Now that kksymoops is in this can be safely merged:
In some cases there is a benchmarkable difference (in the several percent
range) between a compiled in driver and a modular driver. The reason is
the compiled in driver is mapped in 2/4MB pages in the direct mapping,
while modules are mapped wit 4K pages in the vmalloc area. The resulting
TLB misses can make that much difference.
This patch fixes it by just putting the modules into the linear mapping
if possible. This needs unfragmented memory, but modules tend to be loaded
at bootup when there is still plenty of this. When it fails it just falls
back to vmalloc.
It has one drawback in that it makes the load address of modules much more
random than they were previously: you have a bigger chance that
the module is loaded somewhere else after reboot and then ksymoops wouldn't
be able to decode an module oops correctly anymore. With kksymoops this
shouldn't be a problem anymore, so it can be safely merged.
To avoid too much memory wastage it adds a new interface alloc_exact/free_exact
that allocates pages from the page allocator but doesn't round the area size
to the next power of two.
Variants of this patch have run successfully for several years in the -aa
kernels and in the SuSE kernels.
Patch for 2.5.38. Please apply.
-Andi
--- linux-2.5.38-work/include/linux/gfp.h-MODULE 2002-09-25 00:59:27.000000000 +0200
+++ linux-2.5.38-work/include/linux/gfp.h 2002-09-27 15:30:59.000000000 +0200
@@ -85,4 +85,7 @@
#define __free_page(page) __free_pages((page), 0)
#define free_page(addr) free_pages((addr),0)
+extern void *alloc_exact(unsigned int size);
+extern void free_exact(void * addr, unsigned int size);
+
#endif /* __LINUX_GFP_H */
--- linux-2.5.38-work/include/asm-i386/module.h-MODULE 2002-09-25 00:59:28.000000000 +0200
+++ linux-2.5.38-work/include/asm-i386/module.h 2002-09-27 15:30:35.000000000 +0200
@@ -3,9 +3,31 @@
/*
* This file contains the i386 architecture specific module code.
*/
+#include <linux/vmalloc.h>
+#include <linux/highmem.h>
+#include <asm/page.h>
+
+/*
+ * Try to allocate in the linear large page mapping first to conserve
+ * TLB entries.
+ */
+static inline void *module_map(unsigned long size)
+{
+ void *p = alloc_exact(size);
+ if (!p)
+ p = vmalloc(size);
+ return p;
+}
+
+static inline void module_unmap (struct module *mod)
+{
+ unsigned long mptr = (unsigned long)mod;
+ if (mptr >= VMALLOC_START && mptr+mod->size <= VMALLOC_END)
+ vfree(mod);
+ else
+ free_exact(mod, mod->size);
+}
-#define module_map(x) vmalloc(x)
-#define module_unmap(x) vfree(x)
#define module_arch_init(x) (0)
#define arch_init_modules(x) do { } while (0)
--- linux-2.5.38-work/mm/page_alloc.c-MODULE 2002-09-25 00:59:29.000000000 +0200
+++ linux-2.5.38-work/mm/page_alloc.c 2002-09-27 15:34:59.000000000 +0200
@@ -464,6 +464,38 @@
__free_pages(virt_to_page(addr), order);
}
}
+
+void *alloc_exact(unsigned int size)
+{
+ struct page *p, *w;
+ int order = get_order(size);
+
+ p = alloc_pages(GFP_KERNEL, order);
+ if (p) {
+ struct page *end = p + (1UL << order);
+ for (w = p+1; w < end; ++w)
+ set_page_count(w, 1);
+ for (w = p + (size>>PAGE_SHIFT)+1; w < end; ++w)
+ __free_pages(w, 0);
+ return (void *) page_address(p);
+ }
+
+ return NULL;
+}
+
+void free_exact(void * addr, unsigned int size)
+{
+ struct page * w;
+ unsigned long mptr = (unsigned long) addr;
+ int sz;
+
+ w = virt_to_page(addr);
+ for (sz = size; sz > 0; sz -= PAGE_SIZE, ++w) {
+ if (atomic_read(&w->count) != 1)
+ BUG();
+ __free_pages(w, 0);
+ }
+}
/*
* Total amount of free (allocatable) RAM:
--- linux-2.5.38-work/kernel/ksyms.c-MODULE 2002-09-25 00:59:29.000000000 +0200
+++ linux-2.5.38-work/kernel/ksyms.c 2002-09-27 15:51:47.000000000 +0200
@@ -97,6 +97,8 @@
EXPORT_SYMBOL(get_zeroed_page);
EXPORT_SYMBOL(__page_cache_release);
EXPORT_SYMBOL(__free_pages);
+EXPORT_SYMBOL(alloc_exact);
+EXPORT_SYMBOL(free_exact);
EXPORT_SYMBOL(free_pages);
EXPORT_SYMBOL(num_physpages);
EXPORT_SYMBOL(kmem_find_general_cachep);
Hi,
On Fri, 27 Sep 2002, Andi Kleen wrote:
> --- linux-2.5.38-work/include/asm-i386/module.h-MODULE 2002-09-25 00:59:28.000000000 +0200
> +++ linux-2.5.38-work/include/asm-i386/module.h 2002-09-27 15:30:35.000000000 +0200
> +/*
> + * Try to allocate in the linear large page mapping first to conserve
> + * TLB entries.
