| Message ID | 20200723010137.3127584-1-swboyd@chromium.org (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | soc: qcom: rpmh-rsc: Sleep waiting for tcs slots to be free | expand |
On Wed, Jul 22 2020 at 19:01 -0600, Stephen Boyd wrote: >The busy loop in rpmh_rsc_send_data() is written with the assumption >that the udelay will be preempted by the tcs_tx_done() irq handler when >the TCS slots are all full. This doesn't hold true when the calling >thread is an irqthread and the tcs_tx_done() irq is also an irqthread. >That's because kernel irqthreads are SCHED_FIFO and thus need to >voluntarily give up priority by calling into the scheduler so that other >threads can run. > >I see RCU stalls when I boot with irqthreads on the kernel commandline >because the modem remoteproc driver is trying to send an rpmh async >message from an irqthread that needs to give up the CPU for the rpmh >irqthread to run and clear out tcs slots. > Would this be not better, if we we use a threaded IRQ handler or offload tx_done to another waitqueue instead of handling it in IRQ handler? -- Lina > rcu: INFO: rcu_preempt self-detected stall on CPU > rcu: 0-....: (1 GPs behind) idle=402/1/0x4000000000000002 softirq=2108/2109 fqs=4920 > (t=21016 jiffies g=2933 q=590) > Task dump for CPU 0: > irq/11-smp2p R running task 0 148 2 0x00000028 > Call trace: > dump_backtrace+0x0/0x154 > show_stack+0x20/0x2c > sched_show_task+0xfc/0x108 > dump_cpu_task+0x44/0x50 > rcu_dump_cpu_stacks+0xa4/0xf8 > rcu_sched_clock_irq+0x7dc/0xaa8 > update_process_times+0x30/0x54 > tick_sched_handle+0x50/0x64 > tick_sched_timer+0x4c/0x8c > __hrtimer_run_queues+0x21c/0x36c > hrtimer_interrupt+0xf0/0x22c > arch_timer_handler_phys+0x40/0x50 > handle_percpu_devid_irq+0x114/0x25c > __handle_domain_irq+0x84/0xc4 > gic_handle_irq+0xd0/0x178 > el1_irq+0xbc/0x180 > save_return_addr+0x18/0x28 > return_address+0x54/0x88 > preempt_count_sub+0x40/0x88 > _raw_spin_unlock_irqrestore+0x4c/0x6c > ___ratelimit+0xd0/0x128 > rpmh_rsc_send_data+0x24c/0x378 > __rpmh_write+0x1b0/0x208 > rpmh_write_async+0x90/0xbc > rpmhpd_send_corner+0x60/0x8c > rpmhpd_aggregate_corner+0x8c/0x124 > rpmhpd_set_performance_state+0x8c/0xbc > _genpd_set_performance_state+0xdc/0x1b8 > dev_pm_genpd_set_performance_state+0xb8/0xf8 > q6v5_pds_disable+0x34/0x60 [qcom_q6v5_mss] > qcom_msa_handover+0x38/0x44 [qcom_q6v5_mss] > q6v5_handover_interrupt+0x24/0x3c [qcom_q6v5] > handle_nested_irq+0xd0/0x138 > qcom_smp2p_intr+0x188/0x200 > irq_thread_fn+0x2c/0x70 > irq_thread+0xfc/0x14c > kthread+0x11c/0x12c > ret_from_fork+0x10/0x18 > >This busy loop naturally lends itself to using a wait queue so that each >thread that tries to send a message will sleep waiting on the waitqueue >and only be woken up when a free slot is available. This should make >things more predictable too because the scheduler will be able to sleep >tasks that are waiting on a free tcs instead of the busy loop we >currently have today. > >Cc: Douglas Anderson <dianders@chromium.org> >Cc: Maulik Shah <mkshah@codeaurora.org> >Cc: Lina Iyer <ilina@codeaurora.org> >Signed-off-by: Stephen Boyd <swboyd@chromium.org> >--- > drivers/soc/qcom/rpmh-internal.h | 2 + > drivers/soc/qcom/rpmh-rsc.c | 101 ++++++++++++------------------- > 2 files changed, 41 insertions(+), 62 deletions(-) > >diff --git a/drivers/soc/qcom/rpmh-internal.h b/drivers/soc/qcom/rpmh-internal.h >index ef60e790a750..9a325bac58fe 100644 >--- a/drivers/soc/qcom/rpmh-internal.h >+++ b/drivers/soc/qcom/rpmh-internal.h >@@ -8,6 +8,7 @@ > #define __RPM_INTERNAL_H__ > > #include <linux/bitmap.h> >+#include <linux/wait.h> > #include <soc/qcom/tcs.h> > > #define TCS_TYPE_NR 4 >@@ -118,6 +119,7 @@ struct rsc_drv { > struct tcs_group tcs[TCS_TYPE_NR]; > DECLARE_BITMAP(tcs_in_use, MAX_TCS_NR); > spinlock_t lock; >+ wait_queue_head_t tcs_wait; > struct rpmh_ctrlr client; > }; > >diff --git a/drivers/soc/qcom/rpmh-rsc.c b/drivers/soc/qcom/rpmh-rsc.c >index 076fd27f3081..6c758b052c95 100644 >--- a/drivers/soc/qcom/rpmh-rsc.c >+++ b/drivers/soc/qcom/rpmh-rsc.c >@@ -19,6 +19,7 @@ > #include <linux/platform_device.h> > #include <linux/slab.h> > #include <linux/spinlock.h> >+#include <linux/wait.h> > > #include <soc/qcom/cmd-db.h> > #include <soc/qcom/tcs.h> >@@ -444,6 +445,7 @@ static irqreturn_t tcs_tx_done(int irq, void *p) > */ > if (!drv->tcs[ACTIVE_TCS].num_tcs) > enable_tcs_irq(drv, i, false); >+ wake_up(&drv->tcs_wait); > spin_unlock(&drv->lock); > if (req) > rpmh_tx_done(req, err); >@@ -562,44 +564,59 @@ static int find_free_tcs(struct tcs_group *tcs) > return -EBUSY; > } > >+static int claim_tcs_for_req(struct rsc_drv *drv, struct tcs_group *tcs, >+ const struct tcs_request *msg) >+{ >+ int ret; >+ >+ /* >+ * The h/w does not like if we send a request to the same address, >+ * when one is already in-flight or being processed. >+ */ >+ ret = check_for_req_inflight(drv, tcs, msg); >+ if (ret) >+ return ret; >+ >+ return find_free_tcs(tcs); >+} >+ > /** >- * tcs_write() - Store messages into a TCS right now, or return -EBUSY. >+ * rpmh_rsc_send_data() - Write / trigger active-only message. > * @drv: The controller. > * @msg: The data to be sent. > * >- * Grabs a TCS for ACTIVE_ONLY transfers and writes