Skip to contentSkip to conversation

ext4: disable mount with both dioread_nolock and nodelalloc

PATCH In review Xiaoguang Wang 3 participants 5 messages Jul 31, 2019 – Sep 9, 2019 ext4 file system

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 ext4: disable mount with both dioread_nolock and nodelalloc

Mount with both dioread_nolock and nodelalloc will result in huge performance drop, which indeed is an known issue, so before we fix this issue, currently we disable this behaviour. Below test reproducer can reveal this performance drop.

    mount -o remount,dioread_nolock,delalloc /dev/vdb1
    rm -f testfile
    start_time=$(date +%s)
    dd if=/dev/zero of=testfile bs=4096 count=$((1024*256))
    sync
    end_time=$(date +%s)
    echo $((end_time - start_time))

    mount -o remount,dioread_nolock,nodelalloc /dev/vdb1
    rm -f testfile
    start_time=$(date +%s)
    dd if=/dev/zero of=testfile bs=4096 count=$((1024*256))
    sync
    end_time=$(date +%s)
    echo $((end_time - start_time))
---
 fs/ext4/super.c | 6 ++++++
 1 file changed, 6 insertions(+)
fs/ext4/super.c +6 -0
index 4079605d437a..1a2b2c0cd1b8 100644
@@ -2098,6 +2098,12 @@ static int parse_options(char *options, struct super_block *sb,
 		int blocksize =
 			BLOCK_SIZE << le32_to_cpu(sbi->s_es->s_log_block_size);
 
+		if (!test_opt(sb, DELALLOC)) {
+			ext4_msg(sb, KERN_ERR, "can't mount with "
+				 "both dioread_nolock and nodelalloc");
+			return 0;
+		}
+
 		if (blocksize < PAGE_SIZE) {
 			ext4_msg(sb, KERN_ERR, "can't mount with "
 				 "dioread_nolock if block size != PAGE_SIZE");
Signature · 1 line
2.17.2
Reviewed-by on v1 1/1
#1 lore raw
Quoted 31 lines · Xiaoguang Wang

Mount with both dioread_nolock and nodelalloc will result in huge performance drop, which indeed is an known issue, so before we fix this issue, currently we disable this behaviour. Below test reproducer can reveal this performance drop.

    mount -o remount,dioread_nolock,delalloc /dev/vdb1
    rm -f testfile
    start_time=$(date +%s)
    dd if=/dev/zero of=testfile bs=4096 count=$((1024*256))
    sync
    end_time=$(date +%s)
    echo $((end_time - start_time))

    mount -o remount,dioread_nolock,nodelalloc /dev/vdb1
    rm -f testfile
    start_time=$(date +%s)
    dd if=/dev/zero of=testfile bs=4096 count=$((1024*256))
    sync
    end_time=$(date +%s)
    echo $((end_time - start_time))
---
 fs/ext4/super.c | 6 ++++++
 1 file changed, 6 insertions(+)
fs/ext4/super.c +0 -0
index 4079605d437a..1a2b2c0cd1b8 100644

fs/ext4/super.c static int parse_options(char *options, struct super_block *sb, · line 2106

Quoted 9 lines
fs/ext4/super.c +6 -0
@@ -2098,6 +2098,12 @@ static int parse_options(char *options, struct super_block *sb,
 		int blocksize =
 			BLOCK_SIZE << le32_to_cpu(sbi->s_es->s_log_block_size);
 
+		if (!test_opt(sb, DELALLOC)) {
+			ext4_msg(sb, KERN_ERR, "can't mount with "
+				 "both dioread_nolock and nodelalloc");
+			return 0;
+		}
+

I suggest move it down to keep blocksize check logic together.

Other than that, looks good to me.

Quoted 4 lines
fs/ext4/super.c +0 -0
 		if (blocksize < PAGE_SIZE) {
 			ext4_msg(sb, KERN_ERR, "can't mount with "
 				 "dioread_nolock if block size != PAGE_SIZE");
#2 lore raw
In reply to Xiaoguang Wang

hi,

Ping.
For this patch, I'm afraid someone else still takes effort to look into this issue.

Regards,
Xiaoguang Wang

Quoted 44 lines

Mount with both dioread_nolock and nodelalloc will result in huge performance drop, which indeed is an known issue, so before we fix this issue, currently we disable this behaviour. Below test reproducer can reveal this performance drop.

     mount -o remount,dioread_nolock,delalloc /dev/vdb1
     rm -f testfile
     start_time=$(date +%s)
     dd if=/dev/zero of=testfile bs=4096 count=$((1024*256))
     sync
     end_time=$(date +%s)
     echo $((end_time - start_time))

     mount -o remount,dioread_nolock,nodelalloc /dev/vdb1
     rm -f testfile
     start_time=$(date +%s)
     dd if=/dev/zero of=testfile bs=4096 count=$((1024*256))
     sync
     end_time=$(date +%s)
     echo $((end_time - start_time))
---
  fs/ext4/super.c | 6 ++++++
  1 file changed, 6 insertions(+)
fs/ext4/super.c +6 -0
index 4079605d437a..1a2b2c0cd1b8 100644
@@ -2098,6 +2098,12 @@ static int parse_options(char *options, struct super_block *sb,
  		int blocksize =
  			BLOCK_SIZE << le32_to_cpu(sbi->s_es->s_log_block_size);
 
+		if (!test_opt(sb, DELALLOC)) {
+			ext4_msg(sb, KERN_ERR, "can't mount with "
+				 "both dioread_nolock and nodelalloc");
+			return 0;
+		}
+
  		if (blocksize < PAGE_SIZE) {
  			ext4_msg(sb, KERN_ERR, "can't mount with "
  				 "dioread_nolock if block size != PAGE_SIZE");
maintainer
#3 lore raw
In reply to Xiaoguang Wang
Quoted 4 lines · Xiaoguang Wang

Mount with both dioread_nolock and nodelalloc will result in huge performance drop, which indeed is an known issue, so before we fix this issue, currently we disable this behaviour. Below test reproducer can reveal this performance drop.

Is it really worth it to disable this combination? Nothing goes *wrong* per se; it's just slower than we would like.

- Ted

#4 lore raw

hi,

Quoted 8 lines · Xiaoguang Wang
Quoted 4 lines

Mount with both dioread_nolock and nodelalloc will result in huge performance drop, which indeed is an known issue, so before we fix this issue, currently we disable this behaviour. Below test reproducer can reveal this performance drop.

Is it really worth it to disable this combination? Nothing goes *wrong* per se; it's just slower than we would like.

Yes, agree, then should we have some places to record this issue? In case somebody looks into it again.

Regards,
Xiaoguang Wang

Quoted 3 lines

- Ted