kernel: Map most files to use export.h instead of module.h
[linux-3.10.git] / kernel / pid.c
index c7346fb..fa5f722 100644 (file)
  * allocation scenario when all but one out of 1 million PIDs possible are
  * allocated already: the scanning of 32 list entries and at most PAGE_SIZE
  * bytes. The typical fastpath is a single successful setbit. Freeing is O(1).
+ *
+ * Pid namespaces:
+ *    (C) 2007 Pavel Emelyanov <xemul@openvz.org>, OpenVZ, SWsoft Inc.
+ *    (C) 2007 Sukadev Bhattiprolu <sukadev@us.ibm.com>, IBM
+ *     Many thanks to Oleg Nesterov for comments and help
+ *
  */
 
 #include <linux/mm.h>
-#include <linux/module.h>
+#include <linux/export.h>
 #include <linux/slab.h>
 #include <linux/init.h>
+#include <linux/rculist.h>
 #include <linux/bootmem.h>
 #include <linux/hash.h>
 #include <linux/pid_namespace.h>
 #include <linux/init_task.h>
+#include <linux/syscalls.h>
 
-#define pid_hashfn(nr) hash_long((unsigned long)nr, pidhash_shift)
+#define pid_hashfn(nr, ns)     \
+       hash_long((unsigned long)nr + (unsigned long)ns, pidhash_shift)
 static struct hlist_head *pid_hash;
-static int pidhash_shift;
+static unsigned int pidhash_shift = 4;
 struct pid init_struct_pid = INIT_STRUCT_PID;
 
 int pid_max = PID_MAX_DEFAULT;
@@ -70,11 +79,22 @@ struct pid_namespace init_pid_ns = {
        .level = 0,
        .child_reaper = &init_task,
 };
+EXPORT_SYMBOL_GPL(init_pid_ns);
 
-int is_global_init(struct task_struct *tsk)
+int is_container_init(struct task_struct *tsk)
 {
-       return tsk == init_pid_ns.child_reaper;
+       int ret = 0;
+       struct pid *pid;
+
+       rcu_read_lock();
+       pid = task_pid(tsk);
+       if (pid != NULL && pid->numbers[pid->level].nr == 1)
+               ret = 1;
+       rcu_read_unlock();
+
+       return ret;
 }
+EXPORT_SYMBOL(is_container_init);
 
 /*
  * Note: disable interrupts while the pidmap_lock is held as an
@@ -92,15 +112,53 @@ int is_global_init(struct task_struct *tsk)
 
 static  __cacheline_aligned_in_smp DEFINE_SPINLOCK(pidmap_lock);
 
-static fastcall void free_pidmap(struct pid_namespace *pid_ns, int pid)
+static void free_pidmap(struct upid *upid)
 {
-       struct pidmap *map = pid_ns->pidmap + pid / BITS_PER_PAGE;
-       int offset = pid & BITS_PER_PAGE_MASK;
+       int nr = upid->nr;
+       struct pidmap *map = upid->ns->pidmap + nr / BITS_PER_PAGE;
+       int offset = nr & BITS_PER_PAGE_MASK;
 
