Enhance device state machine to better support suspend/resume
v2 cover Enhance device state machine to better support suspend/resume
Recently we were testing suspend/resume functionality with AMD GPUs, we have encountered several resource tracking related bugs, such as double buffer free, use after free and unbalanced irq reference count.
We have tried to solve these issues case by case, but found that may not be the right way. Especially about the unbalanced irq reference count, there will be new issues appear once we fixed the current known issues. After analyzing related source code, we found that there may be some fundamental implementation flaws behind these resource tracking issues.
The amdgpu driver has two major state machines to driver the device management flow, one is for ip blocks, the other is for ras blocks. The hook points defined in struct amd_ip_funcs for device setup/teardown are symmetric, but the implementation is asymmetric, sometime even ambiguous. The most obvious two issues we noticed are:
1) amdgpu_irq_get() are called from .late_init() but amdgpu_irq_put() are called from .hw_fini() instead of .early_fini().
2) the way to reset ip_bloc.status.valid/sw/hw/late_initialized doesn't match the way to set those flags.
When taking device suspend/resume into account, in addition to device probe/remove, things get much more complex. Some issues arise because many suspend/resume implementations directly reuse .hw_init/.hw_fini/ .late_init hook points.
So we try to fix those issues by two enhancements/refinements to current device management state machines.
The first change is to make the ip block state machine and associated status flags work in stack-like way as below:
Callbacks State after successfully execute callback
AMDGPU_IP_STATE_INVALID
.early_init() AMDGPU_IP_STATE_EARLY .sw_init() AMDGPU_IP_STATE_SW .hw_init() AMDGPU_IP_STATE_HW .late_init() AMDGPU_IP_STATE_LATE .early_fini() AMDGPU_IP_STATE_HW .hw_fini() AMDGPU_IP_STATE_SW .sw_fini() AMDGPU_IP_STATE_EARLY .late_fini() AMDGPU_IP_STATE_INVALID
Also do the same thing for ras block state machine, though it's much more simpler.
The second change is fine tune the overall device management work flow as below:
1. amdgpu_driver_load_kms()
amdgpu_device_init() amdgpu_device_ip_early_init() ip_blocks[i].early_init() ip_blocks[i].status.valid = true amdgpu_device_ip_init() amdgpu_ras_init() ip_blocks[i].sw_init() ip_blocks[i].status.sw = true ip_blocks[i].hw_init() ip_blocks[i].status.hw = true amdgpu_device_ip_late_init() ip_blocks[i].late_init() ip_blocks[i].status.late_initialized = true amdgpu_ras_late_init() ras_blocks[i].ras_late_init() amdgpu_ras_feature_enable_on_boot()
2. amdgpu_pmops_suspend()/amdgpu_pmops_freeze()/amdgpu_pmops_poweroff()
amdgpu_device_suspend() amdgpu_ras_early_fini() ras_blocks[i].ras_early_fini() amdgpu_ras_feature_disable() amdgpu_ras_suspend() amdgpu_ras_disable_all_features() +++ ip_blocks[i].early_fini() +++ ip_blocks[i].status.late_initialized = false ip_blocks[i].suspend()
3. amdgpu_pmops_resume()/amdgpu_pmops_thaw()/amdgpu_pmops_restore()
amdgpu_device_resume() amdgpu_device_ip_resume() ip_blocks[i].resume() amdgpu_device_ip_late_init() ip_blocks[i].late_init() ip_blocks[i].status.late_initialized = true amdgpu_ras_late_init() ras_blocks[i].ras_late_init() amdgpu_ras_feature_enable_on_boot() amdgpu_ras_resume() amdgpu_ras_enable_all_features()
4. amdgpu_driver_unload_kms()
amdgpu_device_fini_hw() amdgpu_ras_early_fini() ras_blocks[i].ras_early_fini() +++ ip_blocks[i].early_fini() +++ ip_blocks[i].status.late_initialized = false ip_blocks[i].hw_fini() ip_blocks[i].status.hw = false
5. amdgpu_driver_release_kms()
amdgpu_device_fini_sw() amdgpu_device_ip_fini() ip_blocks[i].sw_fini() ip_blocks[i].status.sw = false --- ip_blocks[i].status.valid = false +++ amdgpu_ras_fini() ip_blocks[i].late_fini()
+++ ip_blocks[i].status.valid = false
--- ip_blocks[i].status.late_initialized = false
--- amdgpu_ras_fini()
The main changes include:
1) invoke ip_blocks[i].early_fini in amdgpu_pmops_suspend(). Currently there's only one ip block which provides early_fini callback. We have add a check of in_s3 to keep current behavior in function amdgpu_dm_early_fini(). So there should be no functional changes.
2) set ip_blocks[i].status.late_initialized to false after calling callback early_fini. We have auditted all usages of the late_initialized flag and no functional changes found.
3) only set ip_blocks[i].status.valid = false after calling the late_fini callback.
4) call amdgpu_ras_fini() before invoking ip_blocks[i].late_fini.
Then we try to refine each subsystem, such as nbio, asic etc, to follow the new design. Currently we have only taken the nbio and asic as examples to show the proposed changes. Once we have confirmed that's the right way to go, we will handle the lefting subsystems.
This is in early stage and requesting for comments, any comments and suggestions are welcomed!
v2:
- remove patch 1 in v1, it already got merged
- convert status bool flags for ip block into enum
- introduce iterators to walk ip blocks
- refine the way to define status markers
- split amdgpu_dm related change into a dedicated patch
- add patch 13 to walk ip blocks in reverse order when shutdown
Jiang Liu (15):
drm/amdgpu: add helper functions to track status for ras manager drm/amdgpu: add a flag to track ras debugfs creation status drm/amdgpu: free all resources on error recovery path of
amdgpu_ras_init()
drm/amdgpu: introduce a flag to track refcount held for features drm/amdgpu: enhance amdgpu_ras_block_late_fini() drm/amdgpu: enhance amdgpu_ras_pre_fini() to better support SR drm/admgpu: rename amdgpu_ras_pre_fini() to amdgpu_ras_early_fini() drm/amdgpu: make IP block state machine works in stack like way drm/amdgpu_dm: enhance amdgpu_dm_early_fini() for PM ops drm/admgpu: make device state machine work in stack like way drm/amdgpu: convert ip block bool flags into an enum drm/amdgpu: introduce IP block iterators to reduce duplicated code drm/amdgpu: walk IP blocks in reverse order when shutdown drm/amdgpu/nbio: improve the way to manage irq reference count drm/amdgpu/asic: make ip block operations symmetric by .early_fini()
drivers/gpu/drm/amd/amdgpu/aldebaran.c | 46 +- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 109 +++- .../gpu/drm/amd/amdgpu/amdgpu_dev_coredump.c | 3 +- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 504 +++++++++--------- drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 10 +- drivers/gpu/drm/amd/amdgpu/amdgpu_jpeg.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c | 18 +- drivers/gpu/drm/amd/amdgpu/amdgpu_nbio.c | 16 +- drivers/gpu/drm/amd/amdgpu/amdgpu_nbio.h | 1 + drivers/gpu/drm/amd/amdgpu/amdgpu_psp.c | 4 +- drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c | 142 +++-- drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h | 16 +- drivers/gpu/drm/amd/amdgpu/amdgpu_reset.c | 14 +- drivers/gpu/drm/amd/amdgpu/amdgpu_sdma.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_vcn.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_xgmi.c | 2 +- drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c | 2 +- drivers/gpu/drm/amd/amdgpu/jpeg_v4_0_3.c | 2 +- drivers/gpu/drm/amd/amdgpu/mmhub_v1_8.c | 2 +- drivers/gpu/drm/amd/amdgpu/nbio_v7_4.c | 1 + drivers/gpu/drm/amd/amdgpu/nbio_v7_9.c | 1 + drivers/gpu/drm/amd/amdgpu/nv.c | 14 +- drivers/gpu/drm/amd/amdgpu/sienna_cichlid.c | 50 +- drivers/gpu/drm/amd/amdgpu/smu_v13_0_10.c | 51 +- drivers/gpu/drm/amd/amdgpu/soc15.c | 38 +- drivers/gpu/drm/amd/amdgpu/soc21.c | 35 +- drivers/gpu/drm/amd/amdgpu/soc24.c | 22 +- drivers/gpu/drm/amd/amdgpu/vcn_v4_0_3.c | 2 +- .../gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c | 3 + drivers/gpu/drm/amd/pm/swsmu/amdgpu_smu.c | 10 +- 32 files changed, 668 insertions(+), 460 deletions(-)
Signature · 1 line
2.43.5
RFC v2 1/15 drm/amdgpu: add helper functions to track status for ras manager
Add helper functions to track status for ras manager and ip blocks.
--- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 38 +++++++++++++++++++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c | 37 ++++++++++++++++++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h | 10 +++++++ 3 files changed, 85 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +38 -0
@@ -377,12 +377,28 @@ int amdgpu_ip_block_resume(struct amdgpu_ip_block *ip_block);#define AMDGPU_MAX_IP_NUM 16 +enum amdgpu_marker { +// Markers for IRQs, used for both ip blocks and ras blocks. +AMDGPU_MARKER_IRQ0 =32 , +AMDGPU_MARKER_IRQ1 , +AMDGPU_MARKER_IRQ2 , +AMDGPU_MARKER_IRQ3 , +AMDGPU_MARKER_IRQ4 , +AMDGPU_MARKER_IRQ5 , +AMDGPU_MARKER_IRQ6 , +AMDGPU_MARKER_IRQ7 , +AMDGPU_MARKER_IRQ_MAX =63 , +}; + +#define AMDGPU_MARKER_IRQ (idx) (AMDGPU_MARKER_IRQ0 + (idx)) +struct amdgpu_ip_block_status {bool valid;bool sw;bool hw;bool late_initialized;bool hang; +uint64_t markers; };struct amdgpu_ip_block_version {
@@ -410,6 +426,28 @@ amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev,int amdgpu_device_ip_block_add (struct amdgpu_device *adev,const struct amdgpu_ip_block_version *ip_block_version); +static inline void amdgpu_ip_block_set_marker (struct amdgpu_ip_block *ip_block, +enum amdgpu_marker marker) +{ +WARN_ON (marker >63 ); +WARN_ON (ip_block->status.markers & (0x1ull << marker)); + ip_block->status.markers |=0x1ull << (int )marker; +} + +static inline bool amdgpu_ip_block_test_and_clear_marker (struct amdgpu_ip_block *ip_block, +enum amdgpu_marker marker) +{ +bool set = false; +uint64_t value =0x1ull << (int )marker; + +if ((ip_block->status.markers & value) !=0 ) { + ip_block->status.markers &= ~value; + set = true; + } + +return set; +} +/* * BIOS. */
drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +37 -0
@@ -5207,3 +5207,40 @@ bool amdgpu_ras_is_rma(struct amdgpu_device *adev)return con->is_rma; } + +bool amdgpu_ras_test_marker (struct amdgpu_device *adev, +struct ras_common_if *head,int marker) +{ +struct ras_manager *obj =amdgpu_ras_find_obj (adev, head); + +if (obj && obj->markers & (0x1ull << marker)) +return true; + +return false; +} + +void amdgpu_ras_set_marker (struct amdgpu_device *adev, +struct ras_common_if *head,int marker) +{ +struct ras_manager *obj =amdgpu_ras_find_obj (adev, head); + +WARN_ON (marker >63 ); +WARN_ON (obj->markers & (0x1ull << marker)); +if (obj) + obj->markers |=0x1ull << marker; +} + +bool amdgpu_ras_test_and_clear_marker (struct amdgpu_device *adev, +struct ras_common_if *head,int marker) +{ +bool set = false; +uint64_t value =0x1ull << marker; +struct ras_manager *obj =amdgpu_ras_find_obj (adev, head); + +if (obj && (obj->markers & value) !=0 ) { + obj->markers &= ~value; + set = true; + } + +return set; +}
drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h +10 -0
@@ -634,6 +634,8 @@ struct ras_manager {struct ras_common_if head;/* reference count */ int use; +/* Flags for status tracking */ +uint64_t markers;/* ras block link */ struct list_head node;/* the device */
@@ -977,4 +979,12 @@ void amdgpu_ras_event_log_print(struct amdgpu_device *adev, u64 event_id, const char *fmt, ...);bool amdgpu_ras_is_rma (struct amdgpu_device *adev); + +bool amdgpu_ras_test_marker (struct amdgpu_device *adev, +struct ras_common_if *head,int marker); +void amdgpu_ras_set_marker (struct amdgpu_device *adev, +struct ras_common_if *head,int marker); +bool amdgpu_ras_test_and_clear_marker (struct amdgpu_device *adev, +struct ras_common_if *head, +int marker);#endif
Signature · 1 line
2.43.5
RFC v2 2/15 drm/amdgpu: add a flag to track ras debugfs creation status
In reply to Jiang LiuAdd a flag to track ras debugfs creation status, to avoid possible incorrect reference count management for ras block object in function amdgpu_ras_aca_is_supported().
--- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 2 ++ drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c | 9 +++++++-- 2 files changed, 9 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +2 -0
@@ -378,6 +378,8 @@ int amdgpu_ip_block_resume(struct amdgpu_ip_block *ip_block);#define AMDGPU_MAX_IP_NUM 16enum amdgpu_marker { +// Markers for ras blocks. +AMDGPU_MARKER_RAS_DEBUGFS ,// Markers for IRQs, used for both ip blocks and ras blocks. AMDGPU_MARKER_IRQ0 =32 ,AMDGPU_MARKER_IRQ1 ,
drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +7 -2
@@ -1997,7 +1997,8 @@ static void amdgpu_ras_debugfs_create(struct amdgpu_device *adev, {struct ras_manager *obj =amdgpu_ras_find_obj (adev, &head->head); -if (!obj || !dir) +if (!obj || !dir || +amdgpu_ras_test_marker (adev, &head->head,AMDGPU_MARKER_RAS_DEBUGFS ))return ;get_obj (obj);
@@ -2008,6 +2009,8 @@ static void amdgpu_ras_debugfs_create(struct amdgpu_device *adev,debugfs_create_file (obj->fs_data.debugfs_name,S_IWUGO |S_IRUGO , dir, obj, &amdgpu_ras_debugfs_ops); + +amdgpu_ras_set_marker (adev, &head->head,AMDGPU_MARKER_RAS_DEBUGFS ); }static bool amdgpu_ras_aca_is_supported (struct amdgpu_device *adev)
@@ -2136,7 +2139,9 @@ static int amdgpu_ras_fs_fini(struct amdgpu_device *adev)if (IS_ENABLED (CONFIG_DEBUG_FS )) {list_for_each_entry_safe (con_obj, tmp, &con->head, node) { ip_obj =amdgpu_ras_find_obj (adev, &con_obj->head); -if (ip_obj) +if (ip_obj && +amdgpu_ras_test_and_clear_marker (adev, &ip_obj->head, +AMDGPU_MARKER_RAS_DEBUGFS ))put_obj (ip_obj); } }
Signature · 1 line
2.43.5
RFC v2 5/15 drm/amdgpu: enhance amdgpu_ras_block_late_fini()
In reply to Jiang LiuEnhance amdgpu_ras_block_late_fini() to revert what has been done by amdgpu_ras_block_late_init(), and fix a possible resource leakage in function amdgpu_ras_block_late_init().
--- drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c | 16 ++++++++++------ 1 file changed, 10 insertions(+), 6 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +10 -6
@@ -4104,13 +4104,13 @@ int amdgpu_ras_block_late_init(struct amdgpu_device *adev, ras_obj->hw_ops->query_ras_error_status)) { r =amdgpu_ras_sysfs_create (adev, ras_block);if (r) - goto interrupt; + goto cleanup;/* Those are the cached values at init. */ query_info =kzalloc (sizeof (*query_info),GFP_KERNEL );if (!query_info) -return -ENOMEM ; + goto cleanup;memcpy (&query_info->head, ras_block,sizeof (struct ras_common_if ));if (amdgpu_ras_query_error_count (adev, &ce_count, &ue_count, query_info) ==0 ) {
@@ -4123,11 +4123,8 @@ int amdgpu_ras_block_late_init(struct amdgpu_device *adev,return 0 ; -interrupt: -if (ras_obj->ras_cb) -amdgpu_ras_interrupt_remove_handler (adev, ras_block); cleanup: -amdgpu_ras_feature_enable (adev, ras_block,0 ); +amdgpu_ras_block_late_fini (adev, ras_block);return r; }
@@ -4142,9 +4139,16 @@ void amdgpu_ras_block_late_fini(struct amdgpu_device *adev, struct ras_common_if *ras_block) {struct amdgpu_ras_block_object *ras_obj; +if (!ras_block)return ; +amdgpu_ras_feature_enable (adev, ras_block,0 ); + +/* in resume/reset phase, no need to delete ras fs node */ +if (adev->in_suspend ||amdgpu_in_reset (adev)) +return ; +amdgpu_ras_sysfs_remove (adev, ras_block); ras_obj = container_of(ras_block, struct amdgpu_ras_block_object, ras_comm);
Signature · 1 line
2.43.5
RFC v2 6/15 drm/amdgpu: enhance amdgpu_ras_pre_fini() to better support SR
In reply to Jiang LiuEnhance amdgpu_ras_pre_fini() to better support suspend/resume by:
1) fix possible resource leakage. amdgpu_release_ras_context() only kfree(con) but doesn't release resources associated with the con object.
2) call amdgpu_ras_pre_fini() in amdgpu_device_suspend() to undo what has been done by amdgpu_ras_late_init(), because amdgpu_device_resume() will invoke amdgpu_ras_late_init() on resume.
3) move amdgpu_ras_recovery_fini() from amdgpu_ras_pre_fini() to amdgpu_ras_fini()
4) move calling of obj->ras_fini() from amdgpu_ras_fini() to amdgpu_ras_pre_fini().
--- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 6 ++- drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c | 44 +++++++++++++--------- 2 files changed, 31 insertions(+), 19 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +4 -2
@@ -4613,6 +4613,8 @@ int amdgpu_device_init(struct amdgpu_device *adev,return 0 ; release_ras_con: +amdgpu_ras_pre_fini (adev); +amdgpu_ras_fini (adev);if (amdgpu_sriov_vf (adev))amdgpu_virt_release_full_gpu (adev, true);
@@ -4627,8 +4629,6 @@ int amdgpu_device_init(struct amdgpu_device *adev, adev->virt.ops =NULL ; r = -EAGAIN ; } -amdgpu_release_ras_context (adev); - failed:amdgpu_vf_error_trans_all (adev);
@@ -4921,6 +4921,8 @@ int amdgpu_device_suspend(struct drm_device *dev, bool notify_clients)cancel_delayed_work_sync (&adev->delayed_init_work); +/* disable ras feature must before hw fini */ +amdgpu_ras_pre_fini (adev);amdgpu_ras_suspend (adev);amdgpu_device_ip_suspend_phase1 (adev);
drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +27 -17
77 lines not on this page · raw message
Signature · 1 line
2.43.5
RFC v2 3/15 drm/amdgpu: free all resources on error recovery path of amdgpu_ras_init()
In reply to Jiang LiuFree all allocated resources on error recovery path in function amdgpu_ras_init().
--- drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c | 19 ++++++++++++++----- 1 file changed, 14 insertions(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +14 -5
@@ -3864,6 +3864,7 @@ static void amdgpu_ras_init_reserved_vram_size(struct amdgpu_device *adev)int amdgpu_ras_init (struct amdgpu_device *adev) {struct amdgpu_ras *con =amdgpu_ras_get_context (adev); +struct amdgpu_ras_block_list *ras_node, *tmp;int r;if (con)
@@ -3953,20 +3954,20 @@ int amdgpu_ras_init(struct amdgpu_device *adev)* to handle fatal error */ r =amdgpu_nbio_ras_sw_init (adev);if (r) -return r; + goto release_con;if (adev->nbio.ras && adev->nbio.ras->init_ras_controller_interrupt) { r = adev->nbio.ras->init_ras_controller_interrupt (adev);if (r) - goto release_con; + goto free_blocks; }if (adev->nbio.ras && adev->nbio.ras->init_ras_err_event_athub_interrupt) { r = adev->nbio.ras->init_ras_err_event_athub_interrupt (adev);if (r) - goto release_con; + goto free_blocks; }/* Packed socket_id to ras feature mask bits[31:29] */
@@ -3982,7 +3983,7 @@ int amdgpu_ras_init(struct amdgpu_device *adev)if (amdgpu_ras_fs_init (adev)) { r = -EINVAL ; - goto release_con; + goto free_blocks; }if (amdgpu_ras_aca_is_supported (adev)) {
@@ -3991,7 +3992,7 @@ int amdgpu_ras_init(struct amdgpu_device *adev)else r =amdgpu_mca_init (adev);if (r) - goto release_con; + goto clear_ras_fs; } dev_info(adev->dev, "RAS INFO: ras initialized successfully, "
@@ -3999,6 +4000,14 @@ int amdgpu_ras_init(struct amdgpu_device *adev) adev->ras_hw_enabled, adev->ras_enabled);return 0 ; + +clear_ras_fs: +amdgpu_ras_fs_fini (adev); +free_blocks: +list_for_each_entry_safe (ras_node, tmp, &adev->ras_list, node) { +list_del (&ras_node->node); +kfree (ras_node); + } release_con:amdgpu_ras_set_context (adev,NULL );kfree (con);
Signature · 1 line
2.43.5
RFC v2 9/15 drm/amdgpu_dm: enhance amdgpu_dm_early_fini() for PM ops
In reply to Jiang LiuEnhance amdgpu_dm_early_fini() so it can be called in power management operations.
--- drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c | 3 +++ 1 file changed, 3 insertions(+)
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c +3 -0
@@ -2175,6 +2175,9 @@ static int amdgpu_dm_early_fini(struct amdgpu_ip_block *ip_block) {struct amdgpu_device *adev = ip_block->adev; +if (adev->in_s0ix || adev->in_s3 || adev->in_s4 || adev->in_suspend) +return 0 ; +amdgpu_dm_audio_fini (adev);return 0 ;
Signature · 1 line
2.43.5
RFC v2 4/15 drm/amdgpu: introduce a flag to track refcount held for features
In reply to Jiang LiuCurrently we track the refcount on ras block object for features by checking if (obj && amdgpu_ras_is_feature_enabled(adev, head)), which is a little unreliable. So introduce a dedicated flag to track the reference count.
--- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 1 + drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c | 9 +++++++-- 2 files changed, 8 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +1 -0
@@ -380,6 +380,7 @@ int amdgpu_ip_block_resume(struct amdgpu_ip_block *ip_block);enum amdgpu_marker {// Markers for ras blocks. AMDGPU_MARKER_RAS_DEBUGFS , +AMDGPU_MARKER_RAS_FEATURE ,// Markers for IRQs, used for both ip blocks and ras blocks. AMDGPU_MARKER_IRQ0 =32 ,AMDGPU_MARKER_IRQ1 ,
drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +7 -2
@@ -775,15 +775,20 @@ static int __amdgpu_ras_feature_enable(struct amdgpu_device *adev, obj =amdgpu_ras_create_obj (adev, head);if (!obj)return -EINVAL ; - }else { +amdgpu_ras_set_marker (adev, head,AMDGPU_MARKER_RAS_FEATURE ); + }else if (!amdgpu_ras_test_marker (adev, head, +AMDGPU_MARKER_RAS_FEATURE )) {/* In case we create obj somewhere else */ get_obj (obj); +amdgpu_ras_set_marker (adev, head,AMDGPU_MARKER_RAS_FEATURE ); } con->features |=BIT (head->block); }else {if (obj &&amdgpu_ras_is_feature_enabled (adev, head)) { con->features &= ~BIT (head->block); -put_obj (obj); +if (amdgpu_ras_test_and_clear_marker (adev, head, +AMDGPU_MARKER_RAS_FEATURE )) +put_obj (obj); } }
Signature · 1 line
2.43.5
RFC v2 7/15 drm/admgpu: rename amdgpu_ras_pre_fini() to amdgpu_ras_early_fini()
In reply to Jiang LiuRename amdgpu_ras_pre_fini() to amdgpu_ras_early_fini(), to keep same style with other code.
--- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 6 +++--- drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_jpeg.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_nbio.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c | 16 ++++++++-------- drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h | 6 +++--- drivers/gpu/drm/amd/amdgpu/amdgpu_sdma.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_vcn.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_xgmi.c | 2 +- drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c | 2 +- drivers/gpu/drm/amd/amdgpu/jpeg_v4_0_3.c | 2 +- drivers/gpu/drm/amd/amdgpu/mmhub_v1_8.c | 2 +- drivers/gpu/drm/amd/amdgpu/vcn_v4_0_3.c | 2 +- 14 files changed, 25 insertions(+), 25 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +3 -3
@@ -4613,7 +4613,7 @@ int amdgpu_device_init(struct amdgpu_device *adev,return 0 ; release_ras_con: -amdgpu_ras_pre_fini (adev); +amdgpu_ras_early_fini (adev);amdgpu_ras_fini (adev);if (amdgpu_sriov_vf (adev))amdgpu_virt_release_full_gpu (adev, true);
@@ -4705,7 +4705,7 @@ void amdgpu_device_fini_hw(struct amdgpu_device *adev)amdgpu_xcp_cfg_sysfs_fini (adev);/* disable ras feature must before hw fini */ -amdgpu_ras_pre_fini (adev); +amdgpu_ras_early_fini (adev);amdgpu_ttm_set_buffer_funcs_status (adev, false);
@@ -4922,7 +4922,7 @@ int amdgpu_device_suspend(struct drm_device *dev, bool notify_clients)cancel_delayed_work_sync (&adev->delayed_init_work);/* disable ras feature must before hw fini */ -amdgpu_ras_pre_fini (adev); +amdgpu_ras_early_fini (adev);amdgpu_ras_suspend (adev);amdgpu_device_ip_suspend_phase1 (adev);
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +1 -1
@@ -918,7 +918,7 @@ int amdgpu_gfx_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *rreturn 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }
drivers/gpu/drm/amd/amdgpu/amdgpu_jpeg.c +1 -1
@@ -300,7 +300,7 @@ int amdgpu_jpeg_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *return 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }
drivers/gpu/drm/amd/amdgpu/amdgpu_nbio.c +1 -1
@@ -71,6 +71,6 @@ int amdgpu_nbio_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *return 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }
drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +8 -8
@@ -4124,7 +4124,7 @@ int amdgpu_ras_block_late_init(struct amdgpu_device *adev,return 0 ; cleanup: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }
@@ -4135,7 +4135,7 @@ static int amdgpu_ras_block_late_init_default(struct amdgpu_device *adev, }/* helper function to remove ras fs node and interrupt handler */ -void amdgpu_ras_block_late_fini (struct amdgpu_device *adev, +void amdgpu_ras_block_early_fini (struct amdgpu_device *adev,struct ras_common_if *ras_block) {struct amdgpu_ras_block_object *ras_obj;
@@ -4156,10 +4156,10 @@ void amdgpu_ras_block_late_fini(struct amdgpu_device *adev,amdgpu_ras_interrupt_remove_handler (adev, ras_block); } -static void amdgpu_ras_block_late_fini_default (struct amdgpu_device *adev, +static void amdgpu_ras_block_early_fini_default (struct amdgpu_device *adev,struct ras_common_if *ras_block) { -return amdgpu_ras_block_late_fini (adev, ras_block); +return amdgpu_ras_block_early_fini (adev, ras_block); }/* do some init work after IP late init as dependence.
