Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Tue, Nov 20, 2018 at 10:30:55AM +0800, Kenneth Lee wrote: > Date: Tue, 20 Nov 2018 10:30:55 +0800 > From: Kenneth Lee > To: Leon Romanovsky > CC: Tim Sell , linux-...@vger.kernel.org, > Alexander Shishkin , Zaibo Xu > , zhangfei@foxmail.com, linux...@huawei.com, > haojian.zhu...@linaro.org, Christoph Lameter , Hao Fang > , Gavin Schenk , RDMA mailing > list , Zhou Wang , > Jason Gunthorpe , Doug Ledford , Uwe > Kleine-König , David Kershner > , Kenneth Lee , Johan > Hovold , Jerome Glisse , Cyrille > Pitchen , Sagar Dharia > , Jens Axboe , > guodong...@linaro.org, linux-netdev , Randy Dunlap > , linux-ker...@vger.kernel.org, Vinod Koul > , linux-crypto@vger.kernel.org, Philippe Ombredanne > , Sanyog Kale , "David S. > Miller" , linux-accelerat...@lists.ozlabs.org > Subject: Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce > User-Agent: Mutt/1.5.21 (2010-09-15) > Message-ID: <20181120023055.GG157308@Turing-Arch-b> > > On Mon, Nov 19, 2018 at 12:48:01PM +0200, Leon Romanovsky wrote: > > Date: Mon, 19 Nov 2018 12:48:01 +0200 > > From: Leon Romanovsky > > To: Kenneth Lee > > CC: Tim Sell , linux-...@vger.kernel.org, > > Alexander Shishkin , Zaibo Xu > > , zhangfei@foxmail.com, linux...@huawei.com, > > haojian.zhu...@linaro.org, Christoph Lameter , Hao Fang > > , Gavin Schenk , RDMA mailing > > list , Vinod Koul , Jason > > Gunthorpe , Doug Ledford , Uwe > > Kleine-König , David Kershner > > , Kenneth Lee , Johan > > Hovold , Cyrille Pitchen > > , Sagar Dharia > > , Jens Axboe , > > guodong...@linaro.org, linux-netdev , Randy Dunlap > > , linux-ker...@vger.kernel.org, Zhou Wang > > , linux-crypto@vger.kernel.org, Philippe > > Ombredanne , Sanyog Kale , > > "David S. Miller" , > > linux-accelerat...@lists.ozlabs.org, Jerome Glisse > > Subject: Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce > > User-Agent: Mutt/1.10.1 (2018-07-13) > > Message-ID: <20181119104801.gf8...@mtr-leonro.mtl.com> > > > > On Mon, Nov 19, 2018 at 05:19:10PM +0800, Kenneth Lee wrote: > > > On Mon, Nov 19, 2018 at 05:14:05PM +0800, Kenneth Lee wrote: > > > > Date: Mon, 19 Nov 2018 17:14:05 +0800 > > > > From: Kenneth Lee > > > > To: Leon Romanovsky > > > > CC: Tim Sell , linux-...@vger.kernel.org, > > > > Alexander Shishkin , Zaibo Xu > > > > , zhangfei@foxmail.com, linux...@huawei.com, > > > > haojian.zhu...@linaro.org, Christoph Lameter , Hao Fang > > > > , Gavin Schenk , RDMA > > > > mailing > > > > list , Vinod Koul , Jason > > > > Gunthorpe , Doug Ledford , Uwe > > > > Kleine-König , David Kershner > > > > , Kenneth Lee , Johan > > > > Hovold , Cyrille Pitchen > > > > , Sagar Dharia > > > > , Jens Axboe , > > > > guodong...@linaro.org, linux-netdev , Randy > > > > Dunlap > > > > , linux-ker...@vger.kernel.org, Zhou Wang > > > > , linux-crypto@vger.kernel.org, Philippe > > > > Ombredanne , Sanyog Kale > > > > , > > > > "David S. Miller" , > > > > linux-accelerat...@lists.ozlabs.org > > > > Subject: Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce > > > > User-Agent: Mutt/1.5.21 (2010-09-15) > > > > Message-ID: <20181119091405.GE157308@Turing-Arch-b> > > > > > > > > On Thu, Nov 15, 2018 at 04:54:55PM +0200, Leon Romanovsky wrote: > > > > > Date: Thu, 15 Nov 2018 16:54:55 +0200 > > > > > From: Leon Romanovsky > > > > > To: Kenneth Lee > > > > > CC: Kenneth Lee , Tim Sell > > > > > , > > > > > linux-...@vger.kernel.org, Alexander Shishkin > > > > > , Zaibo Xu , > > > > > zhangfei@foxmail.com, linux...@huawei.com, > > > > > haojian.zhu...@linaro.org, > > > > > Christoph Lameter , Hao Fang , > > > > > Gavin > > > > > Schenk , RDMA mailing list > > > > > , Zhou Wang , > > > > > Jason > > > > > Gunthorpe , Doug Ledford , Uwe > > > > > Kleine-König , David Kershner > > > > > , Johan Hovold , Cyrille > > > > > Pitchen , Sagar Dharia > > > > > , Jens Axboe , > > > > > guodong...@linaro.org, linux-netdev , Randy > > > > > Dunlap > > > > > , linux-ker...@v
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Fri, Nov 23, 2018 at 11:05:04AM -0700, Jason Gunthorpe wrote: > Date: Fri, 23 Nov 2018 11:05:04 -0700 > From: Jason Gunthorpe > To: Kenneth Lee > CC: Leon Romanovsky , Kenneth Lee , > Tim Sell , linux-...@vger.kernel.org, Alexander > Shishkin , Zaibo Xu > , zhangfei@foxmail.com, linux...@huawei.com, > haojian.zhu...@linaro.org, Christoph Lameter , Hao Fang > , Gavin Schenk , RDMA mailing > list , Zhou Wang , > Doug Ledford , Uwe Kleine-König > , David Kershner > , Johan Hovold , Cyrille > Pitchen , Sagar Dharia > , Jens Axboe , > guodong...@linaro.org, linux-netdev , Randy Dunlap > , linux-ker...@vger.kernel.org, Vinod Koul > , linux-crypto@vger.kernel.org, Philippe Ombredanne > , Sanyog Kale , "David S. > Miller" , linux-accelerat...@lists.ozlabs.org > Subject: Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce > User-Agent: Mutt/1.9.4 (2018-02-28) > Message-ID: <20181123180504.ga3...@ziepe.ca> > > On Fri, Nov 23, 2018 at 04:02:42PM +0800, Kenneth Lee wrote: > > > It is already part of Jean's patchset. And that's why I built my solution on > > VFIO in the first place. But I think the concept of SVA and PASID is not > > compatible with the original VFIO concept space. You would not share your > > whole > > address space to a device at all in a virtual machine manager, > > wouldn't you? > > Why not? That seems to fit VFIO's space just fine to me.. You might > need a new upcall to create a full MM registration, but that doesn't > seem unsuited. Because the VM manager (such as qemu) do not want to share its whole space to the device. It is a security problem. > > Part of the point here is you should try to make sensible revisions to > existing subsystems before just inventing a new thing... > > VFIO is deeply connected to the IOMMU, so enabling more general IOMMU > based approache seems perfectly fine to me.. > > > > Once the VFIO driver knows about this as a generic capability then the > > > device it exposes to userspace would use CPU addresses instead of DMA > > > addresses. > > > > > > The question is if your driver needs much more than the device > > > agnostic generic services VFIO provides. > > > > > > I'm not sure what you have in mind with resource management.. It is > > > hard to revoke resources from userspace, unless you are doing > > > kernel syscalls, but then why do all this? > > > > Say, I have 1024 queues in my accelerator. I can get one by opening the > > device > > and attach it with the fd. If the process exit by any means, the queue can > > be > > returned with the release of the fd. But if it is mdev, it will still be > > there > > and some one should tell the allocator it is available again. This is not > > easy > > to design in user space. > > ?? why wouldn't the mdev track the queues assigned using the existing > open/close/ioctl callbacks? > > That is basic flow I would expect: > > open(/dev/vfio) > ioctl(unity map entire process MM to mdev with IOMMU) > > // Create a HQ queue and link the PASID in the HW to this HW queue > struct hw queue[..]; > ioctl(create HW queue) > > // Get BAR doorbell memory for the queue > bar = mmap() > > // Submit work to the queue using CPU addresses > queue[0] = ... > writel(bar [..], &queue); > > // Queue, SVA, etc is cleaned up when the VFIO closes > close() This is not the way that you can use mdev. To use mdev, you have to: 1. unbind kernel driver from the device, and rebind it to vfio driver 2. for 0 to 1204: uuid > /sys/.../the_dev/mdev/create to create all the mdev 3. a virtual iommu_group will be created in /dev/vfio/* from every mdev now you can do this in you application (even without considering the pasid) : container = open(/dev/vfio); ioctl(container, settting); group = open(/dev/vfio/my_group_for_particular_mdev); ioctl(container, attach_group, group); device = ioctl(group, get_device); mmap(device); ioctl(container, set_dma_operation); Then you have to make a decision, how can you find a available mdev for use and how to return it. We have considered creating only one mdev and allocating queue when the device is openned. But the VFIO maintainer, Alex, did not agree and said it broke the VFIO origin idea. -Kenneth > > Presumably the kernel has to handle the PASID and related for security > reasons, so they shouldn't go to userspace? > > If there is something missing in vfio to do this is it looks pretty > small to me.. > > Jason -- -K
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Fri, Nov 23, 2018 at 04:02:42PM +0800, Kenneth Lee wrote: > It is already part of Jean's patchset. And that's why I built my solution on > VFIO in the first place. But I think the concept of SVA and PASID is not > compatible with the original VFIO concept space. You would not share your > whole > address space to a device at all in a virtual machine manager, > wouldn't you? Why not? That seems to fit VFIO's space just fine to me.. You might need a new upcall to create a full MM registration, but that doesn't seem unsuited. Part of the point here is you should try to make sensible revisions to existing subsystems before just inventing a new thing... VFIO is deeply connected to the IOMMU, so enabling more general IOMMU based approache seems perfectly fine to me.. > > Once the VFIO driver knows about this as a generic capability then the > > device it exposes to userspace would use CPU addresses instead of DMA > > addresses. > > > > The question is if your driver needs much more than the device > > agnostic generic services VFIO provides. > > > > I'm not sure what you have in mind with resource management.. It is > > hard to revoke resources from userspace, unless you are doing > > kernel syscalls, but then why do all this? > > Say, I have 1024 queues in my accelerator. I can get one by opening the device > and attach it with the fd. If the process exit by any means, the queue can be > returned with the release of the fd. But if it is mdev, it will still be there > and some one should tell the allocator it is available again. This is not easy > to design in user space. ?? why wouldn't the mdev track the queues assigned using the existing open/close/ioctl callbacks? That is basic flow I would expect: open(/dev/vfio) ioctl(unity map entire process MM to mdev with IOMMU) // Create a HQ queue and link the PASID in the HW to this HW queue struct hw queue[..]; ioctl(create HW queue) // Get BAR doorbell memory for the queue bar = mmap() // Submit work to the queue using CPU addresses queue[0] = ... writel(bar [..], &queue); // Queue, SVA, etc is cleaned up when the VFIO closes close() Presumably the kernel has to handle the PASID and related for security reasons, so they shouldn't go to userspace? If there is something missing in vfio to do this is it looks pretty small to me.. Jason
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Wed, Nov 21, 2018 at 07:58:40PM -0700, Jason Gunthorpe wrote: > Date: Wed, 21 Nov 2018 19:58:40 -0700 > From: Jason Gunthorpe > To: Kenneth Lee > CC: Leon Romanovsky , Kenneth Lee , > Tim Sell , linux-...@vger.kernel.org, Alexander > Shishkin , Zaibo Xu > , zhangfei@foxmail.com, linux...@huawei.com, > haojian.zhu...@linaro.org, Christoph Lameter , Hao Fang > , Gavin Schenk , RDMA mailing > list , Zhou Wang , > Doug Ledford , Uwe Kleine-König > , David Kershner > , Johan Hovold , Cyrille > Pitchen , Sagar Dharia > , Jens Axboe , > guodong...@linaro.org, linux-netdev , Randy Dunlap > , linux-ker...@vger.kernel.org, Vinod Koul > , linux-crypto@vger.kernel.org, Philippe Ombredanne > , Sanyog Kale , "David S. > Miller" , linux-accelerat...@lists.ozlabs.org > Subject: Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce > User-Agent: Mutt/1.9.4 (2018-02-28) > Message-ID: <20181122025840.gb19...@ziepe.ca> > > On Wed, Nov 21, 2018 at 02:08:05PM +0800, Kenneth Lee wrote: > > > > But considering Jean's SVA stuff seems based on mmu notifiers, I have > > > a hard time believing that it has any different behavior from RDMA's > > > ODP, and if it does have different behavior, then it is probably just > > > a bug in the ODP implementation. > > > > As Jean has explained, his solution is based on page table sharing. I think > > ODP > > should also consider this new feature. > > Shared page tables would require the HW to walk the page table format > of the CPU directly, not sure how that would be possible for ODP? > > Presumably the implementation for ARM relies on the IOMMU hardware > doing this? Yes, that is the idea. And since Jean is merging the AMD and Intel solution together, I assume they can do the same. This is also the reason I want to solve my problem on top of IOMMU directly. But anyway, let me try to see if I can merge the logic with ODP. > > > > > > If all your driver needs is to mmap some PCI bar space, route > > > > > interrupts and do DMA mapping then mediated VFIO is probably a good > > > > > choice. > > > > > > > > Yes. That is what is done in our RFCv1/v2. But we accepted Jerome's > > > > opinion and > > > > try not to add complexity to the mm subsystem. > > > > > > Why would a mediated VFIO driver touch the mm subsystem? Sounds like > > > you don't have a VFIO driver if it needs to do stuff like that... > > > > VFIO has no ODP-like solution, and if we want to solve the fork problem, we > > have > > to make some change to iommu and the fork procedure. Further, VFIO takes > > every > > queue as a independent device. This create a lot of trouble on resource > > management. For example, you will need a manager process to withdraw the > > unused > > device and you need to let the user process know about PASID of the queue, > > and > > so on. > > Well, I would think you'd add SVA support to the VFIO driver as a > generic capability - it seems pretty useful for any VFIO user as it > avoids all the kernel upcalls to do memory pinning and DMA address > translation. It is already part of Jean's patchset. And that's why I built my solution on VFIO in the first place. But I think the concept of SVA and PASID is not compatible with the original VFIO concept space. You would not share your whole address space to a device at all in a virtual machine manager, wouldn't you? And if you can manage to have a separated mdev for your virtual machine, why bother to set a PASID to it? The answer to those problem, I think, will be Intel's Scalable IO Virtualization. For accelerator, the requirement is simply: getting a handle to device, attaching the process's mm with the handle by sharing the process's page table with its iommu indexed by PASID, and start the communication... > > Once the VFIO driver knows about this as a generic capability then the > device it exposes to userspace would use CPU addresses instead of DMA > addresses. > > The question is if your driver needs much more than the device > agnostic generic services VFIO provides. > > I'm not sure what you have in mind with resource management.. It is > hard to revoke resources from userspace, unless you are doing > kernel syscalls, but then why do all this? Say, I have 1024 queues in my accelerator. I can get one by opening the device and attach it with the fd. If the process exit by any means, the queue can be returned with the release of the fd. But if it is mdev, it will still be there and some one should tell the allocator it is available again. This is not easy to design in user space. > > Jason --
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Wed, Nov 21, 2018 at 02:08:05PM +0800, Kenneth Lee wrote: > > But considering Jean's SVA stuff seems based on mmu notifiers, I have > > a hard time believing that it has any different behavior from RDMA's > > ODP, and if it does have different behavior, then it is probably just > > a bug in the ODP implementation. > > As Jean has explained, his solution is based on page table sharing. I think > ODP > should also consider this new feature. Shared page tables would require the HW to walk the page table format of the CPU directly, not sure how that would be possible for ODP? Presumably the implementation for ARM relies on the IOMMU hardware doing this? > > > > If all your driver needs is to mmap some PCI bar space, route > > > > interrupts and do DMA mapping then mediated VFIO is probably a good > > > > choice. > > > > > > Yes. That is what is done in our RFCv1/v2. But we accepted Jerome's > > > opinion and > > > try not to add complexity to the mm subsystem. > > > > Why would a mediated VFIO driver touch the mm subsystem? Sounds like > > you don't have a VFIO driver if it needs to do stuff like that... > > VFIO has no ODP-like solution, and if we want to solve the fork problem, we > have > to make some change to iommu and the fork procedure. Further, VFIO takes every > queue as a independent device. This create a lot of trouble on resource > management. For example, you will need a manager process to withdraw the > unused > device and you need to let the user process know about PASID of the queue, and > so on. Well, I would think you'd add SVA support to the VFIO driver as a generic capability - it seems pretty useful for any VFIO user as it avoids all the kernel upcalls to do memory pinning and DMA address translation. Once the VFIO driver knows about this as a generic capability then the device it exposes to userspace would use CPU addresses instead of DMA addresses. The question is if your driver needs much more than the device agnostic generic services VFIO provides. I'm not sure what you have in mind with resource management.. It is hard to revoke resources from userspace, unless you are doing kernel syscalls, but then why do all this? Jason
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Mon, Nov 19, 2018 at 08:29:39PM -0700, Jason Gunthorpe wrote: > Date: Mon, 19 Nov 2018 20:29:39 -0700 > From: Jason Gunthorpe > To: Kenneth Lee > CC: Leon Romanovsky , Kenneth Lee , > Tim Sell , linux-...@vger.kernel.org, Alexander > Shishkin , Zaibo Xu > , zhangfei@foxmail.com, linux...@huawei.com, > haojian.zhu...@linaro.org, Christoph Lameter , Hao Fang > , Gavin Schenk , RDMA mailing > list , Zhou Wang , > Doug Ledford , Uwe Kleine-König > , David Kershner > , Johan Hovold , Cyrille > Pitchen , Sagar Dharia > , Jens Axboe , > guodong...@linaro.org, linux-netdev , Randy Dunlap > , linux-ker...@vger.kernel.org, Vinod Koul > , linux-crypto@vger.kernel.org, Philippe Ombredanne > , Sanyog Kale , "David S. > Miller" , linux-accelerat...@lists.ozlabs.org > Subject: Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce > User-Agent: Mutt/1.9.4 (2018-02-28) > Message-ID: <20181120032939.gr4...@ziepe.ca> > > On Tue, Nov 20, 2018 at 11:07:02AM +0800, Kenneth Lee wrote: > > On Mon, Nov 19, 2018 at 11:49:54AM -0700, Jason Gunthorpe wrote: > > > Date: Mon, 19 Nov 2018 11:49:54 -0700 > > > From: Jason Gunthorpe > > > To: Kenneth Lee > > > CC: Leon Romanovsky , Kenneth Lee , > > > Tim Sell , linux-...@vger.kernel.org, Alexander > > > Shishkin , Zaibo Xu > > > , zhangfei@foxmail.com, linux...@huawei.com, > > > haojian.zhu...@linaro.org, Christoph Lameter , Hao Fang > > > , Gavin Schenk , RDMA > > > mailing > > > list , Zhou Wang , > > > Doug Ledford , Uwe Kleine-König > > > , David Kershner > > > , Johan Hovold , Cyrille > > > Pitchen , Sagar Dharia > > > , Jens Axboe , > > > guodong...@linaro.org, linux-netdev , Randy > > > Dunlap > > > , linux-ker...@vger.kernel.org, Vinod Koul > > > , linux-crypto@vger.kernel.org, Philippe Ombredanne > > > , Sanyog Kale , "David S. > > > Miller" , linux-accelerat...@lists.ozlabs.org > > > Subject: Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce > > > User-Agent: Mutt/1.9.4 (2018-02-28) > > > Message-ID: <20181119184954.gb4...@ziepe.ca> > > > > > > On Mon, Nov 19, 2018 at 05:14:05PM +0800, Kenneth Lee wrote: > > > > > > > If the hardware cannot share page table with the CPU, we then need to > > > > have > > > > some way to change the device page table. This is what happen in ODP. It > > > > invalidates the page table in device upon mmu_notifier call back. But > > > > this cannot > > > > solve the COW problem: if the user process A share a page P with > > > > device, and A > > > > forks a new process B, and it continue to write to the page. By COW, the > > > > process B will keep the page P, while A will get a new page P'. But you > > > > have > > > > no way to let the device know it should use P' rather than P. > > > > > > Is this true? I thought mmu_notifiers covered all these cases. > > > > > > The mm_notifier for A should fire if B causes the physical address of > > > A's pages to change via COW. > > > > > > And this causes the device page tables to re-synchronize. > > > > I don't see such code. The current do_cow_fault() implemenation has nothing > > to > > do with mm_notifer. > > Well, that sure sounds like it would be a bug in mmu_notifiers.. Yes, it can be taken that way:) But it is going to be a tough bug. > > But considering Jean's SVA stuff seems based on mmu notifiers, I have > a hard time believing that it has any different behavior from RDMA's > ODP, and if it does have different behavior, then it is probably just > a bug in the ODP implementation. As Jean has explained, his solution is based on page table sharing. I think ODP should also consider this new feature. > > > > > In WarpDrive/uacce, we make this simple. If you support IOMMU and it > > > > support > > > > SVM/SVA. Everything will be fine just like ODP implicit mode. And you > > > > don't need > > > > to write any code for that. Because it has been done by IOMMU > > > > framework. If it > > > > > > Looks like the IOMMU code uses mmu_notifier, so it is identical to > > > IB's ODP. The only difference is that IB tends to have the IOMMU page > > > table in the device, not in the CPU. > > > > > > The only case I know if that is different is th
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Tue, Nov 20, 2018 at 07:17:44AM +0200, Leon Romanovsky wrote: > Date: Tue, 20 Nov 2018 07:17:44 +0200 > From: Leon Romanovsky > To: Kenneth Lee > CC: Jason Gunthorpe , Kenneth Lee , Tim > Sell , linux-...@vger.kernel.org, Alexander > Shishkin , Zaibo Xu > , zhangfei@foxmail.com, linux...@huawei.com, > haojian.zhu...@linaro.org, Christoph Lameter , Hao Fang > , Gavin Schenk , RDMA mailing > list , Zhou Wang , > Doug Ledford , Uwe Kleine-König > , David Kershner > , Johan Hovold , Cyrille > Pitchen , Sagar Dharia > , Jens Axboe , > guodong...@linaro.org, linux-netdev , Randy Dunlap > , linux-ker...@vger.kernel.org, Vinod Koul > , linux-crypto@vger.kernel.org, Philippe Ombredanne > , Sanyog Kale , "David S. > Miller" , linux-accelerat...@lists.ozlabs.org > Subject: Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce > User-Agent: Mutt/1.10.1 (2018-07-13) > Message-ID: <20181120051743.gd25...@mtr-leonro.mtl.com> > > On Tue, Nov 20, 2018 at 11:07:02AM +0800, Kenneth Lee wrote: > > On Mon, Nov 19, 2018 at 11:49:54AM -0700, Jason Gunthorpe wrote: > > > Date: Mon, 19 Nov 2018 11:49:54 -0700 > > > From: Jason Gunthorpe > > > To: Kenneth Lee > > > CC: Leon Romanovsky , Kenneth Lee , > > > Tim Sell , linux-...@vger.kernel.org, Alexander > > > Shishkin , Zaibo Xu > > > , zhangfei@foxmail.com, linux...@huawei.com, > > > haojian.zhu...@linaro.org, Christoph Lameter , Hao Fang > > > , Gavin Schenk , RDMA > > > mailing > > > list , Zhou Wang , > > > Doug Ledford , Uwe Kleine-König > > > , David Kershner > > > , Johan Hovold , Cyrille > > > Pitchen , Sagar Dharia > > > , Jens Axboe , > > > guodong...@linaro.org, linux-netdev , Randy > > > Dunlap > > > , linux-ker...@vger.kernel.org, Vinod Koul > > > , linux-crypto@vger.kernel.org, Philippe Ombredanne > > > , Sanyog Kale , "David S. > > > Miller" , linux-accelerat...@lists.ozlabs.org > > > Subject: Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce > > > User-Agent: Mutt/1.9.4 (2018-02-28) > > > Message-ID: <20181119184954.gb4...@ziepe.ca> > > > > > > On Mon, Nov 19, 2018 at 05:14:05PM +0800, Kenneth Lee wrote: > > > > > > > If the hardware cannot share page table with the CPU, we then need to > > > > have > > > > some way to change the device page table. This is what happen in ODP. It > > > > invalidates the page table in device upon mmu_notifier call back. But > > > > this cannot > > > > solve the COW problem: if the user process A share a page P with > > > > device, and A > > > > forks a new process B, and it continue to write to the page. By COW, the > > > > process B will keep the page P, while A will get a new page P'. But you > > > > have > > > > no way to let the device know it should use P' rather than P. > > > > > > Is this true? I thought mmu_notifiers covered all these cases. > > > > > > The mm_notifier for A should fire if B causes the physical address of > > > A's pages to change via COW. > > > > > > And this causes the device page tables to re-synchronize. > > > > I don't see such code. The current do_cow_fault() implemenation has nothing > > to > > do with mm_notifer. > > > > > > > > > In WarpDrive/uacce, we make this simple. If you support IOMMU and it > > > > support > > > > SVM/SVA. Everything will be fine just like ODP implicit mode. And you > > > > don't need > > > > to write any code for that. Because it has been done by IOMMU > > > > framework. If it > > > > > > Looks like the IOMMU code uses mmu_notifier, so it is identical to > > > IB's ODP. The only difference is that IB tends to have the IOMMU page > > > table in the device, not in the CPU. > > > > > > The only case I know if that is different is the new-fangled CAPI > > > stuff where the IOMMU can directly use the CPU's page table and the > > > IOMMU page table (in device or CPU) is eliminated. > > > > > > > Yes. We are not focusing on the current implementation. As mentioned in the > > cover letter. We are expecting Jean Philips' SVA patch: > > git://linux-arm.org/linux-jpb. > > > > > Anyhow, I don't think a single instance of hardware should justify an > > > entire new subsystem. Su
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On 20/11/2018 09:16, Jonathan Cameron wrote: > +CC Jean-Phillipe and iommu list. Thanks for the Cc, sorry I don't have enough bandwidth to follow this thread at the moment. > In WarpDrive/uacce, we make this simple. If you support IOMMU and it > support > SVM/SVA. Everything will be fine just like ODP implicit mode. And you > don't need > to write any code for that. Because it has been done by IOMMU framework. > If it Looks like the IOMMU code uses mmu_notifier, so it is identical to IB's ODP. The only difference is that IB tends to have the IOMMU page table in the device, not in the CPU. The only case I know if that is different is the new-fangled CAPI stuff where the IOMMU can directly use the CPU's page table and the IOMMU page table (in device or CPU) is eliminated. >>> >>> Yes. We are not focusing on the current implementation. As mentioned in the >>> cover letter. We are expecting Jean Philips' SVA patch: >>> git://linux-arm.org/linux-jpb. >> >> This SVA stuff does not look comparable to CAPI as it still requires >> maintaining seperate IOMMU page tables. With SVA, we use the same page tables in the IOMMU and CPU. It's the same pgd pointer, there is no mirroring of mappings. We bind the process page tables with the device using a PASID (Process Address Space ID). After fork(), the child's mm is different from the parent's one, and is not automatically bound to the device. The device driver will have to issue a new bind() request, and the child mm will be bound with a different PASID. There could be a problem if the child inherits the parent's device handle. Then depending on the device, the child could be able to program DMA and possibly access the parent's address space. The parent needs to be aware of that when using the bind() API, and close the device fd in the child after fork(). We use MMU notifiers for some address space changes: * The IOTLB needs to be invalidated after any unmap() to the process address space. On Arm systems the SMMU IOTLBs can be invalidated by the CPU TLBI instructions, but we still need to invalidate TLBs private to devices that are arch-agnostic (Address Translation Cache in PCI ATS). * When the process mm exits, we need to remove the associated PASID configuration in the IOMMU and invalidate the TLBs. Thanks, Jean
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
+CC Jean-Phillipe and iommu list. On Mon, 19 Nov 2018 20:29:39 -0700 Jason Gunthorpe wrote: > On Tue, Nov 20, 2018 at 11:07:02AM +0800, Kenneth Lee wrote: > > On Mon, Nov 19, 2018 at 11:49:54AM -0700, Jason Gunthorpe wrote: > > > Date: Mon, 19 Nov 2018 11:49:54 -0700 > > > From: Jason Gunthorpe > > > To: Kenneth Lee > > > CC: Leon Romanovsky , Kenneth Lee , > > > Tim Sell , linux-...@vger.kernel.org, Alexander > > > Shishkin , Zaibo Xu > > > , zhangfei@foxmail.com, linux...@huawei.com, > > > haojian.zhu...@linaro.org, Christoph Lameter , Hao Fang > > > , Gavin Schenk , RDMA > > > mailing > > > list , Zhou Wang , > > > Doug Ledford , Uwe Kleine-König > > > , David Kershner > > > , Johan Hovold , Cyrille > > > Pitchen , Sagar Dharia > > > , Jens Axboe , > > > guodong...@linaro.org, linux-netdev , Randy > > > Dunlap > > > , linux-ker...@vger.kernel.org, Vinod Koul > > > , linux-crypto@vger.kernel.org, Philippe Ombredanne > > > , Sanyog Kale , "David S. > > > Miller" , linux-accelerat...@lists.ozlabs.org > > > Subject: Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce > > > User-Agent: Mutt/1.9.4 (2018-02-28) > > > Message-ID: <20181119184954.gb4...@ziepe.ca> > > > > > > On Mon, Nov 19, 2018 at 05:14:05PM +0800, Kenneth Lee wrote: > > > > > > > If the hardware cannot share page table with the CPU, we then need to > > > > have > > > > some way to change the device page table. This is what happen in ODP. It > > > > invalidates the page table in device upon mmu_notifier call back. But > > > > this cannot > > > > solve the COW problem: if the user process A share a page P with > > > > device, and A > > > > forks a new process B, and it continue to write to the page. By COW, the > > > > process B will keep the page P, while A will get a new page P'. But you > > > > have > > > > no way to let the device know it should use P' rather than P. > > > > > > Is this true? I thought mmu_notifiers covered all these cases. > > > > > > The mm_notifier for A should fire if B causes the physical address of > > > A's pages to change via COW. > > > > > > And this causes the device page tables to re-synchronize. > > > > I don't see such code. The current do_cow_fault() implemenation has nothing > > to > > do with mm_notifer. > > Well, that sure sounds like it would be a bug in mmu_notifiers.. > > But considering Jean's SVA stuff seems based on mmu notifiers, I have > a hard time believing that it has any different behavior from RDMA's > ODP, and if it does have different behavior, then it is probably just > a bug in the ODP implementation. > > > > > In WarpDrive/uacce, we make this simple. If you support IOMMU and it > > > > support > > > > SVM/SVA. Everything will be fine just like ODP implicit mode. And you > > > > don't need > > > > to write any code for that. Because it has been done by IOMMU > > > > framework. If it > > > > > > Looks like the IOMMU code uses mmu_notifier, so it is identical to > > > IB's ODP. The only difference is that IB tends to have the IOMMU page > > > table in the device, not in the CPU. > > > > > > The only case I know if that is different is the new-fangled CAPI > > > stuff where the IOMMU can directly use the CPU's page table and the > > > IOMMU page table (in device or CPU) is eliminated. > > > > Yes. We are not focusing on the current implementation. As mentioned in the > > cover letter. We are expecting Jean Philips' SVA patch: > > git://linux-arm.org/linux-jpb. > > This SVA stuff does not look comparable to CAPI as it still requires > maintaining seperate IOMMU page tables. > > Also, those patches from Jean have a lot of references to > mmu_notifiers (ie look at iommu_mmu_notifier). > > Are you really sure it is actually any different at all? > > > > Anyhow, I don't think a single instance of hardware should justify an > > > entire new subsystem. Subsystems are hard to make and without multiple > > > hardware examples there is no way to expect that it would cover any > > > future use cases. > > > > Yes. That's our first expectation. We can keep it with our driver. But > > because > > there
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Tue, Nov 20, 2018 at 11:07:02AM +0800, Kenneth Lee wrote: > On Mon, Nov 19, 2018 at 11:49:54AM -0700, Jason Gunthorpe wrote: > > Date: Mon, 19 Nov 2018 11:49:54 -0700 > > From: Jason Gunthorpe > > To: Kenneth Lee > > CC: Leon Romanovsky , Kenneth Lee , > > Tim Sell , linux-...@vger.kernel.org, Alexander > > Shishkin , Zaibo Xu > > , zhangfei@foxmail.com, linux...@huawei.com, > > haojian.zhu...@linaro.org, Christoph Lameter , Hao Fang > > , Gavin Schenk , RDMA mailing > > list , Zhou Wang , > > Doug Ledford , Uwe Kleine-König > > , David Kershner > > , Johan Hovold , Cyrille > > Pitchen , Sagar Dharia > > , Jens Axboe , > > guodong...@linaro.org, linux-netdev , Randy Dunlap > > , linux-ker...@vger.kernel.org, Vinod Koul > > , linux-crypto@vger.kernel.org, Philippe Ombredanne > > , Sanyog Kale , "David S. > > Miller" , linux-accelerat...@lists.ozlabs.org > > Subject: Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce > > User-Agent: Mutt/1.9.4 (2018-02-28) > > Message-ID: <20181119184954.gb4...@ziepe.ca> > > > > On Mon, Nov 19, 2018 at 05:14:05PM +0800, Kenneth Lee wrote: > > > > > If the hardware cannot share page table with the CPU, we then need to have > > > some way to change the device page table. This is what happen in ODP. It > > > invalidates the page table in device upon mmu_notifier call back. But > > > this cannot > > > solve the COW problem: if the user process A share a page P with device, > > > and A > > > forks a new process B, and it continue to write to the page. By COW, the > > > process B will keep the page P, while A will get a new page P'. But you > > > have > > > no way to let the device know it should use P' rather than P. > > > > Is this true? I thought mmu_notifiers covered all these cases. > > > > The mm_notifier for A should fire if B causes the physical address of > > A's pages to change via COW. > > > > And this causes the device page tables to re-synchronize. > > I don't see such code. The current do_cow_fault() implemenation has nothing to > do with mm_notifer. > > > > > > In WarpDrive/uacce, we make this simple. If you support IOMMU and it > > > support > > > SVM/SVA. Everything will be fine just like ODP implicit mode. And you > > > don't need > > > to write any code for that. Because it has been done by IOMMU framework. > > > If it > > > > Looks like the IOMMU code uses mmu_notifier, so it is identical to > > IB's ODP. The only difference is that IB tends to have the IOMMU page > > table in the device, not in the CPU. > > > > The only case I know if that is different is the new-fangled CAPI > > stuff where the IOMMU can directly use the CPU's page table and the > > IOMMU page table (in device or CPU) is eliminated. > > > > Yes. We are not focusing on the current implementation. As mentioned in the > cover letter. We are expecting Jean Philips' SVA patch: > git://linux-arm.org/linux-jpb. > > > Anyhow, I don't think a single instance of hardware should justify an > > entire new subsystem. Subsystems are hard to make and without multiple > > hardware examples there is no way to expect that it would cover any > > future use cases. > > Yes. That's our first expectation. We can keep it with our driver. But because > there is no user driver support for any accelerator in mainline kernel. Even > the > well known QuickAssit has to be maintained out of tree. So we try to see if > people is interested in working together to solve the problem. > > > > > If all your driver needs is to mmap some PCI bar space, route > > interrupts and do DMA mapping then mediated VFIO is probably a good > > choice. > > Yes. That is what is done in our RFCv1/v2. But we accepted Jerome's opinion > and > try not to add complexity to the mm subsystem. > > > > > If it needs to do a bunch of other stuff, not related to PCI bar > > space, interrupts and DMA mapping (ie special code for compression, > > crypto, AI, whatever) then you should probably do what Jerome said and > > make a drivers/char/hisillicon_foo_bar.c that exposes just what your > > hardware does. > > Yes. If no other accelerator driver writer is interested. That is the > expectation:) > > But we really like to have a public solution here. Consider this scenario: > > You create some connections (queues) to NIC, RSA, and AI engine. Then you got > data direct from t
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Tue, Nov 20, 2018 at 11:07:02AM +0800, Kenneth Lee wrote: > On Mon, Nov 19, 2018 at 11:49:54AM -0700, Jason Gunthorpe wrote: > > Date: Mon, 19 Nov 2018 11:49:54 -0700 > > From: Jason Gunthorpe > > To: Kenneth Lee > > CC: Leon Romanovsky , Kenneth Lee , > > Tim Sell , linux-...@vger.kernel.org, Alexander > > Shishkin , Zaibo Xu > > , zhangfei@foxmail.com, linux...@huawei.com, > > haojian.zhu...@linaro.org, Christoph Lameter , Hao Fang > > , Gavin Schenk , RDMA mailing > > list , Zhou Wang , > > Doug Ledford , Uwe Kleine-König > > , David Kershner > > , Johan Hovold , Cyrille > > Pitchen , Sagar Dharia > > , Jens Axboe , > > guodong...@linaro.org, linux-netdev , Randy Dunlap > > , linux-ker...@vger.kernel.org, Vinod Koul > > , linux-crypto@vger.kernel.org, Philippe Ombredanne > > , Sanyog Kale , "David S. > > Miller" , linux-accelerat...@lists.ozlabs.org > > Subject: Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce > > User-Agent: Mutt/1.9.4 (2018-02-28) > > Message-ID: <20181119184954.gb4...@ziepe.ca> > > > > On Mon, Nov 19, 2018 at 05:14:05PM +0800, Kenneth Lee wrote: > > > > > If the hardware cannot share page table with the CPU, we then need to have > > > some way to change the device page table. This is what happen in ODP. It > > > invalidates the page table in device upon mmu_notifier call back. But > > > this cannot > > > solve the COW problem: if the user process A share a page P with device, > > > and A > > > forks a new process B, and it continue to write to the page. By COW, the > > > process B will keep the page P, while A will get a new page P'. But you > > > have > > > no way to let the device know it should use P' rather than P. > > > > Is this true? I thought mmu_notifiers covered all these cases. > > > > The mm_notifier for A should fire if B causes the physical address of > > A's pages to change via COW. > > > > And this causes the device page tables to re-synchronize. > > I don't see such code. The current do_cow_fault() implemenation has nothing to > do with mm_notifer. Well, that sure sounds like it would be a bug in mmu_notifiers.. But considering Jean's SVA stuff seems based on mmu notifiers, I have a hard time believing that it has any different behavior from RDMA's ODP, and if it does have different behavior, then it is probably just a bug in the ODP implementation. > > > In WarpDrive/uacce, we make this simple. If you support IOMMU and it > > > support > > > SVM/SVA. Everything will be fine just like ODP implicit mode. And you > > > don't need > > > to write any code for that. Because it has been done by IOMMU framework. > > > If it > > > > Looks like the IOMMU code uses mmu_notifier, so it is identical to > > IB's ODP. The only difference is that IB tends to have the IOMMU page > > table in the device, not in the CPU. > > > > The only case I know if that is different is the new-fangled CAPI > > stuff where the IOMMU can directly use the CPU's page table and the > > IOMMU page table (in device or CPU) is eliminated. > > Yes. We are not focusing on the current implementation. As mentioned in the > cover letter. We are expecting Jean Philips' SVA patch: > git://linux-arm.org/linux-jpb. This SVA stuff does not look comparable to CAPI as it still requires maintaining seperate IOMMU page tables. Also, those patches from Jean have a lot of references to mmu_notifiers (ie look at iommu_mmu_notifier). Are you really sure it is actually any different at all? > > Anyhow, I don't think a single instance of hardware should justify an > > entire new subsystem. Subsystems are hard to make and without multiple > > hardware examples there is no way to expect that it would cover any > > future use cases. > > Yes. That's our first expectation. We can keep it with our driver. But because > there is no user driver support for any accelerator in mainline kernel. Even > the > well known QuickAssit has to be maintained out of tree. So we try to see if > people is interested in working together to solve the problem. Well, you should come with patches ack'ed by these other groups. > > If all your driver needs is to mmap some PCI bar space, route > > interrupts and do DMA mapping then mediated VFIO is probably a good > > choice. > > Yes. That is what is done in our RFCv1/v2. But we accepted Jerome's opinion > and > try not to add complexity to the mm subsystem. Why would
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Mon, Nov 19, 2018 at 11:49:54AM -0700, Jason Gunthorpe wrote: > Date: Mon, 19 Nov 2018 11:49:54 -0700 > From: Jason Gunthorpe > To: Kenneth Lee > CC: Leon Romanovsky , Kenneth Lee , > Tim Sell , linux-...@vger.kernel.org, Alexander > Shishkin , Zaibo Xu > , zhangfei@foxmail.com, linux...@huawei.com, > haojian.zhu...@linaro.org, Christoph Lameter , Hao Fang > , Gavin Schenk , RDMA mailing > list , Zhou Wang , > Doug Ledford , Uwe Kleine-König > , David Kershner > , Johan Hovold , Cyrille > Pitchen , Sagar Dharia > , Jens Axboe , > guodong...@linaro.org, linux-netdev , Randy Dunlap > , linux-ker...@vger.kernel.org, Vinod Koul > , linux-crypto@vger.kernel.org, Philippe Ombredanne > , Sanyog Kale , "David S. > Miller" , linux-accelerat...@lists.ozlabs.org > Subject: Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce > User-Agent: Mutt/1.9.4 (2018-02-28) > Message-ID: <20181119184954.gb4...@ziepe.ca> > > On Mon, Nov 19, 2018 at 05:14:05PM +0800, Kenneth Lee wrote: > > > If the hardware cannot share page table with the CPU, we then need to have > > some way to change the device page table. This is what happen in ODP. It > > invalidates the page table in device upon mmu_notifier call back. But this > > cannot > > solve the COW problem: if the user process A share a page P with device, > > and A > > forks a new process B, and it continue to write to the page. By COW, the > > process B will keep the page P, while A will get a new page P'. But you have > > no way to let the device know it should use P' rather than P. > > Is this true? I thought mmu_notifiers covered all these cases. > > The mm_notifier for A should fire if B causes the physical address of > A's pages to change via COW. > > And this causes the device page tables to re-synchronize. I don't see such code. The current do_cow_fault() implemenation has nothing to do with mm_notifer. > > > In WarpDrive/uacce, we make this simple. If you support IOMMU and it support > > SVM/SVA. Everything will be fine just like ODP implicit mode. And you don't > > need > > to write any code for that. Because it has been done by IOMMU framework. If > > it > > Looks like the IOMMU code uses mmu_notifier, so it is identical to > IB's ODP. The only difference is that IB tends to have the IOMMU page > table in the device, not in the CPU. > > The only case I know if that is different is the new-fangled CAPI > stuff where the IOMMU can directly use the CPU's page table and the > IOMMU page table (in device or CPU) is eliminated. > Yes. We are not focusing on the current implementation. As mentioned in the cover letter. We are expecting Jean Philips' SVA patch: git://linux-arm.org/linux-jpb. > Anyhow, I don't think a single instance of hardware should justify an > entire new subsystem. Subsystems are hard to make and without multiple > hardware examples there is no way to expect that it would cover any > future use cases. Yes. That's our first expectation. We can keep it with our driver. But because there is no user driver support for any accelerator in mainline kernel. Even the well known QuickAssit has to be maintained out of tree. So we try to see if people is interested in working together to solve the problem. > > If all your driver needs is to mmap some PCI bar space, route > interrupts and do DMA mapping then mediated VFIO is probably a good > choice. Yes. That is what is done in our RFCv1/v2. But we accepted Jerome's opinion and try not to add complexity to the mm subsystem. > > If it needs to do a bunch of other stuff, not related to PCI bar > space, interrupts and DMA mapping (ie special code for compression, > crypto, AI, whatever) then you should probably do what Jerome said and > make a drivers/char/hisillicon_foo_bar.c that exposes just what your > hardware does. Yes. If no other accelerator driver writer is interested. That is the expectation:) But we really like to have a public solution here. Consider this scenario: You create some connections (queues) to NIC, RSA, and AI engine. Then you got data direct from the NIC and pass the pointer to RSA engine for decryption. The CPU then finish some data taking or operation and then pass through to the AI engine for CNN calculationThis will need a place to maintain the same address space by some means. It is not complex, but it is helpful. > > If you have networking involved in here then consider RDMA, > particularly if this functionality is already part of the same > hardware that the hns infiniband driver is servicing. > > 'computational MRs' are a reasonable approach to a side-car offload of > already existing RDMA support. OK. Thanks. I will spend some time on it. But personally, I really don't like RDMA's complexity. I cannot even try one single function without a...some expensive hardwares and complexity connection in the lab. This is not like a open source way. > > Jason
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Mon, Nov 19, 2018 at 12:48:01PM +0200, Leon Romanovsky wrote: > Date: Mon, 19 Nov 2018 12:48:01 +0200 > From: Leon Romanovsky > To: Kenneth Lee > CC: Tim Sell , linux-...@vger.kernel.org, > Alexander Shishkin , Zaibo Xu > , zhangfei@foxmail.com, linux...@huawei.com, > haojian.zhu...@linaro.org, Christoph Lameter , Hao Fang > , Gavin Schenk , RDMA mailing > list , Vinod Koul , Jason > Gunthorpe , Doug Ledford , Uwe > Kleine-König , David Kershner > , Kenneth Lee , Johan > Hovold , Cyrille Pitchen > , Sagar Dharia > , Jens Axboe , > guodong...@linaro.org, linux-netdev , Randy Dunlap > , linux-ker...@vger.kernel.org, Zhou Wang > , linux-crypto@vger.kernel.org, Philippe > Ombredanne , Sanyog Kale , > "David S. Miller" , > linux-accelerat...@lists.ozlabs.org, Jerome Glisse > Subject: Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce > User-Agent: Mutt/1.10.1 (2018-07-13) > Message-ID: <20181119104801.gf8...@mtr-leonro.mtl.com> > > On Mon, Nov 19, 2018 at 05:19:10PM +0800, Kenneth Lee wrote: > > On Mon, Nov 19, 2018 at 05:14:05PM +0800, Kenneth Lee wrote: > > > Date: Mon, 19 Nov 2018 17:14:05 +0800 > > > From: Kenneth Lee > > > To: Leon Romanovsky > > > CC: Tim Sell , linux-...@vger.kernel.org, > > > Alexander Shishkin , Zaibo Xu > > > , zhangfei@foxmail.com, linux...@huawei.com, > > > haojian.zhu...@linaro.org, Christoph Lameter , Hao Fang > > > , Gavin Schenk , RDMA > > > mailing > > > list , Vinod Koul , Jason > > > Gunthorpe , Doug Ledford , Uwe > > > Kleine-König , David Kershner > > > , Kenneth Lee , Johan > > > Hovold , Cyrille Pitchen > > > , Sagar Dharia > > > , Jens Axboe , > > > guodong...@linaro.org, linux-netdev , Randy > > > Dunlap > > > , linux-ker...@vger.kernel.org, Zhou Wang > > > , linux-crypto@vger.kernel.org, Philippe > > > Ombredanne , Sanyog Kale , > > > "David S. Miller" , > > > linux-accelerat...@lists.ozlabs.org > > > Subject: Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce > > > User-Agent: Mutt/1.5.21 (2010-09-15) > > > Message-ID: <20181119091405.GE157308@Turing-Arch-b> > > > > > > On Thu, Nov 15, 2018 at 04:54:55PM +0200, Leon Romanovsky wrote: > > > > Date: Thu, 15 Nov 2018 16:54:55 +0200 > > > > From: Leon Romanovsky > > > > To: Kenneth Lee > > > > CC: Kenneth Lee , Tim Sell > > > > , > > > > linux-...@vger.kernel.org, Alexander Shishkin > > > > , Zaibo Xu , > > > > zhangfei@foxmail.com, linux...@huawei.com, > > > > haojian.zhu...@linaro.org, > > > > Christoph Lameter , Hao Fang , > > > > Gavin > > > > Schenk , RDMA mailing list > > > > , Zhou Wang , > > > > Jason > > > > Gunthorpe , Doug Ledford , Uwe > > > > Kleine-König , David Kershner > > > > , Johan Hovold , Cyrille > > > > Pitchen , Sagar Dharia > > > > , Jens Axboe , > > > > guodong...@linaro.org, linux-netdev , Randy > > > > Dunlap > > > > , linux-ker...@vger.kernel.org, Vinod Koul > > > > , linux-crypto@vger.kernel.org, Philippe Ombredanne > > > > , Sanyog Kale , "David > > > > S. > > > > Miller" , linux-accelerat...@lists.ozlabs.org > > > > Subject: Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce > > > > User-Agent: Mutt/1.10.1 (2018-07-13) > > > > Message-ID: <20181115145455.gn3...@mtr-leonro.mtl.com> > > > > > > > > On Thu, Nov 15, 2018 at 04:51:09PM +0800, Kenneth Lee wrote: > > > > > On Wed, Nov 14, 2018 at 06:00:17PM +0200, Leon Romanovsky wrote: > > > > > > Date: Wed, 14 Nov 2018 18:00:17 +0200 > > > > > > From: Leon Romanovsky > > > > > > To: Kenneth Lee > > > > > > CC: Tim Sell , linux-...@vger.kernel.org, > > > > > > Alexander Shishkin , Zaibo Xu > > > > > > , zhangfei@foxmail.com, > > > > > > linux...@huawei.com, > > > > > > haojian.zhu...@linaro.org, Christoph Lameter , Hao > > > > > > Fang > > > > > > , Gavin Schenk , RDMA > > > > > > mailing > > > > > > list , Zhou Wang > > > > > > , > > > > > > Jason Gunthorpe , Doug Ledford > > > > > > , U
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Mon, Nov 19, 2018 at 04:33:20PM -0500, Jerome Glisse wrote: > On Mon, Nov 19, 2018 at 02:26:38PM -0700, Jason Gunthorpe wrote: > > On Mon, Nov 19, 2018 at 03:26:15PM -0500, Jerome Glisse wrote: > > > On Mon, Nov 19, 2018 at 01:11:56PM -0700, Jason Gunthorpe wrote: > > > > On Mon, Nov 19, 2018 at 02:46:32PM -0500, Jerome Glisse wrote: > > > > > > > > > > ?? How can O_DIRECT be fine but RDMA not? They use exactly the same > > > > > > get_user_pages flow, right? Can we do what O_DIRECT does in RDMA and > > > > > > be fine too? > > > > > > > > > > > > AFAIK the only difference is the length of the race window. You'd > > > > > > have > > > > > > to fork and fault during the shorter time O_DIRECT has > > > > > > get_user_pages > > > > > > open. > > > > > > > > > > Well in O_DIRECT case there is only one page table, the CPU > > > > > page table and it gets updated during fork() so there is an > > > > > ordering there and the race window is small. > > > > > > > > Not really, in O_DIRECT case there is another 'page table', we just > > > > call it a DMA scatter/gather list and it is sent directly to the block > > > > device's DMA HW. The sgl plays exactly the same role as the various HW > > > > page list data structures that underly RDMA MRs. > > > > > > > > It is not a page table that matters here, it is if the DMA address of > > > > the page is active for DMA on HW. > > > > > > > > Like you say, the only difference is that the race is hopefully small > > > > with O_DIRECT (though that is not really small, NVMeof for instance > > > > has windows as large as connection timeouts, if you try hard enough) > > > > > > > > So we probably can trigger this trouble with O_DIRECT and fork(), and > > > > I would call it a bug :( > > > > > > I can not think of any scenario that would be a bug with O_DIRECT. > > > Do you have one in mind ? When you fork() and do other syscall that > > > affect the memory of your process in another thread you should > > > expect non consistant results. Kernel is not here to provide a fully > > > safe environement to user, user can shoot itself in the foot and > > > that's fine as long as it only affect the process itself and no one > > > else. We should not be in the business of making everything baby > > > proof :) > > > > Sure, I setup AIO with O_DIRECT and launch a read. > > > > Then I fork and dirty the READ target memory using the CPU in the > > child. > > > > As you described in this case the fork will retain the physical page > > that is undergoing O_DIRECT DMA, and the parent gets a new copy'd page. > > > > The DMA completes, and the child gets the DMA'd to page. The parent > > gets an unchanged copy'd page. > > > > The parent gets the AIO completion, but can't see the data. > > > > I'd call that a bug with O_DIRECT. The only correct outcome is that > > the parent will always see the O_DIRECT data. Fork should not cause > > the *parent* to malfunction. I agree the child cannot make any > > prediction what memory it will see. > > > > I assume the same flow is possible using threads and read().. > > > > It is really no different than the RDMA bug with fork. > > > > Yes and that's expected behavior :) If you fork() and have anything > still in flight at time of fork that can change your process address > space (including data in it) then all bets are of. > > At least this is my reading of fork() syscall. Not mine.. I can't think of anything else that would have this behavior. All traditional syscalls, will properly dirty the pages of the parent. ie if I call read() in a thread and do fork in another thread, then not seeing the data after read() completes is clearly a bug. All other syscalls are the same. It is bonkers that opening the file with O_DIRECT would change this basic behavior. I'm calling it a bug :) Jason
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Mon, Nov 19, 2018 at 02:26:38PM -0700, Jason Gunthorpe wrote: > On Mon, Nov 19, 2018 at 03:26:15PM -0500, Jerome Glisse wrote: > > On Mon, Nov 19, 2018 at 01:11:56PM -0700, Jason Gunthorpe wrote: > > > On Mon, Nov 19, 2018 at 02:46:32PM -0500, Jerome Glisse wrote: > > > > > > > > ?? How can O_DIRECT be fine but RDMA not? They use exactly the same > > > > > get_user_pages flow, right? Can we do what O_DIRECT does in RDMA and > > > > > be fine too? > > > > > > > > > > AFAIK the only difference is the length of the race window. You'd have > > > > > to fork and fault during the shorter time O_DIRECT has get_user_pages > > > > > open. > > > > > > > > Well in O_DIRECT case there is only one page table, the CPU > > > > page table and it gets updated during fork() so there is an > > > > ordering there and the race window is small. > > > > > > Not really, in O_DIRECT case there is another 'page table', we just > > > call it a DMA scatter/gather list and it is sent directly to the block > > > device's DMA HW. The sgl plays exactly the same role as the various HW > > > page list data structures that underly RDMA MRs. > > > > > > It is not a page table that matters here, it is if the DMA address of > > > the page is active for DMA on HW. > > > > > > Like you say, the only difference is that the race is hopefully small > > > with O_DIRECT (though that is not really small, NVMeof for instance > > > has windows as large as connection timeouts, if you try hard enough) > > > > > > So we probably can trigger this trouble with O_DIRECT and fork(), and > > > I would call it a bug :( > > > > I can not think of any scenario that would be a bug with O_DIRECT. > > Do you have one in mind ? When you fork() and do other syscall that > > affect the memory of your process in another thread you should > > expect non consistant results. Kernel is not here to provide a fully > > safe environement to user, user can shoot itself in the foot and > > that's fine as long as it only affect the process itself and no one > > else. We should not be in the business of making everything baby > > proof :) > > Sure, I setup AIO with O_DIRECT and launch a read. > > Then I fork and dirty the READ target memory using the CPU in the > child. > > As you described in this case the fork will retain the physical page > that is undergoing O_DIRECT DMA, and the parent gets a new copy'd page. > > The DMA completes, and the child gets the DMA'd to page. The parent > gets an unchanged copy'd page. > > The parent gets the AIO completion, but can't see the data. > > I'd call that a bug with O_DIRECT. The only correct outcome is that > the parent will always see the O_DIRECT data. Fork should not cause > the *parent* to malfunction. I agree the child cannot make any > prediction what memory it will see. > > I assume the same flow is possible using threads and read().. > > It is really no different than the RDMA bug with fork. > Yes and that's expected behavior :) If you fork() and have anything still in flight at time of fork that can change your process address space (including data in it) then all bets are of. At least this is my reading of fork() syscall. Cheers, Jérôme
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Mon, Nov 19, 2018 at 03:26:15PM -0500, Jerome Glisse wrote: > On Mon, Nov 19, 2018 at 01:11:56PM -0700, Jason Gunthorpe wrote: > > On Mon, Nov 19, 2018 at 02:46:32PM -0500, Jerome Glisse wrote: > > > > > > ?? How can O_DIRECT be fine but RDMA not? They use exactly the same > > > > get_user_pages flow, right? Can we do what O_DIRECT does in RDMA and > > > > be fine too? > > > > > > > > AFAIK the only difference is the length of the race window. You'd have > > > > to fork and fault during the shorter time O_DIRECT has get_user_pages > > > > open. > > > > > > Well in O_DIRECT case there is only one page table, the CPU > > > page table and it gets updated during fork() so there is an > > > ordering there and the race window is small. > > > > Not really, in O_DIRECT case there is another 'page table', we just > > call it a DMA scatter/gather list and it is sent directly to the block > > device's DMA HW. The sgl plays exactly the same role as the various HW > > page list data structures that underly RDMA MRs. > > > > It is not a page table that matters here, it is if the DMA address of > > the page is active for DMA on HW. > > > > Like you say, the only difference is that the race is hopefully small > > with O_DIRECT (though that is not really small, NVMeof for instance > > has windows as large as connection timeouts, if you try hard enough) > > > > So we probably can trigger this trouble with O_DIRECT and fork(), and > > I would call it a bug :( > > I can not think of any scenario that would be a bug with O_DIRECT. > Do you have one in mind ? When you fork() and do other syscall that > affect the memory of your process in another thread you should > expect non consistant results. Kernel is not here to provide a fully > safe environement to user, user can shoot itself in the foot and > that's fine as long as it only affect the process itself and no one > else. We should not be in the business of making everything baby > proof :) Sure, I setup AIO with O_DIRECT and launch a read. Then I fork and dirty the READ target memory using the CPU in the child. As you described in this case the fork will retain the physical page that is undergoing O_DIRECT DMA, and the parent gets a new copy'd page. The DMA completes, and the child gets the DMA'd to page. The parent gets an unchanged copy'd page. The parent gets the AIO completion, but can't see the data. I'd call that a bug with O_DIRECT. The only correct outcome is that the parent will always see the O_DIRECT data. Fork should not cause the *parent* to malfunction. I agree the child cannot make any prediction what memory it will see. I assume the same flow is possible using threads and read().. It is really no different than the RDMA bug with fork. Jason
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Mon, Nov 19, 2018 at 01:11:56PM -0700, Jason Gunthorpe wrote: > On Mon, Nov 19, 2018 at 02:46:32PM -0500, Jerome Glisse wrote: > > > > ?? How can O_DIRECT be fine but RDMA not? They use exactly the same > > > get_user_pages flow, right? Can we do what O_DIRECT does in RDMA and > > > be fine too? > > > > > > AFAIK the only difference is the length of the race window. You'd have > > > to fork and fault during the shorter time O_DIRECT has get_user_pages > > > open. > > > > Well in O_DIRECT case there is only one page table, the CPU > > page table and it gets updated during fork() so there is an > > ordering there and the race window is small. > > Not really, in O_DIRECT case there is another 'page table', we just > call it a DMA scatter/gather list and it is sent directly to the block > device's DMA HW. The sgl plays exactly the same role as the various HW > page list data structures that underly RDMA MRs. > > It is not a page table that matters here, it is if the DMA address of > the page is active for DMA on HW. > > Like you say, the only difference is that the race is hopefully small > with O_DIRECT (though that is not really small, NVMeof for instance > has windows as large as connection timeouts, if you try hard enough) > > So we probably can trigger this trouble with O_DIRECT and fork(), and > I would call it a bug :( I can not think of any scenario that would be a bug with O_DIRECT. Do you have one in mind ? When you fork() and do other syscall that affect the memory of your process in another thread you should expect non consistant results. Kernel is not here to provide a fully safe environement to user, user can shoot itself in the foot and that's fine as long as it only affect the process itself and no one else. We should not be in the business of making everything baby proof :) > > > > Why? Keep track in each mm if there are any active get_user_pages > > > FOLL_WRITE pages in the mm, if yes then sweep the VMAs and fix the > > > issue for the FOLL_WRITE pages. > > > > This has a cost and you don't want to do it for O_DIRECT. I am pretty > > sure that any such patch to modify fork() code path would be rejected. > > At least i would not like it and vote against. > > I was thinking the incremental cost on top of what John is already > doing would be very small in the common case and only be triggered in > cases that matter (which apps should avoid anyhow). What John is addressing has nothing to do with fork() it has to do with GUP and filesystem page. More specificaly that after page_mkclean() all filesystem expect that the page content is stable (ie no one write to the page) with GUP and hardware (DIRECT_IO too) this is not necessarily the case. So John is trying to fix that. Not trying to make fork() baby proof AFAICT :) I rather keep saying that you should expect weird thing with RDMA and VFIO when doing fork() than trying to work around this in the kernel. Better behavior through hardware is what we should aim for (CAPI, ODP, ...). Jérôme
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Mon, Nov 19, 2018 at 02:46:32PM -0500, Jerome Glisse wrote: > > ?? How can O_DIRECT be fine but RDMA not? They use exactly the same > > get_user_pages flow, right? Can we do what O_DIRECT does in RDMA and > > be fine too? > > > > AFAIK the only difference is the length of the race window. You'd have > > to fork and fault during the shorter time O_DIRECT has get_user_pages > > open. > > Well in O_DIRECT case there is only one page table, the CPU > page table and it gets updated during fork() so there is an > ordering there and the race window is small. Not really, in O_DIRECT case there is another 'page table', we just call it a DMA scatter/gather list and it is sent directly to the block device's DMA HW. The sgl plays exactly the same role as the various HW page list data structures that underly RDMA MRs. It is not a page table that matters here, it is if the DMA address of the page is active for DMA on HW. Like you say, the only difference is that the race is hopefully small with O_DIRECT (though that is not really small, NVMeof for instance has windows as large as connection timeouts, if you try hard enough) So we probably can trigger this trouble with O_DIRECT and fork(), and I would call it a bug :( > > Why? Keep track in each mm if there are any active get_user_pages > > FOLL_WRITE pages in the mm, if yes then sweep the VMAs and fix the > > issue for the FOLL_WRITE pages. > > This has a cost and you don't want to do it for O_DIRECT. I am pretty > sure that any such patch to modify fork() code path would be rejected. > At least i would not like it and vote against. I was thinking the incremental cost on top of what John is already doing would be very small in the common case and only be triggered in cases that matter (which apps should avoid anyhow). Jason
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Mon, Nov 19, 2018 at 12:27:02PM -0700, Jason Gunthorpe wrote: > On Mon, Nov 19, 2018 at 02:17:21PM -0500, Jerome Glisse wrote: > > On Mon, Nov 19, 2018 at 11:53:33AM -0700, Jason Gunthorpe wrote: > > > On Mon, Nov 19, 2018 at 01:42:16PM -0500, Jerome Glisse wrote: > > > > On Mon, Nov 19, 2018 at 11:27:52AM -0700, Jason Gunthorpe wrote: > > > > > On Mon, Nov 19, 2018 at 11:48:54AM -0500, Jerome Glisse wrote: > > > > > > > > > > > Just to comment on this, any infiniband driver which use umem and do > > > > > > not have ODP (here ODP for me means listening to mmu notifier so all > > > > > > infiniband driver except mlx5) will be affected by same issue > > > > > > AFAICT. > > > > > > > > > > > > AFAICT there is no special thing happening after fork() inside any > > > > > > of > > > > > > those driver. So if parent create a umem mr before fork() and > > > > > > program > > > > > > hardware with it then after fork() the parent might start using new > > > > > > page for the umem range while the old memory is use by the child. > > > > > > The > > > > > > reverse is also true (parent using old memory and child new memory) > > > > > > bottom line you can not predict which memory the child or the parent > > > > > > will use for the range after fork(). > > > > > > > > > > > > So no matter what you consider the child or the parent, what the hw > > > > > > will use for the mr is unlikely to match what the CPU use for the > > > > > > same virtual address. In other word: > > > > > > > > > > > > Before fork: > > > > > > CPU parent: virtual addr ptr1 -> physical address = 0xCAFE > > > > > > HARDWARE: virtual addr ptr1 -> physical address = 0xCAFE > > > > > > > > > > > > Case 1: > > > > > > CPU parent: virtual addr ptr1 -> physical address = 0xCAFE > > > > > > CPU child: virtual addr ptr1 -> physical address = 0xDEAD > > > > > > HARDWARE: virtual addr ptr1 -> physical address = 0xCAFE > > > > > > > > > > > > Case 2: > > > > > > CPU parent: virtual addr ptr1 -> physical address = 0xBEEF > > > > > > CPU child: virtual addr ptr1 -> physical address = 0xCAFE > > > > > > HARDWARE: virtual addr ptr1 -> physical address = 0xCAFE > > > > > > > > > > IIRC this is solved in IB by automatically calling > > > > > madvise(MADV_DONTFORK) before creating the MR. > > > > > > > > > > MADV_DONTFORK > > > > > .. This is useful to prevent copy-on-write semantics from changing > > > > > the > > > > > physical location of a page if the parent writes to it after a > > > > > fork(2) .. > > > > > > > > This would work around the issue but this is not transparent ie > > > > range marked with DONTFORK no longer behave as expected from the > > > > application point of view. > > > > > > Do you know what the difference is? The man page really gives no > > > hint.. > > > > > > Does it sometimes unmap the pages during fork? > > > > It is handled in kernel/fork.c look for DONTCOPY, basicaly it just > > leave empty page table in the child process so child will have to > > fault in new page. This also means that child will get 0 as initial > > value for all memory address under DONTCOPY/DONTFORK which breaks > > application expectation of what fork() do. > > Hum, I wonder why this API was selected then.. Because there is nothing else ? :) > > > > I actually wonder if the kernel is a bit broken here, we have the same > > > problem with O_DIRECT and other stuff, right? > > > > No it is not, O_DIRECT is fine. The only corner case i can think > > of with O_DIRECT is one thread launching an O_DIRECT that write > > to private anonymous memory (other O_DIRECT case do not matter) > > while another thread call fork() then what the child get can be > > undefined ie either it get the data before the O_DIRECT finish > > or it gets the result of the O_DIRECT. But this is realy what > > you should expect when doing such thing without synchronization. > > > > So O_DIRECT is fine. > > ?? How can O_DIRECT be fine but RDMA not? They use exactly the same > get_user_pages flow, right? Can we do what O_DIRECT does in RDMA and > be fine too? > > AFAIK the only difference is the length of the race window. You'd have > to fork and fault during the shorter time O_DIRECT has get_user_pages > open. Well in O_DIRECT case there is only one page table, the CPU page table and it gets updated during fork() so there is an ordering there and the race window is small. More over programmer knows that can get in trouble if they do thing like fork() and don't synchronize their threads with each other. So while some weird thing can happen with O_DIRECT, it is unlikely (very small race window) and if it happens its well within the expected behavior. For hardware the race window is the same as the process lifetime so it can be days, months, years ... Once the hardware has programmed its page table they will never see any update (again mlx5 ODP is the exception here). This is where "issues" weird behavior can arise. Because
