[PATCH] fs/select: add vmalloc fallback for select(2)

The select(2) syscall performs a kmalloc(size, GFP_KERNEL) where size grows
with the number of fds passed. We had a customer report page allocation
failures of order-4 for this allocation. This is a costly order, so it might
easily fail, as the VM expects such allocation to have a lower-order fallback.

Such trivial fallback is vmalloc(), as the memory doesn't have to be
physically contiguous. Also the allocation is temporary for the duration of the
syscall, so it's unlikely to stress vmalloc too much.

Note that the poll(2) syscall seems to use a linked list of order-0 pages, so
it doesn't need this kind of fallback.

Signed-off-by: Vlastimil Babka <redacted>
---
 fs/select.c | 15 +++++++++++----
 1 file changed, 11 insertions(+), 4 deletions(-)

diff --git a/fs/select.c b/fs/select.c
index 8ed9da50896a..8fe5bddbe99b 100644
--- a/fs/select.c
+++ b/fs/select.c

@@ -29,6 +29,7 @@
 #include <linux/sched/rt.h>
 #include <linux/freezer.h>
 #include <net/busy_poll.h>
+#include <linux/vmalloc.h>
 
 #include <asm/uaccess.h>
 

@@ -558,6 +559,7 @@ int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
 	struct fdtable *fdt;
 	/* Allocate small arguments on the stack to save memory and be faster */
 	long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
+	unsigned long alloc_size;
 
 	ret = -EINVAL;
 	if (n < 0)

@@ -580,10 +582,15 @@ int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
 	bits = stack_fds;
 	if (size > sizeof(stack_fds) / 6) {
 		/* Not enough space in on-stack array; must use kmalloc */
+		alloc_size = 6 * size;
 		ret = -ENOMEM;
-		bits = kmalloc(6 * size, GFP_KERNEL);
-		if (!bits)
-			goto out_nofds;
+		bits = kmalloc(alloc_size, GFP_KERNEL|__GFP_NOWARN);
+		if (!bits && alloc_size > PAGE_SIZE) {
+			bits = vmalloc(alloc_size);
+
+			if (!bits)
+				goto out_nofds;
+		}
 	}
 	fds.in      = bits;
 	fds.out     = bits +   size;

@@ -618,7 +625,7 @@ int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
 
 out:
 	if (bits != stack_fds)
-		kfree(bits);
+		kvfree(bits);
 out_nofds:
 	return ret;
 }

-- 
2.10.0

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

On Thu, 2016-09-22 at 17:28 +0200, Vlastimil Babka wrote:

Test should happen if alloc_size <= PAGE_SIZE

+		}

if (!bits && alloc_size > PAGE_SIZE)
    bits = vmalloc(alloc_size);

if (!bits)
      goto out_nofds;

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

On 09/22/2016 06:24 PM, Eric Dumazet wrote:

quoted

+		bits = kmalloc(alloc_size, GFP_KERNEL|__GFP_NOWARN);
+		if (!bits && alloc_size > PAGE_SIZE) {
+			bits = vmalloc(alloc_size);
+
+			if (!bits)
+				goto out_nofds;

Test should happen if alloc_size <= PAGE_SIZE

quoted

+		}

if (!bits && alloc_size > PAGE_SIZE)
    bits = vmalloc(alloc_size);

if (!bits)
      goto out_nofds;

Thanks... stupid last-minute changes.

Looks like we also have to bail out if kmalloc fails with 
alloc_size less than PAGE_SIZE.

thanks
Hillf

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

On Fri, 23 Sep 2016 14:42:53 +0800
"Hillf Danton" [off-list ref] wrote:

quoted

The select(2) syscall performs a kmalloc(size, GFP_KERNEL) where size grows
with the number of fds passed. We had a customer report page allocation
failures of order-4 for this allocation. This is a costly order, so it might
easily fail, as the VM expects such allocation to have a lower-order fallback.

Such trivial fallback is vmalloc(), as the memory doesn't have to be
physically contiguous. Also the allocation is temporary for the duration of the
syscall, so it's unlikely to stress vmalloc too much.

Note that the poll(2) syscall seems to use a linked list of order-0 pages, so
it doesn't need this kind of fallback.

How about something like this? (untested)

Eric isn't wrong about vmalloc sucking :)

Thanks,
Nick


---
 fs/select.c | 57 +++++++++++++++++++++++++++++++++++++++++++--------------
 1 file changed, 43 insertions(+), 14 deletions(-)

diff --git a/fs/select.c b/fs/select.c
index 8ed9da5..3b4834c 100644
--- a/fs/select.c
+++ b/fs/select.c

@@ -555,6 +555,7 @@ int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
 	void *bits;
 	int ret, max_fds;
 	unsigned int size;
+	size_t nr_bytes;
 	struct fdtable *fdt;
 	/* Allocate small arguments on the stack to save memory and be faster */
 	long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];

@@ -576,21 +577,39 @@ int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
 	 * since we used fdset we need to allocate memory in units of
 	 * long-words. 
 	 */
-	size = FDS_BYTES(n);
+	ret = -ENOMEM;
 	bits = stack_fds;
-	if (size > sizeof(stack_fds) / 6) {
-		/* Not enough space in on-stack array; must use kmalloc */
+	size = FDS_BYTES(n);
+	nr_bytes = 6 * size;
+
+	if (unlikely(nr_bytes > PAGE_SIZE)) {
+		/* Avoid multi-page allocation if possible */
 		ret = -ENOMEM;
-		bits = kmalloc(6 * size, GFP_KERNEL);
-		if (!bits)
-			goto out_nofds;
+		fds.in = kmalloc(size, GFP_KERNEL);
+		fds.out = kmalloc(size, GFP_KERNEL);
+		fds.ex = kmalloc(size, GFP_KERNEL);
+		fds.res_in = kmalloc(size, GFP_KERNEL);
+		fds.res_out = kmalloc(size, GFP_KERNEL);
+		fds.res_ex = kmalloc(size, GFP_KERNEL);
+
+		if (!(fds.in && fds.out && fds.ex &&
+				fds.res_in && fds.res_out && fds.res_ex))
+			goto out;
+	} else {
+		if (nr_bytes > sizeof(stack_fds)) {
+			/* Not enough space in on-stack array */
+			if (nr_bytes > PAGE_SIZE * 2)
+			bits = kmalloc(nr_bytes, GFP_KERNEL);
+			if (!bits)
+				goto out_nofds;
+		}
+		fds.in      = bits;
+		fds.out     = bits +   size;
+		fds.ex      = bits + 2*size;
+		fds.res_in  = bits + 3*size;
+		fds.res_out = bits + 4*size;
+		fds.res_ex  = bits + 5*size;
 	}
-	fds.in      = bits;
-	fds.out     = bits +   size;
-	fds.ex      = bits + 2*size;
-	fds.res_in  = bits + 3*size;
-	fds.res_out = bits + 4*size;
-	fds.res_ex  = bits + 5*size;
 
 	if ((ret = get_fd_set(n, inp, fds.in)) ||
 	    (ret = get_fd_set(n, outp, fds.out)) ||

@@ -617,8 +636,18 @@ int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
 		ret = -EFAULT;
 
 out:
-	if (bits != stack_fds)
-		kfree(bits);
+	if (unlikely(nr_bytes > PAGE_SIZE)) {
+		kfree(fds.in);
+		kfree(fds.out);
+		kfree(fds.ex);
+		kfree(fds.res_in);
+		kfree(fds.res_out);
+		kfree(fds.res_ex);
+	} else {
+		if (bits != stack_fds)
+			kfree(bits);
+	}
+
 out_nofds:
 	return ret;
 }

-- 
2.9.3

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Hi,

On 09/23/2016 03:24 AM, Nicholas Piggin wrote:

The 'if' looks extraneous?

Also, I wonder if we can just avoid some allocations altogether by checking by if the user fd_set pointers are NULL? That can avoid failures :)

Thanks,

-Jason

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

On 09/23/2016 06:47 PM, Jason Baron wrote:

Hi,

On 09/23/2016 03:24 AM, Nicholas Piggin wrote:

quoted

On Fri, 23 Sep 2016 14:42:53 +0800
"Hillf Danton" [off-list ref] wrote:

quoted

quoted

The select(2) syscall performs a kmalloc(size, GFP_KERNEL) where size grows
with the number of fds passed. We had a customer report page allocation
failures of order-4 for this allocation. This is a costly order, so it might
easily fail, as the VM expects such allocation to have a lower-order fallback.

Such trivial fallback is vmalloc(), as the memory doesn't have to be
physically contiguous. Also the allocation is temporary for the duration of the
syscall, so it's unlikely to stress vmalloc too much.

Note that the poll(2) syscall seems to use a linked list of order-0 pages, so
it doesn't need this kind of fallback.

How about something like this? (untested)

This pushes the limit further, but might just delay the problem. Could be an optimization on top if there's enough interest, though.

[...]

quoted

+
+		if (!(fds.in && fds.out && fds.ex &&
+				fds.res_in && fds.res_out && fds.res_ex))
+			goto out;
+	} else {
+		if (nr_bytes > sizeof(stack_fds)) {
+			/* Not enough space in on-stack array */
+			if (nr_bytes > PAGE_SIZE * 2)

The 'if' looks extraneous?

Also, I wonder if we can just avoid some allocations altogether by
checking by if the user fd_set pointers are NULL? That can avoid failures :)

That would be a more major rewrite, as the core algorithm doesn't expect NULLs.

Thanks,

-Jason

On Tue, 27 Sep 2016 10:44:04 +0200
Vlastimil Babka [off-list ref] wrote:

On 09/23/2016 06:47 PM, Jason Baron wrote:

quoted

Hi,

On 09/23/2016 03:24 AM, Nicholas Piggin wrote:  

quoted

On Fri, 23 Sep 2016 14:42:53 +0800
"Hillf Danton" [off-list ref] wrote:
 

quoted

quoted

The select(2) syscall performs a kmalloc(size, GFP_KERNEL) where size grows
with the number of fds passed. We had a customer report page allocation
failures of order-4 for this allocation. This is a costly order, so it might
easily fail, as the VM expects such allocation to have a lower-order fallback.

Such trivial fallback is vmalloc(), as the memory doesn't have to be
physically contiguous. Also the allocation is temporary for the duration of the
syscall, so it's unlikely to stress vmalloc too much.

Note that the poll(2) syscall seems to use a linked list of order-0 pages, so
it doesn't need this kind of fallback.  

How about something like this? (untested)  

This pushes the limit further, but might just delay the problem. Could be an 
optimization on top if there's enough interest, though.

What's your customer doing with those selects? If they care at all about
performance, I doubt they want select to attempt order-4 allocations, fail,
then use vmalloc :)

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

From: Nicholas Piggin

Sent: 27 September 2016 12:25
On Tue, 27 Sep 2016 10:44:04 +0200
Vlastimil Babka [off-list ref] wrote:

quoted

On 09/23/2016 06:47 PM, Jason Baron wrote:

quoted

Hi,

On 09/23/2016 03:24 AM, Nicholas Piggin wrote:

quoted

On Fri, 23 Sep 2016 14:42:53 +0800
"Hillf Danton" [off-list ref] wrote:

quoted

quoted

The select(2) syscall performs a kmalloc(size, GFP_KERNEL) where size grows
with the number of fds passed. We had a customer report page allocation
failures of order-4 for this allocation. This is a costly order, so it might
easily fail, as the VM expects such allocation to have a lower-order fallback.

Such trivial fallback is vmalloc(), as the memory doesn't have to be
physically contiguous. Also the allocation is temporary for the duration of the
syscall, so it's unlikely to stress vmalloc too much.

Note that the poll(2) syscall seems to use a linked list of order-0 pages, so
it doesn't need this kind of fallback.

How about something like this? (untested)

This pushes the limit further, but might just delay the problem. Could be an
optimization on top if there's enough interest, though.

What's your customer doing with those selects? If they care at all about
performance, I doubt they want select to attempt order-4 allocations, fail,
then use vmalloc :)

If they care about performance they shouldn't be passing select() lists that
are anywhere near that large.
If the number of actual fd is small - use poll().

Otherwise you want one of the 'event' mechanisms in order to avoid setting
the markers on every fd after every event (can't remember how you do that
in Linux).

At least this isn't SYSV - poll() was O(n^2) in the number of fd
(because the fd were on a linked list).

	David

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

On Tue, 27 Sep 2016 11:37:24 +0000
David Laight [off-list ref] wrote:

From: Nicholas Piggin

quoted

Sent: 27 September 2016 12:25
On Tue, 27 Sep 2016 10:44:04 +0200
Vlastimil Babka [off-list ref] wrote:
  

quoted

On 09/23/2016 06:47 PM, Jason Baron wrote:  

quoted

Hi,

On 09/23/2016 03:24 AM, Nicholas Piggin wrote:  

quoted

On Fri, 23 Sep 2016 14:42:53 +0800
"Hillf Danton" [off-list ref] wrote:
 

quoted

quoted

The select(2) syscall performs a kmalloc(size, GFP_KERNEL) where size grows
with the number of fds passed. We had a customer report page allocation
failures of order-4 for this allocation. This is a costly order, so it might
easily fail, as the VM expects such allocation to have a lower-order fallback.

Such trivial fallback is vmalloc(), as the memory doesn't have to be
physically contiguous. Also the allocation is temporary for the duration of the
syscall, so it's unlikely to stress vmalloc too much.

Note that the poll(2) syscall seems to use a linked list of order-0 pages, so
it doesn't need this kind of fallback.  

How about something like this? (untested)  

This pushes the limit further, but might just delay the problem. Could be an
optimization on top if there's enough interest, though.  

What's your customer doing with those selects? If they care at all about
performance, I doubt they want select to attempt order-4 allocations, fail,
then use vmalloc :)  

If they care about performance they shouldn't be passing select() lists that
are anywhere near that large.
If the number of actual fd is small - use poll().

Right. Presumably it's some old app they're still using, no?

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

On 09/27/2016 01:42 PM, Nicholas Piggin wrote:

On Tue, 27 Sep 2016 11:37:24 +0000
David Laight [off-list ref] wrote:

quoted

From: Nicholas Piggin

quoted

Sent: 27 September 2016 12:25
On Tue, 27 Sep 2016 10:44:04 +0200
Vlastimil Babka [off-list ref] wrote:


What's your customer doing with those selects? If they care at all about
performance, I doubt they want select to attempt order-4 allocations, fail,
then use vmalloc :)

If they care about performance they shouldn't be passing select() lists that
are anywhere near that large.
If the number of actual fd is small - use poll().

Right. Presumably it's some old app they're still using, no?

Process name suggests it's part of db2 database. It seems it has to implement its own interface to select() syscall, because glibc itself seems to have a FD_SETSIZE limit of 1024, which is probably why this wasn't an issue for all the years...


--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

From: Vlastimil Babka

Sent: 27 September 2016 12:51

...

Process name suggests it's part of db2 database. It seems it has to implement
its own interface to select() syscall, because glibc itself seems to have a
FD_SETSIZE limit of 1024, which is probably why this wasn't an issue for all the
years...

ISTR the canonical way to increase the size being to set FD_SETSIZE
to a larger value before including any of the headers.

Or doesn't that work with linux and glibc ??

	David

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

On 09/28/2016 06:30 PM, David Laight wrote:

From: Vlastimil Babka

quoted

Sent: 27 September 2016 12:51

...

quoted

Process name suggests it's part of db2 database. It seems it has to implement
its own interface to select() syscall, because glibc itself seems to have a
FD_SETSIZE limit of 1024, which is probably why this wasn't an issue for all the
years...

ISTR the canonical way to increase the size being to set FD_SETSIZE
to a larger value before including any of the headers.

Or doesn't that work with linux and glibc ??

Doesn't seem so.

	David