Skip to contentSkip to conversation

drm/amdkfd: Provide SMI events watch

PATCH Superseded Amber Lin 3 participants 9 messages Mar 17, 2020 – Apr 1, 2020

Subsystems (3)

v1 of 7 Superseded Latest is v7 →

lore
Share & tools
lore

Amber Lin

author
#0 lore raw

PATCH 1/1 drm/amdkfd: Provide SMI events watch

When the compute is malfunctioning or performance drops, the system admin will use SMI (System Management Interface) tool to monitor/diagnostic what went wrong. This patch provides an event watch interface for the user space to register events they are interested. After the event is registered, the user can use annoymous file descriptor's pull function with wait-time specified to wait for the event to happen. Once the event happens, the user can use read() to retrieve information related to the event.

VM fault event is done in this patch.

---
 drivers/gpu/drm/amd/amdkfd/Makefile              |   3 +-
 drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c |   2 +
 drivers/gpu/drm/amd/amdkfd/kfd_chardev.c         |  38 ++++++
 drivers/gpu/drm/amd/amdkfd/kfd_device.c          |   1 +
 drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c  |   2 +
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h            |  10 ++
 drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c      | 143 +++++++++++++++++++++++
 drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h      |  41 +++++++
 include/uapi/linux/kfd_ioctl.h                   |  27 ++++-
 9 files changed, 265 insertions(+), 2 deletions(-)
 create mode 100644 drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c
 create mode 100644 drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h
drivers/gpu/drm/amd/amdkfd/Makefile +2 -1
index 6147462..cc98b4a 100644
@@ -53,7 +53,8 @@ AMDKFD_FILES	:= $(AMDKFD_PATH)/kfd_module.o \
 		$(AMDKFD_PATH)/kfd_int_process_v9.o \
 		$(AMDKFD_PATH)/kfd_dbgdev.o \
 		$(AMDKFD_PATH)/kfd_dbgmgr.o \
-		$(AMDKFD_PATH)/kfd_crat.o
+		$(AMDKFD_PATH)/kfd_crat.o \
+		$(AMDKFD_PATH)/kfd_smi_events.o
 
 ifneq ($(CONFIG_AMD_IOMMU_V2),)
 AMDKFD_FILES += $(AMDKFD_PATH)/kfd_iommu.o
drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c +2 -0
index 9f59ba9..24b4717 100644
@@ -24,6 +24,7 @@
 #include "kfd_events.h"
 #include "cik_int.h"
 #include "amdgpu_amdkfd.h"
+#include "kfd_smi_events.h"
 
 static bool cik_event_interrupt_isr(struct kfd_dev *dev,
 					const uint32_t *ih_ring_entry,
@@ -107,6 +108,7 @@ static void cik_event_interrupt_wq(struct kfd_dev *dev,
 		ihre->source_id == CIK_INTSRC_GFX_MEM_PROT_FAULT) {
 		struct kfd_vm_fault_info info;
 
+		kfd_smi_event_update_vmfault(dev, pasid);
 		kfd_process_vm_fault(dev->dqm, pasid);
 
 		memset(&info, 0, sizeof(info));
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +38 -0
index f8fa03a..8e92956 100644
@@ -39,6 +39,7 @@
 #include "kfd_device_queue_manager.h"
 #include "kfd_dbgmgr.h"
 #include "amdgpu_amdkfd.h"
+#include "kfd_smi_events.h"
 
 static long kfd_ioctl(struct file *, unsigned int, unsigned long);
 static int kfd_open(struct inode *, struct file *);
@@ -1243,6 +1244,40 @@ static int kfd_ioctl_acquire_vm(struct file *filep, struct kfd_process *p,
 	return ret;
 }
 
+/* Handle requests for watching SMI events */
+static int kfd_ioctl_smi_events(struct file *filep,
+				struct kfd_process *p, void *data)
+{
+	struct kfd_ioctl_smi_events_args *args = data;
+	struct kfd_dev *dev;
+	int ret = 0;
+
+	dev = kfd_device_by_id(args->gpu_id);
+	if (!dev)
+		return -EINVAL;
+
+	switch (args->op) {
+	case KFD_SMI_EVENTS_REGISTER:
+		ret = kfd_smi_event_register(dev, args->events);
+		if (ret >= 0) {
+			/* When the registration is successful, it returns the
+			 * annoymous inode. Pass it to the user in data1
+			 */
+			args->data1 = ret;
+			ret = 0;
+		}
+		break;
+	case KFD_SMI_EVENTS_UNREGISTER:
+		kfd_smi_event_unregister(dev, args->events);
+		break;
+	default:
+		ret = -EINVAL;
+		break;
+	}
+
+	return ret;
+}
+
 bool kfd_dev_is_large_bar(struct kfd_dev *dev)
 {
 	struct kfd_local_mem_info mem_info;
@@ -1827,6 +1862,9 @@ static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
 
 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_QUEUE_GWS,
 			kfd_ioctl_alloc_queue_gws, 0),
+
+	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SMI_EVENTS,
+			kfd_ioctl_smi_events, 0),
 };
 
 #define AMDKFD_CORE_IOCTL_COUNT	ARRAY_SIZE(amdkfd_ioctls)
drivers/gpu/drm/amd/amdkfd/kfd_device.c +1 -0
index 7866cd06..450368c 100644
@@ -532,6 +532,7 @@ struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd,
 	kfd->device_info = device_info;
 	kfd->pdev = pdev;
 	kfd->init_complete = false;
+	kfd->smi_events.events = 0;
 	kfd->kfd2kgd = f2g;
 	atomic_set(&kfd->compute_profile, 0);
 
drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c +2 -0
index e05d75e..151e83e 100644
@@ -24,6 +24,7 @@
 #include "kfd_events.h"
 #include "soc15_int.h"
 #include "kfd_device_queue_manager.h"
+#include "kfd_smi_events.h"
 
 static bool event_interrupt_isr_v9(struct kfd_dev *dev,
 					const uint32_t *ih_ring_entry,
@@ -117,6 +118,7 @@ static void event_interrupt_wq_v9(struct kfd_dev *dev,
 		info.prot_read  = ring_id & 0x10;
 		info.prot_write = ring_id & 0x20;
 
+		kfd_smi_event_update_vmfault(dev, pasid);
 		kfd_process_vm_fault(dev->dqm, pasid);
 		kfd_signal_vm_fault_event(dev, pasid, &info);
 	}
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +10 -0
index 43b888b..fdb51de 100644
@@ -223,6 +223,13 @@ struct kfd_vmid_info {
 	uint32_t vmid_num_kfd;
 };
 
+struct kfd_smi_events {
+	uint64_t events;
+	struct kfifo fifo;
+	wait_queue_head_t wait_queue;
+	uint32_t max_events;
+};
+
 struct kfd_dev {
 	struct kgd_dev *kgd;
 
@@ -309,6 +316,9 @@ struct kfd_dev {
 
 	/* Global GWS resource shared b/t processes*/
 	void *gws;
+
+	/* if this device is in SMI events watch */
+	struct kfd_smi_events smi_events;
 };
 
 enum kfd_mempool {
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c new file +143 -0

143 lines not on this page · raw message

drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h new file +41 -0

41 lines not on this page · raw message

include/uapi/linux/kfd_ioctl.h +26 -1
index 4f66764..6ce7c69 100644
@@ -442,6 +442,28 @@ struct kfd_ioctl_import_dmabuf_args {
 	__u32 dmabuf_fd;	/* to KFD */
 };
 
+/*
+ * KFD SMI(System Management Interface) events
+ */
+enum kfd_smi_events_op {
+	KFD_SMI_EVENTS_REGISTER = 1,
+	KFD_SMI_EVENTS_UNREGISTER
+};
+
+/* Event ID (mask) */
+#define KFD_SMI_EV_VMFAULT     0x00000001
+
+struct kfd_ioctl_smi_events_args {
+	__u32 op;       /* to KFD */
+	/* upper 32 bits: group. lower 32 bits: event IDs */
+	__u64 events;   /* to KFD */
+	__u32 gpu_id;   /* to KFD */
+	pid_t pid;      /* to KFD */
+	__u32 data1;    /* from KFD */
+	__u32 data2;
+	__u32 data3;
+};
+
 /* Register offset inside the remapped mmio page
  */
 enum kfd_mmio_remap {
@@ -546,7 +568,10 @@ enum kfd_mmio_remap {
 #define AMDKFD_IOC_ALLOC_QUEUE_GWS		\
 		AMDKFD_IOWR(0x1E, struct kfd_ioctl_alloc_queue_gws_args)
 
+#define AMDKFD_IOC_SMI_EVENTS			\
+		AMDKFD_IOWR(0x1F, struct kfd_ioctl_smi_events_args)
+
 #define AMDKFD_COMMAND_START		0x01
-#define AMDKFD_COMMAND_END		0x1F
+#define AMDKFD_COMMAND_END		0x20
 
 #endif
Signature · 6 lines
2.7.4

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
#1 lore raw
Quoted 10 lines · Amber Lin

When the compute is malfunctioning or performance drops, the system admin will use SMI (System Management Interface) tool to monitor/diagnostic what went wrong. This patch provides an event watch interface for the user space to register events they are interested. After the event is registered, the user can use annoymous file descriptor's pull function with wait-time specified to wait for the event to happen. Once the event happens, the user can use read() to retrieve information related to the event.

Quoted 3 lines

VM fault event is done in this patch.

Can you provide a link to the userspace tools that make use of this interface?

Thanks,

Alex

Quoted 415 lines
---
 drivers/gpu/drm/amd/amdkfd/Makefile              |   3 +-
 drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c |   2 +
 drivers/gpu/drm/amd/amdkfd/kfd_chardev.c         |  38 ++++++
 drivers/gpu/drm/amd/amdkfd/kfd_device.c          |   1 +
 drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c  |   2 +
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h            |  10 ++
 drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c      | 143 +++++++++++++++++++++++
 drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h      |  41 +++++++
 include/uapi/linux/kfd_ioctl.h                   |  27 ++++-
 9 files changed, 265 insertions(+), 2 deletions(-)
 create mode 100644 drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c
 create mode 100644 drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h
drivers/gpu/drm/amd/amdkfd/Makefile +2 -1
index 6147462..cc98b4a 100644
@@ -53,7 +53,8 @@ AMDKFD_FILES  := $(AMDKFD_PATH)/kfd_module.o \
                $(AMDKFD_PATH)/kfd_int_process_v9.o \
                $(AMDKFD_PATH)/kfd_dbgdev.o \
                $(AMDKFD_PATH)/kfd_dbgmgr.o \
-               $(AMDKFD_PATH)/kfd_crat.o
+               $(AMDKFD_PATH)/kfd_crat.o \
+               $(AMDKFD_PATH)/kfd_smi_events.o
 
 ifneq ($(CONFIG_AMD_IOMMU_V2),)
 AMDKFD_FILES += $(AMDKFD_PATH)/kfd_iommu.o
drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c +2 -0
index 9f59ba9..24b4717 100644
@@ -24,6 +24,7 @@
 #include "kfd_events.h"
 #include "cik_int.h"
 #include "amdgpu_amdkfd.h"
+#include "kfd_smi_events.h"
 
 static bool cik_event_interrupt_isr(struct kfd_dev *dev,
                                        const uint32_t *ih_ring_entry,
@@ -107,6 +108,7 @@ static void cik_event_interrupt_wq(struct kfd_dev *dev,
                ihre->source_id == CIK_INTSRC_GFX_MEM_PROT_FAULT) {
                struct kfd_vm_fault_info info;
 
+               kfd_smi_event_update_vmfault(dev, pasid);
                kfd_process_vm_fault(dev->dqm, pasid);
 
                memset(&info, 0, sizeof(info));
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +38 -0
index f8fa03a..8e92956 100644
@@ -39,6 +39,7 @@
 #include "kfd_device_queue_manager.h"
 #include "kfd_dbgmgr.h"
 #include "amdgpu_amdkfd.h"
+#include "kfd_smi_events.h"
 
 static long kfd_ioctl(struct file *, unsigned int, unsigned long);
 static int kfd_open(struct inode *, struct file *);
@@ -1243,6 +1244,40 @@ static int kfd_ioctl_acquire_vm(struct file *filep, struct kfd_process *p,
        return ret;
 }
 
+/* Handle requests for watching SMI events */
+static int kfd_ioctl_smi_events(struct file *filep,
+                               struct kfd_process *p, void *data)
+{
+       struct kfd_ioctl_smi_events_args *args = data;
+       struct kfd_dev *dev;
+       int ret = 0;
+
+       dev = kfd_device_by_id(args->gpu_id);
+       if (!dev)
+               return -EINVAL;
+
+       switch (args->op) {
+       case KFD_SMI_EVENTS_REGISTER:
+               ret = kfd_smi_event_register(dev, args->events);
+               if (ret >= 0) {
+                       /* When the registration is successful, it returns the
+                        * annoymous inode. Pass it to the user in data1
+                        */
+                       args->data1 = ret;
+                       ret = 0;
+               }
+               break;
+       case KFD_SMI_EVENTS_UNREGISTER:
+               kfd_smi_event_unregister(dev, args->events);
+               break;
+       default:
+               ret = -EINVAL;
+               break;
+       }
+
+       return ret;
+}
+
 bool kfd_dev_is_large_bar(struct kfd_dev *dev)
 {
        struct kfd_local_mem_info mem_info;
@@ -1827,6 +1862,9 @@ static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
 
        AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_QUEUE_GWS,
                        kfd_ioctl_alloc_queue_gws, 0),
+
+       AMDKFD_IOCTL_DEF(AMDKFD_IOC_SMI_EVENTS,
+                       kfd_ioctl_smi_events, 0),
 };
 
 #define AMDKFD_CORE_IOCTL_COUNT        ARRAY_SIZE(amdkfd_ioctls)
drivers/gpu/drm/amd/amdkfd/kfd_device.c +1 -0
index 7866cd06..450368c 100644
@@ -532,6 +532,7 @@ struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd,
        kfd->device_info = device_info;
        kfd->pdev = pdev;
        kfd->init_complete = false;
+       kfd->smi_events.events = 0;
        kfd->kfd2kgd = f2g;
        atomic_set(&kfd->compute_profile, 0);
 
drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c +2 -0
index e05d75e..151e83e 100644
@@ -24,6 +24,7 @@
 #include "kfd_events.h"
 #include "soc15_int.h"
 #include "kfd_device_queue_manager.h"
+#include "kfd_smi_events.h"
 
 static bool event_interrupt_isr_v9(struct kfd_dev *dev,
                                        const uint32_t *ih_ring_entry,
@@ -117,6 +118,7 @@ static void event_interrupt_wq_v9(struct kfd_dev *dev,
                info.prot_read  = ring_id & 0x10;
                info.prot_write = ring_id & 0x20;
 
+               kfd_smi_event_update_vmfault(dev, pasid);
                kfd_process_vm_fault(dev->dqm, pasid);
                kfd_signal_vm_fault_event(dev, pasid, &info);
        }
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +10 -0
index 43b888b..fdb51de 100644
@@ -223,6 +223,13 @@ struct kfd_vmid_info {
        uint32_t vmid_num_kfd;
 };
 
+struct kfd_smi_events {
+       uint64_t events;
+       struct kfifo fifo;
+       wait_queue_head_t wait_queue;
+       uint32_t max_events;
+};
+
 struct kfd_dev {
        struct kgd_dev *kgd;
 
@@ -309,6 +316,9 @@ struct kfd_dev {
 
        /* Global GWS resource shared b/t processes*/
        void *gws;
+
+       /* if this device is in SMI events watch */
+       struct kfd_smi_events smi_events;
 };
 
