Skip to contentSkip to conversation

task work cleanups

PATCH In review Jens Axboe 2 participants 12 messages Nov 22, 2024 – Nov 23, 2024 io_uring

lore
Share & tools
lore

Jens Axboe

author maintainer
#0 lore raw

cover task work cleanups

Hi,

This patchset gets rid of using llist for handling task_work, for both local and normal task_work. Instead of a lockless list, a normal io_wq_work_list is used and protected by a spinlock. I've done some benchmarking with this, and I only see wins with this - the act of adding or iterating task_work is the same cost, but we get rid of the need to reverse the task_work list, which can be substantial for bursty applications or just generally busy task_work usages.

Patch 2 implements io_wq_work_list handling for deferred task_work, and patch 4 does the same for normal task_work. Patch 6 then also switches SQPOLL to use this scheme, which eliminates the passing around of io_wq_work_node for its retry logic.

Outside of cleaning up this code, it also enables us to potentially implement task_work run capping for normal task_work in the future.

Git tree can be found here:

https://git.kernel.dk/cgit/linux/log/?h=io_uring-defer-tw

 include/linux/io_uring_types.h |  17 +-
 io_uring/io_uring.c            | 293 ++++++++++++++++++---------------
 io_uring/io_uring.h            |  20 ++-
 io_uring/slist.h               |  16 ++
 io_uring/sqpoll.c              |  20 ++-
 io_uring/tctx.c                |   3 +-
 6 files changed, 211 insertions(+), 158 deletions(-)
Signature · 1 line
Jens Axboe
author maintainer
#1 lore raw

PATCH 1/6 io_uring: make task_work pending check dependent on ring type

There's no need to check for generic task_work for DEFER_TASKRUN, if we have local task_work pending. This avoids dipping into the huge task_struct, if we have normal task_work pending.

---
 io_uring/io_uring.h | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)
io_uring/io_uring.h +3 -1
index 12abee607e4a..214f9f175102 100644
@@ -354,7 +354,9 @@ static inline bool io_local_work_pending(struct io_ring_ctx *ctx)
 
 static inline bool io_task_work_pending(struct io_ring_ctx *ctx)
 {
-	return task_work_pending(current) || io_local_work_pending(ctx);
+	if (ctx->flags & IORING_SETUP_DEFER_TASKRUN && io_local_work_pending(ctx))
+		return true;
+	return task_work_pending(current);
 }
 
 static inline void io_tw_lock(struct io_ring_ctx *ctx, struct io_tw_state *ts)
Signature · 1 line
2.45.2
author maintainer
#2 lore raw

PATCH 2/6 io_uring: replace defer task_work llist with io_wq_work_list

In reply to Jens Axboe

Add a spinlock for the list, and replace the lockless llist with the work list instead. This avoids needing to reverse items in the list before running them, as the io_wq_work_list is FIFO by nature whereas the llist is LIFO.

---
 include/linux/io_uring_types.h |  13 ++-
 io_uring/io_uring.c            | 194 ++++++++++++++++-----------------
 io_uring/io_uring.h            |   2 +-
 3 files changed, 104 insertions(+), 105 deletions(-)
include/linux/io_uring_types.h +8 -5
index 011860ade268..e9ba99cb0ed0 100644
@@ -335,8 +335,9 @@ struct io_ring_ctx {
 	 * regularly bounce b/w CPUs.
 	 */
 	struct {
-		struct llist_head	work_llist;
-		struct llist_head	retry_llist;
+		struct io_wq_work_list	work_list;
+		spinlock_t		work_lock;
+		int			work_items;
 		unsigned long		check_cq;
 		atomic_t		cq_wait_nr;
 		atomic_t		cq_timeouts;
@@ -566,7 +567,11 @@ enum {
 typedef void (*io_req_tw_func_t)(struct io_kiocb *req, struct io_tw_state *ts);
 
 struct io_task_work {
-	struct llist_node		node;
+	/* DEFER_TASKRUN uses work_node, regular task_work node */
+	union {
+		struct io_wq_work_node	work_node;
+		struct llist_node	node;
+	};
 	io_req_tw_func_t		func;
 };
 
@@ -622,8 +627,6 @@ struct io_kiocb {
 	 */
 	u16				buf_index;
 
-	unsigned			nr_tw;
-
 	/* REQ_F_* flags */
 	io_req_flags_t			flags;
 
io_uring/io_uring.c +95 -99

275 lines not on this page · raw message

io_uring/io_uring.h +1 -1
index 214f9f175102..2fae27803116 100644
@@ -349,7 +349,7 @@ static inline int io_run_task_work(void)
 
 static inline bool io_local_work_pending(struct io_ring_ctx *ctx)
 {
-	return !llist_empty(&ctx->work_llist) || !llist_empty(&ctx->retry_llist);
+	return READ_ONCE(ctx->work_list.first);
 }
 
 static inline bool io_task_work_pending(struct io_ring_ctx *ctx)
Signature · 1 line
2.45.2
author maintainer
#8 lore raw

PATCH 3/6 io_uring/slist: add list-to-list list splice helper

In reply to Jens Axboe

Add a helper to splice a source list to a destination list.

