Re: Packable data structures found by pahole

From: Paul E. McKenney
Date: Fri Feb 20 2009 - 12:14:29 EST

On Wed, Feb 11, 2009 at 11:50:32AM -0200, Arnaldo Carvalho de Melo wrote:
> Em Wed, Feb 11, 2009 at 01:22:36PM +0100, Ingo Molnar escreveu:
> > Is there anything packable in core kernel structures like task struct?
>
> I still haven't added an heuristic to avoid reporting members with
> explicit __alignment attributes, as these are not encoded in DWARF. I'll
> work on that soon, but till then we can use this as an starting point.
>
> struct name, current size, --reorganized size, savings
>
> $ pahole --packable ../build/blkftrace/vmlinux | sort -k4 -nr
> vc_data 432 176 256
> is this exported to userspace?
>
> rcu_ctrlblk 128 64 64
> has ____cacheline_internodealigned_in_smp

For 900 bytes of memory-footprint reduction on uniprocessor builds,
I suggest rcutiny.c (http://lkml.org/lkml/2009/2/3/333).

That said, I don't know of any functional problems that would result
from packing rcu_ctrlblk.

Thanx, Paul

> timex 208 152 56
> syscall interface
>
> hh_cache 128 72 56
> has ____cacheline_aligned_in_smp
>
> cpu_workqueue_struct 128 72 56
> is ____cacheline_aligned
>
> rchan_buf 256 216 40
> is ____cacheline_aligned
>
> tty_struct 1328 1296 32
> this one doesn't have any annotation, looks ripe for --reorganize
>
> task_struct 6008 5976 32
>
> Printing this one here, the rest of the possibly packable data
> structures are after it:
>
> struct task_struct {
> volatile long int state; /* 0 8 */
> void * stack; /* 8 8 */
> atomic_t usage; /* 16 4 */
> unsigned int flags; /* 20 4 */
> unsigned int ptrace; /* 24 4 */
> int lock_depth; /* 28 4 */
> int prio; /* 32 4 */
> int static_prio; /* 36 4 */
> int normal_prio; /* 40 4 */
> unsigned int rt_priority; /* 44 4 */
> const struct sched_class * sched_class; /* 48 8 */
> struct sched_entity se; /* 56 368 */
> /* --- cacheline 6 boundary (384 bytes) was 40 bytes ago --- */
> struct sched_rt_entity rt; /* 424 64 */
> /* --- cacheline 7 boundary (448 bytes) was 40 bytes ago --- */
> unsigned char fpu_counter; /* 488 1 */
> s8 oomkilladj; /* 489 1 */
>
> /* XXX 2 bytes hole, try to pack */
>
> unsigned int btrace_seq; /* 492 4 */
> unsigned int policy; /* 496 4 */
>
> /* XXX 4 bytes hole, try to pack */
>
> cpumask_t cpus_allowed; /* 504 8 */
> /* --- cacheline 8 boundary (512 bytes) --- */
> struct sched_info sched_info; /* 512 40 */
>
> /* XXX last struct has 4 bytes of padding */
>
> struct list_head tasks; /* 552 16 */
> struct plist_node pushable_tasks; /* 568 40 */
> /* --- cacheline 9 boundary (576 bytes) was 32 bytes ago --- */
> struct mm_struct * mm; /* 608 8 */
> struct mm_struct * active_mm; /* 616 8 */
> struct linux_binfmt * binfmt; /* 624 8 */
> int exit_state; /* 632 4 */
> int exit_code; /* 636 4 */
> /* --- cacheline 10 boundary (640 bytes) --- */
> int exit_signal; /* 640 4 */
> int pdeath_signal; /* 644 4 */
> unsigned int personality; /* 648 4 */
> unsigned int did_exec:1; /* 652:31 4 */
>
> /* XXX 31 bits hole, try to pack */
>
> pid_t pid; /* 656 4 */
> pid_t tgid; /* 660 4 */
> long unsigned int stack_canary; /* 664 8 */
