diff mbox series

[v2,1/2] ath10k: Keep track of which interrupts fired, don't poll them

Message ID 20200709082024.v2.1.I4d2f85ffa06f38532631e864a3125691ef5ffe06@changeid (mailing list archive)
State Accepted
Commit d66d24ac300cf41c6b88367fc9b4b6348679273d
Headers show
Series [v2,1/2] ath10k: Keep track of which interrupts fired, don't poll them | expand

Commit Message

Doug Anderson July 9, 2020, 3:21 p.m. UTC
If we have a per CE (Copy Engine) IRQ then we have no summary
register.  Right now the code generates a summary register by
iterating over all copy engines and seeing if they have an interrupt
pending.

This has a problem.  Specifically if _none_ if the Copy Engines have
an interrupt pending then they might go into low power mode and
reading from their address space will cause a full system crash.  This
was seen to happen when two interrupts went off at nearly the same
time.  Both were handled by a single call of ath10k_snoc_napi_poll()
but, because there were two interrupts handled and thus two calls to
napi_schedule() there was still a second call to
ath10k_snoc_napi_poll() which ran with no interrupts pending.

Instead of iterating over all the copy engines, let's just keep track
of the IRQs that fire.  Then we can effectively generate our own
summary without ever needing to read the Copy Engines.

Tested-on: WCN3990 SNOC WLAN.HL.3.2.2-00490-QCAHLSWMTPL-1

Signed-off-by: Douglas Anderson <dianders@chromium.org>
Reviewed-by: Rakesh Pillai <pillair@codeaurora.org>
Reviewed-by: Brian Norris <briannorris@chromium.org>
---
This patch continues work to try to squash all instances of the crash
we've been seeing while reading CE registers and hopefully this patch
addresses the true root of the issue.

The first patch that attempted to address these problems landed as
commit 8f9ed93d09a9 ("ath10k: Wait until copy complete is actually
done before completing").  After that Rakesh Pillai posted ("ath10k:
Add interrupt summary based CE processing") [1] and this patch is
based atop that one.  Both of those patches significantly reduced the
instances of problems but didn't fully eliminate them.  Crossing my
fingers that they're all gone now.

[1] https://lore.kernel.org/r/1593193967-29897-1-git-send-email-pillair@codeaurora.org

Changes in v2:
- Add bitmap_clear() in ath10k_snoc_hif_start().

 drivers/net/wireless/ath/ath10k/ce.c   | 84 ++++++++++----------------
 drivers/net/wireless/ath/ath10k/ce.h   | 14 ++---
 drivers/net/wireless/ath/ath10k/snoc.c | 19 ++++--
 drivers/net/wireless/ath/ath10k/snoc.h |  1 +
 4 files changed, 52 insertions(+), 66 deletions(-)

Comments

Doug Anderson Aug. 21, 2020, 9:25 p.m. UTC | #1
Kalle,

On Thu, Jul 9, 2020 at 8:22 AM Douglas Anderson <dianders@chromium.org> wrote:
>
> If we have a per CE (Copy Engine) IRQ then we have no summary
> register.  Right now the code generates a summary register by
> iterating over all copy engines and seeing if they have an interrupt
> pending.
>
> This has a problem.  Specifically if _none_ if the Copy Engines have
> an interrupt pending then they might go into low power mode and
> reading from their address space will cause a full system crash.  This
> was seen to happen when two interrupts went off at nearly the same
> time.  Both were handled by a single call of ath10k_snoc_napi_poll()
> but, because there were two interrupts handled and thus two calls to
> napi_schedule() there was still a second call to
> ath10k_snoc_napi_poll() which ran with no interrupts pending.
>
> Instead of iterating over all the copy engines, let's just keep track
> of the IRQs that fire.  Then we can effectively generate our own
> summary without ever needing to read the Copy Engines.
>
> Tested-on: WCN3990 SNOC WLAN.HL.3.2.2-00490-QCAHLSWMTPL-1
>
> Signed-off-by: Douglas Anderson <dianders@chromium.org>
> Reviewed-by: Rakesh Pillai <pillair@codeaurora.org>
> Reviewed-by: Brian Norris <briannorris@chromium.org>
> ---
> This patch continues work to try to squash all instances of the crash
> we've been seeing while reading CE registers and hopefully this patch
> addresses the true root of the issue.
>
> The first patch that attempted to address these problems landed as
> commit 8f9ed93d09a9 ("ath10k: Wait until copy complete is actually
> done before completing").  After that Rakesh Pillai posted ("ath10k:
> Add interrupt summary based CE processing") [1] and this patch is
> based atop that one.  Both of those patches significantly reduced the
> instances of problems but didn't fully eliminate them.  Crossing my
> fingers that they're all gone now.
>
> [1] https://lore.kernel.org/r/1593193967-29897-1-git-send-email-pillair@codeaurora.org
>
> Changes in v2:
> - Add bitmap_clear() in ath10k_snoc_hif_start().
>
>  drivers/net/wireless/ath/ath10k/ce.c   | 84 ++++++++++----------------
>  drivers/net/wireless/ath/ath10k/ce.h   | 14 ++---
>  drivers/net/wireless/ath/ath10k/snoc.c | 19 ++++--
>  drivers/net/wireless/ath/ath10k/snoc.h |  1 +
>  4 files changed, 52 insertions(+), 66 deletions(-)

I'm wondering if there's anything else you're looking for here.  If I
just need to sit tight that's fine, but I want to make sure this patch
isn't lost and you're not waiting for any actions on my part.  The
patch it depends on from Rakesh (see above or patchwork ID 11628289)
is also still marked as "Under Review".

We have been using this patch for the last few months and we haven't
hit a single crash like we were getting before.  At the same time, we
haven't found any regressions that have been attributed to this patch.

Anyway, just figured I'd check in.  Thanks!

-Doug
Marc Gonzalez Aug. 24, 2020, 8:36 a.m. UTC | #2
On 09/07/2020 17:21, Douglas Anderson wrote:

> If we have a per CE (Copy Engine) IRQ then we have no summary
> register.  Right now the code generates a summary register by
> iterating over all copy engines and seeing if they have an interrupt
> pending.
> 
> This has a problem.  Specifically if _none_ if the Copy Engines have

s/_none_ if/_none_ of

Hopefully, Kalle can just fixup before applying :-)

> an interrupt pending then they might go into low power mode and
> reading from their address space will cause a full system crash.  This
> was seen to happen when two interrupts went off at nearly the same
> time.  Both were handled by a single call of ath10k_snoc_napi_poll()
> but, because there were two interrupts handled and thus two calls to
> napi_schedule() there was still a second call to
> ath10k_snoc_napi_poll() which ran with no interrupts pending.
> 
> Instead of iterating over all the copy engines, let's just keep track
> of the IRQs that fire.  Then we can effectively generate our own
> summary without ever needing to read the Copy Engines.
Kalle Valo Aug. 26, 2020, 2:50 p.m. UTC | #3
Douglas Anderson <dianders@chromium.org> wrote:

> If we have a per CE (Copy Engine) IRQ then we have no summary
> register.  Right now the code generates a summary register by
> iterating over all copy engines and seeing if they have an interrupt
> pending.
> 
> This has a problem.  Specifically if _none_ if the Copy Engines have
> an interrupt pending then they might go into low power mode and
> reading from their address space will cause a full system crash.  This
> was seen to happen when two interrupts went off at nearly the same
> time.  Both were handled by a single call of ath10k_snoc_napi_poll()
> but, because there were two interrupts handled and thus two calls to
> napi_schedule() there was still a second call to
> ath10k_snoc_napi_poll() which ran with no interrupts pending.
> 
> Instead of iterating over all the copy engines, let's just keep track
> of the IRQs that fire.  Then we can effectively generate our own
> summary without ever needing to read the Copy Engines.
> 
> Tested-on: WCN3990 SNOC WLAN.HL.3.2.2-00490-QCAHLSWMTPL-1
> 
> Signed-off-by: Douglas Anderson <dianders@chromium.org>
> Reviewed-by: Rakesh Pillai <pillair@codeaurora.org>
> Reviewed-by: Brian Norris <briannorris@chromium.org>
> Signed-off-by: Kalle Valo <kvalo@codeaurora.org>

My main concern of this patch is that there's no info how it works on other
hardware families. For example, QCA9984 is very different from WCN3990. The
best would be if someone can provide a Tested-on tags for other hardware (even
some of them).

https://wireless.wiki.kernel.org/en/users/drivers/ath10k/submittingpatches#hardware_families
Doug Anderson Aug. 26, 2020, 2:59 p.m. UTC | #4
Hi,