 enum kfd_mempool {
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c new file +143 -0
new file mode 100644
index 0000000..ba9d036
@@ -0,0 +1,143 @@
+/*
+ * Copyright 2020 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+#include <linux/poll.h>
+#include <linux/wait.h>
+#include <linux/anon_inodes.h>
+#include <uapi/linux/kfd_ioctl.h>
+#include "amdgpu_vm.h"
+#include "kfd_priv.h"
+#include "kfd_smi_events.h"
+
+static DEFINE_MUTEX(kfd_smi_mutex);
+
+struct mutex *kfd_get_smi_mutex(void)
+{
+       return &kfd_smi_mutex;
+}
+
+static __poll_t kfd_smi_ev_poll(struct file *, struct poll_table_struct *);
+static ssize_t kfd_smi_ev_read(struct file *, char __user *, size_t, loff_t *);
+static int kfd_smi_ev_release(struct inode *, struct file *);
+
+static const char kfd_smi_name[] = "kfd_smi_ev";
+
+static const struct file_operations kfd_smi_ev_fops = {
+       .owner = THIS_MODULE,
+       .poll = kfd_smi_ev_poll,
+       .read = kfd_smi_ev_read,
+       .release = kfd_smi_ev_release
+};
+
+static __poll_t kfd_smi_ev_poll(struct file *filep,
+                               struct poll_table_struct *wait)
+{
+       struct kfd_dev *dev = filep->private_data;
+       struct kfd_smi_events *ev = &dev->smi_events;
+
+       __poll_t mask = 0;
+
+       poll_wait(filep, &ev->wait_queue, wait);
+       mask |= !kfifo_is_empty(&ev->fifo) ? POLLIN | POLLRDNORM : mask;
+
+       return mask;
+}
+
+static ssize_t kfd_smi_ev_read(struct file *filep, char __user *user,
+                              size_t size, loff_t *offset)
+{
+       int ret, copied = 0;
+       struct kfd_dev *dev = filep->private_data;
+
+       ret = kfifo_to_user(&dev->smi_events.fifo, user, size, &copied);
+       if (ret || !copied) {
+               pr_debug("kfd smi-events: Fail to read fd (%i) (%i)\n",
+                               ret, copied);
+               return ret ? ret : -EAGAIN;
+       }
+
+       return copied;
+}
+
+static int kfd_smi_ev_release(struct inode *inode, struct file *filep)
+{
+       struct kfd_dev *dev = filep->private_data;
+
+       kfifo_free(&dev->smi_events.fifo);
+       return 0;
+}
+
+void kfd_smi_event_update_vmfault(struct kfd_dev *kdev, uint16_t pasid)
+{
+       struct kfd_smi_vmfault_fifo fifo_out;
+       struct amdgpu_device *adev = (struct amdgpu_device *)kdev->kgd;
+       struct amdgpu_task_info task_info;
+
+       if (!kdev->smi_events.events)
+               return;
+
+       if (!(kdev->smi_events.events & KFD_SMI_EV_VMFAULT))
+               return;
+
+       memset(&task_info, 0, sizeof(struct amdgpu_task_info));
+       amdgpu_vm_get_task_info(adev, pasid, &task_info);
+
+       fifo_out.group = 0;
+       fifo_out.event = KFD_SMI_EV_VMFAULT;
+       fifo_out.gpu_id = kdev->id;
+       fifo_out.pid = task_info.pid;
+       strcpy(fifo_out.task_name, task_info.task_name);
+       kfifo_in(&kdev->smi_events.fifo, &fifo_out, sizeof(fifo_out));
+       wake_up_all(&kdev->smi_events.wait_queue);
+}
+
+void kfd_smi_event_unregister(struct kfd_dev *dev, uint64_t events)
+{
+       mutex_lock(kfd_get_smi_mutex());
+       dev->smi_events.events &= ~events;
+       mutex_unlock(kfd_get_smi_mutex());
+}
+
+int kfd_smi_event_register(struct kfd_dev *dev, uint64_t events)
+{
+       int ret;
+
+       mutex_lock(kfd_get_smi_mutex());
+       dev->smi_events.events |= events;
+       mutex_unlock(kfd_get_smi_mutex());
+
+       /* We use the lower 32 bits for now. Each bit represents one event. If
+        * featured events are increased to more than 32, we'll use the upper
+        * bits as groups so the total number of events can be up to 32*32.
+        */
+       dev->smi_events.max_events = 32;
+       ret = kfifo_alloc(&dev->smi_events.fifo, dev->smi_events.max_events,
+                        GFP_KERNEL);
+       if (ret) {
+               pr_err("fail to allocate kfifo\n");
+               return ret;
+       }
+       init_waitqueue_head(&dev->smi_events.wait_queue);
+
+       return anon_inode_getfd(kfd_smi_name, &kfd_smi_ev_fops,
+                               (void *)dev, 0);
+}
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h new file +41 -0
new file mode 100644
index 0000000..2e320d3
@@ -0,0 +1,41 @@
+/*
+ * Copyright 2020 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+#ifndef KFD_SMI_EVENTS_H_INCLUDED
+#define KFD_SMI_EVENTS_H_INCLUDED
+
+int kfd_smi_event_register(struct kfd_dev *dev, uint64_t events);
+void kfd_smi_event_unregister(struct kfd_dev *dev, uint64_t events);
+void kfd_smi_event_update_vmfault(struct kfd_dev *dev, uint16_t pasid);
+
+/* All FIFO must start with "uint32_t group" and "uint32_t event" so the user
+ * can read the first 8 bytes to determine the next read length.
+ */
+struct kfd_smi_vmfault_fifo {
+       uint32_t group;
+       uint32_t event;
+       unsigned int gpu_id;
+       pid_t pid;
+       char task_name[TASK_COMM_LEN];
+};
+
+#endif
include/uapi/linux/kfd_ioctl.h +26 -2
index 4f66764..6ce7c69 100644
@@ -442,6 +442,28 @@ struct kfd_ioctl_import_dmabuf_args {
        __u32 dmabuf_fd;        /* to KFD */
 };
 
+/*
+ * KFD SMI(System Management Interface) events
+ */
+enum kfd_smi_events_op {
+       KFD_SMI_EVENTS_REGISTER = 1,
+       KFD_SMI_EVENTS_UNREGISTER
+};
+
+/* Event ID (mask) */
+#define KFD_SMI_EV_VMFAULT     0x00000001
+
+struct kfd_ioctl_smi_events_args {
+       __u32 op;       /* to KFD */
+       /* upper 32 bits: group. lower 32 bits: event IDs */
+       __u64 events;   /* to KFD */
+       __u32 gpu_id;   /* to KFD */
+       pid_t pid;      /* to KFD */
+       __u32 data1;    /* from KFD */
+       __u32 data2;
+       __u32 data3;
+};
+
 /* Register offset inside the remapped mmio page
  */
 enum kfd_mmio_remap {
@@ -546,7 +568,10 @@ enum kfd_mmio_remap {
 #define AMDKFD_IOC_ALLOC_QUEUE_GWS             \
                AMDKFD_IOWR(0x1E, struct kfd_ioctl_alloc_queue_gws_args)
 
+#define AMDKFD_IOC_SMI_EVENTS                  \
+               AMDKFD_IOWR(0x1F, struct kfd_ioctl_smi_events_args)
+
 #define AMDKFD_COMMAND_START           0x01
-#define AMDKFD_COMMAND_END             0x1F
+#define AMDKFD_COMMAND_END             0x20
 
 #endif
--

2.7.4

Quoted 4 lines

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

author
#2 lore raw

[AMD Official Use Only - Internal Distribution Only]

________________________________
From: Alex Deucher <[email protected]>
Sent: Tuesday, March 17, 2020 3:03 PM
To: Lin, Amber <[email protected]>
Cc: amd-gfx list <[email protected]>
Subject: Re: [PATCH] drm/amdkfd: Provide SMI events watch

Quoted 10 lines · Amber Lin

When the compute is malfunctioning or performance drops, the system admin will use SMI (System Management Interface) tool to monitor/diagnostic what went wrong. This patch provides an event watch interface for the user space to register events they are interested. After the event is registered, the user can use annoymous file descriptor's pull function with wait-time specified to wait for the event to happen. Once the event happens, the user can use read() to retrieve information related to the event.

Quoted 3 lines

VM fault event is done in this patch.

Can you provide a link to the userspace tools that make use of this interface?

Thanks,

Alex
=====
Hi Alex,

https://github.com/RadeonOpenCompute/rocm_smi_lib will use this interface. Those functions will be added to this library:

/* Get a handler for watching events */
rsmi_status_t rsmi_event_init(rsmi_event_handle_t *handle); /* Register events for the device using the handler from init */ rsmi_status_t rsmi_event_register(uint32_t dv_ind, uint32_t events,

                                rsmi_event_handle_t *handle);

/* Wait for events. If one of the events happens, a success is returned with
* with details in data.
*/
rsmi_status_t rsmi_event_wait(rsmi_event_handle_t handle, uint32_t timeout_ms,

                                rsmi_event_data_t *data);

/* Stop watching events */
rsmi_status_t rsmi_event_free(rsmi_event_handle_t handle);

Regards,
Amber

Quoted 415 lines
---
 drivers/gpu/drm/amd/amdkfd/Makefile              |   3 +-
 drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c |   2 +
 drivers/gpu/drm/amd/amdkfd/kfd_chardev.c         |  38 ++++++
 drivers/gpu/drm/amd/amdkfd/kfd_device.c          |   1 +
 drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c  |   2 +
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h            |  10 ++
 drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c      | 143 +++++++++++++++++++++++
 drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h      |  41 +++++++
 include/uapi/linux/kfd_ioctl.h                   |  27 ++++-
 9 files changed, 265 insertions(+), 2 deletions(-)
 create mode 100644 drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c
 create mode 100644 drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h
drivers/gpu/drm/amd/amdkfd/Makefile +2 -1
index 6147462..cc98b4a 100644
@@ -53,7 +53,8 @@ AMDKFD_FILES  := $(AMDKFD_PATH)/kfd_module.o \
                $(AMDKFD_PATH)/kfd_int_process_v9.o \
                $(AMDKFD_PATH)/kfd_dbgdev.o \
                $(AMDKFD_PATH)/kfd_dbgmgr.o \
-               $(AMDKFD_PATH)/kfd_crat.o
+               $(AMDKFD_PATH)/kfd_crat.o \
+               $(AMDKFD_PATH)/kfd_smi_events.o
 
 ifneq ($(CONFIG_AMD_IOMMU_V2),)
 AMDKFD_FILES += $(AMDKFD_PATH)/kfd_iommu.o
drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c +2 -0
index 9f59ba9..24b4717 100644
@@ -24,6 +24,7 @@
 #include "kfd_events.h"
 #include "cik_int.h"
 #include "amdgpu_amdkfd.h"
+#include "kfd_smi_events.h"
 
 static bool cik_event_interrupt_isr(struct kfd_dev *dev,
                                        const uint32_t *ih_ring_entry,
@@ -107,6 +108,7 @@ static void cik_event_interrupt_wq(struct kfd_dev *dev,
                ihre->source_id == CIK_INTSRC_GFX_MEM_PROT_FAULT) {
                struct kfd_vm_fault_info info;
 
+               kfd_smi_event_update_vmfault(dev, pasid);
                kfd_process_vm_fault(dev->dqm, pasid);
 
                memset(&info, 0, sizeof(info));
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +38 -0
index f8fa03a..8e92956 100644
@@ -39,6 +39,7 @@
 #include "kfd_device_queue_manager.h"
 #include "kfd_dbgmgr.h"
 #include "amdgpu_amdkfd.h"
+#include "kfd_smi_events.h"
 
 static long kfd_ioctl(struct file *, unsigned int, unsigned long);
 static int kfd_open(struct inode *, struct file *);
@@ -1243,6 +1244,40 @@ static int kfd_ioctl_acquire_vm(struct file *filep, struct kfd_process *p,
        return ret;
 }
 
+/* Handle requests for watching SMI events */
+static int kfd_ioctl_smi_events(struct file *filep,
+                               struct kfd_process *p, void *data)
+{
+       struct kfd_ioctl_smi_events_args *args = data;
+       struct kfd_dev *dev;
+       int ret = 0;
+
+       dev = kfd_device_by_id(args->gpu_id);
+       if (!dev)
+               return -EINVAL;
+
+       switch (args->op) {
+       case KFD_SMI_EVENTS_REGISTER:
+               ret = kfd_smi_event_register(dev, args->events);
+               if (ret >= 0) {
+                       /* When the registration is successful, it returns the
+                        * annoymous inode. Pass it to the user in data1
+                        */
+                       args->data1 = ret;
+                       ret = 0;
+               }
+               break;
+       case KFD_SMI_EVENTS_UNREGISTER:
+               kfd_smi_event_unregister(dev, args->events);
+               break;
+       default:
+               ret = -EINVAL;
+               break;
+       }
+
+       return ret;
+}
+
 bool kfd_dev_is_large_bar(struct kfd_dev *dev)
 {
        struct kfd_local_mem_info mem_info;
@@ -1827,6 +1862,9 @@ static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
 
        AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_QUEUE_GWS,
                        kfd_ioctl_alloc_queue_gws, 0),
+
+       AMDKFD_IOCTL_DEF(AMDKFD_IOC_SMI_EVENTS,
+                       kfd_ioctl_smi_events, 0),
 };
 
 #define AMDKFD_CORE_IOCTL_COUNT        ARRAY_SIZE(amdkfd_ioctls)
drivers/gpu/drm/amd/amdkfd/kfd_device.c +1 -0
index 7866cd06..450368c 100644
@@ -532,6 +532,7 @@ struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd,
        kfd->device_info = device_info;
        kfd->pdev = pdev;
        kfd->init_complete = false;
+       kfd->smi_events.events = 0;
        kfd->kfd2kgd = f2g;
        atomic_set(&kfd->compute_profile, 0);
 
drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c +2 -0
index e05d75e..151e83e 100644
@@ -24,6 +24,7 @@
 #include "kfd_events.h"
 #include "soc15_int.h"
 #include "kfd_device_queue_manager.h"
+#include "kfd_smi_events.h"
 
 static bool event_interrupt_isr_v9(struct kfd_dev *dev,
                                        const uint32_t *ih_ring_entry,
@@ -117,6 +118,7 @@ static void event_interrupt_wq_v9(struct kfd_dev *dev,
                info.prot_read  = ring_id & 0x10;
                info.prot_write = ring_id & 0x20;
 
+               kfd_smi_event_update_vmfault(dev, pasid);
                kfd_process_vm_fault(dev->dqm, pasid);
                kfd_signal_vm_fault_event(dev, pasid, &info);
        }
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +10 -0
index 43b888b..fdb51de 100644
@@ -223,6 +223,13 @@ struct kfd_vmid_info {
        uint32_t vmid_num_kfd;
 };
 
+struct kfd_smi_events {
+       uint64_t events;
+       struct kfifo fifo;
+       wait_queue_head_t wait_queue;
+       uint32_t max_events;
+};
+
 struct kfd_dev {
        struct kgd_dev *kgd;
 
@@ -309,6 +316,9 @@ struct kfd_dev {
 
        /* Global GWS resource shared b/t processes*/
        void *gws;
+
+       /* if this device is in SMI events watch */
+       struct kfd_smi_events smi_events;
 };
 
 enum kfd_mempool {
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c new file +143 -0
new file mode 100644
index 0000000..ba9d036
@@ -0,0 +1,143 @@
+/*
+ * Copyright 2020 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+#include <linux/poll.h>
+#include <linux/wait.h>
+#include <linux/anon_inodes.h>
+#include <uapi/linux/kfd_ioctl.h>
+#include "amdgpu_vm.h"
+#include "kfd_priv.h"
+#include "kfd_smi_events.h"
+
+static DEFINE_MUTEX(kfd_smi_mutex);
+
+struct mutex *kfd_get_smi_mutex(void)
+{
+       return &kfd_smi_mutex;
+}
+
+static __poll_t kfd_smi_ev_poll(struct file *, struct poll_table_struct *);
+static ssize_t kfd_smi_ev_read(struct file *, char __user *, size_t, loff_t *);
+static int kfd_smi_ev_release(struct inode *, struct file *);
+
+static const char kfd_smi_name[] = "kfd_smi_ev";
+
+static const struct file_operations kfd_smi_ev_fops = {
+       .owner = THIS_MODULE,
+       .poll = kfd_smi_ev_poll,
+       .read = kfd_smi_ev_read,
+       .release = kfd_smi_ev_release
+};
+
+static __poll_t kfd_smi_ev_poll(struct file *filep,
+                               struct poll_table_struct *wait)
+{
+       struct kfd_dev *dev = filep->private_data;
+       struct kfd_smi_events *ev = &dev->smi_events;
+
+       __poll_t mask = 0;
+
+       poll_wait(filep, &ev->wait_queue, wait);
+       mask |= !kfifo_is_empty(&ev->fifo) ? POLLIN | POLLRDNORM : mask;
+
+       return mask;
+}
+
+static ssize_t kfd_smi_ev_read(struct file *filep, char __user *user,
+                              size_t size, loff_t *offset)
+{
+       int ret, copied = 0;
+       struct kfd_dev *dev = filep->private_data;
+
+       ret = kfifo_to_user(&dev->smi_events.fifo, user, size, &copied);
+       if (ret || !copied) {
+               pr_debug("kfd smi-events: Fail to read fd (%i) (%i)\n",
+                               ret, copied);
+               return ret ? ret : -EAGAIN;
+       }
+
+       return copied;
+}
+
+static int kfd_smi_ev_release(struct inode *inode, struct file *filep)
+{
+       struct kfd_dev *dev = filep->private_data;
+
+       kfifo_free(&dev->smi_events.fifo);
+       return 0;
+}
+
+void kfd_smi_event_update_vmfault(struct kfd_dev *kdev, uint16_t pasid)
+{
+       struct kfd_smi_vmfault_fifo fifo_out;
+       struct amdgpu_device *adev = (struct amdgpu_device *)kdev->kgd;
+       struct amdgpu_task_info task_info;
+
+       if (!kdev->smi_events.events)
+               return;
+
+       if (!(kdev->smi_events.events & KFD_SMI_EV_VMFAULT))
+               return;
+
+       memset(&task_info, 0, sizeof(struct amdgpu_task_info));
+       amdgpu_vm_get_task_info(adev, pasid, &task_info);
+
+       fifo_out.group = 0;
+       fifo_out.event = KFD_SMI_EV_VMFAULT;
+       fifo_out.gpu_id = kdev->id;
+       fifo_out.pid = task_info.pid;
+       strcpy(fifo_out.task_name, task_info.task_name);
+       kfifo_in(&kdev->smi_events.fifo, &fifo_out, sizeof(fifo_out));
+       wake_up_all(&kdev->smi_events.wait_queue);
+}
+
+void kfd_smi_event_unregister(struct kfd_dev *dev, uint64_t events)
+{
+       mutex_lock(kfd_get_smi_mutex());
+       dev->smi_events.events &= ~events;
+       mutex_unlock(kfd_get_smi_mutex());
+}
+
+int kfd_smi_event_register(struct kfd_dev *dev, uint64_t events)
+{
+       int ret;
+
+       mutex_lock(kfd_get_smi_mutex());
+       dev->smi_events.events |= events;
+       mutex_unlock(kfd_get_smi_mutex());
+
+       /* We use the lower 32 bits for now. Each bit represents one event. If
+        * featured events are increased to more than 32, we'll use the upper
+        * bits as groups so the total number of events can be up to 32*32.
+        */
+       dev->smi_events.max_events = 32;
+       ret = kfifo_alloc(&dev->smi_events.fifo, dev->smi_events.max_events,
+                        GFP_KERNEL);
+       if (ret) {
+               pr_err("fail to allocate kfifo\n");
+               return ret;
+       }
+       init_waitqueue_head(&dev->smi_events.wait_queue);
+
+       return anon_inode_getfd(kfd_smi_name, &kfd_smi_ev_fops,
+                               (void *)dev, 0);
+}
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h new file +41 -0
new file mode 100644
index 0000000..2e320d3
@@ -0,0 +1,41 @@
+/*
+ * Copyright 2020 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+#ifndef KFD_SMI_EVENTS_H_INCLUDED
+#define KFD_SMI_EVENTS_H_INCLUDED
+
+int kfd_smi_event_register(struct kfd_dev *dev, uint64_t events);
+void kfd_smi_event_unregister(struct kfd_dev *dev, uint64_t events);
+void kfd_smi_event_update_vmfault(struct kfd_dev *dev, uint16_t pasid);
+
+/* All FIFO must start with "uint32_t group" and "uint32_t event" so the user
+ * can read the first 8 bytes to determine the next read length.
+ */
+struct kfd_smi_vmfault_fifo {
+       uint32_t group;
+       uint32_t event;
+       unsigned int gpu_id;
+       pid_t pid;
+       char task_name[TASK_COMM_LEN];
+};
+
+#endif
include/uapi/linux/kfd_ioctl.h +26 -2
index 4f66764..6ce7c69 100644
@@ -442,6 +442,28 @@ struct kfd_ioctl_import_dmabuf_args {
        __u32 dmabuf_fd;        /* to KFD */
 };
 
