On Sun, 2020-12-20 at 21:20 +0000, Song Bao Hua (Barry Song) wrote:
>
> > -----Original Message-----
> > From: Mike Galbraith [mailto:efa...@gmx.de]
> > Sent: Sunday, December 20, 2020 8:48 PM
> > To: Vitaly Wool <vitaly.w...@konsulko.com>; LKML
> > <linux-kernel@vger.kernel.org>; linux-mm <linux...@kvack.org>
> > Cc: Song Bao Hua (Barry Song) <song.bao....@hisilicon.com>; Sebastian
> > Andrzej
> > Siewior <bige...@linutronix.de>; Minchan Kim <minc...@kernel.org>;
> > NitinGupta
> > <ngu...@vflare.org>
> > Subject: Re: [PATCH] zsmalloc: do not use bit_spin_lock
> >
> > On Sun, 2020-12-20 at 02:23 +0100, Mike Galbraith wrote:
> > > On Sun, 2020-12-20 at 02:22 +0200, Vitaly Wool wrote:
> > > > zsmalloc takes bit spinlock in its _map() callback and releases it
> > > > only in unmap() which is unsafe and leads to zswap complaining
> > > > about scheduling in atomic context.
> > > >
> > > > To fix that and to improve RT properties of zsmalloc, remove that
> > > > bit spinlock completely and use a bit flag instead.
> > >
> > > It also does get_cpu_var() in map(), put_cpu_var() in unmap().
> >
> > That aside, the bit spinlock removal seems to hold up to beating in RT.
> > I stripped out the RT changes to replace the bit spinlocks, applied the
> > still needed atm might_sleep() fix, and ltp zram and zswap test are
> > running in a loop with no signs that it's a bad idea, so I hope that
> > makes it in (minus the preempt disabled spin which I whacked), as it
> > makes zsmalloc markedly more RT friendly.
> >
> > RT changes go from:
> > 1 file changed, 79 insertions(+), 6 deletions(-)
> > to:
> > 1 file changed, 8 insertions(+), 3 deletions(-)
> >
>
> Sorry, would you like to show the change for
> "8 insertions(+), 3 deletions(-)"?
Sure.
---
mm/zsmalloc.c | 11 ++++++++---
1 file changed, 8 insertions(+), 3 deletions(-)
--- a/mm/zsmalloc.c
+++ b/mm/zsmalloc.c
@@ -57,6 +57,7 @@
#include <linux/wait.h>
#include <linux/pagemap.h>
#include <linux/fs.h>
+#include <linux/local_lock.h>
#define ZSPAGE_MAGIC 0x58
@@ -293,6 +294,7 @@ struct zspage {
};
struct mapping_area {
+ local_lock_t lock;
char *vm_buf; /* copy buffer for objects that span pages */
char *vm_addr; /* address of kmap_atomic()'ed pages */
enum zs_mapmode vm_mm; /* mapping mode */
@@ -455,7 +457,9 @@ MODULE_ALIAS("zpool-zsmalloc");
#endif /* CONFIG_ZPOOL */
/* per-cpu VM mapping areas for zspage accesses that cross page
boundaries */
-static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
+static DEFINE_PER_CPU(struct mapping_area, zs_map_area) = {
+ .lock = INIT_LOCAL_LOCK(lock),
+};
static bool is_zspage_isolated(struct zspage *zspage)
{
@@ -1276,7 +1280,8 @@ void *zs_map_object(struct zs_pool *pool
class = pool->size_class[class_idx];
off = (class->size * obj_idx) & ~PAGE_MASK;
- area = &get_cpu_var(zs_map_area);
+ local_lock(&zs_map_area.lock);
+ area = this_cpu_ptr(&zs_map_area);
area->vm_mm = mm;
if (off + class->size <= PAGE_SIZE) {
/* this object is contained entirely within a page */
@@ -1330,7 +1335,7 @@ void zs_unmap_object(struct zs_pool *poo
__zs_unmap_object(area, pages, off, class->size);
}
- put_cpu_var(zs_map_area);
+ local_unlock(&zs_map_area.lock);
migrate_read_unlock(zspage);
unpin_tag(handle);
> BTW, your original patch looks not right as
> crypto_wait_req(crypto_acomp_decompress()...)
> can sleep too.
>
> [copy from your original patch with comment]
> --- a/mm/zswap.c
> +++ b/mm/zswap.c
> @@ -1258,20 +1258,20 @@ static int zswap_frontswap_load(unsigned
>
> /* decompress */
> dlen = PAGE_SIZE;
> + acomp_ctx = raw_cpu_ptr(entry->pool->acomp_ctx);
> + mutex_lock(acomp_ctx->mutex);
> src = zpool_map_handle(entry->pool->zpool, entry->handle, ZPOOL_MM_RO);
> if (zpool_evictable(entry->pool->zpool))
> src += sizeof(struct zswap_header);
>
> - acomp_ctx = raw_cpu_ptr(entry->pool->acomp_ctx);
> - mutex_lock(acomp_ctx->mutex);
> sg_init_one(&input, src, entry->length);
> sg_init_table(&output, 1);
> sg_set_page(&output, page, PAGE_SIZE, 0);
> acomp_request_set_params(acomp_ctx->req, &input, &output,
> entry->length, dlen);
>
> /*!!!!!!!!!!!!!!!!
> * here crypto could sleep
> !!!!!!!!!!!!!!*/
Hohum, another one for my Bitmaster-9000 patch shredder.
-Mike
