)]}' { "log": [ { "commit": "d4e8b7193d87d80f3364d7924692988c974780aa", "tree": "04f54438fab0d64e5cfe23debbd622c9cfc34f4b", "parents": [ "9ae7e01de1cf03c77054da44d135a7e85863fcb0" ], "author": { "name": "Daniel Vetter", "email": "daniel.vetter@ffwll.ch", "time": "Wed Apr 18 15:52:26 2012 +0200" }, "committer": { "name": "Rebecca Schultz Zavin", "email": "rebecca@android.com", "time": "Tue May 22 14:01:06 2012 -0700" }, "message": "dma-buf: mmap support\n\nCompared to Rob Clark\u0027s RFC I\u0027ve ditched the prepare/finish hooks\nand corresponding ioctls on the dma_buf file. The major reason for\nthat is that many people seem to be under the impression that this is\nalso for synchronization with outstanding asynchronous processsing.\nI\u0027m pretty massively opposed to this because:\n\n- It boils down reinventing a new rather general-purpose userspace\n synchronization interface. If we look at things like futexes, this\n is hard to get right.\n- Furthermore a lot of kernel code has to interact with this\n synchronization primitive. This smells a look like the dri1 hw_lock,\n a horror show I prefer not to reinvent.\n- Even more fun is that multiple different subsystems would interact\n here, so we have plenty of opportunities to create funny deadlock\n scenarios.\n\nI think synchronization is a wholesale different problem from data\nsharing and should be tackled as an orthogonal problem.\n\nNow we could demand that prepare/finish may only ensure cache\ncoherency (as Rob intended), but that runs up into the next problem:\nWe not only need mmap support to facilitate sw-only processing nodes\nin a pipeline (without jumping through hoops by importing the dma_buf\ninto some sw-access only importer), which allows for a nicer\nION-\u003edma-buf upgrade path for existing Android userspace. We also need\nmmap support for existing importing subsystems to support existing\nuserspace libraries. And a loot of these subsystems are expected to\nexport coherent userspace mappings.\n\nSo prepare/finish can only ever be optional and the exporter /needs/\nto support coherent mappings. Given that mmap access is always\nsomewhat fallback-y in nature I\u0027ve decided to drop this optimization,\ninstead of just making it optional. If we demonstrate a clear need for\nthis, supported by benchmark results, we can always add it in again\nlater as an optional extension.\n\nOther differences compared to Rob\u0027s RFC is the above mentioned support\nfor mapping a dma-buf through facilities provided by the importer.\nWhich results in mmap support no longer being optional.\n\nNote that this dma-buf mmap patch does _not_ support every possible\ninsanity an existing subsystem could pull of with mmap: Because it\ndoes not allow to intercept pagefaults and shoot down ptes importing\nsubsystems can\u0027t add some magic of their own at these points (e.g. to\nautomatically synchronize with outstanding rendering or set up some\nspecial resources). I\u0027ve done a cursory read through a few mmap\nimplementions of various subsytems and I\u0027m hopeful that we can avoid\nthis (and the complexity it\u0027d bring with it).\n\nAdditonally I\u0027ve extended the documentation a bit to explain the hows\nand whys of this mmap extension.