AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

libhsakmt: update thunk api for exception handling

Browse Source 
The debugger and debug agent no longer use the Thunk API.
Remove all deprecated functions and keep commented
references for future KFD tests.

Update and the keep the version checks for future use
and hsaKmtRuntimeEnable/Disable.

Signed-off-by: Jonathan Kim <jonathan.kim@amd.com>
Signed-off-by: Laurent Morichetti <laurent.morichetti@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
Change-Id: Ia2f10d82f5ac36d0bd1bda233810f26e8a154d55
This commit is contained in:
Jonathan Kim
2021-05-25 11:14:26 -04:00
committed by Jonathan Kim
parent 96c7a5c9dc
commit 31ac82617c
5 changed files with 324 additions and 974 deletions
Download Patch File Download Diff File Expand all files Collapse all files
include/linux/kfd_ioctl.h
+295 -59
View File
@@ -23,7 +23,7 @@
#ifndef KFD_IOCTL_H_INCLUDED
#define KFD_IOCTL_H_INCLUDED
#include <linux/types.h>
#include <drm/drm.h>
#include <linux/ioctl.h>
/*
@@ -44,15 +44,29 @@
* 1.1 - Allow attaching to processes that have not opened /dev/kfd yet
* 1.2 - Allow flag option to clear queue status on queue suspend
* 1.3 - Fix race condition between clear on suspend and trap event handling
* 1.3 - Fix race condition between clear on suspend and trap event handling
* 1.4 - Fix bad kfifo free
* 1.5 - Fix ABA issue between queue snapshot and suspend
* 2.0 - Return number of queues suspended/resumed and mask invalid/error
* array slots
* array slots
* 2.1 - Add Set Address Watch, and Clear Address Watch support.
* 3.0 - Overhaul set wave launch override API
* 3.1 - Add support for GFX10
* 3.2 - Add support for GFX10.3
* 3.3 - Add precise memory operations enable
* 4.0 - Remove gpu_id from api
* 5.0 - Report exception codes to the debugger
* 6.0 - Pass event file descriptor from userspace.
* 6.1 - Add KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO.
* 6.2 - Add KFD_IOC_DBG_TRAP_DEVICE_SNAPSHOT.
* 7.0 - Redefine exception codes
* 7.1 - Add KFD_IOC_DBG_TRAP_RUNTIME_ENABLE
* 7.2 - Add KFD_IOC_DBG_TRAP_SEND_RUNTIME_EVENT
* 8.0 - Expand runtime information given to the debugger
* 8.1 - Allow the debugger to set the exception mask
* 9.0 - Handle multiple exceptions from single trap signal
*/
#define KFD_IOCTL_DBG_MAJOR_VERSION 2
#define KFD_IOCTL_DBG_MINOR_VERSION 1
#define KFD_IOCTL_DBG_MAJOR_VERSION 9
#define KFD_IOCTL_DBG_MINOR_VERSION 0
struct kfd_ioctl_get_version_args {
__u32 major_version; /* from KFD */
@@ -117,6 +131,7 @@ struct kfd_ioctl_get_queue_wave_state_args {
};
struct kfd_queue_snapshot_entry {
__u64 exception_status;
__u64 ring_base_address;
__u64 write_pointer_address;
__u64 read_pointer_address;
@@ -125,8 +140,19 @@ struct kfd_queue_snapshot_entry {
