Skip to contentSkip to conversation

hotplug memory: minor updates

RFC PATCH In review Jeremy Fitzhardinge 7 participants 30 messages Mar 28, 2008

Subsystems (6)
Memory management x86 MM core MGLRU +2 more subsystems
lore
Share & tools
lore
#2 lore raw

Re: [PATCH 3 of 4] sparsemem: reduce i386 PAE section size

In reply to Jeremy Fitzhardinge
Quoted 3 lines · Jeremy Fitzhardinge

A 1G section size makes memory hotplug too coarse in a virtual environment. Retuce it by a factor of 2 to 512M. I would have liked to make it smaller, but it runs out of reserved flags in the page flags.

There are patches in mm which changes that situation. You should have 5 or 6 more flags to play with.

Jeremy Fitzhardinge

author
#0 lore raw

RFC PATCH 1/1 hotplug memory: minor updates

Hi all,

Here's a small set of patches for memory hotplug which I needed to get the Xen balloon driver working.

1. add add_memory_resource(), which allow a caller to provide the resource structure for memory being added,
2. add warnings when you try to add non-section-aligned and sized memory since it has undesireable results when you do so, and
3. decrease the i386 PAE section size to 2^29 (=512MB), since 1G is a bit too large for virtual memory ballooning
(4. don't steal all memory in paravirt_disable_iospace)

Still unresolved is how to prevent the user from onlining balloon memory via sysfs, since that will almost certainly cause the kernel to crash.

Thanks,
J

author
#9 lore raw

PATCH 1/1 paravirt_ops: don't steal memory resources in paravirt_disable_iospace

The memory resource is also used for main memory, and we need it to allocate physical addresses for memory hotplug. Knobbling io space is enough to get the job done anyway.

Signed-off-by Jeremy Fitzhardinge <[email protected]>
Cc Rusty Russell <[email protected]>
---
 arch/x86/kernel/paravirt.c |   18 +-----------------
 1 file changed, 1 insertion(+), 17 deletions(-)
arch/x86/kernel/paravirt.c +1 -17
@@ -206,13 +206,6 @@
 	.flags = IORESOURCE_IO | IORESOURCE_BUSY,
 };
 
-static struct resource reserve_iomem = {
-	.start = 0,
-	.end = -1,
-	.name = "paravirt-iomem",
-	.flags = IORESOURCE_MEM | IORESOURCE_BUSY,
-};
-
 /*
  * Reserve the whole legacy IO space to prevent any legacy drivers
  * from wasting time probing for their hardware.  This is a fairly

@@ -222,16 +215,7 @@
  */
 int paravirt_disable_iospace(void)
 {
-	int ret;
-
-	ret = request_resource(&ioport_resource, &reserve_ioports);
-	if (ret == 0) {
-		ret = request_resource(&iomem_resource, &reserve_iomem);
-		if (ret)
-			release_resource(&reserve_ioports);
-	}
-
-	return ret;
+	return request_resource(&ioport_resource, &reserve_ioports);
 }
 
 static DEFINE_PER_CPU(enum paravirt_lazy_mode, paravirt_lazy_mode) = PARAVIRT_LAZY_NONE;
author
#11 lore raw

PATCH 1/1 hotplug-memory: add add_memory_resource

In reply to Jeremy Fitzhardinge

The Xen balloon driver uses hotplug memory to extend the amount of pseudo-physical memory available to a virtual domain. Since this memory is virtual, it can go anywhere within the kernel's pseudo-physical address space.

This patch adds a new hotplug memory call, add_memory_resource(), which adds memory corresponding to a pre-constructed resource. This allows callers to use allocate_resource() to allocate a suitable chunk of address space for the new memory (as well as cosmetic details, like a descriptive name).

