mirrors/scst
1
0
Fork 0
mirror of https://github.com/SCST-project/scst.git synced 2026-10-22 14:12:33 +00:00
Code Issues Packages Projects Releases Wiki Activity

scst_user: yield cleanup while shared SGV remains checked out

Browse Source 
With SGV sharing, device B can reuse a buffer whose allocation owner is
a command from device A. If A's user handle closes while B still has the
SGV checked out, A's command remains in its hash. The unjam pass sees
that command, but there is no ready A command, so get_next_cmd() returns
-EAGAIN. Once cleanup_done is set, the old inner loop keeps retrying A
and prevents the single cleanup worker from reaching B.

The post-unjam pool flush added by PR #378 frees SGV objects that have
already returned to the recycling lists. It cannot free the object
still checked out by B, so it cannot resolve this dependency.

Yield A to the outer cleanup loop when no ready command exists but the
unjam pass still reports commands. The worker can process B and retry A
after its normal sleep. Signal A's cleanup completion only when its
release is done and no commands remain.

Link: https://github.com/SCST-project/scst/issues/279
This commit is contained in:
wenlxie
2026-10-02 09:40:21 +03:00
committed by Gleb Chesnokov
parent 5908a8da46
commit 6ddf82829e
1 changed files with 10 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
scst/src/dev_handlers/scst_user.c
+10 -5
View File
@@ -2738,7 +2738,8 @@ repeat:
* The pre-unjam flush above misses these objects. Without this second
* flush, dev_user_free_sg_entries() never fires, the alloc_pages
* ucmd_get() ref is never balanced, and the ucmd stays in ucmd_hash
* indefinitely — causing dev_user_process_cleanup() to loop forever.
* indefinitely. A checked-out object cannot be flushed until its user
* returns it; cleanup must allow that user to make progress.
*/
sgv_pool_flush(dev->pool);
sgv_pool_flush(dev->pool_clust);
@@ -3883,8 +3884,6 @@ static int dev_user_process_cleanup(struct scst_user_dev *dev)
TRACE_DBG("Cleanuping dev %p", dev);
rc1 = dev_user_unjam_dev(dev);
if (rc1 == 0 && rc == -EAGAIN && dev->cleanup_done)
break;
spin_lock_irq(&dev->udev_cmd_threads.cmd_list_lock);
@@ -3895,10 +3894,16 @@ static int dev_user_process_cleanup(struct scst_user_dev *dev)
spin_unlock_irq(&dev->udev_cmd_threads.cmd_list_lock);
if (rc == -EAGAIN) {
if (!dev->cleanup_done) {
TRACE_DBG("No more commands (dev %p)", dev);
if (!dev->cleanup_done || rc1 != 0) {
/*
* An SGV object may still be checked out from a shared pool.
* Let other devices finish cleanup before retrying this one.
*/
TRACE_DBG("Waiting for cleanup (dev %p, pending %d)",
dev, rc1);
goto out;
}
break;
}
}