APU堪用否？

缘起

想在卧室里面玩玩游戏。 APU出了这么多年了，HAS, GNN…这些名词噱头很多。但是网络上一般对AMD嘲讽很多，所谓『农企』、『i3默全秒』。

AMD加速处理器 wiki page

https://en.wikipedia.org/wiki/List_of_AMD_accelerated_processing_unit_microprocessors

革命性的 AMD Mantle http://www.amd.com/zh-cn/innovations/software-technologies/mantle

安装

我搞了两台机器，一台APU A10 7800，一台Athlon x840。都是白菜价。x840甚至只要200多点。主板：A88 ITX太贵，m-atx就便宜了很多。高频内存：如果用APU的话就上把，毕竟这个是APU的亮点。

a88 主板未知PCI设备驱动. Use PCI-Z and find it is IOMMU.

FM2A88M Pro3+ asrock page.

https://www.asus.com/Motherboards/A88XMPLUS/ asrock的必须接上显卡才能开机，太搞了。换成asus的可以headless。

散热

这APU发热比较大，特别是游戏的时候。如果用普通的下压那种，容易搞的整个机箱都很烫。搞个侧吹的微塔散热就很容易压制了。

自己买的无牌机箱似乎散热效果不佳。只有在侧板打开的时候才能降到36度，加了前置的风扇也没用。

MATX机箱有银欣的PS07，风道好，但是因为有硬盘位，depth有40cm。安钛克的p180也是类似情况。乔思伯RM2外观好，不过也是下压的。其实乔思伯有专注MATX机箱，比如我给弟配置时买的C3。不过似乎U3更好。官网。http://www.chiphell.com/article-7693-3.html　不适合安装独立显卡，所以很适合ＡＰＵ。接下来，是否水冷是最佳选择？

虚拟化

之所以搞了个Athlon是因为不仅CPU便宜，而且A88的主板BIOS里面都有IOMMU开关，这样就能用PCI-E passthrough了。但是我发现只有华硕的可以不接显示器启动，asrock FM2A88M Pro3+必须接一个显卡。

kvm pci device not assignable. Check /var/log/syslog(which is configured at /etc/libvirt/libvirtd.conf). It say libvirtd[1900]: Failed to open config space file ‘/sys/bus/pci/devices/0000:00:00.0/config’: Permission denied. Can’t find way. Run “virt-manager” with root. Then error “please ensure all devices within the iommu_group are bound to their vfio bus driver” “Device ‘vfio-pci’ could not be initialized”. http://vfio.blogspot.com/2014/08/vfiovga-faq.html Try many way and fix by changing to pci-x16 slot. Maybe the pci-x1 was shared with other pci-e device. If I change to i350 4*1Gb net, I have to add all 4 pci-e device to vm guest.

Then get error “igb 0000:02:00.0: Detected Tx Unit Hang”. OK after I change to i350 pci-e*4 device. But still get this issue when I visit luci console web page.

http://www.linux-kvm.org/page/How_to_assign_devices_with_VT-d_in_KVM

Change to old Xeon Linux kernel version, work! Also I found have to use 3.13 kernel. 3.19 kernel has reported “tx unit hang” error.

Use 3.13 kernel, not need add any parameter to kernel CMD linux.

Other cheaper CPU like Athlon II X4 840 290￥ SHOULD support AMD-Vi also. YES.

X4 840 is 28nm, TDP 65 W.

http://www.amd.com/en-us/products/processors/desktop/athlon-cpu

cpu-world(AMD site tell nothing) tell it support “AMD-V / AMD Virtualization technology”.

Does other AMD CPU Onboard http://www.asrock.com/mb/index.asp?s=AMD%20CPU support AMD-Vi?

http://www.asrock.com/mb/AMD/QC5000M/ this use “AMD FT3 Kabini A4-5000 Quad-Core APU”. Looks it also support AMD-v?

ftp://asrock.cn/manual/QC5000M-ITXPH.pdf say

SVM

When this option is set to [Enabled], a VMM (Virtual Machine Architecture) can

utilize the additional hardware capabilities provided by AMD-V. he default value is

[Enabled]. Coniguration options: [Enabled] and [Disabled].

There is no IOMMU entry.

