On Tue, Jan 28, 2014 at 11:40:57AM -0700, Bjorn Helgaas wrote: > On Mon, Jan 27, 2014 at 6:30 PM, David Cohen > <david.a.cohen@xxxxxxxxxxxxxxx> wrote: > > Hi Bjorn, > > > > On Mon, Jan 27, 2014 at 05:52:30PM -0700, Bjorn Helgaas wrote: > >> On Mon, Dec 16, 2013 at 1:07 PM, David Cohen > >> <david.a.cohen@xxxxxxxxxxxxxxx> wrote: > >> > This code was partially based on Mark Brown's previous work. > >> > > >> > Signed-off-by: David Cohen <david.a.cohen@xxxxxxxxxxxxxxx> > >> > Signed-off-by: Fei Yang <fei.yang@xxxxxxxxx> > >> > Cc: Mark F. Brown <mark.f.brown@xxxxxxxxx> > >> > Cc: Kuppuswamy Sathyanarayanan <sathyanarayanan.kuppuswamy@xxxxxxxxxxxxxxx> > >> > >> I know this has already been merged to Linus' tree, but it looks funny to me. > >> > >> > --- a/arch/x86/platform/intel-mid/intel_mid_weak_decls.h > >> > +++ b/arch/x86/platform/intel-mid/intel_mid_weak_decls.h > >> > @@ -16,3 +16,4 @@ > >> > */ > >> > extern void * __cpuinit get_penwell_ops(void) __attribute__((weak)); > >> > extern void * __cpuinit get_cloverview_ops(void) __attribute__((weak)); > >> > +extern void * __init get_tangier_ops(void) __attribute__((weak)); > >> > >> We should use "__weak" instead of the gcc-specific "__attribute__((weak))". > >> > >> I don't think it's a good idea to use __weak on a declaration in a > >> header file. If there are ever multiple definitions of the symbol, > >> they are *all* made weak symbols, and one is chosen based on link > >> order, which is error-prone. I only see one definition now, but the > >> whole point of weak is to allow multiple definitions, so this looks > >> like a problem waiting to happen. See 10629d711ed, for example. > >> > >> It look me a bit to figure out that these get_*_ops() functions are > >> used by INTEL_MID_OPS_INIT, which constructs the name using a macro, > >> so grep/cscope/etc. don't see any users. A comment pointing to > >> INTEL_MID_OPS_INIT would be helpful. > >> > >> What's the reason for making these symbols weak? Normally we use weak > >> to make a generic default version of a function, while allowing > >> architectures to replace the default with their own version if > >> necessary. But I don't see that happening here. Maybe I'm just > >> missing it, like I missed the uses of get_tangier_ops(), et al. > > > > Intel mid was implemented in such way that we should select which soc to > > be used in compilation time. Depending on the selection, mfld.c or > > mrfl.c could not be compiled then some symbols wouldn't be available. > > > > But IMHO this is a bad legacy design that exists in there, so I started > > to rework it as you can see in this commit: > > > > commit 4cb9b00f42e07830310319a07e6c91413ee8153e > > Author: David Cohen <david.a.cohen@xxxxxxxxxxxxxxx> > > Date: Mon Dec 16 17:37:26 2013 -0800 > > > > x86, intel-mid: Remove deprecated X86_MDFLD and X86_WANT_INTEL_MID > > configs > > > > I'm sending more patches soon and getting rid of intel_mid_weak_decls.h > > file is in my TODO list. > > Sounds good. While you're looking at it, I have similar questions > about ipc_device_handler() and msic_generic_platform_data(). It's not > clear to me why they should be weak. I'm afraid that's gargabe I missed. It supposed to be removed already. The original upstreamed patch set needed it, since all platform data were gathered in a board.c file and some of them could not be compiled. You can see it here: http://us.generation-nt.com/answer/patch-v2-00-10-rework-arch-x86-platform-mrst-intel-mid-help-212689892.html But I reworked this approach and added a sfi_device() macro to let compiler to gather all the platform data, thus board.c file doesn't exist. It means it is not necessary anymore to be weak. I can send a patch right away fixing it. Thanks for pointing that out. Br, David > > Bjorn -- To unsubscribe from this list: send the line "unsubscribe linux-pci" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html