- 16 Oct, 2010 2 commits
-
-
Eric Dumazet authored
commit b30973f8 (node-aware skb allocation) spread a wrong habit of allocating net drivers skbs on a given memory node : The one closest to the NIC hardware. This is wrong because as soon as we try to scale network stack, we need to use many cpus to handle traffic and hit slub/slab management on cross-node allocations/frees when these cpus have to alloc/free skbs bound to a central node. skb allocated in RX path are ephemeral, they have a very short lifetime : Extra cost to maintain NUMA affinity is too expensive. What appeared as a nice idea four years ago is in fact a bad one. In 2010, NIC hardwares are multiqueue, or we use RPS to spread the load, and two 10Gb NIC might deliver more than 28 million packets per second, needing all the available cpus. Cost of cross-node handling in network and vm stacks outperforms the small benefit hardware had when doing its DMA transfert in its 'local' memory node at RX time. Even trying to differentiate the two allocations done for one skb (the sk_buff on local node, the data part on NIC hardware node) is not enough to bring good performance. Signed-off-by:
Eric Dumazet <eric.dumazet@gmail.com> Acked-by:
Tom Herbert <therbert@google.com> Signed-off-by:
David S. Miller <davem@davemloft.net>
-
Eric Dumazet authored
Using standard skb allocations in r8169 leads to order-3 allocations (if PAGE_SIZE=4096), because NIC needs 16383 bytes, and skb overhead makes this bigger than 16384 -> 32768 bytes per "skb" Using kmalloc() permits to reduce memory requirements of one r8169 nic by 4Mbytes. (256 frames * 16Kbytes). This is fine since a hardware bug requires us to copy incoming frames, so we build real skb when doing this copy. Signed-off-by:
-
- 15 Oct, 2010 1 commit
-
-
John Fastabend authored
Remove a DCB check config from DCB configuration we continue to configure DCB even if it fails so don't even bother to check. Plus user space (lldpad) checks this before programming the hw anyways. Worse case is we program some values into the hw that don't make total sense resulting in incorrect bandwidth allocation. Signed-off-by:
John Fastabend <john.r.fastabend@intel.com> Signed-off-by:
Jeff Kirsher <jeffrey.t.kirsher@intel.com> Signed-off-by:
-
- 14 Oct, 2010 11 commits
-
-
Carolyn Wyborny authored
Signed-off-by:
Carolyn Wyborny <carolyn.wyborny@intel.com> Tested-by:
Emil Tantilov <emil.s.tantilov@intel.com> Signed-off-by:
-
Emil Tantilov authored
Following patch fixes warnings reported by `make namespacecheck` Reported by Stephen Hemminger CC: Stephen Hemminger <shemminger@vyatta.com> Signed-off-by:
Stephen Ko <stephen.s.ko@intel.com> Signed-off-by:
-
Emil Tantilov authored
Remove functions that are declared, but not used in the driver. This patch fixes warnings reported by `make namespacecheck` Reported by Stephen Hemminger CC: Stephen Hemminger <shemminger@vyatta.com> Signed-off-by:
-
Michael Chan authored
Add new interrupt ack functions and other hardware interface logic to support the new device. Update version to 2.2.6. Signed-off-by:
Michael Chan <mchan@broadcom.com> Signed-off-by:
-
Michael Chan authored
During cnic shutdown, the original driver code requires userspace to close the uio device within a few seconds. This doesn't always happen as the userapp may be hung or otherwise take a long time to close. The system may crash when this happens. We fix the problem by decoupling the uio structures from the cnic structures during cnic shutdown. We do not unregister the uio device until the cnic driver is unloaded. This eliminates the unreliable wait loop for uio to close. All uio structures are kept in a linked list. If the device is shutdown and later brought back up again, the uio strcture will be found in the linked list and coupled back to the cnic structures. Signed-off-by:
-
Michael Chan authored
and put all uio related structures and ring buffers in it. This allows uio operations to be done independent of the cnic device structures. Signed-off-by:
-
Michael Chan authored
to free all UIO related structures. Signed-off-by:
-
Michael Chan authored
The bnx2x devices require a 2 second quiet time before sending the last RAMROD command to destroy a connection. This sleep wait adds up to a long delay when iscsid is serially destroying maultiple connections. Create a workqueue to perform the final connection cleanup in the background to speed up the process. This significantly speeds up the process as the wait time can be done in parallel for multiple connections. Signed-off-by:
-
Michael Chan authored
Refactoring code for the next patch to defer connection clean up. Signed-off-by:
-
Michael Chan authored
so that we can additional bit definitions without requiring a spinlock. Signed-off-by:
-
Michael Chan authored
to reduce some duplicate code. Also, use tasklet_kill() in cnic_free_irq() to wait for the cnic_irq_task to complete. Signed-off-by:
-
- 13 Oct, 2010 4 commits
-
-
Kumar Sanghvi authored
Updates to Phonet doc for Pipe controller 'connect' socket implementation and changes related to socket options. Signed-off-by:
Kumar Sanghvi <kumar.sanghvi@stericsson.com> Signed-off-by:
-
Kumar Sanghvi authored
Based on suggestion by Rémi Denis-Courmont to implement 'connect' for Pipe controller logic, this patch implements 'connect' socket call for the Pipe controller logic. The patch does following:- - Removes setsockopts for PNPIPE_CREATE and PNPIPE_DESTROY - Adds setsockopt for setting the Pipe handle value - Implements connect socket call - Updates the Pipe controller logic User-space should now follow below sequence with Pipe controller:- -socket -bind -setsockopt for PNPIPE_PIPE_HANDLE -connect -setsockopt for PNPIPE_ENCAP_IP -setsockopt for PNPIPE_ENABLE GPRS/3G data has been tested working fine with this. Signed-off-by:
Rémi Denis-Courmont <remi.denis-courmont@nokia.com> Signed-off-by:
-
Paul Gortmaker authored
Remove all instances of legacy, or as yet to be implemented code that is currently living within an #if 0 ... #endif block. In the rare instance that some of it be needed in the future, it can still be dragged out of history, but there is no need for it to sit in mainline. Signed-off-by:
Paul Gortmaker <paul.gortmaker@windriver.com> Acked-by:
Neil Horman <nhorman@tuxdriver.com> Signed-off-by:
-
David S. Miller authored
Reported-by:
Sachin Sant <sachinp@in.ibm.com> Signed-off-by:
-
- 12 Oct, 2010 9 commits
-
-
Eric Dumazet authored
We tried very hard to remove all possible dev_hold()/dev_put() pairs in network stack, using RCU conversions. There is still an unavoidable device refcount change for every dst we create/destroy, and this can slow down some workloads (routers or some app servers, mmap af_packet) We can switch to a percpu refcount implementation, now dynamic per_cpu infrastructure is mature. On a 64 cpus machine, this consumes 256 bytes per device. On x86, dev_hold(dev) code : before lock incl 0x280(%ebx) after: movl 0x260(%ebx),%eax incl fs:(%eax) Stress bench : (Sending 160.000.000 UDP frames, IP route cache disabled, dual E5540 @2.53GHz, 32bit kernel, FIB_TRIE) Before: real 1m1.662s user 0m14.373s sys 12m55.960s After: real 0m51.179s user 0m15.329s sys 10m15.942s Signed-off-by: -
Dmitry Kravkov authored
Reported-by:
Dmitry Kravkov <dmitry@broadcom.com> Signed-off-by:
Vladislav Zolotarov <vladz@broadcom.com> Tested-by:
-
-
Gerrit Renker authored
This omits the redundant "DCCP:" in warning messages, since DCCP_WARN() already echoes the function name, avoiding messages like kernel: [10988.766503] dccp_close: DCCP: ABORT -- 209 bytes unread Signed-off-by:Gerrit Renker <gerrit@erg.abdn.ac.uk>
