RE: [PATCH] zsmalloc: do not use bit_spin_lock

[Song Bao Hua (Barry Song)]

> -----Original Message-----
> From: Vitaly Wool [mailto:vitaly.wool@xxxxxxxxxxxx]
> Sent: Tuesday, December 22, 2020 11:12 AM
> To: Song Bao Hua (Barry Song) <song.bao.hua@xxxxxxxxxxxxx>
> Cc: Shakeel Butt <shakeelb@xxxxxxxxxx>; Minchan Kim <minchan@xxxxxxxxxx>; Mike
> Galbraith <efault@xxxxxx>; LKML <linux-kernel@xxxxxxxxxxxxxxx>; linux-mm
> <linux-mm@xxxxxxxxx>; Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx>;
> NitinGupta <ngupta@xxxxxxxxxx>; Sergey Senozhatsky
> <sergey.senozhatsky.work@xxxxxxxxx>; Andrew Morton
> <akpm@xxxxxxxxxxxxxxxxxxxx>
> Subject: Re: [PATCH] zsmalloc: do not use bit_spin_lock
> 
> On Mon, Dec 21, 2020 at 10:30 PM Song Bao Hua (Barry Song)
> <song.bao.hua@xxxxxxxxxxxxx> wrote:
> >
> >
> >
> > > -----Original Message-----
> > > From: Shakeel Butt [mailto:shakeelb@xxxxxxxxxx]
> > > Sent: Tuesday, December 22, 2020 10:03 AM
> > > To: Song Bao Hua (Barry Song) <song.bao.hua@xxxxxxxxxxxxx>
> > > Cc: Vitaly Wool <vitaly.wool@xxxxxxxxxxxx>; Minchan Kim
> <minchan@xxxxxxxxxx>;
> > > Mike Galbraith <efault@xxxxxx>; LKML <linux-kernel@xxxxxxxxxxxxxxx>;
> linux-mm
> > > <linux-mm@xxxxxxxxx>; Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx>;
> > > NitinGupta <ngupta@xxxxxxxxxx>; Sergey Senozhatsky
> > > <sergey.senozhatsky.work@xxxxxxxxx>; Andrew Morton
> > > <akpm@xxxxxxxxxxxxxxxxxxxx>
> > > Subject: Re: [PATCH] zsmalloc: do not use bit_spin_lock
> > >
> > > On Mon, Dec 21, 2020 at 12:06 PM Song Bao Hua (Barry Song)
> > > <song.bao.hua@xxxxxxxxxxxxx> wrote:
> > > >
> > > >
> > > >
> > > > > -----Original Message-----
> > > > > From: Shakeel Butt [mailto:shakeelb@xxxxxxxxxx]
> > > > > Sent: Tuesday, December 22, 2020 8:50 AM
> > > > > To: Vitaly Wool <vitaly.wool@xxxxxxxxxxxx>
> > > > > Cc: Minchan Kim <minchan@xxxxxxxxxx>; Mike Galbraith <efault@xxxxxx>;
> LKML
> > > > > <linux-kernel@xxxxxxxxxxxxxxx>; linux-mm <linux-mm@xxxxxxxxx>; Song
> Bao
> > > Hua
> > > > > (Barry Song) <song.bao.hua@xxxxxxxxxxxxx>; Sebastian Andrzej Siewior
> > > > > <bigeasy@xxxxxxxxxxxxx>; NitinGupta <ngupta@xxxxxxxxxx>; Sergey
> > > Senozhatsky
> > > > > <sergey.senozhatsky.work@xxxxxxxxx>; Andrew Morton
> > > > > <akpm@xxxxxxxxxxxxxxxxxxxx>
> > > > > Subject: Re: [PATCH] zsmalloc: do not use bit_spin_lock
> > > > >
> > > > > On Mon, Dec 21, 2020 at 11:20 AM Vitaly Wool <vitaly.wool@xxxxxxxxxxxx>
> > > wrote:
> > > > > >
> > > > > > On Mon, Dec 21, 2020 at 6:24 PM Minchan Kim <minchan@xxxxxxxxxx> wrote:
> > > > > > >
> > > > > > > On Sun, Dec 20, 2020 at 02:22:28AM +0200, Vitaly Wool wrote:
> > > > > > > > zsmalloc takes bit spinlock in its _map() callback and releases
> it
> > > > > > > > only in unmap() which is unsafe and leads to zswap complaining
> > > > > > > > about scheduling in atomic context.
> > > > > > > >
> > > > > > > > To fix that and to improve RT properties of zsmalloc, remove that
> > > > > > > > bit spinlock completely and use a bit flag instead.
> > > > > > >
> > > > > > > I don't want to use such open code for the lock.
> > > > > > >
> > > > > > > I see from Mike's patch, recent zswap change introduced the lockdep
> > > > > > > splat bug and you want to improve zsmalloc to fix the zswap bug
> and
> > > > > > > introduce this patch with allowing preemption enabling.
> > > > > >
> > > > > > This understanding is upside down. The code in zswap you are referring
> > > > > > to is not buggy.  You may claim that it is suboptimal but there is
> > > > > > nothing wrong in taking a mutex.
> > > > > >
> > > > >
> > > > > Is this suboptimal for all or just the hardware accelerators? Sorry,
> I
> > > > > am not very familiar with the crypto API. If I select lzo or lz4 as
> a
> > > > > zswap compressor will the [de]compression be async or sync?
> > > >
> > > > Right now, in crypto subsystem, new drivers are required to write based
> on
> > > > async APIs. The old sync API can't work in new accelerator drivers as
> they
> > > > are not supported at all.
> > > >
> > > > Old drivers are used to sync, but they've got async wrappers to support
> async
> > > > APIs. Eg.
> > > > crypto: acomp - add support for lz4 via scomp
> > > >
> > >
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/
> > > crypto/lz4.c?id=8cd9330e0a615c931037d4def98b5ce0d540f08d
> > > >
> > > > crypto: acomp - add support for lzo via scomp
> > > >
> > >
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/
> > > crypto/lzo.c?id=ac9d2c4b39e022d2c61486bfc33b730cfd02898e
> > > >
> > > > so they are supporting async APIs but they are still working in sync mode
> > > as
> > > > those old drivers don't sleep.
> > > >
> > >
> > > Good to know that those are sync because I want them to be sync.
> > > Please note that zswap is a cache in front of a real swap and the load
> > > operation is latency sensitive as it comes in the page fault path and
> > > directly impacts the applications. I doubt decompressing synchronously
> > > a 4k page on a cpu will be costlier than asynchronously decompressing
> > > the same page from hardware accelerators.
> >
> > If you read the old paper:
> >
> https://www.ibm.com/support/pages/new-linux-zswap-compression-functionalit
> y
> > Because the hardware accelerator speeds up compression, looking at the zswap
> > metrics we observed that there were more store and load requests in a given
> > amount of time, which filled up the zswap pool faster than a software
> > compression run. Because of this behavior, we set the max_pool_percent
> > parameter to 30 for the hardware compression runs - this means that zswap
> > can use up to 30% of the 10GB of total memory.
> >
> > So using hardware accelerators, we get a chance to speed up compression
> > while decreasing cpu utilization.
> >
> > BTW, If it is not easy to change zsmalloc, one quick workaround we might do
> > in zswap is adding the below after applying Mike's original patch:
> >
> > if(in_atomic()) /* for zsmalloc */
> >         while(!try_wait_for_completion(&req->done);
> > else /* for zbud, z3fold */
> >         crypto_wait_req(....);
> 
> I don't think I'm going to ack this, sorry.
> 

