PATCH (2.4.5): /dev/poll support (3rd time lucky)

PATCH (2.4.5): /dev/poll support (3rd time lucky)

[Zarjazz]

[-- Attachment #1: Type: text/plain, Size: 803 bytes --]

Not my day it seems ! Hopefully I remembered to attach the file this time :)

--

Hi,
    this patch adds Solaris 7/8 like /dev/poll support to the kernel.

I can claim no real credit for this as basically this is a fixed version of
a patch available from http://www.citi.umich.edu/projects/linux-scalability/
to compile correctly with 2.4.5 that only seemed to work with the 2.3.x
devel branch. The reason for this is so I can compile & test an application
on my home linux pc when I'm not around my nice work Solaris boxes :)

Please note, I have not got the knowledge of kernel development to know if
this patch is broken or badly written. It may be bugged and/or worse than
the standard poll() call but my application works so I'll leave profiling
etc to people more knowledgable than me.

Vince.



[-- Attachment #2: devpoll-2.4.5.patch --]
[-- Type: application/octet-stream, Size: 31617 bytes --]

diff -rNu linux.orig/drivers/char/Config.in linux/drivers/char/Config.in
--- linux.orig/drivers/char/Config.in	Wed Mar  7 03:44:34 2001
+++ linux/drivers/char/Config.in	Wed Jun 27 16:41:00 2001
@@ -158,6 +158,7 @@
 
