Skip to contentSkip to conversation

Adapt poll layer behaviour to match the epoll layer

PATCH New Yannick Lamarre 3 participants 12 messages Mar 21, 2019 – Mar 25, 2019 6 messages not in the archive

lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): [email protected]

#1 lore raw

PATCH v2 1/1 Adapt poll layer behaviour to match the epoll layer

---
 src/common/compat/compat-poll.c | 28 ++++++++++++++++++----------
 src/common/compat/poll.h        |  1 +
 2 files changed, 19 insertions(+), 10 deletions(-)
src/common/compat/compat-poll.c +18 -10
index b45b39dc..b7c17df7 100644
@@ -1,5 +1,6 @@
 /*
  * Copyright (C)  2011 - David Goulet <[email protected]>
+ * Copyright (C)  2019 - Yannick Lamarre <[email protected]>
  *
  * This program is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License, version 2 only,
@@ -204,10 +205,10 @@ int compat_poll_mod(struct lttng_poll_event *events, int fd,
 		uint32_t req_events)
 {
 	int i;
-	bool fd_found = false;
 	struct compat_poll_event_array *current;
 
-	if (events == NULL || events->current.events == NULL || fd < 0) {
+	if (events == NULL || events->current.nb_fd == 0 || events->current.events
+			== NULL || fd < 0) {
 		ERR("Bad compat poll mod arguments");
 		goto error;
 	}
@@ -216,16 +217,15 @@ int compat_poll_mod(struct lttng_poll_event *events, int fd,
 
 	for (i = 0; i < current->nb_fd; i++) {
 		if (current->events[i].fd == fd) {
-			fd_found = true;
 			current->events[i].events = req_events;
 			events->need_update = 1;
 			break;
 		}
 	}
 
-	if (!fd_found) {
-		goto error;
-	}
+	/* 
+	 * The epoll flavor doesn't flag modifying a non-included FD as an error.
+	 */
 
 	return 0;
 
@@ -238,11 +238,12 @@ error:
  */
 int compat_poll_del(struct lttng_poll_event *events, int fd)
 {
-	int new_size, i, count = 0, ret;
+	int i, count = 0, ret;
+	uint32_t new_size;
 	struct compat_poll_event_array *current;
 
-	if (events == NULL || events->current.events == NULL || fd < 0) {
-		ERR("Wrong arguments for poll del");
+	if (events == NULL || events->current.nb_fd == 0 || events->current.events
+			== NULL || fd < 0) {
 		goto error;
 	}
 
@@ -257,13 +258,19 @@ int compat_poll_del(struct lttng_poll_event *events, int fd)
 			count++;
 		}
 	}
+
+	/* The fd was not in our set, return no error as with epoll. */
+	if (current->nb_fd == count) {
+		goto skip;
+	}
+
 	/* No fd duplicate should be ever added into array. */
 	assert(current->nb_fd - 1 == count);
 	current->nb_fd = count;
 
 	/* Resize array if needed. */
 	new_size = 1U << utils_get_count_order_u32(current->nb_fd);
-	if (new_size != current->alloc_size && new_size >= current->init_size) {
+	if (new_size != current->alloc_size && new_size >= current->init_size && current->nb_fd != 0) {
 		ret = resize_poll_event(current, new_size);
 		if (ret < 0) {
 			goto error;
@@ -272,6 +279,7 @@ int compat_poll_del(struct lttng_poll_event *events, int fd)
 
 	events->need_update = 1;
 
+skip:
 	return 0;
 
 error:
src/common/compat/poll.h +1 -0
index 5eb7ff9c..d2826b7a 100644
@@ -1,5 +1,6 @@
 /*
  * Copyright (C) 2011 - David Goulet <[email protected]>
+ * Copyright (C) 2019 - Yannick Lamarre <[email protected]>
  *
  * This program is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License, version 2 only,
Signature · 1 line
2.11.0
#2 lore raw

PATCH v2 1/1 Fix hang in thread_rotation when using compat-poll

In reply to a message not in the archive

Add missing verification to prevent a blocking read on an empty fd.

---
 src/bin/lttng-sessiond/rotation-thread.c | 4 ++++
 1 file changed, 4 insertions(+)
src/bin/lttng-sessiond/rotation-thread.c +4 -0
index 6669372d..b86b1668 100644
@@ -974,6 +974,10 @@ void *thread_rotation(void *data)
 			int fd = LTTNG_POLL_GETFD(&thread.events, i);
 			uint32_t revents = LTTNG_POLL_GETEV(&thread.events, i);
 
+			if (!revents) {
+				/* No activity for this FD (poll implementation). */
+				continue;
+			}
 			DBG("[rotation-thread] Handling fd (%i) activity (%u)",
 					fd, revents);
 
Signature · 1 line
2.11.0
#3 lore raw

PATCH v2 1/1 Change lttng_poll_wait behaviour of compat-poll to match compat-epoll

In reply to a message not in the archive

This removes the need to verify for eventless file descriptors and mitigates risks of bug due to behaviour mismatch.

