diff mbox series

[RFC,09/15] epoll: introduce stand-alone helpers for polling from userspace

Message ID 20190109164025.24554-10-rpenyaev@suse.de (mailing list archive)
State New, archived
Headers show
Series epoll: support pollable epoll from userspace | expand

Commit Message

Roman Penyaev Jan. 9, 2019, 4:40 p.m. UTC
ep_vrealloc*()
    realloc user header, user index or bitmap memory

ep_get_bit()
    gets free bit from bitmap, if free bit is not found - bitmap is
    expanded on PAGE_SIZE

ep_expand_user_is_required()
    helper which returna true if expand for different memory chunks
    is required

ep_shrink_user_is_required()
    helper which returna new size if shrink for different memory chunks
    is required

ep_expand_user_*()
    expands user header or user index

ep_shrink_user_*()
    shrinks user header, user index or bitmaps.  In case of srink there
    is an important procedure of moving sparsed bits at the end to the
    beginning of the bitmap, in order to free pages at the end.

ep_route_events_to_*()
    routes events to klists or to uring.  Should be called under write
    lock, when all events are stopped.

ep_free_user_item()
    marks item inside user pointer as freed, i.e. atomically exchanges
    ready_events to 0.  Also puts item bit or postponed it to period,
    when user goes to kernel.

ep_add_event_to_uring()
    adds new event to user ring.  Firstly mark user item as ready and if
    item was observed as not ready - fill in user index.

ep_transfer_events_and_shrunk_uring()
    shrinks if needed and transfers events in klists to uring under the
    write lock.

Signed-off-by: Roman Penyaev <rpenyaev@suse.de>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Davidlohr Bueso <dbueso@suse.de>
Cc: Jason Baron <jbaron@akamai.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrea Parri <andrea.parri@amarulasolutions.com>
Cc: linux-fsdevel@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
---
 fs/eventpoll.c | 420 +++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 420 insertions(+)

Comments

Linus Torvalds Jan. 9, 2019, 5:29 p.m. UTC | #1
On Wed, Jan 9, 2019 at 8:40 AM Roman Penyaev <rpenyaev@suse.de> wrote:
>
> ep_vrealloc*()
>     realloc user header, user index or bitmap memory

What? No.

This is wrong, it's much too complicated. And because your
'vrealloc()' doesn't follow the normal realloc rules, it looks both
confusing and buggy, and people have to remember that "oh, vrealloc()
isn't actually vrealloc(), it's really vdupalloc()".

Your other patch to allow users to apparently also do mremap of these
things seems entirely wrongheaded too. Especially when you then have
magical rules for vm_pgoff, which is one of the things that unmapping
parts of a mmap will touch.

So I say no, no, no. This is all *much* too complicated, and the
interfaces are mis-designed to be overly generous to people doing odd
and pointless things.

If you can't have a fixed-size user buffer that stays in one place,
don't even bother.

             Linus
Roman Penyaev Jan. 10, 2019, 10:03 a.m. UTC | #2
On 2019-01-09 18:29, Linus Torvalds wrote:
> On Wed, Jan 9, 2019 at 8:40 AM Roman Penyaev <rpenyaev@suse.de> wrote:
>> 
>> ep_vrealloc*()
>>     realloc user header, user index or bitmap memory
> 
> What? No.
> 
> This is wrong, it's much too complicated. And because your
> 'vrealloc()' doesn't follow the normal realloc rules, it looks both
> confusing and buggy, and people have to remember that "oh, vrealloc()
> isn't actually vrealloc(), it's really vdupalloc()".
> 
> Your other patch to allow users to apparently also do mremap of these
> things seems entirely wrongheaded too. Especially when you then have
> magical rules for vm_pgoff, which is one of the things that unmapping
> parts of a mmap will touch.
> 
> So I say no, no, no. This is all *much* too complicated, and the
> interfaces are mis-designed to be overly generous to people doing odd
> and pointless things.
> 
> If you can't have a fixed-size user buffer that stays in one place,
> don't even bother.

I agree that set of "rules" for this interface is indeed complicated.
The goal was to solve the problem with a constantly changing set of
items (which can be increased / decreased from another thread) without
adding new ctl calls or any limitations.

To fix the size of a user buffer is seems easy to do.  One way is still
to support expand with, say, epoll_ctl(EPOLL_CTL_EXPAND) call and user
has to react explicitly on ENOSPC from epoll_ctl(EPOLL_CTL_ADD).  Thus
reallocation happens, but by user request.