the messages to it. >- * >- * If there are no free TCSes for ACTIVE_ONLY transfers or if a command for >- * the same address is already transferring returns -EBUSY which means the >- * client should retry shortly. >+ * NOTES: >+ * - This is only used for "ACTIVE_ONLY" since the limitations of this >+ * function don't make sense for sleep/wake cases. >+ * - To do the transfer, we will grab a whole TCS for ourselves--we don't >+ * try to share. If there are none available we'll wait indefinitely >+ * for a free one. >+ * - This function will not wait for the commands to be finished, only for >+ * data to be programmed into the RPMh. See rpmh_tx_done() which will >+ * be called when the transfer is fully complete. >+ * - This function must be called with interrupts enabled. If the hardware >+ * is busy doing someone else's transfer we need that transfer to fully >+ * finish so that we can have the hardware, and to fully finish it needs >+ * the interrupt handler to run. If the interrupts is set to run on the >+ * active CPU this can never happen if interrupts are disabled. > * >- * Return: 0 on success, -EBUSY if client should retry, or an error. >- * Client should have interrupts enabled for a bit before retrying. >+ * Return: 0 on success, -EINVAL on error. > */ >-static int tcs_write(struct rsc_drv *drv, const struct tcs_request *msg) >+int rpmh_rsc_send_data(struct rsc_drv *drv, const struct tcs_request *msg) > { > struct tcs_group *tcs; > int tcs_id; > unsigned long flags; >- int ret; > > tcs = get_tcs_for_msg(drv, msg); > if (IS_ERR(tcs)) > return PTR_ERR(tcs); > > spin_lock_irqsave(&drv->lock, flags); >- /* >- * The h/w does not like if we send a request to the same address, >- * when one is already in-flight or being processed. >- */ >- ret = check_for_req_inflight(drv, tcs, msg); >- if (ret) >- goto unlock; > >- ret = find_free_tcs(tcs); >- if (ret < 0) >- goto unlock; >- tcs_id = ret; >+ wait_event_lock_irq(drv->tcs_wait, >+ (tcs_id = claim_tcs_for_req(drv, tcs, msg)) >= 0, >+ drv->lock); > > tcs->req[tcs_id - tcs->offset] = msg; > set_bit(tcs_id, drv->tcs_in_use); >@@ -627,47 +644,6 @@ static int tcs_write(struct rsc_drv *drv, const struct tcs_request *msg) > __tcs_set_trigger(drv, tcs_id, true); > > return 0; >-unlock: >- spin_unlock_irqrestore(&drv->lock, flags); >- return ret; >-} >- >-/** >- * rpmh_rsc_send_data() - Write / trigger active-only message. >- * @drv: The controller. >- * @msg: The data to be sent. >- * >- * NOTES: >- * - This is only used for "ACTIVE_ONLY" since the limitations of this >- * function don't make sense for sleep/wake cases. >- * - To do the transfer, we will grab a whole TCS for ourselves--we don't >- * try to share. If there are none available we'll wait indefinitely >- * for a free one. >- * - This function will not wait for the commands to be finished, only for >- * data to be programmed into the RPMh. See rpmh_tx_done() which will >- * be called when the transfer is fully complete. >- * - This function must be called with interrupts enabled. If the hardware >- * is busy doing someone else's transfer we need that transfer to fully >- * finish so that we can have the hardware, and to fully finish it needs >- * the interrupt handler to run. If the interrupts is set to run on the >- * active CPU this can never happen if interrupts are disabled. >- * >- * Return: 0 on success, -EINVAL on error. >- */ >-int rpmh_rsc_send_data(struct rsc_drv *drv, const struct tcs_request *msg) >-{ >- int ret; >- >- do { >- ret = tcs_write(drv, msg); >- if (ret == -EBUSY) { >- pr_info_ratelimited("TCS Busy, retrying RPMH message send: addr=%#x\n", >- msg->cmds[0].addr); >- udelay(10); >- } >- } while (ret == -EBUSY); >- >- return ret; > } > > /** >@@ -975,6 +951,7 @@ static int rpmh_rsc_probe(struct platform_device *pdev) > return ret; > > spin_lock_init(&drv->lock); >+ init_waitqueue_head(&drv->tcs_wait); > bitmap_zero(drv->tcs_in_use, MAX_TCS_NR); > > irq = platform_get_irq(pdev, drv->id); > >base-commit: b3a9e3b9622ae10064826dccb4f7a52bd88c7407 >-- >Sent by a computer, using git, on the internet >
Quoting Lina Iyer (2020-07-23 10:42:54) > On Wed, Jul 22 2020 at 19:01 -0600, Stephen Boyd wrote: > >The busy loop in rpmh_rsc_send_data() is written with the assumption > >that the udelay will be preempted by the tcs_tx_done() irq handler when > >the TCS slots are all full. This doesn't hold true when the calling > >thread is an irqthread and the tcs_tx_done() irq is also an irqthread. > >That's because kernel irqthreads are SCHED_FIFO and thus need to > >voluntarily give up priority by calling into the scheduler so that other > >threads can run. > > > >I see RCU stalls when I boot with irqthreads on the kernel commandline > >because the modem remoteproc driver is trying to send an rpmh async > >message from an irqthread that needs to give up the CPU for the rpmh > >irqthread to run and clear out tcs slots. > > > Would this be not better, if we we use a threaded IRQ handler or offload > tx_done to another waitqueue instead of handling it in IRQ handler? > Are you asking if jitter is reduced when the rpmh irq is made into a threaded irq? I haven't done any benchmarking to see if it improves things.