---
 src/common/compat/compat-poll.c | 33 ++++++++++++++++++++++++++++++---
 1 file changed, 30 insertions(+), 3 deletions(-)
src/common/compat/compat-poll.c +30 -3
index b7c17df7..6aa9414d 100644
@@ -292,6 +292,7 @@ error:
 int compat_poll_wait(struct lttng_poll_event *events, int timeout)
 {
 	int ret;
+	int eidx = 0;
 
 	if (events == NULL || events->current.events == NULL) {
 		ERR("poll wait arguments error");
@@ -324,10 +325,36 @@ int compat_poll_wait(struct lttng_poll_event *events, int timeout)
 	}
 
 	/*
-	 * poll() should always iterate on all FDs since we handle the pollset in
-	 * user space and after poll returns, we have to try every fd for a match.
+	 * Swap all nonzero revents pollfd structs to the beginning of the array to
+	 * emulate compat-epoll behaviour. This algorithm takes advantage of poll's
+	 * returned value and the burst nature of waking events on the file
+	 * descriptors. The while loop garantees that idlepfd will always point to
+	 * a pollfd with revents == 0.
 	 */
-	return events->wait.nb_fd;
+	if (ret == events->wait.nb_fd) {
+		goto skip;
+	}
+	while (eidx < ret && events->wait.events[eidx].revents != 0) {
+		eidx += 1;
+	}
+
+	for (int idx = eidx + 1; eidx < ret; idx++) {
+		struct pollfd swapfd;
+		struct pollfd *idlepfd = &events->wait.events[eidx];
+		struct pollfd *ipfd = &events->wait.events[idx];
+
+		assert(idx < events->wait.nb_fd);
+
+		if (ipfd->revents != 0) {
+			swapfd = *ipfd;
+			*ipfd = *idlepfd;
+			*idlepfd = swapfd;
+			eidx += 1;
+		}
+	}
+
+skip:
+	return ret;
 
 error:
 	return -1;
Signature · 1 line
2.11.0
#4 lore raw

PATCH v2 1/1 Fix typo

In reply to a message not in the archive
---
 src/common/compat/poll.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
src/common/compat/poll.h +1 -1
index d2826b7a..c7bd0606 100644
@@ -359,7 +359,7 @@ extern int compat_poll_del(struct lttng_poll_event *events, int fd);
 	compat_poll_del(events, fd)
 
 /*
- * Modify an fd's events in the epoll set.
+ * Modify an fd's events in the poll set.
  */
 extern int compat_poll_mod(struct lttng_poll_event *events,
 		int fd, uint32_t req_events);
Signature · 1 line
2.11.0
#5 lore raw

PATCH v2 1/1 Add Unit test to poll compatibility layer

In reply to a message not in the archive
---
 .gitignore                          |   1 +
 tests/unit/Makefile.am              |   9 +-
 tests/unit/test_utils_compat_poll.c | 236 ++++++++++++++++++++++++++++++++++++
 3 files changed, 245 insertions(+), 1 deletion(-)
 create mode 100644 tests/unit/test_utils_compat_poll.c
.gitignore +1 -0
index 19276012..b5d2a55a 100644
@@ -73,6 +73,7 @@ TAGS
 /tests/unit/test_session
 /tests/unit/test_uri
 /tests/unit/test_ust_data
+/tests/unit/test_utils_compat_poll
 /tests/unit/test_utils_parse_size_suffix
 /tests/unit/test_utils_parse_time_suffix
 /tests/unit/test_utils_expand_path
tests/unit/Makefile.am +8 -1
index 5f485f00..bcda6a5c 100644
@@ -12,6 +12,7 @@ TESTS = test_kernel_data \
 	test_utils_parse_size_suffix \
 	test_utils_parse_time_suffix \
 	test_utils_expand_path \
+	test_utils_compat_poll \
 	test_string_utils \
 	test_notification \
 	ini_config/test_ini_config
@@ -28,7 +29,8 @@ LIBLTTNG_CTL=$(top_builddir)/src/lib/lttng-ctl/liblttng-ctl.la
 # Define test programs
 noinst_PROGRAMS = test_uri test_session test_kernel_data \
                   test_utils_parse_size_suffix test_utils_parse_time_suffix \
-                  test_utils_expand_path test_string_utils test_notification
+                  test_utils_expand_path test_utils_compat_poll \
+		  test_string_utils test_notification
 
 if HAVE_LIBLTTNG_UST_CTL
 noinst_PROGRAMS += test_ust_data
@@ -145,6 +147,11 @@ test_utils_parse_size_suffix_LDADD = $(LIBTAP) $(LIBHASHTABLE) $(LIBCOMMON) $(DL
 test_utils_parse_time_suffix_SOURCES = test_utils_parse_time_suffix.c
 test_utils_parse_time_suffix_LDADD = $(LIBTAP) $(LIBHASHTABLE) $(LIBCOMMON)
 