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Mon, Nov 19, 2018 at 02:17:21PM -0500, Jerome Glisse wrote: > On Mon, Nov 19, 2018 at 11:53:33AM -0700, Jason Gunthorpe wrote: > > On Mon, Nov 19, 2018 at 01:42:16PM -0500, Jerome Glisse wrote: > > > On Mon, Nov 19, 2018 at 11:27:52AM -0700, Jason Gunthorpe wrote: > > > > On Mon, Nov 19, 2018 at 11:48:54AM -0500, Jerome Glisse wrote: > > > > > > > > > Just to comment on this, any infiniband driver which use umem and do > > > > > not have ODP (here ODP for me means listening to mmu notifier so all > > > > > infiniband driver except mlx5) will be affected by same issue AFAICT. > > > > > > > > > > AFAICT there is no special thing happening after fork() inside any of > > > > > those driver. So if parent create a umem mr before fork() and program > > > > > hardware with it then after fork() the parent might start using new > > > > > page for the umem range while the old memory is use by the child. The > > > > > reverse is also true (parent using old memory and child new memory) > > > > > bottom line you can not predict which memory the child or the parent > > > > > will use for the range after fork(). > > > > > > > > > > So no matter what you consider the child or the parent, what the hw > > > > > will use for the mr is unlikely to match what the CPU use for the > > > > > same virtual address. In other word: > > > > > > > > > > Before fork: > > > > > CPU parent: virtual addr ptr1 -> physical address = 0xCAFE > > > > > HARDWARE: virtual addr ptr1 -> physical address = 0xCAFE > > > > > > > > > > Case 1: > > > > > CPU parent: virtual addr ptr1 -> physical address = 0xCAFE > > > > > CPU child: virtual addr ptr1 -> physical address = 0xDEAD > > > > > HARDWARE: virtual addr ptr1 -> physical address = 0xCAFE > > > > > > > > > > Case 2: > > > > > CPU parent: virtual addr ptr1 -> physical address = 0xBEEF > > > > > CPU child: virtual addr ptr1 -> physical address = 0xCAFE > > > > > HARDWARE: virtual addr ptr1 -> physical address = 0xCAFE > > > > > > > > IIRC this is solved in IB by automatically calling > > > > madvise(MADV_DONTFORK) before creating the MR. > > > > > > > > MADV_DONTFORK > > > > .. This is useful to prevent copy-on-write semantics from changing the > > > > physical location of a page if the parent writes to it after a > > > > fork(2) .. > > > > > > This would work around the issue but this is not transparent ie > > > range marked with DONTFORK no longer behave as expected from the > > > application point of view. > > > > Do you know what the difference is? The man page really gives no > > hint.. > > > > Does it sometimes unmap the pages during fork? > > It is handled in kernel/fork.c look for DONTCOPY, basicaly it just > leave empty page table in the child process so child will have to > fault in new page. This also means that child will get 0 as initial > value for all memory address under DONTCOPY/DONTFORK which breaks > application expectation of what fork() do. Hum, I wonder why this API was selected then.. > > I actually wonder if the kernel is a bit broken here, we have the same > > problem with O_DIRECT and other stuff, right? > > No it is not, O_DIRECT is fine. The only corner case i can think > of with O_DIRECT is one thread launching an O_DIRECT that write > to private anonymous memory (other O_DIRECT case do not matter) > while another thread call fork() then what the child get can be > undefined ie either it get the data before the O_DIRECT finish > or it gets the result of the O_DIRECT. But this is realy what > you should expect when doing such thing without synchronization. > > So O_DIRECT is fine. ?? How can O_DIRECT be fine but RDMA not? They use exactly the same get_user_pages flow, right? Can we do what O_DIRECT does in RDMA and be fine too? AFAIK the only difference is the length of the race window. You'd have to fork and fault during the shorter time O_DIRECT has get_user_pages open. > > Really, if I have a get_user_pages FOLL_WRITE on a page and we fork, > > then shouldn't the COW immediately be broken during the fork? > > > > The kernel can't guarentee that an ongoing DMA will not write to those > > pages, and it breaks the fork semantic to write to both processes. > > Fixing that would incur a high cost: need to grow struct page, need > to copy potentialy gigabyte of memory during fork() ... this would be > a serious performance regression for many folks just to work around an > abuse of device driver. So i don't think anything on that front would > be welcome. Why? Keep track in each mm if there are any active get_user_pages FOLL_WRITE pages in the mm, if yes then sweep the VMAs and fix the issue for the FOLL_WRITE pages. John is already working on being able to detect pages under GUP, so it seems like a small step.. Since nearly all cases of fork don't have a GUP FOLL_WRITE active there would be no performance hit. > umem without proper ODP and VFIO are the only bad user i know of (for > VFIO you can argue
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Mon, Nov 19, 2018 at 07:19:04PM +, Christopher Lameter wrote: > On Mon, 19 Nov 2018, Jerome Glisse wrote: > > > > IIRC this is solved in IB by automatically calling > > > madvise(MADV_DONTFORK) before creating the MR. > > > > > > MADV_DONTFORK > > > .. This is useful to prevent copy-on-write semantics from changing the > > > physical location of a page if the parent writes to it after a > > > fork(2) .. > > > > This would work around the issue but this is not transparent ie > > range marked with DONTFORK no longer behave as expected from the > > application point of view. > > Why would anyone expect a range registered via MR behave as normal? Device > I/O is going on into that range. Memory is already special. > > > > Also it relies on userspace doing the right thing (which is not > > something i usualy trust :)). > > This has been established practice for 15 or so years in a couple of use > cases. Again user space already has to be doing special things in order to > handle RDMA is that area. Yes RDMA as an existing historical track record and thus people should now be aware of its limitation. What i am fighting against is new addition to kernel that pretend to do SVA (share virtual address) while their hardware is not realy doing SVA. SVA with IOMMU and ATS/PASID is fine, SVA in software with device driver that abide to mmu notifier is fine. Anything else is not. So i am just worrying about new user and making sure they under- stand what is happening and not sell to their user something false. Cheers, Jérôme
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Mon, 19 Nov 2018, Jerome Glisse wrote: > > IIRC this is solved in IB by automatically calling > > madvise(MADV_DONTFORK) before creating the MR. > > > > MADV_DONTFORK > > .. This is useful to prevent copy-on-write semantics from changing the > > physical location of a page if the parent writes to it after a > > fork(2) .. > > This would work around the issue but this is not transparent ie > range marked with DONTFORK no longer behave as expected from the > application point of view. Why would anyone expect a range registered via MR behave as normal? Device I/O is going on into that range. Memory is already special. > Also it relies on userspace doing the right thing (which is not > something i usualy trust :)). This has been established practice for 15 or so years in a couple of use cases. Again user space already has to be doing special things in order to handle RDMA is that area.
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Mon, Nov 19, 2018 at 11:53:33AM -0700, Jason Gunthorpe wrote: > On Mon, Nov 19, 2018 at 01:42:16PM -0500, Jerome Glisse wrote: > > On Mon, Nov 19, 2018 at 11:27:52AM -0700, Jason Gunthorpe wrote: > > > On Mon, Nov 19, 2018 at 11:48:54AM -0500, Jerome Glisse wrote: > > > > > > > Just to comment on this, any infiniband driver which use umem and do > > > > not have ODP (here ODP for me means listening to mmu notifier so all > > > > infiniband driver except mlx5) will be affected by same issue AFAICT. > > > > > > > > AFAICT there is no special thing happening after fork() inside any of > > > > those driver. So if parent create a umem mr before fork() and program > > > > hardware with it then after fork() the parent might start using new > > > > page for the umem range while the old memory is use by the child. The > > > > reverse is also true (parent using old memory and child new memory) > > > > bottom line you can not predict which memory the child or the parent > > > > will use for the range after fork(). > > > > > > > > So no matter what you consider the child or the parent, what the hw > > > > will use for the mr is unlikely to match what the CPU use for the > > > > same virtual address. In other word: > > > > > > > > Before fork: > > > > CPU parent: virtual addr ptr1 -> physical address = 0xCAFE > > > > HARDWARE: virtual addr ptr1 -> physical address = 0xCAFE > > > > > > > > Case 1: > > > > CPU parent: virtual addr ptr1 -> physical address = 0xCAFE > > > > CPU child: virtual addr ptr1 -> physical address = 0xDEAD > > > > HARDWARE: virtual addr ptr1 -> physical address = 0xCAFE > > > > > > > > Case 2: > > > > CPU parent: virtual addr ptr1 -> physical address = 0xBEEF > > > > CPU child: virtual addr ptr1 -> physical address = 0xCAFE > > > > HARDWARE: virtual addr ptr1 -> physical address = 0xCAFE > > > > > > IIRC this is solved in IB by automatically calling > > > madvise(MADV_DONTFORK) before creating the MR. > > > > > > MADV_DONTFORK > > > .. This is useful to prevent copy-on-write semantics from changing the > > > physical location of a page if the parent writes to it after a > > > fork(2) .. > > > > This would work around the issue but this is not transparent ie > > range marked with DONTFORK no longer behave as expected from the > > application point of view. > > Do you know what the difference is? The man page really gives no > hint.. > > Does it sometimes unmap the pages during fork? It is handled in kernel/fork.c look for DONTCOPY, basicaly it just leave empty page table in the child process so child will have to fault in new page. This also means that child will get 0 as initial value for all memory address under DONTCOPY/DONTFORK which breaks application expectation of what fork() do. > > I actually wonder if the kernel is a bit broken here, we have the same > problem with O_DIRECT and other stuff, right? No it is not, O_DIRECT is fine. The only corner case i can think of with O_DIRECT is one thread launching an O_DIRECT that write to private anonymous memory (other O_DIRECT case do not matter) while another thread call fork() then what the child get can be undefined ie either it get the data before the O_DIRECT finish or it gets the result of the O_DIRECT. But this is realy what you should expect when doing such thing without synchronization. So O_DIRECT is fine. > > Really, if I have a get_user_pages FOLL_WRITE on a page and we fork, > then shouldn't the COW immediately be broken during the fork? > > The kernel can't guarentee that an ongoing DMA will not write to those > pages, and it breaks the fork semantic to write to both processes. Fixing that would incur a high cost: need to grow struct page, need to copy potentialy gigabyte of memory during fork() ... this would be a serious performance regression for many folks just to work around an abuse of device driver. So i don't think anything on that front would be welcome. umem without proper ODP and VFIO are the only bad user i know of (for VFIO you can argue that it is part of the API contract and thus that it is not an abuse but it is not spell out loud in documentation). I have been trying to push back on any people trying to push thing that would make the same mistake or at least making sure they understand what is happening. What really need to happen is people fixing their hardware and do the right thing (good software engineer versus evil hardware engineer ;)) Cheers, Jérôme
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Mon, Nov 19, 2018 at 01:42:16PM -0500, Jerome Glisse wrote: > On Mon, Nov 19, 2018 at 11:27:52AM -0700, Jason Gunthorpe wrote: > > On Mon, Nov 19, 2018 at 11:48:54AM -0500, Jerome Glisse wrote: > > > > > Just to comment on this, any infiniband driver which use umem and do > > > not have ODP (here ODP for me means listening to mmu notifier so all > > > infiniband driver except mlx5) will be affected by same issue AFAICT. > > > > > > AFAICT there is no special thing happening after fork() inside any of > > > those driver. So if parent create a umem mr before fork() and program > > > hardware with it then after fork() the parent might start using new > > > page for the umem range while the old memory is use by the child. The > > > reverse is also true (parent using old memory and child new memory) > > > bottom line you can not predict which memory the child or the parent > > > will use for the range after fork(). > > > > > > So no matter what you consider the child or the parent, what the hw > > > will use for the mr is unlikely to match what the CPU use for the > > > same virtual address. In other word: > > > > > > Before fork: > > > CPU parent: virtual addr ptr1 -> physical address = 0xCAFE > > > HARDWARE: virtual addr ptr1 -> physical address = 0xCAFE > > > > > > Case 1: > > > CPU parent: virtual addr ptr1 -> physical address = 0xCAFE > > > CPU child: virtual addr ptr1 -> physical address = 0xDEAD > > > HARDWARE: virtual addr ptr1 -> physical address = 0xCAFE > > > > > > Case 2: > > > CPU parent: virtual addr ptr1 -> physical address = 0xBEEF > > > CPU child: virtual addr ptr1 -> physical address = 0xCAFE > > > HARDWARE: virtual addr ptr1 -> physical address = 0xCAFE > > > > IIRC this is solved in IB by automatically calling > > madvise(MADV_DONTFORK) before creating the MR. > > > > MADV_DONTFORK > > .. This is useful to prevent copy-on-write semantics from changing the > > physical location of a page if the parent writes to it after a > > fork(2) .. > > This would work around the issue but this is not transparent ie > range marked with DONTFORK no longer behave as expected from the > application point of view. > > Also it relies on userspace doing the right thing (which is not > something i usualy trust :)). The good thing that we didn't see anyone who succeeded to run IB stack without our user space, which does right thing under the hood :). > > Cheers, > Jérôme signature.asc Description: PGP signature