> struct task_struct * real_parent; /* 672 8 */
> struct task_struct * parent; /* 680 8 */
> struct list_head children; /* 688 16 */
> /* --- cacheline 11 boundary (704 bytes) --- */
> struct list_head sibling; /* 704 16 */
> struct task_struct * group_leader; /* 720 8 */
> struct list_head ptraced; /* 728 16 */
> struct list_head ptrace_entry; /* 744 16 */
> struct bts_tracer * bts; /* 760 8 */
> /* --- cacheline 12 boundary (768 bytes) --- */
> void * bts_buffer; /* 768 8 */
> size_t bts_size; /* 776 8 */
> struct pid_link pids[3]; /* 784 72 */
> /* --- cacheline 13 boundary (832 bytes) was 24 bytes ago --- */
> struct list_head thread_group; /* 856 16 */
> struct completion * vfork_done; /* 872 8 */
> int * set_child_tid; /* 880 8 */
> int * clear_child_tid; /* 888 8 */
> /* --- cacheline 14 boundary (896 bytes) --- */
> cputime_t utime; /* 896 8 */
> cputime_t stime; /* 904 8 */
> cputime_t utimescaled; /* 912 8 */
> cputime_t stimescaled; /* 920 8 */
> cputime_t gtime; /* 928 8 */
> cputime_t prev_utime; /* 936 8 */
> cputime_t prev_stime; /* 944 8 */
> long unsigned int nvcsw; /* 952 8 */
> /* --- cacheline 15 boundary (960 bytes) --- */
> long unsigned int nivcsw; /* 960 8 */
> struct timespec start_time; /* 968 16 */
> struct timespec real_start_time; /* 984 16 */
> long unsigned int min_flt; /* 1000 8 */
> long unsigned int maj_flt; /* 1008 8 */
> struct task_cputime cputime_expires; /* 1016 24 */
> /* --- cacheline 16 boundary (1024 bytes) was 16 bytes ago --- */
> struct list_head cpu_timers[3]; /* 1040 48 */
> /* --- cacheline 17 boundary (1088 bytes) --- */
> const struct cred * real_cred; /* 1088 8 */
> const struct cred * cred; /* 1096 8 */
> struct mutex cred_exec_mutex; /* 1104 32 */
> char comm[16]; /* 1136 16 */
> /* --- cacheline 18 boundary (1152 bytes) --- */
> int link_count; /* 1152 4 */
> int total_link_count; /* 1156 4 */
> struct sysv_sem sysvsem; /* 1160 8 */
> long unsigned int last_switch_count; /* 1168 8 */
> struct thread_struct thread; /* 1176 208 */
>
> /* XXX last struct has 4 bytes of padding */
>
> /* --- cacheline 21 boundary (1344 bytes) was 40 bytes ago --- */
> struct fs_struct * fs; /* 1384 8 */
> struct files_struct * files; /* 1392 8 */
> struct nsproxy * nsproxy; /* 1400 8 */
> /* --- cacheline 22 boundary (1408 bytes) --- */
> struct signal_struct * signal; /* 1408 8 */
> struct sighand_struct * sighand; /* 1416 8 */
> sigset_t blocked; /* 1424 8 */
> sigset_t real_blocked; /* 1432 8 */
> sigset_t saved_sigmask; /* 1440 8 */
> struct sigpending pending; /* 1448 24 */
> /* --- cacheline 23 boundary (1472 bytes) --- */
> long unsigned int sas_ss_sp; /* 1472 8 */
> size_t sas_ss_size; /* 1480 8 */
> int (*notifier)(void *); /* 1488 8 */
> void * notifier_data; /* 1496 8 */
> sigset_t * notifier_mask; /* 1504 8 */
> struct audit_context * audit_context; /* 1512 8 */
> uid_t loginuid; /* 1520 4 */
> unsigned int sessionid; /* 1524 4 */
> seccomp_t seccomp; /* 1528 4 */
> u32 parent_exec_id; /* 1532 4 */
> /* --- cacheline 24 boundary (1536 bytes) --- */
> u32 self_exec_id; /* 1536 4 */
> spinlock_t alloc_lock; /* 1540 4 */
> spinlock_t pi_lock; /* 1544 4 */
>