@@ -4267,7 +4267,7 @@ int amdgpu_ras_late_init(struct amdgpu_device *adev) }/* do some fini work before IP fini as dependence */ -int amdgpu_ras_pre_fini (struct amdgpu_device *adev) +int amdgpu_ras_early_fini (struct amdgpu_device *adev) {struct amdgpu_ras *con =amdgpu_ras_get_context (adev);struct amdgpu_ras_block_list *node, *tmp;
@@ -4284,10 +4284,10 @@ int amdgpu_ras_pre_fini(struct amdgpu_device *adev)if (!amdgpu_ras_is_supported (adev, obj->ras_comm.block))continue ; -if (obj->ras_fini) - obj->ras_fini (adev, &obj->ras_comm); +if (obj->ras_early_fini) + obj->ras_early_fini (adev, &obj->ras_comm);else -amdgpu_ras_block_late_fini_default (adev, &obj->ras_comm); +amdgpu_ras_block_early_fini_default (adev, &obj->ras_comm); } disable:
drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h +3 -3
@@ -711,7 +711,7 @@ struct amdgpu_ras_block_object {int (*ras_block_match)(struct amdgpu_ras_block_object *block_obj,enum amdgpu_ras_block block,uint32_t sub_block_index);int (*ras_late_init)(struct amdgpu_device *adev,struct ras_common_if *ras_block); -void (*ras_fini)(struct amdgpu_device *adev,struct ras_common_if *ras_block); +void (*ras_early_fini)(struct amdgpu_device *adev,struct ras_common_if *ras_block);ras_ih_cb ras_cb;const struct amdgpu_ras_block_hw_ops *hw_ops; };
@@ -825,13 +825,13 @@ amdgpu_ras_error_to_ta(enum amdgpu_ras_error_type error) {/* called in ip_init and ip_fini */ int amdgpu_ras_init (struct amdgpu_device *adev);int amdgpu_ras_late_init (struct amdgpu_device *adev); +int amdgpu_ras_early_fini (struct amdgpu_device *adev);int amdgpu_ras_fini (struct amdgpu_device *adev); -int amdgpu_ras_pre_fini (struct amdgpu_device *adev);int amdgpu_ras_block_late_init (struct amdgpu_device *adev,struct ras_common_if *ras_block); -void amdgpu_ras_block_late_fini (struct amdgpu_device *adev, +void amdgpu_ras_block_early_fini (struct amdgpu_device *adev,struct ras_common_if *ras_block); int amdgpu_ras_feature_enable(struct amdgpu_device *adev,
drivers/gpu/drm/amd/amdgpu/amdgpu_sdma.c +1 -1
@@ -116,7 +116,7 @@ int amdgpu_sdma_ras_late_init(struct amdgpu_device *adev,return 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }
drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c +1 -1
@@ -336,7 +336,7 @@ int amdgpu_umc_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *rreturn 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }
drivers/gpu/drm/amd/amdgpu/amdgpu_vcn.c +1 -1
@@ -1248,7 +1248,7 @@ int amdgpu_vcn_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *rreturn 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }
drivers/gpu/drm/amd/amdgpu/amdgpu_xgmi.c +1 -1
@@ -1174,7 +1174,7 @@ static int amdgpu_xgmi_ras_late_init(struct amdgpu_device *adev, struct ras_commreturn 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }
drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c +1 -1
@@ -5080,7 +5080,7 @@ static int gfx_v9_4_3_ras_late_init(struct amdgpu_device *adev, struct ras_commoreturn 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }
drivers/gpu/drm/amd/amdgpu/jpeg_v4_0_3.c +1 -1
@@ -1311,7 +1311,7 @@ static int jpeg_v4_0_3_ras_late_init(struct amdgpu_device *adev, struct ras_commreturn 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }
drivers/gpu/drm/amd/amdgpu/mmhub_v1_8.c +1 -1
@@ -789,7 +789,7 @@ static int mmhub_v1_8_ras_late_init(struct amdgpu_device *adev, struct ras_commoreturn 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }
drivers/gpu/drm/amd/amdgpu/vcn_v4_0_3.c +1 -1
@@ -1996,7 +1996,7 @@ static int vcn_v4_0_3_ras_late_init(struct amdgpu_device *adev, struct ras_commoreturn 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }
Signature · 1 line
2.43.5
RFC v2 8/15 drm/amdgpu: make IP block state machine works in stack like way
In reply to Jiang LiuThere are some mismatches between IP block state machine and its associated status flags, especially about the meaning of status.late_initialized. So let's make the state machine and associated status flas work in stack-like way as below:
Callback Statusearly_init: valid = truesw_init: sw = truehw_init: hw = truelate_init: late_initialized = trueearly_fini: late_initialized = falsehw_fini: hw = falsesw_fini: sw = falselate_fini: valid = false
--- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +4 -3
@@ -3376,6 +3376,8 @@ static int amdgpu_device_ip_fini_early(struct amdgpu_device *adev)DRM_DEBUG ("early_fini of IP block <%s> failed %d\n" , adev->ip_blocks[i].version->funcs->name, r); } + + adev->ip_blocks[i].status.late_initialized = false; }amdgpu_device_set_pg_state (adev,AMD_PG_STATE_UNGATE );
@@ -3445,15 +3447,14 @@ static int amdgpu_device_ip_fini(struct amdgpu_device *adev) } } adev->ip_blocks[i].status.sw = false; - adev->ip_blocks[i].status.valid = false; }for (i = adev->num_ip_blocks -1 ; i >=0 ; i--) { -if (!adev->ip_blocks[i].status.late_initialized) +if (!adev->ip_blocks[i].status.valid)continue ;if (adev->ip_blocks[i].version->funcs->late_fini) adev->ip_blocks[i].version->funcs->late_fini (&adev->ip_blocks[i]); - adev->ip_blocks[i].status.late_initialized = false; + adev->ip_blocks[i].status.valid = false; }amdgpu_ras_fini (adev);
Signature · 1 line
2.43.5
RFC v2 10/15 drm/admgpu: make device state machine work in stack like way
In reply to Jiang LiuMake the device state machine work in stack like way to better support suspend/resume by following changes:
1. amdgpu_driver_load_kms()
amdgpu_device_init() amdgpu_device_ip_early_init() ip_blocks[i].early_init() ip_blocks[i].status.valid = true amdgpu_device_ip_init() amdgpu_ras_init() ip_blocks[i].sw_init() ip_blocks[i].status.sw = true ip_blocks[i].hw_init() ip_blocks[i].status.hw = true amdgpu_device_ip_late_init() ip_blocks[i].late_init() ip_blocks[i].status.late_initialized = true amdgpu_ras_late_init() ras_blocks[i].ras_late_init() amdgpu_ras_feature_enable_on_boot()
2. amdgpu_pmops_suspend()/amdgpu_pmops_freeze()/amdgpu_pmops_poweroff()
amdgpu_device_suspend() amdgpu_ras_early_fini() ras_blocks[i].ras_early_fini() amdgpu_ras_feature_disable() amdgpu_ras_suspend() amdgpu_ras_disable_all_features() +++ ip_blocks[i].early_fini() +++ ip_blocks[i].status.late_initialized = false ip_blocks[i].suspend()
3. amdgpu_pmops_resume()/amdgpu_pmops_thaw()/amdgpu_pmops_restore()
amdgpu_device_resume() amdgpu_device_ip_resume() ip_blocks[i].resume() amdgpu_device_ip_late_init() ip_blocks[i].late_init() ip_blocks[i].status.late_initialized = true amdgpu_ras_late_init() ras_blocks[i].ras_late_init() amdgpu_ras_feature_enable_on_boot() amdgpu_ras_resume() amdgpu_ras_enable_all_features()
4. amdgpu_driver_unload_kms()
amdgpu_device_fini_hw() amdgpu_ras_early_fini() ras_blocks[i].ras_early_fini() +++ ip_blocks[i].early_fini() +++ ip_blocks[i].status.late_initialized = false ip_blocks[i].hw_fini() ip_blocks[i].status.hw = false
5. amdgpu_driver_release_kms()
amdgpu_device_fini_sw() amdgpu_device_ip_fini() ip_blocks[i].sw_fini() ip_blocks[i].status.sw = false --- ip_blocks[i].status.valid = false +++ amdgpu_ras_fini() ip_blocks[i].late_fini()
+++ ip_blocks[i].status.valid = false
--- ip_blocks[i].status.late_initialized = false
--- amdgpu_ras_fini()
The main changes include:
1) invoke ip_blocks[i].early_fini in amdgpu_pmops_suspend().
2) set ip_blocks[i].status.late_initialized to false after calling callback early_fini. We have auditted all usages of the late_initialized flag and no functional changes found.
3) only set ip_blocks[i].status.valid = false after calling the late_fini callback.
4) call amdgpu_ras_fini() before invoking ip_blocks[i].late_fini.
There's one more task left to analyze GPU reset related state machine transitions.
--- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 22 ++++++++++++++++++++-- 1 file changed, 20 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +20 -2
@@ -3449,6 +3449,8 @@ static int amdgpu_device_ip_fini(struct amdgpu_device *adev) adev->ip_blocks[i].status.sw = false; } +amdgpu_ras_fini (adev); +for (i = adev->num_ip_blocks -1 ; i >=0 ; i--) {if (!adev->ip_blocks[i].status.valid)continue ;
@@ -3457,8 +3459,6 @@ static int amdgpu_device_ip_fini(struct amdgpu_device *adev) adev->ip_blocks[i].status.valid = false; } -amdgpu_ras_fini (adev); -return 0 ; }
@@ -3516,6 +3516,24 @@ static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev)if (amdgpu_dpm_set_df_cstate (adev,DF_CSTATE_DISALLOW ))dev_warn (adev->dev,"Failed to disallow df cstate" ); +for (i = adev->num_ip_blocks -1 ; i >=0 ; i--) { +if (!adev->ip_blocks[i].status.valid) +continue ; +if (!adev->ip_blocks[i].status.late_initialized) +continue ; + +if (adev->ip_blocks[i].version->funcs->early_fini) { + r = adev->ip_blocks[i].version->funcs->early_fini (&adev->ip_blocks[i]); +if (r) { +DRM_ERROR (" of IP block <%s> failed %d\n" , + adev->ip_blocks[i].version->funcs->name, r); +return r; + } + } + + adev->ip_blocks[i].status.late_initialized = false; + } +for (i = adev->num_ip_blocks -1 ; i >=0 ; i--) {if (!adev->ip_blocks[i].status.valid)continue ;
Signature · 1 line
2.43.5
RFC v2 11/15 drm/amdgpu: convert ip block bool flags into an enum
In reply to Jiang LiuConvert ip block bool flags into an enumeration, to explicitly mark current state of the ip block. Also introduce helper functions to manipulate the ip block state. Now the state machine works as below:
Callbacks State after successfully execute callback
AMDGPU_IP_STATE_INVALID
.early_init() AMDGPU_IP_STATE_EARLY .sw_init() AMDGPU_IP_STATE_SW .hw_init() AMDGPU_IP_STATE_HW .late_init() AMDGPU_IP_STATE_LATE .early_fini() AMDGPU_IP_STATE_HW .hw_fini() AMDGPU_IP_STATE_SW .sw_fini() AMDGPU_IP_STATE_EARLY .late_fini() AMDGPU_IP_STATE_INVALID
--- drivers/gpu/drm/amd/amdgpu/aldebaran.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 48 ++++++- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 144 ++++++++++++-------- drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c | 4 +- drivers/gpu/drm/amd/amdgpu/amdgpu_psp.c | 6 +- drivers/gpu/drm/amd/amdgpu/amdgpu_reset.c | 6 +- drivers/gpu/drm/amd/amdgpu/sienna_cichlid.c | 2 +- drivers/gpu/drm/amd/amdgpu/smu_v13_0_10.c | 2 +- drivers/gpu/drm/amd/pm/swsmu/amdgpu_smu.c | 2 +- 9 files changed, 143 insertions(+), 73 deletions(-)
drivers/gpu/drm/amd/amdgpu/aldebaran.c +1 -1
@@ -301,7 +301,7 @@ static int aldebaran_mode2_restore_ip(struct amdgpu_device *adev)return r; } } - adev->ip_blocks[i].status.late_initialized = true; +amdgpu_device_ip_set_state (adev, i,AMDGPU_IP_STATE_LATE ); }amdgpu_device_set_cg_state (adev,AMD_CG_STATE_GATE );
drivers/gpu/drm/amd/amdgpu/amdgpu.h +44 -4
61 lines not on this page · raw message
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +86 -58
463 lines not on this page · raw message
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +2 -2
@@ -515,7 +515,7 @@ static int amdgpu_hw_ip_info(struct amdgpu_device *adev,for (i =0 ; i < adev->num_ip_blocks; i++)if (adev->ip_blocks[i].version->type == type && - adev->ip_blocks[i].status.valid) +amdgpu_device_ip_valid (adev, i))break ;if (i == adev->num_ip_blocks)
@@ -636,7 +636,7 @@ int amdgpu_info_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) type =amdgpu_ip_get_block_type (adev, info->query_hw_ip.type); ip_block =amdgpu_device_ip_get_ip_block (adev, type); -if (!ip_block || !ip_block->status.valid) +if (!ip_block || !amdgpu_ip_block_valid (ip_block))return -EINVAL ; if (adev->xcp_mgr && adev->xcp_mgr->num_xcps > 0 &&
drivers/gpu/drm/amd/amdgpu/amdgpu_psp.c +4 -2
@@ -3897,7 +3897,8 @@ static ssize_t psp_usbc_pd_fw_sysfs_read(struct device *dev,int ret; ip_block =amdgpu_device_ip_get_ip_block (adev,AMD_IP_BLOCK_TYPE_PSP ); -if (!ip_block || !ip_block->status.late_initialized) { +if (!ip_block || +amdgpu_ip_block_state (ip_block) <AMDGPU_IP_STATE_LATE ) {dev_info (adev->dev,"PSP block is not ready yet\n." );return -EBUSY ; }
@@ -3929,7 +3930,8 @@ static ssize_t psp_usbc_pd_fw_sysfs_write(struct device *dev,struct amdgpu_ip_block *ip_block; ip_block =amdgpu_device_ip_get_ip_block (adev,AMD_IP_BLOCK_TYPE_PSP ); -if (!ip_block || !ip_block->status.late_initialized) { +if (!ip_block || +amdgpu_ip_block_state (ip_block) <AMDGPU_IP_STATE_LATE ) {dev_err (adev->dev,"PSP block is not ready yet." );return -EBUSY ; }
drivers/gpu/drm/amd/amdgpu/amdgpu_reset.c +3 -3
@@ -31,9 +31,9 @@ static int amdgpu_reset_xgmi_reset_on_init_suspend(struct amdgpu_device *adev)int i;for (i = adev->num_ip_blocks -1 ; i >=0 ; i--) { -if (!adev->ip_blocks[i].status.valid) +if (!amdgpu_device_ip_valid (adev, i))continue ; -if (!adev->ip_blocks[i].status.hw) +if (amdgpu_device_ip_state (adev, i) <=AMDGPU_IP_STATE_HW )continue ;/* displays are handled in phase1 */ if (adev->ip_blocks[i].version->type ==AMD_IP_BLOCK_TYPE_DCE )
@@ -41,7 +41,7 @@ static int amdgpu_reset_xgmi_reset_on_init_suspend(struct amdgpu_device *adev)/* XXX handle errors */ amdgpu_ip_block_suspend (&adev->ip_blocks[i]); - adev->ip_blocks[i].status.hw = false; +amdgpu_device_ip_set_state (adev, i,AMDGPU_IP_STATE_SW ); }/* VCN FW shared region is in frambuffer, there are some flags
drivers/gpu/drm/amd/amdgpu/sienna_cichlid.c +1 -1
@@ -204,7 +204,7 @@ static int sienna_cichlid_mode2_restore_ip(struct amdgpu_device *adev)return r; } } - adev->ip_blocks[i].status.late_initialized = true; +amdgpu_device_ip_set_state (adev, i,AMDGPU_IP_STATE_LATE ); }amdgpu_device_set_cg_state (adev,AMD_CG_STATE_GATE );
drivers/gpu/drm/amd/amdgpu/smu_v13_0_10.c +1 -1
@@ -205,7 +205,7 @@ static int smu_v13_0_10_mode2_restore_ip(struct amdgpu_device *adev)return r; } } - adev->ip_blocks[i].status.late_initialized = true; +amdgpu_device_ip_set_state (adev, i,AMDGPU_IP_STATE_LATE ); }amdgpu_device_set_cg_state (adev,AMD_CG_STATE_GATE );
drivers/gpu/drm/amd/pm/swsmu/amdgpu_smu.c +1 -1
@@ -230,7 +230,7 @@ static bool is_vcn_enabled(struct amdgpu_device *adev)for (i =0 ; i < adev->num_ip_blocks; i++) {if ((adev->ip_blocks[i].version->type ==AMD_IP_BLOCK_TYPE_VCN || adev->ip_blocks[i].version->type ==AMD_IP_BLOCK_TYPE_JPEG ) && - !adev->ip_blocks[i].status.valid) + !amdgpu_device_ip_valid (adev, i))return false; }
Signature · 1 line
2.43.5
RFC v2 12/15 drm/amdgpu: introduce IP block iterators to reduce duplicated code
In reply to Jiang LiuIntroduce following IP block iterators to reduce duplicated code:
- amdgpu_for_each_ip_block
- amdgpu_for_each_ip_block_reverse
- amdgpu_for_each_ip_block_valid
- amdgpu_for_each_ip_block_valid_reverse
--- drivers/gpu/drm/amd/amdgpu/aldebaran.c | 46 +- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 30 +- .../gpu/drm/amd/amdgpu/amdgpu_dev_coredump.c | 3 +- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 479 ++++++++---------- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 10 +- drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c | 16 +- drivers/gpu/drm/amd/amdgpu/amdgpu_psp.c | 6 +- drivers/gpu/drm/amd/amdgpu/amdgpu_reset.c | 14 +- drivers/gpu/drm/amd/amdgpu/sienna_cichlid.c | 50 +- drivers/gpu/drm/amd/amdgpu/smu_v13_0_10.c | 51 +- drivers/gpu/drm/amd/pm/swsmu/amdgpu_smu.c | 10 +- 11 files changed, 337 insertions(+), 378 deletions(-)
drivers/gpu/drm/amd/amdgpu/aldebaran.c +19 -27
@@ -73,19 +73,18 @@ aldebaran_get_reset_handler(struct amdgpu_reset_control *reset_ctl,static int aldebaran_mode2_suspend_ip (struct amdgpu_device *adev) { -int r, i; +int r; +struct amdgpu_ip_block *ip_block;amdgpu_device_set_pg_state (adev,AMD_PG_STATE_UNGATE );amdgpu_device_set_cg_state (adev,AMD_CG_STATE_UNGATE ); -for (i = adev->num_ip_blocks -1 ; i >=0 ; i--) { -if (!(adev->ip_blocks[i].version->type == -AMD_IP_BLOCK_TYPE_GFX || - adev->ip_blocks[i].version->type == -AMD_IP_BLOCK_TYPE_SDMA )) +amdgpu_for_each_ip_block_reverse (adev, ip_block) { +if (!(ip_block->version->type ==AMD_IP_BLOCK_TYPE_GFX || + ip_block->version->type ==AMD_IP_BLOCK_TYPE_SDMA ))continue ; - r =amdgpu_ip_block_suspend (&adev->ip_blocks[i]); + r =amdgpu_ip_block_suspend (ip_block);if (r)return r; }
@@ -201,7 +200,7 @@ static int aldebaran_mode2_restore_ip(struct amdgpu_device *adev) {struct amdgpu_firmware_info *ucode_list[AMDGPU_UCODE_ID_MAXIMUM ];struct amdgpu_firmware_info *ucode; -struct amdgpu_ip_block *cmn_block; +struct amdgpu_ip_block *cmn_block, *ip_block;int ucode_count =0 ;int i, r;
@@ -269,39 +268,32 @@ static int aldebaran_mode2_restore_ip(struct amdgpu_device *adev)return r; } -for (i =0 ; i < adev->num_ip_blocks; i++) { -if (!(adev->ip_blocks[i].version->type == -AMD_IP_BLOCK_TYPE_GFX || - adev->ip_blocks[i].version->type == -AMD_IP_BLOCK_TYPE_SDMA )) +amdgpu_for_each_ip_block (adev, ip_block) { +if (!(ip_block->version->type ==AMD_IP_BLOCK_TYPE_GFX || + ip_block->version->type ==AMD_IP_BLOCK_TYPE_SDMA ))continue ; - r =amdgpu_ip_block_resume (&adev->ip_blocks[i]); + r =amdgpu_ip_block_resume (ip_block);if (r)return r; } -for (i =0 ; i < adev->num_ip_blocks; i++) { -if (!(adev->ip_blocks[i].version->type == -AMD_IP_BLOCK_TYPE_GFX || - adev->ip_blocks[i].version->type == -AMD_IP_BLOCK_TYPE_SDMA || - adev->ip_blocks[i].version->type == -AMD_IP_BLOCK_TYPE_COMMON )) +amdgpu_for_each_ip_block (adev, ip_block) { +if (!(ip_block->version->type ==AMD_IP_BLOCK_TYPE_GFX || + ip_block->version->type ==AMD_IP_BLOCK_TYPE_SDMA || + ip_block->version->type ==AMD_IP_BLOCK_TYPE_COMMON ))continue ; -if (adev->ip_blocks[i].version->funcs->late_init) { - r = adev->ip_blocks[i].version->funcs->late_init ( - &adev->ip_blocks[i]); +if (ip_block->version->funcs->late_init) { + r = ip_block->version->funcs->late_init (ip_block);if (r) {dev_err (adev->dev,"late_init of IP block <%s> failed %d after reset\n" , - adev->ip_blocks[i].version->funcs->name, - r); + ip_block->version->funcs->name, r);return r; } } -amdgpu_device_ip_set_state (adev, i,AMDGPU_IP_STATE_LATE ); +amdgpu_ip_block_set_state (ip_block,AMDGPU_IP_STATE_LATE ); }amdgpu_device_set_cg_state (adev,AMD_CG_STATE_GATE );
drivers/gpu/drm/amd/amdgpu/amdgpu.h +25 -5
@@ -450,11 +450,9 @@ amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev,int amdgpu_device_ip_block_add (struct amdgpu_device *adev,const struct amdgpu_ip_block_version *ip_block_version); -void amdgpu_device_ip_set_state (struct amdgpu_device *adev,int index, -enum amdgpu_device_ip_state state); -enum amdgpu_device_ip_state amdgpu_device_ip_state (struct amdgpu_device *adev, -int index); -bool amdgpu_device_ip_valid (struct amdgpu_device *adev,int index); +enum amdgpu_device_ip_state +amdgpu_device_ip_state (struct amdgpu_device *adev,int index); +void amdgpu_ip_block_set_state (struct amdgpu_ip_block *ip_block,enum amdgpu_device_ip_state state);
@@ -491,6 +489,28 @@ static inline bool amdgpu_ip_block_test_and_clear_marker(struct amdgpu_ip_blockreturn set; } +#define amdgpu_for_each_ip_block (adev, entry) \ + for (int __i = 0; \ + __i < (adev)->num_ip_blocks && ((entry) = &(adev)->ip_blocks[__i]); \ + __i++) + +#define amdgpu_for_each_ip_block_reverse (adev, entry) \ + for (int __i = (adev)->num_ip_blocks; \ + __i >= 0 && ((entry) = &(adev)->ip_blocks[__i]); \ + __i--) + +#define amdgpu_for_each_ip_block_valid (adev, entry) \ + for (int __i = (adev)->num_ip_blocks; \ + __i >= 0 && ((entry) = &(adev)->ip_blocks[__i]); \ + __i--) \ + if (amdgpu_ip_block_valid((entry))) + +#define amdgpu_for_each_ip_block_valid_reverse (adev, entry) \ + for (int __i = (adev)->num_ip_blocks; \ + __i >= 0 && (entry = &(adev)->ip_blocks[__i]); \ + __i--) \ + if (amdgpu_ip_block_valid((entry))) +/* * BIOS. */
drivers/gpu/drm/amd/amdgpu/amdgpu_dev_coredump.c +1 -2
@@ -282,8 +282,7 @@ amdgpu_devcoredump_read(char *buffer, loff_t offset, size_t count,/* dump the ip state for each ip */ drm_printf (&p,"IP Dump\n" ); -for (int i =0 ; i < coredump->adev->num_ip_blocks; i++) { - ip_block = &coredump->adev->ip_blocks[i]; +amdgpu_for_each_ip_block (coredump->adev, ip_block) {if (ip_block->version->funcs->print_ip_state) {drm_printf (&p,"IP: %s\n" , ip_block->version->funcs->name); ip_block->version->funcs->print_ip_state (ip_block, &p);
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +221 -258
913 lines not on this page · raw message
drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +5 -5
@@ -36,13 +36,13 @@static void amdgpu_job_do_core_dump (struct amdgpu_device *adev,struct amdgpu_job *job) { -int i; +struct amdgpu_ip_block *block;dev_info (adev->dev,"Dumping IP State\n" ); -for (i =0 ; i < adev->num_ip_blocks; i++) -if (adev->ip_blocks[i].version->funcs->dump_ip_state) - adev->ip_blocks[i].version->funcs - ->dump_ip_state ((void *)&adev->ip_blocks[i]); +amdgpu_for_each_ip_block (adev, block) { +if (block->version->funcs->dump_ip_state) + block->version->funcs->dump_ip_state ((void *)block); + }dev_info (adev->dev,"Dumping IP State Completed\n" );amdgpu_coredump (adev, true, false, job);
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +10 -6
@@ -400,6 +400,8 @@ static int amdgpu_hw_ip_info(struct amdgpu_device *adev,enum amd_ip_block_type type;unsigned int num_rings =0 ;unsigned int i, j; +bool found = false; +struct amdgpu_ip_block *ip_block;if (info->query_hw_ip.ip_instance >=AMDGPU_HW_IP_INSTANCE_MAX_COUNT )return -EINVAL ;
@@ -513,19 +515,21 @@ static int amdgpu_hw_ip_info(struct amdgpu_device *adev,return -EINVAL ; } -for (i =0 ; i < adev->num_ip_blocks; i++) -if (adev->ip_blocks[i].version->type == type && -amdgpu_device_ip_valid (adev, i)) +amdgpu_for_each_ip_block_valid (adev, ip_block) { +if (ip_block->version->type == type) { + found = true;break ; + } + } -if (i == adev->num_ip_blocks) +if (!found)return 0 ; num_rings =min (amdgpu_ctx_num_entities[info->query_hw_ip.type], num_rings); - result->hw_ip_version_major = adev->ip_blocks[i].version->major; - result->hw_ip_version_minor = adev->ip_blocks[i].version->minor; + result->hw_ip_version_major = ip_block->version->major; + result->hw_ip_version_minor = ip_block->version->minor;if (adev->asic_type >=CHIP_VEGA10 ) {switch (type) {
drivers/gpu/drm/amd/amdgpu/amdgpu_psp.c +2 -4
@@ -3897,8 +3897,7 @@ static ssize_t psp_usbc_pd_fw_sysfs_read(struct device *dev,int ret; ip_block =amdgpu_device_ip_get_ip_block (adev,AMD_IP_BLOCK_TYPE_PSP ); -if (!ip_block || -amdgpu_ip_block_state (ip_block) <AMDGPU_IP_STATE_LATE ) { +if (!ip_block ||amdgpu_ip_block_state (ip_block) <AMDGPU_IP_STATE_LATE ) {dev_info (adev->dev,"PSP block is not ready yet\n." );return -EBUSY ; }
@@ -3930,8 +3929,7 @@ static ssize_t psp_usbc_pd_fw_sysfs_write(struct device *dev,struct amdgpu_ip_block *ip_block; ip_block =amdgpu_device_ip_get_ip_block (adev,AMD_IP_BLOCK_TYPE_PSP ); -if (!ip_block || -amdgpu_ip_block_state (ip_block) <AMDGPU_IP_STATE_LATE ) { +if (!ip_block ||amdgpu_ip_block_state (ip_block) <AMDGPU_IP_STATE_LATE ) {dev_err (adev->dev,"PSP block is not ready yet." );return -EBUSY ; }
drivers/gpu/drm/amd/amdgpu/amdgpu_reset.c +6 -8
@@ -28,20 +28,18 @@static int amdgpu_reset_xgmi_reset_on_init_suspend (struct amdgpu_device *adev) { -int i; +struct amdgpu_ip_block *ip_block; -for (i = adev->num_ip_blocks -1 ; i >=0 ; i--) { -if (!amdgpu_device_ip_valid (adev, i)) -continue ; -if (amdgpu_device_ip_state (adev, i) <=AMDGPU_IP_STATE_HW ) +amdgpu_for_each_ip_block_valid_reverse (adev, ip_block) { +if (amdgpu_ip_block_state (ip_block) <AMDGPU_IP_STATE_HW )continue ;/* displays are handled in phase1 */ -if (adev->ip_blocks[i].version->type ==AMD_IP_BLOCK_TYPE_DCE ) +if (ip_block->version->type ==AMD_IP_BLOCK_TYPE_DCE )continue ;/* XXX handle errors */ -amdgpu_ip_block_suspend (&adev->ip_blocks[i]); -amdgpu_device_ip_set_state (adev, i,AMDGPU_IP_STATE_SW ); +amdgpu_ip_block_suspend (ip_block); +amdgpu_ip_block_set_state (ip_block,AMDGPU_IP_STATE_SW ); }/* VCN FW shared region is in frambuffer, there are some flags
drivers/gpu/drm/amd/amdgpu/sienna_cichlid.c +22 -28
93 lines not on this page · raw message
drivers/gpu/drm/amd/amdgpu/smu_v13_0_10.c +21 -30
88 lines not on this page · raw message
drivers/gpu/drm/amd/pm/swsmu/amdgpu_smu.c +5 -5
@@ -225,12 +225,12 @@ static int smu_set_gfx_imu_enable(struct smu_context *smu)static bool is_vcn_enabled (struct amdgpu_device *adev) { -int i; +struct amdgpu_ip_block *ip_block; -for (i =0 ; i < adev->num_ip_blocks; i++) { -if ((adev->ip_blocks[i].version->type ==AMD_IP_BLOCK_TYPE_VCN || - adev->ip_blocks[i].version->type ==AMD_IP_BLOCK_TYPE_JPEG ) && - !amdgpu_device_ip_valid (adev, i)) +amdgpu_for_each_ip_block (adev, ip_block) { +if ((ip_block->version->type ==AMD_IP_BLOCK_TYPE_VCN || + ip_block->version->type ==AMD_IP_BLOCK_TYPE_JPEG ) && + !amdgpu_ip_block_valid (ip_block))return false; }
Signature · 1 line
2.43.5
RFC v2 14/15 drm/amdgpu/nbio: improve the way to manage irq reference count
In reply to Jiang LiuRefactor nbio related code to improve the way to manage irq reference count. Originally amdgpu_irq_get() is called from ip_blocks[].late_init and amdgpu_irq_put is called from ip_blocks[].hw_fini. The asymmetric design may cause issue under certain conditions. So
1) introduce amdgpu_nbio_ras_early_fini() to undo work done by amdgpu_nbio_ras_late_init().
