Message-ID: <20180911134013.GA3932@redhat.com>
Date:   Tue, 11 Sep 2018 09:40:14 -0400
From:   Jerome Glisse <jglisse@...hat.com>
To:     Kenneth Lee <liguozhu@...ilicon.com>
Cc:     Kenneth Lee <nek.in.cn@...il.com>,
        Alex Williamson <alex.williamson@...hat.com>,
        Herbert Xu <herbert@...dor.apana.org.au>, kvm@...r.kernel.org,
        Jonathan Corbet <corbet@....net>,
        Greg Kroah-Hartman <gregkh@...uxfoundation.org>,
        Zaibo Xu <xuzaibo@...wei.com>, linux-doc@...r.kernel.org,
        Sanjay Kumar <sanjay.k.kumar@...el.com>,
        Hao Fang <fanghao11@...wei.com>, linux-kernel@...r.kernel.org,
        linuxarm@...wei.com, iommu@...ts.linux-foundation.org,
        "David S . Miller" <davem@...emloft.net>,
        linux-crypto@...r.kernel.org, Zhou Wang <wangzhou1@...ilicon.com>,
        Philippe Ombredanne <pombredanne@...b.com>,
        Thomas Gleixner <tglx@...utronix.de>,
        Joerg Roedel <joro@...tes.org>,
        linux-accelerators@...ts.ozlabs.org,
        Lu Baolu <baolu.lu@...ux.intel.com>
Subject: Re: [RFCv2 PATCH 0/7] A General Accelerator Framework, WarpDrive

On Tue, Sep 11, 2018 at 02:40:43PM +0800, Kenneth Lee wrote:
> On Mon, Sep 10, 2018 at 11:33:59PM -0400, Jerome Glisse wrote:
> > On Tue, Sep 11, 2018 at 10:42:09AM +0800, Kenneth Lee wrote:
> > > On Mon, Sep 10, 2018 at 10:54:23AM -0400, Jerome Glisse wrote:
> > > > On Mon, Sep 10, 2018 at 11:28:09AM +0800, Kenneth Lee wrote:
> > > > > On Fri, Sep 07, 2018 at 12:53:06PM -0400, Jerome Glisse wrote:
> > > > > > On Fri, Sep 07, 2018 at 12:01:38PM +0800, Kenneth Lee wrote:
> > > > > > > On Thu, Sep 06, 2018 at 09:31:33AM -0400, Jerome Glisse wrote:
> > > > > > > > On Thu, Sep 06, 2018 at 05:45:32PM +0800, Kenneth Lee wrote:
> > > > > > > > > On Tue, Sep 04, 2018 at 10:15:09AM -0600, Alex Williamson wrote:
> > > > > > > > > > On Tue, 4 Sep 2018 11:00:19 -0400 Jerome Glisse <jglisse@...hat.com> wrote:
> > > > > > > > > > > On Mon, Sep 03, 2018 at 08:51:57AM +0800, Kenneth Lee wrote:
> > > > 
> > > > [...]
> > > > 
> > > > > > > I took a look at i915_gem_execbuffer_ioctl(). It seems it "copy_from_user" the
> > > > > > > user memory to the kernel. That is not what we need. What we try to get is: the
> > > > > > > user application do something on its data, and push it away to the accelerator,
> > > > > > > and says: "I'm tied, it is your turn to do the job...". Then the accelerator has
> > > > > > > the memory, referring any portion of it with the same VAs of the application,
> > > > > > > even the VAs are stored inside the memory itself.
> > > > > > 
> > > > > > You were not looking at right place see drivers/gpu/drm/i915/i915_gem_userptr.c
> > > > > > It does GUP and create GEM object AFAICR you can wrap that GEM object into a
> > > > > > dma buffer object.
> > > > > > 
> > > > > 
> > > > > Thank you for directing me to this implementation. It is interesting:).
> > > > > 
> > > > > But it is not yet solve my problem. If I understand it right, the userptr in
> > > > > i915 do the following:
> > > > > 
> > > > > 1. The user process sets a user pointer with size to the kernel via ioctl.
> > > > > 2. The kernel wraps it as a dma-buf and keeps the process's mm for further
> > > > >    reference.
> > > > > 3. The user pages are allocated, GUPed or DMA mapped to the device. So the data
> > > > >    can be shared between the user space and the hardware.
> > > > > 
> > > > > But my scenario is: 
> > > > > 
> > > > > 1. The user process has some data in the user space, pointed by a pointer, say
> > > > >    ptr1. And within the memory, there may be some other pointers, let's say one
> > > > >    of them is ptr2.
> > > > > 2. Now I need to assign ptr1 *directly* to the hardware MMIO space. And the
> > > > >    hardware must refer ptr1 and ptr2 *directly* for data.
> > > > > 
> > > > > Userptr lets the hardware and process share the same memory space. But I need
> > > > > them to share the same *address space*. So IOMMU is a MUST for WarpDrive,
> > > > > NOIOMMU mode, as Jean said, is just for verifying some of the procedure is OK.
> > > > 
> > > > So to be 100% clear should we _ignore_ the non SVA/SVM case ?
> > > > If so then wait for necessary SVA/SVM to land and do warp drive
> > > > without non SVA/SVM path.
> > > > 
> > > 
> > > I think we should clear the concept of SVA/SVM here. As my understanding, Share
> > > Virtual Address/Memory means: any virtual address in a process can be used by
> > > device at the same time. This requires IOMMU device to support PASID. And
> > > optionally, it requires the feature of page-fault-from-device.
> > 
> > Yes we agree on what SVA/SVM is. There is a one gotcha thought, access
> > to range that are MMIO map ie CPU page table pointing to IO memory, IIRC
> > it is undefined what happens on some platform for a device trying to
> > access those using SVA/SVM.
> > 
> > 
> > > But before the feature is settled down, IOMMU can be used immediately in the
> > > current kernel. That make it possible to assign ONE process's virtual addresses
> > > to the device's IOMMU page table with GUP. This make WarpDrive work well for one
> > > process.
> > 
> > UH ? How ? You want to GUP _every_ single valid address in the process
> > and map it to the device ? How do you handle new vma, page being replace
> > (despite GUP because of things that utimately calls zap pte) ...
> > 
> > Again here you said that the device must be able to access _any_ valid
> > pointer. With GUP this is insane.
> > 
> > So i am assuming this is not what you want to do without SVA/SVM ie with
> > GUP you have a different programming model, one in which the userspace
> > must first bind _range_ of memory to the device and get a DMA address
> > for the range.
> > 
> > Again, GUP range of process address space to map it to a device so that
> > userspace can use the device on the mapped range is something that do
> > exist in various places in the kernel.
> > 
> 
> Yes same as your expectation, in WarpDrive, we use the concept of "sharing" to
> do so. If some memory is going to be shared among process and devices, we use
> wd_share_mem(queue, ptr, size) to share those memory. When the queue is working
> in this mode, the point is valid in those memory segments. The wd_share_mem call
> vfio dma map syscall which will do GUP. 
> 
> If SVA/SVM is enabled, user space can set SHARE_ALL flags to the queue. Then
> wd_share_mem() is not necessary.
> 
> This is really not popular when we started the work on WarpDrive. The GUP
> document said it should be put within the scope of mm_sem is locked. Because GUP
> simply increase the page refcount, not keep the mapping between the page and the
> vma. We keep our work together with VFIO to make sure the problem can be solved
> in one deal.

