Skip to contentSkip to conversation

ceph: fix blindly expanding the readahead windows

PATCH Superseded [email protected] 2 participants 6 messages May 15, 2023 – Jun 5, 2023

Subsystems (2)

v6 of 7 Superseded Latest is v7 →

lore
Share & tools
lore
#0 lore raw

v6 cover ceph: fix blindly expanding the readahead windows

From: Xiubo Li <[email protected]>

V6:
- Fixed the ctx issue when reading
- Fixed a potential use-after-free bug

Xiubo Li (2):
ceph: add a dedicated private data for netfs rreq
ceph: fix blindly expanding the readahead windows

 fs/ceph/addr.c  | 85 ++++++++++++++++++++++++++++++++++++++-----------
 fs/ceph/super.h | 13 ++++++++
 2 files changed, 80 insertions(+), 18 deletions(-)
Signature · 1 line
2.40.1
#1 lore raw

PATCH v6 1/2 ceph: add a dedicated private data for netfs rreq

From: Xiubo Li <[email protected]>

We need to save the 'f_ra.ra_pages' to expand the readahead window later.

Fixes 49870056005c ("ceph: convert ceph_readpages to ceph_readahead")

URL: https://lore.kernel.org/ceph-devel/[email protected]
URL: https://www.spinics.net/lists/ceph-users/msg76183.html

---
 fs/ceph/addr.c  | 45 ++++++++++++++++++++++++++++++++++-----------
 fs/ceph/super.h | 13 +++++++++++++
 2 files changed, 47 insertions(+), 11 deletions(-)
fs/ceph/addr.c +34 -11
index 3b20873733af..93fff1a7373f 100644
@@ -404,18 +404,28 @@ static int ceph_init_request(struct netfs_io_request *rreq, struct file *file)
 {
 	struct inode *inode = rreq->inode;
 	int got = 0, want = CEPH_CAP_FILE_CACHE;
+	struct ceph_netfs_request_data *priv;
 	int ret = 0;
 
 	if (rreq->origin != NETFS_READAHEAD)
 		return 0;
 
+	priv = kzalloc(sizeof(*priv), GFP_NOFS);
+	if (!priv)
+		return -ENOMEM;
+
 	if (file) {
 		struct ceph_rw_context *rw_ctx;
 		struct ceph_file_info *fi = file->private_data;
 
+		priv->file_ra_pages = file->f_ra.ra_pages;
+		priv->file_ra_disabled = file->f_mode & FMODE_RANDOM;
+
 		rw_ctx = ceph_find_rw_context(fi);
-		if (rw_ctx)
+		if (rw_ctx) {
+			rreq->netfs_priv = priv;
 			return 0;
+		}
 	}
 
 	/*
@@ -425,27 +435,40 @@ static int ceph_init_request(struct netfs_io_request *rreq, struct file *file)
 	ret = ceph_try_get_caps(inode, CEPH_CAP_FILE_RD, want, true, &got);
 	if (ret < 0) {
 		dout("start_read %p, error getting cap\n", inode);
-		return ret;
+		goto out;
 	}
 
 	if (!(got & want)) {
 		dout("start_read %p, no cache cap\n", inode);
-		return -EACCES;
+		ret = -EACCES;
+		goto out;
+	}
+	if (ret == 0) {
+		ret = -EACCES;
+		goto out;
 	}
-	if (ret == 0)
-		return -EACCES;
 
-	rreq->netfs_priv = (void *)(uintptr_t)got;
-	return 0;
+	priv->caps = got;
+	rreq->netfs_priv = priv;
+
+out:
+	if (ret < 0)
+		kfree(priv);
+
+	return ret;
 }
 
 static void ceph_netfs_free_request(struct netfs_io_request *rreq)
 {
-	struct ceph_inode_info *ci = ceph_inode(rreq->inode);
-	int got = (uintptr_t)rreq->netfs_priv;
+	struct ceph_netfs_request_data *priv = rreq->netfs_priv;
+
+	if (!priv)
+		return;
 
-	if (got)
-		ceph_put_cap_refs(ci, got);
+	if (priv->caps)
+		ceph_put_cap_refs(ceph_inode(rreq->inode), priv->caps);
+	kfree(priv);
+	rreq->netfs_priv = NULL;
 }
 
 const struct netfs_request_ops ceph_netfs_ops = {
fs/ceph/super.h +13 -0
index a226d36b3ecb..3a24b7974d46 100644
@@ -470,6 +470,19 @@ struct ceph_inode_info {
 #endif
 };
 
