This patch makes UP and SMP do the same thing as far as module per-cpu
data go.
Unfortunately it affects core code.
To repeat the problem:
IA64 keeps per-cpu data in a small data area that is referenced by a
22-bit offset, for both UP and SMP cases. If a module defines
per-cpu data, it too will end up in the small-data area. But the
module loader at present special-cases the UP treatment of per-cpu
data, assumes that it is in the GP-relative data area, and does
nothing (for SMP it allocates space, and copies initialised data
items into it)
The effect is that modules defining per-cpu data fail to load if
they're built UP, because of an impossible relocation.
The appended patch makes the treatment of per-cpu data uniform
between UP and SMP cases. For most architectures, the per-cpu data
section will be empty for UP, and so the per-cpu setup code will not
be invoked.
Signed-off-by: Peter Chubb <[email protected]>
diff --git a/arch/ia64/kernel/module.c b/arch/ia64/kernel/module.c
--- a/arch/ia64/kernel/module.c
+++ b/arch/ia64/kernel/module.c
@@ -951,4 +951,10 @@ percpu_modcopy (void *pcpudst, const voi
if (cpu_possible(i))
memcpy(pcpudst + __per_cpu_offset[i], src, size);
}
+#else
+void
+percpu_modcopy (void *pcpudst, const void *src, unsigned long size)
+{
+ memcpy(pcpudst, src, size);
+}
#endif /* CONFIG_SMP */
diff --git a/kernel/module.c b/kernel/module.c
--- a/kernel/module.c
+++ b/kernel/module.c
@@ -209,7 +209,6 @@ static struct module *find_module(const
return NULL;
}
-#ifdef CONFIG_SMP
/* Number of blocks used and allocated. */
static unsigned int pcpu_num_used, pcpu_num_allocated;
/* Size of each block. -ve means used. */
@@ -352,29 +351,7 @@ static int percpu_modinit(void)
return 0;
}
__initcall(percpu_modinit);
-#else /* ... !CONFIG_SMP */
-static inline void *percpu_modalloc(unsigned long size, unsigned long align,
- const char *name)
-{
- return NULL;
-}
-static inline void percpu_modfree(void *pcpuptr)
-{
- BUG();
-}
-static inline unsigned int find_pcpusec(Elf_Ehdr *hdr,
- Elf_Shdr *sechdrs,
- const char *secstrings)
-{
- return 0;
-}
-static inline void percpu_modcopy(void *pcpudst, const void *src,
- unsigned long size)
-{
- /* pcpusec should be 0, and size of that section should be 0. */
- BUG_ON(size != 0);
-}
-#endif /* CONFIG_SMP */
+
#ifdef CONFIG_MODULE_UNLOAD
#define MODINFO_ATTR(field) \
On Thu, 2005-09-01 at 14:59 +1000, Peter Chubb wrote:
>
> This patch makes UP and SMP do the same thing as far as module per-cpu
> data go.
>
> Unfortunately it affects core code.
It causes 2.6.13/x86 to fail to link:
kernel/built-in.o: In function `load_module':
: undefined reference to `percpu_modcopy'
make: *** [.tmp_vmlinux1] Error 1
fyi, this is a problem we're seeing in the Debian UP packages:
http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=325070
On Thu, Sep 22, 2005 at 04:04:59PM -0600, dann frazier wrote:
> On Thu, 2005-09-01 at 14:59 +1000, Peter Chubb wrote:
> >
> > This patch makes UP and SMP do the same thing as far as module per-cpu
> > data go.
> >
> > Unfortunately it affects core code.
