Re: [PATCH v5 0/8] iio: new DMABUF based API, v5
Hi Andrew, Le jeudi 11 janvier 2024 à 11:30 -0600, Andrew Davis a écrit : > On 1/11/24 3:20 AM, Paul Cercueil wrote: > > Hi Andrew, > > > > Le lundi 08 janvier 2024 à 15:12 -0600, Andrew Davis a écrit : > > > On 12/19/23 11:50 AM, Paul Cercueil wrote: > > > > [V4 was: "iio: Add buffer write() support"][1] > > > > > > > > Hi Jonathan, > > > > > > > > This is a respin of the V3 of my patchset that introduced a new > > > > interface based on DMABUF objects [2]. > > > > > > > > The V4 was a split of the patchset, to attempt to upstream > > > > buffer > > > > write() support first. But since there is no current user > > > > upstream, > > > > it > > > > was not merged. This V5 is about doing the opposite, and > > > > contains > > > > the > > > > new DMABUF interface, without adding the buffer write() > > > > support. It > > > > can > > > > already be used with the upstream adi-axi-adc driver. > > > > > > > > In user-space, Libiio uses it to transfer back and forth blocks > > > > of > > > > samples between the hardware and the applications, without > > > > having > > > > to > > > > copy the data. > > > > > > > > On a ZCU102 with a FMComms3 daughter board, running Libiio from > > > > the > > > > pcercuei/dev-new-dmabuf-api branch [3], compiled with > > > > WITH_LOCAL_DMABUF_API=OFF (so that it uses fileio): > > > > sudo utils/iio_rwdev -b 4096 -B cf-ad9361-lpc > > > > Throughput: 116 MiB/s > > > > > > > > Same hardware, with the DMABUF API (WITH_LOCAL_DMABUF_API=ON): > > > > sudo utils/iio_rwdev -b 4096 -B cf-ad9361-lpc > > > > Throughput: 475 MiB/s > > > > > > > > This benchmark only measures the speed at which the data can be > > > > fetched > > > > to iio_rwdev's internal buffers, and does not actually try to > > > > read > > > > the > > > > data (e.g. to pipe it to stdout). It shows that fetching the > > > > data > > > > is > > > > more than 4x faster using the new interface. > > > > > > > > When actually reading the data, the performance difference > > > > isn't > > > > that > > > > impressive (maybe because in case of DMABUF the data is not in > > > > cache): > > > > > > > > WITH_LOCAL_DMABUF_API=OFF (so that it uses fileio): > > > > sudo utils/iio_rwdev -b 4096 cf-ad9361-lpc | dd > > > > of=/dev/zero > > > > status=progress > > > > 2446422528 bytes (2.4 GB, 2.3 GiB) copied, 22 s, 111 MB/s > > > > > > > > WITH_LOCAL_DMABUF_API=ON: > > > > sudo utils/iio_rwdev -b 4096 cf-ad9361-lpc | dd > > > > of=/dev/zero > > > > status=progress > > > > 2334388736 bytes (2.3 GB, 2.2 GiB) copied, 21 s, 114 MB/s > > > > > > > > One interesting thing to note is that fileio is (currently) > > > > actually > > > > faster than the DMABUF interface if you increase a lot the > > > > buffer > > > > size. > > > > My explanation is that the cache invalidation routine takes > > > > more > > > > and > > > > more time the bigger the DMABUF gets. This is because the > > > > DMABUF is > > > > backed by small-size pages, so a (e.g.) 64 MiB DMABUF is backed > > > > by > > > > up > > > > to 16 thousands pages, that have to be invalidated one by one. > > > > This > > > > can > > > > be addressed by using huge pages, but the udmabuf driver does > > > > not > > > > (yet) > > > > support creating DMABUFs backed by huge pages. > > > > > > > > > > Have you tried DMABUFs created using the DMABUF System heap > > > exporter? > > > (drivers/dma-buf/heaps/system_heap.c) It should be able to handle > > > larger allocation better here, and if you don't have any active > > > mmaps or vmaps then it can skip CPU-side coherency maintenance > > > (useful for device to device transfers). > > > > I didn't know about it! > > > > But udmabuf also allows you to skip CPU-side coherency maintenance, > > since DMABUFs have two ioctls to start/finish CPU access anyway. > > > > The only way it lets you skip that is if your application just > doesn't > call those begin/end ioctls, which is wrong. That may work on a > system > where CPU caches can be snooped by all devices that could attach to > a buffer(x86), but that might not work on others(ARM). So calling > those begin/end ioctls is required[0]. If maintenance is not actually > needed then the kernel will turn those calls into NOPs for you, but > only > the kernel can know when that is correct (based on the running system > and the devices attached to that buffer), not userspace. My application only calls these begin/end IOCTLs when the DMABUF's data is accessed (through its mmapped address), and not ie. when I just pass around the DMABUF to another device driver. In that case I don't care that the CPU caches aren't sync'd. > > > > Allocating DMABUFs out of user pages has a bunch of other issues > > > you > > > might run into also. I'd argue udmabuf is now completely > > > superseded > > > by DMABUF system heaps. Try it out :) > > > > I'm curious, what other issues? > > > > For starters the {begin,end}_cpu_access() callbacks don't actually > sync the pages for
