pid namespace .text overhead

FYI,

I just did a compile test on a 2.6.23-rc4-mm1 kernel with and without the following patches on a x86_64 defconfig (I also had to remove CONFIG_IPV6 for some compile reason) :

+ pid-namespaces-rework-forget_original_parent.patch
+ pid-namespaces-move-exit_task_namespaces.patch
+ pid-namespaces-introduce-ms_kernmount-flag.patch
+ pid-namespaces-prepare-proc_flust_task-to-flush-entries-from-multiple-proc-trees.patch
+ pid-namespaces-introduce-struct-upid.patch
+ pid-namespaces-add-support-for-pid-namespaces-hierarchy.patch
+ pid-namespaces-make-alloc_pid-free_pid-and-put_pid-work-with-struct-upid.patch
+ pid-namespaces-helpers-to-obtain-pid-numbers.patch
+ pid-namespaces-helpers-to-find-the-task-by-its-numerical-ids.patch
+ pid-namespaces-helpers-to-find-the-task-by-its-numerical-ids-fix.patch
+ pid-namespaces-move-alloc_pid-lower-in-copy_process.patch
+ pid-namespaces-make-proc-have-multiple-superblocks-one-for-each-namespace.patch
+ pid-namespaces-miscelaneous-preparations-for-pid-namespaces.patch
+ pid-namespaces-allow-cloning-of-new-namespace.patch
+ pid-namespaces-allow-cloning-of-new-namespace-fix-check-for-return-value-of-create_pid_namespace.patch
+ pid-namespaces-make-proc_flush_task-actually-from-entries-from-multiple-namespaces.patch
+ pid-namespaces-initialize-the-namespaces-proc_mnt.patch
+ pid-namespaces-create-a-slab-cache-for-struct-pid_namespace.patch
+ pid-namespaces-allow-signalling-container-init.patch
+ pid-namespaces-destroy-pid-namespace-on-inits-death.patch
+ pid-namespaces-changes-to-show-virtual-ids-to-user.patch
+ pid-namespaces-changes-to-show-virtual-ids-to-user-fix-the-return-value-of-sys_set_tid_address.patch
+ pid-namespaces-changes-to-show-virtual-ids-to-user-use-find_task_by_pid_ns-in-places-that-operate-with-virtual.patch
+ pid-namespaces-changes-to-show-virtual-ids-to-user-use-find_task_by_pid_ns-in-places-that-operate-with-virtual-fix.patch
+ pid-namespaces-changes-to-show-virtual-ids-to-user-use-find_task_by_pid_ns-in-places-that-operate-with-virtual-fix-2.patch
+ pid-namespaces-changes-to-show-virtual-ids-to-user-use-find_task_by_pid_ns-in-places-that-operate-with-virtual-fix-3.patch
+ pid-namespaces-changes-to-show-virtual-ids-to-user-sys_getsid-sys_getpgid-return-wrong-id-for-task-from-another.patch
+ pid-namespaces-changes-to-show-virtual-ids-to-user-fix-the-sys_setpgrp-to-work-between-namespaces.patch
+ pid-namespaces-changes-to-show-virtual-ids-to-user-fix.patch
+ pid-namespaces-remove-the-struct-pid-unneeded-fields.patch
+ isolate-some-explicit-usage-of-task-tgid.patch
+ isolate-some-explicit-usage-of-task-tgid-fix.patch
+ isolate-some-explicit-usage-of-task-tgid-fix-fix.patch

I got a less than *6k* difference in .text. Here are the detailed size results.

size without:
text data bss dec hex filename
5067718 892082 725544 6685344 6602a0 vmlinux

size with:
text data bss dec hex filename
5073314 892210 725544 6691068 6618fc vmlinux

section without with
.text 3383305 3388569

__ex_table                  16704     16704
__bug_table                 41136     41208
.rodata                   1276129   1276469
.pci_fixup                   4032      4032
__ksymtab                   43984     43984
__ksymtab_gpl               13024     13056
__ksymtab_gpl_future           48        48
__ksymtab_strings           83388     83420
__param                      7800      7800
.data                      482928    483056
.data.cacheline_aligned    187008    187008
.data.read_mostly           41600     41600
.vsyscall_0                   227       227
.vsyscall_fn                   54        54
.vsyscall_gtod_data            80        80
.vsyscall_1                    52        52
.vsyscall_2                    91        91
.vgetcpu_mode                   4         4
.jiffies                        8         8
.vsyscall_3                     8         8
.data.init_task              8192      8192
.data.page_aligned           4096      4096
.smp_locks                  33584     33624
.init.text                 153713    153529
.init.data                 137714    137714
.init.setup                  3168      3168
.initcall.init               2200      2200

.con_initcall.init 16 16
.altinstructions 1243 1243
.altinstr_replacement 244 244
.exit.text 4739 4739
.note 24 24
.vdso 4056 4056
.init.ramfs 133 133
.data.percpu 25064 25064
.data_nosave 4 4
.bss 725544 725544
.comment 44352 44352
Total 6729696 6735420

kernel/pid.o .text increases a lot (more than a 1K) but some inline would also need to be checked (that i didn't do yet)

Cheers,

C.

Cedric Le Goater [clg-NmTC/[email protected]] wrote:

| FYI,
| 
| I just did a compile test on a 2.6.23-rc4-mm1 kernel with and without 
| the following patches on a x86_64 defconfig  (I also had to remove 
| CONFIG_IPV6 for some compile reason) :

Thats a good point.

We have been a bit liberal with "inline" given that lot of the code was going to affect the "fast-path" clone().

Should we shoot for a time/space trade-off or can we make some inlines conditional (i.e inline when not configured for say embedded linux) ?

Suka

are are some figures :

with without

   text    data |   text    data | filename
   1186       4 |   1099       4 | kernel/capability.o
  10390       0 |   9941       0 | kernel/exit.o
  10611     140 |  10434     140 | kernel/fork.o
  10765      72 |  10518      72 | kernel/futex.o
    957      56 |    939      56 | kernel/nsproxy.o
   3446    2292 |   2350    2228 | kernel/pid.o
  13930      45 |  13641      45 | kernel/signal.o
  10177     544 |   9819     544 | kernel/sys.o
   5083      56 |   4809      56 | fs/proc/array.o
  17097     184 |  16748     184 | fs/proc/base.o
  51471    1532 |  50474    1532 | fs/proc/proc.o

I'm sure the embedded guys will appreciate :)

Thanks !

C.

Can we make them switch somehow on -Os? That's what the embedded guys use.

-- Dave