| Message ID | 20190816100903.7549-1-jian-hong@endlessm.com (mailing list archive) |
|---|---|
| State | Changes Requested |
| Delegated to: | Kalle Valo |
| Headers | show |
| Series | [v2] rtw88: pci: Move a mass of jobs in hw IRQ to soft IRQ | expand |
> From: Jian-Hong Pan > > There is a mass of jobs between spin lock and unlock in the hardware > IRQ which will occupy much time originally. To make system work more > efficiently, this patch moves the jobs to the soft IRQ (bottom half) to > reduce the time in hardware IRQ. > > Signed-off-by: Jian-Hong Pan <jian-hong@endlessm.com> > --- > v2: > Change the spin_lock_irqsave/unlock_irqrestore to spin_lock/unlock in > rtw_pci_interrupt_handler. Because the interrupts are already disabled > in the hardware interrupt handler. > > drivers/net/wireless/realtek/rtw88/pci.c | 33 +++++++++++++++++++----- > 1 file changed, 27 insertions(+), 6 deletions(-) > > diff --git a/drivers/net/wireless/realtek/rtw88/pci.c > b/drivers/net/wireless/realtek/rtw88/pci.c > index 00ef229552d5..0740140d7e46 100644 > --- a/drivers/net/wireless/realtek/rtw88/pci.c > +++ b/drivers/net/wireless/realtek/rtw88/pci.c > @@ -866,12 +866,28 @@ static irqreturn_t rtw_pci_interrupt_handler(int irq, > void *dev) > { > struct rtw_dev *rtwdev = dev; > struct rtw_pci *rtwpci = (struct rtw_pci *)rtwdev->priv; > - u32 irq_status[4]; > > spin_lock(&rtwpci->irq_lock); > if (!rtwpci->irq_enabled) > goto out; > > + /* disable RTW PCI interrupt to avoid more interrupts before the end of > + * thread function > + */ > + rtw_pci_disable_interrupt(rtwdev, rtwpci); So basically it's to prevent back-to-back interrupts. Nothing wrong about it, I just wondering why we don't like back-to-back interrupts. Does it means that those interrupts fired between irq_handler and threadfin would increase much more time to consume them. > +out: > + spin_unlock(&rtwpci->irq_lock); > + > + return IRQ_WAKE_THREAD; > +} > + > +static irqreturn_t rtw_pci_interrupt_threadfn(int irq, void *dev) > +{ > + struct rtw_dev *rtwdev = dev; > + struct rtw_pci *rtwpci = (struct rtw_pci *)rtwdev->priv; > + unsigned long flags; > + u32 irq_status[4]; > + > rtw_pci_irq_recognized(rtwdev, rtwpci, irq_status); > > if (irq_status[0] & IMR_MGNTDOK) > @@ -891,8 +907,11 @@ static irqreturn_t rtw_pci_interrupt_handler(int irq, > void *dev) > if (irq_status[0] & IMR_ROK) > rtw_pci_rx_isr(rtwdev, rtwpci, RTW_RX_QUEUE_MPDU); > > -out: > - spin_unlock(&rtwpci->irq_lock); > + /* all of the jobs for this interrupt have been done */ > + spin_lock_irqsave(&rtwpci->irq_lock, flags); I suggest to protect the ISRs. Because next patches will require to check if the TX DMA path is empty. This means I will also add this rtwpci->irq_lock to the TX path, and check if the skb_queue does not have any pending SKBs not DMAed successfully. > + if (rtw_flag_check(rtwdev, RTW_FLAG_RUNNING)) Why check the flag here? Is there any racing or something? Otherwise it looks to break the symmetry. > + rtw_pci_enable_interrupt(rtwdev, rtwpci); > + spin_unlock_irqrestore(&rtwpci->irq_lock, flags); > > return IRQ_HANDLED; > } > @@ -1152,8 +1171,10 @@ static int rtw_pci_probe(struct pci_dev *pdev, > goto err_destroy_pci; > } > > - ret = request_irq(pdev->irq, &rtw_pci_interrupt_handler, > - IRQF_SHARED, KBUILD_MODNAME, rtwdev); > + ret = devm_request_threaded_irq(rtwdev->dev, pdev->irq, > + rtw_pci_interrupt_handler, > + rtw_pci_interrupt_threadfn, > + IRQF_SHARED, KBUILD_MODNAME, rtwdev); > if (ret) { > ieee80211_unregister_hw(hw); > goto err_destroy_pci; > @@ -1192,7 +1213,7 @@ static void rtw_pci_remove(struct pci_dev *pdev) > rtw_pci_disable_interrupt(rtwdev, rtwpci); > rtw_pci_destroy(rtwdev, pdev); > rtw_pci_declaim(rtwdev, pdev); > - free_irq(rtwpci->pdev->irq, rtwdev); > + devm_free_irq(rtwdev->dev, rtwpci->pdev->irq, rtwdev); > rtw_core_deinit(rtwdev); > ieee80211_free_hw(hw); > } > -- > 2.20.1 Yan-Hsuan
