* [pidhash] [0/4] patch series vs. 2.5.50-mm1 @ 2002-12-05 7:48 wli 2002-12-05 7:48 ` [pidhash] [1/4] Make __do_SAK() use for_each_task_pid() wli 2002-12-05 8:50 ` [pidhash] [0/4] patch series vs. 2.5.50-mm1 William Lee Irwin III 0 siblings, 2 replies; 6+ messages in thread From: wli @ 2002-12-05 7:48 UTC (permalink / raw) To: linux-kernel; +Cc: mingo This is a series of patches against 2.5.50-mm1. They primarily have to do with pidhashing mechanisms, converting some less-important users to the new API and eliminating some tasklist iterations not covered by the original pidhash patches. These should apply to 2.5.50-CURRENT; 2.5.50-mm1 includes bk snapshots. mm1-2.5.50-2 [1/4] Make __do_SAK() use for_each_task_pid(). The tty invariant involving ->tty being strictly associated with ->session has long been established. This converts __do_SAK() to use the pidhashing instead of iteration over the global tasklist. mm1-2.5.50-5 [2/4] cap_set_pg() uses do_each_thread() to search for tasks in a given pgrp; convert it to for_each_task_pid() wli-2.5.50-3 [3/5] eliminate proc_fill_super() tasklist iteration by means of introducing a nr_processes() per-cpu counter. wli-2.5.50-6 [4/4] make get_pid_list() scan the PID bitmap instead of the tasklist. Tested on 16x/16GB NUMA-Q with up to 16384 distinct processes. Bill ^ permalink raw reply [flat|nested] 6+ messages in thread
* [pidhash] [1/4] Make __do_SAK() use for_each_task_pid(). 2002-12-05 7:48 [pidhash] [0/4] patch series vs. 2.5.50-mm1 wli @ 2002-12-05 7:48 ` wli 2002-12-05 7:48 ` [pidhash] [2/4] make cap_set_pg() " wli 2002-12-05 8:50 ` [pidhash] [0/4] patch series vs. 2.5.50-mm1 William Lee Irwin III 1 sibling, 1 reply; 6+ messages in thread From: wli @ 2002-12-05 7:48 UTC (permalink / raw) To: linux-kernel; +Cc: mingo Kill off the for_each_process() in __do_SAK(): it really just wants to do a hangup using SIGKILL. tty_io.c | 7 ++++--- 1 files changed, 4 insertions(+), 3 deletions(-) diff -urpN mm1-2.5.50-1/drivers/char/tty_io.c mm1-2.5.50-2/drivers/char/tty_io.c --- mm1-2.5.50-1/drivers/char/tty_io.c 2002-12-04 12:49:46.000000000 -0800 +++ mm1-2.5.50-2/drivers/char/tty_io.c 2002-12-04 13:06:20.000000000 -0800 @@ -1838,6 +1838,8 @@ static void __do_SAK(void *arg) #else struct tty_struct *tty = arg; struct task_struct *p; + struct list_head *l; + struct pid *pid; int session; int i; struct file *filp; @@ -1850,9 +1852,8 @@ static void __do_SAK(void *arg) if (tty->driver.flush_buffer) tty->driver.flush_buffer(tty); read_lock(&tasklist_lock); - for_each_process(p) { - if ((p->tty == tty) || - ((session > 0) && (p->session == session))) { + for_each_task_pid(session, PIDTYPE_SID, p, l, pid) { + if (p->tty == tty || session > 0) { printk(KERN_NOTICE "SAK: killed process %d" " (%s): p->session==tty->session\n", p->pid, p->comm); ^ permalink raw reply [flat|nested] 6+ messages in thread