>
> It causes 2.6.13/x86 to fail to link:
> kernel/built-in.o: In function `load_module':
> : undefined reference to `percpu_modcopy'
> make: *** [.tmp_vmlinux1] Error 1
>
> fyi, this is a problem we're seeing in the Debian UP packages:
> http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=325070
Another possible solution is to make ia64 more like other
architectures and make per-cpu variables just turn into
ordinary variables on UP. There are some pros and cons to
this:
+) Being more like other architectures makes it less likely that
we'll be burned by changes in generic code/tools that depend
on implementation details
-) We probably get worse code to access per-cpu variables from
C-compiled code, and definitely get worse code in a couple of
critical paths in assembler (where an "addl" becomes a "movl")
Here's the patch ... lightly tested (just booted and checked that
I could load the ip_conntrack module).
-Tony
diff --git a/arch/ia64/kernel/entry.S b/arch/ia64/kernel/entry.S
index 0741b06..c678224 100644
--- a/arch/ia64/kernel/entry.S
+++ b/arch/ia64/kernel/entry.S
@@ -766,7 +766,11 @@ ENTRY(ia64_leave_syscall)
ld8.fill r15=[r3] // M0|1 restore r15
mov b6=r18 // I0 restore b6
+#ifdef CONFIG_SMP
addl r17=THIS_CPU(ia64_phys_stacked_size_p8),r0 // A
+#else
+ movl r17=THIS_CPU(ia64_phys_stacked_size_p8)
+#endif
mov f9=f0 // F clear f9
(pKStk) br.cond.dpnt.many skip_rbs_switch // B
@@ -952,7 +956,11 @@ GLOBAL_ENTRY(ia64_leave_kernel)
shr.u r18=r19,16 // get byte size of existing "dirty" partition
;;
mov r16=ar.bsp // get existing backing store pointer
+#ifdef CONFIG_SMP
addl r17=THIS_CPU(ia64_phys_stacked_size_p8),r0
+#else
+ movl r17=THIS_CPU(ia64_phys_stacked_size_p8)
+#endif
;;
ld4 r17=[r17] // r17 = cpu_data->phys_stacked_size_p8
(pKStk) br.cond.dpnt skip_rbs_switch
diff --git a/arch/ia64/kernel/head.S b/arch/ia64/kernel/head.S
index bfe65b2..4414970 100644
--- a/arch/ia64/kernel/head.S
+++ b/arch/ia64/kernel/head.S
@@ -980,7 +980,11 @@ END(ia64_delay_loop)
* intermediate precision so that we can produce a full 64-bit result.
*/
GLOBAL_ENTRY(sched_clock)
+#ifdef CONFIG_SMP
addl r8=THIS_CPU(cpu_info) + IA64_CPUINFO_NSEC_PER_CYC_OFFSET,r0
+#else
+ movl r8=THIS_CPU(cpu_info) + IA64_CPUINFO_NSEC_PER_CYC_OFFSET
+#endif
mov.m r9=ar.itc // fetch cycle-counter (35 cyc)
;;
ldf8 f8=[r8]
diff --git a/arch/ia64/kernel/mca.c b/arch/ia64/kernel/mca.c
index 355af15..b6391c9 100644
--- a/arch/ia64/kernel/mca.c
+++ b/arch/ia64/kernel/mca.c
@@ -1474,12 +1474,14 @@ ia64_mca_cpu_init(void *cpu_data)
*/
__get_cpu_var(ia64_mca_data) = __per_cpu_mca[smp_processor_id()];
+#ifdef CONFIG_SMP
/*
* Stash away a copy of the PTE needed to map the per-CPU page.
* We may need it during MCA recovery.
*/
__get_cpu_var(ia64_mca_per_cpu_pte) =
pte_val(mk_pte_phys(__pa(cpu_data), PAGE_KERNEL));
+#endif
/*
* Also, stash away a copy of the PAL address and the PTE
diff --git a/arch/ia64/kernel/mca_asm.S b/arch/ia64/kernel/mca_asm.S
index db32fc1..45d8b30 100644
--- a/arch/ia64/kernel/mca_asm.S
+++ b/arch/ia64/kernel/mca_asm.S
@@ -102,6 +102,7 @@ ia64_do_tlb_purge:
;;
srlz.d
;;
+#ifdef CONFIG_SMP
// 2. Purge DTR for PERCPU data.
movl r16=PERCPU_ADDR
mov r18=PERCPU_PAGE_SHIFT<<2
@@ -110,6 +111,7 @@ ia64_do_tlb_purge:
;;
srlz.d
;;
+#endif
// 3. Purge ITR for PAL code.
