[PATCH 3/2] section garbage collection for i386 - BONUS TRACK
From: Denys Vlasenko
Date: Wed Sep 12 2007 - 16:46:26 EST
Hi Sam,
This patch makes i386 vmlinux linker script simpler
by minimizing number of generated sections.
For example, these sections:
21 .init.text 000204a0 78701000 00701000 00502000 2**0
CONTENTS, ALLOC, LOAD, READONLY, CODE
22 .init.data 000268db 787214a0 007214a0 005224a0 2**3
CONTENTS, ALLOC, LOAD, DATA
23 .init.setup 00000654 78747d80 00747d80 00548d80 2**2
CONTENTS, ALLOC, LOAD, DATA
24 .initcall.init 000004c4 787483d4 007483d4 005493d4 2**2
CONTENTS, ALLOC, LOAD, DATA
25 .con_initcall.init 00000008 78748898 00748898 00549898 2**2
CONTENTS, ALLOC, LOAD, DATA
26 .altinstructions 00009c23 787488a0 007488a0 005498a0 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
27 .altinstr_replacement 000027de 787524c3 007524c3 005534c3 2**0
CONTENTS, ALLOC, LOAD, READONLY, CODE
28 .exit.text 00000f31 78754ca4 00754ca4 00555ca4 2**0
CONTENTS, ALLOC, LOAD, READONLY, CODE
29 .init.ramfs 00000086 78756000 00756000 00557000 2**0
CONTENTS, ALLOC, LOAD, READONLY, DATA
code, data, data, data, rodata, code, code, data... -
doesn't look clean.
Can more sanely be grouped like this:
15 .init.text 00023baf 78701000 00701000 00502000 2**0
CONTENTS, ALLOC, LOAD, READONLY, CODE
16 .init.rodata 00009ca8 78725000 00725000 00526000 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
17 .init.data 00027408 7872eca8 0072eca8 0052fca8 2**3
CONTENTS, ALLOC, LOAD, DATA
All in all, we have 20 sections instead of 32, and it still boots okay.
Full listing:
Before:
$ objdump -h
linux-2.6.23-rc4.gc6.t/vmlinux: file format elf32-i386
Sections:
Idx Name Size VMA LMA File off Algn
0 .head.text 00000379 78200000 00200000 00001000 2**2
CONTENTS, ALLOC, LOAD, READONLY, CODE
1 .text 003646da 78200380 00200380 00001380 2**4
CONTENTS, ALLOC, LOAD, READONLY, CODE
2 __ex_table 00000ee8 78564a60 00564a60 00365a60 2**3
CONTENTS, ALLOC, LOAD, READONLY, DATA
3 .notes 00000024 78565948 00565948 00366948 2**2
CONTENTS, ALLOC, LOAD, READONLY, CODE
4 .rodata 000e85fe 78566000 00566000 00367000 2**3
CONTENTS, ALLOC, LOAD, READONLY, DATA
5 .pci_fixup 00000700 7864e600 0064e600 0044f600 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
6 __ksymtab 000057a0 7864ed00 0064ed00 0044fd00 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
7 __ksymtab_gpl 00001ac0 786544a0 006544a0 004554a0 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
8 __ksymtab_gpl_future 00000018 78655f60 00655f60 00456f60 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
9 __kcrctab 00002bd0 78655f78 00655f78 00456f78 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
10 __kcrctab_gpl 00000d60 78658b48 00658b48 00459b48 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
11 __kcrctab_gpl_future 0000000c 786598a8 006598a8 0045a8a8 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
12 __ksymtab_strings 00010d8b 786598b4 006598b4 0045a8b4 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
13 __param 00001838 7866a640 0066a640 0046b640 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
14 .data 0007fe99 7866c000 0066c000 0046d000 2**7
CONTENTS, ALLOC, LOAD, DATA
15 .data_nosave 00001000 786ec000 006ec000 004ed000 2**12
CONTENTS, ALLOC, LOAD, DATA
16 .page_aligned.data 00000800 786ed000 006ed000 004ee000 2**2
CONTENTS, ALLOC, LOAD, DATA
17 .cacheline_aligned.data 0000a000 786ed800 006ed800 004ee800 2**7
CONTENTS, ALLOC, LOAD, DATA
18 .read_mostly.data 000020d0 786f7800 006f7800 004f8800 2**7
CONTENTS, ALLOC, LOAD, DATA
