mm/page_alloc: fix min_free_kbytes calculation regarding ZONE_MOVABLE

[jamieNguyenNVIDIA]

The current calculation of min_free_kbytes only uses ZONE_DMA and ZONE_NORMAL pages,but the ZONE_MOVABLE zone->_watermark[WMARK_MIN] will also divide part of min_free_kbytes.This will cause the min watermark of ZONE_NORMAL to be too small in the presence of ZONE_MOVEABLE.

__GFP_HIGH and PF_MEMALLOC allocations usually don't need movable zone pages, so just like ZONE_HIGHMEM, cap pages_min to a small value in __setup_per_zone_wmarks().

On my testing machine with 16GB of memory (transparent hugepage is turned off by default, and movablecore=12G is configured) The following is a comparative test data of watermark_min

	no patch	add patch

ZONE_DMA 1 8
ZONE_DMA32 151 709
ZONE_NORMAL 233 1113
ZONE_MOVABLE 1434 128
min_free_kbytes 7288 7326

Link: https://lkml.kernel.org/r/[email protected]

Reviewed-by: "Huang, Ying" [email protected]

(cherry picked from commit 416ef04 linux)

[jamieNguyenNVIDIA]

Internal bug: 4452840

Test case:

• Download and extract tests-Linux-aarch64.tar from //builds/linuxbuilds/release/display/aarch64/<GPU_DRIVER_VER>/tests-Linux-aarch64.tar
• Load UVM with tests enabled:

sudo nvidia-smi -pm 0
sudo rmmod nvidia_uvm
sudo modprobe nvidia_uvm uvm_enable_builtin_tests=1 uvm_perf_access_counter_mimc_migration_enable=1
sudo nvidia-smi -pm 1

• Run ./tests-Linux-aarch64/uvm/uvm_test/uvm_test -t oversubscription_multiprocess

It will fail with the following output without this change:

$ ./uvm_test -t oversubscription_multiprocess
^^^^ RUNNING oversubscription_multiprocess
pid: 200601; started at: Fri Jan 12 02:44:53 UTC 2024
repro cmd: oversubscription --printKmsg -n -t multiprocess --seed 2863276720
(thread 281465001818176 [t0]) At oversubscription.cu:671:
   Value of : cudaMalloc(&dummy, mapGpuEvictAllocSize[gpu.cudaId])
   Which is : 2 (cudaErrorMemoryAllocation)
   Should be: cudaSuccess
   Which is : 0 (cudaSuccess) free_mem = 9557 MB
Test state is:
 Thread 281465001818176 [t0]:
   oversubscription.cu:668: alloc_size = 5768 MB
   oversubscription.cu:657: gpu = { GH100GL-B  GPU-77083a13-fb9f-2097-eac0-81ee8034773d CUDA ID  0 }
   oversubscription.cu:588: child = 0
Waiting 100ms for CUDA API calls in other threads to finish.
Failed to run cleanup action: wait for CUDA API calls.
Forcing termination due to failed cleanup actions.
(thread 281465001818176 [t0]) At utils.cpp:1310:
   expected !spawnedChildDied spawned child 200602 died unexpectedly

It should pass with this patch applied.

[jamieNguyenNVIDIA]

This was broken in 6.3+ kernels that contain the following patch:

[1ebbb21811b76c3b932959787f37985af36f62fa] mm/page_alloc: explicitly define how __GFP_HIGH non-blocking allocations accesses reserves