        clear_bit(offset, map->page);
        atomic_inc(&map->nr_free);
 }
 
+/*
+ * If we started walking pids at 'base', is 'a' seen before 'b'?
+ */
+static int pid_before(int base, int a, int b)
+{
+       /*
+        * This is the same as saying
+        *
+        * (a - base + MAXUINT) % MAXUINT < (b - base + MAXUINT) % MAXUINT
+        * and that mapping orders 'a' and 'b' with respect to 'base'.
+        */
+       return (unsigned)(a - base) < (unsigned)(b - base);
+}
+
+/*
+ * We might be racing with someone else trying to set pid_ns->last_pid.
+ * We want the winner to have the "later" value, because if the
+ * "earlier" value prevails, then a pid may get reused immediately.
+ *
+ * Since pids rollover, it is not sufficient to just pick the bigger
+ * value.  We have to consider where we started counting from.
+ *
+ * 'base' is the value of pid_ns->last_pid that we observed when
+ * we started looking for a pid.
+ *
+ * 'pid' is the pid that we eventually found.
+ */
+static void set_last_pid(struct pid_namespace *pid_ns, int base, int pid)
+{
+       int prev;
+       int last_write = base;
+       do {
+               prev = last_write;
+               last_write = cmpxchg(&pid_ns->last_pid, prev, pid);
+       } while ((prev != last_write) && (pid_before(base, last_write, pid)));
+}
+
 static int alloc_pidmap(struct pid_namespace *pid_ns)
 {
        int i, offset, max_scan, pid, last = pid_ns->last_pid;
@@ -111,7 +169,12 @@ static int alloc_pidmap(struct pid_namespace *pid_ns)
                pid = RESERVED_PIDS;
        offset = pid & BITS_PER_PAGE_MASK;
        map = &pid_ns->pidmap[pid/BITS_PER_PAGE];
-       max_scan = (pid_max + BITS_PER_PAGE - 1)/BITS_PER_PAGE - !offset;
+       /*
+        * If last_pid points into the middle of the map->page we
+        * want to scan this bitmap block twice, the second time
+        * we start with offset == 0 (or RESERVED_PIDS).
+        */
+       max_scan = DIV_ROUND_UP(pid_max, BITS_PER_PAGE) - !offset;
        for (i = 0; i <= max_scan; ++i) {
                if (unlikely(!map->page)) {
                        void *page = kzalloc(PAGE_SIZE, GFP_KERNEL);
@@ -120,11 +183,12 @@ static int alloc_pidmap(struct pid_namespace *pid_ns)
                         * installing it:
                         */
                        spin_lock_irq(&pidmap_lock);
-                       if (map->page)
-                               kfree(page);
-                       else
+                       if (!map->page) {
                                map->page = page;
+                               page = NULL;
+                       }
                        spin_unlock_irq(&pidmap_lock);
+                       kfree(page);
                        if (unlikely(!map->page))
                                break;
                }
@@ -132,20 +196,12 @@ static int alloc_pidmap(struct pid_namespace *pid_ns)
                        do {
                                if (!test_and_set_bit(offset, map->page)) {
                                        atomic_dec(&map->nr_free);
-                                       pid_ns->last_pid = pid;
+                                       set_last_pid(pid_ns, last, pid);
                                        return pid;
                                }
                                offset = find_next_offset(map, offset);
                                pid = mk_pid(pid_ns, map, offset);
-                       /*
-                        * find_next_offset() found a bit, the pid from it
-                        * is in-bounds, and if we fell back to the last
-                        * bitmap block and the final block was the same
-                        * as the starting point, pid is before last_pid.
-                        */
-                       } while (offset < BITS_PER_PAGE && pid < pid_max &&
-                                       (i != max_scan || pid < last ||
-                                           !((last+1) & BITS_PER_PAGE_MASK)));
+                       } while (offset < BITS_PER_PAGE && pid < pid_max);
                }
                if (map < &pid_ns->pidmap[(pid_max-1)/BITS_PER_PAGE]) {
                        ++map;
@@ -161,11 +217,14 @@ static int alloc_pidmap(struct pid_namespace *pid_ns)
        return -1;
 }
 
-static int next_pidmap(struct pid_namespace *pid_ns, int last)
+int next_pidmap(struct pid_namespace *pid_ns, unsigned int last)
 {
        int offset;
        struct pidmap *map, *end;
 
+       if (last >= PID_MAX_LIMIT)
+               return -1;
+
        offset = (last + 1) & BITS_PER_PAGE_MASK;
        map = &pid_ns->pidmap[(last + 1)/BITS_PER_PAGE];
        end = &pid_ns->pidmap[PIDMAP_ENTRIES];
@@ -179,18 +238,19 @@ static int next_pidmap(struct pid_namespace *pid_ns, int last)
        return -1;
 }
 
-fastcall void put_pid(struct pid *pid)
+void put_pid(struct pid *pid)
 {
        struct pid_namespace *ns;
 
        if (!pid)
                return;
 