 dep_tristate 'Intel i8x0 Random Number Generator support' CONFIG_INTEL_RNG $CONFIG_PCI
 tristate '/dev/nvram support' CONFIG_NVRAM
+tristate '/dev/poll support' CONFIG_DEVPOLL
 tristate 'Enhanced Real Time Clock Support' CONFIG_RTC
 if [ "$CONFIG_IA64" = "y" ]; then
    bool 'EFI Real Time Clock Services' CONFIG_EFI_RTC
diff -rNu linux.orig/drivers/char/Makefile linux/drivers/char/Makefile
--- linux.orig/drivers/char/Makefile	Wed May 16 18:27:02 2001
+++ linux/drivers/char/Makefile	Wed Jun 27 16:43:07 2001
@@ -170,6 +170,7 @@
 obj-$(CONFIG_PC110_PAD) += pc110pad.o
 obj-$(CONFIG_RTC) += rtc.o
 obj-$(CONFIG_EFI_RTC) += efirtc.o
+obj-$(CONFIG_DEVPOLL) += devpoll.o
 ifeq ($(CONFIG_PPC),)
   obj-$(CONFIG_NVRAM) += nvram.o
 endif
diff -rNu linux.orig/drivers/char/devpoll.c linux/drivers/char/devpoll.c
--- linux.orig/drivers/char/devpoll.c	Thu Jan  1 01:00:00 1970
+++ linux/drivers/char/devpoll.c	Wed Jun 27 18:55:30 2001
@@ -0,0 +1,756 @@
+/*
+ * /dev/poll
+ * by Niels Provos <provos@citi.umich.edu>
+ *
+ * provides poll() support via /dev/poll as in Solaris.
+ *
+ * Linux 2.3.x port by Michal Ostrowski
+ * Linux 2.4.x patches by Vincent Sweeney <v.sweeney@dexterus.com>
+ */
+
+#include <linux/module.h>
+#include <linux/init.h>
+#include <linux/kernel.h>
+#include <linux/sched.h>
+#include <linux/file.h>
+#include <linux/signal.h>
+#include <linux/errno.h>
+#include <linux/mm.h>
+#include <linux/malloc.h>
+#include <linux/vmalloc.h>
+#include <linux/poll.h>
+#include <linux/miscdevice.h>
+#include <linux/random.h>
+#include <linux/smp_lock.h>
+#include <linux/wrapper.h>
+
+#include <linux/devpoll.h>
+
+#include <asm/uaccess.h>
+#include <asm/system.h>
+#include <asm/io.h>
+
+/*#define DEBUG 1 */
+#ifdef DEBUG
+#define DPRINTK(x)	printk x
+#define DNPRINTK(n,x)	if (n <= DEBUG) printk x
+#else
+#define DPRINTK(x)
+#define DNPRINTK(n,x)
+#endif
+
+/* Various utility functions */
+
+#define DEFAULT_POLLMASK (POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM)
+
+/* Do dynamic hashing */
+
+#define INITIAL_BUCKET_BITS 6
+#define MAX_BUCKET_BITS 16
+#define RESIZE_LENGTH	2
+
+static void free_pg_vec(struct devpoll *dp);
+
+/* Initalize the hash table */
+
+int
+dp_init(struct devpoll *dp)
+{
+	int i;
+	int num_buckets;
+	DNPRINTK(3, (KERN_INFO "/dev/poll: dp_init\n"));
+
+	dp->dp_lock = RW_LOCK_UNLOCKED;
+	dp->dp_entries = 0;
+	dp->dp_max = 0;
+	dp->dp_avg = dp->dp_count = 0;
+	dp->dp_cached = dp->dp_calls = 0;
+	dp->dp_bucket_bits = INITIAL_BUCKET_BITS;
+	dp->dp_bucket_mask = (1 << INITIAL_BUCKET_BITS) - 1;
+
+	num_buckets = (dp->dp_bucket_mask + 1);
+	dp->dp_tab = kmalloc(num_buckets * sizeof (struct list_head),
+			     GFP_KERNEL);
+
+	if (!dp->dp_tab)
+		return -ENOMEM;
+
+	for (i = 0; i < num_buckets; i++) {
+		INIT_LIST_HEAD(&dp->dp_tab[i]);
+	}
+
+	return (0);
+}
+
+int
+dp_resize(struct devpoll *dp)
+{
+	u_int16_t new_mask, old_mask;
+	int i;
+	struct list_head *new_tab, *old_tab;
+	struct dp_fd *dpfd;
+	unsigned long flags;
+	int num_buckets;
+
+	old_mask = dp->dp_bucket_mask;
+	new_mask = (old_mask + 1) * 2 - 1;
+	num_buckets = new_mask + 1;
+
+	DPRINTK((KERN_INFO "/dev/poll: resize %d -> %d\n", old_mask, new_mask));
+
+	new_tab = kmalloc(num_buckets * sizeof (struct list_head), GFP_KERNEL);
+	if (!new_tab)
+		return -ENOMEM;
+
+	for (i = 0; i < num_buckets; i++) {
+		INIT_LIST_HEAD(&new_tab[i]);
+	}
+
+	old_tab = dp->dp_tab;
+
+	/* Rehash all entries */
+	write_lock_irqsave(&dp->dp_lock, flags);
+	for (i = 0; i <= old_mask; i++) {
+		while (!list_empty(&old_tab[i])) {
+			dpfd = list_entry(old_tab[i].next, struct dp_fd, next);
+			list_del(&dpfd->next);
+			INIT_LIST_HEAD(&dpfd->next);
+			list_add(&dpfd->next,
+				 &new_tab[dpfd->pfd.fd & new_mask]);
+		}
+	}
+
+	dp->dp_tab = new_tab;
+	dp->dp_bucket_bits++;
+	dp->dp_bucket_mask = new_mask;
+	write_unlock_irqrestore(&dp->dp_lock, flags);
+
+	kfree(old_tab);
+
+	return (0);
+}
+
+int
+dp_insert(struct devpoll *dp, struct pollfd *pfd)
+{
+	struct dp_fd *dpfd;
+	u_int16_t bucket = pfd->fd & dp->dp_bucket_mask;
+	unsigned long flags;
+	struct file *file;
+
+	dpfd = kmalloc(sizeof (struct dp_fd), GFP_KERNEL);
+	if (!dpfd)
+		return -ENOMEM;
+
+	dpfd->flags = 0;
+	set_bit(DPH_DIRTY, &dpfd->flags);
+	dpfd->pfd = *pfd;
+	dpfd->pfd.revents = 0;
+	INIT_LIST_HEAD(&dpfd->next);
+
+	write_lock_irqsave(&dp->dp_lock, flags);
+
+	list_add(&dpfd->next, &dp->dp_tab[bucket]);
+
+	file = fcheck(pfd->fd);
+	if (file != NULL) {
+		write_lock(&(file)->f_dplock);
