Size of 2.6.20 task_struct on x86_64 machines
This past week I was playing around with that pahole tool
(http://oops.ghostprotocols.net:81/acme/dwarves/) and looking at the size of various struct in the kernel. I was surprised by the size of the task_struct on x86_64, approaching 4K. I looked through the fields in task_struct and found that a number of them were declared as "unsigned long" rather than "unsigned int" despite them appearing okay as 32-bit sized fields. On x86_64 "unsigned long" ends up being 8 bytes in size and forces 8 byte alignment. Is there a reason there a reason they are "unsigned long"?
The patch below drops the size of the struct from 3808 bytes (60 64-byte cachelines) to 3760 bytes (59 64-byte cachelines). A couple other fields in the task struct take a signficant amount of space:
struct thread_struct thread; 688 struct held_lock held_locks[30]; 1680
CONFIG_LOCKDEP is turned on in the .config
-Will
linux/sched.h renamed from linux/sched.h.compress +6 -6
@@ -802,8 +802,8 @@volatile long state;/* -1 unrunnable, 0 runnable, >0 stopped */ struct thread_info *thread_info;atomic_t usage; -unsigned long flags;/* per process flags, defined below */ -unsigned long ptrace; +unsigned int flags;/* per process flags, defined below */ +unsigned int ptrace;int lock_depth;/* BKL lock depth */
@@ -826,7 +826,7 @@unsigned long long sched_time;/* sched_clock time spent running */ enum sleep_type sleep_type; -unsigned long policy; +unsigned int policy;cpumask_t cpus_allowed;unsigned int time_slice, first_time_slice;
@@ -846,11 +846,11 @@/* task state */ struct linux_binfmt *binfmt; -long exit_state; +int exit_state;int exit_code, exit_signal;int pdeath_signal;/* The signal sent when the parent dies */ /* ??? */ -unsigned long personality; +unsigned int personality; unsigned did_exec:1;pid_t pid;pid_t tgid;
@@ -882,7 +882,7 @@int __user *set_child_tid;/* CLONE_CHILD_SETTID */ int __user *clear_child_tid;/* CLONE_CHILD_CLEARTID */ -unsigned long rt_priority; +unsigned int rt_priority;cputime_t utime, stime;unsigned long nvcsw, nivcsw;/* context switch counts */ struct timespec start_time;
From: William Cohen <[email protected]>
Date: Thu, 08 Feb 2007 11:14:13 -0500
Quoted 9 lines
This past week I was playing around with that pahole tool
(http://oops.ghostprotocols.net:81/acme/dwarves/) and looking at the size of various struct in the kernel. I was surprised by the size of thetask_structonx86_64, approaching 4K. I looked through the fields intask_structand found that a number of them were declared as "unsigned long" rather than "unsigned int" despite them appearing okay as 32-bit sized fields. Onx86_64"unsigned long" ends up being 8 bytes in size and forces 8 byte alignment. Is there a reason there a reason they are "unsigned long"?
I think at one point we used the atomic bit operations to operate on things like tsk->flags, and those interfaces require unsigned long as the type.
That doesn't appear to be the case any longer, so at a minimum your tsk->flags conversion to unsigned int should be ok.
Quoted 13 lines · On Thu, 08 Feb 2007 12:19:45 -0800 (PST) David Miller <[email protected]>…
From: William Cohen <[email protected]>
Date: Thu, 08 Feb 2007 11:14:13 -0500Quoted 9 lines
This past week I was playing around with that pahole tool
(http://oops.ghostprotocols.net:81/acme/dwarves/) and looking at the size of various struct in the kernel. I was surprised by the size of thetask_structonx86_64, approaching 4K. I looked through the fields intask_structand found that a number of them were declared as "unsigned long" rather than "unsigned int" despite them appearing okay as 32-bit sized fields. Onx86_64"unsigned long" ends up being 8 bytes in size and forces 8 byte alignment. Is there a reason there a reason they are "unsigned long"?
Quoted 6 lines
I think at one point we used the atomic bit operations to operate on things like
tsk->flags, and those interfaces require unsigned long as the type.That doesn't appear to be the case any longer, so at a minimum your
tsk->flagsconversion to unsigned int should be ok.
Yeah, afacit everything in there is OK and happily all the converted-to-32-bit quantities happen to be contiguous with other 32-bit quantities.
Most architectures' bitops functions take unsigned long * so if anyone is using bitops on these things we should get to hear about it.
Quoted 10 lines · William Cohen
This past week I was playing around with that pahole tool
(http://oops.ghostprotocols.net:81/acme/dwarves/) and looking at the size of various struct in the kernel. I was surprised by the size of thetask_structonx86_64, approaching 4K. I looked through the fields intask_structand found that a number of them were declared as "unsigned long" rather than "unsigned int" despite them appearing okay as 32-bit sized fields. Onx86_64"unsigned long" ends up being 8 bytes in size and forces 8 byte alignment. Is there a reason there a reason they are "unsigned long"?
Quoted 8 lines
The patch below drops the size of the struct from 3808 bytes (60 64-byte cachelines) to 3760 bytes (59 64-byte cachelines). A couple other fields in the task struct take a signficant amount of space:
struct thread_struct thread; 688 struct held_lock held_locks[30]; 1680
CONFIG_LOCKDEPis turned on in the .config
I sent this .. http://lkml.org/lkml/2007/1/2/299
last month which shrinks task struct by 480 bytes when lockdep is enabled. Ingo acked it, but then it fell on the floor.
Here it is again..
Dave
Shrink the held_lock struct by using bitfields.
This shrinks task_struct on lockdep enabled kernels by 480 bytes.
include/linux/lockdep.h +5 -5
@@ -175,11 +175,11 @@ struct held_lock {* The following field is used to detect when we cross into an * interrupt context: */ -int irq_context; -int trylock; -int read; -int check; -int hardirqs_off; + unsigned char irq_context:1; + unsigned char trylock:1; + unsigned char read:2; + unsigned char check:1; + unsigned char hardirqs_off:1; };/*
Signature · 1 line
http://www.codemonkey.org.uk
Quoted 3 lines · Dave Jones
Shrink the
held_lockstruct by using bitfields.
This shrinkstask_structon lockdep enabled kernels by 480 bytes.
Are we sure that there are no users that depend on accessing the different fields under different locks?
Having them as separate "int" fields means that they don't have any interaction, and normal cache coherency will "just work". Once they are fields in the same word in memory, updating one field automatically will do a read-write cycle on the other fields, and if _they_ are updated by interrupts or other CPU's at the same time, a write can get lost..
So I'd like this to be ack'ed by Ingo.
Ingo?
Linus
---
Quoted 27 lines
Signed-off-by Dave Jones <[email protected]>include/linux/lockdep.h +5
-5@@ -175,11 +175,11 @@ struct held_lock {* The following field is used to detect when we cross into an * interrupt context: */ -int irq_context; -int trylock; -int read; -int check; -int hardirqs_off; + unsigned char irq_context:1; + unsigned char trylock:1; + unsigned char read:2; + unsigned char check:1; + unsigned char hardirqs_off:1; };/* Signature · 2 lines
-- http://www.codemonkey.org.uk