Reported memory usage exceeds maximum values

[mi12078]

Kernel: 5.9.1-zen1-1-zen OS: Arch Linux GPU: Radeon R5 230 MoBo: Asus PRIME-B350-PLUS

07:00.0 VGA compatible controller [0300]: Advanced Micro Devices, Inc. [AMD/ATI] Caicos PRO [Radeon HD 7450] [1002:677b]
	Subsystem: Micro-Star International Co., Ltd. [MSI] Device [1462:8099]
	Kernel driver in use: radeon
	Kernel modules: radeon
07:00.1 Audio device [0403]: Advanced Micro Devices, Inc. [AMD/ATI] Caicos HDMI Audio [Radeon HD 6450 / 7450/8450/8490 OEM / R5 230/235/235X OEM] [1002:a...
	Subsystem: Micro-Star International Co., Ltd. [MSI] Device [1462:aa98]
	Kernel driver in use: snd_hda_intel
	Kernel modules: snd_hda_intel

[clbr]

Oh my. And no errors like "Failed to get VRAM usage"? Can you build from git, and add printfs around the vram getting in radeon.c?

[mi12078]

Yes, there was no "Failed to get VRAM usage" reported when I exited the program. Not sure whether you meant to print the following since I am not familiar with the code, but I suspect the out64 holds the value (although it has the unused attribute)

diff --git a/radeon.c b/radeon.c
index 3beaf8b..a81f5a8 100644
--- a/radeon.c
+++ b/radeon.c
@@ -122,8 +122,11 @@ void init_radeon(int fd, int drm_major, int drm_minor) {
                vramsize = gem.vram_size;
                gttsize = gem.gart_size;
 
-               if (!(ret = getvram_radeon(&out64)))
+               fprintf(stdout, "%lu\n", out64);
+               if (!(ret = getvram_radeon(&out64))) {
                        getvram = getvram_radeon;
+                       fprintf(stdout, "%lu\n", out64);
+               }
                else
                        drmError(ret, _("Failed to get VRAM usage"));

The values printed out were the following:

139911984272822
1603707379712

[clbr]

Well, looks like the kernel is reporting success while giving wrong values. Not much radeontop can do there. Either a kernel bug or the kernel has changed the ABI, which is against the guarantees, and so again a bug.

[clbr]

If you'd like to report a kernel bug, a git bisect would help, but even a coarse "version 5.x works, 5.x+1 fails" would be enough to start the report.

[chromer030]

It seems that the issue is resolved on kernel 5.11.1. I tested on Arch Linux, it shows true value now , but now, in-use VRAM/GTT is always zero !

[mi12078]

Similar if not exact behaviour as @chromer030 described regarding VRAM and GTT values on my end, also running arch on v5.11.1. Last time I ran the arch linux-lts kernel radeontop did report values that seemed normal, i.e. not pegged to zero and not exceeding the upper limit. I do not quite remember whether that was the linux-lts version that is currently available (5.10.17) or was it an older one.

[chromer030]

@mi12078 Based on Arch Git, they switched to 5.10 LTS branch on Feb 14, 2021.

Prior to 5.9.x. series, everything is ok.

I do not quite remember whether that was the linux-lts version that is currently available (5.10.17) or was it an older one.

I think you were on 5.4.x series which was ok with Radeon's value reporting.

[bonze82]

debian 11 bullseye 5.10.0-14-amd64 just started doing the same thing today. Last update of any type was 10 days ago.