On 2024-10-04 09:56:56 [+0000], Arnd Bergmann wrote:
> The problem is the additional size overhead from local_lock in
> struct kmem_cache_cpu. Avoid this by preallocating a larger area.
The worst case would be enabling additionally MEMCG so NR_KMALLOC_TYPES
increases by one. And then we have:
PERCPU_DYNAMIC_EARLY_SIZE < NR_KMALLOC_TYPES * KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
There is more to it than just RANDOM_KMALLOC_CACHES and PREEMPT_RT.
There is additionally CONFIG_LOCKDEP which increases the size of
local_lock_t further. Plus CONFIG_LOCK_STAT. The last one a kind of bad
in terms of required pad area. Then we have CONFIG_PAGE_SIZE_64KB set
which is the culprit. But 16K_PAGES also fail in this full blown case.
PERCPU_DYNAMIC_EARLY_SIZE < NR_KMALLOC_TYPES * KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
4KiB 20 << 12 < 19 * (12 + 1) * 288
80KiB < 69.46875
16KiB 20 << 12 < 19 * (14 + 1) * 288
80KiB < 80.15625
64KiB 20 << 12 < 19 * (16 + 1) * 288
80KiB < 90.84375
128KiB 20 << 12 < 19 * (17 + 1) * 288
80KiB < 96.1875
Just disabling CONFIG_LOCK_STAT the 16KiB PAGE_SIZE case works again
(75.703125), the 64KiB still fails (85.796875).
> There is a good chance that there is a better way to address this, this
> version was the first I came up with and I verified that it fixes all of
> the broken configs.
How bad is it, to have PERCPU_DYNAMIC_SIZE_SHIFT unconditionally set to
13? If it is bad could we restrict it with LOCKDEP and PAGE_SIZE > 4KiB?
So maybe something like this:
diff --git a/include/linux/percpu.h b/include/linux/percpu.h
index b6321fc491598..52b5ea663b9f0 100644
--- a/include/linux/percpu.h
+++ b/include/linux/percpu.h
@@ -41,7 +41,11 @@
PCPU_MIN_ALLOC_SHIFT)
#ifdef CONFIG_RANDOM_KMALLOC_CACHES
-#define PERCPU_DYNAMIC_SIZE_SHIFT 12
+# if defined(CONFIG_LOCKDEP) && !defined(CONFIG_PAGE_SIZE_4KB)
+# define PERCPU_DYNAMIC_SIZE_SHIFT 13
+# else
+# define PERCPU_DYNAMIC_SIZE_SHIFT 12
+#endif /* LOCKDEP and PAGE_SIZE > 4KiB */
#else
#define PERCPU_DYNAMIC_SIZE_SHIFT 10
#endif
Sebastian