\n\nIn case we ever want to add support for explicitly cache maneged\nuserspace mmap with a prepare/finish ioctl pair, we could specify that\nuserspace needs to mmap a different part of the dma_buf, e.g. the\nrange starting at dma_buf-\u003esize up to dma_buf-\u003esize*2. This works\nbecause the size of a dma_buf is invariant over it\u0027s lifetime. The\nexporter would obviously need to fall back to coherent mappings for\nboth ranges if a legacy clients maps the coherent range and the\narchitecture cannot suppor conflicting caching policies. Also, this\nwould obviously be optional and userspace needs to be able to fall\nback to coherent mappings.\n\nv2:\n- Spelling fixes from Rob Clark.\n- Compile fix for !DMA_BUF from Rob Clark.\n- Extend commit message to explain how explicitly cache managed mmap\n support could be added later.\n- Extend the documentation with implementations notes for exporters\n that need to manually fake coherency.\n\nChange-Id: Ia8f2ae5d8a1b1c87ed12ca1c89d7bf2067239ee4\nCc: Rob Clark \u003crob.clark@linaro.org\u003e\nCc: Rebecca Schultz Zavin \u003crebecca@android.com\u003e\nSigned-Off-by: Daniel Vetter \u003cdaniel.vetter@ffwll.ch\u003e\n" }, { "commit": "ed0bb8ea059764c3fc882fb135473afd347335e9", "tree": "5274b8335afe85f76d1eb945eb03ffe4040737b4", "parents": [ "47b816ff7d520509176154748713e7d66b3ad6ac", "3e0b2a1993c06e646d90d71e163d03869a211a4c" ], "author": { "name": "Linus Torvalds", "email": "torvalds@linux-foundation.org", "time": "Wed Mar 28 15:02:41 2012 -0700" }, "committer": { "name": "Linus Torvalds", "email": "torvalds@linux-foundation.org", "time": "Wed Mar 28 15:02:41 2012 -0700" }, "message": "Merge branch \u0027for-linus-3.4\u0027 of git://git.linaro.org/people/sumitsemwal/linux-dma-buf\n\nPull dma-buf updates from Sumit Semwal:\n \"This includes the following key items:\n\n - kernel cpu access support,\n - flag-passing to dma_buf_fd,\n - relevant Documentation updates, and\n - some minor cleanups and fixes.\n\n These changes are needed for the drm prime/dma-buf interface code that\n Dave Airlie plans to submit in this merge window.\"\n\n* \u0027for-linus-3.4\u0027 of git://git.linaro.org/people/sumitsemwal/linux-dma-buf:\n dma-buf: correct dummy function declarations.\n dma-buf: document fd flags and O_CLOEXEC requirement\n dma_buf: Add documentation for the new cpu access support\n dma-buf: add support for kernel cpu access\n dma-buf: don\u0027t hold the mutex around map/unmap calls\n dma-buf: add get_dma_buf()\n dma-buf: pass flags into dma_buf_fd.\n dma-buf: add dma_data_direction to unmap dma_buf_op\n dma-buf: Move code out of mutex-protected section in dma_buf_attach()\n dma-buf: Return error instead of using a goto statement when possible\n dma-buf: Remove unneeded sanity checks\n dma-buf: Constify ops argument to dma_buf_export()\n" }, { "commit": "3e0b2a1993c06e646d90d71e163d03869a211a4c", "tree": "40b950985b5559159f0c9a1dbd5f28a43c51fa77", "parents": [ "fbb231e1a98cb0360b681b6a6195a619e98d7077" ], "author": { "name": "Sumit Semwal", "email": "sumit.semwal@ti.com", "time": "Mon Mar 26 11:29:19 2012 +0530" }, "committer": { "name": "Sumit Semwal", "email": "sumit.semwal@ti.com", "time": "Mon Mar 26 11:33:34 2012 +0530" }, "message": "dma-buf: correct dummy function declarations.\n\nDummy functions for the newly added cpu access ops need variable names for\narguments.\nAlso, the introduction of flags in dma_buf_fd needs to be added to dummy\nfunctions as well.