19 .init_task.data 00002000 786fa000 006fa000 004fb000 2**2
CONTENTS, ALLOC, LOAD, DATA
20 .smp_locks 00004798 786fc000 006fc000 004fd000 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
21 .init.text 000204a0 78701000 00701000 00502000 2**0
CONTENTS, ALLOC, LOAD, READONLY, CODE
22 .init.data 000268db 787214a0 007214a0 005224a0 2**3
CONTENTS, ALLOC, LOAD, DATA
23 .init.setup 00000654 78747d80 00747d80 00548d80 2**2
CONTENTS, ALLOC, LOAD, DATA
24 .initcall.init 000004c4 787483d4 007483d4 005493d4 2**2
CONTENTS, ALLOC, LOAD, DATA
25 .con_initcall.init 00000008 78748898 00748898 00549898 2**2
CONTENTS, ALLOC, LOAD, DATA
26 .altinstructions 00009c23 787488a0 007488a0 005498a0 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
27 .altinstr_replacement 000027de 787524c3 007524c3 005534c3 2**0
CONTENTS, ALLOC, LOAD, READONLY, CODE
28 .exit.text 00000f31 78754ca4 00754ca4 00555ca4 2**0
CONTENTS, ALLOC, LOAD, READONLY, CODE
29 .init.ramfs 00000086 78756000 00756000 00557000 2**0
CONTENTS, ALLOC, LOAD, READONLY, DATA
30 .percpu.data 00005f38 78757000 00757000 00558000 2**12
CONTENTS, ALLOC, LOAD, DATA
31 .bss 00047000 7875d000 0075d000 0055df38 2**12
ALLOC
32 .comment 00005004 00000000 00000000 0055df38 2**0
CONTENTS, READONLY
After:
linux-2.6.23-rc4.gc7.t/vmlinux: file format elf32-i386
Sections:
Idx Name Size VMA LMA File off Algn
0 .text 00364a5a 78200000 00200000 00001000 2**4
CONTENTS, ALLOC, LOAD, READONLY, CODE
1 __ex_table 00000ee8 78564a60 00564a60 00365a60 2**3
CONTENTS, ALLOC, LOAD, READONLY, DATA
2 .notes 00000024 78565948 00565948 00366948 2**2
CONTENTS, ALLOC, LOAD, READONLY, CODE
3 .rodata 000e8602 78566000 00566000 00367000 2**3
CONTENTS, ALLOC, LOAD, READONLY, DATA
4 .pci_fixup 00000700 7864e604 0064e604 0044f604 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
5 __ksymtab 000057a0 7864ed04 0064ed04 0044fd04 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
6 __ksymtab_gpl 00001ac0 786544a4 006544a4 004554a4 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
7 __ksymtab_gpl_future 00000018 78655f64 00655f64 00456f64 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
8 __kcrctab 00002bd0 78655f7c 00655f7c 00456f7c 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
9 __kcrctab_gpl 00000d60 78658b4c 00658b4c 00459b4c 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
10 __kcrctab_gpl_future 0000000c 786598ac 006598ac 0045a8ac 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
11 __ksymtab_strings 00010d8b 786598b8 006598b8 0045a8b8 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
12 __param 00001838 7866a644 0066a644 0046b644 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
13 .data 0008f7d0 7866c000 0066c000 0046d000 2**12
CONTENTS, ALLOC, LOAD, DATA
14 .smp_locks 00004798 786fc000 006fc000 004fd000 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
15 .init.text 00023baf 78701000 00701000 00502000 2**0
CONTENTS, ALLOC, LOAD, READONLY, CODE
16 .init.rodata 00009ca8 78725000 00725000 00526000 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
17 .init.data 00027408 7872eca8 0072eca8 0052fca8 2**3
CONTENTS, ALLOC, LOAD, DATA
18 .percpu.data 00005f38 78757000 00757000 00558000 2**12
CONTENTS, ALLOC, LOAD, DATA
19 .bss 00047000 7875d000 0075d000 0055df38 2**12
ALLOC
20 .comment 00005004 00000000 00000000 0055df38 2**0
CONTENTS, READONLY
Patch is run-tested.
It's purely cleanup work, it's not bringing any noticeable
size wins (well, maybe statistically it will have a bit less
alignment padding because I placed together a few page-aligned
chunks, but that's all).