> + */
> +static inline void *module_map(unsigned long size)
> +{
> + void *p = alloc_exact(size);
> + if (!p)
> + p = vmalloc(size);
> + return p;
> +}
Why is i386 only? This is generic code and other archs will benefit from
it as well (or at least it won't hurt).
$ grep module_map include/asm-*/*.h
include/asm-alpha/module.h:#define module_map(x) vmalloc(x)
include/asm-arm/module.h:#define module_map(x) vmalloc(x)
include/asm-cris/module.h:#define module_map(x) vmalloc(x)
include/asm-i386/module.h:#define module_map(x) vmalloc(x)
include/asm-ia64/module.h:#define module_map(x) vmalloc(x)
include/asm-m68k/module.h:#define module_map(x) vmalloc(x)
include/asm-mips/module.h:#define module_map(x) vmalloc(x)
include/asm-mips64/module.h:#define module_map(x) vmalloc(x)
include/asm-parisc/pgtable.h:#define module_map vmalloc
include/asm-ppc/module.h:#define module_map(x) vmalloc(x)
include/asm-ppc64/module.h:#define module_map(x) vmalloc(x)
include/asm-s390/module.h:#define module_map(x) vmalloc(x)
include/asm-s390x/module.h:#define module_map(x) vmalloc(x)
include/asm-sh/module.h:#define module_map(x) vmalloc(x)
include/asm-sparc/module.h:#define module_map(x) vmalloc(x)
include/asm-sparc64/module.h:extern void * module_map (unsigned long size);
include/asm-x86_64/module.h:extern void *module_map(unsigned long);
The x86_64 version is currently broken, as it uses vmalloc_area_pages
which doesn't exist anymore, but it did the same as vmalloc. The sparc64
version also looks like vmalloc reimplementation.
> +
> +void *alloc_exact(unsigned int size)
> +{
> + struct page *p, *w;
> + int order = get_order(size);
> +
> + p = alloc_pages(GFP_KERNEL, order);
Wouldn't it be better to add a gfp argument?
bye, Roman
On Fri, Sep 27, 2002 at 04:09:30PM +0200, Andi Kleen wrote:
> + unsigned long mptr = (unsigned long)mod;
> + if (mptr >= VMALLOC_START && mptr+mod->size <= VMALLOC_END)
Any chance you could add is is_vmalloc() macro to vmalloc.h for this?
Roman Zippel <[email protected]> writes:
> Why is i386 only? This is generic code and other archs will benefit from
> it as well (or at least it won't hurt).
Because some arcitectures have a different module_map() (e.g. x86-64 or
sparc64) and because the VMALLOC_START/END trick doesn't work on all.
> > +
> > +void *alloc_exact(unsigned int size)
> > +{
> > + struct page *p, *w;
> > + int order = get_order(size);
> > +
> > + p = alloc_pages(GFP_KERNEL, order);
>
> Wouldn't it be better to add a gfp argument?