+struct ceph_netfs_request_data {
+	int caps;
+
+	/*
+	 * Maximum size of a file readahead request.
+	 * The fadvise could update the bdi's default ra_pages.
+	 */
+	unsigned int file_ra_pages;
+
+	/* Set it if fadvise disables file readahead entirely */
+	bool file_ra_disabled;
+};
+
 static inline struct ceph_inode_info *
 ceph_inode(const struct inode *inode)
 {
Signature · 1 line
2.40.1
Reviewed-by Hu Weiwen on v6 1/2
#2 lore raw

Re: [PATCH v6 1/2] ceph: add a dedicated private data for netfs rreq

Quoted 128 lines · On Mon, May 15, 2023 at 09:20:43AM +0800, [email protected] wrote:

From: Xiubo Li <[email protected]>

We need to save the 'f_ra.ra_pages' to expand the readahead window later.

Fixes 49870056005c ("ceph: convert ceph_readpages to ceph_readahead")

URL: https://lore.kernel.org/ceph-devel/[email protected]
URL: https://www.spinics.net/lists/ceph-users/msg76183.html

---
 fs/ceph/addr.c  | 45 ++++++++++++++++++++++++++++++++++-----------
 fs/ceph/super.h | 13 +++++++++++++
 2 files changed, 47 insertions(+), 11 deletions(-)
fs/ceph/addr.c +34 -11
index 3b20873733af..93fff1a7373f 100644
@@ -404,18 +404,28 @@ static int ceph_init_request(struct netfs_io_request *rreq, struct file *file)
 {
 	struct inode *inode = rreq->inode;
 	int got = 0, want = CEPH_CAP_FILE_CACHE;
+	struct ceph_netfs_request_data *priv;
 	int ret = 0;
 
 	if (rreq->origin != NETFS_READAHEAD)
 		return 0;
 
+	priv = kzalloc(sizeof(*priv), GFP_NOFS);
+	if (!priv)
+		return -ENOMEM;
+
 	if (file) {
 		struct ceph_rw_context *rw_ctx;
 		struct ceph_file_info *fi = file->private_data;
 
+		priv->file_ra_pages = file->f_ra.ra_pages;
+		priv->file_ra_disabled = file->f_mode & FMODE_RANDOM;
+
 		rw_ctx = ceph_find_rw_context(fi);
-		if (rw_ctx)
+		if (rw_ctx) {
+			rreq->netfs_priv = priv;
 			return 0;
+		}
 	}
 
 	/*
@@ -425,27 +435,40 @@ static int ceph_init_request(struct netfs_io_request *rreq, struct file *file)
 	ret = ceph_try_get_caps(inode, CEPH_CAP_FILE_RD, want, true, &got);
 	if (ret < 0) {
 		dout("start_read %p, error getting cap\n", inode);
-		return ret;
+		goto out;
 	}
 
 	if (!(got & want)) {
 		dout("start_read %p, no cache cap\n", inode);
-		return -EACCES;
+		ret = -EACCES;
+		goto out;
+	}
+	if (ret == 0) {
+		ret = -EACCES;
+		goto out;
 	}
-	if (ret == 0)
-		return -EACCES;
 
-	rreq->netfs_priv = (void *)(uintptr_t)got;
-	return 0;
+	priv->caps = got;
+	rreq->netfs_priv = priv;
+
+out:
+	if (ret < 0)
+		kfree(priv);
+
+	return ret;
 }
 
 static void ceph_netfs_free_request(struct netfs_io_request *rreq)
 {
-	struct ceph_inode_info *ci = ceph_inode(rreq->inode);
-	int got = (uintptr_t)rreq->netfs_priv;
+	struct ceph_netfs_request_data *priv = rreq->netfs_priv;
+
+	if (!priv)
+		return;
 
-	if (got)
-		ceph_put_cap_refs(ci, got);
+	if (priv->caps)
+		ceph_put_cap_refs(ceph_inode(rreq->inode), priv->caps);
+	kfree(priv);
+	rreq->netfs_priv = NULL;
 }
 
 const struct netfs_request_ops ceph_netfs_ops = {
fs/ceph/super.h +13 -0
index a226d36b3ecb..3a24b7974d46 100644
@@ -470,6 +470,19 @@ struct ceph_inode_info {
 #endif
 };
 
