Re: [RFC PATCH net] net/mlx5: Perform DMA operations cleanup before pci_disable_device()

[Saeed Mahameed <saeed@kernel.org>]

On 12 Oct 13:51, Niklas Schnelle wrote:
>On Wed, 2023-10-11 at 11:45 -0700, Saeed Mahameed wrote:
>> From: Shay Drory <shayd@nvidia.com>
>>
>> The cited patch change mlx5 driver so that during probe, DMA
>> operations were performed before pci_enable_device() and during
>> teardown, DMA operations were performed after pci_disable_device().
>> DMA operations require PCI to be enabled. Hence, The above leads to
>> the following oops in PPC systems[2].
>>
>> Fix it by performing the DMA operations during probe, after
>> pci_enable_device() and during teardown, before pci_disable_device().
>>
>> This also fixes a problem reported by Niklas Schnelle [1]
>>
>> [1] https://lore.kernel.org/lkml/20231011-mlx5_init_fix-v3-1-787ffb9183c6@linux.ibm.com/
>>
>> [2]
>> Oops: Kernel access of bad area, sig: 11 [#1]
>> LE PAGE_SIZE=64K MMU=Radix SMP NR_CPUS=2048 NUMA pSeries
>> Modules linked in: xt_MASQUERADE nf_conntrack_netlink
>> nfnetlink xfrm_user iptable_nat xt_addrtype xt_conntrack nf_nat
>> nf_conntrack nf_defrag_ipv6 nf_defrag_ipv4 netconsole rpcsec_gss_krb5
>> auth_rpcgss oid_registry overlay rpcrdma rdma_ucm ib_iser ib_umad
>> rdma_cm ib_ipoib iw_cm libiscsi scsi_transport_iscsi ib_cm ib_uverbs
>> ib_core mlx5_core(-) ptp pps_core fuse vmx_crypto crc32c_vpmsum [last
>> unloaded: mlx5_ib]
>> CPU: 1 PID: 8937 Comm: modprobe Not tainted 6.5.0-rc3_for_upstream_min_debug_2023_07_31_16_02 #1
>> Hardware name: IBM pSeries (emulated by qemu) POWER9 (raw) 0x4e1202 0xf000005 of:SLOF,HEAD hv:linux,kvm pSeries
>> NIP:  c000000000423388 LR: c0000000001e733c CTR: c0000000001e4720
>> REGS: c0000000055636d0 TRAP: 0380   Not tainted (6.5.0-rc3_for_upstream_min_debug_2023_07_31_16_02)
>> MSR:  8000000000009033  CR: 24008884  XER: 20040000
>> CFAR: c0000000001e7338 IRQMASK: 0
>> NIP [c000000000423388] __free_pages+0x28/0x160
>> LR [c0000000001e733c] dma_direct_free+0xac/0x190
>> Call Trace:
>> [c000000005563970] [5deadbeef0000100] 0x5deadbeef0000100 (unreliable)
>> [c0000000055639b0] [c0000000003d46cc] kfree+0x7c/0x150
>> [c000000005563a40] [c0000000001e47c8] dma_free_attrs+0xa8/0x1a0
>> [c000000005563aa0] [c008000000d0064c] mlx5_cmd_cleanup+0xa4/0x100 [mlx5_core]
>> [c000000005563ad0] [c008000000cf629c] mlx5_mdev_uninit+0xf4/0x140 [mlx5_core]
>> [c000000005563b00] [c008000000cf6448] remove_one+0x160/0x1d0 [mlx5_core]
>> [c000000005563b40] [c000000000958540] pci_device_remove+0x60/0x110
>> [c000000005563b80] [c000000000a35e80] device_remove+0x70/0xd0
>> [c000000005563bb0] [c000000000a37a38] device_release_driver_internal+0x2a8/0x330
>> [c000000005563c00] [c000000000a37b8c] driver_detach+0x8c/0x160
>> [c000000005563c40] [c000000000a35350] bus_remove_driver+0x90/0x110
>> [c000000005563c80] [c000000000a38948] driver_unregister+0x48/0x90
>> [c000000005563cf0] [c000000000957e38] pci_unregister_driver+0x38/0x150
>> [c000000005563d40] [c008000000eb6140] mlx5_cleanup+0x38/0x90 [mlx5_core]
>>
>> Fixes: 06cd555f73ca ("net/mlx5: split mlx5_cmd_init() to probe and reload routines")
>> Signed-off-by: Shay Drory <shayd@nvidia.com>
>> Reviewed-by: Moshe Shemesh <moshe@nvidia.com>
>> Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
>> Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
>> CC: Leon Romanovsky <leon@kernel.org>
>> CC: Niklas Schnelle <schnelle@linux.ibm.com>
>>
>> ---
>>  drivers/net/ethernet/mellanox/mlx5/core/cmd.c | 62 ++++++++-----------
>>  1 file changed, 27 insertions(+), 35 deletions(-)
>
>
>I can confirm that this indeed fixes the problem I was seeing. I also
>tested hot unplug and re-plug as well as some smoke tests with the
>devices at hand across multiple ConnectX generations.
>So feel free to add my:
>
>Tested-by: Niklas Schnelle <schnelle@linux.ibm.com>
>
>>
>> diff --git a/drivers/net/ethernet/mellanox/mlx5/core/cmd.c b/drivers/net/ethernet/mellanox/mlx5/core/cmd.c
>> index afb348579577..dd36d9cba62f 100644
>> --- a/drivers/net/ethernet/mellanox/mlx5/core/cmd.c
>> +++ b/drivers/net/ethernet/mellanox/mlx5/core/cmd.c
>> @@ -2186,52 +2186,23 @@ static u16 cmdif_rev(struct mlx5_core_dev *dev)
>>
>>  int mlx5_cmd_init(struct mlx5_core_dev *dev)
>>  {
>> -	int size = sizeof(struct mlx5_cmd_prot_block);
>> -	int align = roundup_pow_of_two(size);
>>  	struct mlx5_cmd *cmd = &dev->cmd;
>> -	u32 cmd_l;
>> -	int err;
>> -
>> -	cmd->pool = dma_pool_create("mlx5_cmd", mlx5_core_dma_dev(dev), size, align, 0);
>> -	if (!cmd->pool)
>> -		return -ENOMEM;
>>
>> -	err = alloc_cmd_page(dev, cmd);
>> -	if (err)
>> -		goto err_free_pool;
>> -
>> -	cmd_l = (u32)(cmd->dma);
>> -	if (cmd_l & 0xfff) {
>> -		mlx5_core_err(dev, "invalid command queue address\n");
>> -		err = -ENOMEM;
>> -		goto err_cmd_page;
>> -	}
>>  	cmd->checksum_disabled = 1;
>>
>>  	spin_lock_init(&cmd->alloc_lock);
>>  	spin_lock_init(&cmd->token_lock);
>>
>> -	create_msg_cache(dev);
>> -
>>  	set_wqname(dev);
>>  	cmd->wq = create_singlethread_workqueue(cmd->wq_name);
>>  	if (!cmd->wq) {
>>  		mlx5_core_err(dev, "failed to create command workqueue\n");
>> -		err = -ENOMEM;
>> -		goto err_cache;
>> +		return -ENOMEM;
>>  	}
>>
>>  	mlx5_cmdif_debugfs_init(dev);
>>
>>  	return 0;
>> -
>> -err_cache:
>> -	destroy_msg_cache(dev);
>> -err_cmd_page:
>> -	free_cmd_page(dev, cmd);
>> -err_free_pool:
>> -	dma_pool_destroy(cmd->pool);
>> -	return err;
>>  }
>
>I like that this leaves mlx5_cmd_init() simpler.
>
>>
>>  void mlx5_cmd_cleanup(struct mlx5_core_dev *dev)
>> @@ -2240,15 +2211,15 @@ void mlx5_cmd_cleanup(struct mlx5_core_dev *dev)
>>
>>  	mlx5_cmdif_debugfs_cleanup(dev);
>>  	destroy_workqueue(cmd->wq);
>> -	destroy_msg_cache(dev);
>> -	free_cmd_page(dev, cmd);
>> -	dma_pool_destroy(cmd->pool);
>>  }
>>
>>  int mlx5_cmd_enable(struct mlx5_core_dev *dev)
>>  {
>> +	int size = sizeof(struct mlx5_cmd_prot_block);
>> +	int align = roundup_pow_of_two(size);
>>  	struct mlx5_cmd *cmd = &dev->cmd;
>>  	u32 cmd_h, cmd_l;
>> +	int err;
>>
>>  	memset(&cmd->vars, 0, sizeof(cmd->vars));
>>  	cmd->vars.cmdif_rev = cmdif_rev(dev);
>> @@ -2281,10 +2252,21 @@ int mlx5_cmd_enable(struct mlx5_core_dev *dev)
>>  	sema_init(&cmd->vars.pages_sem, 1);
>>  	sema_init(&cmd->vars.throttle_sem, DIV_ROUND_UP(cmd->vars.max_reg_cmds, 2));
>>
>> +	cmd->pool = dma_pool_create("mlx5_cmd", mlx5_core_dma_dev(dev), size, align, 0);
>> +	if (!cmd->pool)
>> +		return -ENOMEM;
>> +
>> +	err = alloc_cmd_page(dev, cmd);
>> +	if (err)
>> +		goto err_free_pool;
>> +
>>  	cmd_h = (u32)((u64)(cmd->dma) >> 32);
>>  	cmd_l = (u32)(cmd->dma);
>> -	if (WARN_ON(cmd_l & 0xfff))
>> -		return -EINVAL;
>> +	if (cmd_l & 0xfff) {
>> +		mlx5_core_err(dev, "invalid command queue address\n");
>> +		err = -ENOMEM;
>> +		goto err_cmd_page;
>> +	}
>>
>>  	iowrite32be(cmd_h, &dev->iseg->cmdq_addr_h);
>>  	iowrite32be(cmd_l, &dev->iseg->cmdq_addr_l_sz);
>> @@ -2297,16 +2279,26 @@ int mlx5_cmd_enable(struct mlx5_core_dev *dev)
>>  	cmd->mode = CMD_MODE_POLLING;
>>  	cmd->allowed_opcode = CMD_ALLOWED_OPCODE_ALL;
>>
>> +	create_msg_cache(dev);
>>  	create_debugfs_files(dev);
>>
>>  	return 0;
>> +
>> +err_cmd_page:
>> +	free_cmd_page(dev, cmd);
>> +err_free_pool:
>> +	dma_pool_destroy(cmd->pool);
>> +	return err;
>>  }
>>
>>  void mlx5_cmd_disable(struct mlx5_core_dev *dev)
>>  {
>>  	struct mlx5_cmd *cmd = &dev->cmd;
>>
>> +	destroy_msg_cache(dev);
>>  	clean_debug_files(dev);
>> +	free_cmd_page(dev, cmd);
>> +	dma_pool_destroy(cmd->pool);
>>  	flush_workqueue(cmd->wq);
>>  }
>
>I do like that this tears down the DMA stuff before
>pci_disable_device() I don't think this fixes a problem on s390x but to
>me it's the more natural order. That said, are we sure that the
>flush_workqueue() won't still need DMA and cmd page?
>

Agreed, tt is safer to flush_workqueue first thing in the function.

>>
>> --
>> 2.41.0
>>
>