On Wed, Apr 17, 2019 at 11:03:51AM +0100, Catalin Marinas wrote:
On Wed, Apr 17, 2019 at 10:41:37AM +0100, Russell King wrote:
On Sun, Apr 14, 2019 at 10:51:09PM -0700, Christoph Hellwig wrote:
On Sat, Apr 13, 2019 at 06:30:52PM +0200, Heinrich Schuchardt wrote:
This patch avoids ERROR: "__sync_icache_dcache" [drivers/xen/xen-privcmd.ko] undefined! observed when compiling v4.19.34.
The xen-privcmd driver, which can be modular, calls set_pte_at() which in turn may call __sync_icache_dcache().
Maybe that really is a sign that it should not be modular..
This issue has been discussed several times, and this URL gives a list of all messages on linux-arm-kernel that mention __sync_icache_dcache:
https://archive.armlinux.org.uk/lurker/search/20380101.000000.00000000@ml:li...
At the beginning of March, Boris Ostrovsky pointed that another solution is available that does not need architecture private symbols to be exported.
Since I've already said in a previous thread that I don't want this function exported, it seems the way forward is pretty obvious.
Arnd submitted a patch, not sure what happened to it:
https://lore.kernel.org/lkml/20190304204826.2414365-1-arnd@arndb.de/
Thanks Catalin, this is the "another solution" that I referred to above. (You'll find the exact same URL in Boris' email.)
I think this boils down to whether set_pte_at() is allowed to be called from loadable modules. If yes, __sync_icache_dcache() needs exporting, otherwise Arnd's patch makes the relevant code built-in (and, if merged, I'm happy to revert the similar export on arm64).
I think I've clearly expressed my opinion on that. Hence why "the way forward is pretty obvious." I don't understand why we still have the problem.