Skip to contentSkip to conversation

modules: Use a better scheme for refcounting

PATCH In review Eric Dumazet 5 participants 13 messages May 15, 2008 – May 19, 2008 Module support

lore
Share & tools
lore

Eric Dumazet

author
#0 lore raw

PATCH 1/1 modules: Use a better scheme for refcounting

Current refcounting for modules (done if CONFIG_MODULE_UNLOAD=y) is using a lot of memory.

Each 'struct module' contains an [NR_CPUS] array of full cache lines.

This patch uses existing infrastructure (percpu_modalloc() & percpu_modfree()) to allocate percpu space for the refcount storage.

Instead of wasting NR_CPUS*128 bytes (on i386), we now use num_possible_cpus*sizeof(local_t) bytes.

On a typical distro, where NR_CPUS=8, shiping 2000 modules, we reduce size of module files by about 2 Mbytes. (1Kb per module)

Instead of having all refcounters in the same memory node - with TLB misses because of vmalloc() - this new implementation permits to have better NUMA properties, since each CPU will use storage on its preferred node, thanks to percpu storage.

Signed-off-by Eric Dumazet <[email protected]>

---

include/linux/module.h |   25 ++++++++++++++++---------
kernel/module.c        |   37 +++++++++++++++++++++++++++----------
2 files changed, 43 insertions(+), 19 deletions(-)
include/linux/module.h +16 -9
index 3e03b1a..72db614 100644
@@ -217,11 +217,6 @@ void *__symbol_get_gpl(const char *symbol);
 
 #endif
 
-struct module_ref
-{
-	local_t count;
-} ____cacheline_aligned;
-
 enum module_state
 {
 	MODULE_STATE_LIVE,

@@ -307,8 +302,11 @@ struct module
 
 #ifdef CONFIG_MODULE_UNLOAD
 	/* Reference counts */
-	struct module_ref ref[NR_CPUS];
-
+#ifdef CONFIG_SMP
+	char *refptr;
+#else
+	local_t ref;
+#endif
 	/* What modules depend on me? */
 	struct list_head modules_which_use_me;
 

@@ -378,13 +376,22 @@ void __symbol_put(const char *symbol);
 #define symbol_put(x) __symbol_put(MODULE_SYMBOL_PREFIX #x)
 void symbol_put_addr(void *addr);
 
+static inline local_t *__module_ref_addr(struct module *mod, int cpu)
+{
+#ifdef CONFIG_SMP
+	return (local_t *) (mod->refptr + per_cpu_offset(cpu));
+#else
+	return &mod->ref;
+#endif
+}
+
 /* Sometimes we know we already have a refcount, and it's easier not
    to handle the error case (which only happens with rmmod --wait). */
 static inline void __module_get(struct module *module)
 {
 	if (module) {
 		BUG_ON(module_refcount(module) == 0);
-		local_inc(&module->ref[get_cpu()].count);
+		local_inc(__module_ref_addr(module, get_cpu()));
 		put_cpu();
 	}
 }

@@ -396,7 +403,7 @@ static inline int try_module_get(struct module *module)
 	if (module) {
 		unsigned int cpu = get_cpu();
 		if (likely(module_is_live(module)))
-			local_inc(&module->ref[cpu].count);
+			local_inc(__module_ref_addr(module, cpu));
 		else
 			ret = 0;
 		put_cpu();
kernel/module.c +27 -10
index f5e9491..e0e8712 100644
@@ -522,13 +522,13 @@ static char last_unloaded_module[MODULE_NAME_LEN+1];
 /* Init the unload section of the module. */
 static void module_unload_init(struct module *mod)
 {
-	unsigned int i;
+	int cpu;
 
 	INIT_LIST_HEAD(&mod->modules_which_use_me);
-	for (i = 0; i < NR_CPUS; i++)
-		local_set(&mod->ref[i].count, 0);
+	for_each_possible_cpu(cpu)
+		local_set(__module_ref_addr(mod, cpu), 0);
 	/* Hold reference count during initialization. */
-	local_set(&mod->ref[raw_smp_processor_id()].count, 1);
+	local_set(__module_ref_addr(mod, raw_smp_processor_id()), 1);
 	/* Backwards compatibility macros put refcount during init. */
 	mod->waiter = current;
 }