Re: [PATCH v5 0/8] iio: new DMABUF based API, v5
On 1/11/24 3:20 AM, Paul Cercueil wrote: Hi Andrew, Le lundi 08 janvier 2024 à 15:12 -0600, Andrew Davis a écrit : On 12/19/23 11:50 AM, Paul Cercueil wrote: [V4 was: "iio: Add buffer write() support"][1] Hi Jonathan, This is a respin of the V3 of my patchset that introduced a new interface based on DMABUF objects [2]. The V4 was a split of the patchset, to attempt to upstream buffer write() support first. But since there is no current user upstream, it was not merged. This V5 is about doing the opposite, and contains the new DMABUF interface, without adding the buffer write() support. It can already be used with the upstream adi-axi-adc driver. In user-space, Libiio uses it to transfer back and forth blocks of samples between the hardware and the applications, without having to copy the data. On a ZCU102 with a FMComms3 daughter board, running Libiio from the pcercuei/dev-new-dmabuf-api branch [3], compiled with WITH_LOCAL_DMABUF_API=OFF (so that it uses fileio): sudo utils/iio_rwdev -b 4096 -B cf-ad9361-lpc Throughput: 116 MiB/s Same hardware, with the DMABUF API (WITH_LOCAL_DMABUF_API=ON): sudo utils/iio_rwdev -b 4096 -B cf-ad9361-lpc Throughput: 475 MiB/s This benchmark only measures the speed at which the data can be fetched to iio_rwdev's internal buffers, and does not actually try to read the data (e.g. to pipe it to stdout). It shows that fetching the data is more than 4x faster using the new interface. When actually reading the data, the performance difference isn't that impressive (maybe because in case of DMABUF the data is not in cache): WITH_LOCAL_DMABUF_API=OFF (so that it uses fileio): sudo utils/iio_rwdev -b 4096 cf-ad9361-lpc | dd of=/dev/zero status=progress 2446422528 bytes (2.4 GB, 2.3 GiB) copied, 22 s, 111 MB/s WITH_LOCAL_DMABUF_API=ON: sudo utils/iio_rwdev -b 4096 cf-ad9361-lpc | dd of=/dev/zero status=progress 2334388736 bytes (2.3 GB, 2.2 GiB) copied, 21 s, 114 MB/s One interesting thing to note is that fileio is (currently) actually faster than the DMABUF interface if you increase a lot the buffer size. My explanation is that the cache invalidation routine takes more and more time the bigger the DMABUF gets. This is because the DMABUF is backed by small-size pages, so a (e.g.) 64 MiB DMABUF is backed by up to 16 thousands pages, that have to be invalidated one by one. This can be addressed by using huge pages, but the udmabuf driver does not (yet) support creating DMABUFs backed by huge pages. Have you tried DMABUFs created using the DMABUF System heap exporter? (drivers/dma-buf/heaps/system_heap.c) It should be able to handle larger allocation better here, and if you don't have any active mmaps or vmaps then it can skip CPU-side coherency maintenance (useful for device to device transfers). I didn't know about it! But udmabuf also allows you to skip CPU-side coherency maintenance, since DMABUFs have two ioctls to start/finish CPU access anyway. The only way it lets you skip that is if your application just doesn't call those begin/end ioctls, which is wrong. That may work on a system where CPU caches can be snooped by all devices that could attach to a buffer(x86), but that might not work on others(ARM). So calling those begin/end ioctls is required[0]. If maintenance is not actually needed then the kernel will turn those calls into NOPs for you, but only the kernel can know when that is correct (based on the running system and the devices attached to that buffer), not userspace. Allocating DMABUFs out of user pages has a bunch of other issues you might run into also. I'd argue udmabuf is now completely superseded by DMABUF system heaps. Try it out :) I'm curious, what other issues? For starters the {begin,end}_cpu_access() callbacks don't actually sync the pages for any of the devices attached to the DMABUF, it only makes a fake mapping for the misc device(CPU) then syncs with that. That probably works for the QEMU case it was designed for where the device is always a VM instance running on the same CPU, but for any real devices the sync never happens towards them. I have some patches fixing the above I'll post this cycle, but it wont help with folks doing reads/wrties on the original shmem/memfd outside of the begin/end ioctls. So there is a fundamental issue with the buffer's backing memory's ownership/lifecycle that makes udmabuf broken by design. The DMABUF System Heap owns the backing memory and manages that memory's lifecycle as all correct DMABUF exporters must. The good thing about udmabuf is that the memory is backed by pages, so we can use MSG_ZEROCOPY on sockets to transfer the mmapped data over the network (having a DMABUF interface to the network stack would be better, but I'm not opening that can of worms). Yes, having a DMABUF importer interface for the network stack would be the best long-term solution here, and one will probably end up being needed for zero-copy buffer passing directly
