[PATCH v7 00/16] mm: jit/text allocator

[Liviu Dudau]

On Fri, May 03, 2024 at 09:28:25AM +0300, Mike Rapoport wrote:
> On Fri, May 03, 2024 at 01:23:30AM +0100, Liviu Dudau wrote:
> > On Thu, May 02, 2024 at 04:07:05PM -0700, Luis Chamberlain wrote:
> > > On Thu, May 02, 2024 at 11:50:36PM +0100, Liviu Dudau wrote:
> > > > On Mon, Apr 29, 2024 at 09:29:20AM -0700, Luis Chamberlain wrote:
> > > > > On Mon, Apr 29, 2024 at 03:16:04PM +0300, Mike Rapoport wrote:
> > > > > > From: "Mike Rapoport (IBM)" <rppt at kernel.org>
> > > > > > 
> > > > > > Hi,
> > > > > > 
> > > > > > The patches are also available in git:
> > > > > > https://git.kernel.org/pub/scm/linux/kernel/git/rppt/linux.git/log/?h=execmem/v7
> > > > > > 
> > > > > > v7 changes:
> > > > > > * define MODULE_{VADDR,END} for riscv32 to fix the build and avoid
> > > > > >   #ifdefs in a function body
> > > > > > * add Acks, thanks everybody
> > > > > 
> > > > > Thanks, I've pushed this to modules-next for further exposure / testing.
> > > > > Given the status of testing so far with prior revisions, in that only a
> > > > > few issues were found and that those were fixed, and the status of
> > > > > reviews, this just might be ripe for v6.10.
> > > > 
> > > > Looks like there is still some work needed. I've picked up next-20240501
> > > > and on arch/mips with CONFIG_MODULE_COMPRESS_XZ=y and CONFIG_MODULE_DECOMPRESS=y
> > > > I fail to load any module:
> > > > 
> > > > # modprobe rfkill
> > > > [11746.539090] Invalid ELF header magic: != ELF
> > > > [11746.587149] execmem: unable to allocate memory
> > > > modprobe: can't load module rfkill (kernel/net/rfkill/rfkill.ko.xz): Out of memory
> > > > 
> > > > The (hopefully) relevant parts of my .config:
> > > 
> > > Thanks for the report! Any chance we can get you to try a bisection? I
> > > think it should take 2-3 test boots. To help reduce scope you try modules-next:
> > > 
> > > https://git.kernel.org/pub/scm/linux/kernel/git/mcgrof/linux.git/log/?h=modules-next
> > > 
> > > Then can you check by resetting your tree to commmit 3fbe6c2f820a76 (mm:
> > > introduce execmem_alloc() and execmem_free()"). I suspect that should
> > > boot, so your bad commit would be the tip 3c2c250cb3a5fbb ("bpf: remove
> > > CONFIG_BPF_JIT dependency on CONFIG_MODULES of").
> > > 
> > > That gives us only a few commits to bisect:
> > > 
> > > git log --oneline 3fbe6c2f820a76bc36d5546bda85832f57c8fce2..
> > > 3c2c250cb3a5 (HEAD -> modules-next, korg/modules-next) bpf: remove CONFIG_BPF_JIT dependency on CONFIG_MODULES of
> > > 11e8e65cce5c kprobes: remove dependency on CONFIG_MODULES
> > > e10cbc38697b powerpc: use CONFIG_EXECMEM instead of CONFIG_MODULES where appropriate
> > > 4da3d38f24c5 x86/ftrace: enable dynamic ftrace without CONFIG_MODULES
> > > 13ae3d74ee70 arch: make execmem setup available regardless of CONFIG_MODULES
> > > 460bbbc70a47 powerpc: extend execmem_params for kprobes allocations
> > > e1a14069b5b4 arm64: extend execmem_info for generated code allocations
> > > 971e181c6585 riscv: extend execmem_params for generated code allocations
> > > 0fa276f26721 mm/execmem, arch: convert remaining overrides of module_alloc to execmem
> > > 022cef244287 mm/execmem, arch: convert simple overrides of module_alloc to execmem
> > > 
> > > With 2-3 boots we should be to tell which is the bad commit.
> > 
> > Looks like 0fa276f26721 is the first bad commit.
> > 
> > $ git bisect log
> > # bad: [3c2c250cb3a5fbbccc4a4ff4c9354c54af91f02c] bpf: remove CONFIG_BPF_JIT dependency on CONFIG_MODULES of
> > # good: [3fbe6c2f820a76bc36d5546bda85832f57c8fce2] mm: introduce execmem_alloc() and execmem_free()
> > git bisect start '3c2c250cb3a5' '3fbe6c2f820a76'
> > # bad: [460bbbc70a47e929b1936ca68979f3b79f168fc6] powerpc: extend execmem_params for kprobes allocations
> > git bisect bad 460bbbc70a47e929b1936ca68979f3b79f168fc6
> > # bad: [0fa276f26721e0ffc2ae9c7cf67dcc005b43c67e] mm/execmem, arch: convert remaining overrides of module_alloc to execmem
> > git bisect bad 0fa276f26721e0ffc2ae9c7cf67dcc005b43c67e
> > # good: [022cef2442870db738a366d3b7a636040c081859] mm/execmem, arch: convert simple overrides of module_alloc to execmem
> > git bisect good 022cef2442870db738a366d3b7a636040c081859
> > # first bad commit: [0fa276f26721e0ffc2ae9c7cf67dcc005b43c67e] mm/execmem, arch: convert remaining overrides of module_alloc to execmem
> > 
> > Maybe MIPS also needs a ARCH_WANTS_EXECMEM_LATE?
> 
> I don't think so. It rather seems there's a bug in the initialization of
> the defaults in execmem. This should fix it:
> 
> diff --git a/mm/execmem.c b/mm/execmem.c
> index f6dc3fabc1ca..0c4b36bc6d10 100644
> --- a/mm/execmem.c
> +++ b/mm/execmem.c
> @@ -118,7 +118,6 @@ static void __init __execmem_init(void)
>  		info->ranges[EXECMEM_DEFAULT].end = VMALLOC_END;
>  		info->ranges[EXECMEM_DEFAULT].pgprot = PAGE_KERNEL_EXEC;
>  		info->ranges[EXECMEM_DEFAULT].alignment = 1;
> -		return;
>  	}
>  
>  	if (!execmem_validate(info))
>

That does, indeed, fix the issue!

You can add my Tested-by when you submit the patch.

Best regards,
Liviu


> 
> -- 
> Sincerely yours,
> Mike.
> 

-- 
Everyone who uses computers frequently has had, from time to time,
a mad desire to attack the precocious abacus with an axe.
       	   	      	     	  -- John D. Clark, Ignition!