On Sun, Jul 28, 2019 at 2:44 PM Greg KH <[email protected]> wrote: > > On Sun, Jul 28, 2019 at 02:28:18PM +0300, Oded Gabbay wrote: > > This patch removes the limitation of a single process that can open the > > device. > > > > Now, there is no limitation on the number of processes that can open the > > device and have a valid FD. > > > > However, only a single process can perform compute operations. This is > > enforced by allowing only a single process to have a compute context. > > > > Signed-off-by: Oded Gabbay <[email protected]> > > --- > > drivers/misc/habanalabs/context.c | 100 +++++++++++++++------ > > drivers/misc/habanalabs/device.c | 18 ++-- > > drivers/misc/habanalabs/habanalabs.h | 1 - > > drivers/misc/habanalabs/habanalabs_drv.c | 8 -- > > drivers/misc/habanalabs/habanalabs_ioctl.c | 7 +- > > 5 files changed, 85 insertions(+), 49 deletions(-) > > > > diff --git a/drivers/misc/habanalabs/context.c > > b/drivers/misc/habanalabs/context.c > > index 57bbe59da9b6..f64220fc3a55 100644 > > --- a/drivers/misc/habanalabs/context.c > > +++ b/drivers/misc/habanalabs/context.c > > @@ -56,7 +56,7 @@ void hl_ctx_do_release(struct kref *ref) > > kfree(ctx); > > } > > > > -int hl_ctx_create(struct hl_device *hdev, struct hl_fpriv *hpriv) > > +static int hl_ctx_create(struct hl_device *hdev, struct hl_fpriv *hpriv) > > { > > struct hl_ctx_mgr *mgr = &hpriv->ctx_mgr; > > struct hl_ctx *ctx; > > @@ -89,9 +89,6 @@ int hl_ctx_create(struct hl_device *hdev, struct hl_fpriv > > *hpriv) > > /* TODO: remove for multiple contexts per process */ > > hpriv->ctx = ctx; > > > > - /* TODO: remove the following line for multiple process support */ > > - hdev->compute_ctx = ctx; > > - > > return 0; > > > > remove_from_idr: > > @@ -206,13 +203,22 @@ bool hl_ctx_is_valid(struct hl_fpriv *hpriv, bool > > requires_compute_ctx) > > int rc; > > > > /* First thing, to minimize latency impact, check if context exists. > > - * Also check if it matches the requirements. If so, exit immediately > > + * This is relevant for the "steady state", where a process context > > + * already exists, and we want to minimize the latency in command > > + * submissions. In that case, we want to see if we can quickly exit > > + * with a valid answer. > > + * > > + * If a context doesn't exists, we must grab the mutex. Otherwise, > > + * there can be nasty races in case of multi-threaded application. > > + * > > + * So, if the context exists and we don't need a compute context, > > + * that's fine. If it exists and the context we have is the compute > > + * context, that's also fine. Other then that, we can't check anything > > + * without the mutex. > > */ > > - if (hpriv->ctx) { > > - if ((requires_compute_ctx) && (hdev->compute_ctx != > > hpriv->ctx)) > > - return false; > > + if ((hpriv->ctx) && ((!requires_compute_ctx) || > > + (hdev->compute_ctx == hpriv->ctx))) > > return true; > > - } > > > > mutex_lock(&hdev->lazy_ctx_creation_lock); > > > > @@ -222,35 +228,73 @@ bool hl_ctx_is_valid(struct hl_fpriv *hpriv, bool > > requires_compute_ctx) > > * creation of a context > > */ > > if (hpriv->ctx) { > > - if ((requires_compute_ctx) && (hdev->compute_ctx != > > hpriv->ctx)) > > + if ((!requires_compute_ctx) || > > + (hdev->compute_ctx == hpriv->ctx)) > > + goto unlock_mutex; > > + > > + if (hdev->compute_ctx) { > > valid = false; > > - goto unlock_mutex; > > - } > > + goto unlock_mutex; > > + } > > > > - /* If we already have a compute context, there is no point > > - * of creating one in case we are called from ioctl that needs > > - * a compute context > > - */ > > - if ((hdev->compute_ctx) && (requires_compute_ctx)) { > > + /* If we reached here, it means we have a non-compute context, > > + * but there is no compute context on the device. Therefore, > > + * we can try to "upgrade" the existing context to a compute > > + * context > > + */ > > + dev_dbg_ratelimited(hdev->dev, > > + "Non-compute context %d exists\n", > > + hpriv->ctx->asid); > > + > > + } else if ((hdev->compute_ctx) && (requires_compute_ctx)) { > > + > > + /* If we already have a compute context in the device, there > > is > > + * no point of creating one in case we are called from ioctl > > + * that needs a compute context > > + */ > > dev_err(hdev->dev, > > "Can't create new compute context as one already > > exists\n"); > > valid = false; > > goto unlock_mutex; > > - } > > + } else { > > + /* If we reached here it is because there isn't a context for > > + * the process AND there is no compute context or compute > > + * context wasn't required. In any case, must create a context > > + * for the process > > + */ > > > > - rc = hl_ctx_create(hdev, hpriv); > > - if (rc) { > > - dev_err(hdev->dev, "Failed to create context %d\n", rc); > > - valid = false; > > - goto unlock_mutex; > > + rc = hl_ctx_create(hdev, hpriv); > > + if (rc) { > > + dev_err(hdev->dev, "Failed to create context %d\n", > > rc); > > + valid = false; > > + goto unlock_mutex; > > + } > > + > > + dev_dbg_ratelimited(hdev->dev, "Created context %d\n", > > + hpriv->ctx->asid); > > } > > > > - /* Device is IDLE at this point so it is legal to change PLLs. > > - * There is no need to check anything because if the PLL is > > - * already HIGH, the set function will return without doing > > - * anything > > + /* If we reached here then either we have a new context, or we can > > + * upgrade a non-compute context to a compute context. Do the upgrade > > + * only if the caller required a compute context > > */ > > - hl_device_set_frequency(hdev, PLL_HIGH); > > + if (requires_compute_ctx) { > > + WARN(hdev->compute_ctx, > > + "Compute context exists but driver is setting a new > > one"); > > This will trigger syzbot and will reboot machines that have > 'panic-on-warn' set (i.e. all cloud systems). So be _VERY_ careful > about this. > > If a user can trigger this, do not use WARN(), that's not what it is > for. > > thanks, > > greg k-h
I see... I'll replace it with dev_crit, but I wanted to ask if you recommend to never use WARN in drivers ? Just use it in kernel core code ? Thanks, Oded