Re: [PATCH v5 0/8] iio: new DMABUF based API, v5
Hi Andrew, Le lundi 08 janvier 2024 à 15:12 -0600, Andrew Davis a écrit : > On 12/19/23 11:50 AM, Paul Cercueil wrote: > > [V4 was: "iio: Add buffer write() support"][1] > > > > Hi Jonathan, > > > > This is a respin of the V3 of my patchset that introduced a new > > interface based on DMABUF objects [2]. > > > > The V4 was a split of the patchset, to attempt to upstream buffer > > write() support first. But since there is no current user upstream, > > it > > was not merged. This V5 is about doing the opposite, and contains > > the > > new DMABUF interface, without adding the buffer write() support. It > > can > > already be used with the upstream adi-axi-adc driver. > > > > In user-space, Libiio uses it to transfer back and forth blocks of > > samples between the hardware and the applications, without having > > to > > copy the data. > > > > On a ZCU102 with a FMComms3 daughter board, running Libiio from the > > pcercuei/dev-new-dmabuf-api branch [3], compiled with > > WITH_LOCAL_DMABUF_API=OFF (so that it uses fileio): > > sudo utils/iio_rwdev -b 4096 -B cf-ad9361-lpc > > Throughput: 116 MiB/s > > > > Same hardware, with the DMABUF API (WITH_LOCAL_DMABUF_API=ON): > > sudo utils/iio_rwdev -b 4096 -B cf-ad9361-lpc > > Throughput: 475 MiB/s > > > > This benchmark only measures the speed at which the data can be > > fetched > > to iio_rwdev's internal buffers, and does not actually try to read > > the > > data (e.g. to pipe it to stdout). It shows that fetching the data > > is > > more than 4x faster using the new interface. > > > > When actually reading the data, the performance difference isn't > > that > > impressive (maybe because in case of DMABUF the data is not in > > cache): > > > > WITH_LOCAL_DMABUF_API=OFF (so that it uses fileio): > > sudo utils/iio_rwdev -b 4096 cf-ad9361-lpc | dd of=/dev/zero > > status=progress > > 2446422528 bytes (2.4 GB, 2.3 GiB) copied, 22 s, 111 MB/s > > > > WITH_LOCAL_DMABUF_API=ON: > > sudo utils/iio_rwdev -b 4096 cf-ad9361-lpc | dd of=/dev/zero > > status=progress > > 2334388736 bytes (2.3 GB, 2.2 GiB) copied, 21 s, 114 MB/s > > > > One interesting thing to note is that fileio is (currently) > > actually > > faster than the DMABUF interface if you increase a lot the buffer > > size. > > My explanation is that the cache invalidation routine takes more > > and > > more time the bigger the DMABUF gets. This is because the DMABUF is > > backed by small-size pages, so a (e.g.) 64 MiB DMABUF is backed by > > up > > to 16 thousands pages, that have to be invalidated one by one. This > > can > > be addressed by using huge pages, but the udmabuf driver does not > > (yet) > > support creating DMABUFs backed by huge pages. > > > > Have you tried DMABUFs created using the DMABUF System heap exporter? > (drivers/dma-buf/heaps/system_heap.c) It should be able to handle > larger allocation better here, and if you don't have any active > mmaps or vmaps then it can skip CPU-side coherency maintenance > (useful for device to device transfers). I didn't know about it! But udmabuf also allows you to skip CPU-side coherency maintenance, since DMABUFs have two ioctls to start/finish CPU access anyway. > Allocating DMABUFs out of user pages has a bunch of other issues you > might run into also. I'd argue udmabuf is now completely superseded > by DMABUF system heaps. Try it out :) I'm curious, what other issues? The good thing about udmabuf is that the memory is backed by pages, so we can use MSG_ZEROCOPY on sockets to transfer the mmapped data over the network (having a DMABUF interface to the network stack would be better, but I'm not opening that can of worms). > Andrew Cheers, -Paul > > Anyway, the real benefits happen when the DMABUFs are either shared > > between IIO devices, or between the IIO subsystem and another > > filesystem. In that case, the DMABUFs are simply passed around > > drivers, > > without the data being copied at any moment. > > > > We use that feature to transfer samples from our transceivers to > > USB, > > using a DMABUF interface to FunctionFS [4]. > > > > This drastically increases the throughput, to about 274 MiB/s over > > a > > USB3 link, vs. 127 MiB/s using IIO's fileio interface + write() to > > the > > FunctionFS endpoints, for a lower CPU usage (0.85 vs. 0.65 load > > avg.). > > > > Based on linux-next/next-20231219. > > > > Cheers, > > -Paul > > > > [1] > > https://lore.kernel.org/all/20230807112113.47157-1-p...@crapouillou.net/ > > [2] > > https://lore.kernel.org/all/20230403154800.215924-1-p...@crapouillou.net/ > > [3] > > https://github.com/analogdevicesinc/libiio/tree/pcercuei/dev-new-dmabuf-api > > [4] > > https://lore.kernel.org/all/20230322092118.9213-1-p...@crapouillou.net/ > > > > --- > > Changelog: > > - [3/8]: Replace V3's dmaengine_prep_slave_dma_array() with a new > > dmaengine_prep_slave_dma_vec(), which uses a new 'dma_vec' > > struct. > > Note that at some point we will need to