Signed-off-by: Denys Vlasenko <vda.linux@xxxxxxxxxxxxxx>
--
vda
diff -urpN linux-2.6.23-rc4.gc6/arch/i386/kernel/vmlinux.lds.S linux-2.6.23-rc4.gc7/arch/i386/kernel/vmlinux.lds.S
--- linux-2.6.23-rc4.gc6/arch/i386/kernel/vmlinux.lds.S 2007-09-12 19:51:42.000000000 +0100
+++ linux-2.6.23-rc4.gc7/arch/i386/kernel/vmlinux.lds.S 2007-09-12 21:10:38.000000000 +0100
@@ -6,7 +6,19 @@
* at run time. Absolute symbols are not relocated. If symbol value should
* change if kernel is relocated, make the symbol section relative
* by putting it inside the section definition.
+ *
+ * Try to avoid creating output sections needlessly.
+ * Group pieces with same attributes (ro/rw, code/data,
+ * persistent/discarded-after-init) together in one section.
+ *
+ * Statistically, padding is smaller if input sections
+ * within output section are sorted by alignment in descending order.
+ *
+ * If you need to ALIGN a section, putting ALIGN immediately before section
+ * can result in smaller on-disk image than when you put it immeadiately
+ * after section's opening '{'.
*/
+
#define LOAD_OFFSET __PAGE_OFFSET
#include <asm-generic/vmlinux.lds.h>
@@ -18,6 +30,7 @@
OUTPUT_FORMAT("elf32-i386", "elf32-i386", "elf32-i386")
OUTPUT_ARCH(i386)
ENTRY(phys_startup_32)
+
jiffies = jiffies_64;
PHDRS {
@@ -30,13 +43,11 @@ SECTIONS
. = LOAD_OFFSET + LOAD_PHYSICAL_ADDR;
phys_startup_32 = startup_32 - LOAD_OFFSET;
- .head.text : AT(ADDR(.head.text) - LOAD_OFFSET) {
- _text = .; /* Text and read-only data */
- KEEP(*(.head.text))
- } :text = 0x9090
-
/* read-only */
.text : AT(ADDR(.text) - LOAD_OFFSET) {
+ _text = .; /* Text and read-only data */
+ KEEP(*(.head.text))
+
TEXT_TEXT
SCHED_TEXT
LOCK_TEXT
@@ -54,6 +65,7 @@ SECTIONS
__stop___ex_table = .;
}
+ /* Hm? NOTES end up marked as CODE in vmlinux, but they are not. FIXME */
NOTES :text :note
BUG_TABLE :text
@@ -69,45 +81,39 @@ SECTIONS
RODATA
/* writeable */
- . = ALIGN(4096);
+ . = ALIGN(THREAD_SIZE); /* at least 4k */
.data : AT(ADDR(.data) - LOAD_OFFSET) { /* Data */
- DATA_DATA
- CONSTRUCTORS
- } :data
+ /* init_task */
+ *(.init_task.data)
- . = ALIGN(4096);
- .data_nosave : AT(ADDR(.data_nosave) - LOAD_OFFSET) {
+ . = ALIGN(4096);
+ *(.page_aligned.data)
+
+ . = ALIGN(4096);
__nosave_begin = .;
/* No need to KEEP */
*(.nosave.data)
. = ALIGN(4096);
__nosave_end = .;
- }
- . = ALIGN(4096);
- .page_aligned.data : AT(ADDR(.page_aligned.data) - LOAD_OFFSET) {
- *(.page_aligned.data)
+ /* The IDT has to be page-aligned because of F0 0F bug */
*(.idt.data)
- }
- . = ALIGN(32);
- .cacheline_aligned.data : AT(ADDR(.cacheline_aligned.data) - LOAD_OFFSET) {
+ DATA_DATA
+ CONSTRUCTORS
+
+ . = ALIGN(32);
*(.cacheline_aligned.data)
- }
- /* rarely changed data like cpu maps */
- . = ALIGN(32);
- .read_mostly.data : AT(ADDR(.read_mostly.data) - LOAD_OFFSET) {
+ /* rarely changed data like cpu maps */
+ . = ALIGN(32);
*(.read_mostly.data)
- _edata = .; /* End of data section */
- }
- . = ALIGN(THREAD_SIZE); /* init_task */
- .init_task.data : AT(ADDR(.init_task.data) - LOAD_OFFSET) {
- *(.init_task.data)
- }
+ _edata = .;
+ } :data
+
+ /* Might get freed after init */
- /* might get freed after init */
. = ALIGN(4096);
