x86/head/64: Skip fixmap page fixup
Borislav Petkov
maintainerPATCH 1/1 x86/head/64: Skip fixmap page fixup
From: Borislav Petkov <[email protected]>
Single-stepping through head_64.S made me look at the fixmap page PTEs fixup loop:
So we're going through the whole level2_fixmap_pgt 4K page, looking at whether PAGE_PRESENT is set in those PTEs and add the delta between where we're compiled to run and where we actually end up running.
However, if that delta is 0 (most cases) we go through all those 512 PTEs for no reason at all. Oh well, we add 0 but that's no reason to me.
Skipping that useless fixup gives us a boot speedup of 0.004 seconds in my guest. Not a lot but considering how cheap it is, I'll take it. Here is the printk time difference:
before:
...
[ 0.000000] tsc: Marking TSC unstable due to TSCs unsynchronized [ 0.013590] Calibrating delay loop (skipped), value calculated using timer frequency.. 8027.17 BogoMIPS (lpj=16054348) [ 0.017094] pid_max: default: 32768 minimum: 301 ...
after:
...
[ 0.000000] tsc: Marking TSC unstable due to TSCs unsynchronized [ 0.009587] Calibrating delay loop (skipped), value calculated using timer frequency.. 8026.86 BogoMIPS (lpj=16053724) [ 0.013090] pid_max: default: 32768 minimum: 301 ...
For the other two changes converting naked numbers to defines:
# arch/x86/kernel/head_64.o:
text data bss dec hex filename
1124 290864 4096 296084 48494 head_64.o.before 1124 290864 4096 296084 48494 head_64.o.after
md5:
87086e202588939296f66e892414ffe2 head_64.o.before.asm 87086e202588939296f66e892414ffe2 head_64.o.after.asm
--- arch/x86/kernel/head_64.S | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-)
arch/x86/kernel/head_64.S +6 -2
@@ -142,6 +142,9 @@ startup_64: decl %ecx jnz 1b + test %rbp, %rbp + jz .Lskip_fixup + /* * Fixup the kernel text+data virtual addresses. Note that * we might write invalid pmds, when the kernel is relocated
@@ -149,9 +152,9 @@ startup_64: * beyond _end. */ leaq level2_kernel_pgt(%rip), %rdi - leaq 4096(%rdi), %r8 + leaq PAGE_SIZE(%rdi), %r8 /* See if it is a valid page table entry */ -1: testb $1, 0(%rdi) +1: testb $_PAGE_PRESENT, 0(%rdi) jz 2f addq %rbp, 0(%rdi) /* Go to the next page */
@@ -162,6 +165,7 @@ startup_64: /* Fixup phys_base */ addq %rbp, phys_base(%rip) +.Lskip_fixup: movq $(early_level4_pgt - __START_KERNEL_map), %rax jmp 1f ENTRY(secondary_startup_64)
Signature · 1 line
2.10.0
tip-bot for Borislav Petkov
merge bot[tip:x86/boot] x86/boot/64: Optimize fixmap page fixup
x86/boot/64: Optimize fixmap page fixup
Commit-ID: 6248f4567442081994ad61c63bd9870e147983e0 Gitweb: http://git.kernel.org/tip/6248f4567442081994ad61c63bd9870e147983e0 Author: Borislav Petkov <[email protected]> AuthorDate: Fri, 25 Nov 2016 12:14:48 +0100 Committer: Ingo Molnar <[email protected]> CommitDate: Mon, 28 Nov 2016 07:45:17 +0100
x86/boot/64: Optimize fixmap page fixup
Single-stepping through head_64.S made me look at the fixmap page PTEs fixup loop:
So we're going through the whole level2_fixmap_pgt 4K page, looking at whether PAGE_PRESENT is set in those PTEs and add the delta between where we're compiled to run and where we actually end up running.
However, if that delta is 0 (most cases) we go through all those 512 PTEs for no reason at all. Oh well, we add 0 but that's no reason to me.
Skipping that useless fixup gives us a boot speedup of 0.004 seconds in my guest. Not a lot but considering how cheap it is, I'll take it. Here is the printk time difference:
before:
...
[ 0.000000] tsc: Marking TSC unstable due to TSCs unsynchronized [ 0.013590] Calibrating delay loop (skipped), value calculated using timer frequency.. 8027.17 BogoMIPS (lpj=16054348) [ 0.017094] pid_max: default: 32768 minimum: 301 ...
after:
...
[ 0.000000] tsc: Marking TSC unstable due to TSCs unsynchronized [ 0.009587] Calibrating delay loop (skipped), value calculated using timer frequency.. 8026.86 BogoMIPS (lpj=16053724) [ 0.013090] pid_max: default: 32768 minimum: 301 ...
For the other two changes converting naked numbers to defines:
# arch/x86/kernel/head_64.o:
text data bss dec hex filename
1124 290864 4096 296084 48494 head_64.o.before 1124 290864 4096 296084 48494 head_64.o.after
md5:
87086e202588939296f66e892414ffe2 head_64.o.before.asm 87086e202588939296f66e892414ffe2 head_64.o.after.asm
--- arch/x86/kernel/head_64.S | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-)
arch/x86/kernel/head_64.S +6 -2
@@ -147,6 +147,9 @@ startup_64: decl %ecx jnz 1b + test %rbp, %rbp + jz .Lskip_fixup + /* * Fixup the kernel text+data virtual addresses. Note that * we might write invalid pmds, when the kernel is relocated
@@ -154,9 +157,9 @@ startup_64: * beyond _end. */ leaq level2_kernel_pgt(%rip), %rdi - leaq 4096(%rdi), %r8 + leaq PAGE_SIZE(%rdi), %r8 /* See if it is a valid page table entry */ -1: testb $1, 0(%rdi) +1: testb $_PAGE_PRESENT, 0(%rdi) jz 2f addq %rbp, 0(%rdi) /* Go to the next page */
@@ -167,6 +170,7 @@ startup_64: /* Fixup phys_base */ addq %rbp, phys_base(%rip) +.Lskip_fixup: movq $(early_level4_pgt - __START_KERNEL_map), %rax jmp 1f ENTRY(secondary_startup_64)