[PATCH net] netvsc: fix calculation of available send sections

Subsystems: hyper-v/azure core and drivers, networking drivers, the rest

STALE3390d

3 messages, 3 authors, 2017-04-25 · open the first message on its own page

[PATCH net] netvsc: fix calculation of available send sections

From: Stephen Hemminger <stephen@networkplumber.org>
Date: 2017-04-25 01:33:47

My change (introduced in 4.11) to use find_first_clear_bit
incorrectly assumed that the size argument was words, not bits.
The effect was only a small limited number of the available send
sections were being actually used. This can cause performance loss
with some workloads.

Since map_words is now used only during initialization, it can
be on stack instead of in per-device data.

Fixes: b58a185801da ("netvsc: simplify get next send section")
Signed-off-by: Stephen Hemminger <redacted>
---
 drivers/net/hyperv/hyperv_net.h | 1 -
 drivers/net/hyperv/netvsc.c     | 9 ++++-----
 2 files changed, 4 insertions(+), 6 deletions(-)
diff --git a/drivers/net/hyperv/hyperv_net.h b/drivers/net/hyperv/hyperv_net.h
index f9f3dba7a588..db23cb36ae5c 100644
--- a/drivers/net/hyperv/hyperv_net.h
+++ b/drivers/net/hyperv/hyperv_net.h
@@ -751,7 +751,6 @@ struct netvsc_device {
 	u32 send_section_cnt;
 	u32 send_section_size;
 	unsigned long *send_section_map;
-	int map_words;
 
 	/* Used for NetVSP initialization protocol */
 	struct completion channel_init_wait;
diff --git a/drivers/net/hyperv/netvsc.c b/drivers/net/hyperv/netvsc.c
index 8dd0b8770328..15ef713d96c0 100644
--- a/drivers/net/hyperv/netvsc.c
+++ b/drivers/net/hyperv/netvsc.c
@@ -236,6 +236,7 @@ static int netvsc_init_buf(struct hv_device *device)
 	struct netvsc_device *net_device;
 	struct nvsp_message *init_packet;
 	struct net_device *ndev;
+	size_t map_words;
 	int node;
 
 	net_device = get_outbound_net_device(device);
@@ -401,11 +402,9 @@ static int netvsc_init_buf(struct hv_device *device)
 		   net_device->send_section_size, net_device->send_section_cnt);
 
 	/* Setup state for managing the send buffer. */
-	net_device->map_words = DIV_ROUND_UP(net_device->send_section_cnt,
-					     BITS_PER_LONG);
+	map_words = DIV_ROUND_UP(net_device->send_section_cnt, BITS_PER_LONG);
 
-	net_device->send_section_map = kcalloc(net_device->map_words,
-					       sizeof(ulong), GFP_KERNEL);
+	net_device->send_section_map = kcalloc(map_words, sizeof(ulong), GFP_KERNEL);
 	if (net_device->send_section_map == NULL) {
 		ret = -ENOMEM;
 		goto cleanup;
@@ -683,7 +682,7 @@ static u32 netvsc_get_next_send_section(struct netvsc_device *net_device)
 	unsigned long *map_addr = net_device->send_section_map;
 	unsigned int i;
 
-	for_each_clear_bit(i, map_addr, net_device->map_words) {
+	for_each_clear_bit(i, map_addr, net_device->send_section_cnt) {
 		if (sync_test_and_set_bit(i, map_addr) == 0)
 			return i;
 	}
-- 
2.11.0

Re: [PATCH net] netvsc: fix calculation of available send sections

From: Greg Rose <hidden>
Date: 2017-04-25 03:21:02

On Mon, 2017-04-24 at 18:33 -0700, Stephen Hemminger wrote:
My change (introduced in 4.11) to use find_first_clear_bit
incorrectly assumed that the size argument was words, not bits.
Oops...
The effect was only a small limited number of the available send
sections were being actually used. This can cause performance loss
with some workloads.

Since map_words is now used only during initialization, it can
be on stack instead of in per-device data.

Fixes: b58a185801da ("netvsc: simplify get next send section")
Signed-off-by: Stephen Hemminger <redacted>
Looks good to me.

Reviewed by Greg Rose [off-list ref]
quoted hunk
---
 drivers/net/hyperv/hyperv_net.h | 1 -
 drivers/net/hyperv/netvsc.c     | 9 ++++-----
 2 files changed, 4 insertions(+), 6 deletions(-)
diff --git a/drivers/net/hyperv/hyperv_net.h b/drivers/net/hyperv/hyperv_net.h
index f9f3dba7a588..db23cb36ae5c 100644
--- a/drivers/net/hyperv/hyperv_net.h
+++ b/drivers/net/hyperv/hyperv_net.h
@@ -751,7 +751,6 @@ struct netvsc_device {
 	u32 send_section_cnt;
 	u32 send_section_size;
 	unsigned long *send_section_map;
-	int map_words;
 
 	/* Used for NetVSP initialization protocol */
 	struct completion channel_init_wait;
diff --git a/drivers/net/hyperv/netvsc.c b/drivers/net/hyperv/netvsc.c
index 8dd0b8770328..15ef713d96c0 100644
--- a/drivers/net/hyperv/netvsc.c
+++ b/drivers/net/hyperv/netvsc.c
@@ -236,6 +236,7 @@ static int netvsc_init_buf(struct hv_device *device)
 	struct netvsc_device *net_device;
 	struct nvsp_message *init_packet;
 	struct net_device *ndev;
+	size_t map_words;
 	int node;
 
 	net_device = get_outbound_net_device(device);
@@ -401,11 +402,9 @@ static int netvsc_init_buf(struct hv_device *device)
 		   net_device->send_section_size, net_device->send_section_cnt);
 
 	/* Setup state for managing the send buffer. */
-	net_device->map_words = DIV_ROUND_UP(net_device->send_section_cnt,
-					     BITS_PER_LONG);
+	map_words = DIV_ROUND_UP(net_device->send_section_cnt, BITS_PER_LONG);
 
-	net_device->send_section_map = kcalloc(net_device->map_words,
-					       sizeof(ulong), GFP_KERNEL);
+	net_device->send_section_map = kcalloc(map_words, sizeof(ulong), GFP_KERNEL);
 	if (net_device->send_section_map == NULL) {
 		ret = -ENOMEM;
 		goto cleanup;
@@ -683,7 +682,7 @@ static u32 netvsc_get_next_send_section(struct netvsc_device *net_device)
 	unsigned long *map_addr = net_device->send_section_map;
 	unsigned int i;
 
-	for_each_clear_bit(i, map_addr, net_device->map_words) {
+	for_each_clear_bit(i, map_addr, net_device->send_section_cnt) {
 		if (sync_test_and_set_bit(i, map_addr) == 0)
 			return i;
 	}

Re: [PATCH net] netvsc: fix calculation of available send sections

From: David Miller <davem@davemloft.net>
Date: 2017-04-25 15:57:22

From: Stephen Hemminger <stephen@networkplumber.org>
Date: Mon, 24 Apr 2017 18:33:38 -0700
My change (introduced in 4.11) to use find_first_clear_bit
incorrectly assumed that the size argument was words, not bits.
The effect was only a small limited number of the available send
sections were being actually used. This can cause performance loss
with some workloads.

Since map_words is now used only during initialization, it can
be on stack instead of in per-device data.

Fixes: b58a185801da ("netvsc: simplify get next send section")
Signed-off-by: Stephen Hemminger <redacted>
Applied, thanks.
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help