[PATCH v4] drm/amdkfd: Provide SMI events watch

[Lin, Amber]

[AMD Official Use Only - Internal Distribution Only]

Hi Felix,

That was my assumption too that each registration will get different file descriptor, but it turns out not. When I started two process and both register gpu0 and gpu1, they both got fd=15. If I have process A register gpu0+gpu1, and process B only register gpu0, process A gets fd=15 and process B gets fd=9. That's why I added client ID.

By multiple clients, I mean multiple processes. The ask is users want to have multiple tools and those different tools can use rsmi lib to watch events at the same time. Due to the reason above that two processes can actually get the same fd and I need to add client ID to distinguish the registration, I don't see the point of limiting one registration per process unless I use pid to distinguish the client instead, which was in my consideration too when I was writing the code. But second thought is why adding this restriction when client ID can allow the tool to watch different events on different devices if they want to. Maybe client ID is a bad term and it misleads you. I should call it register ID.

Regards,
Amber

From: Kuehling, Felix <Felix.Kuehling at amd.com>
Sent: Tuesday, April 14, 2020 7:04 PM
To: Lin, Amber <Amber.Lin at amd.com>; amd-gfx at lists.freedesktop.org
Subject: Re: [PATCH v4] drm/amdkfd: Provide SMI events watch


Hi Amber,

Some general remarks about the multi-client support. You added a global client id that's separate from the file descriptor. That's problematic for two reasons:

  1.  A process could change a different process' event mask
  2.  The FD should already be unique per process, no need to invent another ID

If we want to allow one process to register for events multiple times (multiple FDs per process), then the list of clients should be per process. Each process should only be allowed to change the event masks of its own clients. The client could be identified by its FD. No need for another client ID.

But you could also simplify it further by allowing only one event client per process. Then you don't need the client ID lookup at all. Just have a single event client in the kfd_process.

Another approach would be to make enable/disable functions of the event FD, rather than the KFD FD ioctl. It could be an ioctl of the event FD, or even simpler, you could use the write file-operation to write an event mask (of arbitrary length if you want to enable growth in the future). That way everything would be neatly encapsulated in the event FD private data.

Two more comments inline ...


Am 2020-04-14 um 5:30 p.m. schrieb 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 devices and subscribe events they are interested. After

registered, the user can use annoymous file descriptor's poll function

with wait-time specified and wait for events to happen. Once an event

happens, the user can use read() to retrieve information related to the

event.



VM fault event is done in this patch.



v2: - remove UNREGISTER and add event ENABLE/DISABLE

    - correct kfifo usage

    - move event message API to kfd_ioctl.h

v3: send the event msg in text than in binary

v4: support multiple clients



Signed-off-by: Amber Lin <Amber.Lin at amd.com><mailto:Amber.Lin at amd.com>

[snip]

diff --git a/include/uapi/linux/kfd_ioctl.h b/include/uapi/linux/kfd_ioctl.h

index 4f66764..8146437 100644

--- a/include/uapi/linux/kfd_ioctl.h

+++ b/include/uapi/linux/kfd_ioctl.h

@@ -442,6 +442,36 @@ 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_ENABLE,

+ KFD_SMI_EVENTS_DISABLE

+};

+

+/* Event type (defined by bitmask) */

+#define KFD_SMI_EVENT_VMFAULT     0x0000000000000001

+

+struct kfd_ioctl_smi_events_args {

+ __u32 op;              /* to KFD */

+ __u64 events;          /* to KFD */

The binary layout of the ioctl args structure should be the same on 32/64-bit. That means the 64-bit members should be 64-bit aligned. The best way to ensure this is to put all the 64-bit members first.




+ __u64 gpuids_array_ptr;        /* to KFD */

+ __u32 num_gpuids;      /* to KFD */

+ __u32 anon_fd;         /* from KFD */

+ __u32 client_id;       /* to/from KFD */

+};

+

+/* 1. All messages must start with (hex)uint64_event(16) + space(1) +

+ *    (hex)gpuid(8) + space(1) =  26 bytes

+ * 2. VmFault msg = (hex)uint32_pid(8) + space(1) + task name(16) = 25

+ *    When a new event msg uses more memory, change the calculation here.

+ * 3. End with \n(1)

+ * 26 + 25 + 1 = 52

+ */

+#define KFD_SMI_MAX_EVENT_MSG 52

If you define the maximum message length here, clients may start depending on it, and it gets harder to change it later. I'd not define this in the API header. It's not necessary to write correct clients. And if used badly, it may encourage writing incorrect clients that break with longer messages in the future.

Regards,
  Felix





+

 /* Register offset inside the remapped mmio page

  */

 enum kfd_mmio_remap {

@@ -546,7 +576,10 @@ 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
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/amd-gfx/attachments/20200415/98f4a2ec/attachment-0001.htm>