-       /* FIXME - this must be the namespace this pid lives in */
-       ns = &init_pid_ns;
+       ns = pid->numbers[pid->level].ns;
        if ((atomic_read(&pid->count) == 1) ||
-            atomic_dec_and_test(&pid->count))
+            atomic_dec_and_test(&pid->count)) {
                kmem_cache_free(ns->pid_cachep, pid);
+               put_pid_ns(ns);
+       }
 }
 EXPORT_SYMBOL_GPL(put_pid);
 
@@ -200,71 +260,96 @@ static void delayed_put_pid(struct rcu_head *rhp)
        put_pid(pid);
 }
 
-fastcall void free_pid(struct pid *pid)
+void free_pid(struct pid *pid)
 {
        /* We can be called with write_lock_irq(&tasklist_lock) held */
+       int i;
        unsigned long flags;
 
        spin_lock_irqsave(&pidmap_lock, flags);
-       hlist_del_rcu(&pid->pid_chain);
+       for (i = 0; i <= pid->level; i++)
+               hlist_del_rcu(&pid->numbers[i].pid_chain);
        spin_unlock_irqrestore(&pidmap_lock, flags);
 
-       free_pidmap(&init_pid_ns, pid->nr);
+       for (i = 0; i <= pid->level; i++)
+               free_pidmap(pid->numbers + i);
+
        call_rcu(&pid->rcu, delayed_put_pid);
 }
 
-struct pid *alloc_pid(void)
+struct pid *alloc_pid(struct pid_namespace *ns)
 {
        struct pid *pid;
        enum pid_type type;
-       int nr = -1;
-       struct pid_namespace *ns;
+       int i, nr;
+       struct pid_namespace *tmp;
+       struct upid *upid;
 
-       ns = task_active_pid_ns(current);
        pid = kmem_cache_alloc(ns->pid_cachep, GFP_KERNEL);
        if (!pid)
                goto out;
 
-       nr = alloc_pidmap(ns);
-       if (nr < 0)
-               goto out_free;
+       tmp = ns;
+       for (i = ns->level; i >= 0; i--) {
+               nr = alloc_pidmap(tmp);
+               if (nr < 0)
+                       goto out_free;
 
+               pid->numbers[i].nr = nr;
+               pid->numbers[i].ns = tmp;
+               tmp = tmp->parent;
+       }
+
+       get_pid_ns(ns);
+       pid->level = ns->level;
        atomic_set(&pid->count, 1);
-       pid->nr = nr;
        for (type = 0; type < PIDTYPE_MAX; ++type)
                INIT_HLIST_HEAD(&pid->tasks[type]);
 
+       upid = pid->numbers + ns->level;
        spin_lock_irq(&pidmap_lock);
-       hlist_add_head_rcu(&pid->pid_chain, &pid_hash[pid_hashfn(pid->nr)]);
+       for ( ; upid >= pid->numbers; --upid)
+               hlist_add_head_rcu(&upid->pid_chain,
+                               &pid_hash[pid_hashfn(upid->nr, upid->ns)]);
        spin_unlock_irq(&pidmap_lock);
 
 out:
        return pid;
 
 out_free:
+       while (++i <= ns->level)
+               free_pidmap(pid->numbers + i);
+
        kmem_cache_free(ns->pid_cachep, pid);
        pid = NULL;
        goto out;
 }
 
-struct pid * fastcall find_pid(int nr)
+struct pid *find_pid_ns(int nr, struct pid_namespace *ns)
 {
        struct hlist_node *elem;
-       struct pid *pid;
+       struct upid *pnr;
+
+       hlist_for_each_entry_rcu(pnr, elem,
+                       &pid_hash[pid_hashfn(nr, ns)], pid_chain)
+               if (pnr->nr == nr && pnr->ns == ns)
+                       return container_of(pnr, struct pid,
+                                       numbers[ns->level]);
 
