Printing with overflow checks can cause modpost errors
Subsystems (4)
Hello,
Recently some of have been running into modpost errors more frequently. Ahead of Kangrejos, I am trying to study them, the one I looked at today is truly weird, below are more details.
I narrowed it down to the print statement and specifically the FFI call to printk bindings. This was first reported by Timur Tabi on CC.
With CONFIG_RUST_OVERFLOW_CHECKS=y and CONFIG_RUST_BUILD_ASSERT_ALLOW=y, the following patch when applied to nova-core will fail to build with following errors. The question is why does the overflow checking fail since the arithmetic is valid, and why only during printing (and say not during the call to write32).
MODPOST Module.symvers
ERROR: modpost: "rust_build_error" [drivers/gpu/nova-core/nova_core.ko] undefined! make[2]: *** [scripts/Makefile.modpost:147: Module.symvers] Error 1 make[1]: *** [/home/joelaf/repo/linux-nova-rm-call/Makefile:1961: modpost] Error 2
make: *** [Makefile:248: __sub-make] Error 2
Any comments or thoughts?
---8<-----------------------
From: Joel Fernandes <[email protected]>
Subject: [PATCH] Reproduce modpost error due to print
--- drivers/gpu/nova-core/nova_core.rs | 1 + drivers/gpu/nova-core/test_modpost.rs | 60 +++++++++++++++++++++++++++ 2 files changed, 61 insertions(+) create mode 100644 drivers/gpu/nova-core/test_modpost.rs
drivers/gpu/nova-core/nova_core.rs +1 -0
@@ -15,6 +15,7 @@mod sbuffer;mod util;mod vbios; +mod test_modpost;pub (crate )const MODULE_NAME : &kernel::str::CStr = <LocalModule as kernel::ModuleMetadata >::NAME ;
drivers/gpu/nova-core/test_modpost.rs new file +60 -0
60 lines not on this page · raw message
Signature · 1 line
2.34.1
Quoted 14 lines · Joel Fernandes
Hello,
Recently some of have been running into modpost errors more frequently. Ahead of Kangrejos, I am trying to study them, the one I looked at today is truly weird, below are more details.I narrowed it down to the print statement and specifically the FFI call to printk bindings. This was first reported by Timur Tabi on CC.
With
CONFIG_RUST_OVERFLOW_CHECKS=yandCONFIG_RUST_BUILD_ASSERT_ALLOW=y, the following patch when applied to nova-core will fail to build with following errors. The question is why does the overflow checking fail since the arithmetic is valid, and why only during printing (and say not during the call to write32).
Quoted 8 lines
MODPOST Module.symvers
ERROR: modpost: "rust_build_error" [drivers/gpu/nova-core/nova_core.ko] undefined! make[2]: *** [scripts/Makefile.modpost:147: Module.symvers] Error 1 make[1]: *** [/home/joelaf/repo/linux-nova-rm-call/Makefile:1961: modpost] Error 2
make: *** [Makefile:248: __sub-make] Error 2Any comments or thoughts?
Io::write32 tries to do a bounds check at compile time and if it cannot be done it causes a build error. it looks like because a pointer to offset is passed across a ffi boundary, rustc makes no assumptions about the value of offset. so it cannot do the bounds check at compile time and causes a build error.
Best regards,
Andrew Ballance
Quoted 21 lines · Andrew Ballance
Quoted 20 lines · Joel Fernandes
Hello,
Recently some of have been running into modpost errors more frequently. Ahead of Kangrejos, I am trying to study them, the one I looked at today is truly weird, below are more details.I narrowed it down to the print statement and specifically the FFI call to printk bindings. This was first reported by Timur Tabi on CC.
With
CONFIG_RUST_OVERFLOW_CHECKS=yandCONFIG_RUST_BUILD_ASSERT_ALLOW=y, the following patch when applied to nova-core will fail to build with following errors. The question is why does the overflow checking fail since the arithmetic is valid, and why only during printing (and say not during the call to write32).MODPOST Module.symvers
ERROR: modpost: "rust_build_error" [drivers/gpu/nova-core/nova_core.ko] undefined! make[2]: *** [scripts/Makefile.modpost:147: Module.symvers] Error 1 make[1]: *** [/home/joelaf/repo/linux-nova-rm-call/Makefile:1961: modpost] Error 2
make: *** [Makefile:248: __sub-make] Error 2
Quoted 8 lines
Quoted 2 lines
Any comments or thoughts?
