pm24.git/drivers/net/ethernet/intel, branch perf-mon

Merge branch '100GbE' of git://git.kernel.org/pub/scm/linux/kernel/git/tnguy/net-queue

2025-02-03T21:18:44Z

Tony Nguyen says: ==================== ice: fix Rx data path for heavy 9k MTU traffic Maciej Fijalkowski says: This patchset fixes a pretty nasty issue that was reported by RedHat folks which occurred after ~30 minutes (this value varied, just trying here to state that it was not observed immediately but rather after a considerable longer amount of time) when ice driver was tortured with jumbo frames via mix of iperf traffic executed simultaneously with wrk/nginx on client/server sides (HTTP and TCP workloads basically). The reported splats were spanning across all the bad things that can happen to the state of page - refcount underflow, use-after-free, etc. One of these looked as follows: [ 2084.019891] BUG: Bad page state in process swapper/34 pfn:97fcd0 [ 2084.025990] page:00000000a60ee772 refcount:-1 mapcount:0 mapping:0000000000000000 index:0x0 pfn:0x97fcd0 [ 2084.035462] flags: 0x17ffffc0000000(node=0|zone=2|lastcpupid=0x1fffff) [ 2084.041990] raw: 0017ffffc0000000 dead000000000100 dead000000000122 0000000000000000 [ 2084.049730] raw: 0000000000000000 0000000000000000 ffffffffffffffff 0000000000000000 [ 2084.057468] page dumped because: nonzero _refcount [ 2084.062260] Modules linked in: bonding tls sunrpc intel_rapl_msr intel_rapl_common intel_uncore_frequency intel_uncore_frequency_common i10nm_edac nfit libnvdimm x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm mgag200 irqd [ 2084.137829] CPU: 34 PID: 0 Comm: swapper/34 Kdump: loaded Not tainted 5.14.0-427.37.1.el9_4.x86_64 #1 [ 2084.147039] Hardware name: Dell Inc. PowerEdge R750/0216NK, BIOS 1.13.2 12/19/2023 [ 2084.154604] Call Trace: [ 2084.157058] [ 2084.159080] dump_stack_lvl+0x34/0x48 [ 2084.162752] bad_page.cold+0x63/0x94 [ 2084.166333] check_new_pages+0xb3/0xe0 [ 2084.170083] rmqueue_bulk+0x2d2/0x9e0 [ 2084.173749] ? ktime_get+0x35/0xa0 [ 2084.177159] rmqueue_pcplist+0x13b/0x210 [ 2084.181081] rmqueue+0x7d3/0xd40 [ 2084.184316] ? xas_load+0x9/0xa0 [ 2084.187547] ? xas_find+0x183/0x1d0 [ 2084.191041] ? xa_find_after+0xd0/0x130 [ 2084.194879] ? intel_iommu_iotlb_sync_map+0x89/0xe0 [ 2084.199759] get_page_from_freelist+0x11f/0x530 [ 2084.204291] __alloc_pages+0xf2/0x250 [ 2084.207958] ice_alloc_rx_bufs+0xcc/0x1c0 [ice] [ 2084.212543] ice_clean_rx_irq+0x631/0xa20 [ice] [ 2084.217111] ice_napi_poll+0xdf/0x2a0 [ice] [ 2084.221330] __napi_poll+0x27/0x170 [ 2084.224824] net_rx_action+0x233/0x2f0 [ 2084.228575] __do_softirq+0xc7/0x2ac [ 2084.232155] __irq_exit_rcu+0xa1/0xc0 [ 2084.235821] common_interrupt+0x80/0xa0 [ 2084.239662] [ 2084.241768] The fix is mostly about reverting what was done in commit 1dc1a7e7f410 ("ice: Centrallize Rx buffer recycling") followed by proper timing on page_count() storage and then removing the ice_rx_buf::act related logic (which was mostly introduced for purposes from cited commit). Special thanks to Xu Du for providing reproducer and Jacob Keller for initial extensive analysis. * '100GbE' of git://git.kernel.org/pub/scm/linux/kernel/git/tnguy/net-queue: ice: stop storing XDP verdict within ice_rx_buf ice: gather page_count()'s of each frag right before XDP prog call ice: put Rx buffers after being done with current frame ==================== Link: https://patch.msgid.link/20250131185415.3741532-1-anthony.l.nguyen@intel.com Signed-off-by: Jakub Kicinski

ice: Add check for devm_kzalloc()

