[PATCH 4/9] drm/amdkfd: Separate doorbell allocation from PASID

Felix Kuehling Felix.Kuehling at amd.com
Sat Aug 26 07:19:04 UTC 2017


PASID management is moving into KGD. Limiting the PASID range to the
number of doorbell pages is no longer practical.

Change-Id: I2ddee0c481d3cb54c779ef6f180e0188a9ca53ca
Signed-off-by: Felix Kuehling <Felix.Kuehling at amd.com>
---
 drivers/gpu/drm/amd/amdkfd/kfd_device.c   |  7 -----
 drivers/gpu/drm/amd/amdkfd/kfd_doorbell.c | 48 ++++++++++++++++++++-----------
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h     | 10 +++----
 drivers/gpu/drm/amd/amdkfd/kfd_process.c  |  6 ++++
 4 files changed, 43 insertions(+), 28 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
index d102134..378a9cf 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
@@ -413,13 +413,6 @@ static bool device_iommu_pasid_init(struct kfd_dev *kfd)
 	pasid_limit = min_t(unsigned int,
 			(unsigned int)(1 << kfd->device_info->max_pasid_bits),
 			iommu_info.max_pasids);
-	/*
-	 * last pasid is used for kernel queues doorbells
-	 * in the future the last pasid might be used for a kernel thread.
-	 */
-	pasid_limit = min_t(unsigned int,
-				pasid_limit,
-				kfd->doorbell_process_limit - 1);
 
 	if (!kfd_set_pasid_limit(pasid_limit)) {
 		dev_err(kfd_device, "error setting pasid limit\n");
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_doorbell.c b/drivers/gpu/drm/amd/amdkfd/kfd_doorbell.c
index 008d258..17a0b97 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_doorbell.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_doorbell.c
@@ -26,14 +26,12 @@
 #include <linux/io.h>
 
 /*
- * This extension supports a kernel level doorbells management for
- * the kernel queues.
- * Basically the last doorbells page is devoted to kernel queues
- * and that's assures that any user process won't get access to the
- * kernel doorbells page
+ * This extension supports a kernel level doorbells management for the
+ * kernel queues using the first doorbell page reserved for the kernel.
  */
 
-#define KERNEL_DOORBELL_PASID 1
+static DEFINE_IDA(doorbell_ida);
+static unsigned int max_doorbell_slices;
 
 /*
  * Each device exposes a doorbell aperture, a PCI MMIO aperture that
@@ -83,13 +81,15 @@ int kfd_doorbell_init(struct kfd_dev *kfd)
 			(doorbell_aperture_size - doorbell_start_offset) /
 						kfd_doorbell_process_slice(kfd);
 	else
-		doorbell_process_limit = 0;
+		return -ENOSPC;
+
+	if (!max_doorbell_slices || doorbell_process_limit < max_doorbell_slices)
+		max_doorbell_slices = doorbell_process_limit;
 
 	kfd->doorbell_base = kfd->shared_resources.doorbell_physical_address +
 				doorbell_start_offset;
 
 	kfd->doorbell_id_offset = doorbell_start_offset / sizeof(u32);
-	kfd->doorbell_process_limit = doorbell_process_limit - 1;
 
 	kfd->doorbell_kernel_ptr = ioremap(kfd->doorbell_base,
 					   kfd_doorbell_process_slice(kfd));
@@ -181,12 +181,11 @@ void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
 	inx *= kfd->device_info->doorbell_size / sizeof(u32);
 
 	/*
-	 * Calculating the kernel doorbell offset using "faked" kernel
-	 * pasid that allocated for kernel queues only. Offset is in
-	 * dword units regardless of the ASIC-dependent doorbell size.
+	 * Calculating the kernel doorbell offset using the first
+	 * doorbell page. Offset is in dword units regardless of the
+	 * ASIC-dependent doorbell size.
 	 */
-	*doorbell_off = KERNEL_DOORBELL_PASID *
-		(kfd_doorbell_process_slice(kfd) / sizeof(u32)) + inx;
+	*doorbell_off = kfd->doorbell_id_offset + inx;
 
 	pr_debug("Get kernel queue doorbell\n"
 			 "     doorbell offset   == 0x%08X\n"
@@ -235,12 +234,13 @@ unsigned int kfd_doorbell_id_to_offset(struct kfd_dev *kfd,
 {
 	/*
 	 * doorbell_id_offset accounts for doorbells taken by KGD.
-	 * pasid * kfd_doorbell_process_slice/sizeof(u32) adjusts to
+	 * index * kfd_doorbell_process_slice/sizeof(u32) adjusts to
 	 * the process's doorbells. The offset returned is in dword
 	 * units regardless of the ASIC-dependent doorbell size.
 	 */
 	return kfd->doorbell_id_offset +
-		process->pasid * (kfd_doorbell_process_slice(kfd)/sizeof(u32)) +
+		process->doorbell_index
+		* kfd_doorbell_process_slice(kfd) / sizeof(u32) +
 		doorbell_id * kfd->device_info->doorbell_size / sizeof(u32);
 }
 
@@ -258,5 +258,21 @@ phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev,
 					struct kfd_process *process)
 {
 	return dev->doorbell_base +
-		process->pasid * kfd_doorbell_process_slice(dev);
+		process->doorbell_index * kfd_doorbell_process_slice(dev);
+}
+
+int kfd_alloc_process_doorbells(struct kfd_process *process)
+{
+	int r = ida_simple_get(&doorbell_ida, 1, max_doorbell_slices,
+				GFP_KERNEL);
+	if (r > 0)
+		process->doorbell_index = r;
+
+	return r;
+}
+
+void kfd_free_process_doorbells(struct kfd_process *process)
+{
+	if (process->doorbell_index)
+		ida_simple_remove(&doorbell_ida, process->doorbell_index);
 }
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
index f0d0995..325ef81 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
@@ -225,9 +225,6 @@ struct kfd_dev {
 					 * to HW doorbell, GFX reserved some
 					 * at the start)
 					 */
-	size_t doorbell_process_limit;	/* Number of processes we have doorbell
-					 * space for.
-					 */
 	u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells
 					   * page used by kernel queue
 					   */
@@ -692,6 +689,7 @@ struct kfd_process {
 	struct rcu_head	rcu;
 
 	unsigned int pasid;
+	unsigned int doorbell_index;
 
 	/*
 	 * List of kfd_process_device structures,
@@ -826,6 +824,10 @@ void write_kernel_doorbell64(void __iomem *db, u64 value);
 unsigned int kfd_doorbell_id_to_offset(struct kfd_dev *kfd,
 					struct kfd_process *process,
 					unsigned int doorbell_id);
+phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev,
+					struct kfd_process *process);
+int kfd_alloc_process_doorbells(struct kfd_process *process);
+void kfd_free_process_doorbells(struct kfd_process *process);
 
 /* GTT Sub-Allocator */
 
@@ -1020,8 +1022,6 @@ void kfd_pm_func_init_v9(struct packet_manager *pm, uint16_t fw_ver);
 
 
 uint64_t kfd_get_number_elems(struct kfd_dev *kfd);
-phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev,
-					struct kfd_process *process);
 int amdkfd_fence_wait_timeout(unsigned int *fence_addr,
 				unsigned int fence_value,
 				unsigned long timeout_ms);
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
index e3ecc9a..a20ced0 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
@@ -386,6 +386,7 @@ static void kfd_process_wq_release(struct work_struct *work)
 	kfd_event_free_process(p);
 
 	kfd_pasid_free(p->pasid);
+	kfd_free_process_doorbells(p);
 
 	mutex_destroy(&p->mutex);
 
@@ -562,6 +563,9 @@ static struct kfd_process *create_process(const struct task_struct *thread,
 	if (process->pasid == 0)
 		goto err_alloc_pasid;
 
+	if (kfd_alloc_process_doorbells(process) < 0)
+		goto err_alloc_doorbells;
+
 	kref_init(&process->ref);
 	mutex_init(&process->mutex);
 
@@ -623,6 +627,8 @@ static struct kfd_process *create_process(const struct task_struct *thread,
 	mmu_notifier_unregister_no_release(&process->mmu_notifier, process->mm);
 err_mmu_notifier:
 	mutex_destroy(&process->mutex);
+	kfd_free_process_doorbells(process);
+err_alloc_doorbells:
 	kfd_pasid_free(process->pasid);
 err_alloc_pasid:
 	kfree(process);
-- 
2.7.4



More information about the amd-gfx mailing list