Io::write32tries to do a bounds check at compile time and if it cannot be done it causes a build error. it looks like because a pointer to offset is passed across a ffi boundary, rustc makes no assumptions about the value of offset. so it cannot do the bounds check at compile time and causes a build error.
Are you saying this issue is related to iowrite32? I don't think so because the issue does not happen if you comment out the pr_err in my example and leave the write32 as it is. So it is something with the call to printk (FFI).
Why can't it assume the value of offset? All the values to compute it are available at compile time right?
thanks,
- Joel
Quoted 35 lines · Joel Fernandes
Quoted 21 lines · Andrew Ballance
Quoted 20 lines · Joel Fernandes
Hello,
Recently some of have been running into modpost errors more frequently. Ahead of Kangrejos, I am trying to study them, the one I looked at today is truly weird, below are more details.I narrowed it down to the print statement and specifically the FFI call to printk bindings. This was first reported by Timur Tabi on CC.
With
CONFIG_RUST_OVERFLOW_CHECKS=yandCONFIG_RUST_BUILD_ASSERT_ALLOW=y, the following patch when applied to nova-core will fail to build with following errors. The question is why does the overflow checking fail since the arithmetic is valid, and why only during printing (and say not during the call to write32).MODPOST Module.symvers
ERROR: modpost: "rust_build_error" [drivers/gpu/nova-core/nova_core.ko] undefined! make[2]: *** [scripts/Makefile.modpost:147: Module.symvers] Error 1 make[1]: *** [/home/joelaf/repo/linux-nova-rm-call/Makefile:1961: modpost] Error 2
make: *** [Makefile:248: __sub-make] Error 2Quoted 8 lines
Quoted 2 lines
Any comments or thoughts?
Io::write32tries to do a bounds check at compile time and if it cannot be done it causes a build error. it looks like because a pointer to offset is passed across a ffi boundary, rustc makes no assumptions about the value of offset. so it cannot do the bounds check at compile time and causes a build error.Are you saying this issue is related to iowrite32? I don't think so because the issue does not happen if you comment out the
pr_errin my example and leave the write32 as it is. So it is something with the call to printk (FFI).
Quoted 7 lines
Why can't it assume the value of offset? All the values to compute it are available at compile time right?
thanks,
- Joel
This is a resend because I forgot to cc the mailing list.
it has to do with the FFI call. The value of offset can be found out at compile time, but because a pointer is passed through, the c side could theoretically change the value before write32 is called.
The pointer passed is const so rustc should assume that the c side does not change offset, but looks like rustc does not do that.
as a test i created a version where a copy of offset is passed to printk instead of offset and it compiles.
e.g:
// SNIP
let offset = <B as kernel::io::register::RegisterBase<$base>>::BASE
+ Self::OFFSET
+ (idx * Self::STRIDE);
let offset_copy = offset;
pr_err!("{}", offset_copy);io.write32(self.0, offset);
// SNIP
Best regards,
Andrew Ballance
Quoted 3 lines · Andrew Ballance
The pointer passed is const so rustc should assume that the c side does not change offset, but looks like rustc does not do that.
That is not possible -- a const pointer does not guarantee the value will not be changed.
Cheers,
Miguel
Alice Ryhl
maintainerQuoted 9 lines · Miguel Ojeda
Quoted 3 lines · Andrew Ballance
The pointer passed is const so rustc should assume that the c side does not change offset, but looks like rustc does not do that.
That is not possible -- a const pointer does not guarantee the value will not be changed.
I believe this code is using an immutable reference and not just a pointer, so it would be UB to use it to write to offset, and so it would be valid to assume it has not changed. But I think that in most scenarios, Rust only optimizes using that information when the reference appears as a function argument, which is not the case here.
Alice
Quoted 6 lines · Alice Ryhl
I believe this code is using an immutable reference and not just a pointer, so it would be UB to use it to write to
offset, and so it would be valid to assume it has not changed. But I think that in most scenarios, Rust only optimizes using that information when the reference appears as a function argument, which is not the case here.
I understood Andrew as talking about the C side, i.e. a guarantee coming from the C side.
Callers can guarantee in other ways, but I wanted to clarify that the C const pointer doesn't really do anything.