GET_THIS_PADDR(r2, ia64_mca_pal_base)
;;
@@ -197,6 +199,7 @@ ia64_reload_tr:
srlz.i
srlz.d
;;
+#ifdef CONFIG_SMP
// 2. Reload DTR register for PERCPU data.
GET_THIS_PADDR(r2, ia64_mca_per_cpu_pte)
;;
@@ -213,6 +216,7 @@ ia64_reload_tr:
;;
srlz.d
;;
+#endif
// 3. Reload ITR for PAL code.
GET_THIS_PADDR(r2, ia64_mca_pal_pte)
;;
diff --git a/arch/ia64/kernel/setup.c b/arch/ia64/kernel/setup.c
index 5add0bc..ef2dc6c 100644
--- a/arch/ia64/kernel/setup.c
+++ b/arch/ia64/kernel/setup.c
@@ -752,6 +752,7 @@ cpu_init (void)
struct cpuinfo_ia64 *cpu_info;
void *cpu_data;
+#ifdef CONFIG_SMP
cpu_data = per_cpu_init();
/*
@@ -761,9 +762,11 @@ cpu_init (void)
*/
ia64_set_kr(IA64_KR_PER_CPU_DATA,
ia64_tpa(cpu_data) - (long) __per_cpu_start);
+#endif
get_max_cacheline_size();
+#ifdef CONFIG_SMP
/*
* We can't pass "local_cpu_data" to identify_cpu() because we haven't called
* ia64_mmu_init() yet. And we can't call ia64_mmu_init() first because it
@@ -771,6 +774,11 @@ cpu_init (void)
* accessing cpu_data() through the canonical per-CPU address.
*/
cpu_info = cpu_data + ((char *) &__ia64_per_cpu_var(cpu_info) - __per_cpu_start);
+ cpu_data = ia64_imva(cpu_data);
+#else
+ cpu_info = &per_cpu(cpu_info, 0);
+ cpu_data = 0;
+#endif
identify_cpu(cpu_info);
#ifdef CONFIG_MCKINLEY
@@ -817,8 +825,8 @@ cpu_init (void)
if (current->mm)
BUG();
- ia64_mmu_init(ia64_imva(cpu_data));
- ia64_mca_cpu_init(ia64_imva(cpu_data));
+ ia64_mmu_init(cpu_data);
+ ia64_mca_cpu_init(cpu_data);
#ifdef CONFIG_IA32_SUPPORT
ia32_cpu_init();
diff --git a/arch/ia64/kernel/vmlinux.lds.S b/arch/ia64/kernel/vmlinux.lds.S
index 30d8564..2b3fa83 100644
--- a/arch/ia64/kernel/vmlinux.lds.S
+++ b/arch/ia64/kernel/vmlinux.lds.S
@@ -19,7 +19,9 @@ ENTRY(phys_start)
jiffies = jiffies_64;
PHDRS {
code PT_LOAD;
+#ifdef CONFIG_SMP
percpu PT_LOAD;
+#endif
data PT_LOAD;
}
SECTIONS
@@ -180,6 +182,7 @@ SECTIONS
.data.cacheline_aligned : AT(ADDR(.data.cacheline_aligned) - LOAD_OFFSET)
{ *(.data.cacheline_aligned) }
+#ifdef CONFIG_SMP
/* Per-cpu data: */
percpu : { } :percpu
. = ALIGN(PERCPU_PAGE_SIZE);
@@ -191,6 +194,7 @@ SECTIONS
__per_cpu_end = .;
}
. = __phys_per_cpu_start + PERCPU_PAGE_SIZE; /* ensure percpu data fits into percpu page size */
+#endif
data : { } :data
.data : AT(ADDR(.data) - LOAD_OFFSET)
diff --git a/arch/ia64/mm/init.c b/arch/ia64/mm/init.c
index e3215ba..5def017 100644
--- a/arch/ia64/mm/init.c
+++ b/arch/ia64/mm/init.c
@@ -332,7 +332,7 @@ setup_gate (void)
void __devinit
ia64_mmu_init (void *my_cpu_data)
{