Signed-off-by Jeremy Fitzhardinge <[email protected]>
Cc KAMEZAWA Hiroyuki <[email protected]>
Cc Yasunori Goto <[email protected]>
Cc Christoph Lameter <[email protected]>
Cc Dave Hansen <[email protected]>
---
 include/linux/memory_hotplug.h |    3 +++
 mm/memory_hotplug.c            |   20 ++++++++++++++++----
 2 files changed, 19 insertions(+), 4 deletions(-)
include/linux/memory_hotplug.h +3 -0
@@ -171,7 +171,10 @@
 
 #endif /* ! CONFIG_MEMORY_HOTPLUG */
 
+struct resource;
+
 extern int add_memory(int nid, u64 start, u64 size);
+extern int add_memory_resource(int nid, struct resource *res);
 extern int arch_add_memory(int nid, u64 start, u64 size);
 extern int remove_memory(u64 start, u64 size);
 extern int sparse_add_one_section(struct zone *zone, unsigned long start_pfn,
mm/memory_hotplug.c +16 -4
@@ -278,14 +278,28 @@
 
 int add_memory(int nid, u64 start, u64 size)
 {
-	pg_data_t *pgdat = NULL;
-	int new_pgdat = 0;
 	struct resource *res;
 	int ret;
 
 	res = register_memory_resource(start, size);
 	if (!res)
 		return -EEXIST;
+
+	ret = add_memory_resource(nid, res);
+
+	if (ret)
+		release_memory_resource(res);
+
+	return ret;
+}
+
+int add_memory_resource(int nid, struct resource *res)
+{
+	pg_data_t *pgdat = NULL;
+	int new_pgdat = 0;
+	int ret;
+	u64 start = res->start;
+	u64 size = res->end - res->start + 1;
 
 	if (!node_online(nid)) {
 		pgdat = hotadd_new_pgdat(nid, start);

@@ -320,8 +334,6 @@
 	/* rollback pgdat allocation and others */
 	if (new_pgdat)
 		rollback_node_hotadd(nid, pgdat);
-	if (res)
-		release_memory_resource(res);
 
 	return ret;
 }
author
#12 lore raw

PATCH 1/1 hotplug-memory: adding non-section-aligned memory is bad

In reply to Jeremy Fitzhardinge

Adding non-section-aligned memory will cause bad results, because page structures are only built on section-aligned boundaries. Also, memory whose size isn't a multiple of the section size is silently ignored.

This patch adds a couple of WARN_ONs to help confused programmers work out what's going wrong when they hotplug memory without being aware of these constraints.

Signed-off-by Jeremy Fitzhardinge <[email protected]>
Cc KAMEZAWA Hiroyuki <[email protected]>
Cc Yasunori Goto <[email protected]>
Cc Christoph Lameter <[email protected]>
Cc Dave Hansen <[email protected]>
include/linux/mmzone.h +3 -0
@@ -784,6 +784,9 @@
 #define PAGES_PER_SECTION       (1UL << PFN_SECTION_SHIFT)
 #define PAGE_SECTION_MASK	(~(PAGES_PER_SECTION-1))
 
+#define SECTION_SIZE		(1UL << SECTION_SIZE_BITS)
+#define SECTION_SIZE_MASK	(SECTION_SIZE - 1)
+
 #define SECTION_BLOCKFLAGS_BITS \
 	((1UL << (PFN_SECTION_SHIFT - pageblock_order)) * NR_PAGEBLOCK_BITS)
mm/memory_hotplug.c +5 -0
@@ -300,6 +300,11 @@
 	int ret;
 	u64 start = res->start;
 	u64 size = res->end - res->start + 1;
+
+	/* Adding non-section-aligned memory will give unexpected
+	   and unintuitive results. */
+	WARN_ON((start & SECTION_SIZE_MASK) != 0);
+	WARN_ON((size & SECTION_SIZE_MASK) != 0);
 
 	if (!node_online(nid)) {
 		pgdat = hotadd_new_pgdat(nid, start);
author
#1 lore raw

PATCH 1/1 sparsemem: reduce i386 PAE section size

In reply to Jeremy Fitzhardinge

A 1G section size makes memory hotplug too coarse in a virtual environment. Retuce it by a factor of 2 to 512M. I would have liked to make it smaller, but it runs out of reserved flags in the page flags.

Signed-off-by Jeremy Fitzhardinge <[email protected]>
Cc KAMEZAWA Hiroyuki <[email protected]>
Cc Yasunori Goto <[email protected]>
Cc Christoph Lameter <[email protected]>
Cc Dave Hansen <[email protected]>
---
 include/asm-x86/sparsemem.h |    2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
include/asm-x86/sparsemem.h +1 -1
@@ -16,7 +16,7 @@
 
 #ifdef CONFIG_X86_32
 # ifdef CONFIG_X86_PAE
-#  define SECTION_SIZE_BITS	30
+#  define SECTION_SIZE_BITS	29
 #  define MAX_PHYSADDR_BITS	36
 #  define MAX_PHYSMEM_BITS	36
 # else
#13 lore raw

Re: [PATCH 2 of 4] hotplug-memory: adding non-section-aligned memory is bad

In reply to Jeremy Fitzhardinge
Quoted 17 lines · On Thu, 27 Mar 2008 17:28:38 -0700 Jeremy Fitzhardinge <[email protected]> wrote:
include/linux/mmzone.h +3 -0
@@ -784,6 +784,9 @@
 #define PAGES_PER_SECTION       (1UL << PFN_SECTION_SHIFT)
 #define PAGE_SECTION_MASK	(~(PAGES_PER_SECTION-1))
 
+#define SECTION_SIZE		(1UL << SECTION_SIZE_BITS)
+#define SECTION_SIZE_MASK	(SECTION_SIZE - 1)
+
 #define SECTION_BLOCKFLAGS_BITS \
 	((1UL << (PFN_SECTION_SHIFT - pageblock_order)) * NR_PAGEBLOCK_BITS)

mm/memory_hotplug.c line 308

Quoted 9 lines
mm/memory_hotplug.c +5 -0
@@ -300,6 +300,11 @@
 	int ret;
 	u64 start = res->start;
 	u64 size = res->end - res->start + 1;
+
+	/* Adding non-section-aligned memory will give unexpected
+	   and unintuitive results. */
+	WARN_ON((start & SECTION_SIZE_MASK) != 0);
+	WARN_ON((size & SECTION_SIZE_MASK) != 0);

Why just WARNING ? not BUG_ON?

Thanks,
-Kame

#14 lore raw

Re: [PATCH 2 of 4] hotplug-memory: adding non-section-aligned memory is bad

Quoted 15 lines
Quoted 13 lines · On Thu, 27 Mar 2008 17:28:38 -0700 Jeremy Fitzhardinge <[email protected]> wrote:
include/linux/mmzone.h +3 -0
@@ -784,6 +784,9 @@
 #define PAGES_PER_SECTION       (1UL << PFN_SECTION_SHIFT)
 #define PAGE_SECTION_MASK	(~(PAGES_PER_SECTION-1))
 
+#define SECTION_SIZE		(1UL << SECTION_SIZE_BITS)
+#define SECTION_SIZE_MASK	(SECTION_SIZE - 1)
+
 #define SECTION_BLOCKFLAGS_BITS \
 	((1UL << (PFN_SECTION_SHIFT - pageblock_order)) * NR_PAGEBLOCK_BITS)
Quoted 14 lines

mm/memory_hotplug.c line 308

Quoted 13 lines
mm/memory_hotplug.c +5 -0
@@ -300,6 +300,11 @@
 	int ret;
 	u64 start = res->start;
 	u64 size = res->end - res->start + 1;
+
+	/* Adding non-section-aligned memory will give unexpected
+	   and unintuitive results. */
+	WARN_ON((start & SECTION_SIZE_MASK) != 0);
+	WARN_ON((size & SECTION_SIZE_MASK) != 0);

Why just WARNING ? not BUG_ON?

Both Nack.

Because, firmware may occupy some area in the section. Firmware must exclude those area to notify kernel. So, E820, EFI, or _CRS of ACPI may return not aligned address and size. register_memory_resource() and walk_memory_resource() are to skip them silently. This is intended.

Bye.

Signature · 1 line
Yasunori Goto
#15 lore raw

Re: [PATCH 2 of 4] hotplug-memory: adding non-section-aligned memory is bad

Quoted 15 lines · On Fri, 28 Mar 2008 10:34:51 +0900 Yasunori Goto <[email protected]> wrote:
Quoted 6 lines
Quoted 6 lines
mm/memory_hotplug.c +0 -0
@@ -300,6 +300,11 @@
 	int ret;
 	u64 start = res->start;
Quoted 8 lines

mm/memory_hotplug.c line 308

Quoted 7 lines
mm/memory_hotplug.c +5 -0
 	u64 size = res->end - res->start + 1;
+
+	/* Adding non-section-aligned memory will give unexpected
+	   and unintuitive results. */
+	WARN_ON((start & SECTION_SIZE_MASK) != 0);
+	WARN_ON((size & SECTION_SIZE_MASK) != 0);

Why just WARNING ? not BUG_ON?

Quoted 8 lines

Both Nack.

Because, firmware may occupy some area in the section. Firmware must exclude those area to notify kernel. So, E820, EFI, or _CRS of ACPI may return not aligned address and size. register_memory_resource() and walk_memory_resource() are to skip them silently. This is intended.

Ah, ok. sorry.

Jeremy, I think you can check whether you have 'struct page' or not by pfn_valid().

If pfn_valid() == false, you should call add_memory() and create a section/mem_map. If pfn_valid() == true, you should just remove PG_reserved bit in mem_map by online_page().

Thanks,
-Kame

author
#16 lore raw

Re: [PATCH 2 of 4] hotplug-memory: adding non-section-aligned memory is bad

Quoted 28 lines · KAMEZAWA Hiroyuki
Quoted 16 lines · On Fri, 28 Mar 2008 10:34:51 +0900 Yasunori Goto <[email protected]> wrote:
Quoted 7 lines
Quoted 7 lines
mm/memory_hotplug.c +0 -0
@@ -300,6 +300,11 @@
 	int ret;
 	u64 start = res->start;
 	u64 size = res->end - res->start + 1;
Quoted 9 lines

mm/memory_hotplug.c

Quoted 7 lines
mm/memory_hotplug.c +5 -0
+
+	/* Adding non-section-aligned memory will give unexpected
+	   and unintuitive results. */
+	WARN_ON((start & SECTION_SIZE_MASK) != 0);
+	WARN_ON((size & SECTION_SIZE_MASK) != 0);

Why just WARNING ? not BUG_ON?

Quoted 9 lines

Both Nack.

Because, firmware may occupy some area in the section. Firmware must exclude those area to notify kernel. So, E820, EFI, or _CRS of ACPI may return not aligned address and size. register_memory_resource() and walk_memory_resource() are to skip them silently. This is intended.

Quoted 8 lines

Ah, ok. sorry.

Jeremy, I think you can check whether you have 'struct page' or not by pfn_valid().

If pfn_valid() == false, you should call add_memory() and create a section/mem_map. If pfn_valid() == true, you should just remove PG_reserved bit in mem_map by online_page().

OK. Would that ever be necessary if I explicitly align my start and size?

J

author
#3 lore raw

Re: [PATCH 3 of 4] sparsemem: reduce i386 PAE section size

In reply to Christoph Lameter
Quoted 3 lines · Christoph Lameter

There are patches in mm which changes that situation. You should have 5 or 6 more flags to play with.

Oh, good. We can play with it some more when everything gets merged together.

J

#4 lore raw

Re: [PATCH 3 of 4] sparsemem: reduce i386 PAE section size

Quoted 7 lines · Jeremy Fitzhardinge
Quoted 4 lines · Christoph Lameter

There are patches in mm which changes that situation. You should have 5 or 6 more flags to play with.

Oh, good. We can play with it some more when everything gets merged together.

Only works on i386 if you either use sparsemem / vmemmap or not sparsemem though. Is there any need for the other sparsemem memory models? Or could we disable them like on x86_64?

#17 lore raw

Re: [PATCH 2 of 4] hotplug-memory: adding non-section-aligned memory is bad

In reply to Jeremy Fitzhardinge
Quoted 9 lines · On Thu, 27 Mar 2008 19:13:20 -0700 Jeremy Fitzhardinge <[email protected]> wrote:
Quoted 9 lines
Quoted 7 lines

Because, firmware may occupy some area in the section. Firmware must exclude those area to notify kernel. So, E820, EFI, or _CRS of ACPI may return not aligned address and size. register_memory_resource() and walk_memory_resource() are to skip them silently. This is intended.

Ah, ok. sorry.

Quoted 9 lines
Quoted 6 lines

Jeremy, I think you can check whether you have 'struct page' or not by pfn_valid().

