On 10/11/23 1:36 PM, Nathan Lynch wrote:
Haren Myneni <ha...@linux.ibm.com> writes:
Haren Myneni <ha...@linux.ibm.com> writes:
The hypervisor returns migration failure if all VAS windows are not
closed. During pre-migration stage, vas_migration_handler() sets
migration_in_progress flag and closes all windows from the list.
The allocate VAS window routine checks the migration flag, setup
the window and then add it to the list. So there is possibility of
the migration handler missing the window that is still in the
process of setup.
t1: Allocate and open VAS t2: Migration event
window
lock vas_pseries_mutex
If migration_in_progress set
unlock vas_pseries_mutex
return
open window HCALL
unlock vas_pseries_mutex
Modify window HCALL lock vas_pseries_mutex
setup window migration_in_progress=true
Closes all windows from
the list
unlock vas_pseries_mutex
lock vas_pseries_mutex return
if nr_closed_windows == 0
// No DLPAR CPU or migration
add to the list
unlock vas_pseries_mutex
return
unlock vas_pseries_mutex
Close VAS window
// due to DLPAR CPU or migration
return -EBUSY
Could the the path t1 takes simply hold the mutex for the duration of
its execution instead of dropping and reacquiring it in the middle?
Here's the relevant code from vas_allocate_window():
mutex_lock(&vas_pseries_mutex);
if (migration_in_progress)
rc = -EBUSY;
else
rc = allocate_setup_window(txwin, (u64 *)&domain[0],
cop_feat_caps->win_type);
mutex_unlock(&vas_pseries_mutex);
if (rc)
goto out;
rc = h_modify_vas_window(txwin);
if (!rc)
rc = get_vas_user_win_ref(&txwin->vas_win.task_ref);
if (rc)
goto out_free;
txwin->win_type = cop_feat_caps->win_type;
mutex_lock(&vas_pseries_mutex);
if (!caps->nr_close_wins) {
list_add(&txwin->win_list, &caps->list);
caps->nr_open_windows++;
mutex_unlock(&vas_pseries_mutex);
vas_user_win_add_mm_context(&txwin->vas_win.task_ref);
return &txwin->vas_win;
}
mutex_unlock(&vas_pseries_mutex);
Is there something about h_modify_vas_window() or get_vas_user_win_ref()
that requires temporarily dropping the lock?
Thanks Nathan for your comments.
vas_pseries_mutex protects window ID and IRQ allocation between alloc
and free window HCALLs, and window list. Generally try to not using
mutex in HCALLs, but we need this mutex with these HCALLs.
We can add h_modify_vas_window() or get_vas_user_win_ref() with in the
mutex context, but not needed.
Hmm. I contend that it would fix your bug in a simpler way that
eliminates the race instead of coping with it by adding more state and
complicating the locking model. With your change, readers of the
migration_in_progress flag check it under the mutex, but the writer
updates it outside of the mutex, which seems strange and unlikely to be
correct.
The migration thread is the only writer which changes
migration_in_progress flag. The setting this flag in moved outside of
mutex in this patch. The window open is only reader of this flag but
within mutex.
Reason for moved the setting outside of mutex:
Suppose many threads are called open window and waiting on mutex and
later the migration thread started. In this case the migration thread
has to wait on mutex for all window open threads has to complete open
window HCALLs in the hypervisor. Then the migration thread has to close
all these windows immediately. So if the setting is done outside of
mutex, the later open window threads can exit from this function quickly
without opening windows.
Reason for keeping the migration_in_progress check inside of mutex
section with the above change (setting outside of mutex):
If the reader threads waits on mutex after checking this flag (before
holding mutex), end up opening windows which will be closed anyway by
the migration thread. Also the later open threads can return with -EBUSY
quickly.
Also free HCALL can take longer depends
on pending NX requests since the hypervisor waits for all requests to be
completed before closing the window.
Applications can issue window open / free calls at the same time which
can experience mutex contention especially on the large system where
100's of credits are available. Another example: The migration event can
wait longer (or timeout) to get this mutex if many threads issue
open/free window calls. Hence added h_modify_vas_window() (modify HCALL)
or get_vas_user_win_ref() outside of mutex.
OK. I believe you're referring to this code, which can run under the lock:
static long hcall_return_busy_check(long rc)
{
/* Check if we are stalled for some time */
if (H_IS_LONG_BUSY(rc)) {
msleep(get_longbusy_msecs(rc));
rc = H_BUSY;
} else if (rc == H_BUSY) {
cond_resched();
}
return rc;
}
...
static int h_deallocate_vas_window(u64 winid)
{
long rc;
do {
rc = plpar_hcall_norets(H_DEALLOCATE_VAS_WINDOW, winid);
rc = hcall_return_busy_check(rc);
} while (rc == H_BUSY);
...
[ with similar loops in the window allocate and modify functions ]
If get_longbusy_msecs() typically returns low values (<20), then you
should prefer usleep_range() over msleep() to avoid over-sleeping. For
example, msleep(1) can schedule away for much longer than 1ms -- often
more like 20ms.
If mutex hold times have been a problem in this code, then it's probably
worth improving the way it handles the busy/retry statuses under the
lock.
Not only mutex hold time longer for contention, also affects the
application performance if we have large mutex section. The application
thread can make several window open and close calls. In the worst case
it can be for each NX request and someone have similar workload on BML
for performance considerations (if not distribute windows on all VAS
engines). In these kind of scenarios when multiple threads are opening
windows, there could be performance bottleneck on mutex if we have large
mutex section. So to have the flexibility, created two sections one with
allocate window (ID and IRQ) and the next one is to protect the window
list.
Thanks
Haren
