At Sun, 06 Nov 2011 18:31:42 +0200, Avi Kivity wrote:
On 11/06/2011 06:15 PM, Denis V. Lunev wrote:
On 11/6/11 6:51 PM, Avi Kivity wrote:
The recently merged 228cf79376f1 ("ALSA: intel8x0: Improve performance in virtual environment") is hacky and somewhat wrong.
First, the detection code
if (inside_vm< 0) {/* detect KVM and Parallels virtual environments */inside_vm = kvm_para_available();+#if defined(__i386__) || defined(__x86_64__)
inside_vm = inside_vm ||boot_cpu_has(X86_FEATURE_HYPERVISOR); +#endif
}is incorrect. It detects that you're running in a guest, but that doesn't imply that the device you're accessing is emulated. It may be a host device assigned to the guest; presumably the optimization you apply doesn't work for real devices.
Second, the optimization itself looks fishy:
spin_lock(&chip->reg_lock); do { civ = igetbyte(chip, ichdev->reg_offset +ICH_REG_OFF_CIV); ptr1 = igetword(chip, ichdev->reg_offset + ichdev->roff_picb); position = ichdev->position; if (ptr1 == 0) { udelay(10); continue; }
if (civ == igetbyte(chip, ichdev->reg_offset +ICH_REG_OFF_CIV)&&
ptr1 == igetword(chip, ichdev->reg_offset +ichdev->roff_picb))
if (civ != igetbyte(chip, ichdev->reg_offset +ICH_REG_OFF_CIV))
continue;if (chip->inside_vm)break;if (ptr1 == igetword(chip, ichdev->reg_offset +ichdev->roff_picb)) break; } while (timeout--);
Why is the emulated device timing out? Can't the emulation be fixed to behave like real hardware?
Last, please copy kvm@vger.kernel.org on such issues.
The problem is that emulation can not be fixed.
How this is working for real hardware? You get data from real sound card register. The scheduling is off at the moment thus you can not be re-scheduled.
In the virtual environment the situation is different. Any IO emulation is expensive. The processor is switched from guest to hypervisor and further to emulation process takes a lot of time. This time is enough to obtain different value on next register read. That's why this code is really timed out. Please also note that host scheduler also plays his games and could schedule out VCPU thread.
Note on kvm this is rare, since the guest thread and the emulator thread are the same.
The problem could be potentially fixed reducing precision of PICB emulation, but this results in lower sound quality.
This kludge has been written this way in order not to break legacy card for which we do not have an access. The code reading PICB/CIV registers second time was added to fix issues on unknown for now platform and it looks not possible how to find/test against this platform now. We have checked Windows drivers written by Intel/AMD (32/64 bit) and MacOS ones. There is no second reading of CIV/PICB inside. We hope that this is relay needed only for some rare hadware devices.
Ok, so if I understand correctly, this loop is a hack for broken hardware, and this patch basically unhacks it back, assuming that the emulated (or assigned) hardware is not broken.
Exactly. The loop itself is still needed, but the check can be less restrictive.
The only thing we can is to improve detection code. Suggestions are welcome.
I think it's fine to assume that you're not assigning a 2004 era sound card to a guest. So I think the code is fine as it is, and can only suggest to add a comment explaining the mess.
True. Denis, care to submit a patch?
thanks,
Takashi