+/*
+ * KFD SMI(System Management Interface) events
+ */
+enum kfd_smi_events_op {
+       KFD_SMI_EVENTS_REGISTER = 1,
+       KFD_SMI_EVENTS_UNREGISTER
+};
+
+/* Event ID (mask) */
+#define KFD_SMI_EV_VMFAULT     0x00000001
+
+struct kfd_ioctl_smi_events_args {
+       __u32 op;       /* to KFD */
+       /* upper 32 bits: group. lower 32 bits: event IDs */
+       __u64 events;   /* to KFD */
+       __u32 gpu_id;   /* to KFD */
+       pid_t pid;      /* to KFD */
+       __u32 data1;    /* from KFD */
+       __u32 data2;
+       __u32 data3;
+};
+
 /* Register offset inside the remapped mmio page
  */
 enum kfd_mmio_remap {
@@ -546,7 +568,10 @@ enum kfd_mmio_remap {
 #define AMDKFD_IOC_ALLOC_QUEUE_GWS             \
                AMDKFD_IOWR(0x1E, struct kfd_ioctl_alloc_queue_gws_args)
 
+#define AMDKFD_IOC_SMI_EVENTS                  \
+               AMDKFD_IOWR(0x1F, struct kfd_ioctl_smi_events_args)
+
 #define AMDKFD_COMMAND_START           0x01
-#define AMDKFD_COMMAND_END             0x1F
+#define AMDKFD_COMMAND_END             0x20
 
 #endif
--

2.7.4

Quoted 4 lines

_______________________________________________
amd-gfx mailing list
[email protected]
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7CAmber.Lin%40amd.com%7C2f64059b66554d40898208d7caa5e374%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637200686149934820&amp;sdata=vBpnnq1xhSxEMjEWffwSLVIcSymSemDQTiEoYXj5lEE%3D&amp;reserved=0

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

author
#3 lore raw
In reply to Alex Deucher

[AMD Public Use]

Hi Alex,

https://github.com/RadeonOpenCompute/rocm_smi_lib will use this interface. Those functions will be added to this library:

/* Get a handler for watching events */
rsmi_status_t rsmi_event_init(rsmi_event_handle_t *handle); /* Register events for the device using the handler from init */ rsmi_status_t rsmi_event_register(uint32_t dv_ind, uint32_t events,

                                rsmi_event_handle_t *handle);

/* Wait for events. If one of the events happens, a success is returned with
* with details in data.
*/
rsmi_status_t rsmi_event_wait(rsmi_event_handle_t handle, uint32_t timeout_ms,

                                rsmi_event_data_t *data);

/* Stop watching events */
rsmi_status_t rsmi_event_free(rsmi_event_handle_t handle);

I add the ioctl to /dev/kfd with a debate if it should be in /dev/dri/card* or /dev/dri/renderD* instead. The first event to report is VM fault in this patch. Other events like RAS errors, PCIe errors, GPU reset… etc will be added for the system admin to diagnose the system health. I see this as a system feature so I use /dev/kfd. I’ll like to hear if people think differently. Thanks.

Regards,
Amber

-----Original Message-----
From: Alex Deucher <[email protected]>
Sent: Tuesday, March 17, 2020 3:03 PM
To: Lin, Amber <[email protected]>
Cc: amd-gfx list <[email protected]>
Subject: Re: [PATCH] drm/amdkfd: Provide SMI events watch

Quoted 10 lines · Amber Lin

When the compute is malfunctioning or performance drops, the system admin will use SMI (System Management Interface) tool to monitor/diagnostic what went wrong. This patch provides an event watch interface for the user space to register events they are interested. After the event is registered, the user can use annoymous file descriptor's pull function with wait-time specified to wait for the event to happen. Once the event happens, the user can use read() to retrieve information related to the event.

Quoted 3 lines

VM fault event is done in this patch.

Can you provide a link to the userspace tools that make use of this interface?

Thanks,

Alex

Quoted 442 lines
---
 drivers/gpu/drm/amd/amdkfd/Makefile              |   3 +-
 drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c |   2 +
 drivers/gpu/drm/amd/amdkfd/kfd_chardev.c         |  38 ++++++
 drivers/gpu/drm/amd/amdkfd/kfd_device.c          |   1 +
 drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c  |   2 +
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h            |  10 ++
 drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c      | 143 +++++++++++++++++++++++
 drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h      |  41 +++++++
 include/uapi/linux/kfd_ioctl.h                   |  27 ++++-
 9 files changed, 265 insertions(+), 2 deletions(-)  create mode 
100644 drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c
 create mode 100644 drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h
drivers/gpu/drm/amd/amdkfd/Makefile +2 -1
index 6147462..cc98b4a 100644
@@ -53,7 +53,8 @@ AMDKFD_FILES  := $(AMDKFD_PATH)/kfd_module.o \
                $(AMDKFD_PATH)/kfd_int_process_v9.o \
                $(AMDKFD_PATH)/kfd_dbgdev.o \
                $(AMDKFD_PATH)/kfd_dbgmgr.o \
-               $(AMDKFD_PATH)/kfd_crat.o
+               $(AMDKFD_PATH)/kfd_crat.o \
+               $(AMDKFD_PATH)/kfd_smi_events.o
 
 ifneq ($(CONFIG_AMD_IOMMU_V2),)
 AMDKFD_FILES += $(AMDKFD_PATH)/kfd_iommu.o diff --git 

a/drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c b/drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c index 9f59ba9..24b4717 100644

drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c +2 -0
@@ -24,6 +24,7 @@
 #include "kfd_events.h"
 #include "cik_int.h"
 #include "amdgpu_amdkfd.h"
+#include "kfd_smi_events.h"
 
 static bool cik_event_interrupt_isr(struct kfd_dev *dev,
                                        const uint32_t *ih_ring_entry, 
@@ -107,6 +108,7 @@ static void cik_event_interrupt_wq(struct kfd_dev *dev,
                ihre->source_id == CIK_INTSRC_GFX_MEM_PROT_FAULT) {
                struct kfd_vm_fault_info info;
 
+               kfd_smi_event_update_vmfault(dev, pasid);
                kfd_process_vm_fault(dev->dqm, pasid);
 
                memset(&info, 0, sizeof(info)); diff --git 

a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
index f8fa03a..8e92956 100644

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +1 -0
@@ -39,6 +39,7 @@
 #include "kfd_device_queue_manager.h"
 #include "kfd_dbgmgr.h"
 #include "amdgpu_amdkfd.h"
+#include "kfd_smi_events.h"
 
 static long kfd_ioctl(struct file *, unsigned int, unsigned long);  

static int kfd_open(struct inode *, struct file *); @@ -1243,6
+1244,40 @@ static int kfd_ioctl_acquire_vm(struct file *filep, struct kfd_process *p,

        return ret;

}

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +33 -0
+/* Handle requests for watching SMI events */ static int 
+kfd_ioctl_smi_events(struct file *filep,
+                               struct kfd_process *p, void *data) {
+       struct kfd_ioctl_smi_events_args *args = data;
+       struct kfd_dev *dev;
+       int ret = 0;
+
+       dev = kfd_device_by_id(args->gpu_id);
+       if (!dev)
+               return -EINVAL;
+
+       switch (args->op) {
+       case KFD_SMI_EVENTS_REGISTER:
+               ret = kfd_smi_event_register(dev, args->events);
+               if (ret >= 0) {
+                       /* When the registration is successful, it returns the
+                        * annoymous inode. Pass it to the user in data1
+                        */
+                       args->data1 = ret;
+                       ret = 0;
+               }
+               break;
+       case KFD_SMI_EVENTS_UNREGISTER:
+               kfd_smi_event_unregister(dev, args->events);
+               break;
+       default:
+               ret = -EINVAL;
+               break;
+       }
+
+       return ret;
+}
+
 bool kfd_dev_is_large_bar(struct kfd_dev *dev)  {
        struct kfd_local_mem_info mem_info; @@ -1827,6 +1862,9 @@ 

static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {

        AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_QUEUE_GWS,
                        kfd_ioctl_alloc_queue_gws, 0),
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +3 -0
+
+       AMDKFD_IOCTL_DEF(AMDKFD_IOC_SMI_EVENTS,
+                       kfd_ioctl_smi_events, 0),
 };
 
 #define AMDKFD_CORE_IOCTL_COUNT        ARRAY_SIZE(amdkfd_ioctls)
drivers/gpu/drm/amd/amdkfd/kfd_device.c +1 -0
index 7866cd06..450368c 100644
@@ -532,6 +532,7 @@ struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd,
        kfd->device_info = device_info;
        kfd->pdev = pdev;
        kfd->init_complete = false;
+       kfd->smi_events.events = 0;
        kfd->kfd2kgd = f2g;
        atomic_set(&kfd->compute_profile, 0);
 
drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c +2 -0
index e05d75e..151e83e 100644
@@ -24,6 +24,7 @@
 #include "kfd_events.h"
 #include "soc15_int.h"
 #include "kfd_device_queue_manager.h"
+#include "kfd_smi_events.h"
 
 static bool event_interrupt_isr_v9(struct kfd_dev *dev,
                                        const uint32_t *ih_ring_entry, 
@@ -117,6 +118,7 @@ static void event_interrupt_wq_v9(struct kfd_dev *dev,
                info.prot_read  = ring_id & 0x10;
                info.prot_write = ring_id & 0x20;
 
+               kfd_smi_event_update_vmfault(dev, pasid);
                kfd_process_vm_fault(dev->dqm, pasid);
                kfd_signal_vm_fault_event(dev, pasid, &info);
        }
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +10 -0
index 43b888b..fdb51de 100644
@@ -223,6 +223,13 @@ struct kfd_vmid_info {
        uint32_t vmid_num_kfd;
 };
 
+struct kfd_smi_events {
+       uint64_t events;
+       struct kfifo fifo;
+       wait_queue_head_t wait_queue;
+       uint32_t max_events;
+};
+
 struct kfd_dev {
        struct kgd_dev *kgd;
 
@@ -309,6 +316,9 @@ struct kfd_dev {
 
        /* Global GWS resource shared b/t processes*/
        void *gws;
+
+       /* if this device is in SMI events watch */
+       struct kfd_smi_events smi_events;
 };
 
 enum kfd_mempool {
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c new file +151 -0
new file mode 100644
index 0000000..ba9d036
@@ -0,0 +1,143 @@
+/*
+ * Copyright 2020 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person 
+obtaining a
+ * copy of this software and associated documentation files (the 
+"Software"),
+ * to deal in the Software without restriction, including without 
+limitation
+ * the rights to use, copy, modify, merge, publish, distribute, 
+sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom 
+the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be 
+included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 
+EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 
+MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT 
+SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, 
+DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR 
+OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE 
+OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+#include <linux/poll.h>
+#include <linux/wait.h>
+#include <linux/anon_inodes.h>
+#include <uapi/linux/kfd_ioctl.h>
+#include "amdgpu_vm.h"
+#include "kfd_priv.h"
+#include "kfd_smi_events.h"
+
+static DEFINE_MUTEX(kfd_smi_mutex);
+
+struct mutex *kfd_get_smi_mutex(void) {
+       return &kfd_smi_mutex;
+}
+
+static __poll_t kfd_smi_ev_poll(struct file *, struct 
+poll_table_struct *); static ssize_t kfd_smi_ev_read(struct file *, 
+char __user *, size_t, loff_t *); static int 
+kfd_smi_ev_release(struct inode *, struct file *);
+
+static const char kfd_smi_name[] = "kfd_smi_ev";
+
+static const struct file_operations kfd_smi_ev_fops = {
+       .owner = THIS_MODULE,
+       .poll = kfd_smi_ev_poll,
+       .read = kfd_smi_ev_read,
+       .release = kfd_smi_ev_release
+};
+
+static __poll_t kfd_smi_ev_poll(struct file *filep,
+                               struct poll_table_struct *wait) {
+       struct kfd_dev *dev = filep->private_data;
+       struct kfd_smi_events *ev = &dev->smi_events;
+
+       __poll_t mask = 0;
+
+       poll_wait(filep, &ev->wait_queue, wait);
+       mask |= !kfifo_is_empty(&ev->fifo) ? POLLIN | POLLRDNORM : 
+ mask;
+
+       return mask;
+}
+
+static ssize_t kfd_smi_ev_read(struct file *filep, char __user *user,
+                              size_t size, loff_t *offset) {
+       int ret, copied = 0;
+       struct kfd_dev *dev = filep->private_data;
+
+       ret = kfifo_to_user(&dev->smi_events.fifo, user, size, &copied);
+       if (ret || !copied) {
+               pr_debug("kfd smi-events: Fail to read fd (%i) (%i)\n",
+                               ret, copied);
+               return ret ? ret : -EAGAIN;
+       }
+
+       return copied;
+}
+
+static int kfd_smi_ev_release(struct inode *inode, struct file 
+*filep) {
+       struct kfd_dev *dev = filep->private_data;
+
+       kfifo_free(&dev->smi_events.fifo);
+       return 0;
+}
+
+void kfd_smi_event_update_vmfault(struct kfd_dev *kdev, uint16_t 
+pasid) {
+       struct kfd_smi_vmfault_fifo fifo_out;
+       struct amdgpu_device *adev = (struct amdgpu_device *)kdev->kgd;
+       struct amdgpu_task_info task_info;
+
+       if (!kdev->smi_events.events)
+               return;
+
+       if (!(kdev->smi_events.events & KFD_SMI_EV_VMFAULT))
+               return;
+
+       memset(&task_info, 0, sizeof(struct amdgpu_task_info));
+       amdgpu_vm_get_task_info(adev, pasid, &task_info);
+
+       fifo_out.group = 0;
+       fifo_out.event = KFD_SMI_EV_VMFAULT;
+       fifo_out.gpu_id = kdev->id;
+       fifo_out.pid = task_info.pid;
+       strcpy(fifo_out.task_name, task_info.task_name);
+       kfifo_in(&kdev->smi_events.fifo, &fifo_out, sizeof(fifo_out));
+       wake_up_all(&kdev->smi_events.wait_queue);
+}
+
+void kfd_smi_event_unregister(struct kfd_dev *dev, uint64_t events) {
+       mutex_lock(kfd_get_smi_mutex());
+       dev->smi_events.events &= ~events;
+       mutex_unlock(kfd_get_smi_mutex());
+}
+
+int kfd_smi_event_register(struct kfd_dev *dev, uint64_t events) {
+       int ret;
+
+       mutex_lock(kfd_get_smi_mutex());
+       dev->smi_events.events |= events;
+       mutex_unlock(kfd_get_smi_mutex());
+
+       /* We use the lower 32 bits for now. Each bit represents one event. If
+        * featured events are increased to more than 32, we'll use the upper
+        * bits as groups so the total number of events can be up to 32*32.
+        */
+       dev->smi_events.max_events = 32;
+       ret = kfifo_alloc(&dev->smi_events.fifo, dev->smi_events.max_events,
+                        GFP_KERNEL);
+       if (ret) {
+               pr_err("fail to allocate kfifo\n");
+               return ret;
+       }
+       init_waitqueue_head(&dev->smi_events.wait_queue);
+
+       return anon_inode_getfd(kfd_smi_name, &kfd_smi_ev_fops,
+                               (void *)dev, 0); }
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h new file +54 -0
new file mode 100644
index 0000000..2e320d3
@@ -0,0 +1,41 @@
+/*
+ * Copyright 2020 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person 
+obtaining a
+ * copy of this software and associated documentation files (the 
+"Software"),
+ * to deal in the Software without restriction, including without 
+limitation
+ * the rights to use, copy, modify, merge, publish, distribute, 
+sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom 
+the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be 
+included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 
+EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 
+MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT 
+SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, 
+DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR 
+OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE 
+OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+#ifndef KFD_SMI_EVENTS_H_INCLUDED
+#define KFD_SMI_EVENTS_H_INCLUDED
+
+int kfd_smi_event_register(struct kfd_dev *dev, uint64_t events); 
+void kfd_smi_event_unregister(struct kfd_dev *dev, uint64_t events); 
+void kfd_smi_event_update_vmfault(struct kfd_dev *dev, uint16_t 
+pasid);
+
+/* All FIFO must start with "uint32_t group" and "uint32_t event" so 
+the user
+ * can read the first 8 bytes to determine the next read length.
+ */
+struct kfd_smi_vmfault_fifo {
+       uint32_t group;
+       uint32_t event;
+       unsigned int gpu_id;
+       pid_t pid;
+       char task_name[TASK_COMM_LEN]; };
+
+#endif
(quoted diff) +0 -0

b/include/uapi/linux/kfd_ioctl.h index 4f66764..6ce7c69 100644

include/uapi/linux/kfd_ioctl.h +25 -1
@@ -442,6 +442,28 @@ struct kfd_ioctl_import_dmabuf_args {
        __u32 dmabuf_fd;        /* to KFD */
 };
 
+/*
+ * KFD SMI(System Management Interface) events  */ enum 
+kfd_smi_events_op {
+       KFD_SMI_EVENTS_REGISTER = 1,
+       KFD_SMI_EVENTS_UNREGISTER
+};
+
+/* Event ID (mask) */
+#define KFD_SMI_EV_VMFAULT     0x00000001
+
+struct kfd_ioctl_smi_events_args {
+       __u32 op;       /* to KFD */
+       /* upper 32 bits: group. lower 32 bits: event IDs */
+       __u64 events;   /* to KFD */
+       __u32 gpu_id;   /* to KFD */
+       pid_t pid;      /* to KFD */
+       __u32 data1;    /* from KFD */
+       __u32 data2;
+       __u32 data3;
+};
+
 /* Register offset inside the remapped mmio page
  */
 enum kfd_mmio_remap {
@@ -546,7 +568,10 @@ enum kfd_mmio_remap {
 #define AMDKFD_IOC_ALLOC_QUEUE_GWS             \
                AMDKFD_IOWR(0x1E, struct 
 kfd_ioctl_alloc_queue_gws_args)
 
+#define AMDKFD_IOC_SMI_EVENTS                  \
+               AMDKFD_IOWR(0x1F, struct kfd_ioctl_smi_events_args)
+
 #define AMDKFD_COMMAND_START           0x01
-#define AMDKFD_COMMAND_END             0x1F
+#define AMDKFD_COMMAND_END             0x20
 

#endif
--
2.7.4

Quoted 8 lines

_______________________________________________
amd-gfx mailing list
[email protected]
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flist
s.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7CAm ber.Lin%40amd.com%7C2f64059b66554d40898208d7caa5e374%7C3dd8961fe4884e6 08e11a82d994e183d%7C0%7C0%7C637200686149934820&amp;sdata=vBpnnq1xhSxEM jEWffwSLVIcSymSemDQTiEoYXj5lEE%3D&amp;reserved=0

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

author
#4 lore raw
In reply to Alex Deucher

Somehow my reply didn't seem to reach the mailing list...

Hi Alex,

https://github.com/RadeonOpenCompute/rocm_smi_lib will use this interface. Those functions will be added to this library:

/* Get a handler for watching events */
rsmi_status_t rsmi_event_init(rsmi_event_handle_t *handle);
/* Register events for the device using the handler from init */ rsmi_status_t rsmi_event_register(uint32_t dv_ind, uint32_t events, rsmi_event_handle_t *handle);
/* Wait for events. If one of the events happens, a success is returned with
 * with details in data.
 */
rsmi_status_t rsmi_event_wait(rsmi_event_handle_t handle, uint32_t timeout_ms,
                                rsmi_event_data_t *data);
/* Stop watching events */
rsmi_status_t rsmi_event_free(rsmi_event_handle_t handle);

I add the ioctl to /dev/kfd with a debate if it should be in /dev/dri/card* or /dev/dri/renderD* instead. The first event to report is VM fault in this patch. Other events like RAS errors, PCIe errors, GPU reset… etc will be added for the system admin to diagnose the system health. I see this as a system feature so I use /dev/kfd. I’ll like to hear if people think differently. Thanks.

Thanks.

Amber

Quoted 430 lines · Alex Deucher
Quoted 9 lines · Amber Lin

When the compute is malfunctioning or performance drops, the system admin will use SMI (System Management Interface) tool to monitor/diagnostic what went wrong. This patch provides an event watch interface for the user space to register events they are interested. After the event is registered, the user can use annoymous file descriptor's pull function with wait-time specified to wait for the event to happen. Once the event happens, the user can use read() to retrieve information related to the event.

Quoted 3 lines

VM fault event is done in this patch.

Can you provide a link to the userspace tools that make use of this interface?

Thanks,

Alex

Quoted 411 lines
---
  drivers/gpu/drm/amd/amdkfd/Makefile              |   3 +-
  drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c |   2 +
  drivers/gpu/drm/amd/amdkfd/kfd_chardev.c         |  38 ++++++
  drivers/gpu/drm/amd/amdkfd/kfd_device.c          |   1 +
  drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c  |   2 +
  drivers/gpu/drm/amd/amdkfd/kfd_priv.h            |  10 ++
  drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c      | 143 +++++++++++++++++++++++
  drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h      |  41 +++++++
  include/uapi/linux/kfd_ioctl.h                   |  27 ++++-
  9 files changed, 265 insertions(+), 2 deletions(-)

create mode 100644 drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c create mode 100644 drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h

drivers/gpu/drm/amd/amdkfd/Makefile +2 -1
index 6147462..cc98b4a 100644
@@ -53,7 +53,8 @@ AMDKFD_FILES  := $(AMDKFD_PATH)/kfd_module.o \
                 $(AMDKFD_PATH)/kfd_int_process_v9.o \
                 $(AMDKFD_PATH)/kfd_dbgdev.o \
                 $(AMDKFD_PATH)/kfd_dbgmgr.o \
-               $(AMDKFD_PATH)/kfd_crat.o
+               $(AMDKFD_PATH)/kfd_crat.o \
+               $(AMDKFD_PATH)/kfd_smi_events.o
 
  ifneq ($(CONFIG_AMD_IOMMU_V2),)
  AMDKFD_FILES += $(AMDKFD_PATH)/kfd_iommu.o
drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c +2 -0
index 9f59ba9..24b4717 100644
@@ -24,6 +24,7 @@
  #include "kfd_events.h"
  #include "cik_int.h"
  #include "amdgpu_amdkfd.h"
+#include "kfd_smi_events.h"
 
  static bool cik_event_interrupt_isr(struct kfd_dev *dev,
                                         const uint32_t *ih_ring_entry,
@@ -107,6 +108,7 @@ static void cik_event_interrupt_wq(struct kfd_dev *dev,
                 ihre->source_id == CIK_INTSRC_GFX_MEM_PROT_FAULT) {
                 struct kfd_vm_fault_info info;
 
+               kfd_smi_event_update_vmfault(dev, pasid);
                 kfd_process_vm_fault(dev->dqm, pasid);
 
                 memset(&info, 0, sizeof(info));
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +38 -0
index f8fa03a..8e92956 100644
@@ -39,6 +39,7 @@
  #include "kfd_device_queue_manager.h"
  #include "kfd_dbgmgr.h"
  #include "amdgpu_amdkfd.h"
+#include "kfd_smi_events.h"
 
  static long kfd_ioctl(struct file *, unsigned int, unsigned long);
  static int kfd_open(struct inode *, struct file *);
@@ -1243,6 +1244,40 @@ static int kfd_ioctl_acquire_vm(struct file *filep, struct kfd_process *p,
         return ret;
  }
 
+/* Handle requests for watching SMI events */
+static int kfd_ioctl_smi_events(struct file *filep,
+                               struct kfd_process *p, void *data)
+{
+       struct kfd_ioctl_smi_events_args *args = data;
+       struct kfd_dev *dev;
+       int ret = 0;
+
+       dev = kfd_device_by_id(args->gpu_id);
+       if (!dev)
+               return -EINVAL;
+
+       switch (args->op) {
+       case KFD_SMI_EVENTS_REGISTER:
+               ret = kfd_smi_event_register(dev, args->events);
+               if (ret >= 0) {
+                       /* When the registration is successful, it returns the
+                        * annoymous inode. Pass it to the user in data1
+                        */
+                       args->data1 = ret;
+                       ret = 0;
+               }
+               break;
+       case KFD_SMI_EVENTS_UNREGISTER:
+               kfd_smi_event_unregister(dev, args->events);
+               break;
+       default:
+               ret = -EINVAL;
+               break;
+       }
+
+       return ret;
+}
+
  bool kfd_dev_is_large_bar(struct kfd_dev *dev)
  {
         struct kfd_local_mem_info mem_info;
@@ -1827,6 +1862,9 @@ static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
 
         AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_QUEUE_GWS,
                         kfd_ioctl_alloc_queue_gws, 0),
+
+       AMDKFD_IOCTL_DEF(AMDKFD_IOC_SMI_EVENTS,
+                       kfd_ioctl_smi_events, 0),
  };
 
  #define AMDKFD_CORE_IOCTL_COUNT        ARRAY_SIZE(amdkfd_ioctls)
drivers/gpu/drm/amd/amdkfd/kfd_device.c +1 -0
index 7866cd06..450368c 100644
@@ -532,6 +532,7 @@ struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd,
         kfd->device_info = device_info;
         kfd->pdev = pdev;
         kfd->init_complete = false;
+       kfd->smi_events.events = 0;
         kfd->kfd2kgd = f2g;
         atomic_set(&kfd->compute_profile, 0);
 
drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c +2 -0
index e05d75e..151e83e 100644
@@ -24,6 +24,7 @@
  #include "kfd_events.h"
  #include "soc15_int.h"
  #include "kfd_device_queue_manager.h"
+#include "kfd_smi_events.h"
 
  static bool event_interrupt_isr_v9(struct kfd_dev *dev,
                                         const uint32_t *ih_ring_entry,
@@ -117,6 +118,7 @@ static void event_interrupt_wq_v9(struct kfd_dev *dev,
                 info.prot_read  = ring_id & 0x10;
                 info.prot_write = ring_id & 0x20;
 
+               kfd_smi_event_update_vmfault(dev, pasid);
                 kfd_process_vm_fault(dev->dqm, pasid);
                 kfd_signal_vm_fault_event(dev, pasid, &info);
         }
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +10 -0
index 43b888b..fdb51de 100644
@@ -223,6 +223,13 @@ struct kfd_vmid_info {
         uint32_t vmid_num_kfd;
  };
 
+struct kfd_smi_events {
+       uint64_t events;
+       struct kfifo fifo;
+       wait_queue_head_t wait_queue;
+       uint32_t max_events;
+};
+
  struct kfd_dev {
         struct kgd_dev *kgd;
 
@@ -309,6 +316,9 @@ struct kfd_dev {
 
         /* Global GWS resource shared b/t processes*/
         void *gws;
+
+       /* if this device is in SMI events watch */
+       struct kfd_smi_events smi_events;
  };
 
  enum kfd_mempool {
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c new file +143 -0
new file mode 100644
index 0000000..ba9d036
@@ -0,0 +1,143 @@
+/*
+ * Copyright 2020 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+#include <linux/poll.h>
+#include <linux/wait.h>
+#include <linux/anon_inodes.h>
+#include <uapi/linux/kfd_ioctl.h>
+#include "amdgpu_vm.h"
+#include "kfd_priv.h"
+#include "kfd_smi_events.h"
+
+static DEFINE_MUTEX(kfd_smi_mutex);
+
+struct mutex *kfd_get_smi_mutex(void)
+{
+       return &kfd_smi_mutex;
+}
+
+static __poll_t kfd_smi_ev_poll(struct file *, struct poll_table_struct *);
+static ssize_t kfd_smi_ev_read(struct file *, char __user *, size_t, loff_t *);
+static int kfd_smi_ev_release(struct inode *, struct file *);
+
+static const char kfd_smi_name[] = "kfd_smi_ev";
+
+static const struct file_operations kfd_smi_ev_fops = {
+       .owner = THIS_MODULE,
+       .poll = kfd_smi_ev_poll,
+       .read = kfd_smi_ev_read,
+       .release = kfd_smi_ev_release
+};
+
+static __poll_t kfd_smi_ev_poll(struct file *filep,
+                               struct poll_table_struct *wait)
+{
+       struct kfd_dev *dev = filep->private_data;
+       struct kfd_smi_events *ev = &dev->smi_events;
+
+       __poll_t mask = 0;
+
+       poll_wait(filep, &ev->wait_queue, wait);
+       mask |= !kfifo_is_empty(&ev->fifo) ? POLLIN | POLLRDNORM : mask;
+
+       return mask;
+}
+
+static ssize_t kfd_smi_ev_read(struct file *filep, char __user *user,
+                              size_t size, loff_t *offset)
+{
+       int ret, copied = 0;
+       struct kfd_dev *dev = filep->private_data;
+
+       ret = kfifo_to_user(&dev->smi_events.fifo, user, size, &copied);
+       if (ret || !copied) {
+               pr_debug("kfd smi-events: Fail to read fd (%i) (%i)\n",
+                               ret, copied);
+               return ret ? ret : -EAGAIN;
+       }
+
+       return copied;
+}
+
+static int kfd_smi_ev_release(struct inode *inode, struct file *filep)
+{
+       struct kfd_dev *dev = filep->private_data;
+
+       kfifo_free(&dev->smi_events.fifo);
+       return 0;
+}
+
+void kfd_smi_event_update_vmfault(struct kfd_dev *kdev, uint16_t pasid)
+{
+       struct kfd_smi_vmfault_fifo fifo_out;
+       struct amdgpu_device *adev = (struct amdgpu_device *)kdev->kgd;
+       struct amdgpu_task_info task_info;
+
+       if (!kdev->smi_events.events)
+               return;
+
+       if (!(kdev->smi_events.events & KFD_SMI_EV_VMFAULT))
+               return;
+
+       memset(&task_info, 0, sizeof(struct amdgpu_task_info));
+       amdgpu_vm_get_task_info(adev, pasid, &task_info);
+
+       fifo_out.group = 0;
+       fifo_out.event = KFD_SMI_EV_VMFAULT;
+       fifo_out.gpu_id = kdev->id;
+       fifo_out.pid = task_info.pid;
+       strcpy(fifo_out.task_name, task_info.task_name);
+       kfifo_in(&kdev->smi_events.fifo, &fifo_out, sizeof(fifo_out));
+       wake_up_all(&kdev->smi_events.wait_queue);
+}
+
+void kfd_smi_event_unregister(struct kfd_dev *dev, uint64_t events)
+{
+       mutex_lock(kfd_get_smi_mutex());
+       dev->smi_events.events &= ~events;
+       mutex_unlock(kfd_get_smi_mutex());
+}
+
+int kfd_smi_event_register(struct kfd_dev *dev, uint64_t events)
+{
+       int ret;
+
+       mutex_lock(kfd_get_smi_mutex());
+       dev->smi_events.events |= events;
+       mutex_unlock(kfd_get_smi_mutex());
+
+       /* We use the lower 32 bits for now. Each bit represents one event. If
+        * featured events are increased to more than 32, we'll use the upper
+        * bits as groups so the total number of events can be up to 32*32.
+        */
+       dev->smi_events.max_events = 32;
+       ret = kfifo_alloc(&dev->smi_events.fifo, dev->smi_events.max_events,
+                        GFP_KERNEL);
+       if (ret) {
+               pr_err("fail to allocate kfifo\n");
+               return ret;
+       }
+       init_waitqueue_head(&dev->smi_events.wait_queue);
+
+       return anon_inode_getfd(kfd_smi_name, &kfd_smi_ev_fops,
+                               (void *)dev, 0);
+}
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h new file +41 -0
new file mode 100644
index 0000000..2e320d3
@@ -0,0 +1,41 @@
+/*
+ * Copyright 2020 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+#ifndef KFD_SMI_EVENTS_H_INCLUDED
+#define KFD_SMI_EVENTS_H_INCLUDED
+
+int kfd_smi_event_register(struct kfd_dev *dev, uint64_t events);
+void kfd_smi_event_unregister(struct kfd_dev *dev, uint64_t events);
+void kfd_smi_event_update_vmfault(struct kfd_dev *dev, uint16_t pasid);
+
+/* All FIFO must start with "uint32_t group" and "uint32_t event" so the user
+ * can read the first 8 bytes to determine the next read length.
+ */
+struct kfd_smi_vmfault_fifo {
+       uint32_t group;
+       uint32_t event;
+       unsigned int gpu_id;
+       pid_t pid;
+       char task_name[TASK_COMM_LEN];
+};
+
+#endif
include/uapi/linux/kfd_ioctl.h +26 -1
index 4f66764..6ce7c69 100644
@@ -442,6 +442,28 @@ struct kfd_ioctl_import_dmabuf_args {
         __u32 dmabuf_fd;        /* to KFD */
  };
 
+/*
+ * KFD SMI(System Management Interface) events
+ */
+enum kfd_smi_events_op {
+       KFD_SMI_EVENTS_REGISTER = 1,
+       KFD_SMI_EVENTS_UNREGISTER
+};
+
+/* Event ID (mask) */
+#define KFD_SMI_EV_VMFAULT     0x00000001
+
+struct kfd_ioctl_smi_events_args {
+       __u32 op;       /* to KFD */
+       /* upper 32 bits: group. lower 32 bits: event IDs */
+       __u64 events;   /* to KFD */
+       __u32 gpu_id;   /* to KFD */
+       pid_t pid;      /* to KFD */
+       __u32 data1;    /* from KFD */
+       __u32 data2;
+       __u32 data3;
+};
+
  /* Register offset inside the remapped mmio page
   */
  enum kfd_mmio_remap {
@@ -546,7 +568,10 @@ enum kfd_mmio_remap {
  #define AMDKFD_IOC_ALLOC_QUEUE_GWS             \
                 AMDKFD_IOWR(0x1E, struct kfd_ioctl_alloc_queue_gws_args)
 
+#define AMDKFD_IOC_SMI_EVENTS                  \
+               AMDKFD_IOWR(0x1F, struct kfd_ioctl_smi_events_args)
+
  #define AMDKFD_COMMAND_START           0x01
-#define AMDKFD_COMMAND_END             0x1F
+#define AMDKFD_COMMAND_END             0x20
 
Quoted 8 lines
Quoted 8 lines
include/uapi/linux/kfd_ioctl.h +0 -1
  #endif
--

2.7.4

_______________________________________________
amd-gfx mailing list
[email protected]
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7CAmber.Lin%40amd.com%7C2f64059b66554d40898208d7caa5e374%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637200686149934820&amp;sdata=vBpnnq1xhSxEMjEWffwSLVIcSymSemDQTiEoYXj5lEE%3D&amp;reserved=0

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

author
#5 lore raw

Sorry for the messed-up link. This is the link (rocm-smi-lib) which makes use of the interface
https://github.com/RadeonOpenCompute/rocm_smi_lib

Quoted 503 lines · Amber Lin

Somehow my reply didn't seem to reach the mailing list...

Hi Alex,

https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2FRadeonOpenCompute%2Frocm_smi_lib&amp;data=02%7C01%7Camber.lin%40amd.com%7C37d1a82d9e734d9fec6d08d7cf56ce36%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637205844045641423&amp;sdata=I%2BVkN3VKYFUiZ0xGW0Yst70rcqrMRXUTcd995RgfRa4%3D&amp;reserved=0 will use this interface. Those functions will be added to this library:

/* Get a handler for watching events */
rsmi_status_t rsmi_event_init(rsmi_event_handle_t *handle);
/* Register events for the device using the handler from init */ rsmi_status_t rsmi_event_register(uint32_t dv_ind, uint32_t events, rsmi_event_handle_t *handle);
/* Wait for events. If one of the events happens, a success is returned with
 * with details in data.
 */
rsmi_status_t rsmi_event_wait(rsmi_event_handle_t handle, uint32_t timeout_ms,
                                rsmi_event_data_t *data);
/* Stop watching events */
rsmi_status_t rsmi_event_free(rsmi_event_handle_t handle);

I add the ioctl to /dev/kfd with a debate if it should be in /dev/dri/card* or /dev/dri/renderD* instead. The first event to report is VM fault in this patch. Other events like RAS errors, PCIe errors, GPU reset… etc will be added for the system admin to diagnose the system health. I see this as a system feature so I use /dev/kfd. I’ll like to hear if people think differently. Thanks.

