On 5 Aug 2021, at 15:16, Christian König wrote:

Am 05.08.21 um 21:02 schrieb Zi Yan:

quoted

From: Zi Yan <ziy@nvidia.com>

This prepares for the upcoming changes to make MAX_ORDER a boot time
parameter instead of compilation time constant. All static arrays with
MAX_ORDER size are converted to pointers and their memory is allocated
at runtime.

Well in general I strongly suggest to not use the patter kmalloc(sizeof(some struct) * MAX_ORDER,...) instead use kmalloc_array, kcalloc etc..

Then when a array is embedded at the end of a structure you can use a trailing array and the struct_size() macro to determine the allocation size.

Sure. Will fix it.

Additional to that separating the patch into changes for TTM to make the maximum allocation order independent from MAX_ORDER would be rather good to have I think.

Can you elaborate a little bit more on “make the maximum allocation order independent from MAX_ORDER”? From what I understand of ttm_pool_alloc(), it tries to get num_pages pages by allocating as large pages as possible, starting from MAX_ORDER. What is the rationale behind this algorithm? Why not just call alloc_page(order=0) num_pages times? Is it mean to reduce the number of calls to alloc_page? The allocated pages do not need to get as high as MAX_ORDER, is that the case? If yes, I probably can keep ttm pool as static arrays with length of MIN_MAX_ORDER, which I introduce in Patch 14 as the lower bound of boot time parameter MAX_ORDER. Let me know your thoughts.

Thanks.

—
Best Regards,
Yan, Zi