+		poll_backmap(pfd->fd, dpfd, &(file)->f_backmap);
+		write_unlock(&(file)->f_dplock);
+		set_bit(DPH_BACKMAP, &(dpfd)->flags);
+	}
+	write_unlock_irqrestore(&dp->dp_lock, flags);
+
+	dp->dp_entries++;
+	/* Check if we need to resize the hash table */
+	if ((dp->dp_entries >> dp->dp_bucket_bits) > RESIZE_LENGTH &&
+	    dp->dp_bucket_bits < MAX_BUCKET_BITS)
+		dp_resize(dp);
+
+	return (0);
+}
+
+struct dp_fd *
+dp_find(struct devpoll *dp, int fd)
+{
+	struct dp_fd *dpfd = NULL;
+	struct list_head *lh;
+	u_int16_t bucket = fd & dp->dp_bucket_mask;
+
+	read_lock(&dp->dp_lock);
+	list_for_each(lh, &dp->dp_tab[bucket]) {
+		dpfd = list_entry(lh, struct dp_fd, next);
+		if (dpfd->pfd.fd == fd)
+			break;
+		dpfd = NULL;
+	}
+
+	read_unlock(&dp->dp_lock);
+	DNPRINTK(2, (KERN_INFO "dp_find: %d -> %p\n", fd, dpfd));
+
+	return dpfd;
+}
+
+void
+dp_delete(struct devpoll *dp, struct dp_fd *dpfd)
+{
+	unsigned long flags;
+	int fd;
+	struct file *filp;
+
+	write_lock_irqsave(&dp->dp_lock, flags);
+	list_del(&dpfd->next);
+
+	INIT_LIST_HEAD(&dpfd->next);
+
+	/* Remove backmaps if necessary */
+	if (current->files) {
+		fd = dpfd->pfd.fd;
+		filp = fcheck(fd);
+
+		if (test_bit(DPH_BACKMAP, &dpfd->flags) &&
+		    filp && filp->f_backmap) {
+			write_lock(&filp->f_dplock);
+			poll_remove_backmap(&filp->f_backmap, fd,
+					    current->files);
+			write_unlock(&filp->f_dplock);
+		}
+	}
+	write_unlock_irqrestore(&dp->dp_lock, flags);
+
+	kfree(dpfd);
+
+	dp->dp_entries--;
+}
+
+void
+dp_free(struct devpoll *dp)
+{
+	int i;
+	struct dp_fd *dpfd = NULL;
+
+	lock_kernel();
+	for (i = 0; i <= dp->dp_bucket_mask; i++) {
+		while (!list_empty(&dp->dp_tab[i])) {
+			dpfd =
+			    list_entry(dp->dp_tab[i].next, struct dp_fd, next);
+			dp_delete(dp, dpfd);
+		}
+	}
+	unlock_kernel();
+
+	kfree(dp->dp_tab);
+}
+
+/*
+ * poll the fds that we keep in our state, return after we reached
+ * max changed fds or are done.
+ * XXX - I do not like how the wait table stuff is done.
+ */
+
+int
+dp_poll(struct devpoll *dp, int max, poll_table * wait,
+	long timeout, struct pollfd *rfds, int usemmap)
+{
+	int count = 0;
+	lock_kernel();
+	read_lock(&dp->dp_lock);
+	for (;;) {
+		unsigned int j = 0;
+		struct dp_fd *dpfd = NULL;
+		struct pollfd *fdpnt, pfd;
+		struct file *file;
+
+		set_current_state(TASK_INTERRUPTIBLE);
+		for (j = 0; (j <= dp->dp_bucket_mask) && count < max; j++) {
+			struct list_head *lh;
+			list_for_each(lh, &dp->dp_tab[j]) {
+
+				int fd;
+				unsigned int mask = 0;
+				dpfd = list_entry(lh, struct dp_fd, next);
+
+				if (count >= max) {
+					break;
+				}
+
+				fdpnt = &dpfd->pfd;
+				fd = fdpnt->fd;
+
+				/* poll_wait increments f_count if needed */
+				file = fcheck(fd);
+				if (file == NULL) {
+					/* Got to move backward first;
+					 * dp_delete will remove lh from
+					 * the list otherwise
+					 */
+					lh = lh->prev;
+					dp_delete(dp, dpfd);
+					dpfd = NULL;
+					continue;
+				}
+
+				mask = fdpnt->revents;
+				if (test_and_clear_bit(DPH_DIRTY,
+						       &dpfd->flags) ||
+				    wait != NULL || (mask & fdpnt->events)) {
+
+					mask = DEFAULT_POLLMASK;
+					if (file->f_op && file->f_op->poll)
+						mask =
+						    file->f_op->poll(file,
+								     wait);
+					/* if POLLHINT not supported by file
+					 * then set bit to dirty ---
+					 * must poll this file every time,
+					 * otherwise bit will be set by
+					 * calls to dp_add_hint
+					 */
+					if (!(mask & POLLHINT))
+						set_bit(DPH_DIRTY,
+							&dpfd->flags);
+					fdpnt->revents = mask;
+				} else
+					dp->dp_cached++;
+
+				dp->dp_calls++;
+
+				mask &= fdpnt->events | POLLERR | POLLHUP;
+				if (mask) {
+					wait = NULL;
+					count++;
+
+					if (usemmap) {
+						*rfds = *fdpnt;
+						rfds->revents = mask;
+					} else {
+						pfd = *fdpnt;
+						pfd.revents = mask;
+						__copy_to_user(rfds, &pfd,
+							       sizeof (struct
+								       pollfd));
+					}
+
+					rfds++;
+				}
+			}
+		}
+
+		wait = NULL;
+		if (count || !timeout || signal_pending(current))
+			break;
+		read_unlock(&dp->dp_lock);
+		timeout = schedule_timeout(timeout);
+		read_lock(&dp->dp_lock);
+	}
+	set_current_state(TASK_RUNNING);
+	read_unlock(&dp->dp_lock);
+	unlock_kernel();
+
+	if (!count && signal_pending(current))
+		return -EINTR;
+
+	return count;
+}
+
+/*
+ * close a /dev/poll
+ */
+
+static int
+close_devpoll(struct inode *inode, struct file *file)
+{
+	struct devpoll *dp = file->private_data;
+
+	DNPRINTK(1,
+		 (KERN_INFO "close /dev/poll, max: %d, avg: %d(%d/%d) %d/%d\n",
+		  dp->dp_max, dp->dp_avg / dp->dp_count, dp->dp_avg,
+		  dp->dp_count, dp->dp_cached, dp->dp_calls));
+
+	/* free allocated memory */
+	if (dp->dp_memvec)
+		free_pg_vec(dp);
+
+	/* Free the hash table */
+	dp_free(dp);
+
+	kfree(dp);
+
+	MOD_DEC_USE_COUNT;
+	return 0;
+}
+
+/*
+ * open a /dev/poll
+ */
+
+static int
+open_devpoll(struct inode *inode, struct file *file)
+{
+	struct devpoll *dp;
+	int r;
+
+	/* allocated state */
+	dp = kmalloc(sizeof (struct devpoll), GFP_KERNEL);
+	if (dp == NULL)
+		return -ENOMEM;
+
+	memset(dp, 0, sizeof (struct devpoll));
+	if ((r = dp_init(dp))) {
+		kfree(dp);
+		return r;
+	}
+
+	file->private_data = dp;
+
+	MOD_INC_USE_COUNT;
+
+	DNPRINTK(3, (KERN_INFO "open /dev/poll\n"));
+
+	return 0;
+}
+
+/*
+ * write to /dev/poll:
+ * a user writes struct pollfds and we add them to our list, or remove
+ * them if (events & POLLREMOVE) is true
+ */
+
+static int
+write_devpoll(struct file *file, const char *buffer, size_t count,
+	      loff_t * ppos)
+{
+	int r, rcount;
+	struct devpoll *dp = file->private_data;
+	struct pollfd pfd;
+	struct dp_fd *dpfd;
+#ifdef DEBUG
+	int add = 0, delete = 0, change = 0;
+#endif
+
+	DNPRINTK(3, (KERN_INFO "write /dev/poll %i\n", count));
+
+	if (count % sizeof (struct pollfd))
+		return -EINVAL;
+
+	if ((r = verify_area(VERIFY_READ, buffer, count)))
+		return r;
+
+	rcount = count;
+
+	lock_kernel();
+
+	while (count > 0) {
+		__copy_from_user(&pfd, buffer, sizeof (pfd));	/* no check */
+
+		dpfd = dp_find(dp, pfd.fd);
+
+		if (pfd.fd >= current->files->max_fds ||
+		    current->files->fd[pfd.fd] == NULL) {
+			/* Be tolerant, maybe the close happened already */
+			pfd.events = POLLREMOVE;
+		}
+		/* See if we need to remove the file descriptor.  If it
+		 * already exists OR the event fields, otherwise insert
+		 */
+		if (pfd.events & POLLREMOVE) {
+			if (dpfd)
+				dp_delete(dp, dpfd);
+#ifdef DEBUG
+			delete++;
+#endif
+		} else if (dpfd) {
+			/* XXX dpfd->pfd.events |= pfd.events; */
+			dpfd->pfd.events = pfd.events;
+#ifdef DEBUG
+			change++;
+#endif
+		} else {
+			dp_insert(dp, &pfd);
+#ifdef DEBUG
+			add++;
+#endif
+		}
+
+		buffer += sizeof (pfd);
+		count -= sizeof (pfd);
+	}
+
+	unlock_kernel();
+
+	if (dp->dp_max < dp->dp_entries) {
+		dp->dp_max = dp->dp_entries;
+		DNPRINTK(2, (KERN_INFO "/dev/poll: new max %d\n", dp->dp_max));
+	}
+
+	DNPRINTK(3, (KERN_INFO "write /dev/poll: %d entries (%d/%d/%d)\n",
+		     dp->dp_entries, add, delete, change));
+
+	return (rcount);
+}
+
+static int
+ioctl_devpoll(struct inode *inode, struct file *file,
+	      unsigned int cmd, unsigned long arg)
+{
+	struct devpoll *dp = file->private_data;
+	unsigned mapsize = 0;
+	unsigned num_pages = 0;
+	int i = 0;
+	switch (cmd) {
+	case DP_ALLOC:
+		if (arg > current->rlim[RLIMIT_NOFILE].rlim_cur)
+			return -EINVAL;
+		if (dp->dp_mmap)
+			return -EPERM;
+
+		mapsize = DP_MMAP_SIZE(arg);
+		num_pages = (PAGE_ALIGN(mapsize) >> PAGE_SHIFT);
+
+		dp->dp_memvec = kmalloc(num_pages * sizeof (unsigned long *),
+					GFP_KERNEL);
+
+		if (dp->dp_memvec == NULL)
+			return -EINVAL;
+
+		memset(dp->dp_memvec, 0, num_pages * sizeof (unsigned long *));
+
+		for (i = 0; i < num_pages; ++i) {
+			struct page *page, *page_end;
+
+			dp->dp_memvec[i] =
+			    (u_char *) __get_free_pages(GFP_KERNEL, 0);
+			if (!dp->dp_memvec[i]) {
+				free_pg_vec(dp);
+				return -ENOMEM;
+			}
+
+			page_end =
+			    virt_to_page(dp->dp_memvec[i] + PAGE_SIZE - 1);
+			for (page = virt_to_page(dp->dp_memvec[i]);
+			     page <= page_end; page++)
+				set_bit(PG_reserved, &page->flags);
+
+			++dp->dp_numvec;
+		}
+
+		dp->dp_nfds = arg;
+
+		DPRINTK((KERN_INFO "allocated %d pollfds\n", dp->dp_nfds));
+
+		return 0;
+	case DP_FREE:
+		if (atomic_read(&dp->dp_mmapped))
+			return -EBUSY;
+
+		if (dp->dp_memvec[i]) {
+			free_pg_vec(dp);
+		}
+
+		DPRINTK((KERN_INFO "freed %d pollfds\n", dp->dp_nfds));
+		dp->dp_nfds = 0;
+
+		return 0;
+	case DP_ISPOLLED:{
+			struct pollfd pfd;
+			struct dp_fd *dpfd;
+
+			if (copy_from_user(&pfd, (void *) arg, sizeof (pfd)))
+				return -EFAULT;
+			dpfd = dp_find(dp, pfd.fd);
+			if (dpfd == NULL)
+				return (0);
+
+			/* We poll this fd, return the evens we poll on */
+			pfd.events = dpfd->pfd.events;
+			pfd.revents = 0;
+
+			if (copy_to_user((void *) arg, &pfd, sizeof (pfd)))
+				return -EFAULT;
+			return (1);
+		}
