habanalabs: disable CPU access on timeouts

This patch provides a workaround for a bug in the F/W where the response
time for a request from KMD may take more then 100ms. This could cause the
queue between KMD and the F/W to get out of sync.

The WA is to:
1. Increase the timeout of ALL requests to 1s.
2. In case a request isn't answered in time, mark the state as
"cpu_disabled" and prevent sending further requests from KMD to the F/W.
This will eventually lead to a heartbeat failure and hard reset of the
device.
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

drivers/misc/habanalabs/debugfs.c

@@ -723,7 +723,7 @@ static ssize_t hl_device_read(struct file *f, char __user *buf,
 		return 0;
 
 	sprintf(tmp_buf,
-		"Valid values are: disable, enable, suspend, resume\n");
+		"Valid values: disable, enable, suspend, resume, cpu_timeout\n");
 	rc = simple_read_from_buffer(buf, strlen(tmp_buf) + 1, ppos, tmp_buf,
 			strlen(tmp_buf) + 1);
 
@@ -751,9 +751,11 @@ static ssize_t hl_device_write(struct file *f, const char __user *buf,
 		hdev->asic_funcs->suspend(hdev);
 	} else if (strncmp("resume", data, strlen("resume")) == 0) {
 		hdev->asic_funcs->resume(hdev);
+	} else if (strncmp("cpu_timeout", data, strlen("cpu_timeout")) == 0) {
+		hdev->device_cpu_disabled = true;
 	} else {
 		dev_err(hdev->dev,
-			"Valid values are: disable, enable, suspend, resume\n");
+			"Valid values: disable, enable, suspend, resume, cpu_timeout\n");
 		count = -EINVAL;
 	}
 

drivers/misc/habanalabs/device.c

@@ -636,6 +636,8 @@ int hl_device_reset(struct hl_device *hdev, bool hard_reset,
 	/* Finished tear-down, starting to re-initialize */
 
 	if (hard_reset) {
+		hdev->device_cpu_disabled = false;
+
 		/* Allocate the kernel context */
 		hdev->kernel_ctx = kzalloc(sizeof(*hdev->kernel_ctx),
 						GFP_KERNEL);

drivers/misc/habanalabs/goya/goya.c

@@ -3232,6 +3232,11 @@ int goya_send_cpu_message(struct hl_device *hdev, u32 *msg, u16 len,
 	if (hdev->disabled)
 		goto out;
 
+	if (hdev->device_cpu_disabled) {
+		rc = -EIO;
+		goto out;
+	}
+
 	rc = hl_hw_queue_send_cb_no_cmpl(hdev, GOYA_QUEUE_ID_CPU_PQ, len,
 			pkt_dma_addr);
 	if (rc) {
@@ -3245,8 +3250,8 @@ int goya_send_cpu_message(struct hl_device *hdev, u32 *msg, u16 len,
 	hl_hw_queue_inc_ci_kernel(hdev, GOYA_QUEUE_ID_CPU_PQ);
 
 	if (rc == -ETIMEDOUT) {
-		dev_err(hdev->dev,
-			"Timeout while waiting for CPU packet fence\n");
+		dev_err(hdev->dev, "Timeout while waiting for device CPU\n");
+		hdev->device_cpu_disabled = true;
 		goto out;
 	}
 

drivers/misc/habanalabs/habanalabs.h

@@ -1079,6 +1079,7 @@ struct hl_device_reset_work {
  * @dram_default_page_mapping: is DRAM default page mapping enabled.
  * @init_done: is the initialization of the device done.
  * @mmu_enable: is MMU enabled.
+ * @device_cpu_disabled: is the device CPU disabled (due to timeouts)
  */
 struct hl_device {
 	struct pci_dev			*pdev;
@@ -1146,6 +1147,7 @@ struct hl_device {
 	u8				dram_supports_virtual_memory;
 	u8				dram_default_page_mapping;
 	u8				init_done;
+	u8				device_cpu_disabled;
 
 	/* Parameters for bring-up */
 	u8				mmu_enable;

drivers/misc/habanalabs/hwmon.c

@@ -10,7 +10,7 @@
 #include <linux/pci.h>
 #include <linux/hwmon.h>
 
-#define SENSORS_PKT_TIMEOUT		100000	/* 100ms */
+#define SENSORS_PKT_TIMEOUT		1000000	/* 1s */
 #define HWMON_NR_SENSOR_TYPES		(hwmon_pwm + 1)
 
 int hl_build_hwmon_channel_info(struct hl_device *hdev,

drivers/misc/habanalabs/sysfs.c

@@ -9,8 +9,8 @@
 
 #include <linux/pci.h>
 
-#define SET_CLK_PKT_TIMEOUT	200000	/* 200ms */
-#define SET_PWR_PKT_TIMEOUT	400000	/* 400ms */
+#define SET_CLK_PKT_TIMEOUT	1000000	/* 1s */
+#define SET_PWR_PKT_TIMEOUT	1000000	/* 1s */
 
 long hl_get_frequency(struct hl_device *hdev, u32 pll_index, bool curr)
 {