> /* XXX 4 bytes hole, try to pack */
>
> struct plist_head pi_waiters; /* 1552 32 */
> struct rt_mutex_waiter * pi_blocked_on; /* 1584 8 */
> unsigned int irq_events; /* 1592 4 */
> int hardirqs_enabled; /* 1596 4 */
> /* --- cacheline 25 boundary (1600 bytes) --- */
> long unsigned int hardirq_enable_ip; /* 1600 8 */
> unsigned int hardirq_enable_event; /* 1608 4 */
>
> /* XXX 4 bytes hole, try to pack */
>
> long unsigned int hardirq_disable_ip; /* 1616 8 */
> unsigned int hardirq_disable_event; /* 1624 4 */
> int softirqs_enabled; /* 1628 4 */
> long unsigned int softirq_disable_ip; /* 1632 8 */
> unsigned int softirq_disable_event; /* 1640 4 */
>
> /* XXX 4 bytes hole, try to pack */
>
> long unsigned int softirq_enable_ip; /* 1648 8 */
> unsigned int softirq_enable_event; /* 1656 4 */
> int hardirq_context; /* 1660 4 */
> /* --- cacheline 26 boundary (1664 bytes) --- */
> int softirq_context; /* 1664 4 */
>
> /* XXX 4 bytes hole, try to pack */
>
> void * journal_info; /* 1672 8 */
> struct bio * bio_list; /* 1680 8 */
> struct bio * * bio_tail; /* 1688 8 */
> struct reclaim_state * reclaim_state; /* 1696 8 */
> struct backing_dev_info * backing_dev_info; /* 1704 8 */
> struct io_context * io_context; /* 1712 8 */
> long unsigned int ptrace_message; /* 1720 8 */
> /* --- cacheline 27 boundary (1728 bytes) --- */
> siginfo_t * last_siginfo; /* 1728 8 */
> struct task_io_accounting ioac; /* 1736 56 */
> /* --- cacheline 28 boundary (1792 bytes) --- */
> u64 acct_rss_mem1; /* 1792 8 */
> u64 acct_vm_mem1; /* 1800 8 */
> cputime_t acct_timexpd; /* 1808 8 */
> nodemask_t mems_allowed; /* 1816 64 */
> /* --- cacheline 29 boundary (1856 bytes) was 24 bytes ago --- */
> int cpuset_mems_generation; /* 1880 4 */
> int cpuset_mem_spread_rotor; /* 1884 4 */
> struct css_set * cgroups; /* 1888 8 */
> struct list_head cg_list; /* 1896 16 */
> struct robust_list_head * robust_list; /* 1912 8 */
> /* --- cacheline 30 boundary (1920 bytes) --- */
> struct compat_robust_list_head * compat_robust_list; /* 1920 8 */
> struct list_head pi_state_list; /* 1928 16 */
> struct futex_pi_state * pi_state_cache; /* 1944 8 */
> struct perf_counter_context perf_counter_ctx; /* 1952 80 */
> /* --- cacheline 31 boundary (1984 bytes) was 48 bytes ago --- */
> struct mempolicy * mempolicy; /* 2032 8 */
> short int il_next; /* 2040 2 */
>
> /* XXX 2 bytes hole, try to pack */
>
> atomic_t fs_excl; /* 2044 4 */
> /* --- cacheline 32 boundary (2048 bytes) --- */
> struct rcu_head rcu; /* 2048 16 */
> struct pipe_inode_info * splice_pipe; /* 2064 8 */
> struct task_delay_info * delays; /* 2072 8 */
> struct prop_local_single dirties; /* 2080 24 */
> int latency_record_count; /* 2104 4 */
>
> /* XXX 4 bytes hole, try to pack */
>
> /* --- cacheline 33 boundary (2112 bytes) --- */
> struct latency_record latency_record[32]; /* 2112 3840 */
> /* --- cacheline 93 boundary (5952 bytes) --- */
> long unsigned int timer_slack_ns; /* 5952 8 */
> long unsigned int default_timer_slack_ns; /* 5960 8 */
> struct list_head * scm_work_list; /* 5968 8 */
> int curr_ret_stack; /* 5976 4 */
>
> /* XXX 4 bytes hole, try to pack */
>
> struct ftrace_ret_stack * ret_stack; /* 5984 8 */