+	case DP_POLL:{
+			struct dvpoll dopoll;
+			int nfds, usemmap = 0;
+			unsigned long timeout;
+			poll_table wait;
+			struct pollfd *rpfds = NULL;
+
+			if (copy_from_user
+			    (&dopoll, (void *) arg, sizeof (dopoll)))
+				return -EFAULT;
+
+			/* We do not need to check this value, its user space */
+			nfds = dopoll.dp_nfds;
+			if (nfds <= 0)
+				return -EINVAL;
+
+			if (dopoll.dp_fds == NULL) {
+				if (dp->dp_mmap == NULL)
+					return -EINVAL;
+				rpfds = (struct pollfd *) dp->dp_mmap;
+				usemmap = 1;
+			} else {
+				rpfds = dopoll.dp_fds;
+				if (verify_area(VERIFY_WRITE, rpfds,
+						nfds * sizeof (struct pollfd)))
+					return -EFAULT;
+				usemmap = 0;
+			}
+
+			timeout = dopoll.dp_timeout;
+			if (timeout) {
+				/* Careful about overflow in the intermediate values */
+				if ((unsigned long) timeout <
+				    MAX_SCHEDULE_TIMEOUT / HZ)
+					timeout =
+					    (timeout * HZ + 999) / 1000 + 1;
+				else	/* Negative or overflow */
+					timeout = MAX_SCHEDULE_TIMEOUT;
+			}
+
+			/* Initalize wait table */
+			poll_initwait(&wait);
+
+			nfds =
+			    dp_poll(dp, nfds, &wait, timeout, rpfds, usemmap);
+
+			DNPRINTK(2,
+				 (KERN_INFO "poll time %ld -> %d\n", timeout,
+				  nfds));
+
+			poll_freewait(&wait);
+
+			dp->dp_avg += dp->dp_entries;
+			dp->dp_count++;
+
+			return nfds;
+		}
+	default:
+		DPRINTK((KERN_INFO "ioctl(%x) /dev/poll\n", cmd));
+		break;
+	}
+
+	return -EINVAL;
+}
+
+static void
+free_pg_vec(struct devpoll *dp)
+{
+	int i;
+
+	for (i = 0; i < dp->dp_numvec; i++) {
+		if (dp->dp_memvec[i]) {
+			struct page *page, *page_end;
+
+			page_end =
+			    virt_to_page(dp->dp_memvec[i] + PAGE_SIZE - 1);
+			for (page = virt_to_page(dp->dp_memvec[i]);
+			     page <= page_end; page++)
+				clear_bit(PG_reserved, &page->flags);
+
+			free_pages((unsigned) dp->dp_memvec[i], 0);
+		}
+	}
+	kfree(dp->dp_memvec);
+	dp->dp_numvec = 0;
+}
+
+static void
+devpoll_mm_open(struct vm_area_struct *vma)
+{
+	struct file *file = vma->vm_file;
+	struct devpoll *dp = file->private_data;
+	if (dp)
+		atomic_inc(&dp->dp_mmapped);
+}
+
+static void
+devpoll_mm_close(struct vm_area_struct *vma)
+{
+	struct file *file = vma->vm_file;
+	struct devpoll *dp = file->private_data;
+	if (dp)
+		atomic_dec(&dp->dp_mmapped);
+}
+
+static struct vm_operations_struct devpoll_mmap_ops = {
+	open:devpoll_mm_open,
+	close:devpoll_mm_close,
+};
+
+/*
+ * mmap shared memory.  the first half is an array  of struct pollfd,
+ * followed by an array of ints to indicate which file descriptors
+ * changed status.
+ */
+
+static int
+mmap_devpoll(struct file *file, struct vm_area_struct *vma)
+{
+	struct devpoll *dp = file->private_data;
+	unsigned long start;	/* Evil type to remap_page_range */
+	int i = 0;
+	int num_pages = 0;
+	size_t size, mapsize;
+
+	DPRINTK((KERN_INFO "mmap /dev/poll: %lx %lx\n",
+		 vma->vm_start, vma->vm_pgoff << PAGE_SHIFT));
+
+	if ((vma->vm_pgoff << PAGE_SHIFT) != 0)
+		return -EINVAL;
+
+	/* Calculate how much memory we can map */
+	size = PAGE_ALIGN(DP_MMAP_SIZE(dp->dp_nfds));
+	mapsize = PAGE_ALIGN(vma->vm_end - vma->vm_start);
+	num_pages = mapsize >> PAGE_SHIFT;
+
+	/* Check if the requested size is within our size */
+	if (mapsize > dp->dp_numvec << PAGE_SHIFT)
+		return -EINVAL;
+
+	start = vma->vm_start;
+	atomic_set(&dp->dp_mmapped, 1);
+	for (i = 0; i < num_pages; ++i) {
+		if (remap_page_range(start, __pa(dp->dp_memvec[i]),
+				     PAGE_SIZE, vma->vm_page_prot))
+			return -EINVAL;
+		start += PAGE_SIZE;
+	}
+	dp->dp_mmap = (u_char *) vma->vm_start;
+	vma->vm_ops = &devpoll_mmap_ops;
+
+	DPRINTK((KERN_INFO "mmap /dev/poll: %lx %x\n", page, mapsize));
+	return 0;
+}
+
+struct file_operations devpoll_fops = {
+	write:write_devpoll,
+	ioctl:ioctl_devpoll,
+	mmap:mmap_devpoll,
+	open:open_devpoll,
+	release:close_devpoll
+};
+
+static struct miscdevice devpoll = {
+	DEVPOLL_MINOR, "devpoll", &devpoll_fops
+};
+
+int __init
+devpoll_init(void)
+{
+	printk(KERN_INFO "/dev/poll driver installed.\n");
+	misc_register(&devpoll);
+
+	return 0;
+}
+
+module_init(devpoll_init);
+#ifdef MODULE
+
+void
+cleanup_module(void)
+{
+	misc_deregister(&devpoll);
+}
+#endif
diff -rNu linux.orig/fs/file_table.c linux/fs/file_table.c
--- linux.orig/fs/file_table.c	Wed Apr 18 19:49:12 2001
+++ linux/fs/file_table.c	Wed Jun 27 16:49:49 2001
@@ -11,6 +11,7 @@
 #include <linux/init.h>
 #include <linux/module.h>
 #include <linux/smp_lock.h>
+#include <linux/spinlock.h>
 
