Skip to contentSkip to conversation

block/ps3: Fix slow VRAM IO

PATCH In review Geoff Levand 6 participants 7 messages Oct 19, 2009 – Nov 28, 2009

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 block/ps3: Fix slow VRAM IO

From: Hideyuki Sasaki <[email protected]>

The current PS3 VRAM driver uses msleep() to wait for completion of RSX DMA transfers between system memory and VRAM. Depending on the system timing, the processing delay and overhead of this msleep() call can significantly impact VRAM driver IO.

To avoid the condition, add a short duration (200 usec max) udelay() polling loop before entering the msleep() polling loop.

---

 drivers/block/ps3vram.c |   10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)
drivers/block/ps3vram.c +9 -1
@@ -123,7 +123,15 @@ static int ps3vram_notifier_wait(struct 
 {
 	struct ps3vram_priv *priv = ps3_system_bus_get_drvdata(dev);
 	u32 *notify = ps3vram_get_notifier(priv->reports, NOTIFIER);
-	unsigned long timeout = jiffies + msecs_to_jiffies(timeout_ms);
+	unsigned long timeout;
+
+	for (timeout = 20; timeout; timeout--) {
+		if (!notify[3])
+			return 0;
+		udelay(10);
+	}
+
+	timeout = jiffies + msecs_to_jiffies(timeout_ms);
 
 	do {
 		if (!notify[3])
maintainer Acked-by on v1 1/1
#1 lore raw
Quoted 8 lines · Geoff Levand

From: Hideyuki Sasaki <[email protected]>

The current PS3 VRAM driver uses msleep() to wait for completion of RSX DMA transfers between system memory and VRAM. Depending on the system timing, the processing delay and overhead of this msleep() call can significantly impact VRAM driver IO.

Quoted 6 lines

To avoid the condition, add a short duration (200 usec max) udelay() polling loop before entering the msleep() polling loop.

Thanks for tracking this down.

-jim

Quoted 24 lines

---

 drivers/block/ps3vram.c |   10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)
drivers/block/ps3vram.c +9 -1
@@ -123,7 +123,15 @@ static int ps3vram_notifier_wait(struct 
 {
 	struct ps3vram_priv *priv = ps3_system_bus_get_drvdata(dev);
 	u32 *notify = ps3vram_get_notifier(priv->reports, NOTIFIER);
-	unsigned long timeout = jiffies + msecs_to_jiffies(timeout_ms);
+	unsigned long timeout;
+
+	for (timeout = 20; timeout; timeout--) {
+		if (!notify[3])
+			return 0;
+		udelay(10);
+	}
+
+	timeout = jiffies + msecs_to_jiffies(timeout_ms);
 
 	do {
 		if (!notify[3])
#2 lore raw
In reply to Geoff Levand
Quoted 12 lines · On Mon, 19 Oct 2009 12:58:27 -0700 Geoff Levand <[email protected]>…

From: Hideyuki Sasaki <[email protected]>

The current PS3 VRAM driver uses msleep() to wait for completion of RSX DMA transfers between system memory and VRAM. Depending on the system timing, the processing delay and overhead of this msleep() call can significantly impact VRAM driver IO.

To avoid the condition, add a short duration (200 usec max) udelay() polling loop before entering the msleep() polling loop.

When raising a performance-based patch, please always try to include before-and-after performance measurements in the changelog. People want to know the magnitude of the improvement.

Quoted 7 lines
 drivers/block/ps3vram.c |   10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

drivers/block/ps3vram.c static int ps3vram_notifier_wait(struct · line 128

Quoted 7 lines
drivers/block/ps3vram.c +3 -1
@@ -123,7 +123,15 @@ static int ps3vram_notifier_wait(struct 
 {
 	struct ps3vram_priv *priv = ps3_system_bus_get_drvdata(dev);
 	u32 *notify = ps3vram_get_notifier(priv->reports, NOTIFIER);
-	unsigned long timeout = jiffies + msecs_to_jiffies(timeout_ms);
+	unsigned long timeout;
+
+	for (timeout = 20; timeout; timeout--) {
	for (timeout = 0; timeout < 20; timeout++) {

would be simpler.

drivers/block/ps3vram.c static int ps3vram_notifier_wait(struct · line 132

Quoted 4 lines
drivers/block/ps3vram.c +4 -0
+		if (!notify[3])
+			return 0;
+		udelay(10);
+	}

You might as well do a udelay(1) here. The additional cost will be negligible, and it will reduce latency.

drivers/block/ps3vram.c static int ps3vram_notifier_wait(struct · line 134

Quoted 1 line
drivers/block/ps3vram.c +1 -0
+	timeout = jiffies + msecs_to_jiffies(timeout_ms);

The maximum latency is now timout_ms + 200usec.

That's OK with the current constants, but if someone later changes a constant, the error could become significant.

Perhaps that isn't worth bothering about though.

