Skip to contentSkip to conversation

prevent AoE causing cache aliases

PATCH In review Peter Horton 4 participants 7 messages Oct 22, 2009 – Nov 4, 2009

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 prevent AoE causing cache aliases

This patch prevents the AoE block driver from creating cache aliases of page cache pages on machines with virtually indexed caches.

Building kernels on an AT91SAM9G20 board without this patch fails with segmentation faults after a couple of passes.

drivers/block/aoe/aoecmd.c +21 -2
Index: linux-2.6.31/drivers/block/aoe/aoecmd.c
===================================================================
@@ -735,6 +735,21 @@
 	part_stat_unlock();
 }
 
+/*
+ * Ensure we don't create aliases in VI caches
+ */
+static inline void
+killalias(struct bio *bio)
+{
+	struct bio_vec *bv;
+	int i;
+
+	if (bio_data_dir(bio) == READ)
+		__bio_for_each_segment(bv, bio, i, 0) {
+			flush_dcache_page(bv->bv_page);
+		}
+}
+
 void
 aoecmd_ata_rsp(struct sk_buff *skb)
 {
@@ -853,8 +868,12 @@
 
 	if (buf && --buf->nframesout == 0 && buf->resid == 0) {
 		diskstats(d->gd, buf->bio, jiffies - buf->stime, buf->sector);
-		n = (buf->flags & BUFFL_FAIL) ? -EIO : 0;
-		bio_endio(buf->bio, n);
+		if (buf->flags & BUFFL_FAIL)
+			bio_endio(buf->bio, -EIO);
+		else {
+			killalias(buf->bio);
+			bio_endio(buf->bio, 0);
+		}
 		mempool_free(buf, d->bufpool);
 	}
 
#1 lore raw
Quoted 1 line · On Thu, 22 Oct 2009 15:22:28 +0100 [email protected] (Peter Horton) wrote:

To: [email protected]

Have you heard back from Ed on this?

Quoted 43 lines

Cc: [email protected]
Subject: [PATCH] prevent AoE causing cache aliases
Date: Thu, 22 Oct 2009 15:22:28 +0100
Sender: [email protected]
User-Agent: Mutt/1.5.9i

This patch prevents the AoE block driver from creating cache aliases of page cache pages on machines with virtually indexed caches.

Building kernels on an AT91SAM9G20 board without this patch fails with segmentation faults after a couple of passes.

Index: linux-2.6.31/drivers/block/aoe/aoecmd.c
===================================================================

drivers/block/aoe/aoecmd.c +15 -0
@@ -735,6 +735,21 @@
 	part_stat_unlock();
 }
 
+/*
+ * Ensure we don't create aliases in VI caches
+ */
+static inline void
+killalias(struct bio *bio)
+{
+	struct bio_vec *bv;
+	int i;
+
+	if (bio_data_dir(bio) == READ)
+		__bio_for_each_segment(bv, bio, i, 0) {
+			flush_dcache_page(bv->bv_page);
+		}
+}
+
 void
 aoecmd_ata_rsp(struct sk_buff *skb)
 {
@@ -853,8 +868,12 @@
 
 	if (buf && --buf->nframesout == 0 && buf->resid == 0) {
 		diskstats(d->gd, buf->bio, jiffies - buf->stime, buf->sector);

drivers/block/aoe/aoecmd.c line 878

Quoted 10 lines
drivers/block/aoe/aoecmd.c +6 -2
-		n = (buf->flags & BUFFL_FAIL) ? -EIO : 0;
-		bio_endio(buf->bio, n);
+		if (buf->flags & BUFFL_FAIL)
+			bio_endio(buf->bio, -EIO);
+		else {
+			killalias(buf->bio);
+			bio_endio(buf->bio, 0);
+		}
 		mempool_free(buf, d->bufpool);
 	}

Looks OK.

This bugfix will cause a pointless __bio_for_each_segment() busywait loop to be executed on architectures for which flush_dcache_page() is a no-op.

We don't have infrastructure to fix that.

#2 lore raw
Quoted 7 lines · Andrew Morton
Quoted 1 line · On Thu, 22 Oct 2009 15:22:28 +0100 [email protected] (Peter Horton) wrote:

To: [email protected]

Have you heard back from Ed on this?

No.

