From: Eric Dumazet Date: Fri, 07 Jan 2011 21:26:58 +0100 Subject: [ath9k-devel] [PATCH 1/3] ath9k: Decrease skb size to fit into one page. In-Reply-To: <20110107200913.GF21588@tux> References: <1294361165-15308-1-git-send-email-greearb@candelatech.com> <1294397880.3467.1.camel@jlt3.sipsolutions.net> <4D275CCC.90905@candelatech.com> <20110107200913.GF21588@tux> Message-ID: <1294432018.2709.2.camel@edumazet-laptop> List-Id: MIME-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit To: ath9k-devel@lists.ath9k.org Le vendredi 07 janvier 2011 ? 12:09 -0800, Luis R. Rodriguez a ?crit : > On Fri, Jan 07, 2011 at 10:34:52AM -0800, Ben Greear wrote: > > On 01/07/2011 02:58 AM, Johannes Berg wrote: > > > On Thu, 2011-01-06 at 16:46 -0800, greearb at candelatech.com wrote: > > >> From: Ben Greear > > >> > > >> Patch is from Eric Dumazet, as described here: > > >> https://patchwork.kernel.org/patch/104271/ > > >> > > >> Reported-by: Michael Guntsche > > >> Signed-off-by: Eric Dumazet > > >> Signed-off-by: Ben Greear > > >> --- > > >> > > >> NOTE: This needs review by ath9k and/or other informed > > >> people. > > > > > > This doesn't make sense. It might help, but it'll probably lead to not > > > being able to receive all frames off the air. > > > > > > If this is an issue, ath9k should do paged RX like iwlwifi. > > > > Ok, I backed this out..but now I'm back to getting buffer allocation > > failures (and this is on a system with 2GB RAM). > > > > Seems it's coming from mac80211 instead of ath9k, at least most of > > the time (I'm using 60 stations, so it probably needs to make lots of > > copies in the rx path). The traffic I'm generating/receiving is 1024 byte UDP > > payloads. > > > > Does this mean I really received a packet that was 3872 bytes long, > > or is the skb_copy allocating/copying empty data? > > Good question. The buffer we setup for DMA should be large since we > need to support AMSDU RX up to a certain bytes of RX data for the frame. > Harwdware should tell us the right size for the RX'd data and the skb > should be set with that size, respectively. Following this logic, > skb_copy() should only allocate on the order of the required skb->len. > > Remember that trick we did to force the older memory leak issues by > forcing an skb_copy() on every RX'd frame and then just discarding that > buffer immediately? You can try to do the same and print the skb->len > there, just to check what's going on. Using skb_copy() is wrong then, since it makes a copy (order-1 allocations) It should use : skb_alloc( actual_size_of_frame not the 3840 thing ...) copy(data)