Hi Nick,
Le 17/03/2021 à 07:24, Nicholas Piggin a écrit :
This reduces TLB misses by nearly 30x on a `git diff` workload on a
2-node POWER9 (59,800 -> 2,100) and reduces CPU cycles by 0.54%, due
to vfs hashes being allocated with 2MB pages.
Cc: linuxppc-dev@xxxxxxxxxxxxxxxx
Acked-by: Michael Ellerman <mpe@xxxxxxxxxxxxxx>
Signed-off-by: Nicholas Piggin <npiggin@xxxxxxxxx>
---
.../admin-guide/kernel-parameters.txt | 2 ++
arch/powerpc/Kconfig | 1 +
arch/powerpc/kernel/module.c | 22 +++++++++++++++----
3 files changed, 21 insertions(+), 4 deletions(-)
--- a/arch/powerpc/kernel/module.c
+++ b/arch/powerpc/kernel/module.c
@@ -8,6 +8,7 @@
#include <linux/moduleloader.h>
#include <linux/err.h>
#include <linux/vmalloc.h>
+#include <linux/mm.h>
#include <linux/bug.h>
#include <asm/module.h>
#include <linux/uaccess.h>
@@ -87,13 +88,26 @@ int module_finalize(const Elf_Ehdr *hdr,
return 0;
}
-#ifdef MODULES_VADDR
void *module_alloc(unsigned long size)
{
+ unsigned long start = VMALLOC_START;
+ unsigned long end = VMALLOC_END;
+
+#ifdef MODULES_VADDR
BUILD_BUG_ON(TASK_SIZE > MODULES_VADDR);
+ start = MODULES_VADDR;
+ end = MODULES_END;
+#endif
+
+ /*
+ * Don't do huge page allocations for modules yet until more testing
+ * is done. STRICT_MODULE_RWX may require extra work to support this
+ * too.
+ */
- return __vmalloc_node_range(size, 1, MODULES_VADDR, MODULES_END, GFP_KERNEL,
- PAGE_KERNEL_EXEC, VM_FLUSH_RESET_PERMS, NUMA_NO_NODE,
I think you should add the following in <asm/pgtable.h>
#ifndef MODULES_VADDR
#define MODULES_VADDR VMALLOC_START
#define MODULES_END VMALLOC_END
#endif
And leave module_alloc() as is (just removing the enclosing #ifdef MODULES_VADDR and adding the
VM_NO_HUGE_VMAP flag)
This would minimise the conflits with the changes I did in powerpc/next reported by Stephen R.
+ return __vmalloc_node_range(size, 1, start, end, GFP_KERNEL,
+ PAGE_KERNEL_EXEC,
+ VM_NO_HUGE_VMAP | VM_FLUSH_RESET_PERMS,
+ NUMA_NO_NODE,
__builtin_return_address(0));
}
-#endif