+# compat_poll unit test
+test_utils_compat_poll_SOURCES = test_utils_compat_poll.c
+test_utils_compat_poll_LDADD  = $(LIBTAP) $(LIBHASHTABLE) $(DL_LIBS) \
+		      $(top_builddir)/src/common/compat/libcompat.la $(LIBCOMMON)
+
 # expand_path unit test
 test_utils_expand_path_SOURCES = test_utils_expand_path.c
 test_utils_expand_path_LDADD = $(LIBTAP) $(LIBHASHTABLE) $(LIBCOMMON) $(DL_LIBS)
tests/unit/test_utils_compat_poll.c new file +236 -0

236 lines not on this page · raw message

Signature · 1 line
2.11.0
#6 lore raw

PATCH v2 1/1 Change LTTNG_POLL_GETNB behaviour for poll flavor

In reply to a message not in the archive

Modify LTTNG_POLL_GETNB to provide compatibility with the epoll flavor. Since it is only used after a lttng_poll_wait call with no modification (add, del, mod) between, this change does not modify the behaviour in its current usage while ensuring similar API behavior between compatibility layer implementations.

---
 src/common/compat/poll.h | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)
src/common/compat/poll.h +3 -1
index fde54ddb..5eb7ff9c 100644
@@ -315,10 +315,12 @@ static inline int __lttng_poll_get_prev_fd(struct lttng_poll_event *events,
 /*
  * For the following calls, consider 'e' to be a lttng_poll_event pointer and i
  * being the index of the events array.
+ * LTTNG_POLL_GETNB is always used after lttng_poll_wait, thus we can use the
+ * current list for test compatibility purposes.
  */
 #define LTTNG_POLL_GETFD(e, i) LTTNG_REF(e)->wait.events[i].fd
 #define LTTNG_POLL_GETEV(e, i) LTTNG_REF(e)->wait.events[i].revents
-#define LTTNG_POLL_GETNB(e) LTTNG_REF(e)->wait.nb_fd
+#define LTTNG_POLL_GETNB(e) LTTNG_REF(e)->current.nb_fd
 #define LTTNG_POLL_GETSZ(e) LTTNG_REF(e)->wait.events_size
 #define LTTNG_POLL_GET_PREV_FD(e, i, nb_fd) \
 	__lttng_poll_get_prev_fd(LTTNG_REF(e), i, nb_fd)
Signature · 1 line
2.11.0
#7 lore raw

PATCH v2 1/1 Clean code base from redundant verification

In reply to a message not in the archive

Remove redundant verification for file descriptors with 0 revents in code base.

---
 src/bin/lttng-consumerd/health-consumerd.c   |  5 -----
 src/bin/lttng-relayd/health-relayd.c         |  5 -----
 src/bin/lttng-relayd/live.c                  | 10 ----------
 src/bin/lttng-relayd/main.c                  | 16 ----------------
 src/bin/lttng-sessiond/agent-thread.c        |  5 -----
 src/bin/lttng-sessiond/client.c              |  5 -----
 src/bin/lttng-sessiond/dispatch.c            |  5 -----
 src/bin/lttng-sessiond/health.c              |  5 -----
 src/bin/lttng-sessiond/ht-cleanup.c          |  5 -----
 src/bin/lttng-sessiond/manage-apps.c         |  5 -----
 src/bin/lttng-sessiond/manage-consumer.c     | 10 ----------
 src/bin/lttng-sessiond/manage-kernel.c       |  5 -----
 src/bin/lttng-sessiond/notification-thread.c |  3 ---
 src/bin/lttng-sessiond/notify-apps.c         |  5 -----
 src/bin/lttng-sessiond/register.c            |  5 -----
 src/bin/lttng-sessiond/rotation-thread.c     |  4 ----
 src/common/consumer/consumer.c               | 10 ----------
 17 files changed, 108 deletions(-)
src/bin/lttng-consumerd/health-consumerd.c +0 -5
index 1e2f31e4..7d60e54e 100644
@@ -259,11 +259,6 @@ restart:
 			revents = LTTNG_POLL_GETEV(&events, i);
 			pollfd = LTTNG_POLL_GETFD(&events, i);
 