The problem can not be solved in one deal, you can not maintain vaddr
pointing to same page after a fork() this can not be solve without the
use of mmu notifier and device dma mapping invalidation ! So being part
of VFIO will not help you there.

AFAIK VFIO is fine with the way it is as QEMU do not fork() once it
is running a guest and thus the COW that would invalidate vaddr to
physical page assumption is not broken. So i doubt VFIO folks have
any incentive to go down the mmu notifier path and invalidate device
mapping. They also have the replay thing that probably handle some
of fork cases by trusting user space program to do it. In your case
you can not trust the user space program.

In your case AFAICT i do not see any warning or gotcha so the following
scenario is broken (in non SVA/SVM):
    1) program setup the device (open container, mdev, setup queue, ...)
    2) program map some range of its address space wih VFIO_IOMMU_MAP_DMA
    3) program start using the device using map setup in 2)
    ...
    4) program fork()
    5) parent trigger COW inside the range setup in 2)

    At this point it is the child process that can write to the page that
    are access by the device (which was map by the parent in 2)). The
    parent can no longer access that memory from the CPU.

There is just no sane way to fix this beside invalidating device mapping
on fork (and you can not rely on userspace to do so) and thus stopping
the device on fork (SVA/SVM case do not have any issue here).

> And now we have GUP-longterm and many accounting work in VFIO, we don't want to
> do that again.

GUP-longterm does not solve any GUP problem, it just block people to
do GUP on DAX backed vma to avoid pining persistent memory as it is
a nightmare to handle in the block device driver and file system code.

The accounting is the rt limit thing and is litteraly 10 lines of
code so i would not see that as hard to replicate.


> > > Now We are talking about SVA and PASID, just to make sure WarpDrive can benefit
> > > from the feature in the future. It dose not means WarpDrive is useless before
> > > that. And it works for our Zip and RSA accelerators in physical world.
> > 
> > Just not with random process address ...
> > 
> > > > If you still want non SVA/SVM path what you want to do only works
> > > > if both ptr1 and ptr2 are in a range that is DMA mapped to the
> > > > device (moreover you need DMA address to match process address
> > > > which is not an easy feat).
> > > > 
> > > > Now even if you only want SVA/SVM, i do not see what is the point
> > > > of doing this inside VFIO. AMD GPU driver does not and there would
> > > > be no benefit for them to be there. Well a AMD VFIO mdev device
> > > > driver for QEMU guest might be useful but they have SVIO IIRC.
> > > > 
> > > > For SVA/SVM your usage model is:
> > > > 
> > > > Setup:
> > > >     - user space create a warp drive context for the process
> > > >     - user space create a device specific context for the process
> > > >     - user space create a user space command queue for the device
> > > >     - user space bind command queue
> > > > 
> > > >     At this point the kernel driver has bound the process address
> > > >     space to the device with a command queue and userspace
> > > > 
> > > > Usage:
> > > >     - user space schedule work and call appropriate flush/update
> > > >       ioctl from time to time. Might be optional depends on the
> > > >       hardware, but probably a good idea to enforce so that kernel
> > > >       can unbind the command queue to bind another process command
> > > >       queue.
> > > >     ...
> > > > 
> > > > Cleanup:
> > > >     - user space unbind command queue
> > > >     - user space destroy device specific context
> > > >     - user space destroy warp drive context
> > > >     All the above can be implicit when closing the device file.
> > > > 
> > > > So again in the above model i do not see anywhere something from
> > > > VFIO that would benefit this model.
> > > > 
> > > 
> > > Let me show you how the model will be if I use VFIO:
> > > 
> > > Setup (Kernel part)
> > > 	- Kernel driver do every as usual to serve the other functionality, NIC
> > > 	  can still be registered to netdev, encryptor can still be registered
> > > 	  to crypto...
> > > 	- At the same time, the driver can devote some of its hardware resource
> > > 	  and register them as a mdev creator to the VFIO framework. This just
> > > 	  need limited change to the VFIO type1 driver.
> > 
> > In the above VFIO does not help you one bit ... you can do that with
> > as much code with new common device as front end.
> > 
> > > Setup (User space)
> > > 	- System administrator create mdev via the mdev creator interface.
> > > 	- Following VFIO setup routine, user space open the mdev's group, there is
> > > 	  only one group for one device.
> > > 	- Without PASID support, you don't need to do anything. With PASID, bind
> > > 	  the PASID to the device via VFIO interface.
> > > 	- Get the device from the group via VFIO interface and mmap it the user
> > > 	  space for device's MMIO access (for the queue).
> > > 	- Map whatever memory you need to share with the device with VFIO
> > > 	  interface.
> > > 	- (opt) Add more devices into the container if you want to share the
> > > 	  same address space with them
> > 
> > So all VFIO buys you here is boiler plate code that does insert_pfn()
> > to handle MMIO mapping. Which is just couple hundred lines of boiler
> > plate code.
> > 
> 
> No. With VFIO, I don't need to:
> 
> 1. GUP and accounting for RLIMIT_MEMLOCK

That's 10 line of code ...

> 2. Keep all GUP pages for releasing (VFIO uses the rb_tree to do so)

GUP pages are not part of rb_tree and what you want to do can be done
in few lines of code here is pseudo code:

warp_dma_map_range(ulong vaddr, ulong npages)
{
    struct page *pages = kvzalloc(npages);

    for (i = 0; i < npages; ++i, vaddr += PAGE_SIZE) {
        GUP(vaddr, &pages[i]);
        iommu_map(vaddr, page_to_pfn(pages[i]));
    }
    kvfree(pages);
}

warp_dma_unmap_range(ulong vaddr, ulong npages)
{
    for (i = 0; i < npages; ++i, vaddr += PAGE_SIZE) {
        unsigned long pfn;

        pfn = iommu_iova_to_phys(vaddr);
        iommu_unmap(vaddr);
        put_page(pfn_to_page(page)); /* set dirty if mapped write */
    }
}

Add locking, error handling, dirtying and comments and you are barely
looking at couple hundred lines of code. You do not need any of the
complexity of VFIO as you do not have the same requirements. Namely
VFIO have to keep track of iova and physical mapping for things like
migration (migrating guest between host) and few others very
virtualization centric requirements.


> 2. Handle the PASID on SMMU (ARM's IOMMU) myself.

