| Message ID | 20200609082015.1.Ife398994e5a0a6830e4d4a16306ef36e0144e7ba@changeid (mailing list archive) |
|---|---|
| State | Accepted |
| Commit | 8f9ed93d09a97444733d492a3bbf66bcb786a777 |
| Delegated to: | Kalle Valo |
| Headers | show |
| Series | ath10k: Wait until copy complete is actually done before completing | expand |
Hi Doug, The send callback for the CEs do check for hw_index/SRRI before trying to free the buffer. But adding a check for copy-complete (before calling the individual CE callbacks) seems to be the better approach. Hence I agree that this patch should be added. Thanks, Rakesh Pillai. On 2020-06-09 20:50, Douglas Anderson wrote: > On wcn3990 we have "per_ce_irq = true". That makes the > ath10k_ce_interrupt_summary() function always return 0xfff. The > ath10k_ce_per_engine_service_any() function will see this and think > that _all_ copy engines have an interrupt. Without checking, the > ath10k_ce_per_engine_service() assumes that if it's called that the > "copy complete" (cc) interrupt fired. This combination seems bad. > > Let's add a check to make sure that the "copy complete" interrupt > actually fired in ath10k_ce_per_engine_service(). > > This might fix a hard-to-reproduce failure where it appears that the > copy complete handlers run before the copy is really complete. > Specifically a symptom was that we were seeing this on a Qualcomm > sc7180 board: > arm-smmu 15000000.iommu: Unhandled context fault: > fsr=0x402, iova=0x7fdd45780, fsynr=0x30003, cbfrsynra=0xc1, cb=10 > > Even on platforms that don't have wcn3990 this still seems like it > would be a sane thing to do. Specifically the current IRQ handler > comments indicate that there might be other misc interrupt sources > firing that need to be cleared. If one of those sources was the one > that caused the IRQ handler to be called it would also be important to > double-check that the interrupt we cared about actually fired. > > Signed-off-by: Douglas Anderson <dianders@chromium.org> Reviewed-by: Rakesh Pillai <pillair@codeaurora.org> > --- > > drivers/net/wireless/ath/ath10k/ce.c | 30 +++++++++++++++++++--------- > 1 file changed, 21 insertions(+), 9 deletions(-) > > diff --git a/drivers/net/wireless/ath/ath10k/ce.c > b/drivers/net/wireless/ath/ath10k/ce.c > index 294fbc1e89ab..ffdd4b995f33 100644 > --- a/drivers/net/wireless/ath/ath10k/ce.c > +++ b/drivers/net/wireless/ath/ath10k/ce.c > @@ -481,6 +481,15 @@ static inline void > ath10k_ce_engine_int_status_clear(struct ath10k *ar, > ath10k_ce_write32(ar, ce_ctrl_addr + wm_regs->addr, mask); > } > > +static inline 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; > +} > + > /* > * Guts of ath10k_ce_send. > * The caller takes responsibility for any needed locking. > @@ -1301,19 +1310,22 @@ void ath10k_ce_per_engine_service(struct > ath10k *ar, unsigned int ce_id) > > spin_lock_bh(&ce->ce_lock); > > - /* Clear the copy-complete interrupts that will be handled here. */ > - ath10k_ce_engine_int_status_clear(ar, ctrl_addr, > - wm_regs->cc_mask); > + if (ath10k_ce_engine_int_status_check(ar, ctrl_addr, > + wm_regs->cc_mask)) { > + /* Clear before handling */ > + ath10k_ce_engine_int_status_clear(ar, ctrl_addr, > + wm_regs->cc_mask); > > - spin_unlock_bh(&ce->ce_lock); > + spin_unlock_bh(&ce->ce_lock); > > - if (ce_state->recv_cb) > - ce_state->recv_cb(ce_state); > + if (ce_state->recv_cb) > + ce_state->recv_cb(ce_state); > > - if (ce_state->send_cb) > - ce_state->send_cb(ce_state); > + if (ce_state->send_cb) > + ce_state->send_cb(ce_state); > > - spin_lock_bh(&ce->ce_lock); > + spin_lock_bh(&ce->ce_lock); > + } > > /* > * Misc CE interrupts are not being handled, but still need
Douglas Anderson <dianders@chromium.org> wrote: > On wcn3990 we have "per_ce_irq = true". That makes the > ath10k_ce_interrupt_summary() function always return 0xfff. The > ath10k_ce_per_engine_service_any() function will see this and think > that _all_ copy engines have an interrupt. Without checking, the > ath10k_ce_per_engine_service() assumes that if it's called that the > "copy complete" (cc) interrupt fired. This combination seems bad. > > Let's add a check to make sure that the "copy complete" interrupt > actually fired in ath10k_ce_per_engine_service(). > > This might fix a hard-to-reproduce failure where it appears that the > copy complete handlers run before the copy is really complete. > Specifically a symptom was that we were seeing this on a Qualcomm > sc7180 board: > arm-smmu 15000000.iommu: Unhandled context fault: > fsr=0x402, iova=0x7fdd45780, fsynr=0x30003, cbfrsynra=0xc1, cb=10 > > Even on platforms that don't have wcn3990 this still seems like it > would be a sane thing to do. Specifically the current IRQ handler > comments indicate that there might be other misc interrupt sources > firing that need to be cleared. If one of those sources was the one > that caused the IRQ handler to be called it would also be important to > double-check that the interrupt we cared about actually fired. > > Signed-off-by: Douglas Anderson <dianders@chromium.org> > Signed-off-by: Kalle Valo <kvalo@codeaurora.org> ath10k firmwares work very differently, on what hardware and firmware did you test this? I'll add that information to the commit log.
Hi, On Mon, Jun 15, 2020 at 7:32 AM Kalle Valo <kvalo@codeaurora.org> wrote: > > Douglas Anderson <dianders@chromium.org> wrote: > > > On wcn3990 we have "per_ce_irq = true". That makes the > > ath10k_ce_interrupt_summary() function always return 0xfff. The > > ath10k_ce_per_engine_service_any() function will see this and think > > that _all_ copy engines have an interrupt. Without checking, the > > ath10k_ce_per_engine_service() assumes that if it's called that the > > "copy complete" (cc) interrupt fired. This combination seems bad. > > > > Let's add a check to make sure that the "copy complete" interrupt > > actually fired in ath10k_ce_per_engine_service(). > > > > This might fix a hard-to-reproduce failure where it appears that the > > copy complete handlers run before the copy is really complete. > > Specifically a symptom was that we were seeing this on a Qualcomm > > sc7180 board: > > arm-smmu 15000000.iommu: Unhandled context fault: > > fsr=0x402, iova=0x7fdd45780, fsynr=0x30003, cbfrsynra=0xc1, cb=10 > > > > Even on platforms that don't have wcn3990 this still seems like it > > would be a sane thing to do. Specifically the current IRQ handler > > comments indicate that there might be other misc interrupt sources > > firing that need to be cleared. If one of those sources was the one > > that caused the IRQ handler to be called it would also be important to > > double-check that the interrupt we cared about actually fired. > > > > Signed-off-by: Douglas Anderson <dianders@chromium.org> > > Signed-off-by: Kalle Valo <kvalo@codeaurora.org> > > ath10k firmwares work very differently, on what hardware and firmware did you > test this? I'll add that information to the commit log. I am running on a Qualcomm sc7180 SoC. > -- > https://patchwork.kernel.org/patch/11595887/ > > https://wireless.wiki.kernel.org/en/developers/documentation/submittingpatches >
Doug Anderson <dianders@chromium.org> writes: > On Mon, Jun 15, 2020 at 7:32 AM Kalle Valo <kvalo@codeaurora.org> wrote: >> >> Douglas Anderson <dianders@chromium.org> wrote: >> >> > On wcn3990 we have "per_ce_irq = true". That makes the >> > ath10k_ce_interrupt_summary() function always return 0xfff. The >> > ath10k_ce_per_engine_service_any() function will see this and think >> > that _all_ copy engines have an interrupt. Without checking, the >> > ath10k_ce_per_engine_service() assumes that if it's called that the >> > "copy complete" (cc) interrupt fired. This combination seems bad. >> > >> > Let's add a check to make sure that the "copy complete" interrupt >> > actually fired in ath10k_ce_per_engine_service(). >> > >> > This might fix a hard-to-reproduce failure where it appears that the >> > copy complete handlers run before the copy is really complete. >> > Specifically a symptom was that we were seeing this on a Qualcomm >> > sc7180 board: >> > arm-smmu 15000000.iommu: Unhandled context fault: >> > fsr=0x402, iova=0x7fdd45780, fsynr=0x30003, cbfrsynra=0xc1, cb=10 >> > >> > Even on platforms that don't have wcn3990 this still seems like it >> > would be a sane thing to do. Specifically the current IRQ handler >> > comments indicate that there might be other misc interrupt sources >> > firing that need to be cleared. If one of those sources was the one >> > that caused the IRQ handler to be called it would also be important to >> > double-check that the interrupt we cared about actually fired. >> > >> > Signed-off-by: Douglas Anderson <dianders@chromium.org> >> > Signed-off-by: Kalle Valo <kvalo@codeaurora.org> >> >> ath10k firmwares work very differently, on what hardware and firmware did you >> test this? I'll add that information to the commit log. > > I am running on a Qualcomm sc7180 SoC. Sorry, I was unclear, I meant the ath10k hardware :) I guess WCN3990 but what firmware version?
