Commit bf23729c authored by Dave Airlie's avatar Dave Airlie

Merge tag 'amd-drm-next-5.19-2022-05-26-2' of...

Merge tag 'amd-drm-next-5.19-2022-05-26-2' of https://gitlab.freedesktop.org/agd5f/linux into drm-next

amd-drm-next-5.19-2022-05-26-2:

amdgpu:
- Update fdinfo to the common drm format

UAPI:
- Add VM_NOALLOC GPUVM attribute to prevent buffers for going into the MALL
  Add AMDGPU_GEM_CREATE_DISCARDABLE flag to create buffers that can be discarded on eviction
  Mesa code which uses these: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16466Signed-off-by: default avatarDave Airlie <airlied@redhat.com>
From: Alex Deucher <alexander.deucher@amd.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220526202546.66860-1-alexander.deucher@amd.com
parents e573a3cd 9bdc1992
......@@ -162,17 +162,50 @@ static unsigned int amdgpu_ctx_get_hw_prio(struct amdgpu_ctx *ctx, u32 hw_ip)
return hw_prio;
}
/* Calculate the time spend on the hw */
static ktime_t amdgpu_ctx_fence_time(struct dma_fence *fence)
{
struct drm_sched_fence *s_fence;
if (!fence)
return ns_to_ktime(0);
/* When the fence is not even scheduled it can't have spend time */
s_fence = to_drm_sched_fence(fence);
if (!test_bit(DMA_FENCE_FLAG_TIMESTAMP_BIT, &s_fence->scheduled.flags))
return ns_to_ktime(0);
/* When it is still running account how much already spend */
if (!test_bit(DMA_FENCE_FLAG_TIMESTAMP_BIT, &s_fence->finished.flags))
return ktime_sub(ktime_get(), s_fence->scheduled.timestamp);
return ktime_sub(s_fence->finished.timestamp,
s_fence->scheduled.timestamp);
}
static ktime_t amdgpu_ctx_entity_time(struct amdgpu_ctx *ctx,
struct amdgpu_ctx_entity *centity)
{
ktime_t res = ns_to_ktime(0);
uint32_t i;
spin_lock(&ctx->ring_lock);
for (i = 0; i < amdgpu_sched_jobs; i++) {
res = ktime_add(res, amdgpu_ctx_fence_time(centity->fences[i]));
}
spin_unlock(&ctx->ring_lock);
return res;
}
static int amdgpu_ctx_init_entity(struct amdgpu_ctx *ctx, u32 hw_ip,
const u32 ring)
{
struct drm_gpu_scheduler **scheds = NULL, *sched = NULL;
struct amdgpu_device *adev = ctx->mgr->adev;
struct amdgpu_ctx_entity *entity;
struct drm_gpu_scheduler **scheds = NULL, *sched = NULL;
unsigned num_scheds = 0;
int32_t ctx_prio;
unsigned int hw_prio;
enum drm_sched_priority drm_prio;
unsigned int hw_prio, num_scheds;
int32_t ctx_prio;
int r;
entity = kzalloc(struct_size(entity, fences, amdgpu_sched_jobs),
......@@ -182,6 +215,7 @@ static int amdgpu_ctx_init_entity(struct amdgpu_ctx *ctx, u32 hw_ip,
ctx_prio = (ctx->override_priority == AMDGPU_CTX_PRIORITY_UNSET) ?
