linux-zen-server/drivers/accel/habanalabs/common/command_buffer.c

555 lines
12 KiB
C
Raw Permalink Normal View History

2023-08-30 17:53:23 +02:00
// SPDX-License-Identifier: GPL-2.0
/*
* Copyright 2016-2019 HabanaLabs, Ltd.
* All Rights Reserved.
*/
#include <uapi/drm/habanalabs_accel.h>
#include "habanalabs.h"
#include <linux/mm.h>
#include <linux/slab.h>
#include <linux/uaccess.h>
#define CB_VA_POOL_SIZE (4UL * SZ_1G)
static int cb_map_mem(struct hl_ctx *ctx, struct hl_cb *cb)
{
struct hl_device *hdev = ctx->hdev;
struct asic_fixed_properties *prop = &hdev->asic_prop;
u32 page_size = prop->pmmu.page_size;
int rc;
if (!hdev->supports_cb_mapping) {
dev_err_ratelimited(hdev->dev,
"Mapping a CB to the device's MMU is not supported\n");
return -EINVAL;
}
if (!hdev->mmu_enable) {
dev_err_ratelimited(hdev->dev,
"Cannot map CB because MMU is disabled\n");
return -EINVAL;
}
if (cb->is_mmu_mapped)
return 0;
cb->roundup_size = roundup(cb->size, page_size);
cb->virtual_addr = (u64) gen_pool_alloc(ctx->cb_va_pool, cb->roundup_size);
if (!cb->virtual_addr) {
dev_err(hdev->dev, "Failed to allocate device virtual address for CB\n");
return -ENOMEM;
}
mutex_lock(&hdev->mmu_lock);
rc = hl_mmu_map_contiguous(ctx, cb->virtual_addr, cb->bus_address, cb->roundup_size);
if (rc) {
dev_err(hdev->dev, "Failed to map VA %#llx to CB\n", cb->virtual_addr);
goto err_va_umap;
}
rc = hl_mmu_invalidate_cache(hdev, false, MMU_OP_USERPTR | MMU_OP_SKIP_LOW_CACHE_INV);
mutex_unlock(&hdev->mmu_lock);
cb->is_mmu_mapped = true;
return rc;
err_va_umap:
mutex_unlock(&hdev->mmu_lock);
gen_pool_free(ctx->cb_va_pool, cb->virtual_addr, cb->roundup_size);
return rc;
}
static void cb_unmap_mem(struct hl_ctx *ctx, struct hl_cb *cb)
{
struct hl_device *hdev = ctx->hdev;
mutex_lock(&hdev->mmu_lock);
hl_mmu_unmap_contiguous(ctx, cb->virtual_addr, cb->roundup_size);
hl_mmu_invalidate_cache(hdev, true, MMU_OP_USERPTR);
mutex_unlock(&hdev->mmu_lock);
gen_pool_free(ctx->cb_va_pool, cb->virtual_addr, cb->roundup_size);
}
static void cb_fini(struct hl_device *hdev, struct hl_cb *cb)
{
if (cb->is_internal)
gen_pool_free(hdev->internal_cb_pool,
(uintptr_t)cb->kernel_address, cb->size);
else
hl_asic_dma_free_coherent(hdev, cb->size, cb->kernel_address, cb->bus_address);
kfree(cb);
}
static void cb_do_release(struct hl_device *hdev, struct hl_cb *cb)
{
if (cb->is_pool) {
atomic_set(&cb->is_handle_destroyed, 0);
spin_lock(&hdev->cb_pool_lock);
list_add(&cb->pool_list, &hdev->cb_pool);
spin_unlock(&hdev->cb_pool_lock);
} else {
cb_fini(hdev, cb);
}
}
static struct hl_cb *hl_cb_alloc(struct hl_device *hdev, u32 cb_size,
int ctx_id, bool internal_cb)
{
struct hl_cb *cb = NULL;
u32 cb_offset;
void *p;
/*
* We use of GFP_ATOMIC here because this function can be called from
* the latency-sensitive code path for command submission. Due to H/W
* limitations in some of the ASICs, the kernel must copy the user CB
* that is designated for an external queue and actually enqueue
* the kernel's copy. Hence, we must never sleep in this code section
* and must use GFP_ATOMIC for all memory allocations.
*/
if (ctx_id == HL_KERNEL_ASID_ID && !hdev->disabled)
cb = kzalloc(sizeof(*cb), GFP_ATOMIC);
if (!cb)
cb = kzalloc(sizeof(*cb), GFP_KERNEL);
if (!cb)
return NULL;
if (internal_cb) {
p = (void *) gen_pool_alloc(hdev->internal_cb_pool, cb_size);
if (!p) {
kfree(cb);
return NULL;
}
cb_offset = p - hdev->internal_cb_pool_virt_addr;
cb->is_internal = true;
cb->bus_address = hdev->internal_cb_va_base + cb_offset;
} else if (ctx_id == HL_KERNEL_ASID_ID) {
p = hl_asic_dma_alloc_coherent(hdev, cb_size, &cb->bus_address, GFP_ATOMIC);
if (!p)
p = hl_asic_dma_alloc_coherent(hdev, cb_size, &cb->bus_address, GFP_KERNEL);
} else {
p = hl_asic_dma_alloc_coherent(hdev, cb_size, &cb->bus_address,
GFP_USER | __GFP_ZERO);
}
if (!p) {
dev_err(hdev->dev,
"failed to allocate %d of dma memory for CB\n",
cb_size);
kfree(cb);
return NULL;
}
cb->kernel_address = p;
cb->size = cb_size;
return cb;
}
struct hl_cb_mmap_mem_alloc_args {
struct hl_device *hdev;
struct hl_ctx *ctx;
u32 cb_size;
bool internal_cb;
bool map_cb;
};
static void hl_cb_mmap_mem_release(struct hl_mmap_mem_buf *buf)
{
struct hl_cb *cb = buf->private;
hl_debugfs_remove_cb(cb);
if (cb->is_mmu_mapped)
cb_unmap_mem(cb->ctx, cb);
hl_ctx_put(cb->ctx);
cb_do_release(cb->hdev, cb);
}
static int hl_cb_mmap_mem_alloc(struct hl_mmap_mem_buf *buf, gfp_t gfp, void *args)
{
struct hl_cb_mmap_mem_alloc_args *cb_args = args;
struct hl_cb *cb;
int rc, ctx_id = cb_args->ctx->asid;
bool alloc_new_cb = true;
if (!cb_args->internal_cb) {
/* Minimum allocation must be PAGE SIZE */
if (cb_args->cb_size < PAGE_SIZE)
cb_args->cb_size = PAGE_SIZE;
if (ctx_id == HL_KERNEL_ASID_ID &&
cb_args->cb_size <= cb_args->hdev->asic_prop.cb_pool_cb_size) {
spin_lock(&cb_args->hdev->cb_pool_lock);
if (!list_empty(&cb_args->hdev->cb_pool)) {
cb = list_first_entry(&cb_args->hdev->cb_pool,
typeof(*cb), pool_list);
list_del(&cb->pool_list);
spin_unlock(&cb_args->hdev->cb_pool_lock);
alloc_new_cb = false;
} else {
spin_unlock(&cb_args->hdev->cb_pool_lock);
dev_dbg(cb_args->hdev->dev, "CB pool is empty\n");
}
}
}
if (alloc_new_cb) {
cb = hl_cb_alloc(cb_args->hdev, cb_args->cb_size, ctx_id, cb_args->internal_cb);
if (!cb)
return -ENOMEM;
}
cb->hdev = cb_args->hdev;
cb->ctx = cb_args->ctx;
cb->buf = buf;
cb->buf->mappable_size = cb->size;
cb->buf->private = cb;
hl_ctx_get(cb->ctx);
if (cb_args->map_cb) {
if (ctx_id == HL_KERNEL_ASID_ID) {
dev_err(cb_args->hdev->dev,
"CB mapping is not supported for kernel context\n");
rc = -EINVAL;
goto release_cb;
}
rc = cb_map_mem(cb_args->ctx, cb);
if (rc)
goto release_cb;
}
hl_debugfs_add_cb(cb);
return 0;
release_cb:
hl_ctx_put(cb->ctx);
cb_do_release(cb_args->hdev, cb);
return rc;
}
static int hl_cb_mmap(struct hl_mmap_mem_buf *buf,
struct vm_area_struct *vma, void *args)
{
struct hl_cb *cb = buf->private;
return cb->hdev->asic_funcs->mmap(cb->hdev, vma, cb->kernel_address,
cb->bus_address, cb->size);
}
static struct hl_mmap_mem_buf_behavior cb_behavior = {
.topic = "CB",
.mem_id = HL_MMAP_TYPE_CB,
.alloc = hl_cb_mmap_mem_alloc,
.release = hl_cb_mmap_mem_release,
.mmap = hl_cb_mmap,
};
int hl_cb_create(struct hl_device *hdev, struct hl_mem_mgr *mmg,
struct hl_ctx *ctx, u32 cb_size, bool internal_cb,
bool map_cb, u64 *handle)
{
struct hl_cb_mmap_mem_alloc_args args = {
.hdev = hdev,
.ctx = ctx,
.cb_size = cb_size,
.internal_cb = internal_cb,
.map_cb = map_cb,
};
struct hl_mmap_mem_buf *buf;
int ctx_id = ctx->asid;
if ((hdev->disabled) || (hdev->reset_info.in_reset && (ctx_id != HL_KERNEL_ASID_ID))) {
dev_warn_ratelimited(hdev->dev,
"Device is disabled or in reset. Can't create new CBs\n");
return -EBUSY;
}
if (cb_size > SZ_2M) {
dev_err(hdev->dev, "CB size %d must be less than %d\n",
cb_size, SZ_2M);
return -EINVAL;
}
buf = hl_mmap_mem_buf_alloc(
mmg, &cb_behavior,
ctx_id == HL_KERNEL_ASID_ID ? GFP_ATOMIC : GFP_KERNEL, &args);
if (!buf)
return -ENOMEM;
*handle = buf->handle;
return 0;
}
int hl_cb_destroy(struct hl_mem_mgr *mmg, u64 cb_handle)
{
struct hl_cb *cb;
int rc;
cb = hl_cb_get(mmg, cb_handle);
if (!cb) {
dev_dbg(mmg->dev, "CB destroy failed, no CB was found for handle %#llx\n",
cb_handle);
return -EINVAL;
}
/* Make sure that CB handle isn't destroyed more than once */
rc = atomic_cmpxchg(&cb->is_handle_destroyed, 0, 1);
hl_cb_put(cb);
if (rc) {
dev_dbg(mmg->dev, "CB destroy failed, handle %#llx was already destroyed\n",
cb_handle);
return -EINVAL;
}
rc = hl_mmap_mem_buf_put_handle(mmg, cb_handle);
if (rc < 0)
return rc; /* Invalid handle */
if (rc == 0)
dev_dbg(mmg->dev, "CB 0x%llx is destroyed while still in use\n", cb_handle);
return 0;
}
static int hl_cb_info(struct hl_mem_mgr *mmg,
u64 handle, u32 flags, u32 *usage_cnt, u64 *device_va)
{
struct hl_cb *cb;
int rc = 0;
cb = hl_cb_get(mmg, handle);
if (!cb) {
dev_err(mmg->dev,
"CB info failed, no match to handle 0x%llx\n", handle);
return -EINVAL;
}
if (flags & HL_CB_FLAGS_GET_DEVICE_VA) {
if (cb->is_mmu_mapped) {
*device_va = cb->virtual_addr;
} else {
dev_err(mmg->dev, "CB is not mapped to the device's MMU\n");
rc = -EINVAL;
goto out;
}
} else {
*usage_cnt = atomic_read(&cb->cs_cnt);
}
out:
hl_cb_put(cb);
return rc;
}
int hl_cb_ioctl(struct hl_fpriv *hpriv, void *data)
{
union hl_cb_args *args = data;
struct hl_device *hdev = hpriv->hdev;
u64 handle = 0, device_va = 0;
enum hl_device_status status;
u32 usage_cnt = 0;
int rc;
if (!hl_device_operational(hdev, &status)) {
dev_dbg_ratelimited(hdev->dev,
"Device is %s. Can't execute CB IOCTL\n",
hdev->status[status]);
return -EBUSY;
}
switch (args->in.op) {
case HL_CB_OP_CREATE:
if (args->in.cb_size > HL_MAX_CB_SIZE) {
dev_err(hdev->dev,
"User requested CB size %d must be less than %d\n",
args->in.cb_size, HL_MAX_CB_SIZE);
rc = -EINVAL;
} else {
rc = hl_cb_create(hdev, &hpriv->mem_mgr, hpriv->ctx,
args->in.cb_size, false,
!!(args->in.flags & HL_CB_FLAGS_MAP),
&handle);
}
memset(args, 0, sizeof(*args));
args->out.cb_handle = handle;
break;
case HL_CB_OP_DESTROY:
rc = hl_cb_destroy(&hpriv->mem_mgr,
args->in.cb_handle);
break;
case HL_CB_OP_INFO:
rc = hl_cb_info(&hpriv->mem_mgr, args->in.cb_handle,
args->in.flags,
&usage_cnt,
&device_va);
if (rc)
break;
memset(&args->out, 0, sizeof(args->out));
if (args->in.flags & HL_CB_FLAGS_GET_DEVICE_VA)
args->out.device_va = device_va;
else
args->out.usage_cnt = usage_cnt;
break;
default:
rc = -EINVAL;
break;
}
return rc;
}
struct hl_cb *hl_cb_get(struct hl_mem_mgr *mmg, u64 handle)
{
struct hl_mmap_mem_buf *buf;
buf = hl_mmap_mem_buf_get(mmg, handle);
if (!buf)
return NULL;
return buf->private;
}
void hl_cb_put(struct hl_cb *cb)
{
hl_mmap_mem_buf_put(cb->buf);
}
struct hl_cb *hl_cb_kernel_create(struct hl_device *hdev, u32 cb_size,
bool internal_cb)
{
u64 cb_handle;
struct hl_cb *cb;
int rc;
rc = hl_cb_create(hdev, &hdev->kernel_mem_mgr, hdev->kernel_ctx, cb_size,
internal_cb, false, &cb_handle);
if (rc) {
dev_err(hdev->dev,
"Failed to allocate CB for the kernel driver %d\n", rc);
return NULL;
}
cb = hl_cb_get(&hdev->kernel_mem_mgr, cb_handle);
/* hl_cb_get should never fail here */
if (!cb) {
dev_crit(hdev->dev, "Kernel CB handle invalid 0x%x\n",
(u32) cb_handle);
goto destroy_cb;
}
return cb;
destroy_cb:
hl_cb_destroy(&hdev->kernel_mem_mgr, cb_handle);
return NULL;
}
int hl_cb_pool_init(struct hl_device *hdev)
{
struct hl_cb *cb;
int i;
INIT_LIST_HEAD(&hdev->cb_pool);
spin_lock_init(&hdev->cb_pool_lock);
for (i = 0 ; i < hdev->asic_prop.cb_pool_cb_cnt ; i++) {
cb = hl_cb_alloc(hdev, hdev->asic_prop.cb_pool_cb_size,
HL_KERNEL_ASID_ID, false);
if (cb) {
cb->is_pool = true;
list_add(&cb->pool_list, &hdev->cb_pool);
} else {
hl_cb_pool_fini(hdev);
return -ENOMEM;
}
}
return 0;
}
int hl_cb_pool_fini(struct hl_device *hdev)
{
struct hl_cb *cb, *tmp;
list_for_each_entry_safe(cb, tmp, &hdev->cb_pool, pool_list) {
list_del(&cb->pool_list);
cb_fini(hdev, cb);
}
return 0;
}
int hl_cb_va_pool_init(struct hl_ctx *ctx)
{
struct hl_device *hdev = ctx->hdev;
struct asic_fixed_properties *prop = &hdev->asic_prop;
int rc;
if (!hdev->supports_cb_mapping)
return 0;
ctx->cb_va_pool = gen_pool_create(__ffs(prop->pmmu.page_size), -1);
if (!ctx->cb_va_pool) {
dev_err(hdev->dev,
"Failed to create VA gen pool for CB mapping\n");
return -ENOMEM;
}
ctx->cb_va_pool_base = hl_reserve_va_block(hdev, ctx, HL_VA_RANGE_TYPE_HOST,
CB_VA_POOL_SIZE, HL_MMU_VA_ALIGNMENT_NOT_NEEDED);
if (!ctx->cb_va_pool_base) {
rc = -ENOMEM;
goto err_pool_destroy;
}
rc = gen_pool_add(ctx->cb_va_pool, ctx->cb_va_pool_base, CB_VA_POOL_SIZE, -1);
if (rc) {
dev_err(hdev->dev,
"Failed to add memory to VA gen pool for CB mapping\n");
goto err_unreserve_va_block;
}
return 0;
err_unreserve_va_block:
hl_unreserve_va_block(hdev, ctx, ctx->cb_va_pool_base, CB_VA_POOL_SIZE);
err_pool_destroy:
gen_pool_destroy(ctx->cb_va_pool);
return rc;
}
void hl_cb_va_pool_fini(struct hl_ctx *ctx)
{
struct hl_device *hdev = ctx->hdev;
if (!hdev->supports_cb_mapping)
return;
gen_pool_destroy(ctx->cb_va_pool);
hl_unreserve_va_block(hdev, ctx, ctx->cb_va_pool_base, CB_VA_POOL_SIZE);
}