---
 io_uring/slist.h | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)
io_uring/slist.h +16 -0
index 0eb194817242..7ac7c136b702 100644
@@ -85,6 +85,22 @@ static inline bool wq_list_splice(struct io_wq_work_list *list,
 	return false;
 }
 
+static inline bool wq_list_splice_list(struct io_wq_work_list *src,
+				       struct io_wq_work_list *dst)
+{
+	bool ret = false;
+
+	if (wq_list_empty(dst)) {
+		*dst = *src;
+	} else {
+		dst->last->next = src->first;
+		dst->last = src->last;
+		ret = true;
+	}
+	INIT_WQ_LIST(src);
+	return false;
+}
+
 static inline void wq_stack_add_head(struct io_wq_work_node *node,
 				     struct io_wq_work_node *stack)
 {
Signature · 1 line
2.45.2
author maintainer
#9 lore raw

PATCH 4/6 io_uring: switch non-defer task_work to io_wq_work_list

In reply to Jens Axboe

Switch the normal task_work to io_wq_work_list as well, to both unify with defer task_work, but also to avoid needing to reverse the ordering of the list when running it.

Note that this still keeps the manual retry list for SQPOLL task_work. That could go away as well, as now the task_work list is fully ordered and SQPOLL could just leave entries on there when it chops up the running of the list.

---
 include/linux/io_uring_types.h |  14 ++-
 io_uring/io_uring.c            | 167 ++++++++++++++++++++-------------
 io_uring/io_uring.h            |   6 +-
 io_uring/sqpoll.c              |   8 +-
 io_uring/tctx.c                |   3 +-
 5 files changed, 116 insertions(+), 82 deletions(-)
include/linux/io_uring_types.h +6 -8
index e9ba99cb0ed0..7ddac4d1d4b3 100644
@@ -102,7 +102,8 @@ struct io_uring_task {
 	struct percpu_counter		inflight;
 
 	struct { /* task_work */
-		struct llist_head	task_list;
+		struct io_wq_work_list	task_list;
+		spinlock_t		task_lock;
 		struct callback_head	task_work;
 	} ____cacheline_aligned_in_smp;
 };
@@ -390,8 +391,9 @@ struct io_ring_ctx {
 	struct mm_struct		*mm_account;
 
 	/* ctx exit and cancelation */
-	struct llist_head		fallback_llist;
-	struct delayed_work		fallback_work;
+	struct io_wq_work_list		fallback_list;
+	spinlock_t			fallback_lock;
+	struct work_struct		fallback_work;
 	struct work_struct		exit_work;
 	struct list_head		tctx_list;
 	struct completion		ref_comp;
@@ -567,11 +569,7 @@ enum {
 typedef void (*io_req_tw_func_t)(struct io_kiocb *req, struct io_tw_state *ts);
 
 struct io_task_work {
-	/* DEFER_TASKRUN uses work_node, regular task_work node */
-	union {
-		struct io_wq_work_node	work_node;
-		struct llist_node	node;
-	};
+	struct io_wq_work_node		node;
 	io_req_tw_func_t		func;
 };
 
io_uring/io_uring.c +100 -67

269 lines not on this page · raw message

io_uring/io_uring.h +4 -2
index 2fae27803116..0b5181b128aa 100644
@@ -91,8 +91,10 @@ void io_req_task_queue(struct io_kiocb *req);
 void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts);
 void io_req_task_queue_fail(struct io_kiocb *req, int ret);
 void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts);
-struct llist_node *io_handle_tw_list(struct llist_node *node, unsigned int *count, unsigned int max_entries);
-struct llist_node *tctx_task_work_run(struct io_uring_task *tctx, unsigned int max_entries, unsigned int *count);
+struct io_wq_work_node *io_handle_tw_list(struct io_wq_work_node *node,
+	unsigned int *count, unsigned int max_entries);
+struct io_wq_work_node *tctx_task_work_run(struct io_uring_task *tctx,
+	unsigned int max_entries, unsigned int *count);
 void tctx_task_work(struct callback_head *cb);
 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
 int io_uring_alloc_task_context(struct task_struct *task,
io_uring/sqpoll.c +4 -4
index 6df5e649c413..615707260f25 100644
@@ -221,7 +221,7 @@ static bool io_sqd_handle_event(struct io_sq_data *sqd)
  * than we were asked to process. Newly queued task_work isn't run until the
  * retry list has been fully processed.
  */
-static unsigned int io_sq_tw(struct llist_node **retry_list, int max_entries)
+static unsigned int io_sq_tw(struct io_wq_work_node **retry_list, int max_entries)
 {
 	struct io_uring_task *tctx = current->io_uring;
 	unsigned int count = 0;
@@ -239,11 +239,11 @@ static unsigned int io_sq_tw(struct llist_node **retry_list, int max_entries)
 	return count;
 }
 
-static bool io_sq_tw_pending(struct llist_node *retry_list)
+static bool io_sq_tw_pending(struct io_wq_work_node *retry_list)
 {
 	struct io_uring_task *tctx = current->io_uring;
 