2025-02-02T00:57:50Z

Add check for the return value of devm_kzalloc() to guarantee the success of allocation. Fixes: 42c2eb6b1f43 ("ice: Implement devlink-rate API") Signed-off-by: Jiasheng Jiang Reviewed-by: Michal Swiatkowski Link: https://patch.msgid.link/20250131013832.24805-1-jiashengjiangcool@gmail.com Signed-off-by: Jakub Kicinski

ice: stop storing XDP verdict within ice_rx_buf

2025-01-31T18:07:46Z

Idea behind having ice_rx_buf::act was to simplify and speed up the Rx data path by walking through buffers that were representing cleaned HW Rx descriptors. Since it caused us a major headache recently and we rolled back to old approach that 'puts' Rx buffers right after running XDP prog/creating skb, this is useless now and should be removed. Get rid of ice_rx_buf::act and related logic. We still need to take care of a corner case where XDP program releases a particular fragment. Make ice_run_xdp() to return its result and use it within ice_put_rx_mbuf(). Fixes: 2fba7dc5157b ("ice: Add support for XDP multi-buffer on Rx side") Reviewed-by: Przemek Kitszel Reviewed-by: Simon Horman Signed-off-by: Maciej Fijalkowski Tested-by: Chandan Kumar Rout (A Contingent Worker at Intel) Signed-off-by: Tony Nguyen

ice: gather page_count()'s of each frag right before XDP prog call

2025-01-31T18:06:26Z

If we store the pgcnt on few fragments while being in the middle of gathering the whole frame and we stumbled upon DD bit not being set, we terminate the NAPI Rx processing loop and come back later on. Then on next NAPI execution we work on previously stored pgcnt. Imagine that second half of page was used actively by networking stack and by the time we came back, stack is not busy with this page anymore and decremented the refcnt. The page reuse algorithm in this case should be good to reuse the page but given the old refcnt it will not do so and attempt to release the page via page_frag_cache_drain() with pagecnt_bias used as an arg. This in turn will result in negative refcnt on struct page, which was initially observed by Xu Du. Therefore, move the page count storage from ice_get_rx_buf() to a place where we are sure that whole frame has been collected, but before calling XDP program as it internally can also change the page count of fragments belonging to xdp_buff. Fixes: ac0753391195 ("ice: Store page count inside ice_rx_buf") Reported-and-tested-by: Xu Du Reviewed-by: Przemek Kitszel Reviewed-by: Simon Horman Co-developed-by: Jacob Keller Signed-off-by: Jacob Keller Signed-off-by: Maciej Fijalkowski Tested-by: Chandan Kumar Rout (A Contingent Worker at Intel) Signed-off-by: Tony Nguyen

ice: put Rx buffers after being done with current frame

2025-01-31T17:51:47Z

Introduce a new helper ice_put_rx_mbuf() that will go through gathered frags from current frame and will call ice_put_rx_buf() on them. Current logic that was supposed to simplify and optimize the driver where we go through a batch of all buffers processed in current NAPI instance turned out to be broken for jumbo frames and very heavy load that was coming from both multi-thread iperf and nginx/wrk pair between server and client. The delay introduced by approach that we are dropping is simply too big and we need to take the decision regarding page recycling/releasing as quick as we can. While at it, address an error path of ice_add_xdp_frag() - we were missing buffer putting from day 1 there. As a nice side effect we get rid of annoying and repetitive three-liner: xdp->data = NULL; rx_ring->first_desc = ntc; rx_ring->nr_frags = 0; by embedding it within introduced routine. Fixes: 1dc1a7e7f410 ("ice: Centrallize Rx buffer recycling") Reported-and-tested-by: Xu Du Reviewed-by: Przemek Kitszel Reviewed-by: Simon Horman Co-developed-by: Jacob Keller Signed-off-by: Jacob Keller Signed-off-by: Maciej Fijalkowski Tested-by: Chandan Kumar Rout (A Contingent Worker at Intel) Signed-off-by: Tony Nguyen