Quoted 56 lines
Quoted 43 lines

Cc: [email protected]
Subject: [PATCH] prevent AoE causing cache aliases
Date: Thu, 22 Oct 2009 15:22:28 +0100
Sender: [email protected]
User-Agent: Mutt/1.5.9i

This patch prevents the AoE block driver from creating cache aliases of page cache pages on machines with virtually indexed caches.

Building kernels on an AT91SAM9G20 board without this patch fails with segmentation faults after a couple of passes.

Index: linux-2.6.31/drivers/block/aoe/aoecmd.c
===================================================================

drivers/block/aoe/aoecmd.c +15 -0
@@ -735,6 +735,21 @@
 	part_stat_unlock();
 }
 
+/*
+ * Ensure we don't create aliases in VI caches
+ */
+static inline void
+killalias(struct bio *bio)
+{
+	struct bio_vec *bv;
+	int i;
+
+	if (bio_data_dir(bio) == READ)
+		__bio_for_each_segment(bv, bio, i, 0) {
+			flush_dcache_page(bv->bv_page);
+		}
+}
+
 void
 aoecmd_ata_rsp(struct sk_buff *skb)
 {
@@ -853,8 +868,12 @@
 
 	if (buf && --buf->nframesout == 0 && buf->resid == 0) {
 		diskstats(d->gd, buf->bio, jiffies - buf->stime, buf->sector);

drivers/block/aoe/aoecmd.c line 878

Quoted 10 lines
drivers/block/aoe/aoecmd.c +6 -2
-		n = (buf->flags & BUFFL_FAIL) ? -EIO : 0;
-		bio_endio(buf->bio, n);
+		if (buf->flags & BUFFL_FAIL)
+			bio_endio(buf->bio, -EIO);
+		else {
+			killalias(buf->bio);
+			bio_endio(buf->bio, 0);
+		}
 		mempool_free(buf, d->bufpool);
 	}

Looks OK.

Quoted 5 lines

This bugfix will cause a pointless __bio_for_each_segment() busywait loop to be executed on architectures for which flush_dcache_page() is a no-op.

We don't have infrastructure to fix that.

Couldn't we add a flag to the bio that users could set to indicate that they are not house trained with respect to the D-cache (i.e non-DMA drivers). Architectures that needed to could then flush the relevant pages in the bio_endio() path somewhere. At the moment all the non-DMA block drivers need to be aware of the cache aliasing issue which means this problem keeps arising ...

P.

#3 lore raw
Quoted 71 lines · On Wed, 04 Nov 2009 10:54:34 +0000 Peter Horton <[email protected]> wrote:
Quoted 7 lines · Andrew Morton
Quoted 1 line · On Thu, 22 Oct 2009 15:22:28 +0100 [email protected] (Peter Horton) wrote:

To: [email protected]

Have you heard back from Ed on this?

No.

Quoted 60 lines
Quoted 43 lines

Cc: [email protected]
Subject: [PATCH] prevent AoE causing cache aliases
Date: Thu, 22 Oct 2009 15:22:28 +0100
Sender: [email protected]
User-Agent: Mutt/1.5.9i

This patch prevents the AoE block driver from creating cache aliases of page cache pages on machines with virtually indexed caches.

Building kernels on an AT91SAM9G20 board without this patch fails with segmentation faults after a couple of passes.

Index: linux-2.6.31/drivers/block/aoe/aoecmd.c
===================================================================

drivers/block/aoe/aoecmd.c +15 -0
@@ -735,6 +735,21 @@
 	part_stat_unlock();
 }
 
+/*
+ * Ensure we don't create aliases in VI caches
+ */
+static inline void
+killalias(struct bio *bio)
+{
+	struct bio_vec *bv;
+	int i;
+
+	if (bio_data_dir(bio) == READ)
+		__bio_for_each_segment(bv, bio, i, 0) {
+			flush_dcache_page(bv->bv_page);
+		}
+}
+
 void
 aoecmd_ata_rsp(struct sk_buff *skb)
 {
@@ -853,8 +868,12 @@
 
 	if (buf && --buf->nframesout == 0 && buf->resid == 0) {
 		diskstats(d->gd, buf->bio, jiffies - buf->stime, buf->sector);

drivers/block/aoe/aoecmd.c line 878

Quoted 10 lines
drivers/block/aoe/aoecmd.c +6 -2
-		n = (buf->flags & BUFFL_FAIL) ? -EIO : 0;
-		bio_endio(buf->bio, n);
+		if (buf->flags & BUFFL_FAIL)
+			bio_endio(buf->bio, -EIO);
+		else {
+			killalias(buf->bio);
+			bio_endio(buf->bio, 0);
+		}
 		mempool_free(buf, d->bufpool);
 	}

Looks OK.

