ocfs2: Don't repeat ocfs2_xattr_block_find()
[linux-2.6.git] / fs / select.c
index bc60a3e..87df51e 100644 (file)
  *     of fds to overcome nfds < 16390 descriptors limit (Tigran Aivazian).
  */
 
+#include <linux/kernel.h>
 #include <linux/syscalls.h>
 #include <linux/module.h>
 #include <linux/slab.h>
-#include <linux/smp_lock.h>
 #include <linux/poll.h>
 #include <linux/personality.h> /* for STICKY_TIMEOUTS */
 #include <linux/file.h>
+#include <linux/fdtable.h>
 #include <linux/fs.h>
 #include <linux/rcupdate.h>
+#include <linux/hrtimer.h>
 
 #include <asm/uaccess.h>
 
-#define ROUND_UP(x,y) (((x)+(y)-1)/(y))
-#define DEFAULT_POLLMASK (POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM)
 
-struct poll_table_entry {
-       struct file * filp;
-       wait_queue_t wait;
-       wait_queue_head_t * wait_address;
-};
+/*
+ * Estimate expected accuracy in ns from a timeval.
+ *
+ * After quite a bit of churning around, we've settled on
+ * a simple thing of taking 0.1% of the timeout as the
+ * slack, with a cap of 100 msec.
+ * "nice" tasks get a 0.5% slack instead.
+ *
+ * Consider this comment an open invitation to come up with even
+ * better solutions..
+ */
+
+static long __estimate_accuracy(struct timespec *tv)
+{
+       long slack;
+       int divfactor = 1000;
+
+       if (task_nice(current) > 0)
+               divfactor = divfactor / 5;
+
+       slack = tv->tv_nsec / divfactor;
+       slack += tv->tv_sec * (NSEC_PER_SEC/divfactor);
+
+       if (slack > 100 * NSEC_PER_MSEC)
+               slack =  100 * NSEC_PER_MSEC;
+
+       if (slack < 0)
+               slack = 0;
+       return slack;
+}
+
+static long estimate_accuracy(struct timespec *tv)
+{
+       unsigned long ret;
+       struct timespec now;
+
+       /*
+        * Realtime tasks get a slack of 0 for obvious reasons.
+        */
+
+       if (rt_task(current))
+               return 0;
+
+       ktime_get_ts(&now);
+       now = timespec_sub(*tv, now);
+       ret = __estimate_accuracy(&now);
+       if (ret < current->timer_slack_ns)
+               return current->timer_slack_ns;
+       return ret;
+}
+
+
 
 struct poll_table_page {
        struct poll_table_page * next;
@@ -64,13 +111,23 @@ void poll_initwait(struct poll_wqueues *pwq)
        init_poll_funcptr(&pwq->pt, __pollwait);
        pwq->error = 0;
        pwq->table = NULL;
+       pwq->inline_index = 0;
 }
 
 EXPORT_SYMBOL(poll_initwait);
 
+static void free_poll_entry(struct poll_table_entry *entry)
+{
+       remove_wait_queue(entry->wait_address, &entry->wait);
+       fput(entry->filp);
+}
+
 void poll_freewait(struct poll_wqueues *pwq)
 {
        struct poll_table_page * p = pwq->table;
+       int i;
+       for (i = 0; i < pwq->inline_index; i++)
+               free_poll_entry(pwq->inline_entries + i);
        while (p) {
                struct poll_table_entry * entry;
                struct poll_table_page *old;
@@ -78,8 +135,7 @@ void poll_freewait(struct poll_wqueues *pwq)
                entry = p->entry;
                do {
                        entry--;
-                       remove_wait_queue(entry->wait_address,&entry->wait);
-                       fput(entry->filp);
+                       free_poll_entry(entry);
                } while (entry > p->entries);
                old = p;
                p = p->next;
@@ -89,12 +145,14 @@ void poll_freewait(struct poll_wqueues *pwq)
 
 EXPORT_SYMBOL(poll_freewait);
 
-static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
-                      poll_table *_p)
+static struct poll_table_entry *poll_get_entry(poll_table *_p)
 {
        struct poll_wqueues *p = container_of(_p, struct poll_wqueues, pt);
        struct poll_table_page *table = p->table;
 
+       if (p->inline_index < N_INLINE_POLL_ENTRIES)
+               return p->inline_entries + p->inline_index++;
+
        if (!table || POLL_TABLE_FULL(table)) {
                struct poll_table_page *new_table;
 
@@ -102,7 +160,7 @@ static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
                if (!new_table) {
                        p->error = -ENOMEM;
                        __set_current_state(TASK_RUNNING);
-                       return;
+                       return NULL;
                }
                new_table->entry = new_table->entries;
                new_table->next = table;
@@ -110,16 +168,94 @@ static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
                table = new_table;
        }
 