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Mon, Nov 19, 2018 at 01:42:16PM -0500, Jerome Glisse wrote: > On Mon, Nov 19, 2018 at 11:27:52AM -0700, Jason Gunthorpe wrote: > > On Mon, Nov 19, 2018 at 11:48:54AM -0500, Jerome Glisse wrote: > > > > > Just to comment on this, any infiniband driver which use umem and do > > > not have ODP (here ODP for me means listening to mmu notifier so all > > > infiniband driver except mlx5) will be affected by same issue AFAICT. > > > > > > AFAICT there is no special thing happening after fork() inside any of > > > those driver. So if parent create a umem mr before fork() and program > > > hardware with it then after fork() the parent might start using new > > > page for the umem range while the old memory is use by the child. The > > > reverse is also true (parent using old memory and child new memory) > > > bottom line you can not predict which memory the child or the parent > > > will use for the range after fork(). > > > > > > So no matter what you consider the child or the parent, what the hw > > > will use for the mr is unlikely to match what the CPU use for the > > > same virtual address. In other word: > > > > > > Before fork: > > > CPU parent: virtual addr ptr1 -> physical address = 0xCAFE > > > HARDWARE: virtual addr ptr1 -> physical address = 0xCAFE > > > > > > Case 1: > > > CPU parent: virtual addr ptr1 -> physical address = 0xCAFE > > > CPU child: virtual addr ptr1 -> physical address = 0xDEAD > > > HARDWARE: virtual addr ptr1 -> physical address = 0xCAFE > > > > > > Case 2: > > > CPU parent: virtual addr ptr1 -> physical address = 0xBEEF > > > CPU child: virtual addr ptr1 -> physical address = 0xCAFE > > > HARDWARE: virtual addr ptr1 -> physical address = 0xCAFE > > > > IIRC this is solved in IB by automatically calling > > madvise(MADV_DONTFORK) before creating the MR. > > > > MADV_DONTFORK > > .. This is useful to prevent copy-on-write semantics from changing the > > physical location of a page if the parent writes to it after a > > fork(2) .. > > This would work around the issue but this is not transparent ie > range marked with DONTFORK no longer behave as expected from the > application point of view. Do you know what the difference is? The man page really gives no hint.. Does it sometimes unmap the pages during fork? I actually wonder if the kernel is a bit broken here, we have the same problem with O_DIRECT and other stuff, right? Really, if I have a get_user_pages FOLL_WRITE on a page and we fork, then shouldn't the COW immediately be broken during the fork? The kernel can't guarentee that an ongoing DMA will not write to those pages, and it breaks the fork semantic to write to both processes. > Also it relies on userspace doing the right thing (which is not > something i usualy trust :)). Well, if they do it wrong they get to keep all the broken bits :) Jason
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Mon, Nov 19, 2018 at 05:14:05PM +0800, Kenneth Lee wrote: > If the hardware cannot share page table with the CPU, we then need to have > some way to change the device page table. This is what happen in ODP. It > invalidates the page table in device upon mmu_notifier call back. But this > cannot > solve the COW problem: if the user process A share a page P with device, and > A > forks a new process B, and it continue to write to the page. By COW, the > process B will keep the page P, while A will get a new page P'. But you have > no way to let the device know it should use P' rather than P. Is this true? I thought mmu_notifiers covered all these cases. The mm_notifier for A should fire if B causes the physical address of A's pages to change via COW. And this causes the device page tables to re-synchronize. > In WarpDrive/uacce, we make this simple. If you support IOMMU and it support > SVM/SVA. Everything will be fine just like ODP implicit mode. And you don't > need > to write any code for that. Because it has been done by IOMMU framework. If it Looks like the IOMMU code uses mmu_notifier, so it is identical to IB's ODP. The only difference is that IB tends to have the IOMMU page table in the device, not in the CPU. The only case I know if that is different is the new-fangled CAPI stuff where the IOMMU can directly use the CPU's page table and the IOMMU page table (in device or CPU) is eliminated. Anyhow, I don't think a single instance of hardware should justify an entire new subsystem. Subsystems are hard to make and without multiple hardware examples there is no way to expect that it would cover any future use cases. If all your driver needs is to mmap some PCI bar space, route interrupts and do DMA mapping then mediated VFIO is probably a good choice. If it needs to do a bunch of other stuff, not related to PCI bar space, interrupts and DMA mapping (ie special code for compression, crypto, AI, whatever) then you should probably do what Jerome said and make a drivers/char/hisillicon_foo_bar.c that exposes just what your hardware does. If you have networking involved in here then consider RDMA, particularly if this functionality is already part of the same hardware that the hns infiniband driver is servicing. 'computational MRs' are a reasonable approach to a side-car offload of already existing RDMA support. Jason
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Mon, Nov 19, 2018 at 11:27:52AM -0700, Jason Gunthorpe wrote: > On Mon, Nov 19, 2018 at 11:48:54AM -0500, Jerome Glisse wrote: > > > Just to comment on this, any infiniband driver which use umem and do > > not have ODP (here ODP for me means listening to mmu notifier so all > > infiniband driver except mlx5) will be affected by same issue AFAICT. > > > > AFAICT there is no special thing happening after fork() inside any of > > those driver. So if parent create a umem mr before fork() and program > > hardware with it then after fork() the parent might start using new > > page for the umem range while the old memory is use by the child. The > > reverse is also true (parent using old memory and child new memory) > > bottom line you can not predict which memory the child or the parent > > will use for the range after fork(). > > > > So no matter what you consider the child or the parent, what the hw > > will use for the mr is unlikely to match what the CPU use for the > > same virtual address. In other word: > > > > Before fork: > > CPU parent: virtual addr ptr1 -> physical address = 0xCAFE > > HARDWARE: virtual addr ptr1 -> physical address = 0xCAFE > > > > Case 1: > > CPU parent: virtual addr ptr1 -> physical address = 0xCAFE > > CPU child: virtual addr ptr1 -> physical address = 0xDEAD > > HARDWARE: virtual addr ptr1 -> physical address = 0xCAFE > > > > Case 2: > > CPU parent: virtual addr ptr1 -> physical address = 0xBEEF > > CPU child: virtual addr ptr1 -> physical address = 0xCAFE > > HARDWARE: virtual addr ptr1 -> physical address = 0xCAFE > > IIRC this is solved in IB by automatically calling > madvise(MADV_DONTFORK) before creating the MR. > > MADV_DONTFORK > .. This is useful to prevent copy-on-write semantics from changing the > physical location of a page if the parent writes to it after a > fork(2) .. This would work around the issue but this is not transparent ie range marked with DONTFORK no longer behave as expected from the application point of view. Also it relies on userspace doing the right thing (which is not something i usualy trust :)). Cheers, Jérôme
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Mon, Nov 19, 2018 at 11:48:54AM -0500, Jerome Glisse wrote: > Just to comment on this, any infiniband driver which use umem and do > not have ODP (here ODP for me means listening to mmu notifier so all > infiniband driver except mlx5) will be affected by same issue AFAICT. > > AFAICT there is no special thing happening after fork() inside any of > those driver. So if parent create a umem mr before fork() and program > hardware with it then after fork() the parent might start using new > page for the umem range while the old memory is use by the child. The > reverse is also true (parent using old memory and child new memory) > bottom line you can not predict which memory the child or the parent > will use for the range after fork(). > > So no matter what you consider the child or the parent, what the hw > will use for the mr is unlikely to match what the CPU use for the > same virtual address. In other word: > > Before fork: > CPU parent: virtual addr ptr1 -> physical address = 0xCAFE > HARDWARE: virtual addr ptr1 -> physical address = 0xCAFE > > Case 1: > CPU parent: virtual addr ptr1 -> physical address = 0xCAFE > CPU child: virtual addr ptr1 -> physical address = 0xDEAD > HARDWARE: virtual addr ptr1 -> physical address = 0xCAFE > > Case 2: > CPU parent: virtual addr ptr1 -> physical address = 0xBEEF > CPU child: virtual addr ptr1 -> physical address = 0xCAFE > HARDWARE: virtual addr ptr1 -> physical address = 0xCAFE IIRC this is solved in IB by automatically calling madvise(MADV_DONTFORK) before creating the MR. MADV_DONTFORK .. This is useful to prevent copy-on-write semantics from changing the physical location of a page if the parent writes to it after a fork(2) .. Jason
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Mon, Nov 19, 2018 at 12:48:01PM +0200, Leon Romanovsky wrote: > On Mon, Nov 19, 2018 at 05:19:10PM +0800, Kenneth Lee wrote: > > On Mon, Nov 19, 2018 at 05:14:05PM +0800, Kenneth Lee wrote: > > > On Thu, Nov 15, 2018 at 04:54:55PM +0200, Leon Romanovsky wrote: > > > > On Thu, Nov 15, 2018 at 04:51:09PM +0800, Kenneth Lee wrote: > > > > > On Wed, Nov 14, 2018 at 06:00:17PM +0200, Leon Romanovsky wrote: > > > > > > On Wed, Nov 14, 2018 at 10:58:09AM +0800, Kenneth Lee wrote: > > > > > > > > On Mon, Nov 12, 2018 at 03:58:02PM +0800, Kenneth Lee wrote: [...] > > > > memory exposed to user is properly protected from security point of > > > > view. > > > > 3. "stop using the page for a while for the copying" - I'm not fully > > > > understand this claim, maybe this article will help you to better > > > > describe : https://lwn.net/Articles/753027/ > > > > > > This topic was being discussed in RFCv2. The key problem here is that: > > > > > > The device need to hold the memory for its own calculation, but the > > > CPU/software > > > want to stop it for a while for synchronizing with disk or COW. > > > > > > If the hardware support SVM/SVA (Shared Virtual Memory/Address), it is > > > easy, the > > > device share page table with CPU, the device will raise a page fault when > > > the > > > CPU downgrade the PTE to read-only. > > > > > > If the hardware cannot share page table with the CPU, we then need to have > > > some way to change the device page table. This is what happen in ODP. It > > > invalidates the page table in device upon mmu_notifier call back. But > > > this cannot > > > solve the COW problem: if the user process A share a page P with device, > > > and A > > > forks a new process B, and it continue to write to the page. By COW, the > > > process B will keep the page P, while A will get a new page P'. But you > > > have > > > no way to let the device know it should use P' rather than P. > > I didn't hear about such issue and we supported fork for a long time. > Just to comment on this, any infiniband driver which use umem and do not have ODP (here ODP for me means listening to mmu notifier so all infiniband driver except mlx5) will be affected by same issue AFAICT. AFAICT there is no special thing happening after fork() inside any of those driver. So if parent create a umem mr before fork() and program hardware with it then after fork() the parent might start using new page for the umem range while the old memory is use by the child. The reverse is also true (parent using old memory and child new memory) bottom line you can not predict which memory the child or the parent will use for the range after fork(). So no matter what you consider the child or the parent, what the hw will use for the mr is unlikely to match what the CPU use for the same virtual address. In other word: Before fork: CPU parent: virtual addr ptr1 -> physical address = 0xCAFE HARDWARE: virtual addr ptr1 -> physical address = 0xCAFE Case 1: CPU parent: virtual addr ptr1 -> physical address = 0xCAFE CPU child: virtual addr ptr1 -> physical address = 0xDEAD HARDWARE: virtual addr ptr1 -> physical address = 0xCAFE Case 2: CPU parent: virtual addr ptr1 -> physical address = 0xBEEF CPU child: virtual addr ptr1 -> physical address = 0xCAFE HARDWARE: virtual addr ptr1 -> physical address = 0xCAFE This apply for every single page and is not predictable. This only apply to private memory (mmap() with MAP_PRIVATE) I am not familiar enough with RDMA user space API contract to know if this is an issue or not. Note that this can not be fix, no one should have done umem without ODP like mlx5. For this to work properly you need sane hardware like mlx5. Cheers, Jérôme
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Mon, Nov 19, 2018 at 05:19:10PM +0800, Kenneth Lee wrote: > On Mon, Nov 19, 2018 at 05:14:05PM +0800, Kenneth Lee wrote: > > Date: Mon, 19 Nov 2018 17:14:05 +0800 > > From: Kenneth Lee > > To: Leon Romanovsky > > CC: Tim Sell , linux-...@vger.kernel.org, > > Alexander Shishkin , Zaibo Xu > > , zhangfei@foxmail.com, linux...@huawei.com, > > haojian.zhu...@linaro.org, Christoph Lameter , Hao Fang > > , Gavin Schenk , RDMA mailing > > list , Vinod Koul , Jason > > Gunthorpe , Doug Ledford , Uwe > > Kleine-König , David Kershner > > , Kenneth Lee , Johan > > Hovold , Cyrille Pitchen > > , Sagar Dharia > > , Jens Axboe , > > guodong...@linaro.org, linux-netdev , Randy Dunlap > > , linux-ker...@vger.kernel.org, Zhou Wang > > , linux-crypto@vger.kernel.org, Philippe > > Ombredanne , Sanyog Kale , > > "David S. Miller" , > > linux-accelerat...@lists.ozlabs.org > > Subject: Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce > > User-Agent: Mutt/1.5.21 (2010-09-15) > > Message-ID: <20181119091405.GE157308@Turing-Arch-b> > > > > On Thu, Nov 15, 2018 at 04:54:55PM +0200, Leon Romanovsky wrote: > > > Date: Thu, 15 Nov 2018 16:54:55 +0200 > > > From: Leon Romanovsky > > > To: Kenneth Lee > > > CC: Kenneth Lee , Tim Sell , > > > linux-...@vger.kernel.org, Alexander Shishkin > > > , Zaibo Xu , > > > zhangfei@foxmail.com, linux...@huawei.com, haojian.zhu...@linaro.org, > > > Christoph Lameter , Hao Fang , > > > Gavin > > > Schenk , RDMA mailing list > > > , Zhou Wang , Jason > > > Gunthorpe , Doug Ledford , Uwe > > > Kleine-König , David Kershner > > > , Johan Hovold , Cyrille > > > Pitchen , Sagar Dharia > > > , Jens Axboe , > > > guodong...@linaro.org, linux-netdev , Randy > > > Dunlap > > > , linux-ker...@vger.kernel.org, Vinod Koul > > > , linux-crypto@vger.kernel.org, Philippe Ombredanne > > > , Sanyog Kale , "David S. > > > Miller" , linux-accelerat...@lists.ozlabs.org > > > Subject: Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce > > > User-Agent: Mutt/1.10.1 (2018-07-13) > > > Message-ID: <20181115145455.gn3...@mtr-leonro.mtl.com> > > > > > > On Thu, Nov 15, 2018 at 04:51:09PM +0800, Kenneth Lee wrote: > > > > On Wed, Nov 14, 2018 at 06:00:17PM +0200, Leon Romanovsky wrote: > > > > > Date: Wed, 14 Nov 2018 18:00:17 +0200 > > > > > From: Leon Romanovsky > > > > > To: Kenneth Lee > > > > > CC: Tim Sell , linux-...@vger.kernel.org, > > > > > Alexander Shishkin , Zaibo Xu > > > > > , zhangfei@foxmail.com, linux...@huawei.com, > > > > > haojian.zhu...@linaro.org, Christoph Lameter , Hao > > > > > Fang > > > > > , Gavin Schenk , RDMA > > > > > mailing > > > > > list , Zhou Wang > > > > > , > > > > > Jason Gunthorpe , Doug Ledford , > > > > > Uwe > > > > > Kleine-König , David Kershner > > > > > , Johan Hovold , Cyrille > > > > > Pitchen , Sagar Dharia > > > > > , Jens Axboe , > > > > > guodong...@linaro.org, linux-netdev , Randy > > > > > Dunlap > > > > > , linux-ker...@vger.kernel.org, Vinod Koul > > > > > , linux-crypto@vger.kernel.org, Philippe Ombredanne > > > > > , Sanyog Kale , > > > > > Kenneth Lee > > > > > , "David S. Miller" , > > > > > linux-accelerat...@lists.ozlabs.org > > > > > Subject: Re: [RFCv3 PATCH 1/6] uacce: Add documents for > > > > > WarpDrive/uacce > > > > > User-Agent: Mutt/1.10.1 (2018-07-13) > > > > > Message-ID: <20181114160017.gi3...@mtr-leonro.mtl.com> > > > > > > > > > > On Wed, Nov 14, 2018 at 10:58:09AM +0800, Kenneth Lee wrote: > > > > > > > > > > > > 在 2018/11/13 上午8:23, Leon Romanovsky 写道: > > > > > > > On Mon, Nov 12, 2018 at 03:58:02PM +0800, Kenneth Lee wrote: > > > > > > > > From: Kenneth Lee > > > > > > > > > > > > > > > > WarpDrive is a general accelerator framework for the user > > > > > > > > application to > > > > > > > > access the hardware without going through the ker
