On Sat, Sep 16, 2017 at 2:42 AM, Felix Kuehling <Felix.Kuehling at amd.com> wrote:
> From: Yong Zhao <yong.zhao at amd.com>
>
> When we do suspend/resume through "sudo pm-suspend" while there is
> HSA activity running, upon resume we will encounter HWS hanging, which
> is caused by memory read/write failures. The root cause is that when
> suspend, we neglected to unbind pasid from kfd device.
>
> Another major change is that the bind/unbinding is changed to be
> performed on a per process basis, instead of whether there are queues
> in dqm.
>
> Signed-off-by: Yong Zhao <yong.zhao at amd.com>
> Signed-off-by: Felix Kuehling <Felix.Kuehling at amd.com>
> ---
> drivers/gpu/drm/amd/amdkfd/kfd_device.c | 22 ++++--
> .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 13 ----
> drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 15 +++-
> drivers/gpu/drm/amd/amdkfd/kfd_process.c | 89 ++++++++++++++++++----
> 4 files changed, 101 insertions(+), 38 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
> index cc8af11..ff3f97c 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
> @@ -191,7 +191,7 @@ static void iommu_pasid_shutdown_callback(struct pci_dev *pdev, int pasid)
> struct kfd_dev *dev = kfd_device_by_pci_dev(pdev);
>
> if (dev)
> - kfd_unbind_process_from_device(dev, pasid);
> + kfd_process_iommu_unbind_callback(dev, pasid);
> }
>
> /*
> @@ -339,12 +339,16 @@ void kgd2kfd_device_exit(struct kfd_dev *kfd)
>
> void kgd2kfd_suspend(struct kfd_dev *kfd)
> {
> - if (kfd->init_complete) {
> - kfd->dqm->ops.stop(kfd->dqm);
> - amd_iommu_set_invalidate_ctx_cb(kfd->pdev, NULL);
> - amd_iommu_set_invalid_ppr_cb(kfd->pdev, NULL);
> - amd_iommu_free_device(kfd->pdev);
> - }
> + if (!kfd->init_complete)
> + return;
> +
> + kfd->dqm->ops.stop(kfd->dqm);
> +
> + kfd_unbind_processes_from_device(kfd);
> +
> + amd_iommu_set_invalidate_ctx_cb(kfd->pdev, NULL);
> + amd_iommu_set_invalid_ppr_cb(kfd->pdev, NULL);
> + amd_iommu_free_device(kfd->pdev);
> }
>
> int kgd2kfd_resume(struct kfd_dev *kfd)
> @@ -369,6 +373,10 @@ static int kfd_resume(struct kfd_dev *kfd)
> amd_iommu_set_invalid_ppr_cb(kfd->pdev,
> iommu_invalid_ppr_cb);
>
> + err = kfd_bind_processes_to_device(kfd);
> + if (err)
> + return -ENXIO;
You need to undo previous initialization in case
kfd_bind_processes_to_device fails, i.e. call amd_iommu_free_device()
> +
> err = kfd->dqm->ops.start(kfd->dqm);
> if (err) {
> dev_err(kfd_device,
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
> index 53a66e8..5db82b8 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
> @@ -670,7 +670,6 @@ static int initialize_cpsch(struct device_queue_manager *dqm)
>
> static int start_cpsch(struct device_queue_manager *dqm)
> {
> - struct device_process_node *node;
> int retval;
>
> retval = 0;
> @@ -697,11 +696,6 @@ static int start_cpsch(struct device_queue_manager *dqm)
>
> init_interrupts(dqm);
>
> - list_for_each_entry(node, &dqm->queues, list)
> - if (node->qpd->pqm->process && dqm->dev)
> - kfd_bind_process_to_device(dqm->dev,
> - node->qpd->pqm->process);
> -
> execute_queues_cpsch(dqm, true);
>
> return 0;
> @@ -714,15 +708,8 @@ static int start_cpsch(struct device_queue_manager *dqm)
>
> static int stop_cpsch(struct device_queue_manager *dqm)
> {
> - struct device_process_node *node;
> - struct kfd_process_device *pdd;
> -
> destroy_queues_cpsch(dqm, true, true);
>
> - list_for_each_entry(node, &dqm->queues, list) {
> - pdd = qpd_to_pdd(node->qpd);
> - pdd->bound = false;
> - }
> kfd_gtt_sa_free(dqm->dev, dqm->fence_mem);
> pm_uninit(&dqm->packets);
>
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
> index b397ec7..ef582cc 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
> @@ -435,6 +435,13 @@ struct qcm_process_device {
> uint32_t sh_hidden_private_base;
> };
>
> +
> +enum kfd_pdd_bound {
> + PDD_UNBOUND = 0,
> + PDD_BOUND,
> + PDD_BOUND_SUSPENDED,
> +};
> +
> /* Data that is per-process-per device. */
> struct kfd_process_device {
> /*
> @@ -459,7 +466,7 @@ struct kfd_process_device {
> uint64_t scratch_limit;
>
> /* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */
> - bool bound;
> + enum kfd_pdd_bound bound;
>
> /* This flag tells if we should reset all
> * wavefronts on process termination
> @@ -548,8 +555,10 @@ struct kfd_process *kfd_get_process(const struct task_struct *);
> struct kfd_process *kfd_lookup_process_by_pasid(unsigned int pasid);
>
> struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev,
> - struct kfd_process *p);
> -void kfd_unbind_process_from_device(struct kfd_dev *dev, unsigned int pasid);
> + struct kfd_process *p);
> +int kfd_bind_processes_to_device(struct kfd_dev *dev);
> +void kfd_unbind_processes_from_device(struct kfd_dev *dev);
> +void kfd_process_iommu_unbind_callback(struct kfd_dev *dev, unsigned int pasid);
> struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev,
> struct kfd_process *p);
> struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
> index c74cf22..3ffaac3 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
> @@ -174,7 +174,10 @@ static void kfd_process_wq_release(struct work_struct *work)
> if (pdd->reset_wavefronts)
> dbgdev_wave_reset_wavefronts(pdd->dev, p);
>
> - amd_iommu_unbind_pasid(pdd->dev->pdev, p->pasid);
> + if (pdd->bound == PDD_BOUND) {
> + amd_iommu_unbind_pasid(pdd->dev->pdev, p->pasid);
> + pdd->bound = PDD_UNBOUND;
I think setting the bound property here is redundant because we do
kfree(pdd) in two more statements.
> + }
> list_del(&pdd->per_device_list);
>
> kfree(pdd);
> @@ -345,9 +348,9 @@ struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev,
>
> list_for_each_entry(pdd, &p->per_device_data, per_device_list)
> if (pdd->dev == dev)
> - break;
> + return pdd;
>
> - return pdd;
> + return NULL;
> }
>
> struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
> @@ -362,6 +365,7 @@ struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
> INIT_LIST_HEAD(&pdd->qpd.priv_queue_list);
> pdd->qpd.dqm = dev->dqm;
> pdd->reset_wavefronts = false;
> + pdd->bound = PDD_UNBOUND;
> list_add(&pdd->per_device_list, &p->per_device_data);
> }
>
> @@ -387,19 +391,83 @@ struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev,
> return ERR_PTR(-ENOMEM);
> }
>
> - if (pdd->bound)
> + if (pdd->bound == PDD_BOUND)
> return pdd;
>
> + if (pdd->bound == PDD_BOUND_SUSPENDED) {
because we check the same variable, we should do it as "else if"
> + pr_err("Binding PDD_BOUND_SUSPENDED pdd is unexpected!\n");
> + return ERR_PTR(-EINVAL);
> + }
> +
> err = amd_iommu_bind_pasid(dev->pdev, p->pasid, p->lead_thread);
> if (err < 0)
> return ERR_PTR(err);
>
> - pdd->bound = true;
> + pdd->bound = PDD_BOUND;
>
> return pdd;
> }
>
> -void kfd_unbind_process_from_device(struct kfd_dev *dev, unsigned int pasid)
> +int kfd_bind_processes_to_device(struct kfd_dev *dev)
> +{
> + struct kfd_process_device *pdd;
> + struct kfd_process *p;
> + unsigned int temp;
> + int err = 0;
> +
> + int idx = srcu_read_lock(&kfd_processes_srcu);
> +
> + hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
> + mutex_lock(&p->mutex);
> + pdd = kfd_get_process_device_data(dev, p);
> + if (pdd->bound != PDD_BOUND_SUSPENDED) {
> + mutex_unlock(&p->mutex);
> + continue;
> + }
> +
> + err = amd_iommu_bind_pasid(dev->pdev, p->pasid,
> + p->lead_thread);
> + if (err < 0) {
> + pr_err("unexpected pasid %d binding failure\n",
> + p->pasid);
> + mutex_unlock(&p->mutex);
> + break;
> + }
> +
> + pdd->bound = PDD_BOUND;
> + mutex_unlock(&p->mutex);
> + }
> +
> + srcu_read_unlock(&kfd_processes_srcu, idx);
> +
> + return err;
> +}
> +
> +void kfd_unbind_processes_from_device(struct kfd_dev *dev)
> +{
> + struct kfd_process_device *pdd;
> + struct kfd_process *p;
> + unsigned int temp, temp_bound, temp_pasid;
> +
> + int idx = srcu_read_lock(&kfd_processes_srcu);
> +
> + hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
> + mutex_lock(&p->mutex);
> + pdd = kfd_get_process_device_data(dev, p);
> + temp_bound = pdd->bound;
> + temp_pasid = p->pasid;
> + if (pdd->bound == PDD_BOUND)
> + pdd->bound = PDD_BOUND_SUSPENDED;
I would add a comment here explaining why we set pdd->bound to
PDD_BOUND_SUSPENDED and not to PDD_UNBOUND. It took me a couple of
minutes to traces all the paths in the code in order to understand it.
> + mutex_unlock(&p->mutex);
> +
> + if (temp_bound == PDD_BOUND)
> + amd_iommu_unbind_pasid(dev->pdev, temp_pasid);
> + }
> +
> + srcu_read_unlock(&kfd_processes_srcu, idx);
> +}
> +
> +void kfd_process_iommu_unbind_callback(struct kfd_dev *dev, unsigned int pasid)
> {
> struct kfd_process *p;
> struct kfd_process_device *pdd;
> @@ -432,15 +500,6 @@ void kfd_unbind_process_from_device(struct kfd_dev *dev, unsigned int pasid)
> pdd->reset_wavefronts = false;
> }
>
> - /*
> - * Just mark pdd as unbound, because we still need it
> - * to call amd_iommu_unbind_pasid() in when the
> - * process exits.
> - * We don't call amd_iommu_unbind_pasid() here
> - * because the IOMMU called us.
> - */
> - pdd->bound = false;
> -
> mutex_unlock(&p->mutex);
> }
>
> --
> 2.7.4
>
> _______________________________________________
> amd-gfx mailing list
> amd-gfx at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/amd-gfx
With the above minor comments fixed, this patch is:
Reviewed-by: Oded Gabbay <oded.gabbay at gmail.com>