* [pidhash] [2/4] make cap_set_pg() use for_each_task_pid() 2002-12-05 7:48 ` [pidhash] [1/4] Make __do_SAK() use for_each_task_pid() wli @ 2002-12-05 7:48 ` wli 2002-12-05 7:48 ` [pidhash] [3/4] eliminate proc_fill_super() tasklist iteration wli 0 siblings, 1 reply; 6+ messages in thread From: wli @ 2002-12-05 7:48 UTC (permalink / raw) To: linux-kernel; +Cc: mingo capset_set_pg() hunts for a task with the right pgrp. It neither wants nor needs to examine all tasks. capability.c | 16 +++++++++------- 1 files changed, 9 insertions(+), 7 deletions(-) diff -urpN mm1-2.5.50-4/kernel/capability.c mm1-2.5.50-5/kernel/capability.c --- mm1-2.5.50-4/kernel/capability.c 2002-11-27 14:35:50.000000000 -0800 +++ mm1-2.5.50-5/kernel/capability.c 2002-12-04 14:55:10.000000000 -0800 @@ -84,13 +84,15 @@ static inline void cap_set_pg(int pgrp, kernel_cap_t *inheritable, kernel_cap_t *permitted) { - task_t *g, *target; - - do_each_thread(g, target) { - if (target->pgrp != pgrp) - continue; - security_capset_set(target, effective, inheritable, permitted); - } while_each_thread(g, target); + task_t *g, *target; + struct list_head *l; + struct pid *pid; + + for_each_task_pid(pgrp, PIDTYPE_PGID, g, l, pid) { + target = g; + while_each_thread(g, target) + security_capset_set(target, effective, inheritable, permitted); + } } /* ^ permalink raw reply [flat|nested] 6+ messages in thread
* [pidhash] [3/4] eliminate proc_fill_super() tasklist iteration 2002-12-05 7:48 ` [pidhash] [2/4] make cap_set_pg() " wli @ 2002-12-05 7:48 ` wli 2002-12-05 7:48 ` [pidhash] [5/4] make get_pid_list() scan the PID bitmap wli 0 siblings, 1 reply; 6+ messages in thread From: wli @ 2002-12-05 7:48 UTC (permalink / raw) To: linux-kernel; +Cc: mingo Introduce a per-cpu counter for the number of processes on the system, and use that to set ->i_nlink in proc_fill_super() in constant time. fs/proc/inode.c | 6 +----- include/linux/sched.h | 3 +++ kernel/exit.c | 2 ++ kernel/fork.c | 14 ++++++++++++++ 4 files changed, 20 insertions(+), 5 deletions(-) diff -urpN mm1-2.5.50-2/fs/proc/inode.c wli-2.5.50-3/fs/proc/inode.c --- mm1-2.5.50-2/fs/proc/inode.c 2002-11-27 14:36:19.000000000 -0800 +++ wli-2.5.50-3/fs/proc/inode.c 2002-12-04 18:40:59.000000000 -0800 @@ -234,11 +234,7 @@ int proc_fill_super(struct super_block * /* * Fixup the root inode's nlink value */ - read_lock(&tasklist_lock); - for_each_process(p) - if (p->pid) - root_inode->i_nlink++; - read_unlock(&tasklist_lock); + root_inode->i_nlink += nr_processes(); s->s_root = d_alloc_root(root_inode); if (!s->s_root) goto out_no_root; diff -urpN mm1-2.5.50-2/include/linux/sched.h wli-2.5.50-3/include/linux/sched.h --- mm1-2.5.50-2/include/linux/sched.h 2002-12-04 12:49:47.000000000 -0800 +++ wli-2.5.50-3/include/linux/sched.h 2002-12-04 18:42:39.000000000 -0800 @@ -27,6 +27,7 @@ #include <linux/compiler.h> #include <linux/completion.h> #include <linux/pid.h> +#include <linux/percpu.h> struct exec_domain; @@ -87,6 +88,8 @@ extern unsigned long avenrun[]; /* Load extern int nr_threads; extern int last_pid; +DECLARE_PER_CPU(unsigned long, process_counts); +extern int nr_processes(void); extern unsigned long nr_running(void); extern unsigned long nr_uninterruptible(void); extern unsigned long nr_iowait(void); diff -urpN mm1-2.5.50-2/kernel/exit.c wli-2.5.50-3/kernel/exit.c --- mm1-2.5.50-2/kernel/exit.c 2002-12-04 12:49:47.000000000 -0800 +++ wli-2.5.50-3/kernel/exit.c 2002-12-04 18:57:08.000000000 -0800 @@ -41,6 +41,8 @@ static struct dentry * __unhash_process( if (thread_group_leader(p)) { detach_pid(p, PIDTYPE_PGID); detach_pid(p, PIDTYPE_SID); + if (p->pid) + per_cpu(process_counts, smp_processor_id())--; } REMOVE_LINKS(p); diff -urpN mm1-2.5.50-2/kernel/fork.c wli-2.5.50-3/kernel/fork.c --- mm1-2.5.50-2/kernel/fork.c 2002-12-04 12:49:47.000000000 -0800 +++ wli-2.5.50-3/kernel/fork.c 2002-12-04 18:56:55.000000000 -0800 @@ -48,6 +48,8 @@ int nr_threads; int max_threads; unsigned long total_forks; /* Handle normal Linux uptimes. */ +DEFINE_PER_CPU(unsigned long, process_counts); + rwlock_t tasklist_lock __cacheline_aligned = RW_LOCK_UNLOCKED; /* outer */ /* @@ -57,6 +59,16 @@ rwlock_t tasklist_lock __cacheline_align */ static task_t *task_cache[NR_CPUS] __cacheline_aligned; +int nr_processes(void) +{ + int cpu; + unsigned long total = 0; + + for (cpu = 0; cpu < NR_CPUS; ++cpu) + total += per_cpu(process_counts, cpu); + return (int)total; +} + void __put_task_struct(struct task_struct *tsk) { if (tsk != current) { @@ -934,6 +946,8 @@ static struct task_struct *copy_process( attach_pid(p, PIDTYPE_TGID, p->tgid); attach_pid(p, PIDTYPE_PGID, p->pgrp); attach_pid(p, PIDTYPE_SID, p->session); + if (p->pid) + per_cpu(process_counts, smp_processor_id())++; } else link_pid(p, p->pids + PIDTYPE_TGID, &p->group_leader->pids[PIDTYPE_TGID].pid); ^ permalink raw reply [flat|nested] 6+ messages in thread
* [pidhash] [5/4] make get_pid_list() scan the PID bitmap 2002-12-05 7:48 ` [pidhash] [3/4] eliminate proc_fill_super() tasklist iteration wli @ 2002-12-05 7:48 ` wli 0 siblings, 0 replies; 6+ messages in thread From: wli @ 2002-12-05 7:48 UTC (permalink / raw) To: linux-kernel; +Cc: mingo Make get_pid_list() walk the PID bitmap instead of the tasklist. This is not quite as good as it could be, because the PID search structures are not appropriately arranged for range queries. fs/proc/base.c | 22 +--------------------- kernel/pid.c | 53 +++++++++++++++++++++++++++++++++++++++++++++++++++++ 2 files changed, 54 insertions(+), 21 deletions(-) diff -urpN wli-2.5.50-5/fs/proc/base.c wli-2.5.50-6/fs/proc/base.c --- wli-2.5.50-5/fs/proc/base.c 2002-11-27 14:36:06.000000000 -0800 +++ wli-2.5.50-6/fs/proc/base.c 2002-12-04 19:15:00.000000000 -0800 @@ -1153,28 +1153,8 @@ out: * tasklist lock while doing this, and we must release it before * we actually do the filldir itself, so we use a temp buffer.. */ -static int get_pid_list(int index, unsigned int *pids) -{ - struct task_struct *p; - int nr_pids = 0; - - index--; - read_lock(&tasklist_lock); - for_each_process(p) { - int pid = p->pid; - if (!pid) - continue; - if (--index >= 0) - continue; - pids[nr_pids] = pid; - nr_pids++; - if (nr_pids >= PROC_MAXPIDS) - break; - } - read_unlock(&tasklist_lock); - return nr_pids; -} +int get_pid_list(int ordinal_of_pid, unsigned int *pid_buf); int proc_pid_readdir(struct file * filp, void * dirent, filldir_t filldir) { unsigned int pid_array[PROC_MAXPIDS]; diff -urpN wli-2.5.50-5/kernel/pid.c wli-2.5.50-6/kernel/pid.c --- wli-2.5.50-5/kernel/pid.c 2002-11-27 14:36:03.000000000 -0800 +++ wli-2.5.50-6/kernel/pid.c 2002-12-04 19:15:59.000000000 -0800 @@ -258,6 +258,59 @@ void switch_exec_pids(task_t *leader, ta attach_pid(leader, PIDTYPE_SID, leader->session); } +int get_pid_list(int ordinal, unsigned int *pid_buf) +{ + pidmap_t *map = pidmap_array; + int count = 0, k = 0, pos = 0; + + BUG_ON(ordinal <= 0); + ordinal--; + + while (pos + BITS_PER_PAGE < ordinal && map - pidmap_array < PIDMAP_ENTRIES) { + pos += BITS_PER_PAGE - atomic_read(&map->nr_free); + ++map; + } + + if (map - pidmap_array >= PIDMAP_ENTRIES || !map->page) + return 0; + + while (pos < ordinal && k < BITS_PER_PAGE) { + task_t *task; + if (test_bit(k, map->page) + && (map - pidmap_array)*BITS_PER_PAGE + k) { + int pid = (map-pidmap_array)*BITS_PER_PAGE + k; + read_lock(&tasklist_lock); + task = find_task_by_pid(pid); + if (task && thread_group_leader(task)) + ++pos; + read_unlock(&tasklist_lock); + } + ++k; + } + + while (map - pidmap_array < PIDMAP_ENTRIES && map->page && count < 20) { + while (k < BITS_PER_PAGE && count < 20) { + task_t *task; + + if (test_bit(k, map->page) + && (map - pidmap_array)*BITS_PER_PAGE + k) { + int pid = (map-pidmap_array)*BITS_PER_PAGE + k; + read_lock(&tasklist_lock); + task = find_task_by_pid(pid); + if (task && thread_group_leader(task)) { + pid_buf[count] = task->pid; + ++count; + } + read_unlock(&tasklist_lock); + } + ++k; + } + ++map; + k = 0; + } + return count; +} + void __init pidhash_init(void) { int i, j; ^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [pidhash] [0/4] patch series vs. 2.5.50-mm1 2002-12-05 7:48 [pidhash] [0/4] patch series vs. 2.5.50-mm1 wli 2002-12-05 7:48 ` [pidhash] [1/4] Make __do_SAK() use for_each_task_pid() wli @ 2002-12-05 8:50 ` William Lee Irwin III 1 sibling, 0 replies; 6+ messages in thread From: William Lee Irwin III @ 2002-12-05 8:50 UTC (permalink / raw) To: linux-kernel; +Cc: mingo On Wed, Dec 04, 2002 at 11:48:27PM -0800, wli@holomorphy.com wrote: [...] Sorry about the misnumbering; there were about half a dozen other pidhashing related things that didn't make the cut (omitted largely due to theoretical issues or not having enough testing bandwidth for tonight's run) along with a fair amount of other non-pidhash-related fat included in the original numbering scheme before the patch series got renumbered and refined down to bits relevant to single maintainers. Some were also excluded due to mutual interference. Bill ^ permalink raw reply [flat|nested] 6+ messages in thread
end of thread, other threads:[~2002-12-05 8:44 UTC | newest] Thread overview: 6+ messages (download: mbox.gz / follow: Atom feed) -- links below jump to the message on this page -- 2002-12-05 7:48 [pidhash] [0/4] patch series vs. 2.5.50-mm1 wli 2002-12-05 7:48 ` [pidhash] [1/4] Make __do_SAK() use for_each_task_pid() wli 2002-12-05 7:48 ` [pidhash] [2/4] make cap_set_pg() " wli 2002-12-05 7:48 ` [pidhash] [3/4] eliminate proc_fill_super() tasklist iteration wli 2002-12-05 7:48 ` [pidhash] [5/4] make get_pid_list() scan the PID bitmap wli 2002-12-05 8:50 ` [pidhash] [0/4] patch series vs. 2.5.50-mm1 William Lee Irwin III
This is a public inbox, see mirroring instructions for how to clone and mirror all data and code used for this inbox
all inboxes | Powered by JetHome®