On Sun, Jan 18, 2026 at 05:30:36PM -0800, Andrew Morton wrote: > On Mon, 1 Dec 2025 13:18:48 -0500 Audra Mitchell <[email protected]> wrote: > > > If PAGE_SIZE is larger than 4k and if you have a system with a > > large number of CPUs, this test can require a very large amount > > of memory leading to oom-killer firing. Given the type of allocation, > > the kernel won't have anything to kill, causing the system to > > stall. Add a parameter to the test_vmalloc driver to represent the > > number of times a percpu object will be allocated. Calculate this > > in test_vmalloc.sh to be 90% of available memory or the current > > default of 35000, whichever is smaller. > > > > ... > > > > --- a/lib/test_vmalloc.c > > +++ b/lib/test_vmalloc.c > > @@ -57,6 +57,9 @@ __param(int, run_test_mask, 7, > > /* Add a new test case description here. */ > > ); > > > > +__param(int, nr_pcpu_objects, 35000, > > + "Number of pcpu objects to allocate for pcpu_alloc_test"); > > + > > /* > > * This is for synchronization of setup phase. > > */ > > @@ -292,24 +295,24 @@ pcpu_alloc_test(void) > > size_t size, align; > > int i; > > > > - pcpu = vmalloc(sizeof(void __percpu *) * 35000); > > + pcpu = vmalloc(sizeof(void __percpu *) * nr_pcpu_objects); > > Could have used vmalloc_array() here. Otherwise lgtm, thanks. > We can also reduce the default number from 35 000 to smaller one since this patch makes it possible to control it via parameter.
-- Uladzislau Rezki