Re: [PATCH v5 0/8] iio: new DMABUF based API, v5
On 12/19/23 11:50 AM, Paul Cercueil wrote: [V4 was: "iio: Add buffer write() support"][1] Hi Jonathan, This is a respin of the V3 of my patchset that introduced a new interface based on DMABUF objects [2]. The V4 was a split of the patchset, to attempt to upstream buffer write() support first. But since there is no current user upstream, it was not merged. This V5 is about doing the opposite, and contains the new DMABUF interface, without adding the buffer write() support. It can already be used with the upstream adi-axi-adc driver. In user-space, Libiio uses it to transfer back and forth blocks of samples between the hardware and the applications, without having to copy the data. On a ZCU102 with a FMComms3 daughter board, running Libiio from the pcercuei/dev-new-dmabuf-api branch [3], compiled with WITH_LOCAL_DMABUF_API=OFF (so that it uses fileio): sudo utils/iio_rwdev -b 4096 -B cf-ad9361-lpc Throughput: 116 MiB/s Same hardware, with the DMABUF API (WITH_LOCAL_DMABUF_API=ON): sudo utils/iio_rwdev -b 4096 -B cf-ad9361-lpc Throughput: 475 MiB/s This benchmark only measures the speed at which the data can be fetched to iio_rwdev's internal buffers, and does not actually try to read the data (e.g. to pipe it to stdout). It shows that fetching the data is more than 4x faster using the new interface. When actually reading the data, the performance difference isn't that impressive (maybe because in case of DMABUF the data is not in cache): WITH_LOCAL_DMABUF_API=OFF (so that it uses fileio): sudo utils/iio_rwdev -b 4096 cf-ad9361-lpc | dd of=/dev/zero status=progress 2446422528 bytes (2.4 GB, 2.3 GiB) copied, 22 s, 111 MB/s WITH_LOCAL_DMABUF_API=ON: sudo utils/iio_rwdev -b 4096 cf-ad9361-lpc | dd of=/dev/zero status=progress 2334388736 bytes (2.3 GB, 2.2 GiB) copied, 21 s, 114 MB/s One interesting thing to note is that fileio is (currently) actually faster than the DMABUF interface if you increase a lot the buffer size. My explanation is that the cache invalidation routine takes more and more time the bigger the DMABUF gets. This is because the DMABUF is backed by small-size pages, so a (e.g.) 64 MiB DMABUF is backed by up to 16 thousands pages, that have to be invalidated one by one. This can be addressed by using huge pages, but the udmabuf driver does not (yet) support creating DMABUFs backed by huge pages. Have you tried DMABUFs created using the DMABUF System heap exporter? (drivers/dma-buf/heaps/system_heap.c) It should be able to handle larger allocation better here, and if you don't have any active mmaps or vmaps then it can skip CPU-side coherency maintenance (useful for device to device transfers). Allocating DMABUFs out of user pages has a bunch of other issues you might run into also. I'd argue udmabuf is now completely superseded by DMABUF system heaps. Try it out :) Andrew Anyway, the real benefits happen when the DMABUFs are either shared between IIO devices, or between the IIO subsystem and another filesystem. In that case, the DMABUFs are simply passed around drivers, without the data being copied at any moment. We use that feature to transfer samples from our transceivers to USB, using a DMABUF interface to FunctionFS [4]. This drastically increases the throughput, to about 274 MiB/s over a USB3 link, vs. 127 MiB/s using IIO's fileio interface + write() to the FunctionFS endpoints, for a lower CPU usage (0.85 vs. 0.65 load avg.). Based on linux-next/next-20231219. Cheers, -Paul [1] https://lore.kernel.org/all/20230807112113.47157-1-p...@crapouillou.net/ [2] https://lore.kernel.org/all/20230403154800.215924-1-p...@crapouillou.net/ [3] https://github.com/analogdevicesinc/libiio/tree/pcercuei/dev-new-dmabuf-api [4] https://lore.kernel.org/all/20230322092118.9213-1-p...@crapouillou.net/ --- Changelog: - [3/8]: Replace V3's dmaengine_prep_slave_dma_array() with a new dmaengine_prep_slave_dma_vec(), which uses a new 'dma_vec' struct. Note that at some point we will need to support cyclic transfers using dmaengine_prep_slave_dma_vec(). Maybe with a new "flags" parameter to the function? - [4/8]: Implement .device_prep_slave_dma_vec() instead of V3's .device_prep_slave_dma_array(). @Vinod: this patch will cause a small conflict with my other patchset adding scatter-gather support to the axi-dmac driver. This patch adds a call to axi_dmac_alloc_desc(num_sgs), but the prototype of this function changed in my other patchset - it would have to be passed the "chan" variable. I don't know how you prefer it to be resolved. Worst case scenario (and if @Jonathan is okay with that) this one patch can be re-sent later, but it would make this patchset less "atomic". - [5/8]: - Use dev_err() instead of pr_err() - Inline to_iio_dma_fence() - Add comment to explain why we unref twice when detaching dmabuf - Remove TODO comment. It is actually safe to free the file's private data even when transfers are
