When QMAN_FQ_FLAG_DYNAMIC_FQID is set, there's a race condition between
fq_table[fq->idx] state and freeing/allocating from the pool and
WARN_ON(fq_table[fq->idx]) in qman_create_fq() gets triggered.

Indeed, we can have:
         Thread A                             Thread B
    qman_destroy_fq()                    qman_create_fq()
      qman_release_fqid()
        qman_shutdown_fq()
        gen_pool_free()
           -- At this point, the fqid is available again --
                                           qman_alloc_fqid()
           -- so, we can get the just-freed fqid in thread B --
                                           fq->fqid = fqid;
                                           fq->idx = fqid * 2;
                                           WARN_ON(fq_table[fq->idx]);
                                           fq_table[fq->idx] = fq;
     fq_table[fq->idx] = NULL;

And adding some logs between qman_release_fqid() and
fq_table[fq->idx] = NULL makes the WARN_ON() trigger a lot more.

To prevent that, ensure that fq_table[fq->idx] is set to NULL before
gen_pool_free() is called by using smp_wmb().

Fixes: c535e923bb97 ("soc/fsl: Introduce DPAA 1.x QMan device driver")
Signed-off-by: Richard Genoud <[email protected]>
---
 drivers/soc/fsl/qbman/qman.c | 24 ++++++++++++++++++++++--
 1 file changed, 22 insertions(+), 2 deletions(-)

NB: I'm not 100% sure of the need of a barrier here, since even without
it, the WARN_ON() wasn't triggered any more.

diff --git a/drivers/soc/fsl/qbman/qman.c b/drivers/soc/fsl/qbman/qman.c
index 6b392b3ad4b1..39a3e7aab6ff 100644
--- a/drivers/soc/fsl/qbman/qman.c
+++ b/drivers/soc/fsl/qbman/qman.c
@@ -1827,6 +1827,8 @@ EXPORT_SYMBOL(qman_create_fq);
 
 void qman_destroy_fq(struct qman_fq *fq)
 {
+       int leaked;
+
        /*
         * We don't need to lock the FQ as it is a pre-condition that the FQ be
         * quiesced. Instead, run some checks.
@@ -1834,11 +1836,29 @@ void qman_destroy_fq(struct qman_fq *fq)
        switch (fq->state) {
        case qman_fq_state_parked:
        case qman_fq_state_oos:
-               if (fq_isset(fq, QMAN_FQ_FLAG_DYNAMIC_FQID))
-                       qman_release_fqid(fq->fqid);
+               /*
+                * There's a race condition here on releasing the fqid,
+                * setting the fq_table to NULL, and freeing the fqid.
+                * To prevent it, this order should be respected:
+                */
+               if (fq_isset(fq, QMAN_FQ_FLAG_DYNAMIC_FQID)) {
+                       leaked = qman_shutdown_fq(fq->fqid);
+                       if (leaked)
+                               pr_debug("FQID %d leaked\n", fq->fqid);
+               }
 
                DPAA_ASSERT(fq_table[fq->idx]);
                fq_table[fq->idx] = NULL;
+
+               if (fq_isset(fq, QMAN_FQ_FLAG_DYNAMIC_FQID) && !leaked) {
+                       /*
+                        * fq_table[fq->idx] should be set to null before
+                        * freeing fq->fqid otherwise it could by allocated by
+                        * qman_alloc_fqid() while still being !NULL
+                        */
+                       smp_wmb();
+                       gen_pool_free(qm_fqalloc, fq->fqid | DPAA_GENALLOC_OFF, 
1);
+               }
                return;
        default:
                break;

base-commit: 9448598b22c50c8a5bb77a9103e2d49f134c9578
-- 
2.47.3


Reply via email to