[PATCH] zram: move init_done check to the beginning of disksize_store

[PATCH] zram: move init_done check to the beginning of disksize_store

[Geliang Tang]

If we set disksize when disksize has been set, we will get the following
error report: "write error: Cannot allocate memory". This is because
disksize_store fails at zram_meta_alloc. This is not what we expect.
To solve this problem, this patch moves init_done check to the beginning
of disksize_store.

Signed-off-by: Geliang Tang <geliangtang@163.com>
---
 drivers/block/zram/zram_drv.c | 17 ++++++++---------
 1 file changed, 8 insertions(+), 9 deletions(-)

diff --git a/drivers/block/zram/zram_drv.c b/drivers/block/zram/zram_drv.c
index 81a557c..609fc2b 100644
--- a/drivers/block/zram/zram_drv.c
+++ b/drivers/block/zram/zram_drv.c
@@ -1028,6 +1028,14 @@ static ssize_t disksize_store(struct device *dev,
 	struct zram *zram = dev_to_zram(dev);
 	int err;
 
+	down_read(&zram->init_lock);
+	if (init_done(zram)) {
+		up_read(&zram->init_lock);
+		pr_info("Cannot change disksize for initialized device\n");
+		return -EBUSY;
+	}
+	up_read(&zram->init_lock);
+
 	disksize = memparse(buf, NULL);
 	if (!disksize)
 		return -EINVAL;
@@ -1046,12 +1054,6 @@ static ssize_t disksize_store(struct device *dev,
 	}
 
 	down_write(&zram->init_lock);
-	if (init_done(zram)) {
-		pr_info("Cannot change disksize for initialized device\n");
-		err = -EBUSY;
-		goto out_destroy_comp;
-	}
-
 	init_waitqueue_head(&zram->io_done);
 	atomic_set(&zram->refcount, 1);
 	zram->meta = meta;
@@ -1069,9 +1071,6 @@ static ssize_t disksize_store(struct device *dev,
 
 	return len;
 
-out_destroy_comp:
-	up_write(&zram->init_lock);
-	zcomp_destroy(comp);
 out_free_meta:
 	zram_meta_free(meta, disksize);
 	return err;
-- 
2.4.3

Re: [PATCH] zram: move init_done check to the beginning of disksize_store

[Sergey Senozhatsky]

On (10/31/15 22:50), Geliang Tang wrote:
> 
> If we set disksize when disksize has been set, we will get the following
> error report: "write error: Cannot allocate memory". This is because
> disksize_store fails at zram_meta_alloc.

Those things are not connected, absolutely. zram_meta_alloc() can fail
even on un-initialized device. In any case disksize_store() does not end
up changing the state of the device and reports the error back, so I don't
see any real value in this change. Seems that we come across this change
something like once a year (this patch is not the first).

I believe it has to stay the way it is, see https://lkml.org/lkml/2014/2/27/674

	-ss

> This is not what we expect.
> To solve this problem, this patch moves init_done check to the beginning
> of disksize_store.
> 
> Signed-off-by: Geliang Tang <geliangtang@163.com>
> ---
>  drivers/block/zram/zram_drv.c | 17 ++++++++---------
>  1 file changed, 8 insertions(+), 9 deletions(-)
> 
> diff --git a/drivers/block/zram/zram_drv.c b/drivers/block/zram/zram_drv.c
> index 81a557c..609fc2b 100644
> --- a/drivers/block/zram/zram_drv.c
> +++ b/drivers/block/zram/zram_drv.c
> @@ -1028,6 +1028,14 @@ static ssize_t disksize_store(struct device *dev,
>  	struct zram *zram = dev_to_zram(dev);
>  	int err;
>  
> +	down_read(&zram->init_lock);
> +	if (init_done(zram)) {
> +		up_read(&zram->init_lock);
> +		pr_info("Cannot change disksize for initialized device\n");
> +		return -EBUSY;
> +	}
> +	up_read(&zram->init_lock);
> +
>  	disksize = memparse(buf, NULL);
>  	if (!disksize)
>  		return -EINVAL;
> @@ -1046,12 +1054,6 @@ static ssize_t disksize_store(struct device *dev,
>  	}
>  
>  	down_write(&zram->init_lock);
> -	if (init_done(zram)) {
> -		pr_info("Cannot change disksize for initialized device\n");
> -		err = -EBUSY;
> -		goto out_destroy_comp;
> -	}
> -
>  	init_waitqueue_head(&zram->io_done);
>  	atomic_set(&zram->refcount, 1);
>  	zram->meta = meta;
> @@ -1069,9 +1071,6 @@ static ssize_t disksize_store(struct device *dev,
>  
>  	return len;
>  
> -out_destroy_comp:
> -	up_write(&zram->init_lock);
> -	zcomp_destroy(comp);
>  out_free_meta:
>  	zram_meta_free(meta, disksize);
>  	return err;
> -- 
> 2.4.3
> 
> 