Hi, On Wed, Jul 22, 2020 at 6:01 PM Stephen Boyd <swboyd@chromium.org> wrote: > > The busy loop in rpmh_rsc_send_data() is written with the assumption > that the udelay will be preempted by the tcs_tx_done() irq handler when > the TCS slots are all full. This doesn't hold true when the calling > thread is an irqthread and the tcs_tx_done() irq is also an irqthread. > That's because kernel irqthreads are SCHED_FIFO and thus need to > voluntarily give up priority by calling into the scheduler so that other > threads can run. > > I see RCU stalls when I boot with irqthreads on the kernel commandline > because the modem remoteproc driver is trying to send an rpmh async > message from an irqthread that needs to give up the CPU for the rpmh > irqthread to run and clear out tcs slots. > > rcu: INFO: rcu_preempt self-detected stall on CPU > rcu: 0-....: (1 GPs behind) idle=402/1/0x4000000000000002 softirq=2108/2109 fqs=4920 > (t=21016 jiffies g=2933 q=590) > Task dump for CPU 0: > irq/11-smp2p R running task 0 148 2 0x00000028 > Call trace: > dump_backtrace+0x0/0x154 > show_stack+0x20/0x2c > sched_show_task+0xfc/0x108 > dump_cpu_task+0x44/0x50 > rcu_dump_cpu_stacks+0xa4/0xf8 > rcu_sched_clock_irq+0x7dc/0xaa8 > update_process_times+0x30/0x54 > tick_sched_handle+0x50/0x64 > tick_sched_timer+0x4c/0x8c > __hrtimer_run_queues+0x21c/0x36c > hrtimer_interrupt+0xf0/0x22c > arch_timer_handler_phys+0x40/0x50 > handle_percpu_devid_irq+0x114/0x25c > __handle_domain_irq+0x84/0xc4 > gic_handle_irq+0xd0/0x178 > el1_irq+0xbc/0x180 > save_return_addr+0x18/0x28 > return_address+0x54/0x88 > preempt_count_sub+0x40/0x88 > _raw_spin_unlock_irqrestore+0x4c/0x6c > ___ratelimit+0xd0/0x128 > rpmh_rsc_send_data+0x24c/0x378 > __rpmh_write+0x1b0/0x208 > rpmh_write_async+0x90/0xbc > rpmhpd_send_corner+0x60/0x8c > rpmhpd_aggregate_corner+0x8c/0x124 > rpmhpd_set_performance_state+0x8c/0xbc > _genpd_set_performance_state+0xdc/0x1b8 > dev_pm_genpd_set_performance_state+0xb8/0xf8 > q6v5_pds_disable+0x34/0x60 [qcom_q6v5_mss] > qcom_msa_handover+0x38/0x44 [qcom_q6v5_mss] > q6v5_handover_interrupt+0x24/0x3c [qcom_q6v5] > handle_nested_irq+0xd0/0x138 > qcom_smp2p_intr+0x188/0x200 > irq_thread_fn+0x2c/0x70 > irq_thread+0xfc/0x14c > kthread+0x11c/0x12c > ret_from_fork+0x10/0x18 > > This busy loop naturally lends itself to using a wait queue so that each > thread that tries to send a message will sleep waiting on the waitqueue > and only be woken up when a free slot is available. This should make > things more predictable too because the scheduler will be able to sleep > tasks that are waiting on a free tcs instead of the busy loop we > currently have today. > > Cc: Douglas Anderson <dianders@chromium.org> > Cc: Maulik Shah <mkshah@codeaurora.org> > Cc: Lina Iyer <ilina@codeaurora.org> > Signed-off-by: Stephen Boyd <swboyd@chromium.org> > --- > drivers/soc/qcom/rpmh-internal.h | 2 + > drivers/soc/qcom/rpmh-rsc.c | 101 ++++++++++++------------------- > 2 files changed, 41 insertions(+), 62 deletions(-) > > diff --git a/drivers/soc/qcom/rpmh-internal.h b/drivers/soc/qcom/rpmh-internal.h > index ef60e790a750..9a325bac58fe 100644 > --- a/drivers/soc/qcom/rpmh-internal.h > +++ b/drivers/soc/qcom/rpmh-internal.h > @@ -8,6 +8,7 @@ > #define __RPM_INTERNAL_H__ > > #include <linux/bitmap.h> > +#include <linux/wait.h> > #include <soc/qcom/tcs.h> > > #define TCS_TYPE_NR 4 > @@ -118,6 +119,7 @@ struct rsc_drv { > struct tcs_group tcs[TCS_TYPE_NR]; > DECLARE_BITMAP(tcs_in_use, MAX_TCS_NR); > spinlock_t lock; > + wait_queue_head_t tcs_wait; nit: this structure has a kernel-doc comment above it describing the elements. Could you add yours? > struct rpmh_ctrlr client; > }; > > diff --git a/drivers/soc/qcom/rpmh-rsc.c b/drivers/soc/qcom/rpmh-rsc.c > index 076fd27f3081..6c758b052c95 100644 > --- a/drivers/soc/qcom/rpmh-rsc.c > +++ b/drivers/soc/qcom/rpmh-rsc.c > @@ -19,6 +19,7 @@ > #include <linux/platform_device.h> > #include <linux/slab.h> > #include <linux/spinlock.h> > +#include <linux/wait.h> > > #include <soc/qcom/cmd-db.h> > #include <soc/qcom/tcs.h> > @@ -444,6 +445,7 @@ static irqreturn_t tcs_tx_done(int irq, void *p) > */ > if (!drv->tcs[ACTIVE_TCS].num_tcs) > enable_tcs_irq(drv, i, false); > + wake_up(&drv->tcs_wait); > spin_unlock(&drv->lock); nit: I think it's slightly better to do the wake_up() after the spin_unlock(), no? The first thing the other task will do is to try to grab the spinlock and we might as well give it a chance of succeeding without looping. I don't see any reason why we'd need to be holding the lock while calling wake_up(). > if (req) > rpmh_tx_done(req, err); > @@ -562,44 +564,59 @@ static int find_free_tcs(struct tcs_group *tcs) > return -EBUSY; > } > > +static int claim_tcs_for_req(struct rsc_drv *drv, struct tcs_group *tcs, > + const struct tcs_request *msg) nit: I know this is a short function and kernel convention doesn't strictly require comments in front of all functions. However, every other function in this file has a comment and I had a really hard time dealing with the rpmh-rsc code before the comments. Could you add one for your function, even if it's short? One thing that would be nice to note is that the only error it returns is -EBUSY. See below. > +{ > + int ret; > + > + /* > + * The h/w does not like if we send a request to the same address, > + * when one is already in-flight or being processed. > + */ > + ret = check_for_req_inflight(drv, tcs, msg); > + if (ret) > + return ret; > + > + return find_free_tcs(tcs); > +} > + > /** > - * tcs_write() - Store messages into a TCS right now, or return -EBUSY. > + * rpmh_rsc_send_data() - Write / trigger active-only message. > * @drv: The controller. > * @msg: The data to be sent. > * > - * Grabs a TCS for ACTIVE_ONLY transfers and writes the messages to it. > - * > - * If there are no free TCSes for ACTIVE_ONLY transfers or if a command for > - * the same address is already transferring returns -EBUSY which means the > - * client should retry shortly. > + * NOTES: > + * - This is only used for "ACTIVE_ONLY" since the limitations of this > + * function don't make sense for sleep/wake cases. > + * - To do the transfer, we will grab a whole TCS for ourselves--we don't > + * try to share. If there are none available we'll wait indefinitely > + * for a free one. > + * - This function will not wait for the commands to be finished, only for > + * data to be programmed into the RPMh. See rpmh_tx_done() which will > + * be called when the transfer is fully complete. > + * - This function must be called with interrupts enabled. If the hardware > + * is busy doing someone else's transfer we need that transfer to fully > + * finish so that we can have the hardware, and to fully finish it needs > + * the interrupt handler to run. If the interrupts is set to run on the > + * active CPU this can never happen if interrupts are disabled. > * > - * Return: 0 on success, -EBUSY if client should retry, or an error. > - * Client should have interrupts enabled for a bit before retrying. > + * Return: 0 on success, -EINVAL on error. > */ > -static int tcs_write(struct rsc_drv *drv, const struct tcs_request *msg) > +int rpmh_rsc_send_data(struct rsc_drv *drv, const struct tcs_request *msg) > { > struct tcs_group *tcs; > int tcs_id; > unsigned long flags; > - int ret; > > tcs = get_tcs_for_msg(drv, msg); > if (IS_ERR(tcs)) > return PTR_ERR(tcs); > > spin_lock_irqsave(&drv->lock, flags); > - /* > - * The h/w does not like if we send a request to the same address, > - * when one is already in-flight or being processed. > - */ > - ret = check_for_req_inflight(drv, tcs, msg); > - if (ret) > - goto unlock; > > - ret = find_free_tcs(tcs); > - if (ret < 0) > - goto unlock; > - tcs_id = ret; > + wait_event_lock_irq(drv->tcs_wait, > + (tcs_id = claim_tcs_for_req(drv, tcs, msg)) >= 0, Even though claim_tcs_for_req() only returns 0 or -EBUSY today (IOW it never returns error codes other than -EBUSY), should we handle it? If we don't, claim_tcs_for_req() should be very clear that it shouldn't return any errors other than -EBUSY. Other than small nits this looks great to me, thanks! -Doug
Quoting Doug Anderson (2020-07-24 10:42:55) > Hi, > > On Wed, Jul 22, 2020 at 6:01 PM Stephen Boyd <swboyd@chromium.org> wrote: > > diff --git a/drivers/soc/qcom/rpmh-internal.h b/drivers/soc/qcom/rpmh-internal.h > > index ef60e790a750..9a325bac58fe 100644 > > --- a/drivers/soc/qcom/rpmh-internal.h > > +++ b/drivers/soc/qcom/rpmh-internal.h > > @@ -118,6 +119,7 @@ struct rsc_drv { > > struct tcs_group tcs[TCS_TYPE_NR]; > > DECLARE_BITMAP(tcs_in_use, MAX_TCS_NR); > > spinlock_t lock; > > + wait_queue_head_t tcs_wait; > > nit: this structure has a kernel-doc comment above it describing the > elements. Could you add yours? Sure. > > > > struct rpmh_ctrlr client; > > }; > > > > diff --git a/drivers/soc/qcom/rpmh-rsc.c b/drivers/soc/qcom/rpmh-rsc.c > > index 076fd27f3081..6c758b052c95 100644 > > --- a/drivers/soc/qcom/rpmh-rsc.c > > +++ b/drivers/soc/qcom/rpmh-rsc.c > > @@ -19,6 +19,7 @@ > > #include <linux/platform_device.h> > > #include <linux/slab.h> > > #include <linux/spinlock.h> > > +#include <linux/wait.h> > > > > #include <soc/qcom/cmd-db.h> > > #include <soc/qcom/tcs.h> > > @@ -444,6 +445,7 @@ static irqreturn_t tcs_tx_done(int irq, void *p) > > */ > > if (!drv->tcs[ACTIVE_TCS].num_tcs) > > enable_tcs_irq(drv, i, false); > > + wake_up(&drv->tcs_wait); > > spin_unlock(&drv->lock); > > nit: I think it's slightly better to do the