-       hlist_for_each_entry_rcu(pid, elem,
-                       &pid_hash[pid_hashfn(nr)], pid_chain) {
-               if (pid->nr == nr)
-                       return pid;
-       }
        return NULL;
 }
-EXPORT_SYMBOL_GPL(find_pid);
+EXPORT_SYMBOL_GPL(find_pid_ns);
+
+struct pid *find_vpid(int nr)
+{
+       return find_pid_ns(nr, current->nsproxy->pid_ns);
+}
+EXPORT_SYMBOL_GPL(find_vpid);
 
 /*
  * attach_pid() must be called with the tasklist_lock write-held.
  */
-int fastcall attach_pid(struct task_struct *task, enum pid_type type,
+void attach_pid(struct task_struct *task, enum pid_type type,
                struct pid *pid)
 {
        struct pid_link *link;
@@ -272,11 +357,10 @@ int fastcall attach_pid(struct task_struct *task, enum pid_type type,
        link = &task->pids[type];
        link->pid = pid;
        hlist_add_head_rcu(&link->node, &pid->tasks[type]);
-
-       return 0;
 }
 
-void fastcall detach_pid(struct task_struct *task, enum pid_type type)
+static void __change_pid(struct task_struct *task, enum pid_type type,
+                       struct pid *new)
 {
        struct pid_link *link;
        struct pid *pid;
@@ -286,7 +370,7 @@ void fastcall detach_pid(struct task_struct *task, enum pid_type type)
        pid = link->pid;
 
        hlist_del_rcu(&link->node);
-       link->pid = NULL;
+       link->pid = new;
 
        for (tmp = PIDTYPE_MAX; --tmp >= 0; )
                if (!hlist_empty(&pid->tasks[tmp]))
@@ -295,47 +379,69 @@ void fastcall detach_pid(struct task_struct *task, enum pid_type type)
        free_pid(pid);
 }
 
+void detach_pid(struct task_struct *task, enum pid_type type)
+{
+       __change_pid(task, type, NULL);
+}
+
+void change_pid(struct task_struct *task, enum pid_type type,
+               struct pid *pid)
+{
+       __change_pid(task, type, pid);
+       attach_pid(task, type, pid);
+}
+
 /* transfer_pid is an optimization of attach_pid(new), detach_pid(old) */
-void fastcall transfer_pid(struct task_struct *old, struct task_struct *new,
+void transfer_pid(struct task_struct *old, struct task_struct *new,
                           enum pid_type type)
 {
        new->pids[type].pid = old->pids[type].pid;
        hlist_replace_rcu(&old->pids[type].node, &new->pids[type].node);
-       old->pids[type].pid = NULL;
 }
 
-struct task_struct * fastcall pid_task(struct pid *pid, enum pid_type type)
+struct task_struct *pid_task(struct pid *pid, enum pid_type type)
 {
        struct task_struct *result = NULL;
        if (pid) {
                struct hlist_node *first;
-               first = rcu_dereference(pid->tasks[type].first);
+               first = rcu_dereference_check(hlist_first_rcu(&pid->tasks[type]),
+                                             lockdep_tasklist_lock_is_held());
                if (first)
                        result = hlist_entry(first, struct task_struct, pids[(type)].node);
        }
        return result;
 }
+EXPORT_SYMBOL(pid_task);
 
 /*
- * Must be called under rcu_read_lock() or with tasklist_lock read-held.
+ * Must be called under rcu_read_lock().
  */
-struct task_struct *find_task_by_pid_type(int type, int nr)
+struct task_struct *find_task_by_pid_ns(pid_t nr, struct pid_namespace *ns)
 {
-       return pid_task(find_pid(nr), type);
+       rcu_lockdep_assert(rcu_read_lock_held(),
+                          "find_task_by_pid_ns() needs rcu_read_lock()"
+                          " protection");
+       return pid_task(find_pid_ns(nr, ns), PIDTYPE_PID);
 }
 
-EXPORT_SYMBOL(find_task_by_pid_type);
+struct task_struct *find_task_by_vpid(pid_t vnr)
+{
+       return find_task_by_pid_ns(vnr, current->nsproxy->pid_ns);
+}
 
 struct pid *get_task_pid(struct task_struct *task, enum pid_type type)
 {
        struct pid *pid;
        rcu_read_lock();
+       if (type != PIDTYPE_PID)
+               task = task->group_leader;
        pid = get_pid(task->pids[type].pid);
        rcu_read_unlock();
        return pid;
 }
+EXPORT_SYMBOL_GPL(get_task_pid);
 
