habanalabs: expose state dump

To improve the user's ability to debug the case where a workload that
is part of executing training/inference of a topology is getting stuck,
we need to add a 'core dump' each time a CS times-out. The 'core dump'
shall contain all relevant Sync Manager information and corresponding
fence values.

The most recent dumps shall be accessible via debugfs, under
'state_dump' node. Reading from the node will provide the oldest dump
available. Writing an integer value X will discard X dumps, starting
with the oldest one, i.e. subsequent read will now return newer
dumps.
Signed-off-by: Yuri Nudelman <ynudelman@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>

Documentation/ABI/testing/debugfs-driver-habanalabs

@@ -215,6 +215,17 @@ Description:    Sets the skip reset on timeout option for the device. Value of
                 "0" means device will be reset in case some CS has timed out,
                 otherwise it will not be reset.
 
+What:           /sys/kernel/debug/habanalabs/hl<n>/state_dump
+Date:           Oct 2021
+KernelVersion:  5.15
+Contact:        ynudelman@habana.ai
+Description:    Gets the state dump occurring on a CS timeout or failure.
+                State dump is used for debug and is created each time in case of
+                a problem in a CS execution, before reset.
+                Reading from the node returns the newest state dump available.
+                Writing an integer X discards X state dumps, so that the
+                next read would return X+1-st newest state dump.
+
 What:           /sys/kernel/debug/habanalabs/hl<n>/stop_on_err
 Date:           Mar 2020
 KernelVersion:  5.6

drivers/misc/habanalabs/common/Makefile

@@ -10,4 +10,5 @@ HL_COMMON_FILES := common/habanalabs_drv.o common/device.o common/context.o \
 		common/asid.o common/habanalabs_ioctl.o \
 		common/command_buffer.o common/hw_queue.o common/irq.o \
 		common/sysfs.o common/hwmon.o common/memory.o \
-		common/command_submission.o common/firmware_if.o
+		common/command_submission.o common/firmware_if.o \
+		common/state_dump.o

drivers/misc/habanalabs/common/command_submission.c

@@ -621,6 +621,10 @@ static void cs_timedout(struct work_struct *work)
 		break;
 	}
 
+	rc = hl_state_dump(hdev);
+	if (rc)
+		dev_err(hdev->dev, "Error during system state dump %d\n", rc);
+
 	cs_put(cs);
 
 	if (likely(!skip_reset_on_timeout)) {

drivers/misc/habanalabs/common/debugfs.c

@@ -1043,6 +1043,60 @@ static ssize_t hl_security_violations_read(struct file *f, char __user *buf,
 	return 0;
 }
 
+static ssize_t hl_state_dump_read(struct file *f, char __user *buf,
+					size_t count, loff_t *ppos)
+{
+	struct hl_dbg_device_entry *entry = file_inode(f)->i_private;
+	ssize_t rc;
+
+	down_read(&entry->state_dump_sem);
+	if (!entry->state_dump[entry->state_dump_head])
+		rc = 0;
+	else
+		rc = simple_read_from_buffer(
+			buf, count, ppos,
+			entry->state_dump[entry->state_dump_head],
+			strlen(entry->state_dump[entry->state_dump_head]));
+	up_read(&entry->state_dump_sem);
+
+	return rc;
+}
+
+static ssize_t hl_state_dump_write(struct file *f, const char __user *buf,
+					size_t count, loff_t *ppos)
+{
+	struct hl_dbg_device_entry *entry = file_inode(f)->i_private;
+	struct hl_device *hdev = entry->hdev;
+	ssize_t rc;
+	u32 size;
+	int i;
+
+	rc = kstrtouint_from_user(buf, count, 10, &size);
+	if (rc)
+		return rc;
+
+	if (size <= 0 || size >= ARRAY_SIZE(entry->state_dump)) {
+		dev_err(hdev->dev, "Invalid number of dumps to skip\n");
+		return -EINVAL;
+	}
+
+	if (entry->state_dump[entry->state_dump_head]) {
+		down_write(&entry->state_dump_sem);
+		for (i = 0; i < size; ++i) {
+			vfree(entry->state_dump[entry->state_dump_head]);
+			entry->state_dump[entry->state_dump_head] = NULL;
+			if (entry->state_dump_head > 0)
+				entry->state_dump_head--;
+			else
+				entry->state_dump_head =
+					ARRAY_SIZE(entry->state_dump) - 1;
+		}
+		up_write(&entry->state_dump_sem);
+	}
+
+	return count;
+}
+
 static const struct file_operations hl_data32b_fops = {
 	.owner = THIS_MODULE,
 	.read = hl_data_read32,
@@ -1110,6 +1164,12 @@ static const struct file_operations hl_security_violations_fops = {
 	.read = hl_security_violations_read
 };
 
