Re: [Intel-gfx] [PATCH] drm-buf: Add debug option
On Thu, Jan 14, 2021 at 1:03 AM Daniel Vetter wrote: > > On Wed, Jan 13, 2021 at 10:08 PM Chris Wilson > wrote: > > Quoting Daniel Vetter (2021-01-13 20:50:11) > > > On Wed, Jan 13, 2021 at 4:43 PM Chris Wilson > > > wrote: > > > > > > > > Quoting Daniel Vetter (2021-01-13 14:06:04) > > > > > We have too many people abusing the struct page they can get at but > > > > > really shouldn't in importers. Aside from that the backing page might > > > > > simply not exist (for dynamic p2p mappings) looking at it and using it > > > > > e.g. for mmap can also wreak the page handling of the exporter > > > > > completely. Importers really must go through the proper interface like > > > > > dma_buf_mmap for everything. > > > > > > > > If the exporter doesn't want to expose the struct page, why are they > > > > setting it in the exported sg_table? > > > > > > You need to store it somewhere, otherwise the dma-api doesn't work. > > > Essentially this achieves clearing/resetting the struct page pointer, > > > without additional allocations somewhere, or tons of driver changes > > > (since presumably the driver does keep track of the struct page > > > somewhere too). > > > > Only for mapping, and that's before the export -- if there's even a > > struct page to begin with. > > > > > Also as long as we have random importers looking at struct page we > > > can't just remove it, or crashes everywhere. So it has to be some > > > debug option you can disable. > > > > Totally agreed that nothing generic can rely on pages being transported > > via dma-buf, and memfd is there if you do want a suitable transport. The > > one I don't know about is dma-buf heap, do both parties there consent to > > transport pages via the dma-buf? i.e. do they have special cases for > > import/export between heaps? > > heaps shouldn't be any different wrt the interface exposed to > importers. Adding John just in case I missed something. I'm not aware of how this would be an issue right off for dma-buf heaps. Obviously there may be some corner cases with things like secure heaps, but I've not gotten to work on any of those yet and there's none in-tree. I did test out the patch on HiKey960 (using the cma and system heap for display and gpu buffers - admittedly not particularly complex) and didn't see any issues with it enabled. I've added Suren and Hridya for more input but don't have any objections right off. thanks -john ___ dri-devel mailing list dri-devel@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/dri-devel
Re: [Intel-gfx] [PATCH] drm-buf: Add debug option
Quoting Daniel Vetter (2021-01-14 09:47:40) > On Thu, Jan 14, 2021 at 09:45:37AM +, Chris Wilson wrote: > > Quoting Daniel Vetter (2021-01-14 09:30:32) > > > On Thu, Jan 14, 2021 at 10:23 AM Chris Wilson > > > wrote: > > > > The only other problem I see with the implementation is that there's > > > > nothing that says that each dmabuf->ops->map_dma_buf() returns a new > > > > sg_table, so we may end up undoing the xor. Or should each dma-buf > > > > return a fresh dma-mapping for iommu isolation? > > > > > > Maybe I screwed it up, but that's why I extracted the little helpers: > > > We scramble when we get the sgtable from exporter, and unscramble > > > before we pass it back. dma-buf.c does some caching and will hand back > > > the same sgtable, but for that case we don't re-scramble. > > > > The attachment is only mapped once, but there can be more than one > > attachment, and the backend could return the same sg_table for each > > mapping. Conceivably, it could return its own private sg_table where it > > wants to maintain the struct page. Seems like just adding a sentence to > > @map_dma_buf to clarify that each call should return a new sg_table will > > suffice. > > Ah yes good point, will augment (once CI stops being angry at me). Fwiw, with a quick explanation of "don't do this" in the docs, Reviewed-by: Chris Wilson -Chris ___ dri-devel mailing list dri-devel@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/dri-devel
Re: [Intel-gfx] [PATCH] drm-buf: Add debug option
On Thu, Jan 14, 2021 at 09:45:37AM +, Chris Wilson wrote: > Quoting Daniel Vetter (2021-01-14 09:30:32) > > On Thu, Jan 14, 2021 at 10:23 AM Chris Wilson > > wrote: > > > The only other problem I see with the implementation is that there's > > > nothing that says that each dmabuf->ops->map_dma_buf() returns a new > > > sg_table, so we may end up undoing the xor. Or should each dma-buf > > > return a fresh dma-mapping for iommu isolation? > > > > Maybe I screwed it up, but that's why I extracted the little helpers: > > We scramble when we get the sgtable from exporter, and unscramble > > before we pass it back. dma-buf.c does some caching and will hand back > > the same sgtable, but for that case we don't re-scramble. > > The attachment is only mapped once, but there can be more than one > attachment, and the backend could return the same sg_table for each > mapping. Conceivably, it could return its own private sg_table where it > wants to maintain the struct page. Seems like just adding a sentence to > @map_dma_buf to clarify that each call should return a new sg_table will > suffice. Ah yes good point, will augment (once CI stops being angry at me). -Daniel -- Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch ___ dri-devel mailing list dri-devel@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/dri-devel