Re: [PATCH v5 0/8] iio: new DMABUF based API, v5
On Thu, 21 Dec 2023 18:56:52 +0100 Paul Cercueil wrote: > Hi Jonathan, > > Le jeudi 21 décembre 2023 à 16:30 +, Jonathan Cameron a écrit : > > On Tue, 19 Dec 2023 18:50:01 +0100 > > Paul Cercueil wrote: > > > > > [V4 was: "iio: Add buffer write() support"][1] > > > > > > Hi Jonathan, > > > > > Hi Paul, > > > > > This is a respin of the V3 of my patchset that introduced a new > > > interface based on DMABUF objects [2]. > > > > Great to see this moving forwards. > > > > > > > > The V4 was a split of the patchset, to attempt to upstream buffer > > > write() support first. But since there is no current user upstream, > > > it > > > was not merged. This V5 is about doing the opposite, and contains > > > the > > > new DMABUF interface, without adding the buffer write() support. It > > > can > > > already be used with the upstream adi-axi-adc driver. > > > > Seems like a sensible path in the short term. > > > > > > > > In user-space, Libiio uses it to transfer back and forth blocks of > > > samples between the hardware and the applications, without having > > > to > > > copy the data. > > > > > > On a ZCU102 with a FMComms3 daughter board, running Libiio from the > > > pcercuei/dev-new-dmabuf-api branch [3], compiled with > > > WITH_LOCAL_DMABUF_API=OFF (so that it uses fileio): > > > sudo utils/iio_rwdev -b 4096 -B cf-ad9361-lpc > > > Throughput: 116 MiB/s > > > > > > Same hardware, with the DMABUF API (WITH_LOCAL_DMABUF_API=ON): > > > sudo utils/iio_rwdev -b 4096 -B cf-ad9361-lpc > > > Throughput: 475 MiB/s > > > > > > This benchmark only measures the speed at which the data can be > > > fetched > > > to iio_rwdev's internal buffers, and does not actually try to read > > > the > > > data (e.g. to pipe it to stdout). It shows that fetching the data > > > is > > > more than 4x faster using the new interface. > > > > > > When actually reading the data, the performance difference isn't > > > that > > > impressive (maybe because in case of DMABUF the data is not in > > > cache): > > > > This needs a bit more investigation ideally. Perhaps perf counters > > can be > > used to establish that cache misses are the main different between > > dropping it on the floor and actually reading the data. > > Yes, we'll work on it. The other big difference is that fileio uses > dma_alloc_coherent() while the DMABUFs use non-coherent mappings. I > guess coherent memory is faster for the typical access pattern (which > is "read/write everything sequentially once"). Long time since I last worked much with a platform that wasn't always IO coherent, so I've forgotten how all this works (all ends up as no-ops on platforms I tend to use these days!) Good luck, I'll be interested to see what this turns out to be. > > > > > > > WITH_LOCAL_DMABUF_API=OFF (so that it uses fileio): > > > sudo utils/iio_rwdev -b 4096 cf-ad9361-lpc | dd of=/dev/zero > > > status=progress > > > 2446422528 bytes (2.4 GB, 2.3 GiB) copied, 22 s, 111 MB/s > > > > > > WITH_LOCAL_DMABUF_API=ON: > > > sudo utils/iio_rwdev -b 4096 cf-ad9361-lpc | dd of=/dev/zero > > > status=progress > > > 2334388736 bytes (2.3 GB, 2.2 GiB) copied, 21 s, 114 MB/s > > > > > > One interesting thing to note is that fileio is (currently) > > > actually > > > faster than the DMABUF interface if you increase a lot the buffer > > > size. > > > My explanation is that the cache invalidation routine takes more > > > and > > > more time the bigger the DMABUF gets. This is because the DMABUF is > > > backed by small-size pages, so a (e.g.) 64 MiB DMABUF is backed by > > > up > > > to 16 thousands pages, that have to be invalidated one by one. This > > > can > > > be addressed by using huge pages, but the udmabuf driver does not > > > (yet) > > > support creating DMABUFs backed by huge pages. > > > > I'd imagine folios of reasonable size will help sort of a huge page > > as then hopefully it will use the flush by va range instructions if > > available. > > > > > > > > Anyway, the real benefits happen when the DMABUFs are either shared > > > between IIO devices, or between the IIO subsystem and another > > > filesystem. In that case, the DMABUFs are simply passed around > > > drivers, > > > without the data being copied at any moment. > > > > > > We use that feature to transfer samples from our transceivers to > > > USB, > > > using a DMABUF interface to FunctionFS [4]. > > > > > > This drastically increases the throughput, to about 274 MiB/s over > > > a > > > USB3 link, vs. 127 MiB/s using IIO's fileio interface + write() to > > > the > > > FunctionFS endpoints, for a lower CPU usage (0.85 vs. 0.65 load > > > avg.). > > > > This is a nice example. Where are you with getting the patch merged? > > I'll send a new version (mostly a [RESEND]...) in the coming days. As > you can see from the review on my last attempt, the main blocker is > that nobody wants to merge a new interface if the rest of the kernel > bits aren't
