We don't need to do register_memory_resource() since it has its own lock and
doesn't make any callbacks.
Also register_memory_resource return NULL on failure so we don't have anything
to cleanup at this point.
The reason for this rfc is I was doing some experiments with hotplugging of
memory on some of our larger systems. While it seems to work, it can be quite
slow. With some preliminary digging I found that lock_memory_hotplug is
clearly ripe for breakup.
It could be broken up per nid or something but it also covers the
online_page_callback. The online_page_callback shouldn't be very hard to break
out.
Also there is the issue of various structures(wmarks come to mind) that are
only updated under the lock_memory_hotplug that would need to be dealt with.
cc: Andrew Morton <[email protected]>
cc: Tang Chen <[email protected]>
cc: Wen Congyang <[email protected]>
cc: Kamezawa Hiroyuki <[email protected]>
cc: Yasuaki Ishimatsu <[email protected]>
cc: "Rafael J. Wysocki" <[email protected]>
cc: Hedi <[email protected]>
cc: Mike Travis <[email protected]>
cc: [email protected]
cc: [email protected]
---
mm/memory_hotplug.c | 7 ++++---
1 file changed, 4 insertions(+), 3 deletions(-)
diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c
index 1ad92b4..62a0cd1 100644
--- a/mm/memory_hotplug.c
+++ b/mm/memory_hotplug.c
@@ -1097,17 +1097,18 @@ int __ref add_memory(int nid, u64 start, u64 size)
struct resource *res;
int ret;
- lock_memory_hotplug();
-
res = register_memory_resource(start, size);
ret = -EEXIST;
if (!res)
- goto out;
+ return ret;
{ /* Stupid hack to suppress address-never-null warning */
void *p = NODE_DATA(nid);
new_pgdat = !p;
}
+
+ lock_memory_hotplug();
+
new_node = !node_online(nid);
if (new_node) {
pgdat = hotadd_new_pgdat(nid, start);
--
1.8.2.1
On Tue, 14 Jan 2014 12:24:34 -0600 Nathan Zimmer <[email protected]> wrote:
> We don't need to do register_memory_resource() since it has its own lock and
> doesn't make any callbacks.
>
> Also register_memory_resource return NULL on failure so we don't have anything
> to cleanup at this point.
>
>
> The reason for this rfc is I was doing some experiments with hotplugging of
> memory on some of our larger systems. While it seems to work, it can be quite
> slow. With some preliminary digging I found that lock_memory_hotplug is
> clearly ripe for breakup.
>
> It could be broken up per nid or something but it also covers the
> online_page_callback. The online_page_callback shouldn't be very hard to break
> out.
>
> Also there is the issue of various structures(wmarks come to mind) that are
> only updated under the lock_memory_hotplug that would need to be dealt with.
>
> ...
>
> --- a/mm/memory_hotplug.c
> +++ b/mm/memory_hotplug.c
> @@ -1097,17 +1097,18 @@ int __ref add_memory(int nid, u64 start, u64 size)
> struct resource *res;
> int ret;
>
> - lock_memory_hotplug();
> -
> res = register_memory_resource(start, size);
> ret = -EEXIST;
> if (!res)
> - goto out;
> + return ret;
>
> { /* Stupid hack to suppress address-never-null warning */
> void *p = NODE_DATA(nid);
> new_pgdat = !p;
> }
> +
> + lock_memory_hotplug();
> +
> new_node = !node_online(nid);
> if (new_node) {
> pgdat = hotadd_new_pgdat(nid, start);
Looks sane to me.
register_memory_resource() makes me cry. Please review:
From: Andrew Morton <[email protected]>
Subject: mm/memory_hotplug.c: register_memory_resource() fixes
- register_memory_resource() should not go BUG on ENOMEM. That's
appropriate at system boot time, but not at memory-hotplug time. Fix.
- register_memory_resource()'s caller is incorrectly replacing
request_resource()'s -EBUSY with -EEXIST. Fix this by propagating
errors appropriately.
Cc: "Rafael J. Wysocki" <[email protected]>
Cc: Hedi <[email protected]>
Cc: Kamezawa Hiroyuki <[email protected]>
Cc: Mike Travis <[email protected]>
Cc: Nathan Zimmer <[email protected]>
Cc: Tang Chen <[email protected]>
Cc: Wen Congyang <[email protected]>
Cc: Yasuaki Ishimatsu <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
---
mm/memory_hotplug.c | 15 +++++++++------
1 file changed, 9 insertions(+), 6 deletions(-)
diff -puN mm/memory_hotplug.c~mm-memory_hotplugc-register_memory_resource-fixes mm/memory_hotplug.c
--- a/mm/memory_hotplug.c~mm-memory_hotplugc-register_memory_resource-fixes
+++ a/mm/memory_hotplug.c
@@ -64,17 +64,21 @@ void unlock_memory_hotplug(void)
static struct resource *register_memory_resource(u64 start, u64 size)
{
struct resource *res;
+ int err;
+
res = kzalloc(sizeof(struct resource), GFP_KERNEL);
- BUG_ON(!res);
+ if (!res)
+ return ERR_PTR(-ENOMEM);
res->name = "System RAM";
res->start = start;
res->end = start + size - 1;
res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
- if (request_resource(&iomem_resource, res) < 0) {
+ err = request_resource(&iomem_resource, res);
+ if (err) {
pr_debug("System RAM resource %pR cannot be added\n", res);
kfree(res);
- res = NULL;
+ res = ERR_PTR(err);
}
return res;
}
@@ -1108,9 +1112,8 @@ int __ref add_memory(int nid, u64 start,
return ret;
res = register_memory_resource(start, size);
- ret = -EEXIST;
- if (!res)
- return ret;
+ if (IS_ERR(res))
+ return PTR_ERR(res);
{ /* Stupid hack to suppress address-never-null warning */
void *p = NODE_DATA(nid);
_
On Tue, 14 Jan 2014, Nathan Zimmer wrote:
> We don't need to do register_memory_resource() since it has its own lock and
> doesn't make any callbacks.