-struct task_struct *fastcall get_pid_task(struct pid *pid, enum pid_type type)
+struct task_struct *get_pid_task(struct pid *pid, enum pid_type type)
 {
        struct task_struct *result;
        rcu_read_lock();
@@ -345,101 +451,87 @@ struct task_struct *fastcall get_pid_task(struct pid *pid, enum pid_type type)
        rcu_read_unlock();
        return result;
 }
+EXPORT_SYMBOL_GPL(get_pid_task);
 
 struct pid *find_get_pid(pid_t nr)
 {
        struct pid *pid;
 
        rcu_read_lock();
-       pid = get_pid(find_pid(nr));
+       pid = get_pid(find_vpid(nr));
        rcu_read_unlock();
 
        return pid;
 }
+EXPORT_SYMBOL_GPL(find_get_pid);
 
-/*
- * Used by proc to find the first pid that is greater then or equal to nr.
- *
- * If there is a pid at nr this function is exactly the same as find_pid.
- */
-struct pid *find_ge_pid(int nr)
+pid_t pid_nr_ns(struct pid *pid, struct pid_namespace *ns)
 {
-       struct pid *pid;
-
-       do {
-               pid = find_pid(nr);
-               if (pid)
-                       break;
-               nr = next_pidmap(task_active_pid_ns(current), nr);
-       } while (nr > 0);
+       struct upid *upid;
+       pid_t nr = 0;
 
-       return pid;
+       if (pid && ns->level <= pid->level) {
+               upid = &pid->numbers[ns->level];
+               if (upid->ns == ns)
+                       nr = upid->nr;
+       }
+       return nr;
 }
-EXPORT_SYMBOL_GPL(find_get_pid);
-
-struct pid_cache {
-       int nr_ids;
-       char name[16];
-       struct kmem_cache *cachep;
-       struct list_head list;
-};
 
-static LIST_HEAD(pid_caches_lh);
-static DEFINE_MUTEX(pid_caches_mutex);
-
-/*
- * creates the kmem cache to allocate pids from.
- * @nr_ids: the number of numerical ids this pid will have to carry
- */
-
-static struct kmem_cache *create_pid_cachep(int nr_ids)
+pid_t pid_vnr(struct pid *pid)
 {
-       struct pid_cache *pcache;
-       struct kmem_cache *cachep;
-
-       mutex_lock(&pid_caches_mutex);
-       list_for_each_entry (pcache, &pid_caches_lh, list)
-               if (pcache->nr_ids == nr_ids)
-                       goto out;
+       return pid_nr_ns(pid, current->nsproxy->pid_ns);
+}
+EXPORT_SYMBOL_GPL(pid_vnr);
 
-       pcache = kmalloc(sizeof(struct pid_cache), GFP_KERNEL);
-       if (pcache == NULL)
-               goto err_alloc;
+pid_t __task_pid_nr_ns(struct task_struct *task, enum pid_type type,
+                       struct pid_namespace *ns)
+{
+       pid_t nr = 0;
 
-       snprintf(pcache->name, sizeof(pcache->name), "pid_%d", nr_ids);
-       cachep = kmem_cache_create(pcache->name,
-                       /* FIXME add numerical ids here */
-                       sizeof(struct pid), 0, SLAB_HWCACHE_ALIGN, NULL);
-       if (cachep == NULL)
-               goto err_cachep;
+       rcu_read_lock();
+       if (!ns)
+               ns = current->nsproxy->pid_ns;
+       if (likely(pid_alive(task))) {
+               if (type != PIDTYPE_PID)
+                       task = task->group_leader;
+               nr = pid_nr_ns(task->pids[type].pid, ns);
+       }
+       rcu_read_unlock();
 