 /* sysctl tunables... */
 struct files_stat_struct files_stat = {0, 0, NR_FILE};
@@ -45,6 +46,7 @@
 		f->f_version = ++event;
 		f->f_uid = current->fsuid;
 		f->f_gid = current->fsgid;
+		rwlock_init(&f->f_dplock);
 		list_add(&f->f_list, &anon_list);
 		file_list_unlock();
 		return f;
diff -rNu linux.orig/fs/open.c linux/fs/open.c
--- linux.orig/fs/open.c	Fri Feb  9 19:29:44 2001
+++ linux/fs/open.c	Wed Jun 27 18:01:15 2001
@@ -14,6 +14,8 @@
 #include <linux/module.h>
 #include <linux/slab.h>
 #include <linux/tty.h>
+#include <linux/poll.h>
+#include <linux/devpoll.h>
 
 #include <asm/uaccess.h>
 
@@ -802,6 +805,14 @@
 		retval = filp->f_op->flush(filp);
 		unlock_kernel();
 	}
+
+	if (filp->f_backmap) {
+		unsigned long flags;
+		write_lock_irqsave(&filp->f_dplock,flags);
+		poll_clean_backmap(&filp->f_backmap);
+		write_unlock_irqrestore(&filp->f_dplock,flags);
+	}
+
 	fcntl_dirnotify(0, filp, 0);
 	locks_remove_posix(filp, id);
 	fput(filp);
@@ -828,6 +839,14 @@
 	FD_CLR(fd, files->close_on_exec);
 	__put_unused_fd(files, fd);
 	write_unlock(&files->file_lock);
+
+	if (filp->f_backmap) {
+		unsigned long flags;
+		write_lock_irqsave(&filp->f_dplock,flags);
+		poll_remove_backmap(&filp->f_backmap,fd, files);
+		write_unlock_irqrestore(&filp->f_dplock,flags);
+	}
+	
 	return filp_close(filp, files);
 
 out_unlock:
diff -rNu linux.orig/include/asm-i386/poll.h linux/include/asm-i386/poll.h
--- linux.orig/include/asm-i386/poll.h	Thu Jan 23 19:01:28 1997
+++ linux/include/asm-i386/poll.h	Wed Jun 27 17:16:57 2001
@@ -15,6 +15,8 @@
 #define POLLWRNORM	0x0100
 #define POLLWRBAND	0x0200
 #define POLLMSG		0x0400
+#define POLLREMOVE	0x1000
+#define POLLHINT	0x2000
 
 struct pollfd {
 	int fd;
diff -rNu linux.orig/include/linux/devpoll.h linux/include/linux/devpoll.h
--- linux.orig/include/linux/devpoll.h	Thu Jan  1 01:00:00 1970
+++ linux/include/linux/devpoll.h	Wed Jun 27 19:58:52 2001
@@ -0,0 +1,85 @@
+/*
+ * /dev/poll
+ * by Niels Provos <provos@citi.umich.edu>
+ *
+ * provides poll() support via /dev/poll as in Solaris.
+ *
+ * Linux 2.3.x port by Michal Ostrowski
+ * Linux 2.4.x patches by Vincent Sweeney <v.sweeney@dexterus.com>
+ */
+
+#ifndef _LINUX_DEVPOLL_H
+#define _LINUX_DEVPOLL_H
+
+#include <asm/bitops.h>
+#include <linux/list.h>
+#include <asm/atomic.h>
+
+#define DPH_DIRTY	0	/* entry is dirty - bit */
+#define DPH_BACKMAP	1	/* file has an fd back map - bit */
+#ifdef __KERNEL__
+struct dp_fd {
+	struct list_head next;
+	struct pollfd pfd;
+	int flags;		/* for hinting */
+};
+
+struct devpoll {
+	struct list_head *dp_tab;
+	int dp_entries;		/* Entries in hash table */
+	int dp_max;		/* statistics */
+	int dp_avg;		/* more */
+	int dp_count;
+	int dp_cached;
+	int dp_calls;
+	int dp_bucket_bits;
+	int dp_bucket_mask;
+	int dp_nfds;		/* Number of poll fds */
+	u_char *dp_mmap;	/* vaddr of mapped region */
+	atomic_t dp_mmapped;	/* Are we mmapped */
+	rwlock_t dp_lock;
+	u_char **dp_memvec;	/* Pointer to pages allocated for mmap */
+	int dp_numvec;		/* Size of above array */
+};
+#endif
+/* Match solaris */
+
+struct dvpoll {
+	struct pollfd *dp_fds;	/* Leave this ZERO for mmap */
+	int dp_nfds;
+	int dp_timeout;
+};
+
+#define DEVPOLL_MINOR       125	/* Minor device # for /dev/poll */
+
+#define DP_MMAP_SIZE(x)	((x) * sizeof(struct pollfd))
+
+#define DP_ALLOC	_IOR('P', 1, int)
+#define DP_POLL		_IOWR('P', 2, struct dvpoll)
+#define DP_FREE		_IO('P', 3)
+#define DP_ISPOLLED	_IOWR('P', 4, struct pollfd)
+
+#ifdef __KERNEL__
+extern rwlock_t devpoll_lock;
+/* Function Prototypes */
+
+extern inline void
+dp_add_hint (struct poll_backmap **map, rwlock_t * lock)
+{
+	struct poll_backmap *entry;
+	struct dp_fd *dpfd;
+	if (!map)
+		return;
+
+	read_lock (lock);
+	entry = *map;
+	while (entry) {
+		dpfd = entry->arg;
+		set_bit (DPH_DIRTY, &dpfd->flags);	/* atomic */
+		entry = entry->next;
+	}
+	read_unlock (lock);
+}
+#endif				/* __KERNEL__ */
+
+#endif
diff -rNu linux.orig/include/linux/fs.h linux/include/linux/fs.h
--- linux.orig/include/linux/fs.h	Sat May 26 02:01:28 2001
+++ linux/include/linux/fs.h	Wed Jun 27 19:20:36 2001
@@ -502,6 +502,10 @@
 	int			f_error;
 
 	unsigned long		f_version;
+	
+	/* used by /dev/poll hinting */
+	struct poll_backmap	*f_backmap;
+	rwlock_t		f_dplock;
 
