On Fri, Jul 08, 2022 at 03:08:20PM +0800, Kai Ye wrote:
> Define the device isolation strategy by the device driver. The
> user configures a frequency value by uacce interface. If the
> slot reset frequency exceeds the value of setting for a certain
> period of time, the device will not be available in user space.
> The time window is one hour. The VF device use the PF device
> isolation strategy. All the hardware errors are processed by PF
> driver. This solution can be used for other drivers.
>
> Signed-off-by: Kai Ye <yekai13@xxxxxxxxxx>
> ---
> drivers/crypto/hisilicon/qm.c | 163 +++++++++++++++++++++++++++++++---
> include/linux/hisi_acc_qm.h | 9 ++
> 2 files changed, 160 insertions(+), 12 deletions(-)
>
> diff --git a/drivers/crypto/hisilicon/qm.c b/drivers/crypto/hisilicon/qm.c
> index ad83c194d664..8eb3b790a655 100644
> --- a/drivers/crypto/hisilicon/qm.c
> +++ b/drivers/crypto/hisilicon/qm.c
> @@ -417,6 +417,16 @@ struct hisi_qm_resource {
> struct list_head list;
> };
>
> +/**
> + * struct qm_hw_err - Structure describing the device errors
> + * @list: hardware error list
> + * @timestamp: timestamp when the error occurred
> + */
> +struct qm_hw_err {
> + struct list_head list;
> + unsigned long long timestamp;
> +};
> +
> struct hisi_qm_hw_ops {
> int (*get_vft)(struct hisi_qm *qm, u32 *base, u32 *number);
> void (*qm_db)(struct hisi_qm *qm, u16 qn,
> @@ -3410,6 +3420,111 @@ static long hisi_qm_uacce_ioctl(struct uacce_queue *q, unsigned int cmd,
> return 0;
> }
>
> +/**
> + * qm_hw_err_isolate() - Try to isolate the uacce device with its VFs
> + * according to user's configuration of isolation strategy. Warning: this
> + * API should be called while there the users on this device are suspended
> + * by slot resetting preparation of PCI AER.
> + * @qm: the uacce device
> + */
> +static int qm_hw_err_isolate(struct hisi_qm *qm)
> +{
> + struct qm_hw_err *err, *tmp, *hw_err;
> + struct qm_err_isolate *isolate;
> + u32 count = 0;
> +
> + isolate = &qm->isolate_data;
> +
> +#define SECONDS_PER_HOUR 3600
> +
> + /* All the hw errs are processed by PF driver */
> + if (qm->uacce->is_vf || isolate->is_isolate ||
> + !isolate->hw_err_isolate_hz)
> + return 0;
> +
> + hw_err = kzalloc(sizeof(*hw_err), GFP_ATOMIC);
Why atomic? What lock is held here?
> + if (!hw_err)
> + return -ENOMEM;
> +
> + mutex_lock(&isolate->isolate_lock);
> + hw_err->timestamp = jiffies;
> + list_for_each_entry_safe(err, tmp, &isolate->uacce_hw_errs, list) {
> + if ((hw_err->timestamp - err->timestamp) / HZ >
> + SECONDS_PER_HOUR) {
No possiblity of wrapping the timestamp?
> + list_del(&err->list);
> + kfree(err);
> + } else {
> + count++;
> + }
> + }
> + list_add(&hw_err->list, &isolate->uacce_hw_errs);
> + mutex_unlock(&isolate->isolate_lock);
> +
> + if (count >= isolate->hw_err_isolate_hz)
> + isolate->is_isolate = true;
> +
> + return 0;
> +}
> +
> +static void qm_hw_err_destroy(struct hisi_qm *qm)
> +{
> + struct qm_hw_err *err, *tmp;
> +
> + mutex_lock(&qm->isolate_data.isolate_lock);
> + list_for_each_entry_safe(err, tmp, &qm->isolate_data.uacce_hw_errs, list) {
> + list_del(&err->list);
> + kfree(err);
> + }
> + mutex_unlock(&qm->isolate_data.isolate_lock);
> +}
> +
> +static enum uacce_dev_state hisi_qm_get_isolate_state(struct uacce_device *uacce)
> +{
> + struct hisi_qm *qm = uacce->priv;
> + struct hisi_qm *pf_qm;
> +
> + if (uacce->is_vf)
> + pf_qm = pci_get_drvdata(pci_physfn(qm->pdev));
> + else
> + pf_qm = qm;
> +
> + return pf_qm->isolate_data.is_isolate ?
> + UACCE_DEV_ISOLATE : UACCE_DEV_NORMAL;
> +}
> +
> +static int hisi_qm_isolate_strategy_write(struct uacce_device *uacce,
> + u32 freq)
> +{
> + struct hisi_qm *qm = uacce->priv;
> +
> + /* Must be set by PF */
> + if (uacce->is_vf)
> + return -EINVAL;
But the value passed to you is not invalid, something else went wrong.
Are you sure this is the correct error?
> +
> + if (qm->isolate_data.is_isolate)
> + return -EINVAL;
Same here, why is this correct?
> +
> + qm->isolate_data.hw_err_isolate_hz = freq;
No validation of the value passed to you? It can be anything?
> +
> + /* After the policy is updated, need to reset the hardware err list */
> + qm_hw_err_destroy(qm);
No error checking?
thanks,
greg k-h