- unsigned long psr, pta, impl_va_bits;
+ unsigned long pta, impl_va_bits;
extern void __devinit tlb_init (void);
#ifdef CONFIG_DISABLE_VHPT
@@ -341,6 +341,10 @@ ia64_mmu_init (void *my_cpu_data)
# define VHPT_ENABLE_BIT 1
#endif
+#ifdef CONFIG_SMP
+{
+ unsigned long psr;
+
/* Pin mapping for percpu area into TLB */
psr = ia64_clear_ic();
ia64_itr(0x2, IA64_TR_PERCPU_DATA, PERCPU_ADDR,
@@ -349,6 +353,8 @@ ia64_mmu_init (void *my_cpu_data)
ia64_set_psr(psr);
ia64_srlz_i();
+}
+#endif
/*
* Check if the virtually mapped linear page table (VMLPT) overlaps with a mapped
diff --git a/include/asm-ia64/mca_asm.h b/include/asm-ia64/mca_asm.h
index 27c9203..26d8710 100644
--- a/include/asm-ia64/mca_asm.h
+++ b/include/asm-ia64/mca_asm.h
@@ -48,9 +48,14 @@
mov temp = 0x7 ;; \
dep addr = temp, addr, 61, 3
+#ifdef CONFIG_SMP
#define GET_THIS_PADDR(reg, var) \
mov reg = IA64_KR(PER_CPU_DATA);; \
addl reg = THIS_CPU(var), reg
+#else
+#define GET_THIS_PADDR(reg, var) \
+ LOAD_PHYSICAL(p0, reg, per_cpu__##var)
+#endif
/*
* This macro jumps to the instruction at the given virtual address
diff --git a/include/asm-ia64/percpu.h b/include/asm-ia64/percpu.h
index 2b14dee..9248d17 100644
--- a/include/asm-ia64/percpu.h
+++ b/include/asm-ia64/percpu.h
@@ -16,7 +16,7 @@
#include <linux/threads.h>
-#ifdef HAVE_MODEL_SMALL_ATTRIBUTE
+#if defined(HAVE_MODEL_SMALL_ATTRIBUTE) && defined(CONFIG_SMP)
# define __SMALL_ADDR_AREA __attribute__((__model__ (__small__)))
#else
# define __SMALL_ADDR_AREA
@@ -25,6 +25,7 @@
#define DECLARE_PER_CPU(type, name) \
extern __SMALL_ADDR_AREA __typeof__(type) per_cpu__##name
+#ifdef CONFIG_SMP
/* Separate out the type, so (int[3], foo) works. */
#define DEFINE_PER_CPU(type, name) \
__attribute__((__section__(".data.percpu"))) \
@@ -34,7 +35,6 @@
* Pretty much a literal copy of asm-generic/percpu.h, except that percpu_modcopy() is an
* external routine, to avoid include-hell.
*/
-#ifdef CONFIG_SMP
extern unsigned long __per_cpu_offset[NR_CPUS];
@@ -50,6 +50,9 @@ extern void *per_cpu_init(void);
#else /* ! SMP */
+#define DEFINE_PER_CPU(type, name) \
+ __typeof__(type) per_cpu__##name
+
#define per_cpu(var, cpu) (*((void)(cpu), &per_cpu__##var))
#define __get_cpu_var(var) per_cpu__##var
#define per_cpu_init() (__phys_per_cpu_start)
On Mon, 2005-12-19 at 13:07 -0800, Luck, Tony wrote:
> On Thu, Sep 22, 2005 at 04:04:59PM -0600, dann frazier wrote:
> > On Thu, 2005-09-01 at 14:59 +1000, Peter Chubb wrote:
> > >
> > > This patch makes UP and SMP do the same thing as far as module per-cpu
> > > data go.