> atomic_t trace_overrun; /* 5992 4 */
> atomic_t tracing_graph_pause; /* 5996 4 */
> long unsigned int trace; /* 6000 8 */
>
> /* size: 6008, cachelines: 94, members: 148 */
> /* sum members: 5976, holes: 9, sum holes: 32 */
> /* bit holes: 1, sum bit holes: 31 bits */
> /* paddings: 2, sum paddings: 8 */
> /* last cacheline: 56 bytes */
> }; /* definitions: 742 */
>
> If we ask pahole to reorganize it it would do these steps:
>
> $ pahole -C task_struct --reorganize --show_reorg_steps kernel/sched.o|grep ^\/
> /* Demoting bitfield ('did_exec' ... 'did_exec') from 'unsigned int' to
> * 'unsigned char' */
>
> /* Moving bitfield('did_exec' ... 'did_exec') from after 'personality'
> * to after 'oomkilladj' */
>
> /* Moving 'personality' from after 'pdeath_signal' to after 'policy' */
>
> /* Moving 'hardirq_enable_event' from after 'hardirq_enable_ip' to after
> * 'pi_lock' */
>
> /* Moving 'softirq_context' from after 'hardirq_context' to after
> * 'softirq_disable_event' */
>
> /* Moving 'curr_ret_stack' from after 'scm_work_list' to after
> * 'latency_record_count' */
>
> And the new stats would be:
>
> /* size: 5976, cachelines: 94, members: 148 */
> /* sum members: 5973, holes: 2, sum holes: 3 */
> /* bit holes: 1, sum bit holes: 7 bits */
> /* paddings: 2, sum paddings: 8 */
> /* last cacheline: 24 bytes */
> }; /* saved 32 bytes! */
>
> It would still have these holes/paddings:
>
> <SNIP>
>
> s8 oomkilladj; /* 489 1 */
> unsigned char did_exec:1; /* 490: 7 1 */
>
> /* XXX 7 bits hole, try to pack */
> /* XXX 1 byte hole, try to pack */
>
> unsigned int btrace_seq; /* 492 4 */
>
> <SNIP>
>
> /* --- cacheline 8 boundary (512 bytes) --- */
> struct sched_info sched_info; /* 512 40 */
>
> /* XXX last struct has 4 bytes of padding */
>
> struct list_head tasks; /* 552 16 */
>
> <SNIP>
>
> long unsigned int last_switch_count; /* 1160 8 */
> struct thread_struct thread; /* 1168 208 */
>
> /* XXX last struct has 4 bytes of padding */
>
> /* --- cacheline 21 boundary (1344 bytes) was 32 bytes ago --- */
>
> <SNIP>
>
> /* --- cacheline 31 boundary (1984 bytes) was 24 bytes ago --- */
> struct mempolicy * mempolicy; /* 2008 8 */
> short int il_next; /* 2016 2 */
>
> /* XXX 2 bytes hole, try to pack */
>
> atomic_t fs_excl; /* 2020 4 */
>
> <SNIP>
>
> I put the pahole vmlinux output on
> http://fedorapeople.org/~acme/pahole/vmlinux.pahole.c
>
> zone 1536 1512 24
> super_block 768 744 24
> Scsi_Host 1384 1360 24
> scsi_device 1312 1288 24
> rq 2456 2432 24
> request_queue 2272 2248 24
> net_device 1600 1576 24
> cp_private 1344 1320 24
> clocksource 192 168 24
> ata_port 11184 11160 24
> taskstats 328 312 16
> sock 544 528 16
> rtl8139_private 448 432 16
> rtentry 120 104 16
> pci_dev 1624 1608 16
> packet_sock 760 744 16
> mtd_info 352 336 16
> mousedev 784 768 16
> module 512 496 16
> mm_struct 808 792 16
> loop_device 400 384 16
> journal_s 568 552 16
> gendisk 720 704 16
> floppy_drive_params 128 112 16
> files_struct 704 688 16
> dio 856 840 16
> block_device 248 232 16
> audit_context 1968 1952 16
> xfrm_state 632 624 8
> writeback_control 64 56 8
> vt_spawn_console 24 16 8
