Re: [PATCH net v2] net/smc: fix listen processing for SMC-Rv2

[Tony Lu <tonylu@linux.alibaba.com>]

On Tue, Aug 30, 2022 at 01:58:06PM +0800, liuyacan@corp.netease.com wrote:
> > > From: liuyacan <liuyacan@corp.netease.com>
> > > 
> > > After modifying the QP to the Error state, all RX WR would be
> > > completed with WC in IB_WC_WR_FLUSH_ERR status. Current
> > > implementation does not wait for it is done, but free the link
> > > directly. So there is a risk that accessing the freed link in
> > > tasklet context.
> > > 
> > > Here is a crash example:
> > > 
> > >  BUG: unable to handle page fault for address: ffffffff8f220860
> > >  #PF: supervisor write access in kernel mode
> > >  #PF: error_code(0x0002) - not-present page
> > >  PGD f7300e067 P4D f7300e067 PUD f7300f063 PMD 8c4e45063 PTE 800ffff08c9df060
> > >  Oops: 0002 [#1] SMP PTI
> > >  CPU: 1 PID: 0 Comm: swapper/1 Kdump: loaded Tainted: G S         OE     5.10.0-0607+ #23
> > >  Hardware name: Inspur NF5280M4/YZMB-00689-101, BIOS 4.1.20 07/09/2018
> > >  RIP: 0010:native_queued_spin_lock_slowpath+0x176/0x1b0
> > >  Code: f3 90 48 8b 32 48 85 f6 74 f6 eb d5 c1 ee 12 83 e0 03 83 ee 01 48 c1 e0 05 48 63 f6 48 05 00 c8 02 00 48 03 04 f5 00 09 98 8e <48> 89 10 8b 42 08 85 c0 75 09 f3 90 8b 42 08 85 c0 74 f7 48 8b 32
> > >  RSP: 0018:ffffb3b6c001ebd8 EFLAGS: 00010086
> > >  RAX: ffffffff8f220860 RBX: 0000000000000246 RCX: 0000000000080000
> > >  RDX: ffff91db1f86c800 RSI: 000000000000173c RDI: ffff91db62bace00
> > >  RBP: ffff91db62bacc00 R08: 0000000000000000 R09: c00000010000028b
> > >  R10: 0000000000055198 R11: ffffb3b6c001ea58 R12: ffff91db80e05010
> > >  R13: 000000000000000a R14: 0000000000000006 R15: 0000000000000040
> > >  FS:  0000000000000000(0000) GS:ffff91db1f840000(0000) knlGS:0000000000000000
> > >  CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> > >  CR2: ffffffff8f220860 CR3: 00000001f9580004 CR4: 00000000003706e0
> > >  DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
> > >  DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
> > >  Call Trace:
> > >   <IRQ>
> > >   _raw_spin_lock_irqsave+0x30/0x40
> > >   mlx5_ib_poll_cq+0x4c/0xc50 [mlx5_ib]
> > >   smc_wr_rx_tasklet_fn+0x56/0xa0 [smc]
> > >   tasklet_action_common.isra.21+0x66/0x100
> > >   __do_softirq+0xd5/0x29c
> > >   asm_call_irq_on_stack+0x12/0x20
> > >   </IRQ>
> > >   do_softirq_own_stack+0x37/0x40
> > >   irq_exit_rcu+0x9d/0xa0
> > >   sysvec_call_function_single+0x34/0x80
> > >   asm_sysvec_call_function_single+0x12/0x20
> > > 
> > > Signed-off-by: liuyacan <liuyacan@corp.netease.com>
> > > ---
> > >  net/smc/smc_core.c |  2 ++
> > >  net/smc/smc_core.h |  2 ++
> > >  net/smc/smc_wr.c   | 12 ++++++++++++
> > >  net/smc/smc_wr.h   |  3 +++
> > >  4 files changed, 19 insertions(+)
> > > 
> > > diff --git a/net/smc/smc_core.c b/net/smc/smc_core.c
> > > index ff49a11f5..b632a33f1 100644
> > > --- a/net/smc/smc_core.c
> > > +++ b/net/smc/smc_core.c
> > > @@ -752,6 +752,7 @@ int smcr_link_init(struct smc_link_group *lgr, struct smc_link *lnk,
> > >  	atomic_inc(&lnk->smcibdev->lnk_cnt);
> > >  	refcount_set(&lnk->refcnt, 1); /* link refcnt is set to 1 */
> > >  	lnk->clearing = 0;
> > > +	lnk->rx_drained = 0;
> > >  	lnk->path_mtu = lnk->smcibdev->pattr[lnk->ibport - 1].active_mtu;
> > >  	lnk->link_id = smcr_next_link_id(lgr);
> > >  	lnk->lgr = lgr;
> > > @@ -1269,6 +1270,7 @@ void smcr_link_clear(struct smc_link *lnk, bool log)
> > >  	smcr_buf_unmap_lgr(lnk);
> > >  	smcr_rtoken_clear_link(lnk);
> > >  	smc_ib_modify_qp_error(lnk);
> > > +	smc_wr_drain_cq(lnk);
> > >  	smc_wr_free_link(lnk);
> > >  	smc_ib_destroy_queue_pair(lnk);
> > >  	smc_ib_dealloc_protection_domain(lnk);
> > > diff --git a/net/smc/smc_core.h b/net/smc/smc_core.h
> > > index fe8b524ad..0a469a3e7 100644
> > > --- a/net/smc/smc_core.h
> > > +++ b/net/smc/smc_core.h
> > > @@ -117,6 +117,7 @@ struct smc_link {
> > >  	u64			wr_rx_id;	/* seq # of last recv WR */
> > >  	u32			wr_rx_cnt;	/* number of WR recv buffers */
> > >  	unsigned long		wr_rx_tstamp;	/* jiffies when last buf rx */
> > > +	wait_queue_head_t       wr_rx_drain_wait; /* wait for WR drain */
> > >  
> > >  	struct ib_reg_wr	wr_reg;		/* WR register memory region */
> > >  	wait_queue_head_t	wr_reg_wait;	/* wait for wr_reg result */
> > > @@ -138,6 +139,7 @@ struct smc_link {
> > >  	u8			link_idx;	/* index in lgr link array */
> > >  	u8			link_is_asym;	/* is link asymmetric? */
> > >  	u8			clearing : 1;	/* link is being cleared */
> > > +	u8                      rx_drained : 1; /* link is drained */
> > >  	refcount_t		refcnt;		/* link reference count */
> > >  	struct smc_link_group	*lgr;		/* parent link group */
> > >  	struct work_struct	link_down_wrk;	/* wrk to bring link down */
> > > diff --git a/net/smc/smc_wr.c b/net/smc/smc_wr.c
> > > index 26f8f240d..f9992896a 100644
> > > --- a/net/smc/smc_wr.c
> > > +++ b/net/smc/smc_wr.c
> > > @@ -465,6 +465,10 @@ static inline void smc_wr_rx_process_cqes(struct ib_wc wc[], int num)
> > >  			case IB_WC_RNR_RETRY_EXC_ERR:
> > >  			case IB_WC_WR_FLUSH_ERR:
> > >  				smcr_link_down_cond_sched(link);
> > > +				if (link->clearing && wc[i]->wr_id == link->wr_rx_id) {
> > > +					link->rx_drained = 1;
> > > +					wake_up(&link->wr_rx_drain_wait);
> > > +				}
> > 
> > I am wondering if we should wait for all the wc comes back?
> 
> I think yes, so other processes can safely destroy qp.
> 
> > 
> > >  				break;
> > >  			default:
> > >  				smc_wr_rx_post(link); /* refill WR RX */
> > > @@ -631,6 +635,13 @@ static void smc_wr_init_sge(struct smc_link *lnk)
> > >  	lnk->wr_reg.access = IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE;
> > >  }
> > >  
> > > +void smc_wr_drain_cq(struct smc_link *lnk)
> > > +{
> > > +	wait_event_interruptible_timeout(lnk->wr_rx_drain_wait,
> > > +					 (lnk->drained == 1),
> > > +					 SMC_WR_RX_WAIT_DRAIN_TIME);
> > > +}
> > 
> > Should we wait for it with timeout? It should eventually be wake up
> > normally before freeing link. Waiting for SMC_WR_RX_WAIT_DRAIN_TIME (2s)
> > may also have this issue, although the probability of occurrence is
> > greatly reduced.
> 
> Indeed, there should logically probably be a perpetual wait here. I'm just worried if it 
> will get stuck for some unknown reason.

IMHO, it's better to get stuck rather than to hide unknown issues. So I
think timeout is unnecessary. 

Cheers,
Tony Lu