Re: [PATCH 0/2] scsi: target+fcoe: replace -EEXIST with -EBUSY in module_init() paths

[James Bottomley <James.Bottomley@HansenPartnership.com>]

On Sun, 2025-12-21 at 04:30 +0100, Daniel Gomez wrote:
> On 20/12/2025 05.27, James Bottomley wrote:
> > On Sat, 2025-12-20 at 04:37 +0100, Daniel Gomez wrote:
> > > The error code -EEXIST is reserved by the kernel module loader to
> > > indicate that a module with the same name is already loaded. When
> > > a module's init function returns -EEXIST, kmod interprets this as
> > > "module already loaded" and reports success instead of failure
> > > [1].
> > 
> > That reference doesn't sufficiently explain why this error code
> > should be unique to modules.
> 
> It's unique only to the module initialization. You can find how it's
> used in the kernel module code at module_patient_check_exists() in
> kernel/module/main.c [1].
> 
> In addition, init_module(2) man pages indicates this:
> 
> man 2 init_module | grep EEXIST
>        EEXIST A module with this name is already loaded.
> 
> So, a module that is already loaded will be detected by the kernel
> module loader and the EEXIST error will be returned. This will be
> detected by kmod as success [2]. I think this functionality was added
> very early on in kmod by commit 5f35147 "libkmod-module: probe: add
> flag to stop loading on already loaded" [3]. Prior to that, module-
> init-tools had the same behavior [4]. Even in
> modutils [5], we had back then in insmod/insmod.c:2088:
> 
> 		case EEXIST:
> 			if (dolock) {
> 				/*
> 				 * Assume that we were just invoked
> 				 * simultaneous with another insmod
> 				 * and return success.
> 				 */
> 				exit_status = 0;
> 				goto out;
> 			}
> 			error("a module named %s already exists",
> m_name);
> 			goto out;
> 
> Link:
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/kernel/module/main.c?h=v6.19-rc1#n3206
>  [1]
> Link:
> https://github.com/kmod-project/kmod/blob/v34.2/libkmod/libkmod-module.c#L1088
>  [2]
> Link:
> https://github.com/kmod-project/kmod/commit/5f3514731ef82084c1a24b15445e0f1352681a19
>  [3]
> Link:
> https://git.kernel.org/pub/scm/utils/kernel/module-init-tools/module-init-tools.git/tree/modprobe.c#n1797
>  [4]
> Link:
> https://cdn.kernel.org/pub/linux/utils/kernel/modutils/v2.4/modutils-2.4.27.tar.gz
>  [5]
> 
> > EEXIST is used all over the kernel to indicate
> > that something being attempted has already happened or does already
> > exist and that seems perfectly logical .... please explain why
> > you're
> 
> That is correct but not all are conflicts within the
> module_init()/init_module(2) path. I have detected 40+ cases where
> this error is returned and another 20+ where error is returned but in
> upper layers of the module itself, not propagated back to userspace.
> So far, I've only sent just a few + docs:
> 
> https://lore.kernel.org/all/20251220-dev-module-init-eexists-dm-devel-v1-1-90ed00444ea0@samsung.com
>  
> https://lore.kernel.org/all/20251220-dev-module-init-eexists-keyring-v1-1-a2f23248c300@samsung.com
> https://lore.kernel.org/all/20251220-dev-module-init-eexists-linux-acpi-v1-1-af59b1a0e217@samsung.com
> https://lore.kernel.org/all/20251218-dev-module-init-eexists-modules-docs-v1-0-361569aa782a@samsung.com

None of that answers the why question:  Given that EEXIST is used all
over the kernel, for what appear to be fairly legitimate cases, why
would we suddenly want it to become only for modules?  I get that we
can, as you propose patches above, but why should we bother?  It seems
to be a useful error code outside the module use case, so why the need
to restrict it to being only for modules?

Regards,

James