Fair enough. And I am also thinking if we can move zpool_unmap_handle()
quite after zpool_map_handle() as below:

	dlen = PAGE_SIZE;
	src = zpool_map_handle(entry->pool->zpool, entry->handle, ZPOOL_MM_RO);
	if (zpool_evictable(entry->pool->zpool))
		src += sizeof(struct zswap_header);
+	zpool_unmap_handle(entry->pool->zpool, entry->handle);

	acomp_ctx = raw_cpu_ptr(entry->pool->acomp_ctx);
	mutex_lock(acomp_ctx->mutex);
	sg_init_one(&input, src, entry->length);
	sg_init_table(&output, 1);
	sg_set_page(&output, page, PAGE_SIZE, 0);
	acomp_request_set_params(acomp_ctx->req, &input, &output, entry->length, dlen);
	ret = crypto_wait_req(crypto_acomp_decompress(acomp_ctx->req), &acomp_ctx->wait);
	mutex_unlock(acomp_ctx->mutex);

-	zpool_unmap_handle(entry->pool->zpool, entry->handle);

Since src is always low memory and we only need its virtual address
to get the page of src in sg_init_one(). We don't actually read it
by CPU anywhere.

> Best regards,
>    Vitaly
> 
> > crypto_wait_req() is actually doing wait_for_completion():
> > static inline int crypto_wait_req(int err, struct crypto_wait *wait)
> > {
> >         switch (err) {
> >         case -EINPROGRESS:
> >         case -EBUSY:
> >                 wait_for_completion(&wait->completion);
> >                 reinit_completion(&wait->completion);
> >                 err = wait->err;
> >                 break;
> >         }
> >
> >         return err;
> > }
> >
Thanks
Barry