On Thu, Jul 04, 2024 at 01:04:56PM +0530, Sughosh Ganu wrote:
The aim of this patch series is to fix the current state of incoherence between modules when it comes to memory usage. The primary issue that this series is trying to fix is that the EFI memory module which is responsible for allocating and freeing memory, does not have any visibility of the memory that is being used by the LMB module. This is further complicated by the fact that the LMB allocations are caller specific -- the LMB memory map is not global nor persistent. This means that the memory "allocated" by the LMB module might be relevant only for a given function. Hence one of the requirements for making the memory usage visible across modules is to make LMB allocations persistent and global, and then have means to communicate the use of memory across modules.
The first set of patches in this series work on making the LMB memory map persistent and global. This is being done keeping in mind the usage of LMB memory by platforms where the same memory region can be used to load multiple different images. What is not allowed is to overwrite memory that has been allocated by the other module, currently the EFI memory module. This is being achieved by introducing a new flag, LMB_NOOVERWRITE, which represents memory which cannot be re-requested once allocated.
A review comment on the earlier version was to do away with the static arrays for the LMB lists of free and used memory. This version uses the alloced list data structure for the LMB lists.
The second set of patches are making changes to the EFI memory module to make use of the LMB functions to allocate and free memory. A *_flags() version of LMB API's has been introduced for the same. The earlier version was using notification mechanism from both LMB and EFI modules to maintain memory coherence. This version makes use of the LMB API functions for the memory allocations. This is based on review comments of EFI maintainers.
Please re-work so that the series is bisectable. For example xilinx_zynqmp_r5 fails that check currently. And I found that looking in to why it grows by ~1500 bytes overall. This likely is due to CONFIG_EFI_LOADER=n and so while the case where EFI_LOADER is enabled tends to be a size win (reduction) or wash we need to look at the CONFIG_EFI_LOADER=n case more. The alist code will be a little growth and that's fine enough. But realloc and do_bdinfo are the two big growths at the top, in this case.