[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-Id: <20231119.182544.2069714044528296795.fujita.tomonori@gmail.com>
Date: Sun, 19 Nov 2023 18:25:44 +0900 (JST)
From: FUJITA Tomonori <fujita.tomonori@...il.com>
To: benno.lossin@...ton.me
Cc: andrew@...n.ch, boqun.feng@...il.com, fujita.tomonori@...il.com,
netdev@...r.kernel.org, rust-for-linux@...r.kernel.org, tmgross@...ch.edu,
miguel.ojeda.sandonis@...il.com, wedsonaf@...il.com
Subject: Re: [PATCH net-next v7 2/5] rust: net::phy add module_phy_driver
macro
On Fri, 17 Nov 2023 23:01:58 +0000
Benno Lossin <benno.lossin@...ton.me> wrote:
> On 11/17/23 23:54, Andrew Lunn wrote:
>> Each kernel module should be in its own symbol name space. The only
>> symbols which are visible outside of the module are those exported
>> using EXPORT_SYMBOL_GPL() or EXPORT_SYMBOL(). A PHY driver does not
>> export anything, in general.
>>
>> Being built in also does not change this.
>>
>> Neither drivers/net/phy/ax88796b_rust.o nor
>> rust/doctests_kernel_generated.o should have exported this symbol.
>>
>> I've no idea how this actually works, i guess there are multiple
>> passes through the linker? Maybe once to resolve symbols across object
>> files within a module. Normal global symbols are then made local,
>> leaving only those exported with EXPORT_SYMBOL_GPL() or
>> EXPORT_SYMBOL()? A second pass through linker then links all the
>> exported symbols thorough the kernel?
>
> I brought this issue up in [1], but I was a bit confused by your last
> reply there, as I have no idea how the `EXPORT_SYMBOL` macros work.
>
> IIRC on the Rust side all public items are automatically GPL exported.
Hmm, they are public but doesn't look like exported by EXPORT_SYMBOL()
or EXPORT_SYMBOL_GPL().
> But `#[no_mangle]` is probably a special case, since in userspace it
> is usually used to do interop with C (and therefore the symbol is always
> exported with the name not mangled).
>
> One fix would be to make the name unique by using the type name supplied
> in the macro.
>
> [1]: https://lore.kernel.org/rust-for-linux/1aea7ddb-73b7-8228-161e-e2e4ff5bc98d@proton.me/
I fixed this in a different way; declaring
__mod_mdio__phydev_device_table only when built as a module.
(@device_table [$($dev:expr),+]) => {
+ #[cfg(MODULE)]
#[no_mangle]
static __mod_mdio__phydev_device_table: [::kernel::bindings::mdio_device_id;
$crate::module_phy_driver!(@count_devices $($dev),+) + 1] = [
This is used for dynamic loading so when a phy driver is built-in, we
don't need this.
When a driver is built as a module, the user-space tool converts this
into moduole alias information. This __mod_mdio__phydev_device_table
isn't loaded into kernel so no conflict even when multiple phy drivers
are loaded.
Powered by blists - more mailing lists