-			if (!revents) {
-				/* No activity for this FD (poll implementation). */
-				continue;
-			}
-
 			/* Thread quit pipe has been closed. Killing thread. */
 			ret = check_health_quit_pipe(pollfd, revents);
 			if (ret) {
src/bin/lttng-relayd/health-relayd.c +0 -5
index ba996621..b782c3e7 100644
@@ -329,11 +329,6 @@ restart:
 			revents = LTTNG_POLL_GETEV(&events, i);
 			pollfd = LTTNG_POLL_GETFD(&events, i);
 
-			if (!revents) {
-				/* No activity for this FD (poll implementation). */
-				continue;
-			}
-
 			/* Thread quit pipe has been closed. Killing thread. */
 			ret = check_health_quit_pipe(pollfd, revents);
 			if (ret) {
src/bin/lttng-relayd/live.c +0 -10
index 9efab8a7..9503ba47 100644
@@ -543,11 +543,6 @@ restart:
 			revents = LTTNG_POLL_GETEV(&events, i);
 			pollfd = LTTNG_POLL_GETFD(&events, i);
 
-			if (!revents) {
-				/* No activity for this FD (poll implementation). */
-				continue;
-			}
-
 			/* Thread quit pipe has been closed. Killing thread. */
 			ret = check_thread_quit_pipe(pollfd, revents);
 			if (ret) {
@@ -2000,11 +1995,6 @@ restart:
 
 			health_code_update();
 
-			if (!revents) {
-				/* No activity for this FD (poll implementation). */
-				continue;
-			}
-
 			/* Thread quit pipe has been closed. Killing thread. */
 			ret = check_thread_quit_pipe(pollfd, revents);
 			if (ret) {
src/bin/lttng-relayd/main.c +0 -16
index de14cdb4..42f871b5 100644
@@ -858,14 +858,6 @@ restart:
 			revents = LTTNG_POLL_GETEV(&events, i);
 			pollfd = LTTNG_POLL_GETFD(&events, i);
 
-			if (!revents) {
-				/*
-				 * No activity for this FD (poll
-				 * implementation).
-				 */
-				continue;
-			}
-
 			/* Thread quit pipe has been closed. Killing thread. */
 			ret = check_thread_quit_pipe(pollfd, revents);
 			if (ret) {
@@ -3811,14 +3803,6 @@ restart:
 
 			health_code_update();
 
-			if (!revents) {
-				/*
-				 * No activity for this FD (poll
-				 * implementation).
-				 */
-				continue;
-			}
-
 			/* Thread quit pipe has been closed. Killing thread. */
 			ret = check_thread_quit_pipe(pollfd, revents);
 			if (ret) {
src/bin/lttng-sessiond/agent-thread.c +0 -5
index f7be1ef7..575f6aee 100644
@@ -404,11 +404,6 @@ restart:
 			revents = LTTNG_POLL_GETEV(&events, i);
 			pollfd = LTTNG_POLL_GETFD(&events, i);
 
-			if (!revents) {
-				/* No activity for this FD (poll implementation). */
-				continue;
-			}
-
 			/* Thread quit pipe has been closed. Killing thread. */
 			if (pollfd == quit_pipe_read_fd) {
 				goto exit;
src/bin/lttng-sessiond/client.c +0 -5
index a889529a..0be065ec 100644
@@ -2174,11 +2174,6 @@ static void *thread_manage_clients(void *data)
 
 			health_code_update();
 
-			if (!revents) {
-				/* No activity for this FD (poll implementation). */
-				continue;
-			}
-
 			if (pollfd == thread_quit_pipe_fd) {
 				err = 0;
 				goto exit;
src/bin/lttng-sessiond/dispatch.c +0 -5
index 04b3954c..a637f23d 100644
@@ -144,11 +144,6 @@ static void sanitize_wait_queue(struct ust_reg_wait_queue *wait_queue)
 		uint32_t revents = LTTNG_POLL_GETEV(&events, i);
 		int pollfd = LTTNG_POLL_GETFD(&events, i);
 
-		if (!revents) {
-			/* No activity for this FD (poll implementation). */
-			continue;
-		}
-
 		cds_list_for_each_entry_safe(wait_node, tmp_wait_node,
 				&wait_queue->head, head) {
 			if (pollfd == wait_node->app->sock &&
src/bin/lttng-sessiond/health.c +0 -5
index 921b4526..5b4f5f8f 100644
@@ -158,11 +158,6 @@ restart:
 			revents = LTTNG_POLL_GETEV(&events, i);
 			pollfd = LTTNG_POLL_GETFD(&events, i);
 
