[PATCH 1/5] interrupts: fix number of bytes read for vdev

[PATCH 1/5] interrupts: fix number of bytes read for vdev

[skoteshwar]

From: Satha Rao <skoteshwar@marvell.com>

Correct the number of bytes to read during vdev interrupt processing.
The driver sets this value, and the rxtx interrupt handling already
performs this correctly.

Fixes: 99998feec90 ("eal/linux: add interrupt type for vdev")
Cc: stable@dpdk.org

Signed-off-by: Satha Rao <skoteshwar@marvell.com>
---
 lib/eal/linux/eal_interrupts.c | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/lib/eal/linux/eal_interrupts.c b/lib/eal/linux/eal_interrupts.c
index 6436f796eb..80eef7ec4b 100644
--- a/lib/eal/linux/eal_interrupts.c
+++ b/lib/eal/linux/eal_interrupts.c
@@ -949,6 +949,10 @@ eal_intr_process_interrupts(struct epoll_event *events, int nfds)
 			break;
 #endif
 		case RTE_INTR_HANDLE_VDEV:
+			/* For vdev, number of bytes to read is set by driver */
+			bytes_read = rte_intr_efd_counter_size_get(src->intr_handle);
+			call = true;
+			break;
 		case RTE_INTR_HANDLE_EXT:
 			bytes_read = 0;
 			call = true;
-- 
2.25.1

Re: [PATCH 1/5] interrupts: fix number of bytes read for vdev

[Stephen Hemminger]

On Mon, 7 Oct 2024 20:33:59 +0530
<skoteshwar@marvell.com> wrote:

> From: Satha Rao <skoteshwar@marvell.com>
> 
> Correct the number of bytes to read during vdev interrupt processing.
> The driver sets this value, and the rxtx interrupt handling already
> performs this correctly.
> 
> Fixes: 99998feec90 ("eal/linux: add interrupt type for vdev")
> Cc: stable@dpdk.org
> 
> Signed-off-by: Satha Rao <skoteshwar@marvell.com>

Not sure how this works.
The code in eal_handle_interrupts sits on a epoll_wait() waiting
for next interrupt.

eal_inter_process_interrupts() is called.
The switch you modified decides what needs to be read to clear the interrupt.
If you change RTE_INTR_HANDLE_VDEV to read bytes from the eventfd 
by setting bytes_read then the code in this function will read the counter
value from the event fd immediately.

Then when eal_intr_proc_rxtx_intr() is called it will decide to read again.
And since the event is cleared, this read will get EWOULDBLOCK.

What exactly is the issue? How to reproduce it?

Also setting call = true in the switch is wrong.
The code later in eal_intr_process_interrupts that handles bytes_read > 0
already sets call = true if interrupt was cleared from eventfd.