Another way seems much simpler but has a limitation: user has to specify
expected max limit passing the value to a new epoll_create syscall, e.g.
epoll_create2(EPOLL_USERPOLL, 1000). Further attempt to add 1001 
descriptor
will end with ENOSPC. Period. No magic under the hood. Another 1001
descriptor can be added to a new epoll, which can be nested then (what
is forbidden for "polled from user" descriptors in current 
implementation,
but should not be difficult to allow). Then yes, no remapping / 
reallocating.
But this epoll nesting thing ... Which personally I do not like.

What do you think?

--
Roman
diff mbox series

Patch

diff --git a/fs/eventpoll.c b/fs/eventpoll.c
index bdaec59a847e..36c451c26681 100644
--- a/fs/eventpoll.c
+++ b/fs/eventpoll.c
@@ -929,6 +929,238 @@  static void epi_rcu_free(struct rcu_head *head)
 	kmem_cache_free(epi_cache, epi);
 }
 
+static int ep_vrealloc(void **pptr, unsigned int size)
+{
+	void *old = *pptr, *new;
+
+	new = vrealloc(old, size);
+	if (unlikely(!new))
+		return -ENOMEM;
+	if (unlikely(new == old))
+		return 0;
+
+	*pptr = new;
+	vfree(old);
+
+	return 0;
+}
+
+static int ep_vrealloc_bm(struct eventpoll *ep, unsigned int bm_len)
+{
+	unsigned long *bm, *removed_bm;
+
+	/* Reallocate all at once */
+	bm = vrealloc(ep->items_bm, bm_len);
+	removed_bm = vrealloc(ep->removed_items_bm, bm_len);
+
+	if (unlikely(!bm || !removed_bm)) {
+		vfree(bm);
+		vfree(removed_bm);
+
+		return -ENOMEM;
+	}
+	ep->items_bm = bm;
+	ep->removed_items_bm = removed_bm;
+	ep->items_bm_length = bm_len;
+
+	return 0;
+}
+
+static int ep_get_bit(struct eventpoll *ep)
+{
+	unsigned int max_nr;
+	int bit, start_bit;
+	bool was_set;
+
+	lockdep_assert_held(&ep->mtx);
+
+	start_bit = 0;
+again:
+	max_nr = ep_max_items_bm_nr(ep);
+	bit = find_next_zero_bit(ep->items_bm, max_nr, start_bit);
+	if (bit >= max_nr) {
+		unsigned int bm_len;
+		int rc;
+
+		start_bit = max_nr;
+		bm_len = ep->items_bm_length + PAGE_SIZE;
+
+		rc = ep_vrealloc_bm(ep, bm_len);
+		if (unlikely(rc))
+			return rc;
+
+		goto again;
+	}
+
+	was_set = test_and_set_bit(bit, ep->items_bm);
+	WARN_ON(was_set);
+
+	return bit;
+}
+
+static inline bool ep_expand_user_items_is_required(struct eventpoll *ep)
+{
+	return (ep->items_nr >= ep_max_items_nr(ep));
+}
+
+static inline bool ep_expand_user_index_is_required(struct eventpoll *ep)
+{
+	return (ep->items_nr + EPOLL_USER_EXTRA_INDEX_NR)
+		>= ep_max_index_nr(ep);
+}
+
+static inline bool ep_expand_user_is_required(struct eventpoll *ep)
+{
+	return ep_expand_user_items_is_required(ep) ||
+		ep_expand_user_index_is_required(ep);
+}
+
+static inline unsigned int ep_shrunk_user_index_length(struct eventpoll *ep)
+{
+	unsigned int len, nr;
+
+	nr = ep->items_nr + EPOLL_USER_EXTRA_INDEX_NR;
+	len = PAGE_ALIGN(to_index_length(nr) + (PAGE_SIZE >> 1));
+	if (len < ep->index_length)
+		return len;
+
+	return 0;
+}
+
+static inline unsigned int ep_shrunk_user_items_length(struct eventpoll *ep)
+{
+	unsigned int len;
+
+	len = PAGE_ALIGN(to_items_length(ep->items_nr) + (PAGE_SIZE >> 1));
+	if (len < ep->header_length)
+		return len;
+
+	return 0;
+}
+
+static inline unsigned int ep_shrunk_items_bm_length(struct eventpoll *ep)
+{
+	unsigned int len;
+