wake_up() after the > spin_unlock(), no? The first thing the other task will do is to try > to grab the spinlock and we might as well give it a chance of > succeeding without looping. I don't see any reason why we'd need to > be holding the lock while calling wake_up(). Right that's better. > > > > if (req) > > rpmh_tx_done(req, err); > > @@ -562,44 +564,59 @@ static int find_free_tcs(struct tcs_group *tcs) > > return -EBUSY; > > } > > > > +static int claim_tcs_for_req(struct rsc_drv *drv, struct tcs_group *tcs, > > + const struct tcs_request *msg) > > nit: I know this is a short function and kernel convention doesn't > strictly require comments in front of all functions. However, every > other function in this file has a comment and I had a really hard time > dealing with the rpmh-rsc code before the comments. Could you add one > for your function, even if it's short? One thing that would be nice > to note is that the only error it returns is -EBUSY. See below. Sure I'll write up some kernel-doc. > > > - if (ret) > > - goto unlock; > > > > - ret = find_free_tcs(tcs); > > - if (ret < 0) > > - goto unlock; > > - tcs_id = ret; > > + wait_event_lock_irq(drv->tcs_wait, > > + (tcs_id = claim_tcs_for_req(drv, tcs, msg)) >= 0, > > Even though claim_tcs_for_req() only returns 0 or -EBUSY today (IOW it > never returns error codes other than -EBUSY), should we handle it? If > we don't, claim_tcs_for_req() should be very clear that it shouldn't > return any errors other than -EBUSY. Do you mean you want to change it to be (tcs_id = claim_tcs_for_req(drv, tcs, msg)) != -EBUSY instead of >= 0? It should return the tcs_id that was claimed, not just 0 or -EBUSY.
Hi, On Fri, Jul 24, 2020 at 12:44 PM Stephen Boyd <swboyd@chromium.org> wrote: > > > > - if (ret) > > > - goto unlock; > > > > > > - ret = find_free_tcs(tcs); > > > - if (ret < 0) > > > - goto unlock; > > > - tcs_id = ret; > > > + wait_event_lock_irq(drv->tcs_wait, > > > + (tcs_id = claim_tcs_for_req(drv, tcs, msg)) >= 0, > > > > Even though claim_tcs_for_req() only returns 0 or -EBUSY today (IOW it > > never returns error codes other than -EBUSY), should we handle it? If > > we don't, claim_tcs_for_req() should be very clear that it shouldn't > > return any errors other than -EBUSY. > > Do you mean you want to change it to be > > (tcs_id = claim_tcs_for_req(drv, tcs, msg)) != -EBUSY > > instead of >= 0? It should return the tcs_id that was claimed, not just > 0 or -EBUSY. Ah, right. Yes, you got it right. Of course then we have to add a "if (tcd_id < 0) goto unlock", too. If you think it's not worth adding this then we just need to make sure it's super obvious in claim_tcs_for_req() that it's not allowed to return other errors. -Doug
Quoting Doug Anderson (2020-07-24 12:49:56) > Hi, > > On Fri, Jul 24, 2020 at 12:44 PM Stephen Boyd <swboyd@chromium.org> wrote: > > > > > > - if (ret) > > > > - goto unlock; > > > > > > > > - ret = find_free_tcs(tcs); > > > > - if (ret < 0) > > > > - goto unlock; > > > > - tcs_id = ret; > > > > + wait_event_lock_irq(drv->tcs_wait, > > > > + (tcs_id = claim_tcs_for_req(drv, tcs, msg)) >= 0, > > > > > > Even though claim_tcs_for_req() only returns 0 or -EBUSY today (IOW it > > > never returns error codes other than -EBUSY), should we handle it? If > > > we don't, claim_tcs_for_req() should be very clear that it shouldn't > > > return any errors other than -EBUSY. > > > > Do you mean you want to change it to be > > > > (tcs_id = claim_tcs_for_req(drv, tcs, msg)) != -EBUSY > > > > instead of >= 0? It should return the tcs_id that was claimed, not just > > 0 or -EBUSY. > > Ah, right. Yes, you got it right. Of course then we have to add a > "if (tcd_id < 0) goto unlock", too. If you think it's not worth > adding this then we just need to make sure it's super obvious in > claim_tcs_for_req() that it's not allowed to return other errors. > Hmm right now the code will wait forever for the condition to become true, so it will only ever continue past this point if tcs_id >= 0. We could add a timeout here in another patch, but I didn't want to change the behavior of what is there in this patch. I don't really care if >= 0 or != -EBUSY is used here so I can change it to -EBUSY to provide clarity. If we add a timeout here, it would be better to change this driver to use a pull model instead of the push model it is using today so that the timeout isn't necessary. That would entail making a proper kthread that pulls requests off a queue of messages and then this asnyc call would append messages to the end of the queue and return immediately. That would be necessary if we want the async calls to work from non-sleeping contexts for example. I think Lina was alluding to this earlier in this thread.
