Bug 107375

Created attachment 140810 [details]
HTML output of `sudo ./sleepgraph.py -config config/suspend-callgraph.cfg` (devicelist: amdgpu, maxdepth: 12)

Profiling the suspend and resume time [1] on a MSI B350M MORTAR (MS-7A37) with AMD Ryzen 3 2200G with Radeon Vega Graphics with Linux 4.18-rc6+ and amd-staging-drm-next [2], `pci_pm_suspend()` takes over one second, which is too long [3], cf. bug 107278 [3].

The majority of that, 750 ms, is spent in `amdgpu_device_ip_resume_phase2()`.

•   amdgpu_device_ip_resume_phase2 [amdgpu] (747.134 ms @ 97.889371)
    •   psp_resume [amdgpu] (673.370 ms @ 97.889371)
        […]
        •   psp_hw_start [amdgpu] (88.577 ms @ 97.889379)
        […]
        •   psp_np_fw_load [amdgpu] (584.783 ms @ 97.977957)
    […]
    •   dm_resume [amdgpu] (44.992 ms @ 98.567232)
        […]
        •   dc_link_detect [amdgpu] (17.777 ms @ 98.568212)
            (EDID read takes 8 ms?)
        […]
        •   drm_atomic_helper_resume (25.139 ms @ 98.587030)
        […]
    […]

In the PSP code the problem is `psp_cmd_submit_buf()`, which takes from 45 ms to 141 ms. Something is wrong there.

```
static int
psp_cmd_submit_buf(struct psp_context *psp,
                   struct amdgpu_firmware_info *ucode,
                   struct psp_gfx_cmd_resp *cmd, uint64_t fence_mc_addr,
                   int index)
{
        int ret;

        memset(psp->cmd_buf_mem, 0, PSP_CMD_BUFFER_SIZE);

        memcpy(psp->cmd_buf_mem, cmd, sizeof(struct psp_gfx_cmd_resp));

        ret = psp_cmd_submit(psp, ucode, psp->cmd_buf_mc_addr,
                             fence_mc_addr, index);

        while (*((unsigned int *)psp->fence_buf) != index) {
                msleep(1);
        }

        return ret;
}
```

Also, I wonder why the sleep in the trace is 3 ms instead of 1 ms like in the code.


[1]: https://01.org/suspendresume
[2]: https://cgit.freedesktop.org/~agd5f/linux/log/?h=amd-staging-drm-next
[3]: https://bugs.freedesktop.org/show_bug.cgi?id=107278