From mboxrd@z Thu Jan 1 00:00:00 1970 From: Neil Horman Subject: Re: [net PATCH] atl1c: Fix misuse of netdev_alloc_skb in refilling rx ring Date: Sun, 28 Jul 2013 14:53:18 -0400 Message-ID: <20130728185318.GA10795@neilslaptop.think-freely.org> References: <87k3kbdcmy.fsf@canonical.com> <1374960610.3607.13.camel@deadeye.wl.decadent.org.uk> <1374969583.3669.23.camel@edumazet-glaptop> <20130727.200205.67471633133830510.davem@davemloft.net> <20130728104446.GB9876@neilslaptop.think-freely.org> <1375028154.3669.30.camel@edumazet-glaptop> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: David Miller , bhutchings@solarflare.com, luis.henriques@canonical.com, netdev@vger.kernel.org, jcliburn@gmail.com, stable@vger.kernel.org To: Eric Dumazet Return-path: Content-Disposition: inline In-Reply-To: <1375028154.3669.30.camel@edumazet-glaptop> Sender: stable-owner@vger.kernel.org List-Id: netdev.vger.kernel.org On Sun, Jul 28, 2013 at 09:15:54AM -0700, Eric Dumazet wrote: > On Sun, 2013-07-28 at 06:44 -0400, Neil Horman wrote: > > On Sat, Jul 27, 2013 at 08:02:05PM -0700, David Miller wrote: > > > From: Eric Dumazet > > > Date: Sat, 27 Jul 2013 16:59:43 -0700 > > > > > > > If a hardware needs frame being in a single 4K page, its driver must do > > > > its own allocation, or add appropriate aligning logic. > > > > > > Whatever the cause the original patch in this thread is not the > > > correct fix and I've thus reverted it. > > > > > > > So what exactly is the consensus here? We can certainly modify the patch to > > ensure allocation on a 4k boundary, and within a unique 4k page, but thats just > > a guess at the problem, and the Qualcomm people have been silent on the issue > > for weeks now. > > > > Thats the guess yes, but since MTU can be more than 4096 bytes, it could > be something else, some kind of DMA problem. > Right, thats exactly my concern, its a guess, one which infers problems of its own. I'll certainly code up a patch to test, sure, given what we've discussed here, but I'm not sure it leaves us in any better position than we were in with my origional patch. > I wonder why rx_buffer_len includes VLAN_HLEN as this nic provides > accelerated vlan. > I expect its an overestimation of sizing requirements in the rx ring. The set_features method for the driver doesn't re-allocate the rx ring buffers if the vlan tag accleration feature is disabled, so I suspect that they keep that extra space in place in the event that its ever turned off. > drivers/net/ethernet/atheros/atl1c/atl1c_main.c uses : > hw->dmar_block = atl1c_dma_req_1024; > > while drivers/net/ethernet/atheros/atlx/atl1.c uses : > hw->dmar_block = atl1_dma_req_256; > Another question for the Atheros people I suppose. Although, atl1c takes this value through several other contortions in atl1c_configure_tx, in which it seems to allow the value to be either 2, or 3 (capped by the min_t call, and floored by the check against DEVICE_CTRL_MAXRRS_MIN). By all rights the setting in the atlx driver would be considered invalid in the atl1c driver (if such comparisons can be drawn). > drivers/net/ethernet/atheros/atlx/atl1.c also has this code around line > 1962 : > > /* rrd seems to be bad */ > if (unlikely(i-- > 0)) { > /* rrd may not be DMAed completely */ > udelay(1); > goto chk_rrd; > } > > Hmm, that is interesting, expecially given that atl1c has this code: rfd_num = (rrs->word0 >> RRS_RX_RFD_CNT_SHIFT) & RRS_RX_RFD_CNT_MASK; if (unlikely(rfd_num != 1)) /* TODO support mul rfd*/ if (netif_msg_rx_err(adapter)) dev_warn(&pdev->dev, "Multi rfd not support yet!\n"); goto rrs_checked; It appears that atl1c does a simmilar check to what atlx does, but atl1c just ignores the possibility that rfd_num may be 0. I wonder if I shouldn't add a loop simmilar to what alx does there. > I would add the following debugging aid just to make sure... > > diff --git a/drivers/net/ethernet/atheros/atl1c/atl1c_main.c b/drivers/net/ethernet/atheros/atl1c/atl1c_main.c > index 786a874..4c794f9 100644 > --- a/drivers/net/ethernet/atheros/atl1c/atl1c_main.c > +++ b/drivers/net/ethernet/atheros/atl1c/atl1c_main.c > @@ -1794,6 +1794,7 @@ rrs_checked: > buffer_info->length, PCI_DMA_FROMDEVICE); > skb = buffer_info->skb; > } else { > + WARN_ON_ONCE(1); > /* TODO */ > if (netif_msg_rx_err(adapter)) > dev_warn(&pdev->dev, > > > I agree, I think perhaps a loop higher up prior to the rrs_checked label to ensure that rfd_num is at least 1 prior to going through the rest of the rx path. I'll have something for the reporters to test tomorrow. Thanks! Neil >