-
Gerrit Renker authored
This schedules an Ack when receiving a timestamp, exploiting the existing inet_csk_schedule_ack() function, saving one case in the `dccp_ack_pending()' function. Signed-off-by: -
Ivo Calado authored
This patch generalises the task of determining data loss from RFC 4340, 7.7.1. Let S_A, S_B be sequence numbers such that S_B is "after" S_A, and let N_B be the NDP count of packet S_B. Then, using modulo-2^48 arithmetic, D = S_B - S_A - 1 is an upper bound of the number of lost data packets, D - N_B is an approximation of the number of lost data packets (there are cases where this is not exact). The patch implements this as dccp_loss_count(S_A, S_B, N_B) := max(S_B - S_A - 1 - N_B, 0) Signed-off-by:Ivo Calado <ivocalado@embedded.ufcg.edu.br> Signed-off-by:
Erivaldo Xavier <desadoc@gmail.com> Signed-off-by:
Leandro Sales <leandroal@gmail.com> Signed-off-by:
-
Gerrit Renker authored
This removes the argument `more' from ccid_hc_tx_packet_sent, since it was nowhere used in the entire code. (Btw, this argument was not even used in the original KAME code where the function initially came from; compare the variable moreToSend in the freebsd61-dccp-kame-28.08.2006.patch kept by Emmanuel Lochin.) Signed-off-by: -
Gerrit Renker authored
After moving the assignment of GAR/ISS from dccp_connect_init() to dccp_transmit_skb(), the former function becomes very small, so that a merger with dccp_connect() suggests itself. Signed-off-by: -
Gerrit Renker authored
This fixes a problem and a potential loophole with regard to seqno/ackno validity: currently the initial adjustments to AWL/SWL are only performed once at the begin of the connection, during the handshake. Since the Sequence Window feature is always greater than Wmin=32 (7.5.2), it is however necessary to perform these adjustments at least for the first W/W' (variables as per 7.5.1) packets in the lifetime of a connection. This requirement is complicated by the fact that W/W' can change at any time during the lifetime of a connection. Therefore it is better to perform that safety check each time SWL/AWL are updated, as implemented by the patch. A second problem solved by this patch is that the remote/local Sequence Window feature values (which set the bounds for AWL/SWL/SWH) are undefined until the feature negotiation has completed. During the initial handshake we have more stringent sequence number protection; the changes added by this patch effect that {A,S}W{L,H} are within the correct bounds at the instant that feature negotiation completes (since the SeqWin feature activation handlers call dccp_update_gsr/gss()). Signed-off-by:
-
- 11 Oct, 2010 13 commits
-
-
Michael Chan authored
To prevent unnecessary error message. pci_save_state() is also moved to the end of ->probe() so that all PCI config, including AER state, will be saved. Update version to 2.0.18. Signed-off-by:
-
Michael Chan authored
- Improved flow control and simplified interface - Use hardware RSS indirection table instead of the slower firmware- based table - Lower latency interrupt on 5709 Signed-off-by:
Benjamin Li <benli@broadcom.com> Signed-off-by:
-
Eric Dumazet authored
Le mardi 12 octobre 2010 à 00:02 +0200, Eric Dumazet a écrit : > Here is the followup patch. > > Thanks ! > Oops, this was an old version, the up2date ones also took care of "used" field. I guess its time for a sleep, sorry again. [PATCH net-next V2] neigh: reorder struct neighbour fields (refcnt) and (ha_lock, ha, used, dev, output, ops, primary_key) should be placed on a separate cache lines. refcnt can be often written, while other fields are mostly read. This gave me good result on stress test : before: real 0m45.570s user 0m15.525s sys 9m56.669s After: real 0m41.841s user 0m15.261s sys 8m45.949s Signed-off-by:
-
Eric Dumazet authored
struct dst_ops tracks number of allocated dst in an atomic_t field, subject to high cache line contention in stress workload. Switch to a percpu_counter, to reduce number of time we need to dirty a central location. Place it on a separate cache line to avoid dirtying read only fields. Stress test : (Sending 160.000.000 UDP frames, IP route cache disabled, dual E5540 @2.53GHz, 32bit kernel, FIB_TRIE, SLUB/NUMA) Before: real 0m51.179s user 0m15.329s sys 10m15.942s After: real 0m45.570s user 0m15.525s sys 9m56.669s With a small reordering of struct neighbour fields, subject of a following patch, (to separate refcnt from other read mostly fields) real 0m41.841s user 0m15.261s sys 8m45.949s Signed-off-by:
-
Eric Dumazet authored
Add a seqlock in struct neighbour to protect neigh->ha[], and avoid dirtying neighbour in stress situation (many different flows / dsts) Dirtying takes place because of read_lock(&n->lock) and n->used writes. Switching to a seqlock, and writing n->used only on jiffies changes permits less dirtying. Signed-off-by:
-
David S. Miller authored
Conflicts: net/core/ethtool.c
-
Kees Cook authored
Several other ethtool functions leave heap uncleared (potentially) by drivers. Some interfaces appear safe (eeprom, etc), in that the sizes are well controlled. In some situations (e.g. unchecked error conditions), the heap will remain unchanged in areas before copying back to userspace. Note that these are less of an issue since these all require CAP_NET_ADMIN. Cc: stable@kernel.org Signed-off-by:
Kees Cook <kees.cook@canonical.com> Acked-by:
Ben Hutchings <bhutchings@solarflare.com> Signed-off-by:
-
Jiri Slaby authored
Stanse found that pch_gbe_xmit_frame uses skb after it is freed. Fix that. Signed-off-by:
Jiri Slaby <jslaby@suse.cz> Cc: Masayuki Ohtake <masa-korg@dsn.okisemi.com> Acked-by:
-
Jiri Slaby authored
Stanse found that i2400m_rx frees skb, but still uses skb->len even though it has skb_len defined. So use skb_len properly in the code. And also define it unsinged int rather than size_t to solve compilation warnings. Signed-off-by:
Inaky Perez-Gonzalez <inaky.perez-gonzalez@intel.com> Signed-off-by:
-
Jiri Slaby authored
Stanse found that ia_init_one locks a spinlock and inside of that it calls ia_start which calls: * request_irq * tx_init which does kmalloc(GFP_KERNEL) Both of them can thus sleep and result in a deadlock. I don't see a reason to have a per-device spinlock there which is used only there and inited right before the lock location. So remove it completely. Signed-off-by:
-
Jiri Slaby authored
Stanse found that mpc_push frees skb and then it dereferences it. It is a typo, new_skb should be dereferenced there. Signed-off-by:
-
Jiri Slaby authored
Stanse found we do in console_show: kfree_skb(skb); return skb->len; which is not good. Fix that by remembering the len and use it in the function instead. Signed-off-by:
-
Eric Dumazet authored
When a new dst is used to send a frame, neigh_resolve_output() tries to associate an struct hh_cache to this dst, calling neigh_hh_init() with the neigh rwlock write locked. Most of the time, hh_cache is already known and linked into neighbour, so we find it and increment its refcount. This patch changes the logic so that we call neigh_hh_init() with neighbour lock read locked only, so that fast path can be run in parallel by concurrent cpus. This brings part of the speedup we got with commit c7d4426a (introduce DST_NOCACHE flag) for non cached dsts, even for cached ones, removing one of the contention point that routers hit on multiqueue enabled machines. Further improvements would need to use a seqlock instead of an rwlock to protect neigh->ha[], to not dirty neigh too often and remove two atomic ops. Signed-off-by:
-