Thanks.

Amber

Quoted 478 lines · Alex Deucher
Quoted 9 lines · Amber Lin

When the compute is malfunctioning or performance drops, the system admin will use SMI (System Management Interface) tool to monitor/diagnostic what went wrong. This patch provides an event watch interface for the user space to register events they are interested. After the event is registered, the user can use annoymous file descriptor's pull function with wait-time specified to wait for the event to happen. Once the event happens, the user can use read() to retrieve information related to the event.

Quoted 3 lines

VM fault event is done in this patch.

Can you provide a link to the userspace tools that make use of this interface?

Thanks,

Alex

Quoted 459 lines

---
  drivers/gpu/drm/amd/amdkfd/Makefile              |   3 +- drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c |   2 + drivers/gpu/drm/amd/amdkfd/kfd_chardev.c         |  38 ++++++ drivers/gpu/drm/amd/amdkfd/kfd_device.c          |   1 + drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c  |   2 + drivers/gpu/drm/amd/amdkfd/kfd_priv.h            |  10 ++ drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c      | 143
+++++++++++++++++++++++
  drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h      |  41 +++++++ include/uapi/linux/kfd_ioctl.h                   |  27 ++++-

  9 files changed, 265 insertions(+), 2 deletions(-)

  create mode 100644 drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c create mode 100644 drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h

drivers/gpu/drm/amd/amdkfd/Makefile +2 -1
index 6147462..cc98b4a 100644
@@ -53,7 +53,8 @@ AMDKFD_FILES  := $(AMDKFD_PATH)/kfd_module.o \
                  $(AMDKFD_PATH)/kfd_int_process_v9.o \
                  $(AMDKFD_PATH)/kfd_dbgdev.o \
                  $(AMDKFD_PATH)/kfd_dbgmgr.o \
-               $(AMDKFD_PATH)/kfd_crat.o
+               $(AMDKFD_PATH)/kfd_crat.o \
+               $(AMDKFD_PATH)/kfd_smi_events.o
 
   ifneq ($(CONFIG_AMD_IOMMU_V2),)

  AMDKFD_FILES += $(AMDKFD_PATH)/kfd_iommu.o

drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c +2 -0
index 9f59ba9..24b4717 100644
@@ -24,6 +24,7 @@
   #include "kfd_events.h"
   #include "cik_int.h"
   #include "amdgpu_amdkfd.h"
+#include "kfd_smi_events.h"
 
   static bool cik_event_interrupt_isr(struct kfd_dev *dev,
                                          const uint32_t *ih_ring_entry,
@@ -107,6 +108,7 @@ static void cik_event_interrupt_wq(struct 
 kfd_dev *dev,
                  ihre->source_id == CIK_INTSRC_GFX_MEM_PROT_FAULT) {
                  struct kfd_vm_fault_info info;
 
+               kfd_smi_event_update_vmfault(dev, pasid);
                  kfd_process_vm_fault(dev->dqm, pasid);
 

                 memset(&info, 0, sizeof(info));

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +18 -0
index f8fa03a..8e92956 100644
@@ -39,6 +39,7 @@
   #include "kfd_device_queue_manager.h"
   #include "kfd_dbgmgr.h"
   #include "amdgpu_amdkfd.h"
+#include "kfd_smi_events.h"
 
   static long kfd_ioctl(struct file *, unsigned int, unsigned long);
   static int kfd_open(struct inode *, struct file *);
@@ -1243,6 +1244,40 @@ static int kfd_ioctl_acquire_vm(struct file 
 *filep, struct kfd_process *p,
          return ret;
   }
 
+/* Handle requests for watching SMI events */
+static int kfd_ioctl_smi_events(struct file *filep,
+                               struct kfd_process *p, void *data)
+{
+       struct kfd_ioctl_smi_events_args *args = data;
+       struct kfd_dev *dev;
+       int ret = 0;
+
+       dev = kfd_device_by_id(args->gpu_id);
+       if (!dev)
+               return -EINVAL;
+
+       switch (args->op) {
+       case KFD_SMI_EVENTS_REGISTER:
+               ret = kfd_smi_event_register(dev, args->events);
+               if (ret >= 0) {
+                       /* When the registration is successful, it 

returns the
+                        * annoymous inode. Pass it to the user in data1
+                        */
+                       args->data1 = ret;
+                       ret = 0;
+               }
+               break;
+       case KFD_SMI_EVENTS_UNREGISTER:
+               kfd_smi_event_unregister(dev, args->events); +               break;
+       default:
+               ret = -EINVAL;
+               break;
+       }

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +4 -0
+
+       return ret;
+}
+

  bool kfd_dev_is_large_bar(struct kfd_dev *dev)
  {
         struct kfd_local_mem_info mem_info;

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +3 -0
@@ -1827,6 +1862,9 @@ static const struct amdkfd_ioctl_desc 
 amdkfd_ioctls[] = {
 
          AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_QUEUE_GWS,
                          kfd_ioctl_alloc_queue_gws, 0),
+
+       AMDKFD_IOCTL_DEF(AMDKFD_IOC_SMI_EVENTS,
+                       kfd_ioctl_smi_events, 0),
   };
 

  #define AMDKFD_CORE_IOCTL_COUNT ARRAY_SIZE(amdkfd_ioctls)

drivers/gpu/drm/amd/amdkfd/kfd_device.c +1 -0
index 7866cd06..450368c 100644
@@ -532,6 +532,7 @@ struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd,
          kfd->device_info = device_info;
          kfd->pdev = pdev;
          kfd->init_complete = false;
+       kfd->smi_events.events = 0;
          kfd->kfd2kgd = f2g;
          atomic_set(&kfd->compute_profile, 0);
 
drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c +2 -0
index e05d75e..151e83e 100644
@@ -24,6 +24,7 @@
   #include "kfd_events.h"
   #include "soc15_int.h"
   #include "kfd_device_queue_manager.h"
+#include "kfd_smi_events.h"
 
   static bool event_interrupt_isr_v9(struct kfd_dev *dev,
                                          const uint32_t *ih_ring_entry,
@@ -117,6 +118,7 @@ static void event_interrupt_wq_v9(struct kfd_dev 
 *dev,
                  info.prot_read  = ring_id & 0x10;
                  info.prot_write = ring_id & 0x20;
 
+               kfd_smi_event_update_vmfault(dev, pasid);
                  kfd_process_vm_fault(dev->dqm, pasid);
                  kfd_signal_vm_fault_event(dev, pasid, &info);

         }

drivers/gpu/drm/amd/amdkfd/kfd_priv.h +10 -0
index 43b888b..fdb51de 100644
@@ -223,6 +223,13 @@ struct kfd_vmid_info {
          uint32_t vmid_num_kfd;
   };
 
+struct kfd_smi_events {
+       uint64_t events;
+       struct kfifo fifo;
+       wait_queue_head_t wait_queue;
+       uint32_t max_events;
+};
+
   struct kfd_dev {
          struct kgd_dev *kgd;
 
@@ -309,6 +316,9 @@ struct kfd_dev {
 
          /* Global GWS resource shared b/t processes*/
          void *gws;
+
+       /* if this device is in SMI events watch */
+       struct kfd_smi_events smi_events;
   };
 
   enum kfd_mempool {
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c new file +4 -0
new file mode 100644
index 0000000..ba9d036
@@ -0,0 +1,143 @@
+/*
+ * Copyright 2020 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person 

obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */

drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c new file +17 -0
+
+#include <linux/poll.h>
+#include <linux/wait.h>
+#include <linux/anon_inodes.h>
+#include <uapi/linux/kfd_ioctl.h>
+#include "amdgpu_vm.h"
+#include "kfd_priv.h"
+#include "kfd_smi_events.h"
+
+static DEFINE_MUTEX(kfd_smi_mutex);
+
+struct mutex *kfd_get_smi_mutex(void)
+{
+       return &kfd_smi_mutex;
+}
+
+static __poll_t kfd_smi_ev_poll(struct file *, struct 

poll_table_struct *);
+static ssize_t kfd_smi_ev_read(struct file *, char __user *, size_t, loff_t *);

drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c new file +21 -0
+static int kfd_smi_ev_release(struct inode *, struct file *);
+
+static const char kfd_smi_name[] = "kfd_smi_ev";
+
+static const struct file_operations kfd_smi_ev_fops = {
+       .owner = THIS_MODULE,
+       .poll = kfd_smi_ev_poll,
+       .read = kfd_smi_ev_read,
+       .release = kfd_smi_ev_release
+};
+
+static __poll_t kfd_smi_ev_poll(struct file *filep,
+                               struct poll_table_struct *wait)
+{
+       struct kfd_dev *dev = filep->private_data;
+       struct kfd_smi_events *ev = &dev->smi_events;
+
+       __poll_t mask = 0;
+
+       poll_wait(filep, &ev->wait_queue, wait);
+       mask |= !kfifo_is_empty(&ev->fifo) ? POLLIN | POLLRDNORM : 

mask;

drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c new file +11 -0
+
+       return mask;
+}
+
+static ssize_t kfd_smi_ev_read(struct file *filep, char __user *user,
+                              size_t size, loff_t *offset)
+{
+       int ret, copied = 0;
+       struct kfd_dev *dev = filep->private_data;
+
+       ret = kfifo_to_user(&dev->smi_events.fifo, user, size, 

&copied);
+       if (ret || !copied) {
+               pr_debug("kfd smi-events: Fail to read fd (%i) (%i)\n", +                               ret, copied);
+               return ret ? ret : -EAGAIN;
+       }

drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c new file +13 -0
+
+       return copied;
+}
+
+static int kfd_smi_ev_release(struct inode *inode, struct file *filep)
+{
+       struct kfd_dev *dev = filep->private_data;
+
+       kfifo_free(&dev->smi_events.fifo);
+       return 0;
+}
+
+void kfd_smi_event_update_vmfault(struct kfd_dev *kdev, uint16_t 

pasid)

drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c new file +39 -0
+{
+       struct kfd_smi_vmfault_fifo fifo_out;
+       struct amdgpu_device *adev = (struct amdgpu_device *)kdev->kgd;
+       struct amdgpu_task_info task_info;
+
+       if (!kdev->smi_events.events)
+               return;
+
+       if (!(kdev->smi_events.events & KFD_SMI_EV_VMFAULT))
+               return;
+
+       memset(&task_info, 0, sizeof(struct amdgpu_task_info));
+       amdgpu_vm_get_task_info(adev, pasid, &task_info);
+
+       fifo_out.group = 0;
+       fifo_out.event = KFD_SMI_EV_VMFAULT;
+       fifo_out.gpu_id = kdev->id;
+       fifo_out.pid = task_info.pid;
+       strcpy(fifo_out.task_name, task_info.task_name);
+       kfifo_in(&kdev->smi_events.fifo, &fifo_out, sizeof(fifo_out));
+       wake_up_all(&kdev->smi_events.wait_queue);
+}
+
+void kfd_smi_event_unregister(struct kfd_dev *dev, uint64_t events)
+{
+       mutex_lock(kfd_get_smi_mutex());
+       dev->smi_events.events &= ~events;
+       mutex_unlock(kfd_get_smi_mutex());
+}
+
+int kfd_smi_event_register(struct kfd_dev *dev, uint64_t events)
+{
+       int ret;
+
+       mutex_lock(kfd_get_smi_mutex());
+       dev->smi_events.events |= events;
+       mutex_unlock(kfd_get_smi_mutex());
+
+       /* We use the lower 32 bits for now. Each bit represents one 

event. If
+        * featured events are increased to more than 32, we'll use the upper
+        * bits as groups so the total number of events can be up to 32*32.
+        */
+       dev->smi_events.max_events = 32;
+       ret = kfifo_alloc(&dev->smi_events.fifo, dev->smi_events.max_events,
+                        GFP_KERNEL);
+       if (ret) {
+               pr_err("fail to allocate kfifo\n");
+               return ret;
+       }
+ init_waitqueue_head(&dev->smi_events.wait_queue);

drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c new file +4 -0
+
+       return anon_inode_getfd(kfd_smi_name, &kfd_smi_ev_fops,
+                               (void *)dev, 0);
+}
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h new file +4 -0
new file mode 100644
index 0000000..2e320d3
@@ -0,0 +1,41 @@
+/*
+ * Copyright 2020 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person 

obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */

drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h new file +7 -0
+
+#ifndef KFD_SMI_EVENTS_H_INCLUDED
+#define KFD_SMI_EVENTS_H_INCLUDED
+
+int kfd_smi_event_register(struct kfd_dev *dev, uint64_t events);
+void kfd_smi_event_unregister(struct kfd_dev *dev, uint64_t events);
+void kfd_smi_event_update_vmfault(struct kfd_dev *dev, uint16_t 

pasid);

drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h new file +2 -0
+
+/* All FIFO must start with "uint32_t group" and "uint32_t event" 

so the user
+ * can read the first 8 bytes to determine the next read length.
+ */

drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h new file +9 -0
+struct kfd_smi_vmfault_fifo {
+       uint32_t group;
+       uint32_t event;
+       unsigned int gpu_id;
+       pid_t pid;
+       char task_name[TASK_COMM_LEN];
+};
+
+#endif
include/uapi/linux/kfd_ioctl.h +0 -0
index 4f66764..6ce7c69 100644
@@ -442,6 +442,28 @@ struct kfd_ioctl_import_dmabuf_args {

         __u32 dmabuf_fd;        /* to KFD */
  };

include/uapi/linux/kfd_ioctl.h +22 -0
+/*
+ * KFD SMI(System Management Interface) events
+ */
+enum kfd_smi_events_op {
+       KFD_SMI_EVENTS_REGISTER = 1,
+       KFD_SMI_EVENTS_UNREGISTER
+};
+
+/* Event ID (mask) */
+#define KFD_SMI_EV_VMFAULT     0x00000001
+
+struct kfd_ioctl_smi_events_args {
+       __u32 op;       /* to KFD */
+       /* upper 32 bits: group. lower 32 bits: event IDs */
+       __u64 events;   /* to KFD */
+       __u32 gpu_id;   /* to KFD */
+       pid_t pid;      /* to KFD */
+       __u32 data1;    /* from KFD */
+       __u32 data2;
+       __u32 data3;
+};
+

  /* Register offset inside the remapped mmio page */
  enum kfd_mmio_remap {

include/uapi/linux/kfd_ioctl.h +4 -1
@@ -546,7 +568,10 @@ enum kfd_mmio_remap {
   #define AMDKFD_IOC_ALLOC_QUEUE_GWS             \
                  AMDKFD_IOWR(0x1E, struct kfd_ioctl_alloc_queue_gws_args)
 
+#define AMDKFD_IOC_SMI_EVENTS                  \
+               AMDKFD_IOWR(0x1F, struct kfd_ioctl_smi_events_args)
+
   #define AMDKFD_COMMAND_START           0x01
-#define AMDKFD_COMMAND_END             0x1F
+#define AMDKFD_COMMAND_END             0x20
 
   #endif
Signature · 1 line
2.7.4
Quoted 11 lines
Quoted 5 lines
Quoted 5 lines

_______________________________________________
amd-gfx mailing list
[email protected]
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Camber.lin%40amd.com%7C37d1a82d9e734d9fec6d08d7cf56ce36%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637205844045641423&amp;sdata=U34cf9CypZMAY7us1cbtR7F7VAfP1U1OSVM6L6LYTY4%3D&amp;reserved=0

_______________________________________________
amd-gfx mailing list
[email protected]
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Camber.lin%40amd.com%7C37d1a82d9e734d9fec6d08d7cf56ce36%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637205844045641423&amp;sdata=U34cf9CypZMAY7us1cbtR7F7VAfP1U1OSVM6L6LYTY4%3D&amp;reserved=0

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

maintainer
#6 lore raw
In reply to Amber Lin

Hi Amber,

I see that this is based on the debugger event code. Jon and I are just working through some issues with that code. The lessons from that will need to be applied to this as well. But I think we can define your API to simplify this a bit.

The basic problem is, that we have one Fifo in the kfd_device, but potentially multiple file descriptors referring to it. For the event interface I think we can enforce only a single file descriptor per device. If there is already one, your register call can fail. See more comments inline.

Quoted 5 lines · Amber Lin

When the compute is malfunctioning or performance drops, the system admin will use SMI (System Management Interface) tool to monitor/diagnostic what went wrong. This patch provides an event watch interface for the user space to register events they are interested. After the event is registered, the user can use annoymous file descriptor's pull function

pull -> poll

Quoted 82 lines

with wait-time specified to wait for the event to happen. Once the event happens, the user can use read() to retrieve information related to the event.

VM fault event is done in this patch.

