elf: init pt_regs pointer later
authorAlexey Dobriyan <adobriyan@gmail.com>
Tue, 14 May 2019 22:43:54 +0000 (15:43 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 15 May 2019 02:52:50 +0000 (19:52 -0700)
Get "current_pt_regs" pointer right before usage.

Space savings on x86_64:

add/remove: 0/0 grow/shrink: 0/1 up/down: 0/-180 (-180)
Function                           old     new   delta
load_elf_binary                   5806    5626    -180 !!!

Looks like the compiler doesn't know that "current_pt_regs" is stable
pointer (because it doesn't know ->stack isn't) even though it knows
that "current" is stable pointer.  So it saves it in the very beginning
and then tries to carry it through a lot of code.

Here is what happens here:

load_elf_binary()
...
mov rax,QWORD PTR gs:0x14c00
mov r13,QWORD PTR [rax+0x18] r13 = current->stack
call kmem_cache_alloc # first kmalloc

[980 bytes later!]

# let's spill that sucker because we need a register
# for "load_bias" calculations at
#
# if (interpreter) {
# load_bias = ELF_ET_DYN_BASE;
# if (current->flags & PF_RANDOMIZE)
# load_bias += arch_mmap_rnd();
# elf_flags |= elf_fixed;
# }
mov QWORD PTR [rsp+0x68],r13

If this is not _the_ root cause it is still eeeeh.

After the patch things become much simpler:

mov rax, QWORD PTR gs:0x14c00 # current
mov rdx, QWORD PTR [rax+0x18] # current->stack
movq [rdx+0x3fb8], 0 # fill pt_regs
...
call finalize_exec

Link: http://lkml.kernel.org/r/20190419200343.GA19788@avx2
Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Tested-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
fs/binfmt_elf.c

index 31f264bd5126b941951769afc87476f5b18087c0..1a66b6215c803e9975f65a33e11d0face6a5cc72 100644 (file)
@@ -704,12 +704,12 @@ static int load_elf_binary(struct linux_binprm *bprm)
        unsigned long start_code, end_code, start_data, end_data;
        unsigned long reloc_func_desc __maybe_unused = 0;
        int executable_stack = EXSTACK_DEFAULT;
-       struct pt_regs *regs = current_pt_regs();
        struct {
                struct elfhdr elf_ex;
                struct elfhdr interp_elf_ex;
        } *loc;
        struct arch_elf_state arch_state = INIT_ARCH_ELF_STATE;
+       struct pt_regs *regs;
 
        loc = kmalloc(sizeof(*loc), GFP_KERNEL);
        if (!loc) {
@@ -1150,6 +1150,7 @@ out_free_interp:
                                MAP_FIXED | MAP_PRIVATE, 0);
        }
 
+       regs = current_pt_regs();
 #ifdef ELF_PLAT_INIT
        /*
         * The ABI may specify that certain registers be set up in special