I don't see a need for it. GFP_ATOMIC doesn't make sense for > order 0,
and > order 0 is the only case that is interesting for alloc_exact.
GFP_DMA is not needed here, and GFP_HIGHUSER neither supports > order 0
properly (because of kmap)
-Andi
Hi,
On 27 Sep 2002, Andi Kleen wrote:
> > Why is i386 only? This is generic code and other archs will benefit from
> > it as well (or at least it won't hurt).
>
> Because some arcitectures have a different module_map() (e.g. x86-64 or
> sparc64)
As I already said in the last mail, these functions look like vmalloc
reimplementations.
> and because the VMALLOC_START/END trick doesn't work on all.
Where doesn't it work? vmalloc wouldn't work there either.
> > > +void *alloc_exact(unsigned int size)
> >
> > Wouldn't it be better to add a gfp argument?
>
> I don't see a need for it. GFP_ATOMIC doesn't make sense for > order 0,
> and > order 0 is the only case that is interesting for alloc_exact.
> GFP_DMA is not needed here, and GFP_HIGHUSER neither supports > order 0
> properly (because of kmap)
If it's supposed to be a generic function, it makes sense, otherwise we
could just put it into module.c.
bye, Roman
On Fri, Sep 27, 2002 at 05:55:13PM +0200, Roman Zippel wrote:
> Hi,
>
> On 27 Sep 2002, Andi Kleen wrote:
>
> > > Why is i386 only? This is generic code and other archs will benefit from
> > > it as well (or at least it won't hurt).
> >
> > Because some arcitectures have a different module_map() (e.g. x86-64 or
> > sparc64)
>
> As I already said in the last mail, these functions look like vmalloc
> reimplementations.
No, they aren't. x86-64 uses it because modules need to be in 32bit
range from the main kernel and vmalloc is in a different area. I suspect
it is the same on sparc64.
> > and > order 0 is the only case that is interesting for alloc_exact.
> > GFP_DMA is not needed here, and GFP_HIGHUSER neither supports > order 0
> > properly (because of kmap)
>
> If it's supposed to be a generic function, it makes sense, otherwise we
> could just put it into module.c.
Ok, I will change it.
-Andi
Hi,
On Fri, 27 Sep 2002, Andi Kleen wrote:
> No, they aren't. x86-64 uses it because modules need to be in 32bit
> range from the main kernel and vmalloc is in a different area. I suspect
> it is the same on sparc64.
Ok, makes sense.
> > If it's supposed to be a generic function, it makes sense, otherwise we
> > could just put it into module.c.
>
> Ok, I will change it.
Any chance to use __HAVE_MODULE_MAP, so every arch (except sparc64/x86-64)
can automatically benefit from this?
bye, Roman
On Fri, Sep 27, 2002 at 06:52:28PM +0200, Roman Zippel wrote:
> > > If it's supposed to be a generic function, it makes sense, otherwise we
> > > could just put it into module.c.
> >
> > Ok, I will change it.
>
> Any chance to use __HAVE_MODULE_MAP, so every arch (except sparc64/x86-64)
> can automatically benefit from this?
I can put it in asm-generic/
But people have to check themselves if the vmalloc trick works for them.
-Andi
Hi,
On Fri, 27 Sep 2002, Andi Kleen wrote:
> > Any chance to use __HAVE_MODULE_MAP, so every arch (except sparc64/x86-64)
> > can automatically benefit from this?
>
> I can put it in asm-generic/
>
> But people have to check themselves if the vmalloc trick works for them.
If it's in module.c anyway, you could set a flag in mod->flags, at how
it was allocated.
bye, Roman
On Fri, 27 Sep 2002 16:09:30 +0200,
Andi Kleen <[email protected]> wrote:
>It has one drawback in that it makes the load address of modules much more
>random than they were previously: you have a bigger chance that
>the module is loaded somewhere else after reboot and then ksymoops wouldn't
>be able to decode an module oops correctly anymore. With kksymoops this
>shouldn't be a problem anymore, so it can be safely merged.
man insmod. See 'ksymoops assistance'. mkdir /var/log/ksymoops.