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Mon, Nov 19, 2018 at 05:14:05PM +0800, Kenneth Lee wrote: > Date: Mon, 19 Nov 2018 17:14:05 +0800 > From: Kenneth Lee > To: Leon Romanovsky > CC: Tim Sell , linux-...@vger.kernel.org, > Alexander Shishkin , Zaibo Xu > , zhangfei@foxmail.com, linux...@huawei.com, > haojian.zhu...@linaro.org, Christoph Lameter , Hao Fang > , Gavin Schenk , RDMA mailing > list , Vinod Koul , Jason > Gunthorpe , Doug Ledford , Uwe > Kleine-König , David Kershner > , Kenneth Lee , Johan > Hovold , Cyrille Pitchen > , Sagar Dharia > , Jens Axboe , > guodong...@linaro.org, linux-netdev , Randy Dunlap > , linux-ker...@vger.kernel.org, Zhou Wang > , linux-crypto@vger.kernel.org, Philippe > Ombredanne , Sanyog Kale , > "David S. Miller" , > linux-accelerat...@lists.ozlabs.org > Subject: Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce > User-Agent: Mutt/1.5.21 (2010-09-15) > Message-ID: <20181119091405.GE157308@Turing-Arch-b> > > On Thu, Nov 15, 2018 at 04:54:55PM +0200, Leon Romanovsky wrote: > > Date: Thu, 15 Nov 2018 16:54:55 +0200 > > From: Leon Romanovsky > > To: Kenneth Lee > > CC: Kenneth Lee , Tim Sell , > > linux-...@vger.kernel.org, Alexander Shishkin > > , Zaibo Xu , > > zhangfei@foxmail.com, linux...@huawei.com, haojian.zhu...@linaro.org, > > Christoph Lameter , Hao Fang , Gavin > > Schenk , RDMA mailing list > > , Zhou Wang , Jason > > Gunthorpe , Doug Ledford , Uwe > > Kleine-König , David Kershner > > , Johan Hovold , Cyrille > > Pitchen , Sagar Dharia > > , Jens Axboe , > > guodong...@linaro.org, linux-netdev , Randy Dunlap > > , linux-ker...@vger.kernel.org, Vinod Koul > > , linux-crypto@vger.kernel.org, Philippe Ombredanne > > , Sanyog Kale , "David S. > > Miller" , linux-accelerat...@lists.ozlabs.org > > Subject: Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce > > User-Agent: Mutt/1.10.1 (2018-07-13) > > Message-ID: <20181115145455.gn3...@mtr-leonro.mtl.com> > > > > On Thu, Nov 15, 2018 at 04:51:09PM +0800, Kenneth Lee wrote: > > > On Wed, Nov 14, 2018 at 06:00:17PM +0200, Leon Romanovsky wrote: > > > > Date: Wed, 14 Nov 2018 18:00:17 +0200 > > > > From: Leon Romanovsky > > > > To: Kenneth Lee > > > > CC: Tim Sell , linux-...@vger.kernel.org, > > > > Alexander Shishkin , Zaibo Xu > > > > , zhangfei@foxmail.com, linux...@huawei.com, > > > > haojian.zhu...@linaro.org, Christoph Lameter , Hao Fang > > > > , Gavin Schenk , RDMA > > > > mailing > > > > list , Zhou Wang , > > > > Jason Gunthorpe , Doug Ledford , > > > > Uwe > > > > Kleine-König , David Kershner > > > > , Johan Hovold , Cyrille > > > > Pitchen , Sagar Dharia > > > > , Jens Axboe , > > > > guodong...@linaro.org, linux-netdev , Randy > > > > Dunlap > > > > , linux-ker...@vger.kernel.org, Vinod Koul > > > > , linux-crypto@vger.kernel.org, Philippe Ombredanne > > > > , Sanyog Kale , Kenneth > > > > Lee > > > > , "David S. Miller" , > > > > linux-accelerat...@lists.ozlabs.org > > > > Subject: Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce > > > > User-Agent: Mutt/1.10.1 (2018-07-13) > > > > Message-ID: <20181114160017.gi3...@mtr-leonro.mtl.com> > > > > > > > > On Wed, Nov 14, 2018 at 10:58:09AM +0800, Kenneth Lee wrote: > > > > > > > > > > 在 2018/11/13 上午8:23, Leon Romanovsky 写道: > > > > > > On Mon, Nov 12, 2018 at 03:58:02PM +0800, Kenneth Lee wrote: > > > > > > > From: Kenneth Lee > > > > > > > > > > > > > > WarpDrive is a general accelerator framework for the user > > > > > > > application to > > > > > > > access the hardware without going through the kernel in data path. > > > > > > > > > > > > > > The kernel component to provide kernel facility to driver for > > > > > > > expose the > > > > > > > user interface is called uacce. It a short name for > > > > > > > "Unified/User-space-access-intended Accelerator Framework". > > > > > > > > > > > > > > This patch add document to explain how it works. > > > > > > + RDMA and netdev folks > > > > > > > > > > >
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Thu, Nov 15, 2018 at 04:54:55PM +0200, Leon Romanovsky wrote: > Date: Thu, 15 Nov 2018 16:54:55 +0200 > From: Leon Romanovsky > To: Kenneth Lee > CC: Kenneth Lee , Tim Sell , > linux-...@vger.kernel.org, Alexander Shishkin > , Zaibo Xu , > zhangfei@foxmail.com, linux...@huawei.com, haojian.zhu...@linaro.org, > Christoph Lameter , Hao Fang , Gavin > Schenk , RDMA mailing list > , Zhou Wang , Jason > Gunthorpe , Doug Ledford , Uwe > Kleine-König , David Kershner > , Johan Hovold , Cyrille > Pitchen , Sagar Dharia > , Jens Axboe , > guodong...@linaro.org, linux-netdev , Randy Dunlap > , linux-ker...@vger.kernel.org, Vinod Koul > , linux-crypto@vger.kernel.org, Philippe Ombredanne > , Sanyog Kale , "David S. > Miller" , linux-accelerat...@lists.ozlabs.org > Subject: Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce > User-Agent: Mutt/1.10.1 (2018-07-13) > Message-ID: <20181115145455.gn3...@mtr-leonro.mtl.com> > > On Thu, Nov 15, 2018 at 04:51:09PM +0800, Kenneth Lee wrote: > > On Wed, Nov 14, 2018 at 06:00:17PM +0200, Leon Romanovsky wrote: > > > Date: Wed, 14 Nov 2018 18:00:17 +0200 > > > From: Leon Romanovsky > > > To: Kenneth Lee > > > CC: Tim Sell , linux-...@vger.kernel.org, > > > Alexander Shishkin , Zaibo Xu > > > , zhangfei@foxmail.com, linux...@huawei.com, > > > haojian.zhu...@linaro.org, Christoph Lameter , Hao Fang > > > , Gavin Schenk , RDMA > > > mailing > > > list , Zhou Wang , > > > Jason Gunthorpe , Doug Ledford , Uwe > > > Kleine-König , David Kershner > > > , Johan Hovold , Cyrille > > > Pitchen , Sagar Dharia > > > , Jens Axboe , > > > guodong...@linaro.org, linux-netdev , Randy > > > Dunlap > > > , linux-ker...@vger.kernel.org, Vinod Koul > > > , linux-crypto@vger.kernel.org, Philippe Ombredanne > > > , Sanyog Kale , Kenneth > > > Lee > > > , "David S. Miller" , > > > linux-accelerat...@lists.ozlabs.org > > > Subject: Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce > > > User-Agent: Mutt/1.10.1 (2018-07-13) > > > Message-ID: <20181114160017.gi3...@mtr-leonro.mtl.com> > > > > > > On Wed, Nov 14, 2018 at 10:58:09AM +0800, Kenneth Lee wrote: > > > > > > > > 在 2018/11/13 上午8:23, Leon Romanovsky 写道: > > > > > On Mon, Nov 12, 2018 at 03:58:02PM +0800, Kenneth Lee wrote: > > > > > > From: Kenneth Lee > > > > > > > > > > > > WarpDrive is a general accelerator framework for the user > > > > > > application to > > > > > > access the hardware without going through the kernel in data path. > > > > > > > > > > > > The kernel component to provide kernel facility to driver for > > > > > > expose the > > > > > > user interface is called uacce. It a short name for > > > > > > "Unified/User-space-access-intended Accelerator Framework". > > > > > > > > > > > > This patch add document to explain how it works. > > > > > + RDMA and netdev folks > > > > > > > > > > Sorry, to be late in the game, I don't see other patches, but from > > > > > the description below it seems like you are reinventing RDMA verbs > > > > > model. I have hard time to see the differences in the proposed > > > > > framework to already implemented in drivers/infiniband/* for the > > > > > kernel > > > > > space and for the https://github.com/linux-rdma/rdma-core/ for the > > > > > user > > > > > space parts. > > > > > > > > Thanks Leon, > > > > > > > > Yes, we tried to solve similar problem in RDMA. We also learned a lot > > > > from > > > > the exist code of RDMA. But we we have to make a new one because we > > > > cannot > > > > register accelerators such as AI operation, encryption or compression > > > > to the > > > > RDMA framework:) > > > > > > Assuming that you did everything right and still failed to use RDMA > > > framework, you was supposed to fix it and not to reinvent new exactly > > > same one. It is how we develop kernel, by reusing existing code. > > > > Yes, but we don't force other system such as NIC or GPU into RDMA, do we? > > You don't introduce new NIC or GPU, but proposing a
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Thu, Nov 15, 2018 at 04:51:09PM +0800, Kenneth Lee wrote: > On Wed, Nov 14, 2018 at 06:00:17PM +0200, Leon Romanovsky wrote: > > Date: Wed, 14 Nov 2018 18:00:17 +0200 > > From: Leon Romanovsky > > To: Kenneth Lee > > CC: Tim Sell , linux-...@vger.kernel.org, > > Alexander Shishkin , Zaibo Xu > > , zhangfei@foxmail.com, linux...@huawei.com, > > haojian.zhu...@linaro.org, Christoph Lameter , Hao Fang > > , Gavin Schenk , RDMA mailing > > list , Zhou Wang , > > Jason Gunthorpe , Doug Ledford , Uwe > > Kleine-König , David Kershner > > , Johan Hovold , Cyrille > > Pitchen , Sagar Dharia > > , Jens Axboe , > > guodong...@linaro.org, linux-netdev , Randy Dunlap > > , linux-ker...@vger.kernel.org, Vinod Koul > > , linux-crypto@vger.kernel.org, Philippe Ombredanne > > , Sanyog Kale , Kenneth Lee > > , "David S. Miller" , > > linux-accelerat...@lists.ozlabs.org > > Subject: Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce > > User-Agent: Mutt/1.10.1 (2018-07-13) > > Message-ID: <20181114160017.gi3...@mtr-leonro.mtl.com> > > > > On Wed, Nov 14, 2018 at 10:58:09AM +0800, Kenneth Lee wrote: > > > > > > 在 2018/11/13 上午8:23, Leon Romanovsky 写道: > > > > On Mon, Nov 12, 2018 at 03:58:02PM +0800, Kenneth Lee wrote: > > > > > From: Kenneth Lee > > > > > > > > > > WarpDrive is a general accelerator framework for the user application > > > > > to > > > > > access the hardware without going through the kernel in data path. > > > > > > > > > > The kernel component to provide kernel facility to driver for expose > > > > > the > > > > > user interface is called uacce. It a short name for > > > > > "Unified/User-space-access-intended Accelerator Framework". > > > > > > > > > > This patch add document to explain how it works. > > > > + RDMA and netdev folks > > > > > > > > Sorry, to be late in the game, I don't see other patches, but from > > > > the description below it seems like you are reinventing RDMA verbs > > > > model. I have hard time to see the differences in the proposed > > > > framework to already implemented in drivers/infiniband/* for the kernel > > > > space and for the https://github.com/linux-rdma/rdma-core/ for the user > > > > space parts. > > > > > > Thanks Leon, > > > > > > Yes, we tried to solve similar problem in RDMA. We also learned a lot from > > > the exist code of RDMA. But we we have to make a new one because we cannot > > > register accelerators such as AI operation, encryption or compression to > > > the > > > RDMA framework:) > > > > Assuming that you did everything right and still failed to use RDMA > > framework, you was supposed to fix it and not to reinvent new exactly > > same one. It is how we develop kernel, by reusing existing code. > > Yes, but we don't force other system such as NIC or GPU into RDMA, do we? You don't introduce new NIC or GPU, but proposing another interface to directly access HW memory and bypass kernel for the data path. This is whole idea of RDMA and this is why it is already present in the kernel. Various hardware devices are supported in our stack allow a ton of crazy stuff, including GPUs interconnections and NIC functionalities. > > I assume you would not agree to register a zip accelerator to infiniband? :) "infiniband" name in the "drivers/infiniband/" is legacy one and the current code supports IB, RoCE, iWARP and OmniPath as a transport layers. For a lone time, we wanted to rename that folder to be "drivers/rdma", but didn't find enough brave men/women to do it, due to backport mess for such move. The addition of zip accelerator to RDMA is possible and depends on how you will model such new functionality - new driver, or maybe new ULP. > > Further, I don't think it is wise to break an exist system (RDMA) to fulfill a > totally new scenario. The better choice is to let them run in parallel for > some > time and try to merge them accordingly. Awesome, so please run your code out-of-tree for now and once you are ready for submission let's try to merge it. > > > > > > > > > Another problem we tried to address is the way to pin the memory for dma > > > operation. The RDMA way to pin the memory cannot avoid the page lost due > > > to > > > copy-on-write operation during the memory is used by the device. T
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Wed, Nov 14, 2018 at 06:00:17PM +0200, Leon Romanovsky wrote: > Date: Wed, 14 Nov 2018 18:00:17 +0200 > From: Leon Romanovsky > To: Kenneth Lee > CC: Tim Sell , linux-...@vger.kernel.org, > Alexander Shishkin , Zaibo Xu > , zhangfei@foxmail.com, linux...@huawei.com, > haojian.zhu...@linaro.org, Christoph Lameter , Hao Fang > , Gavin Schenk , RDMA mailing > list , Zhou Wang , > Jason Gunthorpe , Doug Ledford , Uwe > Kleine-König , David Kershner > , Johan Hovold , Cyrille > Pitchen , Sagar Dharia > , Jens Axboe , > guodong...@linaro.org, linux-netdev , Randy Dunlap > , linux-ker...@vger.kernel.org, Vinod Koul > , linux-crypto@vger.kernel.org, Philippe Ombredanne > , Sanyog Kale , Kenneth Lee > , "David S. Miller" , > linux-accelerat...@lists.ozlabs.org > Subject: Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce > User-Agent: Mutt/1.10.1 (2018-07-13) > Message-ID: <20181114160017.gi3...@mtr-leonro.mtl.com> > > On Wed, Nov 14, 2018 at 10:58:09AM +0800, Kenneth Lee wrote: > > > > 在 2018/11/13 上午8:23, Leon Romanovsky 写道: > > > On Mon, Nov 12, 2018 at 03:58:02PM +0800, Kenneth Lee wrote: > > > > From: Kenneth Lee > > > > > > > > WarpDrive is a general accelerator framework for the user application to > > > > access the hardware without going through the kernel in data path. > > > > > > > > The kernel component to provide kernel facility to driver for expose the > > > > user interface is called uacce. It a short name for > > > > "Unified/User-space-access-intended Accelerator Framework". > > > > > > > > This patch add document to explain how it works. > > > + RDMA and netdev folks > > > > > > Sorry, to be late in the game, I don't see other patches, but from > > > the description below it seems like you are reinventing RDMA verbs > > > model. I have hard time to see the differences in the proposed > > > framework to already implemented in drivers/infiniband/* for the kernel > > > space and for the https://github.com/linux-rdma/rdma-core/ for the user > > > space parts. > > > > Thanks Leon, > > > > Yes, we tried to solve similar problem in RDMA. We also learned a lot from > > the exist code of RDMA. But we we have to make a new one because we cannot > > register accelerators such as AI operation, encryption or compression to the > > RDMA framework:) > > Assuming that you did everything right and still failed to use RDMA > framework, you was supposed to fix it and not to reinvent new exactly > same one. It is how we develop kernel, by reusing existing code. Yes, but we don't force other system such as NIC or GPU into RDMA, do we? I assume you would not agree to register a zip accelerator to infiniband? :) Further, I don't think it is wise to break an exist system (RDMA) to fulfill a totally new scenario. The better choice is to let them run in parallel for some time and try to merge them accordingly. > > > > > Another problem we tried to address is the way to pin the memory for dma > > operation. The RDMA way to pin the memory cannot avoid the page lost due to > > copy-on-write operation during the memory is used by the device. This may > > not be important to RDMA library. But it is important to accelerator. > > Such support exists in drivers/infiniband/ from late 2014 and > it is called ODP (on demand paging). I reviewed ODP and I think it is a solution bound to infiniband. It is part of MR semantics and required a infiniband specific hook (ucontext->invalidate_range()). And the hook requires the device to be able to stop using the page for a while for the copying. It is ok for infiniband (actually, only mlx5 uses it). I don't think most accelerators can support this mode. But WarpDrive works fully on top of IOMMU interface, it has no this limitation. > > > > > Hope this can help the understanding. > > Yes, it helped me a lot. > Now, I'm more than before convinced that this whole patchset shouldn't > exist in the first place. Then maybe you can tell me how I can register my accelerator to the user space? > > To be clear, NAK. > > Thanks > > > > > Cheers > > > > > > > > Hard NAK from RDMA side. > > > > > > Thanks > > > > > > > Signed-off-by: Kenneth Lee > > > > --- > > > > Documentation/warpdrive/warpdrive.rst | 260 +++ > > > > Documentation/warpdrive/wd-arch.svg | 764 > > > > Documentation/warpdrive
