AMD Zen/Ryzen thread (part 2)

[trippels]

[Naib]

yes it does... wrc1944 was stating that 7.2 would never compile for him. Now maybe that entire discussion should have been in a different thread BUT it wasn't it was in this thread
wrc1944 was running into "internal compiler" errors and I was providing references that 7.2 might have some specific issues rather than be aggravating Zen
To be fair I am also seeing issues in building 7.2. I had originally stated it was fine but after wrc1944 was stating he wasn't & I was looking around for any specifics with 7.2, I noted I didn't actually switch over.

every single compile of 7.2 fails in EXACTLY the same place, every single time... GCC-7/1 compiles fine for me, mesa (back to back... multiple runs) all compile fine

[trippels]

The Ryzen bug manifests itself only in non-reproducible segfaults.

[NeddySeagoon]

Naib,

I've been using gcc-7.2 since it appeared in the tree on arm64, amd64 and x86.
Not on Ryzen, yet though.

Your snippet appears to be when the random system compiler is building gcc-7.2 for the first time.
The segfault, if it is in gcc, is in your random system compiler, which may not be gcc-7.2.
Its not the gcc-7.2 being compiled anyway.
_________________
Regards,

NeddySeagoon

Computer users fall into two groups:-
those that do backups
those that have never had a hard drive fail.

[Naib]

The thing is as has been just demonstrated ANY mention ryzen being used and "oh its a chip issue, RMA" Maybe it is but something is up
GCC-7.2 will compile with march=x86-64
GCC-7.2 will fail at EXACTLY the same point when march=znver1
GCC-7.2 will fail at EXACTLY the same point when march=native (ie as above)
_________________

[trippels]

______________________________________________________________

[Naib]

A GCC-7.1 attempting to build GCC-7.2 (this will build due to march=x86-64 as a test as I am trying to narrow it down)

[trippels]

_________________________________________________

[NeddySeagoon]

Naib,

I'm using the /17.0/ profile (or its equivalent) everywhere. That gets me USE=(pie) forced on which translates into --default-pie.
Don't go there just yet as it breaks all your static libs, so flipping the pie flag is not something to be done lightly.

Can you build gcc-7.2 with gcc-6.x?
The segfault is happening in gcc-7.1 while it builds gcc-7.2, so no surprises its repeatable. From a systems point of view, that's a good thing :)
The emitted code is not being run yet, so its unlikely gcc-7.2 is the root cause of the segfault. However, its source code could be triggering an issue in gcc-7.1.
_________________
Regards,

NeddySeagoon

Computer users fall into two groups:-
those that do backups
those that have never had a hard drive fail.

[Naib]

[trippels]

________________________

[vaxbrat]

I was noticing fairly regular kernel hangs (about one a day) with this build. Things would lock up so tight, I couldn't even magic-sysreq-key to emergency sync and boot. On a hunch, I went back in and switched my processor type from athlon64/hammer/opteron/k8 to generic amd64 and rebuilt. That choice seemed to be too old to pull in the obsolete bulldozer instructions, but maybe gcc 6 was doing it anyway? I also saw 4.12.12 come in from the last update, so I made the change in there and switched from 4.13.1 to using it. So far so good. I even set up my two 4tb spinners as a btrfs mirror, ran bonnie++ on that and started installing a vm in it to take a look at Kali for work.

I'm going to gut swap another Piledriver box in the next day or so and take an image of this install to put on it as another 1700X. If that one seems stable, I have a couple of Ceph OSD daemons I will be running on it again.

[Naib]

[vaxbrat]

Looks like you need to turn on expert mode to see that RCU stuff. So I did that and then decided to choose the CONFIG_CLASSIC_SRCU just in case:

Before:

[vaxbrat]

When I poked around in the bios I saw c6 mode was enabled so I turned that off.

[krinn]

Another think i have note about the "ryzen gate", is that none report any mce errors on faulty ryzen.

So either nobody use ryzen mce feature, or worst, ryzen mce miserably fail also, which again expose another problem with ryzen.
# CONFIG_X86_MCE_AMD is not set
Anyone having it enable?

[NeddySeagoon]

krinn,

You expect something that is faulty, to work as designed, to detect that its faulty, when its faulty?
That's a definite maybe. :)

Don't expect too much of the MCE anywhere.
_________________
Regards,

NeddySeagoon

Computer users fall into two groups:-
those that do backups
those that have never had a hard drive fail.

[krinn]

i expect too much from something that is suppose to report cpu errors?
a part of the cpu is faulty, but i don't expect the mce to be faulty too

[mir3x]

In my ryzen history I had one mce error, I wrote it to some file, but its on disk I cannot access now, but it was like ( I googled it):

[Naib]

[krinn]

[wrc1944]

Finally got power and internet back on after Irma on Sept 18, so I requested an RMA using
the email form at: http://support.amd.com/en-us/contact/email-form

I filled out all the info on the page, and in the message field I wrote:

[vaxbrat]

Picked up another 1700x at the Microcenter store in Rockville MD and will use it with another 2x16 Corsair vengence 2666 and a Taichi mobo this time to do a gut swap with another fx-8350 box. This one is s/n 9gu8187070117 which is only about 40 away from my other chip assuming they don't play games with randomizing the serial numbers. The fab info on the chip is week 9 (UA 1709SUT) so maybe the people from AMD and Microcenter who might be monitoring Gentoo for Ryzen issues could get together and proactively swap out their inventory?

The changes I made to C6 state and RCU appear to have stopped all freezes since I haven't had one for a day or two now. After going through the thread on the AMD board, I'm going to tweak the SOC voltage the next time I'm poking around in there. I definitely will look at setting up the Taichi and making sure to update the firmware after getting the build together though.

[wrc1944]

FWIW, on page 108 over on https://community.amd.com/thread/215773?start=1605&tstart=0 scorpio810 posted this 0n Sept.22:

[pjp]