-			if (!revents) {
-				/* No activity for this FD (poll implementation). */
-				continue;
-			}
-
 			/* Event on the registration socket */
 			if (pollfd == sock) {
 				if (revents & LPOLLIN) {
src/bin/lttng-sessiond/ht-cleanup.c +0 -5
index 91c0544c..02244f0c 100644
@@ -155,11 +155,6 @@ static void *thread_ht_cleanup(void *data)
 			revents = LTTNG_POLL_GETEV(&events, i);
 			pollfd = LTTNG_POLL_GETFD(&events, i);
 
-			if (!revents) {
-				/* No activity for this FD (poll implementation). */
-				continue;
-			}
-
 			if (pollfd != ht_cleanup_pipe[0]) {
 				continue;
 			}
src/bin/lttng-sessiond/manage-apps.c +0 -5
index f9ec356d..356c147a 100644
@@ -123,11 +123,6 @@ static void *thread_application_management(void *data)
 
 			health_code_update();
 
-			if (!revents) {
-				/* No activity for this FD (poll implementation). */
-				continue;
-			}
-
 			if (pollfd == quit_pipe_read_fd) {
 				err = 0;
 				goto exit;
src/bin/lttng-sessiond/manage-consumer.c +0 -10
index b710c61f..fa802adc 100644
@@ -134,11 +134,6 @@ void *thread_consumer_management(void *data)
 
 		health_code_update();
 
-		if (!revents) {
-			/* No activity for this FD (poll implementation). */
-			continue;
-		}
-
 		/* Thread quit pipe has been closed. Killing thread. */
 		if (pollfd == quit_pipe_read_fd) {
 			err = 0;
@@ -298,11 +293,6 @@ void *thread_consumer_management(void *data)
 
 			health_code_update();
 
-			if (!revents) {
-				/* No activity for this FD (poll implementation). */
-				continue;
-			}
-
 			/*
 			 * Thread quit pipe has been triggered, flag that we should stop
 			 * but continue the current loop to handle potential data from
src/bin/lttng-sessiond/manage-kernel.c +0 -5
index 34887d7b..f656c9f5 100644
@@ -267,11 +267,6 @@ static void *thread_kernel_management(void *data)
 
 			health_code_update();
 
-			if (!revents) {
-				/* No activity for this FD (poll implementation). */
-				continue;
-			}
-
 			if (pollfd == quit_pipe_read_fd) {
 				err = 0;
 				goto exit;
src/bin/lttng-sessiond/notification-thread.c +0 -3
index b42b282e..0a1fd557 100644
@@ -562,9 +562,6 @@ void *thread_notification(void *data)
 			int fd = LTTNG_POLL_GETFD(&state.events, i);
 			uint32_t revents = LTTNG_POLL_GETEV(&state.events, i);
 
-			if (!revents) {
-				continue;
-			}
 			DBG("[notification-thread] Handling fd (%i) activity (%u)", fd, revents);
 
 			if (fd == state.notification_channel_socket) {
src/bin/lttng-sessiond/notify-apps.c +0 -5
index bc7405c7..17f3cd0c 100644
@@ -109,11 +109,6 @@ restart:
 			revents = LTTNG_POLL_GETEV(&events, i);
 			pollfd = LTTNG_POLL_GETFD(&events, i);
 
-			if (!revents) {
-				/* No activity for this FD (poll implementation). */
-				continue;
-			}
-
 			/* Thread quit pipe has been closed. Killing thread. */
 			if (pollfd == quit_pipe_read_fd) {
 				err = 0;
src/bin/lttng-sessiond/register.c +0 -5
index 6ce25ad0..e809834f 100644
@@ -239,11 +239,6 @@ static void *thread_application_registration(void *data)
 			revents = LTTNG_POLL_GETEV(&events, i);
 			pollfd = LTTNG_POLL_GETFD(&events, i);
 
-			if (!revents) {
-				/* No activity for this FD (poll implementation). */
-				continue;
-			}
-
 			/* Thread quit pipe has been closed. Killing thread. */
 			if (pollfd == quit_pipe_read_fd) {
 				err = 0;
src/bin/lttng-sessiond/rotation-thread.c +0 -4
index b86b1668..6669372d 100644
@@ -974,10 +974,6 @@ void *thread_rotation(void *data)
 			int fd = LTTNG_POLL_GETFD(&thread.events, i);
 			uint32_t revents = LTTNG_POLL_GETEV(&thread.events, i);
 
-			if (!revents) {
-				/* No activity for this FD (poll implementation). */
-				continue;
-			}
 			DBG("[rotation-thread] Handling fd (%i) activity (%u)",
 					fd, revents);
 
src/common/consumer/consumer.c +0 -10
index 10273e66..c70d95a5 100644
@@ -2396,11 +2396,6 @@ restart:
 			revents = LTTNG_POLL_GETEV(&events, i);
 			pollfd = LTTNG_POLL_GETFD(&events, i);
 