__u32 gpu_id;
__u32 ring_size;
__u32 queue_type;
__u32 queue_status;
__u32 reserved[19];
__u32 reserved[18];
};
struct kfd_dbg_device_info_entry {
__u64 exception_status;
__u64 lds_base;
__u64 lds_limit;
__u64 scratch_base;
__u64 scratch_limit;
__u64 gpuvm_base;
__u64 gpuvm_limit;
__u32 gpu_id;
__u32 pad;
};
/* For kfd_ioctl_set_memory_policy_args.default_policy and alternate_policy */
@@ -224,114 +250,323 @@ struct kfd_ioctl_dbg_wave_control_args {
__u32 gpu_id; /* to KFD */
__u32 buf_size_in_bytes; /*including gpu_id and buf_size */
};
/* mapping event types to API spec */
#define KFD_DBG_EV_STATUS_TRAP 1
#define KFD_DBG_EV_STATUS_VMFAULT 2
#define KFD_DBG_EV_STATUS_SUSPENDED 4
#define KFD_DBG_EV_STATUS_NEW_QUEUE 8
#define KFD_DBG_EV_STATUS_HOST_TRAP_TIMEDOUT 16
#define KFD_DBG_EV_FLAG_CLEAR_STATUS 1
/* queue states for suspend/resume */
#define KFD_DBG_QUEUE_ERROR_BIT 30
#define KFD_DBG_QUEUE_INVALID_BIT 31
#define KFD_DBG_QUEUE_ERROR_MASK (1 << KFD_DBG_QUEUE_ERROR_BIT)
#define KFD_DBG_QUEUE_INVALID_MASK (1 << KFD_DBG_QUEUE_INVALID_BIT)
#define KFD_INVALID_GPUID 0xffffffff
#define KFD_INVALID_QUEUEID 0xffffffff
enum kfd_dbg_trap_override_mode {
KFD_DBG_TRAP_OVERRIDE_OR = 0,
KFD_DBG_TRAP_OVERRIDE_REPLACE = 1
};
enum kfd_dbg_trap_mask {
KFD_DBG_TRAP_MASK_FP_INVALID = 1,
KFD_DBG_TRAP_MASK_FP_INPUT_DENORMAL = 2,
KFD_DBG_TRAP_MASK_FP_DIVIDE_BY_ZERO = 4,
KFD_DBG_TRAP_MASK_FP_OVERFLOW = 8,
KFD_DBG_TRAP_MASK_FP_UNDERFLOW = 16,
KFD_DBG_TRAP_MASK_FP_INEXACT = 32,
KFD_DBG_TRAP_MASK_INT_DIVIDE_BY_ZERO = 64,
KFD_DBG_TRAP_MASK_DBG_ADDRESS_WATCH = 128,
KFD_DBG_TRAP_MASK_DBG_MEMORY_VIOLATION = 256
};
enum kfd_dbg_trap_exception_code {
EC_NONE = 0,
/* per queue */
EC_QUEUE_WAVE_ABORT = 1,
EC_QUEUE_WAVE_TRAP = 2,
EC_QUEUE_WAVE_MATH_ERROR = 3,
EC_QUEUE_WAVE_ILLEGAL_INSTRUCTION = 4,
EC_QUEUE_WAVE_MEMORY_VIOLATION = 5,
EC_QUEUE_WAVE_APERTURE_VIOLATION = 6,
EC_QUEUE_PACKET_DISPATCH_DIM_INVALID = 16,
EC_QUEUE_PACKET_DISPATCH_GROUP_SEGMENT_SIZE_INVALID = 17,
EC_QUEUE_PACKET_DISPATCH_CODE_INVALID = 18,
EC_QUEUE_PACKET_RESERVED = 19,
EC_QUEUE_PACKET_UNSUPPORTED = 20,
EC_QUEUE_PACKET_DISPATCH_WORK_GROUP_SIZE_INVALID = 21,
EC_QUEUE_PACKET_DISPATCH_REGISTER_INVALID = 22,
EC_QUEUE_PACKET_VENDOR_UNSUPPORTED = 23,
EC_QUEUE_PREEMPTION_ERROR = 30,
EC_QUEUE_NEW = 31,
/* per device */
EC_DEVICE_QUEUE_DELETE = 32,
EC_DEVICE_MEMORY_VIOLATION = 33,
EC_DEVICE_RAS_ERROR = 34,
EC_DEVICE_FATAL_HALT = 35,
EC_DEVICE_NEW = 36,
/* per process */
EC_PROCESS_RUNTIME = 48,
EC_PROCESS_DEVICE_REMOVE = 49,
EC_MAX
};
/* Mask generated by ecode defined in enum above. */
#define KFD_EC_MASK(ecode) (1ULL << (ecode - 1))
/* Masks for exception code type checks below. */
#define KFD_EC_MASK_QUEUE (KFD_EC_MASK(EC_QUEUE_WAVE_ABORT) | \