-	return retry_list || !llist_empty(&tctx->task_list);
+	return retry_list || READ_ONCE(tctx->task_list.first);
 }
 
 static void io_sq_update_worktime(struct io_sq_data *sqd, struct rusage *start)
@@ -259,7 +259,7 @@ static void io_sq_update_worktime(struct io_sq_data *sqd, struct rusage *start)
 
 static int io_sq_thread(void *data)
 {
-	struct llist_node *retry_list = NULL;
+	struct io_wq_work_node *retry_list = NULL;
 	struct io_sq_data *sqd = data;
 	struct io_ring_ctx *ctx;
 	struct rusage start;
io_uring/tctx.c +2 -1
index 503f3ff8bc4f..7155b3c56c85 100644
@@ -87,7 +87,8 @@ __cold int io_uring_alloc_task_context(struct task_struct *task,
 	atomic_set(&tctx->in_cancel, 0);
 	atomic_set(&tctx->inflight_tracked, 0);
 	task->io_uring = tctx;
-	init_llist_head(&tctx->task_list);
+	INIT_WQ_LIST(&tctx->task_list);
+	spin_lock_init(&tctx->task_lock);
 	init_task_work(&tctx->task_work, tctx_task_work);
 	return 0;
 }
Signature · 1 line
2.45.2
author maintainer
#10 lore raw

PATCH 5/6 io_uring: add __tctx_task_work_run() helper

In reply to Jens Axboe

Most use cases only care about running all of the task_work, and they don't need the node passed back or the work capped. Rename the existing helper to __tctx_task_work_run(), and add a wrapper around that for the more basic use cases.

---
 io_uring/io_uring.c | 18 ++++++++++++------
 io_uring/io_uring.h |  9 +++------
 io_uring/sqpoll.c   |  2 +-
 3 files changed, 16 insertions(+), 13 deletions(-)
io_uring/io_uring.c +12 -6
index 3bb93c77ac3f..bc520a67fc03 100644
@@ -1137,9 +1137,9 @@ static void io_fallback_tw(struct io_uring_task *tctx, bool sync)
 	__io_fallback_tw(&tctx->task_list, &tctx->task_lock, sync);
 }
 
-struct io_wq_work_node *tctx_task_work_run(struct io_uring_task *tctx,
-					   unsigned int max_entries,
-					   unsigned int *count)
+struct io_wq_work_node *__tctx_task_work_run(struct io_uring_task *tctx,
+					     unsigned int max_entries,
+					     unsigned int *count)
 {
 	struct io_wq_work_node *node;
 
@@ -1167,14 +1167,20 @@ struct io_wq_work_node *tctx_task_work_run(struct io_uring_task *tctx,
 	return node;
 }
 
+unsigned int tctx_task_work_run(struct io_uring_task *tctx)
+{
+	unsigned int count = 0;
+
+	__tctx_task_work_run(tctx, UINT_MAX, &count);
+	return count;
+}
+
 void tctx_task_work(struct callback_head *cb)
 {
 	struct io_uring_task *tctx;
-	unsigned int count = 0;
 
 	tctx = container_of(cb, struct io_uring_task, task_work);
-	if (tctx_task_work_run(tctx, UINT_MAX, &count))
-		WARN_ON_ONCE(1);
+	tctx_task_work_run(tctx);
 }
 
 static inline void io_req_local_work_add(struct io_kiocb *req,
io_uring/io_uring.h +3 -6
index 0b5181b128aa..2b0e7c5db30d 100644
@@ -93,8 +93,9 @@ void io_req_task_queue_fail(struct io_kiocb *req, int ret);
 void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts);
 struct io_wq_work_node *io_handle_tw_list(struct io_wq_work_node *node,
 	unsigned int *count, unsigned int max_entries);
-struct io_wq_work_node *tctx_task_work_run(struct io_uring_task *tctx,
+struct io_wq_work_node *__tctx_task_work_run(struct io_uring_task *tctx,
 	unsigned int max_entries, unsigned int *count);
+unsigned int tctx_task_work_run(struct io_uring_task *tctx);
 void tctx_task_work(struct callback_head *cb);
 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
 int io_uring_alloc_task_context(struct task_struct *task,
@@ -332,12 +333,8 @@ static inline int io_run_task_work(void)
 			resume_user_mode_work(NULL);
 		}
 		if (current->io_uring) {
-			unsigned int count = 0;
-
 			__set_current_state(TASK_RUNNING);
-			tctx_task_work_run(current->io_uring, UINT_MAX, &count);
-			if (count)
-				ret = true;
+			ret = tctx_task_work_run(current->io_uring) != 0;
 		}
 	}
 	if (task_work_pending(current)) {
io_uring/sqpoll.c +1 -1
index 615707260f25..aec6c2d56910 100644
@@ -232,7 +232,7 @@ static unsigned int io_sq_tw(struct io_wq_work_node **retry_list, int max_entrie
 			goto out;
 		max_entries -= count;
 	}
-	*retry_list = tctx_task_work_run(tctx, max_entries, &count);
+	*retry_list = __tctx_task_work_run(tctx, max_entries, &count);
 out:
 	if (task_work_pending(current))
 		task_work_run();
Signature · 1 line
2.45.2
author maintainer
#11 lore raw

PATCH 6/6 io_uring: make __tctx_task_work_run() take an io_wq_work_list

In reply to Jens Axboe

The normal task_work logic doesn't really need it, as it always runs all of the pending work. But for SQPOLL, it can now pass in its retry_list which simplifies the tracking of split up task_work running.

