Commit 56fc40ab authored by Yong Zhao's avatar Yong Zhao Committed by Alex Deucher

drm/amdkfd: Eliminate get_atc_vmid_pasid_mapping_valid

get_atc_vmid_pasid_mapping_valid() is very similar to
get_atc_vmid_pasid_mapping_pasid(), so they can be merged into a new
function get_atc_vmid_pasid_mapping_info() to reduce register access
times. More importantly, getting the PASID and the valid bit atomically
with a single read fixes some potential race conditions where the
mapping changes between the two reads.
Signed-off-by: default avatarYong Zhao <Yong.Zhao@amd.com>
Reviewed-by: default avatarFelix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: default avatarAlex Deucher <alexander.deucher@amd.com>
parent 3fe023d4
......@@ -278,10 +278,8 @@ static const struct kfd2kgd_calls kfd2kgd = {
.address_watch_execute = kgd_gfx_v9_address_watch_execute,
.wave_control_execute = kgd_gfx_v9_wave_control_execute,
.address_watch_get_offset = kgd_gfx_v9_address_watch_get_offset,
.get_atc_vmid_pasid_mapping_pasid =
kgd_gfx_v9_get_atc_vmid_pasid_mapping_pasid,
.get_atc_vmid_pasid_mapping_valid =
kgd_gfx_v9_get_atc_vmid_pasid_mapping_valid,
.get_atc_vmid_pasid_mapping_info =
kgd_gfx_v9_get_atc_vmid_pasid_mapping_info,
.get_tile_config = kgd_gfx_v9_get_tile_config,
.set_vm_context_page_table_base = kgd_gfx_v9_set_vm_context_page_table_base,
.invalidate_tlbs = kgd_gfx_v9_invalidate_tlbs,
......
......@@ -98,10 +98,8 @@ static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
unsigned int watch_point_id,
unsigned int reg_offset);
static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd,
uint8_t vmid);
static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
uint8_t vmid);
static bool get_atc_vmid_pasid_mapping_info(struct kgd_dev *kgd,
uint8_t vmid, uint16_t *p_pasid);
static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
uint64_t page_table_base);
static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid);
......@@ -155,10 +153,8 @@ static const struct kfd2kgd_calls kfd2kgd = {
.address_watch_execute = kgd_address_watch_execute,
.wave_control_execute = kgd_wave_control_execute,
.address_watch_get_offset = kgd_address_watch_get_offset,
.get_atc_vmid_pasid_mapping_pasid =
get_atc_vmid_pasid_mapping_pasid,
.get_atc_vmid_pasid_mapping_valid =
get_atc_vmid_pasid_mapping_valid,
.get_atc_vmid_pasid_mapping_info =
get_atc_vmid_pasid_mapping_info,
.get_tile_config = amdgpu_amdkfd_get_tile_config,
.set_vm_context_page_table_base = set_vm_context_page_table_base,
.invalidate_tlbs = invalidate_tlbs,
......@@ -775,26 +771,17 @@ static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
return 0;
}
static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd,
uint8_t vmid)
static bool get_atc_vmid_pasid_mapping_info(struct kgd_dev *kgd,
uint8_t vmid, uint16_t *p_pasid)
{
uint32_t reg;
uint32_t value;
struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
reg = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
value = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
+ vmid);
return reg & ATC_VMID0_PASID_MAPPING__VALID_MASK;
}
static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
uint8_t vmid)
{
uint32_t reg;
struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
*p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK;
reg = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
+ vmid);
return reg & ATC_VMID0_PASID_MAPPING__PASID_MASK;
return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK);
}
static int invalidate_tlbs_with_kiq(struct amdgpu_device *adev, uint16_t pasid)
......@@ -826,6 +813,8 @@ static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid)
{
struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
int vmid;
uint16_t queried_pasid;
bool ret;
struct amdgpu_ring *ring = &adev->gfx.kiq.ring;
if (amdgpu_emu_mode == 0 && ring->sched.ready)
......@@ -834,13 +823,13 @@ static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid)
for (vmid = 0; vmid < 16; vmid++) {
if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid))
continue;
if (get_atc_vmid_pasid_mapping_valid(kgd, vmid)) {
if (get_atc_vmid_pasid_mapping_pasid(kgd, vmid)
== pasid) {
amdgpu_gmc_flush_gpu_tlb(adev, vmid,
AMDGPU_GFXHUB_0, 0);
break;
}
ret = get_atc_vmid_pasid_mapping_info(kgd, vmid,
&queried_pasid);
if (ret && queried_pasid == pasid) {
amdgpu_gmc_flush_gpu_tlb(adev, vmid,
AMDGPU_GFXHUB_0, 0);
break;
}
}
......
