Re: [PATCH] zram: Using GFP_ATOMIC instead of GFP_KERNEL to allocate bitmap memory in backing_dev_store

[Dongyun Liu]

On 2023/12/1 22:19, Jens Axboe wrote:
> On 11/30/23 11:51 PM, Dongyun Liu wrote:
> > On 2023/11/30 23:37, Jens Axboe wrote:
> > > On 11/30/23 8:20 AM, Dongyun Liu wrote:
> > > > diff --git a/drivers/block/zram/zram_drv.c b/drivers/block/zram/zram_drv.c
> > > > index d77d3664ca08..ee6c22c50e09 100644
> > > > --- a/drivers/block/zram/zram_drv.c
> > > > +++ b/drivers/block/zram/zram_drv.c
> > > > @@ -514,7 +514,7 @@ static ssize_t backing_dev_store(struct device *dev,
> > > >          nr_pages = i_size_read(inode) >> PAGE_SHIFT;
> > > >        bitmap_sz = BITS_TO_LONGS(nr_pages) * sizeof(long);
> > > > -    bitmap = kvzalloc(bitmap_sz, GFP_KERNEL);
> > > > +    bitmap = kmalloc(bitmap_sz, GFP_ATOMIC);
> > > >        if (!bitmap) {
> > > >            err = -ENOMEM;
> > > >            goto out;
> > >
> > > Outside of this moving from a zeroed alloc to one that does not, the
> > > change looks woefully incomplete. Why does this allocation need to be
> > > GFP_ATOMIC, and:
> >
> > By using GFP_ATOMIC, it indicates that the caller cannot reclaim or
> > sleep, although we can prevent the risk of  deadlock when acquiring
> > the zram->lock again in zram_bvec_write.
>
> Yes, I am very much aware of how gfp allocation flags work and how why
> it's broken. It was a rhetorical question as to why you think you could
> get away with just fixing one of them.
>
> > > 1) file_name = kmalloc(PATH_MAX, GFP_KERNEL); does not
> >
> > There is no zram->init_lock held here, so there is no need to use
> > GFP_ATOMIC.
>
> True
>
> > > 2) filp_open() -> getname_kernel() -> __getname() does not
> > > 3) filp_open() -> getname_kernel() does not
> > > 4) bdev_open_by_dev() does not
> >
> > Missing the use of GFP_ATOMIC.
>
> Indeed!
>
> > > IOW, you have a slew of GFP_KERNEL allocations in there, and you
> > > probably just patched the largest one. But the core issue remains.
> > >
> > > The whole handling of backing_dev_store() looks pretty broken.
> >
> > Indeed, this patch only solves the biggest problem and does not
> > fundamentally solve it, because there are many processes for holding
> > zram->init_lock before allocation memory in backing_dev_store that
> > need to be fully modified, and I did not consider it thoroughly.
> > Obviously, a larger and better patch is needed to eliminate this risk,
> > but it is currently not necessary.
>
> You agree that it doesn't fix the issue, it just happens to fix the one
> that you hit. And then you jump to the conclusion that this is all
> that's needed to fix it. Ehm, confused?

Hi, Jens, Maybe there's something wrong with my expression. You can 
think of it this way: I agree with you that it doesn't fix the issue.