.smp_locks : AT(ADDR(.smp_locks) - LOAD_OFFSET) {
__smp_locks = .;
@@ -116,75 +122,80 @@ SECTIONS
}
/* Will be freed after init */
- /* Init code and data */
+
. = ALIGN(4096);
.init.text : AT(ADDR(.init.text) - LOAD_OFFSET) {
__init_begin = .;
_sinittext = .;
*(.init.text)
_einittext = .;
+
+ /* No need to KEEP - referenced by .altinstructions */
+ *(.altinstr_replacement)
+
+ /* .exit.text is discarded at runtime, not link time, to deal with references
+ from .altinstructions and .eh_frame */
+ *(.exit.text)
+ }
+
+ /* ro, code */
+ SECURITY_INIT
+
+#if defined(CONFIG_BLK_DEV_INITRD)
+ /* ALIGN outside section may make on-disk image smaller */
+ . = ALIGN(4096);
+#endif
+ .init.rodata : AT(ADDR(.init.rodata) - LOAD_OFFSET) {
+#if defined(CONFIG_BLK_DEV_INITRD)
+ __initramfs_start = .;
+ KEEP(*(.init.ramfs))
+ __initramfs_end = .;
+#endif
+ . = ALIGN(4);
+ __alt_instructions = .;
+ KEEP(*(.altinstructions))
+ __alt_instructions_end = .;
+
+ . = ALIGN(4);
+ __parainstructions = .;
+ KEEP(*(.parainstructions))
+ __parainstructions_end = .;
}
- .init.data : AT(ADDR(.init.data) - LOAD_OFFSET) { *(.init.data) }
- . = ALIGN(16);
- .init.setup : AT(ADDR(.init.setup) - LOAD_OFFSET) {
+
+ .init.data : AT(ADDR(.init.data) - LOAD_OFFSET) {
+ *(.init.data)
+
+ . = ALIGN(16);
__setup_start = .;
/* __setup_param() macro produces these,
* obsolete_checksetup() processes them */
KEEP(*(.init.setup))
__setup_end = .;
- }
- .initcall.init : AT(ADDR(.initcall.init) - LOAD_OFFSET) {
+
__initcall_start = .;
INITCALLS
__initcall_end = .;
- }
- .con_initcall.init : AT(ADDR(.con_initcall.init) - LOAD_OFFSET) {
+
__con_initcall_start = .;
KEEP(*(.con_initcall.init))
__con_initcall_end = .;
- }
- SECURITY_INIT
- . = ALIGN(4);
- .altinstructions : AT(ADDR(.altinstructions) - LOAD_OFFSET) {
- __alt_instructions = .;
- KEEP(*(.altinstructions))
- __alt_instructions_end = .;
- }
- .altinstr_replacement : AT(ADDR(.altinstr_replacement) - LOAD_OFFSET) {
- /* No need to KEEP - referenced by .altinstructions */
- *(.altinstr_replacement)
- }
- . = ALIGN(4);
- .parainstructions : AT(ADDR(.parainstructions) - LOAD_OFFSET) {
- __parainstructions = .;
- KEEP(*(.parainstructions))
- __parainstructions_end = .;
- }
- /* .exit.text is discarded at runtime, not link time, to deal with references
- from .altinstructions and .eh_frame */
- .exit.text : AT(ADDR(.exit.text) - LOAD_OFFSET) { *(.exit.text) }
- .exit.data : AT(ADDR(.exit.data) - LOAD_OFFSET) { *(.exit.data) }
-#if defined(CONFIG_BLK_DEV_INITRD)
- . = ALIGN(4096);
- .init.ramfs : AT(ADDR(.init.ramfs) - LOAD_OFFSET) {
- __initramfs_start = .;
- KEEP(*(.init.ramfs))
- __initramfs_end = .;
+ /* TODO: can .exit.data be discarded at link time? */
+ *(.exit.data)
}
-#endif
. = ALIGN(4096);
.percpu.data : AT(ADDR(.percpu.data) - LOAD_OFFSET) {
__per_cpu_start = .;
- *(.percpu.data)
+ *(SORT_BY_ALIGNMENT(.percpu.data))
+ /* Dont move in front of .percpu.data, .percpu.data contains page-aligned stuff */
*(.percpu.shared_aligned.data)
__per_cpu_end = .;
}
- . = ALIGN(4096);
/* End of "freed after init" portion */
+ . = ALIGN(4096);
.bss : AT(ADDR(.bss) - LOAD_OFFSET) {
__init_end = .;
__bss_start = .; /* BSS */