2) remove call of amdgpu_irq_put in xxxx_hw_fini().
3) record the status where reference count is held for specific irq.
--- drivers/gpu/drm/amd/amdgpu/amdgpu_nbio.c | 16 +++++++++++++++- drivers/gpu/drm/amd/amdgpu/amdgpu_nbio.h | 1 + drivers/gpu/drm/amd/amdgpu/nbio_v7_4.c | 1 + drivers/gpu/drm/amd/amdgpu/nbio_v7_9.c | 1 + drivers/gpu/drm/amd/amdgpu/soc15.c | 16 ---------------- 5 files changed, 18 insertions(+), 17 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_nbio.c +15 -1
@@ -64,13 +64,27 @@ int amdgpu_nbio_ras_late_init(struct amdgpu_device *adev, struct ras_common_if * r =amdgpu_irq_get (adev, &adev->nbio.ras_controller_irq,0 );if (r) goto late_fini; +amdgpu_ras_set_marker (adev, ras_block,AMDGPU_MARKER_IRQ0 ); r =amdgpu_irq_get (adev, &adev->nbio.ras_err_event_athub_irq,0 );if (r) goto late_fini; +amdgpu_ras_set_marker (adev, ras_block,AMDGPU_MARKER_IRQ1 ); }return 0 ; late_fini: -amdgpu_ras_block_early_fini (adev, ras_block); +amdgpu_nbio_ras_early_fini (adev, ras_block);return r; } + +void amdgpu_nbio_ras_early_fini (struct amdgpu_device *adev,struct ras_common_if *ras_block) +{ +if (amdgpu_ras_is_supported (adev, adev->nbio.ras_if->block)) { +if (amdgpu_ras_test_and_clear_marker (adev, ras_block,AMDGPU_MARKER_IRQ0 )) +amdgpu_irq_put (adev, &adev->nbio.ras_controller_irq,0 ); +if (amdgpu_ras_test_and_clear_marker (adev, ras_block,AMDGPU_MARKER_IRQ1 )) +amdgpu_irq_put (adev, &adev->nbio.ras_err_event_athub_irq,0 ); + } + +amdgpu_ras_block_early_fini (adev, ras_block); +}
drivers/gpu/drm/amd/amdgpu/amdgpu_nbio.h +1 -0
@@ -117,6 +117,7 @@ struct amdgpu_nbio {int amdgpu_nbio_ras_sw_init (struct amdgpu_device *adev);int amdgpu_nbio_ras_late_init (struct amdgpu_device *adev,struct ras_common_if *ras_block); +void amdgpu_nbio_ras_early_fini (struct amdgpu_device *adev,struct ras_common_if *ras_block);u64 amdgpu_nbio_get_pcie_replay_count (struct amdgpu_device *adev);#endif
drivers/gpu/drm/amd/amdgpu/nbio_v7_4.c +1 -0
@@ -665,6 +665,7 @@ struct amdgpu_nbio_ras nbio_v7_4_ras = { }, .hw_ops = &nbio_v7_4_ras_hw_ops, .ras_late_init = amdgpu_nbio_ras_late_init, + .ras_early_fini = amdgpu_nbio_ras_early_fini, }, .handle_ras_controller_intr_no_bifring = nbio_v7_4_handle_ras_controller_intr_no_bifring, .handle_ras_err_event_athub_intr_no_bifring = nbio_v7_4_handle_ras_err_event_athub_intr_no_bifring,
drivers/gpu/drm/amd/amdgpu/nbio_v7_9.c +1 -0
@@ -703,6 +703,7 @@ struct amdgpu_nbio_ras nbio_v7_9_ras = { }, .hw_ops = &nbio_v7_9_ras_hw_ops, .ras_late_init = amdgpu_nbio_ras_late_init, + .ras_early_fini = amdgpu_nbio_ras_early_fini, }, .handle_ras_controller_intr_no_bifring = nbio_v7_9_handle_ras_controller_intr_no_bifring, .handle_ras_err_event_athub_intr_no_bifring = nbio_v7_9_handle_ras_err_event_athub_intr_no_bifring,
drivers/gpu/drm/amd/amdgpu/soc15.c +0 -16
@@ -1328,22 +1328,6 @@ static int soc15_common_hw_fini(struct amdgpu_ip_block *ip_block)if (amdgpu_sriov_vf (adev))xgpu_ai_mailbox_put_irq (adev); -/* -* For minimal init, late_init is not called, hence RAS irqs are not -* enabled. -*/ -if ((!amdgpu_sriov_vf (adev)) && - (adev->init_lvl->level !=AMDGPU_INIT_LEVEL_MINIMAL_XGMI ) && - adev->nbio.ras_if && -amdgpu_ras_is_supported (adev, adev->nbio.ras_if->block)) { -if (adev->nbio.ras && - adev->nbio.ras->init_ras_controller_interrupt) -amdgpu_irq_put (adev, &adev->nbio.ras_controller_irq,0 ); -if (adev->nbio.ras && - adev->nbio.ras->init_ras_err_event_athub_interrupt) -amdgpu_irq_put (adev, &adev->nbio.ras_err_event_athub_irq,0 ); - } -return 0 ; }
Signature · 1 line
2.43.5
RFC v2 15/15 drm/amdgpu/asic: make ip block operations symmetric by .early_fini()
In reply to Jiang LiuMake ip block operations for asic symmetric by making using of the .early_fini() hook, which will undo work done by the .late_init() hook.
1) introduce xxx_common_early_fini() for nv/soc15/soc21/soc24.
2) move enable_doorbell_selfring_aperture(adev, false) from .hw_init() into .early_fini().
3) call xgpu_nv_mailbox_put_irq() for nv.c to avoid possible resource leakage.
4) use flags to track irq reference count usage.
--- drivers/gpu/drm/amd/amdgpu/nv.c | 14 +++++++++++- drivers/gpu/drm/amd/amdgpu/soc15.c | 22 +++++++++++-------- drivers/gpu/drm/amd/amdgpu/soc21.c | 35 ++++++++++++++++++++---------- drivers/gpu/drm/amd/amdgpu/soc24.c | 22 +++++++++++-------- 4 files changed, 63 insertions(+), 30 deletions(-)
drivers/gpu/drm/amd/amdgpu/nv.c +13 -1
@@ -973,6 +973,18 @@ static int nv_common_late_init(struct amdgpu_ip_block *ip_block)return 0 ; } +static int nv_common_early_fini (struct amdgpu_ip_block *ip_block) +{ +struct amdgpu_device *adev = ip_block->adev; + + adev->nbio.funcs->enable_doorbell_selfring_aperture (adev, false); + +if (amdgpu_sriov_vf (adev)) +xgpu_nv_mailbox_put_irq (adev); + +return 0 ; +} +static int nv_common_sw_init (struct amdgpu_ip_block *ip_block) {struct amdgpu_device *adev = ip_block->adev;
@@ -1019,7 +1031,6 @@ static int nv_common_hw_fini(struct amdgpu_ip_block *ip_block)* selfring doorbell. */ adev->nbio.funcs->enable_doorbell_aperture (adev, false); - adev->nbio.funcs->enable_doorbell_selfring_aperture (adev, false);return 0 ; }
@@ -1095,6 +1106,7 @@ static const struct amd_ip_funcs nv_common_ip_funcs = { .name = "nv_common", .early_init = nv_common_early_init, .late_init = nv_common_late_init, + .early_fini = nv_common_early_fini, .sw_init = nv_common_sw_init, .hw_init = nv_common_hw_init, .hw_fini = nv_common_hw_fini,
drivers/gpu/drm/amd/amdgpu/soc15.c +13 -9
@@ -1246,6 +1246,18 @@ static int soc15_common_late_init(struct amdgpu_ip_block *ip_block)return 0 ; } +static int soc15_common_early_fini (struct amdgpu_ip_block *ip_block) +{ +struct amdgpu_device *adev = ip_block->adev; + + adev->nbio.funcs->enable_doorbell_selfring_aperture (adev, false); + +if (amdgpu_sriov_vf (adev)) +xgpu_ai_mailbox_put_irq (adev); + +return 0 ; +} +static int soc15_common_sw_init (struct amdgpu_ip_block *ip_block) {struct amdgpu_device *adev = ip_block->adev;
@@ -1317,16 +1329,7 @@ static int soc15_common_hw_fini(struct amdgpu_ip_block *ip_block) {struct amdgpu_device *adev = ip_block->adev; -/* Disable the doorbell aperture and selfring doorbell aperture -* separately in hw_fini because soc15_enable_doorbell_aperture -* has been removed and there is no need to delay disabling -* selfring doorbell. -*/ adev->nbio.funcs->enable_doorbell_aperture (adev, false); - adev->nbio.funcs->enable_doorbell_selfring_aperture (adev, false); - -if (amdgpu_sriov_vf (adev)) -xgpu_ai_mailbox_put_irq (adev);return 0 ; }
@@ -1496,6 +1499,7 @@ static const struct amd_ip_funcs soc15_common_ip_funcs = { .name = "soc15_common", .early_init = soc15_common_early_init, .late_init = soc15_common_late_init, + .early_fini = soc15_common_early_fini, .sw_init = soc15_common_sw_init, .sw_fini = soc15_common_sw_fini, .hw_init = soc15_common_hw_init,
drivers/gpu/drm/amd/amdgpu/soc21.c +24 -11
@@ -796,6 +796,7 @@ static int soc21_common_early_init(struct amdgpu_ip_block *ip_block)static int soc21_common_late_init (struct amdgpu_ip_block *ip_block) { +int r;struct amdgpu_device *adev = ip_block->adev;if (amdgpu_sriov_vf (adev)) {
@@ -816,12 +817,16 @@ static int soc21_common_late_init(struct amdgpu_ip_block *ip_block) } }else {if (adev->nbio.ras && - adev->nbio.ras_err_event_athub_irq.funcs) + adev->nbio.ras_err_event_athub_irq.funcs) {/* don't need to fail gpu late init * if enabling athub_err_event interrupt failed * nbio v4_3 only support fatal error hanlding * just enable the interrupt directly */ -amdgpu_irq_get (adev, &adev->nbio.ras_err_event_athub_irq,0 ); + r =amdgpu_irq_get (adev, &adev->nbio.ras_err_event_athub_irq,0 ); +if (r) +return r; +amdgpu_ip_block_set_marker (ip_block,AMDGPU_MARKER_IRQ0 ); + } }/* Enable selfring doorbell aperture late because doorbell BAR
@@ -832,6 +837,22 @@ static int soc21_common_late_init(struct amdgpu_ip_block *ip_block)return 0 ; } +static int soc21_common_early_fini (struct amdgpu_ip_block *ip_block) +{ +struct amdgpu_device *adev = ip_block->adev; + + adev->nbio.funcs->enable_doorbell_selfring_aperture (adev, false); + +if (amdgpu_sriov_vf (adev)) { +xgpu_nv_mailbox_put_irq (adev); + }else { +if (amdgpu_ip_block_test_and_clear_marker (ip_block,AMDGPU_MARKER_IRQ0 )) +amdgpu_irq_put (adev, &adev->nbio.ras_err_event_athub_irq,0 ); + } + +return 0 ; +} +static int soc21_common_sw_init (struct amdgpu_ip_block *ip_block) {struct amdgpu_device *adev = ip_block->adev;
@@ -872,15 +893,6 @@ static int soc21_common_hw_fini(struct amdgpu_ip_block *ip_block)* selfring doorbell. */ adev->nbio.funcs->enable_doorbell_aperture (adev, false); - adev->nbio.funcs->enable_doorbell_selfring_aperture (adev, false); - -if (amdgpu_sriov_vf (adev)) { -xgpu_nv_mailbox_put_irq (adev); - }else { -if (adev->nbio.ras && - adev->nbio.ras_err_event_athub_irq.funcs) -amdgpu_irq_put (adev, &adev->nbio.ras_err_event_athub_irq,0 ); - }return 0 ; }
@@ -985,6 +997,7 @@ static const struct amd_ip_funcs soc21_common_ip_funcs = { .name = "soc21_common", .early_init = soc21_common_early_init, .late_init = soc21_common_late_init, + .early_fini = soc21_common_early_fini, .sw_init = soc21_common_sw_init, .hw_init = soc21_common_hw_init, .hw_fini = soc21_common_hw_fini,
drivers/gpu/drm/amd/amdgpu/soc24.c +13 -9
@@ -465,6 +465,18 @@ static int soc24_common_late_init(struct amdgpu_ip_block *ip_block)return 0 ; } +static int soc24_common_early_fini (struct amdgpu_ip_block *ip_block) +{ +struct amdgpu_device *adev = ip_block->adev; + + adev->nbio.funcs->enable_doorbell_selfring_aperture (adev, false); + +if (amdgpu_sriov_vf (adev)) +xgpu_nv_mailbox_put_irq (adev); + +return 0 ; +} +static int soc24_common_sw_init (struct amdgpu_ip_block *ip_block) {struct amdgpu_device *adev = ip_block->adev;
@@ -509,15 +521,6 @@ static int soc24_common_hw_fini(struct amdgpu_ip_block *ip_block)* selfring doorbell. */ adev->nbio.funcs->enable_doorbell_aperture (adev, false); - adev->nbio.funcs->enable_doorbell_selfring_aperture (adev, false); - -if (amdgpu_sriov_vf (adev)) { -xgpu_nv_mailbox_put_irq (adev); - }else { -if (adev->nbio.ras && - adev->nbio.ras_err_event_athub_irq.funcs) -amdgpu_irq_put (adev, &adev->nbio.ras_err_event_athub_irq,0 ); - }return 0 ; }
@@ -590,6 +593,7 @@ static const struct amd_ip_funcs soc24_common_ip_funcs = { .name = "soc24_common", .early_init = soc24_common_early_init, .late_init = soc24_common_late_init, + .early_fini = soc24_common_early_fini, .sw_init = soc24_common_sw_init, .hw_init = soc24_common_hw_init, .hw_fini = soc24_common_hw_fini,
Signature · 1 line
2.43.5
RFC v2 13/15 drm/amdgpu: walk IP blocks in reverse order when shutdown
In reply to Jiang LiuWalk IP blocks in reverse order in function amdgpu_device_ip_fini_early and amdgpu_device_smu_fini_early, to keep consistence with other finish functions.
--- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +2 -2
@@ -3358,7 +3358,7 @@ static void amdgpu_device_smu_fini_early(struct amdgpu_device *adev)if (amdgpu_ip_version (adev,GC_HWIP ,0 ) >IP_VERSION (9 ,0 ,0 ))return ; -amdgpu_for_each_ip_block (adev, ip_block) { +amdgpu_for_each_ip_block_reverse (adev, ip_block) {if (amdgpu_ip_block_state (ip_block) <AMDGPU_IP_STATE_HW )continue ;if (ip_block->version->type ==AMD_IP_BLOCK_TYPE_SMC ) {
@@ -3373,7 +3373,7 @@ static int amdgpu_device_ip_fini_early(struct amdgpu_device *adev)int r;struct amdgpu_ip_block *ip_block; -amdgpu_for_each_ip_block (adev, ip_block) { +amdgpu_for_each_ip_block_reverse (adev, ip_block) {if (!ip_block->version->funcs->early_fini)continue ;
Signature · 1 line
2.43.5
Re: [RFC v2 10/15] drm/admgpu: make device state machine work in stack like way
In reply to Jiang LiuQuoted 75 lines · Jiang Liu
Make the device state machine work in stack like way to better support suspend/resume by following changes:
1.
amdgpu_driver_load_kms()amdgpu_device_init() amdgpu_device_ip_early_init() ip_blocks[i].early_init() ip_blocks[i].status.valid = true amdgpu_device_ip_init() amdgpu_ras_init() ip_blocks[i].sw_init() ip_blocks[i].status.sw = true ip_blocks[i].hw_init() ip_blocks[i].status.hw = true amdgpu_device_ip_late_init() ip_blocks[i].late_init() ip_blocks[i].status.late_initialized = true amdgpu_ras_late_init() ras_blocks[i].ras_late_init() amdgpu_ras_feature_enable_on_boot()2. amdgpu_pmops_suspend()/amdgpu_pmops_freeze()/amdgpu_pmops_poweroff()
amdgpu_device_suspend() amdgpu_ras_early_fini() ras_blocks[i].ras_early_fini() amdgpu_ras_feature_disable() amdgpu_ras_suspend() amdgpu_ras_disable_all_features() +++ ip_blocks[i].early_fini() +++ ip_blocks[i].status.late_initialized = false ip_blocks[i].suspend()3. amdgpu_pmops_resume()/amdgpu_pmops_thaw()/amdgpu_pmops_restore()
amdgpu_device_resume() amdgpu_device_ip_resume() ip_blocks[i].resume() amdgpu_device_ip_late_init() ip_blocks[i].late_init() ip_blocks[i].status.late_initialized = true amdgpu_ras_late_init() ras_blocks[i].ras_late_init() amdgpu_ras_feature_enable_on_boot() amdgpu_ras_resume() amdgpu_ras_enable_all_features()4.
amdgpu_driver_unload_kms()amdgpu_device_fini_hw() amdgpu_ras_early_fini() ras_blocks[i].ras_early_fini() +++ ip_blocks[i].early_fini() +++ ip_blocks[i].status.late_initialized = false ip_blocks[i].hw_fini() ip_blocks[i].status.hw = false5.
amdgpu_driver_release_kms()amdgpu_device_fini_sw() amdgpu_device_ip_fini() ip_blocks[i].sw_fini() ip_blocks[i].status.sw = false --- ip_blocks[i].status.valid = false +++ amdgpu_ras_fini() ip_blocks[i].late_fini()+++ ip_blocks[i].status.valid = false
--- ip_blocks[i].status.late_initialized = false
---amdgpu_ras_fini()The main changes include:
1) invoke ip_blocks[i].early_fini inamdgpu_pmops_suspend().
2) set ip_blocks[i].status.late_initialized to false after calling callbackearly_fini. We have auditted all usages of thelate_initializedflag and no functional changes found.
3) only set ip_blocks[i].status.valid = false after calling thelate_finicallback.
4) callamdgpu_ras_fini()before invoking ip_blocks[i].late_fini.
Quoted 4 lines
There's one more task left to analyze GPU reset related state machine transitions.
Signed-off-by Jiang Liu <[email protected]>
Ideally I think you should swap the order of patch 10 and 11, what do you think?