---
  drivers/gpu/drm/amd/amdkfd/Makefile              |   3 +-
  drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c |   2 +
  drivers/gpu/drm/amd/amdkfd/kfd_chardev.c         |  38 ++++++
  drivers/gpu/drm/amd/amdkfd/kfd_device.c          |   1 +
  drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c  |   2 +
  drivers/gpu/drm/amd/amdkfd/kfd_priv.h            |  10 ++
  drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c      | 143 +++++++++++++++++++++++
  drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h      |  41 +++++++
  include/uapi/linux/kfd_ioctl.h                   |  27 ++++-
  9 files changed, 265 insertions(+), 2 deletions(-)

create mode 100644 drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c create mode 100644 drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h

drivers/gpu/drm/amd/amdkfd/Makefile +2 -1
index 6147462..cc98b4a 100644
@@ -53,7 +53,8 @@ AMDKFD_FILES	:= $(AMDKFD_PATH)/kfd_module.o \
  		$(AMDKFD_PATH)/kfd_int_process_v9.o \
  		$(AMDKFD_PATH)/kfd_dbgdev.o \
  		$(AMDKFD_PATH)/kfd_dbgmgr.o \
-		$(AMDKFD_PATH)/kfd_crat.o
+		$(AMDKFD_PATH)/kfd_crat.o \
+		$(AMDKFD_PATH)/kfd_smi_events.o
 
  ifneq ($(CONFIG_AMD_IOMMU_V2),)
  AMDKFD_FILES += $(AMDKFD_PATH)/kfd_iommu.o
drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c +2 -0
index 9f59ba9..24b4717 100644
@@ -24,6 +24,7 @@
  #include "kfd_events.h"
  #include "cik_int.h"
  #include "amdgpu_amdkfd.h"
+#include "kfd_smi_events.h"
 
  static bool cik_event_interrupt_isr(struct kfd_dev *dev,
  					const uint32_t *ih_ring_entry,
@@ -107,6 +108,7 @@ static void cik_event_interrupt_wq(struct kfd_dev *dev,
  		ihre->source_id == CIK_INTSRC_GFX_MEM_PROT_FAULT) {
  		struct kfd_vm_fault_info info;
 
+		kfd_smi_event_update_vmfault(dev, pasid);
  		kfd_process_vm_fault(dev->dqm, pasid);
 
  		memset(&info, 0, sizeof(info));
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +12 -0
index f8fa03a..8e92956 100644
@@ -39,6 +39,7 @@
  #include "kfd_device_queue_manager.h"
  #include "kfd_dbgmgr.h"
  #include "amdgpu_amdkfd.h"
+#include "kfd_smi_events.h"
 
  static long kfd_ioctl(struct file *, unsigned int, unsigned long);
  static int kfd_open(struct inode *, struct file *);
@@ -1243,6 +1244,40 @@ static int kfd_ioctl_acquire_vm(struct file *filep, struct kfd_process *p,
  	return ret;
  }
 
+/* Handle requests for watching SMI events */
+static int kfd_ioctl_smi_events(struct file *filep,
+				struct kfd_process *p, void *data)
+{
+	struct kfd_ioctl_smi_events_args *args = data;
+	struct kfd_dev *dev;
+	int ret = 0;
+
+	dev = kfd_device_by_id(args->gpu_id);
+	if (!dev)
+		return -EINVAL;

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_acquire_vm(struct file *filep, struct kfd_process *p, · line 1267

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +10 -0
+
+	switch (args->op) {
+	case KFD_SMI_EVENTS_REGISTER:
+		ret = kfd_smi_event_register(dev, args->events);
+		if (ret >= 0) {
+			/* When the registration is successful, it returns the
+			 * annoymous inode. Pass it to the user in data1
+			 */
+			args->data1 = ret;
+			ret = 0;

You could return the file descriptor as the return value. On success it returns a positive fd. On failure it returns a negative error code.

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_acquire_vm(struct file *filep, struct kfd_process *p, · line 1271

Quoted 4 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +4 -0
+		}
+		break;
+	case KFD_SMI_EVENTS_UNREGISTER:
+		kfd_smi_event_unregister(dev, args->events);

Register seems to do two things: create a file descriptor, and subscribe to specific events. This unregister function only affects the subscription but not the file descriptor. I'd suggest a cleaner API:

* Register: creates the file descriptor
* Subscribe: enables/disables subscription to specific events

The unregistration is done implicitly by closing the file descriptor, so there is no explicit ioctl API call for this.

Quoted 174 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +12 -0
+		break;
+	default:
+		ret = -EINVAL;
+		break;
+	}
+
+	return ret;
+}
+
  bool kfd_dev_is_large_bar(struct kfd_dev *dev)
  {
  	struct kfd_local_mem_info mem_info;
@@ -1827,6 +1862,9 @@ static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
 
  	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_QUEUE_GWS,
  			kfd_ioctl_alloc_queue_gws, 0),
+
+	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SMI_EVENTS,
+			kfd_ioctl_smi_events, 0),
  };
 
  #define AMDKFD_CORE_IOCTL_COUNT	ARRAY_SIZE(amdkfd_ioctls)
drivers/gpu/drm/amd/amdkfd/kfd_device.c +1 -0
index 7866cd06..450368c 100644
@@ -532,6 +532,7 @@ struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd,
  	kfd->device_info = device_info;
  	kfd->pdev = pdev;
  	kfd->init_complete = false;
+	kfd->smi_events.events = 0;
  	kfd->kfd2kgd = f2g;
  	atomic_set(&kfd->compute_profile, 0);
 
drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c +2 -0
index e05d75e..151e83e 100644
@@ -24,6 +24,7 @@
  #include "kfd_events.h"
  #include "soc15_int.h"
  #include "kfd_device_queue_manager.h"
+#include "kfd_smi_events.h"
 
  static bool event_interrupt_isr_v9(struct kfd_dev *dev,
  					const uint32_t *ih_ring_entry,
@@ -117,6 +118,7 @@ static void event_interrupt_wq_v9(struct kfd_dev *dev,
  		info.prot_read  = ring_id & 0x10;
  		info.prot_write = ring_id & 0x20;
 
+		kfd_smi_event_update_vmfault(dev, pasid);
  		kfd_process_vm_fault(dev->dqm, pasid);
  		kfd_signal_vm_fault_event(dev, pasid, &info);
  	}
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +10 -0
index 43b888b..fdb51de 100644
@@ -223,6 +223,13 @@ struct kfd_vmid_info {
  	uint32_t vmid_num_kfd;
  };
 
+struct kfd_smi_events {
+	uint64_t events;
+	struct kfifo fifo;
+	wait_queue_head_t wait_queue;
+	uint32_t max_events;
+};
+
  struct kfd_dev {
  	struct kgd_dev *kgd;
 
@@ -309,6 +316,9 @@ struct kfd_dev {
 
  	/* Global GWS resource shared b/t processes*/
  	void *gws;
+
+	/* if this device is in SMI events watch */
+	struct kfd_smi_events smi_events;
  };
 
  enum kfd_mempool {
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c new file +86 -0
new file mode 100644
index 0000000..ba9d036
@@ -0,0 +1,143 @@
+/*
+ * Copyright 2020 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+#include <linux/poll.h>
+#include <linux/wait.h>
+#include <linux/anon_inodes.h>
+#include <uapi/linux/kfd_ioctl.h>
+#include "amdgpu_vm.h"
+#include "kfd_priv.h"
+#include "kfd_smi_events.h"
+
+static DEFINE_MUTEX(kfd_smi_mutex);
+
+struct mutex *kfd_get_smi_mutex(void)
+{
+	return &kfd_smi_mutex;
+}
+
+static __poll_t kfd_smi_ev_poll(struct file *, struct poll_table_struct *);
+static ssize_t kfd_smi_ev_read(struct file *, char __user *, size_t, loff_t *);
+static int kfd_smi_ev_release(struct inode *, struct file *);
+
+static const char kfd_smi_name[] = "kfd_smi_ev";
+
+static const struct file_operations kfd_smi_ev_fops = {
+	.owner = THIS_MODULE,
+	.poll = kfd_smi_ev_poll,
+	.read = kfd_smi_ev_read,
+	.release = kfd_smi_ev_release
+};
+
+static __poll_t kfd_smi_ev_poll(struct file *filep,
+				struct poll_table_struct *wait)
+{
+	struct kfd_dev *dev = filep->private_data;
+	struct kfd_smi_events *ev = &dev->smi_events;
+
+	__poll_t mask = 0;
+
+	poll_wait(filep, &ev->wait_queue, wait);
+	mask |= !kfifo_is_empty(&ev->fifo) ? POLLIN | POLLRDNORM : mask;
+
+	return mask;
+}
+
+static ssize_t kfd_smi_ev_read(struct file *filep, char __user *user,
+			       size_t size, loff_t *offset)
+{
+	int ret, copied = 0;
+	struct kfd_dev *dev = filep->private_data;
+
+	ret = kfifo_to_user(&dev->smi_events.fifo, user, size, &copied);
+	if (ret || !copied) {
+		pr_debug("kfd smi-events: Fail to read fd (%i) (%i)\n",
+				ret, copied);
+		return ret ? ret : -EAGAIN;
+	}
+
+	return copied;
+}
+
+static int kfd_smi_ev_release(struct inode *inode, struct file *filep)
+{
+	struct kfd_dev *dev = filep->private_data;
+
+	kfifo_free(&dev->smi_events.fifo);
+	return 0;

drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c line 96

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c new file +10 -0
+}
+
+void kfd_smi_event_update_vmfault(struct kfd_dev *kdev, uint16_t pasid)
+{
+	struct kfd_smi_vmfault_fifo fifo_out;
+	struct amdgpu_device *adev = (struct amdgpu_device *)kdev->kgd;
+	struct amdgpu_task_info task_info;
+
+	if (!kdev->smi_events.events)
+		return;

This condition is redundant.

Quoted 26 lines
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c new file +26 -0
+
+	if (!(kdev->smi_events.events & KFD_SMI_EV_VMFAULT))
+		return;
+
+	memset(&task_info, 0, sizeof(struct amdgpu_task_info));
+	amdgpu_vm_get_task_info(adev, pasid, &task_info);
+
+	fifo_out.group = 0;
+	fifo_out.event = KFD_SMI_EV_VMFAULT;
+	fifo_out.gpu_id = kdev->id;
+	fifo_out.pid = task_info.pid;
+	strcpy(fifo_out.task_name, task_info.task_name);
+	kfifo_in(&kdev->smi_events.fifo, &fifo_out, sizeof(fifo_out));
+	wake_up_all(&kdev->smi_events.wait_queue);
+}
+
+void kfd_smi_event_unregister(struct kfd_dev *dev, uint64_t events)
+{
+	mutex_lock(kfd_get_smi_mutex());
+	dev->smi_events.events &= ~events;
+	mutex_unlock(kfd_get_smi_mutex());
+}
+
+int kfd_smi_event_register(struct kfd_dev *dev, uint64_t events)
+{
+	int ret;

drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c line 132

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c new file +10 -0
+
+	mutex_lock(kfd_get_smi_mutex());
+	dev->smi_events.events |= events;
+	mutex_unlock(kfd_get_smi_mutex());
+
+	/* We use the lower 32 bits for now. Each bit represents one event. If
+	 * featured events are increased to more than 32, we'll use the upper
+	 * bits as groups so the total number of events can be up to 32*32.
+	 */
+	dev->smi_events.max_events = 32;

I don't understand the explanation above. It seems to refer to the event subscription mechanism. But you use this as the size of the fifo. That's the size in bytes. Your struct kfd_smi_vmfault_fifo is bigger than that, so even a single entry will overflow your FIFO queue.

Quoted 38 lines
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c new file +11 -0
+	ret = kfifo_alloc(&dev->smi_events.fifo, dev->smi_events.max_events,
+			 GFP_KERNEL);
+	if (ret) {
+		pr_err("fail to allocate kfifo\n");
+		return ret;
+	}
+	init_waitqueue_head(&dev->smi_events.wait_queue);
+
+	return anon_inode_getfd(kfd_smi_name, &kfd_smi_ev_fops,
+				(void *)dev, 0);
+}
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h new file +21 -0
new file mode 100644
index 0000000..2e320d3
@@ -0,0 +1,41 @@
+/*
+ * Copyright 2020 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */

drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h line 31

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h new file +10 -0
+
+#ifndef KFD_SMI_EVENTS_H_INCLUDED
+#define KFD_SMI_EVENTS_H_INCLUDED
+
+int kfd_smi_event_register(struct kfd_dev *dev, uint64_t events);
+void kfd_smi_event_unregister(struct kfd_dev *dev, uint64_t events);
+void kfd_smi_event_update_vmfault(struct kfd_dev *dev, uint16_t pasid);
+
+/* All FIFO must start with "uint32_t group" and "uint32_t event" so the user
+ * can read the first 8 bytes to determine the next read length.

Note about terminology: FIFO refers to a fifo queue. I think you're using it to describe an entry or record in the FIFO.

I don't know why you need to split group and event? The event mask is only 64-bit, so there can be no more than 64 events.

drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h line 36

Quoted 5 lines
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h new file +5 -0
+ */
+struct kfd_smi_vmfault_fifo {
+	uint32_t group;
+	uint32_t event;
+	unsigned int gpu_id;

The gpu_id is redundant because the file descriptor used to read the events is associated with a kfd_dev (GPU). If you want to have a single fifo, you should change the register API to not require a gpu_id and the fifo should be global, not a member of the kfd_dev.

drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h line 39

Quoted 3 lines
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h new file +3 -0
+	pid_t pid;
+	char task_name[TASK_COMM_LEN];
+};

This needs to be part of the user API because usermode needs to parse this structure. So it should be defined in kfd_ioctl.h. In your proposed API I think the size of the FIFO entry is implied by the event type. It would be cleaner to separate the FIFO entries into a header and payload. The header would be the same for all events and doesn't need to be duplicated in each event structure. It would contain the event type and size (to allow variable size events). The payload would depend on the event type.

Quoted 20 lines
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h new file +2 -0
+
+#endif
include/uapi/linux/kfd_ioctl.h +10 -0
index 4f66764..6ce7c69 100644
@@ -442,6 +442,28 @@ struct kfd_ioctl_import_dmabuf_args {
  	__u32 dmabuf_fd;	/* to KFD */
  };
 
+/*
+ * KFD SMI(System Management Interface) events
+ */
+enum kfd_smi_events_op {
+	KFD_SMI_EVENTS_REGISTER = 1,
+	KFD_SMI_EVENTS_UNREGISTER
+};
+
+/* Event ID (mask) */
+#define KFD_SMI_EV_VMFAULT     0x00000001

include/uapi/linux/kfd_ioctl.h struct kfd_ioctl_import_dmabuf_args { · line 464

Quoted 10 lines
include/uapi/linux/kfd_ioctl.h +10 -0
+
+struct kfd_ioctl_smi_events_args {
+	__u32 op;       /* to KFD */
+	/* upper 32 bits: group. lower 32 bits: event IDs */
+	__u64 events;   /* to KFD */
+	__u32 gpu_id;   /* to KFD */
+	pid_t pid;      /* to KFD */
+	__u32 data1;    /* from KFD */
+	__u32 data2;
+	__u32 data3;

This looks like you copied it from the debug API.  pid, data1-3 are not used by your API. I think gpu_id should not be used either if you want the event FIFO to be global.

Regards,
  Felix

Quoted 7 lines
include/uapi/linux/kfd_ioctl.h +2 -0
+};
+
  /* Register offset inside the remapped mmio page
   */
  enum kfd_mmio_remap {
@@ -546,7 +568,10 @@ enum kfd_mmio_remap {
  #define AMDKFD_IOC_ALLOC_QUEUE_GWS		\

include/uapi/linux/kfd_ioctl.h enum kfd_mmio_remap { · line 577

Quoted 10 lines
include/uapi/linux/kfd_ioctl.h +4 -1
  		AMDKFD_IOWR(0x1E, struct kfd_ioctl_alloc_queue_gws_args)
 
+#define AMDKFD_IOC_SMI_EVENTS			\
+		AMDKFD_IOWR(0x1F, struct kfd_ioctl_smi_events_args)
+
  #define AMDKFD_COMMAND_START		0x01
-#define AMDKFD_COMMAND_END		0x1F
+#define AMDKFD_COMMAND_END		0x20
 
  #endif

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

author
#7 lore raw

Thanks Felix for the review. I have a better understanding of how kfifo works now and have changed my code quite a bit. Couple of questions below inline regarding the gpu_id and data arguments.

Thanks.

Amber

Quoted 7 lines · Felix Kuehling

Hi Amber,

I see that this is based on the debugger event code. Jon and I are just working through some issues with that code. The lessons from that will need to be applied to this as well. But I think we can define your API to simplify this a bit.

The basic problem is, that we have one Fifo in the kfd_device, but potentially multiple file descriptors referring to it. For the event interface I think we can enforce only a single file descriptor per device. If there is already one, your register call can fail. See more comments inline.

Quoted 9 lines
Quoted 5 lines · Amber Lin

When the compute is malfunctioning or performance drops, the system admin will use SMI (System Management Interface) tool to monitor/diagnostic what went wrong. This patch provides an event watch interface for the user space to register events they are interested. After the event is registered, the user can use annoymous file descriptor's pull function

pull -> poll

Thank you for spotting the typo. I’ll change that.

Quoted 86 lines
Quoted 86 lines

with wait-time specified to wait for the event to happen. Once the event happens, the user can use read() to retrieve information related to the event.

VM fault event is done in this patch.