-			if (!revents) {
-				/* No activity for this FD (poll implementation). */
-				continue;
-			}
-
 			if (pollfd == lttng_pipe_get_readfd(ctx->consumer_metadata_pipe)) {
 				if (revents & LPOLLIN) {
 					ssize_t pipe_len;
@@ -2990,11 +2985,6 @@ restart:
 			revents = LTTNG_POLL_GETEV(&events, i);
 			pollfd = LTTNG_POLL_GETFD(&events, i);
 
-			if (!revents) {
-				/* No activity for this FD (poll implementation). */
-				continue;
-			}
-
 			if (pollfd == ctx->consumer_channel_pipe[0]) {
 				if (revents & LPOLLIN) {
 					enum consumer_channel_action action;
Signature · 1 line
2.11.0

A message that is not in the archive (sent off-list or not yet received): [email protected]

#9 lore raw

Re: [PATCH lttng-tools v2] Change lttng_poll_wait behaviour of compat-poll to match compat-epoll

Thanks for this patch, it does allow a worthwhile clean-up! Read on for comments about style.

Quoted 14 lines · Yannick Lamarre

This removes the need to verify for eventless file descriptors and mitigates risks of bug due to behaviour mismatch.

---
 src/common/compat/compat-poll.c | 33 ++++++++++++++++++++++++++++++---
 1 file changed, 30 insertions(+), 3 deletions(-)
src/common/compat/compat-poll.c +0 -0
index b7c17df7..6aa9414d 100644

src/common/compat/compat-poll.c error: · line 295

Quoted 4 lines
src/common/compat/compat-poll.c +1 -0
@@ -292,6 +292,7 @@ error:
 int compat_poll_wait(struct lttng_poll_event *events, int timeout)
 {
        int ret;
+       int eidx = 0;

Reading the code, this seems to mean the "index of the first idle file descriptor".

As a general rule, don't use uncommon abbreviations. Abbreviations are okay when they are common-place, unambiguous and save a lot of characters (e.g. cfg for configuration, or lttng for linux_trace_toolkit_next_generation).

When in doubt, prefer verbose names.

In this case, "idle_pfd_index" and a comment (see below) would clarify the intent.

src/common/compat/compat-poll.c int compat_poll_wait(struct lttng_poll_event *events, int timeout) · line 331

Quoted 13 lines
src/common/compat/compat-poll.c +4 -2
        if (events == NULL || events->current.events == NULL) {
                ERR("poll wait arguments error");
@@ -324,10 +325,36 @@ int compat_poll_wait(struct lttng_poll_event *events, int timeout)
        }
 
        /*
-        * poll() should always iterate on all FDs since we handle the pollset in
-        * user space and after poll returns, we have to try every fd for a match.
+        * Swap all nonzero revents pollfd structs to the beginning of the array to
+        * emulate compat-epoll behaviour. This algorithm takes advantage of poll's
+        * returned value and the burst nature of waking events on the file
+        * descriptors. The while loop garantees that idlepfd will always point to

garantees -> guarantees

This comment block exceeds 80 chars. Make sure your editor's tab width is set to 8 characters.

The code, commit message, and comments make use of multiple terms to refer to the same file descriptor state ('idle', 'empty', 'eventless', and 'zero').

Please stick to one of those terms.

Idle seems preferable, but feel free to whichever term the poll() specification uses.

src/common/compat/compat-poll.c int compat_poll_wait(struct lttng_poll_event *events, int timeout) · line 338

Quoted 8 lines
src/common/compat/compat-poll.c +6 -1
+        * a pollfd with revents == 0.
         */
-       return events->wait.nb_fd;
+       if (ret == events->wait.nb_fd) {
+               goto skip;
+       }
+       while (eidx < ret && events->wait.events[eidx].revents != 0) {
+               eidx += 1;

Prefer "eidx++".

Using variables for a single purpose is fine and tends to make the code clearer.

For example, reworking this section of code to:

active_fd_count = ret;
/* Find the first idle pollfd. */
while (idle_pfd_index < active_fd_count &&

events->wait.events[idle_pfd_index].revents != 0) {
        idle_pfd_index++;
}

makes the intent clearer.

src/common/compat/compat-poll.c int compat_poll_wait(struct lttng_poll_event *events, int timeout) · line 341

Quoted 3 lines
src/common/compat/compat-poll.c +3 -0
+       }
+
+       for (int idx = eidx + 1; eidx < ret; idx++) {

Same change here (regarding 'ret').

To maintain uniformity with the rest of the code base, please declare variables at the top of the inner-most scope containing the loop.

When declaring a trivial loop counter/index, use the name 'i'. Here, the difference between 'idx' and 'eidx' is not immediately apparent. Renaming 'eidx' makes the loop clearer.