This avoids passing io_wq_work_node around. Rather than pass in a list, SQPOLL could re-add the leftover items to the generic task_work list. But that requires re-locking the task_lock and using task_list for that, whereas having a separate retry list allows for skipping those steps. The downside is that now two lists need checking, but that's now it was before as well.

---
 io_uring/io_uring.c | 36 ++++++++++++++++--------------------
 io_uring/io_uring.h |  9 +++++----
 io_uring/sqpoll.c   | 20 +++++++++++---------
 3 files changed, 32 insertions(+), 33 deletions(-)
io_uring/io_uring.c +16 -20
index bc520a67fc03..5e52d8db3dca 100644
@@ -1044,20 +1044,20 @@ static void ctx_flush_and_put(struct io_ring_ctx *ctx, struct io_tw_state *ts)
 /*
  * Run queued task_work, returning the number of entries processed in *count.
  * If more entries than max_entries are available, stop processing once this
- * is reached and return the rest of the list.
+ * is reached.
  */
-struct io_wq_work_node *io_handle_tw_list(struct io_wq_work_node *node,
-					  unsigned int *count,
-					  unsigned int max_entries)
+void io_handle_tw_list(struct io_wq_work_list *list, unsigned int *count,
+		       unsigned int max_entries)
 {
 	struct io_ring_ctx *ctx = NULL;
 	struct io_tw_state ts = { };
 
 	do {
-		struct io_wq_work_node *next = node->next;
+		struct io_wq_work_node *node = list->first;
 		struct io_kiocb *req = container_of(node, struct io_kiocb,
 						    io_task_work.node);
 
+		list->first = node->next;
 		if (req->ctx != ctx) {
 			ctx_flush_and_put(ctx, &ts);
 			ctx = req->ctx;
@@ -1067,17 +1067,15 @@ struct io_wq_work_node *io_handle_tw_list(struct io_wq_work_node *node,
 		INDIRECT_CALL_2(req->io_task_work.func,
 				io_poll_task_func, io_req_rw_complete,
 				req, &ts);
-		node = next;
 		(*count)++;
 		if (unlikely(need_resched())) {
 			ctx_flush_and_put(ctx, &ts);
 			ctx = NULL;
 			cond_resched();
 		}
-	} while (node && *count < max_entries);
+	} while (list->first && *count < max_entries);
 
 	ctx_flush_and_put(ctx, &ts);
-	return node;
 }
 
 static __cold void __io_fallback_schedule(struct io_ring_ctx *ctx,
@@ -1137,41 +1135,39 @@ static void io_fallback_tw(struct io_uring_task *tctx, bool sync)
 	__io_fallback_tw(&tctx->task_list, &tctx->task_lock, sync);
 }
 
-struct io_wq_work_node *__tctx_task_work_run(struct io_uring_task *tctx,
-					     unsigned int max_entries,
-					     unsigned int *count)
+void __tctx_task_work_run(struct io_uring_task *tctx,
+			  struct io_wq_work_list *list,
+			  unsigned int max_entries, unsigned int *count)
 {
-	struct io_wq_work_node *node;
-
 	if (unlikely(current->flags & PF_EXITING)) {
 		io_fallback_tw(tctx, true);
-		return NULL;
+		return;
 	}
 
 	if (!READ_ONCE(tctx->task_list.first))
-		return NULL;
+		return;
 
 	spin_lock_irq(&tctx->task_lock);
-	node = tctx->task_list.first;
+	*list = tctx->task_list;
 	INIT_WQ_LIST(&tctx->task_list);
 	spin_unlock_irq(&tctx->task_lock);
 
-	if (node)
-		node = io_handle_tw_list(node, count, max_entries);
+	if (!wq_list_empty(list))
+		io_handle_tw_list(list, count, max_entries);
 
 	/* relaxed read is enough as only the task itself sets ->in_cancel */
 	if (unlikely(atomic_read(&tctx->in_cancel)))
 		io_uring_drop_tctx_refs(current);
 
 	trace_io_uring_task_work_run(tctx, *count);
-	return node;
 }
 
 unsigned int tctx_task_work_run(struct io_uring_task *tctx)
 {
+	struct io_wq_work_list list;
 	unsigned int count = 0;
 