On Fri, Jul 24 2020 at 14:01 -0600, Stephen Boyd wrote: >Quoting Doug Anderson (2020-07-24 12:49:56) >> Hi, >> >> On Fri, Jul 24, 2020 at 12:44 PM Stephen Boyd <swboyd@chromium.org> wrote: >I think Lina was alluding to this earlier in this >thread. I was thinking more of threaded irq handler than a kthread to post the requests. We went away from post-thread approach a few years ago.
Quoting Lina Iyer (2020-07-24 13:08:41) > On Fri, Jul 24 2020 at 14:01 -0600, Stephen Boyd wrote: > >Quoting Doug Anderson (2020-07-24 12:49:56) > >> Hi, > >> > >> On Fri, Jul 24, 2020 at 12:44 PM Stephen Boyd <swboyd@chromium.org> wrote: > >I think Lina was alluding to this earlier in this > >thread. > I was thinking more of threaded irq handler than a kthread to post the > requests. We went away from post-thread approach a few years ago. > Ok, got it. Why was the kthread approach abandoned?
Hi, On Fri, Jul 24, 2020 at 1:01 PM Stephen Boyd <swboyd@chromium.org> wrote: > > Quoting Doug Anderson (2020-07-24 12:49:56) > > Hi, > > > > On Fri, Jul 24, 2020 at 12:44 PM Stephen Boyd <swboyd@chromium.org> wrote: > > > > > > > > - if (ret) > > > > > - goto unlock; > > > > > > > > > > - ret = find_free_tcs(tcs); > > > > > - if (ret < 0) > > > > > - goto unlock; > > > > > - tcs_id = ret; > > > > > + wait_event_lock_irq(drv->tcs_wait, > > > > > + (tcs_id = claim_tcs_for_req(drv, tcs, msg)) >= 0, > > > > > > > > Even though claim_tcs_for_req() only returns 0 or -EBUSY today (IOW it > > > > never returns error codes other than -EBUSY), should we handle it? If > > > > we don't, claim_tcs_for_req() should be very clear that it shouldn't > > > > return any errors other than -EBUSY. > > > > > > Do you mean you want to change it to be > > > > > > (tcs_id = claim_tcs_for_req(drv, tcs, msg)) != -EBUSY > > > > > > instead of >= 0? It should return the tcs_id that was claimed, not just > > > 0 or -EBUSY. > > > > Ah, right. Yes, you got it right. Of course then we have to add a > > "if (tcd_id < 0) goto unlock", too. If you think it's not worth > > adding this then we just need to make sure it's super obvious in > > claim_tcs_for_req() that it's not allowed to return other errors. > > > > Hmm right now the code will wait forever for the condition to become > true, so it will only ever continue past this point if tcs_id >= 0. We > could add a timeout here in another patch, but I didn't want to change > the behavior of what is there in this patch. I don't really care if >= 0 > or != -EBUSY is used here so I can change it to -EBUSY to provide > clarity. > > If we add a timeout here, it would be better to change this driver to > use a pull model instead of the push model it is using today so that the > timeout isn't necessary. That would entail making a proper kthread that > pulls requests off a queue of messages and then this asnyc call would > append messages to the end of the queue and return immediately. That > would be necessary if we want the async calls to work from non-sleeping > contexts for example. I think Lina was alluding to this earlier in this > thread. I wasn't suggesting adding a timeout. I was just saying that if claim_tcs_for_req() were to ever return an error code other than -EBUSY that we'd need a check for it because otherwise we'd interpret the result as a tcs_id. -Doug
On Fri, Jul 24 2020 at 14:11 -0600, Stephen Boyd wrote: >Quoting Lina Iyer (2020-07-24 13:08:41) >> On Fri, Jul 24 2020 at 14:01 -0600, Stephen Boyd wrote: >> >Quoting Doug Anderson (2020-07-24 12:49:56) >> >> Hi, >> >> >> >> On Fri, Jul 24, 2020 at 12:44 PM Stephen Boyd <swboyd@chromium.org> wrote: >> >I think Lina was alluding to this earlier in this >> >thread. >> I was thinking more of threaded irq handler than a kthread to post the >> requests. We went away from post-thread approach a few years ago. >> > >Ok, got it. Why was the kthread approach abandoned? Simplification mostly, I think. Also, somebody requested that when the async call returns they would atleast be guaranteed that the request has been posted not necessarily processed at the remote end.
Quoting Doug Anderson (2020-07-24 13:11:59) > > I wasn't suggesting adding a timeout. I was just saying that if > claim_tcs_for_req() were to ever return an error code other than > -EBUSY that we'd need a check for it because otherwise we'd interpret > the result as a tcs_id. > Ok that sounds like you don't want a check for -EBUSY so I'll leave this as >= 0.
Hi, On Fri, Jul 24, 2020 at 1:27 PM Stephen Boyd <swboyd@chromium.org> wrote: > > Quoting Doug Anderson (2020-07-24 13:11:59) > > > > I wasn't suggesting adding a timeout. I was just saying that if > > claim_tcs_for_req() were to ever return an error code other than > > -EBUSY that we'd need a check for it because otherwise we'd interpret > > the result as a tcs_id. > > > > Ok that sounds like you don't want a check for -EBUSY so I'll leave this > as >= 0. To clarify, I'd be OK with either of these (slight preference towards #2, but not a strong one): 1. Your current code and a REALLY OBVIOUS comment in claim_tcs_for_req() saying that we'd better not return any error codes other than -EBUSY (because we'll just blindly retry on all of them). - or - 2. Handling error codes other than -EBUSY, like this: wait_event_lock_irq(drv->tcs_wait, (tcs_id = claim_tcs_for_req(drv, tcs, msg)) != -EBUSY, drv->lock); if (tcs_id < 0) goto unlock; -Doug
Quoting Doug Anderson (2020-07-24 13:31:39) > Hi, > > On Fri, Jul 24, 2020 at 1:27 PM Stephen Boyd <swboyd@chromium.org> wrote: > > > > Quoting Doug Anderson (2020-07-24 13:11:59) > > > > > > I wasn't suggesting adding a timeout. I was just saying that if > > > claim_tcs_for_req() were to ever return an error code other than > > > -EBUSY that we'd need a check for it because otherwise we'd interpret > > > the result as a tcs_id. > > > > > > > Ok that sounds like you don't want a check for -EBUSY so I'll leave this > > as >= 0. > > To clarify, I'd be OK with either of these (slight preference towards > #2, but not a strong one): > > 1. Your current code and a REALLY OBVIOUS comment in > claim_tcs_for_req() saying that we'd better not return any error codes > other than -EBUSY (because we'll just blindly retry on all of them). > > - or - > > 2. Handling error codes other than -EBUSY, like this: > > wait_event_lock_irq(drv->tcs_wait, > (tcs_id = claim_tcs_for_req(drv, tcs, msg)) != -EBUSY, > drv->lock); > if (tcs_id < 0) > goto unlock; > Ah I think I understand. You're thinking that claim_tcs_for_req() may return an error value that isn't -EBUSY some day and then this wait_event_lock_irq() will keep spinning forever when it isn't busy but invalid or some such? I'd rather not do #2 because it is dead code until claim_tcs_for_req() changes. I'll add a comment indicating that it must return something that is claimed or the caller will keep trying. When the code changes the call sites should be evaluated by the author to make sure that it keeps working. I'm afraid a really big comment won't do much to help with that in the future.
diff --git a/drivers/soc/qcom/rpmh-internal.h b/drivers/soc/qcom/rpmh-internal.h index ef60e790a750..9a325bac58fe 100644 --- a/drivers/soc/qcom/rpmh-internal.h +++ b/drivers/soc/qcom/rpmh-internal.h @@ -8,6 +8,7 @@ #define __RPM_INTERNAL_H__ #include <linux/bitmap.h> +#include <linux/wait.h> #include <soc/qcom/tcs.h> #define TCS_TYPE_NR 4 @@ -118,6 +119,7 @@ struct rsc_drv { struct tcs_group tcs[TCS_TYPE_NR]; DECLARE_BITMAP(tcs_in_use, MAX_TCS_NR); spinlock_t lock; + wait_queue_head_t tcs_wait; struct rpmh_ctrlr client; }; diff --git a/drivers/soc/qcom/rpmh-rsc.c b/drivers/soc/qcom/rpmh-rsc.c index 076fd27f3081..6c758b052c95 100644 --- a/drivers/soc/qcom/rpmh-rsc.c +++ b/drivers/soc/qcom/rpmh-rsc.c @@ -19,6 +19,7 @@ #include <linux/platform_device.h> #include <linux/slab.h> #include <linux/spinlock.h> +#include <linux/wait.h> #include <soc/qcom/cmd-db.h> #include <soc/qcom/tcs.h> @@ -444,6 +445,7 @@ static irqreturn_t tcs_tx_done(int irq, void *p) */ if (!drv->tcs[ACTIVE_TCS].num_tcs) enable_tcs_irq(drv, i, false); + wake_up(&drv->tcs_wait); spin_unlock(&drv->lock); if (req) rpmh_tx_done(req, err); @@ -562,44 +564,59 @@ static int find_free_tcs(struct tcs_group *tcs) return -EBUSY; } +static int claim_tcs_for_req(struct rsc_drv *drv, struct tcs_group *tcs, + const struct tcs_request *msg) +{ + int ret; + + /* + * The h/w does not like if we send a request to the same address, + * when one is already in-flight or being processed. + */ + ret = check_for_req_inflight(drv, tcs, msg); + if (ret) + return ret; + + return find_free_tcs(tcs); +} + /** - * tcs_write() - Store messages into a TCS right now, or return -EBUSY. + * rpmh_rsc_send_data() - Write / trigger active-only message. * @drv: The controller. * @msg: The data to be sent. * - * Grabs a TCS for ACTIVE_ONLY transfers and writes the messages to it. - * - * If there are no free TCSes for ACTIVE_ONLY transfers or if a command for - * the same address is already transferring returns -EBUSY which means the - * client should retry shortly. + * NOTES: + * - This is only used for "ACTIVE_ONLY" since the limitations of this + * function don't make sense for sleep/wake cases. + * - To do the transfer, we will grab a whole TCS for ourselves--we don't + * try to share. If there are none available we'll wait indefinitely + * for a free one. + * - This function will not wait for the commands to be finished, only for + * data to be programmed into the RPMh. See rpmh_tx_done() which will + * be called when the transfer is fully complete. + * - This function must be called with interrupts enabled. If the hardware + * is busy doing someone else's transfer we need that transfer to fully + * finish so that we can have the hardware, and to fully finish it needs + * the interrupt handler to run. If the interrupts is set to run on the + * active CPU this can never happen if interrupts are disabled. * - * Return: 0 on success, -EBUSY if client should retry, or an error. - * Client should have interrupts enabled for a bit before retrying. + * Return: 0 on success, -EINVAL on error. */ -static int tcs_write(struct rsc_drv *drv, const struct tcs_request *msg) +int rpmh_rsc_send_data(struct rsc_drv *drv, const struct tcs_request *msg) { struct tcs_group *tcs; int tcs_id; unsigned long flags; - int ret; tcs = get_tcs_for_msg(drv, msg); if (IS_ERR(tcs)) return PTR_ERR(tcs); spin_lock_irqsave(&drv->lock, flags); - /* - * The h/w does not like if we send a request to the same address, - * when one is already in-flight or being processed. - */ - ret = check_for_req_inflight(drv, tcs, msg); - if (ret) - goto unlock; - ret = find_free_tcs(tcs); - if (ret < 0) - goto unlock; - tcs_id = ret; + wait_event_lock_irq(drv->tcs_wait, + (tcs_id = claim_tcs_for_req(drv, tcs, msg)) >= 0, + drv->lock); tcs->req[tcs_id - tcs->offset] = msg; set_bit(tcs_id, drv->tcs_in_use); @@ -627,47 +644,6 @@ static int tcs_write(struct rsc_drv *drv, const struct tcs_request *msg) __tcs_set_trigger(drv, tcs_id, true); return 0; -unlock: - spin_unlock_irqrestore(&drv->lock, flags); - return ret; -} - -/** - * rpmh_rsc_send_data() - Write / trigger active-only message. - * @drv: The controller. - * @msg: The data to be sent. - * - * NOTES: - * - This is only used for "ACTIVE_ONLY" since the limitations of this - * function don't make sense for sleep/wake cases. - * - To do the transfer, we will grab a whole TCS for ourselves--we don't - * try to share. If there are none available we'll wait indefinitely - * for a free one. - * - This function will not wait for the commands to be finished, only for - * data to be programmed into the RPMh. See rpmh_tx_done() which will - * be called when the transfer is fully complete. - * - This function must be called with interrupts enabled. If the hardware - * is busy doing someone else's transfer we need that transfer to fully - * finish so that we can have the hardware, and to fully finish it needs - * the interrupt handler to run. If the interrupts is set to run on the - * active CPU this can never happen if interrupts are disabled. - * - * Return: 0 on success, -EINVAL on error. - */ -int rpmh_rsc_send_data(struct rsc_drv *drv, const struct tcs_request *msg) -{ - int ret; - - do { - ret = tcs_write(drv, msg); - if (ret == -EBUSY) { - pr_info_ratelimited("TCS Busy, retrying RPMH message send: addr=%#x\n", - msg->cmds[0].addr); - udelay(10); - } - } while (ret == -EBUSY); - - return ret; } /** @@ -975,6 +951,7 @@ static int rpmh_rsc_probe(struct platform_device *pdev) return ret; spin_lock_init(&drv->lock); + init_waitqueue_head(&drv->tcs_wait); bitmap_zero(drv->tcs_in_use, MAX_TCS_NR); irq = platform_get_irq(pdev, drv->id);
The busy loop in rpmh_rsc_send_data() is written with the assumption that the udelay will be preempted by the tcs_tx_done() irq handler when the TCS slots are all full. This doesn't hold true when the calling thread is an irqthread and the tcs_tx_done() irq is also an irqthread. That's because kernel irqthreads are SCHED_FIFO and thus need to voluntarily give up priority by calling into the scheduler so that other threads can run. I see RCU stalls when I boot with irqthreads on the kernel commandline because the modem remoteproc driver is trying to send an rpmh async message from an irqthread that needs to give up the CPU for the rpmh irqthread to run and clear out tcs slots. rcu: INFO: rcu_preempt self-detected stall on CPU rcu: 0-....: (1 GPs behind) idle=402/1/0x4000000000000002 softirq=2108/2109 fqs=4920 (t=21016 jiffies g=2933 q=590) Task dump for CPU 0: irq/11-smp2p R running task 0 148 2 0x00000028 Call trace: dump_backtrace+0x0/0x154 show_stack+0x20/0x2c sched_show_task+0xfc/0x108 dump_cpu_task+0x44/0x50 rcu_dump_cpu_stacks+0xa4/0xf8 rcu_sched_clock_irq+0x7dc/0xaa8 update_process_times+0x30/0x54 tick_sched_handle+0x50/0x64 tick_sched_timer+0x4c/0x8c __hrtimer_run_queues+0x21c/0x36c hrtimer_interrupt+0xf0/0x22c arch_timer_handler_phys+0x40/0x50 handle_percpu_devid_irq+0x114/0x25c __handle_domain_irq+0x84/0xc4 gic_handle_irq+0xd0/0x178 el1_irq+0xbc/0x180 save_return_addr+0x18/0x28 return_address+0x54/0x88 preempt_count_sub+0x40/0x88 _raw_spin_unlock_irqrestore+0x4c/0x6c ___ratelimit+0xd0/0x128 rpmh_rsc_send_data+0x24c/0x378 __rpmh_write+0x1b0/0x208 rpmh_write_async+0x90/0xbc rpmhpd_send_corner+0x60/0x8c rpmhpd_aggregate_corner+0x8c/0x124 rpmhpd_set_performance_state+0x8c/0xbc _genpd_set_performance_state+0xdc/0x1b8 dev_pm_genpd_set_performance_state+0xb8/0xf8 q6v5_pds_disable+0x34/0x60 [qcom_q6v5_mss] qcom_msa_handover+0x38/0x44 [qcom_q6v5_mss] q6v5_handover_interrupt+0x24/0x3c [qcom_q6v5] handle_nested_irq+0xd0/0x138 qcom_smp2p_intr+0x188/0x200 irq_thread_fn+0x2c/0x70 irq_thread+0xfc/0x14c kthread+0x11c/0x12c ret_from_fork+0x10/0x18 This busy loop naturally lends itself to using a wait queue so that each thread that tries to send a message will sleep waiting on the waitqueue and only be woken up when a free slot is available. This should make things more predictable too because the scheduler will be able to sleep tasks that are waiting on a free tcs instead of the busy loop we currently have today. Cc: Douglas Anderson <dianders@chromium.org> Cc: Maulik Shah <mkshah@codeaurora.org> Cc: Lina Iyer <ilina@codeaurora.org> Signed-off-by: Stephen Boyd <swboyd@chromium.org> --- drivers/soc/qcom/rpmh-internal.h | 2 + drivers/soc/qcom/rpmh-rsc.c | 101 ++++++++++++------------------- 2 files changed, 41 insertions(+), 62 deletions(-) base-commit: b3a9e3b9622ae10064826dccb4f7a52bd88c7407