diff options
Diffstat (limited to 'doc/develop/driver-model')
| -rw-r--r-- | doc/develop/driver-model/livetree.rst | 60 |
1 files changed, 56 insertions, 4 deletions
diff --git a/doc/develop/driver-model/livetree.rst b/doc/develop/driver-model/livetree.rst index 9f654f3b894..faf3eb5b5f0 100644 --- a/doc/develop/driver-model/livetree.rst +++ b/doc/develop/driver-model/livetree.rst @@ -211,9 +211,61 @@ using it in new code. Modifying the livetree ---------------------- -This is not currently supported. Once implemented it should provide a much -more efficient implementation for modification of the device tree than using -the flat tree. +This is supported in a limited way, with ofnode_write_prop() and related +functions. + +The unflattening algorithm results in a single block of memory being +allocated for the whole tree. When writing new properties, these are +allocated new memory outside that block. When the block is freed, the +allocated properties remain. This can result in a memory leak. + +The solution to this leak would be to add a flag for properties (and nodes when +support is provided for adding those) that indicates that they should be +freed. Then the tree can be scanned for these 'separately allocated' nodes and +properties before freeing the memory block. + +The ofnode_write\_...() functions also support writing to the flat tree. Care +should be taken however, since this can change the position of node names and +properties in the flat tree, thus affecting the live tree. Generally this does +not matter, since when we fire up the live tree we don't ever use the flat tree +again. But in the case of tests, this can cause a problem. + +The sandbox tests typically run with OF_LIVE enabled but with the actual live +tree either present or absent. This is to make sure that the flat tree functions +work correctly even with OF_LIVE is enabled. But if a test modifies the flat +device tree, then the live tree can become invalid. Any live tree tests that run +after that point will use a corrupted tree, e.g. with an incorrect property name +or worse. To deal with this we use a flag UT_TESTF_LIVE_OR_FLAT then ensures +that tests which write to the flat tree are not run if OF_LIVE is enabled. Only +the live tree version of the test is run, when OF_LIVE is enabled, with +sandbox_flattree running the flat tree version. + +This is of course a work-around, even if a reasonable one. One solution to this +problem would be to make a copy of the flat tree before the test and restore it +afterwards, in the same memory location, so that the live tree pointers work +again. Another would be to regenerate the live tree if a test modified the flat +tree. + +Neither of these solutions is currently implemented, since the situation that +causes the problem can only occur in sandbox tests, is somewhat esoteric and +the UT_TESTF_LIVE_OR_FLAT flag deals with it in a reasonable way. + + +Multiple livetrees +------------------ + +The livetree implementation was originally designed for use with the control +FDT. This means that the FDT fix-ups (ft_board_setup() and the like, must use +a flat tree. + +It would be helpful to use livetree for fixups, since adding a lot of nodes and +properties would involve less memory copying and be more efficient. As a step +towards this, an `oftree` type has been introduced. It is normally set to +oftree_default() but can be set to other values. Eventually this should allow +the use of FDT fixups using the ofnode interface, instead of the low-level +libfdt one. + +See dm_test_ofnode_root() for some examples. Internal implementation @@ -281,6 +333,6 @@ Live tree support was introduced in U-Boot 2017.07. There is still quite a bit of work to do to flesh this out: - tests for all access functions -- support for livetree modification +- more support for livetree modification - addition of more access functions as needed - support for livetree in SPL and before relocation (if desired) |