Also, in this particular case, 'i' is fine since you are only using it to assign the current pollfd, which is fairly descriptive.

src/common/compat/compat-poll.c int compat_poll_wait(struct lttng_poll_event *events, int timeout) · line 344

Quoted 3 lines
src/common/compat/compat-poll.c +3 -0
+               struct pollfd swapfd;
+               struct pollfd *idlepfd = &events->wait.events[eidx];
+               struct pollfd *ipfd = &events->wait.events[idx];

Use an underscore to split variable names, e.g. swap_pfd, idle_pfd, current_pfd. Again, verbose names are preferable.

src/common/compat/compat-poll.c int compat_poll_wait(struct lttng_poll_event *events, int timeout) · line 346

Quoted 2 lines
src/common/compat/compat-poll.c +2 -0
+
+               assert(idx < events->wait.nb_fd);

Not needed. Using assert() is fine to highlight assumptions across interfaces/functions and/or catch catastrophic internal errors. Here, this would be a very local condition that seems to be already enforced.

Reworking the 'for' condition to use 'i < events->wait.nb_fd' and breaking out early of the loop is also an option here.

src/common/compat/compat-poll.c int compat_poll_wait(struct lttng_poll_event *events, int timeout) · line 352

Quoted 6 lines
src/common/compat/compat-poll.c +6 -0
+
+               if (ipfd->revents != 0) {
+                       swapfd = *ipfd;
+                       *ipfd = *idlepfd;
+                       *idlepfd = swapfd;
+                       eidx += 1;

eidx++;

src/common/compat/compat-poll.c int compat_poll_wait(struct lttng_poll_event *events, int timeout) · line 356

Quoted 4 lines
src/common/compat/compat-poll.c +4 -0
+               }
+       }
+
+skip:

Rename label to 'end'.

Thanks!
Jérémie

Quoted 7 lines
src/common/compat/compat-poll.c +1 -1
+       return ret;
 
 error:
        return -1;
--

2.11.0

--
Jérémie Galarneau
EfficiOS Inc.
http://www.efficios.com
_______________________________________________
lttng-dev mailing list
[email protected]
https://lists.lttng.org/cgi-bin/mailman/listinfo/lttng-dev

A message that is not in the archive (sent off-list or not yet received): CA+jJMxui0_5EF6XLH-9G90-yJEE63zdiBifTeF4i9iQp95D7tg@mail.gmail.com

#11 lore raw

Re: [PATCH lttng-tools v2] Change lttng_poll_wait behaviour of compat-poll to match compat-epoll

----- On Mar 21, 2019, at 4:25 PM, Jeremie Galarneau [email protected] wrote:

[...]

Quoted 9 lines
Quoted 1 line

+ assert(idx < events->wait.nb_fd);

Not needed. Using assert() is fine to highlight assumptions across interfaces/functions and/or catch catastrophic internal errors. Here, this would be a very local condition that seems to be already enforced.

Reworking the 'for' condition to use 'i < events->wait.nb_fd' and breaking out early of the loop is also an option here.

The reason for using assert() here is to ensure that there is no discrepancy between the number of active poll fd in the set and the value returned by poll indicating that same number of active fds.

We'd need a kernel bug (or memory corruption) to have a mismatch, but I thought having an assert in there would not hurt, since it might help diagnose issues that cross user-kernel boundaries.

Thoughts ?

Mathieu

Signature · 3 lines
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com

A message that is not in the archive (sent off-list or not yet received): [email protected]

#13 lore raw

Re: [PATCH lttng-tools v2] Change lttng_poll_wait behaviour of compat-poll to match compat-epoll

Quoted 18 lines · Mathieu Desnoyers

----- On Mar 21, 2019, at 4:25 PM, Jeremie Galarneau [email protected] wrote:

[...]

Quoted 9 lines
Quoted 1 line

+ assert(idx < events->wait.nb_fd);

Not needed. Using assert() is fine to highlight assumptions across interfaces/functions and/or catch catastrophic internal errors. Here, this would be a very local condition that seems to be already enforced.

Reworking the 'for' condition to use 'i < events->wait.nb_fd' and breaking out early of the loop is also an option here.

The reason for using assert() here is to ensure that there is no discrepancy between the number of active poll fd in the set and the value returned by poll indicating that same number of active fds.

Quoted 5 lines

We'd need a kernel bug (or memory corruption) to have a mismatch, but I thought having an assert in there would not hurt, since it might help diagnose issues that cross user-kernel boundaries.

Thoughts ?

Hi,

I understand the intent, but I don't feel these type of kernel tests/checks belong in the code.

However, if there is a specific concern that our use of poll() could highlight this type of bug (in the kernel or in the interactions with lttng-modules), we can add tests for that.

