RE: [PATCH] uacce: Add uacce_ctrl misc device

["Song Bao Hua (Barry Song)" <song.bao.hua@hisilicon.com>]

> -----Original Message-----
> From: Greg Kroah-Hartman [mailto:gregkh@linuxfoundation.org]
> Sent: Friday, January 22, 2021 12:19 AM
> To: Song Bao Hua (Barry Song) <song.bao.hua@hisilicon.com>
> Cc: Wangzhou (B) <wangzhou1@hisilicon.com>; Zhangfei Gao
> <zhangfei.gao@linaro.org>; Arnd Bergmann <arnd@arndb.de>;
> linux-accelerators@lists.ozlabs.org; linux-kernel@vger.kernel.org;
> chensihang (A) <chensihang1@hisilicon.com>
> Subject: Re: [PATCH] uacce: Add uacce_ctrl misc device
> 
> On Thu, Jan 21, 2021 at 10:18:24AM +0000, Song Bao Hua (Barry Song) wrote:
> >
> >
> > > -----Original Message-----
> > > From: Greg Kroah-Hartman [mailto:gregkh@linuxfoundation.org]
> > > Sent: Thursday, January 21, 2021 10:46 PM
> > > To: Wangzhou (B) <wangzhou1@hisilicon.com>
> > > Cc: Zhangfei Gao <zhangfei.gao@linaro.org>; Arnd Bergmann <arnd@arndb.de>;
> > > linux-accelerators@lists.ozlabs.org; linux-kernel@vger.kernel.org;
> > > chensihang (A) <chensihang1@hisilicon.com>
> > > Subject: Re: [PATCH] uacce: Add uacce_ctrl misc device
> > >
> > > On Thu, Jan 21, 2021 at 05:09:14PM +0800, Zhou Wang wrote:
> > > > When IO page fault happens, DMA performance will be affected. Pin user
> page
> > > > can avoid IO page fault, this patch introduces a new char device named
> > > > /dev/uacce_ctrl to help to maintain pin/unpin pages. User space can do
> > > > pin/unpin pages by ioctls of an open file of /dev/uacce_ctrl, all pinned
> > > > pages under one file will be unpinned in file release process.
> > >
> > > Also, what are you really trying to do here?  If you need to mess with
> > > memory pages, why can't the existing memory apis work properly for you?
> > > Please work with the linux-mm developers to resolve the issue using the
> > > standard apis and not creating a one-off char device node for this type
> > > of thing.
> >
> > Basically the purpose is implementing a pinned memory poll for userspace
> > DMA to achieve better performance by removing io page fault.
> 
> And what could possibly go wrong with that :)

I think we have resolved this concern while uacce came in :-)
Uacce is based on SVA so devices are accessing memory in userspace
by strict permission control.

> 
> > I really like this can be done in generic mm code. Unfortunately there is
> no
> > this standard API in kernel to support userspace pin. Right now, various
> > subsystems depend on the ioctl of /dev/<name> to implement the pin, for example,
> > v4l2, gpu, infiniband, media etc.
> >
> > I feel it is extremely hard to sell a standard mpin() API like mlock()
> > for this stage as mm could hardly buy this. And it will require
> > huge changes in kernel.
> 
> Why?  This is what mlock() is for, why can't you use it?

mlock() can only guarantee memory won't be swapped out, it doesn't
make sure memory won't move. alloc_pages() can cause memory compaction,
cma, numa balance, huge pages etc can move mlock()-ed pages.
We would still see many I/O page faults for mlock() area.

> 
> > We need a way to manage what pages are pinned by process and ensure the
> > pages can be unpinned while the process is killed abnormally. otherwise,
> > memory gets leaked.
> 
> Can't mlock() handle that?  It works on the process that called it.
> 
> > file_operations release() is a good entry for this kind of things. In
> > this way, we don't have to maintain the pinned page set in task_struct
> > and unpin them during exit().
> >
> > If there is anything to make it better by doing this in a driver. I
> > would believe we could have a generic misc driver for pin like
> > vms_ballon.c for ballon. The driver doesn't have to bind with uacce.
> >
> > In this way, the pinned memory pool implementation in userspace doesn't
> > need to depend on a specific uacce driver any more.
> 
> Please work with the mm developers to get them to agree with this type
> of thing, as well as the dma developers, both of which you didn't cc: on
> this patch :(

Yep.

> 
> Remember, you are creating a new api for Linux that goes around existing
> syscalls, but is in reality, a new syscall, so why not just make it a
> new syscall?

The difficulty would be how to record which pages are pinned for a process
if it is done by a new syscall.

For mlock(), it can be much easier as it will change VMA. Hardly we can
change VMA for pin. On the other hand, if the implementation is done in
driver, with file_operations, we can record pinned pages in the private
data of an opened file.

> 
> thanks,
> 
> greg k-h

Thanks
Barry