问题

有时候板载的eth0会自己关掉，连上TTL看到有这样信息：

[503386.687406] AMD-Vi: Event logged [IO_PAGE_FAULT device=04:00.0 domain=0x001d address=0x0000000000003000 flags=0x0050]
[504193.617255] r8169 0000:04:00.0 eth0: link down

自己重新ifup eth0又可以恢复。https://github.com/HSAFoundation/HSA-Drivers-Linux-AMD/issues/4 https://bugzilla.redhat.com/show_bug.cgi?id=889749 这里看来是r8169网卡驱动的问题了：当数据陡然变大时（比如samba拷贝）会挂掉。看来是我错怪 AMD 了，我原来以为是这机器兼容性不行。

Ryzen

Ryzen R5 2600

万紫千红 4Gx2, R15 1255，同样内存 i5 8400 929。

Ryzen R5 2600 运行老机器上面的 linux 启动超慢

[fan@i5fedora ~]$ systemd-analyze 
Startup finished in 2.005s (kernel) + 2.296s (initrd) + 2min 32.117s (userspace) = 2min 36.419s 
graphical.target reached after 2min 7.963s in userspace 
[fan@i5fedora ~]$ systemd-analyze critical-chain 
The time after the unit is active or started is printed after the "@" character. 
The time the unit takes to start is printed after the "+" character. 
graphical.target @2min 7.963s 
└─multi-user.target @2min 7.963s 
└─docker.service @2min 4.974s +1.041s 
└─docker-containerd.service @2min 4.589s 
└─network.target @2min 4.580s 
└─wpa_supplicant.service @2min 7.098s +20ms 
└─dbus.service @2min 3.725s 
└─basic.target @2min 3.713s 
└─sockets.target @2min 3.713s 
└─dbus.socket @2min 3.713s 
└─sysinit.target @2min 3.704s 
└─sys-fs-fuse-connections.mount @2min 27.201s +7ms 
└─systemd-journald.socket 
└─system.slice 
└─-.slice 

sudo systemctl disable wpa_supplicant 
sudo systemctl disable sys-fs-fuse-connections.mount 
[fan@i5fedora ~]$ systemd-analyze blame 
2min 523ms lvm2-monitor.service 
2min 405ms systemd-udev-settle.service 
25.063s bolt.service 
3.782s NetworkManager-wait-online.service 
3.195s plymouth-quit-wait.service 
1.190s dracut-initqueue.service 
747ms akmods.service 
sudo systemctl disable lvm2-monitor 

禁用了一堆服务还是很慢，这些服务都是相互依赖，难以找出根源。据说 new kernel (4.17+) and graphic drivers are STRONGLY recommended。

sudo systemctl mask systemd-udev-settle.service

这个才是关键，但是这个和上面的 lvm2 有关系。

这个问题修好然后是无法重启的问题：

journalctl -b -1

tail

Hardware watchdog ‘SP5100 TCO timer’, version 0

kernel: watchdog: watchdog0: watchdog did not stop!

似乎是个已知问题 https://wiki.gentoo.org/wiki/Ryzen

进入 text mode方法：在 gurb 里面加上single，比如：

linux /vmlinuz-2.6.32-5 root=/dev/mapper/debian-root ro single

然后重启，发现其实最后卡在这里很久：

systemd-udevd[450]: seq 2689 ‘/devices/system/cpu/cpu10’ killed

好多主板都有这个问题，这里，这里，似乎和bios有关，他们更严重，直接整个机器 freeze 了。

按照这里添加grub的reboot=pci选项 http://michalorman.com/2013/10/fix-ubuntu-freeze-during-restart/，还是没有用。

https://bugzilla.redhat.com/show_bug.cgi?id=1608242 一模一样的BUG，据说是bios的问题，kernel也有 work around。按照其方法修改了systemd-udev-settle.service timeout（禁用服务不大好），启用稍慢点，但是重启还是很慢。

https://www.reddit.com/r/Amd/comments/6vbe6w/threadripper_broken_on_linux_for_pci_passthrough/

Intel/英特尔 545s 256G，但是如果用 usb 硬盘即便接在ryzen gen2 口上速度才 50MB，慢的出奇。硬盘盒还是主板原因？

Turn to Ubuntu 18.04 4.15 kernel，USB live，重启正常，SR-IOV 正常，Ubuntu 的还是兼容性更好些。但是 pci passthrough 一样问题。

In Ubuntu, I ran kvm-ok and get error: kvm: disabled by bios. But I have turn on IOMMU. After Google, I have to go to BIOS M.I.T -> Advanced CPU Core Settings to turn on SVM. Not sure if Fedora will work too after this way. SVM means ` AMD Secure Virtual Machine`.

https://forums.unraid.net/topic/59893-iommu-grouping-issues-for-gpu-even-with-acs-override-enabled/ Here same issue but it success at last. 看了下，好多 pci 都在一个 group 里面，感觉还不如老的 Opteron 设备，或许 Ryzen 本来就不是为服务器准备的。