Thanks,
Jérémie

Quoted 7 lines

Mathieu

--
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com

--
Jérémie Galarneau
EfficiOS Inc.
http://www.efficios.com
_______________________________________________
lttng-dev mailing list
[email protected]
https://lists.lttng.org/cgi-bin/mailman/listinfo/lttng-dev

A message that is not in the archive (sent off-list or not yet received): CA+jJMxsbA0ksOJN9yqbtCVXkTLKOqfEMek6P9_bxSdKH75zEPQ@mail.gmail.com

#15 lore raw

Re: [PATCH lttng-tools v2] Change lttng_poll_wait behaviour of compat-poll to match compat-epoll

Quoted 27 lines · ----- On Mar 25, 2019, at 11:10 AM, Jeremie Galarneau…
Quoted 19 lines · Mathieu Desnoyers

----- On Mar 21, 2019, at 4:25 PM, Jeremie Galarneau [email protected] wrote:

[...]

Quoted 9 lines
Quoted 1 line

+ assert(idx < events->wait.nb_fd);

Not needed. Using assert() is fine to highlight assumptions across interfaces/functions and/or catch catastrophic internal errors. Here, this would be a very local condition that seems to be already enforced.

Reworking the 'for' condition to use 'i < events->wait.nb_fd' and breaking out early of the loop is also an option here.

The reason for using assert() here is to ensure that there is no discrepancy between the number of active poll fd in the set and the value returned by poll indicating that same number of active fds.

Quoted 5 lines

We'd need a kernel bug (or memory corruption) to have a mismatch, but I thought having an assert in there would not hurt, since it might help diagnose issues that cross user-kernel boundaries.

Thoughts ?

Quoted 8 lines

Hi,

I understand the intent, but I don't feel these type of kernel tests/checks belong in the code.

However, if there is a specific concern that our use of poll() could highlight this type of bug (in the kernel or in the interactions with lttng-modules), we can add tests for that.

There is no specific concern. We can remove the assert.

Thanks,

Mathieu

Quoted 17 lines

Thanks,
Jérémie

Quoted 7 lines

Mathieu

--
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com

--
Jérémie Galarneau
EfficiOS Inc.
http://www.efficios.com

Signature · 7 lines
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com
_______________________________________________
lttng-dev mailing list
[email protected]
https://lists.lttng.org/cgi-bin/mailman/listinfo/lttng-dev

A message that is not in the archive (sent off-list or not yet received): [email protected]

#17 lore raw

Re: [PATCH lttng-tools v2] Change lttng_poll_wait behaviour of compat-poll to match compat-epoll

Alright, I'll remove the assert for v4.

----- Original Message -----
From: "Mathieu Desnoyers" <[email protected]>
To: "Jeremie Galarneau" <[email protected]>
Cc: "Yannick Lamarre" <[email protected]>, "lttng-dev" <[email protected]>, "jgalar" <[email protected]>
Sent: Monday, March 25, 2019 11:56:11 AM
Subject: Re: [lttng-dev] [PATCH lttng-tools v2] Change lttng_poll_wait behaviour of compat-poll to match compat-epoll

Quoted 27 lines · ----- On Mar 25, 2019, at 11:10 AM, Jeremie Galarneau…
Quoted 19 lines · Mathieu Desnoyers

----- On Mar 21, 2019, at 4:25 PM, Jeremie Galarneau [email protected] wrote:

[...]

Quoted 9 lines
Quoted 1 line

+ assert(idx < events->wait.nb_fd);

Not needed. Using assert() is fine to highlight assumptions across interfaces/functions and/or catch catastrophic internal errors. Here, this would be a very local condition that seems to be already enforced.

Reworking the 'for' condition to use 'i < events->wait.nb_fd' and breaking out early of the loop is also an option here.

The reason for using assert() here is to ensure that there is no discrepancy between the number of active poll fd in the set and the value returned by poll indicating that same number of active fds.

Quoted 5 lines

We'd need a kernel bug (or memory corruption) to have a mismatch, but I thought having an assert in there would not hurt, since it might help diagnose issues that cross user-kernel boundaries.

Thoughts ?

Quoted 8 lines

Hi,

I understand the intent, but I don't feel these type of kernel tests/checks belong in the code.

However, if there is a specific concern that our use of poll() could highlight this type of bug (in the kernel or in the interactions with lttng-modules), we can add tests for that.

There is no specific concern. We can remove the assert.

Thanks,

Mathieu

Quoted 17 lines

Thanks,
Jérémie

Quoted 7 lines

Mathieu

--
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com

--
Jérémie Galarneau
EfficiOS Inc.
http://www.efficios.com

Signature · 7 lines
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com
_______________________________________________
lttng-dev mailing list
[email protected]
https://lists.lttng.org/cgi-bin/mailman/listinfo/lttng-dev