......@@ -133,9 +133,8 @@ static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
unsigned int watch_point_id,
unsigned int reg_offset);
static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd, uint8_t vmid);
static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
uint8_t vmid);
static bool get_atc_vmid_pasid_mapping_info(struct kgd_dev *kgd,
uint8_t vmid, uint16_t *p_pasid);
static void set_scratch_backing_va(struct kgd_dev *kgd,
uint64_t va, uint32_t vmid);
......@@ -186,8 +185,7 @@ static const struct kfd2kgd_calls kfd2kgd = {
.address_watch_execute = kgd_address_watch_execute,
.wave_control_execute = kgd_wave_control_execute,
.address_watch_get_offset = kgd_address_watch_get_offset,
.get_atc_vmid_pasid_mapping_pasid = get_atc_vmid_pasid_mapping_pasid,
.get_atc_vmid_pasid_mapping_valid = get_atc_vmid_pasid_mapping_valid,
.get_atc_vmid_pasid_mapping_info = get_atc_vmid_pasid_mapping_info,
.set_scratch_backing_va = set_scratch_backing_va,
.get_tile_config = get_tile_config,
.set_vm_context_page_table_base = set_vm_context_page_table_base,
......@@ -753,24 +751,16 @@ static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
return watchRegs[watch_point_id * ADDRESS_WATCH_REG_MAX + reg_offset];
}
static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd,
uint8_t vmid)
static bool get_atc_vmid_pasid_mapping_info(struct kgd_dev *kgd,
uint8_t vmid, uint16_t *p_pasid)
{
uint32_t reg;
uint32_t value;
struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
reg = RREG32(mmATC_VMID0_PASID_MAPPING + vmid);
return reg & ATC_VMID0_PASID_MAPPING__VALID_MASK;
}
static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
uint8_t vmid)
{
uint32_t reg;
struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
value = RREG32(mmATC_VMID0_PASID_MAPPING + vmid);
*p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK;
reg = RREG32(mmATC_VMID0_PASID_MAPPING + vmid);
return reg & ATC_VMID0_PASID_MAPPING__PASID_MASK;
return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK);
}
static void set_scratch_backing_va(struct kgd_dev *kgd,
......
......@@ -89,10 +89,8 @@ static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
unsigned int watch_point_id,
unsigned int reg_offset);
static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd,
uint8_t vmid);
static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
uint8_t vmid);
static bool get_atc_vmid_pasid_mapping_info(struct kgd_dev *kgd,
uint8_t vmid, uint16_t *p_pasid);
static void set_scratch_backing_va(struct kgd_dev *kgd,
uint64_t va, uint32_t vmid);
static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
......@@ -141,10 +139,8 @@ static const struct kfd2kgd_calls kfd2kgd = {
.address_watch_execute = kgd_address_watch_execute,
.wave_control_execute = kgd_wave_control_execute,
.address_watch_get_offset = kgd_address_watch_get_offset,
.get_atc_vmid_pasid_mapping_pasid =
get_atc_vmid_pasid_mapping_pasid,
.get_atc_vmid_pasid_mapping_valid =
get_atc_vmid_pasid_mapping_valid,
.get_atc_vmid_pasid_mapping_info =
get_atc_vmid_pasid_mapping_info,
.set_scratch_backing_va = set_scratch_backing_va,
.get_tile_config = get_tile_config,
.set_vm_context_page_table_base = set_vm_context_page_table_base,
......@@ -667,24 +663,16 @@ static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
return 0;
}
static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd,
uint8_t vmid)
static bool get_atc_vmid_pasid_mapping_info(struct kgd_dev *kgd,
uint8_t vmid, uint16_t *p_pasid)
{
uint32_t reg;
uint32_t value;
struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
reg = RREG32(mmATC_VMID0_PASID_MAPPING + vmid);
return reg & ATC_VMID0_PASID_MAPPING__VALID_MASK;
}
static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
uint8_t vmid)
{
uint32_t reg;
struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
value = RREG32(mmATC_VMID0_PASID_MAPPING + vmid);
*p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK;
reg = RREG32(mmATC_VMID0_PASID_MAPPING + vmid);
return reg & ATC_VMID0_PASID_MAPPING__PASID_MASK;
return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK);
}
static int kgd_address_watch_disable(struct kgd_dev *kgd)
......