+struct ceph_netfs_request_data {
+	int caps;
+
+	/*
+	 * Maximum size of a file readahead request.
+	 * The fadvise could update the bdi's default ra_pages.
+	 */
+	unsigned int file_ra_pages;
+
+	/* Set it if fadvise disables file readahead entirely */
+	bool file_ra_disabled;
+};
+
 static inline struct ceph_inode_info *
 ceph_inode(const struct inode *inode)
 {
Signature · 1 line
2.40.1

LGTM.

#3 lore raw

PATCH v6 2/2 ceph: fix blindly expanding the readahead windows

In reply to [email protected]

From: Xiubo Li <[email protected]>

Blindly expanding the readahead windows will cause unneccessary pagecache thrashing and also will introdue the network workload. We should disable expanding the windows if the readahead is disabled and also shouldn't expand the windows too much.

Expanding forward firstly instead of expanding backward for possible sequential reads.

Bound rreq->len to the actual file size to restore the previous page cache usage.

The posix_fadvise may change the maximum size of a file readahead.

Fixes 49870056005c ("ceph: convert ceph_readpages to ceph_readahead")

URL: https://lore.kernel.org/ceph-devel/[email protected]
URL: https://www.spinics.net/lists/ceph-users/msg76183.html

---
 fs/ceph/addr.c | 40 +++++++++++++++++++++++++++++++++-------
 1 file changed, 33 insertions(+), 7 deletions(-)
fs/ceph/addr.c +33 -7
index 93fff1a7373f..4b29777c01d7 100644
@@ -188,16 +188,42 @@ static void ceph_netfs_expand_readahead(struct netfs_io_request *rreq)
 	struct inode *inode = rreq->inode;
 	struct ceph_inode_info *ci = ceph_inode(inode);
 	struct ceph_file_layout *lo = &ci->i_layout;
+	unsigned long max_pages = inode->i_sb->s_bdi->ra_pages;
+	loff_t end = rreq->start + rreq->len, new_end;
+	struct ceph_netfs_request_data *priv = rreq->netfs_priv;
+	unsigned long max_len;
 	u32 blockoff;
-	u64 blockno;
 
-	/* Expand the start downward */
-	blockno = div_u64_rem(rreq->start, lo->stripe_unit, &blockoff);
-	rreq->start = blockno * lo->stripe_unit;
-	rreq->len += blockoff;
+	if (priv) {
+		/* Readahead is disabled by posix_fadvise POSIX_FADV_RANDOM */
+		if (priv->file_ra_disabled)
+			max_pages = 0;
+		else
+			max_pages = priv->file_ra_pages;
+
+	}
+
+	/* Readahead is disabled */
+	if (!max_pages)
+		return;
 
-	/* Now, round up the length to the next block */
-	rreq->len = roundup(rreq->len, lo->stripe_unit);
+	max_len = max_pages << PAGE_SHIFT;
+
+	/*
+	 * Try to expand the length forward by rounding  up it to the next
+	 * block, but do not exceed the file size, unless the original
+	 * request already exceeds it.
+	 */
+	new_end = min(round_up(end, lo->stripe_unit), rreq->i_size);
+	if (new_end > end && new_end <= rreq->start + max_len)
+		rreq->len = new_end - rreq->start;
+
+	/* Try to expand the start downward */
+	div_u64_rem(rreq->start, lo->stripe_unit, &blockoff);
+	if (rreq->len + blockoff <= max_len) {
+		rreq->start -= blockoff;
+		rreq->len += blockoff;
+	}
 }
 
 static bool ceph_netfs_clamp_length(struct netfs_io_subrequest *subreq)
Signature · 1 line
2.40.1
Reviewed-by Hu Weiwen Tested-by Hu Weiwen on v6 2/2
#4 lore raw
Quoted 4 lines · On Mon, May 15, 2023 at 09:20:44AM +0800, [email protected] wrote:

From: Xiubo Li <[email protected]>

Blindly expanding the readahead windows will cause unneccessary pagecache thrashing and also will introdue the network workload.

s/introdue/introduce/

Quoted 49 lines

We should disable expanding the windows if the readahead is disabled and also shouldn't expand the windows too much.

Expanding forward firstly instead of expanding backward for possible sequential reads.

Bound rreq->len to the actual file size to restore the previous page cache usage.

The posix_fadvise may change the maximum size of a file readahead.

Fixes 49870056005c ("ceph: convert ceph_readpages to ceph_readahead")

URL: https://lore.kernel.org/ceph-devel/[email protected]
URL: https://www.spinics.net/lists/ceph-users/msg76183.html

---
 fs/ceph/addr.c | 40 +++++++++++++++++++++++++++++++++-------
 1 file changed, 33 insertions(+), 7 deletions(-)
fs/ceph/addr.c +13 -5
index 93fff1a7373f..4b29777c01d7 100644
@@ -188,16 +188,42 @@ static void ceph_netfs_expand_readahead(struct netfs_io_request *rreq)
 	struct inode *inode = rreq->inode;
 	struct ceph_inode_info *ci = ceph_inode(inode);
 	struct ceph_file_layout *lo = &ci->i_layout;
+	unsigned long max_pages = inode->i_sb->s_bdi->ra_pages;
+	loff_t end = rreq->start + rreq->len, new_end;
+	struct ceph_netfs_request_data *priv = rreq->netfs_priv;
+	unsigned long max_len;
 	u32 blockoff;
-	u64 blockno;
 
-	/* Expand the start downward */
-	blockno = div_u64_rem(rreq->start, lo->stripe_unit, &blockoff);
-	rreq->start = blockno * lo->stripe_unit;
-	rreq->len += blockoff;
+	if (priv) {
+		/* Readahead is disabled by posix_fadvise POSIX_FADV_RANDOM */
+		if (priv->file_ra_disabled)
+			max_pages = 0;
+		else
+			max_pages = priv->file_ra_pages;
+
+	}
+

fs/ceph/addr.c static void ceph_netfs_expand_readahead(struct netfs_io_request *rreq) · line 213

Quoted 10 lines
fs/ceph/addr.c +7 -2
+	/* Readahead is disabled */
+	if (!max_pages)
+		return;
 
-	/* Now, round up the length to the next block */
-	rreq->len = roundup(rreq->len, lo->stripe_unit);
+	max_len = max_pages << PAGE_SHIFT;
+
+	/*
+	 * Try to expand the length forward by rounding  up it to the next

An extra space between "rounding up".

Apart from above two typo, LGTM.

I also tested this patch with our workload. Reading the first 16k images from ImageNet dataset (1.69GiB) takes about 1.8Gi page cache (as reported by `free -h'). This is expected.

For the fadvise use-case, I use `fio' to do the test:

$ fio --name=rand --size=32M --fadvise_hint=1 --ioengine=libaio --iodepth=128 --rw=randread --bs=4k --filesize=2G

after the test, page cache increased by about 35Mi, which is expected. So if appropriate:

However, also note random reading to a large file without fadvise still suffers from degradation. e.g., this test:

$ fio --name=rand --size=32M --fadvise_hint=0 --ioengine=libaio --iodepth=128 --rw=randread --bs=4k --filesize=2G

will load nearly every page of the 2Gi test file into page cache, although I only need 32Mi of them.

Quoted 19 lines
fs/ceph/addr.c +13 -0
+	 * block, but do not exceed the file size, unless the original
+	 * request already exceeds it.
+	 */
+	new_end = min(round_up(end, lo->stripe_unit), rreq->i_size);
+	if (new_end > end && new_end <= rreq->start + max_len)
+		rreq->len = new_end - rreq->start;
+
+	/* Try to expand the start downward */
+	div_u64_rem(rreq->start, lo->stripe_unit, &blockoff);
+	if (rreq->len + blockoff <= max_len) {
+		rreq->start -= blockoff;
+		rreq->len += blockoff;
+	}
 }
 
 static bool ceph_netfs_clamp_length(struct netfs_io_subrequest *subreq)
Signature · 1 line
2.40.1
author
#5 lore raw
Quoted 6 lines · 胡玮文
Quoted 4 lines · On Mon, May 15, 2023 at 09:20:44AM +0800, [email protected] wrote:

From: Xiubo Li <[email protected]>

Blindly expanding the readahead windows will cause unneccessary pagecache thrashing and also will introdue the network workload.

s/introdue/introduce/

Will fix it.

Quoted 53 lines
Quoted 53 lines

We should disable expanding the windows if the readahead is disabled and also shouldn't expand the windows too much.

Expanding forward firstly instead of expanding backward for possible sequential reads.

Bound rreq->len to the actual file size to restore the previous page cache usage.

The posix_fadvise may change the maximum size of a file readahead.

Fixes 49870056005c ("ceph: convert ceph_readpages to ceph_readahead")

URL: https://lore.kernel.org/ceph-devel/[email protected]
URL: https://www.spinics.net/lists/ceph-users/msg76183.html

---
  fs/ceph/addr.c | 40 +++++++++++++++++++++++++++++++++-------
  1 file changed, 33 insertions(+), 7 deletions(-)
fs/ceph/addr.c +16 -5
index 93fff1a7373f..4b29777c01d7 100644
@@ -188,16 +188,42 @@ static void ceph_netfs_expand_readahead(struct netfs_io_request *rreq)
  	struct inode *inode = rreq->inode;
  	struct ceph_inode_info *ci = ceph_inode(inode);
  	struct ceph_file_layout *lo = &ci->i_layout;
+	unsigned long max_pages = inode->i_sb->s_bdi->ra_pages;
+	loff_t end = rreq->start + rreq->len, new_end;
+	struct ceph_netfs_request_data *priv = rreq->netfs_priv;
+	unsigned long max_len;
  	u32 blockoff;
-	u64 blockno;
 
-	/* Expand the start downward */
-	blockno = div_u64_rem(rreq->start, lo->stripe_unit, &blockoff);
-	rreq->start = blockno * lo->stripe_unit;
-	rreq->len += blockoff;
+	if (priv) {
+		/* Readahead is disabled by posix_fadvise POSIX_FADV_RANDOM */
+		if (priv->file_ra_disabled)
+			max_pages = 0;
+		else
+			max_pages = priv->file_ra_pages;
+
+	}
+
+	/* Readahead is disabled */
+	if (!max_pages)
+		return;
 
Quoted 7 lines

fs/ceph/addr.c static void ceph_netfs_expand_readahead(struct netfs_io_request *rreq) · line 213

Quoted 6 lines
fs/ceph/addr.c +4 -2
-	/* Now, round up the length to the next block */
-	rreq->len = roundup(rreq->len, lo->stripe_unit);
+	max_len = max_pages << PAGE_SHIFT;
+
+	/*
+	 * Try to expand the length forward by rounding  up it to the next

An extra space between "rounding up".

Will fix it.

Quoted 8 lines

Apart from above two typo, LGTM.

I also tested this patch with our workload. Reading the first 16k images from ImageNet dataset (1.69GiB) takes about 1.8Gi page cache (as reported by `free -h'). This is expected.

Quoted 7 lines

For the fadvise use-case, I use `fio' to do the test:

$ fio --name=rand --size=32M --fadvise_hint=1 --ioengine=libaio --iodepth=128 --rw=randread --bs=4k --filesize=2G

after the test, page cache increased by about 35Mi, which is expected. So if appropriate:

Thanks for your tests and reviewing.

Quoted 6 lines

However, also note random reading to a large file without fadvise still suffers from degradation. e.g., this test:

$ fio --name=rand --size=32M --fadvise_hint=0 --ioengine=libaio --iodepth=128 --rw=randread --bs=4k --filesize=2G

will load nearly every page of the 2Gi test file into page cache, although I only need 32Mi of them.

This is another issue since this patch just to fix blindly expanding readahead windows, please send one following patch to fix it.

Thanks

- Xiubo

Quoted 19 lines
Quoted 19 lines
fs/ceph/addr.c +13 -0
+	 * block, but do not exceed the file size, unless the original
+	 * request already exceeds it.
+	 */
+	new_end = min(round_up(end, lo->stripe_unit), rreq->i_size);
+	if (new_end > end && new_end <= rreq->start + max_len)
+		rreq->len = new_end - rreq->start;
+
+	/* Try to expand the start downward */
+	div_u64_rem(rreq->start, lo->stripe_unit, &blockoff);
+	if (rreq->len + blockoff <= max_len) {
+		rreq->start -= blockoff;
+		rreq->len += blockoff;
+	}
  }
 
  static bool ceph_netfs_clamp_length(struct netfs_io_subrequest *subreq)
Signature · 1 line
2.40.1