> > >
> > > Unfortunately it affects core code.
> >
> > It causes 2.6.13/x86 to fail to link:
> > kernel/built-in.o: In function `load_module':
> > : undefined reference to `percpu_modcopy'
> > make: *** [.tmp_vmlinux1] Error 1
> >
> > fyi, this is a problem we're seeing in the Debian UP packages:
> > http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=325070
>
> Another possible solution is to make ia64 more like other
> architectures and make per-cpu variables just turn into
> ordinary variables on UP. There are some pros and cons to
> this:
>
> +) Being more like other architectures makes it less likely that
> we'll be burned by changes in generic code/tools that depend
> on implementation details
>
> -) We probably get worse code to access per-cpu variables from
> C-compiled code, and definitely get worse code in a couple of
> critical paths in assembler (where an "addl" becomes a "movl")
>
> Here's the patch ... lightly tested (just booted and checked that
> I could load the ip_conntrack module).
Thanks Tony; sorry for taking so long to test this. I required an
additional change to discontig.c to get this to build w/ the Debian
config. With this additional patch, a UP kernel boots fine on my
rx2600.
--- build-ia64-none-mckinley/arch/ia64/mm/discontig.c~ 2006-01-02 20:21:10.000000000 -0700
+++ build-ia64-none-mckinley/arch/ia64/mm/discontig.c 2006-01-09 19:56:58.000000000 -0700
@@ -339,8 +339,7 @@
struct cpuinfo_ia64 *cpu0_cpu_info;
cpu = 0;
node = node_cpuid[cpu].nid;
- cpu0_cpu_info = (struct cpuinfo_ia64 *)(__phys_per_cpu_start +
- ((char *)&per_cpu__cpu_info - __per_cpu_start));
+ cpu0_cpu_info = &per_cpu(cpu_info, 0);
cpu0_cpu_info->node_data = mem_data[node].node_data;
}
#endif /* CONFIG_SMP */
On Tue, 2006-01-10 at 14:21 -0700, dann frazier wrote:
> On Mon, 2005-12-19 at 13:07 -0800, Luck, Tony wrote:
> > On Thu, Sep 22, 2005 at 04:04:59PM -0600, dann frazier wrote:
> > > On Thu, 2005-09-01 at 14:59 +1000, Peter Chubb wrote:
> > > >
> > > > This patch makes UP and SMP do the same thing as far as module per-cpu
> > > > data go.
> > > >
> > > > Unfortunately it affects core code.
> > >
> > > It causes 2.6.13/x86 to fail to link:
> > > kernel/built-in.o: In function `load_module':
> > > : undefined reference to `percpu_modcopy'
> > > make: *** [.tmp_vmlinux1] Error 1
> > >
> > > fyi, this is a problem we're seeing in the Debian UP packages:
> > > http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=325070
> >
> > Another possible solution is to make ia64 more like other
> > architectures and make per-cpu variables just turn into
> > ordinary variables on UP. There are some pros and cons to
> > this:
> >
> > +) Being more like other architectures makes it less likely that
> > we'll be burned by changes in generic code/tools that depend
> > on implementation details
> >
> > -) We probably get worse code to access per-cpu variables from
> > C-compiled code, and definitely get worse code in a couple of
> > critical paths in assembler (where an "addl" becomes a "movl")
> >
> > Here's the patch ... lightly tested (just booted and checked that
> > I could load the ip_conntrack module).
>
> Thanks Tony; sorry for taking so long to test this. I required an
> additional change to discontig.c to get this to build w/ the Debian
> config. With this additional patch, a UP kernel boots fine on my
> rx2600.
fyi, just noticed that the efivars module still doesn't load.