+	len = PAGE_ALIGN(to_items_bm_length(ep->items_nr) + (PAGE_SIZE >> 1));
+	if (len < ep->items_bm_length)
+		return len;
+
+	return 0;
+}
+
+static inline bool ep_shrink_user_is_required(struct eventpoll *ep)
+{
+	return ep_shrunk_user_items_length(ep) != 0 ||
+		ep_shrunk_user_index_length(ep) != 0 ||
+		ep_shrunk_items_bm_length(ep) != 0;
+}
+
+static inline void ep_route_events_to_klists(struct eventpoll *ep)
+{
+	WARN_ON(!ep_polled_by_user(ep));
+	ep->events_to_uring = false;
+	ep->user_header->state = EPOLL_USER_POLL_INACTIVE;
+	/* Make sure userspace sees INACTIVE state ASAP */
+	smp_wmb();
+}
+
+static inline void ep_route_events_to_uring(struct eventpoll *ep)
+{
+	WARN_ON(!ep_polled_by_user(ep));
+	ep->events_to_uring = true;
+	/* Commit all previous writes to user header */
+	smp_wmb();
+	ep->user_header->state = EPOLL_USER_POLL_ACTIVE;
+}
+
+static inline bool ep_events_routed_to_klists(struct eventpoll *ep)
+{
+	return !ep->events_to_uring;
+}
+
+static inline bool ep_events_routed_to_uring(struct eventpoll *ep)
+{
+	return ep->events_to_uring;
+}
+
+static inline bool ep_free_user_item(struct epitem *epi)
+{
+	struct eventpoll *ep = epi->ep;
+	struct user_epitem *uitem;
+
+	bool events_to_klist = false;
+
+	lockdep_assert_held(&ep->mtx);
+
+	ep->items_nr--;
+
+	uitem = &ep->user_header->items[epi->bit];
+
+	/* Firstly drop item events passed from userland */
+	memset(&uitem->event, 0, sizeof(uitem->event));
+
+	/*
+	 * If event is not signaled yet and has been already consumed by
+	 * userspace it is safe to reuse the bit immediately, i.e. just
+	 * put it.  If userspace has not been yet consumed this event
+	 * we set the bit in removed bitmap in order to put it later.
+	 */
+	if (xchg(&uitem->ready_events, 0)) {
+		set_bit(epi->bit, ep->removed_items_bm);
+		events_to_klist = true;
+	} else {
+		/*
+		 * Should not be reordered with memset above, thus unlock
+		 * semantics.
+		 */
+		clear_bit_unlock(epi->bit, ep->items_bm);
+		events_to_klist = ep_shrink_user_is_required(ep);
+	}
+
+	return events_to_klist;
+}
+
+static bool ep_add_event_to_uring(struct epitem *epi, __poll_t pollflags)
+{
+	struct eventpoll *ep = epi->ep;
+	struct user_epitem *uitem;
+	bool added = false;
+
+	if (WARN_ON(!pollflags))
+		return false;
+
+	uitem = &ep->user_header->items[epi->bit];
+	if (!__atomic_fetch_or(&uitem->ready_events, pollflags,
+			       __ATOMIC_ACQUIRE)) {
+		unsigned int i, *item_idx, index_mask;
+
+		/*
+		 * Item was not ready before, thus we have to insert
+		 * new index to the ring.
+		 */
+
+		index_mask = ep_max_index_nr(ep) - 1;
+		i = __atomic_fetch_add(&ep->user_header->tail, 1,
+				       __ATOMIC_ACQUIRE);
+		item_idx = &ep->user_index[i & index_mask];
+
+		/* Signal with a bit, which is > 0 */
+		*item_idx = epi->bit + 1;
+
+		/*
+		 * Want index update be flushed from CPU write buffer and
+		 * immediately visible on userspace side to avoid long busy
+		 * loops.
+		 */
+		smp_wmb();
+
+		added = true;
+	}
+
+	return added;
+}
+
 /*
  * Removes a "struct epitem" from the eventpoll RB tree and deallocates
  * all the associated resources. Must be called with "mtx" held.
@@ -1695,6 +1927,44 @@  static noinline void ep_destroy_wakeup_source(struct epitem *epi)
 	wakeup_source_unregister(ws);
 }
 