> vmap_block_queue 48 40 8
> vfsmount 224 216 8
> user_struct 96 88 8
> unix_skb_parms 32 24 8
> unity_map_entry 48 40 8
> uart_port 200 192 8
> tty_ldisc_ops 144 136 8
> tty_bufhead 152 144 8
> tty_audit_buf 72 64 8
> transaction_s 168 160 8
> tick_sched 248 240 8
> thread_struct 208 200 8
> sysfs_dirent 80 72 8
> sk_buff 192 184 8
> signal_struct 944 936 8
> sighand_struct 2088 2080 8
> sg_io_hdr 88 80 8
> serio 704 696 8
> semid_ds 88 80 8
> scsi_target 616 608 8
> scsi_pointer 64 56 8
> scm_cookie 40 32 8
> rt_rq 1760 1752 8
> rtc_device 744 736 8
> root_domain 1704 1696 8
> ring_buffer_per_cpu 112 104 8
> ring_buffer 72 64 8
> request 368 360 8
> rchan 376 368 8
> psmouse_protocol 48 40 8
> proto 336 328 8
> protection_domain 48 40 8
> prop_local_percpu 64 56 8
> proc_dir_entry 160 152 8
> power_supply 112 104 8
> pnp_card 632 624 8
> platform_device 520 512 8
> pid_namespace 2112 2104 8
> pglist_data 80576 80568 8
> perf_counter_context 80 72 8
> perf_counter 4408 4400 8
> pci_root_info 40 32 8
> old_serial_port 40 32 8
> net 592 584 8
> neigh_table 472 464 8
> neighbour 240 232 8
> ncp_mount_data_v4 80 72 8
> mtd_oob_ops 64 56 8
> msghdr 56 48 8
> mnt_namespace 64 56 8
> ml_device 888 880 8
> loop_info 168 160 8
> kprobe 128 120 8
> kparam_array 48 40 8
> kmem_cache 4352 4344 8
> irq_desc 192 184 8
> ip_sf_list 40 32 8
> ip_mc_list 168 160 8
> ipc_namespace 296 288 8
> input_dev 2352 2344 8
> inode 560 552 8
> inet_timewait_death_row 568 560 8
> inet6_ifaddr 184 176 8
> in_device 376 368 8
> i387_soft_struct 136 128 8
> hrtimer_cpu_base 160 152 8
> hid_field 112 104 8
> hid_device 7144 7136 8
> gen_estimator 112 104 8
> fs_quota_stat 80 72 8
> floppy_write_errors 40 32 8
> floppy_fdc_state 40 32 8
> flock 32 24 8
> fb_info 712 704 8
> ext3_sb_info 440 432 8
> ext3_inode_info 768 760 8
> dquot 232 224 8
> cpuinfo_x86 192 184 8
> clock_event_device 128 120 8
> cdrom_generic_command 64 56 8
> cache_detail 224 216 8
> bsg_device 160 152 8
> bsg_class_device 48 40 8
> blk_user_trace_setup 72 64 8
> blk_trace 96 88 8
> blkcipher_walk 112 104 8
> audit_watch 72 64 8
> atkbd 1488 1480 8
> ata_queued_cmd 224 216 8
> ata_host 72 64 8
> ata_device 1168 1160 8
> as_io_context 104 96 8
> amd_iommu 120 112 8
> agp_kern_info 80 72 8
> agp_bridge_data 200 192 8
> acpi_thermal 1472 1464 8
> acpi_pscope_state 56 48 8
> acpi_prt_entry 48 40 8
> acpi_processor_power 2112 2104 8
> acpi_processor_performance 112 104 8
> acpi_processor_cx 136 128 8
> acpi_blacklist_item 56 48 8
> tty_port 136 132 4
> scsi_host_cmd_pool 48 44 4
> rtentry32 84 80 4
> msqid_ds 104 100 4
> inotify_watch 64 60 4
> in6_rtmsg 80 76 4
> fown_struct 32 28 4
> fib_iter_state 56 52 4
> entropy_store 56 52 4
> compat_ncp_mount_data 56 52 4
> compat_loop_info 140 136 4
> compat_floppy_fdc_state 32 28 4
> compat_floppy_drive_params 88 84 4
> agp_allocate 24 20 4
> acpi_parse_obj_named 72 68 4
> fb_monspecs 144 141 3
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@xxxxxxxxxxxxxxx
> More majordomo info at http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at http://www.tux.org/lkml/
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/