virtmgr: Pass hugepages to crosvm
Now crosvm applies block alignment ("arm64: Align RAM region with the
block size") when mmaping the RAM region for a VM, it is possible to
back this memory with THP (transparent-hugepages). This is controlled by
either the VM config option "hugepages" or the "bin/vm" option
"--hugepages"
Enabling --hugepages makes crosvm "madvise(MADV_HUGEPAGE)" that region.
This will have at the moment no effect on Android as the default value
for /sys/kernel/mm/transparent_hugepages/shmem_enabled is [never].
However, it'll enable OEMs to turn on the feature by tweaking the latter
knob.
Using THP brings a significant performance improvement by reducing the
number of fault necessary when using a VM (/512 for a 4K pages Arm
machine) and reducing the pressure on the TLB (for both stage-1 and
stage-2). However, finding huge pages might be a difficult task when the
system has been running for a long time and the memory is quite
fragmented. khugepaged helps promoting pages to huge-pages but running
it has a cost and might delay the memory allocation depending on the
chosen defrag policy.
Bug: 278011447
Change-Id: I954f93df4f08ad015958d36d115d9f9e0c3547b5
diff --git a/vm/src/main.rs b/vm/src/main.rs
index b60f2db..d6ee3a5 100644
--- a/vm/src/main.rs
+++ b/vm/src/main.rs
@@ -53,6 +53,13 @@
/// Run VM in protected mode.
#[arg(short, long)]
protected: bool,
+
+ /// Ask the kernel for transparent huge-pages (THP). This is only a hint and
+ /// the kernel will allocate THP-backed memory only if globally enabled by
+ /// the system and if any can be found. See
+ /// https://docs.kernel.org/admin-guide/mm/transhuge.html
+ #[arg(short, long)]
+ hugepages: bool,
}
#[derive(Args, Default)]