......@@ -612,26 +612,17 @@ static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
return 0;
}
bool kgd_gfx_v9_get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd,
uint8_t vmid)
bool kgd_gfx_v9_get_atc_vmid_pasid_mapping_info(struct kgd_dev *kgd,
uint8_t vmid, uint16_t *p_pasid)
{
uint32_t reg;
uint32_t value;
struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
reg = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
value = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
+ vmid);
return reg & ATC_VMID0_PASID_MAPPING__VALID_MASK;
}
uint16_t kgd_gfx_v9_get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
uint8_t vmid)
{
uint32_t reg;
struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
*p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK;
reg = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
+ vmid);
return reg & ATC_VMID0_PASID_MAPPING__PASID_MASK;
return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK);
}
static int invalidate_tlbs_with_kiq(struct amdgpu_device *adev, uint16_t pasid,
......@@ -666,6 +657,8 @@ int kgd_gfx_v9_invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid)
{
struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
int vmid, i;
uint16_t queried_pasid;
bool ret;
struct amdgpu_ring *ring = &adev->gfx.kiq.ring;
uint32_t flush_type = 0;
......@@ -681,14 +674,14 @@ int kgd_gfx_v9_invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid)
for (vmid = 0; vmid < 16; vmid++) {
if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid))
continue;
if (kgd_gfx_v9_get_atc_vmid_pasid_mapping_valid(kgd, vmid)) {
if (kgd_gfx_v9_get_atc_vmid_pasid_mapping_pasid(kgd, vmid)
== pasid) {
for (i = 0; i < adev->num_vmhubs; i++)
amdgpu_gmc_flush_gpu_tlb(adev, vmid,
i, flush_type);
break;
}
ret = kgd_gfx_v9_get_atc_vmid_pasid_mapping_info(kgd, vmid,
&queried_pasid);
if (ret && queried_pasid == pasid) {
for (i = 0; i < adev->num_vmhubs; i++)
amdgpu_gmc_flush_gpu_tlb(adev, vmid,
i, flush_type);
break;
}
}
......@@ -813,10 +806,8 @@ static const struct kfd2kgd_calls kfd2kgd = {
.address_watch_execute = kgd_gfx_v9_address_watch_execute,
.wave_control_execute = kgd_gfx_v9_wave_control_execute,
.address_watch_get_offset = kgd_gfx_v9_address_watch_get_offset,
.get_atc_vmid_pasid_mapping_pasid =
kgd_gfx_v9_get_atc_vmid_pasid_mapping_pasid,
.get_atc_vmid_pasid_mapping_valid =
kgd_gfx_v9_get_atc_vmid_pasid_mapping_valid,
.get_atc_vmid_pasid_mapping_info =
kgd_gfx_v9_get_atc_vmid_pasid_mapping_info,
.get_tile_config = kgd_gfx_v9_get_tile_config,
.set_vm_context_page_table_base = kgd_gfx_v9_set_vm_context_page_table_base,
.invalidate_tlbs = kgd_gfx_v9_invalidate_tlbs,
......