Hi, On Mon, Jun 15, 2020 at 7:56 AM Kalle Valo <kvalo@codeaurora.org> wrote: > > Doug Anderson <dianders@chromium.org> writes: > > > On Mon, Jun 15, 2020 at 7:32 AM Kalle Valo <kvalo@codeaurora.org> wrote: > >> > >> Douglas Anderson <dianders@chromium.org> wrote: > >> > >> > On wcn3990 we have "per_ce_irq = true". That makes the > >> > ath10k_ce_interrupt_summary() function always return 0xfff. The > >> > ath10k_ce_per_engine_service_any() function will see this and think > >> > that _all_ copy engines have an interrupt. Without checking, the > >> > ath10k_ce_per_engine_service() assumes that if it's called that the > >> > "copy complete" (cc) interrupt fired. This combination seems bad. > >> > > >> > Let's add a check to make sure that the "copy complete" interrupt > >> > actually fired in ath10k_ce_per_engine_service(). > >> > > >> > This might fix a hard-to-reproduce failure where it appears that the > >> > copy complete handlers run before the copy is really complete. > >> > Specifically a symptom was that we were seeing this on a Qualcomm > >> > sc7180 board: > >> > arm-smmu 15000000.iommu: Unhandled context fault: > >> > fsr=0x402, iova=0x7fdd45780, fsynr=0x30003, cbfrsynra=0xc1, cb=10 > >> > > >> > Even on platforms that don't have wcn3990 this still seems like it > >> > would be a sane thing to do. Specifically the current IRQ handler > >> > comments indicate that there might be other misc interrupt sources > >> > firing that need to be cleared. If one of those sources was the one > >> > that caused the IRQ handler to be called it would also be important to > >> > double-check that the interrupt we cared about actually fired. > >> > > >> > Signed-off-by: Douglas Anderson <dianders@chromium.org> > >> > Signed-off-by: Kalle Valo <kvalo@codeaurora.org> > >> > >> ath10k firmwares work very differently, on what hardware and firmware did you > >> test this? I'll add that information to the commit log. > > > > I am running on a Qualcomm sc7180 SoC. > > Sorry, I was unclear, I meant the ath10k hardware :) I guess WCN3990 but > what firmware version? Ah, sorry! Yes, it appears to be wcn3990 based on my device tree: $ git grep -A2 wifi -- arch/arm64/boot/dts/qcom/sc7180.dtsi wifi: wifi@18800000 { compatible = "qcom,wcn3990-wifi"; reg = <0 0x18800000 0 0x800000>; reg-names = "membase"; Firmware isn't final yet, but currently my boot log shows: qmi fw_version 0x322a01ea fw_build_timestamp 2020-05-20 03:47 QC_IMAGE_VERSION_STRING=WLAN.HL.3.2.2-00490-QCAHLSWMTPL-1 -Doug
Douglas Anderson <dianders@chromium.org> wrote: > On wcn3990 we have "per_ce_irq = true". That makes the > ath10k_ce_interrupt_summary() function always return 0xfff. The > ath10k_ce_per_engine_service_any() function will see this and think > that _all_ copy engines have an interrupt. Without checking, the > ath10k_ce_per_engine_service() assumes that if it's called that the > "copy complete" (cc) interrupt fired. This combination seems bad. > > Let's add a check to make sure that the "copy complete" interrupt > actually fired in ath10k_ce_per_engine_service(). > > This might fix a hard-to-reproduce failure where it appears that the > copy complete handlers run before the copy is really complete. > Specifically a symptom was that we were seeing this on a Qualcomm > sc7180 board: > arm-smmu 15000000.iommu: Unhandled context fault: > fsr=0x402, iova=0x7fdd45780, fsynr=0x30003, cbfrsynra=0xc1, cb=10 > > Even on platforms that don't have wcn3990 this