Tony Chuang <yhchuang@realtek.com> 於 2019年8月16日 週五 下午6:44寫道： > > > From: Jian-Hong Pan > > > > There is a mass of jobs between spin lock and unlock in the hardware > > IRQ which will occupy much time originally. To make system work more > > efficiently, this patch moves the jobs to the soft IRQ (bottom half) to > > reduce the time in hardware IRQ. > > > > Signed-off-by: Jian-Hong Pan <jian-hong@endlessm.com> > > --- > > v2: > > Change the spin_lock_irqsave/unlock_irqrestore to spin_lock/unlock in > > rtw_pci_interrupt_handler. Because the interrupts are already disabled > > in the hardware interrupt handler. > > > > drivers/net/wireless/realtek/rtw88/pci.c | 33 +++++++++++++++++++----- > > 1 file changed, 27 insertions(+), 6 deletions(-) > > > > diff --git a/drivers/net/wireless/realtek/rtw88/pci.c > > b/drivers/net/wireless/realtek/rtw88/pci.c > > index 00ef229552d5..0740140d7e46 100644 > > --- a/drivers/net/wireless/realtek/rtw88/pci.c > > +++ b/drivers/net/wireless/realtek/rtw88/pci.c > > @@ -866,12 +866,28 @@ static irqreturn_t rtw_pci_interrupt_handler(int irq, > > void *dev) > > { > > struct rtw_dev *rtwdev = dev; > > struct rtw_pci *rtwpci = (struct rtw_pci *)rtwdev->priv; > > - u32 irq_status[4]; > > > > spin_lock(&rtwpci->irq_lock); > > if (!rtwpci->irq_enabled) > > goto out; > > > > + /* disable RTW PCI interrupt to avoid more interrupts before the end of > > + * thread function > > + */ > > + rtw_pci_disable_interrupt(rtwdev, rtwpci); > > So basically it's to prevent back-to-back interrupts. > > Nothing wrong about it, I just wondering why we don't like > back-to-back interrupts. Does it means that those interrupts > fired between irq_handler and threadfin would increase > much more time to consume them. 1. It is one of the reasons. Besides, the hardware interrupt has higher priority than soft IRQ. If next hardware interrupt is coming faster then the soft IRQ (BH), the software IRQ may always become pended. 2. Keep the data's state until the interrupt is fully dealt. 3. The process of request_threaded_irq is documented: https://www.kernel.org/doc/htmldocs/kernel-api/API-request-threaded-irq.html > > +out: > > + spin_unlock(&rtwpci->irq_lock); > > + > > + return IRQ_WAKE_THREAD; > > +} > > + > > +static irqreturn_t rtw_pci_interrupt_threadfn(int irq, void *dev) > > +{ > > + struct rtw_dev *rtwdev = dev; > > + struct rtw_pci *rtwpci = (struct rtw_pci *)rtwdev->priv; > > + unsigned long flags; > > + u32 irq_status[4]; > > + > > rtw_pci_irq_recognized(rtwdev, rtwpci, irq_status); > > > > if (irq_status[0] & IMR_MGNTDOK) > > @@ -891,8 +907,11 @@ static irqreturn_t rtw_pci_interrupt_handler(int irq, > > void *dev) > > if (irq_status[0] & IMR_ROK) > > rtw_pci_rx_isr(rtwdev, rtwpci, RTW_RX_QUEUE_MPDU); > > > > -out: > > - spin_unlock(&rtwpci->irq_lock); > > + /* all of the jobs for this interrupt have been done */ > > + spin_lock_irqsave(&rtwpci->irq_lock, flags); > > I suggest to protect the ISRs. Because next patches will require > to check if the TX DMA path is empty. This means I will also add > this rtwpci->irq_lock