On Wed, Aug 26, 2020 at 7:51 AM Kalle Valo <kvalo@codeaurora.org> wrote:
>
> Douglas Anderson <dianders@chromium.org> wrote:
>
> > If we have a per CE (Copy Engine) IRQ then we have no summary
> > register.  Right now the code generates a summary register by
> > iterating over all copy engines and seeing if they have an interrupt
> > pending.
> >
> > This has a problem.  Specifically if _none_ if the Copy Engines have
> > an interrupt pending then they might go into low power mode and
> > reading from their address space will cause a full system crash.  This
> > was seen to happen when two interrupts went off at nearly the same
> > time.  Both were handled by a single call of ath10k_snoc_napi_poll()
> > but, because there were two interrupts handled and thus two calls to
> > napi_schedule() there was still a second call to
> > ath10k_snoc_napi_poll() which ran with no interrupts pending.
> >
> > Instead of iterating over all the copy engines, let's just keep track
> > of the IRQs that fire.  Then we can effectively generate our own
> > summary without ever needing to read the Copy Engines.
> >
> > Tested-on: WCN3990 SNOC WLAN.HL.3.2.2-00490-QCAHLSWMTPL-1
> >
> > Signed-off-by: Douglas Anderson <dianders@chromium.org>
> > Reviewed-by: Rakesh Pillai <pillair@codeaurora.org>
> > Reviewed-by: Brian Norris <briannorris@chromium.org>
> > Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
>
> My main concern of this patch is that there's no info how it works on other
> hardware families. For example, QCA9984 is very different from WCN3990. The
> best would be if someone can provide a Tested-on tags for other hardware (even
> some of them).

I simply don't have access to any other Atheros hardware.  Hopefully
others on this thread do, though?  ...but, if nothing else, I believe
code inspection shows that the only places that are affected by the
changes here are:

* Wifi devices that use "snoc.c".  The only compatible string listed
in "snoc.c" is wcn3990.

* Wifi devices that set "per_ce_irq" to true.  The only place in the
table where this is set to true is wcn3990.

While it is certainly possible that I messed up and somehow affected
other WiFi devices, the common bits of code in "ce.c" and "ce.h" are
fairly easy to validate so hopefully they look OK?

-Doug
Kalle Valo Sept. 1, 2020, 12:06 p.m. UTC | #5
Douglas Anderson <dianders@chromium.org> wrote:

> If we have a per CE (Copy Engine) IRQ then we have no summary
> register.  Right now the code generates a summary register by
> iterating over all copy engines and seeing if they have an interrupt
> pending.
> 
> This has a problem.  Specifically if _none_ if the Copy Engines have
> an interrupt pending then they might go into low power mode and
> reading from their address space will cause a full system crash.  This
> was seen to happen when two interrupts went off at nearly the same
> time.  Both were handled by a single call of ath10k_snoc_napi_poll()
> but, because there were two interrupts handled and thus two calls to
> napi_schedule() there was still a second call to
> ath10k_snoc_napi_poll() which ran with no interrupts pending.
> 
> Instead of iterating over all the copy engines, let's just keep track
> of the IRQs that fire.  Then we can effectively generate our own
> summary without ever needing to read the Copy Engines.
> 
> Tested-on: WCN3990 SNOC WLAN.HL.3.2.2-00490-QCAHLSWMTPL-1
> 
> Signed-off-by: Douglas Anderson <dianders@chromium.org>
> Reviewed-by: Rakesh Pillai <pillair@codeaurora.org>
> Reviewed-by: Brian Norris <briannorris@chromium.org>
> Signed-off-by: Kalle Valo <kvalo@codeaurora.org>

2 patches applied to ath-next branch of ath.git, thanks.

d66d24ac300c ath10k: Keep track of which interrupts fired, don't poll them
7f8655166512 ath10k: Get rid of "per_ce_irq" hw param
Kalle Valo Sept. 1, 2020, 12:14 p.m. UTC | #6
Doug Anderson <dianders@chromium.org> writes:

