Fedora kernel-2.6.17-1.2142_FC4 patched with stable patch-2.6.17.4-vs2.0.2-rc26.diff
[linux-2.6.git] / fs / select.c
index b80e7eb..8cd6dc3 100644 (file)
 #include <linux/personality.h> /* for STICKY_TIMEOUTS */
 #include <linux/file.h>
 #include <linux/fs.h>
+#include <linux/rcupdate.h>
+#include <linux/kallsyms.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;
-};
-
 struct poll_table_page {
        struct poll_table_page * next;
        struct poll_table_entry * entry;
@@ -63,13 +59,24 @@ 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)
+{
+       if (remove_wait_queue(entry->wait_address,&entry->wait) < 0)
+               print_symbol("bad poll-entry for %s", (unsigned long) entry->filp->f_op->poll);
+       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;
@@ -77,8 +84,7 @@ 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;
@@ -88,12 +94,14 @@ 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;
 
@@ -101,7 +109,7 @@ 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;
@@ -109,16 +117,21 @@ 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);
 }
 
 #define FDS_IN(fds, n)         (fds->in + n)
@@ -132,11 +145,13 @@ static int max_select_fd(unsigned long n, fd_set_bits *fds)
        unsigned long *open_fds;
        unsigned long set;
        int max;
+       struct fdtable *fdt;
 
        /* handle last in-complete long-word first */
        set = ~(~0UL << (n & (__NFDBITS-1)));
        n /= __NFDBITS;
-       open_fds = current->files->open_fds->fds_bits+n;
+       fdt = files_fdtable(current->files);
+       open_fds = fdt->open_fds->fds_bits+n;
        max = 0;
        if (set) {
                set &= BITS(fds, n);
@@ -176,16 +191,15 @@ get_max:
 #define POLLOUT_SET (POLLWRBAND | POLLWRNORM | POLLOUT | POLLERR)
 #define POLLEX_SET (POLLPRI)
 
-int do_select(int n, fd_set_bits *fds, long *timeout)
+int do_select(int n, fd_set_bits *fds, s64 *timeout)
 {
        struct poll_wqueues table;
        poll_table *wait;
        int retval, i;
-       long __timeout = *timeout;
 
-       spin_lock(&current->files->file_lock);
+       rcu_read_lock();
        retval = max_select_fd(n, fds);
-       spin_unlock(&current->files->file_lock);
+       rcu_read_unlock();
 
        if (retval < 0)
                return retval;
@@ -193,11 +207,12 @@ int do_select(int n, fd_set_bits *fds, long *timeout)
 
        poll_initwait(&table);
        wait = &table.pt;
-       if (!__timeout)
+       if (!*timeout)
                wait = NULL;
        retval = 0;
        for (;;) {
                unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp;
+               long __timeout;
 
                set_current_state(TASK_INTERRUPTIBLE);
 
@@ -207,7 +222,7 @@ int do_select(int n, fd_set_bits *fds, long *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++;
@@ -218,17 +233,18 @@ int do_select(int n, fd_set_bits *fds, long *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++;
@@ -252,35 +268,35 @@ int do_select(int n, fd_set_bits *fds, long *timeout)
                                *rexp = res_ex;
                }
                wait = NULL;
-               if (retval || !__timeout || signal_pending(current))
+               if (retval || !*timeout || signal_pending(current))
                        break;
                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;
+               }
                __timeout = schedule_timeout(__timeout);
+               if (*timeout >= 0)
+                       *timeout += __timeout;
        }
        __set_current_state(TASK_RUNNING);
 
        poll_freewait(&table);
 
-       /*
-        * Up-to-date the caller timeout.
-        */
-       *timeout = __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
@@ -292,41 +308,26 @@ static void select_bits_free(void *bits, int size)
 #define MAX_SELECT_SECONDS \
        ((unsigned long) (MAX_SCHEDULE_TIMEOUT / HZ)-1)
 
-asmlinkage long
-sys_select(int n, fd_set __user *inp, fd_set __user *outp, fd_set __user *exp, struct timeval __user *tvp)
+static int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
+                          fd_set __user *exp, s64 *timeout)
 {
        fd_set_bits fds;
-       char *bits;
-       long timeout;
-       int ret, size, max_fdset;
-
-       timeout = MAX_SCHEDULE_TIMEOUT;
-       if (tvp) {
-               time_t sec, usec;
-
-               if (!access_ok(VERIFY_READ, tvp, sizeof(*tvp))
-                   || __get_user(sec, &tvp->tv_sec)
-                   || __get_user(usec, &tvp->tv_usec)) {
-                       ret = -EFAULT;
-                       goto out_nofds;
-               }
-
-               ret = -EINVAL;
-               if (sec < 0 || usec < 0)
-                       goto out_nofds;
-
-               if ((unsigned long) sec < MAX_SELECT_SECONDS) {
-                       timeout = ROUND_UP(usec, 1000000/HZ);
-                       timeout += sec * (unsigned long) HZ;
-               }
-       }
+       void *bits;
+       int ret, max_fdset;
+       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_fdset = current->files->max_fdset;
+       rcu_read_lock();
+       fdt = files_fdtable(current->files);
+       max_fdset = fdt->max_fdset;
+       rcu_read_unlock();
        if (n > max_fdset)
                n = max_fdset;
 
@@ -335,17 +336,21 @@ sys_select(int n, fd_set __user *inp, fd_set __user *outp, fd_set __user *exp, s
         * 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)) ||
@@ -355,18 +360,7 @@ sys_select(int n, fd_set __user *inp, fd_set __user *outp, fd_set __user *exp, s
        zero_fd_set(n, fds.res_out);
        zero_fd_set(n, fds.res_ex);
 
-       ret = do_select(n, &fds, &timeout);
-
-       if (tvp && !(current->personality & STICKY_TIMEOUTS)) {
-               time_t sec = 0, usec = 0;
-               if (timeout) {
-                       sec = timeout / HZ;
-                       usec = timeout % HZ;
-                       usec *= (1000000/HZ);
-               }
-               put_user(sec, &tvp->tv_sec);
-               put_user(usec, &tvp->tv_usec);
-       }
+       ret = do_select(n, &fds, timeout);
 
        if (ret < 0)
                goto out;
@@ -383,11 +377,169 @@ sys_select(int n, fd_set __user *inp, fd_set __user *outp, fd_set __user *exp, s
                ret = -EFAULT;
 
 out:
-       select_bits_free(bits, size);
+       if (bits != stack_fds)
+               kfree(bits);
 out_nofds:
        return ret;
 }
 
+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 timeval tv;
+       int ret;
+
+       if (tvp) {
+               if (copy_from_user(&tv, tvp, sizeof(tv)))
+                       return -EFAULT;
+
+               if (tv.tv_sec < 0 || tv.tv_usec < 0)
+                       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) {
+               struct timeval rtv;
+
+               if (current->personality & STICKY_TIMEOUTS)
+                       goto sticky;
+               rtv.tv_usec = jiffies_to_usecs(do_div((*(u64*)&timeout), HZ));
+               rtv.tv_sec = timeout;
+               if (timeval_compare(&rtv, &tv) >= 0)
+                       rtv = tv;
+               if (copy_to_user(tvp, &rtv, sizeof(rtv))) {
+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;
+               }
+       }
+
+       return ret;
+}
+
+#ifdef TIF_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;
+       int ret;
+
+       if (tsp) {
+               if (copy_from_user(&ts, tsp, sizeof(ts)))
+                       return -EFAULT;
+
+               if (ts.tv_sec < 0 || ts.tv_nsec < 0)
+                       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) {
+               /* XXX: Don't preclude handling different sized sigset_t's.  */
+               if (sigsetsize != sizeof(sigset_t))
+                       return -EINVAL;
+               if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
+                       return -EFAULT;
+
+               sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
+               sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
+       }
+
+       ret = core_sys_select(n, inp, outp, exp, &timeout);
+
+       if (tsp) {
+               struct timespec rts;
+
+               if (current->personality & STICKY_TIMEOUTS)
+                       goto sticky;
+               rts.tv_nsec = jiffies_to_usecs(do_div((*(u64*)&timeout), HZ)) *
+                                               1000;
+               rts.tv_sec = timeout;
+               if (timespec_compare(&rts, &ts) >= 0)
+                       rts = ts;
+               if (copy_to_user(tsp, &rts, sizeof(rts))) {
+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;
+               }
+       }
+
+       if (ret == -ERESTARTNOHAND) {
+               /*
+                * Don't restore the signal mask yet. Let do_signal() deliver
+                * the signal on the way back to userspace, before the signal
+                * mask is restored.
+                */
+               if (sigmask) {
+                       memcpy(&current->saved_sigmask, &sigsaved,
+                                       sizeof(sigsaved));
+                       set_thread_flag(TIF_RESTORE_SIGMASK);
+               }
+       } else if (sigmask)
+               sigprocmask(SIG_SETMASK, &sigsaved, NULL);
+
+       return ret;
+}
+
+/*
+ * Most architectures can't handle 7-argument syscalls. So we provide a
+ * 6-argument version where the sixth argument is a pointer to a structure
+ * which has a pointer to the sigset_t itself followed by a size_t containing
+ * the sigset size.
+ */
+asmlinkage long sys_pselect6(int n, fd_set __user *inp, fd_set __user *outp,
+       fd_set __user *exp, struct timespec __user *tsp, void __user *sig)
+{
+       size_t sigsetsize = 0;
+       sigset_t __user *up = NULL;
+
+       if (sig) {
+               if (!access_ok(VERIFY_READ, sig, sizeof(void *)+sizeof(size_t))
+                   || __get_user(up, (sigset_t __user * __user *)sig)
+                   || __get_user(sigsetsize,
+                               (size_t __user *)(sig+sizeof(void *))))
+                       return -EFAULT;
+       }
+
+       return sys_pselect7(n, inp, outp, exp, tsp, up, sigsetsize);
+}
+#endif /* TIF_RESTORE_SIGMASK */
+
 struct poll_list {
        struct poll_list *next;
        int len;
@@ -410,14 +562,15 @@ static void do_pollfd(unsigned int num, struct pollfd * fdpage,
                fdp = fdpage+i;
                fd = fdp->fd;
                if (fd >= 0) {
-                       struct file * file = fget(fd);
+                       int fput_needed;
+                       struct 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 &= fdp->events | POLLERR | POLLHUP;
-                               fput(file);
+                               fput_light(file, fput_needed);
                        }
                        if (mask) {
                                *pwait = NULL;
@@ -429,16 +582,19 @@ static void do_pollfd(unsigned int num, struct pollfd * fdpage,
 }
 
 static int do_poll(unsigned int nfds,  struct poll_list *list,
-                       struct poll_wqueues *wait, long timeout)
+                  struct poll_wqueues *wait, s64 *timeout)
 {
        int count = 0;
        poll_table* pt = &wait->pt;
 
-       if (!timeout)
+       /* Optimise the no-wait case */
+       if (!(*timeout))
                pt = NULL;
  
        for (;;) {
                struct poll_list *walk;
+               long __timeout;
+
                set_current_state(TASK_INTERRUPTIBLE);
                walk = list;
                while(walk != NULL) {
@@ -446,37 +602,61 @@ static int do_poll(unsigned int nfds,  struct poll_list *list,
                        walk = walk->next;
                }
                pt = NULL;
-               if (count || !timeout || signal_pending(current))
+               if (count || !*timeout || signal_pending(current))
                        break;
                count = wait->error;
                if (count)
                        break;
-               timeout = schedule_timeout(timeout);
+
+               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;
+               }
+
+               __timeout = schedule_timeout(__timeout);
+               if (*timeout >= 0)
+                       *timeout += __timeout;
        }
        __set_current_state(TASK_RUNNING);
        return count;
 }
 
-asmlinkage long sys_poll(struct pollfd __user * ufds, unsigned int nfds, long 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, s64 *timeout)
 {
        struct poll_wqueues table;
        int fdcount, err;
        unsigned int i;
        struct poll_list *head;
        struct poll_list *walk;
+       struct fdtable *fdt;
+       int max_fdset;
+       /* 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 *stack_pp = NULL;
 
        /* Do a sanity check on nfds ... */
-       if (nfds > current->files->max_fdset && nfds > OPEN_MAX)
+       rcu_read_lock();
+       fdt = files_fdtable(current->files);
+       max_fdset = fdt->max_fdset;
+       rcu_read_unlock();
+       if (nfds > max_fdset && nfds > OPEN_MAX)
                return -EINVAL;
 
-       if (timeout) {
-               /* Careful about overflow in the intermediate values */
-               if ((unsigned long) timeout < MAX_SCHEDULE_TIMEOUT / HZ)
-                       timeout = (unsigned long)(timeout*HZ+999)/1000+1;
-               else /* Negative or overflow */
-                       timeout = MAX_SCHEDULE_TIMEOUT;
-       }
-
        poll_initwait(&table);
 
        head = NULL;
@@ -485,14 +665,23 @@ asmlinkage long sys_poll(struct pollfd __user * ufds, unsigned int nfds, long ti
        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)
-                       goto out_fds;
+               int num, size;
+               if (stack_pp == NULL)
+                       num = N_STACK_PPS;
+               else
+                       num = POLLFD_PER_PAGE;
+               if (num > i)
+                       num = i;
+               size = sizeof(struct poll_list) + sizeof(struct pollfd)*num;
+               if (!stack_pp)
+                       stack_pp = pp = (struct poll_list *)stack_pps;
+               else {
+                       pp = kmalloc(size, GFP_KERNEL);
+                       if (!pp)
+                               goto out_fds;
+               }
                pp->next=NULL;
-               pp->len = (i>POLLFD_PER_PAGE?POLLFD_PER_PAGE:i);
+               pp->len = num;
                if (head == NULL)
                        head = pp;
                else
@@ -500,12 +689,13 @@ asmlinkage long sys_poll(struct pollfd __user * ufds, unsigned int nfds, long ti
 
                walk = pp;
                if (copy_from_user(pp->entries, ufds + nfds-i, 
-                               sizeof(struct pollfd)*pp->len)) {
+                               sizeof(struct pollfd)*num)) {
                        err = -EFAULT;
                        goto out_fds;
                }
                i -= pp->len;
        }
+
        fdcount = do_poll(nfds, head, &table, timeout);
 
        /* OK, now copy the revents fields back to user space. */
@@ -528,9 +718,110 @@ out_fds:
        walk = head;
        while(walk!=NULL) {
                struct poll_list *pp = walk->next;
-               kfree(walk);
+               if (walk != stack_pp)
+                       kfree(walk);
                walk = pp;
        }
        poll_freewait(&table);
        return err;
 }
+
+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);
+       }
+
+       return do_sys_poll(ufds, nfds, &timeout_jiffies);
+}
+
+#ifdef TIF_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;
+       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;
+               }
+       }
+
+       if (sigmask) {
+               /* XXX: Don't preclude handling different sized sigset_t's.  */
+               if (sigsetsize != sizeof(sigset_t))
+                       return -EINVAL;
+               if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
+                       return -EFAULT;
+
+               sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
+               sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
+       }
+
+       ret = do_sys_poll(ufds, nfds, &timeout);
+
+       /* We can restart this syscall, usually */
+       if (ret == -EINTR) {
+               /*
+                * Don't restore the signal mask yet. Let do_signal() deliver
+                * the signal on the way back to userspace, before the signal
+                * mask is restored.
+                */
+               if (sigmask) {
+                       memcpy(&current->saved_sigmask, &sigsaved,
+                                       sizeof(sigsaved));
+                       set_thread_flag(TIF_RESTORE_SIGMASK);
+               }
+               ret = -ERESTARTNOHAND;
+       } else if (sigmask)
+               sigprocmask(SIG_SETMASK, &sigsaved, NULL);
+
+       if (tsp && timeout >= 0) {
+               struct timespec rts;
+
+               if (current->personality & STICKY_TIMEOUTS)
+                       goto sticky;
+               /* Yes, we know it's actually an s64, but it's also positive. */
+               rts.tv_nsec = jiffies_to_usecs(do_div((*(u64*)&timeout), HZ)) *
+                                               1000;
+               rts.tv_sec = timeout;
+               if (timespec_compare(&rts, &ts) >= 0)
+                       rts = ts;
+               if (copy_to_user(tsp, &rts, sizeof(rts))) {
+               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;
+               }
+       }
+
+       return ret;
+}
+#endif /* TIF_RESTORE_SIGMASK */