If pfn_valid() == false, you should call add_memory() and create a section/mem_map. If pfn_valid() == true, you should just remove PG_reserved bit in mem_map by online_page().

OK. Would that ever be necessary if I explicitly align my start and size?

Maybe no. but be carefull not to register resource in overlapped manner. (I wrote online_page() in above, but online_pages() is maybe better.
It does all what you want.)

Start/Size are automatically alined to section in __add_pages().

See below.
==
110 int __add_pages(struct zone *zone, unsigned long phys_start_pfn,
111 unsigned long nr_pages)
112 {

 113         unsigned long i;
 114         int err = 0;
 115         int start_sec, end_sec;
 116         /* during initialize mem_map, align hot-added range to section */
 117         start_sec = pfn_to_section_nr(phys_start_pfn);
 118         end_sec = pfn_to_section_nr(phys_start_pfn + nr_pages - 1);

==

And online_pages(), which onlines pages in [pfn, pfn + size), will see registerred resources within [pfn, pfn + size).
==
184 int online_pages(unsigned long pfn, unsigned long nr_pages) <snip>
227 walk_memory_resource(pfn, nr_pages, &onlined_pages,
228 online_pages_range);
==

One of my concern is how-to-handle sysfs status in this case.

Another concerns is, currently, I think no one tried to online a section twice to online reserved pages in a section. so, you may see bug.
For example, mem_notify() in online_pages() will be called several times against a section.

Thanks,
-Kame

#10 lore raw

Re: [PATCH 4 of 4] paravirt_ops: don't steal memory resources in paravirt_disable_iospace

In reply to Jeremy Fitzhardinge
Quoted 3 lines · On Friday 28 March 2008 11:28:40 Jeremy Fitzhardinge wrote:

The memory resource is also used for main memory, and we need it to allocate physical addresses for memory hotplug. Knobbling io space is enough to get the job done anyway.

Indeed, tested with lguest and it works for me. I've put this in my patchqueue, but if it all gets taken by someone else I can easily drop it.

Thanks,
Rusty.

author
#18 lore raw

Re: [PATCH 2 of 4] hotplug-memory: adding non-section-aligned memory is bad

In reply to KAMEZAWA Hiroyuki
Quoted 17 lines · KAMEZAWA Hiroyuki
Quoted 13 lines · On Thu, 27 Mar 2008 19:13:20 -0700 Jeremy Fitzhardinge <[email protected]> wrote:
Quoted 13 lines
Quoted 8 lines

Because, firmware may occupy some area in the section. Firmware must exclude those area to notify kernel. So, E820, EFI, or _CRS of ACPI may return not aligned address and size. register_memory_resource() and walk_memory_resource() are to skip them silently. This is intended.

Ah, ok. sorry.

Jeremy, I think you can check whether you have 'struct page' or not by pfn_valid().

Quoted 9 lines
Quoted 7 lines
Quoted 4 lines

If pfn_valid() == false, you should call add_memory() and create a section/mem_map. If pfn_valid() == true, you should just remove PG_reserved bit in mem_map by online_page().

OK. Would that ever be necessary if I explicitly align my start and size?

Maybe no. but be carefull not to register resource in overlapped manner.

Yes. That's why I added add_memory_resource(), so I could use allocate_resource() to find a non-overlapping range to put the new memory.

Quoted 3 lines

(I wrote online_page() in above, but online_pages() is maybe better.
It does all what you want.)

No, for my use-case the pages must be onlined one by one as they get some physical memory assigned to them. At the time I do add_memory(), I'm just allocating page structures, but there's no memory backing that range.

That's why I need to disable the sysfs onlining interface, because it bulk onlines the pages before there's anything behind them.

Quoted 24 lines

Start/Size are automatically alined to section in __add_pages().