This bugfix will cause a pointless __bio_for_each_segment() busywait loop to be executed on architectures for which flush_dcache_page() is a no-op.

Quoted 9 lines
Quoted 1 line

We don't have infrastructure to fix that.

Couldn't we add a flag to the bio that users could set to indicate that they are not house trained with respect to the D-cache (i.e non-DMA drivers). Architectures that needed to could then flush the relevant pages in the bio_endio() path somewhere. At the moment all the non-DMA block drivers need to be aware of the cache aliasing issue which means this problem keeps arising ...

Could. We'll need to change each arch _somehow_. Even if it's a matter of adding `#define i_am_not_house_trained' to the troublesome ones or something, then ifdeffing existing code.

I was thinking that a general bio_flush_dcache_pages() in block core (or in each arch) would be a suitable way to handle this but I was unable to find other drivers which needed it after a brief search.

#4 lore raw
Quoted 74 lines · Andrew Morton
Quoted 64 lines · On Wed, 04 Nov 2009 10:54:34 +0000 Peter Horton <[email protected]> wrote:
Quoted 6 lines · Andrew Morton
Quoted 1 line · On Thu, 22 Oct 2009 15:22:28 +0100 [email protected] (Peter Horton) wrote:

To: [email protected]

Have you heard back from Ed on this?

No.

Quoted 55 lines
Quoted 43 lines

Cc: [email protected]
Subject: [PATCH] prevent AoE causing cache aliases
Date: Thu, 22 Oct 2009 15:22:28 +0100
Sender: [email protected]
User-Agent: Mutt/1.5.9i

This patch prevents the AoE block driver from creating cache aliases of page cache pages on machines with virtually indexed caches.

Building kernels on an AT91SAM9G20 board without this patch fails with segmentation faults after a couple of passes.

Index: linux-2.6.31/drivers/block/aoe/aoecmd.c
===================================================================

drivers/block/aoe/aoecmd.c +15 -0
@@ -735,6 +735,21 @@
 	part_stat_unlock();
 }
 
+/*
+ * Ensure we don't create aliases in VI caches
+ */
+static inline void
+killalias(struct bio *bio)
+{
+	struct bio_vec *bv;
+	int i;
+
+	if (bio_data_dir(bio) == READ)
+		__bio_for_each_segment(bv, bio, i, 0) {
+			flush_dcache_page(bv->bv_page);
+		}
+}
+
 void
 aoecmd_ata_rsp(struct sk_buff *skb)
 {
@@ -853,8 +868,12 @@
 
 	if (buf && --buf->nframesout == 0 && buf->resid == 0) {
 		diskstats(d->gd, buf->bio, jiffies - buf->stime, buf->sector);

drivers/block/aoe/aoecmd.c line 878

Quoted 10 lines
drivers/block/aoe/aoecmd.c +6 -2
-		n = (buf->flags & BUFFL_FAIL) ? -EIO : 0;
-		bio_endio(buf->bio, n);
+		if (buf->flags & BUFFL_FAIL)
+			bio_endio(buf->bio, -EIO);
+		else {
+			killalias(buf->bio);
+			bio_endio(buf->bio, 0);
+		}
 		mempool_free(buf, d->bufpool);
 	}

Looks OK.

Quoted 7 lines
Quoted 5 lines

This bugfix will cause a pointless __bio_for_each_segment() busywait loop to be executed on architectures for which flush_dcache_page() is a no-op.

We don't have infrastructure to fix that.

Couldn't we add a flag to the bio that users could set to indicate that they are not house trained with respect to the D-cache (i.e non-DMA drivers). Architectures that needed to could then flush the relevant pages in the bio_endio() path somewhere. At the moment all the non-DMA block drivers need to be aware of the cache aliasing issue which means this problem keeps arising ...

Quoted 9 lines

Could. We'll need to change each arch _somehow_. Even if it's a matter of adding `#define i_am_not_house_trained' to the troublesome ones or something, then ifdeffing existing code.

I was thinking that a general bio_flush_dcache_pages() in block core (or in each arch) would be a suitable way to handle this but I was unable to find other drivers which needed it after a brief search.

IDE does it at a lower level (arch/mips/include/asm/mach-generic/ide.h for example).

