Re: [PATCH v2 06/10] module: introduce MODULE_STATE_GONE

[Mike Rapoport <rppt@kernel.org>]

On Thu, Jan 23, 2025 at 03:16:28PM +0100, Petr Pavlu wrote:
> On 1/21/25 10:57, Mike Rapoport wrote:
> > In order to use execmem's API for temporal remapping of the memory
> > allocated from ROX cache as writable, there is a need to distinguish
> > between the state when the module is being formed and the state when it is
> > deconstructed and freed so that when module_memory_free() is called from
> > error paths during module loading it could restore ROX mappings.
> > 
> > Replace open coded checks for MODULE_STATE_UNFORMED with a helper
> > function module_is_formed() and add a new MODULE_STATE_GONE that will be
> > set when the module is deconstructed and freed.
> 
> I don't fully follow why this case requires a new module state. My
> understanding it that the function load_module() has the necessary
> context that after calling layout_and_allocate(), the updated ROX
> mappings need to be restored. I would then expect the function to be
> appropriately able to unwind this operation in case of an error. It
> could be done by having a helper that walks the mappings and calls
> execmem_restore_rox(), or if you want to keep it in module_memory_free()
> as done in the patch #7 then a flag could be passed down to
> module_deallocate() -> free_mod_mem() -> module_memory_free()?

Initially I wanted to track ROX <-> RW transitions in struct module_memory
so that module_memory_free() could do the right thing depending on memory
state. But that meant either ugly games with const'ness in strict_rwx.c,
an additional helper or a new global module state. The latter seemed the
most elegant to me.
If a new global module state is really that intrusive, I can drop it in
favor a helper that will be called from error handling paths. E.g.
something like the patch below (on top of this series and with this patch
reverted)

diff --git a/kernel/module/main.c b/kernel/module/main.c
index 7164cd353a78..4a02503836d7 100644
--- a/kernel/module/main.c
+++ b/kernel/module/main.c
@@ -1268,13 +1268,20 @@ static int module_memory_alloc(struct module *mod, enum mod_mem_type type)
 	return 0;
 }
 
+static void module_memory_restore_rox(struct module *mod)
+{
+	for_class_mod_mem_type(type, text) {
+		struct module_memory *mem = &mod->mem[type];
+
+		if (mem->is_rox)
+			execmem_restore_rox(mem->base, mem->size);
+	}
+}
+
 static void module_memory_free(struct module *mod, enum mod_mem_type type)
 {
 	struct module_memory *mem = &mod->mem[type];
 
-	if (mod->state == MODULE_STATE_UNFORMED && mem->is_rox)
-		execmem_restore_rox(mem->base, mem->size);
-
 	execmem_free(mem->base);
 }
 
@@ -2617,6 +2624,7 @@ static int move_module(struct module *mod, struct load_info *info)
 
 	return 0;
 out_err:
+	module_memory_restore_rox(mod);
 	for (t--; t >= 0; t--)
 		module_memory_free(mod, t);
 	if (codetag_section_found)
@@ -3372,6 +3380,7 @@ static int load_module(struct load_info *info, const char __user *uargs,
 				       mod->mem[type].size);
 	}
 
+	module_memory_restore_rox(mod);
 	module_deallocate(mod, info);
  free_copy:
 	/*
 
> It is at least good that MODULE_STATE_GONE is only set in free_module()
> past the sysfs teardown, so it never shows in
> /sys/module/<mod>/initstate. Otherwise, this would require teaching kmod
> about this state as well.
> 
> -- 
> Thanks,
> Petr

-- 
Sincerely yours,
Mike.