+static int ep_expand_user_items(struct eventpoll *ep)
+{
+	unsigned int len;
+	int rc;
+
+	if (!ep_expand_user_items_is_required(ep))
+		/* Expanding is not needed */
+		return 0;
+
+	len = ep->header_length + PAGE_SIZE;
+	rc = ep_vrealloc((void **)&ep->user_header, len);
+	if (unlikely(rc))
+		return rc;
+
+	ep->header_length = len;
+
+	return 0;
+}
+
+static int ep_expand_user_index(struct eventpoll *ep)
+{
+	unsigned int len;
+	int rc;
+
+	if (!ep_expand_user_index_is_required(ep))
+		/* Expanding is not needed */
+		return 0;
+
+	len = ep->index_length + PAGE_SIZE;
+	rc = ep_vrealloc((void **)&ep->user_index, len);
+	if (unlikely(rc))
+		return rc;
+
+	ep->index_length = len;
+
+	return 0;
+}
+
 /*
  * Must be called with "mtx" held.
  */
@@ -2010,6 +2280,156 @@  static inline struct timespec64 ep_set_mstimeout(long ms)
 	return timespec64_add_safe(now, ts);
 }
 
+static int ep_shrink_user_index(struct eventpoll *ep)
+{
+	unsigned int len;
+	int rc;
+
+	len = ep_shrunk_user_index_length(ep);
+	if (!len)
+		/* Shrinking is not needed */
+		return 0;
+
+	rc = ep_vrealloc((void **)&ep->user_index, len);
+	if (unlikely(rc))
+		return rc;
+
+	ep->index_length = len;
+
+	return 0;
+}
+
+static int ep_shrink_user_items_and_bm(struct eventpoll *ep)
+{
+	unsigned int header_len, bm_len;
+	unsigned int bit, last_bit = UINT_MAX;
+	int rc;
+
+	struct rb_node *rbp;
+	struct epitem *epi;
+
+	lockdep_assert_held(&ep->mtx);
+
+	header_len = ep_shrunk_user_items_length(ep);
+	bm_len = ep_shrunk_items_bm_length(ep);
+	if (!header_len && !bm_len)
+		/* Shrinking is not needed */
+		return 0;
+
+	/*
+	 * Find left most last bit
+	 */
+	if (header_len)
+		last_bit = to_items_nr(header_len);
+	if (bm_len)
+		last_bit = min(last_bit, to_items_bm_nr(header_len));
+
+	if (WARN_ON(last_bit <= ep->items_nr))
+		return -EINVAL;
+
+	/*
+	 * Find bits from the right and move them to the left in order to
+	 * free space on the right.
+	 *
+	 * This is not nice, because O(n), but frankly this operation should
+	 * be quite rare.  If not - let's switch to idr or something similar
+	 * (but that obviously will consume more memory).
+	 *
+	 */
+	bit = 0;
+	for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
+		epi = rb_entry(rbp, struct epitem, rbn);
+
+		if (epi->bit >= last_bit) {
+			/* Find first available bit from left */
+			bit = find_next_zero_bit(ep->items_bm, last_bit, bit);
+			if (WARN_ON(bit >= last_bit))
+				return -EINVAL;
+
+			/* Clear old bit from right */
+			clear_bit(epi->bit, ep->items_bm);
+
+			/*
+			 * Set item bit and advance an iterator for the
+			 * following find_next_zero_bit() call.
+			 */
+			epi->bit = bit++;
+		}
+	}
+
+	/*
+	 * Reallocate memory and commit lengths
+	 */
+	if (header_len) {
+		rc = ep_vrealloc((void **)&ep->user_header, header_len);
+		if (unlikely(rc))
+			return rc;
+
+		ep->header_length = header_len;
+	}
+	if (bm_len) {
+		rc = ep_vrealloc_bm(ep, bm_len);
+		if (unlikely(rc))
+			return rc;
+	}
+
+	return 0;
+}
+
+static inline void ep_put_postponed_user_items_bits(struct eventpoll *ep)
+{
+	size_t sz, i;
+
+	lockdep_assert_held(&ep->mtx);
+
+	sz = ep->items_bm_length >> ilog2(sizeof(ep->items_bm[0]));
+	for (i = 0; i < sz; i++) {
+		ep->items_bm[i] &= ~(ep->removed_items_bm[i]);
+		ep->removed_items_bm[i] = 0ul;
+	}
+}
+
+static int ep_transfer_events_and_shrink_uring(struct eventpoll *ep)
+{
+	struct epitem *epi, *tmp;
+	int rc = 0;
+
+	mutex_lock(&ep->mtx);
+	if (ep_events_routed_to_uring(ep))
+		/* A bit late */
+		goto unlock;
+
+	/* Here at this point we are sure uring is empty */
+	ep_put_postponed_user_items_bits(ep);
+
+	rc = ep_shrink_user_index(ep);
+	if (unlikely(rc))
+		goto unlock;
+
+	rc = ep_shrink_user_items_and_bm(ep);
+	if (unlikely(rc))
+		goto unlock;
+
+	/* Commit lengths to userspace, but state is not yet ACTIVE */
+	ep->user_header->index_length = ep->index_length;
+	ep->user_header->header_length = ep->header_length;
+
+	write_lock_irq(&ep->lock);
+	/* Atomically transfer events from klists to uring */
+	list_for_each_entry_safe(epi, tmp, &ep->rdllist, rdllink) {
+		ep_add_event_to_uring(epi, epi->ready_events);
+		list_del_init(&epi->rdllink);
+		epi->ready_events = 0;
+	}
+	ep_route_events_to_uring(ep);
+	write_unlock_irq(&ep->lock);
+
+unlock:
+	mutex_unlock(&ep->mtx);
+
+	return rc;
+}
+
 /**
  * ep_poll - Retrieves ready events, and delivers them to the caller supplied
  *           event buffer.