[RH7,v2] cgroup: add export_operations to cgroup super block

Submitted by Andrey Zhadchenko on July 24, 2020, 3:50 p.m.

Details

Message ID 1595605816-582804-1-git-send-email-andrey.zhadchenko@virtuozzo.com
State New
Series "cgroup: add export_operations to cgroup super block"
Headers show

Commit Message

Andrey Zhadchenko July 24, 2020, 3:50 p.m.
criu uses fhandle from fdinfo to dump inotify objects. cgroup super block has
no export operations, but .encode_fh and .fh_to_dentry are needed for
inotify_fdinfo function and open_by_handle_at syscall in order to correctly
open files located on cgroupfs by fhandle.
Add hash table as a storage for inodes with exported fhandle.

https://jira.sw.ru/browse/PSBM-105889
Signed-off-by: Andrey Zhadchenko <andrey.zhadchenko@virtuozzo.com>
---
 kernel/cgroup.c | 112 +++++++++++++++++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 111 insertions(+), 1 deletion(-)

Patch hide | download patch | download mbox

diff --git a/kernel/cgroup.c b/kernel/cgroup.c
index 9fdba79..a459c56 100644
--- a/kernel/cgroup.c
+++ b/kernel/cgroup.c
@@ -62,6 +62,7 @@ 
 #include <linux/kthread.h>
 #include <linux/ve.h>
 #include <linux/stacktrace.h>
+#include <linux/exportfs.h>
 
 #include <linux/atomic.h>
 
@@ -1390,9 +1391,117 @@  out:
 }
 #endif
 
+/*
+ * hashtable for inodes that have exported fhandles.
+ * When we export fhandle, we add it's inode into
+ * hashtable so we can find it fast
+ */
+
+#define CGROUP_INODE_HASH_BITS 10
+static DEFINE_HASHTABLE(cgroup_inode_table, CGROUP_INODE_HASH_BITS);
+static DEFINE_SPINLOCK(cgroup_inode_table_lock);
+
+struct cg_inode_hitem {
+	struct inode *inode;
+	struct hlist_node hlist;
+};
+
+static inline unsigned long cgroup_inode_get_hash(unsigned int i_ino)
+{
+	return hash_32(i_ino, CGROUP_INODE_HASH_BITS);
+}
+
+static struct cg_inode_hitem *cgroup_inode_hash_find(unsigned int i_ino)
+{
+	struct cg_inode_hitem *i;
+	struct hlist_head *head = cgroup_inode_table
+		+ cgroup_inode_get_hash(i_ino);
+	struct cg_inode_hitem *found = 0;
+
+	spin_lock(&cgroup_inode_table_lock);
+	hlist_for_each_entry(i, head, hlist) {
+		if (i->inode->i_ino == i_ino) {
+			found = i;
+			break;
+		}
+	}
+	spin_unlock(&cgroup_inode_table_lock);
+
+	return found;
+}
+
+static struct dentry *cgroup_fh_to_dentry(struct super_block *sb,
+		struct fid *fid, int fh_len, int fh_type)
+{
+	struct cg_inode_hitem *item;
+	struct dentry *dentry = ERR_PTR(-ENOENT);
+
+	if (fh_len < 1)
+		return NULL;
+
+	item = cgroup_inode_hash_find(fid->raw[0]);
+	if (item)
+		dentry = d_find_alias(item->inode);
+
+	return dentry;
+}
+
+static int cgroup_encode_fh(struct inode *inode, __u32 *fh, int *len,
+				struct inode *parent)
+{
+	struct hlist_head *head = cgroup_inode_table
+		+ cgroup_inode_get_hash(inode->i_ino);
+	struct cg_inode_hitem *item;
+
+	if (*len < 1) {
+		*len = 1;
+		return FILEID_INVALID;
+	}
+
+	if (cgroup_inode_hash_find(inode->i_ino) == 0) {
+		item = kmalloc(sizeof(struct cg_inode_hitem),
+			GFP_KERNEL);
+		/*
+		 * encode_fh is expected to return 255 (FILEID_INVALID)
+		 * in case of failure. We can't return ENOMEM, so
+		 * return FILEID_INVALID at least
+		 */
+		if (!item)
+			return FILEID_INVALID;
+		item->inode = inode;
+
+		spin_lock(&cgroup_inode_table_lock);
+		hlist_add_head(&item->hlist, head);
+		spin_unlock(&cgroup_inode_table_lock);
+	}
+
+	fh[0] = inode->i_ino;
+	*len = 1;
+	return 1;
+}
+
+static const struct export_operations cgroup_export_ops = {
+	.encode_fh      = cgroup_encode_fh,
+	.fh_to_dentry	= cgroup_fh_to_dentry,
+};
+
+static int cgroup_delete_inode(struct inode *inode)
+{
+	struct cg_inode_hitem *item = cgroup_inode_hash_find(inode->i_ino);
+	if (item) {
+		spin_lock(&cgroup_inode_table_lock);
+		hlist_del(&item->hlist);
+		spin_unlock(&cgroup_inode_table_lock);
+
+		kfree(item);
+	}
+
+	return generic_delete_inode(inode);
+}
+
 static const struct super_operations cgroup_ops = {
 	.statfs = simple_statfs,
-	.drop_inode = generic_delete_inode,
+	.drop_inode = cgroup_delete_inode,
 	.show_options = cgroup_show_options,
 #ifdef CONFIG_VE
 	.show_path = cgroup_show_path,
@@ -1539,6 +1648,7 @@  static int cgroup_set_super(struct super_block *sb, void *data)
 	sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
 	sb->s_magic = CGROUP_SUPER_MAGIC;
 	sb->s_op = &cgroup_ops;
+	sb->s_export_op = &cgroup_export_ops;
 
 	return 0;
 }

Comments

Pavel Tikhomirov July 27, 2020, 9:43 a.m.
On 7/24/20 6:50 PM, Andrey Zhadchenko wrote:
> criu uses fhandle from fdinfo to dump inotify objects. cgroup super block has
> no export operations, but .encode_fh and .fh_to_dentry are needed for
> inotify_fdinfo function and open_by_handle_at syscall in order to correctly
> open files located on cgroupfs by fhandle.
> Add hash table as a storage for inodes with exported fhandle.
> 
> https://jira.sw.ru/browse/PSBM-105889
> Signed-off-by: Andrey Zhadchenko <andrey.zhadchenko@virtuozzo.com>
> ---
>   kernel/cgroup.c | 112 +++++++++++++++++++++++++++++++++++++++++++++++++++++++-
>   1 file changed, 111 insertions(+), 1 deletion(-)
> 
> diff --git a/kernel/cgroup.c b/kernel/cgroup.c
> index 9fdba79..a459c56 100644
> --- a/kernel/cgroup.c
> +++ b/kernel/cgroup.c
> @@ -62,6 +62,7 @@
>   #include <linux/kthread.h>
>   #include <linux/ve.h>
>   #include <linux/stacktrace.h>
> +#include <linux/exportfs.h>
>   
>   #include <linux/atomic.h>
>   
> @@ -1390,9 +1391,117 @@ out:
>   }
>   #endif
>   
> +/*
> + * hashtable for inodes that have exported fhandles.
> + * When we export fhandle, we add it's inode into
> + * hashtable so we can find it fast
> + */
> +
> +#define CGROUP_INODE_HASH_BITS 10
> +static DEFINE_HASHTABLE(cgroup_inode_table, CGROUP_INODE_HASH_BITS);
> +static DEFINE_SPINLOCK(cgroup_inode_table_lock);
> +
> +struct cg_inode_hitem {
> +	struct inode *inode;
> +	struct hlist_node hlist;
> +};
> +
> +static inline unsigned long cgroup_inode_get_hash(unsigned int i_ino)
> +{
> +	return hash_32(i_ino, CGROUP_INODE_HASH_BITS);
> +}
> +
> +static struct cg_inode_hitem *cgroup_inode_hash_find(unsigned int i_ino)
> +{
> +	struct cg_inode_hitem *i;
> +	struct hlist_head *head = cgroup_inode_table
> +		+ cgroup_inode_get_hash(i_ino);
> +	struct cg_inode_hitem *found = 0;

NULL would look better than 0 for pointer.

> +
> +	spin_lock(&cgroup_inode_table_lock);
> +	hlist_for_each_entry(i, head, hlist) {
> +		if (i->inode->i_ino == i_ino) {
> +			found = i;
> +			break;
> +		}
> +	}
> +	spin_unlock(&cgroup_inode_table_lock);

The cg_inode_hitem can be freed from other thread just after you release 
the lock. Inode can be evicted and freed just after you release the 
lock. We should iget a reference on inode before releasing lock and 
return inode instead of cg_inode_hitem.

> +
> +	return found;
> +}
> +
> +static struct dentry *cgroup_fh_to_dentry(struct super_block *sb,
> +		struct fid *fid, int fh_len, int fh_type)
> +{
> +	struct cg_inode_hitem *item;
> +	struct dentry *dentry = ERR_PTR(-ENOENT);
> +
> +	if (fh_len < 1)
> +		return NULL;
> +
> +	item = cgroup_inode_hash_find(fid->raw[0]);
> +	if (item)
> +		dentry = d_find_alias(item->inode);

When you've got dentry you can put the inode reference back. See how 
it's done in shmem_fh_to_dentry.

> +
> +	return dentry;
> +}
> +
> +static int cgroup_encode_fh(struct inode *inode, __u32 *fh, int *len,
> +				struct inode *parent)
> +{
> +	struct hlist_head *head = cgroup_inode_table
> +		+ cgroup_inode_get_hash(inode->i_ino);
> +	struct cg_inode_hitem *item;
> +
> +	if (*len < 1) {
> +		*len = 1;
> +		return FILEID_INVALID;
> +	}
> +
> +	if (cgroup_inode_hash_find(inode->i_ino) == 0) {

Would look better:
	if (!cgroup_inode_hash_find(inode->i_ino)) {

> +		item = kmalloc(sizeof(struct cg_inode_hitem),
> +			GFP_KERNEL);
> +		/*
> +		 * encode_fh is expected to return 255 (FILEID_INVALID)
> +		 * in case of failure. We can't return ENOMEM, so
> +		 * return FILEID_INVALID at least
> +		 */
> +		if (!item)
> +			return FILEID_INVALID;
> +		item->inode = inode;
> +
> +		spin_lock(&cgroup_inode_table_lock);
> +		hlist_add_head(&item->hlist, head);
> +		spin_unlock(&cgroup_inode_table_lock);
> +	}
> +
> +	fh[0] = inode->i_ino;
> +	*len = 1;
> +	return 1;
> +}
> +
> +static const struct export_operations cgroup_export_ops = {
> +	.encode_fh      = cgroup_encode_fh,
> +	.fh_to_dentry	= cgroup_fh_to_dentry,
> +};
> +
> +static int cgroup_delete_inode(struct inode *inode)
> +{
> +	struct cg_inode_hitem *item = cgroup_inode_hash_find(inode->i_ino);

I don't like this. Lets only merge variable initialization and 
declaration if You don't expect any fail. And put empty line between 
declaration and everything else.

