Re: [PATCH] net/mlx5: Add affinity for each irq

[Shay Drory <shayd@nvidia.com>]

On 6/6/2022 10:13, Yajun Deng wrote:
> The mlx5 would allocate no less than one irq for per cpu, we can bond each
> irq to a cpu to improve interrupt performance.

The maximum number of affinity set is hard coded to 4. in case nvec > 4 
* (num_CPUs)[1]
we will hit the following WARN[2].
Also, we hit an oops following this WARN...

[1]
mlx5 support up to 2K MSIX (depends on the HW). e.g.: if we max out mlx5 
MSIX capability,
we will cross this limit on any machine, at least that I know of.

[2]

This is a machine with 10 CPUs and 350 MSIX

[    1.633436] ------------[ cut here ]------------
  [    1.633437] WARNING: CPU: 2 PID: 194 at kernel/irq/affinity.c:443 irq_create_affinity_masks+0x175/0x270
  [    1.633467] Modules linked in: mlx5_core(+)
  [    1.633474] CPU: 2 PID: 194 Comm: systemd-modules Not tainted 5.18.0+ #1
  [    1.633480] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.13.0-0-gf21b5a4aeb02-prebuilt.qemu.org 04/01/2014
  [    1.633483] RIP: 0010:irq_create_affinity_masks+0x175/0x270
  [    1.633492] Code: 5c 41 5d 41 5e 41 5f c3 48 c7 46 20 90 6d 19 81 48 c7 c0 90 6d 19 81 8b 34 24 4c 89 ef ff d0 41 83 7d 08 04 0f 86 de fe ff ff <0f> 0b 45 31 f6 eb c5 45 8b 5d 00 8b 34 24 43 8d 04 1f 42 8d 0c 1e
  [    1.633497] RSP: 0018:ffff88810716bac0 EFLAGS: 00010202
  [    1.633501] RAX: 000000000000000a RBX: 0000000000000001 RCX: 0000000000000200
  [    1.633504] RDX: ffffffff82605000 RSI: ffffffff82605000 RDI: 0000000000000000
  [    1.633507] RBP: ffff88810716bbd0 R08: 000000000000000a R09: ffffffff82604fc0
  [    1.633510] R10: 0000000000000008 R11: 000ffffffffff000 R12: 0000000000000000
  [    1.633513] R13: ffff88810716bbd0 R14: 0000000000000160 R15: 0000000000000160
  [    1.633516] FS:  00007f8d72994b80(0000) GS:ffff88852c900000(0000) knlGS:0000000000000000
  [    1.633525] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
  [    1.633528] CR2: 00007f8d73ba4490 CR3: 0000000103fce001 CR4: 0000000000370ea0
  [    1.633531] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
  [    1.633534] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
  [    1.633536] Call Trace:
  [    1.633549]  <TASK>
  [    1.633553]  __pci_enable_msix_range+0x2b9/0x4c0
  [    1.633572]  pci_alloc_irq_vectors_affinity+0xa5/0x100
  [    1.633579]  mlx5_irq_table_create.cold+0x6d/0x22f [mlx5_core]
  [    1.634032]  ? probe_one+0x1aa/0x280 [mlx5_core]
  [    1.634193]  ? pci_device_probe+0xa4/0x140
  [    1.634201]  ? really_probe+0xc9/0x350
  [    1.634205]  ? pm_runtime_barrier+0x43/0x80
  [    1.634213]  ? __driver_probe_device+0x80/0x170
  [    1.634218]  ? driver_probe_device+0x1e/0x90
  [    1.634223]  ? __driver_attach+0xcd/0x1b0
  [    1.634226]  ? __device_attach_driver+0xf0/0xf0
  [    1.634231]  ? __device_attach_driver+0xf0/0xf0
  [    1.634235]  ? bus_for_each_dev+0x77/0xc0
  [    1.634243]  ? bus_add_driver+0x184/0x1f0
  [    1.634247]  ? driver_register+0x8f/0xe0
  [    1.634251]  ? 0xffffffffa0180000
  [    1.634256]  ? init+0x62/0x1000 [mlx5_core]
  [    1.634413]  ? do_one_initcall+0x4a/0x1e0
  [    1.634418]  ? kmem_cache_alloc_trace+0x33/0x420
  [    1.634426]  ? do_init_module+0x72/0x260
  [    1.634434]  ? __do_sys_finit_module+0xbb/0x130
  [    1.634443]  ? do_syscall_64+0x3d/0x90
  [    1.634452]  ? entry_SYSCALL_64_after_hwframe+0x46/0xb0
  [    1.634461]  </TASK>
  [    1.634463] ---[ end trace 0000000000000000 ]---
  [^[[0;32m  OK  ^[[0m] Finished ^[[0;1;39mudev Coldplug all Devices^[[0m.
  [    1.713428] Kernel panic - not syncing: stack-protector: Kernel stack is corrupted in: mlx5_irq_table_create+0x9c/0xa0 [mlx5_core]
  [    1.715521] CPU: 2 PID: 194 Comm: systemd-modules Tainted: G        W         5.18.0+ #1
  [    1.715524] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.13.0-0-gf21b5a4aeb02-prebuilt.qemu.org 04/01/2014
  [    1.715525] Call Trace:
  [    1.715532]  <TASK>
  [    1.715533]  dump_stack_lvl+0x34/0x44
  [    1.715538]  panic+0x100/0x255
  [    1.715542]  ? mlx5_irq_table_create+0x9c/0xa0 [mlx5_core]
  [    1.715602]  __stack_chk_fail+0x10/0x10
  [    1.715607]  mlx5_irq_table_create+0x9c/0xa0 [mlx5_core]
  [    1.715662]  ? probe_one+0x1aa/0x280 [mlx5_core]
  [    1.715709]  ? pci_device_probe+0xa4/0x140
  [    1.715712]  ? really_probe+0xc9/0x350
  [    1.715715]  ? pm_runtime_barrier+0x43/0x80
  [    1.715718]  ? __driver_probe_device+0x80/0x170
  [    1.715719]  ? driver_probe_device+0x1e/0x90
  [    1.715721]  ? __driver_attach+0xcd/0x1b0
  [    1.715722]  ? __device_attach_driver+0xf0/0xf0
  [    1.715723]  ? __device_attach_driver+0xf0/0xf0
  [    1.715724]  ? bus_for_each_dev+0x77/0xc0
  [    1.715727]  ? bus_add_driver+0x184/0x1f0
  [    1.715728]  ? driver_register+0x8f/0xe0
  [    1.715730]  ? 0xffffffffa0180000
  [    1.715731]  ? init+0x62/0x1000 [mlx5_core]
  [    1.715778]  ? do_one_initcall+0x4a/0x1e0
  [    1.715781]  ? kmem_cache_alloc_trace+0x33/0x420
  [    1.715784]  ? do_init_module+0x72/0x260
  [    1.715788]  ? __do_sys_finit_module+0xbb/0x130
  [    1.715790]  ? do_syscall_64+0x3d/0x90
  [    1.715792]  ? entry_SYSCALL_64_after_hwframe+0x46/0xb0
  [    1.715796]  </TASK>
  [    1.715938] Kernel Offset: disabled
  [    1.732563] ---[ end Kernel panic - not syncing: stack-protector: Kernel stack is corrupted in: mlx5_irq_table_create+0x9c/0xa0 [mlx5_core] ]---

>
> Signed-off-by: Yajun Deng <yajun.deng@linux.dev>
> ---
>   .../net/ethernet/mellanox/mlx5/core/pci_irq.c | 19 ++++++++++++++++++-
>   1 file changed, 18 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/net/ethernet/mellanox/mlx5/core/pci_irq.c b/drivers/net/ethernet/mellanox/mlx5/core/pci_irq.c
> index 662f1d55e30e..d13fc403fe78 100644
> --- a/drivers/net/ethernet/mellanox/mlx5/core/pci_irq.c
> +++ b/drivers/net/ethernet/mellanox/mlx5/core/pci_irq.c
> @@ -624,11 +624,27 @@ int mlx5_irq_table_get_num_comp(struct mlx5_irq_table *table)
>   	return table->pf_pool->xa_num_irqs.max - table->pf_pool->xa_num_irqs.min;
>   }
>   
> +static void mlx5_calc_sets(struct irq_affinity *affd, unsigned int nvecs)
> +{
> +	int i;
> +
> +	affd->nr_sets = (nvecs - 1) / num_possible_cpus() + 1;
> +
> +	for (i = 0; i < affd->nr_sets; i++) {
> +		affd->set_size[i] = min(nvecs, num_possible_cpus());
> +		nvecs -= num_possible_cpus();
> +	}
> +}
> +
>   int mlx5_irq_table_create(struct mlx5_core_dev *dev)
>   {
>   	int num_eqs = MLX5_CAP_GEN(dev, max_num_eqs) ?
>   		      MLX5_CAP_GEN(dev, max_num_eqs) :
>   		      1 << MLX5_CAP_GEN(dev, log_max_eq);
> +	struct irq_affinity affd = {
> +		.pre_vectors = 0,
> +		.calc_sets   = mlx5_calc_sets,
> +	};
>   	int total_vec;
>   	int pf_vec;
>   	int err;
> @@ -644,7 +660,8 @@ int mlx5_irq_table_create(struct mlx5_core_dev *dev)
>   		total_vec += MLX5_IRQ_CTRL_SF_MAX +
>   			MLX5_COMP_EQS_PER_SF * mlx5_sf_max_functions(dev);
>   
> -	total_vec = pci_alloc_irq_vectors(dev->pdev, 1, total_vec, PCI_IRQ_MSIX);
> +	total_vec = pci_alloc_irq_vectors_affinity(dev->pdev, 1, total_vec,
> +						   PCI_IRQ_MSIX | PCI_IRQ_AFFINITY, &affd);
>   	if (total_vec < 0)
>   		return total_vec;
>   	pf_vec = min(pf_vec, total_vec);