still seems like it > would be a sane thing to do. Specifically the current IRQ handler > comments indicate that there might be other misc interrupt sources > firing that need to be cleared. If one of those sources was the one > that caused the IRQ handler to be called it would also be important to > double-check that the interrupt we cared about actually fired. > > Tested-on: WCN3990 SNOC WLAN.HL.3.2.2-00490-QCAHLSWMTPL-1 > > Signed-off-by: Douglas Anderson <dianders@chromium.org> > Signed-off-by: Kalle Valo <kvalo@codeaurora.org> Patch applied to ath-next branch of ath.git, thanks. 8f9ed93d09a9 ath10k: Wait until copy complete is actually done before completing
diff --git a/drivers/net/wireless/ath/ath10k/ce.c b/drivers/net/wireless/ath/ath10k/ce.c index 294fbc1e89ab..ffdd4b995f33 100644 --- a/drivers/net/wireless/ath/ath10k/ce.c +++ b/drivers/net/wireless/ath/ath10k/ce.c @@ -481,6 +481,15 @@ static inline void ath10k_ce_engine_int_status_clear(struct ath10k *ar, ath10k_ce_write32(ar, ce_ctrl_addr + wm_regs->addr, mask); } +static inline 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; +} + /* * Guts of ath10k_ce_send. * The caller takes responsibility for any needed locking. @@ -1301,19 +1310,22 @@ void ath10k_ce_per_engine_service(struct ath10k *ar, unsigned int ce_id) spin_lock_bh(&ce->ce_lock); - /* Clear the copy-complete interrupts that will be handled here. */ - ath10k_ce_engine_int_status_clear(ar, ctrl_addr, - wm_regs->cc_mask); + if (ath10k_ce_engine_int_status_check(ar, ctrl_addr, + wm_regs->cc_mask)) { + /* Clear before handling */ + ath10k_ce_engine_int_status_clear(ar, ctrl_addr, + wm_regs->cc_mask); - spin_unlock_bh(&ce->ce_lock); + spin_unlock_bh(&ce->ce_lock); - if (ce_state->recv_cb) - ce_state->recv_cb(ce_state); + if (ce_state->recv_cb) + ce_state->recv_cb(ce_state); - if (ce_state->send_cb) - ce_state->send_cb(ce_state); + if (ce_state->send_cb) + ce_state->send_cb(ce_state); - spin_lock_bh(&ce->ce_lock); + spin_lock_bh(&ce->ce_lock); + } /* * Misc CE interrupts are not being handled, but still need
On wcn3990 we have "per_ce_irq = true". That makes the ath10k_ce_interrupt_summary() function always return 0xfff. The ath10k_ce_per_engine_service_any() function will see this and think that _all_ copy engines have an interrupt. Without checking, the ath10k_ce_per_engine_service() assumes that if it's called that the "copy complete" (cc) interrupt fired. This combination seems bad. Let's add a check to make sure that the "copy complete" interrupt actually fired in ath10k_ce_per_engine_service(). This might fix a hard-to-reproduce failure where it appears that the copy complete handlers run before the copy is really complete. Specifically a symptom was that we were seeing this on a Qualcomm sc7180 board: arm-smmu 15000000.iommu: Unhandled context fault: fsr=0x402, iova=0x7fdd45780, fsynr=0x30003, cbfrsynra=0xc1, cb=10 Even on platforms that don't have wcn3990 this still seems like it would be a sane thing to do. Specifically the current IRQ handler comments indicate that there might be other misc interrupt sources firing that need to be cleared. If one of those sources was the one that caused the IRQ handler to be called it would also be important to double-check that the interrupt we cared about actually fired. Signed-off-by: Douglas Anderson <dianders@chromium.org> --- drivers/net/wireless/ath/ath10k/ce.c | 30 +++++++++++++++++++--------- 1 file changed, 21 insertions(+), 9 deletions(-)