[PATCH v6 1/2] erofs: avoid memory allocation failure during rolling decompression

[PATCH v6 1/2] erofs: avoid memory allocation failure during rolling decompression

[Huang Jianan via Linux-erofs]

Currently, err would be treated as io error. Therefore, it'd be
better to ensure memory allocation during rolling decompression
to avoid such io error.

In the long term, we might consider adding another !Uptodate case
for such case.

Signed-off-by: Huang Jianan <huangjianan@oppo.com>
Signed-off-by: Guo Weichao <guoweichao@oppo.com>
Reviewed-by: Gao Xiang <hsiangkao@redhat.com>
---
 fs/erofs/decompressor.c | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/fs/erofs/decompressor.c b/fs/erofs/decompressor.c
index 1cb1ffd10569..34e73ff76f89 100644
--- a/fs/erofs/decompressor.c
+++ b/fs/erofs/decompressor.c
@@ -73,9 +73,8 @@ static int z_erofs_lz4_prepare_destpages(struct z_erofs_decompress_req *rq,
 			victim = availables[--top];
 			get_page(victim);
 		} else {
-			victim = erofs_allocpage(pagepool, GFP_KERNEL);
-			if (!victim)
-				return -ENOMEM;
+			victim = erofs_allocpage(pagepool,
+						 GFP_KERNEL | __GFP_NOFAIL);
 			set_page_private(victim, Z_EROFS_SHORTLIVED_PAGE);
 		}
 		rq->out[i] = victim;
-- 
2.25.1

[PATCH v6 2/2] erofs: decompress in endio if possible

[Huang Jianan via Linux-erofs]

z_erofs_decompressqueue_endio may not be executed in the atomic
context, for example, when dm-verity is turned on. In this scenario,
data can be decompressed directly to get rid of additional kworker
scheduling overhead. Also, it makes no sense to apply synchronous
decompression for such case.

Signed-off-by: Huang Jianan <huangjianan@oppo.com>
Signed-off-by: Guo Weichao <guoweichao@oppo.com>
Reviewed-by: Gao Xiang <hsiangkao@redhat.com>
---
 fs/erofs/internal.h |  2 ++
 fs/erofs/super.c    |  1 +
 fs/erofs/zdata.c    | 15 +++++++++++++--
 3 files changed, 16 insertions(+), 2 deletions(-)

diff --git a/fs/erofs/internal.h b/fs/erofs/internal.h
index 67a7ec945686..fbc4040715be 100644
--- a/fs/erofs/internal.h
+++ b/fs/erofs/internal.h
@@ -50,6 +50,8 @@ struct erofs_fs_context {
 #ifdef CONFIG_EROFS_FS_ZIP
 	/* current strategy of how to use managed cache */
 	unsigned char cache_strategy;
+	/* strategy of sync decompression (false - auto, true - force on) */
+	bool readahead_sync_decompress;
 
 	/* threshold for decompression synchronously */
 	unsigned int max_sync_decompress_pages;
diff --git a/fs/erofs/super.c b/fs/erofs/super.c
index d5a6b9b888a5..0445d09b6331 100644
--- a/fs/erofs/super.c
+++ b/fs/erofs/super.c
@@ -200,6 +200,7 @@ static void erofs_default_options(struct erofs_fs_context *ctx)
 #ifdef CONFIG_EROFS_FS_ZIP
 	ctx->cache_strategy = EROFS_ZIP_CACHE_READAROUND;
 	ctx->max_sync_decompress_pages = 3;
+	ctx->readahead_sync_decompress = false;
 #endif
 #ifdef CONFIG_EROFS_FS_XATTR
 	set_opt(ctx, XATTR_USER);
diff --git a/fs/erofs/zdata.c b/fs/erofs/zdata.c
index 6cb356c4217b..25a0c4890d0a 100644
--- a/fs/erofs/zdata.c
+++ b/fs/erofs/zdata.c
@@ -706,9 +706,12 @@ static int z_erofs_do_read_page(struct z_erofs_decompress_frontend *fe,
 	goto out;
 }
 