Re: [Intel-gfx] [PATCH] drm-buf: Add debug option
Quoting Daniel Vetter (2021-01-14 09:30:32) > On Thu, Jan 14, 2021 at 10:23 AM Chris Wilson > wrote: > > The only other problem I see with the implementation is that there's > > nothing that says that each dmabuf->ops->map_dma_buf() returns a new > > sg_table, so we may end up undoing the xor. Or should each dma-buf > > return a fresh dma-mapping for iommu isolation? > > Maybe I screwed it up, but that's why I extracted the little helpers: > We scramble when we get the sgtable from exporter, and unscramble > before we pass it back. dma-buf.c does some caching and will hand back > the same sgtable, but for that case we don't re-scramble. The attachment is only mapped once, but there can be more than one attachment, and the backend could return the same sg_table for each mapping. Conceivably, it could return its own private sg_table where it wants to maintain the struct page. Seems like just adding a sentence to @map_dma_buf to clarify that each call should return a new sg_table will suffice. -Chris ___ dri-devel mailing list dri-devel@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/dri-devel
Re: [Intel-gfx] [PATCH] drm-buf: Add debug option
On Thu, Jan 14, 2021 at 10:23 AM Chris Wilson wrote: > > Quoting Daniel Vetter (2021-01-14 09:02:57) > > On Wed, Jan 13, 2021 at 10:08 PM Chris Wilson > > wrote: > > > Quoting Daniel Vetter (2021-01-13 20:50:11) > > > > On Wed, Jan 13, 2021 at 4:43 PM Chris Wilson > > > > wrote: > > > > > > > > > > Quoting Daniel Vetter (2021-01-13 14:06:04) > > > > > > We have too many people abusing the struct page they can get at but > > > > > > really shouldn't in importers. Aside from that the backing page > > > > > > might > > > > > > simply not exist (for dynamic p2p mappings) looking at it and using > > > > > > it > > > > > > e.g. for mmap can also wreak the page handling of the exporter > > > > > > completely. Importers really must go through the proper interface > > > > > > like > > > > > > dma_buf_mmap for everything. > > > > > > > > > > If the exporter doesn't want to expose the struct page, why are they > > > > > setting it in the exported sg_table? > > > > > > > > You need to store it somewhere, otherwise the dma-api doesn't work. > > > > Essentially this achieves clearing/resetting the struct page pointer, > > > > without additional allocations somewhere, or tons of driver changes > > > > (since presumably the driver does keep track of the struct page > > > > somewhere too). > > > > > > Only for mapping, and that's before the export -- if there's even a > > > struct page to begin with. > > > > > > > Also as long as we have random importers looking at struct page we > > > > can't just remove it, or crashes everywhere. So it has to be some > > > > debug option you can disable. > > > > > > Totally agreed that nothing generic can rely on pages being transported > > > via dma-buf, and memfd is there if you do want a suitable transport. The > > > one I don't know about is dma-buf heap, do both parties there consent to > > > transport pages via the dma-buf? i.e. do they have special cases for > > > import/export between heaps? > > > > heaps shouldn't be any different wrt the interface exposed to > > importers. Adding John just in case I missed something. > > > > I think the only problem we have is that the first import for ttm > > simply pulled out the struct page and ignored the sgtable otherwise, > > then that copypasted to places and we're still have some of that left. > > Although it's a lot better. So largely the problem is importers being > > a bit silly. > > > > I also think I should change the defaulty y to default y if > > DMA_API_DEBUG or something like that, to make sure it's actually > > enabled often enough. > > It felt overly draconian, but other than the open question of dma-buf > heaps (which I realise that we need some CI coverage for), I can't > think of a good reason to argue for hiding a struct page transport > within dma-buf. Yeah there's the occasional idea floating around to split sgtable into the page and the dma side completely. But aside from the bikeshed no one volunteered for the massive amount of work rolling that out would mean, so I'm trying to go with a cheap trick here meanwhile. > The only other problem I see with the implementation is that there's > nothing that says that each dmabuf->ops->map_dma_buf() returns a new > sg_table, so we may end up undoing the xor. Or should each dma-buf > return a fresh dma-mapping for iommu isolation? Maybe I screwed it up, but that's why I extracted the little helpers: We scramble when we get the sgtable from exporter, and unscramble before we pass it back. dma-buf.c does some caching and will hand back the same sgtable, but for that case we don't re-scramble. -Daniel -- Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch ___ dri-devel mailing list dri-devel@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/dri-devel