Re: [PATCH v5 0/8] iio: new DMABUF based API, v5
Hi Jonathan, Le jeudi 21 décembre 2023 à 16:30 +, Jonathan Cameron a écrit : > On Tue, 19 Dec 2023 18:50:01 +0100 > Paul Cercueil wrote: > > > [V4 was: "iio: Add buffer write() support"][1] > > > > Hi Jonathan, > > > Hi Paul, > > > This is a respin of the V3 of my patchset that introduced a new > > interface based on DMABUF objects [2]. > > Great to see this moving forwards. > > > > > The V4 was a split of the patchset, to attempt to upstream buffer > > write() support first. But since there is no current user upstream, > > it > > was not merged. This V5 is about doing the opposite, and contains > > the > > new DMABUF interface, without adding the buffer write() support. It > > can > > already be used with the upstream adi-axi-adc driver. > > Seems like a sensible path in the short term. > > > > > In user-space, Libiio uses it to transfer back and forth blocks of > > samples between the hardware and the applications, without having > > to > > copy the data. > > > > On a ZCU102 with a FMComms3 daughter board, running Libiio from the > > pcercuei/dev-new-dmabuf-api branch [3], compiled with > > WITH_LOCAL_DMABUF_API=OFF (so that it uses fileio): > > sudo utils/iio_rwdev -b 4096 -B cf-ad9361-lpc > > Throughput: 116 MiB/s > > > > Same hardware, with the DMABUF API (WITH_LOCAL_DMABUF_API=ON): > > sudo utils/iio_rwdev -b 4096 -B cf-ad9361-lpc > > Throughput: 475 MiB/s > > > > This benchmark only measures the speed at which the data can be > > fetched > > to iio_rwdev's internal buffers, and does not actually try to read > > the > > data (e.g. to pipe it to stdout). It shows that fetching the data > > is > > more than 4x faster using the new interface. > > > > When actually reading the data, the performance difference isn't > > that > > impressive (maybe because in case of DMABUF the data is not in > > cache): > > This needs a bit more investigation ideally. Perhaps perf counters > can be > used to establish that cache misses are the main different between > dropping it on the floor and actually reading the data. Yes, we'll work on it. The other big difference is that fileio uses dma_alloc_coherent() while the DMABUFs use non-coherent mappings. I guess coherent memory is faster for the typical access pattern (which is "read/write everything sequentially once"). > > > > WITH_LOCAL_DMABUF_API=OFF (so that it uses fileio): > > sudo utils/iio_rwdev -b 4096 cf-ad9361-lpc | dd of=/dev/zero > > status=progress > > 2446422528 bytes (2.4 GB, 2.3 GiB) copied, 22 s, 111 MB/s > > > > WITH_LOCAL_DMABUF_API=ON: > > sudo utils/iio_rwdev -b 4096 cf-ad9361-lpc | dd of=/dev/zero > > status=progress > > 2334388736 bytes (2.3 GB, 2.2 GiB) copied, 21 s, 114 MB/s > > > > One interesting thing to note is that fileio is (currently) > > actually > > faster than the DMABUF interface if you increase a lot the buffer > > size. > > My explanation is that the cache invalidation routine takes more > > and > > more time the bigger the DMABUF gets. This is because the DMABUF is > > backed by small-size pages, so a (e.g.) 64 MiB DMABUF is backed by > > up > > to 16 thousands pages, that have to be invalidated one by one. This > > can > > be addressed by using huge pages, but the udmabuf driver does not > > (yet) > > support creating DMABUFs backed by huge pages. > > I'd imagine folios of reasonable size will help sort of a huge page > as then hopefully it will use the flush by va range instructions if > available. > > > > > Anyway, the real benefits happen when the DMABUFs are either shared > > between IIO devices, or between the IIO subsystem and another > > filesystem. In that case, the DMABUFs are simply passed around > > drivers, > > without the data being copied at any moment. > > > > We use that feature to transfer samples from our transceivers to > > USB, > > using a DMABUF interface to FunctionFS [4]. > > > > This drastically increases the throughput, to about 274 MiB/s over > > a > > USB3 link, vs. 127 MiB/s using IIO's fileio interface + write() to > > the > > FunctionFS endpoints, for a lower CPU usage (0.85 vs. 0.65 load > > avg.). > > This is a nice example. Where are you with getting the patch merged? I'll send a new version (mostly a [RESEND]...) in the coming days. As you can see from the review on my last attempt, the main blocker is that nobody wants to merge a new interface if the rest of the kernel bits aren't upstream yet. Kind of a chicken-and-egg problem :) > Overall, this code looks fine to me, though there are some parts that > need review by other maintainers (e.g. Vinod for the dmaengine > callback) > and I'd like a 'looks fine' at least form those who know a lot more > about dmabuf than I do. > > To actually make this useful sounds like either udmabuf needs some > perf > improvements, or there has to be an upstream case of sharing it > without > something else (e.g your functionfs patches). So what do we need to > get in before the positive benefit becomes worth
