On 05.11.2020 02:45, Jakub Kicinski wrote: > On Tue, 3 Nov 2020 18:52:18 +0100 Heiner Kallweit wrote: >> Network problems with RTL8125B have been reported [0] and with help >> from Realtek it turned out that this chip version has a hw problem >> with short packets (similar to RTL8168evl). Having said that activate >> the same workaround as for RTL8168evl. >> Realtek suggested to activate the workaround for RTL8125A too, even >> though they're not 100% sure yet which RTL8125 versions are affected. >> >> [0] https://bugzilla.kernel.org/show_bug.cgi?id=209839 >> >> Fixes: 0439297be951 ("r8169: add support for RTL8125B") >> Reported-by: Maxim Plotnikov <w...@torlan.ru> >> Tested-by: Maxim Plotnikov <w...@torlan.ru> >> Signed-off-by: Heiner Kallweit <hkallwe...@gmail.com> > > Applied, thanks! > >> @@ -4125,7 +4133,7 @@ static bool rtl8169_tso_csum_v2(struct rtl8169_private >> *tp, >> >> opts[1] |= transport_offset << TCPHO_SHIFT; >> } else { >> - if (unlikely(rtl_test_hw_pad_bug(tp, skb))) >> + if (unlikely(skb->len < ETH_ZLEN && rtl_test_hw_pad_bug(tp))) >> return !eth_skb_pad(skb); >> } > > But looks like we may have another bug here - looks like this function > treas skb_cow_head() and eth_skb_pad() failures the same, but former > doesn't free the skb on error, while the latter does. > Thanks for the hint, indeed we have an issue. The caller of rtl8169_tso_csum_v2() also frees the skb if false is returned, therefore we have a double free if eth_skb_pad() fails.
When checking eth_skb_pad() I saw that it uses kfree_skb() to free the skb on error. Kernel documentation say about ndo_start_xmit context: Process with BHs disabled or BH (timer), will be called with interrupts disabled by netconsole. Is it safe to use kfree_skb() if interrupts are disabled? I'm asking because dev_kfree_skb_any() uses the irq path if irqs_disabled() is true.