Existing driver do that with 20 lines of with comments and error
handling (see kfd_iommu_bind_process_to_device() for instance) i
doubt you need much more than that.


> 3. Multiple devices menagement (VFIO uses container to manage this)

All the vfio_group* stuff ? OK that's boiler plate code, note that
hard to replicate thought.

> And even as a boiler plate, it is valueable, the memory thing is sensitive
> interface to user space, it can easily become a security problem. If I can
> achieve my target within the scope of VFIO, why not? At lease it has been
> proved to be safe for the time being.

The thing is being part of VFIO impose things on you, things that you
do not need. Like one device per group (maybe it is you imposing this,
i am loosing track here). Or the complex dma mapping tracking ...


> > > Cleanup:
> > > 	- User space close the group file handler
> > > 	- There will be a problem to let the other process know the mdev is
> > > 	  freed to be used again. My RFCv1 choose a file handler solution. Alex
> > > 	  dose not like it. But it is not a big problem. We can always have a
> > > 	  scheduler process to manage the state of the mdev or even we can
> > > 	  switch back to the RFCv1 solution without too much effort if we like
> > > 	  in the future.
> > 
> > If you were outside VFIO you would have more freedom on how to do that.
> > For instance process opening the device file can be placed on queue and
> > first one in the queue get to use the device until it closes/release the
> > device. Then next one in queue get the device ...
> 
> Yes. I do like the file handle solution. But I hope the solution become mature
> as soon as possible. Many of our products, and as I know include some of our
> partners, are waiting for a long term solution as direction. If I rely on some
> unmature solution, they may choose some deviated, customized solution. That will
> be much harmful. Compare to this, the freedom is not so important...

I do not see how being part of VFIO protect you from people doing crazy
thing to their kernel ... Time to market being key in this world, i doubt
that being part of VFIO would make anyone think twice before taking a
shortcut.

I have seen horrible things on that front and only players like Google
can impose a minimum level of sanity.


> > > Except for the minimum update to the type1 driver and use sdmdev to manage the
> > > interrupt sharing, I don't need any extra code to gain the address sharing
> > > capability. And the capability will be strengthen along with the upgrade of VFIO.
> > > 
> > > > 
> > > > > > > And I don't understand why I should avoid to use VFIO? As Alex said, VFIO is the
> > > > > > > user driver framework. And I need exactly a user driver interface. Why should I
> > > > > > > invent another wheel? It has most of stuff I need:
> > > > > > > 
> > > > > > > 1. Connecting multiple devices to the same application space
> > > > > > > 2. Pinning and DMA from the application space to the whole set of device
> > > > > > > 3. Managing hardware resource by device
> > > > > > > 
> > > > > > > We just need the last step: make sure multiple applications and the kernel can
> > > > > > > share the same IOMMU. Then why shouldn't we use VFIO?
> > > > > > 
> > > > > > Because tons of other drivers already do all of the above outside VFIO. Many
> > > > > > driver have a sizeable userspace side to them (anything with ioctl do) so they
> > > > > > can be construded as userspace driver too.
> > > > > > 
> > > > > 
> > > > > Ignoring if there are *tons* of drivers are doing that;), even I do the same as
> > > > > i915 and solve the address space problem. And if I don't need to with VFIO, why
> > > > > should I spend so much effort to do it again?
> > > > 
> > > > Because you do not need any code from VFIO, nor do you need to reinvent
> > > > things. If non SVA/SVM matters to you then use dma buffer. If not then
> > > > i do not see anything in VFIO that you need.
> > > > 
> > > 
> > > As I have explain, if I don't use VFIO, at lease I have to do all that has been
> > > done in i915 or even more than that.
> > 
> > So beside the MMIO mmap() handling and dma mapping of range of user space
> > address space (again all very boiler plate code duplicated accross the
> > kernel several time in different forms). You do not gain anything being
> > inside VFIO right ?
> > 
> 
> As I said, rb-tree for gup, rlimit accounting, cooperation on SMMU, and mature
> user interface are our concern.
> > 
> > > > > > So there is no reasons to do that under VFIO. Especialy as in your example
> > > > > > it is not a real user space device driver, the userspace portion only knows
> > > > > > about writting command into command buffer AFAICT.
> > > > > > 
> > > > > > VFIO is for real userspace driver where interrupt, configurations, ... ie
> > > > > > all the driver is handled in userspace. This means that the userspace have
> > > > > > to be trusted as it could program the device to do DMA to anywhere (if
> > > > > > IOMMU is disabled at boot which is still the default configuration in the
> > > > > > kernel).
> > > > > > 
> > > > > 
> > > > > But as Alex explained, VFIO is not simply used by VM. So it need not to have all
> > > > > stuffs as a driver in host system. And I do need to share the user space as DMA
> > > > > buffer to the hardware. And I can get it with just a little update, then it can
> > > > > service me perfectly. I don't understand why I should choose a long route.
> > > > 
> > > > Again this is not the long route i do not see anything in VFIO that
> > > > benefit you in the SVA/SVM case. A basic character device driver can
> > > > do that.
> > > > 
> > > > 
> > > > > > So i do not see any reasons to do anything you want inside VFIO. All you
> > > > > > want to do can be done outside as easily. Moreover it would be better if
> > > > > > you define clearly each scenario because from where i sit it looks like
> > > > > > you are opening the door wide open to userspace to DMA anywhere when IOMMU
> > > > > > is disabled.
> > > > > > 
> > > > > > When IOMMU is disabled you can _not_ expose command queue to userspace
> > > > > > unless your device has its own page table and all commands are relative
> > > > > > to that page table and the device page table is populated by kernel driver
> > > > > > in secure way (ie by checking that what is populated can be access).
> > > > > > 
> > > > > > I do not believe your example device to have such page table nor do i see
> > > > > > a fallback path when IOMMU is disabled that force user to do ioctl for
> > > > > > each commands.
> > > > > > 
> > > > > > Yes i understand that you target SVA/SVM but still you claim to support
> > > > > > non SVA/SVM. The point is that userspace can not be trusted if you want
> > > > > > to have random program use your device. I am pretty sure that all user
> > > > > > of VFIO are trusted process (like QEMU).
> > > > > > 
> > > > > > 
> > > > > > Finaly i am convince that the IOMMU grouping stuff related to VFIO is
> > > > > > useless for your usecase. I really do not see the point of that, it
> > > > > > does complicate things for you for no reasons AFAICT.
> > > > > 
> > > > > Indeed, I don't like the group thing. I believe VFIO's maintains would not like
> > > > > it very much either;). But the problem is, the group reflects to the same
> > > > > IOMMU(unit), which may shared with other devices.  It is a security problem. I
> > > > > cannot ignore it. I have to take it into account event I don't use VFIO.
> > > > 
> > > > To me it seems you are making a policy decission in kernel space ie
> > > > wether the device should be isolated in its own group or not is a
> > > > decission that is up to the sys admin or something in userspace.
> > > > Right now existing user of SVA/SVM don't (at least AFAICT).
> > > > 
> > > > Do we really want to force such isolation ?
> > > > 
> > > 
> > > But it is not my decision, that how the iommu subsystem is designed. Personally
> > > I don't like it at all, because all our hardwares have their own stream id
> > > (device id). I don't need the group concept at all. But the iommu subsystem
> > > assume some devices may share the name device ID to a single IOMMU.
> > 
> > My question was do you really want to force group isolation for the
> > device ? Existing SVA/SVM capable driver do not force that, they let
> > the userspace decide this (sysadm, distributions, ...). Being part of
> > VFIO (in the way you do, likely ways to avoid this inside VFIO too)
> > force this decision ie make a policy decision without userspace having
> > anything to say about it.

You still do not answer my question, do you really want to force group
isolation for device in your framework ? Which is a policy decision from
my POV and thus belong to userspace and should not be enforce by kernel.


> > The IOMMU group thing as always been doubt full to me, it is advertise
> > as allowing to share resources (ie IOMMU page table) between devices.
> > But this assume that all device driver in the group have some way of
> > communicating with each other to share common DMA address that point
> > to memory devices care. I believe only VFIO does that and probably
> > only when use by QEMU.
> > 
> > 
> > Anyway my question is:
> > 
> > Is it that much useful to be inside VFIO (to avoid few hundred lines
> > of boiler plate code) given that it forces you into a model (group
> > isolation) that so far have never been the prefered way for all
> > existing device driver that already do what you want to achieve ?
> > 
> 
> You mean to say I create another framework and copy most of the code from VFIO?
> It is hard to believe the mainline kernel will take my code. So how about let me
> try the VFIO way first and try that if it won't work? ;)

There is no trying, this is the kernel, once you expose something to
userspace you have to keep supporting it forever ... There is no, hey
let's add this new framework and see how it goes and removing it few
kernel version latter ...

That is why i am being pedantic :) on making sure there is good reasons
to do what you do inside VFIO. I do believe that we want a common frame-
work like the one you are proposing but i do not believe it should be
part of VFIO given the baggages it comes with and that are not relevant
to the use cases for this kind of devices.

Cheers,
Jérôme

Powered by blists - more mailing lists