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Wed, Nov 14, 2018 at 10:58:09AM +0800, Kenneth Lee wrote: > > 在 2018/11/13 上午8:23, Leon Romanovsky 写道: > > On Mon, Nov 12, 2018 at 03:58:02PM +0800, Kenneth Lee wrote: > > > From: Kenneth Lee > > > > > > WarpDrive is a general accelerator framework for the user application to > > > access the hardware without going through the kernel in data path. > > > > > > The kernel component to provide kernel facility to driver for expose the > > > user interface is called uacce. It a short name for > > > "Unified/User-space-access-intended Accelerator Framework". > > > > > > This patch add document to explain how it works. > > + RDMA and netdev folks > > > > Sorry, to be late in the game, I don't see other patches, but from > > the description below it seems like you are reinventing RDMA verbs > > model. I have hard time to see the differences in the proposed > > framework to already implemented in drivers/infiniband/* for the kernel > > space and for the https://github.com/linux-rdma/rdma-core/ for the user > > space parts. > > Thanks Leon, > > Yes, we tried to solve similar problem in RDMA. We also learned a lot from > the exist code of RDMA. But we we have to make a new one because we cannot > register accelerators such as AI operation, encryption or compression to the > RDMA framework:) Assuming that you did everything right and still failed to use RDMA framework, you was supposed to fix it and not to reinvent new exactly same one. It is how we develop kernel, by reusing existing code. > > Another problem we tried to address is the way to pin the memory for dma > operation. The RDMA way to pin the memory cannot avoid the page lost due to > copy-on-write operation during the memory is used by the device. This may > not be important to RDMA library. But it is important to accelerator. Such support exists in drivers/infiniband/ from late 2014 and it is called ODP (on demand paging). > > Hope this can help the understanding. Yes, it helped me a lot. Now, I'm more than before convinced that this whole patchset shouldn't exist in the first place. To be clear, NAK. Thanks > > Cheers > > > > > Hard NAK from RDMA side. > > > > Thanks > > > > > Signed-off-by: Kenneth Lee > > > --- > > > Documentation/warpdrive/warpdrive.rst | 260 +++ > > > Documentation/warpdrive/wd-arch.svg | 764 > > > Documentation/warpdrive/wd.svg | 526 ++ > > > Documentation/warpdrive/wd_q_addr_space.svg | 359 + > > > 4 files changed, 1909 insertions(+) > > > create mode 100644 Documentation/warpdrive/warpdrive.rst > > > create mode 100644 Documentation/warpdrive/wd-arch.svg > > > create mode 100644 Documentation/warpdrive/wd.svg > > > create mode 100644 Documentation/warpdrive/wd_q_addr_space.svg > > > > > > diff --git a/Documentation/warpdrive/warpdrive.rst > > > b/Documentation/warpdrive/warpdrive.rst > > > new file mode 100644 > > > index ..ef84d3a2d462 > > > --- /dev/null > > > +++ b/Documentation/warpdrive/warpdrive.rst > > > @@ -0,0 +1,260 @@ > > > +Introduction of WarpDrive > > > += > > > + > > > +*WarpDrive* is a general accelerator framework for the user application > > > to > > > +access the hardware without going through the kernel in data path. > > > + > > > +It can be used as the quick channel for accelerators, network adaptors or > > > +other hardware for application in user space. > > > + > > > +This may make some implementation simpler. E.g. you can reuse most of > > > the > > > +*netdev* driver in kernel and just share some ring buffer to the user > > > space > > > +driver for *DPDK* [4] or *ODP* [5]. Or you can combine the RSA > > > accelerator with > > > +the *netdev* in the user space as a https reversed proxy, etc. > > > + > > > +*WarpDrive* takes the hardware accelerator as a heterogeneous processor > > > which > > > +can share particular load from the CPU: > > > + > > > +.. image:: wd.svg > > > +:alt: WarpDrive Concept > > > + > > > +The virtual concept, queue, is used to manage the requests sent to the > > > +accelerator. The application send requests to the queue by writing to > > > some > > > +particular address, while the hardware takes the requests directly from > > > the > > > +address and send feedback accordingly. > > > + > > > +The format of the queue may differ from hardware to hardware. But the > > > +application need not to make any system call for the communication. > > > + > > > +*WarpDrive* tries to create a shared virtual address space for all > > > involved > > > +accelerators. Within this space, the requests sent to queue can refer to > > > any > > > +virtual address, which will be valid to the application and all involved > > > +accelerators. > > > + > > > +The name *WarpDrive* is simply a cool and general name meaning the > > > framework > > > +makes the application faster. It includes general user library, kernel > > > +management module and drivers f
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
在 2018/11/13 上午8:23, Leon Romanovsky 写道: On Mon, Nov 12, 2018 at 03:58:02PM +0800, Kenneth Lee wrote: From: Kenneth Lee WarpDrive is a general accelerator framework for the user application to access the hardware without going through the kernel in data path. The kernel component to provide kernel facility to driver for expose the user interface is called uacce. It a short name for "Unified/User-space-access-intended Accelerator Framework". This patch add document to explain how it works. + RDMA and netdev folks Sorry, to be late in the game, I don't see other patches, but from the description below it seems like you are reinventing RDMA verbs model. I have hard time to see the differences in the proposed framework to already implemented in drivers/infiniband/* for the kernel space and for the https://github.com/linux-rdma/rdma-core/ for the user space parts. Thanks Leon, Yes, we tried to solve similar problem in RDMA. We also learned a lot from the exist code of RDMA. But we we have to make a new one because we cannot register accelerators such as AI operation, encryption or compression to the RDMA framework:) Another problem we tried to address is the way to pin the memory for dma operation. The RDMA way to pin the memory cannot avoid the page lost due to copy-on-write operation during the memory is used by the device. This may not be important to RDMA library. But it is important to accelerator. Hope this can help the understanding. Cheers Hard NAK from RDMA side. Thanks Signed-off-by: Kenneth Lee --- Documentation/warpdrive/warpdrive.rst | 260 +++ Documentation/warpdrive/wd-arch.svg | 764 Documentation/warpdrive/wd.svg | 526 ++ Documentation/warpdrive/wd_q_addr_space.svg | 359 + 4 files changed, 1909 insertions(+) create mode 100644 Documentation/warpdrive/warpdrive.rst create mode 100644 Documentation/warpdrive/wd-arch.svg create mode 100644 Documentation/warpdrive/wd.svg create mode 100644 Documentation/warpdrive/wd_q_addr_space.svg diff --git a/Documentation/warpdrive/warpdrive.rst b/Documentation/warpdrive/warpdrive.rst new file mode 100644 index ..ef84d3a2d462 --- /dev/null +++ b/Documentation/warpdrive/warpdrive.rst @@ -0,0 +1,260 @@ +Introduction of WarpDrive += + +*WarpDrive* is a general accelerator framework for the user application to +access the hardware without going through the kernel in data path. + +It can be used as the quick channel for accelerators, network adaptors or +other hardware for application in user space. + +This may make some implementation simpler. E.g. you can reuse most of the +*netdev* driver in kernel and just share some ring buffer to the user space +driver for *DPDK* [4] or *ODP* [5]. Or you can combine the RSA accelerator with +the *netdev* in the user space as a https reversed proxy, etc. + +*WarpDrive* takes the hardware accelerator as a heterogeneous processor which +can share particular load from the CPU: + +.. image:: wd.svg +:alt: WarpDrive Concept + +The virtual concept, queue, is used to manage the requests sent to the +accelerator. The application send requests to the queue by writing to some +particular address, while the hardware takes the requests directly from the +address and send feedback accordingly. + +The format of the queue may differ from hardware to hardware. But the +application need not to make any system call for the communication. + +*WarpDrive* tries to create a shared virtual address space for all involved +accelerators. Within this space, the requests sent to queue can refer to any +virtual address, which will be valid to the application and all involved +accelerators. + +The name *WarpDrive* is simply a cool and general name meaning the framework +makes the application faster. It includes general user library, kernel +management module and drivers for the hardware. In kernel, the management +module is called *uacce*, meaning "Unified/User-space-access-intended +Accelerator Framework". + + +How does it work + + +*WarpDrive* uses *mmap* and *IOMMU* to play the trick. + +*Uacce* creates a chrdev for the device registered to it. A "queue" will be +created when the chrdev is opened. The application access the queue by mmap +different address region of the queue file. + +The following figure demonstrated the queue file address space: + +.. image:: wd_q_addr_space.svg +:alt: WarpDrive Queue Address Space + +The first region of the space, device region, is used for the application to +write request or read answer to or from the hardware. + +Normally, there can be three types of device regions mmio and memory regions. +It is recommended to use common memory for request/answer descriptors and use +the mmio space for device notification, such as doorbell. But of course, this +is all up to the interface designer. + +There can be two types of device memory re
Re: [RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
On Mon, Nov 12, 2018 at 03:58:02PM +0800, Kenneth Lee wrote: > From: Kenneth Lee > > WarpDrive is a general accelerator framework for the user application to > access the hardware without going through the kernel in data path. > > The kernel component to provide kernel facility to driver for expose the > user interface is called uacce. It a short name for > "Unified/User-space-access-intended Accelerator Framework". > > This patch add document to explain how it works. + RDMA and netdev folks Sorry, to be late in the game, I don't see other patches, but from the description below it seems like you are reinventing RDMA verbs model. I have hard time to see the differences in the proposed framework to already implemented in drivers/infiniband/* for the kernel space and for the https://github.com/linux-rdma/rdma-core/ for the user space parts. Hard NAK from RDMA side. Thanks > > Signed-off-by: Kenneth Lee > --- > Documentation/warpdrive/warpdrive.rst | 260 +++ > Documentation/warpdrive/wd-arch.svg | 764 > Documentation/warpdrive/wd.svg | 526 ++ > Documentation/warpdrive/wd_q_addr_space.svg | 359 + > 4 files changed, 1909 insertions(+) > create mode 100644 Documentation/warpdrive/warpdrive.rst > create mode 100644 Documentation/warpdrive/wd-arch.svg > create mode 100644 Documentation/warpdrive/wd.svg > create mode 100644 Documentation/warpdrive/wd_q_addr_space.svg > > diff --git a/Documentation/warpdrive/warpdrive.rst > b/Documentation/warpdrive/warpdrive.rst > new file mode 100644 > index ..ef84d3a2d462 > --- /dev/null > +++ b/Documentation/warpdrive/warpdrive.rst > @@ -0,0 +1,260 @@ > +Introduction of WarpDrive > += > + > +*WarpDrive* is a general accelerator framework for the user application to > +access the hardware without going through the kernel in data path. > + > +It can be used as the quick channel for accelerators, network adaptors or > +other hardware for application in user space. > + > +This may make some implementation simpler. E.g. you can reuse most of the > +*netdev* driver in kernel and just share some ring buffer to the user space > +driver for *DPDK* [4] or *ODP* [5]. Or you can combine the RSA accelerator > with > +the *netdev* in the user space as a https reversed proxy, etc. > + > +*WarpDrive* takes the hardware accelerator as a heterogeneous processor which > +can share particular load from the CPU: > + > +.. image:: wd.svg > +:alt: WarpDrive Concept > + > +The virtual concept, queue, is used to manage the requests sent to the > +accelerator. The application send requests to the queue by writing to some > +particular address, while the hardware takes the requests directly from the > +address and send feedback accordingly. > + > +The format of the queue may differ from hardware to hardware. But the > +application need not to make any system call for the communication. > + > +*WarpDrive* tries to create a shared virtual address space for all involved > +accelerators. Within this space, the requests sent to queue can refer to any > +virtual address, which will be valid to the application and all involved > +accelerators. > + > +The name *WarpDrive* is simply a cool and general name meaning the framework > +makes the application faster. It includes general user library, kernel > +management module and drivers for the hardware. In kernel, the management > +module is called *uacce*, meaning "Unified/User-space-access-intended > +Accelerator Framework". > + > + > +How does it work > + > + > +*WarpDrive* uses *mmap* and *IOMMU* to play the trick. > + > +*Uacce* creates a chrdev for the device registered to it. A "queue" will be > +created when the chrdev is opened. The application access the queue by mmap > +different address region of the queue file. > + > +The following figure demonstrated the queue file address space: > + > +.. image:: wd_q_addr_space.svg > +:alt: WarpDrive Queue Address Space > + > +The first region of the space, device region, is used for the application to > +write request or read answer to or from the hardware. > + > +Normally, there can be three types of device regions mmio and memory regions. > +It is recommended to use common memory for request/answer descriptors and use > +the mmio space for device notification, such as doorbell. But of course, this > +is all up to the interface designer. > + > +There can be two types of device memory regions, kernel-only and user-shared. > +This will be explained in the "kernel APIs" section. > + > +The Static Share Virtual Memory region is necessary only when the device > IOMMU > +does not support "Share Virtual Memory". This will be explained after the > +*IOMMU* idea. > + > + > +Architecture > + > + > +The full *WarpDrive* architecture is represented in the following class > +diagram: > + > +.. image:: wd-arch.svg > +:alt: WarpDrive Architecture > + >
[RFCv3 PATCH 1/6] uacce: Add documents for WarpDrive/uacce
From: Kenneth Lee WarpDrive is a general accelerator framework for the user application to access the hardware without going through the kernel in data path. The kernel component to provide kernel facility to driver for expose the user interface is called uacce. It a short name for "Unified/User-space-access-intended Accelerator Framework". This patch add document to explain how it works. Signed-off-by: Kenneth Lee --- Documentation/warpdrive/warpdrive.rst | 260 +++ Documentation/warpdrive/wd-arch.svg | 764 Documentation/warpdrive/wd.svg | 526 ++ Documentation/warpdrive/wd_q_addr_space.svg | 359 + 4 files changed, 1909 insertions(+) create mode 100644 Documentation/warpdrive/warpdrive.rst create mode 100644 Documentation/warpdrive/wd-arch.svg create mode 100644 Documentation/warpdrive/wd.svg create mode 100644 Documentation/warpdrive/wd_q_addr_space.svg diff --git a/Documentation/warpdrive/warpdrive.rst b/Documentation/warpdrive/warpdrive.rst new file mode 100644 index ..ef84d3a2d462 --- /dev/null +++ b/Documentation/warpdrive/warpdrive.rst @@ -0,0 +1,260 @@ +Introduction of WarpDrive += + +*WarpDrive* is a general accelerator framework for the user application to +access the hardware without going through the kernel in data path. + +It can be used as the quick channel for accelerators, network adaptors or +other hardware for application in user space. + +This may make some implementation simpler. E.g. you can reuse most of the +*netdev* driver in kernel and just share some ring buffer to the user space +driver for *DPDK* [4] or *ODP* [5]. Or you can combine the RSA accelerator with +the *netdev* in the user space as a https reversed proxy, etc. + +*WarpDrive* takes the hardware accelerator as a heterogeneous processor which +can share particular load from the CPU: + +.. image:: wd.svg +:alt: WarpDrive Concept + +The virtual concept, queue, is used to manage the requests sent to the +accelerator. The application send requests to the queue by writing to some +particular address, while the hardware takes the requests directly from the +address and send feedback accordingly. + +The format of the queue may differ from hardware to hardware. But the +application need not to make any system call for the communication. + +*WarpDrive* tries to create a shared virtual address space for all involved +accelerators. Within this space, the requests sent to queue can refer to any +virtual address, which will be valid to the application and all involved +accelerators. + +The name *WarpDrive* is simply a cool and general name meaning the framework +makes the application faster. It includes general user library, kernel +management module and drivers for the hardware. In kernel, the management +module is called *uacce*, meaning "Unified/User-space-access-intended +Accelerator Framework". + + +How does it work + + +*WarpDrive* uses *mmap* and *IOMMU* to play the trick. + +*Uacce* creates a chrdev for the device registered to it. A "queue" will be +created when the chrdev is opened. The application access the queue by mmap +different address region of the queue file. + +The following figure demonstrated the queue file address space: + +.. image:: wd_q_addr_space.svg +:alt: WarpDrive Queue Address Space + +The first region of the space, device region, is used for the application to +write request or read answer to or from the hardware. + +Normally, there can be three types of device regions mmio and memory regions. +It is recommended to use common memory for request/answer descriptors and use +the mmio space for device notification, such as doorbell. But of course, this +is all up to the interface designer. + +There can be two types of device memory regions, kernel-only and user-shared. +This will be explained in the "kernel APIs" section. + +The Static Share Virtual Memory region is necessary only when the device IOMMU +does not support "Share Virtual Memory". This will be explained after the +*IOMMU* idea. + + +Architecture + + +The full *WarpDrive* architecture is represented in the following class +diagram: + +.. image:: wd-arch.svg +:alt: WarpDrive Architecture + + +The user API + + +We adopt a polling style interface in the user space: :: + +int wd_request_queue(struct wd_queue *q); +void wd_release_queue(struct wd_queue *q); + +int wd_send(struct wd_queue *q, void *req); +int wd_recv(struct wd_queue *q, void **req); +int wd_recv_sync(struct wd_queue *q, void **req); +void wd_flush(struct wd_queue *q); + +wd_recv_sync() is a wrapper to its non-sync version. It will trapped into +kernel and waits until the queue become available. + +If the queue do not support SVA/SVM. The following helper function +can be used to create Static Virtual Share Memory: :: + +void *wd_preserve_share_m