---
  drivers/gpu/drm/amd/amdkfd/Makefile              |   3 +-
  drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c |   2 +
  drivers/gpu/drm/amd/amdkfd/kfd_chardev.c         |  38 ++++++
  drivers/gpu/drm/amd/amdkfd/kfd_device.c          |   1 +
  drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c  |   2 +
  drivers/gpu/drm/amd/amdkfd/kfd_priv.h            |  10 ++
  drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c      | 143 +++++++++++++++++++++++
  drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h      |  41 +++++++
  include/uapi/linux/kfd_ioctl.h                   |  27 ++++-
  9 files changed, 265 insertions(+), 2 deletions(-)

create mode 100644 drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c create mode 100644 drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h

drivers/gpu/drm/amd/amdkfd/Makefile +2 -1
index 6147462..cc98b4a 100644
@@ -53,7 +53,8 @@ AMDKFD_FILES	:= $(AMDKFD_PATH)/kfd_module.o \
  		$(AMDKFD_PATH)/kfd_int_process_v9.o \
  		$(AMDKFD_PATH)/kfd_dbgdev.o \
  		$(AMDKFD_PATH)/kfd_dbgmgr.o \
-		$(AMDKFD_PATH)/kfd_crat.o
+		$(AMDKFD_PATH)/kfd_crat.o \
+		$(AMDKFD_PATH)/kfd_smi_events.o
 
  ifneq ($(CONFIG_AMD_IOMMU_V2),)
  AMDKFD_FILES += $(AMDKFD_PATH)/kfd_iommu.o
drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c +2 -0
index 9f59ba9..24b4717 100644
@@ -24,6 +24,7 @@
  #include "kfd_events.h"
  #include "cik_int.h"
  #include "amdgpu_amdkfd.h"
+#include "kfd_smi_events.h"
 
  static bool cik_event_interrupt_isr(struct kfd_dev *dev,
  					const uint32_t *ih_ring_entry,
@@ -107,6 +108,7 @@ static void cik_event_interrupt_wq(struct kfd_dev *dev,
  		ihre->source_id == CIK_INTSRC_GFX_MEM_PROT_FAULT) {
  		struct kfd_vm_fault_info info;
 
+		kfd_smi_event_update_vmfault(dev, pasid);
  		kfd_process_vm_fault(dev->dqm, pasid);
 
  		memset(&info, 0, sizeof(info));
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +16 -0
index f8fa03a..8e92956 100644
@@ -39,6 +39,7 @@
  #include "kfd_device_queue_manager.h"
  #include "kfd_dbgmgr.h"
  #include "amdgpu_amdkfd.h"
+#include "kfd_smi_events.h"
 
  static long kfd_ioctl(struct file *, unsigned int, unsigned long);
  static int kfd_open(struct inode *, struct file *);
@@ -1243,6 +1244,40 @@ static int kfd_ioctl_acquire_vm(struct file *filep, struct kfd_process *p,
  	return ret;
  }
 
+/* Handle requests for watching SMI events */
+static int kfd_ioctl_smi_events(struct file *filep,
+				struct kfd_process *p, void *data)
+{
+	struct kfd_ioctl_smi_events_args *args = data;
+	struct kfd_dev *dev;
+	int ret = 0;
+
+	dev = kfd_device_by_id(args->gpu_id);
+	if (!dev)
+		return -EINVAL;
+
+	switch (args->op) {
+	case KFD_SMI_EVENTS_REGISTER:
+		ret = kfd_smi_event_register(dev, args->events);
Quoted 9 lines

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_acquire_vm(struct file *filep, struct kfd_process *p, · line 1267

Quoted 6 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +6 -0
+		if (ret >= 0) {
+			/* When the registration is successful, it returns the
+			 * annoymous inode. Pass it to the user in data1
+			 */
+			args->data1 = ret;
+			ret = 0;

You could return the file descriptor as the return value. On success it returns a positive fd. On failure it returns a negative error code.

I'm thinking to have a consistent return value regardless of the argument content. When args->op is not REGISTER, the return value as 0 is success.

Quoted 8 lines

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_acquire_vm(struct file *filep, struct kfd_process *p, · line 1271

Quoted 5 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +4 -0
+		}
+		break;
+	case KFD_SMI_EVENTS_UNREGISTER:
+		kfd_smi_event_unregister(dev, args->events);

Register seems to do two things: create a file descriptor, and subscribe to specific events. This unregister function only affects the subscription but not the file descriptor. I'd suggest a cleaner API:

Quoted 5 lines

* Register: creates the file descriptor
* Subscribe: enables/disables subscription to specific events

The unregistration is done implicitly by closing the file descriptor, so there is no explicit ioctl API call for this.

Make sense. I've changed this.

Quoted 179 lines
Quoted 179 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +12 -0
+		break;
+	default:
+		ret = -EINVAL;
+		break;
+	}
+
+	return ret;
+}
+
  bool kfd_dev_is_large_bar(struct kfd_dev *dev)
  {
  	struct kfd_local_mem_info mem_info;
@@ -1827,6 +1862,9 @@ static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
 
  	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_QUEUE_GWS,
  			kfd_ioctl_alloc_queue_gws, 0),
+
+	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SMI_EVENTS,
+			kfd_ioctl_smi_events, 0),
  };
 
  #define AMDKFD_CORE_IOCTL_COUNT	ARRAY_SIZE(amdkfd_ioctls)
drivers/gpu/drm/amd/amdkfd/kfd_device.c +1 -0
index 7866cd06..450368c 100644
@@ -532,6 +532,7 @@ struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd,
  	kfd->device_info = device_info;
  	kfd->pdev = pdev;
  	kfd->init_complete = false;
+	kfd->smi_events.events = 0;
  	kfd->kfd2kgd = f2g;
  	atomic_set(&kfd->compute_profile, 0);
 
drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c +2 -0
index e05d75e..151e83e 100644
@@ -24,6 +24,7 @@
  #include "kfd_events.h"
  #include "soc15_int.h"
  #include "kfd_device_queue_manager.h"
+#include "kfd_smi_events.h"
 
  static bool event_interrupt_isr_v9(struct kfd_dev *dev,
  					const uint32_t *ih_ring_entry,
@@ -117,6 +118,7 @@ static void event_interrupt_wq_v9(struct kfd_dev *dev,
  		info.prot_read  = ring_id & 0x10;
  		info.prot_write = ring_id & 0x20;
 
+		kfd_smi_event_update_vmfault(dev, pasid);
  		kfd_process_vm_fault(dev->dqm, pasid);
  		kfd_signal_vm_fault_event(dev, pasid, &info);
  	}
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +10 -0
index 43b888b..fdb51de 100644
@@ -223,6 +223,13 @@ struct kfd_vmid_info {
  	uint32_t vmid_num_kfd;
  };
 
+struct kfd_smi_events {
+	uint64_t events;
+	struct kfifo fifo;
+	wait_queue_head_t wait_queue;
+	uint32_t max_events;
+};
+
  struct kfd_dev {
  	struct kgd_dev *kgd;
 
@@ -309,6 +316,9 @@ struct kfd_dev {
 
  	/* Global GWS resource shared b/t processes*/
  	void *gws;
+
+	/* if this device is in SMI events watch */
+	struct kfd_smi_events smi_events;
  };
 
  enum kfd_mempool {
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c new file +90 -0
new file mode 100644
index 0000000..ba9d036
@@ -0,0 +1,143 @@
+/*
+ * Copyright 2020 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+#include <linux/poll.h>
+#include <linux/wait.h>
+#include <linux/anon_inodes.h>
+#include <uapi/linux/kfd_ioctl.h>
+#include "amdgpu_vm.h"
+#include "kfd_priv.h"
+#include "kfd_smi_events.h"
+
+static DEFINE_MUTEX(kfd_smi_mutex);
+
+struct mutex *kfd_get_smi_mutex(void)
+{
+	return &kfd_smi_mutex;
+}
+
+static __poll_t kfd_smi_ev_poll(struct file *, struct poll_table_struct *);
+static ssize_t kfd_smi_ev_read(struct file *, char __user *, size_t, loff_t *);
+static int kfd_smi_ev_release(struct inode *, struct file *);
+
+static const char kfd_smi_name[] = "kfd_smi_ev";
+
+static const struct file_operations kfd_smi_ev_fops = {
+	.owner = THIS_MODULE,
+	.poll = kfd_smi_ev_poll,
+	.read = kfd_smi_ev_read,
+	.release = kfd_smi_ev_release
+};
+
+static __poll_t kfd_smi_ev_poll(struct file *filep,
+				struct poll_table_struct *wait)
+{
+	struct kfd_dev *dev = filep->private_data;
+	struct kfd_smi_events *ev = &dev->smi_events;
+
+	__poll_t mask = 0;
+
+	poll_wait(filep, &ev->wait_queue, wait);
+	mask |= !kfifo_is_empty(&ev->fifo) ? POLLIN | POLLRDNORM : mask;
+
+	return mask;
+}
+
+static ssize_t kfd_smi_ev_read(struct file *filep, char __user *user,
+			       size_t size, loff_t *offset)
+{
+	int ret, copied = 0;
+	struct kfd_dev *dev = filep->private_data;
+
+	ret = kfifo_to_user(&dev->smi_events.fifo, user, size, &copied);
+	if (ret || !copied) {
+		pr_debug("kfd smi-events: Fail to read fd (%i) (%i)\n",
+				ret, copied);
+		return ret ? ret : -EAGAIN;
+	}
+
+	return copied;
+}
+
+static int kfd_smi_ev_release(struct inode *inode, struct file *filep)
+{
+	struct kfd_dev *dev = filep->private_data;
+
+	kfifo_free(&dev->smi_events.fifo);
+	return 0;
+}
+
+void kfd_smi_event_update_vmfault(struct kfd_dev *kdev, uint16_t pasid)
+{
Quoted 9 lines

drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c line 96

Quoted 6 lines
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c new file +6 -0
+	struct kfd_smi_vmfault_fifo fifo_out;
+	struct amdgpu_device *adev = (struct amdgpu_device *)kdev->kgd;
+	struct amdgpu_task_info task_info;
+
+	if (!kdev->smi_events.events)
+		return;

This condition is redundant.

Removed it. It was from my original plan to handle multiple events in one function but after realizing not do-able and changing the function name, I forgot to remove this check.

Quoted 31 lines
Quoted 31 lines
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c new file +30 -0
+
+	if (!(kdev->smi_events.events & KFD_SMI_EV_VMFAULT))
+		return;
+
+	memset(&task_info, 0, sizeof(struct amdgpu_task_info));
+	amdgpu_vm_get_task_info(adev, pasid, &task_info);
+
+	fifo_out.group = 0;
+	fifo_out.event = KFD_SMI_EV_VMFAULT;
+	fifo_out.gpu_id = kdev->id;
+	fifo_out.pid = task_info.pid;
+	strcpy(fifo_out.task_name, task_info.task_name);
+	kfifo_in(&kdev->smi_events.fifo, &fifo_out, sizeof(fifo_out));
+	wake_up_all(&kdev->smi_events.wait_queue);
+}
+
+void kfd_smi_event_unregister(struct kfd_dev *dev, uint64_t events)
+{
+	mutex_lock(kfd_get_smi_mutex());
+	dev->smi_events.events &= ~events;
+	mutex_unlock(kfd_get_smi_mutex());
+}
+
+int kfd_smi_event_register(struct kfd_dev *dev, uint64_t events)
+{
+	int ret;
+
+	mutex_lock(kfd_get_smi_mutex());
+	dev->smi_events.events |= events;
+	mutex_unlock(kfd_get_smi_mutex());
Quoted 9 lines

drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c line 132

Quoted 6 lines
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c new file +6 -0
+
+	/* We use the lower 32 bits for now. Each bit represents one event. If
+	 * featured events are increased to more than 32, we'll use the upper
+	 * bits as groups so the total number of events can be up to 32*32.
+	 */
+	dev->smi_events.max_events = 32;

I don't understand the explanation above. It seems to refer to the event subscription mechanism. But you use this as the size of the fifo. That's the size in bytes. Your struct kfd_smi_vmfault_fifo is bigger than that, so even a single entry will overflow your FIFO queue.

I’m afraid we’ll eventually need to support more than 64 events, so I want to reserve the upper 32 bits to serve as “group”. Once we reach 32 events, 0x00000000_80000000 will be the last event in group 0, and 0x00000001_00000001 is the first event in group 1. This way we can expand the 64 bits to support more than 64 events.

This seems too complicated to explain in code. I'll just use the events parameter as event IDs. Once the featured events grows to more than 64, we can use "data" in the argument to express group when the op is ENABLE/DISABLE events.

Quoted 51 lines
Quoted 38 lines
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c new file +11 -0
+	ret = kfifo_alloc(&dev->smi_events.fifo, dev->smi_events.max_events,
+			 GFP_KERNEL);
+	if (ret) {
+		pr_err("fail to allocate kfifo\n");
+		return ret;
+	}
+	init_waitqueue_head(&dev->smi_events.wait_queue);
+
+	return anon_inode_getfd(kfd_smi_name, &kfd_smi_ev_fops,
+				(void *)dev, 0);
+}
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h new file +21 -0
new file mode 100644
index 0000000..2e320d3
@@ -0,0 +1,41 @@
+/*
+ * Copyright 2020 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */

drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h line 31

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h new file +10 -0
+
+#ifndef KFD_SMI_EVENTS_H_INCLUDED
+#define KFD_SMI_EVENTS_H_INCLUDED
+
+int kfd_smi_event_register(struct kfd_dev *dev, uint64_t events);
+void kfd_smi_event_unregister(struct kfd_dev *dev, uint64_t events);
+void kfd_smi_event_update_vmfault(struct kfd_dev *dev, uint16_t pasid);
+
+/* All FIFO must start with "uint32_t group" and "uint32_t event" so the user
+ * can read the first 8 bytes to determine the next read length.

Note about terminology: FIFO refers to a fifo queue. I think you're using it to describe an entry or record in the FIFO.

Quoted 2 lines

I don't know why you need to split group and event? The event mask is only 64-bit, so there can be no more than 64 events.

I've removed group

Quoted 9 lines

drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h line 36

Quoted 6 lines
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h new file +5 -0
+ */
+struct kfd_smi_vmfault_fifo {
+	uint32_t group;
+	uint32_t event;
+	unsigned int gpu_id;

The gpu_id is redundant because the file descriptor used to read the events is associated with a kfd_dev (GPU). If you want to have a single fifo, you should change the register API to not require a gpu_id and the fifo should be global, not a member of the kfd_dev.

True

Quoted 7 lines

drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h line 39

Quoted 4 lines
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h new file +3 -0
+	pid_t pid;
+	char task_name[TASK_COMM_LEN];
+};

This needs to be part of the user API because usermode needs to parse this structure. So it should be defined in kfd_ioctl.h. In your proposed API I think the size of the FIFO entry is implied by the event type. It would be cleaner to separate the FIFO entries into a header and payload. The header would be the same for all events and doesn't need to be duplicated in each event structure. It would contain the event type and size (to allow variable size events). The payload would depend on the event type.

I'll change this

Quoted 34 lines
Quoted 21 lines
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h new file +2 -0
+
+#endif
include/uapi/linux/kfd_ioctl.h +10 -0
index 4f66764..6ce7c69 100644
@@ -442,6 +442,28 @@ struct kfd_ioctl_import_dmabuf_args {
  	__u32 dmabuf_fd;	/* to KFD */
  };
 
+/*
+ * KFD SMI(System Management Interface) events
+ */
+enum kfd_smi_events_op {
+	KFD_SMI_EVENTS_REGISTER = 1,
+	KFD_SMI_EVENTS_UNREGISTER
+};
+
+/* Event ID (mask) */
+#define KFD_SMI_EV_VMFAULT     0x00000001

include/uapi/linux/kfd_ioctl.h struct kfd_ioctl_import_dmabuf_args { · line 464

Quoted 10 lines
include/uapi/linux/kfd_ioctl.h +10 -0
+
+struct kfd_ioctl_smi_events_args {
+	__u32 op;       /* to KFD */
+	/* upper 32 bits: group. lower 32 bits: event IDs */
+	__u64 events;   /* to KFD */
+	__u32 gpu_id;   /* to KFD */
+	pid_t pid;      /* to KFD */
+	__u32 data1;    /* from KFD */
+	__u32 data2;
+	__u32 data3;

This looks like you copied it from the debug API.  pid, data1-3 are not used by your API. I think gpu_id should not be used either if you want the event FIFO to be global.

Quoted 3 lines

Regards,
  Felix

gpu_id is for the user space to tell KFD which device it wants to watch events. I'll use separate fd/fifo for each device. I want to reserve some extra parameters in case I need to feedback extra information for possible future events. Further thinking of it, that will be handled in the anonymous fd's output, not here. I'll remove data 2 and 3 and rename data1 as data. I want to use it for anon_fd in REGISTRATION and for group in DISABLE/ENABLE when we need to support more than 64 events.

Quoted 10 lines
Quoted 10 lines
include/uapi/linux/kfd_ioctl.h +2 -0
+};
+
  /* Register offset inside the remapped mmio page
   */
  enum kfd_mmio_remap {
@@ -546,7 +568,10 @@ enum kfd_mmio_remap {
  #define AMDKFD_IOC_ALLOC_QUEUE_GWS		\
  		AMDKFD_IOWR(0x1E, struct kfd_ioctl_alloc_queue_gws_args)
 
Quoted 8 lines

include/uapi/linux/kfd_ioctl.h enum kfd_mmio_remap { · line 577

Quoted 8 lines
include/uapi/linux/kfd_ioctl.h +4 -1
+#define AMDKFD_IOC_SMI_EVENTS			\
+		AMDKFD_IOWR(0x1F, struct kfd_ioctl_smi_events_args)
+
  #define AMDKFD_COMMAND_START		0x01
-#define AMDKFD_COMMAND_END		0x1F
+#define AMDKFD_COMMAND_END		0x20
 
  #endif

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

maintainer
#8 lore raw

Am 2020-04-01 um 9:10 a.m. schrieb Amber Lin:

Quoted 5 lines

Thanks Felix for the review. I have a better understanding of how kfifo works now and have changed my code quite a bit. Couple of questions below inline regarding the gpu_id and data arguments.

Replies inline ...

Quoted 126 lines

Thanks.

Amber

Quoted 14 lines · Felix Kuehling

Hi Amber,

I see that this is based on the debugger event code. Jon and I are just working through some issues with that code. The lessons from that will need to be applied to this as well. But I think we can define your API to simplify this a bit.

The basic problem is, that we have one Fifo in the kfd_device, but potentially multiple file descriptors referring to it. For the event interface I think we can enforce only a single file descriptor per device. If there is already one, your register call can fail. See more comments inline.

Quoted 9 lines
Quoted 5 lines · Amber Lin

When the compute is malfunctioning or performance drops, the system admin will use SMI (System Management Interface) tool to monitor/diagnostic what went wrong. This patch provides an event watch interface for the user space to register events they are interested. After the event is registered, the user can use annoymous file descriptor's pull function

pull -> poll

Thank you for spotting the typo. I’ll change that.

Quoted 87 lines
Quoted 87 lines

with wait-time specified to wait for the event to happen. Once the event happens, the user can use read() to retrieve information related to the event.

VM fault event is done in this patch.