-	__tctx_task_work_run(tctx, UINT_MAX, &count);
+	__tctx_task_work_run(tctx, &list, UINT_MAX, &count);
 	return count;
 }
 
io_uring/io_uring.h +5 -4
index 2b0e7c5db30d..74b1468aefda 100644
@@ -91,10 +91,11 @@ void io_req_task_queue(struct io_kiocb *req);
 void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts);
 void io_req_task_queue_fail(struct io_kiocb *req, int ret);
 void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts);
-struct io_wq_work_node *io_handle_tw_list(struct io_wq_work_node *node,
-	unsigned int *count, unsigned int max_entries);
-struct io_wq_work_node *__tctx_task_work_run(struct io_uring_task *tctx,
-	unsigned int max_entries, unsigned int *count);
+void io_handle_tw_list(struct io_wq_work_list *list, unsigned int *count,
+	unsigned int max_entries);
+void __tctx_task_work_run(struct io_uring_task *tctx,
+	struct io_wq_work_list *list, unsigned int max_entries,
+	unsigned int *count);
 unsigned int tctx_task_work_run(struct io_uring_task *tctx);
 void tctx_task_work(struct callback_head *cb);
 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
io_uring/sqpoll.c +11 -9
index aec6c2d56910..3cd50369db5a 100644
@@ -221,29 +221,29 @@ static bool io_sqd_handle_event(struct io_sq_data *sqd)
  * than we were asked to process. Newly queued task_work isn't run until the
  * retry list has been fully processed.
  */
-static unsigned int io_sq_tw(struct io_wq_work_node **retry_list, int max_entries)
+static unsigned int io_sq_tw(struct io_wq_work_list *retry_list, int max_entries)
 {
 	struct io_uring_task *tctx = current->io_uring;
 	unsigned int count = 0;
 
-	if (*retry_list) {
-		*retry_list = io_handle_tw_list(*retry_list, &count, max_entries);
+	if (!wq_list_empty(retry_list)) {
+		io_handle_tw_list(retry_list, &count, max_entries);
 		if (count >= max_entries)
 			goto out;
 		max_entries -= count;
 	}
-	*retry_list = __tctx_task_work_run(tctx, max_entries, &count);
+	__tctx_task_work_run(tctx, retry_list, max_entries, &count);
 out:
 	if (task_work_pending(current))
 		task_work_run();
 	return count;
 }
 
-static bool io_sq_tw_pending(struct io_wq_work_node *retry_list)
+static bool io_sq_tw_pending(struct io_wq_work_list *retry_list)
 {
 	struct io_uring_task *tctx = current->io_uring;
 
-	return retry_list || READ_ONCE(tctx->task_list.first);
+	return !wq_list_empty(retry_list) || !wq_list_empty(&tctx->task_list);
 }
 
 static void io_sq_update_worktime(struct io_sq_data *sqd, struct rusage *start)
@@ -259,7 +259,7 @@ static void io_sq_update_worktime(struct io_sq_data *sqd, struct rusage *start)
 
 static int io_sq_thread(void *data)
 {
-	struct io_wq_work_node *retry_list = NULL;
+	struct io_wq_work_list retry_list;
 	struct io_sq_data *sqd = data;
 	struct io_ring_ctx *ctx;
 	struct rusage start;
@@ -292,6 +292,7 @@ static int io_sq_thread(void *data)
 	audit_uring_entry(IORING_OP_NOP);
 	audit_uring_exit(true, 0);
 
+	INIT_WQ_LIST(&retry_list);
 	mutex_lock(&sqd->lock);
 	while (1) {
 		bool cap_entries, sqt_spin = false;
@@ -332,7 +333,8 @@ static int io_sq_thread(void *data)
 		}
 
 		prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
-		if (!io_sqd_events_pending(sqd) && !io_sq_tw_pending(retry_list)) {
+		if (!io_sqd_events_pending(sqd) &&
+		    !io_sq_tw_pending(&retry_list)) {
 			bool needs_sched = true;
 
 			list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
@@ -371,7 +373,7 @@ static int io_sq_thread(void *data)
 		timeout = jiffies + sqd->sq_thread_idle;
 	}
 
-	if (retry_list)
+	if (!wq_list_empty(&retry_list))
 		io_sq_tw(&retry_list, UINT_MAX);
 
 	io_uring_cancel_generic(true, sqd);
Signature · 1 line
2.45.2
#3 lore raw

Re: [PATCH 2/6] io_uring: replace defer task_work llist with io_wq_work_list

In reply to Jens Axboe

On 11/22/24 16:12, Jens Axboe wrote:
...

Quoted 14 lines

static inline void io_req_local_work_add(struct io_kiocb *req,
struct io_ring_ctx *ctx,

io_uring/io_uring.c +3 -3
-					 unsigned flags)
+					 unsigned tw_flags)
  {
-	unsigned nr_wait, nr_tw, nr_tw_prev;
-	struct llist_node *head;
+	unsigned nr_tw, nr_tw_prev, nr_wait;
+	unsigned long flags;
 
  	/* See comment above IO_CQ_WAKE_INIT */
  	BUILD_BUG_ON(IO_CQ_WAKE_FORCE <= IORING_MAX_CQ_ENTRIES);
 
  	/*

io_uring/io_uring.c void tctx_task_work(struct callback_head *cb) · line 1157

Quoted 10 lines
io_uring/io_uring.c +3 -4
-	 * We don't know how many reuqests is there in the link and whether
-	 * they can even be queued lazily, fall back to non-lazy.
+	 * We don't know how many requests are in the link and whether they can
+	 * even be queued lazily, fall back to non-lazy.
  	 */
  	if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK))
-		flags &= ~IOU_F_TWQ_LAZY_WAKE;
+		tw_flags &= ~IOU_F_TWQ_LAZY_WAKE;
 
-	guard(rcu)();

protects against ctx->task deallocation, see a comment in io_ring_exit_work() -> synchronize_rcu()

io_uring/io_uring.c void tctx_task_work(struct callback_head *cb) · line 1159

Quoted 3 lines
io_uring/io_uring.c +3 -0
+	spin_lock_irqsave(&ctx->work_lock, flags);
+	wq_list_add_tail(&req->io_task_work.work_node, &ctx->work_list);
+	nr_tw_prev = ctx->work_items++;

Is there a good reason why it changes the semantics of what's stored across adds? It was assigning a corrected nr_tw, this one will start heavily spamming with wake_up() in some cases.

Quoted 60 lines
io_uring/io_uring.c +13 -35
+	spin_unlock_irqrestore(&ctx->work_lock, flags);
 
-	head = READ_ONCE(ctx->work_llist.first);
-	do {
-		nr_tw_prev = 0;
-		if (head) {
-			struct io_kiocb *first_req = container_of(head,
-							struct io_kiocb,
-							io_task_work.node);
-			/*
-			 * Might be executed at any moment, rely on
-			 * SLAB_TYPESAFE_BY_RCU to keep it alive.
-			 */
-			nr_tw_prev = READ_ONCE(first_req->nr_tw);
-		}
-
-		/*
-		 * Theoretically, it can overflow, but that's fine as one of
-		 * previous adds should've tried to wake the task.
-		 */
-		nr_tw = nr_tw_prev + 1;
-		if (!(flags & IOU_F_TWQ_LAZY_WAKE))
-			nr_tw = IO_CQ_WAKE_FORCE;
-
-		req->nr_tw = nr_tw;
-		req->io_task_work.node.next = head;
-	} while (!try_cmpxchg(&ctx->work_llist.first, &head,
-			      &req->io_task_work.node));
-
-	/*
-	 * cmpxchg implies a full barrier, which pairs with the barrier
-	 * in set_current_state() on the io_cqring_wait() side. It's used
-	 * to ensure that either we see updated ->cq_wait_nr, or waiters
-	 * going to sleep will observe the work added to the list, which
-	 * is similar to the wait/wawke task state sync.
-	 */
+	nr_tw = nr_tw_prev + 1;
+	if (!(tw_flags & IOU_F_TWQ_LAZY_WAKE))
+		nr_tw = IO_CQ_WAKE_FORCE;
 
-	if (!head) {
+	if (!nr_tw_prev) {
  		if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
  			atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
  		if (ctx->has_evfd)
  			io_eventfd_signal(ctx);
  	}
 
+	/*
+	 * We need a barrier after unlock, which pairs with the barrier
+	 * in set_current_state() on the io_cqring_wait() side. It's used
+	 * to ensure that either we see updated ->cq_wait_nr, or waiters
+	 * going to sleep will observe the work added to the list, which
+	 * is similar to the wait/wake task state sync.
+	 */
+	smp_mb();
  	nr_wait = atomic_read(&ctx->cq_wait_nr);
  	/* not enough or no one is waiting */
  	if (nr_tw < nr_wait)
@@ -1253,11 +1233,27 @@ void io_req_task_work_add_remote(struct io_kiocb *req, struct io_ring_ctx *ctx,
Signature · 1 line
Pavel Begunkov
author maintainer
#4 lore raw

Re: [PATCH 2/6] io_uring: replace defer task_work llist with io_wq_work_list

Quoted 24 lines · Pavel Begunkov

On 11/22/24 16:12, Jens Axboe wrote:
...

Quoted 11 lines

static inline void io_req_local_work_add(struct io_kiocb *req,
struct io_ring_ctx *ctx,

io_uring/io_uring.c +3 -3
-                     unsigned flags)
+                     unsigned tw_flags)
  {
-    unsigned nr_wait, nr_tw, nr_tw_prev;
-    struct llist_node *head;
+    unsigned nr_tw, nr_tw_prev, nr_wait;
+    unsigned long flags;
        /* See comment above IO_CQ_WAKE_INIT */
      BUILD_BUG_ON(IO_CQ_WAKE_FORCE <= IORING_MAX_CQ_ENTRIES);

io_uring/io_uring.c void tctx_task_work(struct callback_head *cb)

Quoted 10 lines
io_uring/io_uring.c +3 -3
        /*
-     * We don't know how many reuqests is there in the link and whether
-     * they can even be queued lazily, fall back to non-lazy.
+     * We don't know how many requests are in the link and whether they can
+     * even be queued lazily, fall back to non-lazy.
       */
      if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK))
-        flags &= ~IOU_F_TWQ_LAZY_WAKE;
+        tw_flags &= ~IOU_F_TWQ_LAZY_WAKE;
  -    guard(rcu)();
Quoted 2 lines

protects against ctx->task deallocation, see a comment in io_ring_exit_work() -> synchronize_rcu()

Yeah that's just an editing mistake.

Quoted 8 lines

io_uring/io_uring.c void tctx_task_work(struct callback_head *cb) · line 1159

Quoted 3 lines
io_uring/io_uring.c +3 -0
+    spin_lock_irqsave(&ctx->work_lock, flags);
+    wq_list_add_tail(&req->io_task_work.work_node, &ctx->work_list);
+    nr_tw_prev = ctx->work_items++;

Is there a good reason why it changes the semantics of what's stored across adds? It was assigning a corrected nr_tw, this one will start heavily spamming with wake_up() in some cases.

Not sure I follow, how so? nr_tw_prev will be the previous count, just like before. Except we won't need to dig into the list to find it, we have it readily available. nr_tw will be the current code, or force wake if needed. As before.

Signature · 1 line
Jens Axboe
#5 lore raw

Re: [PATCH 2/6] io_uring: replace defer task_work llist with io_wq_work_list

Quoted 39 lines · Jens Axboe
Quoted 24 lines · Pavel Begunkov

On 11/22/24 16:12, Jens Axboe wrote:
...

Quoted 11 lines

static inline void io_req_local_work_add(struct io_kiocb *req,
struct io_ring_ctx *ctx,

io_uring/io_uring.c +3 -3
-                     unsigned flags)
+                     unsigned tw_flags)
   {
-    unsigned nr_wait, nr_tw, nr_tw_prev;
-    struct llist_node *head;
+    unsigned nr_tw, nr_tw_prev, nr_wait;
+    unsigned long flags;
         /* See comment above IO_CQ_WAKE_INIT */
       BUILD_BUG_ON(IO_CQ_WAKE_FORCE <= IORING_MAX_CQ_ENTRIES);

io_uring/io_uring.c void tctx_task_work(struct callback_head *cb)

Quoted 10 lines
io_uring/io_uring.c +3 -3
         /*
-     * We don't know how many reuqests is there in the link and whether
-     * they can even be queued lazily, fall back to non-lazy.
+     * We don't know how many requests are in the link and whether they can
+     * even be queued lazily, fall back to non-lazy.
        */
       if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK))
-        flags &= ~IOU_F_TWQ_LAZY_WAKE;
+        tw_flags &= ~IOU_F_TWQ_LAZY_WAKE;
   -    guard(rcu)();
Quoted 2 lines

protects against ctx->task deallocation, see a comment in io_ring_exit_work() -> synchronize_rcu()

Yeah that's just an editing mistake.

Quoted 8 lines

io_uring/io_uring.c void tctx_task_work(struct callback_head *cb) · line 1159

Quoted 3 lines
io_uring/io_uring.c +3 -0
+    spin_lock_irqsave(&ctx->work_lock, flags);
+    wq_list_add_tail(&req->io_task_work.work_node, &ctx->work_list);
+    nr_tw_prev = ctx->work_items++;

Is there a good reason why it changes the semantics of what's stored across adds? It was assigning a corrected nr_tw, this one will start heavily spamming with wake_up() in some cases.

Quoted 4 lines

Not sure I follow, how so? nr_tw_prev will be the previous count, just like before. Except we won't need to dig into the list to find it, we have it readily available. nr_tw will be the current code, or force wake if needed. As before.

The problem is what it stores, not how and where. Before req->nr_tw could've been set to IO_CQ_WAKE_FORCE, in which case following requests are not going to attempt waking up the task, now work_items is just a counter.

Let's say you've got a bunch of non-lazy adds coming close to each other. The first sets IO_CQ_WAKE_FORCE and wakes the task, and others just queue themselves in the list. Now, every single one of them will try to wake_up() as long as ->cq_wait_nr is large enough.

Signature · 1 line
Pavel Begunkov
author maintainer
#6 lore raw

Re: [PATCH 2/6] io_uring: replace defer task_work llist with io_wq_work_list

Quoted 50 lines · Pavel Begunkov
Quoted 39 lines · Jens Axboe
Quoted 24 lines · Pavel Begunkov

On 11/22/24 16:12, Jens Axboe wrote:
...

Quoted 11 lines

static inline void io_req_local_work_add(struct io_kiocb *req,
struct io_ring_ctx *ctx,

io_uring/io_uring.c +3 -3
-                     unsigned flags)
+                     unsigned tw_flags)
   {
-    unsigned nr_wait, nr_tw, nr_tw_prev;
-    struct llist_node *head;
+    unsigned nr_tw, nr_tw_prev, nr_wait;
+    unsigned long flags;
         /* See comment above IO_CQ_WAKE_INIT */
       BUILD_BUG_ON(IO_CQ_WAKE_FORCE <= IORING_MAX_CQ_ENTRIES);

io_uring/io_uring.c void tctx_task_work(struct callback_head *cb)

Quoted 10 lines
io_uring/io_uring.c +3 -3
         /*
-     * We don't know how many reuqests is there in the link and whether
-     * they can even be queued lazily, fall back to non-lazy.
+     * We don't know how many requests are in the link and whether they can
+     * even be queued lazily, fall back to non-lazy.
        */
       if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK))
-        flags &= ~IOU_F_TWQ_LAZY_WAKE;
+        tw_flags &= ~IOU_F_TWQ_LAZY_WAKE;
   -    guard(rcu)();
Quoted 2 lines

protects against ctx->task deallocation, see a comment in io_ring_exit_work() -> synchronize_rcu()

Yeah that's just an editing mistake.

Quoted 8 lines

io_uring/io_uring.c void tctx_task_work(struct callback_head *cb) · line 1159

Quoted 3 lines
io_uring/io_uring.c +3 -0
+    spin_lock_irqsave(&ctx->work_lock, flags);
+    wq_list_add_tail(&req->io_task_work.work_node, &ctx->work_list);
+    nr_tw_prev = ctx->work_items++;

Is there a good reason why it changes the semantics of what's stored across adds? It was assigning a corrected nr_tw, this one will start heavily spamming with wake_up() in some cases.

Quoted 4 lines

Not sure I follow, how so? nr_tw_prev will be the previous count, just like before. Except we won't need to dig into the list to find it, we have it readily available. nr_tw will be the current code, or force wake if needed. As before.

The problem is what it stores, not how and where. Before req->nr_tw could've been set to IO_CQ_WAKE_FORCE, in which case following requests are not going to attempt waking up the task, now work_items is just a counter.

Quoted 5 lines

Let's say you've got a bunch of non-lazy adds coming close to each other. The first sets IO_CQ_WAKE_FORCE and wakes the task, and others just queue themselves in the list. Now, every single one of them will try to wake_up() as long as ->cq_wait_nr is large enough.

If we really care about the non-lazy path as much, we can just use the same storing scheme as we did in req->nr_tw, except in ->work_items instead. Not a big deal imho.

Signature · 1 line
Jens Axboe
#7 lore raw

Re: [PATCH 2/6] io_uring: replace defer task_work llist with io_wq_work_list

Quoted 51 lines · Jens Axboe
Quoted 50 lines · Pavel Begunkov
Quoted 39 lines · Jens Axboe
Quoted 24 lines · Pavel Begunkov

On 11/22/24 16:12, Jens Axboe wrote:
...

Quoted 11 lines

static inline void io_req_local_work_add(struct io_kiocb *req,
struct io_ring_ctx *ctx,

io_uring/io_uring.c +3 -3
-                     unsigned flags)
+                     unsigned tw_flags)
    {
-    unsigned nr_wait, nr_tw, nr_tw_prev;
-    struct llist_node *head;
+    unsigned nr_tw, nr_tw_prev, nr_wait;
+    unsigned long flags;
          /* See comment above IO_CQ_WAKE_INIT */
        BUILD_BUG_ON(IO_CQ_WAKE_FORCE <= IORING_MAX_CQ_ENTRIES);

io_uring/io_uring.c void tctx_task_work(struct callback_head *cb)

Quoted 10 lines
io_uring/io_uring.c +3 -3
          /*
-     * We don't know how many reuqests is there in the link and whether
-     * they can even be queued lazily, fall back to non-lazy.
+     * We don't know how many requests are in the link and whether they can
+     * even be queued lazily, fall back to non-lazy.
         */
        if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK))
-        flags &= ~IOU_F_TWQ_LAZY_WAKE;
+        tw_flags &= ~IOU_F_TWQ_LAZY_WAKE;
    -    guard(rcu)();
Quoted 2 lines

protects against ctx->task deallocation, see a comment in io_ring_exit_work() -> synchronize_rcu()

Yeah that's just an editing mistake.

Quoted 8 lines

io_uring/io_uring.c void tctx_task_work(struct callback_head *cb) · line 1159

Quoted 3 lines
io_uring/io_uring.c +3 -0
+    spin_lock_irqsave(&ctx->work_lock, flags);
+    wq_list_add_tail(&req->io_task_work.work_node, &ctx->work_list);
+    nr_tw_prev = ctx->work_items++;

Is there a good reason why it changes the semantics of what's stored across adds? It was assigning a corrected nr_tw, this one will start heavily spamming with wake_up() in some cases.

Quoted 4 lines

Not sure I follow, how so? nr_tw_prev will be the previous count, just like before. Except we won't need to dig into the list to find it, we have it readily available. nr_tw will be the current code, or force wake if needed. As before.

The problem is what it stores, not how and where. Before req->nr_tw could've been set to IO_CQ_WAKE_FORCE, in which case following requests are not going to attempt waking up the task, now work_items is just a counter.

Quoted 7 lines
Quoted 5 lines

Let's say you've got a bunch of non-lazy adds coming close to each other. The first sets IO_CQ_WAKE_FORCE and wakes the task, and others just queue themselves in the list. Now, every single one of them will try to wake_up() as long as ->cq_wait_nr is large enough.

If we really care about the non-lazy path as much, we can just use the

Well, it's all linked requests, some of sendzc notif until I optimise it, maybe something else?

Quoted 2 lines

same storing scheme as we did in req->nr_tw, except in ->work_items instead. Not a big deal imho.

Yes please. It wouldn't be great sneaking them in the same commit either way.

Signature · 1 line
Pavel Begunkov