iavf: allow changing VLAN state without calling PF

2025-01-24T18:52:09Z

First case: > ip l a l $VF name vlanx type vlan id 100 > ip l d vlanx > ip l a l $VF name vlanx type vlan id 100 As workqueue can be execute after sometime, there is a window to have call trace like that: - iavf_del_vlan - iavf_add_vlan - iavf_del_vlans (wq) It means that our VLAN 100 will change the state from IAVF_VLAN_ACTIVE to IAVF_VLAN_REMOVE (iavf_del_vlan). After that in iavf_add_vlan state won't be changed because VLAN 100 is on the filter list. The final result is that the VLAN 100 filter isn't added in hardware (no iavf_add_vlans call). To fix that change the state if the filter wasn't removed yet directly to active. It is save as IAVF_VLAN_REMOVE means that virtchnl message wasn't sent yet. Second case: > ip l a l $VF name vlanx type vlan id 100 Any type of VF reset ex. change trust > ip l s $PF vf $VF_NUM trust on > ip l d vlanx > ip l a l $VF name vlanx type vlan id 100 In case of reset iavf driver is responsible for readding all filters that are being used. To do that all VLAN filters state are changed to IAVF_VLAN_ADD. Here is even longer window for changing VLAN state from kernel side, as workqueue isn't called immediately. We can have call trace like that: - changing to IAVF_VLAN_ADD (after reset) - iavf_del_vlan (called from kernel ops) - iavf_del_vlans (wq) Not exsisitng VLAN filters will be removed from hardware. It isn't a bug, ice driver will handle it fine. However, we can have call trace like that: - changing to IAVF_VLAN_ADD (after reset) - iavf_del_vlan (called from kernel ops) - iavf_add_vlan (called from kernel ops) - iavf_del_vlans (wq) With fix for previous case we end up with no VLAN filters in hardware. We have to remove VLAN filters if the state is IAVF_VLAN_ADD and delete VLAN was called. It is save as IAVF_VLAN_ADD means that virtchnl message wasn't sent yet. Fixes: 0c0da0e95105 ("iavf: refactor VLAN filter states") Signed-off-by: Michal Swiatkowski Reviewed-by: Przemek Kitszel Tested-by: Rafal Romanowski Signed-off-by: Tony Nguyen

ice: remove invalid parameter of equalizer

2025-01-24T18:49:42Z

It occurred that in the commit 70838938e89c ("ice: Implement driver functionality to dump serdes equalizer values") the invalid DRATE parameter for reading has been added. The output of the command: $ ethtool -d returns the garbage value in the place where DRATE value should be stored. Remove mentioned parameter to prevent return of corrupted data to userspace. Fixes: 70838938e89c ("ice: Implement driver functionality to dump serdes equalizer values") Signed-off-by: Mateusz Polchlopek Reviewed-by: Michal Swiatkowski Tested-by: Rinitha S (A Contingent worker at Intel) Signed-off-by: Tony Nguyen

