idle_mask);
} else {
dev_dbg(hdev->dev, "closing kernel context\n");
+ hdev->asic_funcs->ctx_fini(ctx);
+ hl_vm_ctx_fini(ctx);
hl_mmu_ctx_fini(ctx);
}
}
if (is_kernel_ctx) {
ctx->asid = HL_KERNEL_ASID_ID; /* Kernel driver gets ASID 0 */
- rc = hl_mmu_ctx_init(ctx);
+ rc = hl_vm_ctx_init(ctx);
if (rc) {
- dev_err(hdev->dev, "Failed to init mmu ctx module\n");
+ dev_err(hdev->dev, "Failed to init mem ctx module\n");
+ rc = -ENOMEM;
goto err_free_cs_pending;
}
+
+ rc = hdev->asic_funcs->ctx_init(ctx);
+ if (rc) {
+ dev_err(hdev->dev, "ctx_init failed\n");
+ goto err_vm_ctx_fini;
+ }
} else {
ctx->asid = hl_asid_alloc(hdev);
if (!ctx->asid) {
err_vm_ctx_fini:
hl_vm_ctx_fini(ctx);
err_asid_free:
- hl_asid_free(hdev, ctx->asid);
+ if (ctx->asid != HL_KERNEL_ASID_ID)
+ hl_asid_free(hdev, ctx->asid);
err_free_cs_pending:
kfree(ctx->cs_pending);
hdev->compute_ctx = NULL;
+ hl_debugfs_add_device(hdev);
+
+ /* debugfs nodes are created in hl_ctx_init so it must be called after
+ * hl_debugfs_add_device.
+ */
rc = hl_ctx_init(hdev, hdev->kernel_ctx, true);
if (rc) {
dev_err(hdev->dev, "failed to initialize kernel context\n");
kfree(hdev->kernel_ctx);
- goto mmu_fini;
+ goto remove_device_from_debugfs;
}
rc = hl_cb_pool_init(hdev);
goto release_ctx;
}
- hl_debugfs_add_device(hdev);
-
/*
* From this point, in case of an error, add char devices and create
* sysfs nodes as part of the error flow, to allow debugging.
if (hl_ctx_put(hdev->kernel_ctx) != 1)
dev_err(hdev->dev,
"kernel ctx is still alive on initialization failure\n");
+remove_device_from_debugfs:
+ hl_debugfs_remove_device(hdev);
mmu_fini:
hl_mmu_fini(hdev);
eq_fini:
device_late_fini(hdev);
- hl_debugfs_remove_device(hdev);
-
/*
* Halt the engines and disable interrupts so we won't get any more
* completions from H/W and we won't have any accesses from the
if ((hdev->kernel_ctx) && (hl_ctx_put(hdev->kernel_ctx) != 1))
dev_err(hdev->dev, "kernel ctx is still alive\n");
+ hl_debugfs_remove_device(hdev);
+
hl_vm_fini(hdev);
hl_mmu_fini(hdev);
* because the user notifies us on allocations. If the user is no more,
* all DRAM is available
*/
- if (!ctx->hdev->asic_prop.dram_supports_virtual_memory)
+ if (ctx->asid != HL_KERNEL_ASID_ID &&
+ !ctx->hdev->asic_prop.dram_supports_virtual_memory)
atomic64_set(&ctx->hdev->dram_used_mem, 0);
}
static int gaudi_ctx_init(struct hl_ctx *ctx)
{
+ if (ctx->asid == HL_KERNEL_ASID_ID)
+ return 0;
+
gaudi_mmu_prepare(ctx->hdev, ctx->asid);
return gaudi_internal_cb_pool_init(ctx->hdev, ctx);
}
static void gaudi_ctx_fini(struct hl_ctx *ctx)
{
- struct hl_device *hdev = ctx->hdev;
-
- /* Gaudi will NEVER support more then a single compute context.
- * Therefore, don't clear anything unless it is the compute context
- */
- if (hdev->compute_ctx != ctx)
+ if (ctx->asid == HL_KERNEL_ASID_ID)
return;
gaudi_internal_cb_pool_fini(ctx->hdev, ctx);