Quoted 2 lines
drivers/block/ps3vram.c +0 -0
 	do {
 		if (!notify[3])
#3 lore raw

Thank you for the review!

Quoted 5 lines
Quoted 5 lines

The current PS3 VRAM driver uses msleep() to wait for completion of RSX DMA transfers between system memory and VRAM. Depending on the system timing, the processing delay and overhead of this msleep() call can significantly impact VRAM driver IO.

Quoted 8 lines
Quoted 4 lines

To avoid the condition, add a short duration (200 usec max) udelay() polling loop before entering the msleep() polling loop.

When raising a performance-based patch, please always try to include before-and-after performance measurements in the changelog. People want to know the magnitude of the improvement.

No problem we will add the difference of improvement in the changelog. This is the results. Pretty impressive.
Before
Reading: 33MB/s
Writing: 16MB/s
After
Reading: 370MB/s
Writing: 238MB/s

Quoted 7 lines

drivers/block/ps3vram.c static int ps3vram_notifier_wait(struct · line 132

Quoted 4 lines
drivers/block/ps3vram.c +4 -0
+		if (!notify[3])
+			return 0;
+		udelay(10);
+	}

You might as well do a udelay(1) here. The additional cost will be negligible, and it will reduce latency.

Are you mentioning adding udelay(1) in the between udelay polling and msleep polling? Or are you mentioning to change udelay(10) to udelay(1) inside the udelay polling?

The former is no problem, but the later has impact on performance of PS3 system.
Because Cell/B.E.(consists of PPE and SPEs cores) and GPU are connected with ring bus called EIB and every issuing notify[3] to check VRAM-DMA results will generate data transfer to the bus.
There are only one EIB bus in PS3 and other devices connected on the bus such as SPEs will be affected if the bus is occupied by many notify[3] and as a result it will decrease the over all system performance.

The udelay(10) was the most reasonable distance not to overcrowd the bus and not to wait too long for checking DMA on VRAM.
We have tried udelay(5) but did not improve the VRAM IO speed.

Quoted 6 lines

drivers/block/ps3vram.c static int ps3vram_notifier_wait(struct · line 134

Quoted 1 line
drivers/block/ps3vram.c +1 -0
+	timeout = jiffies + msecs_to_jiffies(timeout_ms);

The maximum latency is now timout_ms + 200usec.

That's OK with the current constants, but if someone later changes a constant, the error could become significant.

Yes, I think so too. Probably reconstructing the design entirely based on usec instead of msec might be ideal but adding 200usec loops fixes the current slow VRAM driver, so I thought it is acceptable work around.

Quoted 4 lines

Perhaps that isn't worth bothering about though.

Quoted 2 lines
drivers/block/ps3vram.c +0 -0
 	do {
 		if (!notify[3])
Signature · 4 lines
Akira Tsukamoto
Sony Computer Entertainment Inc. 
Architecture Lab.
Japan
#4 lore raw

Hello Andrew Morton,

Ping?

This patch is pretty important to improve the performance of PS3. I really appreciate for your reply.

Thanks,

Akira

On Mon, 09 Nov 2009 15:40:42 +0900,
Akira Tsukamoto <[email protected]> mentioned:

Quoted 77 lines

Thank you for the review!

Quoted 5 lines
Quoted 5 lines

The current PS3 VRAM driver uses msleep() to wait for completion of RSX DMA transfers between system memory and VRAM. Depending on the system timing, the processing delay and overhead of this msleep() call can significantly impact VRAM driver IO.

Quoted 8 lines
Quoted 4 lines

To avoid the condition, add a short duration (200 usec max) udelay() polling loop before entering the msleep() polling loop.

When raising a performance-based patch, please always try to include before-and-after performance measurements in the changelog. People want to know the magnitude of the improvement.

No problem we will add the difference of improvement in the changelog. This is the results. Pretty impressive.
Before
Reading: 33MB/s
Writing: 16MB/s
After
Reading: 370MB/s
Writing: 238MB/s

Quoted 7 lines

drivers/block/ps3vram.c static int ps3vram_notifier_wait(struct · line 132

Quoted 4 lines
drivers/block/ps3vram.c +4 -0
+		if (!notify[3])
+			return 0;
+		udelay(10);
+	}

You might as well do a udelay(1) here. The additional cost will be negligible, and it will reduce latency.

Are you mentioning adding udelay(1) in the between udelay polling and msleep polling? Or are you mentioning to change udelay(10) to udelay(1) inside the udelay polling?

The former is no problem, but the later has impact on performance of PS3 system.
Because Cell/B.E.(consists of PPE and SPEs cores) and GPU are connected with ring bus called EIB and every issuing notify[3] to check VRAM-DMA results will generate data transfer to the bus.
There are only one EIB bus in PS3 and other devices connected on the bus such as SPEs will be affected if the bus is occupied by many notify[3] and as a result it will decrease the over all system performance.

The udelay(10) was the most reasonable distance not to overcrowd the bus and not to wait too long for checking DMA on VRAM.
We have tried udelay(5) but did not improve the VRAM IO speed.

Quoted 6 lines

drivers/block/ps3vram.c static int ps3vram_notifier_wait(struct · line 134

Quoted 1 line
drivers/block/ps3vram.c +1 -0
+	timeout = jiffies + msecs_to_jiffies(timeout_ms);

The maximum latency is now timout_ms + 200usec.

That's OK with the current constants, but if someone later changes a constant, the error could become significant.

Yes, I think so too. Probably reconstructing the design entirely based on usec instead of msec might be ideal but adding 200usec loops fixes the current slow VRAM driver, so I thought it is acceptable work around.

Quoted 4 lines

Perhaps that isn't worth bothering about though.

Quoted 2 lines
drivers/block/ps3vram.c +0 -0
 	do {
 		if (!notify[3])
Signature · 10 lines
Akira Tsukamoto
Sony Computer Entertainment Inc. 
Architecture Lab.
Japan

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/
Signature · 4 lines
Akira Tsukamoto
Sony Computer Entertainment Inc. 
Architecture Lab.
Japan
#5 lore raw
Quoted 6 lines · Akira Tsukamoto

Hello Andrew Morton,

Ping?

This patch is pretty important to improve the performance of PS3. I really appreciate for your reply.

I queued it up for 2.6.33 some time ago, so it's not lost.

Signature · 1 line
Jens Axboe
#6 lore raw
In reply to Akira Tsukamoto
Quoted 53 lines · Akira Tsukamoto

Thank you for the review!

Quoted 5 lines
Quoted 5 lines

The current PS3 VRAM driver uses msleep() to wait for completion of RSX DMA transfers between system memory and VRAM. Depending on the system timing, the processing delay and overhead of this msleep() call can significantly impact VRAM driver IO.

Quoted 7 lines
Quoted 3 lines

To avoid the condition, add a short duration (200 usec max) udelay() polling loop before entering the msleep() polling loop.

When raising a performance-based patch, please always try to include before-and-after performance measurements in the changelog. People want to know the magnitude of the improvement.

No problem we will add the difference of improvement in the changelog. This is the results. Pretty impressive.
Before
Reading: 33MB/s
Writing: 16MB/s
After
Reading: 370MB/s
Writing: 238MB/s

Quoted 7 lines

drivers/block/ps3vram.c static int ps3vram_notifier_wait(struct · line 132

Quoted 4 lines
drivers/block/ps3vram.c +4 -0
+		if (!notify[3])
+			return 0;
+		udelay(10);
+	}

You might as well do a udelay(1) here. The additional cost will be negligible, and it will reduce latency.

Are you mentioning adding udelay(1) in the between udelay polling and msleep polling? Or are you mentioning to change udelay(10) to udelay(1) inside the udelay polling?

The former is no problem, but the later has impact on performance of PS3 system.
Because Cell/B.E.(consists of PPE and SPEs cores) and GPU are connected with ring bus called EIB and every issuing notify[3] to check VRAM-DMA results will generate data transfer to the bus.
There are only one EIB bus in PS3 and other devices connected on the bus such as SPEs will be affected if the bus is occupied by many notify[3] and as a result it will decrease the over all system performance.

The udelay(10) was the most reasonable distance not to overcrowd the bus and not to wait too long for checking DMA on VRAM.
We have tried udelay(5) but did not improve the VRAM IO speed.

Quoted 4 lines

drivers/block/ps3vram.c static int ps3vram_notifier_wait(struct · line 134

Quoted 1 line
drivers/block/ps3vram.c +1 -0
+	timeout = jiffies + msecs_to_jiffies(timeout_ms);

The maximum latency is now timout_ms + 200usec.

Quoted 6 lines
Quoted 2 lines

That's OK with the current constants, but if someone later changes a constant, the error could become significant.

Yes, I think so too. Probably reconstructing the design entirely based on usec instead of msec might be ideal but adding 200usec loops fixes the current slow VRAM driver, so I thought it is acceptable work around.

Thanks for the detailed explanations. I wonder if it makes sense to change 200usec magic number to something more flexible. If I understand it correctly, 200usec is just about twice the time that is needed to transfer 256KiB sized ps3vram internal cache page from or to VRAM via DMA with otherwise idle EIB bus. I guess it is done so that msleep is only ever reachable when EIB bus is heavily overloaded. Reaching msleep in the code means getting all the same 33MB/s or 16MB/s for ps3vram performance.

If somebody tries to play with tweaking ps3vram constants like CACHE_PAGE_SIZE, the magic 200usec delay may need to be changed to something more appropriate, but right now it is not very obvious from the patch description or comments in the code.

So probably some constant, based on DMA throughput and cache page size would be better here?

Signature · 2 lines
Best regards,
Siarhei Siamashka