCV_SETUP_H */
--
2.20.1
On Tue, 16 Mar 2021 12:34:20 PDT (-0700), [email protected] wrote:
> Increase COMMAND_LINE_SIZE as the current default value is too low
> for syzbot kernel command line.
>
> Reported-by: Dmitry Vyukov <[email protected]>
> Signed-off-by: Alexandre Ghiti <[email protected]>
> ---
> arch/riscv/include/uapi/asm/setup.h | 8 ++++++++
> 1 file changed, 8 insertions(+)
> create mode 100644 arch/riscv/include/uapi/asm/setup.h
>
> diff --git a/arch/riscv/include/uapi/asm/setup.h b/arch/riscv/include/uapi/asm/setup.h
> new file mode 100644
> index 000000000000..66b13a522880
> --- /dev/null
> +++ b/arch/riscv/include/uapi/asm/setup.h
> @@ -0,0 +1,8 @@
> +/* SPDX-License-Identifier: GPL-2.0-only WITH Linux-syscall-note */
> +
> +#ifndef _UAPI_ASM_RISCV_SETUP_H
> +#define _UAPI_ASM_RISCV_SETUP_H
> +
> +#define COMMAND_LINE_SIZE 1024
> +
> +#endif /* _UAPI_ASM_RISCV_SETUP_H */
I put this on fixes, but it seemes like this should really be a Kconfig
enttry. Either way, ours was quite a bit smaller than most
architectures and it's great that syzbot has started to find bugs, so
I'd rather get this in sooner.
On Mon, 29 Mar 2021, Palmer Dabbelt wrote:
> > --- /dev/null
> > +++ b/arch/riscv/include/uapi/asm/setup.h
> > @@ -0,0 +1,8 @@
> > +/* SPDX-License-Identifier: GPL-2.0-only WITH Linux-syscall-note */
> > +
> > +#ifndef _UAPI_ASM_RISCV_SETUP_H
> > +#define _UAPI_ASM_RISCV_SETUP_H
> > +
> > +#define COMMAND_LINE_SIZE 1024
> > +
> > +#endif /* _UAPI_ASM_RISCV_SETUP_H */
>
> I put this on fixes, but it seemes like this should really be a Kconfig
> enttry. Either way, ours was quite a bit smaller than most architectures and
> it's great that syzbot has started to find bugs, so I'd rather get this in
> sooner.
This macro is exported as a part of the user API so it must not depend on
Kconfig. Also changing it (rather than say adding COMMAND_LINE_SIZE_V2 or
switching to an entirely new data object that has its dimension set in a
different way) requires careful evaluation as external binaries have and
will have the value it expands to compiled in, so it's a part of the ABI
too.
Maciej
On Tue, 30 Mar 2021 13:31:45 PDT (-0700), [email protected] wrote:
> On Mon, 29 Mar 2021, Palmer Dabbelt wrote:
>
>> > --- /dev/null
>> > +++ b/arch/riscv/include/uapi/asm/setup.h
>> > @@ -0,0 +1,8 @@
>> > +/* SPDX-License-Identifier: GPL-2.0-only WITH Linux-syscall-note */
>> > +
>> > +#ifndef _UAPI_ASM_RISCV_SETUP_H
>> > +#define _UAPI_ASM_RISCV_SETUP_H
>> > +
>> > +#define COMMAND_LINE_SIZE 1024
>> > +
>> > +#endif /* _UAPI_ASM_RISCV_SETUP_H */
>>
>> I put this on fixes, but it seemes like this should really be a Kconfig
>> enttry. Either way, ours was quite a bit smaller than most architectures and
>> it's great that syzbot has started to find bugs, so I'd rather get this in
>> sooner.
>
> This macro is exported as a part of the user API so it must not depend on
> Kconfig. Also changing it (rather than say adding COMMAND_LINE_SIZE_V2 or
> switching to an entirely new data object that has its dimension set in a
> different way) requires careful evaluation as external binaries have and
> will have the value it expands to compiled in, so it's a part of the ABI
> too.
Thanks, I didn't realize this was part of the user BI. In that case we
really can't chage it, so we'll have to sort out some other way do fix
whatever is going on.
I've dropped this from fixes.
On Fri, Apr 2, 2021 at 6:37 AM Palmer Dabbelt <[email protected]> wrote:
>
> On Tue, 30 Mar 2021 13:31:45 PDT (-0700), [email protected] wrote:
> > On Mon, 29 Mar 2021, Palmer Dabbelt wrote:
> >
> >> > --- /dev/null
> >> > +++ b/arch/riscv/include/uapi/asm/setup.h
> >> > @@ -0,0 +1,8 @@
> >> > +/* SPDX-License-Identifier: GPL-2.0-only WITH Linux-syscall-note */
> >> > +
> >> > +#ifndef _UAPI_ASM_RISCV_SETUP_H
> >> > +#define _UAPI_ASM_RISCV_SETUP_H
> >> > +
> >> > +#define COMMAND_LINE_SIZE 1024
> >> > +
> >> > +#endif /* _UAPI_ASM_RISCV_SETUP_H */
> >>
> >> I put this on fixes, but it seemes like this should really be a Kconfig
> >> enttry. Either way, ours was quite a bit smaller than most architectures and
> >> it's great that syzbot has started to find bugs, so I'd rather get this in
> >> sooner.
> >
> > This macro is exported as a part of the user API so it must not depend on
> > Kconfig. Also changing it (rather than say adding COMMAND_LINE_SIZE_V2 or
> > switching to an entirely new data object that has its dimension set in a
> > different way) requires careful evaluation as external binaries have and
> > will have the value it expands to compiled in, so it's a part of the ABI
> > too.
>
> Thanks, I didn't realize this was part of the user BI. In that case we
> really can't chage it, so we'll have to sort out some other way do fix
> whatever is going on.
>
> I've dropped this from fixes.
Does increasing COMMAND_LINE_SIZE break user-space binaries? I would
expect it to work the same way as adding new enum values, or adding
fields at the end of versioned structs, etc.
I would assume the old bootloaders/etc will only support up to the
old, smaller max command line size, while the kernel will support
larger command line size, which is fine.
However, if something copies /proc/cmdline into a fixed-size buffer
and expects that to work, that will break... that's quite unfortunate
user-space code... is it what we afraid of?
Alternatively, could expose the same COMMAND_LINE_SIZE, but internally
support a larger command line?
On Fri, Apr 2, 2021 at 11:43 AM Dmitry Vyukov <[email protected]> wrote:
>
> On Fri, Apr 2, 2021 at 6:37 AM Palmer Dabbelt <[email protected]> wrote:
> >
> > On Tue, 30 Mar 2021 13:31:45 PDT (-0700), [email protected] wrote:
> > > On Mon, 29 Mar 2021, Palmer Dabbelt wrote:
> > >
> > >> > --- /dev/null
> > >> > +++ b/arch/riscv/include/uapi/asm/setup.h
> > >> > @@ -0,0 +1,8 @@
> > >> > +/* SPDX-License-Identifier: GPL-2.0-only WITH Linux-syscall-note */
> > >> > +
> > >> > +#ifndef _UAPI_ASM_RISCV_SETUP_H
> > >> > +#define _UAPI_ASM_RISCV_SETUP_H
> > >> > +
> > >> > +#define COMMAND_LINE_SIZE 1024
> > >> > +
> > >> > +#endif /* _UAPI_ASM_RISCV_SETUP_H */
> > >>
> > >> I put this on fixes, but it seemes like this should really be a Kconfig
> > >> enttry. Either way, ours was quite a bit smaller than most architectures and
> > >> it's great that syzbot has started to find bugs, so I'd rather get this in
> > >> sooner.
> > >
> > > This macro is exported as a part of the user API so it must not depend on
> > > Kconfig. Also changing it (rather than say adding COMMAND_LINE_SIZE_V2 or
> > > switching to an entirely new data object that has its dimension set in a
> > > different way) requires careful evaluation as external binaries have and
> > > will have the value it expands to compiled in, so it's a part of the ABI
> > > too.
> >
> > Thanks, I didn't realize this was part of the user BI. In that case we
> > really can't chage it, so we'll have to sort out some other way do fix
> > whatever is going on.
> >
> > I've dropped this from fixes.
>
> Does increasing COMMAND_LINE_SIZE break user-space binaries? I would
> expect it to work the same way as adding new enum values, or adding
> fields at the end of versioned structs, etc.
> I would assume the old bootloaders/etc will only support up to the
> old, smaller max command line size, while the kernel will support
> larger command line size, which is fine.
> However, if something copies /proc/cmdline into a fixed-size buffer
> and expects that to work, that will break... that's quite unfortunate
> user-space code... is it what we afraid of?
>
> Alternatively, could expose the same COMMAND_LINE_SIZE, but internally
> support a larger command line?
Looking at kernel commit history I see PowerPC switched from 512 to
2048, and I don't see complaints about the ABI on the mailing list.
If COMMAND_LINE_SIZE is used by user space applications and we
increase it there shouldn't be problems. I would expect things to
work, but just get truncated boot args? That is the application will
continue only to look at the initial 512 chars.
https://linuxppc-dev.ozlabs.narkive.com/m4cj8nBa/patch-1-1-powerpc-increase-command-line-size-to-2048-from-512
On Fri, 2 Apr 2021, David Abdurachmanov wrote:
> > > > This macro is exported as a part of the user API so it must not depend on
> > > > Kconfig. Also changing it (rather than say adding COMMAND_LINE_SIZE_V2 or
> > > > switching to an entirely new data object that has its dimension set in a
> > > > different way) requires careful evaluation as external binaries have and
> > > > will have the value it expands to compiled in, so it's a part of the ABI
> > > > too.
> > >
> > > Thanks, I didn't realize this was part of the user BI. In that case we
> > > really can't chage it, so we'll have to sort out some other way do fix
> > > whatever is going on.
> > >
> > > I've dropped this from fixes.
> >
> > Does increasing COMMAND_LINE_SIZE break user-space binaries? I would
> > expect it to work the same way as adding new enum values, or adding
> > fields at the end of versioned structs, etc.
> > I would assume the old bootloaders/etc will only support up to the
> > old, smaller max command line size, while the kernel will support
> > larger command line size, which is fine.
> > However, if something copies /proc/cmdline into a fixed-size buffer
> > and expects that to work, that will break... that's quite unfortunate
> > user-space code... is it what we afraid of?
> >
> > Alternatively, could expose the same COMMAND_LINE_SIZE, but internally
> > support a larger command line?
>
> Looking at kernel commit history I see PowerPC switched from 512 to
> 2048, and I don't see complaints about the ABI on the mailing list.
>
> If COMMAND_LINE_SIZE is used by user space applications and we
> increase it there shouldn't be problems. I would expect things to
> work, but just get truncated boot args? That is the application will
> continue only to look at the initial 512 chars.
The macro is in an include/uapi header, so it's exported to the userland
and a part of the user API. I don't know what the consequences are for
the RISC-V port specifically, but it has raised my attention, and I think
it has to be investigated.
Perhaps it's OK to change it after all, but you'd have to go through
known/potential users of this macro. I guess there shouldn't be that many
of them.
In any case it cannot depend on Kconfig, because the userland won't have
access to the configuration, and then presumably wants to handle any and
all.
Maciej
On Fri, 02 Apr 2021 11:33:30 PDT (-0700), [email protected] wrote:
> On Fri, 2 Apr 2021, David Abdurachmanov wrote:
>
>> > > > This macro is exported as a part of the user API so it must not depend on
>> > > > Kconfig. Also changing it (rather than say adding COMMAND_LINE_SIZE_V2 or
>> > > > switching to an entirely new data object that has its dimension set in a
>> > > > different way) requires careful evaluation as external binaries have and
>> > > > will have the value it expands to compiled in, so it's a part of the ABI
>> > > > too.
>> > >
>> > > Thanks, I didn't realize this was part of the user BI. In that case we
>> > > really can't chage it, so we'll have to sort out some other way do fix
>> > > whatever is going on.
>> > >
>> > > I've dropped this from fixes.
>> >
>> > Does increasing COMMAND_LINE_SIZE break user-space binaries? I would
>> > expect it to work the same way as adding new enum values, or adding
>> > fields at the end of versioned structs, etc.
>> > I would assume the old bootloaders/etc will only support up to the
>> > old, smaller max command line size, while the kernel will support
>> > larger command line size, which is fine.
>> > However, if something copies /proc/cmdline into a fixed-size buffer
>> > and expects that to work, that will break... that's quite unfortunate
>> > user-space code... is it what we afraid of?
>> >
>> > Alternatively, could expose the same COMMAND_LINE_SIZE, but internally
>> > support a larger command line?
>>
>> Looking at kernel commit history I see PowerPC switched from 512 to
>> 2048, and I don't see complaints about the ABI on the mailing list.
>>
>> If COMMAND_LINE_SIZE is used by user space applications and we
>> increase it there shouldn't be problems. I would expect things to
>> work, but just get truncated boot args? That is the application will
>> continue only to look at the initial 512 chars.
>
> The macro is in an include/uapi header, so it's exported to the userland
> and a part of the user API. I don't know what the consequences are for
> the RISC-V port specifically, but it has raised my attention, and I think
> it has to be investigated.
>
> Perhaps it's OK to change it after all, but you'd have to go through
> known/potential users of this macro. I guess there shouldn't be that many
> of them.
>
> In any case it cannot depend on Kconfig, because the userland won't have
> access to the configuration, and then presumably wants to handle any and
> all.
It kind of feels to me like COMMAND_LINE_SIZE shouldn't have been part
of the UABI to begin with. I sent a patch to remove it from the
asm-generic UABI, let's see if anyone knows of a reason it should be
UABI:
https://lore.kernel.org/linux-arch/[email protected]/T/#u
Hi Palmer,
Le 23/04/2021 ? 04:57, Palmer Dabbelt a ?crit?:
> On Fri, 02 Apr 2021 11:33:30 PDT (-0700), [email protected] wrote:
>> On Fri, 2 Apr 2021, David Abdurachmanov wrote:
>>
>>> > > >? This macro is exported as a part of the user API so it must
>>> not depend on
>>> > > > Kconfig.? Also changing it (rather than say adding
>>> COMMAND_LINE_SIZE_V2 or
>>> > > > switching to an entirely new data object that has its dimension
>>> set in a
>>> > > > different way) requires careful evaluation as external binaries
>>> have and
>>> > > > will have the value it expands to compiled in, so it's a part
>>> of the ABI
>>> > > > too.
>>> > >
>>> > > Thanks, I didn't realize this was part of the user BI.? In that
>>> case we
>>> > > really can't chage it, so we'll have to sort out some other way
>>> do fix
>>> > > whatever is going on.
>>> > >
>>> > > I've dropped this from fixes.
>>> >
>>> > Does increasing COMMAND_LINE_SIZE break user-space binaries? I would
>>> > expect it to work the same way as adding new enum values, or adding
>>> > fields at the end of versioned structs, etc.
>>> > I would assume the old bootloaders/etc will only support up to the
>>> > old, smaller max command line size, while the kernel will support
>>> > larger command line size, which is fine.
>>> > However, if something copies /proc/cmdline into a fixed-size buffer
>>> > and expects that to work, that will break... that's quite unfortunate
>>> > user-space code... is it what we afraid of?
>>> >
>>> > Alternatively, could expose the same COMMAND_LINE_SIZE, but internally
>>> > support a larger command line?
>>>
>>> Looking at kernel commit history I see PowerPC switched from 512 to
>>> 2048, and I don't see complaints about the ABI on the mailing list.
>>>
>>> If COMMAND_LINE_SIZE is used by user space applications and we
>>> increase it there shouldn't be problems. I would expect things to
>>> work, but just get truncated boot args? That is the application will
>>> continue only to look at the initial 512 chars.
>>
>> ?The macro is in an include/uapi header, so it's exported to the userland
>> and a part of the user API.? I don't know what the consequences are for
>> the RISC-V port specifically, but it has raised my attention, and I think
>> it has to be investigated.
>>
>> ?Perhaps it's OK to change it after all, but you'd have to go through
>> known/potential users of this macro.? I guess there shouldn't be that
>> many
>> of them.
>>
>> ?In any case it cannot depend on Kconfig, because the userland won't have
>> access to the configuration, and then presumably wants to handle any and
>> all.
>
> It kind of feels to me like COMMAND_LINE_SIZE shouldn't have been part
> of the UABI to begin with.? I sent a patch to remove it from the
> asm-generic UABI, let's see if anyone knows of a reason it should be UABI:
>
> https://lore.kernel.org/linux-arch/[email protected]/T/#u
Arnd seemed to agree with you about removing COMMAND_LINE_SIZE from the
UABI, any progress on your side?
Thanks,
Alex
>
>
> _______________________________________________
> linux-riscv mailing list
> [email protected]
> http://lists.infradead.org/mailman/listinfo/linux-riscv
On Mon, 21 Jun 2021 at 00:11, Alex Ghiti <[email protected]> wrote:
>
> Hi Palmer,
>
> Le 23/04/2021 à 04:57, Palmer Dabbelt a écrit :
> > On Fri, 02 Apr 2021 11:33:30 PDT (-0700), [email protected] wrote:
> >> On Fri, 2 Apr 2021, David Abdurachmanov wrote:
> >>
> >>> > > > This macro is exported as a part of the user API so it must
> >>> not depend on
> >>> > > > Kconfig. Also changing it (rather than say adding
> >>> COMMAND_LINE_SIZE_V2 or
> >>> > > > switching to an entirely new data object that has its dimension
> >>> set in a
> >>> > > > different way) requires careful evaluation as external binaries
> >>> have and
> >>> > > > will have the value it expands to compiled in, so it's a part
> >>> of the ABI
> >>> > > > too.
> >>> > >
> >>> > > Thanks, I didn't realize this was part of the user BI. In that
> >>> case we
> >>> > > really can't chage it, so we'll have to sort out some other way
> >>> do fix
> >>> > > whatever is going on.
> >>> > >
> >>> > > I've dropped this from fixes.
> >>> >
> >>> > Does increasing COMMAND_LINE_SIZE break user-space binaries? I would
> >>> > expect it to work the same way as adding new enum values, or adding
> >>> > fields at the end of versioned structs, etc.
> >>> > I would assume the old bootloaders/etc will only support up to the
> >>> > old, smaller max command line size, while the kernel will support
> >>> > larger command line size, which is fine.
> >>> > However, if something copies /proc/cmdline into a fixed-size buffer
> >>> > and expects that to work, that will break... that's quite unfortunate
> >>> > user-space code... is it what we afraid of?
> >>> >
> >>> > Alternatively, could expose the same COMMAND_LINE_SIZE, but internally
> >>> > support a larger command line?
> >>>
> >>> Looking at kernel commit history I see PowerPC switched from 512 to
> >>> 2048, and I don't see complaints about the ABI on the mailing list.
> >>>
> >>> If COMMAND_LINE_SIZE is used by user space applications and we
> >>> increase it there shouldn't be problems. I would expect things to
> >>> work, but just get truncated boot args? That is the application will
> >>> continue only to look at the initial 512 chars.
> >>
> >> The macro is in an include/uapi header, so it's exported to the userland
> >> and a part of the user API. I don't know what the consequences are for
> >> the RISC-V port specifically, but it has raised my attention, and I think
> >> it has to be investigated.
> >>
> >> Perhaps it's OK to change it after all, but you'd have to go through
> >> known/potential users of this macro. I guess there shouldn't be that
> >> many
> >> of them.
> >>
> >> In any case it cannot depend on Kconfig, because the userland won't have
> >> access to the configuration, and then presumably wants to handle any and
> >> all.
> >
> > It kind of feels to me like COMMAND_LINE_SIZE shouldn't have been part
> > of the UABI to begin with. I sent a patch to remove it from the
> > asm-generic UABI, let's see if anyone knows of a reason it should be UABI:
> >
> > https://lore.kernel.org/linux-arch/[email protected]/T/#u
>
> Arnd seemed to agree with you about removing COMMAND_LINE_SIZE from the
> UABI, any progress on your side?
Was this ever merged? Don't see this even in linux-next.
On Thu, 10 Nov 2022 at 22:01, Dmitry Vyukov <[email protected]> wrote:
>
> On Mon, 21 Jun 2021 at 00:11, Alex Ghiti <[email protected]> wrote:
> >
> > Hi Palmer,
> >
> > Le 23/04/2021 à 04:57, Palmer Dabbelt a écrit :
> > > On Fri, 02 Apr 2021 11:33:30 PDT (-0700), [email protected] wrote:
> > >> On Fri, 2 Apr 2021, David Abdurachmanov wrote:
> > >>
> > >>> > > > This macro is exported as a part of the user API so it must
> > >>> not depend on
> > >>> > > > Kconfig. Also changing it (rather than say adding
> > >>> COMMAND_LINE_SIZE_V2 or
> > >>> > > > switching to an entirely new data object that has its dimension
> > >>> set in a
> > >>> > > > different way) requires careful evaluation as external binaries
> > >>> have and
> > >>> > > > will have the value it expands to compiled in, so it's a part
> > >>> of the ABI
> > >>> > > > too.
> > >>> > >
> > >>> > > Thanks, I didn't realize this was part of the user BI. In that
> > >>> case we
> > >>> > > really can't chage it, so we'll have to sort out some other way
> > >>> do fix
> > >>> > > whatever is going on.
> > >>> > >
> > >>> > > I've dropped this from fixes.
> > >>> >
> > >>> > Does increasing COMMAND_LINE_SIZE break user-space binaries? I would
> > >>> > expect it to work the same way as adding new enum values, or adding
> > >>> > fields at the end of versioned structs, etc.
> > >>> > I would assume the old bootloaders/etc will only support up to the
> > >>> > old, smaller max command line size, while the kernel will support
> > >>> > larger command line size, which is fine.
> > >>> > However, if something copies /proc/cmdline into a fixed-size buffer
> > >>> > and expects that to work, that will break... that's quite unfortunate
> > >>> > user-space code... is it what we afraid of?
> > >>> >
> > >>> > Alternatively, could expose the same COMMAND_LINE_SIZE, but internally
> > >>> > support a larger command line?
> > >>>
> > >>> Looking at kernel commit history I see PowerPC switched from 512 to
> > >>> 2048, and I don't see complaints about the ABI on the mailing list.
> > >>>
> > >>> If COMMAND_LINE_SIZE is used by user space applications and we
> > >>> increase it there shouldn't be problems. I would expect things to
> > >>> work, but just get truncated boot args? That is the application will
> > >>> continue only to look at the initial 512 chars.
> > >>
> > >> The macro is in an include/uapi header, so it's exported to the userland
> > >> and a part of the user API. I don't know what the consequences are for
> > >> the RISC-V port specifically, but it has raised my attention, and I think
> > >> it has to be investigated.
> > >>
> > >> Perhaps it's OK to change it after all, but you'd have to go through
> > >> known/potential users of this macro. I guess there shouldn't be that
> > >> many
> > >> of them.
> > >>
> > >> In any case it cannot depend on Kconfig, because the userland won't have
> > >> access to the configuration, and then presumably wants to handle any and
> > >> all.
> > >
> > > It kind of feels to me like COMMAND_LINE_SIZE shouldn't have been part
> > > of the UABI to begin with. I sent a patch to remove it from the
> > > asm-generic UABI, let's see if anyone knows of a reason it should be UABI:
> > >
> > > https://lore.kernel.org/linux-arch/[email protected]/T/#u
> >
> > Arnd seemed to agree with you about removing COMMAND_LINE_SIZE from the
> > UABI, any progress on your side?
>
> Was this ever merged? Don't see this even in linux-next.
Ping. Still an issue at least for syzbot.
Le 9/02/2023 à 12:37, Dmitry Vyukov a écrit :
> On Thu, 10 Nov 2022 at 22:01, Dmitry Vyukov <[email protected]> wrote:
>>
>> On Mon, 21 Jun 2021 at 00:11, Alex Ghiti <[email protected]> wrote:
>>>
>>> Hi Palmer,
>>>
>>> Le 23/04/2021 à 04:57, Palmer Dabbelt a écrit :
>>>> On Fri, 02 Apr 2021 11:33:30 PDT (-0700), [email protected] wrote:
>>>>> On Fri, 2 Apr 2021, David Abdurachmanov wrote:
>>>>>
>>>>>>>>> This macro is exported as a part of the user API so it must
>>>>>> not depend on
>>>>>>>>> Kconfig. Also changing it (rather than say adding
>>>>>> COMMAND_LINE_SIZE_V2 or
>>>>>>>>> switching to an entirely new data object that has its dimension
>>>>>> set in a
>>>>>>>>> different way) requires careful evaluation as external binaries
>>>>>> have and
>>>>>>>>> will have the value it expands to compiled in, so it's a part
>>>>>> of the ABI
>>>>>>>>> too.
>>>>>>>>
>>>>>>>> Thanks, I didn't realize this was part of the user BI. In that
>>>>>> case we
>>>>>>>> really can't chage it, so we'll have to sort out some other way
>>>>>> do fix
>>>>>>>> whatever is going on.
>>>>>>>>
>>>>>>>> I've dropped this from fixes.
>>>>>>>
>>>>>>> Does increasing COMMAND_LINE_SIZE break user-space binaries? I would
>>>>>>> expect it to work the same way as adding new enum values, or adding
>>>>>>> fields at the end of versioned structs, etc.
>>>>>>> I would assume the old bootloaders/etc will only support up to the
>>>>>>> old, smaller max command line size, while the kernel will support
>>>>>>> larger command line size, which is fine.
>>>>>>> However, if something copies /proc/cmdline into a fixed-size buffer
>>>>>>> and expects that to work, that will break... that's quite unfortunate
>>>>>>> user-space code... is it what we afraid of?
>>>>>>>
>>>>>>> Alternatively, could expose the same COMMAND_LINE_SIZE, but internally
>>>>>>> support a larger command line?
>>>>>>
>>>>>> Looking at kernel commit history I see PowerPC switched from 512 to
>>>>>> 2048, and I don't see complaints about the ABI on the mailing list.
>>>>>>
>>>>>> If COMMAND_LINE_SIZE is used by user space applications and we
>>>>>> increase it there shouldn't be problems. I would expect things to
>>>>>> work, but just get truncated boot args? That is the application will
>>>>>> continue only to look at the initial 512 chars.
>>>>>
>>>>> The macro is in an include/uapi header, so it's exported to the userland
>>>>> and a part of the user API. I don't know what the consequences are for
>>>>> the RISC-V port specifically, but it has raised my attention, and I think
>>>>> it has to be investigated.
>>>>>
>>>>> Perhaps it's OK to change it after all, but you'd have to go through
>>>>> known/potential users of this macro. I guess there shouldn't be that
>>>>> many
>>>>> of them.
>>>>>
>>>>> In any case it cannot depend on Kconfig, because the userland won't have
>>>>> access to the configuration, and then presumably wants to handle any and
>>>>> all.
>>>>
>>>> It kind of feels to me like COMMAND_LINE_SIZE shouldn't have been part
>>>> of the UABI to begin with. I sent a patch to remove it from the
>>>> asm-generic UABI, let's see if anyone knows of a reason it should be UABI:
>>>>
>>>> https://lore.kernel.org/linux-arch/[email protected]/T/#u
>>>
>>> Arnd seemed to agree with you about removing COMMAND_LINE_SIZE from the
>>> UABI, any progress on your side?
>>
>> Was this ever merged? Don't see this even in linux-next.
>
> Ping. Still an issue at least for syzbot.
Yes, agreed, Palmer proposed the following instead since blindly
increasing the command line size would break userspace ABI:
https://lore.kernel.org/lkml/[email protected]/T/
I will bump this thread to make progress, thanks for the ping.
Alex
>
> _______________________________________________
> linux-riscv mailing list
> [email protected]
> http://lists.infradead.org/mailman/listinfo/linux-riscv
On Tue, 16 Mar 2021 12:34:20 PDT (-0700), [email protected] wrote:
> Increase COMMAND_LINE_SIZE as the current default value is too low
> for syzbot kernel command line.
>
> Reported-by: Dmitry Vyukov <[email protected]>
> Signed-off-by: Alexandre Ghiti <[email protected]>
> ---
> arch/riscv/include/uapi/asm/setup.h | 8 ++++++++
> 1 file changed, 8 insertions(+)
> create mode 100644 arch/riscv/include/uapi/asm/setup.h
>
> diff --git a/arch/riscv/include/uapi/asm/setup.h b/arch/riscv/include/uapi/asm/setup.h
> new file mode 100644
> index 000000000000..66b13a522880
> --- /dev/null
> +++ b/arch/riscv/include/uapi/asm/setup.h
> @@ -0,0 +1,8 @@
> +/* SPDX-License-Identifier: GPL-2.0-only WITH Linux-syscall-note */
> +
> +#ifndef _UAPI_ASM_RISCV_SETUP_H
> +#define _UAPI_ASM_RISCV_SETUP_H
> +
> +#define COMMAND_LINE_SIZE 1024
> +
> +#endif /* _UAPI_ASM_RISCV_SETUP_H */
This is now back on for-next, with some commit text explaining why it's
not a uABI change as per Arnd's message
<https://lore.kernel.org/linux-riscv/[email protected]/#t>.
I intend on sending this one for 6.3 as syzkaller depends on it.
Thanks!