Quoted 51 lines
--- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 22 ++++++++++++++++++++-- 1 file changed, 20 insertions(+), 2 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +20
-2@@ -3449,6 +3449,8 @@ static int amdgpu_device_ip_fini(struct amdgpu_device *adev) adev->ip_blocks[i].status.sw = false; } +amdgpu_ras_fini (adev); +for (i = adev->num_ip_blocks -1 ; i >=0 ; i--) {if (!adev->ip_blocks[i].status.valid)continue ;@@ -3457,8 +3459,6 @@ static int amdgpu_device_ip_fini(struct amdgpu_device *adev) adev->ip_blocks[i].status.valid = false; } -amdgpu_ras_fini (adev); -return 0 ; }@@ -3516,6 +3516,24 @@ static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev)if (amdgpu_dpm_set_df_cstate (adev,DF_CSTATE_DISALLOW ))dev_warn (adev->dev,"Failed to disallow df cstate" ); +for (i = adev->num_ip_blocks -1 ; i >=0 ; i--) { +if (!adev->ip_blocks[i].status.valid) +continue ; +if (!adev->ip_blocks[i].status.late_initialized) +continue ; + +if (adev->ip_blocks[i].version->funcs->early_fini) { + r = adev->ip_blocks[i].version->funcs->early_fini (&adev->ip_blocks[i]); +if (r) { +DRM_ERROR (" of IP block <%s> failed %d\n" , + adev->ip_blocks[i].version->funcs->name, r); +return r; + } + } + + adev->ip_blocks[i].status.late_initialized = false; + } +for (i = adev->num_ip_blocks -1 ; i >=0 ; i--) {if (!adev->ip_blocks[i].status.valid)continue ;
Re: [RFC v2 13/15] drm/amdgpu: walk IP blocks in reverse order when shutdown
In reply to Jiang LiuQuoted 5 lines · Jiang Liu
Walk IP blocks in reverse order in function
amdgpu_device_ip_fini_earlyandamdgpu_device_smu_fini_early, to keep consistence with other finish functions.Signed-off-by Jiang Liu <[email protected]>
Quoted 26 lines
--- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +2
-2@@ -3358,7 +3358,7 @@ static void amdgpu_device_smu_fini_early(struct amdgpu_device *adev)if (amdgpu_ip_version (adev,GC_HWIP ,0 ) >IP_VERSION (9 ,0 ,0 ))return ; -amdgpu_for_each_ip_block (adev, ip_block) { +amdgpu_for_each_ip_block_reverse (adev, ip_block) {if (amdgpu_ip_block_state (ip_block) <AMDGPU_IP_STATE_HW )continue ;if (ip_block->version->type ==AMD_IP_BLOCK_TYPE_SMC ) {@@ -3373,7 +3373,7 @@ static int amdgpu_device_ip_fini_early(struct amdgpu_device *adev)int r;struct amdgpu_ip_block *ip_block; -amdgpu_for_each_ip_block (adev, ip_block) { +amdgpu_for_each_ip_block_reverse (adev, ip_block) {if (!ip_block->version->funcs->early_fini)continue ;
Re: [RFC v2 10/15] drm/admgpu: make device state machine work in stack like way
In reply to Mario LimoncielloQuoted 77 lines
2025年1月14日 06:27,Mario Limonciello <[email protected]> 写道:
Quoted 70 lines · Jiang Liu
Make the device state machine work in stack like way to better support suspend/resume by following changes:
1.amdgpu_driver_load_kms()amdgpu_device_init() amdgpu_device_ip_early_init() ip_blocks[i].early_init() ip_blocks[i].status.valid = true amdgpu_device_ip_init() amdgpu_ras_init() ip_blocks[i].sw_init() ip_blocks[i].status.sw = true ip_blocks[i].hw_init() ip_blocks[i].status.hw = true amdgpu_device_ip_late_init() ip_blocks[i].late_init() ip_blocks[i].status.late_initialized = true amdgpu_ras_late_init() ras_blocks[i].ras_late_init() amdgpu_ras_feature_enable_on_boot() 2. amdgpu_pmops_suspend()/amdgpu_pmops_freeze()/amdgpu_pmops_poweroff() amdgpu_device_suspend() amdgpu_ras_early_fini() ras_blocks[i].ras_early_fini() amdgpu_ras_feature_disable() amdgpu_ras_suspend() amdgpu_ras_disable_all_features() +++ ip_blocks[i].early_fini() +++ ip_blocks[i].status.late_initialized = false ip_blocks[i].suspend() 3. amdgpu_pmops_resume()/amdgpu_pmops_thaw()/amdgpu_pmops_restore() amdgpu_device_resume() amdgpu_device_ip_resume() ip_blocks[i].resume() amdgpu_device_ip_late_init() ip_blocks[i].late_init() ip_blocks[i].status.late_initialized = true amdgpu_ras_late_init() ras_blocks[i].ras_late_init() amdgpu_ras_feature_enable_on_boot() amdgpu_ras_resume() amdgpu_ras_enable_all_features() 4. amdgpu_driver_unload_kms() amdgpu_device_fini_hw() amdgpu_ras_early_fini() ras_blocks[i].ras_early_fini() +++ ip_blocks[i].early_fini() +++ ip_blocks[i].status.late_initialized = false ip_blocks[i].hw_fini() ip_blocks[i].status.hw = false 5. amdgpu_driver_release_kms() amdgpu_device_fini_sw() amdgpu_device_ip_fini() ip_blocks[i].sw_fini() ip_blocks[i].status.sw = false --- ip_blocks[i].status.valid = false +++ amdgpu_ras_fini() ip_blocks[i].late_fini()+++ ip_blocks[i].status.valid = false
--- ip_blocks[i].status.late_initialized = false
---amdgpu_ras_fini()
The main changes include:
1) invoke ip_blocks[i].early_fini inamdgpu_pmops_suspend().
2) set ip_blocks[i].status.late_initialized to false after calling callbackearly_fini. We have auditted all usages of thelate_initializedflag and no functional changes found.
3) only set ip_blocks[i].status.valid = false after calling thelate_finicallback.
4) callamdgpu_ras_fini()before invoking ip_blocks[i].late_fini. There's one more task left to analyze GPU reset related state machine transitions.Quoted 1 line
Signed-off-by Jiang Liu <[email protected]>
Quoted 1 line
Ideally I think you should swap the order of patch 10 and 11, what do you think?
I realized this when working patch 11, many changes introduced by patch 10 are changed again by patch 11. But swapping these patches will cause too much rework. How about folding these two patches instead?
Quoted 47 lines
Quoted 47 lines
--- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 22 ++++++++++++++++++++-- 1 file changed, 20 insertions(+), 2 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +18
-1@@ -3449,6 +3449,8 @@ static int amdgpu_device_ip_fini(struct amdgpu_device *adev) adev->ip_blocks[i].status.sw = false; } +amdgpu_ras_fini (adev); +for (i = adev->num_ip_blocks -1 ; i >=0 ; i--) {if (!adev->ip_blocks[i].status.valid)continue ;@@ -3457,8 +3459,6 @@ static int amdgpu_device_ip_fini(struct amdgpu_device *adev) adev->ip_blocks[i].status.valid = false; } -amdgpu_ras_fini (adev); -return 0 ; } @@ -3516,6 +3516,24 @@ static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev) if (amdgpu_dpm_set_df_cstate(adev, DF_CSTATE_DISALLOW)) dev_warn(adev->dev, "Failed to disallow df cstate"); + for (i = adev->num_ip_blocks - 1; i >= 0; i--) { +if (!adev->ip_blocks[i].status.valid) +continue ; +if (!adev->ip_blocks[i].status.late_initialized) +continue ; + +if (adev->ip_blocks[i].version->funcs->early_fini) { + r = adev->ip_blocks[i].version->funcs->early_fini (&adev->ip_blocks[i]); +if (r) { +DRM_ERROR (" of IP block <%s> failed %d\n" , + adev->ip_blocks[i].version->funcs->name, r); +return r; + } + } + + adev->ip_blocks[i].status.late_initialized = false; + } +for (i = adev->num_ip_blocks -1 ; i >=0 ; i--) {if (!adev->ip_blocks[i].status.valid)continue ;
Re: [RFC v2 10/15] drm/admgpu: make device state machine work in stack like way
Quoted 77 lines · Gerry Liu
Quoted 77 lines
2025年1月14日 06:27,Mario Limonciello <[email protected]> 写道:
Quoted 71 lines · Jiang Liu
Make the device state machine work in stack like way to better support suspend/resume by following changes:
1.amdgpu_driver_load_kms()amdgpu_device_init() amdgpu_device_ip_early_init() ip_blocks[i].early_init() ip_blocks[i].status.valid = true amdgpu_device_ip_init() amdgpu_ras_init() ip_blocks[i].sw_init() ip_blocks[i].status.sw = true ip_blocks[i].hw_init() ip_blocks[i].status.hw = true amdgpu_device_ip_late_init() ip_blocks[i].late_init() ip_blocks[i].status.late_initialized = true amdgpu_ras_late_init() ras_blocks[i].ras_late_init() amdgpu_ras_feature_enable_on_boot() 2. amdgpu_pmops_suspend()/amdgpu_pmops_freeze()/amdgpu_pmops_poweroff() amdgpu_device_suspend() amdgpu_ras_early_fini() ras_blocks[i].ras_early_fini() amdgpu_ras_feature_disable() amdgpu_ras_suspend() amdgpu_ras_disable_all_features() +++ ip_blocks[i].early_fini() +++ ip_blocks[i].status.late_initialized = false ip_blocks[i].suspend() 3. amdgpu_pmops_resume()/amdgpu_pmops_thaw()/amdgpu_pmops_restore() amdgpu_device_resume() amdgpu_device_ip_resume() ip_blocks[i].resume() amdgpu_device_ip_late_init() ip_blocks[i].late_init() ip_blocks[i].status.late_initialized = true amdgpu_ras_late_init() ras_blocks[i].ras_late_init() amdgpu_ras_feature_enable_on_boot() amdgpu_ras_resume() amdgpu_ras_enable_all_features() 4. amdgpu_driver_unload_kms() amdgpu_device_fini_hw() amdgpu_ras_early_fini() ras_blocks[i].ras_early_fini() +++ ip_blocks[i].early_fini() +++ ip_blocks[i].status.late_initialized = false ip_blocks[i].hw_fini() ip_blocks[i].status.hw = false 5. amdgpu_driver_release_kms() amdgpu_device_fini_sw() amdgpu_device_ip_fini() ip_blocks[i].sw_fini() ip_blocks[i].status.sw = false --- ip_blocks[i].status.valid = false +++ amdgpu_ras_fini() ip_blocks[i].late_fini()+++ ip_blocks[i].status.valid = false
--- ip_blocks[i].status.late_initialized = false
---amdgpu_ras_fini()
The main changes include:
1) invoke ip_blocks[i].early_fini inamdgpu_pmops_suspend().
2) set ip_blocks[i].status.late_initialized to false after calling callbackearly_fini. We have auditted all usages of thelate_initializedflag and no functional changes found.
3) only set ip_blocks[i].status.valid = false after calling thelate_finicallback.
4) callamdgpu_ras_fini()before invoking ip_blocks[i].late_fini. There's one more task left to analyze GPU reset related state machine transitions.Signed-off-by Jiang Liu <[email protected]>
Quoted 3 lines
Quoted 1 line
Ideally I think you should swap the order of patch 10 and 11, what do you think?
I realized this when working patch 11, many changes introduced by patch 10 are changed again by patch 11. But swapping these patches will cause too much rework. How about folding these two patches instead?
It might be too big of a patch because it changes a lot all at same time.
Re-ordering is painful but leads to more reable patches and less ping-pong of code.
I think you can do something like this (I've done this myself on big patch series):
1) squash the two patches together
git rebase -i HEAD~15
2) Spit it out as a patch file
git format-patch -1 $SHA
3) Check out the commit before this combined one
git checkout -b gerry/rebase $SHA~1
4) Apply the patch using patch -p1 (NOT git am)
patch -p1 < foo.patch
5) Use vscode (specifically) to stage the lines that should go into patch 10.
6) Commit patch 10
7) Stage the lines that go into patch 11.
8) Commit patch 11
9) Note those two commit hashes
10) Go back to your original branch
git checkout gerry/original
11) Run a rebase again, but swap out the "squashed" hash with the two hashes you made.
IE if your two hashes are aaaabbb and bbbccc and it's currently
pick abc123
change it to
pick aaaabbb
pick bbbccc
Then finish the rebase and it should swap them all out for you!
Quoted 50 lines
Quoted 49 lines
Quoted 49 lines
--- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 22 ++++++++++++++++++++-- 1 file changed, 20 insertions(+), 2 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +18
-1@@ -3449,6 +3449,8 @@ static int amdgpu_device_ip_fini(struct amdgpu_device *adev) adev->ip_blocks[i].status.sw = false; } +amdgpu_ras_fini (adev); +for (i = adev->num_ip_blocks -1 ; i >=0 ; i--) {if (!adev->ip_blocks[i].status.valid)continue ;@@ -3457,8 +3459,6 @@ static int amdgpu_device_ip_fini(struct amdgpu_device *adev) adev->ip_blocks[i].status.valid = false; } -amdgpu_ras_fini (adev); -return 0 ; } @@ -3516,6 +3516,24 @@ static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev) if (amdgpu_dpm_set_df_cstate(adev, DF_CSTATE_DISALLOW)) dev_warn(adev->dev, "Failed to disallow df cstate"); + for (i = adev->num_ip_blocks - 1; i >= 0; i--) { +if (!adev->ip_blocks[i].status.valid) +continue ; +if (!adev->ip_blocks[i].status.late_initialized) +continue ; + +if (adev->ip_blocks[i].version->funcs->early_fini) { + r = adev->ip_blocks[i].version->funcs->early_fini (&adev->ip_blocks[i]); +if (r) { +DRM_ERROR (" of IP block <%s> failed %d\n" , + adev->ip_blocks[i].version->funcs->name, r); +return r; + } + } + + adev->ip_blocks[i].status.late_initialized = false; + } +for (i = adev->num_ip_blocks -1 ; i >=0 ; i--) {if (!adev->ip_blocks[i].status.valid)continue ;
Re: [RFC v2 03/15] drm/amdgpu: free all resources on error recovery path of amdgpu_ras_init()
In reply to Jiang LiuQuoted 46 lines · Jiang Liu
Free all allocated resources on error recovery path in function
amdgpu_ras_init().Signed-off-by Jiang Liu <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c | 19 ++++++++++++++----- 1 file changed, 14 insertions(+), 5 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +4
-3@@ -3864,6 +3864,7 @@ static void amdgpu_ras_init_reserved_vram_size(struct amdgpu_device *adev)int amdgpu_ras_init (struct amdgpu_device *adev) {struct amdgpu_ras *con =amdgpu_ras_get_context (adev); +struct amdgpu_ras_block_list *ras_node, *tmp;int r;if (con)@@ -3953,20 +3954,20 @@ int amdgpu_ras_init(struct amdgpu_device *adev)* to handle fatal error */ r =amdgpu_nbio_ras_sw_init (adev);if (r) -return r; + goto release_con;if (adev->nbio.ras && adev->nbio.ras->init_ras_controller_interrupt) { r = adev->nbio.ras->init_ras_controller_interrupt (adev);if (r) - goto release_con; + goto free_blocks; }if (adev->nbio.ras && adev->nbio.ras->init_ras_err_event_athub_interrupt) { r = adev->nbio.ras->init_ras_err_event_athub_interrupt (adev);if (r) - goto release_con; + goto free_blocks; }/* Packed socket_id to ras feature mask bits[31:29] */
drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c int amdgpu_ras_init(struct amdgpu_device *adev) · line 3986
Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +1
-1@@ -3982,7 +3983,7 @@ int amdgpu_ras_init(struct amdgpu_device *adev)if (amdgpu_ras_fs_init (adev)) { r = -EINVAL ; - goto release_con; + goto free_blocks;
Reviewing this shows there is a mistake in amdgpu_ras_fs_init(). If sysfs fails to init there it still returns 0.
Please modify amdgpu_ras_fs_init() to 'return r' and then use that value. IE:
r = amdgpu_ras_fs_init(adev);
if (r)
goto free_blocks;
Quoted 27 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +9
-1}if (amdgpu_ras_aca_is_supported (adev)) {@@ -3991,7 +3992,7 @@ int amdgpu_ras_init(struct amdgpu_device *adev)else r =amdgpu_mca_init (adev);if (r) - goto release_con; + goto clear_ras_fs; } dev_info(adev->dev, "RAS INFO: ras initialized successfully, "@@ -3999,6 +4000,14 @@ int amdgpu_ras_init(struct amdgpu_device *adev) adev->ras_hw_enabled, adev->ras_enabled);return 0 ; + +clear_ras_fs: +amdgpu_ras_fs_fini (adev); +free_blocks: +list_for_each_entry_safe (ras_node, tmp, &adev->ras_list, node) { +list_del (&ras_node->node); +kfree (ras_node); + } release_con:amdgpu_ras_set_context (adev,NULL );kfree (con);
Re: [RFC v2 05/15] drm/amdgpu: enhance amdgpu_ras_block_late_fini()
In reply to Jiang LiuQuoted 16 lines · Jiang Liu
Enhance
amdgpu_ras_block_late_fini()to revert what has been done byamdgpu_ras_block_late_init(), and fix a possible resource leakage in functionamdgpu_ras_block_late_init().Signed-off-by Jiang Liu <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c | 16 ++++++++++------ 1 file changed, 10 insertions(+), 6 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +0
-0@@ -4104,13 +4104,13 @@ int amdgpu_ras_block_late_init(struct amdgpu_device *adev, ras_obj->hw_ops->query_ras_error_status)) { r =amdgpu_ras_sysfs_create (adev, ras_block);
drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c int amdgpu_ras_block_late_init(struct amdgpu_device *adev, · line 4113
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +2
-2if (r) - goto interrupt; + goto cleanup;/* Those are the cached values at init. */ query_info =kzalloc (sizeof (*query_info),GFP_KERNEL );if (!query_info) -return -ENOMEM ; + goto cleanup;
AFAICT you still need to set "r = -ENOMEM" here for this error flow.
Quoted 33 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +8
-4memcpy (&query_info->head, ras_block,sizeof (struct ras_common_if ));if (amdgpu_ras_query_error_count (adev, &ce_count, &ue_count, query_info) ==0 ) {@@ -4123,11 +4123,8 @@ int amdgpu_ras_block_late_init(struct amdgpu_device *adev,return 0 ; -interrupt: -if (ras_obj->ras_cb) -amdgpu_ras_interrupt_remove_handler (adev, ras_block); cleanup: -amdgpu_ras_feature_enable (adev, ras_block,0 ); +amdgpu_ras_block_late_fini (adev, ras_block);return r; }@@ -4142,9 +4139,16 @@ void amdgpu_ras_block_late_fini(struct amdgpu_device *adev, struct ras_common_if *ras_block) {struct amdgpu_ras_block_object *ras_obj; +if (!ras_block)return ; +amdgpu_ras_feature_enable (adev, ras_block,0 ); + +/* in resume/reset phase, no need to delete ras fs node */ +if (adev->in_suspend ||amdgpu_in_reset (adev)) +return ; +amdgpu_ras_sysfs_remove (adev, ras_block); ras_obj = container_of(ras_block, struct amdgpu_ras_block_object, ras_comm);
Re: [RFC v2 06/15] drm/amdgpu: enhance amdgpu_ras_pre_fini() to better support SR
In reply to Jiang LiuQuoted 88 lines · Jiang Liu
Enhance
amdgpu_ras_pre_fini()to better support suspend/resume by:
1) fix possible resource leakage.amdgpu_release_ras_context()only kfree(con) but doesn't release resources associated with the con object.
2) callamdgpu_ras_pre_fini()inamdgpu_device_suspend()to undo what has been done byamdgpu_ras_late_init(), becauseamdgpu_device_resume()will invokeamdgpu_ras_late_init()on resume.
3) moveamdgpu_ras_recovery_fini()fromamdgpu_ras_pre_fini()toamdgpu_ras_fini()
4) move calling ofobj->ras_fini()fromamdgpu_ras_fini()toamdgpu_ras_pre_fini().Signed-off-by Jiang Liu <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 6 ++- drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c | 44 +++++++++++++--------- 2 files changed, 31 insertions(+), 19 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +4
-2@@ -4613,6 +4613,8 @@ int amdgpu_device_init(struct amdgpu_device *adev,return 0 ; release_ras_con: +amdgpu_ras_pre_fini (adev); +amdgpu_ras_fini (adev);if (amdgpu_sriov_vf (adev))amdgpu_virt_release_full_gpu (adev, true);@@ -4627,8 +4629,6 @@ int amdgpu_device_init(struct amdgpu_device *adev, adev->virt.ops =NULL ; r = -EAGAIN ; } -amdgpu_release_ras_context (adev); - failed:amdgpu_vf_error_trans_all (adev);@@ -4921,6 +4921,8 @@ int amdgpu_device_suspend(struct drm_device *dev, bool notify_clients)cancel_delayed_work_sync (&adev->delayed_init_work); +/* disable ras feature must before hw fini */ +amdgpu_ras_pre_fini (adev);amdgpu_ras_suspend (adev);amdgpu_device_ip_suspend_phase1 (adev);drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +20
-4@@ -4270,42 +4270,49 @@ int amdgpu_ras_late_init(struct amdgpu_device *adev)int amdgpu_ras_pre_fini (struct amdgpu_device *adev) {struct amdgpu_ras *con =amdgpu_ras_get_context (adev); +struct amdgpu_ras_block_list *node, *tmp; +struct amdgpu_ras_block_object *obj; -if (!adev->ras_enabled || !con) -return 0 ; +if (amdgpu_sriov_vf (adev) && !amdgpu_sriov_ras_telemetry_en (adev)) + goto disable; +list_for_each_entry_safe (node, tmp, &adev->ras_list, node) { + obj = node->ras_obj; +if (!obj) +continue ; + +if (!amdgpu_ras_is_supported (adev, obj->ras_comm.block)) +continue ; + +if (obj->ras_fini) + obj->ras_fini (adev, &obj->ras_comm); +else +amdgpu_ras_block_late_fini_default (adev, &obj->ras_comm); + } +disable:/* Need disable ras on all IPs here before ip [hw/sw]fini */ -if (AMDGPU_RAS_GET_FEATURES (con->features)) +if (con &&AMDGPU_RAS_GET_FEATURES (con->features))amdgpu_ras_disable_all_features (adev,0 ); -amdgpu_ras_recovery_fini (adev); +return 0 ; }
drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c int amdgpu_ras_late_init(struct amdgpu_device *adev) · line 4307
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +1
-2int amdgpu_ras_fini (struct amdgpu_device *adev) {struct amdgpu_ras_block_list *ras_node, *tmp; -struct amdgpu_ras_block_object *obj =NULL ;struct amdgpu_ras *con =amdgpu_ras_get_context (adev);if (!adev->ras_enabled || !con) -return 0 ; + goto out_free_context;
This block doesn't make sense for two reasons.
1) Why would you have a context if adev->ras_enabled was false?
2) If amdgpu_ras_get_context() returned NULL, then jumping to out_free_context and checking con again doesn't make sense - you already knew it was NULL.
Quoted 34 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +6
-11list_for_each_entry_safe (ras_node, tmp, &adev->ras_list, node) { -if (ras_node->ras_obj) { - obj = ras_node->ras_obj; -if (amdgpu_ras_is_supported (adev, obj->ras_comm.block) && - obj->ras_fini) - obj->ras_fini (adev, &obj->ras_comm); -else -amdgpu_ras_block_late_fini_default (adev, &obj->ras_comm); - } -/* Clear ras blocks from ras_list and free ras block list node */ list_del (&ras_node->node);kfree (ras_node); } +amdgpu_ras_recovery_fini (adev);amdgpu_ras_fs_fini (adev);amdgpu_ras_interrupt_remove_all (adev);@@ -4323,8 +4330,11 @@ int amdgpu_ras_fini(struct amdgpu_device *adev)cancel_delayed_work_sync (&con->ras_counte_delay_work); -amdgpu_ras_set_context (adev,NULL ); -kfree (con); +out_free_context: +if (con) { +amdgpu_ras_set_context (adev,NULL ); +kfree (con); + }return 0 ; }
Re: [RFC v2 07/15] drm/admgpu: rename amdgpu_ras_pre_fini() to amdgpu_ras_early_fini()
In reply to Jiang LiuQuoted 2 lines · Jiang Liu
Rename
amdgpu_ras_pre_fini()toamdgpu_ras_early_fini(), to keep same style with other code.
Besides amdgpu_ras_pre_fini() -> amdgpu_ras_early_fini() you also changed amdgpu_ras_block_late_fini() -> amdgpu_ras_early_fini().
Is that really intended? If so; the commit message needs to be amended.