+static const struct file_operations hl_state_dump_fops = {
+	.owner = THIS_MODULE,
+	.read = hl_state_dump_read,
+	.write = hl_state_dump_write
+};
+
 static const struct hl_info_list hl_debugfs_list[] = {
 	{"command_buffers", command_buffers_show, NULL},
 	{"command_submission", command_submission_show, NULL},
@@ -1172,6 +1232,7 @@ void hl_debugfs_add_device(struct hl_device *hdev)
 	INIT_LIST_HEAD(&dev_entry->userptr_list);
 	INIT_LIST_HEAD(&dev_entry->ctx_mem_hash_list);
 	mutex_init(&dev_entry->file_mutex);
+	init_rwsem(&dev_entry->state_dump_sem);
 	spin_lock_init(&dev_entry->cb_spinlock);
 	spin_lock_init(&dev_entry->cs_spinlock);
 	spin_lock_init(&dev_entry->cs_job_spinlock);
@@ -1283,6 +1344,12 @@ void hl_debugfs_add_device(struct hl_device *hdev)
 				dev_entry->root,
 				&hdev->skip_reset_on_timeout);
 
+	debugfs_create_file("state_dump",
+				0600,
+				dev_entry->root,
+				dev_entry,
+				&hl_state_dump_fops);
+
 	for (i = 0, entry = dev_entry->entry_arr ; i < count ; i++, entry++) {
 		debugfs_create_file(hl_debugfs_list[i].name,
 					0444,
@@ -1297,6 +1364,7 @@ void hl_debugfs_add_device(struct hl_device *hdev)
 void hl_debugfs_remove_device(struct hl_device *hdev)
 {
 	struct hl_dbg_device_entry *entry = &hdev->hl_debugfs;
+	int i;
 
 	debugfs_remove_recursive(entry->root);
 
@@ -1304,6 +1372,9 @@ void hl_debugfs_remove_device(struct hl_device *hdev)
 
 	vfree(entry->blob_desc.data);
 
+	for (i = 0; i < ARRAY_SIZE(entry->state_dump); ++i)
+		vfree(entry->state_dump[i]);
+
 	kfree(entry->entry_arr);
 }
 
@@ -1416,6 +1487,28 @@ void hl_debugfs_remove_ctx_mem_hash(struct hl_device *hdev, struct hl_ctx *ctx)
 	spin_unlock(&dev_entry->ctx_mem_hash_spinlock);
 }
 
+/**
+ * hl_debugfs_set_state_dump - register state dump making it accessible via
+ *                             debugfs
+ * @hdev: pointer to the device structure
+ * @data: the actual dump data
+ * @length: the length of the data
+ */
+void hl_debugfs_set_state_dump(struct hl_device *hdev, char *data,
+					unsigned long length)
+{
+	struct hl_dbg_device_entry *dev_entry = &hdev->hl_debugfs;
+
+	down_write(&dev_entry->state_dump_sem);
+
+	dev_entry->state_dump_head = (dev_entry->state_dump_head + 1) %
+					ARRAY_SIZE(dev_entry->state_dump);
+	vfree(dev_entry->state_dump[dev_entry->state_dump_head]);
+	dev_entry->state_dump[dev_entry->state_dump_head] = data;
+
+	up_write(&dev_entry->state_dump_sem);
+}
+
 void __init hl_debugfs_init(void)
 {
 	hl_debug_root = debugfs_create_dir("habanalabs", NULL);

drivers/misc/habanalabs/common/device.c

@@ -1375,6 +1375,8 @@ int hl_device_init(struct hl_device *hdev, struct class *hclass)
 
 	hdev->compute_ctx = NULL;
 
+	hdev->asic_funcs->state_dump_init(hdev);
+
 	hl_debugfs_add_device(hdev);
 
 	/* debugfs nodes are created in hl_ctx_init so it must be called after

drivers/misc/habanalabs/common/habanalabs.h

@@ -20,6 +20,7 @@
 #include <linux/scatterlist.h>
 #include <linux/hashtable.h>
 #include <linux/debugfs.h>
+#include <linux/rwsem.h>
 #include <linux/bitfield.h>
 #include <linux/genalloc.h>
 #include <linux/sched/signal.h>
@@ -65,6 +66,11 @@
 