\n\nSigned-off-by: Sumit Semwal \u003csumit.semwal@ti.com\u003e\nSigned-off-by: Sumit Semwal \u003csumit.semwal@linaro.org\u003e\n" }, { "commit": "fc13020e086bfedf2afb95c91c026d5af1f80107", "tree": "a06be5c3eff6865a42046b51270e4d60e30609f4", "parents": [ "6b607e3a658fee490bdabfdeb739a3eb498b1bff" ], "author": { "name": "Daniel Vetter", "email": "daniel.vetter@ffwll.ch", "time": "Tue Mar 20 00:02:37 2012 +0100" }, "committer": { "name": "Sumit Semwal", "email": "sumit.semwal@ti.com", "time": "Mon Mar 26 11:33:02 2012 +0530" }, "message": "dma-buf: add support for kernel cpu access\n\nBig differences to other contenders in the field (like ion) is\nthat this also supports highmem, so we have to split up the cpu\naccess from the kernel side into a prepare and a kmap step.\n\nPrepare is allowed to fail and should do everything required so that\nthe kmap calls can succeed (like swapin/backing storage allocation,\nflushing, ...).\n\nMore in-depth explanations will follow in the follow-up documentation\npatch.\n\nChanges in v2:\n\n- Clear up begin_cpu_access confusion noticed by Sumit Semwal.\n- Don\u0027t automatically fallback from the _atomic variants to the\n non-atomic variants. The _atomic callbacks are not allowed to\n sleep, so we want exporters to make this decision explicit. The\n function signatures are explicit, so simpler exporters can still\n use the same function for both.\n- Make the unmap functions optional. Simpler exporters with permanent\n mappings don\u0027t need to do anything at unmap time.\n\nChanges in v3:\n\n- Adjust the WARN_ON checks for the new -\u003eops functions as suggested\n by Rob Clark and Sumit Semwal.\n- Rebased on top of latest dma-buf-next git.\n\nChanges in v4:\n\n- Fixup a missing - in a return -EINVAL; statement.\n\nSigned-Off-by: Daniel Vetter \u003cdaniel.vetter@ffwll.ch\u003e\nSigned-off-by: Rob Clark \u003crob@ti.com\u003e\nSigned-off-by: Sumit Semwal \u003csumit.semwal@linaro.org\u003e\n" }, { "commit": "6b607e3a658fee490bdabfdeb739a3eb498b1bff", "tree": "a10ad6f713586f9dc7584fe54110d1c44f87b1d5", "parents": [ "f9a24d1ac9cb82baf5ec5efdb6580a9ce0bd5bfc" ], "author": { "name": "Daniel Vetter", "email": "daniel.vetter@ffwll.ch", "time": "Mon Mar 19 00:34:25 2012 +0100" }, "committer": { "name": "Sumit Semwal", "email": "sumit.semwal@ti.com", "time": "Mon Mar 26 11:32:50 2012 +0530" }, "message": "dma-buf: don\u0027t hold the mutex around map/unmap calls\n\nThe mutex protects the attachment list and hence needs to be held\naround the callbakc to the exporters (optional) attach/detach\nfunctions.\n\nHolding the mutex around the map/unmap calls doesn\u0027t protect any\ndma_buf state. Exporters need to properly protect any of their own\nstate anyway (to protect against calls from their own interfaces).\nSo this only makes the locking messier (and lockdep easier to anger).\n\nTherefore let\u0027s just drop this.\n\nv2: Rebased on top of latest dma-buf-next git.\n\nSigned-off-by: Daniel Vetter \u003cdaniel.vetter@ffwll.ch\u003e\nReviewed-by: Rob Clark \u003crob.clark@linaro.org\u003e\nSigned-off-by: Sumit Semwal \u003csumit.semwal@linaro.org\u003e\n" }, { "commit": "f9a24d1ac9cb82baf5ec5efdb6580a9ce0bd5bfc", "tree": "6dd135509fa938eb3c9f8adab6919110a6f01f2b", "parents": [ "55c1c4ca23d0f2736ef7c219d0fb005323ff8ee0" ], "author": { "name": "Rob Clark", "email": "rob@ti.com", "time": "Fri Mar 16 11:04:41 2012 -0500" }, "committer": { "name": "Sumit Semwal", "email": "sumit.semwal@ti.com", "time": "Mon Mar 26 11:32:40 2012 +0530" }, "message": "dma-buf: add get_dma_buf()\n\nWorks in a similar way to get_file(), and is needed in cases such as\nwhen the exporter needs to also keep a reference to the dmabuf (that\nis later released with a dma_buf_put()), and possibly other similar\ncases.