See below.
==
110 int __add_pages(struct zone *zone, unsigned long phys_start_pfn,
111 unsigned long nr_pages)
112 {

 113         unsigned long i;
 114         int err = 0;
 115         int start_sec, end_sec;
 116         /* during initialize mem_map, align hot-added range to section */
 117         start_sec = pfn_to_section_nr(phys_start_pfn);
 118         end_sec = pfn_to_section_nr(phys_start_pfn + nr_pages - 1);

==

And online_pages(), which onlines pages in [pfn, pfn + size), will see registerred resources within [pfn, pfn + size).
==
184 int online_pages(unsigned long pfn, unsigned long nr_pages) <snip>
227 walk_memory_resource(pfn, nr_pages, &onlined_pages,
228 online_pages_range);
==

Quoted 7 lines

One of my concern is how-to-handle sysfs status in this case.

Another concerns is, currently, I think no one tried to online a section twice to online reserved pages in a section. so, you may see bug.
For example, mem_notify() in online_pages() will be called several times against a section.

I'd really rather prevent online_pages from happening at all, since it can only cause havoc.

Thanks,
J

#19 lore raw

Re: [PATCH 2 of 4] hotplug-memory: adding non-section-aligned memory is bad

Quoted 9 lines · On Thu, 27 Mar 2008 19:51:26 -0700 Jeremy Fitzhardinge <[email protected]> wrote:
Quoted 3 lines

(I wrote online_page() in above, but online_pages() is maybe better.
It does all what you want.)

No, for my use-case the pages must be onlined one by one as they get some physical memory assigned to them. At the time I do add_memory(), I'm just allocating page structures, but there's no memory backing that range.

yes, I see.

Quoted 3 lines

That's why I need to disable the sysfs onlining interface, because it bulk onlines the pages before there's anything behind them.

My point is. online_pages() does following things.

- call notifier
- update information , total_pages etc...
- re-configure zonelist if necessary...

But online_page() not. Hmm...

How about capturing online_page() by balloon ?
ex.)
==
call add_memory() to create mem_map
call online_pages() against the whole section. <=== call this without sysfs.
online_pages() do misc. jobs
call online_page() one by one (arch dependent) called by walk_memory_resource.
online_page() will finally call free_page(page). <=========== Xen capture here.
Don't free onlined page and swallow them into baloon driver.
==

Thanks,
-Kame

author
#20 lore raw

Re: [PATCH 2 of 4] hotplug-memory: adding non-section-aligned memory is bad

In reply to KAMEZAWA Hiroyuki
Quoted 20 lines · KAMEZAWA Hiroyuki
Quoted 7 lines · On Thu, 27 Mar 2008 19:51:26 -0700 Jeremy Fitzhardinge <[email protected]> wrote:
Quoted 4 lines

(I wrote online_page() in above, but online_pages() is maybe better.
It does all what you want.)

No, for my use-case the pages must be onlined one by one as they get some physical memory assigned to them. At the time I do add_memory(), I'm just allocating page structures, but there's no memory backing that range.

yes, I see.

Quoted 3 lines

That's why I need to disable the sysfs onlining interface, because it bulk onlines the pages before there's anything behind them.

My point is. online_pages() does following things.

Quoted 8 lines

- call notifier
- update information , total_pages etc...
- re-configure zonelist if necessary...

But online_page() not. Hmm...

How about capturing online_page() by balloon ?

You're saying that using online_page() on each page on its own is not sufficient?

Quoted 11 lines

ex.)
==
call add_memory() to create mem_map
call online_pages() against the whole section. <=== call this without sysfs.
online_pages() do misc. jobs
call online_page() one by one (arch dependent) called by walk_memory_resource.
online_page() will finally call free_page(page). <=========== Xen capture here.
Don't free onlined page and swallow them into baloon driver.
==

I'm not sure what you mean by "capture" here. Do you mean a hook? I'd rather not have to put some Xen-specific hook in here.

What would happen if I did online_pages(pfn, 1) on each page as I populate it?

J

#21 lore raw

Re: [PATCH 2 of 4] hotplug-memory: adding non-section-aligned memory is bad

Quoted 6 lines · On Thu, 27 Mar 2008 21:20:18 -0700 Jeremy Fitzhardinge <[email protected]> wrote:
Quoted 2 lines

How about capturing online_page() by balloon ?

You're saying that using online_page() on each page on its own is not sufficient?

online_page() just remove PG_reserved flags and free it. Not dealing with global data, like total pages, zonelist, etc.