Re: [Intel-gfx] [PATCH] drm-buf: Add debug option
Quoting Daniel Vetter (2021-01-14 09:02:57) > On Wed, Jan 13, 2021 at 10:08 PM Chris Wilson > wrote: > > Quoting Daniel Vetter (2021-01-13 20:50:11) > > > On Wed, Jan 13, 2021 at 4:43 PM Chris Wilson > > > wrote: > > > > > > > > Quoting Daniel Vetter (2021-01-13 14:06:04) > > > > > We have too many people abusing the struct page they can get at but > > > > > really shouldn't in importers. Aside from that the backing page might > > > > > simply not exist (for dynamic p2p mappings) looking at it and using it > > > > > e.g. for mmap can also wreak the page handling of the exporter > > > > > completely. Importers really must go through the proper interface like > > > > > dma_buf_mmap for everything. > > > > > > > > If the exporter doesn't want to expose the struct page, why are they > > > > setting it in the exported sg_table? > > > > > > You need to store it somewhere, otherwise the dma-api doesn't work. > > > Essentially this achieves clearing/resetting the struct page pointer, > > > without additional allocations somewhere, or tons of driver changes > > > (since presumably the driver does keep track of the struct page > > > somewhere too). > > > > Only for mapping, and that's before the export -- if there's even a > > struct page to begin with. > > > > > Also as long as we have random importers looking at struct page we > > > can't just remove it, or crashes everywhere. So it has to be some > > > debug option you can disable. > > > > Totally agreed that nothing generic can rely on pages being transported > > via dma-buf, and memfd is there if you do want a suitable transport. The > > one I don't know about is dma-buf heap, do both parties there consent to > > transport pages via the dma-buf? i.e. do they have special cases for > > import/export between heaps? > > heaps shouldn't be any different wrt the interface exposed to > importers. Adding John just in case I missed something. > > I think the only problem we have is that the first import for ttm > simply pulled out the struct page and ignored the sgtable otherwise, > then that copypasted to places and we're still have some of that left. > Although it's a lot better. So largely the problem is importers being > a bit silly. > > I also think I should change the defaulty y to default y if > DMA_API_DEBUG or something like that, to make sure it's actually > enabled often enough. It felt overly draconian, but other than the open question of dma-buf heaps (which I realise that we need some CI coverage for), I can't think of a good reason to argue for hiding a struct page transport within dma-buf. The only other problem I see with the implementation is that there's nothing that says that each dmabuf->ops->map_dma_buf() returns a new sg_table, so we may end up undoing the xor. Or should each dma-buf return a fresh dma-mapping for iommu isolation? -Chris ___ dri-devel mailing list dri-devel@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/dri-devel
Re: [Intel-gfx] [PATCH] drm-buf: Add debug option
On Wed, Jan 13, 2021 at 10:08 PM Chris Wilson wrote: > Quoting Daniel Vetter (2021-01-13 20:50:11) > > On Wed, Jan 13, 2021 at 4:43 PM Chris Wilson > > wrote: > > > > > > Quoting Daniel Vetter (2021-01-13 14:06:04) > > > > We have too many people abusing the struct page they can get at but > > > > really shouldn't in importers. Aside from that the backing page might > > > > simply not exist (for dynamic p2p mappings) looking at it and using it > > > > e.g. for mmap can also wreak the page handling of the exporter > > > > completely. Importers really must go through the proper interface like > > > > dma_buf_mmap for everything. > > > > > > If the exporter doesn't want to expose the struct page, why are they > > > setting it in the exported sg_table? > > > > You need to store it somewhere, otherwise the dma-api doesn't work. > > Essentially this achieves clearing/resetting the struct page pointer, > > without additional allocations somewhere, or tons of driver changes > > (since presumably the driver does keep track of the struct page > > somewhere too). > > Only for mapping, and that's before the export -- if there's even a > struct page to begin with. > > > Also as long as we have random importers looking at struct page we > > can't just remove it, or crashes everywhere. So it has to be some > > debug option you can disable. > > Totally agreed that nothing generic can rely on pages being transported > via dma-buf, and memfd is there if you do want a suitable transport. The > one I don't know about is dma-buf heap, do both parties there consent to > transport pages via the dma-buf? i.e. do they have special cases for > import/export between heaps? heaps shouldn't be any different wrt the interface exposed to importers. Adding John just in case I missed something. I think the only problem we have is that the first import for ttm simply pulled out the struct page and ignored the sgtable otherwise, then that copypasted to places and we're still have some of that left. Although it's a lot better. So largely the problem is importers being a bit silly. I also think I should change the defaulty y to default y if DMA_API_DEBUG or something like that, to make sure it's actually enabled often enough. -Daniel -- Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch ___ dri-devel mailing list dri-devel@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/dri-devel
Re: [Intel-gfx] [PATCH] drm-buf: Add debug option
Quoting Daniel Vetter (2021-01-13 20:50:11) > On Wed, Jan 13, 2021 at 4:43 PM Chris Wilson wrote: > > > > Quoting Daniel Vetter (2021-01-13 14:06:04) > > > We have too many people abusing the struct page they can get at but > > > really shouldn't in importers. Aside from that the backing page might > > > simply not exist (for dynamic p2p mappings) looking at it and using it > > > e.g. for mmap can also wreak the page handling of the exporter > > > completely. Importers really must go through the proper interface like > > > dma_buf_mmap for everything. > > > > If the exporter doesn't want to expose the struct page, why are they > > setting it in the exported sg_table? > > You need to store it somewhere, otherwise the dma-api doesn't work. > Essentially this achieves clearing/resetting the struct page pointer, > without additional allocations somewhere, or tons of driver changes > (since presumably the driver does keep track of the struct page > somewhere too). Only for mapping, and that's before the export -- if there's even a struct page to begin with. > Also as long as we have random importers looking at struct page we > can't just remove it, or crashes everywhere. So it has to be some > debug option you can disable. Totally agreed that nothing generic can rely on pages being transported via dma-buf, and memfd is there if you do want a suitable transport. The one I don't know about is dma-buf heap, do both parties there consent to transport pages via the dma-buf? i.e. do they have special cases for import/export between heaps? -Chris ___ dri-devel mailing list dri-devel@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/dri-devel
Re: [Intel-gfx] [PATCH] drm-buf: Add debug option
On Wed, Jan 13, 2021 at 4:43 PM Chris Wilson wrote: > > Quoting Daniel Vetter (2021-01-13 14:06:04) > > We have too many people abusing the struct page they can get at but > > really shouldn't in importers. Aside from that the backing page might > > simply not exist (for dynamic p2p mappings) looking at it and using it > > e.g. for mmap can also wreak the page handling of the exporter > > completely. Importers really must go through the proper interface like > > dma_buf_mmap for everything. > > If the exporter doesn't want to expose the struct page, why are they > setting it in the exported sg_table? You need to store it somewhere, otherwise the dma-api doesn't work. Essentially this achieves clearing/resetting the struct page pointer, without additional allocations somewhere, or tons of driver changes (since presumably the driver does keep track of the struct page somewhere too). Also as long as we have random importers looking at struct page we can't just remove it, or crashes everywhere. So it has to be some debug option you can disable. -Daniel -- Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch ___ dri-devel mailing list dri-devel@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/dri-devel
Re: [Intel-gfx] [PATCH] drm-buf: Add debug option
Quoting Daniel Vetter (2021-01-13 14:06:04) > We have too many people abusing the struct page they can get at but > really shouldn't in importers. Aside from that the backing page might > simply not exist (for dynamic p2p mappings) looking at it and using it > e.g. for mmap can also wreak the page handling of the exporter > completely. Importers really must go through the proper interface like > dma_buf_mmap for everything. If the exporter doesn't want to expose the struct page, why are they setting it in the exported sg_table? -Chris ___ dri-devel mailing list dri-devel@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/dri-devel