Looks like the generic PIO ops in drivers/ata/libata-sff.c could cause problems too.

I think the problem is often masked by the small cache sizes on the platforms with VI caches. The original AoE problem I see is on an ARM926 with 32K D-cache, I don't see the problem at all on another ARM926 with 16K D-cache.

P.

#5 lore raw
In reply to Andrew Morton
Quoted 83 lines · Andrew Morton
Quoted 71 lines · On Wed, 04 Nov 2009 10:54:34 +0000 Peter Horton <[email protected]> wrote:
Quoted 7 lines · Andrew Morton
Quoted 1 line · On Thu, 22 Oct 2009 15:22:28 +0100 [email protected] (Peter Horton) wrote:

To: [email protected]

Have you heard back from Ed on this?

No.

Quoted 60 lines
Quoted 43 lines

Cc: [email protected]
Subject: [PATCH] prevent AoE causing cache aliases
Date: Thu, 22 Oct 2009 15:22:28 +0100
Sender: [email protected]
User-Agent: Mutt/1.5.9i

This patch prevents the AoE block driver from creating cache aliases of page cache pages on machines with virtually indexed caches.

Building kernels on an AT91SAM9G20 board without this patch fails with segmentation faults after a couple of passes.

Index: linux-2.6.31/drivers/block/aoe/aoecmd.c
===================================================================

drivers/block/aoe/aoecmd.c +15 -0
@@ -735,6 +735,21 @@
 	part_stat_unlock();
 }
 
+/*
+ * Ensure we don't create aliases in VI caches
+ */
+static inline void
+killalias(struct bio *bio)
+{
+	struct bio_vec *bv;
+	int i;
+
+	if (bio_data_dir(bio) == READ)
+		__bio_for_each_segment(bv, bio, i, 0) {
+			flush_dcache_page(bv->bv_page);
+		}
+}
+
 void
 aoecmd_ata_rsp(struct sk_buff *skb)
 {
@@ -853,8 +868,12 @@
 
 	if (buf && --buf->nframesout == 0 && buf->resid == 0) {
 		diskstats(d->gd, buf->bio, jiffies - buf->stime, buf->sector);

drivers/block/aoe/aoecmd.c line 878

Quoted 10 lines
drivers/block/aoe/aoecmd.c +6 -2
-		n = (buf->flags & BUFFL_FAIL) ? -EIO : 0;
-		bio_endio(buf->bio, n);
+		if (buf->flags & BUFFL_FAIL)
+			bio_endio(buf->bio, -EIO);
+		else {
+			killalias(buf->bio);
+			bio_endio(buf->bio, 0);
+		}
 		mempool_free(buf, d->bufpool);
 	}

Looks OK.

This bugfix will cause a pointless __bio_for_each_segment() busywait loop to be executed on architectures for which flush_dcache_page() is a no-op.

Quoted 9 lines
Quoted 1 line

We don't have infrastructure to fix that.

Couldn't we add a flag to the bio that users could set to indicate that they are not house trained with respect to the D-cache (i.e non-DMA drivers). Architectures that needed to could then flush the relevant pages in the bio_endio() path somewhere. At the moment all the non-DMA block drivers need to be aware of the cache aliasing issue which means this problem keeps arising ...

Quoted 7 lines

Could. We'll need to change each arch _somehow_. Even if it's a matter of adding `#define i_am_not_house_trained' to the troublesome ones or something, then ifdeffing existing code.

I was thinking that a general bio_flush_dcache_pages() in block core (or in each arch) would be a suitable way to handle this but I was unable to find other drivers which needed it after a brief search.

Indeed, we should have such a helper. I can't find any ARCH define that tells us when we need to do this. Easiest is probably to grep in arch/ for non-empty definitions of flush_dcache_page() and add such a define.

I'll hack one up.

Signature · 1 line
Jens Axboe
#6 lore raw
In reply to Andrew Morton
Quoted 6 lines · Andrew Morton
Quoted 1 line · On Thu, 22 Oct 2009 15:22:28 +0100 [email protected] (Peter Horton) wrote:

To: [email protected]

Have you heard back from Ed on this?

Sorry, I didn't comment because I don't have much experience with virtually indexed caches, and while the fix seems to make sense, I was hoping to learn more from the discussion to follow.

I do think that it would be helpful to spell out "virtually indexed" in the comment. "VI" might not be enough to jog the memory of folks who haven't used such an architecture recently.

Signature · 1 line
  Ed