Re: [PATCH v5 0/8] iio: new DMABUF based API, v5
On Tue, 19 Dec 2023 18:50:01 +0100 Paul Cercueil wrote: > [V4 was: "iio: Add buffer write() support"][1] > > Hi Jonathan, > Hi Paul, > This is a respin of the V3 of my patchset that introduced a new > interface based on DMABUF objects [2]. Great to see this moving forwards. > > The V4 was a split of the patchset, to attempt to upstream buffer > write() support first. But since there is no current user upstream, it > was not merged. This V5 is about doing the opposite, and contains the > new DMABUF interface, without adding the buffer write() support. It can > already be used with the upstream adi-axi-adc driver. Seems like a sensible path in the short term. > > In user-space, Libiio uses it to transfer back and forth blocks of > samples between the hardware and the applications, without having to > copy the data. > > On a ZCU102 with a FMComms3 daughter board, running Libiio from the > pcercuei/dev-new-dmabuf-api branch [3], compiled with > WITH_LOCAL_DMABUF_API=OFF (so that it uses fileio): > sudo utils/iio_rwdev -b 4096 -B cf-ad9361-lpc > Throughput: 116 MiB/s > > Same hardware, with the DMABUF API (WITH_LOCAL_DMABUF_API=ON): > sudo utils/iio_rwdev -b 4096 -B cf-ad9361-lpc > Throughput: 475 MiB/s > > This benchmark only measures the speed at which the data can be fetched > to iio_rwdev's internal buffers, and does not actually try to read the > data (e.g. to pipe it to stdout). It shows that fetching the data is > more than 4x faster using the new interface. > > When actually reading the data, the performance difference isn't that > impressive (maybe because in case of DMABUF the data is not in cache): This needs a bit more investigation ideally. Perhaps perf counters can be used to establish that cache misses are the main different between dropping it on the floor and actually reading the data. > > WITH_LOCAL_DMABUF_API=OFF (so that it uses fileio): > sudo utils/iio_rwdev -b 4096 cf-ad9361-lpc | dd of=/dev/zero status=progress > 2446422528 bytes (2.4 GB, 2.3 GiB) copied, 22 s, 111 MB/s > > WITH_LOCAL_DMABUF_API=ON: > sudo utils/iio_rwdev -b 4096 cf-ad9361-lpc | dd of=/dev/zero status=progress > 2334388736 bytes (2.3 GB, 2.2 GiB) copied, 21 s, 114 MB/s > > One interesting thing to note is that fileio is (currently) actually > faster than the DMABUF interface if you increase a lot the buffer size. > My explanation is that the cache invalidation routine takes more and > more time the bigger the DMABUF gets. This is because the DMABUF is > backed by small-size pages, so a (e.g.) 64 MiB DMABUF is backed by up > to 16 thousands pages, that have to be invalidated one by one. This can > be addressed by using huge pages, but the udmabuf driver does not (yet) > support creating DMABUFs backed by huge pages. I'd imagine folios of reasonable size will help sort of a huge page as then hopefully it will use the flush by va range instructions if available. > > Anyway, the real benefits happen when the DMABUFs are either shared > between IIO devices, or between the IIO subsystem and another > filesystem. In that case, the DMABUFs are simply passed around drivers, > without the data being copied at any moment. > > We use that feature to transfer samples from our transceivers to USB, > using a DMABUF interface to FunctionFS [4]. > > This drastically increases the throughput, to about 274 MiB/s over a > USB3 link, vs. 127 MiB/s using IIO's fileio interface + write() to the > FunctionFS endpoints, for a lower CPU usage (0.85 vs. 0.65 load avg.). This is a nice example. Where are you with getting the patch merged? Overall, this code looks fine to me, though there are some parts that need review by other maintainers (e.g. Vinod for the dmaengine callback) and I'd like a 'looks fine' at least form those who know a lot more about dmabuf than I do. To actually make this useful sounds like either udmabuf needs some perf improvements, or there has to be an upstream case of sharing it without something else (e.g your functionfs patches). So what do we need to get in before the positive benefit becomes worth carrying this extra complexity? (which isn't too bad so I'm fine with a small benefit and promises of riches :) Jonathan > > Based on linux-next/next-20231219. > > Cheers, > -Paul > > [1] https://lore.kernel.org/all/20230807112113.47157-1-p...@crapouillou.net/ > [2] https://lore.kernel.org/all/20230403154800.215924-1-p...@crapouillou.net/ > [3] > https://github.com/analogdevicesinc/libiio/tree/pcercuei/dev-new-dmabuf-api > [4] https://lore.kernel.org/all/20230322092118.9213-1-p...@crapouillou.net/ > > --- > Changelog: > - [3/8]: Replace V3's dmaengine_prep_slave_dma_array() with a new > dmaengine_prep_slave_dma_vec(), which uses a new 'dma_vec' struct. > Note that at some point we will need to support cyclic transfers > using dmaengine_prep_slave_dma_vec(). Maybe with a new "flags" > parameter to the function? > > - [4/8]: Implement .device_prep_slave_dma_vec()