Quoted 279 lines
Signed-off-by Jiang Liu <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 6 +++--- drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_jpeg.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_nbio.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c | 16 ++++++++-------- drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h | 6 +++--- drivers/gpu/drm/amd/amdgpu/amdgpu_sdma.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_vcn.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_xgmi.c | 2 +- drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c | 2 +- drivers/gpu/drm/amd/amdgpu/jpeg_v4_0_3.c | 2 +- drivers/gpu/drm/amd/amdgpu/mmhub_v1_8.c | 2 +- drivers/gpu/drm/amd/amdgpu/vcn_v4_0_3.c | 2 +- 14 files changed, 25 insertions(+), 25 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +3
-3@@ -4613,7 +4613,7 @@ int amdgpu_device_init(struct amdgpu_device *adev,return 0 ; release_ras_con: -amdgpu_ras_pre_fini (adev); +amdgpu_ras_early_fini (adev);amdgpu_ras_fini (adev);if (amdgpu_sriov_vf (adev))amdgpu_virt_release_full_gpu (adev, true);@@ -4705,7 +4705,7 @@ void amdgpu_device_fini_hw(struct amdgpu_device *adev)amdgpu_xcp_cfg_sysfs_fini (adev);/* disable ras feature must before hw fini */ -amdgpu_ras_pre_fini (adev); +amdgpu_ras_early_fini (adev);amdgpu_ttm_set_buffer_funcs_status (adev, false);@@ -4922,7 +4922,7 @@ int amdgpu_device_suspend(struct drm_device *dev, bool notify_clients)cancel_delayed_work_sync (&adev->delayed_init_work);/* disable ras feature must before hw fini */ -amdgpu_ras_pre_fini (adev); +amdgpu_ras_early_fini (adev);amdgpu_ras_suspend (adev);amdgpu_device_ip_suspend_phase1 (adev);drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +1
-1@@ -918,7 +918,7 @@ int amdgpu_gfx_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *rreturn 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }drivers/gpu/drm/amd/amdgpu/amdgpu_jpeg.c +1
-1@@ -300,7 +300,7 @@ int amdgpu_jpeg_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *return 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }drivers/gpu/drm/amd/amdgpu/amdgpu_nbio.c +1
-1@@ -71,6 +71,6 @@ int amdgpu_nbio_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *return 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +8
-8@@ -4124,7 +4124,7 @@ int amdgpu_ras_block_late_init(struct amdgpu_device *adev,return 0 ; cleanup: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }@@ -4135,7 +4135,7 @@ static int amdgpu_ras_block_late_init_default(struct amdgpu_device *adev, }/* helper function to remove ras fs node and interrupt handler */ -void amdgpu_ras_block_late_fini (struct amdgpu_device *adev, +void amdgpu_ras_block_early_fini (struct amdgpu_device *adev,struct ras_common_if *ras_block) {struct amdgpu_ras_block_object *ras_obj;@@ -4156,10 +4156,10 @@ void amdgpu_ras_block_late_fini(struct amdgpu_device *adev,amdgpu_ras_interrupt_remove_handler (adev, ras_block); } -static void amdgpu_ras_block_late_fini_default (struct amdgpu_device *adev, +static void amdgpu_ras_block_early_fini_default (struct amdgpu_device *adev,struct ras_common_if *ras_block) { -return amdgpu_ras_block_late_fini (adev, ras_block); +return amdgpu_ras_block_early_fini (adev, ras_block); }/* do some init work after IP late init as dependence. @@ -4267,7 +4267,7 @@ int amdgpu_ras_late_init(struct amdgpu_device *adev) }/* do some fini work before IP fini as dependence */ -int amdgpu_ras_pre_fini (struct amdgpu_device *adev) +int amdgpu_ras_early_fini (struct amdgpu_device *adev) {struct amdgpu_ras *con =amdgpu_ras_get_context (adev);struct amdgpu_ras_block_list *node, *tmp;@@ -4284,10 +4284,10 @@ int amdgpu_ras_pre_fini(struct amdgpu_device *adev)if (!amdgpu_ras_is_supported (adev, obj->ras_comm.block))continue ; -if (obj->ras_fini) - obj->ras_fini (adev, &obj->ras_comm); +if (obj->ras_early_fini) + obj->ras_early_fini (adev, &obj->ras_comm);else -amdgpu_ras_block_late_fini_default (adev, &obj->ras_comm); +amdgpu_ras_block_early_fini_default (adev, &obj->ras_comm); } disable:drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h +3
-3@@ -711,7 +711,7 @@ struct amdgpu_ras_block_object {int (*ras_block_match)(struct amdgpu_ras_block_object *block_obj,enum amdgpu_ras_block block,uint32_t sub_block_index);int (*ras_late_init)(struct amdgpu_device *adev,struct ras_common_if *ras_block); -void (*ras_fini)(struct amdgpu_device *adev,struct ras_common_if *ras_block); +void (*ras_early_fini)(struct amdgpu_device *adev,struct ras_common_if *ras_block);ras_ih_cb ras_cb;const struct amdgpu_ras_block_hw_ops *hw_ops; };@@ -825,13 +825,13 @@ amdgpu_ras_error_to_ta(enum amdgpu_ras_error_type error) {/* called in ip_init and ip_fini */ int amdgpu_ras_init (struct amdgpu_device *adev);int amdgpu_ras_late_init (struct amdgpu_device *adev); +int amdgpu_ras_early_fini (struct amdgpu_device *adev);int amdgpu_ras_fini (struct amdgpu_device *adev); -int amdgpu_ras_pre_fini (struct amdgpu_device *adev);int amdgpu_ras_block_late_init (struct amdgpu_device *adev,struct ras_common_if *ras_block); -void amdgpu_ras_block_late_fini (struct amdgpu_device *adev, +void amdgpu_ras_block_early_fini (struct amdgpu_device *adev,struct ras_common_if *ras_block); int amdgpu_ras_feature_enable(struct amdgpu_device *adev,drivers/gpu/drm/amd/amdgpu/amdgpu_sdma.c +1
-1@@ -116,7 +116,7 @@ int amdgpu_sdma_ras_late_init(struct amdgpu_device *adev,return 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c +1
-1@@ -336,7 +336,7 @@ int amdgpu_umc_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *rreturn 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }drivers/gpu/drm/amd/amdgpu/amdgpu_vcn.c +1
-1@@ -1248,7 +1248,7 @@ int amdgpu_vcn_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *rreturn 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }drivers/gpu/drm/amd/amdgpu/amdgpu_xgmi.c +1
-1@@ -1174,7 +1174,7 @@ static int amdgpu_xgmi_ras_late_init(struct amdgpu_device *adev, struct ras_commreturn 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c +1
-1@@ -5080,7 +5080,7 @@ static int gfx_v9_4_3_ras_late_init(struct amdgpu_device *adev, struct ras_commoreturn 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }drivers/gpu/drm/amd/amdgpu/jpeg_v4_0_3.c +1
-1@@ -1311,7 +1311,7 @@ static int jpeg_v4_0_3_ras_late_init(struct amdgpu_device *adev, struct ras_commreturn 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }drivers/gpu/drm/amd/amdgpu/mmhub_v1_8.c +1
-1@@ -789,7 +789,7 @@ static int mmhub_v1_8_ras_late_init(struct amdgpu_device *adev, struct ras_commoreturn 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }drivers/gpu/drm/amd/amdgpu/vcn_v4_0_3.c +1
-1@@ -1996,7 +1996,7 @@ static int vcn_v4_0_3_ras_late_init(struct amdgpu_device *adev, struct ras_commoreturn 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }
Re: [RFC v2 09/15] drm/amdgpu_dm: enhance amdgpu_dm_early_fini() for PM ops
In reply to Jiang LiuQuoted 13 lines · Jiang Liu
Enhance
amdgpu_dm_early_fini()so it can be called in power management operations.Signed-off-by Jiang Liu <[email protected]>--- drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c | 3 +++ 1 file changed, 3 insertions(+)drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c +0
-0
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c static int amdgpu_dm_early_fini(struct amdgpu_ip_block *ip_block) · line 2180
Quoted 6 lines
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c +3
-0@@ -2175,6 +2175,9 @@ static int amdgpu_dm_early_fini(struct amdgpu_ip_block *ip_block) {struct amdgpu_device *adev = ip_block->adev; +if (adev->in_s0ix || adev->in_s3 || adev->in_s4 || adev->in_suspend) +return 0 ; +
I'm a bit confused how this has even happened over suspend. The call stack I see is:
amdgpu_pci_remove() ->amdgpu_driver_unload_kms() ->->amdgpu_device_fini_hw()
->->-> amdgpu_device_ip_fini_early()
->->->-> adev->ip_blocks[i].version->funcs->early_fini()
Is there another call stack I'm missing?
Quoted 3 lines
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c +0
-0amdgpu_dm_audio_fini (adev);return 0 ;
RE: [RFC v2 01/15] drm/amdgpu: add helper functions to track status for ras manager
In reply to Jiang Liu[AMD Official Use Only - AMD Internal Distribution Only]
-----Original Message-----
From: amd-gfx <[email protected]> On Behalf Of Jiang Liu
Sent: Monday, January 13, 2025 09:42
To: Deucher, Alexander <[email protected]>; Koenig, Christian <[email protected]>; Pan, Xinhui <[email protected]>; [email protected]; [email protected]; Khatri, Sunil <[email protected]>; Lazar, Lijo <[email protected]>; Zhang, Hawking <[email protected]>; Limonciello, Mario <[email protected]>; Chen, Xiaogang <[email protected]>; Russell, Kent <[email protected]>; [email protected]; [email protected]
Cc: Jiang Liu <[email protected]>
Subject: [RFC v2 01/15] drm/amdgpu: add helper functions to track status for ras manager
Add helper functions to track status for ras manager and ip blocks.
--- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 38 +++++++++++++++++++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c | 37 ++++++++++++++++++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h | 10 +++++++ 3 files changed, 85 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +37 -0
@@ -377,12 +377,28 @@ int amdgpu_ip_block_resume(struct amdgpu_ip_block *ip_block);#define AMDGPU_MAX_IP_NUM 16 +enum amdgpu_marker { +// Markers for IRQs, used for both ip blocks and ras blocks. +AMDGPU_MARKER_IRQ0 =32 , +AMDGPU_MARKER_IRQ1 , +AMDGPU_MARKER_IRQ2 , +AMDGPU_MARKER_IRQ3 , +AMDGPU_MARKER_IRQ4 , +AMDGPU_MARKER_IRQ5 , +AMDGPU_MARKER_IRQ6 , +AMDGPU_MARKER_IRQ7 , +AMDGPU_MARKER_IRQ_MAX =63 , +}; + +#define AMDGPU_MARKER_IRQ (idx) (AMDGPU_MARKER_IRQ0 + (idx)) +struct amdgpu_ip_block_status {bool valid;bool sw;bool hw;bool late_initialized;bool hang; +uint64_t markers; };struct amdgpu_ip_block_version {
@@ -410,6 +426,28 @@ amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev, int amdgpu_device_ip_block_add(struct amdgpu_device *adev, const struct amdgpu_ip_block_version *ip_block_version); +static inline void amdgpu_ip_block_set_marker (struct amdgpu_ip_block *ip_block, +enum amdgpu_marker marker) +{ +WARN_ON (marker >63 ); +WARN_ON (ip_block->status.markers & (0x1ull << marker)); + ip_block->status.markers |=0x1ull << (int )marker; } + +static inline bool amdgpu_ip_block_test_and_clear_marker (struct amdgpu_ip_block *ip_block, +enum amdgpu_marker marker) +{ +bool set = false; +uint64_t value =0x1ull << (int )marker; + +if ((ip_block->status.markers & value) !=0 ) { + ip_block->status.markers &= ~value; + set = true; + } + +return set; +} +/* * BIOS. */
drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +17 -0
@@ -5207,3 +5207,40 @@ bool amdgpu_ras_is_rma(struct amdgpu_device *adev)return con->is_rma; } + +bool amdgpu_ras_test_marker (struct amdgpu_device *adev, +struct ras_common_if *head,int marker) { +struct ras_manager *obj =amdgpu_ras_find_obj (adev, head); + +if (obj && obj->markers & (0x1ull << marker)) +return true; + +return false; +} + +void amdgpu_ras_set_marker (struct amdgpu_device *adev, +struct ras_common_if *head,int marker) { +struct ras_manager *obj =amdgpu_ras_find_obj (adev, head); + +WARN_ON (marker >63 ); +WARN_ON (obj->markers & (0x1ull << marker));
[kevin]:
It's best to check for null pointers before accessing member variables Best Regards,
Kevin
+ if (obj)
+ obj->markers |= 0x1ull << marker;
+}
+
+bool amdgpu_ras_test_and_clear_marker(struct amdgpu_device *adev,
+ struct ras_common_if *head, int marker) {
+ bool set = false; + uint64_t value = 0x1ull << marker; + struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
+
+ if (obj && (obj->markers & value) != 0) {
+ obj->markers &= ~value;
+ set = true;
+ }
+
+ return set;
+}
drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h +10 -0
@@ -634,6 +634,8 @@ struct ras_manager {struct ras_common_if head;/* reference count */ int use; +/* Flags for status tracking */ +uint64_t markers;/* ras block link */ struct list_head node;/* the device */
@@ -977,4 +979,12 @@ void amdgpu_ras_event_log_print(struct amdgpu_device *adev, u64 event_id, const char *fmt, ...);bool amdgpu_ras_is_rma (struct amdgpu_device *adev); + +bool amdgpu_ras_test_marker (struct amdgpu_device *adev, +struct ras_common_if *head,int marker);void +amdgpu_ras_set_marker (struct amdgpu_device *adev, +struct ras_common_if *head,int marker);bool +amdgpu_ras_test_and_clear_marker (struct amdgpu_device *adev, +struct ras_common_if *head, +int marker);#endif
--
2.43.5
RE: [RFC v2 07/15] drm/admgpu: rename amdgpu_ras_pre_fini() to amdgpu_ras_early_fini()
In reply to Jiang Liu[AMD Official Use Only - AMD Internal Distribution Only]
-----Original Message-----
From: amd-gfx <[email protected]> On Behalf Of Jiang Liu
Sent: Monday, January 13, 2025 09:42
To: Deucher, Alexander <[email protected]>; Koenig, Christian <[email protected]>; Pan, Xinhui <[email protected]>; [email protected]; [email protected]; Khatri, Sunil <[email protected]>; Lazar, Lijo <[email protected]>; Zhang, Hawking <[email protected]>; Limonciello, Mario <[email protected]>; Chen, Xiaogang <[email protected]>; Russell, Kent <[email protected]>; [email protected]; [email protected]
Cc: Jiang Liu <[email protected]>
Subject: [RFC v2 07/15] drm/admgpu: rename amdgpu_ras_pre_fini() to amdgpu_ras_early_fini()
Rename amdgpu_ras_pre_fini() to amdgpu_ras_early_fini(), to keep same style with other code.
--- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 6 +++--- drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_jpeg.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_nbio.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c | 16 ++++++++-------- drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h | 6 +++--- drivers/gpu/drm/amd/amdgpu/amdgpu_sdma.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_vcn.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_xgmi.c | 2 +- drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c | 2 +- drivers/gpu/drm/amd/amdgpu/jpeg_v4_0_3.c | 2 +- drivers/gpu/drm/amd/amdgpu/mmhub_v1_8.c | 2 +- drivers/gpu/drm/amd/amdgpu/vcn_v4_0_3.c | 2 +- 14 files changed, 25 insertions(+), 25 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -1
@@ -4613,7 +4613,7 @@ int amdgpu_device_init(struct amdgpu_device *adev,return 0 ; release_ras_con: -amdgpu_ras_pre_fini (adev); +amdgpu_ras_early_fini (adev);amdgpu_ras_fini (adev);if (amdgpu_sriov_vf (adev)) amdgpu_virt_release_full_gpu(adev, true); @@ -4705,7 +4705,7 @@ void amdgpu_device_fini_hw(struct amdgpu_device *adev)
amdgpu_xcp_cfg_sysfs_fini(adev);
/* disable ras feature must before hw fini */- amdgpu_ras_pre_fini(adev);
+ amdgpu_ras_early_fini(adev);
amdgpu_ttm_set_buffer_funcs_status(adev, false);
@@ -4922,7 +4922,7 @@ int amdgpu_device_suspend(struct drm_device *dev, bool notify_clients)
cancel_delayed_work_sync(&adev->delayed_init_work);
/* disable ras feature must before hw fini */
- amdgpu_ras_pre_fini(adev);
+ amdgpu_ras_early_fini(adev);
amdgpu_ras_suspend(adev);
amdgpu_device_ip_suspend_phase1(adev);drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +0 -0
@@ -918,7 +918,7 @@ int amdgpu_gfx_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *rreturn 0 ; late_fini:
[kevin]:
It's better to rename above label name together.
Best Regards,
Kevin
- amdgpu_ras_block_late_fini(adev, ras_block);
+ amdgpu_ras_block_early_fini(adev, ras_block);
return r;}
drivers/gpu/drm/amd/amdgpu/amdgpu_jpeg.c +1 -1
@@ -300,7 +300,7 @@ int amdgpu_jpeg_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *return 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }
drivers/gpu/drm/amd/amdgpu/amdgpu_nbio.c +1 -1
@@ -71,6 +71,6 @@ int amdgpu_nbio_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *return 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }
drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +2 -1
@@ -4124,7 +4124,7 @@ int amdgpu_ras_block_late_init(struct amdgpu_device *adev,return 0 ; cleanup: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }
@@ -4135,7 +4135,7 @@ static int amdgpu_ras_block_late_init_default(struct amdgpu_device *adev, } /* helper function to remove ras fs node and interrupt handler */ -void amdgpu_ras_block_late_fini(struct amdgpu_device *adev, +void amdgpu_ras_block_early_fini(struct amdgpu_device *adev, struct ras_common_if *ras_block) { struct amdgpu_ras_block_object *ras_obj; @@ -4156,10 +4156,10 @@ void amdgpu_ras_block_late_fini(struct amdgpu_device *adev, amdgpu_ras_interrupt_remove_handler(adev, ras_block); }
-static void amdgpu_ras_block_late_fini_default(struct amdgpu_device *adev, +static void amdgpu_ras_block_early_fini_default(struct amdgpu_device +*adev,
struct ras_common_if *ras_block)
{
- return amdgpu_ras_block_late_fini(adev, ras_block);
+ return amdgpu_ras_block_early_fini(adev, ras_block);
}
/* do some init work after IP late init as dependence.
@@ -4267,7 +4267,7 @@ int amdgpu_ras_late_init(struct amdgpu_device *adev) }
/* do some fini work before IP fini as dependence */ -int amdgpu_ras_pre_fini(struct amdgpu_device *adev)
+int amdgpu_ras_early_fini(struct amdgpu_device *adev)
{
struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
struct amdgpu_ras_block_list *node, *tmp; @@ -4284,10 +4284,10 @@ int amdgpu_ras_pre_fini(struct amdgpu_device *adev)
if (!amdgpu_ras_is_supported(adev, obj->ras_comm.block))
continue;- if (obj->ras_fini)
- obj->ras_fini(adev, &obj->ras_comm);
+ if (obj->ras_early_fini)
+ obj->ras_early_fini(adev, &obj->ras_comm);
else
- amdgpu_ras_block_late_fini_default(adev, &obj->ras_comm);
+ amdgpu_ras_block_early_fini_default(adev, &obj->ras_comm);
}disable:
drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h +2 -1
@@ -711,7 +711,7 @@ struct amdgpu_ras_block_object {int (*ras_block_match)(struct amdgpu_ras_block_object *block_obj,enum amdgpu_ras_block block,uint32_t sub_block_index);int (*ras_late_init)(struct amdgpu_device *adev,struct ras_common_if *ras_block); -void (*ras_fini)(struct amdgpu_device *adev,struct ras_common_if *ras_block); +void (*ras_early_fini)(struct amdgpu_device *adev,struct +ras_common_if *ras_block);ras_ih_cb ras_cb; const struct amdgpu_ras_block_hw_ops *hw_ops; }; @@ -825,13 +825,13 @@ amdgpu_ras_error_to_ta(enum amdgpu_ras_error_type error) {
/* called in ip_init and ip_fini */
int amdgpu_ras_init(struct amdgpu_device *adev); int amdgpu_ras_late_init(struct amdgpu_device *adev);
+int amdgpu_ras_early_fini(struct amdgpu_device *adev);
int amdgpu_ras_fini(struct amdgpu_device *adev); -int amdgpu_ras_pre_fini(struct amdgpu_device *adev);
int amdgpu_ras_block_late_init(struct amdgpu_device *adev,
struct ras_common_if *ras_block);
-void amdgpu_ras_block_late_fini(struct amdgpu_device *adev, +void amdgpu_ras_block_early_fini(struct amdgpu_device *adev,
struct ras_common_if *ras_block);
int amdgpu_ras_feature_enable(struct amdgpu_device *adev, diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_sdma.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_sdma.c
index 174badca27e7..1a1834e47b50 100644
drivers/gpu/drm/amd/amdgpu/amdgpu_sdma.c +1 -1
@@ -116,7 +116,7 @@ int amdgpu_sdma_ras_late_init(struct amdgpu_device *adev,return 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }
drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c +1 -1
@@ -336,7 +336,7 @@ int amdgpu_umc_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *rreturn 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }
drivers/gpu/drm/amd/amdgpu/amdgpu_vcn.c +1 -1
@@ -1248,7 +1248,7 @@ int amdgpu_vcn_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *rreturn 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }
drivers/gpu/drm/amd/amdgpu/amdgpu_xgmi.c +1 -1
@@ -1174,7 +1174,7 @@ static int amdgpu_xgmi_ras_late_init(struct amdgpu_device *adev, struct ras_commreturn 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }
drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c +1 -1
@@ -5080,7 +5080,7 @@ static int gfx_v9_4_3_ras_late_init(struct amdgpu_device *adev, struct ras_commoreturn 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }
drivers/gpu/drm/amd/amdgpu/jpeg_v4_0_3.c +1 -1
@@ -1311,7 +1311,7 @@ static int jpeg_v4_0_3_ras_late_init(struct amdgpu_device *adev, struct ras_commreturn 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }
drivers/gpu/drm/amd/amdgpu/mmhub_v1_8.c +1 -1
@@ -789,7 +789,7 @@ static int mmhub_v1_8_ras_late_init(struct amdgpu_device *adev, struct ras_commoreturn 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }
drivers/gpu/drm/amd/amdgpu/vcn_v4_0_3.c +1 -1
@@ -1996,7 +1996,7 @@ static int vcn_v4_0_3_ras_late_init(struct amdgpu_device *adev, struct ras_commoreturn 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }
--
2.43.5
Re: [RFC v2 01/15] drm/amdgpu: add helper functions to track status for ras manager
In reply to Jiang LiuQuoted 31 lines · Jiang Liu
Add helper functions to track status for ras manager and ip blocks.
Signed-off-by Jiang Liu <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 38 +++++++++++++++++++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c | 37 ++++++++++++++++++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h | 10 +++++++ 3 files changed, 85 insertions(+)drivers/gpu/drm/amd/amdgpu/amdgpu.h +14
-0@@ -377,12 +377,28 @@ int amdgpu_ip_block_resume(struct amdgpu_ip_block *ip_block);#define AMDGPU_MAX_IP_NUM 16 +enum amdgpu_marker { +// Markers for IRQs, used for both ip blocks and ras blocks. +AMDGPU_MARKER_IRQ0 =32 , +AMDGPU_MARKER_IRQ1 , +AMDGPU_MARKER_IRQ2 , +AMDGPU_MARKER_IRQ3 , +AMDGPU_MARKER_IRQ4 , +AMDGPU_MARKER_IRQ5 , +AMDGPU_MARKER_IRQ6 , +AMDGPU_MARKER_IRQ7 , +AMDGPU_MARKER_IRQ_MAX =63 , +}; + +#define AMDGPU_MARKER_IRQ (idx) (AMDGPU_MARKER_IRQ0 + (idx))
drivers/gpu/drm/amd/amdgpu/amdgpu.h int amdgpu_ip_block_resume(struct amdgpu_ip_block *ip_block); · line 403
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu.h +2
-0+struct amdgpu_ip_block_status {bool valid;bool sw;bool hw;bool late_initialized;bool hang; +uint64_t markers; };
This fine grained levels maintained at IP layer doesn't look like a proper solution. It's either IP or RAS block has the required IRQs enabled or disabled. Unwinding them needs to be tracked at IRQ object layer and not here.
Thanks,
Lijo
Quoted 101 lines
drivers/gpu/drm/amd/amdgpu/amdgpu.h +22
-0struct amdgpu_ip_block_version {@@ -410,6 +426,28 @@ amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev,int amdgpu_device_ip_block_add (struct amdgpu_device *adev,const struct amdgpu_ip_block_version *ip_block_version); +static inline void amdgpu_ip_block_set_marker (struct amdgpu_ip_block *ip_block, +enum amdgpu_marker marker) +{ +WARN_ON (marker >63 ); +WARN_ON (ip_block->status.markers & (0x1ull << marker)); + ip_block->status.markers |=0x1ull << (int )marker; +} + +static inline bool amdgpu_ip_block_test_and_clear_marker (struct amdgpu_ip_block *ip_block, +enum amdgpu_marker marker) +{ +bool set = false; +uint64_t value =0x1ull << (int )marker; + +if ((ip_block->status.markers & value) !=0 ) { + ip_block->status.markers &= ~value; + set = true; + } + +return set; +} +/* * BIOS. */ drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +37
-0@@ -5207,3 +5207,40 @@ bool amdgpu_ras_is_rma(struct amdgpu_device *adev)return con->is_rma; } + +bool amdgpu_ras_test_marker (struct amdgpu_device *adev, +struct ras_common_if *head,int marker) +{ +struct ras_manager *obj =amdgpu_ras_find_obj (adev, head); + +if (obj && obj->markers & (0x1ull << marker)) +return true; + +return false; +} + +void amdgpu_ras_set_marker (struct amdgpu_device *adev, +struct ras_common_if *head,int marker) +{ +struct ras_manager *obj =amdgpu_ras_find_obj (adev, head); + +WARN_ON (marker >63 ); +WARN_ON (obj->markers & (0x1ull << marker)); +if (obj) + obj->markers |=0x1ull << marker; +} + +bool amdgpu_ras_test_and_clear_marker (struct amdgpu_device *adev, +struct ras_common_if *head,int marker) +{ +bool set = false; +uint64_t value =0x1ull << marker; +struct ras_manager *obj =amdgpu_ras_find_obj (adev, head); + +if (obj && (obj->markers & value) !=0 ) { + obj->markers &= ~value; + set = true; + } + +return set; +}drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h +10
-0@@ -634,6 +634,8 @@ struct ras_manager {struct ras_common_if head;/* reference count */ int use; +/* Flags for status tracking */ +uint64_t markers;/* ras block link */ struct list_head node;/* the device */ @@ -977,4 +979,12 @@ void amdgpu_ras_event_log_print(struct amdgpu_device *adev, u64 event_id, const char *fmt, ...);bool amdgpu_ras_is_rma (struct amdgpu_device *adev); + +bool amdgpu_ras_test_marker (struct amdgpu_device *adev, +struct ras_common_if *head,int marker); +void amdgpu_ras_set_marker (struct amdgpu_device *adev, +struct ras_common_if *head,int marker); +bool amdgpu_ras_test_and_clear_marker (struct amdgpu_device *adev, +struct ras_common_if *head, +int marker);#endif
Re: [RFC v2 02/15] drm/amdgpu: add a flag to track ras debugfs creation status
In reply to Jiang LiuQuoted 15 lines · Jiang Liu
Add a flag to track ras debugfs creation status, to avoid possible incorrect reference count management for ras block object in function
amdgpu_ras_aca_is_supported().Signed-off-by Jiang Liu <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 2 ++ drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c | 9 +++++++-- 2 files changed, 9 insertions(+), 2 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu.h +0
-0
drivers/gpu/drm/amd/amdgpu/amdgpu.h int amdgpu_ip_block_resume(struct amdgpu_ip_block *ip_block); · line 382
Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/amdgpu.h +2
-0@@ -378,6 +378,8 @@ int amdgpu_ip_block_resume(struct amdgpu_ip_block *ip_block);#define AMDGPU_MAX_IP_NUM 16enum amdgpu_marker { +// Markers for ras blocks. +AMDGPU_MARKER_RAS_DEBUGFS ,
As mentioned in patch 1, keeping a global tracker at this layer is not the right solution. The expectation is each object/IP block to keep the state info.