Cheers,
Miguel
Quoted 61 lines · Andrew Ballance
Quoted 35 lines · Joel Fernandes
Quoted 21 lines · Andrew Ballance
Quoted 20 lines · Joel Fernandes
Hello,
Recently some of have been running into modpost errors more frequently. Ahead of Kangrejos, I am trying to study them, the one I looked at today is truly weird, below are more details.I narrowed it down to the print statement and specifically the FFI call to printk bindings. This was first reported by Timur Tabi on CC.
With
CONFIG_RUST_OVERFLOW_CHECKS=yandCONFIG_RUST_BUILD_ASSERT_ALLOW=y, the following patch when applied to nova-core will fail to build with following errors. The question is why does the overflow checking fail since the arithmetic is valid, and why only during printing (and say not during the call to write32).MODPOST Module.symvers
ERROR: modpost: "rust_build_error" [drivers/gpu/nova-core/nova_core.ko] undefined! make[2]: *** [scripts/Makefile.modpost:147: Module.symvers] Error 1 make[1]: *** [/home/joelaf/repo/linux-nova-rm-call/Makefile:1961: modpost] Error 2
make: *** [Makefile:248: __sub-make] Error 2Quoted 8 lines
Quoted 2 lines
Any comments or thoughts?
Io::write32tries to do a bounds check at compile time and if it cannot be done it causes a build error. it looks like because a pointer to offset is passed across a ffi boundary, rustc makes no assumptions about the value of offset. so it cannot do the bounds check at compile time and causes a build error.Are you saying this issue is related to iowrite32? I don't think so because the issue does not happen if you comment out the
pr_errin my example and leave the write32 as it is. So it is something with the call to printk (FFI).Quoted 7 lines
Why can't it assume the value of offset? All the values to compute it are available at compile time right?
thanks,
- Joel
This is a resend because I forgot to cc the mailing list.
it has to do with the FFI call. The value of offset can be found out at compile time, but because a pointer is passed through, the c side could theoretically change the value before write32 is called.
The pointer passed is const so rustc should assume that the c side does not change offset, but looks like rustc does not do that.as a test i created a version where a copy of offset is passed to printk instead of offset and it compiles.
e.g:
// SNIP
let offset = <B as kernel::io::register::RegisterBase<$base>>::BASE
+Self::OFFSET
+ (idx *Self::STRIDE);
letoffset_copy= offset;
Quoted 3 lines
pr_err!("{}",
offset_copy);io.write32(self.0, offset);
// SNIP
Andrew,
Thanks, I came to the same conclusion. After the first FFI call, the compiler has to redo the overflow checking and cannot optimize it away. The issue does not happen if either drop the print, or the io.write32, so it is their combination that causes the issue.
So I guess how do we fix it? One crude way could be for the print macro to alias its arguments automatically. But that does not fix the general problem as it could occur with other FFI calls as well, not just printing.
thanks,
- Joel
Quoted 72 lines · Joel Fernandes
Quoted 61 lines · Andrew Ballance
Quoted 35 lines · Joel Fernandes
Quoted 21 lines · Andrew Ballance
Quoted 20 lines · Joel Fernandes
Hello,
Recently some of have been running into modpost errors more frequently. Ahead of Kangrejos, I am trying to study them, the one I looked at today is truly weird, below are more details.I narrowed it down to the print statement and specifically the FFI call to printk bindings. This was first reported by Timur Tabi on CC.
With
CONFIG_RUST_OVERFLOW_CHECKS=yandCONFIG_RUST_BUILD_ASSERT_ALLOW=y, the following patch when applied to nova-core will fail to build with following errors. The question is why does the overflow checking fail since the arithmetic is valid, and why only during printing (and say not during the call to write32).MODPOST Module.symvers
ERROR: modpost: "rust_build_error" [drivers/gpu/nova-core/nova_core.ko] undefined! make[2]: *** [scripts/Makefile.modpost:147: Module.symvers] Error 1 make[1]: *** [/home/joelaf/repo/linux-nova-rm-call/Makefile:1961: modpost] Error 2
make: *** [Makefile:248: __sub-make] Error 2Quoted 8 lines
Quoted 2 lines
Any comments or thoughts?