Quoted 12 lines
Quoted 11 lines

ex.)
==
call add_memory() to create mem_map
call online_pages() against the whole section. <=== call this without sysfs.
online_pages() do misc. jobs
call online_page() one by one (arch dependent) called by walk_memory_resource.
online_page() will finally call free_page(page). <=========== Xen capture here.
Don't free onlined page and swallow them into baloon driver.
==

Quoted 3 lines

I'm not sure what you mean by "capture" here. Do you mean a hook? I'd rather not have to put some Xen-specific hook in here.

ok.

Quoted 3 lines

What would happen if I did online_pages(pfn, 1) on each page as I populate it?

I think (hope) it works well. But it seems no one tries to do that.

Thanks,
-Kame

author
#5 lore raw

Re: [PATCH 3 of 4] sparsemem: reduce i386 PAE section size

In reply to Christoph Lameter
Quoted 1 line · Christoph Lameter

Only works on i386 if you either use sparsemem / vmemmap or not sparsemem though. Is there any need for the other sparsemem memory models? Or could we disable them like on x86_64?

Probably, but I don't know what the tradeoffs are. All I want is hotplug memory.

J

#27 lore raw
In reply to Jeremy Fitzhardinge
Quoted 4 lines · Jeremy Fitzhardinge

Hi all,

Here's a small set of patches for memory hotplug which I needed to get the Xen balloon driver working.

the x86 bits look small and trivial, so i'd suggest you get this into -mm to after all the other VM patches, because this seems primarily a VM issue.

i've applied #3 to x86.git straight away. Does #4 have dependencies on the other patches? I suspect it does so it cannot be applied standalone.

Ingo

author
#28 lore raw
Quoted 4 lines · Ingo Molnar

the x86 bits look small and trivial, so i'd suggest you get this into -mm to after all the other VM patches, because this seems primarily a VM issue.

I would do that, but I'm planning on posting some Xen balloon driver patches which depend on at least one of these patches.

I'll drop "hotplug-memory: adding non-section-aligned memory is bad" because it was just warnings and it was NACKed. "hotplug-memory: add add_memory_resource()" is just a refactor, and should be a noop as far as the existing add_memory() is concerned.

Quoted 2 lines

i've applied #3 to x86.git straight away. Does #4 have dependencies on the other patches? I suspect it does so it cannot be applied standalone.

You mean the "paravirt_ops: don't steal memory resources in paravirt_disable_iospace"? That shouldn't have any dependencies. It's the only patch in my queue which even references that file. Rusty said he picked it up too; are you exchanging patches with him?

J

#6 lore raw

Re: [PATCH 3 of 4] sparsemem: reduce i386 PAE section size

In reply to Jeremy Fitzhardinge
Quoted 7 lines · Jeremy Fitzhardinge
Quoted 3 lines · Christoph Lameter

Only works on i386 if you either use sparsemem / vmemmap or not sparsemem though. Is there any need for the other sparsemem memory models? Or could we disable them like on x86_64?

Probably, but I don't know what the tradeoffs are. All I want is hotplug memory.

Developers should probably learn the tradeoffs before they go start poking around in code and sending patches. ;)

Maybe Christoph has a good idea what the tradeoffs are for the vmemmap variant, but the table-based one is pretty simple. Smaller sections mean that it costs you larger sparsemem structures and more bits in page->flags. But, the smaller the section, the more mem_map[] you might waste.

BTW, the current 1G sections were picked because all the hardware we knew of at the time had 512MB dimms that had to be added in pairs. So, we didn't ever have physical hotplug of less than that.

-- Dave

#8 lore raw

Re: [PATCH 3 of 4] sparsemem: reduce i386 PAE section size

In reply to Jeremy Fitzhardinge

include/asm-x86/sparsemem.h line 22

Quoted 12 lines · Jeremy Fitzhardinge
include/asm-x86/sparsemem.h +1 -1
@@ -16,7 +16,7 @@
 
 #ifdef CONFIG_X86_32
 # ifdef CONFIG_X86_PAE
-#  define SECTION_SIZE_BITS    30
+#  define SECTION_SIZE_BITS    29
 #  define MAX_PHYSADDR_BITS    36
 #  define MAX_PHYSMEM_BITS     36
 # else

This is fine with me as long as we have the room in page->flags. I probably need to look at Christoph's new patches, too.

-- Dave

author
#7 lore raw

Re: [PATCH 3 of 4] sparsemem: reduce i386 PAE section size

In reply to Dave Hansen
Quoted 4 lines · Dave Hansen

Developers should probably learn the tradeoffs before they go start poking around in code and sending patches. ;)

I don't know any better way of learning the tradeoffs than by poking around and sending patches ;)

Quoted 4 lines

BTW, the current 1G sections were picked because all the hardware we knew of at the time had 512MB dimms that had to be added in pairs. So, we didn't ever have physical hotplug of less than that.

I guessed it was something like that. Even for my case 1G isn't too bad. It was bad when I was trying to hotplug early and it was having allocation failures in the bootmem allocator, but I think it should be OK now. On the other hand, when you're hotplugging new memory is precisely when you don't want to be allocating lots of new memory...

J

author
#22 lore raw

Re: [PATCH 2 of 4] hotplug-memory: adding non-section-aligned memory is bad

In reply to KAMEZAWA Hiroyuki
Quoted 4 lines · KAMEZAWA Hiroyuki
Quoted 3 lines

What would happen if I did online_pages(pfn, 1) on each page as I populate it?

I think (hope) it works well. But it seems no one tries to do that.

Well I had immediate problems because I try to use it under spinlock and it calls online_pages -> build_all_zonelists -> stop_machine_run. I can easily rearrange to fix that, but it seems to me that stop_machine_run() is probably too expensive to call thousands of times (one for each page, rather than once per section).

I'm just looking at how I can refactor online_pages() into a function which does the general zone setup stuff, and one for actually onlining pages. The function of online_pages() would remain unchanged for existing users, but I could call the pieces separately at the appropriate times.

J

#23 lore raw

Re: [PATCH 2 of 4] hotplug-memory: adding non-section-aligned memory is bad

Quoted 5 lines · Jeremy Fitzhardinge
Quoted 4 lines · KAMEZAWA Hiroyuki
Quoted 4 lines

What would happen if I did online_pages(pfn, 1) on each page as I populate it?

Quoted 7 lines
Quoted 1 line

I think (hope) it works well. But it seems no one tries to do that.

Well I had immediate problems because I try to use it under spinlock and it calls online_pages -> build_all_zonelists -> stop_machine_run. I can easily rearrange to fix that, but it seems to me that stop_machine_run() is probably too expensive to call thousands of times (one for each page, rather than once per section).

Yeah, you certainly don't want to be messing with things like the zone boundaries for each page online operation.

All that you really want to do is hook into add_one_highpage_hotplug() and keep the pages from going back into the allocator, right? Then, you can have the Xen code go and actually free_page() on them, later.

I don't think *any* of this code actually touches the new physical pages themselves. The issue that you're probably seeing is because they get stuck into the allocator and some other user grabs them right away and *does* touch them.

-- Dave

author
#26 lore raw

Re: [PATCH 2 of 4] hotplug-memory: adding non-section-aligned memory is bad

In reply to KAMEZAWA Hiroyuki
Quoted 3 lines · KAMEZAWA Hiroyuki

For example, mem_notify() in online_pages() will be called several times against a section.

It seems the only in-tree user of the notifier is mm/slub.c, and it only cares about MEM_GOING_ONLINE, GOING_OFFLINE, OFFLINE and CANCEL_ONLINE (not ONLINE or CANCEL_OFFLINE).

Also,

	if (onlined_pages)
		memory_notify(MEM_ONLINE, &arg);

doesn't seem correct, because arg.nr_pages still has the original nr_pages, and not the number of pages actually onlined (onlined_pages). Also arg.start_pfn isn't terribly meaningful if a discontiguous set of pages can be actually onlined.

J

#29 lore raw
In reply to Jeremy Fitzhardinge
Quoted 6 lines · Jeremy Fitzhardinge
Quoted 3 lines

the x86 bits look small and trivial, so i'd suggest you get this into -mm to after all the other VM patches, because this seems primarily a VM issue.

I would do that, but I'm planning on posting some Xen balloon driver patches which depend on at least one of these patches.

ok. Will wait patiently then :)

Ingo