From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mx2.mail.elte.hu (mx2.mail.elte.hu [157.181.151.9]) (using TLSv1 with cipher DHE-RSA-AES256-SHA (256/256 bits)) (Client did not present a certificate) by ozlabs.org (Postfix) with ESMTP id 0DBD9DE253 for ; Thu, 31 Jan 2008 07:59:34 +1100 (EST) Date: Wed, 30 Jan 2008 21:59:15 +0100 From: Ingo Molnar To: "Luck, Tony" Subject: Re: x86/non-x86: percpu, node ids, apic ids x86.git fixup Message-ID: <20080130205915.GA4562@elte.hu> References: <20080130161036.GA12293@elte.hu> <1FE6DD409037234FAB833C420AA843EC757C72@orsmsx424.amr.corp.intel.com> <20080130180623.GA24881@elte.hu> <1FE6DD409037234FAB833C420AA843EC757DEA@orsmsx424.amr.corp.intel.com> <47A0CD4B.5040706@sgi.com> <1FE6DD409037234FAB833C420AA843EC757F12@orsmsx424.amr.corp.intel.com> <20080130194616.GA19426@elte.hu> <1FE6DD409037234FAB833C420AA843EC757F7C@orsmsx424.amr.corp.intel.com> <20080130200221.GA28073@elte.hu> <1FE6DD409037234FAB833C420AA843EC757FC2@orsmsx424.amr.corp.intel.com> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii In-Reply-To: <1FE6DD409037234FAB833C420AA843EC757FC2@orsmsx424.amr.corp.intel.com> Cc: sparclinux@vger.kernel.org, linux-ia64@vger.kernel.org, Linux Kernel Development , Mike Travis , Linux/PPC Development , Geert Uytterhoeven , Thomas Gleixner , Linus Torvalds List-Id: Linux on PowerPC Developers Mail List List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , * Luck, Tony wrote: > > could you send the .config you are using? > > Ok. Attached. thanks a ton - this produced a link error here too. after half an hour of head scratching, the updated patch below solves the build problem. The problem i believe is this code in arch/ia64/kernel/mca_asm.S: #define GET_IA64_MCA_DATA(reg) \ GET_THIS_PADDR(reg, ia64_mca_data) \ ;; \ ld8 reg=[reg] this i believe builds an implicit dependency between the mca_asm.o position within the image and the ia64_mca_data percpu variable it accesses - it relies on the immediate 22 addressing mode that has 4MB of scope. Per chance, the .config you sent creates a 14MB image, and the percpu variables moved too far away for the linker to be able to fulfill this constraint. The workaround is to define PER_CPU_ATTRIBUTES to link percpu variables back into the .percpu section on UP too - which ia64 links specially into its vmlinux.lds. But ultimately i think the better solution would be to remove this dependency between arch/ia64/kernel/mca_asm.S and the position of the percpu data. Is my analysis correct? Do you like my fix and does the patch build and boot on your system? Thanks, Ingo ---------------> Subject: ia64: on UP percpu variables are not small memory model From: Ingo Molnar Tony says: | The CONFIG_SMP=n path in ia64 makes quite radical changes ... rather | than putting all the per-cpu stuff into the top 64K of address space | and providing a per-cpu TLB mapping for that range to a different | physical address ... it just makes all the per-cpu stuff link as ordinary | variables in .data. the new generic percpu code got confused about this as PER_CPU_ATTRIBUTES was defined even on UP, so it picked up that small memory model - which was not possible to get linked. The right fix is to only define that on SMP. This resolved the build failures in my cross-compiling environment. also link these variables into the .percpu section - some assembly code has offset dependencies. Signed-off-by: Ingo Molnar --- include/asm-ia64/percpu.h | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) Index: linux-x86.q/include/asm-ia64/percpu.h =================================================================== --- linux-x86.q.orig/include/asm-ia64/percpu.h +++ linux-x86.q/include/asm-ia64/percpu.h @@ -15,18 +15,20 @@ #include +#ifdef CONFIG_SMP + #ifdef HAVE_MODEL_SMALL_ATTRIBUTE # define PER_CPU_ATTRIBUTES __attribute__((__model__ (__small__))) #endif -#ifdef CONFIG_SMP - #define __my_cpu_offset __ia64_per_cpu_var(local_per_cpu_offset) extern void *per_cpu_init(void); #else /* ! SMP */ +#define PER_CPU_ATTRIBUTES __attribute__((__section__(".data.percpu"))) + #define per_cpu_init() (__phys_per_cpu_start) #endif /* SMP */