......@@ -55,10 +55,8 @@ uint32_t kgd_gfx_v9_address_watch_get_offset(struct kgd_dev *kgd,
unsigned int watch_point_id,
unsigned int reg_offset);
bool kgd_gfx_v9_get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd,
uint8_t vmid);
uint16_t kgd_gfx_v9_get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
uint8_t vmid);
bool kgd_gfx_v9_get_atc_vmid_pasid_mapping_info(struct kgd_dev *kgd,
uint8_t vmid, uint16_t *p_pasid);
void kgd_gfx_v9_set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
uint64_t page_table_base);
int kgd_gfx_v9_invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid);
......
......@@ -33,7 +33,9 @@ static bool cik_event_interrupt_isr(struct kfd_dev *dev,
const struct cik_ih_ring_entry *ihre =
(const struct cik_ih_ring_entry *)ih_ring_entry;
const struct kfd2kgd_calls *f2g = dev->kfd2kgd;
unsigned int vmid, pasid;
unsigned int vmid;
uint16_t pasid;
bool ret;
/* This workaround is due to HW/FW limitation on Hawaii that
* VMID and PASID are not written into ih_ring_entry
......@@ -48,13 +50,13 @@ static bool cik_event_interrupt_isr(struct kfd_dev *dev,
*tmp_ihre = *ihre;
vmid = f2g->read_vmid_from_vmfault_reg(dev->kgd);
pasid = f2g->get_atc_vmid_pasid_mapping_pasid(dev->kgd, vmid);
ret = f2g->get_atc_vmid_pasid_mapping_info(dev->kgd, vmid, &pasid);
tmp_ihre->ring_id &= 0x000000ff;
tmp_ihre->ring_id |= vmid << 8;
tmp_ihre->ring_id |= pasid << 16;
return (pasid != 0) &&
return ret && (pasid != 0) &&
vmid >= dev->vm_info.first_vmid_kfd &&
vmid <= dev->vm_info.last_vmid_kfd;
}
......
......@@ -761,6 +761,7 @@ int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p)
{
int status = 0;
unsigned int vmid;
uint16_t queried_pasid;
union SQ_CMD_BITS reg_sq_cmd;
union GRBM_GFX_INDEX_BITS reg_gfx_index;
struct kfd_process_device *pdd;
......@@ -782,14 +783,13 @@ int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p)
*/
for (vmid = first_vmid_to_scan; vmid <= last_vmid_to_scan; vmid++) {
if (dev->kfd2kgd->get_atc_vmid_pasid_mapping_valid
(dev->kgd, vmid)) {
if (dev->kfd2kgd->get_atc_vmid_pasid_mapping_pasid
(dev->kgd, vmid) == p->pasid) {
pr_debug("Killing wave fronts of vmid %d and pasid 0x%x\n",
vmid, p->pasid);
break;
}
status = dev->kfd2kgd->get_atc_vmid_pasid_mapping_info
(dev->kgd, vmid, &queried_pasid);
if (status && queried_pasid == p->pasid) {
pr_debug("Killing wave fronts of vmid %d and pasid 0x%x\n",
vmid, p->pasid);
break;
}
}
......
......@@ -291,12 +291,10 @@ struct kfd2kgd_calls {
uint32_t (*address_watch_get_offset)(struct kgd_dev *kgd,
unsigned int watch_point_id,
unsigned int reg_offset);
bool (*get_atc_vmid_pasid_mapping_valid)(
bool (*get_atc_vmid_pasid_mapping_info)(
struct kgd_dev *kgd,
uint8_t vmid);
uint16_t (*get_atc_vmid_pasid_mapping_pasid)(
struct kgd_dev *kgd,
uint8_t vmid);
uint8_t vmid,
uint16_t *p_pasid);
/* No longer needed from GFXv9 onward. The scratch base address is
* passed to the shader by the CP. It's the user mode driver's
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment