zram: free meta out of init_lock

The zram->stat handling should be procted by init_lock. Otherwise, user could see stale value from the stat.

Signed-off-by Minchan Kim <[email protected]>

The zram->stat handling should be procted by init_lock. Otherwise, user could see stale value from the stat.

Signed-off-by Minchan Kim <[email protected]>

---

I don't think it's stable material. The race is rare in real practice and this stale stat value read is not a critical.

 drivers/block/zram/zram_drv.c | 37 ++++++++++++++++++++++++++++---------
 1 file changed, 28 insertions(+), 9 deletions(-)

drivers/block/zram/zram_drv.c +0 -0

index 0299d82275e7..53f176f590b0 100644

drivers/block/zram/zram_drv.c +28 -9

 	struct zram *zram = dev_to_zram(d);				\
-	return scnprintf(b, PAGE_SIZE, "%llu\n",			\
-		(u64)atomic64_read(&zram->stats.name));			\
+	u64 val = 0;							\
+									\
+	down_read(&zram->init_lock);					\
+	if (init_done(zram))						\
+		val = atomic64_read(&zram->stats.name);			\
+	up_read(&zram->init_lock);					\
+	return scnprintf(b, PAGE_SIZE, "%llu\n", val);			\
 }									\
 static DEVICE_ATTR_RO(name);
 

@@ -67,8 +72,14 @@ static ssize_t disksize_show(struct device *dev,
 		struct device_attribute *attr, char *buf)
 {
 	struct zram *zram = dev_to_zram(dev);
+	u64 val = 0;
+
+	down_read(&zram->init_lock);
+	if (init_done(zram))
+		val = zram->disksize;
+	up_read(&zram->init_lock);
 
-	return scnprintf(buf, PAGE_SIZE, "%llu\n", zram->disksize);
+	return scnprintf(buf, PAGE_SIZE, "%llu\n", val);
 }
 
 static ssize_t initstate_show(struct device *dev,

@@ -88,9 +99,14 @@ static ssize_t orig_data_size_show(struct device *dev,
 		struct device_attribute *attr, char *buf)
 {
 	struct zram *zram = dev_to_zram(dev);
+	u64 val = 0;
+
+	down_read(&zram->init_lock);
+	if (init_done(zram))
+		val = atomic64_read(&zram->stats.pages_stored) << PAGE_SHIFT;
+	up_read(&zram->init_lock);
 
-	return scnprintf(buf, PAGE_SIZE, "%llu\n",
-		(u64)(atomic64_read(&zram->stats.pages_stored)) << PAGE_SHIFT);
+	return scnprintf(buf, PAGE_SIZE, "%llu\n", val);
 }
 
 static ssize_t mem_used_total_show(struct device *dev,

@@ -957,10 +973,6 @@ static int zram_rw_page(struct block_device *bdev, sector_t sector,
 	struct bio_vec bv;
 
 	zram = bdev->bd_disk->private_data;
-	if (!valid_io_request(zram, sector, PAGE_SIZE)) {
-		atomic64_inc(&zram->stats.invalid_io);
-		return -EINVAL;
-	}
 
 	down_read(&zram->init_lock);
 	if (unlikely(!init_done(zram))) {

@@ -968,6 +980,13 @@ static int zram_rw_page(struct block_device *bdev, sector_t sector,
 		goto out_unlock;
 	}
 
+	if (!valid_io_request(zram, sector, PAGE_SIZE)) {
+		atomic64_inc(&zram->stats.invalid_io);
+		err = -EINVAL;
+		goto out_unlock;
+	}
+
+
 	index = sector >> SECTORS_PER_PAGE_SHIFT;
 	offset = sector & (SECTORS_PER_PAGE - 1) << SECTOR_SHIFT;
 

Signature · 1 line

1.9.1

Quoted 12 lines

The zram->stat handling should be procted by init_lock. Otherwise, user could see stale value from the stat.

Signed-off-by Minchan Kim <[email protected]>

---

I don't think it's stable material. The race is rare in real practice and this stale stat value read is not a critical.

 drivers/block/zram/zram_drv.c | 37 ++++++++++++++++++++++++++++---------
 1 file changed, 28 insertions(+), 9 deletions(-)

drivers/block/zram/zram_drv.c static ssize_t name##_show(struct device *d, \ · line 49

Quoted 7 lines

drivers/block/zram/zram_drv.c +0 -0

index 0299d82275e7..53f176f590b0 100644
@@ -48,8 +48,13 @@ static ssize_t name##_show(struct device *d,               \
                              struct device_attribute *attr, char *b) \
 {                                                                    \

a side note: I wasn't Cc'd in that patchset and found out it only when it's been merged. I'm not sure I understand, why it has been renamed from specific zram_X_show to X_show. what gives?

can't help, catches my eye every time, that rename has broken the original formatting:

drivers/block/zram/zram_drv.c +1 -1

index 9250b3f..c567af5 100644
@@ -44,7 +44,7 @@ static const char *default_compressor = "lzo";
 static unsigned int num_devices = 1;
 
 #define ZRAM_ATTR_RO(name)                                             \
-static ssize_t name##_show(struct device *d,           \
+static ssize_t name##_show(struct device *d,                           \
                                struct device_attribute *attr, char *b) \
 {                                                                      \
        struct zram *zram = dev_to_zram(d);                             \

I don't have any objections. but do we really want to wrap atomic ops in semaphore? it is really such serious race?

        -ss

Quoted 71 lines

drivers/block/zram/zram_drv.c +28 -9

      struct zram *zram = dev_to_zram(d);                             \
-     return scnprintf(b, PAGE_SIZE, "%llu\n",                        \
-             (u64)atomic64_read(&zram->stats.name));                 \
+     u64 val = 0;                                                    \
+                                                                     \
+     down_read(&zram->init_lock);                                    \
+     if (init_done(zram))                                            \
+             val = atomic64_read(&zram->stats.name);                 \
+     up_read(&zram->init_lock);                                      \
+     return scnprintf(b, PAGE_SIZE, "%llu\n", val);                  \
 }                                                                    \
 static DEVICE_ATTR_RO(name);
 

@@ -67,8 +72,14 @@ static ssize_t disksize_show(struct device *dev,
              struct device_attribute *attr, char *buf)
 {
      struct zram *zram = dev_to_zram(dev);
+     u64 val = 0;
+
+     down_read(&zram->init_lock);
+     if (init_done(zram))
+             val = zram->disksize;
+     up_read(&zram->init_lock);
 
-     return scnprintf(buf, PAGE_SIZE, "%llu\n", zram->disksize);
+     return scnprintf(buf, PAGE_SIZE, "%llu\n", val);
 }
 
 static ssize_t initstate_show(struct device *dev,

@@ -88,9 +99,14 @@ static ssize_t orig_data_size_show(struct device *dev,
              struct device_attribute *attr, char *buf)
 {
      struct zram *zram = dev_to_zram(dev);
+     u64 val = 0;
+
+     down_read(&zram->init_lock);
+     if (init_done(zram))
+             val = atomic64_read(&zram->stats.pages_stored) << PAGE_SHIFT;
+     up_read(&zram->init_lock);
 
-     return scnprintf(buf, PAGE_SIZE, "%llu\n",
-             (u64)(atomic64_read(&zram->stats.pages_stored)) << PAGE_SHIFT);
+     return scnprintf(buf, PAGE_SIZE, "%llu\n", val);
 }
 
 static ssize_t mem_used_total_show(struct device *dev,

@@ -957,10 +973,6 @@ static int zram_rw_page(struct block_device *bdev, sector_t sector,
      struct bio_vec bv;
 
      zram = bdev->bd_disk->private_data;
-     if (!valid_io_request(zram, sector, PAGE_SIZE)) {
-             atomic64_inc(&zram->stats.invalid_io);
-             return -EINVAL;
-     }
 
      down_read(&zram->init_lock);
      if (unlikely(!init_done(zram))) {

@@ -968,6 +980,13 @@ static int zram_rw_page(struct block_device *bdev, sector_t sector,
              goto out_unlock;
      }
 
+     if (!valid_io_request(zram, sector, PAGE_SIZE)) {
+             atomic64_inc(&zram->stats.invalid_io);
+             err = -EINVAL;
+             goto out_unlock;
+     }
+
+
      index = sector >> SECTORS_PER_PAGE_SHIFT;
      offset = sector & (SECTORS_PER_PAGE - 1) << SECTOR_SHIFT;
 

Quoted 3 lines

--
1.9.1

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

Hello Sergey

2015-01-23 22:38 GMT+08:00 Sergey Senozhatsky <[email protected]>:

Quoted 13 lines · Minchan Kim

Quoted 12 lines

The zram->stat handling should be procted by init_lock. Otherwise, user could see stale value from the stat.

Signed-off-by Minchan Kim <[email protected]>

---

I don't think it's stable material. The race is rare in real practice and this stale stat value read is not a critical.

 drivers/block/zram/zram_drv.c | 37 ++++++++++++++++++++++++++++---------
 1 file changed, 28 insertions(+), 9 deletions(-)

Quoted 11 lines

drivers/block/zram/zram_drv.c static ssize_t name##_show(struct device *d, \ · line 49

Quoted 7 lines

drivers/block/zram/zram_drv.c +0 -0

index 0299d82275e7..53f176f590b0 100644
@@ -48,8 +48,13 @@ static ssize_t name##_show(struct device *d,               \
                              struct device_attribute *attr, char *b) \
 {                                                                    \

a side note: I wasn't Cc'd in that patchset and found out it only when it's been merged. I'm not sure I understand, why it has been renamed from specific zram_X_show to X_show. what gives?

I changed from zram_attr_##name##_show to name##_show in commit:
fcf1bce zram: use DEVICE_ATTR_[RW|RO|WO] to define zram sys device attribute

I just want to keep the name consistent with others, like disksize_show(), initstate_show().

Thanks.

Quoted 107 lines

can't help, catches my eye every time, that rename has broken the original formatting:

drivers/block/zram/zram_drv.c +1 -1

index 9250b3f..c567af5 100644
@@ -44,7 +44,7 @@ static const char *default_compressor = "lzo";
 static unsigned int num_devices = 1;
 
 #define ZRAM_ATTR_RO(name)                                             \
-static ssize_t name##_show(struct device *d,           \
+static ssize_t name##_show(struct device *d,                           \
                                struct device_attribute *attr, char *b) \
 {                                                                      \
        struct zram *zram = dev_to_zram(d);                             \

I don't have any objections. but do we really want to wrap atomic ops in semaphore? it is really such serious race?

        -ss

Quoted 72 lines

drivers/block/zram/zram_drv.c +28 -10

      struct zram *zram = dev_to_zram(d);                             \
-     return scnprintf(b, PAGE_SIZE, "%llu\n",                        \
-             (u64)atomic64_read(&zram->stats.name));                 \
+     u64 val = 0;                                                    \
+                                                                     \
+     down_read(&zram->init_lock);                                    \
+     if (init_done(zram))                                            \
+             val = atomic64_read(&zram->stats.name);                 \
+     up_read(&zram->init_lock);                                      \
+     return scnprintf(b, PAGE_SIZE, "%llu\n", val);                  \
 }                                                                    \
 static DEVICE_ATTR_RO(name);
 

@@ -67,8 +72,14 @@ static ssize_t disksize_show(struct device *dev,
              struct device_attribute *attr, char *buf)
 {
      struct zram *zram = dev_to_zram(dev);
+     u64 val = 0;
+
+     down_read(&zram->init_lock);
+     if (init_done(zram))
+             val = zram->disksize;
+     up_read(&zram->init_lock);
 
-     return scnprintf(buf, PAGE_SIZE, "%llu\n", zram->disksize);
+     return scnprintf(buf, PAGE_SIZE, "%llu\n", val);
 }
 
 static ssize_t initstate_show(struct device *dev,

@@ -88,9 +99,14 @@ static ssize_t orig_data_size_show(struct device *dev,
              struct device_attribute *attr, char *buf)
 {
      struct zram *zram = dev_to_zram(dev);
+     u64 val = 0;
+
+     down_read(&zram->init_lock);
+     if (init_done(zram))
+             val = atomic64_read(&zram->stats.pages_stored) << PAGE_SHIFT;
+     up_read(&zram->init_lock);
 
-     return scnprintf(buf, PAGE_SIZE, "%llu\n",
-             (u64)(atomic64_read(&zram->stats.pages_stored)) << PAGE_SHIFT);
+     return scnprintf(buf, PAGE_SIZE, "%llu\n", val);
 }
 
 static ssize_t mem_used_total_show(struct device *dev,

@@ -957,10 +973,6 @@ static int zram_rw_page(struct block_device *bdev, sector_t sector,
      struct bio_vec bv;
 
      zram = bdev->bd_disk->private_data;
-     if (!valid_io_request(zram, sector, PAGE_SIZE)) {
-             atomic64_inc(&zram->stats.invalid_io);
-             return -EINVAL;
-     }
 
      down_read(&zram->init_lock);
      if (unlikely(!init_done(zram))) {

@@ -968,6 +980,13 @@ static int zram_rw_page(struct block_device *bdev, sector_t sector,
              goto out_unlock;
      }
 
+     if (!valid_io_request(zram, sector, PAGE_SIZE)) {
+             atomic64_inc(&zram->stats.invalid_io);
+             err = -EINVAL;
+             goto out_unlock;
+     }
+
+
      index = sector >> SECTORS_PER_PAGE_SHIFT;
      offset = sector & (SECTORS_PER_PAGE - 1) << SECTOR_SHIFT;
 
--

Quoted 2 lines

1.9.1

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

We don't need to call zram_meta_free, zcomp_destroy and zs_free under init_lock. What we need to prevent race with init_lock in reset is setting NULL into zram->meta (ie, init_done). This patch does it.

Signed-off-by Minchan Kim <[email protected]>

We don't need to call zram_meta_free, zcomp_destroy and zs_free under init_lock. What we need to prevent race with init_lock in reset is setting NULL into zram->meta (ie, init_done). This patch does it.

Signed-off-by Minchan Kim <[email protected]>

---
 drivers/block/zram/zram_drv.c | 28 ++++++++++++++++------------
 1 file changed, 16 insertions(+), 12 deletions(-)

drivers/block/zram/zram_drv.c +1 -0

index 9250b3f54a8f..0299d82275e7 100644
@@ -708,6 +708,7 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 {
 	size_t index;
 	struct zram_meta *meta;
+	struct zcomp *comp;
 
 	down_write(&zram->init_lock);
 

@@ -719,20 +720,10 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 	}
 
 	meta = zram->meta;

drivers/block/zram/zram_drv.c +15 -2

+	comp = zram->comp;
+	zram->meta = NULL;
 	zram->max_comp_streams = 1;
 
-	zram_meta_free(zram->meta);
-	zram->meta = NULL;
 	/* Reset stats */
 	memset(&zram->stats, 0, sizeof(zram->stats));
 

@@ -742,6 +733,19 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 
 	up_write(&zram->init_lock);
 
+	/* Free all pages that are still in this zram device */
+	for (index = 0; index < zram->disksize >> PAGE_SHIFT; index++) {
+		unsigned long handle = meta->table[index].handle;
+
+		if (!handle)
+			continue;
+
+		zs_free(meta->mem_pool, handle);
+	}
+
+	zcomp_destroy(comp);
+	zram_meta_free(meta);
+
 	/*
 	 * Revalidate disk out of the init_lock to avoid lockdep splat.
 	 * It's okay because disk's capacity is protected by init_lock

Signature · 1 line

1.9.1

Quoted 26 lines · Minchan Kim

We don't need to call zram_meta_free, zcomp_destroy and zs_free under init_lock. What we need to prevent race with init_lock in reset is setting NULL into zram->meta (ie, init_done). This patch does it.

Signed-off-by Minchan Kim <[email protected]>

---
 drivers/block/zram/zram_drv.c | 28 ++++++++++++++++------------
 1 file changed, 16 insertions(+), 12 deletions(-)

drivers/block/zram/zram_drv.c +1 -0

index 9250b3f54a8f..0299d82275e7 100644
@@ -708,6 +708,7 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 {
 	size_t index;
 	struct zram_meta *meta;
+	struct zcomp *comp;
 
 	down_write(&zram->init_lock);
 

@@ -719,20 +720,10 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 	}
 
 	meta = zram->meta;

drivers/block/zram/zram_drv.c static void zram_reset_device(struct zram *zram, bool reset_capacity) · line 731

Quoted 10 lines

drivers/block/zram/zram_drv.c +0 -10

-	/* Free all pages that are still in this zram device */
-	for (index = 0; index < zram->disksize >> PAGE_SHIFT; index++) {
-		unsigned long handle = meta->table[index].handle;
-		if (!handle)
-			continue;
-
-		zs_free(meta->mem_pool, handle);
-	}
-
-	zcomp_destroy(zram->comp);

I'm not so sure about moving zcomp destruction. if we would have detached it from zram, then yes. otherwise, think of zram ->destoy vs ->init race.

suppose,
CPU1 waits for down_write() init lock in disksize_store() with new comp already allocated; CPU0 detaches ->meta and releases write init lock;
CPU1 grabs the lock and does zram->comp = comp;
CPU0 reaches the point of zcomp_destroy(zram->comp);

I'd probably prefer to keep zcomp destruction on its current place. I see a little real value in introducing zcomp detaching and moving destruction out of init_lock.

	-ss

Quoted 32 lines

drivers/block/zram/zram_drv.c +15 -2

+	comp = zram->comp;
+	zram->meta = NULL;
 	zram->max_comp_streams = 1;
 
-	zram_meta_free(zram->meta);
-	zram->meta = NULL;
 	/* Reset stats */
 	memset(&zram->stats, 0, sizeof(zram->stats));
 

@@ -742,6 +733,19 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 
 	up_write(&zram->init_lock);
 
+	/* Free all pages that are still in this zram device */
+	for (index = 0; index < zram->disksize >> PAGE_SHIFT; index++) {
+		unsigned long handle = meta->table[index].handle;
+
+		if (!handle)
+			continue;
+
+		zs_free(meta->mem_pool, handle);
+	}
+
+	zcomp_destroy(comp);
+	zram_meta_free(meta);
+
 	/*
 	 * Revalidate disk out of the init_lock to avoid lockdep splat.
 	 * It's okay because disk's capacity is protected by init_lock

Signature · 1 line

1.9.1

Date: Fri, 23 Jan 2015 15:48:05 +0100
From: Jerome Marchand <[email protected]>
To: Sergey Senozhatsky <[email protected]>, Minchan Kim
<[email protected]>
CC: Andrew Morton <[email protected]>,
[email protected], [email protected], Nitin Gupta <[email protected]>
Subject: Re: [PATCH 1/2] zram: free meta out of init_lock
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:31.0) Gecko/20100101
Thunderbird/31.3.0

Quoted 41 lines · Sergey Senozhatsky

Quoted 26 lines · Minchan Kim

We don't need to call zram_meta_free, zcomp_destroy and zs_free under init_lock. What we need to prevent race with init_lock in reset is setting NULL into zram->meta (ie, init_done). This patch does it.

Signed-off-by Minchan Kim <[email protected]>

---
 drivers/block/zram/zram_drv.c | 28 ++++++++++++++++------------
 1 file changed, 16 insertions(+), 12 deletions(-)

drivers/block/zram/zram_drv.c +1 -0

index 9250b3f54a8f..0299d82275e7 100644
@@ -708,6 +708,7 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 {
 	size_t index;
 	struct zram_meta *meta;
+	struct zcomp *comp;
 
 	down_write(&zram->init_lock);
 

@@ -719,20 +720,10 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 	}
 
 	meta = zram->meta;

drivers/block/zram/zram_drv.c static void zram_reset_device(struct zram *zram, bool reset_capacity) · line 731

Quoted 10 lines

drivers/block/zram/zram_drv.c +0 -10

-	/* Free all pages that are still in this zram device */
-	for (index = 0; index < zram->disksize >> PAGE_SHIFT; index++) {
-		unsigned long handle = meta->table[index].handle;
-		if (!handle)
-			continue;
-
-		zs_free(meta->mem_pool, handle);
-	}
-
-	zcomp_destroy(zram->comp);

I'm not so sure about moving zcomp destruction. if we would have detached it from zram, then yes. otherwise, think of zram ->destoy vs ->init race.

Quoted 5 lines

suppose,
CPU1 waits for down_write() init lock in disksize_store() with new comp already allocated; CPU0 detaches ->meta and releases write init lock;
CPU1 grabs the lock and does zram->comp = comp;
CPU0 reaches the point of zcomp_destroy(zram->comp);

I don't see your point: this patch does not call zcomp_destroy(zram->comp) anymore, but zram_destroy(comp), where comp is the old zram->comp.

Quoted 59 lines · Jerome Marchand

Date: Fri, 23 Jan 2015 15:48:05 +0100
From: Jerome Marchand <[email protected]>
To: Sergey Senozhatsky <[email protected]>, Minchan Kim
<[email protected]>
CC: Andrew Morton <[email protected]>,
[email protected], [email protected], Nitin Gupta <[email protected]>
Subject: Re: [PATCH 1/2] zram: free meta out of init_lock
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:31.0) Gecko/20100101
Thunderbird/31.3.0

Quoted 41 lines · Sergey Senozhatsky

Quoted 26 lines · Minchan Kim

We don't need to call zram_meta_free, zcomp_destroy and zs_free under init_lock. What we need to prevent race with init_lock in reset is setting NULL into zram->meta (ie, init_done). This patch does it.

Signed-off-by Minchan Kim <[email protected]>

---
 drivers/block/zram/zram_drv.c | 28 ++++++++++++++++------------
 1 file changed, 16 insertions(+), 12 deletions(-)

drivers/block/zram/zram_drv.c +1 -0

index 9250b3f54a8f..0299d82275e7 100644
@@ -708,6 +708,7 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 {
 	size_t index;
 	struct zram_meta *meta;
+	struct zcomp *comp;
 
 	down_write(&zram->init_lock);
 

@@ -719,20 +720,10 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 	}
 
 	meta = zram->meta;

drivers/block/zram/zram_drv.c static void zram_reset_device(struct zram *zram, bool reset_capacity) · line 731

Quoted 10 lines

drivers/block/zram/zram_drv.c +0 -10

-	/* Free all pages that are still in this zram device */
-	for (index = 0; index < zram->disksize >> PAGE_SHIFT; index++) {
-		unsigned long handle = meta->table[index].handle;
-		if (!handle)
-			continue;
-
-		zs_free(meta->mem_pool, handle);
-	}
-
-	zcomp_destroy(zram->comp);

I'm not so sure about moving zcomp destruction. if we would have detached it from zram, then yes. otherwise, think of zram ->destoy vs ->init race.

Quoted 5 lines

suppose,
CPU1 waits for down_write() init lock in disksize_store() with new comp already allocated; CPU0 detaches ->meta and releases write init lock;
CPU1 grabs the lock and does zram->comp = comp;
CPU0 reaches the point of zcomp_destroy(zram->comp);

Quoted 3 lines

I don't see your point: this patch does not call zcomp_destroy(zram->comp) anymore, but zram_destroy(comp), where comp is the old zram->comp.

oh... yes. sorry! my bad.

anyway, on a second thought, do we even want to destoy meta out of init_lock?

I mean, it will let you init new device quicker. but... assume, you have 30G zram (or any other bad-enough number). on CPU0 you reset device -- iterate over 30G meta->table, etc. out of init_lock.
on CPU1 you concurrently re-init device and request again 30G.

how bad that can be?

diskstore called on already initialised device is also not so perfect. we first will try to allocate ->meta (vmalloc pages for another 30G), then allocate comp, then down_write() init lock to find out that device is initialised and we need to release allocated memory.

may be we better keep ->meta destruction under init_lock and additionally move ->meta and ->comp allocation under init_lock in disksize_store()?

like the following one:

---

 drivers/block/zram/zram_drv.c | 25 +++++++++++++------------
 1 file changed, 13 insertions(+), 12 deletions(-)

drivers/block/zram/zram_drv.c +11 -9

index 9250b3f..827ab21 100644
@@ -765,9 +765,18 @@ static ssize_t disksize_store(struct device *dev,
 		return -EINVAL;
 
 	disksize = PAGE_ALIGN(disksize);
+	down_write(&zram->init_lock);
+	if (init_done(zram)) {
+		up_write(&zram->init_lock);
+		pr_info("Cannot change disksize for initialized device\n");
+		return -EBUSY;
+	}
+
 	meta = zram_meta_alloc(zram->disk->first_minor, disksize);
-	if (!meta)
-		return -ENOMEM;
+	if (!meta) {
+		err = -ENOMEM;
+		goto out_unlock;
+	}
 
 	comp = zcomp_create(zram->compressor, zram->max_comp_streams);
 	if (IS_ERR(comp)) {

@@ -777,13 +786,6 @@ static ssize_t disksize_store(struct device *dev,
 		goto out_free_meta;
 	}
 
-	down_write(&zram->init_lock);
-	if (init_done(zram)) {
-		pr_info("Cannot change disksize for initialized device\n");
-		err = -EBUSY;
-		goto out_destroy_comp;
-	}
-
 	zram->meta = meta;
 	zram->comp = comp;
 	zram->disksize = disksize;

@@ -799,11 +801,10 @@ static ssize_t disksize_store(struct device *dev,
 
 	return len;
 

Hello,

Quoted 129 lines · Sergey Senozhatsky

Quoted 48 lines · Jerome Marchand

Quoted 41 lines · Sergey Senozhatsky

Quoted 26 lines · Minchan Kim

We don't need to call zram_meta_free, zcomp_destroy and zs_free under init_lock. What we need to prevent race with init_lock in reset is setting NULL into zram->meta (ie, init_done). This patch does it.

Signed-off-by Minchan Kim <[email protected]>

---
 drivers/block/zram/zram_drv.c | 28 ++++++++++++++++------------
 1 file changed, 16 insertions(+), 12 deletions(-)

drivers/block/zram/zram_drv.c +1 -0

index 9250b3f54a8f..0299d82275e7 100644
@@ -708,6 +708,7 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 {
 	size_t index;
 	struct zram_meta *meta;
+	struct zcomp *comp;
 
 	down_write(&zram->init_lock);
 

@@ -719,20 +720,10 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 	}
 
 	meta = zram->meta;

drivers/block/zram/zram_drv.c static void zram_reset_device(struct zram *zram, bool reset_capacity) · line 731

Quoted 10 lines

drivers/block/zram/zram_drv.c +0 -10

-	/* Free all pages that are still in this zram device */
-	for (index = 0; index < zram->disksize >> PAGE_SHIFT; index++) {
-		unsigned long handle = meta->table[index].handle;
-		if (!handle)
-			continue;
-
-		zs_free(meta->mem_pool, handle);
-	}
-
-	zcomp_destroy(zram->comp);

I'm not so sure about moving zcomp destruction. if we would have detached it from zram, then yes. otherwise, think of zram ->destoy vs ->init race.

Quoted 5 lines

suppose,
CPU1 waits for down_write() init lock in disksize_store() with new comp already allocated; CPU0 detaches ->meta and releases write init lock;
CPU1 grabs the lock and does zram->comp = comp;
CPU0 reaches the point of zcomp_destroy(zram->comp);

Quoted 3 lines

I don't see your point: this patch does not call zcomp_destroy(zram->comp) anymore, but zram_destroy(comp), where comp is the old zram->comp.

oh... yes. sorry! my bad.

anyway, on a second thought, do we even want to destoy meta out of init_lock?

I mean, it will let you init new device quicker. but... assume, you have 30G zram (or any other bad-enough number). on CPU0 you reset device -- iterate over 30G meta->table, etc. out of init_lock.
on CPU1 you concurrently re-init device and request again 30G.

how bad that can be?

diskstore called on already initialised device is also not so perfect. we first will try to allocate ->meta (vmalloc pages for another 30G), then allocate comp, then down_write() init lock to find out that device is initialised and we need to release allocated memory.

may be we better keep ->meta destruction under init_lock and additionally move ->meta and ->comp allocation under init_lock in disksize_store()?

like the following one:

---

 drivers/block/zram/zram_drv.c | 25 +++++++++++++------------
 1 file changed, 13 insertions(+), 12 deletions(-)

drivers/block/zram/zram_drv.c +11 -9

index 9250b3f..827ab21 100644
@@ -765,9 +765,18 @@ static ssize_t disksize_store(struct device *dev,
 		return -EINVAL;
 
 	disksize = PAGE_ALIGN(disksize);
+	down_write(&zram->init_lock);
+	if (init_done(zram)) {
+		up_write(&zram->init_lock);
+		pr_info("Cannot change disksize for initialized device\n");
+		return -EBUSY;
+	}
+
 	meta = zram_meta_alloc(zram->disk->first_minor, disksize);
-	if (!meta)
-		return -ENOMEM;
+	if (!meta) {
+		err = -ENOMEM;
+		goto out_unlock;
+	}
 
 	comp = zcomp_create(zram->compressor, zram->max_comp_streams);
 	if (IS_ERR(comp)) {

@@ -777,13 +786,6 @@ static ssize_t disksize_store(struct device *dev,
 		goto out_free_meta;
 	}
 
-	down_write(&zram->init_lock);
-	if (init_done(zram)) {
-		pr_info("Cannot change disksize for initialized device\n");
-		err = -EBUSY;
-		goto out_destroy_comp;
-	}
-
 	zram->meta = meta;
 	zram->comp = comp;
 	zram->disksize = disksize;

@@ -799,11 +801,10 @@ static ssize_t disksize_store(struct device *dev,
 
 	return len;
 

drivers/block/zram/zram_drv.c static ssize_t disksize_store(struct device *dev, · line 810

Quoted 10 lines

drivers/block/zram/zram_drv.c +2 -3

-out_destroy_comp:
-	up_write(&zram->init_lock);
-	zcomp_destroy(comp);
 out_free_meta:
 	zram_meta_free(meta);
+out_unlock:
+	up_write(&zram->init_lock);
 	return err;
 }
 

The init_lock is really troublesome. We can't do call zram_meta_alloc under init_lock due to lockdep report. Please keep in mind.

The zram_rw_page is one of the function under reclaim path and hold it as read_lock while here holds it as write_lock.
It's a false positive so that we might could make shut lockdep up by annotation but I don't want it but want to work with lockdep rather than disable. As well, there are other pathes to use init_lock to protect other data where would be victims of lockdep.

I didn't tell the motivation of this patch because it made you busy guys wasted. Let me tell it now. It was another lockdep report by kmem_cache_destroy for zsmalloc compaction about init_lock. That's why the patchset was one of the patch in compaction.

Yes, the ideal is to remove horrible init_lock of zram in this phase and make code more simple and clear but I don't want to stuck zsmalloc compaction by the work.

Having said that, I feel it's time to revisit
to remove init_lock.
At least, I will think over to find a solution to kill init_lock.

Hello,

Quoted 132 lines · Minchan Kim

Hello,

Quoted 129 lines · Sergey Senozhatsky

Quoted 48 lines · Jerome Marchand

Quoted 41 lines · Sergey Senozhatsky

Quoted 26 lines · Minchan Kim

We don't need to call zram_meta_free, zcomp_destroy and zs_free under init_lock. What we need to prevent race with init_lock in reset is setting NULL into zram->meta (ie, init_done). This patch does it.

Signed-off-by Minchan Kim <[email protected]>

---
 drivers/block/zram/zram_drv.c | 28 ++++++++++++++++------------
 1 file changed, 16 insertions(+), 12 deletions(-)

drivers/block/zram/zram_drv.c +1 -0

index 9250b3f54a8f..0299d82275e7 100644
@@ -708,6 +708,7 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 {
 	size_t index;
 	struct zram_meta *meta;
+	struct zcomp *comp;
 
 	down_write(&zram->init_lock);
 

@@ -719,20 +720,10 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 	}
 
 	meta = zram->meta;

drivers/block/zram/zram_drv.c static void zram_reset_device(struct zram *zram, bool reset_capacity) · line 731

Quoted 10 lines

drivers/block/zram/zram_drv.c +0 -10

-	/* Free all pages that are still in this zram device */
-	for (index = 0; index < zram->disksize >> PAGE_SHIFT; index++) {
-		unsigned long handle = meta->table[index].handle;
-		if (!handle)
-			continue;
-
-		zs_free(meta->mem_pool, handle);
-	}
-
-	zcomp_destroy(zram->comp);

I'm not so sure about moving zcomp destruction. if we would have detached it from zram, then yes. otherwise, think of zram ->destoy vs ->init race.

Quoted 5 lines

suppose,
CPU1 waits for down_write() init lock in disksize_store() with new comp already allocated; CPU0 detaches ->meta and releases write init lock;
CPU1 grabs the lock and does zram->comp = comp;
CPU0 reaches the point of zcomp_destroy(zram->comp);

Quoted 3 lines

I don't see your point: this patch does not call zcomp_destroy(zram->comp) anymore, but zram_destroy(comp), where comp is the old zram->comp.

oh... yes. sorry! my bad.

anyway, on a second thought, do we even want to destoy meta out of init_lock?

I mean, it will let you init new device quicker. but... assume, you have 30G zram (or any other bad-enough number). on CPU0 you reset device -- iterate over 30G meta->table, etc. out of init_lock.
on CPU1 you concurrently re-init device and request again 30G.

how bad that can be?

diskstore called on already initialised device is also not so perfect. we first will try to allocate ->meta (vmalloc pages for another 30G), then allocate comp, then down_write() init lock to find out that device is initialised and we need to release allocated memory.

may be we better keep ->meta destruction under init_lock and additionally move ->meta and ->comp allocation under init_lock in disksize_store()?

like the following one:

---

 drivers/block/zram/zram_drv.c | 25 +++++++++++++------------
 1 file changed, 13 insertions(+), 12 deletions(-)

drivers/block/zram/zram_drv.c +11 -9

index 9250b3f..827ab21 100644
@@ -765,9 +765,18 @@ static ssize_t disksize_store(struct device *dev,
 		return -EINVAL;
 
 	disksize = PAGE_ALIGN(disksize);
+	down_write(&zram->init_lock);
+	if (init_done(zram)) {
+		up_write(&zram->init_lock);
+		pr_info("Cannot change disksize for initialized device\n");
+		return -EBUSY;
+	}
+
 	meta = zram_meta_alloc(zram->disk->first_minor, disksize);
-	if (!meta)
-		return -ENOMEM;
+	if (!meta) {
+		err = -ENOMEM;
+		goto out_unlock;
+	}
 
 	comp = zcomp_create(zram->compressor, zram->max_comp_streams);
 	if (IS_ERR(comp)) {

@@ -777,13 +786,6 @@ static ssize_t disksize_store(struct device *dev,
 		goto out_free_meta;
 	}
 
-	down_write(&zram->init_lock);
-	if (init_done(zram)) {
-		pr_info("Cannot change disksize for initialized device\n");
-		err = -EBUSY;
-		goto out_destroy_comp;
-	}
-
 	zram->meta = meta;
 	zram->comp = comp;
 	zram->disksize = disksize;

@@ -799,11 +801,10 @@ static ssize_t disksize_store(struct device *dev,
 
 	return len;
 

Quoted 14 lines

drivers/block/zram/zram_drv.c static ssize_t disksize_store(struct device *dev, · line 810

Quoted 10 lines

drivers/block/zram/zram_drv.c +2 -3

-out_destroy_comp:
-	up_write(&zram->init_lock);
-	zcomp_destroy(comp);
 out_free_meta:
 	zram_meta_free(meta);
+out_unlock:
+	up_write(&zram->init_lock);
 	return err;
 }
 

The init_lock is really troublesome. We can't do call zram_meta_alloc under init_lock due to lockdep report. Please keep in mind.

ah... I do recall it, thanks for your reminder.

Quoted 15 lines

The zram_rw_page is one of the function under reclaim path and hold it as read_lock while here holds it as write_lock.
It's a false positive so that we might could make shut lockdep up by annotation but I don't want it but want to work with lockdep rather than disable. As well, there are other pathes to use init_lock to protect other data where would be victims of lockdep.

I didn't tell the motivation of this patch because it made you busy guys wasted. Let me tell it now. It was another lockdep report by kmem_cache_destroy for zsmalloc compaction about init_lock. That's why the patchset was one of the patch in compaction.

Yes, the ideal is to remove horrible init_lock of zram in this phase and make code more simple and clear but I don't want to stuck zsmalloc compaction by the work.

Quoted 3 lines

Having said that, I feel it's time to revisit
to remove init_lock.
At least, I will think over to find a solution to kill init_lock.

hm, can't think of anything quick...

	-ss

Quoted 176 lines · Sergey Senozhatsky

Hello,

Quoted 132 lines · Minchan Kim

Hello,

Quoted 129 lines · Sergey Senozhatsky

Quoted 48 lines · Jerome Marchand

Quoted 41 lines · Sergey Senozhatsky

Quoted 26 lines · Minchan Kim

We don't need to call zram_meta_free, zcomp_destroy and zs_free under init_lock. What we need to prevent race with init_lock in reset is setting NULL into zram->meta (ie, init_done). This patch does it.

Signed-off-by Minchan Kim <[email protected]>

---
 drivers/block/zram/zram_drv.c | 28 ++++++++++++++++------------
 1 file changed, 16 insertions(+), 12 deletions(-)

drivers/block/zram/zram_drv.c +1 -0

index 9250b3f54a8f..0299d82275e7 100644
@@ -708,6 +708,7 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 {
 	size_t index;
 	struct zram_meta *meta;
+	struct zcomp *comp;
 
 	down_write(&zram->init_lock);
 

@@ -719,20 +720,10 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 	}
 
 	meta = zram->meta;

drivers/block/zram/zram_drv.c static void zram_reset_device(struct zram *zram, bool reset_capacity) · line 731

Quoted 10 lines

drivers/block/zram/zram_drv.c +0 -10

-	/* Free all pages that are still in this zram device */
-	for (index = 0; index < zram->disksize >> PAGE_SHIFT; index++) {
-		unsigned long handle = meta->table[index].handle;
-		if (!handle)
-			continue;
-
-		zs_free(meta->mem_pool, handle);
-	}
-
-	zcomp_destroy(zram->comp);

I'm not so sure about moving zcomp destruction. if we would have detached it from zram, then yes. otherwise, think of zram ->destoy vs ->init race.

Quoted 5 lines

suppose,
CPU1 waits for down_write() init lock in disksize_store() with new comp already allocated; CPU0 detaches ->meta and releases write init lock;
CPU1 grabs the lock and does zram->comp = comp;
CPU0 reaches the point of zcomp_destroy(zram->comp);

Quoted 3 lines

I don't see your point: this patch does not call zcomp_destroy(zram->comp) anymore, but zram_destroy(comp), where comp is the old zram->comp.

oh... yes. sorry! my bad.

anyway, on a second thought, do we even want to destoy meta out of init_lock?

I mean, it will let you init new device quicker. but... assume, you have 30G zram (or any other bad-enough number). on CPU0 you reset device -- iterate over 30G meta->table, etc. out of init_lock.
on CPU1 you concurrently re-init device and request again 30G.

how bad that can be?

diskstore called on already initialised device is also not so perfect. we first will try to allocate ->meta (vmalloc pages for another 30G), then allocate comp, then down_write() init lock to find out that device is initialised and we need to release allocated memory.

may be we better keep ->meta destruction under init_lock and additionally move ->meta and ->comp allocation under init_lock in disksize_store()?

like the following one:

---

 drivers/block/zram/zram_drv.c | 25 +++++++++++++------------
 1 file changed, 13 insertions(+), 12 deletions(-)

drivers/block/zram/zram_drv.c +11 -9

index 9250b3f..827ab21 100644
@@ -765,9 +765,18 @@ static ssize_t disksize_store(struct device *dev,
 		return -EINVAL;
 
 	disksize = PAGE_ALIGN(disksize);
+	down_write(&zram->init_lock);
+	if (init_done(zram)) {
+		up_write(&zram->init_lock);
+		pr_info("Cannot change disksize for initialized device\n");
+		return -EBUSY;
+	}
+
 	meta = zram_meta_alloc(zram->disk->first_minor, disksize);
-	if (!meta)
-		return -ENOMEM;
+	if (!meta) {
+		err = -ENOMEM;
+		goto out_unlock;
+	}
 
 	comp = zcomp_create(zram->compressor, zram->max_comp_streams);
 	if (IS_ERR(comp)) {

@@ -777,13 +786,6 @@ static ssize_t disksize_store(struct device *dev,
 		goto out_free_meta;
 	}
 
-	down_write(&zram->init_lock);
-	if (init_done(zram)) {
-		pr_info("Cannot change disksize for initialized device\n");
-		err = -EBUSY;
-		goto out_destroy_comp;
-	}
-
 	zram->meta = meta;
 	zram->comp = comp;
 	zram->disksize = disksize;

@@ -799,11 +801,10 @@ static ssize_t disksize_store(struct device *dev,
 
 	return len;
 

Quoted 14 lines

drivers/block/zram/zram_drv.c static ssize_t disksize_store(struct device *dev, · line 810

Quoted 10 lines

drivers/block/zram/zram_drv.c +2 -3

-out_destroy_comp:
-	up_write(&zram->init_lock);
-	zcomp_destroy(comp);
 out_free_meta:
 	zram_meta_free(meta);
+out_unlock:
+	up_write(&zram->init_lock);
 	return err;
 }
 

The init_lock is really troublesome. We can't do call zram_meta_alloc under init_lock due to lockdep report. Please keep in mind.

ah... I do recall it, thanks for your reminder.

Quoted 12 lines

The zram_rw_page is one of the function under reclaim path and hold it as read_lock while here holds it as write_lock.
It's a false positive so that we might could make shut lockdep up by annotation but I don't want it but want to work with lockdep rather than disable. As well, there are other pathes to use init_lock to protect other data where would be victims of lockdep.

I didn't tell the motivation of this patch because it made you busy guys wasted. Let me tell it now. It was another lockdep report by kmem_cache_destroy for zsmalloc compaction about init_lock. That's why the patchset was one of the patch in compaction.

Quoted 8 lines

Yes, the ideal is to remove horrible init_lock of zram in this phase and make code more simple and clear but I don't want to stuck zsmalloc compaction by the work.

Having said that, I feel it's time to revisit
to remove init_lock.
At least, I will think over to find a solution to kill init_lock.

hm, can't think of anything quick...

Quoted 1 line

	-ss

Hello guys,

How about this?

It's based on Ganesh's patch.
https://lkml.org/lkml/2015/1/24/50

From afda9fd2f6c40dd0745d8a6babe78c5cbdceddf5 Mon Sep 17 00:00:00 2001
From: Minchan Kim <[email protected]>
Date: Mon, 26 Jan 2015 14:34:10 +0900
Subject: [RFC] zram: remove init_lock in zram_make_request

Admin could reset zram during I/O operation going on so we have used zram->init_lock as read-side lock in I/O path to prevent sudden zram meta freeing.

However, the init_lock is really troublesome.
We can't do call zram_meta_alloc under init_lock due to lockdep splat because zram_rw_page is one of the function under reclaim path and hold it as read_lock while other places in process context hold it as write_lock. So, we have used allocation out of the lock to avoid lockdep warn but it's not good for readability and fainally, I met another lockdep splat between init_lock and cpu_hotpulug from kmem_cache_destroy during wokring zsmalloc compaction. :(

Yes, the ideal is to remove horrible init_lock of zram in rw path. This patch removes it in rw path and instead, put init_done bool variable to check initialization done with smp_[wmb|rmb] and srcu_[un]read_lock to prevent sudden zram meta freeing during I/O operation.

Signed-off-by Minchan Kim <[email protected]>

---
 drivers/block/zram/zram_drv.c | 76 +++++++++++++++++++++++++++++--------------
 drivers/block/zram/zram_drv.h |  5 +++
 2 files changed, 57 insertions(+), 24 deletions(-)

drivers/block/zram/zram_drv.c +2 -1

index a598ada817f0..e06ff975f997 100644
@@ -32,6 +32,7 @@
 #include <linux/string.h>
 #include <linux/vmalloc.h>
 #include <linux/err.h>
+#include <linux/srcu.h>
 
 #include "zram_drv.h"
 

@@ -53,9 +54,16 @@ static ssize_t name##_show(struct device *d,		\
 }									\
 static DEVICE_ATTR_RO(name);
 
-static inline int init_done(struct zram *zram)
+static inline bool init_done(struct zram *zram)

drivers/block/zram/zram_drv.c +4 -3

 }
 
 static inline struct zram *dev_to_zram(struct device *dev)

@@ -326,6 +334,10 @@ static void zram_meta_free(struct zram_meta *meta)
 	kfree(meta);
 }
 
+static void rcu_zram_do_nothing(struct rcu_head *unused)
+{
+}
+
 static struct zram_meta *zram_meta_alloc(int device_id, u64 disksize)
 {
 	char pool_name[8];

@@ -726,11 +738,8 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 		return;
 	}
 
-	zcomp_destroy(zram->comp);
 	zram->max_comp_streams = 1;
 
-	zram_meta_free(zram->meta);
-	zram->meta = NULL;
 	/* Reset stats */
 	memset(&zram->stats, 0, sizeof(zram->stats));
 

drivers/block/zram/zram_drv.c +21 -9

+	call_srcu(&zram->srcu, &zram->rcu, rcu_zram_do_nothing);
+	synchronize_srcu(&zram->srcu);
+	zram_meta_free(zram->meta);
+	zcomp_destroy(zram->comp);
 	up_write(&zram->init_lock);
-
 	/*
 	 * Revalidate disk out of the init_lock to avoid lockdep splat.
 	 * It's okay because disk's capacity is protected by init_lock

@@ -762,10 +775,19 @@ static ssize_t disksize_store(struct device *dev,
 	if (!disksize)
 		return -EINVAL;
 
+	down_write(&zram->init_lock);
+	if (init_done(zram)) {
+		pr_info("Cannot change disksize for initialized device\n");
+		up_write(&zram->init_lock);
+		return -EBUSY;
+	}
+
 	disksize = PAGE_ALIGN(disksize);
 	meta = zram_meta_alloc(zram->disk->first_minor, disksize);
-	if (!meta)
+	if (!meta) {
+		up_write(&zram->init_lock);
 		return -ENOMEM;
+	}
 
 	comp = zcomp_create(zram->compressor, zram->max_comp_streams);
 	if (IS_ERR(comp)) {

@@ -775,17 +797,17 @@ static ssize_t disksize_store(struct device *dev,
 		goto out_free_meta;
 	}
 
-	down_write(&zram->init_lock);
-	if (init_done(zram)) {
-		pr_info("Cannot change disksize for initialized device\n");
-		err = -EBUSY;
-		goto out_destroy_comp;
-	}
-
 	zram->meta = meta;
 	zram->comp = comp;
 	zram->disksize = disksize;
 	set_capacity(zram->disk, zram->disksize >> SECTOR_SHIFT);
+	/*
+	 * Store operation of struct zram fields should complete
+	 * before init_done set up because zram_bvec_rw doesn't
+	 * hold an zram->init_lock.
+	 */
+	smp_wmb();
+	zram->init_done = true;
 	up_write(&zram->init_lock);
 
 	/*

drivers/block/zram/zram_drv.c +15 -7

 	return err;
 }

@@ -905,9 +925,10 @@ out:
  */
 static void zram_make_request(struct request_queue *queue, struct bio *bio)
 {
+	int idx;
 	struct zram *zram = queue->queuedata;
 
-	down_read(&zram->init_lock);
+	idx = srcu_read_lock(&zram->srcu);
 	if (unlikely(!init_done(zram)))
 		goto error;
 

@@ -918,12 +939,12 @@ static void zram_make_request(struct request_queue *queue, struct bio *bio)
 	}
 
 	__zram_make_request(zram, bio);
-	up_read(&zram->init_lock);
+	srcu_read_unlock(&zram->srcu, idx);
 
 	return;
 
 error:
-	up_read(&zram->init_lock);
+	srcu_read_unlock(&zram->srcu, idx);
 	bio_io_error(bio);
 }
 

@@ -945,18 +966,20 @@ static void zram_slot_free_notify(struct block_device *bdev,
 static int zram_rw_page(struct block_device *bdev, sector_t sector,
 		       struct page *page, int rw)
 {
-	int offset, err;
+	int offset, err, idx;
 	u32 index;
 	struct zram *zram;
 	struct bio_vec bv;
 
 	zram = bdev->bd_disk->private_data;
+	idx = srcu_read_lock(&zram->srcu);
+
 	if (!valid_io_request(zram, sector, PAGE_SIZE)) {
 		atomic64_inc(&zram->stats.invalid_io);
+		srcu_read_unlock(&zram->srcu, idx);
 		return -EINVAL;
 	}
 
-	down_read(&zram->init_lock);
 	if (unlikely(!init_done(zram))) {
 		err = -EIO;
 		goto out_unlock;

@@ -971,7 +994,7 @@ static int zram_rw_page(struct block_device *bdev, sector_t sector,
 
 	err = zram_bvec_rw(zram, &bv, index, offset, rw);
 out_unlock:
-	up_read(&zram->init_lock);
+	srcu_read_unlock(&zram->srcu, idx);
 	/*
 	 * If I/O fails, just return error(ie, non-zero) without
 	 * calling page_endio.

@@ -1041,6 +1064,11 @@ static int create_device(struct zram *zram, int device_id)
 
 	init_rwsem(&zram->init_lock);
 
+	if (init_srcu_struct(&zram->srcu)) {
+		pr_err("Error initialize srcu for device %d\n", device_id);
+		goto out;
+	}
+
 	zram->queue = blk_alloc_queue(GFP_KERNEL);
 	if (!zram->queue) {
 		pr_err("Error allocating disk queue for device %d\n",

@@ -1125,8 +1153,8 @@ static void destroy_device(struct zram *zram)
 
 	del_gendisk(zram->disk);
 	put_disk(zram->disk);
-
 	blk_cleanup_queue(zram->queue);
+	cleanup_srcu_struct(&zram->srcu);
 }
 
 static int __init zram_init(void)

drivers/block/zram/zram_drv.h +5 -0

index e492f6bf11f1..2042c310aea8 100644
@@ -105,8 +105,13 @@ struct zram {
 	struct gendisk *disk;
 	struct zcomp *comp;
 
+	struct srcu_struct srcu;
+	struct rcu_head rcu;
+
 	/* Prevent concurrent execution of device init, reset and R/W request */
 	struct rw_semaphore init_lock;
+	bool init_done;
+
 	/*
 	 * This is the limit on amount of *uncompressed* worth of data
 	 * we can store in a disk.

Signature · 1 line

1.9.1

Quoted 177 lines · Minchan Kim

Quoted 176 lines · Sergey Senozhatsky

Hello,

Quoted 132 lines · Minchan Kim

Hello,

Quoted 129 lines · Sergey Senozhatsky

Quoted 48 lines · Jerome Marchand

Quoted 41 lines · Sergey Senozhatsky

Quoted 26 lines · Minchan Kim

We don't need to call zram_meta_free, zcomp_destroy and zs_free under init_lock. What we need to prevent race with init_lock in reset is setting NULL into zram->meta (ie, init_done). This patch does it.

Signed-off-by Minchan Kim <[email protected]>

---
 drivers/block/zram/zram_drv.c | 28 ++++++++++++++++------------
 1 file changed, 16 insertions(+), 12 deletions(-)

drivers/block/zram/zram_drv.c +1 -0

index 9250b3f54a8f..0299d82275e7 100644
@@ -708,6 +708,7 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 {
 	size_t index;
 	struct zram_meta *meta;
+	struct zcomp *comp;
 
 	down_write(&zram->init_lock);
 

@@ -719,20 +720,10 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 	}
 
 	meta = zram->meta;

drivers/block/zram/zram_drv.c static void zram_reset_device(struct zram *zram, bool reset_capacity) · line 731

Quoted 10 lines

drivers/block/zram/zram_drv.c +0 -10

-	/* Free all pages that are still in this zram device */
-	for (index = 0; index < zram->disksize >> PAGE_SHIFT; index++) {
-		unsigned long handle = meta->table[index].handle;
-		if (!handle)
-			continue;
-
-		zs_free(meta->mem_pool, handle);
-	}
-
-	zcomp_destroy(zram->comp);

I'm not so sure about moving zcomp destruction. if we would have detached it from zram, then yes. otherwise, think of zram ->destoy vs ->init race.

Quoted 5 lines

suppose,
CPU1 waits for down_write() init lock in disksize_store() with new comp already allocated; CPU0 detaches ->meta and releases write init lock;
CPU1 grabs the lock and does zram->comp = comp;
CPU0 reaches the point of zcomp_destroy(zram->comp);

Quoted 3 lines

I don't see your point: this patch does not call zcomp_destroy(zram->comp) anymore, but zram_destroy(comp), where comp is the old zram->comp.

oh... yes. sorry! my bad.

anyway, on a second thought, do we even want to destoy meta out of init_lock?

I mean, it will let you init new device quicker. but... assume, you have 30G zram (or any other bad-enough number). on CPU0 you reset device -- iterate over 30G meta->table, etc. out of init_lock.
on CPU1 you concurrently re-init device and request again 30G.

how bad that can be?

diskstore called on already initialised device is also not so perfect. we first will try to allocate ->meta (vmalloc pages for another 30G), then allocate comp, then down_write() init lock to find out that device is initialised and we need to release allocated memory.

may be we better keep ->meta destruction under init_lock and additionally move ->meta and ->comp allocation under init_lock in disksize_store()?

like the following one:

---

 drivers/block/zram/zram_drv.c | 25 +++++++++++++------------
 1 file changed, 13 insertions(+), 12 deletions(-)

drivers/block/zram/zram_drv.c +11 -9

index 9250b3f..827ab21 100644
@@ -765,9 +765,18 @@ static ssize_t disksize_store(struct device *dev,
 		return -EINVAL;
 
 	disksize = PAGE_ALIGN(disksize);
+	down_write(&zram->init_lock);
+	if (init_done(zram)) {
+		up_write(&zram->init_lock);
+		pr_info("Cannot change disksize for initialized device\n");
+		return -EBUSY;
+	}
+
 	meta = zram_meta_alloc(zram->disk->first_minor, disksize);
-	if (!meta)
-		return -ENOMEM;
+	if (!meta) {
+		err = -ENOMEM;
+		goto out_unlock;
+	}
 
 	comp = zcomp_create(zram->compressor, zram->max_comp_streams);
 	if (IS_ERR(comp)) {

@@ -777,13 +786,6 @@ static ssize_t disksize_store(struct device *dev,
 		goto out_free_meta;
 	}
 
-	down_write(&zram->init_lock);
-	if (init_done(zram)) {
-		pr_info("Cannot change disksize for initialized device\n");
-		err = -EBUSY;
-		goto out_destroy_comp;
-	}
-
 	zram->meta = meta;
 	zram->comp = comp;
 	zram->disksize = disksize;

@@ -799,11 +801,10 @@ static ssize_t disksize_store(struct device *dev,
 
 	return len;
 

Quoted 14 lines

drivers/block/zram/zram_drv.c static ssize_t disksize_store(struct device *dev, · line 810

Quoted 10 lines

drivers/block/zram/zram_drv.c +2 -3

-out_destroy_comp:
-	up_write(&zram->init_lock);
-	zcomp_destroy(comp);
 out_free_meta:
 	zram_meta_free(meta);
+out_unlock:
+	up_write(&zram->init_lock);
 	return err;
 }
 

The init_lock is really troublesome. We can't do call zram_meta_alloc under init_lock due to lockdep report. Please keep in mind.

ah... I do recall it, thanks for your reminder.

Quoted 12 lines

The zram_rw_page is one of the function under reclaim path and hold it as read_lock while here holds it as write_lock.
It's a false positive so that we might could make shut lockdep up by annotation but I don't want it but want to work with lockdep rather than disable. As well, there are other pathes to use init_lock to protect other data where would be victims of lockdep.

I didn't tell the motivation of this patch because it made you busy guys wasted. Let me tell it now. It was another lockdep report by kmem_cache_destroy for zsmalloc compaction about init_lock. That's why the patchset was one of the patch in compaction.

Quoted 8 lines

Yes, the ideal is to remove horrible init_lock of zram in this phase and make code more simple and clear but I don't want to stuck zsmalloc compaction by the work.

Having said that, I feel it's time to revisit
to remove init_lock.
At least, I will think over to find a solution to kill init_lock.

hm, can't think of anything quick...

Quoted 8 lines

Quoted 1 line

	-ss

Hello guys,

How about this?

It's based on Ganesh's patch.
https://lkml.org/lkml/2015/1/24/50

(I see no similarities with Ganesh's patch)

hm, you probably meant this one https://lkml.org/lkml/2015/1/23/406

at glance this makes things a bit more complicated, so I need to think more.

Quoted 48 lines

From afda9fd2f6c40dd0745d8a6babe78c5cbdceddf5 Mon Sep 17 00:00:00 2001
From: Minchan Kim <[email protected]>
Date: Mon, 26 Jan 2015 14:34:10 +0900
Subject: [RFC] zram: remove init_lock in zram_make_request

Admin could reset zram during I/O operation going on so we have used zram->init_lock as read-side lock in I/O path to prevent sudden zram meta freeing.

However, the init_lock is really troublesome.
We can't do call zram_meta_alloc under init_lock due to lockdep splat because zram_rw_page is one of the function under reclaim path and hold it as read_lock while other places in process context hold it as write_lock. So, we have used allocation out of the lock to avoid lockdep warn but it's not good for readability and fainally, I met another lockdep splat between init_lock and cpu_hotpulug from kmem_cache_destroy during wokring zsmalloc compaction. :(

Yes, the ideal is to remove horrible init_lock of zram in rw path. This patch removes it in rw path and instead, put init_done bool variable to check initialization done with smp_[wmb|rmb] and srcu_[un]read_lock to prevent sudden zram meta freeing during I/O operation.

Signed-off-by Minchan Kim <[email protected]>

---
 drivers/block/zram/zram_drv.c | 76 +++++++++++++++++++++++++++++--------------
 drivers/block/zram/zram_drv.h |  5 +++
 2 files changed, 57 insertions(+), 24 deletions(-)

drivers/block/zram/zram_drv.c +2 -1

index a598ada817f0..e06ff975f997 100644
@@ -32,6 +32,7 @@
 #include <linux/string.h>
 #include <linux/vmalloc.h>
 #include <linux/err.h>
+#include <linux/srcu.h>
 
 #include "zram_drv.h"
 

@@ -53,9 +54,16 @@ static ssize_t name##_show(struct device *d,		\
 }									\
 static DEVICE_ATTR_RO(name);
 
-static inline int init_done(struct zram *zram)
+static inline bool init_done(struct zram *zram)

drivers/block/zram/zram_drv.c static ssize_t name##_show(struct device *d, \ · line 66

Quoted 10 lines

drivers/block/zram/zram_drv.c +8 -1

 {
-	return zram->meta != NULL;
+	/*
+	 * init_done can be used without holding zram->init_lock in
+	 * read/write handler(ie, zram_make_request) but we should make sure
+	 * that zram->init_done should set up after meta initialization is
+	 * done. Look at disksize_store.
+	 */
+	smp_rmb();
+	return zram->init_done;

->init_done returns back :)

can we rely on write ->meta; wmb; --- rmb; read ->meta?

how much performance do we lose on barriers?

Quoted 27 lines

drivers/block/zram/zram_drv.c +4 -3

 }
 
 static inline struct zram *dev_to_zram(struct device *dev)

@@ -326,6 +334,10 @@ static void zram_meta_free(struct zram_meta *meta)
 	kfree(meta);
 }
 
+static void rcu_zram_do_nothing(struct rcu_head *unused)
+{
+}
+
 static struct zram_meta *zram_meta_alloc(int device_id, u64 disksize)
 {
 	char pool_name[8];

@@ -726,11 +738,8 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 		return;
 	}
 
-	zcomp_destroy(zram->comp);
 	zram->max_comp_streams = 1;
 
-	zram_meta_free(zram->meta);
-	zram->meta = NULL;
 	/* Reset stats */
 	memset(&zram->stats, 0, sizeof(zram->stats));
 

drivers/block/zram/zram_drv.c static void zram_reset_device(struct zram *zram, bool reset_capacity) · line 750

Quoted 5 lines

drivers/block/zram/zram_drv.c +1 -0

@@ -738,8 +747,12 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 	if (reset_capacity)
 		set_capacity(zram->disk, 0);
 
+	zram->init_done = false;

missing wmb?

I think we also better put comments after every wmb/rmb. like

	smp_wmb(); /* pairs with rmb() in foo() */

Quoted 58 lines

drivers/block/zram/zram_drv.c +21 -9

+	call_srcu(&zram->srcu, &zram->rcu, rcu_zram_do_nothing);
+	synchronize_srcu(&zram->srcu);
+	zram_meta_free(zram->meta);
+	zcomp_destroy(zram->comp);
 	up_write(&zram->init_lock);
-
 	/*
 	 * Revalidate disk out of the init_lock to avoid lockdep splat.
 	 * It's okay because disk's capacity is protected by init_lock

@@ -762,10 +775,19 @@ static ssize_t disksize_store(struct device *dev,
 	if (!disksize)
 		return -EINVAL;
 
+	down_write(&zram->init_lock);
+	if (init_done(zram)) {
+		pr_info("Cannot change disksize for initialized device\n");
+		up_write(&zram->init_lock);
+		return -EBUSY;
+	}
+
 	disksize = PAGE_ALIGN(disksize);
 	meta = zram_meta_alloc(zram->disk->first_minor, disksize);
-	if (!meta)
+	if (!meta) {
+		up_write(&zram->init_lock);
 		return -ENOMEM;
+	}
 
 	comp = zcomp_create(zram->compressor, zram->max_comp_streams);
 	if (IS_ERR(comp)) {

@@ -775,17 +797,17 @@ static ssize_t disksize_store(struct device *dev,
 		goto out_free_meta;
 	}
 
-	down_write(&zram->init_lock);
-	if (init_done(zram)) {
-		pr_info("Cannot change disksize for initialized device\n");
-		err = -EBUSY;
-		goto out_destroy_comp;
-	}
-
 	zram->meta = meta;
 	zram->comp = comp;
 	zram->disksize = disksize;
 	set_capacity(zram->disk, zram->disksize >> SECTOR_SHIFT);
+	/*
+	 * Store operation of struct zram fields should complete
+	 * before init_done set up because zram_bvec_rw doesn't
+	 * hold an zram->init_lock.
+	 */
+	smp_wmb();
+	zram->init_done = true;
 	up_write(&zram->init_lock);
 
 	/*

drivers/block/zram/zram_drv.c static ssize_t disksize_store(struct device *dev, · line 824

Quoted 9 lines

drivers/block/zram/zram_drv.c +1 -3

@@ -797,10 +819,8 @@ static ssize_t disksize_store(struct device *dev,
 
 	return len;
 
-out_destroy_comp:
-	up_write(&zram->init_lock);
-	zcomp_destroy(comp);
 out_free_meta:
+	up_write(&zram->init_lock);
 	zram_meta_free(meta);

zram_meta_free(meta);
up_write(&zram->init_lock);

?

Quoted 105 lines

drivers/block/zram/zram_drv.c +15 -7

 	return err;
 }

@@ -905,9 +925,10 @@ out:
  */
 static void zram_make_request(struct request_queue *queue, struct bio *bio)
 {
+	int idx;
 	struct zram *zram = queue->queuedata;
 
-	down_read(&zram->init_lock);
+	idx = srcu_read_lock(&zram->srcu);
 	if (unlikely(!init_done(zram)))
 		goto error;
 

@@ -918,12 +939,12 @@ static void zram_make_request(struct request_queue *queue, struct bio *bio)
 	}
 
 	__zram_make_request(zram, bio);
-	up_read(&zram->init_lock);
+	srcu_read_unlock(&zram->srcu, idx);
 
 	return;
 
 error:
-	up_read(&zram->init_lock);
+	srcu_read_unlock(&zram->srcu, idx);
 	bio_io_error(bio);
 }
 

@@ -945,18 +966,20 @@ static void zram_slot_free_notify(struct block_device *bdev,
 static int zram_rw_page(struct block_device *bdev, sector_t sector,
 		       struct page *page, int rw)
 {
-	int offset, err;
+	int offset, err, idx;
 	u32 index;
 	struct zram *zram;
 	struct bio_vec bv;
 
 	zram = bdev->bd_disk->private_data;
+	idx = srcu_read_lock(&zram->srcu);
+
 	if (!valid_io_request(zram, sector, PAGE_SIZE)) {
 		atomic64_inc(&zram->stats.invalid_io);
+		srcu_read_unlock(&zram->srcu, idx);
 		return -EINVAL;
 	}
 
-	down_read(&zram->init_lock);
 	if (unlikely(!init_done(zram))) {
 		err = -EIO;
 		goto out_unlock;

@@ -971,7 +994,7 @@ static int zram_rw_page(struct block_device *bdev, sector_t sector,
 
 	err = zram_bvec_rw(zram, &bv, index, offset, rw);
 out_unlock:
-	up_read(&zram->init_lock);
+	srcu_read_unlock(&zram->srcu, idx);
 	/*
 	 * If I/O fails, just return error(ie, non-zero) without
 	 * calling page_endio.

@@ -1041,6 +1064,11 @@ static int create_device(struct zram *zram, int device_id)
 
 	init_rwsem(&zram->init_lock);
 
+	if (init_srcu_struct(&zram->srcu)) {
+		pr_err("Error initialize srcu for device %d\n", device_id);
+		goto out;
+	}
+
 	zram->queue = blk_alloc_queue(GFP_KERNEL);
 	if (!zram->queue) {
 		pr_err("Error allocating disk queue for device %d\n",

@@ -1125,8 +1153,8 @@ static void destroy_device(struct zram *zram)
 
 	del_gendisk(zram->disk);
 	put_disk(zram->disk);
-
 	blk_cleanup_queue(zram->queue);
+	cleanup_srcu_struct(&zram->srcu);
 }
 
 static int __init zram_init(void)

drivers/block/zram/zram_drv.h +5 -0

index e492f6bf11f1..2042c310aea8 100644
@@ -105,8 +105,13 @@ struct zram {
 	struct gendisk *disk;
 	struct zcomp *comp;
 
+	struct srcu_struct srcu;
+	struct rcu_head rcu;
+
 	/* Prevent concurrent execution of device init, reset and R/W request */
 	struct rw_semaphore init_lock;
+	bool init_done;
+
 	/*
 	 * This is the limit on amount of *uncompressed* worth of data
 	 * we can store in a disk.

Signature · 1 line

1.9.1

Hello Sergey,

Quoted 255 lines · Sergey Senozhatsky

Quoted 177 lines · Minchan Kim

Quoted 176 lines · Sergey Senozhatsky

Hello,

Quoted 132 lines · Minchan Kim

Hello,

Quoted 129 lines · Sergey Senozhatsky

Quoted 48 lines · Jerome Marchand

Quoted 41 lines · Sergey Senozhatsky

Quoted 26 lines · Minchan Kim

We don't need to call zram_meta_free, zcomp_destroy and zs_free under init_lock. What we need to prevent race with init_lock in reset is setting NULL into zram->meta (ie, init_done). This patch does it.

Signed-off-by Minchan Kim <[email protected]>

---
 drivers/block/zram/zram_drv.c | 28 ++++++++++++++++------------
 1 file changed, 16 insertions(+), 12 deletions(-)

drivers/block/zram/zram_drv.c +1 -0

index 9250b3f54a8f..0299d82275e7 100644
@@ -708,6 +708,7 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 {
 	size_t index;
 	struct zram_meta *meta;
+	struct zcomp *comp;
 
 	down_write(&zram->init_lock);
 

@@ -719,20 +720,10 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 	}
 
 	meta = zram->meta;

drivers/block/zram/zram_drv.c static void zram_reset_device(struct zram *zram, bool reset_capacity) · line 731

Quoted 10 lines

drivers/block/zram/zram_drv.c +0 -10

-	/* Free all pages that are still in this zram device */
-	for (index = 0; index < zram->disksize >> PAGE_SHIFT; index++) {
-		unsigned long handle = meta->table[index].handle;
-		if (!handle)
-			continue;
-
-		zs_free(meta->mem_pool, handle);
-	}
-
-	zcomp_destroy(zram->comp);

I'm not so sure about moving zcomp destruction. if we would have detached it from zram, then yes. otherwise, think of zram ->destoy vs ->init race.

Quoted 5 lines

suppose,
CPU1 waits for down_write() init lock in disksize_store() with new comp already allocated; CPU0 detaches ->meta and releases write init lock;
CPU1 grabs the lock and does zram->comp = comp;
CPU0 reaches the point of zcomp_destroy(zram->comp);

Quoted 3 lines

I don't see your point: this patch does not call zcomp_destroy(zram->comp) anymore, but zram_destroy(comp), where comp is the old zram->comp.

oh... yes. sorry! my bad.

anyway, on a second thought, do we even want to destoy meta out of init_lock?

I mean, it will let you init new device quicker. but... assume, you have 30G zram (or any other bad-enough number). on CPU0 you reset device -- iterate over 30G meta->table, etc. out of init_lock.
on CPU1 you concurrently re-init device and request again 30G.

how bad that can be?

diskstore called on already initialised device is also not so perfect. we first will try to allocate ->meta (vmalloc pages for another 30G), then allocate comp, then down_write() init lock to find out that device is initialised and we need to release allocated memory.

may be we better keep ->meta destruction under init_lock and additionally move ->meta and ->comp allocation under init_lock in disksize_store()?

like the following one:

---

 drivers/block/zram/zram_drv.c | 25 +++++++++++++------------
 1 file changed, 13 insertions(+), 12 deletions(-)

drivers/block/zram/zram_drv.c +11 -9

index 9250b3f..827ab21 100644
@@ -765,9 +765,18 @@ static ssize_t disksize_store(struct device *dev,
 		return -EINVAL;
 
 	disksize = PAGE_ALIGN(disksize);
+	down_write(&zram->init_lock);
+	if (init_done(zram)) {
+		up_write(&zram->init_lock);
+		pr_info("Cannot change disksize for initialized device\n");
+		return -EBUSY;
+	}
+
 	meta = zram_meta_alloc(zram->disk->first_minor, disksize);
-	if (!meta)
-		return -ENOMEM;
+	if (!meta) {
+		err = -ENOMEM;
+		goto out_unlock;
+	}
 
 	comp = zcomp_create(zram->compressor, zram->max_comp_streams);
 	if (IS_ERR(comp)) {

@@ -777,13 +786,6 @@ static ssize_t disksize_store(struct device *dev,
 		goto out_free_meta;
 	}
 
-	down_write(&zram->init_lock);
-	if (init_done(zram)) {
-		pr_info("Cannot change disksize for initialized device\n");
-		err = -EBUSY;
-		goto out_destroy_comp;
-	}
-
 	zram->meta = meta;
 	zram->comp = comp;
 	zram->disksize = disksize;

@@ -799,11 +801,10 @@ static ssize_t disksize_store(struct device *dev,
 
 	return len;
 

Quoted 14 lines

drivers/block/zram/zram_drv.c static ssize_t disksize_store(struct device *dev, · line 810

Quoted 10 lines

drivers/block/zram/zram_drv.c +2 -3

-out_destroy_comp:
-	up_write(&zram->init_lock);
-	zcomp_destroy(comp);
 out_free_meta:
 	zram_meta_free(meta);
+out_unlock:
+	up_write(&zram->init_lock);
 	return err;
 }
 

The init_lock is really troublesome. We can't do call zram_meta_alloc under init_lock due to lockdep report. Please keep in mind.

ah... I do recall it, thanks for your reminder.

Quoted 12 lines

The zram_rw_page is one of the function under reclaim path and hold it as read_lock while here holds it as write_lock.
It's a false positive so that we might could make shut lockdep up by annotation but I don't want it but want to work with lockdep rather than disable. As well, there are other pathes to use init_lock to protect other data where would be victims of lockdep.

I didn't tell the motivation of this patch because it made you busy guys wasted. Let me tell it now. It was another lockdep report by kmem_cache_destroy for zsmalloc compaction about init_lock. That's why the patchset was one of the patch in compaction.

Quoted 8 lines

Yes, the ideal is to remove horrible init_lock of zram in this phase and make code more simple and clear but I don't want to stuck zsmalloc compaction by the work.

Having said that, I feel it's time to revisit
to remove init_lock.
At least, I will think over to find a solution to kill init_lock.

hm, can't think of anything quick...

Quoted 8 lines

Quoted 1 line

	-ss

Hello guys,

How about this?

It's based on Ganesh's patch.
https://lkml.org/lkml/2015/1/24/50

(I see no similarities with Ganesh's patch)

hm, you probably meant this one https://lkml.org/lkml/2015/1/23/406

at glance this makes things a bit more complicated, so I need to think more.

Quoted 48 lines

From afda9fd2f6c40dd0745d8a6babe78c5cbdceddf5 Mon Sep 17 00:00:00 2001
From: Minchan Kim <[email protected]>
Date: Mon, 26 Jan 2015 14:34:10 +0900
Subject: [RFC] zram: remove init_lock in zram_make_request

Admin could reset zram during I/O operation going on so we have used zram->init_lock as read-side lock in I/O path to prevent sudden zram meta freeing.

However, the init_lock is really troublesome.
We can't do call zram_meta_alloc under init_lock due to lockdep splat because zram_rw_page is one of the function under reclaim path and hold it as read_lock while other places in process context hold it as write_lock. So, we have used allocation out of the lock to avoid lockdep warn but it's not good for readability and fainally, I met another lockdep splat between init_lock and cpu_hotpulug from kmem_cache_destroy during wokring zsmalloc compaction. :(

Yes, the ideal is to remove horrible init_lock of zram in rw path. This patch removes it in rw path and instead, put init_done bool variable to check initialization done with smp_[wmb|rmb] and srcu_[un]read_lock to prevent sudden zram meta freeing during I/O operation.

Signed-off-by Minchan Kim <[email protected]>

---
 drivers/block/zram/zram_drv.c | 76 +++++++++++++++++++++++++++++--------------
 drivers/block/zram/zram_drv.h |  5 +++
 2 files changed, 57 insertions(+), 24 deletions(-)

drivers/block/zram/zram_drv.c +2 -1

index a598ada817f0..e06ff975f997 100644
@@ -32,6 +32,7 @@
 #include <linux/string.h>
 #include <linux/vmalloc.h>
 #include <linux/err.h>
+#include <linux/srcu.h>
 
 #include "zram_drv.h"
 

@@ -53,9 +54,16 @@ static ssize_t name##_show(struct device *d,		\
 }									\
 static DEVICE_ATTR_RO(name);
 
-static inline int init_done(struct zram *zram)
+static inline bool init_done(struct zram *zram)

drivers/block/zram/zram_drv.c static ssize_t name##_show(struct device *d, \ · line 66

Quoted 10 lines

drivers/block/zram/zram_drv.c +8 -1

 {
-	return zram->meta != NULL;
+	/*
+	 * init_done can be used without holding zram->init_lock in
+	 * read/write handler(ie, zram_make_request) but we should make sure
+	 * that zram->init_done should set up after meta initialization is
+	 * done. Look at disksize_store.
+	 */
+	smp_rmb();
+	return zram->init_done;

->init_done returns back :)

Quoted 1 line

can we rely on write ->meta; wmb; --- rmb; read ->meta?

Might be possible.

Quoted 2 lines

how much performance do we lose on barriers?

I think it's not too much than locking which does more than(ie, barrier, fairness, spin on owner and so on) such simple barrier.

Quoted 31 lines

Quoted 31 lines

drivers/block/zram/zram_drv.c +4 -3

 }
 
 static inline struct zram *dev_to_zram(struct device *dev)

@@ -326,6 +334,10 @@ static void zram_meta_free(struct zram_meta *meta)
 	kfree(meta);
 }
 
+static void rcu_zram_do_nothing(struct rcu_head *unused)
+{
+}
+
 static struct zram_meta *zram_meta_alloc(int device_id, u64 disksize)
 {
 	char pool_name[8];

@@ -726,11 +738,8 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 		return;
 	}
 
-	zcomp_destroy(zram->comp);
 	zram->max_comp_streams = 1;
 
-	zram_meta_free(zram->meta);
-	zram->meta = NULL;
 	/* Reset stats */
 	memset(&zram->stats, 0, sizeof(zram->stats));
 

@@ -738,8 +747,12 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 	if (reset_capacity)
 		set_capacity(zram->disk, 0);
 

Quoted 3 lines

drivers/block/zram/zram_drv.c static void zram_reset_device(struct zram *zram, bool reset_capacity) · line 750

Quoted 1 line

drivers/block/zram/zram_drv.c +1 -0

+	zram->init_done = false;

missing wmb?

I thouht about it but when I read comment from call_srcu as follows "each cpu is guaranteed to have executed a full memory barrier", I decided we don't need it. Right? (ie, double check)

Quoted 4 lines

I think we also better put comments after every wmb/rmb. like

	smp_wmb(); /* pairs with rmb() in foo() */

I already put the comment in other smp_rmb/wmb.
If it's not what you want, please suggest me. :)

Quoted 71 lines

Quoted 58 lines

drivers/block/zram/zram_drv.c +21 -9

+	call_srcu(&zram->srcu, &zram->rcu, rcu_zram_do_nothing);
+	synchronize_srcu(&zram->srcu);
+	zram_meta_free(zram->meta);
+	zcomp_destroy(zram->comp);
 	up_write(&zram->init_lock);
-
 	/*
 	 * Revalidate disk out of the init_lock to avoid lockdep splat.
 	 * It's okay because disk's capacity is protected by init_lock

@@ -762,10 +775,19 @@ static ssize_t disksize_store(struct device *dev,
 	if (!disksize)
 		return -EINVAL;
 
+	down_write(&zram->init_lock);
+	if (init_done(zram)) {
+		pr_info("Cannot change disksize for initialized device\n");
+		up_write(&zram->init_lock);
+		return -EBUSY;
+	}
+
 	disksize = PAGE_ALIGN(disksize);
 	meta = zram_meta_alloc(zram->disk->first_minor, disksize);
-	if (!meta)
+	if (!meta) {
+		up_write(&zram->init_lock);
 		return -ENOMEM;
+	}
 
 	comp = zcomp_create(zram->compressor, zram->max_comp_streams);
 	if (IS_ERR(comp)) {

@@ -775,17 +797,17 @@ static ssize_t disksize_store(struct device *dev,
 		goto out_free_meta;
 	}
 
-	down_write(&zram->init_lock);
-	if (init_done(zram)) {
-		pr_info("Cannot change disksize for initialized device\n");
-		err = -EBUSY;
-		goto out_destroy_comp;
-	}
-
 	zram->meta = meta;
 	zram->comp = comp;
 	zram->disksize = disksize;
 	set_capacity(zram->disk, zram->disksize >> SECTOR_SHIFT);
+	/*
+	 * Store operation of struct zram fields should complete
+	 * before init_done set up because zram_bvec_rw doesn't
+	 * hold an zram->init_lock.
+	 */
+	smp_wmb();
+	zram->init_done = true;
 	up_write(&zram->init_lock);
 
 	/*

drivers/block/zram/zram_drv.c static ssize_t disksize_store(struct device *dev, · line 824

Quoted 9 lines

drivers/block/zram/zram_drv.c +1 -3

@@ -797,10 +819,8 @@ static ssize_t disksize_store(struct device *dev,
 
 	return len;
 
-out_destroy_comp:
-	up_write(&zram->init_lock);
-	zcomp_destroy(comp);
 out_free_meta:
+	up_write(&zram->init_lock);
 	zram_meta_free(meta);

zram_meta_free(meta);
up_write(&zram->init_lock);

Quoted 1 line

?

I don't think we should release meta under init_lock. Do you have any reason I am missing?

Quoted 92 lines

Quoted 92 lines

drivers/block/zram/zram_drv.c +15 -7

 	return err;
 }

@@ -905,9 +925,10 @@ out:
  */
 static void zram_make_request(struct request_queue *queue, struct bio *bio)
 {
+	int idx;
 	struct zram *zram = queue->queuedata;
 
-	down_read(&zram->init_lock);
+	idx = srcu_read_lock(&zram->srcu);
 	if (unlikely(!init_done(zram)))
 		goto error;
 

@@ -918,12 +939,12 @@ static void zram_make_request(struct request_queue *queue, struct bio *bio)
 	}
 
 	__zram_make_request(zram, bio);
-	up_read(&zram->init_lock);
+	srcu_read_unlock(&zram->srcu, idx);
 
 	return;
 
 error:
-	up_read(&zram->init_lock);
+	srcu_read_unlock(&zram->srcu, idx);
 	bio_io_error(bio);
 }
 

@@ -945,18 +966,20 @@ static void zram_slot_free_notify(struct block_device *bdev,
 static int zram_rw_page(struct block_device *bdev, sector_t sector,
 		       struct page *page, int rw)
 {
-	int offset, err;
+	int offset, err, idx;
 	u32 index;
 	struct zram *zram;
 	struct bio_vec bv;
 
 	zram = bdev->bd_disk->private_data;
+	idx = srcu_read_lock(&zram->srcu);
+
 	if (!valid_io_request(zram, sector, PAGE_SIZE)) {
 		atomic64_inc(&zram->stats.invalid_io);
+		srcu_read_unlock(&zram->srcu, idx);
 		return -EINVAL;
 	}
 
-	down_read(&zram->init_lock);
 	if (unlikely(!init_done(zram))) {
 		err = -EIO;
 		goto out_unlock;

@@ -971,7 +994,7 @@ static int zram_rw_page(struct block_device *bdev, sector_t sector,
 
 	err = zram_bvec_rw(zram, &bv, index, offset, rw);
 out_unlock:
-	up_read(&zram->init_lock);
+	srcu_read_unlock(&zram->srcu, idx);
 	/*
 	 * If I/O fails, just return error(ie, non-zero) without
 	 * calling page_endio.

@@ -1041,6 +1064,11 @@ static int create_device(struct zram *zram, int device_id)
 
 	init_rwsem(&zram->init_lock);
 
+	if (init_srcu_struct(&zram->srcu)) {
+		pr_err("Error initialize srcu for device %d\n", device_id);
+		goto out;
+	}
+
 	zram->queue = blk_alloc_queue(GFP_KERNEL);
 	if (!zram->queue) {
 		pr_err("Error allocating disk queue for device %d\n",

@@ -1125,8 +1153,8 @@ static void destroy_device(struct zram *zram)
 
 	del_gendisk(zram->disk);
 	put_disk(zram->disk);
-
 	blk_cleanup_queue(zram->queue);
+	cleanup_srcu_struct(&zram->srcu);
 }
 
 static int __init zram_init(void)

drivers/block/zram/zram_drv.h +0 -0

index e492f6bf11f1..2042c310aea8 100644
@@ -105,8 +105,13 @@ struct zram {
 	struct gendisk *disk;
 	struct zcomp *comp;
 

Quoted 13 lines

Quoted 13 lines

drivers/block/zram/zram_drv.h +5 -0

+	struct srcu_struct srcu;
+	struct rcu_head rcu;
+
 	/* Prevent concurrent execution of device init, reset and R/W request */
 	struct rw_semaphore init_lock;
+	bool init_done;
+
 	/*
 	 * This is the limit on amount of *uncompressed* worth of data
 	 * we can store in a disk.

Signature · 1 line

1.9.1

Signature · 2 lines

Kind regards,
Minchan Kim

Hello,

Quoted 261 lines · Minchan Kim

Hello Sergey,

Quoted 255 lines · Sergey Senozhatsky

Quoted 177 lines · Minchan Kim

Quoted 176 lines · Sergey Senozhatsky

Hello,

Quoted 132 lines · Minchan Kim

Hello,

Quoted 129 lines · Sergey Senozhatsky

Quoted 48 lines · Jerome Marchand

Quoted 41 lines · Sergey Senozhatsky

Quoted 26 lines · Minchan Kim

We don't need to call zram_meta_free, zcomp_destroy and zs_free under init_lock. What we need to prevent race with init_lock in reset is setting NULL into zram->meta (ie, init_done). This patch does it.

Signed-off-by Minchan Kim <[email protected]>

---
 drivers/block/zram/zram_drv.c | 28 ++++++++++++++++------------
 1 file changed, 16 insertions(+), 12 deletions(-)

drivers/block/zram/zram_drv.c +1 -0

index 9250b3f54a8f..0299d82275e7 100644
@@ -708,6 +708,7 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 {
 	size_t index;
 	struct zram_meta *meta;
+	struct zcomp *comp;
 
 	down_write(&zram->init_lock);
 

@@ -719,20 +720,10 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 	}
 
 	meta = zram->meta;

drivers/block/zram/zram_drv.c static void zram_reset_device(struct zram *zram, bool reset_capacity) · line 731

Quoted 10 lines

drivers/block/zram/zram_drv.c +0 -10

-	/* Free all pages that are still in this zram device */
-	for (index = 0; index < zram->disksize >> PAGE_SHIFT; index++) {
-		unsigned long handle = meta->table[index].handle;
-		if (!handle)
-			continue;
-
-		zs_free(meta->mem_pool, handle);
-	}
-
-	zcomp_destroy(zram->comp);

I'm not so sure about moving zcomp destruction. if we would have detached it from zram, then yes. otherwise, think of zram ->destoy vs ->init race.

Quoted 5 lines

suppose,
CPU1 waits for down_write() init lock in disksize_store() with new comp already allocated; CPU0 detaches ->meta and releases write init lock;
CPU1 grabs the lock and does zram->comp = comp;
CPU0 reaches the point of zcomp_destroy(zram->comp);

Quoted 3 lines

I don't see your point: this patch does not call zcomp_destroy(zram->comp) anymore, but zram_destroy(comp), where comp is the old zram->comp.

oh... yes. sorry! my bad.

anyway, on a second thought, do we even want to destoy meta out of init_lock?

I mean, it will let you init new device quicker. but... assume, you have 30G zram (or any other bad-enough number). on CPU0 you reset device -- iterate over 30G meta->table, etc. out of init_lock.
on CPU1 you concurrently re-init device and request again 30G.

how bad that can be?

diskstore called on already initialised device is also not so perfect. we first will try to allocate ->meta (vmalloc pages for another 30G), then allocate comp, then down_write() init lock to find out that device is initialised and we need to release allocated memory.

may be we better keep ->meta destruction under init_lock and additionally move ->meta and ->comp allocation under init_lock in disksize_store()?

like the following one:

---

 drivers/block/zram/zram_drv.c | 25 +++++++++++++------------
 1 file changed, 13 insertions(+), 12 deletions(-)

drivers/block/zram/zram_drv.c +11 -9

index 9250b3f..827ab21 100644
@@ -765,9 +765,18 @@ static ssize_t disksize_store(struct device *dev,
 		return -EINVAL;
 
 	disksize = PAGE_ALIGN(disksize);
+	down_write(&zram->init_lock);
+	if (init_done(zram)) {
+		up_write(&zram->init_lock);
+		pr_info("Cannot change disksize for initialized device\n");
+		return -EBUSY;
+	}
+
 	meta = zram_meta_alloc(zram->disk->first_minor, disksize);
-	if (!meta)
-		return -ENOMEM;
+	if (!meta) {
+		err = -ENOMEM;
+		goto out_unlock;
+	}
 
 	comp = zcomp_create(zram->compressor, zram->max_comp_streams);
 	if (IS_ERR(comp)) {

@@ -777,13 +786,6 @@ static ssize_t disksize_store(struct device *dev,
 		goto out_free_meta;
 	}
 
-	down_write(&zram->init_lock);
-	if (init_done(zram)) {
-		pr_info("Cannot change disksize for initialized device\n");
-		err = -EBUSY;
-		goto out_destroy_comp;
-	}
-
 	zram->meta = meta;
 	zram->comp = comp;
 	zram->disksize = disksize;

@@ -799,11 +801,10 @@ static ssize_t disksize_store(struct device *dev,
 
 	return len;
 

Quoted 14 lines

drivers/block/zram/zram_drv.c static ssize_t disksize_store(struct device *dev, · line 810

Quoted 10 lines

drivers/block/zram/zram_drv.c +2 -3

-out_destroy_comp:
-	up_write(&zram->init_lock);
-	zcomp_destroy(comp);
 out_free_meta:
 	zram_meta_free(meta);
+out_unlock:
+	up_write(&zram->init_lock);
 	return err;
 }
 

The init_lock is really troublesome. We can't do call zram_meta_alloc under init_lock due to lockdep report. Please keep in mind.

ah... I do recall it, thanks for your reminder.

Quoted 12 lines

The zram_rw_page is one of the function under reclaim path and hold it as read_lock while here holds it as write_lock.
It's a false positive so that we might could make shut lockdep up by annotation but I don't want it but want to work with lockdep rather than disable. As well, there are other pathes to use init_lock to protect other data where would be victims of lockdep.

I didn't tell the motivation of this patch because it made you busy guys wasted. Let me tell it now. It was another lockdep report by kmem_cache_destroy for zsmalloc compaction about init_lock. That's why the patchset was one of the patch in compaction.

Quoted 8 lines

Yes, the ideal is to remove horrible init_lock of zram in this phase and make code more simple and clear but I don't want to stuck zsmalloc compaction by the work.

Having said that, I feel it's time to revisit
to remove init_lock.
At least, I will think over to find a solution to kill init_lock.

hm, can't think of anything quick...

Quoted 8 lines

Quoted 1 line

	-ss

Hello guys,

How about this?

It's based on Ganesh's patch.
https://lkml.org/lkml/2015/1/24/50

(I see no similarities with Ganesh's patch)

hm, you probably meant this one https://lkml.org/lkml/2015/1/23/406

at glance this makes things a bit more complicated, so I need to think more.

Quoted 48 lines

From afda9fd2f6c40dd0745d8a6babe78c5cbdceddf5 Mon Sep 17 00:00:00 2001
From: Minchan Kim <[email protected]>
Date: Mon, 26 Jan 2015 14:34:10 +0900
Subject: [RFC] zram: remove init_lock in zram_make_request

Admin could reset zram during I/O operation going on so we have used zram->init_lock as read-side lock in I/O path to prevent sudden zram meta freeing.

However, the init_lock is really troublesome.
We can't do call zram_meta_alloc under init_lock due to lockdep splat because zram_rw_page is one of the function under reclaim path and hold it as read_lock while other places in process context hold it as write_lock. So, we have used allocation out of the lock to avoid lockdep warn but it's not good for readability and fainally, I met another lockdep splat between init_lock and cpu_hotpulug from kmem_cache_destroy during wokring zsmalloc compaction. :(

Yes, the ideal is to remove horrible init_lock of zram in rw path. This patch removes it in rw path and instead, put init_done bool variable to check initialization done with smp_[wmb|rmb] and srcu_[un]read_lock to prevent sudden zram meta freeing during I/O operation.

Signed-off-by Minchan Kim <[email protected]>

---
 drivers/block/zram/zram_drv.c | 76 +++++++++++++++++++++++++++++--------------
 drivers/block/zram/zram_drv.h |  5 +++
 2 files changed, 57 insertions(+), 24 deletions(-)

drivers/block/zram/zram_drv.c +2 -1

index a598ada817f0..e06ff975f997 100644
@@ -32,6 +32,7 @@
 #include <linux/string.h>
 #include <linux/vmalloc.h>
 #include <linux/err.h>
+#include <linux/srcu.h>
 
 #include "zram_drv.h"
 

@@ -53,9 +54,16 @@ static ssize_t name##_show(struct device *d,		\
 }									\
 static DEVICE_ATTR_RO(name);
 
-static inline int init_done(struct zram *zram)
+static inline bool init_done(struct zram *zram)

drivers/block/zram/zram_drv.c static ssize_t name##_show(struct device *d, \ · line 66

Quoted 10 lines

drivers/block/zram/zram_drv.c +8 -1

 {
-	return zram->meta != NULL;
+	/*
+	 * init_done can be used without holding zram->init_lock in
+	 * read/write handler(ie, zram_make_request) but we should make sure
+	 * that zram->init_done should set up after meta initialization is
+	 * done. Look at disksize_store.
+	 */
+	smp_rmb();
+	return zram->init_done;

->init_done returns back :)

Quoted 1 line

can we rely on write ->meta; wmb; --- rmb; read ->meta?

Might be possible.

Quoted 42 lines

Quoted 3 lines

how much performance do we lose on barriers?

I think it's not too much than locking which does more than(ie, barrier, fairness, spin on owner and so on) such simple barrier.

Quoted 34 lines

Quoted 28 lines

drivers/block/zram/zram_drv.c +4 -3

 }
 
 static inline struct zram *dev_to_zram(struct device *dev)

@@ -326,6 +334,10 @@ static void zram_meta_free(struct zram_meta *meta)
 	kfree(meta);
 }
 
+static void rcu_zram_do_nothing(struct rcu_head *unused)
+{
+}
+
 static struct zram_meta *zram_meta_alloc(int device_id, u64 disksize)
 {
 	char pool_name[8];

@@ -726,11 +738,8 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 		return;
 	}
 
-	zcomp_destroy(zram->comp);
 	zram->max_comp_streams = 1;
 
-	zram_meta_free(zram->meta);
-	zram->meta = NULL;
 	/* Reset stats */
 	memset(&zram->stats, 0, sizeof(zram->stats));
 

drivers/block/zram/zram_drv.c static void zram_reset_device(struct zram *zram, bool reset_capacity) · line 750

Quoted 4 lines

drivers/block/zram/zram_drv.c +1 -0

@@ -738,8 +747,12 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 	if (reset_capacity)
 		set_capacity(zram->disk, 0);
 
+	zram->init_done = false;

missing wmb?

Quoted 4 lines

I thouht about it but when I read comment from call_srcu as follows "each cpu is guaranteed to have executed a full memory barrier", I decided we don't need it. Right? (ie, double check)

hm, need to think about it.

Quoted 9 lines

Quoted 5 lines

I think we also better put comments after every wmb/rmb. like

	smp_wmb(); /* pairs with rmb() in foo() */

I already put the comment in other smp_rmb/wmb.
If it's not what you want, please suggest me. :)

they are fine. it was a minor nitpick.
I just read in the list that guys want to explicitly show which wmb corresponds to which rmb. but we have only two of them, so it's not a big deal.

Quoted 71 lines

Quoted 71 lines

Quoted 58 lines

drivers/block/zram/zram_drv.c +21 -9

+	call_srcu(&zram->srcu, &zram->rcu, rcu_zram_do_nothing);
+	synchronize_srcu(&zram->srcu);
+	zram_meta_free(zram->meta);
+	zcomp_destroy(zram->comp);
 	up_write(&zram->init_lock);
-
 	/*
 	 * Revalidate disk out of the init_lock to avoid lockdep splat.
 	 * It's okay because disk's capacity is protected by init_lock

@@ -762,10 +775,19 @@ static ssize_t disksize_store(struct device *dev,
 	if (!disksize)
 		return -EINVAL;
 
+	down_write(&zram->init_lock);
+	if (init_done(zram)) {
+		pr_info("Cannot change disksize for initialized device\n");
+		up_write(&zram->init_lock);
+		return -EBUSY;
+	}
+
 	disksize = PAGE_ALIGN(disksize);
 	meta = zram_meta_alloc(zram->disk->first_minor, disksize);
-	if (!meta)
+	if (!meta) {
+		up_write(&zram->init_lock);
 		return -ENOMEM;
+	}
 
 	comp = zcomp_create(zram->compressor, zram->max_comp_streams);
 	if (IS_ERR(comp)) {

@@ -775,17 +797,17 @@ static ssize_t disksize_store(struct device *dev,
 		goto out_free_meta;
 	}
 
-	down_write(&zram->init_lock);
-	if (init_done(zram)) {
-		pr_info("Cannot change disksize for initialized device\n");
-		err = -EBUSY;
-		goto out_destroy_comp;
-	}
-
 	zram->meta = meta;
 	zram->comp = comp;
 	zram->disksize = disksize;
 	set_capacity(zram->disk, zram->disksize >> SECTOR_SHIFT);
+	/*
+	 * Store operation of struct zram fields should complete
+	 * before init_done set up because zram_bvec_rw doesn't
+	 * hold an zram->init_lock.
+	 */
+	smp_wmb();
+	zram->init_done = true;
 	up_write(&zram->init_lock);
 
 	/*

drivers/block/zram/zram_drv.c static ssize_t disksize_store(struct device *dev, · line 824

Quoted 9 lines

drivers/block/zram/zram_drv.c +1 -3

@@ -797,10 +819,8 @@ static ssize_t disksize_store(struct device *dev,
 
 	return len;
 
-out_destroy_comp:
-	up_write(&zram->init_lock);
-	zcomp_destroy(comp);
 out_free_meta:
+	up_write(&zram->init_lock);
 	zram_meta_free(meta);

zram_meta_free(meta);
up_write(&zram->init_lock);

Quoted 5 lines

Quoted 1 line

?

I don't think we should release meta under init_lock. Do you have any reason I am missing?

well, just theoretical.
forbid concurrent initialization until we completely rollback.

CPU0 CPU1

echo 30G > /.../zram0/disksize
meta = vmalloc(pages for 30G)

out_free_meta: echo 30G > /.../zram0/disksize

	up_write(&zram->init_lock);         meta = vmalloc(pages for 30G)
	zram_meta_free(meta);               ^^^^ 30G + 30G
                                            out_free_meta:
                                                   ....
	-ss

Quoted 111 lines

Quoted 106 lines

Quoted 106 lines

drivers/block/zram/zram_drv.c +15 -7

 	return err;
 }

@@ -905,9 +925,10 @@ out:
  */
 static void zram_make_request(struct request_queue *queue, struct bio *bio)
 {
+	int idx;
 	struct zram *zram = queue->queuedata;
 
-	down_read(&zram->init_lock);
+	idx = srcu_read_lock(&zram->srcu);
 	if (unlikely(!init_done(zram)))
 		goto error;
 

@@ -918,12 +939,12 @@ static void zram_make_request(struct request_queue *queue, struct bio *bio)
 	}
 
 	__zram_make_request(zram, bio);
-	up_read(&zram->init_lock);
+	srcu_read_unlock(&zram->srcu, idx);
 
 	return;
 
 error:
-	up_read(&zram->init_lock);
+	srcu_read_unlock(&zram->srcu, idx);
 	bio_io_error(bio);
 }
 

@@ -945,18 +966,20 @@ static void zram_slot_free_notify(struct block_device *bdev,
 static int zram_rw_page(struct block_device *bdev, sector_t sector,
 		       struct page *page, int rw)
 {
-	int offset, err;
+	int offset, err, idx;
 	u32 index;
 	struct zram *zram;
 	struct bio_vec bv;
 
 	zram = bdev->bd_disk->private_data;
+	idx = srcu_read_lock(&zram->srcu);
+
 	if (!valid_io_request(zram, sector, PAGE_SIZE)) {
 		atomic64_inc(&zram->stats.invalid_io);
+		srcu_read_unlock(&zram->srcu, idx);
 		return -EINVAL;
 	}
 
-	down_read(&zram->init_lock);
 	if (unlikely(!init_done(zram))) {
 		err = -EIO;
 		goto out_unlock;

@@ -971,7 +994,7 @@ static int zram_rw_page(struct block_device *bdev, sector_t sector,
 
 	err = zram_bvec_rw(zram, &bv, index, offset, rw);
 out_unlock:
-	up_read(&zram->init_lock);
+	srcu_read_unlock(&zram->srcu, idx);
 	/*
 	 * If I/O fails, just return error(ie, non-zero) without
 	 * calling page_endio.

@@ -1041,6 +1064,11 @@ static int create_device(struct zram *zram, int device_id)
 
 	init_rwsem(&zram->init_lock);
 
+	if (init_srcu_struct(&zram->srcu)) {
+		pr_err("Error initialize srcu for device %d\n", device_id);
+		goto out;
+	}
+
 	zram->queue = blk_alloc_queue(GFP_KERNEL);
 	if (!zram->queue) {
 		pr_err("Error allocating disk queue for device %d\n",

@@ -1125,8 +1153,8 @@ static void destroy_device(struct zram *zram)
 
 	del_gendisk(zram->disk);
 	put_disk(zram->disk);
-
 	blk_cleanup_queue(zram->queue);
+	cleanup_srcu_struct(&zram->srcu);
 }
 
 static int __init zram_init(void)

drivers/block/zram/zram_drv.h +5 -0

index e492f6bf11f1..2042c310aea8 100644
@@ -105,8 +105,13 @@ struct zram {
 	struct gendisk *disk;
 	struct zcomp *comp;
 
+	struct srcu_struct srcu;
+	struct rcu_head rcu;
+
 	/* Prevent concurrent execution of device init, reset and R/W request */
 	struct rw_semaphore init_lock;
+	bool init_done;
+
 	/*
 	 * This is the limit on amount of *uncompressed* worth of data
 	 * we can store in a disk.

Signature · 1 line

1.9.1

Signature · 2 lines

Kind regards,
Minchan Kim

Hello Sergey,

Quoted 264 lines · Sergey Senozhatsky

Hello,

Quoted 261 lines · Minchan Kim

Hello Sergey,

Quoted 255 lines · Sergey Senozhatsky

Quoted 177 lines · Minchan Kim

Quoted 176 lines · Sergey Senozhatsky

Hello,

Quoted 132 lines · Minchan Kim

Hello,

Quoted 129 lines · Sergey Senozhatsky

Quoted 48 lines · Jerome Marchand

Quoted 41 lines · Sergey Senozhatsky

Quoted 26 lines · Minchan Kim

We don't need to call zram_meta_free, zcomp_destroy and zs_free under init_lock. What we need to prevent race with init_lock in reset is setting NULL into zram->meta (ie, init_done). This patch does it.

Signed-off-by Minchan Kim <[email protected]>

---
 drivers/block/zram/zram_drv.c | 28 ++++++++++++++++------------
 1 file changed, 16 insertions(+), 12 deletions(-)

drivers/block/zram/zram_drv.c +1 -0

index 9250b3f54a8f..0299d82275e7 100644
@@ -708,6 +708,7 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 {
 	size_t index;
 	struct zram_meta *meta;
+	struct zcomp *comp;
 
 	down_write(&zram->init_lock);
 

@@ -719,20 +720,10 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 	}
 
 	meta = zram->meta;

drivers/block/zram/zram_drv.c static void zram_reset_device(struct zram *zram, bool reset_capacity) · line 731

Quoted 10 lines

drivers/block/zram/zram_drv.c +0 -10

-	/* Free all pages that are still in this zram device */
-	for (index = 0; index < zram->disksize >> PAGE_SHIFT; index++) {
-		unsigned long handle = meta->table[index].handle;
-		if (!handle)
-			continue;
-
-		zs_free(meta->mem_pool, handle);
-	}
-
-	zcomp_destroy(zram->comp);

I'm not so sure about moving zcomp destruction. if we would have detached it from zram, then yes. otherwise, think of zram ->destoy vs ->init race.

Quoted 5 lines

suppose,
CPU1 waits for down_write() init lock in disksize_store() with new comp already allocated; CPU0 detaches ->meta and releases write init lock;
CPU1 grabs the lock and does zram->comp = comp;
CPU0 reaches the point of zcomp_destroy(zram->comp);

Quoted 3 lines

I don't see your point: this patch does not call zcomp_destroy(zram->comp) anymore, but zram_destroy(comp), where comp is the old zram->comp.

oh... yes. sorry! my bad.

anyway, on a second thought, do we even want to destoy meta out of init_lock?

I mean, it will let you init new device quicker. but... assume, you have 30G zram (or any other bad-enough number). on CPU0 you reset device -- iterate over 30G meta->table, etc. out of init_lock.
on CPU1 you concurrently re-init device and request again 30G.

how bad that can be?

diskstore called on already initialised device is also not so perfect. we first will try to allocate ->meta (vmalloc pages for another 30G), then allocate comp, then down_write() init lock to find out that device is initialised and we need to release allocated memory.

may be we better keep ->meta destruction under init_lock and additionally move ->meta and ->comp allocation under init_lock in disksize_store()?

like the following one:

---

 drivers/block/zram/zram_drv.c | 25 +++++++++++++------------
 1 file changed, 13 insertions(+), 12 deletions(-)

drivers/block/zram/zram_drv.c +11 -9

index 9250b3f..827ab21 100644
@@ -765,9 +765,18 @@ static ssize_t disksize_store(struct device *dev,
 		return -EINVAL;
 
 	disksize = PAGE_ALIGN(disksize);
+	down_write(&zram->init_lock);
+	if (init_done(zram)) {
+		up_write(&zram->init_lock);
+		pr_info("Cannot change disksize for initialized device\n");
+		return -EBUSY;
+	}
+
 	meta = zram_meta_alloc(zram->disk->first_minor, disksize);
-	if (!meta)
-		return -ENOMEM;
+	if (!meta) {
+		err = -ENOMEM;
+		goto out_unlock;
+	}
 
 	comp = zcomp_create(zram->compressor, zram->max_comp_streams);
 	if (IS_ERR(comp)) {

@@ -777,13 +786,6 @@ static ssize_t disksize_store(struct device *dev,
 		goto out_free_meta;
 	}
 
-	down_write(&zram->init_lock);
-	if (init_done(zram)) {
-		pr_info("Cannot change disksize for initialized device\n");
-		err = -EBUSY;
-		goto out_destroy_comp;
-	}
-
 	zram->meta = meta;
 	zram->comp = comp;
 	zram->disksize = disksize;

@@ -799,11 +801,10 @@ static ssize_t disksize_store(struct device *dev,
 
 	return len;
 

Quoted 14 lines

drivers/block/zram/zram_drv.c static ssize_t disksize_store(struct device *dev, · line 810

Quoted 10 lines

drivers/block/zram/zram_drv.c +2 -3

-out_destroy_comp:
-	up_write(&zram->init_lock);
-	zcomp_destroy(comp);
 out_free_meta:
 	zram_meta_free(meta);
+out_unlock:
+	up_write(&zram->init_lock);
 	return err;
 }
 

The init_lock is really troublesome. We can't do call zram_meta_alloc under init_lock due to lockdep report. Please keep in mind.

ah... I do recall it, thanks for your reminder.

Quoted 12 lines

The zram_rw_page is one of the function under reclaim path and hold it as read_lock while here holds it as write_lock.
It's a false positive so that we might could make shut lockdep up by annotation but I don't want it but want to work with lockdep rather than disable. As well, there are other pathes to use init_lock to protect other data where would be victims of lockdep.

I didn't tell the motivation of this patch because it made you busy guys wasted. Let me tell it now. It was another lockdep report by kmem_cache_destroy for zsmalloc compaction about init_lock. That's why the patchset was one of the patch in compaction.

Quoted 8 lines

Yes, the ideal is to remove horrible init_lock of zram in this phase and make code more simple and clear but I don't want to stuck zsmalloc compaction by the work.

Having said that, I feel it's time to revisit
to remove init_lock.
At least, I will think over to find a solution to kill init_lock.

hm, can't think of anything quick...

Quoted 8 lines

Quoted 1 line

	-ss

Hello guys,

How about this?

It's based on Ganesh's patch.
https://lkml.org/lkml/2015/1/24/50

(I see no similarities with Ganesh's patch)

hm, you probably meant this one https://lkml.org/lkml/2015/1/23/406

at glance this makes things a bit more complicated, so I need to think more.

Quoted 48 lines

From afda9fd2f6c40dd0745d8a6babe78c5cbdceddf5 Mon Sep 17 00:00:00 2001
From: Minchan Kim <[email protected]>
Date: Mon, 26 Jan 2015 14:34:10 +0900
Subject: [RFC] zram: remove init_lock in zram_make_request

Admin could reset zram during I/O operation going on so we have used zram->init_lock as read-side lock in I/O path to prevent sudden zram meta freeing.

However, the init_lock is really troublesome.
We can't do call zram_meta_alloc under init_lock due to lockdep splat because zram_rw_page is one of the function under reclaim path and hold it as read_lock while other places in process context hold it as write_lock. So, we have used allocation out of the lock to avoid lockdep warn but it's not good for readability and fainally, I met another lockdep splat between init_lock and cpu_hotpulug from kmem_cache_destroy during wokring zsmalloc compaction. :(

Yes, the ideal is to remove horrible init_lock of zram in rw path. This patch removes it in rw path and instead, put init_done bool variable to check initialization done with smp_[wmb|rmb] and srcu_[un]read_lock to prevent sudden zram meta freeing during I/O operation.

Signed-off-by Minchan Kim <[email protected]>

---
 drivers/block/zram/zram_drv.c | 76 +++++++++++++++++++++++++++++--------------
 drivers/block/zram/zram_drv.h |  5 +++
 2 files changed, 57 insertions(+), 24 deletions(-)

drivers/block/zram/zram_drv.c +2 -1

index a598ada817f0..e06ff975f997 100644
@@ -32,6 +32,7 @@
 #include <linux/string.h>
 #include <linux/vmalloc.h>
 #include <linux/err.h>
+#include <linux/srcu.h>
 
 #include "zram_drv.h"
 

@@ -53,9 +54,16 @@ static ssize_t name##_show(struct device *d,		\
 }									\
 static DEVICE_ATTR_RO(name);
 
-static inline int init_done(struct zram *zram)
+static inline bool init_done(struct zram *zram)

drivers/block/zram/zram_drv.c static ssize_t name##_show(struct device *d, \ · line 66

Quoted 10 lines

drivers/block/zram/zram_drv.c +8 -1

 {
-	return zram->meta != NULL;
+	/*
+	 * init_done can be used without holding zram->init_lock in
+	 * read/write handler(ie, zram_make_request) but we should make sure
+	 * that zram->init_done should set up after meta initialization is
+	 * done. Look at disksize_store.
+	 */
+	smp_rmb();
+	return zram->init_done;

->init_done returns back :)

Quoted 1 line

can we rely on write ->meta; wmb; --- rmb; read ->meta?

Might be possible.

Now that I think about it, it's impossible with zram->meta because we need to nullify it before call_srcu but pre-existing SRCU read-side critical sections can access zram->meta.
Anyway, introducing a new variable should be not a party-pooper.

Quoted 42 lines

Quoted 42 lines

Quoted 3 lines

how much performance do we lose on barriers?

I think it's not too much than locking which does more than(ie, barrier, fairness, spin on owner and so on) such simple barrier.

Quoted 34 lines

Quoted 28 lines

drivers/block/zram/zram_drv.c +4 -3

 }
 
 static inline struct zram *dev_to_zram(struct device *dev)

@@ -326,6 +334,10 @@ static void zram_meta_free(struct zram_meta *meta)
 	kfree(meta);
 }
 
+static void rcu_zram_do_nothing(struct rcu_head *unused)
+{
+}
+
 static struct zram_meta *zram_meta_alloc(int device_id, u64 disksize)
 {
 	char pool_name[8];

@@ -726,11 +738,8 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 		return;
 	}
 
-	zcomp_destroy(zram->comp);
 	zram->max_comp_streams = 1;
 
-	zram_meta_free(zram->meta);
-	zram->meta = NULL;
 	/* Reset stats */
 	memset(&zram->stats, 0, sizeof(zram->stats));
 

drivers/block/zram/zram_drv.c static void zram_reset_device(struct zram *zram, bool reset_capacity) · line 750

Quoted 4 lines

drivers/block/zram/zram_drv.c +1 -0

@@ -738,8 +747,12 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 	if (reset_capacity)
 		set_capacity(zram->disk, 0);
 
+	zram->init_done = false;

missing wmb?

Quoted 6 lines

Quoted 4 lines

I thouht about it but when I read comment from call_srcu as follows "each cpu is guaranteed to have executed a full memory barrier", I decided we don't need it. Right? (ie, double check)

hm, need to think about it.

Another idea is to use kick_all_cpus_sync, not srcu. With that, we don't need to add more instruction in rw path. I will try it.

Quoted 320 lines · Minchan Kim

Hello Sergey,

Quoted 264 lines · Sergey Senozhatsky

Hello,

Quoted 261 lines · Minchan Kim

Hello Sergey,

Quoted 255 lines · Sergey Senozhatsky

Quoted 177 lines · Minchan Kim

Quoted 176 lines · Sergey Senozhatsky

Hello,

Quoted 132 lines · Minchan Kim

Hello,

Quoted 129 lines · Sergey Senozhatsky

Quoted 48 lines · Jerome Marchand

Quoted 41 lines · Sergey Senozhatsky

Quoted 26 lines · Minchan Kim

We don't need to call zram_meta_free, zcomp_destroy and zs_free under init_lock. What we need to prevent race with init_lock in reset is setting NULL into zram->meta (ie, init_done). This patch does it.

Signed-off-by Minchan Kim <[email protected]>

---
 drivers/block/zram/zram_drv.c | 28 ++++++++++++++++------------
 1 file changed, 16 insertions(+), 12 deletions(-)

drivers/block/zram/zram_drv.c +1 -0

index 9250b3f54a8f..0299d82275e7 100644
@@ -708,6 +708,7 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 {
 	size_t index;
 	struct zram_meta *meta;
+	struct zcomp *comp;
 
 	down_write(&zram->init_lock);
 

@@ -719,20 +720,10 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 	}
 
 	meta = zram->meta;

drivers/block/zram/zram_drv.c static void zram_reset_device(struct zram *zram, bool reset_capacity) · line 731

Quoted 10 lines

drivers/block/zram/zram_drv.c +0 -10

-	/* Free all pages that are still in this zram device */
-	for (index = 0; index < zram->disksize >> PAGE_SHIFT; index++) {
-		unsigned long handle = meta->table[index].handle;
-		if (!handle)
-			continue;
-
-		zs_free(meta->mem_pool, handle);
-	}
-
-	zcomp_destroy(zram->comp);

I'm not so sure about moving zcomp destruction. if we would have detached it from zram, then yes. otherwise, think of zram ->destoy vs ->init race.

Quoted 5 lines

suppose,
CPU1 waits for down_write() init lock in disksize_store() with new comp already allocated; CPU0 detaches ->meta and releases write init lock;
CPU1 grabs the lock and does zram->comp = comp;
CPU0 reaches the point of zcomp_destroy(zram->comp);

Quoted 3 lines

I don't see your point: this patch does not call zcomp_destroy(zram->comp) anymore, but zram_destroy(comp), where comp is the old zram->comp.

oh... yes. sorry! my bad.

anyway, on a second thought, do we even want to destoy meta out of init_lock?

I mean, it will let you init new device quicker. but... assume, you have 30G zram (or any other bad-enough number). on CPU0 you reset device -- iterate over 30G meta->table, etc. out of init_lock.
on CPU1 you concurrently re-init device and request again 30G.

how bad that can be?

diskstore called on already initialised device is also not so perfect. we first will try to allocate ->meta (vmalloc pages for another 30G), then allocate comp, then down_write() init lock to find out that device is initialised and we need to release allocated memory.

may be we better keep ->meta destruction under init_lock and additionally move ->meta and ->comp allocation under init_lock in disksize_store()?

like the following one:

---

 drivers/block/zram/zram_drv.c | 25 +++++++++++++------------
 1 file changed, 13 insertions(+), 12 deletions(-)

drivers/block/zram/zram_drv.c +11 -9

index 9250b3f..827ab21 100644
@@ -765,9 +765,18 @@ static ssize_t disksize_store(struct device *dev,
 		return -EINVAL;
 
 	disksize = PAGE_ALIGN(disksize);
+	down_write(&zram->init_lock);
+	if (init_done(zram)) {
+		up_write(&zram->init_lock);
+		pr_info("Cannot change disksize for initialized device\n");
+		return -EBUSY;
+	}
+
 	meta = zram_meta_alloc(zram->disk->first_minor, disksize);
-	if (!meta)
-		return -ENOMEM;
+	if (!meta) {
+		err = -ENOMEM;
+		goto out_unlock;
+	}
 
 	comp = zcomp_create(zram->compressor, zram->max_comp_streams);
 	if (IS_ERR(comp)) {

@@ -777,13 +786,6 @@ static ssize_t disksize_store(struct device *dev,
 		goto out_free_meta;
 	}
 
-	down_write(&zram->init_lock);
-	if (init_done(zram)) {
-		pr_info("Cannot change disksize for initialized device\n");
-		err = -EBUSY;
-		goto out_destroy_comp;
-	}
-
 	zram->meta = meta;
 	zram->comp = comp;
 	zram->disksize = disksize;

@@ -799,11 +801,10 @@ static ssize_t disksize_store(struct device *dev,
 
 	return len;
 

Quoted 14 lines

drivers/block/zram/zram_drv.c static ssize_t disksize_store(struct device *dev, · line 810

Quoted 10 lines

drivers/block/zram/zram_drv.c +2 -3

-out_destroy_comp:
-	up_write(&zram->init_lock);
-	zcomp_destroy(comp);
 out_free_meta:
 	zram_meta_free(meta);
+out_unlock:
+	up_write(&zram->init_lock);
 	return err;
 }
 

The init_lock is really troublesome. We can't do call zram_meta_alloc under init_lock due to lockdep report. Please keep in mind.

ah... I do recall it, thanks for your reminder.

Quoted 12 lines

The zram_rw_page is one of the function under reclaim path and hold it as read_lock while here holds it as write_lock.
It's a false positive so that we might could make shut lockdep up by annotation but I don't want it but want to work with lockdep rather than disable. As well, there are other pathes to use init_lock to protect other data where would be victims of lockdep.

I didn't tell the motivation of this patch because it made you busy guys wasted. Let me tell it now. It was another lockdep report by kmem_cache_destroy for zsmalloc compaction about init_lock. That's why the patchset was one of the patch in compaction.

Quoted 8 lines

Yes, the ideal is to remove horrible init_lock of zram in this phase and make code more simple and clear but I don't want to stuck zsmalloc compaction by the work.

Having said that, I feel it's time to revisit
to remove init_lock.
At least, I will think over to find a solution to kill init_lock.

hm, can't think of anything quick...

Quoted 8 lines

Quoted 1 line

	-ss

Hello guys,

How about this?

It's based on Ganesh's patch.
https://lkml.org/lkml/2015/1/24/50

(I see no similarities with Ganesh's patch)

hm, you probably meant this one https://lkml.org/lkml/2015/1/23/406

at glance this makes things a bit more complicated, so I need to think more.

Quoted 48 lines

From afda9fd2f6c40dd0745d8a6babe78c5cbdceddf5 Mon Sep 17 00:00:00 2001
From: Minchan Kim <[email protected]>
Date: Mon, 26 Jan 2015 14:34:10 +0900
Subject: [RFC] zram: remove init_lock in zram_make_request

Admin could reset zram during I/O operation going on so we have used zram->init_lock as read-side lock in I/O path to prevent sudden zram meta freeing.

However, the init_lock is really troublesome.
We can't do call zram_meta_alloc under init_lock due to lockdep splat because zram_rw_page is one of the function under reclaim path and hold it as read_lock while other places in process context hold it as write_lock. So, we have used allocation out of the lock to avoid lockdep warn but it's not good for readability and fainally, I met another lockdep splat between init_lock and cpu_hotpulug from kmem_cache_destroy during wokring zsmalloc compaction. :(

Yes, the ideal is to remove horrible init_lock of zram in rw path. This patch removes it in rw path and instead, put init_done bool variable to check initialization done with smp_[wmb|rmb] and srcu_[un]read_lock to prevent sudden zram meta freeing during I/O operation.

Signed-off-by Minchan Kim <[email protected]>

---
 drivers/block/zram/zram_drv.c | 76 +++++++++++++++++++++++++++++--------------
 drivers/block/zram/zram_drv.h |  5 +++
 2 files changed, 57 insertions(+), 24 deletions(-)

drivers/block/zram/zram_drv.c +2 -1

index a598ada817f0..e06ff975f997 100644
@@ -32,6 +32,7 @@
 #include <linux/string.h>
 #include <linux/vmalloc.h>
 #include <linux/err.h>
+#include <linux/srcu.h>
 
 #include "zram_drv.h"
 

@@ -53,9 +54,16 @@ static ssize_t name##_show(struct device *d,		\
 }									\
 static DEVICE_ATTR_RO(name);
 
-static inline int init_done(struct zram *zram)
+static inline bool init_done(struct zram *zram)

drivers/block/zram/zram_drv.c static ssize_t name##_show(struct device *d, \ · line 66

Quoted 10 lines

drivers/block/zram/zram_drv.c +8 -1

 {
-	return zram->meta != NULL;
+	/*
+	 * init_done can be used without holding zram->init_lock in
+	 * read/write handler(ie, zram_make_request) but we should make sure
+	 * that zram->init_done should set up after meta initialization is
+	 * done. Look at disksize_store.
+	 */
+	smp_rmb();
+	return zram->init_done;

->init_done returns back :)

Quoted 1 line

can we rely on write ->meta; wmb; --- rmb; read ->meta?

Might be possible.

Now that I think about it, it's impossible with zram->meta because we need to nullify it before call_srcu but pre-existing SRCU read-side critical sections can access zram->meta.
Anyway, introducing a new variable should be not a party-pooper.

Quoted 47 lines

Quoted 46 lines

Quoted 3 lines

how much performance do we lose on barriers?

I think it's not too much than locking which does more than(ie, barrier, fairness, spin on owner and so on) such simple barrier.

Quoted 27 lines

Quoted 27 lines

drivers/block/zram/zram_drv.c +4 -3

 }
 
 static inline struct zram *dev_to_zram(struct device *dev)

@@ -326,6 +334,10 @@ static void zram_meta_free(struct zram_meta *meta)
 	kfree(meta);
 }
 
+static void rcu_zram_do_nothing(struct rcu_head *unused)
+{
+}
+
 static struct zram_meta *zram_meta_alloc(int device_id, u64 disksize)
 {
 	char pool_name[8];

@@ -726,11 +738,8 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 		return;
 	}
 
-	zcomp_destroy(zram->comp);
 	zram->max_comp_streams = 1;
 
-	zram_meta_free(zram->meta);
-	zram->meta = NULL;
 	/* Reset stats */
 	memset(&zram->stats, 0, sizeof(zram->stats));
 

Quoted 7 lines

drivers/block/zram/zram_drv.c static void zram_reset_device(struct zram *zram, bool reset_capacity) · line 750

Quoted 5 lines

drivers/block/zram/zram_drv.c +1 -0

@@ -738,8 +747,12 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 	if (reset_capacity)
 		set_capacity(zram->disk, 0);
 
+	zram->init_done = false;

missing wmb?

I thouht about it but when I read comment from call_srcu as follows "each cpu is guaranteed to have executed a full memory barrier", I decided we don't need it. Right? (ie, double check)

Quoted 5 lines

Quoted 1 line

hm, need to think about it.

Another idea is to use kick_all_cpus_sync, not srcu. With that, we don't need to add more instruction in rw path. I will try it.

From 560478040d2e08c61796e67d0c3ee519ae67ac0f Mon Sep 17 00:00:00 2001
From: Minchan Kim <[email protected]>
Date: Mon, 26 Jan 2015 14:34:10 +0900
Subject: [PATCH] zram: remove init_lock in zram_make_request

Admin could reset zram during I/O operation going on so we have used zram->init_lock as read-side lock in I/O path to prevent sudden zram meta freeing.

However, the init_lock is really troublesome.
We can't do call zram_meta_alloc under init_lock due to lockdep splat because zram_rw_page is one of the function under reclaim path and hold it as read_lock while other places in process context hold it as write_lock. So, we have used allocation out of the lock to avoid lockdep warn but it's not good for readability and finally, I met another lockdep splat between init_lock and cpu_hotpulug from kmem_cache_destroy during wokring zsmalloc compaction. :(

Yes, the ideal is to remove horrible init_lock of zram in rw path. This patch removes it in rw path and instead, use kick_all_cpus_sync and a bool init_done variable to check initialization done with smp_[wmb|rmb].

Upon kick_all_cpus_sync returns, any CPU cannot access zram meta any more due to init_done in zram_make_request so it's safe to free meta. So, finally, we avoids init_lock in reclaim context so we are free for deadlock.

Signed-off-by Minchan Kim <[email protected]>

---
 drivers/block/zram/zram_drv.c | 70 +++++++++++++++++++++++++------------------
 drivers/block/zram/zram_drv.h |  2 ++
 2 files changed, 43 insertions(+), 29 deletions(-)

drivers/block/zram/zram_drv.c +1 -2

index a598ada817f0..404602b1932e 100644
@@ -53,9 +53,16 @@ static ssize_t name##_show(struct device *d,		\
 }									\
 static DEVICE_ATTR_RO(name);
 
-static inline int init_done(struct zram *zram)
+static inline bool init_done(struct zram *zram)
 {
-	return zram->meta != NULL;

drivers/block/zram/zram_drv.c +0 -3

 static inline struct zram *dev_to_zram(struct device *dev)

@@ -726,11 +733,8 @@ static void zram_reset_device(struct zram *zram, bool reset_capacity)
 		return;
 	}
 
-	zcomp_destroy(zram->comp);
 	zram->max_comp_streams = 1;
 
-	zram_meta_free(zram->meta);
-	zram->meta = NULL;
 	/* Reset stats */
 	memset(&zram->stats, 0, sizeof(zram->stats));
 

drivers/block/zram/zram_drv.c +16 -9

+	/*
+	 * From now on, any read/write cannot access zram meta data
+	 * by init_done in the handler.
+	 */
+	zram_meta_free(zram->meta);
+	zcomp_destroy(zram->comp);
 	up_write(&zram->init_lock);
-
 	/*
 	 * Revalidate disk out of the init_lock to avoid lockdep splat.
 	 * It's okay because disk's capacity is protected by init_lock

@@ -762,10 +774,19 @@ static ssize_t disksize_store(struct device *dev,
 	if (!disksize)
 		return -EINVAL;
 
+	down_write(&zram->init_lock);
+	if (init_done(zram)) {
+		pr_info("Cannot change disksize for initialized device\n");
+		up_write(&zram->init_lock);
+		return -EBUSY;
+	}
+
 	disksize = PAGE_ALIGN(disksize);
 	meta = zram_meta_alloc(zram->disk->first_minor, disksize);
-	if (!meta)
+	if (!meta) {
+		up_write(&zram->init_lock);
 		return -ENOMEM;
+	}
 
 	comp = zcomp_create(zram->compressor, zram->max_comp_streams);
 	if (IS_ERR(comp)) {

@@ -775,17 +796,17 @@ static ssize_t disksize_store(struct device *dev,
 		goto out_free_meta;
 	}
 
-	down_write(&zram->init_lock);
-	if (init_done(zram)) {
-		pr_info("Cannot change disksize for initialized device\n");
-		err = -EBUSY;
-		goto out_destroy_comp;
-	}
-
 	zram->meta = meta;