@@ -659,10 +659,11 @@ static int try_stop_module(struct module *mod, int flags, int *forced)
 
 unsigned int module_refcount(struct module *mod)
 {
-	unsigned int i, total = 0;
+	unsigned int total = 0;
+	int cpu;
 
-	for (i = 0; i < NR_CPUS; i++)
-		total += local_read(&mod->ref[i].count);
+	for_each_possible_cpu(cpu)
+		total += local_read(__module_ref_addr(mod, cpu));
 	return total;
 }
 EXPORT_SYMBOL(module_refcount);

@@ -824,7 +825,7 @@ void module_put(struct module *module)
 {
 	if (module) {
 		unsigned int cpu = get_cpu();
-		local_dec(&module->ref[cpu].count);
+		local_dec(__module_ref_addr(module, cpu));
 		/* Maybe they're waiting for us to drop reference? */
 		if (unlikely(!module_is_live(module)))
 			wake_up_process(module->waiter);

@@ -1392,7 +1393,10 @@ static void free_module(struct module *mod)
 	kfree(mod->args);
 	if (mod->percpu)
 		percpu_modfree(mod->percpu);
-
+#if defined(CONFIG_MODULE_UNLOAD) && defined(CONFIG_SMP)
+	if (mod->refptr)
+		percpu_modfree(mod->refptr);
+#endif
 	/* Free lock-classes: */
 	lockdep_free_key_range(mod->module_core, mod->core_size);
 

@@ -1760,6 +1764,9 @@ static struct module *load_module(void __user *umod,
 	unsigned int markersstringsindex;
 	struct module *mod;
 	long err = 0;
+#if defined(CONFIG_MODULE_UNLOAD) && defined(CONFIG_SMP)
+	void *refptr = NULL;
+#endif
 	void *percpu = NULL, *ptr = NULL; /* Stops spurious gcc warning */
 	struct exception_table_entry *extable;
 	mm_segment_t old_fs;

@@ -1904,6 +1911,12 @@ static struct module *load_module(void __user *umod,
 	if (err < 0)
 		goto free_mod;
 
+#if defined(CONFIG_MODULE_UNLOAD) && defined(CONFIG_SMP)
+	refptr = percpu_modalloc(sizeof(local_t), sizeof(local_t), mod->name);
+	if (!refptr)
+		goto free_mod;
+	mod->refptr = refptr;
+#endif
 	if (pcpuindex) {
 		/* We have a special allocation for this section. */
 		percpu = percpu_modalloc(sechdrs[pcpuindex].sh_size,

@@ -1911,7 +1924,7 @@ static struct module *load_module(void __user *umod,
 					 mod->name);
 		if (!percpu) {
 			err = -ENOMEM;
-			goto free_mod;
+			goto free_percpu;
 		}
 		sechdrs[pcpuindex].sh_flags &= ~(unsigned long)SHF_ALLOC;
 		mod->percpu = percpu;

@@ -2164,6 +2177,10 @@ static struct module *load_module(void __user *umod,
  free_percpu:
 	if (percpu)
 		percpu_modfree(percpu);
+#if defined(CONFIG_MODULE_UNLOAD) && defined(CONFIG_SMP)
+	if (refptr)
+		percpu_modfree(refptr);
+#endif
  free_mod:
 	kfree(args);
  free_hdr:
#1 lore raw
Quoted 4 lines · Eric Dumazet

Instead of wasting NR_CPUS*128 bytes (on i386), we now use num_possible_cpus*sizeof(local_t) bytes.

On a typical distro, where NR_CPUS=8,

More typical would be NR_CPUS=128, with NR_CPUS=four digits when Mike Travis et.al. are finished

Quoted 1 line

shiping 2000 modules, we reduce

Surely only the loaded modules count? Perhaps 20-30.