> +	if (item) {
> +		spin_lock(&cgroup_inode_table_lock);
> +		hlist_del(&item->hlist);
> +		spin_unlock(&cgroup_inode_table_lock);
> +
> +		kfree(item);
> +	}
> +
> +	return generic_delete_inode(inode);
> +}
> +
>   static const struct super_operations cgroup_ops = {
>   	.statfs = simple_statfs,
> -	.drop_inode = generic_delete_inode,
> +	.drop_inode = cgroup_delete_inode,
>   	.show_options = cgroup_show_options,
>   #ifdef CONFIG_VE
>   	.show_path = cgroup_show_path,
> @@ -1539,6 +1648,7 @@ static int cgroup_set_super(struct super_block *sb, void *data)
>   	sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
>   	sb->s_magic = CGROUP_SUPER_MAGIC;
>   	sb->s_op = &cgroup_ops;
> +	sb->s_export_op = &cgroup_export_ops;
>   
>   	return 0;
>   }
>
Pavel Tikhomirov July 27, 2020, 10:32 a.m.
One more thing: inode can be reused, and I don't see where you handle 
it. E.g.: shmem_fh_to_dentry / shmem_match uses inode generation to 
detect it.

On 7/27/20 12:43 PM, Pavel Tikhomirov wrote:
> 
> 
> On 7/24/20 6:50 PM, Andrey Zhadchenko wrote:
>> criu uses fhandle from fdinfo to dump inotify objects. cgroup super 
>> block has
>> no export operations, but .encode_fh and .fh_to_dentry are needed for
>> inotify_fdinfo function and open_by_handle_at syscall in order to 
>> correctly
>> open files located on cgroupfs by fhandle.
>> Add hash table as a storage for inodes with exported fhandle.
>>
>> https://jira.sw.ru/browse/PSBM-105889
>> Signed-off-by: Andrey Zhadchenko <andrey.zhadchenko@virtuozzo.com>
>> ---
>>   kernel/cgroup.c | 112 
>> +++++++++++++++++++++++++++++++++++++++++++++++++++++++-
>>   1 file changed, 111 insertions(+), 1 deletion(-)
>>
>> diff --git a/kernel/cgroup.c b/kernel/cgroup.c
>> index 9fdba79..a459c56 100644
>> --- a/kernel/cgroup.c
>> +++ b/kernel/cgroup.c
>> @@ -62,6 +62,7 @@
>>   #include <linux/kthread.h>
>>   #include <linux/ve.h>
>>   #include <linux/stacktrace.h>
>> +#include <linux/exportfs.h>
>>   #include <linux/atomic.h>
>> @@ -1390,9 +1391,117 @@ out:
>>   }
>>   #endif
>> +/*
>> + * hashtable for inodes that have exported fhandles.
>> + * When we export fhandle, we add it's inode into
>> + * hashtable so we can find it fast
>> + */
>> +
>> +#define CGROUP_INODE_HASH_BITS 10
>> +static DEFINE_HASHTABLE(cgroup_inode_table, CGROUP_INODE_HASH_BITS);
>> +static DEFINE_SPINLOCK(cgroup_inode_table_lock);
>> +
>> +struct cg_inode_hitem {
>> +    struct inode *inode;
>> +    struct hlist_node hlist;
>> +};
>> +
>> +static inline unsigned long cgroup_inode_get_hash(unsigned int i_ino)
>> +{
>> +    return hash_32(i_ino, CGROUP_INODE_HASH_BITS);
>> +}
>> +
>> +static struct cg_inode_hitem *cgroup_inode_hash_find(unsigned int i_ino)
>> +{
>> +    struct cg_inode_hitem *i;
>> +    struct hlist_head *head = cgroup_inode_table
>> +        + cgroup_inode_get_hash(i_ino);
>> +    struct cg_inode_hitem *found = 0;
> 
> NULL would look better than 0 for pointer.
> 
>> +
>> +    spin_lock(&cgroup_inode_table_lock);
>> +    hlist_for_each_entry(i, head, hlist) {
>> +        if (i->inode->i_ino == i_ino) {
>> +            found = i;
>> +            break;
>> +        }
>> +    }
>> +    spin_unlock(&cgroup_inode_table_lock);
> 
> The cg_inode_hitem can be freed from other thread just after you release 
> the lock. Inode can be evicted and freed just after you release the 
> lock. We should iget a reference on inode before releasing lock and 
> return inode instead of cg_inode_hitem.
> 
>> +
>> +    return found;
>> +}
>> +
>> +static struct dentry *cgroup_fh_to_dentry(struct super_block *sb,
>> +        struct fid *fid, int fh_len, int fh_type)
>> +{
>> +    struct cg_inode_hitem *item;
>> +    struct dentry *dentry = ERR_PTR(-ENOENT);
>> +
>> +    if (fh_len < 1)
>> +        return NULL;
>> +
>> +    item = cgroup_inode_hash_find(fid->raw[0]);
>> +    if (item)
>> +        dentry = d_find_alias(item->inode);
> 
> When you've got dentry you can put the inode reference back. See how 
> it's done in shmem_fh_to_dentry.
> 
>> +
>> +    return dentry;
>> +}
>> +
>> +static int cgroup_encode_fh(struct inode *inode, __u32 *fh, int *len,
>> +                struct inode *parent)
>> +{
>> +    struct hlist_head *head = cgroup_inode_table
>> +        + cgroup_inode_get_hash(inode->i_ino);
>> +    struct cg_inode_hitem *item;
>> +
>> +    if (*len < 1) {
>> +        *len = 1;
>> +        return FILEID_INVALID;
>> +    }
>> +
>> +    if (cgroup_inode_hash_find(inode->i_ino) == 0) {
> 
> Would look better:
>      if (!cgroup_inode_hash_find(inode->i_ino)) {
> 
>> +        item = kmalloc(sizeof(struct cg_inode_hitem),
>> +            GFP_KERNEL);
>> +        /*
>> +         * encode_fh is expected to return 255 (FILEID_INVALID)
>> +         * in case of failure. We can't return ENOMEM, so
>> +         * return FILEID_INVALID at least
>> +         */
>> +        if (!item)
>> +            return FILEID_INVALID;
>> +        item->inode = inode;
>> +
>> +        spin_lock(&cgroup_inode_table_lock);
>> +        hlist_add_head(&item->hlist, head);
>> +        spin_unlock(&cgroup_inode_table_lock);
>> +    }
>> +
>> +    fh[0] = inode->i_ino;
>> +    *len = 1;
>> +    return 1;
>> +}
>> +
>> +static const struct export_operations cgroup_export_ops = {
>> +    .encode_fh      = cgroup_encode_fh,
>> +    .fh_to_dentry    = cgroup_fh_to_dentry,
>> +};
>> +
>> +static int cgroup_delete_inode(struct inode *inode)
>> +{
>> +    struct cg_inode_hitem *item = cgroup_inode_hash_find(inode->i_ino);
> 
> I don't like this. Lets only merge variable initialization and 
> declaration if You don't expect any fail. And put empty line between 
> declaration and everything else.
> 
>> +    if (item) {
>> +        spin_lock(&cgroup_inode_table_lock);
>> +        hlist_del(&item->hlist);
>> +        spin_unlock(&cgroup_inode_table_lock);
>> +
>> +        kfree(item);
>> +    }
>> +
>> +    return generic_delete_inode(inode);
>> +}
>> +
>>   static const struct super_operations cgroup_ops = {
>>       .statfs = simple_statfs,
>> -    .drop_inode = generic_delete_inode,
>> +    .drop_inode = cgroup_delete_inode,
>>       .show_options = cgroup_show_options,
>>   #ifdef CONFIG_VE
>>       .show_path = cgroup_show_path,
>> @@ -1539,6 +1648,7 @@ static int cgroup_set_super(struct super_block 
>> *sb, void *data)
>>       sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
>>       sb->s_magic = CGROUP_SUPER_MAGIC;
>>       sb->s_op = &cgroup_ops;
>> +    sb->s_export_op = &cgroup_export_ops;
>>       return 0;
>>   }
>>
>