KFD_EC_MASK(EC_QUEUE_WAVE_TRAP) | \
KFD_EC_MASK(EC_QUEUE_WAVE_MATH_ERROR) | \
KFD_EC_MASK(EC_QUEUE_WAVE_ILLEGAL_INSTRUCTION) | \
KFD_EC_MASK(EC_QUEUE_WAVE_MEMORY_VIOLATION) | \
KFD_EC_MASK(EC_QUEUE_WAVE_APERTURE_VIOLATION) | \
KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_DIM_INVALID) | \
KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_GROUP_SEGMENT_SIZE_INVALID) | \
KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_CODE_INVALID) | \
KFD_EC_MASK(EC_QUEUE_PACKET_UNSUPPORTED) | \
KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_WORK_GROUP_SIZE_INVALID) | \
KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_REGISTER_INVALID) | \
KFD_EC_MASK(EC_QUEUE_PACKET_VENDOR_UNSUPPORTED) | \
KFD_EC_MASK(EC_QUEUE_PREEMPTION_ERROR) | \
KFD_EC_MASK(EC_QUEUE_NEW))
#define KFD_EC_MASK_DEVICE (KFD_EC_MASK(EC_DEVICE_QUEUE_DELETE) | \
KFD_EC_MASK(EC_DEVICE_RAS_ERROR) | \
KFD_EC_MASK(EC_DEVICE_FATAL_HALT) | \
KFD_EC_MASK(EC_DEVICE_MEMORY_VIOLATION) | \
KFD_EC_MASK(EC_DEVICE_NEW))
#define KFD_EC_MASK_PROCESS (KFD_EC_MASK(EC_PROCESS_RUNTIME) | \
KFD_EC_MASK(EC_PROCESS_DEVICE_REMOVE))
/* Checks for exception code types for KFD search. */
#define KFD_DBG_EC_TYPE_IS_QUEUE(ecode) \
(!!(KFD_EC_MASK(ecode) & KFD_EC_MASK_QUEUE))
#define KFD_DBG_EC_TYPE_IS_DEVICE(ecode) \
(!!(KFD_EC_MASK(ecode) & KFD_EC_MASK_DEVICE))
#define KFD_DBG_EC_TYPE_IS_PROCESS(ecode) \
(!!(KFD_EC_MASK(ecode) & KFD_EC_MASK_PROCESS))
enum kfd_dbg_runtime_state {
DEBUG_RUNTIME_STATE_DISABLED = 0,
DEBUG_RUNTIME_STATE_ENABLED = 1,
DEBUG_RUNTIME_STATE_ENABLED_BUSY = 2,
DEBUG_RUNTIME_STATE_ENABLED_ERROR = 3
};
struct kfd_runtime_info {
__u64 r_debug;
__u32 runtime_state;
__u32 ttmp_setup;
};
/* KFD_IOC_DBG_TRAP_ENABLE:
* ptr: unused
* exception_mask: exceptions to be reported to the debugger
* ptr: runtime info buffer to copy to (IN)
* data1: 0=disable, 1=enable
* data2: queue ID (for future use)
* data3: return value for fd
* data2: return value for fd
* data3: runtime info size
* data4: unused
*
*/
#define KFD_IOC_DBG_TRAP_ENABLE 0
/* KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE:
* exception_mask: unused
* ptr: unused
* data1: override mode: 0=OR, 1=REPLACE
* data2: mask
* data3: unused
* data1: override mode (see enum kfd_dbg_trap_override_mode)
* data2: [in/out] trap mask (see enum kfd_dbg_trap_mask)
* data3: [in] requested mask, [out] supported mask
* data4: unused
*
* May fail with -EPERM if the requested mode is not supported.
*
* May fail with -EACCES if requested trap mask bits are not supported.
* In that case the supported trap mask bits are returned in data3.
*
* If successful, output parameters return the previous trap mask
* value and the hardware-dependent mask of supported trap mask bits.
