Skip to contentSkip to conversation

infiniband: sw: rdmavt: possible ABBA deadlocks in rvt_ruc_loopback()

BUG Jia-Ju Bai 3 participants 3 messages Oct 20, 2022 InfiniBand

lore
Share & tools
lore

Jia-Ju Bai

author
#0 lore raw

Hello,

My static analysis tool reports several possible ABBA deadlock in the rdmavt driver in Linux 5.18:

rvt_ruc_loopback()
  spin_lock_irqsave(&sqp->s_lock, flags); --> Line 3190 (Lock A) spin_lock(&sqp->r_lock); --> Line 3195 (Lock B)

rvt_qp_mr_clean() spin_lock_irq(&qp->r_lock); --> Line 698 (Lock B) spin_lock(&qp->s_lock); --> Line 700 (Lock A)

rvt_rc_timeout()
  spin_lock_irqsave(&qp->r_lock, flags); --> Line 2595 (Lock B) spin_lock(&qp->s_lock); --> Line 2596 (Lock A)

rvt_modify_qp()
  spin_lock_irq(&qp->r_lock); --> Line 1419 (Lock B) spin_lock(&qp->s_lock); --> Line 1421(Lock A)

_rvt_reset_qp()
  spin_lock_irq(&qp->r_lock); --> Line 907 (Lock B) spin_lock(&qp->s_lock); --> Line 909 (Lock A)

rvt_reset_qp()
  spin_lock_irq(&qp->r_lock); --> Line 936 (Lock B) spin_lock(&qp->s_lock); --> Line 938 (Lock A)

When rvt_ruc_loopback() is concurrently executed with rvt_qp_mr_clean(), rvt_rc_timeout(), rvt_modify_qp(), _rvt_reset_qp() or rvt_reset_qp(), the deadlocks can occur.

I am not quite sure whether these possible deadlocks are real and how to fix them if real. Any feedback would be appreciated, thanks :)

Best wishes,
Jia-Ju Bai

#1 lore raw
Quoted 33 lines · Jia-Ju Bai

Hello,

My static analysis tool reports several possible ABBA deadlock in the rdmavt driver in Linux 5.18:

rvt_ruc_loopback()
  spin_lock_irqsave(&sqp->s_lock, flags); --> Line 3190 (Lock A) spin_lock(&sqp->r_lock); --> Line 3195 (Lock B)

rvt_qp_mr_clean()
  spin_lock_irq(&qp->r_lock); --> Line 698 (Lock B) spin_lock(&qp->s_lock); --> Line 700 (Lock A)

rvt_rc_timeout()
  spin_lock_irqsave(&qp->r_lock, flags); --> Line 2595 (Lock B) spin_lock(&qp->s_lock); --> Line 2596 (Lock A)

rvt_modify_qp()
  spin_lock_irq(&qp->r_lock); --> Line 1419 (Lock B) spin_lock(&qp->s_lock); --> Line 1421(Lock A)

_rvt_reset_qp()
  spin_lock_irq(&qp->r_lock); --> Line 907 (Lock B) spin_lock(&qp->s_lock); --> Line 909 (Lock A)

rvt_reset_qp()
  spin_lock_irq(&qp->r_lock); --> Line 936 (Lock B) spin_lock(&qp->s_lock); --> Line 938 (Lock A)

When rvt_ruc_loopback() is concurrently executed with rvt_qp_mr_clean(), rvt_rc_timeout(), rvt_modify_qp(), _rvt_reset_qp() or rvt_reset_qp(), the deadlocks can occur.

Quoted 5 lines

I am not quite sure whether these possible deadlocks are real and how to fix them if real.
Any feedback would be appreciated, thanks :)

Thanks, we'll look into them.

-Denny

#2 lore raw
In reply to Jia-Ju Bai
Quoted 30 lines · Jia-Ju Bai

Hello,

My static analysis tool reports several possible ABBA deadlock in the rdmavt driver in Linux 5.18:

rvt_ruc_loopback()
  spin_lock_irqsave(&sqp->s_lock, flags); --> Line 3190 (Lock A) spin_lock(&sqp->r_lock); --> Line 3195 (Lock B)

rvt_qp_mr_clean() spin_lock_irq(&qp->r_lock); --> Line 698 (Lock B) spin_lock(&qp->s_lock); --> Line 700 (Lock A)

rvt_rc_timeout()
  spin_lock_irqsave(&qp->r_lock, flags); --> Line 2595 (Lock B) spin_lock(&qp->s_lock); --> Line 2596 (Lock A)

rvt_modify_qp()
  spin_lock_irq(&qp->r_lock); --> Line 1419 (Lock B) spin_lock(&qp->s_lock); --> Line 1421(Lock A)

_rvt_reset_qp()
  spin_lock_irq(&qp->r_lock); --> Line 907 (Lock B) spin_lock(&qp->s_lock); --> Line 909 (Lock A)

rvt_reset_qp()
  spin_lock_irq(&qp->r_lock); --> Line 936 (Lock B) spin_lock(&qp->s_lock); --> Line 938 (Lock A)

When rvt_ruc_loopback() is concurrently executed with rvt_qp_mr_clean(), rvt_rc_timeout(), rvt_modify_qp(), _rvt_reset_qp() or rvt_reset_qp(), the deadlocks can occur.

Quoted 8 lines

I am not quite sure whether these possible deadlocks are real and how to fix them if real. Any feedback would be appreciated, thanks :)

Best wishes,
Jia-Ju Bai

Hi

I don't know whether this deadlock is real or not, but I wrote the offending code. If it's a real bug, I'm sorry about that.
If that's the case the bug is also present in 2 other places of rvt_ruc_loopback() in v5.19+:

1)
spin_lock_irqsave(&sqp->s_lock, flags); --> Line 3133 spin_lock(&sqp->r_lock); --> Line 3137

2)
spin_lock_irqsave(&sqp->s_lock, flags); --> Line 3192 spin_lock(&sqp->r_lock); --> Line 3193

Kind regards
Niels