From: Mike Galbraith <[email protected]>
The bit spinlock disables preemption. The spinlock_t lock becomes a sleeping
lock on PREEMPT_RT and it can not be acquired in this context. In this locked
section, zs_free() acquires a zs_pool::lock, and there is access to
zram::wb_limit_lock.
Use a spinlock_t on PREEMPT_RT for locking and set/ clear ZRAM_LOCK bit after
the lock has been acquired/ dropped.
Signed-off-by: Mike Galbraith <[email protected]>
Signed-off-by: Sebastian Andrzej Siewior <[email protected]>
Link: https://lkml.kernel.org/r/[email protected]
---
I'm simply forwarding Mike's patch here. The other alternative is to let
the driver depend on !PREEMPT_RT. I can't tell likely it is that this
driver is used. Mike most likely stumbled upon it while running LTP.
drivers/block/zram/zram_drv.c | 36 ++++++++++++++++++++++++++++++++++++
drivers/block/zram/zram_drv.h | 3 +++
2 files changed, 39 insertions(+)
--- a/drivers/block/zram/zram_drv.c
+++ b/drivers/block/zram/zram_drv.c
@@ -57,6 +57,40 @@ static void zram_free_page(struct zram *
static int zram_bvec_read(struct zram *zram, struct bio_vec *bvec,
u32 index, int offset, struct bio *bio);
+#ifdef CONFIG_PREEMPT_RT
+static void zram_meta_init_table_locks(struct zram *zram, size_t num_pages)
+{
+ size_t index;
+
+ for (index = 0; index < num_pages; index++)
+ spin_lock_init(&zram->table[index].lock);
+}
+
+static int zram_slot_trylock(struct zram *zram, u32 index)
+{
+ int ret;
+
+ ret = spin_trylock(&zram->table[index].lock);
+ if (ret)
+ __set_bit(ZRAM_LOCK, &zram->table[index].flags);
+ return ret;
+}
+
+static void zram_slot_lock(struct zram *zram, u32 index)
+{
+ spin_lock(&zram->table[index].lock);
+ __set_bit(ZRAM_LOCK, &zram->table[index].flags);
+}
+
+static void zram_slot_unlock(struct zram *zram, u32 index)
+{
+ __clear_bit(ZRAM_LOCK, &zram->table[index].flags);
+ spin_unlock(&zram->table[index].lock);
+}
+
+#else
+
+static void zram_meta_init_table_locks(struct zram *zram, size_t num_pages) { }
static int zram_slot_trylock(struct zram *zram, u32 index)
{
@@ -72,6 +106,7 @@ static void zram_slot_unlock(struct zram
{
bit_spin_unlock(ZRAM_LOCK, &zram->table[index].flags);
}
+#endif
static inline bool init_done(struct zram *zram)
{
@@ -1311,6 +1346,7 @@ static bool zram_meta_alloc(struct zram
if (!huge_class_size)
huge_class_size = zs_huge_class_size(zram->mem_pool);
+ zram_meta_init_table_locks(zram, num_pages);
return true;
}
--- a/drivers/block/zram/zram_drv.h
+++ b/drivers/block/zram/zram_drv.h
@@ -69,6 +69,9 @@ struct zram_table_entry {
unsigned long element;
};
unsigned long flags;
+#ifdef CONFIG_PREEMPT_RT
+ spinlock_t lock;
+#endif
#ifdef CONFIG_ZRAM_MEMORY_TRACKING
ktime_t ac_time;
#endif
On (23/03/23 17:18), Sebastian Andrzej Siewior wrote:
> From: Mike Galbraith <[email protected]>
>
> The bit spinlock disables preemption. The spinlock_t lock becomes a sleeping
> lock on PREEMPT_RT and it can not be acquired in this context. In this locked
> section, zs_free() acquires a zs_pool::lock, and there is access to
> zram::wb_limit_lock.
>
> Use a spinlock_t on PREEMPT_RT for locking and set/ clear ZRAM_LOCK bit after
> the lock has been acquired/ dropped.
>
> Signed-off-by: Mike Galbraith <[email protected]>
> Signed-off-by: Sebastian Andrzej Siewior <[email protected]>
> Link: https://lkml.kernel.org/r/[email protected]
> ---
>
> I'm simply forwarding Mike's patch here. The other alternative is to let
> the driver depend on !PREEMPT_RT. I can't tell likely it is that this
> driver is used. Mike most likely stumbled upon it while running LTP.
Yeah, I'm curious if anyone uses zram in preempt-rt systems. I don't
mind this patch but would be nice to add new code when it solves some
real problems. Maybe `depend on !PREEMPT_RT` can be a better option.
On Fri, 2023-03-24 at 13:07 +0900, Sergey Senozhatsky wrote:
> On (23/03/23 17:18), Sebastian Andrzej Siewior wrote:
> > Mike most likely stumbled upon it while running LTP.
>
> Yeah, I'm curious if anyone uses zram in preempt-rt systems. I don't
> mind this patch but would be nice to add new code when it solves some
> real problems. Maybe `depend on !PREEMPT_RT` can be a better option.
Patchlet's job here is only obese config RT vs !RT testing. It can
always move back into local_patches, it won't be lonely ;-)
-Mike