Thanks,
Lijo
Quoted 37 lines
drivers/gpu/drm/amd/amdgpu/amdgpu.h +0
-0// Markers for IRQs, used for both ip blocks and ras blocks. AMDGPU_MARKER_IRQ0 =32 , AMDGPU_MARKER_IRQ1,drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +7
-2@@ -1997,7 +1997,8 @@ static void amdgpu_ras_debugfs_create(struct amdgpu_device *adev, {struct ras_manager *obj =amdgpu_ras_find_obj (adev, &head->head); -if (!obj || !dir) +if (!obj || !dir || +amdgpu_ras_test_marker (adev, &head->head,AMDGPU_MARKER_RAS_DEBUGFS ))return ;get_obj (obj);@@ -2008,6 +2009,8 @@ static void amdgpu_ras_debugfs_create(struct amdgpu_device *adev,debugfs_create_file (obj->fs_data.debugfs_name,S_IWUGO |S_IRUGO , dir, obj, &amdgpu_ras_debugfs_ops); + +amdgpu_ras_set_marker (adev, &head->head,AMDGPU_MARKER_RAS_DEBUGFS ); }static bool amdgpu_ras_aca_is_supported (struct amdgpu_device *adev)@@ -2136,7 +2139,9 @@ static int amdgpu_ras_fs_fini(struct amdgpu_device *adev)if (IS_ENABLED (CONFIG_DEBUG_FS )) {list_for_each_entry_safe (con_obj, tmp, &con->head, node) { ip_obj =amdgpu_ras_find_obj (adev, &con_obj->head); -if (ip_obj) +if (ip_obj && +amdgpu_ras_test_and_clear_marker (adev, &ip_obj->head, +AMDGPU_MARKER_RAS_DEBUGFS ))put_obj (ip_obj); } }
Re: [RFC v2 03/15] drm/amdgpu: free all resources on error recovery path of amdgpu_ras_init()
In reply to Jiang LiuQuoted 63 lines · Jiang Liu
Free all allocated resources on error recovery path in function
amdgpu_ras_init().Signed-off-by Jiang Liu <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c | 19 ++++++++++++++----- 1 file changed, 14 insertions(+), 5 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +6
-5@@ -3864,6 +3864,7 @@ static void amdgpu_ras_init_reserved_vram_size(struct amdgpu_device *adev)int amdgpu_ras_init (struct amdgpu_device *adev) {struct amdgpu_ras *con =amdgpu_ras_get_context (adev); +struct amdgpu_ras_block_list *ras_node, *tmp;int r;if (con)@@ -3953,20 +3954,20 @@ int amdgpu_ras_init(struct amdgpu_device *adev)* to handle fatal error */ r =amdgpu_nbio_ras_sw_init (adev);if (r) -return r; + goto release_con;if (adev->nbio.ras && adev->nbio.ras->init_ras_controller_interrupt) { r = adev->nbio.ras->init_ras_controller_interrupt (adev);if (r) - goto release_con; + goto free_blocks; }if (adev->nbio.ras && adev->nbio.ras->init_ras_err_event_athub_interrupt) { r = adev->nbio.ras->init_ras_err_event_athub_interrupt (adev);if (r) - goto release_con; + goto free_blocks; }/* Packed socket_id to ras feature mask bits[31:29] */ @@ -3982,7 +3983,7 @@ int amdgpu_ras_init(struct amdgpu_device *adev)if (amdgpu_ras_fs_init (adev)) { r = -EINVAL ; - goto release_con; + goto free_blocks; }if (amdgpu_ras_aca_is_supported (adev)) {@@ -3991,7 +3992,7 @@ int amdgpu_ras_init(struct amdgpu_device *adev)else r =amdgpu_mca_init (adev);if (r) - goto release_con; + goto clear_ras_fs; } dev_info(adev->dev, "RAS INFO: ras initialized successfully, "
drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c int amdgpu_ras_init(struct amdgpu_device *adev) · line 4009
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +7
-0@@ -3999,6 +4000,14 @@ int amdgpu_ras_init(struct amdgpu_device *adev) adev->ras_hw_enabled, adev->ras_enabled);return 0 ; + +clear_ras_fs: +amdgpu_ras_fs_fini (adev); +free_blocks: +list_for_each_entry_safe (ras_node, tmp, &adev->ras_list, node) { +list_del (&ras_node->node); +kfree (ras_node);
Suggest to add amdgpu_nbio_ras_sw_fini which calls something like amdgpu_ras_unregister_ras_block instead of this.
Thanks,
Lijo
Quoted 4 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +1
-0+ } release_con:amdgpu_ras_set_context (adev,NULL );kfree (con);
Re: [RFC v2 04/15] drm/amdgpu: introduce a flag to track refcount held for features
In reply to Jiang LiuQuoted 5 lines · Jiang Liu
Currently we track the refcount on ras block object for features by checking
if (obj && amdgpu_ras_is_feature_enabled(adev, head)), which is a little unreliable. So introduce a dedicated flag to track the reference count.
Please clarify more on this. I see con->features available to track the status.
Thanks,
Lijo
Quoted 45 lines
Signed-off-by Jiang Liu <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 1 + drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c | 9 +++++++-- 2 files changed, 8 insertions(+), 2 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu.h +1
-0@@ -380,6 +380,7 @@ int amdgpu_ip_block_resume(struct amdgpu_ip_block *ip_block);enum amdgpu_marker {// Markers for ras blocks. AMDGPU_MARKER_RAS_DEBUGFS , +AMDGPU_MARKER_RAS_FEATURE ,// Markers for IRQs, used for both ip blocks and ras blocks. AMDGPU_MARKER_IRQ0 =32 ,AMDGPU_MARKER_IRQ1 ,drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +7
-2@@ -775,15 +775,20 @@ static int __amdgpu_ras_feature_enable(struct amdgpu_device *adev, obj =amdgpu_ras_create_obj (adev, head);if (!obj)return -EINVAL ; - }else { +amdgpu_ras_set_marker (adev, head,AMDGPU_MARKER_RAS_FEATURE ); + }else if (!amdgpu_ras_test_marker (adev, head, +AMDGPU_MARKER_RAS_FEATURE )) {/* In case we create obj somewhere else */ get_obj (obj); +amdgpu_ras_set_marker (adev, head,AMDGPU_MARKER_RAS_FEATURE ); } con->features |=BIT (head->block); }else {if (obj &&amdgpu_ras_is_feature_enabled (adev, head)) { con->features &= ~BIT (head->block); -put_obj (obj); +if (amdgpu_ras_test_and_clear_marker (adev, head, +AMDGPU_MARKER_RAS_FEATURE )) +put_obj (obj); } }
Re: [RFC v2 05/15] drm/amdgpu: enhance amdgpu_ras_block_late_fini()
In reply to Jiang LiuQuoted 46 lines · Jiang Liu
Enhance
amdgpu_ras_block_late_fini()to revert what has been done byamdgpu_ras_block_late_init(), and fix a possible resource leakage in functionamdgpu_ras_block_late_init().Signed-off-by Jiang Liu <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c | 16 ++++++++++------ 1 file changed, 10 insertions(+), 6 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +3
-6@@ -4104,13 +4104,13 @@ int amdgpu_ras_block_late_init(struct amdgpu_device *adev, ras_obj->hw_ops->query_ras_error_status)) { r =amdgpu_ras_sysfs_create (adev, ras_block);if (r) - goto interrupt; + goto cleanup;/* Those are the cached values at init. */ query_info =kzalloc (sizeof (*query_info),GFP_KERNEL );if (!query_info) -return -ENOMEM ; + goto cleanup;memcpy (&query_info->head, ras_block,sizeof (struct ras_common_if ));if (amdgpu_ras_query_error_count (adev, &ce_count, &ue_count, query_info) ==0 ) {@@ -4123,11 +4123,8 @@ int amdgpu_ras_block_late_init(struct amdgpu_device *adev,return 0 ; -interrupt: -if (ras_obj->ras_cb) -amdgpu_ras_interrupt_remove_handler (adev, ras_block); cleanup: -amdgpu_ras_feature_enable (adev, ras_block,0 ); +amdgpu_ras_block_late_fini (adev, ras_block);return r; }@@ -4142,9 +4139,16 @@ void amdgpu_ras_block_late_fini(struct amdgpu_device *adev, struct ras_common_if *ras_block) {struct amdgpu_ras_block_object *ras_obj;
drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c void amdgpu_ras_block_late_fini(struct amdgpu_device *adev, · line 4151
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +7
-0+if (!ras_block)return ; +amdgpu_ras_feature_enable (adev, ras_block,0 ); + +/* in resume/reset phase, no need to delete ras fs node */ +if (adev->in_suspend ||amdgpu_in_reset (adev)) +return ; +
I guess late_init will get called during resume or after reset. Won't it go to sysfs creation logic that time?
Thanks,
Lijo
Quoted 3 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +0
-0amdgpu_ras_sysfs_remove (adev, ras_block); ras_obj = container_of(ras_block, struct amdgpu_ras_block_object, ras_comm);
Re: [RFC v2 06/15] drm/amdgpu: enhance amdgpu_ras_pre_fini() to better support SR
In reply to Jiang LiuQuoted 41 lines · Jiang Liu
Enhance
amdgpu_ras_pre_fini()to better support suspend/resume by:
1) fix possible resource leakage.amdgpu_release_ras_context()only kfree(con) but doesn't release resources associated with the con object.
2) callamdgpu_ras_pre_fini()inamdgpu_device_suspend()to undo what has been done byamdgpu_ras_late_init(), becauseamdgpu_device_resume()will invokeamdgpu_ras_late_init()on resume.
3) moveamdgpu_ras_recovery_fini()fromamdgpu_ras_pre_fini()toamdgpu_ras_fini()
4) move calling ofobj->ras_fini()fromamdgpu_ras_fini()toamdgpu_ras_pre_fini().Signed-off-by Jiang Liu <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 6 ++- drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c | 44 +++++++++++++--------- 2 files changed, 31 insertions(+), 19 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +2
-2@@ -4613,6 +4613,8 @@ int amdgpu_device_init(struct amdgpu_device *adev,return 0 ; release_ras_con: +amdgpu_ras_pre_fini (adev); +amdgpu_ras_fini (adev);if (amdgpu_sriov_vf (adev))amdgpu_virt_release_full_gpu (adev, true);@@ -4627,8 +4629,6 @@ int amdgpu_device_init(struct amdgpu_device *adev, adev->virt.ops =NULL ; r = -EAGAIN ; } -amdgpu_release_ras_context (adev); - failed:amdgpu_vf_error_trans_all (adev);
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c int amdgpu_device_suspend(struct drm_device *dev, bool notify_clients) · line 4926
Quoted 6 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +2
-0@@ -4921,6 +4921,8 @@ int amdgpu_device_suspend(struct drm_device *dev, bool notify_clients)cancel_delayed_work_sync (&adev->delayed_init_work); +/* disable ras feature must before hw fini */ +amdgpu_ras_pre_fini (adev);amdgpu_ras_suspend (adev);
Based on the usages above, it makes more sense to keep amdgpu_ras_pre_fini as a static function and call in ras_fini/ras_suspend (contain the calls at ras layer and avoid another new public interface).
Copying Tao to take a look.
Thanks,
Lijo
Quoted 85 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0
-0amdgpu_device_ip_suspend_phase1 (adev);drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +27
-17@@ -4270,42 +4270,49 @@ int amdgpu_ras_late_init(struct amdgpu_device *adev)int amdgpu_ras_pre_fini (struct amdgpu_device *adev) {struct amdgpu_ras *con =amdgpu_ras_get_context (adev); +struct amdgpu_ras_block_list *node, *tmp; +struct amdgpu_ras_block_object *obj; -if (!adev->ras_enabled || !con) -return 0 ; +if (amdgpu_sriov_vf (adev) && !amdgpu_sriov_ras_telemetry_en (adev)) + goto disable; +list_for_each_entry_safe (node, tmp, &adev->ras_list, node) { + obj = node->ras_obj; +if (!obj) +continue ; + +if (!amdgpu_ras_is_supported (adev, obj->ras_comm.block)) +continue ; + +if (obj->ras_fini) + obj->ras_fini (adev, &obj->ras_comm); +else +amdgpu_ras_block_late_fini_default (adev, &obj->ras_comm); + } +disable:/* Need disable ras on all IPs here before ip [hw/sw]fini */ -if (AMDGPU_RAS_GET_FEATURES (con->features)) +if (con &&AMDGPU_RAS_GET_FEATURES (con->features))amdgpu_ras_disable_all_features (adev,0 ); -amdgpu_ras_recovery_fini (adev); +return 0 ; }int amdgpu_ras_fini (struct amdgpu_device *adev) {struct amdgpu_ras_block_list *ras_node, *tmp; -struct amdgpu_ras_block_object *obj =NULL ;struct amdgpu_ras *con =amdgpu_ras_get_context (adev);if (!adev->ras_enabled || !con) -return 0 ; + goto out_free_context;list_for_each_entry_safe (ras_node, tmp, &adev->ras_list, node) { -if (ras_node->ras_obj) { - obj = ras_node->ras_obj; -if (amdgpu_ras_is_supported (adev, obj->ras_comm.block) && - obj->ras_fini) - obj->ras_fini (adev, &obj->ras_comm); -else -amdgpu_ras_block_late_fini_default (adev, &obj->ras_comm); - } -/* Clear ras blocks from ras_list and free ras block list node */ list_del (&ras_node->node);kfree (ras_node); } +amdgpu_ras_recovery_fini (adev);amdgpu_ras_fs_fini (adev);amdgpu_ras_interrupt_remove_all (adev);@@ -4323,8 +4330,11 @@ int amdgpu_ras_fini(struct amdgpu_device *adev)cancel_delayed_work_sync (&con->ras_counte_delay_work); -amdgpu_ras_set_context (adev,NULL ); -kfree (con); +out_free_context: +if (con) { +amdgpu_ras_set_context (adev,NULL ); +kfree (con); + }return 0 ; }
Re: [RFC v2 07/15] drm/admgpu: rename amdgpu_ras_pre_fini() to amdgpu_ras_early_fini()
In reply to Jiang LiuQuoted 70 lines · Jiang Liu
Rename
amdgpu_ras_pre_fini()toamdgpu_ras_early_fini(), to keep same style with other code.Signed-off-by Jiang Liu <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 6 +++--- drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_jpeg.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_nbio.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c | 16 ++++++++-------- drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h | 6 +++--- drivers/gpu/drm/amd/amdgpu/amdgpu_sdma.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_vcn.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_xgmi.c | 2 +- drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c | 2 +- drivers/gpu/drm/amd/amdgpu/jpeg_v4_0_3.c | 2 +- drivers/gpu/drm/amd/amdgpu/mmhub_v1_8.c | 2 +- drivers/gpu/drm/amd/amdgpu/vcn_v4_0_3.c | 2 +- 14 files changed, 25 insertions(+), 25 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +3
-3@@ -4613,7 +4613,7 @@ int amdgpu_device_init(struct amdgpu_device *adev,return 0 ; release_ras_con: -amdgpu_ras_pre_fini (adev); +amdgpu_ras_early_fini (adev);amdgpu_ras_fini (adev);if (amdgpu_sriov_vf (adev))amdgpu_virt_release_full_gpu (adev, true);@@ -4705,7 +4705,7 @@ void amdgpu_device_fini_hw(struct amdgpu_device *adev)amdgpu_xcp_cfg_sysfs_fini (adev);/* disable ras feature must before hw fini */ -amdgpu_ras_pre_fini (adev); +amdgpu_ras_early_fini (adev);amdgpu_ttm_set_buffer_funcs_status (adev, false);@@ -4922,7 +4922,7 @@ int amdgpu_device_suspend(struct drm_device *dev, bool notify_clients)cancel_delayed_work_sync (&adev->delayed_init_work);/* disable ras feature must before hw fini */ -amdgpu_ras_pre_fini (adev); +amdgpu_ras_early_fini (adev);amdgpu_ras_suspend (adev);amdgpu_device_ip_suspend_phase1 (adev);drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +1
-1@@ -918,7 +918,7 @@ int amdgpu_gfx_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *rreturn 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }drivers/gpu/drm/amd/amdgpu/amdgpu_jpeg.c +0
-0
drivers/gpu/drm/amd/amdgpu/amdgpu_jpeg.c int amdgpu_jpeg_ras_late_init(struct amdgpu_device *adev, struct ras_common_if * · line 303
Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_jpeg.c +1
-1@@ -300,7 +300,7 @@ int amdgpu_jpeg_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *return 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);
This feels confusing - calling early_fini from late_init
Thanks,
Lijo
Quoted 206 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_jpeg.c +0
-0return r; }drivers/gpu/drm/amd/amdgpu/amdgpu_nbio.c +1
-1@@ -71,6 +71,6 @@ int amdgpu_nbio_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *return 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +8
-8@@ -4124,7 +4124,7 @@ int amdgpu_ras_block_late_init(struct amdgpu_device *adev,return 0 ; cleanup: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }@@ -4135,7 +4135,7 @@ static int amdgpu_ras_block_late_init_default(struct amdgpu_device *adev, }/* helper function to remove ras fs node and interrupt handler */ -void amdgpu_ras_block_late_fini (struct amdgpu_device *adev, +void amdgpu_ras_block_early_fini (struct amdgpu_device *adev,struct ras_common_if *ras_block) {struct amdgpu_ras_block_object *ras_obj;@@ -4156,10 +4156,10 @@ void amdgpu_ras_block_late_fini(struct amdgpu_device *adev,amdgpu_ras_interrupt_remove_handler (adev, ras_block); } -static void amdgpu_ras_block_late_fini_default (struct amdgpu_device *adev, +static void amdgpu_ras_block_early_fini_default (struct amdgpu_device *adev,struct ras_common_if *ras_block) { -return amdgpu_ras_block_late_fini (adev, ras_block); +return amdgpu_ras_block_early_fini (adev, ras_block); }/* do some init work after IP late init as dependence. @@ -4267,7 +4267,7 @@ int amdgpu_ras_late_init(struct amdgpu_device *adev) }/* do some fini work before IP fini as dependence */ -int amdgpu_ras_pre_fini (struct amdgpu_device *adev) +int amdgpu_ras_early_fini (struct amdgpu_device *adev) {struct amdgpu_ras *con =amdgpu_ras_get_context (adev);struct amdgpu_ras_block_list *node, *tmp;@@ -4284,10 +4284,10 @@ int amdgpu_ras_pre_fini(struct amdgpu_device *adev)if (!amdgpu_ras_is_supported (adev, obj->ras_comm.block))continue ; -if (obj->ras_fini) - obj->ras_fini (adev, &obj->ras_comm); +if (obj->ras_early_fini) + obj->ras_early_fini (adev, &obj->ras_comm);else -amdgpu_ras_block_late_fini_default (adev, &obj->ras_comm); +amdgpu_ras_block_early_fini_default (adev, &obj->ras_comm); } disable:drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h +3
-3@@ -711,7 +711,7 @@ struct amdgpu_ras_block_object {int (*ras_block_match)(struct amdgpu_ras_block_object *block_obj,enum amdgpu_ras_block block,uint32_t sub_block_index);int (*ras_late_init)(struct amdgpu_device *adev,struct ras_common_if *ras_block); -void (*ras_fini)(struct amdgpu_device *adev,struct ras_common_if *ras_block); +void (*ras_early_fini)(struct amdgpu_device *adev,struct ras_common_if *ras_block);ras_ih_cb ras_cb;const struct amdgpu_ras_block_hw_ops *hw_ops; };@@ -825,13 +825,13 @@ amdgpu_ras_error_to_ta(enum amdgpu_ras_error_type error) {/* called in ip_init and ip_fini */ int amdgpu_ras_init (struct amdgpu_device *adev);int amdgpu_ras_late_init (struct amdgpu_device *adev); +int amdgpu_ras_early_fini (struct amdgpu_device *adev);int amdgpu_ras_fini (struct amdgpu_device *adev); -int amdgpu_ras_pre_fini (struct amdgpu_device *adev);int amdgpu_ras_block_late_init (struct amdgpu_device *adev,struct ras_common_if *ras_block); -void amdgpu_ras_block_late_fini (struct amdgpu_device *adev, +void amdgpu_ras_block_early_fini (struct amdgpu_device *adev,struct ras_common_if *ras_block); int amdgpu_ras_feature_enable(struct amdgpu_device *adev,drivers/gpu/drm/amd/amdgpu/amdgpu_sdma.c +1
-1@@ -116,7 +116,7 @@ int amdgpu_sdma_ras_late_init(struct amdgpu_device *adev,return 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c +1
-1@@ -336,7 +336,7 @@ int amdgpu_umc_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *rreturn 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }drivers/gpu/drm/amd/amdgpu/amdgpu_vcn.c +1
-1@@ -1248,7 +1248,7 @@ int amdgpu_vcn_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *rreturn 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }drivers/gpu/drm/amd/amdgpu/amdgpu_xgmi.c +1
-1@@ -1174,7 +1174,7 @@ static int amdgpu_xgmi_ras_late_init(struct amdgpu_device *adev, struct ras_commreturn 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c +1
-1@@ -5080,7 +5080,7 @@ static int gfx_v9_4_3_ras_late_init(struct amdgpu_device *adev, struct ras_commoreturn 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }drivers/gpu/drm/amd/amdgpu/jpeg_v4_0_3.c +1
-1@@ -1311,7 +1311,7 @@ static int jpeg_v4_0_3_ras_late_init(struct amdgpu_device *adev, struct ras_commreturn 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }drivers/gpu/drm/amd/amdgpu/mmhub_v1_8.c +1
-1@@ -789,7 +789,7 @@ static int mmhub_v1_8_ras_late_init(struct amdgpu_device *adev, struct ras_commoreturn 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }drivers/gpu/drm/amd/amdgpu/vcn_v4_0_3.c +1
-1@@ -1996,7 +1996,7 @@ static int vcn_v4_0_3_ras_late_init(struct amdgpu_device *adev, struct ras_commoreturn 0 ; late_fini: -amdgpu_ras_block_late_fini (adev, ras_block); +amdgpu_ras_block_early_fini (adev, ras_block);return r; }
Re: [RFC v2 08/15] drm/amdgpu: make IP block state machine works in stack like way
In reply to Jiang LiuQuoted 10 lines · Jiang Liu
There are some mismatches between IP block state machine and its associated status flags, especially about the meaning of
status.late_initialized. So let's make the state machine and associated status flas work in stack-like way as below:
Callback Statusearly_init: valid = truesw_init: sw = truehw_init: hw = truelate_init:late_initialized= trueearly_fini:late_initialized= false
Changing the state like this is confusing. The intention of late_fini is to reverse the steps in late_init. It's straight forward read like if the ip is not late_initialized, no need to late_fini. This is making that complicated.
Thanks,
Lijo
Quoted 40 lines
hw_fini: hw = falsesw_fini: sw = falselate_fini: valid = falseSigned-off-by Jiang Liu <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +4
-3@@ -3376,6 +3376,8 @@ static int amdgpu_device_ip_fini_early(struct amdgpu_device *adev)DRM_DEBUG ("early_fini of IP block <%s> failed %d\n" , adev->ip_blocks[i].version->funcs->name, r); } + + adev->ip_blocks[i].status.late_initialized = false; }amdgpu_device_set_pg_state (adev,AMD_PG_STATE_UNGATE );@@ -3445,15 +3447,14 @@ static int amdgpu_device_ip_fini(struct amdgpu_device *adev) } } adev->ip_blocks[i].status.sw = false; - adev->ip_blocks[i].status.valid = false; }for (i = adev->num_ip_blocks -1 ; i >=0 ; i--) { -if (!adev->ip_blocks[i].status.late_initialized) +if (!adev->ip_blocks[i].status.valid)continue ;if (adev->ip_blocks[i].version->funcs->late_fini) adev->ip_blocks[i].version->funcs->late_fini (&adev->ip_blocks[i]); - adev->ip_blocks[i].status.late_initialized = false; + adev->ip_blocks[i].status.valid = false; }amdgpu_ras_fini (adev);
Re: [RFC v2 10/15] drm/admgpu: make device state machine work in stack like way
In reply to Jiang LiuQuoted 22 lines · Jiang Liu
Make the device state machine work in stack like way to better support suspend/resume by following changes:
1.
amdgpu_driver_load_kms()amdgpu_device_init() amdgpu_device_ip_early_init() ip_blocks[i].early_init() ip_blocks[i].status.valid = true amdgpu_device_ip_init() amdgpu_ras_init() ip_blocks[i].sw_init() ip_blocks[i].status.sw = true ip_blocks[i].hw_init() ip_blocks[i].status.hw = true amdgpu_device_ip_late_init() ip_blocks[i].late_init() ip_blocks[i].status.late_initialized = true amdgpu_ras_late_init() ras_blocks[i].ras_late_init() amdgpu_ras_feature_enable_on_boot()2. amdgpu_pmops_suspend()/amdgpu_pmops_freeze()/amdgpu_pmops_poweroff()
Quoted 8 lines
amdgpu_device_suspend() amdgpu_ras_early_fini() ras_blocks[i].ras_early_fini() amdgpu_ras_feature_disable() amdgpu_ras_suspend() amdgpu_ras_disable_all_features()+++ ip_blocks[i].early_fini()
+++ ip_blocks[i].status.late_initialized = false
As said in the previous patch, please don't add confusion. You could maintain a state machine like early fini done/late fini done etc, but please don't introduce this kind of confusing things.
Thanks,
Lijo
Quoted 100 lines
ip_blocks[i].suspend()3. amdgpu_pmops_resume()/amdgpu_pmops_thaw()/amdgpu_pmops_restore()
amdgpu_device_resume() amdgpu_device_ip_resume() ip_blocks[i].resume() amdgpu_device_ip_late_init() ip_blocks[i].late_init() ip_blocks[i].status.late_initialized = true amdgpu_ras_late_init() ras_blocks[i].ras_late_init() amdgpu_ras_feature_enable_on_boot() amdgpu_ras_resume() amdgpu_ras_enable_all_features()4.
amdgpu_driver_unload_kms()amdgpu_device_fini_hw() amdgpu_ras_early_fini() ras_blocks[i].ras_early_fini() +++ ip_blocks[i].early_fini() +++ ip_blocks[i].status.late_initialized = false ip_blocks[i].hw_fini() ip_blocks[i].status.hw = false5.
amdgpu_driver_release_kms()amdgpu_device_fini_sw() amdgpu_device_ip_fini() ip_blocks[i].sw_fini() ip_blocks[i].status.sw = false --- ip_blocks[i].status.valid = false +++ amdgpu_ras_fini() ip_blocks[i].late_fini()+++ ip_blocks[i].status.valid = false
--- ip_blocks[i].status.late_initialized = false
---amdgpu_ras_fini()The main changes include:
1) invoke ip_blocks[i].early_fini inamdgpu_pmops_suspend().
2) set ip_blocks[i].status.late_initialized to false after calling callbackearly_fini. We have auditted all usages of thelate_initializedflag and no functional changes found.
3) only set ip_blocks[i].status.valid = false after calling thelate_finicallback.
4) callamdgpu_ras_fini()before invoking ip_blocks[i].late_fini.There's one more task left to analyze GPU reset related state machine transitions.
Signed-off-by Jiang Liu <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 22 ++++++++++++++++++++-- 1 file changed, 20 insertions(+), 2 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +20
-2@@ -3449,6 +3449,8 @@ static int amdgpu_device_ip_fini(struct amdgpu_device *adev) adev->ip_blocks[i].status.sw = false; } +amdgpu_ras_fini (adev); +for (i = adev->num_ip_blocks -1 ; i >=0 ; i--) {if (!adev->ip_blocks[i].status.valid)continue ;@@ -3457,8 +3459,6 @@ static int amdgpu_device_ip_fini(struct amdgpu_device *adev) adev->ip_blocks[i].status.valid = false; } -amdgpu_ras_fini (adev); -return 0 ; }@@ -3516,6 +3516,24 @@ static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev)if (amdgpu_dpm_set_df_cstate (adev,DF_CSTATE_DISALLOW ))dev_warn (adev->dev,"Failed to disallow df cstate" ); +for (i = adev->num_ip_blocks -1 ; i >=0 ; i--) { +if (!adev->ip_blocks[i].status.valid) +continue ; +if (!adev->ip_blocks[i].status.late_initialized) +continue ; + +if (adev->ip_blocks[i].version->funcs->early_fini) { + r = adev->ip_blocks[i].version->funcs->early_fini (&adev->ip_blocks[i]); +if (r) { +DRM_ERROR (" of IP block <%s> failed %d\n" , + adev->ip_blocks[i].version->funcs->name, r); +return r; + } + } + + adev->ip_blocks[i].status.late_initialized = false; + } +for (i = adev->num_ip_blocks -1 ; i >=0 ; i--) {if (!adev->ip_blocks[i].status.valid)continue ;
Re: [RFC v2 11/15] drm/amdgpu: convert ip block bool flags into an enum
In reply to Jiang LiuQuoted 60 lines · Jiang Liu
Convert ip block bool flags into an enumeration, to explicitly mark current state of the ip block. Also introduce helper functions to manipulate the ip block state. Now the state machine works as below:
Callbacks State after successfully execute callback
AMDGPU_IP_STATE_INVALID.early_init() AMDGPU_IP_STATE_EARLY .sw_init() AMDGPU_IP_STATE_SW .hw_init() AMDGPU_IP_STATE_HW .late_init() AMDGPU_IP_STATE_LATE .early_fini() AMDGPU_IP_STATE_HW .hw_fini() AMDGPU_IP_STATE_SW .sw_fini() AMDGPU_IP_STATE_EARLY .late_fini() AMDGPU_IP_STATE_INVALIDSigned-off-by Jiang Liu <[email protected]>--- drivers/gpu/drm/amd/amdgpu/aldebaran.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 48 ++++++- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 144 ++++++++++++-------- drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c | 4 +- drivers/gpu/drm/amd/amdgpu/amdgpu_psp.c | 6 +- drivers/gpu/drm/amd/amdgpu/amdgpu_reset.c | 6 +- drivers/gpu/drm/amd/amdgpu/sienna_cichlid.c | 2 +- drivers/gpu/drm/amd/amdgpu/smu_v13_0_10.c | 2 +- drivers/gpu/drm/amd/pm/swsmu/amdgpu_smu.c | 2 +- 9 files changed, 143 insertions(+), 73 deletions(-)drivers/gpu/drm/amd/amdgpu/aldebaran.c +1
-1@@ -301,7 +301,7 @@ static int aldebaran_mode2_restore_ip(struct amdgpu_device *adev)return r; } } - adev->ip_blocks[i].status.late_initialized = true; +amdgpu_device_ip_set_state (adev, i,AMDGPU_IP_STATE_LATE ); }amdgpu_device_set_cg_state (adev,AMD_CG_STATE_GATE );drivers/gpu/drm/amd/amdgpu/amdgpu.h +12
-0@@ -395,11 +395,32 @@ enum amdgpu_marker {#define AMDGPU_MARKER_IRQ (idx) (AMDGPU_MARKER_IRQ0 + (idx)) +/** +* States for ip block state machine. +* +* The IP block state machine has five states and the associated state +* transition works in stack like way as below: +* Callbacks State after successfully execute callback +* AMDGPU_IP_STATE_INVALID +* .early_init() AMDGPU_IP_STATE_EARLY +* .sw_init() AMDGPU_IP_STATE_SW +* .hw_init() AMDGPU_IP_STATE_HW +* .late_init() AMDGPU_IP_STATE_LATE +* .early_fini() AMDGPU_IP_STATE_HW
drivers/gpu/drm/amd/amdgpu/amdgpu.h enum amdgpu_marker { · line 419
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu.h +10
-0+* .hw_fini() AMDGPU_IP_STATE_SW +* .sw_fini() AMDGPU_IP_STATE_EARLY +* .late_fini() AMDGPU_IP_STATE_INVALID +*/ +enum amdgpu_device_ip_state { +AMDGPU_IP_STATE_INVALID =0 , +AMDGPU_IP_STATE_EARLY =1 , +AMDGPU_IP_STATE_SW =2 , +AMDGPU_IP_STATE_HW =3 , +AMDGPU_IP_STATE_LATE =4 ,
Would suggest to add separate INIT/FINI stages for each state - EARLY_INIT_DONE/EARLY_FINI_DONE. Then reading the code will be easier.
Thanks,
Lijo
Quoted 662 lines
drivers/gpu/drm/amd/amdgpu/amdgpu.h +22
-4+}; +struct amdgpu_ip_block_status { -bool valid; -bool sw; -bool hw; -bool late_initialized; +enum amdgpu_device_ip_state state;bool hang;uint64_t markers; };@@ -429,6 +450,25 @@ amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev,int amdgpu_device_ip_block_add (struct amdgpu_device *adev,const struct amdgpu_ip_block_version *ip_block_version); +void amdgpu_device_ip_set_state (struct amdgpu_device *adev,int index, +enum amdgpu_device_ip_state state); +enum amdgpu_device_ip_state amdgpu_device_ip_state (struct amdgpu_device *adev, +int index); +bool amdgpu_device_ip_valid (struct amdgpu_device *adev,int index); +void amdgpu_ip_block_set_state (struct amdgpu_ip_block *ip_block, +enum amdgpu_device_ip_state state); + +static inline enum amdgpu_device_ip_state +amdgpu_ip_block_state (struct amdgpu_ip_block *ip_block) +{ +return ip_block->status.state; +} + +static inline bool amdgpu_ip_block_valid (struct amdgpu_ip_block *ip_block) +{ +return ip_block->status.state !=AMDGPU_IP_STATE_INVALID ; +} +static inline void amdgpu_ip_block_set_marker (struct amdgpu_ip_block *ip_block,enum amdgpu_marker marker) {drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +86
-58@@ -297,7 +297,7 @@ int amdgpu_ip_block_suspend(struct amdgpu_ip_block *ip_block) } } - ip_block->status.hw = false; +amdgpu_ip_block_set_state (ip_block,AMDGPU_IP_STATE_SW );return 0 ; }@@ -315,7 +315,7 @@ int amdgpu_ip_block_resume(struct amdgpu_ip_block *ip_block) } } - ip_block->status.hw = true; +amdgpu_ip_block_set_state (ip_block,AMDGPU_IP_STATE_HW );return 0 ; }@@ -2164,7 +2164,7 @@ int amdgpu_device_ip_set_clockgating_state(void *dev,int i, r =0 ;for (i =0 ; i < adev->num_ip_blocks; i++) { -if (!adev->ip_blocks[i].status.valid) +if (!amdgpu_device_ip_valid (adev, i))continue ;if (adev->ip_blocks[i].version->type != block_type)continue ;@@ -2198,7 +2198,7 @@ int amdgpu_device_ip_set_powergating_state(void *dev,int i, r =0 ;for (i =0 ; i < adev->num_ip_blocks; i++) { -if (!adev->ip_blocks[i].status.valid) +if (!amdgpu_device_ip_valid (adev, i))continue ;if (adev->ip_blocks[i].version->type != block_type)continue ;@@ -2230,7 +2230,7 @@ void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev,int i;for (i =0 ; i < adev->num_ip_blocks; i++) { -if (!adev->ip_blocks[i].status.valid) +if (!amdgpu_device_ip_valid (adev, i))continue ;if (adev->ip_blocks[i].version->funcs->get_clockgating_state) adev->ip_blocks[i].version->funcs->get_clockgating_state ((void *)adev, flags);@@ -2252,7 +2252,7 @@ int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev,int i, r;for (i =0 ; i < adev->num_ip_blocks; i++) { -if (!adev->ip_blocks[i].status.valid) +if (!amdgpu_device_ip_valid (adev, i))continue ;if (adev->ip_blocks[i].version->type == block_type) {if (adev->ip_blocks[i].version->funcs->wait_for_idle) {@@ -2284,7 +2284,7 @@ bool amdgpu_device_ip_is_valid(struct amdgpu_device *adev,for (i =0 ; i < adev->num_ip_blocks; i++) {if (adev->ip_blocks[i].version->type == block_type) -return adev->ip_blocks[i].status.valid; +return amdgpu_device_ip_valid (adev, i); }return false;@@ -2375,6 +2375,29 @@ int amdgpu_device_ip_block_add(struct amdgpu_device *adev,return 0 ; } +void amdgpu_device_ip_set_state (struct amdgpu_device *adev,int index, +enum amdgpu_device_ip_state state) +{ +amdgpu_ip_block_set_state (&adev->ip_blocks[index], state); +} + +enum amdgpu_device_ip_state amdgpu_device_ip_state (struct amdgpu_device *adev, +int index) +{ +return amdgpu_ip_block_state (&adev->ip_blocks[index]); +} + +bool amdgpu_device_ip_valid (struct amdgpu_device *adev,int index) +{ +return amdgpu_ip_block_valid (&adev->ip_blocks[index]); +} + +void amdgpu_ip_block_set_state (struct amdgpu_ip_block *ip_block, +enum amdgpu_device_ip_state state) +{ + ip_block->status.state = state; +} +/** * amdgpu_device_enable_virtual_display - enable virtual display feature * @@ -2671,20 +2694,21 @@ static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)if ((amdgpu_ip_block_mask & (1 << i)) ==0 ) {DRM_WARN ("disabled ip block: %d <%s>\n" , i, adev->ip_blocks[i].version->funcs->name); - adev->ip_blocks[i].status.valid = false; +amdgpu_device_ip_set_state (adev, i,AMDGPU_IP_STATE_INVALID ); }else if (ip_block->version->funcs->early_init) { r = ip_block->version->funcs->early_init (ip_block);if (r == -ENOENT ) { - adev->ip_blocks[i].status.valid = false; +amdgpu_device_ip_set_state (adev, i,AMDGPU_IP_STATE_INVALID ); }else if (r) {DRM_ERROR ("early_init of IP block <%s> failed %d\n" , adev->ip_blocks[i].version->funcs->name, r); total = false; +amdgpu_device_ip_set_state (adev, i,AMDGPU_IP_STATE_INVALID ); }else { - adev->ip_blocks[i].status.valid = true; +amdgpu_device_ip_set_state (adev, i,AMDGPU_IP_STATE_EARLY ); } }else { - adev->ip_blocks[i].status.valid = true; +amdgpu_device_ip_set_state (adev, i,AMDGPU_IP_STATE_EARLY ); }/* get the vbios after the asic_funcs are set up */ if (adev->ip_blocks[i].version->type ==AMD_IP_BLOCK_TYPE_COMMON ) {@@ -2715,7 +2739,7 @@ static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)return -ENODEV ; ip_block =amdgpu_device_ip_get_ip_block (adev,AMD_IP_BLOCK_TYPE_GFX ); -if (ip_block->status.valid != false) +if (amdgpu_ip_block_valid (ip_block))amdgpu_amdkfd_device_probe (adev); adev->cg_flags &= amdgpu_cg_mask;@@ -2729,9 +2753,9 @@ static int amdgpu_device_ip_hw_init_phase1(struct amdgpu_device *adev)int i, r;for (i =0 ; i < adev->num_ip_blocks; i++) { -if (!adev->ip_blocks[i].status.sw) +if (amdgpu_device_ip_state (adev, i) <AMDGPU_IP_STATE_SW )continue ; -if (adev->ip_blocks[i].status.hw) +if (amdgpu_device_ip_state (adev, i) >=AMDGPU_IP_STATE_HW )continue ;if (!amdgpu_ip_member_of_hwini ( adev, adev->ip_blocks[i].version->type))@@ -2745,7 +2769,7 @@ static int amdgpu_device_ip_hw_init_phase1(struct amdgpu_device *adev) adev->ip_blocks[i].version->funcs->name, r);return r; } - adev->ip_blocks[i].status.hw = true; +amdgpu_device_ip_set_state (adev, i,AMDGPU_IP_STATE_HW ); } }@@ -2757,9 +2781,9 @@ static int amdgpu_device_ip_hw_init_phase2(struct amdgpu_device *adev)int i, r;for (i =0 ; i < adev->num_ip_blocks; i++) { -if (!adev->ip_blocks[i].status.sw) +if (amdgpu_device_ip_state (adev, i) <AMDGPU_IP_STATE_SW )continue ; -if (adev->ip_blocks[i].status.hw) +if (amdgpu_device_ip_state (adev, i) >=AMDGPU_IP_STATE_HW )continue ;if (!amdgpu_ip_member_of_hwini ( adev, adev->ip_blocks[i].version->type))@@ -2770,7 +2794,7 @@ static int amdgpu_device_ip_hw_init_phase2(struct amdgpu_device *adev) adev->ip_blocks[i].version->funcs->name, r);return r; } - adev->ip_blocks[i].status.hw = true; +amdgpu_device_ip_set_state (adev, i,AMDGPU_IP_STATE_HW ); }return 0 ;@@ -2791,11 +2815,11 @@ static int amdgpu_device_fw_loading(struct amdgpu_device *adev) AMD_IP_BLOCK_TYPE_PSP))break ; -if (!adev->ip_blocks[i].status.sw) +if (amdgpu_device_ip_state (adev, i) <AMDGPU_IP_STATE_SW )continue ;/* no need to do the fw loading again if already done*/ -if (adev->ip_blocks[i].status.hw == true) +if (amdgpu_device_ip_state (adev, i) >=AMDGPU_IP_STATE_HW )break ;if (amdgpu_in_reset (adev) || adev->in_suspend) {@@ -2809,7 +2833,7 @@ static int amdgpu_device_fw_loading(struct amdgpu_device *adev) adev->ip_blocks[i].version->funcs->name, r);return r; } - adev->ip_blocks[i].status.hw = true; +amdgpu_device_ip_set_state (adev, i,AMDGPU_IP_STATE_HW ); }break ; }@@ -2900,7 +2924,7 @@ static int amdgpu_device_ip_init(struct amdgpu_device *adev)return r;for (i =0 ; i < adev->num_ip_blocks; i++) { -if (!adev->ip_blocks[i].status.valid) +if (!amdgpu_device_ip_valid (adev, i))continue ;if (adev->ip_blocks[i].version->funcs->sw_init) { r = adev->ip_blocks[i].version->funcs->sw_init (&adev->ip_blocks[i]);@@ -2910,7 +2934,7 @@ static int amdgpu_device_ip_init(struct amdgpu_device *adev) goto init_failed; } } - adev->ip_blocks[i].status.sw = true; +amdgpu_device_ip_set_state (adev, i,AMDGPU_IP_STATE_SW );if (!amdgpu_ip_member_of_hwini ( adev, adev->ip_blocks[i].version->type))@@ -2923,7 +2947,7 @@ static int amdgpu_device_ip_init(struct amdgpu_device *adev)DRM_ERROR ("hw_init %d failed %d\n" , i, r); goto init_failed; } - adev->ip_blocks[i].status.hw = true; +amdgpu_device_ip_set_state (adev, i,AMDGPU_IP_STATE_HW ); }else if (adev->ip_blocks[i].version->type ==AMD_IP_BLOCK_TYPE_GMC ) {/* need to do gmc hw init early so we can allocate gpu mem */ /* Try to reserve bad pages early */ @@ -2945,7 +2969,7 @@ static int amdgpu_device_ip_init(struct amdgpu_device *adev)DRM_ERROR ("amdgpu_device_wb_init failed %d\n" , r); goto init_failed; } - adev->ip_blocks[i].status.hw = true; +amdgpu_device_ip_set_state (adev, i,AMDGPU_IP_STATE_HW );/* right after GMC hw init, we create CSA */ if (adev->gfx.mcbp) {@@ -3130,7 +3154,7 @@ int amdgpu_device_set_cg_state(struct amdgpu_device *adev,for (j =0 ; j < adev->num_ip_blocks; j++) { i = state ==AMD_CG_STATE_GATE ? j : adev->num_ip_blocks - j -1 ; -if (!adev->ip_blocks[i].status.late_initialized) +if (amdgpu_device_ip_state (adev, i) <AMDGPU_IP_STATE_LATE )continue ;/* skip CG for GFX, SDMA on S0ix */ if (adev->in_s0ix &&@@ -3167,7 +3191,7 @@ int amdgpu_device_set_pg_state(struct amdgpu_device *adev,for (j =0 ; j < adev->num_ip_blocks; j++) { i = state ==AMD_PG_STATE_GATE ? j : adev->num_ip_blocks - j -1 ; -if (!adev->ip_blocks[i].status.late_initialized) +if (amdgpu_device_ip_state (adev, i) <AMDGPU_IP_STATE_LATE )continue ;/* skip PG for GFX, SDMA on S0ix */ if (adev->in_s0ix &&@@ -3246,7 +3270,7 @@ static int amdgpu_device_ip_late_init(struct amdgpu_device *adev)int i =0 , r;for (i =0 ; i < adev->num_ip_blocks; i++) { -if (!adev->ip_blocks[i].status.hw) +if (amdgpu_device_ip_state (adev, i) <AMDGPU_IP_STATE_HW )continue ;if (adev->ip_blocks[i].version->funcs->late_init) { r = adev->ip_blocks[i].version->funcs->late_init (&adev->ip_blocks[i]);@@ -3256,7 +3280,7 @@ static int amdgpu_device_ip_late_init(struct amdgpu_device *adev)return r; } } - adev->ip_blocks[i].status.late_initialized = true; +amdgpu_device_ip_set_state (adev, i,AMDGPU_IP_STATE_LATE ); } r =amdgpu_ras_late_init (adev);@@ -3336,7 +3360,7 @@ static void amdgpu_ip_block_hw_fini(struct amdgpu_ip_block *ip_block) } } - ip_block->status.hw = false; +amdgpu_ip_block_set_state (ip_block,AMDGPU_IP_STATE_SW ); }/** @@ -3354,7 +3378,7 @@ static void amdgpu_device_smu_fini_early(struct amdgpu_device *adev)return ;for (i =0 ; i < adev->num_ip_blocks; i++) { -if (!adev->ip_blocks[i].status.hw) +if (amdgpu_device_ip_state (adev, i) <AMDGPU_IP_STATE_HW )continue ;if (adev->ip_blocks[i].version->type ==AMD_IP_BLOCK_TYPE_SMC ) {amdgpu_ip_block_hw_fini (&adev->ip_blocks[i]);@@ -3377,7 +3401,7 @@ static int amdgpu_device_ip_fini_early(struct amdgpu_device *adev) adev->ip_blocks[i].version->funcs->name, r); } - adev->ip_blocks[i].status.late_initialized = false; +amdgpu_device_ip_set_state (adev, i,AMDGPU_IP_STATE_HW ); }amdgpu_device_set_pg_state (adev,AMD_PG_STATE_UNGATE );@@ -3389,7 +3413,7 @@ static int amdgpu_device_ip_fini_early(struct amdgpu_device *adev)amdgpu_device_smu_fini_early (adev);for (i = adev->num_ip_blocks -1 ; i >=0 ; i--) { -if (!adev->ip_blocks[i].status.hw) +if (amdgpu_device_ip_state (adev, i) <AMDGPU_IP_STATE_HW )continue ;amdgpu_ip_block_hw_fini (&adev->ip_blocks[i]);@@ -3427,7 +3451,7 @@ static int amdgpu_device_ip_fini(struct amdgpu_device *adev)amdgpu_amdkfd_device_fini_sw (adev);for (i = adev->num_ip_blocks -1 ; i >=0 ; i--) { -if (!adev->ip_blocks[i].status.sw) +if (amdgpu_device_ip_state (adev, i) <AMDGPU_IP_STATE_SW )continue ;if (adev->ip_blocks[i].version->type ==AMD_IP_BLOCK_TYPE_GMC ) {@@ -3446,17 +3470,17 @@ static int amdgpu_device_ip_fini(struct amdgpu_device *adev) adev->ip_blocks[i].version->funcs->name, r); } } - adev->ip_blocks[i].status.sw = false; +amdgpu_device_ip_set_state (adev, i,AMDGPU_IP_STATE_EARLY ); }amdgpu_ras_fini (adev);for (i = adev->num_ip_blocks -1 ; i >=0 ; i--) { -if (!adev->ip_blocks[i].status.valid) +if (!amdgpu_device_ip_valid (adev, i))continue ;if (adev->ip_blocks[i].version->funcs->late_fini) adev->ip_blocks[i].version->funcs->late_fini (&adev->ip_blocks[i]); - adev->ip_blocks[i].status.valid = false; +amdgpu_device_ip_set_state (adev, i,AMDGPU_IP_STATE_INVALID ); }return 0 ;@@ -3517,9 +3541,9 @@ static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev)dev_warn (adev->dev,"Failed to disallow df cstate" );for (i = adev->num_ip_blocks -1 ; i >=0 ; i--) { -if (!adev->ip_blocks[i].status.valid) +if (!amdgpu_device_ip_valid (adev, i))continue ; -if (!adev->ip_blocks[i].status.late_initialized) +if (amdgpu_device_ip_state (adev, i) <AMDGPU_IP_STATE_LATE )continue ;if (adev->ip_blocks[i].version->funcs->early_fini) {@@ -3531,11 +3555,11 @@ static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev) } } - adev->ip_blocks[i].status.late_initialized = false; +amdgpu_device_ip_set_state (adev, i,AMDGPU_IP_STATE_HW ); }for (i = adev->num_ip_blocks -1 ; i >=0 ; i--) { -if (!adev->ip_blocks[i].status.valid) +if (!amdgpu_device_ip_valid (adev, i))continue ;/* displays are handled separately */ @@ -3570,7 +3594,7 @@ static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev)amdgpu_dpm_gfx_state_change (adev, sGpuChangeState_D3Entry);for (i = adev->num_ip_blocks -1 ; i >=0 ; i--) { -if (!adev->ip_blocks[i].status.valid) +if (!amdgpu_device_ip_valid (adev, i))continue ;/* displays are handled in phase1 */ if (adev->ip_blocks[i].version->type ==AMD_IP_BLOCK_TYPE_DCE )@@ -3578,7 +3602,7 @@ static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev)/* PSP lost connection when err_event_athub occurs */ if (amdgpu_ras_intr_triggered () && adev->ip_blocks[i].version->type ==AMD_IP_BLOCK_TYPE_PSP ) { - adev->ip_blocks[i].status.hw = false; +amdgpu_device_ip_set_state (adev, i,AMDGPU_IP_STATE_SW );continue ; }@@ -3620,7 +3644,7 @@ static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev)/* XXX handle errors */ r =amdgpu_ip_block_suspend (&adev->ip_blocks[i]); - adev->ip_blocks[i].status.hw = false; +amdgpu_device_ip_set_state (adev, i,AMDGPU_IP_STATE_SW );/* handle putting the SMC in the appropriate state */ if (!amdgpu_sriov_vf (adev)) {@@ -3687,12 +3711,12 @@ static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)struct amdgpu_ip_block *block; block = &adev->ip_blocks[i]; - block->status.hw = false; +amdgpu_device_ip_set_state (adev, i,AMDGPU_IP_STATE_SW );for (j =0 ; j <ARRAY_SIZE (ip_order); j++) {if (block->version->type != ip_order[j] || - !block->status.valid) + !amdgpu_device_ip_valid (adev, i))continue ; r = block->version->funcs->hw_init (&adev->ip_blocks[i]);@@ -3701,7 +3725,7 @@ static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev) block->version->funcs->name);return r; } - block->status.hw = true; +amdgpu_device_ip_set_state (adev, i,AMDGPU_IP_STATE_HW ); } }@@ -3731,7 +3755,8 @@ static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)if (!block)continue ; -if (block->status.valid && !block->status.hw) { +if (amdgpu_ip_block_valid (block) && +amdgpu_ip_block_state (block) <AMDGPU_IP_STATE_HW ) {if (block->version->type ==AMD_IP_BLOCK_TYPE_SMC ) { r =amdgpu_ip_block_resume (block); }else {@@ -3743,7 +3768,7 @@ static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev) block->version->funcs->name);break ; } - block->status.hw = true; +amdgpu_ip_block_set_state (block,AMDGPU_IP_STATE_HW ); } }@@ -3767,7 +3792,8 @@ static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev)int i, r;for (i =0 ; i < adev->num_ip_blocks; i++) { -if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw) +if (!amdgpu_device_ip_valid (adev, i) || +amdgpu_device_ip_state (adev, i) >=AMDGPU_IP_STATE_HW )continue ; if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||@@ -3801,7 +3827,8 @@ static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev)int i, r;for (i =0 ; i < adev->num_ip_blocks; i++) { -if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw) +if (!amdgpu_device_ip_valid (adev, i) || +amdgpu_device_ip_state (adev, i) >=AMDGPU_IP_STATE_HW )continue ; if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||@@ -3835,7 +3862,8 @@ static int amdgpu_device_ip_resume_phase3(struct amdgpu_device *adev)int i, r;for (i =0 ; i < adev->num_ip_blocks; i++) { -if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw) +if (!amdgpu_device_ip_valid (adev, i) || +amdgpu_device_ip_state (adev, i) >=AMDGPU_IP_STATE_HW )continue ;if (adev->ip_blocks[i].version->type ==AMD_IP_BLOCK_TYPE_DCE ) { r =amdgpu_ip_block_resume (&adev->ip_blocks[i]);@@ -4888,7 +4916,7 @@ int amdgpu_device_prepare(struct drm_device *dev)flush_delayed_work (&adev->gfx.gfx_off_delay_work);for (i =0 ; i < adev->num_ip_blocks; i++) { -if (!adev->ip_blocks[i].status.valid) +if (!amdgpu_device_ip_valid (adev, i))continue ;if (!adev->ip_blocks[i].version->funcs->prepare_suspend)continue ;@@ -5090,7 +5118,7 @@ static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev)return true;for (i =0 ; i < adev->num_ip_blocks; i++) { -if (!adev->ip_blocks[i].status.valid) +if (!amdgpu_device_ip_valid (adev, i))continue ;if (adev->ip_blocks[i].version->funcs->check_soft_reset) adev->ip_blocks[i].status.hang =@@ -5120,7 +5148,7 @@ static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev)int i, r =0 ;for (i =0 ; i < adev->num_ip_blocks; i++) { -if (!adev->ip_blocks[i].status.valid) +if (!amdgpu_device_ip_valid (adev, i))continue ;if (adev->ip_blocks[i].status.hang && adev->ip_blocks[i].version->funcs->pre_soft_reset) {@@ -5150,7 +5178,7 @@ static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev)return true;for (i =0 ; i < adev->num_ip_blocks; i++) { -if (!adev->ip_blocks[i].status.valid) +if (!amdgpu_device_ip_valid (adev, i))continue ; if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) || (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) ||@@ -5182,7 +5210,7 @@ static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev)int i, r =0 ;for (i =0 ; i < adev->num_ip_blocks; i++) { -if (!adev->ip_blocks[i].status.valid) +if (!amdgpu_device_ip_valid (adev, i))continue ;if (adev->ip_blocks[i].status.hang && adev->ip_blocks[i].version->funcs->soft_reset) {@@ -5211,7 +5239,7 @@ static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev)int i, r =0 ;for (i =0 ; i < adev->num_ip_blocks; i++) { -if (!adev->ip_blocks[i].status.valid) +if (!amdgpu_device_ip_valid (adev, i))continue ;if (adev->ip_blocks[i].status.hang && adev->ip_blocks[i].version->funcs->post_soft_reset)drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +2
-2@@ -515,7 +515,7 @@ static int amdgpu_hw_ip_info(struct amdgpu_device *adev,for (i =0 ; i < adev->num_ip_blocks; i++)if (adev->ip_blocks[i].version->type == type && - adev->ip_blocks[i].status.valid) +amdgpu_device_ip_valid (adev, i))break ;if (i == adev->num_ip_blocks)@@ -636,7 +636,7 @@ int amdgpu_info_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) type =amdgpu_ip_get_block_type (adev, info->query_hw_ip.type); ip_block =amdgpu_device_ip_get_ip_block (adev, type); -if (!ip_block || !ip_block->status.valid) +if (!ip_block || !amdgpu_ip_block_valid (ip_block))return -EINVAL ; if (adev->xcp_mgr && adev->xcp_mgr->num_xcps > 0 &&drivers/gpu/drm/amd/amdgpu/amdgpu_psp.c +4
-2@@ -3897,7 +3897,8 @@ static ssize_t psp_usbc_pd_fw_sysfs_read(struct device *dev,int ret; ip_block =amdgpu_device_ip_get_ip_block (adev,AMD_IP_BLOCK_TYPE_PSP ); -if (!ip_block || !ip_block->status.late_initialized) { +if (!ip_block || +amdgpu_ip_block_state (ip_block) <AMDGPU_IP_STATE_LATE ) {dev_info (adev->dev,"PSP block is not ready yet\n." );return -EBUSY ; }@@ -3929,7 +3930,8 @@ static ssize_t psp_usbc_pd_fw_sysfs_write(struct device *dev,struct amdgpu_ip_block *ip_block; ip_block =amdgpu_device_ip_get_ip_block (adev,AMD_IP_BLOCK_TYPE_PSP ); -if (!ip_block || !ip_block->status.late_initialized) { +if (!ip_block || +amdgpu_ip_block_state (ip_block) <AMDGPU_IP_STATE_LATE ) {dev_err (adev->dev,"PSP block is not ready yet." );return -EBUSY ; }drivers/gpu/drm/amd/amdgpu/amdgpu_reset.c +3
-3@@ -31,9 +31,9 @@ static int amdgpu_reset_xgmi_reset_on_init_suspend(struct amdgpu_device *adev)int i;for (i = adev->num_ip_blocks -1 ; i >=0 ; i--) { -if (!adev->ip_blocks[i].status.valid) +if (!amdgpu_device_ip_valid (adev, i))continue ; -if (!adev->ip_blocks[i].status.hw) +if (amdgpu_device_ip_state (adev, i) <=AMDGPU_IP_STATE_HW )continue ;/* displays are handled in phase1 */ if (adev->ip_blocks[i].version->type ==AMD_IP_BLOCK_TYPE_DCE )@@ -41,7 +41,7 @@ static int amdgpu_reset_xgmi_reset_on_init_suspend(struct amdgpu_device *adev)/* XXX handle errors */ amdgpu_ip_block_suspend (&adev->ip_blocks[i]); - adev->ip_blocks[i].status.hw = false; +amdgpu_device_ip_set_state (adev, i,AMDGPU_IP_STATE_SW ); }/* VCN FW shared region is in frambuffer, there are some flags drivers/gpu/drm/amd/amdgpu/sienna_cichlid.c +1
-1@@ -204,7 +204,7 @@ static int sienna_cichlid_mode2_restore_ip(struct amdgpu_device *adev)return r; } } - adev->ip_blocks[i].status.late_initialized = true; +amdgpu_device_ip_set_state (adev, i,AMDGPU_IP_STATE_LATE ); }amdgpu_device_set_cg_state (adev,AMD_CG_STATE_GATE );drivers/gpu/drm/amd/amdgpu/smu_v13_0_10.c +1
-1@@ -205,7 +205,7 @@ static int smu_v13_0_10_mode2_restore_ip(struct amdgpu_device *adev)return r; } } - adev->ip_blocks[i].status.late_initialized = true; +amdgpu_device_ip_set_state (adev, i,AMDGPU_IP_STATE_LATE ); }amdgpu_device_set_cg_state (adev,AMD_CG_STATE_GATE );drivers/gpu/drm/amd/pm/swsmu/amdgpu_smu.c +1
-1@@ -230,7 +230,7 @@ static bool is_vcn_enabled(struct amdgpu_device *adev)for (i =0 ; i < adev->num_ip_blocks; i++) {if ((adev->ip_blocks[i].version->type ==AMD_IP_BLOCK_TYPE_VCN || adev->ip_blocks[i].version->type ==AMD_IP_BLOCK_TYPE_JPEG ) && - !adev->ip_blocks[i].status.valid) + !amdgpu_device_ip_valid (adev, i))return false; }
[AMD Official Use Only - AMD Internal Distribution Only]
Thanks for the patches.
We currently have no plans to include RAS programming as part of the device suspend/resume sequence. Instead, we are focusing on a series of clean up patches and a new RAS software module that will eliminate all legacy code/workarounds and also the changes you proposed here. It is not necessary to make this interim change in the upstream.
Regards,
Hawking
-----Original Message-----
From: Jiang Liu <[email protected]>
Sent: Monday, January 13, 2025 09:42
To: Deucher, Alexander <[email protected]>; Koenig, Christian <[email protected]>; Pan, Xinhui <[email protected]>; [email protected]; [email protected]; Khatri, Sunil <[email protected]>; Lazar, Lijo <[email protected]>; Zhang, Hawking <[email protected]>; Limonciello, Mario <[email protected]>; Chen, Xiaogang <[email protected]>; Russell, Kent <[email protected]>; [email protected]; [email protected]
Cc: Jiang Liu <[email protected]>
Subject: [RFC v2 00/15] Enhance device state machine to better support suspend/resume
Recently we were testing suspend/resume functionality with AMD GPUs, we have encountered several resource tracking related bugs, such as double buffer free, use after free and unbalanced irq reference count.
We have tried to solve these issues case by case, but found that may not be the right way. Especially about the unbalanced irq reference count, there will be new issues appear once we fixed the current known issues. After analyzing related source code, we found that there may be some fundamental implementation flaws behind these resource tracking issues.
The amdgpu driver has two major state machines to driver the device management flow, one is for ip blocks, the other is for ras blocks.
The hook points defined in struct amd_ip_funcs for device setup/teardown are symmetric, but the implementation is asymmetric, sometime even ambiguous. The most obvious two issues we noticed are:
1) amdgpu_irq_get() are called from .late_init() but amdgpu_irq_put()
are called from .hw_fini() instead of .early_fini().
2) the way to reset ip_bloc.status.valid/sw/hw/late_initialized doesn't
match the way to set those flags.
When taking device suspend/resume into account, in addition to device probe/remove, things get much more complex. Some issues arise because many suspend/resume implementations directly reuse .hw_init/.hw_fini/ .late_init hook points.
So we try to fix those issues by two enhancements/refinements to current device management state machines.
The first change is to make the ip block state machine and associated status flags work in stack-like way as below:
Callbacks State after successfully execute callback
AMDGPU_IP_STATE_INVALID
.early_init() AMDGPU_IP_STATE_EARLY .sw_init() AMDGPU_IP_STATE_SW .hw_init() AMDGPU_IP_STATE_HW .late_init() AMDGPU_IP_STATE_LATE .early_fini() AMDGPU_IP_STATE_HW .hw_fini() AMDGPU_IP_STATE_SW .sw_fini() AMDGPU_IP_STATE_EARLY .late_fini() AMDGPU_IP_STATE_INVALID
Also do the same thing for ras block state machine, though it's much more simpler.
The second change is fine tune the overall device management work flow as below:
1. amdgpu_driver_load_kms()
amdgpu_device_init()
amdgpu_device_ip_early_init()
ip_blocks[i].early_init()
ip_blocks[i].status.valid = true
amdgpu_device_ip_init()
amdgpu_ras_init()
ip_blocks[i].sw_init()
ip_blocks[i].status.sw = true
ip_blocks[i].hw_init()
ip_blocks[i].status.hw = true
amdgpu_device_ip_late_init()
ip_blocks[i].late_init()
ip_blocks[i].status.late_initialized = true
amdgpu_ras_late_init()
ras_blocks[i].ras_late_init()
amdgpu_ras_feature_enable_on_boot()2. amdgpu_pmops_suspend()/amdgpu_pmops_freeze()/amdgpu_pmops_poweroff()
amdgpu_device_suspend()
amdgpu_ras_early_fini()
ras_blocks[i].ras_early_fini()
amdgpu_ras_feature_disable()
amdgpu_ras_suspend()
amdgpu_ras_disable_all_features()
+++ ip_blocks[i].early_fini()
+++ ip_blocks[i].status.late_initialized = false
ip_blocks[i].suspend()3. amdgpu_pmops_resume()/amdgpu_pmops_thaw()/amdgpu_pmops_restore()
amdgpu_device_resume()
amdgpu_device_ip_resume()
ip_blocks[i].resume()
amdgpu_device_ip_late_init()
ip_blocks[i].late_init()
ip_blocks[i].status.late_initialized = true
amdgpu_ras_late_init()
ras_blocks[i].ras_late_init()
amdgpu_ras_feature_enable_on_boot()
amdgpu_ras_resume()
amdgpu_ras_enable_all_features()4. amdgpu_driver_unload_kms()
amdgpu_device_fini_hw()
amdgpu_ras_early_fini()
ras_blocks[i].ras_early_fini()+++ ip_blocks[i].early_fini()
+++ ip_blocks[i].status.late_initialized = false
ip_blocks[i].hw_fini()
ip_blocks[i].status.hw = false
5. amdgpu_driver_release_kms() amdgpu_device_fini_sw() amdgpu_device_ip_fini() ip_blocks[i].sw_fini() ip_blocks[i].status.sw = false
--- ip_blocks[i].status.valid = false
+++ amdgpu_ras_fini()
ip_blocks[i].late_fini() +++ ip_blocks[i].status.valid = false --- ip_blocks[i].status.late_initialized = false --- amdgpu_ras_fini()
The main changes include:
1) invoke ip_blocks[i].early_fini in amdgpu_pmops_suspend(). Currently there's only one ip block which provides early_fini callback. We have add a check of in_s3 to keep current behavior in function amdgpu_dm_early_fini(). So there should be no functional changes.
2) set ip_blocks[i].status.late_initialized to false after calling callback early_fini. We have auditted all usages of the late_initialized flag and no functional changes found.
3) only set ip_blocks[i].status.valid = false after calling the late_fini callback.
4) call amdgpu_ras_fini() before invoking ip_blocks[i].late_fini.
Then we try to refine each subsystem, such as nbio, asic etc, to follow the new design. Currently we have only taken the nbio and asic as examples to show the proposed changes. Once we have confirmed that's the right way to go, we will handle the lefting subsystems.
This is in early stage and requesting for comments, any comments and suggestions are welcomed!
v2:
- remove patch 1 in v1, it already got merged
- convert status bool flags for ip block into enum
- introduce iterators to walk ip blocks
- refine the way to define status markers
- split amdgpu_dm related change into a dedicated patch
- add patch 13 to walk ip blocks in reverse order when shutdown
Jiang Liu (15):
drm/amdgpu: add helper functions to track status for ras manager drm/amdgpu: add a flag to track ras debugfs creation status drm/amdgpu: free all resources on error recovery path of
amdgpu_ras_init()
drm/amdgpu: introduce a flag to track refcount held for features drm/amdgpu: enhance amdgpu_ras_block_late_fini() drm/amdgpu: enhance amdgpu_ras_pre_fini() to better support SR drm/admgpu: rename amdgpu_ras_pre_fini() to amdgpu_ras_early_fini() drm/amdgpu: make IP block state machine works in stack like way drm/amdgpu_dm: enhance amdgpu_dm_early_fini() for PM ops drm/admgpu: make device state machine work in stack like way drm/amdgpu: convert ip block bool flags into an enum drm/amdgpu: introduce IP block iterators to reduce duplicated code drm/amdgpu: walk IP blocks in reverse order when shutdown drm/amdgpu/nbio: improve the way to manage irq reference count drm/amdgpu/asic: make ip block operations symmetric by .early_fini()
drivers/gpu/drm/amd/amdgpu/aldebaran.c | 46 +- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 109 +++- .../gpu/drm/amd/amdgpu/amdgpu_dev_coredump.c | 3 +- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 504 +++++++++--------- drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 10 +- drivers/gpu/drm/amd/amdgpu/amdgpu_jpeg.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c | 18 +- drivers/gpu/drm/amd/amdgpu/amdgpu_nbio.c | 16 +- drivers/gpu/drm/amd/amdgpu/amdgpu_nbio.h | 1 + drivers/gpu/drm/amd/amdgpu/amdgpu_psp.c | 4 +- drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c | 142 +++-- drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h | 16 +- drivers/gpu/drm/amd/amdgpu/amdgpu_reset.c | 14 +- drivers/gpu/drm/amd/amdgpu/amdgpu_sdma.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_vcn.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_xgmi.c | 2 +- drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c | 2 +- drivers/gpu/drm/amd/amdgpu/jpeg_v4_0_3.c | 2 +- drivers/gpu/drm/amd/amdgpu/mmhub_v1_8.c | 2 +- drivers/gpu/drm/amd/amdgpu/nbio_v7_4.c | 1 + drivers/gpu/drm/amd/amdgpu/nbio_v7_9.c | 1 + drivers/gpu/drm/amd/amdgpu/nv.c | 14 +- drivers/gpu/drm/amd/amdgpu/sienna_cichlid.c | 50 +- drivers/gpu/drm/amd/amdgpu/smu_v13_0_10.c | 51 +- drivers/gpu/drm/amd/amdgpu/soc15.c | 38 +- drivers/gpu/drm/amd/amdgpu/soc21.c | 35 +- drivers/gpu/drm/amd/amdgpu/soc24.c | 22 +- drivers/gpu/drm/amd/amdgpu/vcn_v4_0_3.c | 2 +- .../gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c | 3 + drivers/gpu/drm/amd/pm/swsmu/amdgpu_smu.c | 10 +- 32 files changed, 668 insertions(+), 460 deletions(-)
--
2.43.5
Is the latest version of this patch series (with possible fixes based on to comments) however maintained/available on some git tree for testing?
Quoted 222 lines · Hawking
[AMD Official Use Only - AMD Internal Distribution Only]
Thanks for the patches.
We currently have no plans to include RAS programming as part of the device suspend/resume sequence. Instead, we are focusing on a series of clean up patches and a new RAS software module that will eliminate all legacy code/workarounds and also the changes you proposed here. It is not necessary to make this interim change in the upstream.
Regards,
Hawking-----Original Message-----
From: Jiang Liu <[email protected]>
Sent: Monday, January 13, 2025 09:42
To: Deucher, Alexander <[email protected]>; Koenig, Christian < [email protected]>; Pan, Xinhui <[email protected]>; [email protected]; [email protected]; Khatri, Sunil <[email protected]>; Lazar, Lijo <[email protected]>; Zhang, Hawking <[email protected]>; Limonciello, Mario <[email protected]>; Chen, Xiaogang < [email protected]>; Russell, Kent <[email protected]>; [email protected]; [email protected]
Cc: Jiang Liu <[email protected]>
Subject: [RFC v2 00/15] Enhance device state machine to better support suspend/resumeRecently we were testing suspend/resume functionality with AMD GPUs, we have encountered several resource tracking related bugs, such as double buffer free, use after free and unbalanced irq reference count.
We have tried to solve these issues case by case, but found that may not be the right way. Especially about the unbalanced irq reference count, there will be new issues appear once we fixed the current known issues. After analyzing related source code, we found that there may be some fundamental implementation flaws behind these resource tracking issues.
The amdgpu driver has two major state machines to driver the device management flow, one is for ip blocks, the other is for ras blocks. The hook points defined in struct
amd_ip_funcsfor device setup/teardown are symmetric, but the implementation is asymmetric, sometime even ambiguous. The most obvious two issues we noticed are:
1)amdgpu_irq_get()are called from .late_init() butamdgpu_irq_put()are called from .hw_fini() instead of .early_fini().
2) the way to reset ip_bloc.status.valid/sw/hw/late_initialized doesn't match the way to set those flags.When taking device suspend/resume into account, in addition to device probe/remove, things get much more complex. Some issues arise because many suspend/resume implementations directly reuse .hw_init/.hw_fini/ .late_init hook points.
So we try to fix those issues by two enhancements/refinements to current device management state machines.
The first change is to make the ip block state machine and associated status flags work in stack-like way as below:
Callbacks State after successfully execute callback
AMDGPU_IP_STATE_INVALID.early_init() AMDGPU_IP_STATE_EARLY .sw_init() AMDGPU_IP_STATE_SW .hw_init() AMDGPU_IP_STATE_HW .late_init() AMDGPU_IP_STATE_LATE .early_fini() AMDGPU_IP_STATE_HW .hw_fini() AMDGPU_IP_STATE_SW .sw_fini() AMDGPU_IP_STATE_EARLY .late_fini() AMDGPU_IP_STATE_INVALIDAlso do the same thing for ras block state machine, though it's much more simpler.
The second change is fine tune the overall device management work flow as below:
1.amdgpu_driver_load_kms()amdgpu_device_init() amdgpu_device_ip_early_init() ip_blocks[i].early_init() ip_blocks[i].status.valid = true amdgpu_device_ip_init() amdgpu_ras_init() ip_blocks[i].sw_init() ip_blocks[i].status.sw = true ip_blocks[i].hw_init() ip_blocks[i].status.hw = true amdgpu_device_ip_late_init() ip_blocks[i].late_init() ip_blocks[i].status.late_initialized = true amdgpu_ras_late_init() ras_blocks[i].ras_late_init() amdgpu_ras_feature_enable_on_boot()2. amdgpu_pmops_suspend()/amdgpu_pmops_freeze()/amdgpu_pmops_poweroff()
amdgpu_device_suspend() amdgpu_ras_early_fini() ras_blocks[i].ras_early_fini() amdgpu_ras_feature_disable() amdgpu_ras_suspend() amdgpu_ras_disable_all_features() +++ ip_blocks[i].early_fini() +++ ip_blocks[i].status.late_initialized = false ip_blocks[i].suspend()3. amdgpu_pmops_resume()/amdgpu_pmops_thaw()/amdgpu_pmops_restore()
amdgpu_device_resume() amdgpu_device_ip_resume() ip_blocks[i].resume() amdgpu_device_ip_late_init() ip_blocks[i].late_init() ip_blocks[i].status.late_initialized = true amdgpu_ras_late_init() ras_blocks[i].ras_late_init() amdgpu_ras_feature_enable_on_boot() amdgpu_ras_resume() amdgpu_ras_enable_all_features()4.
amdgpu_driver_unload_kms()amdgpu_device_fini_hw() amdgpu_ras_early_fini() ras_blocks[i].ras_early_fini()+++ ip_blocks[i].early_fini()
+++ ip_blocks[i].status.late_initialized = false
ip_blocks[i].hw_fini()
ip_blocks[i].status.hw = false5.
amdgpu_driver_release_kms()amdgpu_device_fini_sw()amdgpu_device_ip_fini()ip_blocks[i].sw_fini() ip_blocks[i].status.sw = falseamdgpu_ras_fini() renamed from ip_blocks[i].status.valid = false +1
-2ip_blocks[i].late_fini() +++ ip_blocks[i].status.valid = false --- ip_blocks[i].status.late_initialized = false --- amdgpu_ras_fini()The main changes include:
1) invoke ip_blocks[i].early_fini inamdgpu_pmops_suspend(). Currently there's only one ip block which providesearly_finicallback. We have add a check ofin_s3to keep current behavior in functionamdgpu_dm_early_fini(). So there should be no functional changes.
2) set ip_blocks[i].status.late_initialized to false after calling callbackearly_fini. We have auditted all usages of thelate_initializedflag and no functional changes found.
3) only set ip_blocks[i].status.valid = false after calling thelate_finicallback.
4) callamdgpu_ras_fini()before invoking ip_blocks[i].late_fini.Then we try to refine each subsystem, such as nbio, asic etc, to follow the new design. Currently we have only taken the nbio and asic as examples to show the proposed changes. Once we have confirmed that's the right way to go, we will handle the lefting subsystems.
This is in early stage and requesting for comments, any comments and suggestions are welcomed!
v2:
- remove patch 1 in v1, it already got merged
- convert status bool flags for ip block into enum
- introduce iterators to walk ip blocks
- refine the way to define status markers
- splitamdgpu_dmrelated change into a dedicated patch
- add patch 13 to walk ip blocks in reverse order when shutdownJiang Liu (15):
drm/amdgpu: add helper functions to track status for ras manager drm/amdgpu: add a flag to track ras debugfs creation status drm/amdgpu: free all resources on error recovery path of
amdgpu_ras_init()
drm/amdgpu: introduce a flag to track refcount held for features drm/amdgpu: enhanceamdgpu_ras_block_late_fini()drm/amdgpu: enhanceamdgpu_ras_pre_fini()to better support SR drm/admgpu: renameamdgpu_ras_pre_fini()toamdgpu_ras_early_fini()drm/amdgpu: make IP block state machine works in stack like way drm/amdgpu_dm: enhanceamdgpu_dm_early_fini()for PM ops drm/admgpu: make device state machine work in stack like way drm/amdgpu: convert ip block bool flags into an enum drm/amdgpu: introduce IP block iterators to reduce duplicated code drm/amdgpu: walk IP blocks in reverse order when shutdown drm/amdgpu/nbio: improve the way to manage irq reference count drm/amdgpu/asic: make ip block operations symmetric by .early_fini()drivers/gpu/drm/amd/amdgpu/aldebaran.c | 46 +- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 109 +++- .../gpu/drm/amd/amdgpu/amdgpu_dev_coredump.c | 3 +- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 504 +++++++++--------- drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 10 +- drivers/gpu/drm/amd/amdgpu/amdgpu_jpeg.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c | 18 +- drivers/gpu/drm/amd/amdgpu/amdgpu_nbio.c | 16 +- drivers/gpu/drm/amd/amdgpu/amdgpu_nbio.h | 1 + drivers/gpu/drm/amd/amdgpu/amdgpu_psp.c | 4 +- drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c | 142 +++-- drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h | 16 +- drivers/gpu/drm/amd/amdgpu/amdgpu_reset.c | 14 +- drivers/gpu/drm/amd/amdgpu/amdgpu_sdma.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_vcn.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_xgmi.c | 2 +- drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c | 2 +- drivers/gpu/drm/amd/amdgpu/jpeg_v4_0_3.c | 2 +- drivers/gpu/drm/amd/amdgpu/mmhub_v1_8.c | 2 +- drivers/gpu/drm/amd/amdgpu/nbio_v7_4.c | 1 + drivers/gpu/drm/amd/amdgpu/nbio_v7_9.c | 1 + drivers/gpu/drm/amd/amdgpu/nv.c | 14 +- drivers/gpu/drm/amd/amdgpu/sienna_cichlid.c | 50 +- drivers/gpu/drm/amd/amdgpu/smu_v13_0_10.c | 51 +- drivers/gpu/drm/amd/amdgpu/soc15.c | 38 +- drivers/gpu/drm/amd/amdgpu/soc21.c | 35 +- drivers/gpu/drm/amd/amdgpu/soc24.c | 22 +- drivers/gpu/drm/amd/amdgpu/vcn_v4_0_3.c | 2 +- .../gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c | 3 + drivers/gpu/drm/amd/pm/swsmu/amdgpu_smu.c | 10 +- 32 files changed, 668 insertions(+), 460 deletions(-)--
2.43.5