Io::write32tries to do a bounds check at compile time and if it cannot be done it causes a build error. it looks like because a pointer to offset is passed across a ffi boundary, rustc makes no assumptions about the value of offset. so it cannot do the bounds check at compile time and causes a build error.Are you saying this issue is related to iowrite32? I don't think so because the issue does not happen if you comment out the
pr_errin my example and leave the write32 as it is. So it is something with the call to printk (FFI).Quoted 7 lines
Why can't it assume the value of offset? All the values to compute it are available at compile time right?
thanks,
- Joel
This is a resend because I forgot to cc the mailing list.
it has to do with the FFI call. The value of offset can be found out at compile time, but because a pointer is passed through, the c side could theoretically change the value before write32 is called.
The pointer passed is const so rustc should assume that the c side does not change offset, but looks like rustc does not do that.as a test i created a version where a copy of offset is passed to printk instead of offset and it compiles.
e.g:
// SNIP
let offset = <B as kernel::io::register::RegisterBase<$base>>::BASE
+Self::OFFSET
+ (idx *Self::STRIDE);
letoffset_copy= offset;Quoted 3 lines
pr_err!("{}",
offset_copy);io.write32(self.0, offset);
// SNIPAndrew,
Thanks, I came to the same conclusion. After the first FFI call, the compiler has to redo the overflow checking and cannot optimize it away. The issue does not happen if either drop the print, or the io.write32, so it is their combination that causes the issue.
Quoted 3 lines
So I guess how do we fix it? One crude way could be for the print macro to alias its arguments automatically. But that does not fix the general problem as it could occur with other FFI calls as well, not just printing.
I even see it with the following simple example, just using same variable between safe and unsafe code, here offset is not even going to the C side (there's no print):
let mut offset = 0;
unsafe {
offset = 5;
}
io.write32(self.0, offset);So maybe the issue is that the FFI related to print involves unsafe { }, hence it causes the same issue there too?
thanks,
- Joel
Quoted 81 lines · Joel Fernandes
Quoted 72 lines · Joel Fernandes
Quoted 61 lines · Andrew Ballance
Quoted 35 lines · Joel Fernandes
Quoted 21 lines · Andrew Ballance
Quoted 20 lines · Joel Fernandes
Hello,
Recently some of have been running into modpost errors more frequently. Ahead of Kangrejos, I am trying to study them, the one I looked at today is truly weird, below are more details.I narrowed it down to the print statement and specifically the FFI call to printk bindings. This was first reported by Timur Tabi on CC.
With
CONFIG_RUST_OVERFLOW_CHECKS=yandCONFIG_RUST_BUILD_ASSERT_ALLOW=y, the following patch when applied to nova-core will fail to build with following errors. The question is why does the overflow checking fail since the arithmetic is valid, and why only during printing (and say not during the call to write32).MODPOST Module.symvers
ERROR: modpost: "rust_build_error" [drivers/gpu/nova-core/nova_core.ko] undefined! make[2]: *** [scripts/Makefile.modpost:147: Module.symvers] Error 1 make[1]: *** [/home/joelaf/repo/linux-nova-rm-call/Makefile:1961: modpost] Error 2
make: *** [Makefile:248: __sub-make] Error 2Quoted 8 lines
Quoted 2 lines
Any comments or thoughts?
Io::write32tries to do a bounds check at compile time and if it cannot be done it causes a build error. it looks like because a pointer to offset is passed across a ffi boundary, rustc makes no assumptions about the value of offset. so it cannot do the bounds check at compile time and causes a build error.Are you saying this issue is related to iowrite32? I don't think so because the issue does not happen if you comment out the
pr_errin my example and leave the write32 as it is. So it is something with the call to printk (FFI).Quoted 7 lines
Why can't it assume the value of offset? All the values to compute it are available at compile time right?
thanks,
- Joel
This is a resend because I forgot to cc the mailing list.
it has to do with the FFI call. The value of offset can be found out at compile time, but because a pointer is passed through, the c side could theoretically change the value before write32 is called.
The pointer passed is const so rustc should assume that the c side does not change offset, but looks like rustc does not do that.as a test i created a version where a copy of offset is passed to printk instead of offset and it compiles.
e.g:
// SNIP
let offset = <B as kernel::io::register::RegisterBase<$base>>::BASE
+Self::OFFSET
+ (idx *Self::STRIDE);
letoffset_copy= offset;Quoted 3 lines
pr_err!("{}",
offset_copy);io.write32(self.0, offset);
// SNIPAndrew,
Thanks, I came to the same conclusion. After the first FFI call, the compiler has to redo the overflow checking and cannot optimize it away. The issue does not happen if either drop the print, or the io.write32, so it is their combination that causes the issue.Quoted 3 lines
So I guess how do we fix it? One crude way could be for the print macro to alias its arguments automatically. But that does not fix the general problem as it could occur with other FFI calls as well, not just printing.
I even see it with the following simple example, just using same variable between safe and unsafe code, here offset is not even going to the C side (there's no print):
Quoted 9 lines
let mut offset = 0; unsafe { offset = 5; } io.write32(self.0, offset);So maybe the issue is that the FFI related to print involves unsafe { }, hence it causes the same issue there too?
For the print usecase, if I move the match statement into a closure, the issue disappears. Below is a patch, I can/will send it as a proper patch but any initial thoughts? (Fair warning: Only lightly tested).
---8<-----------------------
From: Joel Fernandes <[email protected]>
Subject: [PATCH] rust: print: Fix issue with rust_build_error
When printing just before calling io.write32(), modpost fails. The issue seems to be that, the printk arguments are used in an unsafe block, thus Rust cannot trust its value. This can be fixed by simply creating a variable alias for each variable.
Fix it in an even simpler way by just call printk in a closure. Rust captures the arguments into the closure's arguments thus breaking the dependencies. Tested with the following snippet and it builds with patch:
let offset = 0;
pr_err!("{}", offset);
io.write32(base, offset);--- rust/kernel/print.rs | 33 ++++++++++++++++++--------------- 1 file changed, 18 insertions(+), 15 deletions(-)
rust/kernel/print.rs +18 -15
@@ -143,27 +143,30 @@ pub fn call_printk_cont(args: fmt::Arguments<'_>) {#[expect(clippy::crate_in_macro_def)] macro_rules! print_macro ( // The non-continuation cases (most of them, e.g. `INFO`). - ($format_string:path, false, $($arg:tt)+) => ( + ($format_string:path, false, $($arg:tt)+) => ({ // To remain sound, `arg`s must be expanded outside the `unsafe` block. // Typically one would use a `let` binding for that; however, `format_args!` // takes borrows on the arguments, but does not extend the scope of temporaries. // Therefore, a `match` expression is used to keep them around, since // the scrutinee is kept until the end of the `match`. - match $crate::prelude::fmt!($($arg)+) { - // SAFETY: This hidden macro should only be called by the documented - // printing macros which ensure the format string is one of the fixed - // ones. All `__LOG_PREFIX`s are null-terminated as they are generated - // by the `module!` proc macro or fixed values defined in a kernel - // crate. - args => unsafe { - $crate::print::call_printk( - &$format_string, - crate::__LOG_PREFIX, - args, - ); + let print_fn = || { + match $crate::prelude::fmt!($($arg)+) { + // SAFETY: This hidden macro should only be called by the documented + // printing macros which ensure the format string is one of the fixed + // ones. All `__LOG_PREFIX`s are null-terminated as they are generated + // by the `module!` proc macro or fixed values defined in a kernel + // crate. + args => unsafe { + $crate::print::call_printk( + &$format_string, + crate::__LOG_PREFIX, + args, + ); + } } - } - ); + }; + print_fn(); + }); // The `CONT` case. ($format_string:path, true, $($arg:tt)+) => (
Signature · 1 line
2.34.1
Timur Tabi
Tested-byQuoted 12 lines · Joel Fernandes
From: Joel Fernandes <[email protected]>
Subject: [PATCH] rust: print: Fix issue withrust_build_errorWhen printing just before calling
io.write32(), modpost fails. The issue seems to be that, the printk arguments are used in an unsafe block, thus Rust cannot trust its value. This can be fixed by simply creating a variable alias for each variable.Fix it in an even simpler way by just call printk in a closure. Rust captures the arguments into the closure's arguments thus breaking the dependencies. Tested with the following snippet and it builds with patch:
Quoted 5 lines
let offset = 0;
pr_err!("{}", offset);
io.write32(base, offset);Signed-off-by Joel Fernandes <[email protected]>
This fixes an issue I had with pr_err!() and io.write32() and causing rust_build_error.