 #define HL_COMMON_USER_INTERRUPT_ID	0xFFF
 
+#define HL_STATE_DUMP_HIST_LEN		5
+
+#define OBJ_NAMES_HASH_TABLE_BITS	7 /* 1 << 7 buckets */
+#define SYNC_TO_ENGINE_HASH_TABLE_BITS	7 /* 1 << 7 buckets */
+
 /* Memory */
 #define MEM_HASH_TABLE_BITS		7 /* 1 << 7 buckets */
 
@@ -1123,6 +1129,7 @@ struct fw_load_mgr {
  *                         generic f/w compatible PLL Indexes
  * @init_firmware_loader: initialize data for FW loader.
  * @init_cpu_scrambler_dram: Enable CPU specific DRAM scrambling
+ * @state_dump_init: initialize constants required for state dump
  */
 struct hl_asic_funcs {
 	int (*early_init)(struct hl_device *hdev);
@@ -1248,6 +1255,7 @@ struct hl_asic_funcs {
 	int (*map_pll_idx_to_fw_idx)(u32 pll_idx);
 	void (*init_firmware_loader)(struct hl_device *hdev);
 	void (*init_cpu_scrambler_dram)(struct hl_device *hdev);
+	void (*state_dump_init)(struct hl_device *hdev);
 };
 
 
@@ -1781,9 +1789,12 @@ struct hl_debugfs_entry {
  * @ctx_mem_hash_list: list of available contexts with MMU mappings.
  * @ctx_mem_hash_spinlock: protects cb_list.
  * @blob_desc: descriptor of blob
+ * @state_dump: data of the system states in case of a bad cs.
+ * @state_dump_sem: protects state_dump.
  * @addr: next address to read/write from/to in read/write32.
  * @mmu_addr: next virtual address to translate to physical address in mmu_show.
  * @mmu_asid: ASID to use while translating in mmu_show.
+ * @state_dump_head: index of the latest state dump
  * @i2c_bus: generic u8 debugfs file for bus value to use in i2c_data_read.
  * @i2c_addr: generic u8 debugfs file for address value to use in i2c_data_read.
  * @i2c_reg: generic u8 debugfs file for register value to use in i2c_data_read.
@@ -1805,14 +1816,117 @@ struct hl_dbg_device_entry {
 	struct list_head		ctx_mem_hash_list;
 	spinlock_t			ctx_mem_hash_spinlock;
 	struct debugfs_blob_wrapper	blob_desc;
+	char				*state_dump[HL_STATE_DUMP_HIST_LEN];
+	struct rw_semaphore		state_dump_sem;
 	u64				addr;
 	u64				mmu_addr;
 	u32				mmu_asid;
+	u32				state_dump_head;
 	u8				i2c_bus;
 	u8				i2c_addr;
 	u8				i2c_reg;
 };
 