> On Wed, Aug 26, 2020 at 7:51 AM Kalle Valo <kvalo@codeaurora.org> wrote:
>>
>> Douglas Anderson <dianders@chromium.org> wrote:
>>
>> > If we have a per CE (Copy Engine) IRQ then we have no summary
>> > register.  Right now the code generates a summary register by
>> > iterating over all copy engines and seeing if they have an interrupt
>> > pending.
>> >
>> > This has a problem.  Specifically if _none_ if the Copy Engines have
>> > an interrupt pending then they might go into low power mode and
>> > reading from their address space will cause a full system crash.  This
>> > was seen to happen when two interrupts went off at nearly the same
>> > time.  Both were handled by a single call of ath10k_snoc_napi_poll()
>> > but, because there were two interrupts handled and thus two calls to
>> > napi_schedule() there was still a second call to
>> > ath10k_snoc_napi_poll() which ran with no interrupts pending.
>> >
>> > Instead of iterating over all the copy engines, let's just keep track
>> > of the IRQs that fire.  Then we can effectively generate our own
>> > summary without ever needing to read the Copy Engines.
>> >
>> > Tested-on: WCN3990 SNOC WLAN.HL.3.2.2-00490-QCAHLSWMTPL-1
>> >
>> > Signed-off-by: Douglas Anderson <dianders@chromium.org>
>> > Reviewed-by: Rakesh Pillai <pillair@codeaurora.org>
>> > Reviewed-by: Brian Norris <briannorris@chromium.org>
>> > Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
>>
>> My main concern of this patch is that there's no info how it works on other
>> hardware families. For example, QCA9984 is very different from WCN3990. The
>> best would be if someone can provide a Tested-on tags for other hardware (even
>> some of them).
>
> I simply don't have access to any other Atheros hardware.  Hopefully
> others on this thread do, though?

I have the hardware but in practise no time to do the testing :/

> ...but, if nothing else, I believe code inspection shows that the only
> places that are affected by the changes here are:
>
> * Wifi devices that use "snoc.c".  The only compatible string listed
> in "snoc.c" is wcn3990.
>
> * Wifi devices that set "per_ce_irq" to true.  The only place in the
> table where this is set to true is wcn3990.
>
> While it is certainly possible that I messed up and somehow affected
> other WiFi devices, the common bits of code in "ce.c" and "ce.h" are
> fairly easy to validate so hopefully they look OK?

Basically I would like to see some evidence in the commit log that _all_
hardware families are taken into account to avoid any regressions, be it
testing or at least thorough review. I see way too many patches where
people are working just on one hardware/firmware combo and not giving a
single thought how it would work on other hardware.

But I applied the three patches now, let's hope they are ok. At least I
was not able to find any problems during review, but of course real
testing would be better than just review.
diff mbox series

Patch

diff --git a/drivers/net/wireless/ath/ath10k/ce.c b/drivers/net/wireless/ath/ath10k/ce.c
index 1e16f263854a..84ec80c6d08f 100644
--- a/drivers/net/wireless/ath/ath10k/ce.c
+++ b/drivers/net/wireless/ath/ath10k/ce.c
@@ -481,38 +481,6 @@  static inline void ath10k_ce_engine_int_status_clear(struct ath10k *ar,
 	ath10k_ce_write32(ar, ce_ctrl_addr + wm_regs->addr, mask);
 }
 
-static bool ath10k_ce_engine_int_status_check(struct ath10k *ar, u32 ce_ctrl_addr,
-					      unsigned int mask)
-{
-	struct ath10k_hw_ce_host_wm_regs *wm_regs = ar->hw_ce_regs->wm_regs;
-
-	return ath10k_ce_read32(ar, ce_ctrl_addr + wm_regs->addr) & mask;
-}
-
-u32 ath10k_ce_gen_interrupt_summary(struct ath10k *ar)
-{
-	struct ath10k_hw_ce_host_wm_regs *wm_regs = ar->hw_ce_regs->wm_regs;
-	struct ath10k_ce_pipe *ce_state;
-	struct ath10k_ce *ce;
-	u32 irq_summary = 0;
-	u32 ctrl_addr;
-	u32 ce_id;
-
-	ce = ath10k_ce_priv(ar);
-
-	for (ce_id = 0; ce_id < CE_COUNT; ce_id++) {
-		ce_state = &ce->ce_states[ce_id];
-		ctrl_addr = ce_state->ctrl_addr;
-		if (ath10k_ce_engine_int_status_check(ar, ctrl_addr,
-						      wm_regs->cc_mask)) {
-			irq_summary |= BIT(ce_id);
-		}
-	}
-
-	return irq_summary;
-}
-EXPORT_SYMBOL(ath10k_ce_gen_interrupt_summary);
-
 /*
  * Guts of ath10k_ce_send.
  * The caller takes responsibility for any needed locking.
@@ -1399,45 +1367,55 @@  static void ath10k_ce_per_engine_handler_adjust(struct ath10k_ce_pipe *ce_state)
 	ath10k_ce_watermark_intr_disable(ar, ctrl_addr);
 }
 
