Messages in this thread | | | From | Oded Gabbay <> | Date | Mon, 28 Jan 2019 14:53:16 +0200 | Subject | Re: [PATCH 10/15] habanalabs: add device reset support |
| |
On Sun, Jan 27, 2019 at 9:51 AM Mike Rapoport <rppt@linux.ibm.com> wrote: > > On Wed, Jan 23, 2019 at 02:00:52AM +0200, Oded Gabbay wrote: > > This patch adds support for doing various on-the-fly reset of Goya. > > > > The driver supports two types of resets: > > 1. soft-reset > > 2. hard-reset > > > > Soft-reset is done when the device detects a timeout of a command > > submission that was given to the device. The soft-reset process only resets > > the engines that are relevant for the submission of compute jobs, i.e. the > > DMA channels, the TPCs and the MME. The purpose is to bring the device as > > fast as possible to a working state. > > > > Hard-reset is done in several cases: > > 1. After soft-reset is done but the device is not responding > > 2. When fatal errors occur inside the device, e.g. ECC error > > 3. When the driver is removed > > > > Hard-reset performs a reset of the entire chip except for the PCI > > controller and the PLLs. It is a much longer process then soft-reset but it > > helps to recover the device without the need to reboot the Host. > > > > After hard-reset, the driver will restore the max power attribute and in > > case of manual power management, the frequencies that were set. > > > > This patch also adds two entries to the sysfs, which allows the root user > > to initiate a soft or hard reset. > > > > Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com> > > --- > > drivers/misc/habanalabs/command_buffer.c | 11 +- > > drivers/misc/habanalabs/device.c | 308 +++++++++++++++++++++- > > drivers/misc/habanalabs/goya/goya.c | 201 ++++++++++++++ > > drivers/misc/habanalabs/goya/goya_hwmgr.c | 18 +- > > drivers/misc/habanalabs/habanalabs.h | 35 +++ > > drivers/misc/habanalabs/habanalabs_drv.c | 9 +- > > drivers/misc/habanalabs/hwmon.c | 4 +- > > drivers/misc/habanalabs/irq.c | 31 +++ > > drivers/misc/habanalabs/sysfs.c | 120 ++++++++- > > 9 files changed, 712 insertions(+), 25 deletions(-) > > > > diff --git a/drivers/misc/habanalabs/command_buffer.c b/drivers/misc/habanalabs/command_buffer.c > > index 535ed6cc5bda..700c6da01188 100644 > > --- a/drivers/misc/habanalabs/command_buffer.c > > +++ b/drivers/misc/habanalabs/command_buffer.c > > @@ -81,9 +81,10 @@ int hl_cb_create(struct hl_device *hdev, struct hl_cb_mgr *mgr, > > bool alloc_new_cb = true; > > int rc; > > > > - if (hdev->disabled) { > > + if ((hdev->disabled) || ((atomic_read(&hdev->in_reset)) && > > + (ctx_id != HL_KERNEL_ASID_ID))) { > > dev_warn_ratelimited(hdev->dev, > > - "Device is disabled !!! Can't create new CBs\n"); > > + "Device is disabled or in reset !!! Can't create new CBs\n"); > > rc = -EBUSY; > > goto out_err; > > } > > @@ -187,6 +188,12 @@ int hl_cb_ioctl(struct hl_fpriv *hpriv, void *data) > > u64 handle; > > int rc; > > > > + if (hdev->hard_reset_pending) { > > + dev_crit_ratelimited(hdev->dev, > > + "Device HARD reset pending !!! Please close FD\n"); > > + return -ENODEV; > > + } > > Probably this check should be done at the top-level ioctl()? fixed > And, what will happen if the devices performs hard reset, but the used > keeps the file descriptor open? I take care of that in the reset function. Basically, I don't do the hard-reset until all user processes (and currently I only support a single one) close their FDs. And if they don't close it after a timeout, I kill the user processes. Take a look at hl_device_hard_reset_pending() > > > + > > switch (args->in.op) { > > case HL_CB_OP_CREATE: > > rc = hl_cb_create(hdev, &hpriv->cb_mgr, args->in.cb_size, > > diff --git a/drivers/misc/habanalabs/device.c b/drivers/misc/habanalabs/device.c > > index ff7b610f18c4..00fde57ce823 100644 > > --- a/drivers/misc/habanalabs/device.c > > +++ b/drivers/misc/habanalabs/device.c > > @@ -188,6 +188,7 @@ static int device_early_init(struct hl_device *hdev) > > > > mutex_init(&hdev->device_open); > > mutex_init(&hdev->send_cpu_message_lock); > > + atomic_set(&hdev->in_reset, 0); > > atomic_set(&hdev->fd_open_cnt, 0); > > > > return 0; > > @@ -238,6 +239,27 @@ static void set_freq_to_low_job(struct work_struct *work) > > usecs_to_jiffies(HL_PLL_LOW_JOB_FREQ_USEC)); > > } > > > > +static void hl_device_heartbeat(struct work_struct *work) > > +{ > > + struct hl_device *hdev = container_of(work, struct hl_device, > > + work_heartbeat.work); > > + > > + if ((hdev->disabled) || (atomic_read(&hdev->in_reset))) > > + goto reschedule; > > + > > + if (!hdev->asic_funcs->send_heartbeat(hdev)) > > + goto reschedule; > > AFAIU, asic_funcs->send_heartbeat() it set once at init time. The work > should not be scheduled it it's NULL, I suppose. I don't check her if the function pointer is NULL. I check the return value of the call to the function. The function itself is always implemented
> > > + > > + dev_err(hdev->dev, "Device heartbeat failed !!!\n"); > > + hl_device_reset(hdev, true, false); > > + > > + return; > > + > > +reschedule: > > + schedule_delayed_work(&hdev->work_heartbeat, > > + usecs_to_jiffies(HL_HEARTBEAT_PER_USEC)); > > +} > > + > > /** > > * device_late_init - do late stuff initialization for the habanalabs device > > * > > @@ -273,6 +295,12 @@ static int device_late_init(struct hl_device *hdev) > > schedule_delayed_work(&hdev->work_freq, > > usecs_to_jiffies(HL_PLL_LOW_JOB_FREQ_USEC)); > > > > + if (hdev->heartbeat) { > > + INIT_DELAYED_WORK(&hdev->work_heartbeat, hl_device_heartbeat); > > + schedule_delayed_work(&hdev->work_heartbeat, > > + usecs_to_jiffies(HL_HEARTBEAT_PER_USEC)); > > + } > > + > > hdev->late_init_done = true; > > > > return 0; > > @@ -290,6 +318,8 @@ static void device_late_fini(struct hl_device *hdev) > > return; > > > > cancel_delayed_work_sync(&hdev->work_freq); > > + if (hdev->heartbeat) > > + cancel_delayed_work_sync(&hdev->work_heartbeat); > > > > if (hdev->asic_funcs->late_fini) > > hdev->asic_funcs->late_fini(hdev); > > @@ -397,6 +427,254 @@ int hl_device_resume(struct hl_device *hdev) > > return 0; > > } > > > > +static void hl_device_hard_reset_pending(struct work_struct *work) > > +{ > > + struct hl_device_reset_work *device_reset_work = > > + container_of(work, struct hl_device_reset_work, reset_work); > > + struct hl_device *hdev = device_reset_work->hdev; > > + u16 pending_cnt = HL_PENDING_RESET_PER_SEC; > > + struct task_struct *task = NULL; > > + > > + /* Flush all processes that are inside hl_open */ > > + mutex_lock(&hdev->device_open); > > + > > + while ((atomic_read(&hdev->fd_open_cnt)) && (pending_cnt)) { > > + > > + pending_cnt--; > > + > > + dev_info(hdev->dev, > > + "Can't HARD reset, waiting for user to close FD\n"); > > + ssleep(1); > > + } > > + > > + if (atomic_read(&hdev->fd_open_cnt)) { > > + task = get_pid_task(hdev->user_ctx->hpriv->taskpid, > > + PIDTYPE_PID); > > + if (task) { > > + dev_info(hdev->dev, "Killing user processes\n"); > > + send_sig(SIGKILL, task, 1); > > Shouldn't the user get a chance for cleanup? I give them 5 seconds - It's eternity :) This is a question where I deliberated with myself a lot about. Should I kill the process to do the hard-reset automatically, or wait until the FD is closed, and potentially never hard-reset because the user will never close the FD. Currently I decided to do the former. I guess that if users won't like this behavior, I may add a kernel parameter to control this behavior.
> > > + msleep(100); > > + > > + put_task_struct(task); > > + } > > + } > > + > > + mutex_unlock(&hdev->device_open); > > + > > + hl_device_reset(hdev, true, true); > > + > > + kfree(device_reset_work); > > +} > > + > > [ ... ] > > > diff --git a/drivers/misc/habanalabs/goya/goya_hwmgr.c b/drivers/misc/habanalabs/goya/goya_hwmgr.c > > index 866d1774b2e4..9482dbb2e03a 100644 > > --- a/drivers/misc/habanalabs/goya/goya_hwmgr.c > > +++ b/drivers/misc/habanalabs/goya/goya_hwmgr.c > > @@ -38,7 +38,7 @@ static ssize_t mme_clk_show(struct device *dev, struct device_attribute *attr, > > struct hl_device *hdev = dev_get_drvdata(dev); > > long value; > > > > - if (hdev->disabled) > > + if ((hdev->disabled) || (atomic_read(&hdev->in_reset))) > > return -ENODEV; > > > > value = hl_get_frequency(hdev, MME_PLL, false); > > @@ -57,7 +57,7 @@ static ssize_t mme_clk_store(struct device *dev, struct device_attribute *attr, > > int rc; > > long value; > > > > - if (hdev->disabled) { > > + if ((hdev->disabled) || (atomic_read(&hdev->in_reset))) { > > There are quite a few of those, maybe split this check to a helper > function? Fixed > > > count = -ENODEV; > > goto fail; > > } > > @@ -87,7 +87,7 @@ static ssize_t tpc_clk_show(struct device *dev, struct device_attribute *attr, > > struct hl_device *hdev = dev_get_drvdata(dev); > > long value; > > > > - if (hdev->disabled) > > + if ((hdev->disabled) || (atomic_read(&hdev->in_reset))) > > return -ENODEV; > > > > value = hl_get_frequency(hdev, TPC_PLL, false); > > @@ -106,7 +106,7 @@ static ssize_t tpc_clk_store(struct device *dev, struct device_attribute *attr, > > int rc; > > long value; > > > > - if (hdev->disabled) { > > + if ((hdev->disabled) || (atomic_read(&hdev->in_reset))) { > > count = -ENODEV; > > goto fail; > > } > > -- > Sincerely yours, > Mike. >
| |