+/**
+ * struct hl_hw_obj_name_entry - single hw object name, member of
+ * hl_state_dump_specs
+ * @node: link to the containing hash table
+ * @name: hw object name
+ * @id: object identifier
+ */
+struct hl_hw_obj_name_entry {
+	struct hlist_node	node;
+	const char		*name;
+	u32			id;
+};
+
+enum hl_state_dump_specs_props {
+	SP_SYNC_OBJ_BASE_ADDR,
+	SP_NEXT_SYNC_OBJ_ADDR,
+	SP_SYNC_OBJ_AMOUNT,
+	SP_MON_OBJ_WR_ADDR_LOW,
+	SP_MON_OBJ_WR_ADDR_HIGH,
+	SP_MON_OBJ_WR_DATA,
+	SP_MON_OBJ_ARM_DATA,
+	SP_MON_OBJ_STATUS,
+	SP_MONITORS_AMOUNT,
+	SP_TPC0_CMDQ,
+	SP_TPC0_CFG_SO,
+	SP_NEXT_TPC,
+	SP_MME_CMDQ,
+	SP_MME_CFG_SO,
+	SP_NEXT_MME,
+	SP_DMA_CMDQ,
+	SP_DMA_CFG_SO,
+	SP_DMA_QUEUES_OFFSET,
+	SP_NUM_OF_MME_ENGINES,
+	SP_SUB_MME_ENG_NUM,
+	SP_NUM_OF_DMA_ENGINES,
+	SP_NUM_OF_TPC_ENGINES,
+	SP_ENGINE_NUM_OF_QUEUES,
+	SP_ENGINE_NUM_OF_STREAMS,
+	SP_ENGINE_NUM_OF_FENCES,
+	SP_FENCE0_CNT_OFFSET,
+	SP_FENCE0_RDATA_OFFSET,
+	SP_CP_STS_OFFSET,
+	SP_NUM_CORES,
+
+	SP_MAX
+};
+
+enum hl_sync_engine_type {
+	ENGINE_TPC,
+	ENGINE_DMA,
+	ENGINE_MME,
+};
+
+/**
+ * struct hl_sync_to_engine_map_entry - sync object id to engine mapping entry
+ * @engine_type: type of the engine
+ * @engine_id: id of the engine
+ * @sync_id: id of the sync object
+ */
+struct hl_sync_to_engine_map_entry {
+	struct hlist_node		node;
+	enum hl_sync_engine_type	engine_type;
+	u32				engine_id;
+	u32				sync_id;
+};
+
+/**
+ * struct hl_sync_to_engine_map - maps sync object id to associated engine id
+ * @tb: hash table containing the mapping, each element is of type
+ *      struct hl_sync_to_engine_map_entry
+ */
+struct hl_sync_to_engine_map {
+	DECLARE_HASHTABLE(tb, SYNC_TO_ENGINE_HASH_TABLE_BITS);
+};
+
+/**
+ * struct hl_state_dump_specs_funcs - virtual functions used by the state dump
+ * @gen_sync_to_engine_map: generate a hash map from sync obj id to its engine
+ */
+struct hl_state_dump_specs_funcs {
+	int (*gen_sync_to_engine_map)(struct hl_device *hdev,
+				struct hl_sync_to_engine_map *map);
+};
+
+/**
+ * struct hl_state_dump_specs - defines ASIC known hw objects names
+ * @so_id_to_str_tb: sync objects names index table
+ * @monitor_id_to_str_tb: monitors names index table
+ * @funcs: virtual functions used for state dump
+ * @sync_namager_names: readable names for sync manager if available (ex: N_E)
+ * @props: pointer to a per asic const props array required for state dump
+ */
+struct hl_state_dump_specs {
+	DECLARE_HASHTABLE(so_id_to_str_tb, OBJ_NAMES_HASH_TABLE_BITS);
+	DECLARE_HASHTABLE(monitor_id_to_str_tb, OBJ_NAMES_HASH_TABLE_BITS);
+	struct hl_state_dump_specs_funcs	funcs;
+	const char * const			*sync_namager_names;
+	s64					*props;
+};
+
 
 /*
  * DEVICES
@@ -2151,6 +2265,7 @@ struct hl_mmu_funcs {
  * @mmu_func: device-related MMU functions.
  * @fw_loader: FW loader manager.
  * @pci_mem_region: array of memory regions in the PCI
+ * @state_dump_specs: constants and dictionaries needed to dump system state.
  * @dram_used_mem: current DRAM memory consumption.
  * @timeout_jiffies: device CS timeout value.
  * @max_power: the max power of the device, as configured by the sysadmin. This
@@ -2295,6 +2410,8 @@ struct hl_device {
 
 	struct pci_mem_region		pci_mem_region[PCI_REGION_NUMBER];
 
+	struct hl_state_dump_specs	state_dump_specs;
+
 	atomic64_t			dram_used_mem;
 	u64				timeout_jiffies;
 	u64				max_power;
@@ -2676,6 +2793,14 @@ void hl_release_pending_user_interrupts(struct hl_device *hdev);
 int hl_cs_signal_sob_wraparound_handler(struct hl_device *hdev, u32 q_idx,
 			struct hl_hw_sob **hw_sob, u32 count);
 
+int hl_state_dump(struct hl_device *hdev);
+const char *hl_state_dump_get_sync_name(struct hl_device *hdev, u32 sync_id);
+void hl_state_dump_free_sync_to_engine_map(struct hl_sync_to_engine_map *map);
+__printf(4, 5) int hl_snprintf_resize(char **buf, size_t *size, size_t *offset,
+					const char *format, ...);
+char *hl_format_as_binary(char *buf, size_t buf_len, u32 n);
+const char *hl_sync_engine_to_string(enum hl_sync_engine_type engine_type);
+
 #ifdef CONFIG_DEBUG_FS
 
 void hl_debugfs_init(void);
@@ -2695,6 +2820,8 @@ void hl_debugfs_remove_userptr(struct hl_device *hdev,
 				struct hl_userptr *userptr);
 void hl_debugfs_add_ctx_mem_hash(struct hl_device *hdev, struct hl_ctx *ctx);
 void hl_debugfs_remove_ctx_mem_hash(struct hl_device *hdev, struct hl_ctx *ctx);
+void hl_debugfs_set_state_dump(struct hl_device *hdev, char *data,
+					unsigned long length);
 
 #else
 
@@ -2768,6 +2895,11 @@ static inline void hl_debugfs_remove_ctx_mem_hash(struct hl_device *hdev,
 {
 }
 
+void hl_debugfs_set_state_dump(struct hl_device *hdev, char *data,
+					unsigned long length)
+{
+}
+
 #endif
 
 /* IOCTLs */

drivers/misc/habanalabs/gaudi/gaudi.c

@@ -348,6 +348,8 @@ static enum hl_queue_type gaudi_queue_type[GAUDI_QUEUE_ID_SIZE] = {
 	QUEUE_TYPE_INT, /* GAUDI_QUEUE_ID_NIC_9_3 */
 };
 
+static s64 gaudi_state_dump_specs_props[SP_MAX] = {0};
+
 struct ecc_info_extract_params {
 	u64 block_address;
 	u32 num_memories;
@@ -8977,6 +8979,25 @@ static int gaudi_map_pll_idx_to_fw_idx(u32 pll_idx)
 	}
 }
 
+static int gaudi_gen_sync_to_engine_map(struct hl_device *hdev,
+				struct hl_sync_to_engine_map *map)
+{
+	/* Not implemented */
+	return 0;
+}
+
+
+static struct hl_state_dump_specs_funcs gaudi_state_dump_funcs = {
+	.gen_sync_to_engine_map = gaudi_gen_sync_to_engine_map,
+};
+
+static void gaudi_state_dump_init(struct hl_device *hdev)
+{
+	/* Not implemented */
+	hdev->state_dump_specs.props = gaudi_state_dump_specs_props;
+	hdev->state_dump_specs.funcs = gaudi_state_dump_funcs;
+}
+
 static const struct hl_asic_funcs gaudi_funcs = {
 	.early_init = gaudi_early_init,
 	.early_fini = gaudi_early_fini,
@@ -9062,7 +9083,8 @@ static const struct hl_asic_funcs gaudi_funcs = {
 	.enable_events_from_fw = gaudi_enable_events_from_fw,
 	.map_pll_idx_to_fw_idx = gaudi_map_pll_idx_to_fw_idx,
 	.init_firmware_loader = gaudi_init_firmware_loader,
-	.init_cpu_scrambler_dram = gaudi_init_scrambler_hbm
+	.init_cpu_scrambler_dram = gaudi_init_scrambler_hbm,
+	.state_dump_init = gaudi_state_dump_init
 };
 
 /**

drivers/misc/habanalabs/goya/goya.c

@@ -350,6 +350,8 @@ static u32 goya_all_events[] = {
 	GOYA_ASYNC_EVENT_ID_FIX_THERMAL_ENV_E
 };
 
+static s64 goya_state_dump_specs_props[SP_MAX] = {0};
+
 static int goya_mmu_clear_pgt_range(struct hl_device *hdev);
 static int goya_mmu_set_dram_default_page(struct hl_device *hdev);
 static int goya_mmu_add_mappings_for_device_cpu(struct hl_device *hdev);
@@ -5524,6 +5526,25 @@ static int goya_map_pll_idx_to_fw_idx(u32 pll_idx)
 	}
 }
 
+static int goya_gen_sync_to_engine_map(struct hl_device *hdev,
+				struct hl_sync_to_engine_map *map)
+{
+	/* Not implemented */
+	return 0;
+}
+
+
+static struct hl_state_dump_specs_funcs goya_state_dump_funcs = {
+	.gen_sync_to_engine_map = goya_gen_sync_to_engine_map,
+};
+
+static void goya_state_dump_init(struct hl_device *hdev)
+{
+	/* Not implemented */
+	hdev->state_dump_specs.props = goya_state_dump_specs_props;
+	hdev->state_dump_specs.funcs = goya_state_dump_funcs;
+}
+
 static const struct hl_asic_funcs goya_funcs = {
 	.early_init = goya_early_init,
 	.early_fini = goya_early_fini,
@@ -5609,7 +5630,8 @@ static const struct hl_asic_funcs goya_funcs = {
 	.enable_events_from_fw = goya_enable_events_from_fw,
 	.map_pll_idx_to_fw_idx = goya_map_pll_idx_to_fw_idx,
 	.init_firmware_loader = goya_init_firmware_loader,
-	.init_cpu_scrambler_dram = goya_cpu_init_scrambler_dram
+	.init_cpu_scrambler_dram = goya_cpu_init_scrambler_dram,
+	.state_dump_init = goya_state_dump_init,
 };
 
 /*