*/
#define KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE 1
/* KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE:
* exception_mask: unused
* ptr: unused
* data1: 0=normal, 1=halt, 2=kill, 3=singlestep, 4=disable
* data2: unused
* data3: unused
* data4: unused
*/
#define KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE 2
/* KFD_IOC_DBG_TRAP_NODE_SUSPEND:
* ptr: pointer to an array of Queues IDs
* data1: flags
* data2: number of queues
* data3: grace period
* exception_mask: exceptions to clear on suspend
* ptr: pointer to an array of Queues IDs (IN/OUT)
* data1: number of queues (IN)
* data2: grace period (IN)
* data3: unused
* data4: unused
*
* Returns the number of queues suspended from array of Queue IDs (ptr).
* Requested queues that fail to suspend are masked in the array:
*
* KFD_DBG_QUEUE_INVALID_MASK - requested queue does not exist or cannot be
* suspended (new or being destroyed).
*
* KFD_DBG_QUEUE_ERROR_MASK - bad internal operation occurred on requested
* queue.
*
* NOTE! All queue destroy requests will be blocked on a suspended queue.
* Queue resume will unblock.
*
* Grace period (data2) is time allowed for waves to complete before CWSR.
* 0 can be entered for immediate preemption.
*/
#define KFD_IOC_DBG_TRAP_NODE_SUSPEND 3
/* KFD_IOC_DBG_TRAP_NODE_RESUME:
* ptr: pointer to an array of Queues IDs
* data1: flags
* data2: number of queues
* exception_mask: unused
* ptr: pointer to an array of Queues IDs (IN/OUT)
* data1: number of queues (IN)
* data2: unused
* data3: unused
* data4: unused
*
* Returns the number of queues resumed from array of Queue IDs (ptr).
* Requested queues that fail to resume are masked in the array:
*
* KFD_DBG_QUEUE_INVALID_MASK - requested queue does not exist.
*
* KFD_DBG_QUEUE_ERROR_MASK - bad internal operation occurred on requested
* queue.
*/
#define KFD_IOC_DBG_TRAP_NODE_RESUME 4
/* KFD_IOC_DBG_TRAP_QUERY_DEBUG_EVENT:
* exception_mask: exception to clear (IN) on query and report (OUT)
* ptr: unused
* data1: queue id (IN/OUT)
* data2: flags (IN)
* data3: suspend[2:2], event type [1:0] (OUT)
* data1: source id (OUT)
* data2: unused
* data3: unused
* data4: unused
*
* Source ID (data 1 - OUT) returns the queue_id for a queue event, the gpu_id
* for a device event or 0 for a process event.
*
* Returns EAGAIN if no event is found and 0 otherwise.
*/
#define KFD_IOC_DBG_TRAP_QUERY_DEBUG_EVENT 5
/* KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT:
* exception_mask: exception to clear on snapshot
* ptr: user buffer (IN)
* data1: flags (IN)
* data2: number of queue snapshot entries (IN/OUT)
* data1: number of queue snapshots (IN/OUT) - 0 for IN ignores buffer writes
* data2: unused
* data3: unused
* data4: unused
*/
#define KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT 6
/* KFD_IOC_DBG_TRAP_GET_VERSION:
* exception_mask: unused
* ptr: unsused
* data1: major version (OUT)
* data2: minor version (OUT)
* data3: unused
* data4: unused
*/
#define KFD_IOC_DBG_TRAP_GET_VERSION 7
/* KFD_IOC_DBG_TRAP_CLEAR_ADDRESS_WATCH:
* exception_mask: unused
* ptr: unused
* data1: watch ID
* data2: unused
* data3: unused
* data4: unused
*/
#define KFD_IOC_DBG_TRAP_CLEAR_ADDRESS_WATCH 8
/* KFD_IOC_DBG_TRAP_SET_ADDRESS_WATCH:
* exception_mask: unused
* ptr: Watch address
* data1: Watch ID (OUT)
* data2: watch_mode: 0=read, 1=nonread, 2=atomic, 3=all
* data3: watch address mask
* data4: unused
*/
#define KFD_IOC_DBG_TRAP_SET_ADDRESS_WATCH 9
/* KFD_IOC_DBG_SEND_HOST_TRAP:
/* KFD_IOC_DBG_TRAP_SET_PRECISE_MEM_OPS
* exception_mask: unused
* ptr: unused
* data1: 0=disable, 1=enable (IN)
* data2: unused
* data3: unused
* data4: unused
*/
#define KFD_IOC_DBG_TRAP_SET_PRECISE_MEM_OPS 10
/* KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO
* exception_mask: unused
* ptr: exception info pointer to copy to
* data1: source_id
* data2: exception_code
* data3: clear_exception (1 == true, 0 == false)
* data4: exception info data size
*
* NOTE: If data2 == EC_PROCESS_RUNTIME, the saved runtime info will be copied
* to the exception info pointer.
*/
#define KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO 11
/* KFD_IOC_DBG_TRAP_DEVICE_SNAPSHOT
* exception_mask: exception to clear on snapshot
* ptr: user buffer for 'struct kfd_dbg_device_info_entry' entries (IN)
* data1: number of devices in snapshot (IN/OUT)
* data2: unused
* data3: unused
* data4: unused
*/
#define KFD_IOC_DBG_TRAP_DEVICE_SNAPSHOT 12
/* KFD_IOC_DBG_TRAP_RUNTIME_ENABLE
* exception_mask: unused
* ptr: r_debug info to save
* data1: enable (0=disable, 1=enable)
* data2: enable ttmp save (0=disable, 1=enable)
* data3: unused
* data4: unused
*
* FIXME: This option is temporary. Future upstream will use a separate IOCTL.
*/
#define KFD_IOC_DBG_TRAP_RUNTIME_ENABLE 13
/* KFD_IOC_DBG_TRAP_SEND_RUNTIME_EVENT
* exception_mask: exception to send
* ptr: unused
* data1: destination device id
* data2: destination queue id
* data3: usused
* data4: unused
*/
#define KFD_IOC_DBG_TRAP_SEND_RUNTIME_EVENT 14
/* KFD_IOC_DBG_TRAP_SET_EXCEPTION_ENABLED
* exception_mask: exception to set
* ptr: unused
* data1: unused
* data2: unused
* data3: unused
* data3: usused
* data4: unused
*/
#define KFD_IOC_DBG_TRAP_SEND_HOST_TRAP 10
#define KFD_IOC_DBG_TRAP_SET_EXCEPTIONS_ENABLED 15
struct kfd_ioctl_dbg_trap_args {
__u64 ptr; /* to KFD -- used for pointer arguments: queue arrays */
__u64 exception_mask; /* to KFD */
__u64 ptr; /* to KFD */
__u32 pid; /* to KFD */
__u32 gpu_id; /* to KFD */
__u32 op; /* to KFD */
__u32 data1; /* to KFD */
__u32 data2; /* to KFD */
__u32 data3; /* to KFD */
__u32 data4; /* to KFD */
};
/* Matching HSA_EVENTTYPE */
@@ -569,7 +804,7 @@ struct kfd_ioctl_alloc_queue_gws_args {
__u32 queue_id; /* to KFD */
__u32 num_gws; /* to KFD */
__u32 first_gws; /* from KFD */
__u32 pad; /* to KFD */
__u32 pad;
};
struct kfd_ioctl_get_dmabuf_info_args {
@@ -737,7 +972,6 @@ struct kfd_ioctl_cross_memory_copy_args {
__u64 bytes_copied;
};
/* Guarantee host access to memory */
#define KFD_IOCTL_SVM_FLAG_HOST_ACCESS 0x00000001
/* Fine grained coherency between all devices with access */
@@ -749,7 +983,7 @@ struct kfd_ioctl_cross_memory_copy_args {
/* Allow execution on GPU */
#define KFD_IOCTL_SVM_FLAG_GPU_EXEC 0x00000010
/* GPUs mostly read, may allow similar optimizations as RO, but writes fault */
#define KFD_IOCTL_SVM_FLAG_GPU_READ_MOSTLY 0x00000020
#define KFD_IOCTL_SVM_FLAG_GPU_READ_MOSTLY 0x00000020
/**
* kfd_ioctl_svm_op - SVM ioctl operations
@@ -762,6 +996,17 @@ enum kfd_ioctl_svm_op {
KFD_IOCTL_SVM_OP_GET_ATTR
};
/** kfd_ioctl_svm_location - Enum for preferred and prefetch locations
*
* GPU IDs are used to specify GPUs as preferred and prefetch locations.
* Below definitions are used for system memory or for leaving the preferred
* location unspecified.
*/
enum kfd_ioctl_svm_location {
KFD_IOCTL_SVM_LOCATION_SYSMEM = 0,
KFD_IOCTL_SVM_LOCATION_UNDEFINED = 0xffffffff
};
/**
* kfd_ioctl_svm_attr_type - SVM attribute types
*
@@ -791,17 +1036,6 @@ enum kfd_ioctl_svm_attr_type {
KFD_IOCTL_SVM_ATTR_GRANULARITY
};
/** kfd_ioctl_svm_location - Enum for preferred and prefetch locations
*
* GPU IDs are used to specify GPUs as preferred and prefetch locations.
* Below definitions are used for system memory or for leaving the preferred
* location unspecified.
*/
enum kfd_ioctl_svm_location {
KFD_IOCTL_SVM_LOCATION_SYSMEM = 0,
KFD_IOCTL_SVM_LOCATION_UNDEFINED = 0xffffffff
};
/**
* kfd_ioctl_svm_attribute - Attributes as pairs of type and value
*
@@ -882,16 +1116,16 @@ struct kfd_ioctl_svm_args {
* functionality.
*
* Enabling XNACK mode requires shader programs to be compiled
* differently. Furthermore, not all GPUs support changing the mode
* differently. Furthermore, not all GPUs support changing the mode
* per-process. Therefore changing the mode is only allowed while no
* user mode queues exist in the process. This ensure that no shader
* code is running that may be compiled for the wrong mode. And GPUs
* that cannot change to the requested mode will be disabled by
* failing subsequent requests to create user mode queues.
*
* This ioctl returns the status of the requested xnack mode.
* that cannot change to the requested mode will prevent the XNACK
* mode from occurring. All GPUs used by the process must be in the
* same XNACK mode.
*
* GFXv8 or older GPUs do not support 48 bit virtual addresses or SVM.
* Therefore those GPUs are not considered for the XNACK mode switch.
*
* Return: 0 on success, -errno on failure
*/
@@ -956,7 +1190,7 @@ struct kfd_ioctl_set_xnack_mode_args {
#define AMDKFD_IOC_SET_SCRATCH_BACKING_VA \
AMDKFD_IOWR(0x11, struct kfd_ioctl_set_scratch_backing_va_args)
#define AMDKFD_IOC_GET_TILE_CONFIG \
#define AMDKFD_IOC_GET_TILE_CONFIG \
AMDKFD_IOWR(0x12, struct kfd_ioctl_get_tile_config_args)
#define AMDKFD_IOC_SET_TRAP_HANDLER \
@@ -996,6 +1230,9 @@ struct kfd_ioctl_set_xnack_mode_args {
#define AMDKFD_IOC_ALLOC_QUEUE_GWS \
AMDKFD_IOWR(0x1E, struct kfd_ioctl_alloc_queue_gws_args)
#define AMDKFD_IOC_SMI_EVENTS \
AMDKFD_IOWR(0x1F, struct kfd_ioctl_smi_events_args)
#define AMDKFD_IOC_SVM AMDKFD_IOWR(0x20, struct kfd_ioctl_svm_args)
#define AMDKFD_IOC_SET_XNACK_MODE \
@@ -1021,7 +1258,6 @@ struct kfd_ioctl_set_xnack_mode_args {
AMDKFD_IOWR(0x84, struct kfd_ioctl_spm_args)
#define AMDKFD_COMMAND_START_2 0x80
#define AMDKFD_COMMAND_END_2 0x85