But it's a cool improvement. Very nice.

-Andi

author
#2 lore raw
Quoted 11 lines · Andi Kleen
Quoted 5 lines · Eric Dumazet

Instead of wasting NR_CPUS*128 bytes (on i386), we now use num_possible_cpus*sizeof(local_t) bytes.

On a typical distro, where NR_CPUS=8,

More typical would be NR_CPUS=128, with NR_CPUS=four digits when Mike Travis et.al. are finished

Yes. With NR_CPUS=4096, we save about half a megabyte per module. (!!!)

Quoted 6 lines
Quoted 2 lines

shiping 2000 modules, we reduce

Surely only the loaded modules count? Perhaps 20-30.

Well, I should have stated that this saving also takes place in the module disk file. (The "struct module" is included in it, in the ".gnu.linkonce.this_module" section) Here on my distro, around 2000 modules are shiped.

Quoted 2 lines

But it's a cool improvement. Very nice.

Thanks

#3 lore raw
In reply to Eric Dumazet
Quoted 2 lines · On Friday 16 May 2008 06:40:37 Eric Dumazet wrote:

Current refcounting for modules (done if CONFIG_MODULE_UNLOAD=y) is using a lot of memory.

Hi Eric,

I like this patch! The plan was always to create a proper dynamic per-cpu
allocator which used the normal per-cpu offsets, but I think module refcounts are worthwhile as a special case.

Any chance I can ask you look at the issue of full dynamic per-cpu
allocation? The problem of allocating memory which is laid out precisely as the original per-cpu alloc is vexing on NUMA, and probably requires reserving virtual address space and remapping into it, but the rewards would be maximally-efficient per-cpu accessors, and getting rid of that boutique allocator in module.c.

Only minor commentry on the patch itself:

include/linux/module.h struct module · line 307

Quoted 3 lines
include/linux/module.h +3 -0
+#ifdef CONFIG_SMP
+       char *refptr;
+#else

void * would seem more natural here (love those gcc'isms)

kernel/module.c static struct module *load_module(void __user *umod, · line 1769

Quoted 3 lines
kernel/module.c +3 -0
+#if defined(CONFIG_MODULE_UNLOAD) && defined(CONFIG_SMP)
+       void *refptr = NULL;
+#endif

Looks like you can skip this if you assign to mod->refptr directly below:

kernel/module.c static struct module *load_module(void __user *umod, · line 1919

Quoted 6 lines
kernel/module.c +6 -0
+#if defined(CONFIG_MODULE_UNLOAD) && defined(CONFIG_SMP)
+       refptr = percpu_modalloc(sizeof(local_t), sizeof(local_t), mod->name);
+       if (!refptr)
+               goto free_mod;
+       mod->refptr = refptr;
+#endif

And finally:

kernel/module.c static struct module *load_module(void __user *umod, · line 2183

Quoted 4 lines
kernel/module.c +4 -0
+#if defined(CONFIG_MODULE_UNLOAD) && defined(CONFIG_SMP)
+       if (refptr)
+               percpu_modfree(refptr);
+#endif

This if (refptr) seems redundant.

Thanks!
Rusty.

author
#4 lore raw
Quoted 12 lines · Rusty Russell
Quoted 3 lines · On Friday 16 May 2008 06:40:37 Eric Dumazet wrote:

Current refcounting for modules (done if CONFIG_MODULE_UNLOAD=y) is using a lot of memory.

Hi Eric,

I like this patch! The plan was always to create a proper dynamic per-cpu
allocator which used the normal per-cpu offsets, but I think module refcounts are worthwhile as a special case.

Quoted 8 lines

Any chance I can ask you look at the issue of full dynamic per-cpu
allocation? The problem of allocating memory which is laid out precisely as the original per-cpu alloc is vexing on NUMA, and probably requires reserving virtual address space and remapping into it, but the rewards would be maximally-efficient per-cpu accessors, and getting rid of that boutique allocator in module.c.

You mean using alloc_percpu() ? Problem is that current implementation is expensive, since it is using an extra array of pointers (struct percpu_data). On x86_64, that means at least a 200% space increase over the solution of using 4 bytes in the static percpu zone. We probably can change this to dynamic per-cpu as soon as Mike or Christopher finish their work on new dynamic per-cpu implementation ?

Quoted 11 lines

Only minor commentry on the patch itself:

Review

Quoted 4 lines
(quoted diff) +3 -0
+#ifdef CONFIG_SMP
+       char *refptr;
+#else

void * would seem more natural here (love those gcc'isms)

Yes, sure.

Quoted 8 lines

Review

Quoted 4 lines
(quoted diff) +3 -0
+#if defined(CONFIG_MODULE_UNLOAD) && defined(CONFIG_SMP)
+       void *refptr = NULL;
+#endif

Looks like you can skip this if you assign to mod->refptr directly below:

Quoted 9 lines

Review

Quoted 7 lines
(quoted diff) +6 -0
+#if defined(CONFIG_MODULE_UNLOAD) && defined(CONFIG_SMP)
+       refptr = percpu_modalloc(sizeof(local_t), sizeof(local_t), mod->name);
+       if (!refptr)
+               goto free_mod;
+       mod->refptr = refptr;
+#endif

Unfortunatly no. I tried it and failed.
This is the problem that at freeing time, we cannot anymore use mod->refptr, since we just unmaped mod, using vfree().
You had same problem with percpu, and had to use a temporaty variable on stack :)

Quoted 12 lines

And finally:

Review

Quoted 5 lines
(quoted diff) +4 -0
+#if defined(CONFIG_MODULE_UNLOAD) && defined(CONFIG_SMP)
+       if (refptr)
+               percpu_modfree(refptr);
+#endif

This if (refptr) seems redundant.

Yes, thanks a lot. Please find an updated patch.

Quoted 5 lines

Thanks!
Rusty.

[PATCH] modules: Use a better scheme for refcounting

Current refcounting for modules (done if CONFIG_MODULE_UNLOAD=y) is using a lot of memory (and disk space)

Each 'struct module' contains an [NR_CPUS] array of full cache lines.

This patch uses existing infrastructure (percpu_modalloc() & percpu_modfree()) to allocate percpu space for the refcount storage.

Instead of wasting NR_CPUS*128 bytes (on i386), we now use num_possible_cpus*sizeof(local_t) bytes.

On a typical distro, where NR_CPUS=8, shiping 2000 modules, we reduce size of module files by about 2 Mbytes. (1Kb per module)

Instead of having all refcounters in the same memory node - with TLB misses because of vmalloc() - this new implementation permits to have better NUMA properties, since each CPU will use storage on its preferred node, thanks to percpu storage.

Signed-off-by Eric Dumazet <[email protected]>

---

include/linux/module.h |   25 +++++++++++++++---------
kernel/module.c        |   40 +++++++++++++++++++++++++++++----------
2 files changed, 46 insertions(+), 19 deletions(-)
include/linux/module.h +16 -9
index 3e03b1a..dfd36ed 100644
@@ -217,11 +217,6 @@ void *__symbol_get_gpl(const char *symbol);
 
 #endif
 
-struct module_ref
-{
-	local_t count;
-} ____cacheline_aligned;
-
 enum module_state
 {
 	MODULE_STATE_LIVE,

@@ -307,8 +302,11 @@ struct module
 
 #ifdef CONFIG_MODULE_UNLOAD
 	/* Reference counts */
-	struct module_ref ref[NR_CPUS];
-
+#ifdef CONFIG_SMP
+	void *refptr;
+#else
+	local_t ref;
+#endif
 	/* What modules depend on me? */
 	struct list_head modules_which_use_me;
 

@@ -378,13 +376,22 @@ void __symbol_put(const char *symbol);
 #define symbol_put(x) __symbol_put(MODULE_SYMBOL_PREFIX #x)
 void symbol_put_addr(void *addr);
 
+static inline local_t *__module_ref_addr(struct module *mod, int cpu)
+{
+#ifdef CONFIG_SMP
+	return (local_t *) (mod->refptr + per_cpu_offset(cpu));
+#else
+	return &mod->ref;
+#endif
+}
+
 /* Sometimes we know we already have a refcount, and it's easier not
    to handle the error case (which only happens with rmmod --wait). */
 static inline void __module_get(struct module *module)
 {
 	if (module) {
 		BUG_ON(module_refcount(module) == 0);
-		local_inc(&module->ref[get_cpu()].count);
+		local_inc(__module_ref_addr(module, get_cpu()));
 		put_cpu();
 	}
 }

@@ -396,7 +403,7 @@ static inline int try_module_get(struct module *module)
 	if (module) {
 		unsigned int cpu = get_cpu();
 		if (likely(module_is_live(module)))
-			local_inc(&module->ref[cpu].count);
+			local_inc(__module_ref_addr(module, cpu));
 		else
 			ret = 0;
 		put_cpu();
kernel/module.c +30 -10
index f5e9491..e82b2d9 100644
@@ -522,13 +522,13 @@ static char last_unloaded_module[MODULE_NAME_LEN+1];
 /* Init the unload section of the module. */
 static void module_unload_init(struct module *mod)
 {
-	unsigned int i;
+	int cpu;
 
 	INIT_LIST_HEAD(&mod->modules_which_use_me);
-	for (i = 0; i < NR_CPUS; i++)
-		local_set(&mod->ref[i].count, 0);
+	for_each_possible_cpu(cpu)
+		local_set(__module_ref_addr(mod, cpu), 0);
 	/* Hold reference count during initialization. */
-	local_set(&mod->ref[raw_smp_processor_id()].count, 1);
+	local_set(__module_ref_addr(mod, raw_smp_processor_id()), 1);
 	/* Backwards compatibility macros put refcount during init. */
 	mod->waiter = current;
 }

@@ -659,10 +659,11 @@ static int try_stop_module(struct module *mod, int flags, int *forced)
 
 unsigned int module_refcount(struct module *mod)
 {
-	unsigned int i, total = 0;
+	unsigned int total = 0;
+	int cpu;
 
-	for (i = 0; i < NR_CPUS; i++)
-		total += local_read(&mod->ref[i].count);
+	for_each_possible_cpu(cpu)
+		total += local_read(__module_ref_addr(mod, cpu));
 	return total;
 }
 EXPORT_SYMBOL(module_refcount);

@@ -824,7 +825,7 @@ void module_put(struct module *module)
 {
 	if (module) {
 		unsigned int cpu = get_cpu();
-		local_dec(&module->ref[cpu].count);
+		local_dec(__module_ref_addr(module, cpu));
 		/* Maybe they're waiting for us to drop reference? */
 		if (unlikely(!module_is_live(module)))
 			wake_up_process(module->waiter);

@@ -1392,7 +1393,9 @@ static void free_module(struct module *mod)
 	kfree(mod->args);
 	if (mod->percpu)
 		percpu_modfree(mod->percpu);
-
+#if defined(CONFIG_MODULE_UNLOAD) && defined(CONFIG_SMP)
+	percpu_modfree(mod->refptr);
+#endif
 	/* Free lock-classes: */
 	lockdep_free_key_range(mod->module_core, mod->core_size);
 

@@ -1760,6 +1763,9 @@ static struct module *load_module(void __user *umod,
 	unsigned int markersstringsindex;
 	struct module *mod;
 	long err = 0;
+#if defined(CONFIG_MODULE_UNLOAD) && defined(CONFIG_SMP)
+	void *refptr = NULL;
+#endif
 	void *percpu = NULL, *ptr = NULL; /* Stops spurious gcc warning */
 	struct exception_table_entry *extable;
 	mm_segment_t old_fs;

@@ -1904,6 +1910,16 @@ static struct module *load_module(void __user *umod,
 	if (err < 0)
 		goto free_mod;
 
+#if defined(CONFIG_MODULE_UNLOAD) && defined(CONFIG_SMP)
+	refptr = percpu_modalloc(sizeof(local_t),
+				 sizeof(local_t),
+				 mod->name);
+	if (!refptr) {
+		err = -ENOMEM;
+		goto free_mod;
+	}
+	mod->refptr = refptr;
+#endif
 	if (pcpuindex) {
 		/* We have a special allocation for this section. */
 		percpu = percpu_modalloc(sechdrs[pcpuindex].sh_size,

@@ -1911,7 +1927,7 @@ static struct module *load_module(void __user *umod,
 					 mod->name);
 		if (!percpu) {
 			err = -ENOMEM;
-			goto free_mod;
+			goto free_refptr;
 		}
 		sechdrs[pcpuindex].sh_flags &= ~(unsigned long)SHF_ALLOC;
 		mod->percpu = percpu;

@@ -2164,6 +2180,10 @@ static struct module *load_module(void __user *umod,
  free_percpu:
 	if (percpu)
 		percpu_modfree(percpu);
+ free_refptr:
+#if defined(CONFIG_MODULE_UNLOAD) && defined(CONFIG_SMP)
+	percpu_modfree(refptr);
+#endif
  free_mod:
 	kfree(args);
  free_hdr:
#5 lore raw
Quoted 1 line · Eric Dumazet

Rusty Russell a écrit :

...

Quoted 17 lines
Quoted 9 lines

Hi Eric,

I like this patch! The plan was always to create a proper dynamic
per-cpu
allocator which used the normal per-cpu offsets, but I think module refcounts
are worthwhile as a special case.

Quoted 8 lines

Any chance I can ask you look at the issue of full dynamic per-cpu
allocation? The problem of allocating memory which is laid out precisely as the original per-cpu alloc is vexing on NUMA, and probably requires reserving virtual address space and remapping into it, but the rewards would be maximally-efficient per-cpu accessors, and getting rid of that boutique allocator in module.c.

Quoted 8 lines

You mean using alloc_percpu() ? Problem is that current implementation is expensive, since it is using
an extra array of pointers (struct percpu_data). On x86_64, that means at least a 200% space increase
over the solution of using 4 bytes in the static percpu zone. We probably can change this to dynamic
per-cpu as soon as Mike or Christopher finish their work on new dynamic per-cpu implementation ?

Yes, the zero-based percpu variables followed by the cpu_alloc patch should provide this and shrink the code quite well, including in some cases removing locking requirements (because the resultant instructions will be atomic.)

Thanks,
Mike

#6 lore raw
Quoted 7 lines · On Friday 16 May 2008 23:41:16 Mike Travis wrote:
Quoted 6 lines · Eric Dumazet

Rusty Russell a écrit :

Quoted 2 lines

Any chance I can ask you look at the issue of full dynamic per-cpu
allocation?

You mean using alloc_percpu() ? Problem is that current implementation is expensive,

I mean rewriting alloc_percpu :)

Quoted 7 lines
Quoted 2 lines

We probably can change this to dynamic per-cpu as soon as Mike or Christopher finish their work on new dynamic per-cpu implementation ?

Yes, the zero-based percpu variables followed by the cpu_alloc patch should provide this and shrink the code quite well, including in some cases removing locking requirements (because the resultant instructions will be atomic.)

Ah, I hadn't realized that Mike was already working on this. Mike, have you published patches already?

Thanks,
Rusty.

author
#7 lore raw
Quoted 19 lines · Rusty Russell
Quoted 10 lines · On Friday 16 May 2008 23:41:16 Mike Travis wrote:
Quoted 8 lines · Eric Dumazet

Rusty Russell a écrit :

Quoted 3 lines

Any chance I can ask you look at the issue of full dynamic per-cpu
allocation?

You mean using alloc_percpu() ? Problem is that current implementation is expensive,

I mean rewriting alloc_percpu :)

Quoted 3 lines
Quoted 3 lines

We probably can change this to dynamic per-cpu as soon as Mike or Christopher finish their work on new dynamic per-cpu implementation ?

Quoted 9 lines
Quoted 5 lines

Yes, the zero-based percpu variables followed by the cpu_alloc patch should provide this and shrink the code quite well, including in some cases removing locking requirements (because the resultant instructions will be atomic.)

Ah, I hadn't realized that Mike was already working on this. Mike, have you published patches already?

Christoph Lameter made good work last year and apparently the path is to :

1) Put pda at the begining of percpu section.
2) Relocate percpu variables to begin at zero
3) Use %gs (or %fs) register to address pda AND percpu section.
4) Reserve an appropriate large virtual zone into percpu section so that it can fullfill :
- percpu "static" allocations needed by module loader.
- per_cpu "dynamic" allocations
(No more need for an array of pointers to find the address of dynamic
per_cpu variable)
- Eventually allocates PAGES dynamically into this zone to satisfy any size of percpu needs. (But with a limit of say 256 MB per cpu on x86_64)

Thats a lot of work and apparently Mike and Christoph are doing it.

Some pointers to previous work:

http://groups.google.pl/group/linux.kernel/browse_thread/thread/f2ff6901ca6ae9fc/b2ed3b7f3612a157?lnk=raot

http://kerneltrap.org/mailarchive/linux-kernel/2008/2/1/683104

http://linux.derkeiler.com/Mailing-Lists/Kernel/2008-01/msg09052.html

#8 lore raw
Quoted 3 lines · On Saturday 17 May 2008 17:36:21 Eric Dumazet wrote:

Christoph Lameter made good work last year and apparently the path is to :

1) Put pda at the begining of percpu section.

Well, we killed the pda on i386: it's basically a limited arch-specific percpu section. Unfortunately, ISTR that doing so kills -fstack-protector on x86-64 (gcc was explicitly patched to support the kernel doing this, but it nailed us to a fixed offset off %fs). But the PDA is a redundant idea which should be removed.

Quoted 2 lines

2) Relocate percpu variables to begin at zero
3) Use %gs (or %fs) register to address pda AND percpu section.

Thanks for the prod Eric. ISTR acking those patches; Christoph, what happened to them?

Cheers,
Rusty.

#9 lore raw
Quoted 2 lines · Rusty Russell

Thanks for the prod Eric. ISTR acking those patches; Christoph, what happened to them?

Mike is handling them.

#10 lore raw
In reply to Eric Dumazet
Quoted 1 line · Eric Dumazet

Thats a lot of work and apparently Mike and Christoph are doing it.

Indeed. Thanks for the support. I hope Mike gets the stuff out soon. I know he was working on one issue that cause the zero based stuff to break. Maybe it would help if he would post what he has so far?

#11 lore raw
Quoted 7 lines · Christoph Lameter
Quoted 1 line · Eric Dumazet

Thats a lot of work and apparently Mike and Christoph are doing it.

Indeed. Thanks for the support. I hope Mike gets the stuff out soon. I know he was working on one issue that cause the zero based stuff to break. Maybe it would help if he would post what he has so far?

I'm re-basing it on the x86/latest tree as sched/latest was failing my tests last week on both the AMD and Intel boxes. As soon as I re-verify everything I'll send it in.

Thanks,
Mike

#12 lore raw
In reply to Rusty Russell
Quoted 4 lines · Rusty Russell
Quoted 2 lines
Quoted 2 lines

You mean using alloc_percpu() ? Problem is that current implementation is expensive,

I mean rewriting alloc_percpu :)

cpu_alooc is going to replace that completely.

Quoted 11 lines
Quoted 8 lines
Quoted 3 lines

We probably can change this to dynamic per-cpu as soon as Mike or Christopher finish their work on new dynamic per-cpu implementation ?

Yes, the zero-based percpu variables followed by the cpu_alloc patch should provide this and shrink the code quite well, including in some cases removing locking requirements (because the resultant instructions will be atomic.)

Ah, I hadn't realized that Mike was already working on this. Mike, have you published patches already?

He is reworking my patches. He can use some encouragement though...