Re: [PATCH] zram: move init_done check to the beginning of disksize_store

[Geliang Tang]

On Mon, Nov 02, 2015 at 08:43:58AM +0900, Sergey Senozhatsky wrote:
> On (10/31/15 22:50), Geliang Tang wrote:
> > 
> > If we set disksize when disksize has been set, we will get the following
> > error report: "write error: Cannot allocate memory". This is because
> > disksize_store fails at zram_meta_alloc.
> 
> Those things are not connected, absolutely. zram_meta_alloc() can fail
> even on un-initialized device. In any case disksize_store() does not end
> up changing the state of the device and reports the error back, so I don't
> see any real value in this change. Seems that we come across this change
> something like once a year (this patch is not the first).

Thanks for your reply.

zram_meta_alloc() always fails on initialized device at debugfs_create_dir():
debugfs dir <zram0> creation failed
zram: Error creating memory pool

I think we should deal with this error.

> 
> I believe it has to stay the way it is, see https://lkml.org/lkml/2014/2/27/674

Unlike above, this patch did not call zram_meta_alloc() in lock. So it
would not cause the deed lock problem.

Geliang Tang

Re: [PATCH] zram: move init_done check to the beginning of disksize_store

[Sergey Senozhatsky]

Hi,

[please don't trim the Cc list]

On (11/02/15 21:53), Geliang Tang wrote:
> > Those things are not connected, absolutely. zram_meta_alloc() can fail
> > even on un-initialized device. In any case disksize_store() does not end
> > up changing the state of the device and reports the error back, so I don't
> > see any real value in this change. Seems that we come across this change
> > something like once a year (this patch is not the first).
> 
> Thanks for your reply.
> 
> zram_meta_alloc() always fails on initialized device at debugfs_create_dir():
> debugfs dir <zram0> creation failed

Aha, so you are talking about calling disksize_store() on
	- already initialized device that
	- uses zsmalloc with
	- CONFIG_ZSMALLOC_STAT as an alacator

(zbud is not affected). that should have been mentioned in the commit message.


hm, does the patch _actually_ make it better? splitting a ->init_lock regions
is *never* a good idea. consider an un-initialized device.

CPU0							CPU1

down_read(&zram->init_lock);
if (init_done(zram))
	return -EBUSY;
up_read(&zram->init_lock);

					down_read(&zram->init_lock);
					if (init_done(zram))
						return -EBUSY;
					up_read(&zram->init_lock);

meta = zram_meta_alloc();
					meta = zram_meta_alloc();
comp = zcomp_create();
					comp = zcomp_create();
down_write(&zram->init_lock);
zram->meta = meta;
...
up_write(&zram->init_lock);

					down_write(&zram->init_lock);
					zram->meta = meta;
					...
					up_write(&zram->init_lock);


So the above code will return a "debugfs dir <zram0> creation failed"
IFF zsmalloc with CONFIG_ZSMALLOC_STAT enabled is being used (regardless the
effort). And it will totally succeed and lead to N (depending on the number of
concurrent disksize_store() calls) pool memory and zcomp_create memory leaks
otherwise. Because of this part:

---
        down_write(&zram->init_lock);
-       if (init_done(zram)) {
-               pr_info("Cannot change disksize for initialized device\n");
-               err = -EBUSY;
-               goto out_destroy_comp;
-       }
-
        init_waitqueue_head(&zram->io_done);
        atomic_set(&zram->refcount, 1);
        zram->meta = meta;
---


> zram: Error creating memory pool
> 
> I think we should deal with this error.

Is it so critical? I mean it does not break the existing device; we
print the error and keep the existing device fully functional. The
error line may be a bit misleading, which is another thing and can
be solved documenting the behavior, *for example*. (I just don't
really want to complicate the code for a corner case which doesn't
even break anything).

	-ss