-int ath10k_ce_disable_interrupts(struct ath10k *ar)
+void ath10k_ce_disable_interrupt(struct ath10k *ar, int ce_id)
 {
 	struct ath10k_ce *ce = ath10k_ce_priv(ar);
 	struct ath10k_ce_pipe *ce_state;
 	u32 ctrl_addr;
-	int ce_id;
 
-	for (ce_id = 0; ce_id < CE_COUNT; ce_id++) {
-		ce_state  = &ce->ce_states[ce_id];
-		if (ce_state->attr_flags & CE_ATTR_POLL)
-			continue;
+	ce_state  = &ce->ce_states[ce_id];
+	if (ce_state->attr_flags & CE_ATTR_POLL)
+		return;
 
-		ctrl_addr = ath10k_ce_base_address(ar, ce_id);
+	ctrl_addr = ath10k_ce_base_address(ar, ce_id);
 
-		ath10k_ce_copy_complete_intr_disable(ar, ctrl_addr);
-		ath10k_ce_error_intr_disable(ar, ctrl_addr);
-		ath10k_ce_watermark_intr_disable(ar, ctrl_addr);
-	}
+	ath10k_ce_copy_complete_intr_disable(ar, ctrl_addr);
+	ath10k_ce_error_intr_disable(ar, ctrl_addr);
+	ath10k_ce_watermark_intr_disable(ar, ctrl_addr);
+}
+EXPORT_SYMBOL(ath10k_ce_disable_interrupt);
 
-	return 0;
+void ath10k_ce_disable_interrupts(struct ath10k *ar)
+{
+	int ce_id;
+
+	for (ce_id = 0; ce_id < CE_COUNT; ce_id++)
+		ath10k_ce_disable_interrupt(ar, ce_id);
 }
 EXPORT_SYMBOL(ath10k_ce_disable_interrupts);
 
-void ath10k_ce_enable_interrupts(struct ath10k *ar)
+void ath10k_ce_enable_interrupt(struct ath10k *ar, int ce_id)
 {
 	struct ath10k_ce *ce = ath10k_ce_priv(ar);
-	int ce_id;
 	struct ath10k_ce_pipe *ce_state;
 
+	ce_state  = &ce->ce_states[ce_id];
+	if (ce_state->attr_flags & CE_ATTR_POLL)
+		return;
+
+	ath10k_ce_per_engine_handler_adjust(ce_state);
+}
+EXPORT_SYMBOL(ath10k_ce_enable_interrupt);
+
+void ath10k_ce_enable_interrupts(struct ath10k *ar)
+{
+	int ce_id;
+
 	/* Enable interrupts for copy engine that
 	 * are not using polling mode.
 	 */
-	for (ce_id = 0; ce_id < CE_COUNT; ce_id++) {
-		ce_state  = &ce->ce_states[ce_id];
-		if (ce_state->attr_flags & CE_ATTR_POLL)
-			continue;
-
-		ath10k_ce_per_engine_handler_adjust(ce_state);
-	}
+	for (ce_id = 0; ce_id < CE_COUNT; ce_id++)
+		ath10k_ce_enable_interrupt(ar, ce_id);
 }
 EXPORT_SYMBOL(ath10k_ce_enable_interrupts);
 
diff --git a/drivers/net/wireless/ath/ath10k/ce.h b/drivers/net/wireless/ath/ath10k/ce.h
index a440aaf74aa4..666ce384a1d8 100644
--- a/drivers/net/wireless/ath/ath10k/ce.h
+++ b/drivers/net/wireless/ath/ath10k/ce.h
@@ -255,12 +255,13 @@  int ath10k_ce_cancel_send_next(struct ath10k_ce_pipe *ce_state,
 /*==================CE Interrupt Handlers====================*/
 void ath10k_ce_per_engine_service_any(struct ath10k *ar);
 void ath10k_ce_per_engine_service(struct ath10k *ar, unsigned int ce_id);
-int ath10k_ce_disable_interrupts(struct ath10k *ar);
+void ath10k_ce_disable_interrupt(struct ath10k *ar, int ce_id);
+void ath10k_ce_disable_interrupts(struct ath10k *ar);
+void ath10k_ce_enable_interrupt(struct ath10k *ar, int ce_id);
 void ath10k_ce_enable_interrupts(struct ath10k *ar);
 void ath10k_ce_dump_registers(struct ath10k *ar,
 			      struct ath10k_fw_crash_data *crash_data);
 
-u32 ath10k_ce_gen_interrupt_summary(struct ath10k *ar);
 void ath10k_ce_alloc_rri(struct ath10k *ar);
 void ath10k_ce_free_rri(struct ath10k *ar);
 