to the TX path, and check if the skb_queue > does not have any pending SKBs not DMAed successfully. Ah ... Okay for the TX path > > + if (rtw_flag_check(rtwdev, RTW_FLAG_RUNNING)) > > Why check the flag here? Is there any racing or something? > Otherwise it looks to break the symmetry. According to backtrace, I notice rtw_pci_stop will be invoked in rtw_power_off, rtw_core_stop which clears the running state. rtw_core_stop is invoked by rtw_enter_ips due to IEEE80211_CONF_IDLE. If the stop command comes between the hardware interrupt and soft IRQ (BH) and there is no start command again, I think user wants the WiFi stop instead of becoming start again. Jian-Hong Pan
diff --git a/drivers/net/wireless/realtek/rtw88/pci.c b/drivers/net/wireless/realtek/rtw88/pci.c index 00ef229552d5..0740140d7e46 100644 --- a/drivers/net/wireless/realtek/rtw88/pci.c +++ b/drivers/net/wireless/realtek/rtw88/pci.c @@ -866,12 +866,28 @@ static irqreturn_t rtw_pci_interrupt_handler(int irq, void *dev) { struct rtw_dev *rtwdev = dev; struct rtw_pci *rtwpci = (struct rtw_pci *)rtwdev->priv; - u32 irq_status[4]; spin_lock(&rtwpci->irq_lock); if (!rtwpci->irq_enabled) goto out; + /* disable RTW PCI interrupt to avoid more interrupts before the end of + * thread function + */ + rtw_pci_disable_interrupt(rtwdev, rtwpci); +out: + spin_unlock(&rtwpci->irq_lock); + + return IRQ_WAKE_THREAD; +} + +static irqreturn_t rtw_pci_interrupt_threadfn(int irq, void *dev) +{ + struct rtw_dev *rtwdev = dev; + struct rtw_pci *rtwpci = (struct rtw_pci *)rtwdev->priv; + unsigned long flags; + u32 irq_status[4]; + rtw_pci_irq_recognized(rtwdev, rtwpci, irq_status); if (irq_status[0] & IMR_MGNTDOK) @@ -891,8 +907,11 @@ static irqreturn_t rtw_pci_interrupt_handler(int irq, void *dev) if (irq_status[0] & IMR_ROK) rtw_pci_rx_isr(rtwdev, rtwpci, RTW_RX_QUEUE_MPDU); -out: - spin_unlock(&rtwpci->irq_lock); + /* all of the jobs for this interrupt have been done */ + spin_lock_irqsave(&rtwpci->irq_lock, flags); + if (rtw_flag_check(rtwdev, RTW_FLAG_RUNNING)) + rtw_pci_enable_interrupt(rtwdev, rtwpci); + spin_unlock_irqrestore(&rtwpci->irq_lock, flags); return IRQ_HANDLED; } @@ -1152,8 +1171,10 @@ static int rtw_pci_probe(struct pci_dev *pdev, goto err_destroy_pci; } - ret = request_irq(pdev->irq, &rtw_pci_interrupt_handler, - IRQF_SHARED, KBUILD_MODNAME, rtwdev); + ret = devm_request_threaded_irq(rtwdev->dev, pdev->irq, + rtw_pci_interrupt_handler, + rtw_pci_interrupt_threadfn, + IRQF_SHARED, KBUILD_MODNAME, rtwdev); if (ret) { ieee80211_unregister_hw(hw); goto err_destroy_pci; @@ -1192,7 +1213,7 @@ static void rtw_pci_remove(struct pci_dev *pdev) rtw_pci_disable_interrupt(rtwdev, rtwpci); rtw_pci_destroy(rtwdev, pdev); rtw_pci_declaim(rtwdev, pdev); - free_irq(rtwpci->pdev->irq, rtwdev); + devm_free_irq(rtwdev->dev, rtwpci->pdev->irq, rtwdev); rtw_core_deinit(rtwdev); ieee80211_free_hw(hw); }
There is a mass of jobs between spin lock and unlock in the hardware IRQ which will occupy much time originally. To make system work more efficiently, this patch moves the jobs to the soft IRQ (bottom half) to reduce the time in hardware IRQ. Signed-off-by: Jian-Hong Pan <jian-hong@endlessm.com> --- v2: Change the spin_lock_irqsave/unlock_irqrestore to spin_lock/unlock in rtw_pci_interrupt_handler. Because the interrupts are already disabled in the hardware interrupt handler. drivers/net/wireless/realtek/rtw88/pci.c | 33 +++++++++++++++++++----- 1 file changed, 27 insertions(+), 6 deletions(-)