-       /* Add a new entry */
-       {
-               struct poll_table_entry * entry = table->entry;
-               table->entry = entry+1;
-               get_file(filp);
-               entry->filp = filp;
-               entry->wait_address = wait_address;
-               init_waitqueue_entry(&entry->wait, current);
-               add_wait_queue(wait_address,&entry->wait);
+       return table->entry++;
+}
+
+/* Add a new entry */
+static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
+                               poll_table *p)
+{
+       struct poll_table_entry *entry = poll_get_entry(p);
+       if (!entry)
+               return;
+       get_file(filp);
+       entry->filp = filp;
+       entry->wait_address = wait_address;
+       init_waitqueue_entry(&entry->wait, current);
+       add_wait_queue(wait_address, &entry->wait);
+}
+
+/**
+ * poll_select_set_timeout - helper function to setup the timeout value
+ * @to:                pointer to timespec variable for the final timeout
+ * @sec:       seconds (from user space)
+ * @nsec:      nanoseconds (from user space)
+ *
+ * Note, we do not use a timespec for the user space value here, That
+ * way we can use the function for timeval and compat interfaces as well.
+ *
+ * Returns -EINVAL if sec/nsec are not normalized. Otherwise 0.
+ */
+int poll_select_set_timeout(struct timespec *to, long sec, long nsec)
+{
+       struct timespec ts = {.tv_sec = sec, .tv_nsec = nsec};
+
+       if (!timespec_valid(&ts))
+               return -EINVAL;
+
+       /* Optimize for the zero timeout value here */
+       if (!sec && !nsec) {
+               to->tv_sec = to->tv_nsec = 0;
+       } else {
+               ktime_get_ts(to);
+               *to = timespec_add_safe(*to, ts);
        }
+       return 0;
+}
+
+static int poll_select_copy_remaining(struct timespec *end_time, void __user *p,
+                                     int timeval, int ret)
+{
+       struct timespec rts;
+       struct timeval rtv;
+
+       if (!p)
+               return ret;
+
+       if (current->personality & STICKY_TIMEOUTS)
+               goto sticky;
+
+       /* No update for zero timeout */
+       if (!end_time->tv_sec && !end_time->tv_nsec)
+               return ret;
+
+       ktime_get_ts(&rts);
+       rts = timespec_sub(*end_time, rts);
+       if (rts.tv_sec < 0)
+               rts.tv_sec = rts.tv_nsec = 0;
+
+       if (timeval) {
+               rtv.tv_sec = rts.tv_sec;
+               rtv.tv_usec = rts.tv_nsec / NSEC_PER_USEC;
+
+               if (!copy_to_user(p, &rtv, sizeof(rtv)))
+                       return ret;
+
+       } else if (!copy_to_user(p, &rts, sizeof(rts)))
+               return ret;
+
+       /*
+        * If an application puts its timeval in read-only memory, we
+        * don't want the Linux-specific update to the timeval to
+        * cause a fault after the select has completed
+        * successfully. However, because we're not updating the
+        * timeval, we can't restart the system call.
+        */
+
+sticky:
+       if (ret == -ERESTARTNOHAND)
+               ret = -EINTR;
+       return ret;
 }
 
 #define FDS_IN(fds, n)         (fds->in + n)
@@ -170,20 +306,17 @@ get_max:
        return max;
 }
 
-#define BIT(i)         (1UL << ((i)&(__NFDBITS-1)))
-#define MEM(i,m)       ((m)+(unsigned)(i)/__NFDBITS)
-#define ISSET(i,m)     (((i)&*(m)) != 0)
-#define SET(i,m)       (*(m) |= (i))
-
 #define POLLIN_SET (POLLRDNORM | POLLRDBAND | POLLIN | POLLHUP | POLLERR)
 #define POLLOUT_SET (POLLWRBAND | POLLWRNORM | POLLOUT | POLLERR)
 #define POLLEX_SET (POLLPRI)
 
-int do_select(int n, fd_set_bits *fds, s64 *timeout)
+int do_select(int n, fd_set_bits *fds, struct timespec *end_time)
 {
+       ktime_t expire, *to = NULL;
        struct poll_wqueues table;
        poll_table *wait;
-       int retval, i;
+       int retval, i, timed_out = 0;
+       unsigned long slack = 0;
 
        rcu_read_lock();
        retval = max_select_fd(n, fds);
@@ -195,12 +328,17 @@ int do_select(int n, fd_set_bits *fds, s64 *timeout)
 
        poll_initwait(&table);
        wait = &table.pt;
-       if (!*timeout)
+       if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
                wait = NULL;
+               timed_out = 1;
+       }
+
+       if (end_time && !timed_out)
+               slack = estimate_accuracy(end_time);
+
        retval = 0;
        for (;;) {
                unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp;
-               long __timeout;
 
                set_current_state(TASK_INTERRUPTIBLE);
 
@@ -210,7 +348,7 @@ int do_select(int n, fd_set_bits *fds, s64 *timeout)
                for (i = 0; i < n; ++rinp, ++routp, ++rexp) {
                        unsigned long in, out, ex, all_bits, bit = 1, mask, j;
                        unsigned long res_in = 0, res_out = 0, res_ex = 0;
-                       struct file_operations *f_op = NULL;
+                       const struct file_operations *f_op = NULL;
                        struct file *file = NULL;
 
                        in = *inp++; out = *outp++; ex = *exp++;
@@ -221,17 +359,18 @@ int do_select(int n, fd_set_bits *fds, s64 *timeout)
                        }
 
                        for (j = 0; j < __NFDBITS; ++j, ++i, bit <<= 1) {
+                               int fput_needed;
                                if (i >= n)
                                        break;
                                if (!(bit & all_bits))
                                        continue;
-                               file = fget(i);
+                               file = fget_light(i, &fput_needed);
                                if (file) {
                                        f_op = file->f_op;
                                        mask = DEFAULT_POLLMASK;
                                        if (f_op && f_op->poll)
                                                mask = (*f_op->poll)(file, retval ? NULL : wait);
-                                       fput(file);
+                                       fput_light(file, fput_needed);
                                        if ((mask & POLLIN_SET) && (in & bit)) {
                                                res_in |= bit;
                                                retval++;
@@ -245,7 +384,6 @@ int do_select(int n, fd_set_bits *fds, s64 *timeout)
                                                retval++;
                                        }
                                }
-                               cond_resched();
                        }
                        if (res_in)
                                *rinp = res_in;
@@ -253,29 +391,28 @@ int do_select(int n, fd_set_bits *fds, s64 *timeout)
                                *routp = res_out;
                        if (res_ex)
                                *rexp = res_ex;
+                       cond_resched();
                }
                wait = NULL;
-               if (retval || !*timeout || signal_pending(current))
+               if (retval || timed_out || signal_pending(current))
                        break;
-               if(table.error) {
+               if (table.error) {
                        retval = table.error;
                        break;
                }
 
-               if (*timeout < 0) {
-                       /* Wait indefinitely */
-                       __timeout = MAX_SCHEDULE_TIMEOUT;
-               } else if (unlikely(*timeout >= (s64)MAX_SCHEDULE_TIMEOUT - 1)) {
-                       /* Wait for longer than MAX_SCHEDULE_TIMEOUT. Do it in a loop */
-                       __timeout = MAX_SCHEDULE_TIMEOUT - 1;
-                       *timeout -= __timeout;
-               } else {
-                       __timeout = *timeout;
-                       *timeout = 0;
+               /*
+                * If this is the first loop and we have a timeout
+                * given, then we convert to ktime_t and set the to
+                * pointer to the expiry value.
+                */
+               if (end_time && !to) {
+                       expire = timespec_to_ktime(*end_time);
+                       to = &expire;
                }
-               __timeout = schedule_timeout(__timeout);
-               if (*timeout >= 0)
-                       *timeout += __timeout;
+
+               if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
+                       timed_out = 1;
        }
        __set_current_state(TASK_RUNNING);
 
@@ -284,16 +421,6 @@ int do_select(int n, fd_set_bits *fds, s64 *timeout)
        return retval;
 }
 
-static void *select_bits_alloc(int size)
-{
-       return kmalloc(6 * size, GFP_KERNEL);
-}
-
-static void select_bits_free(void *bits, int size)
-{
-       kfree(bits);
-}
-
 /*
  * We can actually return ERESTARTSYS instead of EINTR, but I'd
  * like to be certain this leads to no problems. So I return
@@ -305,42 +432,49 @@ static void select_bits_free(void *bits, int size)
 #define MAX_SELECT_SECONDS \
        ((unsigned long) (MAX_SCHEDULE_TIMEOUT / HZ)-1)
 
-static int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
-                          fd_set __user *exp, s64 *timeout)
+int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
+                          fd_set __user *exp, struct timespec *end_time)
 {
        fd_set_bits fds;
-       char *bits;
-       int ret, size, max_fdset;
+       void *bits;
+       int ret, max_fds;
+       unsigned int size;
        struct fdtable *fdt;
+       /* Allocate small arguments on the stack to save memory and be faster */
+       long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
 
        ret = -EINVAL;
        if (n < 0)
                goto out_nofds;
 
-       /* max_fdset can increase, so grab it once to avoid race */
+       /* max_fds can increase, so grab it once to avoid race */
        rcu_read_lock();
        fdt = files_fdtable(current->files);
-       max_fdset = fdt->max_fdset;
+       max_fds = fdt->max_fds;
        rcu_read_unlock();
-       if (n > max_fdset)
-               n = max_fdset;
+       if (n > max_fds)
+               n = max_fds;
 
        /*
         * We need 6 bitmaps (in/out/ex for both incoming and outgoing),
         * since we used fdset we need to allocate memory in units of
         * long-words. 
         */
-       ret = -ENOMEM;
        size = FDS_BYTES(n);
-       bits = select_bits_alloc(size);
-       if (!bits)
-               goto out_nofds;
-       fds.in      = (unsigned long *)  bits;
-       fds.out     = (unsigned long *) (bits +   size);
-       fds.ex      = (unsigned long *) (bits + 2*size);
-       fds.res_in  = (unsigned long *) (bits + 3*size);
-       fds.res_out = (unsigned long *) (bits + 4*size);
-       fds.res_ex  = (unsigned long *) (bits + 5*size);
+       bits = stack_fds;
+       if (size > sizeof(stack_fds) / 6) {
+               /* Not enough space in on-stack array; must use kmalloc */
+               ret = -ENOMEM;
+               bits = kmalloc(6 * size, GFP_KERNEL);
+               if (!bits)
+                       goto out_nofds;
+       }
+       fds.in      = bits;
+       fds.out     = bits +   size;
+       fds.ex      = bits + 2*size;
+       fds.res_in  = bits + 3*size;
+       fds.res_out = bits + 4*size;
+       fds.res_ex  = bits + 5*size;
 
        if ((ret = get_fd_set(n, inp, fds.in)) ||
            (ret = get_fd_set(n, outp, fds.out)) ||
@@ -350,7 +484,7 @@ static int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
        zero_fd_set(n, fds.res_out);
        zero_fd_set(n, fds.res_ex);
 
-       ret = do_select(n, &fds, timeout);
+       ret = do_select(n, &fds, end_time);
 
        if (ret < 0)
                goto out;
@@ -367,7 +501,8 @@ static int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
                ret = -EFAULT;
 
 out:
-       select_bits_free(bits, size);
+       if (bits != stack_fds)
+               kfree(bits);
 out_nofds:
        return ret;
 }
@@ -375,7 +510,7 @@ out_nofds:
 asmlinkage long sys_select(int n, fd_set __user *inp, fd_set __user *outp,
                        fd_set __user *exp, struct timeval __user *tvp)
 {
-       s64 timeout = -1;
+       struct timespec end_time, *to = NULL;
        struct timeval tv;
        int ret;
 
@@ -383,67 +518,35 @@ asmlinkage long sys_select(int n, fd_set __user *inp, fd_set __user *outp,
                if (copy_from_user(&tv, tvp, sizeof(tv)))
                        return -EFAULT;
 
-               if (tv.tv_sec < 0 || tv.tv_usec < 0)
+               to = &end_time;
+               if (poll_select_set_timeout(to,
+                               tv.tv_sec + (tv.tv_usec / USEC_PER_SEC),
+                               (tv.tv_usec % USEC_PER_SEC) * NSEC_PER_USEC))
                        return -EINVAL;
-
-               /* Cast to u64 to make GCC stop complaining */
-               if ((u64)tv.tv_sec >= (u64)MAX_INT64_SECONDS)
-                       timeout = -1;   /* infinite */
-               else {
-                       timeout = ROUND_UP(tv.tv_usec, USEC_PER_SEC/HZ);
-                       timeout += tv.tv_sec * HZ;
-               }
        }
 
-       ret = core_sys_select(n, inp, outp, exp, &timeout);
-
-       if (tvp) {
-               if (current->personality & STICKY_TIMEOUTS)
-                       goto sticky;
-               tv.tv_usec = jiffies_to_usecs(do_div((*(u64*)&timeout), HZ));
-               tv.tv_sec = timeout;
-               if (copy_to_user(tvp, &tv, sizeof(tv))) {
-sticky:
-                       /*
-                        * If an application puts its timeval in read-only
-                        * memory, we don't want the Linux-specific update to
-                        * the timeval to cause a fault after the select has
-                        * completed successfully. However, because we're not
-                        * updating the timeval, we can't restart the system
-                        * call.
-                        */
-                       if (ret == -ERESTARTNOHAND)
-                               ret = -EINTR;
-               }
-       }
+       ret = core_sys_select(n, inp, outp, exp, to);
+       ret = poll_select_copy_remaining(&end_time, tvp, 1, ret);
 
        return ret;
 }
 
-#ifdef TIF_RESTORE_SIGMASK
+#ifdef HAVE_SET_RESTORE_SIGMASK
 asmlinkage long sys_pselect7(int n, fd_set __user *inp, fd_set __user *outp,
                fd_set __user *exp, struct timespec __user *tsp,
                const sigset_t __user *sigmask, size_t sigsetsize)
 {
-       s64 timeout = MAX_SCHEDULE_TIMEOUT;
        sigset_t ksigmask, sigsaved;
-       struct timespec ts;
+       struct timespec ts, end_time, *to = NULL;
        int ret;
 
        if (tsp) {
                if (copy_from_user(&ts, tsp, sizeof(ts)))
                        return -EFAULT;
 
-               if (ts.tv_sec < 0 || ts.tv_nsec < 0)
+               to = &end_time;
+               if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
                        return -EINVAL;
-
-               /* Cast to u64 to make GCC stop complaining */
-               if ((u64)ts.tv_sec >= (u64)MAX_INT64_SECONDS)
-                       timeout = -1;   /* infinite */
-               else {
-                       timeout = ROUND_UP(ts.tv_nsec, NSEC_PER_SEC/HZ);
-                       timeout += ts.tv_sec * HZ;
-               }
        }
 
        if (sigmask) {
@@ -457,27 +560,8 @@ asmlinkage long sys_pselect7(int n, fd_set __user *inp, fd_set __user *outp,
                sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
        }
 
-       ret = core_sys_select(n, inp, outp, exp, &timeout);
-
-       if (tsp) {
-               if (current->personality & STICKY_TIMEOUTS)
-                       goto sticky;
-               ts.tv_nsec = jiffies_to_usecs(do_div((*(u64*)&timeout), HZ)) * 1000;
-               ts.tv_sec = timeout;
-               if (copy_to_user(tsp, &ts, sizeof(ts))) {
-sticky:
-                       /*
-                        * If an application puts its timeval in read-only
-                        * memory, we don't want the Linux-specific update to
-                        * the timeval to cause a fault after the select has
-                        * completed successfully. However, because we're not
-                        * updating the timeval, we can't restart the system
-                        * call.
-                        */
-                       if (ret == -ERESTARTNOHAND)
-                               ret = -EINTR;
-               }
-       }
+       ret = core_sys_select(n, inp, outp, exp, &end_time);
+       ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
 
        if (ret == -ERESTARTNOHAND) {
                /*
@@ -488,7 +572,7 @@ sticky:
                if (sigmask) {
                        memcpy(&current->saved_sigmask, &sigsaved,
                                        sizeof(sigsaved));
-                       set_thread_flag(TIF_RESTORE_SIGMASK);
+                       set_restore_sigmask();
                }
        } else if (sigmask)
                sigprocmask(SIG_SETMASK, &sigsaved, NULL);
@@ -518,7 +602,7 @@ asmlinkage long sys_pselect6(int n, fd_set __user *inp, fd_set __user *outp,
 
        return sys_pselect7(n, inp, outp, exp, tsp, up, sigsetsize);
 }
-#endif /* TIF_RESTORE_SIGMASK */
+#endif /* HAVE_SET_RESTORE_SIGMASK */
 
 struct poll_list {
        struct poll_list *next;
@@ -528,204 +612,250 @@ struct poll_list {
 
 #define POLLFD_PER_PAGE  ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd))
 
-static void do_pollfd(unsigned int num, struct pollfd * fdpage,
-       poll_table ** pwait, int *count)
+/*
+ * Fish for pollable events on the pollfd->fd file descriptor. We're only
+ * interested in events matching the pollfd->events mask, and the result
+ * matching that mask is both recorded in pollfd->revents and returned. The
+ * pwait poll_table will be used by the fd-provided poll handler for waiting,
+ * if non-NULL.
+ */
+static inline unsigned int do_pollfd(struct pollfd *pollfd, poll_table *pwait)
 {
-       int i;
-
-       for (i = 0; i < num; i++) {
-               int fd;
-               unsigned int mask;
-               struct pollfd *fdp;
-
-               mask = 0;
-               fdp = fdpage+i;
-               fd = fdp->fd;
-               if (fd >= 0) {
-                       struct file * file = fget(fd);
-                       mask = POLLNVAL;
-                       if (file != NULL) {
-                               mask = DEFAULT_POLLMASK;
-                               if (file->f_op && file->f_op->poll)
-                                       mask = file->f_op->poll(file, *pwait);
-                               mask &= fdp->events | POLLERR | POLLHUP;
-                               fput(file);
-                       }
-                       if (mask) {
-                               *pwait = NULL;
-                               (*count)++;
-                       }
+       unsigned int mask;
+       int fd;
+
+       mask = 0;
+       fd = pollfd->fd;
+       if (fd >= 0) {
+               int fput_needed;
+               struct file * file;
+
+               file = fget_light(fd, &fput_needed);
+               mask = POLLNVAL;
+               if (file != NULL) {
+                       mask = DEFAULT_POLLMASK;
+                       if (file->f_op && file->f_op->poll)
+                               mask = file->f_op->poll(file, pwait);
+                       /* Mask out unneeded events. */
+                       mask &= pollfd->events | POLLERR | POLLHUP;
+                       fput_light(file, fput_needed);
                }
-               fdp->revents = mask;
        }
+       pollfd->revents = mask;
+
+       return mask;
 }
 
 static int do_poll(unsigned int nfds,  struct poll_list *list,
-                  struct poll_wqueues *wait, s64 *timeout)
+                  struct poll_wqueues *wait, struct timespec *end_time)
 {
-       int count = 0;
        poll_table* pt = &wait->pt;
+       ktime_t expire, *to = NULL;
+       int timed_out = 0, count = 0;
+       unsigned long slack = 0;
 
        /* Optimise the no-wait case */
-       if (!(*timeout))
+       if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
                pt = NULL;
+               timed_out = 1;
+       }
+
+       if (end_time && !timed_out)
+               slack = estimate_accuracy(end_time);
+
        for (;;) {
                struct poll_list *walk;
-               long __timeout;
 
                set_current_state(TASK_INTERRUPTIBLE);
-               walk = list;
-               while(walk != NULL) {
-                       do_pollfd( walk->len, walk->entries, &pt, &count);
-                       walk = walk->next;
+               for (walk = list; walk != NULL; walk = walk->next) {
+                       struct pollfd * pfd, * pfd_end;
+
+                       pfd = walk->entries;
+                       pfd_end = pfd + walk->len;
+                       for (; pfd != pfd_end; pfd++) {
+                               /*
+                                * Fish for events. If we found one, record it
+                                * and kill the poll_table, so we don't
+                                * needlessly register any other waiters after
+                                * this. They'll get immediately deregistered
+                                * when we break out and return.
+                                */
+                               if (do_pollfd(pfd, pt)) {
+                                       count++;
+                                       pt = NULL;
+                               }
+                       }
                }
+               /*
+                * All waiters have already been registered, so don't provide
+                * a poll_table to them on the next loop iteration.
+                */
                pt = NULL;
-               if (count || !*timeout || signal_pending(current))
-                       break;
-               count = wait->error;
-               if (count)
+               if (!count) {
+                       count = wait->error;
+                       if (signal_pending(current))
+                               count = -EINTR;
+               }
+               if (count || timed_out)
                        break;
 
-               if (*timeout < 0) {
-                       /* Wait indefinitely */
-                       __timeout = MAX_SCHEDULE_TIMEOUT;
-               } else if (unlikely(*timeout >= (s64)MAX_SCHEDULE_TIMEOUT-1)) {
-                       /*
-                        * Wait for longer than MAX_SCHEDULE_TIMEOUT. Do it in
-                        * a loop
-                        */
-                       __timeout = MAX_SCHEDULE_TIMEOUT - 1;
-                       *timeout -= __timeout;
-               } else {
-                       __timeout = *timeout;
-                       *timeout = 0;
+               /*
+                * If this is the first loop and we have a timeout
+                * given, then we convert to ktime_t and set the to
+                * pointer to the expiry value.
+                */
+               if (end_time && !to) {
+                       expire = timespec_to_ktime(*end_time);
+                       to = &expire;
                }
 
-               __timeout = schedule_timeout(__timeout);
-               if (*timeout >= 0)
-                       *timeout += __timeout;
+               if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
+                       timed_out = 1;
        }
        __set_current_state(TASK_RUNNING);
        return count;
 }
 
-int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds, s64 *timeout)
+#define N_STACK_PPS ((sizeof(stack_pps) - sizeof(struct poll_list))  / \
+                       sizeof(struct pollfd))
+
+int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds,
+               struct timespec *end_time)
 {
        struct poll_wqueues table;
-       int fdcount, err;
-       unsigned int i;
-       struct poll_list *head;
-       struct poll_list *walk;
-       struct fdtable *fdt;
-       int max_fdset;
-
-       /* Do a sanity check on nfds ... */
-       rcu_read_lock();
-       fdt = files_fdtable(current->files);
-       max_fdset = fdt->max_fdset;
-       rcu_read_unlock();
-       if (nfds > max_fdset && nfds > OPEN_MAX)
+       int err = -EFAULT, fdcount, len, size;
+       /* Allocate small arguments on the stack to save memory and be
+          faster - use long to make sure the buffer is aligned properly
+          on 64 bit archs to avoid unaligned access */
+       long stack_pps[POLL_STACK_ALLOC/sizeof(long)];
+       struct poll_list *const head = (struct poll_list *)stack_pps;
+       struct poll_list *walk = head;
+       unsigned long todo = nfds;
+
+       if (nfds > current->signal->rlim[RLIMIT_NOFILE].rlim_cur)
                return -EINVAL;
 
-       poll_initwait(&table);
+       len = min_t(unsigned int, nfds, N_STACK_PPS);
+       for (;;) {
+               walk->next = NULL;
+               walk->len = len;
+               if (!len)
+                       break;
 
-       head = NULL;
-       walk = NULL;
-       i = nfds;
-       err = -ENOMEM;
-       while(i!=0) {
-               struct poll_list *pp;
-               pp = kmalloc(sizeof(struct poll_list)+
-                               sizeof(struct pollfd)*
-                               (i>POLLFD_PER_PAGE?POLLFD_PER_PAGE:i),
-                                       GFP_KERNEL);
-               if(pp==NULL)
+               if (copy_from_user(walk->entries, ufds + nfds-todo,
+                                       sizeof(struct pollfd) * walk->len))
                        goto out_fds;
-               pp->next=NULL;
-               pp->len = (i>POLLFD_PER_PAGE?POLLFD_PER_PAGE:i);
-               if (head == NULL)
-                       head = pp;
-               else
-                       walk->next = pp;
-
-               walk = pp;
-               if (copy_from_user(pp->entries, ufds + nfds-i, 
-                               sizeof(struct pollfd)*pp->len)) {
-                       err = -EFAULT;
+
+               todo -= walk->len;
+               if (!todo)
+                       break;
+
+               len = min(todo, POLLFD_PER_PAGE);
+               size = sizeof(struct poll_list) + sizeof(struct pollfd) * len;
+               walk = walk->next = kmalloc(size, GFP_KERNEL);
+               if (!walk) {
+                       err = -ENOMEM;
                        goto out_fds;
                }
-               i -= pp->len;
        }
 
-       fdcount = do_poll(nfds, head, &table, timeout);
+       poll_initwait(&table);
+       fdcount = do_poll(nfds, head, &table, end_time);
+       poll_freewait(&table);
 
-       /* OK, now copy the revents fields back to user space. */
-       walk = head;
-       err = -EFAULT;
-       while(walk != NULL) {
+       for (walk = head; walk; walk = walk->next) {
                struct pollfd *fds = walk->entries;
                int j;
 
-               for (j=0; j < walk->len; j++, ufds++) {
-                       if(__put_user(fds[j].revents, &ufds->revents))
+               for (j = 0; j < walk->len; j++, ufds++)
+                       if (__put_user(fds[j].revents, &ufds->revents))
                                goto out_fds;
-               }
-               walk = walk->next;
        }
+
        err = fdcount;
-       if (!fdcount && signal_pending(current))
-               err = -EINTR;
 out_fds:
-       walk = head;
-       while(walk!=NULL) {
-               struct poll_list *pp = walk->next;
-               kfree(walk);
-               walk = pp;
+       walk = head->next;
+       while (walk) {
+               struct poll_list *pos = walk;
+               walk = walk->next;
+               kfree(pos);
        }
-       poll_freewait(&table);
+
        return err;
 }
 
+static long do_restart_poll(struct restart_block *restart_block)
+{
+       struct pollfd __user *ufds = restart_block->poll.ufds;
+       int nfds = restart_block->poll.nfds;
+       struct timespec *to = NULL, end_time;
+       int ret;
+
+       if (restart_block->poll.has_timeout) {
+               end_time.tv_sec = restart_block->poll.tv_sec;
+               end_time.tv_nsec = restart_block->poll.tv_nsec;
+               to = &end_time;
+       }
+
+       ret = do_sys_poll(ufds, nfds, to);
+
+       if (ret == -EINTR) {
+               restart_block->fn = do_restart_poll;
+               ret = -ERESTART_RESTARTBLOCK;
+       }
+       return ret;
+}
+
 asmlinkage long sys_poll(struct pollfd __user *ufds, unsigned int nfds,
                        long timeout_msecs)
 {
-       s64 timeout_jiffies = 0;
-
-       if (timeout_msecs) {
-#if HZ > 1000
-               /* We can only overflow if HZ > 1000 */
-               if (timeout_msecs / 1000 > (s64)0x7fffffffffffffffULL / (s64)HZ)
-                       timeout_jiffies = -1;
-               else
-#endif
-                       timeout_jiffies = msecs_to_jiffies(timeout_msecs);
+       struct timespec end_time, *to = NULL;
+       int ret;
+
+       if (timeout_msecs >= 0) {
+               to = &end_time;
+               poll_select_set_timeout(to, timeout_msecs / MSEC_PER_SEC,
+                       NSEC_PER_MSEC * (timeout_msecs % MSEC_PER_SEC));
        }
 
-       return do_sys_poll(ufds, nfds, &timeout_jiffies);
+       ret = do_sys_poll(ufds, nfds, to);
+
+       if (ret == -EINTR) {
+               struct restart_block *restart_block;
+
+               restart_block = &current_thread_info()->restart_block;
+               restart_block->fn = do_restart_poll;
+               restart_block->poll.ufds = ufds;
+               restart_block->poll.nfds = nfds;
+
+               if (timeout_msecs >= 0) {
+                       restart_block->poll.tv_sec = end_time.tv_sec;
+                       restart_block->poll.tv_nsec = end_time.tv_nsec;
+                       restart_block->poll.has_timeout = 1;
+               } else
+                       restart_block->poll.has_timeout = 0;
+
+               ret = -ERESTART_RESTARTBLOCK;
+       }
+       return ret;
 }
 
-#ifdef TIF_RESTORE_SIGMASK
+#ifdef HAVE_SET_RESTORE_SIGMASK
 asmlinkage long sys_ppoll(struct pollfd __user *ufds, unsigned int nfds,
        struct timespec __user *tsp, const sigset_t __user *sigmask,
        size_t sigsetsize)
 {
        sigset_t ksigmask, sigsaved;
-       struct timespec ts;
-       s64 timeout = -1;
+       struct timespec ts, end_time, *to = NULL;
        int ret;
 
        if (tsp) {
                if (copy_from_user(&ts, tsp, sizeof(ts)))
                        return -EFAULT;
 
-               /* Cast to u64 to make GCC stop complaining */
-               if ((u64)ts.tv_sec >= (u64)MAX_INT64_SECONDS)
-                       timeout = -1;   /* infinite */
-               else {
-                       timeout = ROUND_UP(ts.tv_nsec, NSEC_PER_SEC/HZ);
-                       timeout += ts.tv_sec * HZ;
-               }
+               to = &end_time;
+               if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
+                       return -EINVAL;
        }
 
        if (sigmask) {
@@ -739,7 +869,7 @@ asmlinkage long sys_ppoll(struct pollfd __user *ufds, unsigned int nfds,
                sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
        }
 
-       ret = do_sys_poll(ufds, nfds, &timeout);
+       ret = do_sys_poll(ufds, nfds, to);
 
        /* We can restart this syscall, usually */
        if (ret == -EINTR) {
@@ -751,33 +881,14 @@ asmlinkage long sys_ppoll(struct pollfd __user *ufds, unsigned int nfds,
                if (sigmask) {
                        memcpy(&current->saved_sigmask, &sigsaved,
                                        sizeof(sigsaved));
-                       set_thread_flag(TIF_RESTORE_SIGMASK);
+                       set_restore_sigmask();
                }
                ret = -ERESTARTNOHAND;
        } else if (sigmask)
                sigprocmask(SIG_SETMASK, &sigsaved, NULL);
 
-       if (tsp && timeout >= 0) {
-               if (current->personality & STICKY_TIMEOUTS)
-                       goto sticky;
-               /* Yes, we know it's actually an s64, but it's also positive. */
-               ts.tv_nsec = jiffies_to_usecs(do_div((*(u64*)&timeout), HZ)) * 1000;
-               ts.tv_sec = timeout;
-               if (copy_to_user(tsp, &ts, sizeof(ts))) {
-               sticky:
-                       /*
-                        * If an application puts its timeval in read-only
-                        * memory, we don't want the Linux-specific update to
-                        * the timeval to cause a fault after the select has
-                        * completed successfully. However, because we're not
-                        * updating the timeval, we can't restart the system
-                        * call.
-                        */
-                       if (ret == -ERESTARTNOHAND && timeout >= 0)
-                               ret = -EINTR;
-               }
-       }
+       ret = poll_select_copy_remaining(&end_time, tsp, 0, ret);
 
        return ret;
 }
-#endif /* TIF_RESTORE_SIGMASK */
+#endif /* HAVE_SET_RESTORE_SIGMASK */