>
We need to do it, just not under lock_memory_hotplug() :).
> Also register_memory_resource return NULL on failure so we don't have anything
> to cleanup at this point.
>
>
> The reason for this rfc is I was doing some experiments with hotplugging of
> memory on some of our larger systems. While it seems to work, it can be quite
> slow. With some preliminary digging I found that lock_memory_hotplug is
> clearly ripe for breakup.
>
> It could be broken up per nid or something but it also covers the
> online_page_callback. The online_page_callback shouldn't be very hard to break
> out.
>
> Also there is the issue of various structures(wmarks come to mind) that are
> only updated under the lock_memory_hotplug that would need to be dealt with.
>
>
> cc: Andrew Morton <[email protected]>
> cc: Tang Chen <[email protected]>
> cc: Wen Congyang <[email protected]>
> cc: Kamezawa Hiroyuki <[email protected]>
> cc: Yasuaki Ishimatsu <[email protected]>
> cc: "Rafael J. Wysocki" <[email protected]>
> cc: Hedi <[email protected]>
> cc: Mike Travis <[email protected]>
> cc: [email protected]
> cc: [email protected]
Looks like you're modifying a pre-3.12 kernel version that doesn't have
27356f54c8c3 ("mm/hotplug: verify hotplug memory range").
When your patch is signed off, feel free to add
Acked-by: David Rientjes <[email protected]>
On Tue, 14 Jan 2014, Andrew Morton wrote:
> From: Andrew Morton <[email protected]>
> Subject: mm/memory_hotplug.c: register_memory_resource() fixes
>
> - register_memory_resource() should not go BUG on ENOMEM. That's
> appropriate at system boot time, but not at memory-hotplug time. Fix.
>
> - register_memory_resource()'s caller is incorrectly replacing
> request_resource()'s -EBUSY with -EEXIST. Fix this by propagating
> errors appropriately.
>
Unfortunately, -EEXIST is a special case return value for both
acpi_memory_enable_device() and hv_mem_hot_add(), so they would need to be
modified to agree concurrently with this change.
(2014/01/15 3:24), Nathan Zimmer wrote:
> We don't need to do register_memory_resource() since it has its own lock and
> doesn't make any callbacks.
>
> Also register_memory_resource return NULL on failure so we don't have anything
> to cleanup at this point.
>
>
> The reason for this rfc is I was doing some experiments with hotplugging of
> memory on some of our larger systems. While it seems to work, it can be quite
> slow. With some preliminary digging I found that lock_memory_hotplug is
> clearly ripe for breakup.
>
> It could be broken up per nid or something but it also covers the
> online_page_callback. The online_page_callback shouldn't be very hard to break
> out.
>
> Also there is the issue of various structures(wmarks come to mind) that are
> only updated under the lock_memory_hotplug that would need to be dealt with.
>
>
> cc: Andrew Morton <[email protected]>
> cc: Tang Chen <[email protected]>
> cc: Wen Congyang <[email protected]>
> cc: Kamezawa Hiroyuki <[email protected]>
> cc: Yasuaki Ishimatsu <[email protected]>
> cc: "Rafael J. Wysocki" <[email protected]>
> cc: Hedi <[email protected]>
> cc: Mike Travis <[email protected]>
> cc: [email protected]
> cc: [email protected]
>
>
> ---
The patch seems good to me.
Reviewed-by: Yasuaki Ishimatsu <[email protected]>
Thanks,
Yasuaki Ishimatsu
> mm/memory_hotplug.c | 7 ++++---
> 1 file changed, 4 insertions(+), 3 deletions(-)
>
> diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c
> index 1ad92b4..62a0cd1 100644
> --- a/mm/memory_hotplug.c
> +++ b/mm/memory_hotplug.c
> @@ -1097,17 +1097,18 @@ int __ref add_memory(int nid, u64 start, u64 size)
> struct resource *res;
> int ret;
>
> - lock_memory_hotplug();
> -
> res = register_memory_resource(start, size);
> ret = -EEXIST;
> if (!res)
> - goto out;
> + return ret;
>
> { /* Stupid hack to suppress address-never-null warning */
> void *p = NODE_DATA(nid);
> new_pgdat = !p;
> }
> +
> + lock_memory_hotplug();
> +
> new_node = !node_online(nid);
> if (new_node) {
> pgdat = hotadd_new_pgdat(nid, start);
>
On Tue, 14 Jan 2014 17:05:42 -0800 (PST) David Rientjes <[email protected]> wrote:
> On Tue, 14 Jan 2014, Andrew Morton wrote:
>
> > From: Andrew Morton <[email protected]>
> > Subject: mm/memory_hotplug.c: register_memory_resource() fixes
> >
> > - register_memory_resource() should not go BUG on ENOMEM. That's
> > appropriate at system boot time, but not at memory-hotplug time. Fix.
> >
> > - register_memory_resource()'s caller is incorrectly replacing
> > request_resource()'s -EBUSY with -EEXIST. Fix this by propagating
> > errors appropriately.
> >
>
> Unfortunately, -EEXIST is a special case return value for both
> acpi_memory_enable_device() and hv_mem_hot_add(), so they would need to be
> modified to agree concurrently with this change.
blah, OK, thanks, I'll drop it.