ice: fix ice_parser_rt::bst_key array size

2025-01-24T18:49:30Z

Fix &ice_parser_rt::bst_key size. It was wrongly set to 10 instead of 20 in the initial impl commit (see Fixes tag). All usage code assumed it was of size 20. That was also the initial size present up to v2 of the intro series [2], but halved by v3 [3] refactor described as "Replace magic hardcoded values with macros." The introducing series was so big that some ugliness was unnoticed, same for bugs :/ ICE_BST_KEY_TCAM_SIZE and ICE_BST_TCAM_KEY_SIZE were differing by one. There was tmp variable @j in the scope of edited function, but was not used in all places. This ugliness is now gone. I'm moving ice_parser_rt::pg_prio a few positions up, to fill up one of the holes in order to compensate for the added 10 bytes to the ::bst_key, resulting in the same size of the whole as prior to the fix, and minimal changes in the offsets of the fields. Extend also the debug dump print of the key to cover all bytes. To not have string with 20 "%02x" and 20 params, switch to ice_debug_array_w_prefix(). This fix obsoletes Ahmed's attempt at [1]. [1] https://lore.kernel.org/intel-wired-lan/20240823230847.172295-1-ahmed.zaki@intel.com [2] https://lore.kernel.org/intel-wired-lan/20230605054641.2865142-13-junfeng.guo@intel.com [3] https://lore.kernel.org/intel-wired-lan/20230817093442.2576997-13-junfeng.guo@intel.com Reported-by: Dan Carpenter Closes: https://lore.kernel.org/intel-wired-lan/b1fb6ff9-b69e-4026-9988-3c783d86c2e0@stanley.mountain Fixes: 9a4c07aaa0f5 ("ice: add parser execution main loop") CC: Ahmed Zaki Reviewed-by: Larysa Zaremba Signed-off-by: Przemek Kitszel Reviewed-by: Simon Horman Tested-by: Rafal Romanowski Signed-off-by: Tony Nguyen

idpf: add more info during virtchnl transaction timeout/salt mismatch

2025-01-24T18:03:39Z

Add more information related to the transaction like cookie, vc_op, salt when transaction times out and include similar information when transaction salt does not match. Info output for transaction timeout: ------------------- (op:5015 cookie:45fe vc_op:5015 salt:45 timeout:60000ms) ------------------- before it was: ------------------- (op 5015, 60000ms) ------------------- Signed-off-by: Manoj Vishwanathan Signed-off-by: Brian Vazquez Reviewed-by: Jacob Keller Reviewed-by: Pavan Kumar Linga Reviewed-by: Paul Menzel Tested-by: Krishneil Singh Signed-off-by: Tony Nguyen

idpf: convert workqueues to unbound

2025-01-24T18:03:39Z

When a workqueue is created with `WQ_UNBOUND`, its work items are served by special worker-pools, whose host workers are not bound to any specific CPU. In the default configuration (i.e. when `queue_delayed_work` and friends do not specify which CPU to run the work item on), `WQ_UNBOUND` allows the work item to be executed on any CPU in the same node of the CPU it was enqueued on. While this solution potentially sacrifices locality, it avoids contention with other processes that might dominate the CPU time of the processor the work item was scheduled on. This is not just a theoretical problem: in a particular scenario misconfigured process was hogging most of the time from CPU0, leaving less than 0.5% of its CPU time to the kworker. The IDPF workqueues that were using the kworker on CPU0 suffered large completion delays as a result, causing performance degradation, timeouts and eventual system crash. Tested: * I have also run a manual test to gauge the performance improvement. The test consists of an antagonist process (`./stress --cpu 2`) consuming as much of CPU 0 as possible. This process is run under `taskset 01` to bind it to CPU0, and its priority is changed with `chrt -pQ 9900 10000 ${pid}` and `renice -n -20 ${pid}` after start. Then, the IDPF driver is forced to prefer CPU0 by editing all calls to `queue_delayed_work`, `mod_delayed_work`, etc... to use CPU 0. Finally, `ktraces` for the workqueue events are collected. Without the current patch, the antagonist process can force arbitrary delays between `workqueue_queue_work` and `workqueue_execute_start`, that in my tests were as high as `30ms`. With the current patch applied, the workqueue can be migrated to another unloaded CPU in the same node, and, keeping everything else equal, the maximum delay I could see was `6us`. Fixes: 0fe45467a104 ("idpf: add create vport and netdev configuration") Signed-off-by: Marco Leogrande Signed-off-by: Manoj Vishwanathan Signed-off-by: Brian Vazquez Reviewed-by: Jacob Keller Reviewed-by: Pavan Kumar Linga Tested-by: Krishneil Singh Signed-off-by: Tony Nguyen