+static void z_erofs_decompressqueue_work(struct work_struct *work);
 static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io,
 				       bool sync, int bios)
 {
+	struct erofs_sb_info *const sbi = EROFS_SB(io->sb);
+
 	/* wake up the caller thread for sync decompression */
 	if (sync) {
 		unsigned long flags;
@@ -720,8 +723,15 @@ static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io,
 		return;
 	}
 
-	if (!atomic_add_return(bios, &io->pending_bios))
+	if (atomic_add_return(bios, &io->pending_bios))
+		return;
+	/* Use workqueue and sync decompression for atomic contexts only */
+	if (in_atomic() || irqs_disabled()) {
 		queue_work(z_erofs_workqueue, &io->u.work);
+		sbi->ctx.readahead_sync_decompress = true;
+		return;
+	}
+	z_erofs_decompressqueue_work(&io->u.work);
 }
 
 static bool z_erofs_page_is_invalidated(struct page *page)
@@ -1333,7 +1343,8 @@ static void z_erofs_readahead(struct readahead_control *rac)
 	struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
 
 	unsigned int nr_pages = readahead_count(rac);
-	bool sync = (nr_pages <= sbi->ctx.max_sync_decompress_pages);
+	bool sync = (sbi->ctx.readahead_sync_decompress &&
+			nr_pages <= sbi->ctx.max_sync_decompress_pages);
 	struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
 	struct page *page, *head = NULL;
 	LIST_HEAD(pagepool);
-- 
2.25.1

Re: [PATCH v6 1/2] erofs: avoid memory allocation failure during rolling decompression

[Chao Yu]

On 2021/3/16 11:15, Huang Jianan via Linux-erofs wrote:
> Currently, err would be treated as io error. Therefore, it'd be
> better to ensure memory allocation during rolling decompression
> to avoid such io error.
> 
> In the long term, we might consider adding another !Uptodate case
> for such case.
> 
> Signed-off-by: Huang Jianan <huangjianan@oppo.com>
> Signed-off-by: Guo Weichao <guoweichao@oppo.com>
> Reviewed-by: Gao Xiang <hsiangkao@redhat.com>

Reviewed-by: Chao Yu <yuchao0@huawei.com>

Thanks,

Re: [PATCH v6 2/2] erofs: decompress in endio if possible

[Chao Yu]

Hi Jianan,

On 2021/3/16 11:15, Huang Jianan via Linux-erofs wrote:
> z_erofs_decompressqueue_endio may not be executed in the atomic
> context, for example, when dm-verity is turned on. In this scenario,
> data can be decompressed directly to get rid of additional kworker
> scheduling overhead. Also, it makes no sense to apply synchronous
> decompression for such case.

It looks this patch does more than one things:
- combine dm-verity and erofs workqueue
- change policy of decompression in context of thread

Normally, we do one thing in one patch, by this way, we will be benefit in
scenario of when backporting patches and bisecting problematic patch with
minimum granularity, and also it will help reviewer to focus on reviewing
single code logic by following patch's goal.

So IMO, it would be better to separate this patch into two.

One more thing is could you explain a little bit more about why we need to
change policy of decompression in context of thread? for better performance?

BTW, code looks clean to me. :)

Thanks,

> 
> Signed-off-by: Huang Jianan <huangjianan@oppo.com>
> Signed-off-by: Guo Weichao <guoweichao@oppo.com>
> Reviewed-by: Gao Xiang <hsiangkao@redhat.com>
> ---
>   fs/erofs/internal.h |  2 ++
>   fs/erofs/super.c    |  1 +
>   fs/erofs/zdata.c    | 15 +++++++++++++--
>   3 files changed, 16 insertions(+), 2 deletions(-)
> 
> diff --git a/fs/erofs/internal.h b/fs/erofs/internal.h
> index 67a7ec945686..fbc4040715be 100644
> --- a/fs/erofs/internal.h
> +++ b/fs/erofs/internal.h
> @@ -50,6 +50,8 @@ struct erofs_fs_context {
>   #ifdef CONFIG_EROFS_FS_ZIP
>   	/* current strategy of how to use managed cache */
>   	unsigned char cache_strategy;
> +	/* strategy of sync decompression (false - auto, true - force on) */
> +	bool readahead_sync_decompress;
>   
>   	/* threshold for decompression synchronously */
>   	unsigned int max_sync_decompress_pages;
> diff --git a/fs/erofs/super.c b/fs/erofs/super.c
> index d5a6b9b888a5..0445d09b6331 100644
> --- a/fs/erofs/super.c
> +++ b/fs/erofs/super.c
> @@ -200,6 +200,7 @@ static void erofs_default_options(struct erofs_fs_context *ctx)
>   #ifdef CONFIG_EROFS_FS_ZIP
>   	ctx->cache_strategy = EROFS_ZIP_CACHE_READAROUND;
>   	ctx->max_sync_decompress_pages = 3;
> +	ctx->readahead_sync_decompress = false;
>   #endif
>   #ifdef CONFIG_EROFS_FS_XATTR
>   	set_opt(ctx, XATTR_USER);
> diff --git a/fs/erofs/zdata.c b/fs/erofs/zdata.c
> index 6cb356c4217b..25a0c4890d0a 100644
> --- a/fs/erofs/zdata.c
> +++ b/fs/erofs/zdata.c
> @@ -706,9 +706,12 @@ static int z_erofs_do_read_page(struct z_erofs_decompress_frontend *fe,
>   	goto out;
>   }
>   
> +static void z_erofs_decompressqueue_work(struct work_struct *work);
>   static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io,
>   				       bool sync, int bios)
>   {
> +	struct erofs_sb_info *const sbi = EROFS_SB(io->sb);
> +
>   	/* wake up the caller thread for sync decompression */
>   	if (sync) {
>   		unsigned long flags;
> @@ -720,8 +723,15 @@ static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io,
>   		return;
>   	}
>   
> -	if (!atomic_add_return(bios, &io->pending_bios))
> +	if (atomic_add_return(bios, &io->pending_bios))
> +		return;
> +	/* Use workqueue and sync decompression for atomic contexts only */
> +	if (in_atomic() || irqs_disabled()) {
>   		queue_work(z_erofs_workqueue, &io->u.work);
> +		sbi->ctx.readahead_sync_decompress = true;
> +		return;
> +	}
> +	z_erofs_decompressqueue_work(&io->u.work);
>   }
>   
>   static bool z_erofs_page_is_invalidated(struct page *page)
> @@ -1333,7 +1343,8 @@ static void z_erofs_readahead(struct readahead_control *rac)
>   	struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
>   
>   	unsigned int nr_pages = readahead_count(rac);
> -	bool sync = (nr_pages <= sbi->ctx.max_sync_decompress_pages);
> +	bool sync = (sbi->ctx.readahead_sync_decompress &&
> +			nr_pages <= sbi->ctx.max_sync_decompress_pages);
>   	struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
>   	struct page *page, *head = NULL;
>   	LIST_HEAD(pagepool);
> 

Re: [PATCH v6 2/2] erofs: decompress in endio if possible

[Huang Jianan via Linux-erofs]

On 2021/3/16 16:26, Chao Yu wrote:
> Hi Jianan,
>
> On 2021/3/16 11:15, Huang Jianan via Linux-erofs wrote:
>> z_erofs_decompressqueue_endio may not be executed in the atomic
>> context, for example, when dm-verity is turned on. In this scenario,
>> data can be decompressed directly to get rid of additional kworker
>> scheduling overhead. Also, it makes no sense to apply synchronous
>> decompression for such case.
>
> It looks this patch does more than one things:
> - combine dm-verity and erofs workqueue
> - change policy of decompression in context of thread
>
> Normally, we do one thing in one patch, by this way, we will be 
> benefit in
> scenario of when backporting patches and bisecting problematic patch with
> minimum granularity, and also it will help reviewer to focus on reviewing
> single code logic by following patch's goal.
>
> So IMO, it would be better to separate this patch into two.
>
Thanks for the suggestion, I will send a new patch set.
> One more thing is could you explain a little bit more about why we 
> need to
> change policy of decompression in context of thread? for better 
> performance?
>
Sync decompression was introduced to get rid of additional kworker 
scheduling

overhead. But there is no such overhead in if we try to decompress 
directly in

z_erofs_decompressqueue_endio . Therefore, it  should be better to turn off

sync decompression to avoid the current thread waiting in z_erofs_runqueue.

> BTW, code looks clean to me. :)
>
> Thanks,
>
>>
>> Signed-off-by: Huang Jianan <huangjianan@oppo.com>
>> Signed-off-by: Guo Weichao <guoweichao@oppo.com>
>> Reviewed-by: Gao Xiang <hsiangkao@redhat.com>
>> ---
>>   fs/erofs/internal.h |  2 ++
>>   fs/erofs/super.c    |  1 +
>>   fs/erofs/zdata.c    | 15 +++++++++++++--
>>   3 files changed, 16 insertions(+), 2 deletions(-)
>>
>> diff --git a/fs/erofs/internal.h b/fs/erofs/internal.h
>> index 67a7ec945686..fbc4040715be 100644
>> --- a/fs/erofs/internal.h
>> +++ b/fs/erofs/internal.h
>> @@ -50,6 +50,8 @@ struct erofs_fs_context {
>>   #ifdef CONFIG_EROFS_FS_ZIP
>>       /* current strategy of how to use managed cache */
>>       unsigned char cache_strategy;
>> +    /* strategy of sync decompression (false - auto, true - force 
>> on) */
>> +    bool readahead_sync_decompress;
>>         /* threshold for decompression synchronously */
>>       unsigned int max_sync_decompress_pages;
>> diff --git a/fs/erofs/super.c b/fs/erofs/super.c
>> index d5a6b9b888a5..0445d09b6331 100644
>> --- a/fs/erofs/super.c
>> +++ b/fs/erofs/super.c
>> @@ -200,6 +200,7 @@ static void erofs_default_options(struct 
>> erofs_fs_context *ctx)
>>   #ifdef CONFIG_EROFS_FS_ZIP
>>       ctx->cache_strategy = EROFS_ZIP_CACHE_READAROUND;
>>       ctx->max_sync_decompress_pages = 3;
>> +    ctx->readahead_sync_decompress = false;
>>   #endif
>>   #ifdef CONFIG_EROFS_FS_XATTR
>>       set_opt(ctx, XATTR_USER);
>> diff --git a/fs/erofs/zdata.c b/fs/erofs/zdata.c
>> index 6cb356c4217b..25a0c4890d0a 100644
>> --- a/fs/erofs/zdata.c
>> +++ b/fs/erofs/zdata.c
>> @@ -706,9 +706,12 @@ static int z_erofs_do_read_page(struct 
>> z_erofs_decompress_frontend *fe,
>>       goto out;
>>   }
>>   +static void z_erofs_decompressqueue_work(struct work_struct *work);
>>   static void z_erofs_decompress_kickoff(struct 
>> z_erofs_decompressqueue *io,
>>                          bool sync, int bios)
>>   {
>> +    struct erofs_sb_info *const sbi = EROFS_SB(io->sb);
>> +
>>       /* wake up the caller thread for sync decompression */
>>       if (sync) {
>>           unsigned long flags;
>> @@ -720,8 +723,15 @@ static void z_erofs_decompress_kickoff(struct 
>> z_erofs_decompressqueue *io,
>>           return;
>>       }
>>   -    if (!atomic_add_return(bios, &io->pending_bios))
>> +    if (atomic_add_return(bios, &io->pending_bios))
>> +        return;
>> +    /* Use workqueue and sync decompression for atomic contexts only */
>> +    if (in_atomic() || irqs_disabled()) {
>>           queue_work(z_erofs_workqueue, &io->u.work);
>> +        sbi->ctx.readahead_sync_decompress = true;
>> +        return;
>> +    }
>> +    z_erofs_decompressqueue_work(&io->u.work);
>>   }
>>     static bool z_erofs_page_is_invalidated(struct page *page)
>> @@ -1333,7 +1343,8 @@ static void z_erofs_readahead(struct 
>> readahead_control *rac)
>>       struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
>>         unsigned int nr_pages = readahead_count(rac);
>> -    bool sync = (nr_pages <= sbi->ctx.max_sync_decompress_pages);
>> +    bool sync = (sbi->ctx.readahead_sync_decompress &&
>> +            nr_pages <= sbi->ctx.max_sync_decompress_pages);
>>       struct z_erofs_decompress_frontend f = 
>> DECOMPRESS_FRONTEND_INIT(inode);
>>       struct page *page, *head = NULL;
>>       LIST_HEAD(pagepool);
>>