\n\nSigned-off-by: Rob Clark \u003crob@ti.com\u003e\nReviewed-by: Dave Airlie \u003cairlied@redhat.com\u003e\nReviewed-by: Kyungmin Park \u003ckyungmin.park@samsung.com\u003e\nSigned-off-by: Sumit Semwal \u003csumit.semwal@linaro.org\u003e\n" }, { "commit": "55c1c4ca23d0f2736ef7c219d0fb005323ff8ee0", "tree": "52c2ae868121e09df3d77fc998f233549cf8ae34", "parents": [ "33ea2dcb39ba50b0b69d1b1dc24702f084b46411" ], "author": { "name": "Dave Airlie", "email": "airlied@redhat.com", "time": "Fri Mar 16 10:34:02 2012 +0000" }, "committer": { "name": "Sumit Semwal", "email": "sumit.semwal@ti.com", "time": "Mon Mar 26 11:32:26 2012 +0530" }, "message": "dma-buf: pass flags into dma_buf_fd.\n\nWe need to pass the flags into dma_buf_fd at this point,\nso the flags end up doing the right thing for O_CLOEXEC.\n\nSigned-off-by: Dave Airlie \u003cairlied@redhat.com\u003e\nSigned-off-by: Rob Clark \u003crob@ti.com\u003e\nSigned-off-by: Sumit Semwal \u003csumit.semwal@linaro.org\u003e\n" }, { "commit": "33ea2dcb39ba50b0b69d1b1dc24702f084b46411", "tree": "ec1e1b0c223e09861341f82d2aa0b13ec841be95", "parents": [ "2ed9201bdd9a8e462d768a838fb2bc944c6887a4" ], "author": { "name": "Sumit Semwal", "email": "sumit.semwal@ti.com", "time": "Fri Jan 27 15:09:27 2012 +0530" }, "committer": { "name": "Sumit Semwal", "email": "sumit.semwal@ti.com", "time": "Mon Mar 26 11:31:58 2012 +0530" }, "message": "dma-buf: add dma_data_direction to unmap dma_buf_op\n\nSome exporters may use DMA map/unmap APIs in dma-buf ops, which require\nenum dma_data_direction for both map and unmap operations.\n\nThus, the unmap dma_buf_op also needs to have enum dma_data_direction as\na parameter.\n\nReported-by: Tomasz Stanislawski \u003ct.stanislaws@samsung.com\u003e\nSigned-off-by: Sumit Semwal \u003csumit.semwal@ti.com\u003e\nReviewed-by: Daniel Vetter \u003cdaniel.vetter@ffwll.ch\u003e\nSigned-off-by: Sumit Semwal \u003csumit.semwal@linaro.org\u003e\n" }, { "commit": "5375764f9408b8ef1fb8d6cd1ed0efd97dce4824", "tree": "05a88e01997b609418de73d53e9f6fbf06add73e", "parents": [ "c16fa4f2ad19908a47c63d8fa436a1178438c7e7" ], "author": { "name": "Laurent Pinchart", "email": "laurent.pinchart@ideasonboard.com", "time": "Thu Jan 26 12:27:22 2012 +0100" }, "committer": { "name": "Sumit Semwal", "email": "sumit.semwal@ti.com", "time": "Mon Mar 19 15:42:14 2012 +0530" }, "message": "dma-buf: Constify ops argument to dma_buf_export()\n\nThis allows drivers to make the dma buf operations structure constant.\n\nSigned-off-by: Laurent Pinchart \u003claurent.pinchart@ideasonboard.com\u003e\nReviewed-by: Daniel Vetter \u003cdaniel.vetter@ffwll.ch\u003e\nSigned-off-by: Sumit Semwal \u003csumit.semwal@linaro.org\u003e\n" }, { "commit": "313162d0b83836e2f57e51b9b8650fb4b9c396ea", "tree": "8f39f20515a924d392cf9a72db0918fda07a34c7", "parents": [ "51990e825431089747f8896244b5c17d3a6423f1" ], "author": { "name": "Paul Gortmaker", "email": "paul.gortmaker@windriver.com", "time": "Mon Jan 30 11:46:54 2012 -0500" }, "committer": { "name": "Paul Gortmaker", "email": "paul.gortmaker@windriver.com", "time": "Fri Mar 16 10:38:24 2012 -0400" }, "message": "device.h: audit and cleanup users in main include dir\n\nThe \u003clinux/device.h\u003e header includes a lot of stuff, and\nit in turn gets a lot of use just for the basic \"struct device\"\nwhich appears so often.\n\nClean up the users as follows:\n\n1) For those headers only needing \"struct device\" as a pointer\nin fcn args, replace the include with exactly that.\n\n2) For headers not really using anything from device.h, simply\ndelete the include altogether.