by Mike Galbraith

[permalink] [raw]

On 2016-10-19 17:56:30 [+0200], Mike Galbraith wrote:
> In v4.7, the driver switched to percpu compression streams, disabling
> preemption vai get/put_cpu_ptr(). Use a local lock instead for RT.

since you could have multiple streams which would be then serialized
with the local lock I went for this:

From: Mike Galbraith <[email protected]>
Date: Thu, 20 Oct 2016 11:15:22 +0200
Subject: [PATCH] drivers/zram: Don't disable preemption in
zcomp_stream_get/put()

In v4.7, the driver switched to percpu compression streams, disabling
preemption via get/put_cpu_ptr(). Use a per-zcomp_strm lock here. We
also have to fix an lock order issue in zram_decompress_page() such
that zs_map_object() nests inside of zcomp_stream_put() as it does in
zram_bvec_write().

Signed-off-by: Mike Galbraith <[email protected]>
[bigeasy: get_locked_var() -> per zcomp_strm lock]
Signed-off-by: Sebastian Andrzej Siewior <[email protected]>
---
drivers/block/zram/zcomp.c | 12 ++++++++++--
drivers/block/zram/zcomp.h | 1 +
drivers/block/zram/zram_drv.c | 6 +++---
3 files changed, 14 insertions(+), 5 deletions(-)

--- a/drivers/block/zram/zcomp.c
+++ b/drivers/block/zram/zcomp.c
@@ -118,12 +118,19 @@ ssize_t zcomp_available_show(const char

struct zcomp_strm *zcomp_stream_get(struct zcomp *comp)
{
- return *get_cpu_ptr(comp->stream);
+ struct zcomp_strm *zstrm;
+
+ zstrm = *this_cpu_ptr(comp->stream);
+ spin_lock(&zstrm->zcomp_lock);
+ return zstrm;
}

void zcomp_stream_put(struct zcomp *comp)
{
- put_cpu_ptr(comp->stream);
+ struct zcomp_strm *zstrm;
+
+ zstrm = *this_cpu_ptr(comp->stream);
+ spin_unlock(&zstrm->zcomp_lock);
}

int zcomp_compress(struct zcomp_strm *zstrm,
@@ -174,6 +181,7 @@ static int __zcomp_cpu_notifier(struct z
pr_err("Can't allocate a compression stream\n");
return NOTIFY_BAD;
}
+ spin_lock_init(&zstrm->zcomp_lock);
*per_cpu_ptr(comp->stream, cpu) = zstrm;
break;
case CPU_DEAD:
--- a/drivers/block/zram/zcomp.h
+++ b/drivers/block/zram/zcomp.h
@@ -14,6 +14,7 @@ struct zcomp_strm {
/* compression/decompression buffer */
void *buffer;
struct crypto_comp *tfm;
+ spinlock_t zcomp_lock;
};

/* dynamic per-device compression frontend */
--- a/drivers/block/zram/zram_drv.c
+++ b/drivers/block/zram/zram_drv.c
@@ -568,6 +568,7 @@ static int zram_decompress_page(struct z
struct zram_meta *meta = zram->meta;
unsigned long handle;
unsigned int size;
+ struct zcomp_strm *zstrm;

zram_lock_table(&meta->table[index]);
handle = meta->table[index].handle;
@@ -579,16 +580,15 @@ static int zram_decompress_page(struct z
return 0;
}

+ zstrm = zcomp_stream_get(zram->comp);
cmem = zs_map_object(meta->mem_pool, handle, ZS_MM_RO);
if (size == PAGE_SIZE) {
copy_page(mem, cmem);
} else {
- struct zcomp_strm *zstrm = zcomp_stream_get(zram->comp);
-
ret = zcomp_decompress(zstrm, cmem, size, mem);
- zcomp_stream_put(zram->comp);
}
zs_unmap_object(meta->mem_pool, handle);
+ zcomp_stream_put(zram->comp);
zram_unlock_table(&meta->table[index]);

/* Should NEVER happen. Return bio error if it does. */

Sebastian

2016-10-20 11:21:50

by Sebastian Andrzej Siewior

[permalink] [raw]

Subject: Re: [rfc patch] hotplug: Call mmdrop_delayed() in sched_cpu_dying() if PREEMPT_RT_FULL

On 2016-10-20 11:34:03 [+0200], Mike Galbraith wrote:
> My 64 core box just passed an hour running Steven's hotplug stress
> script along with stockfish and futextests (tip-rt.today w. hotplug
> hacks you saw a while back), and seems content to just keep on grinding
> away. Without it, box quickly becomes a doorstop.

This is new. It is v4.7-rc1 new. The hotplug rework moved the notified
around and after

e9cd8fa4fcfd ("sched/migration: Move calc_load_migrate() into CPU_DYING")
f2785ddb5367 ("sched/hotplug: Move migration CPU_DYING to sched_cpu_dying()")

we have no more CPU_DEAD, just CPU_DYING which is invoked with
interrupts off. I don't see anything wrong with pushing this over to
RCU.

Sebastian