-       pcache->nr_ids = nr_ids;
-       pcache->cachep = cachep;
-       list_add(&pcache->list, &pid_caches_lh);
-out:
-       mutex_unlock(&pid_caches_mutex);
-       return pcache->cachep;
+       return nr;
+}
+EXPORT_SYMBOL(__task_pid_nr_ns);
 
-err_cachep:
-       kfree(pcache);
-err_alloc:
-       mutex_unlock(&pid_caches_mutex);
-       return NULL;
+pid_t task_tgid_nr_ns(struct task_struct *tsk, struct pid_namespace *ns)
+{
+       return pid_nr_ns(task_tgid(tsk), ns);
 }
+EXPORT_SYMBOL(task_tgid_nr_ns);
 
-struct pid_namespace *copy_pid_ns(unsigned long flags, struct pid_namespace *old_ns)
+struct pid_namespace *task_active_pid_ns(struct task_struct *tsk)
 {
-       BUG_ON(!old_ns);
-       get_pid_ns(old_ns);
-       return old_ns;
+       return ns_of_pid(task_pid(tsk));
 }
+EXPORT_SYMBOL_GPL(task_active_pid_ns);
 
-void free_pid_ns(struct kref *kref)
+/*
+ * Used by proc to find the first pid that is greater than or equal to nr.
+ *
+ * If there is a pid at nr this function is exactly the same as find_pid_ns.
+ */
+struct pid *find_ge_pid(int nr, struct pid_namespace *ns)
 {
-       struct pid_namespace *ns;
+       struct pid *pid;
 
-       ns = container_of(kref, struct pid_namespace, kref);
-       kfree(ns);
+       do {
+               pid = find_pid_ns(nr, ns);
+               if (pid)
+                       break;
+               nr = next_pidmap(ns, nr);
+       } while (nr > 0);
+
+       return pid;
 }
 
 /*
@@ -450,31 +542,30 @@ void free_pid_ns(struct kref *kref)
 void __init pidhash_init(void)
 {
        int i, pidhash_size;
-       unsigned long megabytes = nr_kernel_pages >> (20 - PAGE_SHIFT);
 
-       pidhash_shift = max(4, fls(megabytes * 4));
-       pidhash_shift = min(12, pidhash_shift);
+       pid_hash = alloc_large_system_hash("PID", sizeof(*pid_hash), 0, 18,
+                                          HASH_EARLY | HASH_SMALL,
+                                          &pidhash_shift, NULL, 4096);
        pidhash_size = 1 << pidhash_shift;
 
-       printk("PID hash table entries: %d (order: %d, %Zd bytes)\n",
-               pidhash_size, pidhash_shift,
-               pidhash_size * sizeof(struct hlist_head));
-
-       pid_hash = alloc_bootmem(pidhash_size * sizeof(*(pid_hash)));
-       if (!pid_hash)
-               panic("Could not alloc pidhash!\n");
        for (i = 0; i < pidhash_size; i++)
                INIT_HLIST_HEAD(&pid_hash[i]);
 }
 
 void __init pidmap_init(void)
 {
+       /* bump default and minimum pid_max based on number of cpus */
+       pid_max = min(pid_max_max, max_t(int, pid_max,
+                               PIDS_PER_CPU_DEFAULT * num_possible_cpus()));
+       pid_max_min = max_t(int, pid_max_min,
+                               PIDS_PER_CPU_MIN * num_possible_cpus());
+       pr_info("pid_max: default: %u minimum: %u\n", pid_max, pid_max_min);
+
        init_pid_ns.pidmap[0].page = kzalloc(PAGE_SIZE, GFP_KERNEL);
        /* Reserve PID 0. We never call free_pidmap(0) */
        set_bit(0, init_pid_ns.pidmap[0].page);
        atomic_dec(&init_pid_ns.pidmap[0].nr_free);
 
-       init_pid_ns.pid_cachep = create_pid_cachep(1);
-       if (init_pid_ns.pid_cachep == NULL)
-               panic("Can't create pid_1 cachep\n");
+       init_pid_ns.pid_cachep = KMEM_CACHE(pid,
+                       SLAB_HWCACHE_ALIGN | SLAB_PANIC);
 }