On Fri, Jun 29, 2018 at 05:00:51PM -0700, Fenghua Yu wrote:
> On Fri, Jun 29, 2018 at 11:44:44PM +0200, Thomas Gleixner wrote:
> > On Fri, 29 Jun 2018, Dave Hansen wrote:
> >
> > > On 06/29/2018 01:38 PM, Fenghua Yu wrote:
> > > > How to handle data that is used in generic code which can be used on
> > > > non-Intel platform? For exmple, if I do this change for struct efi in
> > > > include/linux/efi.h because set_bit() sets bits in efi.flags:
> > > > - unsigned long flags;
> > > > + unsigned long flags __aligned(unsigned long);
> > > > } efi;
> > > >
> > > > People may argue that the alignment unnecessarily increases size of
> > > > 'efi'
> > > > on non-Intel platform which doesn't have split lock issue. Do we care
> > > > this
> > > > argument?
> > >
> > > Unaligned memory accesses are bad, pretty much universally. This is a
> > > general good practice that we should have been doing anyway. Let folks
> > > complain. Don't let it stop you.
> > >
> > > Also, look at the size of that structure. Look at how many pointers it
> > > has. Do you think *anyone* is going to complain about an extra 4 bytes
> > > in a 400-byte structure?
> >
> > But in the above case the compiler does already the right thing. Why?
> > Because struct members are aligned to their natural alignment unless the
> > struct is explicitely marked 'packed'. In that case the programmer has to
> > take care of the alignment.
> >
> > Just look at it with pahole:
> >
> > struct efi_memory_map memmap; /* 280 56 */
> >
> > /* XXX last struct has 7 bytes of padding */
> >
> > /* --- cacheline 5 boundary (320 bytes) was 16 bytes ago --- */
> > long unsigned int flags; /* 336 8 */
> >
> > The issue with the capability arrays is that the data type is u32 which has
> > the natural alignment of 4 byte, while unsigned long has 8 byte on 64bit.
> >
> > So just slapping blindly aligned(unsigned long) to anything which is
> > accessed by locked instructions is pointless.
> >
>
> Thank you for you education!
>
> Below is part of the future patches that are supposed to fix more potential
> split lock issues.
>
> Could you please take a look and see if the changes are in the
> right direction before I move further?
>
> diff --git a/arch/x86/boot/cpuflags.h b/arch/x86/boot/cpuflags.h
Please ignore the patch in my last email because of some obvious stupid
mistakes. Sorry about that.
Instead, could you please take a look at the following patch and see if
the changes are in the right direction before I move further?
diff --git a/arch/x86/boot/cpuflags.h b/arch/x86/boot/cpuflags.h
index 2e20814d3ce3..29de0ff74351 100644
--- a/arch/x86/boot/cpuflags.h
+++ b/arch/x86/boot/cpuflags.h
@@ -9,7 +9,7 @@ struct cpu_features {
int level; /* Family, or 64 for x86-64 */
int family; /* Family, always */
int model;
- u32 flags[NCAPINTS];
+ u32 flags[NCAPINTS] __aligned(sizeof(unsigned long));
};
extern struct cpu_features cpu;
diff --git a/arch/x86/include/asm/mce.h b/arch/x86/include/asm/mce.h
index 8c7b3e5a2d01..444a2275c1f8 100644
--- a/arch/x86/include/asm/mce.h
+++ b/arch/x86/include/asm/mce.h
@@ -133,7 +133,7 @@ struct mce_log_buffer {
char signature[12]; /* "MACHINECHECK" */
unsigned len; /* = MCE_LOG_LEN */
unsigned next;
- unsigned flags;
+ unsigned flags __aligned(sizeof(unsigned long));
unsigned recordlen; /* length of struct mce */
struct mce entry[MCE_LOG_LEN];
};
diff --git a/arch/x86/kernel/cpu/common.c b/arch/x86/kernel/cpu/common.c
index eb4cb3efd20e..e6a28163e905 100644
--- a/arch/x86/kernel/cpu/common.c
+++ b/arch/x86/kernel/cpu/common.c
@@ -488,8 +488,8 @@ static const char *table_lookup_model(struct cpuinfo_x86 *c)
return NULL; /* Not found */
}
-__u32 cpu_caps_cleared[NCAPINTS + NBUGINTS];
-__u32 cpu_caps_set[NCAPINTS + NBUGINTS];
+__u32 cpu_caps_cleared[NCAPINTS + NBUGINTS] __aligned(sizeof(unsigned long));
+__u32 cpu_caps_set[NCAPINTS + NBUGINTS] __aligned(sizeof(unsigned long));
void load_percpu_segment(int cpu)
{
Thanks.
-Fenghua