 	/* needed for tty driver, and maybe others */
 	void			*private_data;
diff -rNu linux.orig/include/linux/poll.h linux/include/linux/poll.h
--- linux.orig/include/linux/poll.h	Sat May 26 02:01:43 2001
+++ linux/include/linux/poll.h	Wed Jun 27 19:21:05 2001
@@ -8,10 +8,18 @@
 #include <linux/wait.h>
 #include <linux/string.h>
 #include <linux/mm.h>
+#include <linux/malloc.h>
 #include <asm/uaccess.h>
 
 struct poll_table_page;
 
+struct poll_backmap {
+	struct poll_backmap *next;
+	void *arg;			/* pointer to devpoll */
+	struct files_struct *files;	/* files which has this file as */
+	int fd;				/* file descriptor number fd */
+};
+
 typedef struct poll_table_struct {
 	int error;
 	struct poll_table_page * table;
@@ -83,7 +91,88 @@
 	memset(fdset, 0, FDS_BYTES(nr));
 }
 
+extern inline void
+poll_backmap(int fd, void *arg, struct poll_backmap ** entry)
+{
+	struct poll_backmap *tmp;
+
+	if (!entry)
+		return;
+
+	/*
+	 * See if we have an entry in the backmap already, in general
+	 * we expect this linked list to be very short.
+	 */
+	tmp = *entry;
+	while (tmp != NULL) {
+		if (tmp->files == current->files && tmp->fd == fd && 
+		    arg==tmp->arg)
+			return;
+		tmp = tmp->next;
+	}
+
+	tmp = (struct poll_backmap *) kmalloc(sizeof(*entry), GFP_KERNEL);
+	if (tmp == NULL)
+		return;
+
+	tmp->arg = arg;
+	tmp->files = current->files;
+	tmp->fd = fd;
+	tmp->next = *entry;
+
+	*entry = tmp;
+}
+
+extern inline void poll_remove_backmap(struct poll_backmap **map, int fd,
+				       struct files_struct *files)
+{
+	struct poll_backmap *tmp = *map, *old = NULL;
+	
+	while (tmp != NULL) {
+		if (tmp->files == files && tmp->fd == fd) {
+			struct poll_backmap *next = tmp->next;
+			if( old==NULL )
+				*map = next;
+			else
+				old->next = next;
+			kfree(tmp);
+			tmp = next;
+		} else {
+			old = tmp;
+			tmp = tmp->next;
+		}
+	}
+	
+	if (!tmp)
+		return;
+	
+	if (old == NULL)
+		*map = tmp->next;
+	else 
+		old->next = tmp->next;
+
+	kfree (tmp);
+}
+
+extern inline void poll_clean_backmap(struct poll_backmap **map)
+{
+	struct poll_backmap *tmp = *map, *old;
+
+	printk("poll_clean_backmap: map %p\n", map);
+	printk("poll_clean_backmap: *map %p\n", *map);
+
+	while (tmp) {
+	  printk("poll_clean_backmap: tmp %p\n", tmp);
+		old = tmp;
+		tmp = tmp->next;
+		kfree (old);
+	}
+
+	*map = NULL;
+}
+
 extern int do_select(int n, fd_set_bits *fds, long *timeout);
+extern void poll_freewait(poll_table *p);
 
 #endif /* KERNEL */
 
diff -rNu linux.orig/include/net/sock.h linux/include/net/sock.h
--- linux.orig/include/net/sock.h	Sat May 26 02:03:05 2001
+++ linux/include/net/sock.h	Wed Jun 27 19:21:05 2001
@@ -666,6 +666,10 @@
 	/* Identd and reporting IO signals */
 	struct socket		*socket;
 
+	/* For Poll hinting */
+	void			*backmap;
+	void			*dplock;
+
 	/* RPC layer private data */
 	void			*user_data;
   
diff -rNu linux.orig/net/core/datagram.c linux/net/core/datagram.c
--- linux.orig/net/core/datagram.c	Thu Apr 12 20:11:39 2001
+++ linux/net/core/datagram.c	Wed Jun 27 17:28:29 2001
@@ -402,8 +402,6 @@
 	return -EFAULT;
 }
 
-
-
 /*
  *	Datagram poll: Again totally generic. This also handles
  *	sequenced packet sockets providing the socket receive queue
@@ -420,7 +418,10 @@
 	unsigned int mask;
 
 	poll_wait(file, sk->sleep, wait);
-	mask = 0;
+	mask = POLLHINT;
+	
+	sk->backmap = &file->f_backmap;
+	sk->dplock  = &file->f_dplock;
 
 	/* exceptional events? */
 	if (sk->err || !skb_queue_empty(&sk->error_queue))
diff -rNu linux.orig/net/core/sock.c linux/net/core/sock.c
--- linux.orig/net/core/sock.c	Wed Apr 25 22:57:39 2001
+++ linux/net/core/sock.c	Wed Jun 27 18:04:44 2001
@@ -108,6 +108,7 @@
 #include <linux/interrupt.h>
 #include <linux/poll.h>
 #include <linux/init.h>
+#include <linux/devpoll.h>
 
 #include <asm/uaccess.h>
 #include <asm/system.h>
@@ -1100,16 +1101,20 @@
 void sock_def_wakeup(struct sock *sk)
 {
 	read_lock(&sk->callback_lock);
-	if (sk->sleep && waitqueue_active(sk->sleep))
+	if (sk->sleep && waitqueue_active(sk->sleep)) {
+		dp_add_hint(sk->backmap, sk->dplock);
 		wake_up_interruptible_all(sk->sleep);
+	}
 	read_unlock(&sk->callback_lock);
 }
 
 void sock_def_error_report(struct sock *sk)
 {
 	read_lock(&sk->callback_lock);
-	if (sk->sleep && waitqueue_active(sk->sleep))
+	if (sk->sleep && waitqueue_active(sk->sleep)) {
+		dp_add_hint(sk->backmap, sk->dplock);
 		wake_up_interruptible(sk->sleep);
+	}
 	sk_wake_async(sk,0,POLL_ERR); 
 	read_unlock(&sk->callback_lock);
 }
@@ -1117,8 +1122,10 @@
 void sock_def_readable(struct sock *sk, int len)
 {
 	read_lock(&sk->callback_lock);
-	if (sk->sleep && waitqueue_active(sk->sleep))
+	if (sk->sleep && waitqueue_active(sk->sleep)) {
+		dp_add_hint(sk->backmap, sk->dplock);
 		wake_up_interruptible(sk->sleep);
+	}
 	sk_wake_async(sk,1,POLL_IN);
 	read_unlock(&sk->callback_lock);
 }
@@ -1131,8 +1138,10 @@
 	 * progress.  --DaveM
 	 */
 	if((atomic_read(&sk->wmem_alloc) << 1) <= sk->sndbuf) {
-		if (sk->sleep && waitqueue_active(sk->sleep))
+		if (sk->sleep && waitqueue_active(sk->sleep)) {
+			dp_add_hint(sk->backmap, sk->dplock);		
 			wake_up_interruptible(sk->sleep);
+		}
 
 		/* Should agree with poll, otherwise some programs break */
 		if (sock_writeable(sk))
@@ -1163,6 +1172,9 @@
 	sk->zapped	=	1;
 	sk->socket	=	sock;
 
+	sk->backmap	=	NULL;
+	sk->dplock	=	NULL;
+	
 	if(sock)
 	{
 		sk->type	=	sock->type;
diff -rNu linux.orig/net/ipv4/af_inet.c linux/net/ipv4/af_inet.c
--- linux.orig/net/ipv4/af_inet.c	Wed May  2 04:59:24 2001
+++ linux/net/ipv4/af_inet.c	Wed Jun 27 18:06:43 2001
@@ -444,6 +444,7 @@
 		if (sk->linger && !(current->flags & PF_EXITING))
 			timeout = sk->lingertime;
 		sock->sk = NULL;
+		sk->backmap = NULL;
 		sk->prot->close(sk, timeout);
 	}
 	return(0);
diff -rNu linux.orig/net/ipv4/tcp.c linux/net/ipv4/tcp.c
--- linux.orig/net/ipv4/tcp.c	Wed May 16 18:31:27 2001
+++ linux/net/ipv4/tcp.c	Wed Jun 27 17:37:22 2001
@@ -249,6 +249,7 @@
 #include <linux/types.h>
 #include <linux/fcntl.h>
 #include <linux/poll.h>
+#include <linux/devpoll.h>
 #include <linux/init.h>
 #include <linux/smp_lock.h>
 
@@ -380,8 +381,12 @@
 	struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
 
 	poll_wait(file, sk->sleep, wait);
+
+	sk->backmap = &file->f_backmap;
+	sk->dplock  = &file->f_dplock;
+
 	if (sk->state == TCP_LISTEN)
-		return tcp_listen_poll(sk, wait);
+		return tcp_listen_poll(sk, wait) | POLLHINT;
 
 	/* Socket is not locked. We are protected from async events
 	   by poll logic and correct handling of state changes
@@ -454,7 +459,7 @@
 		if (tp->urg_data & TCP_URG_VALID)
 			mask |= POLLPRI;
 	}
-	return mask;
+	return mask | POLLHINT;
 }
 
 /*
@@ -467,8 +472,10 @@
 	if (tcp_wspace(sk) >= tcp_min_write_space(sk) && sock) {
 		clear_bit(SOCK_NOSPACE, &sock->flags);
 
-		if (sk->sleep && waitqueue_active(sk->sleep))
+		if (sk->sleep && waitqueue_active(sk->sleep)) {
+			dp_add_hint(sk->backmap, sk->dplock);
 			wake_up_interruptible(sk->sleep);
+		}
 
 		if (sock->fasync_list && !(sk->shutdown&SEND_SHUTDOWN))
 			sock_wake_async(sock, 2, POLL_OUT);
diff -rNu linux.orig/net/unix/af_unix.c linux/net/unix/af_unix.c
--- linux.orig/net/unix/af_unix.c	Thu Apr 12 20:11:39 2001
+++ linux/net/unix/af_unix.c	Wed Jun 27 17:39:17 2001
@@ -107,6 +107,7 @@
 #include <net/scm.h>
 #include <linux/init.h>
 #include <linux/poll.h>
+#include <linux/devpoll.h>
 #include <linux/smp_lock.h>
 
 #include <asm/checksum.h>
@@ -299,8 +300,10 @@
 {
 	read_lock(&sk->callback_lock);
 	if (unix_writable(sk)) {
-		if (sk->sleep && waitqueue_active(sk->sleep))
+		if (sk->sleep && waitqueue_active(sk->sleep)) {
+			dp_add_hint(sk->backmap,sk->dplock);
 			wake_up_interruptible(sk->sleep);
+		}
 		sk_wake_async(sk, 2, POLL_OUT);
 	}
 	read_unlock(&sk->callback_lock);
@@ -1698,7 +1701,10 @@
 	unsigned int mask;
 
 	poll_wait(file, sk->sleep, wait);
-	mask = 0;
+	mask = POLLHINT;
+
+	sk->backmap = &file->f_backmap;
+	sk->dplock  = &file->f_dplock;
 
 	/* exceptional events? */
 	if (sk->err)