[PATCH v5 0/8] iio: new DMABUF based API, v5
[V4 was: "iio: Add buffer write() support"][1] Hi Jonathan, This is a respin of the V3 of my patchset that introduced a new interface based on DMABUF objects [2]. The V4 was a split of the patchset, to attempt to upstream buffer write() support first. But since there is no current user upstream, it was not merged. This V5 is about doing the opposite, and contains the new DMABUF interface, without adding the buffer write() support. It can already be used with the upstream adi-axi-adc driver. In user-space, Libiio uses it to transfer back and forth blocks of samples between the hardware and the applications, without having to copy the data. On a ZCU102 with a FMComms3 daughter board, running Libiio from the pcercuei/dev-new-dmabuf-api branch [3], compiled with WITH_LOCAL_DMABUF_API=OFF (so that it uses fileio): sudo utils/iio_rwdev -b 4096 -B cf-ad9361-lpc Throughput: 116 MiB/s Same hardware, with the DMABUF API (WITH_LOCAL_DMABUF_API=ON): sudo utils/iio_rwdev -b 4096 -B cf-ad9361-lpc Throughput: 475 MiB/s This benchmark only measures the speed at which the data can be fetched to iio_rwdev's internal buffers, and does not actually try to read the data (e.g. to pipe it to stdout). It shows that fetching the data is more than 4x faster using the new interface. When actually reading the data, the performance difference isn't that impressive (maybe because in case of DMABUF the data is not in cache): WITH_LOCAL_DMABUF_API=OFF (so that it uses fileio): sudo utils/iio_rwdev -b 4096 cf-ad9361-lpc | dd of=/dev/zero status=progress 2446422528 bytes (2.4 GB, 2.3 GiB) copied, 22 s, 111 MB/s WITH_LOCAL_DMABUF_API=ON: sudo utils/iio_rwdev -b 4096 cf-ad9361-lpc | dd of=/dev/zero status=progress 2334388736 bytes (2.3 GB, 2.2 GiB) copied, 21 s, 114 MB/s One interesting thing to note is that fileio is (currently) actually faster than the DMABUF interface if you increase a lot the buffer size. My explanation is that the cache invalidation routine takes more and more time the bigger the DMABUF gets. This is because the DMABUF is backed by small-size pages, so a (e.g.) 64 MiB DMABUF is backed by up to 16 thousands pages, that have to be invalidated one by one. This can be addressed by using huge pages, but the udmabuf driver does not (yet) support creating DMABUFs backed by huge pages. Anyway, the real benefits happen when the DMABUFs are either shared between IIO devices, or between the IIO subsystem and another filesystem. In that case, the DMABUFs are simply passed around drivers, without the data being copied at any moment. We use that feature to transfer samples from our transceivers to USB, using a DMABUF interface to FunctionFS [4]. This drastically increases the throughput, to about 274 MiB/s over a USB3 link, vs. 127 MiB/s using IIO's fileio interface + write() to the FunctionFS endpoints, for a lower CPU usage (0.85 vs. 0.65 load avg.). Based on linux-next/next-20231219. Cheers, -Paul [1] https://lore.kernel.org/all/20230807112113.47157-1-p...@crapouillou.net/ [2] https://lore.kernel.org/all/20230403154800.215924-1-p...@crapouillou.net/ [3] https://github.com/analogdevicesinc/libiio/tree/pcercuei/dev-new-dmabuf-api [4] https://lore.kernel.org/all/20230322092118.9213-1-p...@crapouillou.net/ --- Changelog: - [3/8]: Replace V3's dmaengine_prep_slave_dma_array() with a new dmaengine_prep_slave_dma_vec(), which uses a new 'dma_vec' struct. Note that at some point we will need to support cyclic transfers using dmaengine_prep_slave_dma_vec(). Maybe with a new "flags" parameter to the function? - [4/8]: Implement .device_prep_slave_dma_vec() instead of V3's .device_prep_slave_dma_array(). @Vinod: this patch will cause a small conflict with my other patchset adding scatter-gather support to the axi-dmac driver. This patch adds a call to axi_dmac_alloc_desc(num_sgs), but the prototype of this function changed in my other patchset - it would have to be passed the "chan" variable. I don't know how you prefer it to be resolved. Worst case scenario (and if @Jonathan is okay with that) this one patch can be re-sent later, but it would make this patchset less "atomic". - [5/8]: - Use dev_err() instead of pr_err() - Inline to_iio_dma_fence() - Add comment to explain why we unref twice when detaching dmabuf - Remove TODO comment. It is actually safe to free the file's private data even when transfers are still pending because it won't be accessed. - Fix documentation of new fields in struct iio_buffer_access_funcs - iio_dma_resv_lock() does not need to be exported, make it static - [7/8]: - Use the new dmaengine_prep_slave_dma_vec(). - Restrict to input buffers, since output buffers are not yet supported by IIO buffers. - [8/8]: Use description lists for the documentation of the three new IOCTLs instead of abusing subsections. --- Alexandru Ardelean (1): iio: buffer-dma: split iio_dma_buffer_fileio_free() function Paul Cercueil (7):