---
 drivers/gpu/drm/amd/amdkfd/Makefile              |   3 +-
 drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c |   2 +
 drivers/gpu/drm/amd/amdkfd/kfd_chardev.c         |  38 ++++++
 drivers/gpu/drm/amd/amdkfd/kfd_device.c          |   1 +
 drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c  |   2 +
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h            |  10 ++
 drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c      | 143 +++++++++++++++++++++++
 drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h      |  41 +++++++
 include/uapi/linux/kfd_ioctl.h                   |  27 ++++-
 9 files changed, 265 insertions(+), 2 deletions(-)
 create mode 100644 drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c
 create mode 100644 drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h
drivers/gpu/drm/amd/amdkfd/Makefile +2 -1
index 6147462..cc98b4a 100644
@@ -53,7 +53,8 @@ AMDKFD_FILES	:= $(AMDKFD_PATH)/kfd_module.o \
 		$(AMDKFD_PATH)/kfd_int_process_v9.o \
 		$(AMDKFD_PATH)/kfd_dbgdev.o \
 		$(AMDKFD_PATH)/kfd_dbgmgr.o \
-		$(AMDKFD_PATH)/kfd_crat.o
+		$(AMDKFD_PATH)/kfd_crat.o \
+		$(AMDKFD_PATH)/kfd_smi_events.o
 
 ifneq ($(CONFIG_AMD_IOMMU_V2),)
 AMDKFD_FILES += $(AMDKFD_PATH)/kfd_iommu.o
drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c +2 -0
index 9f59ba9..24b4717 100644
@@ -24,6 +24,7 @@
 #include "kfd_events.h"
 #include "cik_int.h"
 #include "amdgpu_amdkfd.h"
+#include "kfd_smi_events.h"
 
 static bool cik_event_interrupt_isr(struct kfd_dev *dev,
 					const uint32_t *ih_ring_entry,
@@ -107,6 +108,7 @@ static void cik_event_interrupt_wq(struct kfd_dev *dev,
 		ihre->source_id == CIK_INTSRC_GFX_MEM_PROT_FAULT) {
 		struct kfd_vm_fault_info info;
 
+		kfd_smi_event_update_vmfault(dev, pasid);
 		kfd_process_vm_fault(dev->dqm, pasid);
 
 		memset(&info, 0, sizeof(info));
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +17 -0
index f8fa03a..8e92956 100644
@@ -39,6 +39,7 @@
 #include "kfd_device_queue_manager.h"
 #include "kfd_dbgmgr.h"
 #include "amdgpu_amdkfd.h"
+#include "kfd_smi_events.h"
 
 static long kfd_ioctl(struct file *, unsigned int, unsigned long);
 static int kfd_open(struct inode *, struct file *);
@@ -1243,6 +1244,40 @@ static int kfd_ioctl_acquire_vm(struct file *filep, struct kfd_process *p,
 	return ret;
 }
 
+/* Handle requests for watching SMI events */
+static int kfd_ioctl_smi_events(struct file *filep,
+				struct kfd_process *p, void *data)
+{
+	struct kfd_ioctl_smi_events_args *args = data;
+	struct kfd_dev *dev;
+	int ret = 0;
+
+	dev = kfd_device_by_id(args->gpu_id);
+	if (!dev)
+		return -EINVAL;
+
+	switch (args->op) {
+	case KFD_SMI_EVENTS_REGISTER:
+		ret = kfd_smi_event_register(dev, args->events);
+		if (ret >= 0) {
Quoted 9 lines

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_acquire_vm(struct file *filep, struct kfd_process *p, · line 1267

Quoted 5 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +5 -0
+			/* When the registration is successful, it returns the
+			 * annoymous inode. Pass it to the user in data1
+			 */
+			args->data1 = ret;
+			ret = 0;

You could return the file descriptor as the return value. On success it returns a positive fd. On failure it returns a negative error code.

Quoted 3 lines

I'm thinking to have a consistent return value regardless of the argument content. When args->op is not REGISTER, the return value as 0 is success.

OK.

Quoted 250 lines
Quoted 9 lines

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_acquire_vm(struct file *filep, struct kfd_process *p, · line 1271

Quoted 4 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +4 -0
+		}
+		break;
+	case KFD_SMI_EVENTS_UNREGISTER:
+		kfd_smi_event_unregister(dev, args->events);

Register seems to do two things: create a file descriptor, and subscribe to specific events. This unregister function only affects the subscription but not the file descriptor. I'd suggest a cleaner API:

Quoted 6 lines

* Register: creates the file descriptor
* Subscribe: enables/disables subscription to specific events

The unregistration is done implicitly by closing the file descriptor, so there is no explicit ioctl API call for this.

Make sense. I've changed this.

Quoted 179 lines
Quoted 179 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +12 -0
+		break;
+	default:
+		ret = -EINVAL;
+		break;
+	}
+
+	return ret;
+}
+
 bool kfd_dev_is_large_bar(struct kfd_dev *dev)
 {
 	struct kfd_local_mem_info mem_info;
@@ -1827,6 +1862,9 @@ static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
 
 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_QUEUE_GWS,
 			kfd_ioctl_alloc_queue_gws, 0),
+
+	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SMI_EVENTS,
+			kfd_ioctl_smi_events, 0),
 };
 
 #define AMDKFD_CORE_IOCTL_COUNT	ARRAY_SIZE(amdkfd_ioctls)
drivers/gpu/drm/amd/amdkfd/kfd_device.c +1 -0
index 7866cd06..450368c 100644
@@ -532,6 +532,7 @@ struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd,
 	kfd->device_info = device_info;
 	kfd->pdev = pdev;
 	kfd->init_complete = false;
+	kfd->smi_events.events = 0;
 	kfd->kfd2kgd = f2g;
 	atomic_set(&kfd->compute_profile, 0);
 
drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c +2 -0
index e05d75e..151e83e 100644
@@ -24,6 +24,7 @@
 #include "kfd_events.h"
 #include "soc15_int.h"
 #include "kfd_device_queue_manager.h"
+#include "kfd_smi_events.h"
 
 static bool event_interrupt_isr_v9(struct kfd_dev *dev,
 					const uint32_t *ih_ring_entry,
@@ -117,6 +118,7 @@ static void event_interrupt_wq_v9(struct kfd_dev *dev,
 		info.prot_read  = ring_id & 0x10;
 		info.prot_write = ring_id & 0x20;
 
+		kfd_smi_event_update_vmfault(dev, pasid);
 		kfd_process_vm_fault(dev->dqm, pasid);
 		kfd_signal_vm_fault_event(dev, pasid, &info);
 	}
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +10 -0
index 43b888b..fdb51de 100644
@@ -223,6 +223,13 @@ struct kfd_vmid_info {
 	uint32_t vmid_num_kfd;
 };
 
+struct kfd_smi_events {
+	uint64_t events;
+	struct kfifo fifo;
+	wait_queue_head_t wait_queue;
+	uint32_t max_events;
+};
+
 struct kfd_dev {
 	struct kgd_dev *kgd;
 
@@ -309,6 +316,9 @@ struct kfd_dev {
 
 	/* Global GWS resource shared b/t processes*/
 	void *gws;
+
+	/* if this device is in SMI events watch */
+	struct kfd_smi_events smi_events;
 };
 
 enum kfd_mempool {
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c new file +90 -0
new file mode 100644
index 0000000..ba9d036
@@ -0,0 +1,143 @@
+/*
+ * Copyright 2020 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+#include <linux/poll.h>
+#include <linux/wait.h>
+#include <linux/anon_inodes.h>
+#include <uapi/linux/kfd_ioctl.h>
+#include "amdgpu_vm.h"
+#include "kfd_priv.h"
+#include "kfd_smi_events.h"
+
+static DEFINE_MUTEX(kfd_smi_mutex);
+
+struct mutex *kfd_get_smi_mutex(void)
+{
+	return &kfd_smi_mutex;
+}
+
+static __poll_t kfd_smi_ev_poll(struct file *, struct poll_table_struct *);
+static ssize_t kfd_smi_ev_read(struct file *, char __user *, size_t, loff_t *);
+static int kfd_smi_ev_release(struct inode *, struct file *);
+
+static const char kfd_smi_name[] = "kfd_smi_ev";
+
+static const struct file_operations kfd_smi_ev_fops = {
+	.owner = THIS_MODULE,
+	.poll = kfd_smi_ev_poll,
+	.read = kfd_smi_ev_read,
+	.release = kfd_smi_ev_release
+};
+
+static __poll_t kfd_smi_ev_poll(struct file *filep,
+				struct poll_table_struct *wait)
+{
+	struct kfd_dev *dev = filep->private_data;
+	struct kfd_smi_events *ev = &dev->smi_events;
+
+	__poll_t mask = 0;
+
+	poll_wait(filep, &ev->wait_queue, wait);
+	mask |= !kfifo_is_empty(&ev->fifo) ? POLLIN | POLLRDNORM : mask;
+
+	return mask;
+}
+
+static ssize_t kfd_smi_ev_read(struct file *filep, char __user *user,
+			       size_t size, loff_t *offset)
+{
+	int ret, copied = 0;
+	struct kfd_dev *dev = filep->private_data;
+
+	ret = kfifo_to_user(&dev->smi_events.fifo, user, size, &copied);
+	if (ret || !copied) {
+		pr_debug("kfd smi-events: Fail to read fd (%i) (%i)\n",
+				ret, copied);
+		return ret ? ret : -EAGAIN;
+	}
+
+	return copied;
+}
+
+static int kfd_smi_ev_release(struct inode *inode, struct file *filep)
+{
+	struct kfd_dev *dev = filep->private_data;
+
+	kfifo_free(&dev->smi_events.fifo);
+	return 0;
+}
+
+void kfd_smi_event_update_vmfault(struct kfd_dev *kdev, uint16_t pasid)
+{
Quoted 9 lines

drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c line 96

Quoted 6 lines
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c new file +6 -0
+	struct kfd_smi_vmfault_fifo fifo_out;
+	struct amdgpu_device *adev = (struct amdgpu_device *)kdev->kgd;
+	struct amdgpu_task_info task_info;
+
+	if (!kdev->smi_events.events)
+		return;

This condition is redundant.

Removed it. It was from my original plan to handle multiple events in one function but after realizing not do-able and changing the function name, I forgot to remove this check.

Quoted 34 lines
Quoted 34 lines
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c new file +33 -0
+
+	if (!(kdev->smi_events.events & KFD_SMI_EV_VMFAULT))
+		return;
+
+	memset(&task_info, 0, sizeof(struct amdgpu_task_info));
+	amdgpu_vm_get_task_info(adev, pasid, &task_info);
+
+	fifo_out.group = 0;
+	fifo_out.event = KFD_SMI_EV_VMFAULT;
+	fifo_out.gpu_id = kdev->id;
+	fifo_out.pid = task_info.pid;
+	strcpy(fifo_out.task_name, task_info.task_name);
+	kfifo_in(&kdev->smi_events.fifo, &fifo_out, sizeof(fifo_out));
+	wake_up_all(&kdev->smi_events.wait_queue);
+}
+
+void kfd_smi_event_unregister(struct kfd_dev *dev, uint64_t events)
+{
+	mutex_lock(kfd_get_smi_mutex());
+	dev->smi_events.events &= ~events;
+	mutex_unlock(kfd_get_smi_mutex());
+}
+
+int kfd_smi_event_register(struct kfd_dev *dev, uint64_t events)
+{
+	int ret;
+
+	mutex_lock(kfd_get_smi_mutex());
+	dev->smi_events.events |= events;
+	mutex_unlock(kfd_get_smi_mutex());
+
+	/* We use the lower 32 bits for now. Each bit represents one event. If
+	 * featured events are increased to more than 32, we'll use the upper
Quoted 9 lines

drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c line 132

Quoted 3 lines
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c new file +3 -0
+	 * bits as groups so the total number of events can be up to 32*32.
+	 */
+	dev->smi_events.max_events = 32;

I don't understand the explanation above. It seems to refer to the event subscription mechanism. But you use this as the size of the fifo. That's the size in bytes. Your struct kfd_smi_vmfault_fifo is bigger than that, so even a single entry will overflow your FIFO queue.

Quoted 6 lines

I’m afraid we’ll eventually need to support more than 64 events, so I want to reserve the upper 32 bits to serve as “group”. Once we reach 32 events, 0x00000000_80000000 will be the last event in group 0, and 0x00000001_00000001 is the first event in group 1. This way we can expand the 64 bits to support more than 64 events.

It seems you can only use one group at a time. I don't see how you can enable events in different groups at the same time using this method. At least internally you'll need a bitmask that's wide enough to represent all event types. In the subscribe/unsubscribe API you can use different ways to encode the events in a more compact way because you can make multiple calls to enable/disable events.

Quoted 128 lines

This seems too complicated to explain in code. I'll just use the events parameter as event IDs. Once the featured events grows to more than 64, we can use "data" in the argument to express group when the op is ENABLE/DISABLE events.

Quoted 52 lines
Quoted 38 lines
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c new file +11 -0
+	ret = kfifo_alloc(&dev->smi_events.fifo, dev->smi_events.max_events,
+			 GFP_KERNEL);
+	if (ret) {
+		pr_err("fail to allocate kfifo\n");
+		return ret;
+	}
+	init_waitqueue_head(&dev->smi_events.wait_queue);
+
+	return anon_inode_getfd(kfd_smi_name, &kfd_smi_ev_fops,
+				(void *)dev, 0);
+}
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h new file +21 -0
new file mode 100644
index 0000000..2e320d3
@@ -0,0 +1,41 @@
+/*
+ * Copyright 2020 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */

drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h line 31

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h new file +10 -0
+
+#ifndef KFD_SMI_EVENTS_H_INCLUDED
+#define KFD_SMI_EVENTS_H_INCLUDED
+
+int kfd_smi_event_register(struct kfd_dev *dev, uint64_t events);
+void kfd_smi_event_unregister(struct kfd_dev *dev, uint64_t events);
+void kfd_smi_event_update_vmfault(struct kfd_dev *dev, uint16_t pasid);
+
+/* All FIFO must start with "uint32_t group" and "uint32_t event" so the user
+ * can read the first 8 bytes to determine the next read length.

Note about terminology: FIFO refers to a fifo queue. I think you're using it to describe an entry or record in the FIFO.

Quoted 3 lines

I don't know why you need to split group and event? The event mask is only 64-bit, so there can be no more than 64 events.

I've removed group

Quoted 12 lines

drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h line 36

Quoted 6 lines
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h new file +5 -0
+ */
+struct kfd_smi_vmfault_fifo {
+	uint32_t group;
+	uint32_t event;
+	unsigned int gpu_id;

The gpu_id is redundant because the file descriptor used to read the events is associated with a kfd_dev (GPU). If you want to have a single fifo, you should change the register API to not require a gpu_id and the fifo should be global, not a member of the kfd_dev.

True

Quoted 14 lines

drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h line 39

Quoted 4 lines
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h new file +3 -0
+	pid_t pid;
+	char task_name[TASK_COMM_LEN];
+};

This needs to be part of the user API because usermode needs to parse this structure. So it should be defined in kfd_ioctl.h. In your proposed API I think the size of the FIFO entry is implied by the event type. It would be cleaner to separate the FIFO entries into a header and payload. The header would be the same for all events and doesn't need to be duplicated in each event structure. It would contain the event type and size (to allow variable size events). The payload would depend on the event type.

I'll change this

Quoted 36 lines
Quoted 21 lines
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h new file +2 -0
+
+#endif
include/uapi/linux/kfd_ioctl.h +10 -0
index 4f66764..6ce7c69 100644
@@ -442,6 +442,28 @@ struct kfd_ioctl_import_dmabuf_args {
 	__u32 dmabuf_fd;	/* to KFD */
 };
 
+/*
+ * KFD SMI(System Management Interface) events
+ */
+enum kfd_smi_events_op {
+	KFD_SMI_EVENTS_REGISTER = 1,
+	KFD_SMI_EVENTS_UNREGISTER
+};
+
+/* Event ID (mask) */
+#define KFD_SMI_EV_VMFAULT     0x00000001

include/uapi/linux/kfd_ioctl.h struct kfd_ioctl_import_dmabuf_args { · line 464

Quoted 10 lines
include/uapi/linux/kfd_ioctl.h +10 -0
+
+struct kfd_ioctl_smi_events_args {
+	__u32 op;       /* to KFD */
+	/* upper 32 bits: group. lower 32 bits: event IDs */
+	__u64 events;   /* to KFD */
+	__u32 gpu_id;   /* to KFD */
+	pid_t pid;      /* to KFD */
+	__u32 data1;    /* from KFD */
+	__u32 data2;
+	__u32 data3;

This looks like you copied it from the debug API.  pid, data1-3 are not used by your API. I think gpu_id should not be used either if you want the event FIFO to be global.

Quoted 3 lines

Regards,
  Felix

Quoted 8 lines

gpu_id is for the user space to tell KFD which device it wants to watch events. I'll use separate fd/fifo for each device. I want to reserve some extra parameters in case I need to feedback extra information for possible future events. Further thinking of it, that will be handled in the anonymous fd's output, not here. I'll remove data 2 and 3 and rename data1 as data. I want to use it for anon_fd in REGISTRATION and for group in DISABLE/ENABLE when we need to support more than 64 events.

OK.

Regards,
  Felix

Quoted 9 lines
Quoted 9 lines
Quoted 9 lines
include/uapi/linux/kfd_ioctl.h +2 -0
+};
+
 /* Register offset inside the remapped mmio page
  */
 enum kfd_mmio_remap {
@@ -546,7 +568,10 @@ enum kfd_mmio_remap {
 #define AMDKFD_IOC_ALLOC_QUEUE_GWS		\
 		AMDKFD_IOWR(0x1E, struct kfd_ioctl_alloc_queue_gws_args)
 
Quoted 8 lines
Quoted 8 lines

include/uapi/linux/kfd_ioctl.h enum kfd_mmio_remap { · line 577

Quoted 8 lines
include/uapi/linux/kfd_ioctl.h +4 -1
+#define AMDKFD_IOC_SMI_EVENTS			\
+		AMDKFD_IOWR(0x1F, struct kfd_ioctl_smi_events_args)
+
 #define AMDKFD_COMMAND_START		0x01
-#define AMDKFD_COMMAND_END		0x1F
+#define AMDKFD_COMMAND_END		0x20
 
 #endif

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx