Hi Sarosh,
Some comments in-line:
> -----Original Message-----
> From: Sarosh Arif <sarosh.a...@emumba.com>
> Sent: Friday, July 10, 2020 2:00 AM
> To: rsanf...@akamai.com; Carrillo, Erik G <erik.g.carri...@intel.com>;
> dev@dpdk.org
> Cc: sta...@dpdk.org; Sarosh Arif <sarosh.a...@emumba.com>;
> h.mikit...@gmail.com
> Subject: [PATCH v3] lib/librte_timer:fix corruption with reset
The subject is misleading - perhaps wording closer to the title of the Bugzilla
bug would be more helpful.
>
> If the user tries to reset/stop some other timer in it's callback function,
> which
> is also about to expire, using rte_timer_reset_sync/rte_timer_stop_sync the
> application goes into an infinite loop. This happens because
> rte_timer_reset_sync/rte_timer_stop_sync loop until the timer resets/stops
> and there is check inside timer_set_config_state which prevents a running
> timer from being reset/stopped by not it's own timer_cb. Therefore
> timer_set_config_state returns -1 due to which rte_timer_reset returns -1
> and rte_timer_reset_sync goes into an infinite loop.
>
> The soloution to this problem is to return -1 from
> rte_timer_reset_sync/rte_timer_stop_sync in case the user tries to
> reset/stop some other timer in it's callback function.
>
> Bugzilla ID: 491
> Fixes: 20d159f20543 ("timer: fix corruption with reset")
> Cc: h.mikit...@gmail.com
> Signed-off-by: Sarosh Arif <sarosh.a...@emumba.com>
> ---
> v2: remove line continuations
> v3: separate code and declarations
> ---
> lib/librte_timer/rte_timer.c | 26 ++++++++++++++++++++++++--
> lib/librte_timer/rte_timer.h | 4 ++--
> 2 files changed, 26 insertions(+), 4 deletions(-)
>
> diff --git a/lib/librte_timer/rte_timer.c b/lib/librte_timer/rte_timer.c index
> 6d19ce469..0cd3e2c86 100644
> --- a/lib/librte_timer/rte_timer.c
> +++ b/lib/librte_timer/rte_timer.c
> @@ -576,14 +576,24 @@ rte_timer_alt_reset(uint32_t timer_data_id, struct
> rte_timer *tim, }
>
> /* loop until rte_timer_reset() succeed */ -void
> +int
> rte_timer_reset_sync(struct rte_timer *tim, uint64_t ticks,
> enum rte_timer_type type, unsigned tim_lcore,
> rte_timer_cb_t fct, void *arg)
> {
> + struct rte_timer_data *timer_data;
> + TIMER_DATA_VALID_GET_OR_ERR_RET(default_data_id,
> timer_data, -EINVAL);
> +
> + if (tim->status.state == RTE_TIMER_RUNNING &&
> + (tim->status.owner != (uint16_t)tim_lcore ||
> + tim != timer_data->priv_timer[tim_lcore].running_tim))
> + return -1;
> +
As I understand it, Bugzilla 491 describes two scenarios where a hang can occur:
1. A timer's callback tries to synchronously reset/stop another timer in the
same run list
2. A timer's callback tries to synchronously reset/stop another timer in a
different run list whose lcore happens to be running a timer callback that is
synchronously resetting/stopping a timer in the first run list
The if condition from the patch above can be broken up as:
(tim->status.state == RTE_TIMER_RUNNING && tim->status.owner ==
(uint16_t)lcore_id && tim != timer_data->priv_timer[lcore_id].running_tim)
And
(tim->status.state == RTE_TIMER_RUNNING && tim->status.owner !=
(uint16_t)lcore_id)
This second condition could be transient and doesn't necessarily identify
scenario (2) above. In this case, the *_sync() calls could fail unnecessarily.
Offhand, I'm not seeing a way to more precisely detect scenario 2 above. I'm
wondering if some kind of a timeout parameter could be added to avoid hanging
instead. Thoughts?
As Stephen mentioned in another response, it looks like this will require an
API change. I believe this can be announced in the next release via
doc/guides/rel_notes/deprecation.rst. Then, the new API can be added in the
next ABI-breaking release, possibly with versioned symbols
(http://doc.dpdk.org/guides/contributing/abi_versioning.html#versioning-macros).
Thanks,
Erik
> while (rte_timer_reset(tim, ticks, type, tim_lcore,
> fct, arg) != 0)
> rte_pause();
> +
> + return 0;
> }
>
> static int
> @@ -642,11 +652,23 @@ rte_timer_alt_stop(uint32_t timer_data_id, struct
> rte_timer *tim) }
>
> /* loop until rte_timer_stop() succeed */ -void
> +int
> rte_timer_stop_sync(struct rte_timer *tim) {
> + struct rte_timer_data *timer_data;
> + unsigned int lcore_id = rte_lcore_id();
> +
> + TIMER_DATA_VALID_GET_OR_ERR_RET(default_data_id,
> timer_data, -EINVAL);
> +
> + if (tim->status.state == RTE_TIMER_RUNNING &&
> + (tim->status.owner != (uint16_t)lcore_id ||
> + tim != timer_data->priv_timer[lcore_id].running_tim))
> + return -1;
> +
> while (rte_timer_stop(tim) != 0)
> rte_pause();
> +
> + return 0;
> }
>
> /* Test the PENDING status of the timer handle tim */ diff --git
> a/lib/librte_timer/rte_timer.h b/lib/librte_timer/rte_timer.h index
> c6b3d450d..392ca423d 100644
> --- a/lib/librte_timer/rte_timer.h
> +++ b/lib/librte_timer/rte_timer.h
> @@ -275,7 +275,7 @@ int rte_timer_reset(struct rte_timer *tim, uint64_t
> ticks,
> * @param arg
> * The user argument of the callback function.
> */
> -void
> +int
> rte_timer_reset_sync(struct rte_timer *tim, uint64_t ticks,
> enum rte_timer_type type, unsigned tim_lcore,
> rte_timer_cb_t fct, void *arg);
> @@ -314,7 +314,7 @@ int rte_timer_stop(struct rte_timer *tim);
> * @param tim
> * The timer handle.
> */
> -void rte_timer_stop_sync(struct rte_timer *tim);
> +int rte_timer_stop_sync(struct rte_timer *tim);
>
> /**
> * Test if a timer is pending.
> --
> 2.17.1
