On 12 May 2016 at 16:09, Francois Romieu <rom...@fr.zoreil.com> wrote: >> On 12 May 2016 at 01:58, David Miller <da...@davemloft.net> wrote: >> > From: Ard Biesheuvel <ard.biesheu...@linaro.org> >> > Date: Wed, 11 May 2016 09:47:49 +0200 > [...] >> > I think we should just seriously consider changing the default, it's >> > a really outdated reasoning behind the current default setting. Maybe >> > relevant a decade ago, but probably not now. >> > >> > And if the card is completely disfunctional in said configuration, the >> > default is definitely wrong. >> >> The card is indeed completely disfunctional. So we could try to >> resurrect 353176888386 ("r8169: enable 64-bit DMA by default for PCI >> Express devices"), and instead of backing it out again if regressions >> are reported, blacklist the particular chips. This is a much better >> approach, since then we can also print some kind of diagnostic on >> those arm64 systems why such a blacklisted NIC is not supported. > > I doubt there will be much *reporting* from broken systems that > include plain old PCI realtek chipsets (r8169.c::RTL_CFG_0). Changing > the default for those is imnsho asking for troubles without clear > benefit (experimental evidence suggests that smsc etherpower II grows > older more easily than plain pci 8169 :o/ ). >
By resurrecting 353176888386, I mean the patch that changes the default for PCI express devices, so I think we are in agreement here? > I'd rather leave these alone and change the default for the PCI Express > chipsets. Btw, while it does not seem to hurt, they should not need any > CPlusCmd Dual Access Cycle tweak either. Realtek may establish it (Lin ?) > > A few news from the "pathologically better safe than sorry" squad: > I have switched the default on a couple of non-critical production > servers that include 8168c (RTL_GIGA_MAC_VER_22). It should give an hint > for hardware from 2008 ~ 2009. I'll do some basic sanity testing with > different chipsets. > Thanks for testing that. In the mean time, I will dust off the patch and rebase it to today's -next.