提交 a91e70e3 编写于 作者: F Felix Kuehling 提交者: Alex Deucher

drm/amdkfd: Separate doorbell allocation from PASID

PASID management is moving into KGD. Limiting the PASID range to the
number of doorbell pages is no longer practical.
Signed-off-by: NFelix Kuehling <Felix.Kuehling@amd.com>
Acked-by: NAlex Deucher <alexander.deucher@amd.com>
Reviewed-by: NOded Gabbay <oded.gabbay@gmail.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>
上级 f4d6229b
......@@ -168,13 +168,6 @@ 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);
err = amd_iommu_init_device(kfd->pdev, pasid_limit);
if (err < 0) {
......
......@@ -24,16 +24,15 @@
#include <linux/mman.h>
#include <linux/slab.h>
#include <linux/io.h>
#include <linux/idr.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;
#define KFD_SIZE_OF_DOORBELL_IN_BYTES 4
/*
......@@ -84,13 +83,16 @@ int kfd_doorbell_init(struct kfd_dev *kfd)
(doorbell_aperture_size - doorbell_start_offset) /
doorbell_process_allocation();
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,
doorbell_process_allocation());
......@@ -185,11 +187,10 @@ u32 __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
return NULL;
/*
* Calculating the kernel doorbell offset using "faked" kernel
* pasid that allocated for kernel queues only
* Calculating the kernel doorbell offset using the first
* doorbell page.
*/
*doorbell_off = KERNEL_DOORBELL_PASID * (doorbell_process_allocation() /
sizeof(u32)) + inx;
*doorbell_off = kfd->doorbell_id_offset + inx;
pr_debug("Get kernel queue doorbell\n"
" doorbell offset == 0x%08X\n"
......@@ -228,11 +229,12 @@ unsigned int kfd_queue_id_to_doorbell(struct kfd_dev *kfd,
{
/*
* doorbell_id_offset accounts for doorbells taken by KGD.
* pasid * doorbell_process_allocation/sizeof(u32) adjusts
* to the process's doorbells
* index * doorbell_process_allocation/sizeof(u32) adjusts to
* the process's doorbells.
*/
return kfd->doorbell_id_offset +
process->pasid * (doorbell_process_allocation()/sizeof(u32)) +
process->doorbell_index
* doorbell_process_allocation() / sizeof(u32) +
queue_id;
}
......@@ -250,5 +252,21 @@ phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev,
struct kfd_process *process)
{
return dev->doorbell_base +
process->pasid * doorbell_process_allocation();
process->doorbell_index * doorbell_process_allocation();
}
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);
}
......@@ -157,9 +157,6 @@ 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
*/
......@@ -495,6 +492,7 @@ struct kfd_process {
struct rcu_head rcu;
unsigned int pasid;
unsigned int doorbell_index;
/*
* List of kfd_process_device structures,
......@@ -583,6 +581,10 @@ void write_kernel_doorbell(u32 __iomem *db, u32 value);
unsigned int kfd_queue_id_to_doorbell(struct kfd_dev *kfd,
struct kfd_process *process,
unsigned int queue_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 */
......@@ -694,8 +696,6 @@ int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type,
void pm_release_ib(struct packet_manager *pm);
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);
/* Events */
extern const struct kfd_event_interrupt_class event_interrupt_class_cik;
......
......@@ -183,6 +183,7 @@ 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_unlock(&p->mutex);
......@@ -288,6 +289,9 @@ 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;
mutex_init(&process->mutex);
process->mm = thread->mm;
......@@ -329,6 +333,8 @@ 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->queues);
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册