\n\n3) For headers relying on getting device.h implicitly before\nbeing included themselves, now explicitly include device.h\n\n4) For files in which doing #1 or #2 uncovers an implicit\ndependency on some other header, fix by explicitly adding\nthe required header(s).\n\nAny C files that were implicitly relying on device.h to be\npresent have already been dealt with in advance.\n\nTotal removals from #1 and #2: 51. Total additions coming\nfrom #3: 9. Total other implicit dependencies from #4: 7.\n\nAs of 3.3-rc1, there were 110, so a net removal of 42 gives\nabout a 38% reduction in device.h presence in include/*\n\nSigned-off-by: Paul Gortmaker \u003cpaul.gortmaker@windriver.com\u003e\n" }, { "commit": "d15bd7ee445d0702ad801fdaece348fdb79e6581", "tree": "6f18212787e77b9615c955506adda516206ecbac", "parents": [ "805a6af8dba5dfdd35ec35dc52ec0122400b2610" ], "author": { "name": "Sumit Semwal", "email": "sumit.semwal@ti.com", "time": "Mon Dec 26 14:53:15 2011 +0530" }, "committer": { "name": "Dave Airlie", "email": "airlied@redhat.com", "time": "Fri Jan 06 10:20:21 2012 +0000" }, "message": "dma-buf: Introduce dma buffer sharing mechanism\n\nThis is the first step in defining a dma buffer sharing mechanism.\n\nA new buffer object dma_buf is added, with operations and API to allow easy\nsharing of this buffer object across devices.\n\nThe framework allows:\n- creation of a buffer object, its association with a file pointer, and\n associated allocator-defined operations on that buffer. This operation is\n called the \u0027export\u0027 operation.\n- different devices to \u0027attach\u0027 themselves to this exported buffer object, to\n facilitate backing storage negotiation, using dma_buf_attach() API.\n- the exported buffer object to be shared with the other entity by asking for\n its \u0027file-descriptor (fd)\u0027, and sharing the fd across.\n- a received fd to get the buffer object back, where it can be accessed using\n the associated exporter-defined operations.\n- the exporter and user to share the scatterlist associated with this buffer\n object using map_dma_buf and unmap_dma_buf operations.\n\nAtleast one \u0027attach()\u0027 call is required to be made prior to calling the\nmap_dma_buf() operation.\n\nCouple of building blocks in map_dma_buf() are added to ease introduction\nof sync\u0027ing across exporter and users, and late allocation by the exporter.\n\nFor this first version, this framework will work with certain conditions:\n- *ONLY* exporter will be allowed to mmap to userspace (outside of this\n framework - mmap is not a buffer object operation),\n- currently, *ONLY* users that do not need CPU access to the buffer are\n allowed.\n\nMore details are there in the documentation patch.\n\nThis is based on design suggestions from many people at the mini-summits[1],\nmost notably from Arnd Bergmann \u003carnd@arndb.de\u003e, Rob Clark \u003crob@ti.com\u003e and\nDaniel Vetter \u003cdaniel@ffwll.ch\u003e.\n\nThe implementation is inspired from proof-of-concept patch-set from\nTomasz Stanislawski \u003ct.stanislaws@samsung.com\u003e, who demonstrated buffer sharing\nbetween two v4l2 devices. [2]\n\n[1]: https://wiki.linaro.org/OfficeofCTO/MemoryManagement\n[2]: http://lwn.net/Articles/454389\n\nSigned-off-by: Sumit Semwal \u003csumit.semwal@linaro.org\u003e\nSigned-off-by: Sumit Semwal \u003csumit.semwal@ti.com\u003e\nReviewed-by: Daniel Vetter \u003cdaniel.vetter@ffwll.ch\u003e\nReviewed-by: Dave Airlie \u003cairlied@redhat.com\u003e\nReviewed-and-Tested-by: Rob Clark \u003crob.clark@linaro.org\u003e\nSigned-off-by: Dave Airlie \u003cairlied@redhat.com\u003e\n" } ] }