ctx->init_priority : ctx->override_priority;
entity->hw_ip = hw_ip;
entity->sequence = 1;
hw_prio = amdgpu_ctx_get_hw_prio(ctx, hw_ip);
drm_prio = amdgpu_ctx_to_drm_sched_prio(ctx_prio);
......@@ -220,6 +254,23 @@ static int amdgpu_ctx_init_entity(struct amdgpu_ctx *ctx, u32 hw_ip,
return r;
}
static ktime_t amdgpu_ctx_fini_entity(struct amdgpu_ctx_entity *entity)
{
ktime_t res = ns_to_ktime(0);
int i;
if (!entity)
return res;
for (i = 0; i < amdgpu_sched_jobs; ++i) {
res = ktime_add(res, amdgpu_ctx_fence_time(entity->fences[i]));
dma_fence_put(entity->fences[i]);
}
kfree(entity);
return res;
}
static int amdgpu_ctx_init(struct amdgpu_ctx_mgr *mgr, int32_t priority,
struct drm_file *filp, struct amdgpu_ctx *ctx)
{
......@@ -246,20 +297,6 @@ static int amdgpu_ctx_init(struct amdgpu_ctx_mgr *mgr, int32_t priority,
return 0;
}
static void amdgpu_ctx_fini_entity(struct amdgpu_ctx_entity *entity)
{
int i;
if (!entity)
return;
for (i = 0; i < amdgpu_sched_jobs; ++i)
dma_fence_put(entity->fences[i]);
kfree(entity);
}
static int amdgpu_ctx_get_stable_pstate(struct amdgpu_ctx *ctx,
u32 *stable_pstate)
{
......@@ -351,8 +388,10 @@ static void amdgpu_ctx_fini(struct kref *ref)
for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) {
for (j = 0; j < AMDGPU_MAX_ENTITY_NUM; ++j) {
amdgpu_ctx_fini_entity(ctx->entities[i][j]);
ctx->entities[i][j] = NULL;
ktime_t spend;
spend = amdgpu_ctx_fini_entity(ctx->entities[i][j]);
atomic64_add(ktime_to_ns(spend), &mgr->time_spend[i]);
}
}
......@@ -689,6 +728,9 @@ uint64_t amdgpu_ctx_add_fence(struct amdgpu_ctx *ctx,
centity->sequence++;
spin_unlock(&ctx->ring_lock);
atomic64_add(ktime_to_ns(amdgpu_ctx_fence_time(other)),
&ctx->mgr->time_spend[centity->hw_ip]);
dma_fence_put(other);
return seq;
}
......@@ -795,9 +837,14 @@ int amdgpu_ctx_wait_prev_fence(struct amdgpu_ctx *ctx,
void amdgpu_ctx_mgr_init(struct amdgpu_ctx_mgr *mgr,
struct amdgpu_device *adev)
{
unsigned int i;
mgr->adev = adev;
mutex_init(&mgr->lock);
idr_init(&mgr->ctx_handles);
for (i = 0; i < AMDGPU_HW_IP_NUM; ++i)
atomic64_set(&mgr->time_spend[i], 0);
}
long amdgpu_ctx_mgr_entity_flush(struct amdgpu_ctx_mgr *mgr, long timeout)
......@@ -873,80 +920,38 @@ void amdgpu_ctx_mgr_fini(struct amdgpu_ctx_mgr *mgr)
mutex_destroy(&mgr->lock);
}
static void amdgpu_ctx_fence_time(struct amdgpu_ctx *ctx,
struct amdgpu_ctx_entity *centity, ktime_t *total, ktime_t *max)
{
ktime_t now, t1;
uint32_t i;
*total = *max = 0;
now = ktime_get();
for (i = 0; i < amdgpu_sched_jobs; i++) {
struct dma_fence *fence;
struct drm_sched_fence *s_fence;
spin_lock(&ctx->ring_lock);
fence = dma_fence_get(centity->fences[i]);
spin_unlock(&ctx->ring_lock);
if (!fence)
continue;
s_fence = to_drm_sched_fence(fence);
if (!dma_fence_is_signaled(&s_fence->scheduled)) {
dma_fence_put(fence);
continue;
}
t1 = s_fence->scheduled.timestamp;
if (!ktime_before(t1, now)) {
dma_fence_put(fence);
continue;
}
if (dma_fence_is_signaled(&s_fence->finished) &&
s_fence->finished.timestamp < now)
*total += ktime_sub(s_fence->finished.timestamp, t1);
else
*total += ktime_sub(now, t1);
t1 = ktime_sub(now, t1);
dma_fence_put(fence);
*max = max(t1, *max);
}
}
ktime_t amdgpu_ctx_mgr_fence_usage(struct amdgpu_ctx_mgr *mgr, uint32_t hwip,
uint32_t idx, uint64_t *elapsed)
void amdgpu_ctx_mgr_usage(struct amdgpu_ctx_mgr *mgr,
ktime_t usage[AMDGPU_HW_IP_NUM])
{
struct idr *idp;
struct amdgpu_ctx *ctx;
unsigned int hw_ip, i;
uint32_t id;
struct amdgpu_ctx_entity *centity;
ktime_t total = 0, max = 0;
if (idx >= AMDGPU_MAX_ENTITY_NUM)
return 0;
idp = &mgr->ctx_handles;
/*
* This is a little bit racy because it can be that a ctx or a fence are
* destroyed just in the moment we try to account them. But that is ok
* since exactly that case is explicitely allowed by the interface.
*/
mutex_lock(&mgr->lock);
idr_for_each_entry(idp, ctx, id) {
ktime_t ttotal, tmax;
for (hw_ip = 0; hw_ip < AMDGPU_HW_IP_NUM; ++hw_ip) {
uint64_t ns = atomic64_read(&mgr->time_spend[hw_ip]);
if (!ctx->entities[hwip][idx])
continue;
usage[hw_ip] = ns_to_ktime(ns);
}
centity = ctx->entities[hwip][idx];
amdgpu_ctx_fence_time(ctx, centity, &ttotal, &tmax);
idr_for_each_entry(&mgr->ctx_handles, ctx, id) {
for (hw_ip = 0; hw_ip < AMDGPU_HW_IP_NUM; ++hw_ip) {
for (i = 0; i < amdgpu_ctx_num_entities[hw_ip]; ++i) {
struct amdgpu_ctx_entity *centity;
ktime_t spend;
/* Harmonic mean approximation diverges for very small
* values. If ratio < 0.01% ignore
*/
if (AMDGPU_CTX_FENCE_USAGE_MIN_RATIO(tmax, ttotal))
continue;
total = ktime_add(total, ttotal);
max = ktime_after(tmax, max) ? tmax : max;
centity = ctx->entities[hw_ip][i];
if (!centity)
continue;
spend = amdgpu_ctx_entity_time(ctx, centity);
usage[hw_ip] = ktime_add(usage[hw_ip], spend);
}
}
}
mutex_unlock(&mgr->lock);
if (elapsed)
*elapsed = max;
return total;
}
......@@ -23,16 +23,20 @@
#ifndef __AMDGPU_CTX_H__
#define __AMDGPU_CTX_H__
#include <linux/ktime.h>
#include <linux/types.h>
#include "amdgpu_ring.h"
struct drm_device;
struct drm_file;
struct amdgpu_fpriv;
struct amdgpu_ctx_mgr;
#define AMDGPU_MAX_ENTITY_NUM 4
#define AMDGPU_CTX_FENCE_USAGE_MIN_RATIO(max, total) ((max) > 16384ULL*(total))
struct amdgpu_ctx_entity {
uint32_t hw_ip;
uint64_t sequence;
struct drm_sched_entity entity;
struct dma_fence *fences[];
......@@ -61,6 +65,7 @@ struct amdgpu_ctx_mgr {
struct mutex lock;
/* protected by lock */
struct idr ctx_handles;
atomic64_t time_spend[AMDGPU_HW_IP_NUM];
};
extern const unsigned int amdgpu_ctx_num_entities[AMDGPU_HW_IP_NUM];
......@@ -90,6 +95,7 @@ void amdgpu_ctx_mgr_init(struct amdgpu_ctx_mgr *mgr,
void amdgpu_ctx_mgr_entity_fini(struct amdgpu_ctx_mgr *mgr);
long amdgpu_ctx_mgr_entity_flush(struct amdgpu_ctx_mgr *mgr, long timeout);
void amdgpu_ctx_mgr_fini(struct amdgpu_ctx_mgr *mgr);
ktime_t amdgpu_ctx_mgr_fence_usage(struct amdgpu_ctx_mgr *mgr, uint32_t hwip,
uint32_t idx, uint64_t *elapsed);
void amdgpu_ctx_mgr_usage(struct amdgpu_ctx_mgr *mgr,
ktime_t usage[AMDGPU_HW_IP_NUM]);
#endif
......@@ -99,10 +99,11 @@
* - 3.43.0 - Add device hot plug/unplug support
* - 3.44.0 - DCN3 supports DCC independent block settings: !64B && 128B, 64B && 128B
* - 3.45.0 - Add context ioctl stable pstate interface
* * 3.46.0 - To enable hot plug amdgpu tests in libdrm
* - 3.46.0 - To enable hot plug amdgpu tests in libdrm
* * 3.47.0 - Add AMDGPU_GEM_CREATE_DISCARDABLE and AMDGPU_VM_NOALLOC flags
*/
#define KMS_DRIVER_MAJOR 3
#define KMS_DRIVER_MINOR 46
#define KMS_DRIVER_MINOR 47
#define KMS_DRIVER_PATCHLEVEL 0
int amdgpu_vram_limit;
......
......@@ -32,6 +32,7 @@
#include <drm/amdgpu_drm.h>
#include <drm/drm_debugfs.h>
#include <drm/drm_drv.h>
#include "amdgpu.h"
#include "amdgpu_vm.h"
......@@ -54,58 +55,49 @@ static const char *amdgpu_ip_name[AMDGPU_HW_IP_NUM] = {
void amdgpu_show_fdinfo(struct seq_file *m, struct file *f)
{
struct amdgpu_fpriv *fpriv;
uint32_t bus, dev, fn, i, domain;
uint64_t vram_mem = 0, gtt_mem = 0, cpu_mem = 0;
struct drm_file *file = f->private_data;
struct amdgpu_device *adev = drm_to_adev(file->minor->dev);
struct amdgpu_bo *root;
struct amdgpu_fpriv *fpriv = file->driver_priv;
struct amdgpu_vm *vm = &fpriv->vm;
uint64_t vram_mem = 0, gtt_mem = 0, cpu_mem = 0;
ktime_t usage[AMDGPU_HW_IP_NUM];
uint32_t bus, dev, fn, domain;
unsigned int hw_ip;
int ret;
ret = amdgpu_file_to_fpriv(f, &fpriv);
if (ret)
return;
bus = adev->pdev->bus->number;
domain = pci_domain_nr(adev->pdev->bus);
dev = PCI_SLOT(adev->pdev->devfn);
fn = PCI_FUNC(adev->pdev->devfn);
root = amdgpu_bo_ref(fpriv->vm.root.bo);
if (!root)
ret = amdgpu_bo_reserve(vm->root.bo, false);
if (ret)
return;
ret = amdgpu_bo_reserve(root, false);
if (ret) {
DRM_ERROR("Fail to reserve bo\n");
return;
}
amdgpu_vm_get_memory(&fpriv->vm, &vram_mem, &gtt_mem, &cpu_mem);
amdgpu_bo_unreserve(root);
amdgpu_bo_unref(&root);
amdgpu_vm_get_memory(vm, &vram_mem, &gtt_mem, &cpu_mem);
amdgpu_bo_unreserve(vm->root.bo);
seq_printf(m, "pdev:\t%04x:%02x:%02x.%d\npasid:\t%u\n", domain, bus,
dev, fn, fpriv->vm.pasid);
seq_printf(m, "vram mem:\t%llu kB\n", vram_mem/1024UL);
seq_printf(m, "gtt mem:\t%llu kB\n", gtt_mem/1024UL);
seq_printf(m, "cpu mem:\t%llu kB\n", cpu_mem/1024UL);
for (i = 0; i < AMDGPU_HW_IP_NUM; i++) {
uint32_t count = amdgpu_ctx_num_entities[i];
int idx = 0;
uint64_t total = 0, min = 0;
uint32_t perc, frac;
amdgpu_ctx_mgr_usage(&fpriv->ctx_mgr, usage);
for (idx = 0; idx < count; idx++) {
total = amdgpu_ctx_mgr_fence_usage(&fpriv->ctx_mgr,
i, idx, &min);
if ((total == 0) || (min == 0))
continue;
/*
* ******************************************************************
* For text output format description please see drm-usage-stats.rst!
* ******************************************************************
*/
perc = div64_u64(10000 * total, min);
frac = perc % 100;
seq_printf(m, "pasid:\t%u\n", fpriv->vm.pasid);
seq_printf(m, "drm-driver:\t%s\n", file->minor->dev->driver->name);
seq_printf(m, "drm-pdev:\t%04x:%02x:%02x.%d\n", domain, bus, dev, fn);
seq_printf(m, "drm-client-id:\t%Lu\n", vm->immediate.fence_context);
seq_printf(m, "drm-memory-vram:\t%llu KiB\n", vram_mem/1024UL);
seq_printf(m, "drm-memory-gtt: \t%llu KiB\n", gtt_mem/1024UL);
seq_printf(m, "drm-memory-cpu: \t%llu KiB\n", cpu_mem/1024UL);
for (hw_ip = 0; hw_ip < AMDGPU_HW_IP_NUM; ++hw_ip) {
if (!usage[hw_ip])
continue;
seq_printf(m, "%s%d:\t%d.%d%%\n",
amdgpu_ip_name[i],
idx, perc/100, frac);
}
seq_printf(m, "drm-engine-%s:\t%Ld ns\n", amdgpu_ip_name[hw_ip],
ktime_to_ns(usage[hw_ip]));
}
}
......@@ -296,8 +296,8 @@ int amdgpu_gem_create_ioctl(struct drm_device *dev, void *data,
AMDGPU_GEM_CREATE_VRAM_CLEARED |
AMDGPU_GEM_CREATE_VM_ALWAYS_VALID |
AMDGPU_GEM_CREATE_EXPLICIT_SYNC |
AMDGPU_GEM_CREATE_ENCRYPTED))
AMDGPU_GEM_CREATE_ENCRYPTED |
AMDGPU_GEM_CREATE_DISCARDABLE))
return -EINVAL;
/* reject invalid gem domains */
......@@ -645,6 +645,8 @@ uint64_t amdgpu_gem_va_map_flags(struct amdgpu_device *adev, uint32_t flags)
pte_flag |= AMDGPU_PTE_WRITEABLE;
if (flags & AMDGPU_VM_PAGE_PRT)
pte_flag |= AMDGPU_PTE_PRT;
if (flags & AMDGPU_VM_PAGE_NOALLOC)
pte_flag |= AMDGPU_PTE_NOALLOC;
if (adev->gmc.gmc_funcs->map_mtype)
pte_flag |= amdgpu_gmc_map_mtype(adev,
......@@ -658,7 +660,8 @@ int amdgpu_gem_va_ioctl(struct drm_device *dev, void *data,
{
const uint32_t valid_flags = AMDGPU_VM_DELAY_UPDATE |
AMDGPU_VM_PAGE_READABLE | AMDGPU_VM_PAGE_WRITEABLE |
AMDGPU_VM_PAGE_EXECUTABLE | AMDGPU_VM_MTYPE_MASK;
AMDGPU_VM_PAGE_EXECUTABLE | AMDGPU_VM_MTYPE_MASK |
AMDGPU_VM_PAGE_NOALLOC;
const uint32_t prt_flags = AMDGPU_VM_DELAY_UPDATE |
AMDGPU_VM_PAGE_PRT;
......
......@@ -567,6 +567,7 @@ int amdgpu_bo_create(struct amdgpu_device *adev,
bp->domain;
bo->allowed_domains = bo->preferred_domains;
if (bp->type != ttm_bo_type_kernel &&
!(bp->flags & AMDGPU_GEM_CREATE_DISCARDABLE) &&
bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
......
......@@ -41,7 +41,6 @@
/* BO flag to indicate a KFD userptr BO */
#define AMDGPU_AMDKFD_CREATE_USERPTR_BO (1ULL << 63)
#define AMDGPU_AMDKFD_CREATE_SVM_BO (1ULL << 62)
#define to_amdgpu_bo_user(abo) container_of((abo), struct amdgpu_bo_user, bo)
#define to_amdgpu_bo_vm(abo) container_of((abo), struct amdgpu_bo_vm, bo)
......
......@@ -117,7 +117,7 @@ static void amdgpu_evict_flags(struct ttm_buffer_object *bo,
}
abo = ttm_to_amdgpu_bo(bo);
if (abo->flags & AMDGPU_AMDKFD_CREATE_SVM_BO) {
if (abo->flags & AMDGPU_GEM_CREATE_DISCARDABLE) {
placement->num_placement = 0;
placement->num_busy_placement = 0;
return;
......
......@@ -613,6 +613,9 @@ static void gmc_v10_0_get_vm_pte(struct amdgpu_device *adev,
*flags &= ~AMDGPU_PTE_MTYPE_NV10_MASK;
*flags |= (mapping->flags & AMDGPU_PTE_MTYPE_NV10_MASK);
*flags &= ~AMDGPU_PTE_NOALLOC;
*flags |= (mapping->flags & AMDGPU_PTE_NOALLOC);
if (mapping->flags & AMDGPU_PTE_PRT) {
*flags |= AMDGPU_PTE_PRT;
*flags |= AMDGPU_PTE_SNOOPED;
......
......@@ -500,6 +500,9 @@ static void gmc_v11_0_get_vm_pte(struct amdgpu_device *adev,
*flags &= ~AMDGPU_PTE_MTYPE_NV10_MASK;
*flags |= (mapping->flags & AMDGPU_PTE_MTYPE_NV10_MASK);
*flags &= ~AMDGPU_PTE_NOALLOC;
*flags |= (mapping->flags & AMDGPU_PTE_NOALLOC);
if (mapping->flags & AMDGPU_PTE_PRT) {
*flags |= AMDGPU_PTE_PRT;
*flags |= AMDGPU_PTE_SNOOPED;
......
......@@ -531,7 +531,7 @@ svm_range_vram_node_new(struct amdgpu_device *adev, struct svm_range *prange,
bp.domain = AMDGPU_GEM_DOMAIN_VRAM;
bp.flags = AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
bp.flags |= clear ? AMDGPU_GEM_CREATE_VRAM_CLEARED : 0;
bp.flags |= AMDGPU_AMDKFD_CREATE_SVM_BO;
bp.flags |= AMDGPU_GEM_CREATE_DISCARDABLE;
bp.type = ttm_bo_type_device;
bp.resv = NULL;
......
......@@ -140,6 +140,10 @@ extern "C" {
* not require GTT memory accounting
*/
#define AMDGPU_GEM_CREATE_PREEMPTIBLE (1 << 11)
/* Flag that BO can be discarded under memory pressure without keeping the
* content.
*/
#define AMDGPU_GEM_CREATE_DISCARDABLE (1 << 12)
struct drm_amdgpu_gem_create_in {
/** the requested memory size */
......@@ -529,6 +533,8 @@ struct drm_amdgpu_gem_op {
#define AMDGPU_VM_MTYPE_UC (4 << 5)
/* Use Read Write MTYPE instead of default MTYPE */
#define AMDGPU_VM_MTYPE_RW (5 << 5)
/* don't allocate MALL */
#define AMDGPU_VM_PAGE_NOALLOC (1 << 9)
struct drm_amdgpu_gem_va {
/** GEM object handle */
......
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