[v13,2/7] net: introduce napi_busy_loop_rcu()

Commit Message

Stefan Roesch May 18, 2023, 9:17 p.m. UTC

This introduces the napi_busy_loop_rcu() function. If the caller of
napi_busy_loop() function is also taking the rcu read lock, it is possible
that napi_busy_loop() is releasing the read lock if it invokes schedule.
However the caller is expecting that the rcu read lock is not released
until the function completes. This new function avoids that problem. It
expects that the caller MUST hold the rcu_read_lock while calling this
function.

Signed-off-by: Stefan Roesch <[e-mail]>
---
 include/net/busy_poll.h |  4 ++++
 net/core/dev.c          | 38 ++++++++++++++++++++++++++++++++++++++
 2 files changed, 42 insertions(+)

Comments

Jakub Kicinski May 31, 2023, 5:38 p.m. UTC | #1

On Wed, 31 May 2023 10:29:15 -0700 Jakub Kicinski wrote:
> I think it's worth noting that until the non-RCU version this one may
> exit without calling need_resched().

My brain is broken. I meant to say "unlike non-RCU version this one may
exit for reasons other than loop_end()'s decision".

Stefan Roesch June 5, 2023, 5:45 p.m. UTC | #2

Jakub Kicinski <[e-mail]> writes:

> On Thu, 18 May 2023 14:17:46 -0700 Stefan Roesch wrote:
>> + * napi_busy_loop_rcu() - busy poll (caller holds rcu read lock)
>> + * @napi_id         : napi id
>> + * @loop_end        : function to check for loop end
>> + * @loop_end_arg    : argument for loop end function
>> + * @prefer_busy_poll: prefer busy poll
>> + * @budget          : budget for busy poll
>> + *
>> + * Warning: caller must hold rcu read lock.
>
> I think it's worth noting that until the non-RCU version this one may
> exit without calling need_resched(). Otherwise the entire kdoc can go
> IMO, it adds no value. It just restates the arguments and the RCU-iness
> is obviously implied by the suffix.

I fixed the comment.

Patch

diff --git a/include/net/busy_poll.h b/include/net/busy_poll.h
index f90f0021f5f2..622623f5740e 100644
--- a/include/net/busy_poll.h
+++ b/include/net/busy_poll.h
@@ -47,6 +47,10 @@  void napi_busy_loop(unsigned int napi_id,
 		    bool (*loop_end)(void *, unsigned long),
 		    void *loop_end_arg, bool prefer_busy_poll, u16 budget);
 
+void napi_busy_loop_rcu(unsigned int napi_id,
+			bool (*loop_end)(void *, unsigned long),
+			void *loop_end_arg, bool prefer_busy_poll, u16 budget);
+
 #else /* CONFIG_NET_RX_BUSY_POLL */
 static inline unsigned long net_busy_loop_on(void)
 {
diff --git a/net/core/dev.c b/net/core/dev.c
index f4677aa20f84..51a1c70a70a2 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -6213,6 +6213,44 @@  static inline void __napi_busy_poll(struct napi_busy_poll_ctx *ctx,
 				LINUX_MIB_BUSYPOLLRXPACKETS, work);
 	local_bh_enable();
 }
+
+/*
+ * napi_busy_loop_rcu() - busy poll (caller holds rcu read lock)
+ * @napi_id         : napi id
+ * @loop_end        : function to check for loop end
+ * @loop_end_arg    : argument for loop end function
+ * @prefer_busy_poll: prefer busy poll
+ * @budget          : budget for busy poll
+ *
+ * Warning: caller must hold rcu read lock.
+ */
+void napi_busy_loop_rcu(unsigned int napi_id,
+		    bool (*loop_end)(void *, unsigned long),
+		    void *loop_end_arg, bool prefer_busy_poll, u16 budget)
+{
+	unsigned long start_time = loop_end ? busy_loop_current_time() : 0;
+	struct napi_busy_poll_ctx ctx = {};
+
+	ctx.napi = napi_by_id(napi_id);
+	if (!ctx.napi)
+		return;
+
+	preempt_disable();
+	for (;;) {
+		__napi_busy_poll(&ctx, prefer_busy_poll, budget);
+
+		if (!loop_end || loop_end(loop_end_arg, start_time))
+			break;
+		if (unlikely(need_resched()))
+			break;
+
+		cpu_relax();
+	}
+	if (ctx.napi_poll)
+		busy_poll_stop(ctx.napi, ctx.have_poll_lock, prefer_busy_poll, budget);
+	preempt_enable();
+}
+
 void napi_busy_loop(unsigned int napi_id,
 		    bool (*loop_end)(void *, unsigned long),
 		    void *loop_end_arg, bool prefer_busy_poll, u16 budget)