@@ -376,12 +377,9 @@  static inline u32 ath10k_ce_interrupt_summary(struct ath10k *ar)
 {
 	struct ath10k_ce *ce = ath10k_ce_priv(ar);
 
-	if (!ar->hw_params.per_ce_irq)
-		return CE_WRAPPER_INTERRUPT_SUMMARY_HOST_MSI_GET(
-			ce->bus_ops->read32((ar), CE_WRAPPER_BASE_ADDRESS +
-			CE_WRAPPER_INTERRUPT_SUMMARY_ADDRESS));
-	else
-		return ath10k_ce_gen_interrupt_summary(ar);
+	return CE_WRAPPER_INTERRUPT_SUMMARY_HOST_MSI_GET(
+		ce->bus_ops->read32((ar), CE_WRAPPER_BASE_ADDRESS +
+		CE_WRAPPER_INTERRUPT_SUMMARY_ADDRESS));
 }
 
 /* Host software's Copy Engine configuration. */
diff --git a/drivers/net/wireless/ath/ath10k/snoc.c b/drivers/net/wireless/ath/ath10k/snoc.c
index 354d49b1cd45..1ef5fdb8248b 100644
--- a/drivers/net/wireless/ath/ath10k/snoc.c
+++ b/drivers/net/wireless/ath/ath10k/snoc.c
@@ -3,6 +3,7 @@ 
  * Copyright (c) 2018 The Linux Foundation. All rights reserved.
  */
 
+#include <linux/bits.h>
 #include <linux/clk.h>
 #include <linux/kernel.h>
 #include <linux/module.h>
@@ -923,6 +924,7 @@  static int ath10k_snoc_hif_start(struct ath10k *ar)
 {
 	struct ath10k_snoc *ar_snoc = ath10k_snoc_priv(ar);
 
+	bitmap_clear(ar_snoc->pending_ce_irqs, 0, CE_COUNT_MAX);
 	napi_enable(&ar->napi);
 	ath10k_snoc_irq_enable(ar);
 	ath10k_snoc_rx_post(ar);
@@ -1158,7 +1160,9 @@  static irqreturn_t ath10k_snoc_per_engine_handler(int irq, void *arg)
 		return IRQ_HANDLED;
 	}
 
-	ath10k_snoc_irq_disable(ar);
+	ath10k_ce_disable_interrupt(ar, ce_id);
+	set_bit(ce_id, ar_snoc->pending_ce_irqs);
+
 	napi_schedule(&ar->napi);
 
 	return IRQ_HANDLED;
@@ -1167,20 +1171,25 @@  static irqreturn_t ath10k_snoc_per_engine_handler(int irq, void *arg)
 static int ath10k_snoc_napi_poll(struct napi_struct *ctx, int budget)
 {
 	struct ath10k *ar = container_of(ctx, struct ath10k, napi);
+	struct ath10k_snoc *ar_snoc = ath10k_snoc_priv(ar);
 	int done = 0;
+	int ce_id;
 
 	if (test_bit(ATH10K_FLAG_CRASH_FLUSH, &ar->dev_flags)) {
 		napi_complete(ctx);
 		return done;
 	}
 
-	ath10k_ce_per_engine_service_any(ar);
+	for (ce_id = 0; ce_id < CE_COUNT; ce_id++)
+		if (test_and_clear_bit(ce_id, ar_snoc->pending_ce_irqs)) {
+			ath10k_ce_per_engine_service(ar, ce_id);
+			ath10k_ce_enable_interrupt(ar, ce_id);
+		}
+
 	done = ath10k_htt_txrx_compl_task(ar, budget);
 
-	if (done < budget) {
+	if (done < budget)
 		napi_complete(ctx);
-		ath10k_snoc_irq_enable(ar);
-	}
 
 	return done;
 }
diff --git a/drivers/net/wireless/ath/ath10k/snoc.h b/drivers/net/wireless/ath/ath10k/snoc.h
index a3dd06f6ac62..5095d1893681 100644
--- a/drivers/net/wireless/ath/ath10k/snoc.h
+++ b/drivers/net/wireless/ath/ath10k/snoc.h
@@ -78,6 +78,7 @@  struct ath10k_snoc {
 	unsigned long flags;
 	bool xo_cal_supported;
 	u32 xo_cal_data;
+	DECLARE_BITMAP(pending_ce_irqs, CE_COUNT_MAX);
 };
 
 static inline struct ath10k_snoc *ath10k_snoc_priv(struct ath10k *ar)