shmem: prepare shmem quota infrastructure
authorCarlos Maiolino <cem@kernel.org>
Tue, 25 Jul 2023 14:45:07 +0000 (16:45 +0200)
committerChristian Brauner <brauner@kernel.org>
Wed, 9 Aug 2023 07:15:39 +0000 (09:15 +0200)
Add new shmem quota format, its quota_format_ops together with
dquot_operations

Signed-off-by: Lukas Czerner <lczerner@redhat.com>
Signed-off-by: Carlos Maiolino <cmaiolino@redhat.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Message-Id: <20230725144510.253763-5-cem@kernel.org>
Signed-off-by: Christian Brauner <brauner@kernel.org>
fs/Kconfig
include/linux/shmem_fs.h
include/uapi/linux/quota.h
mm/Makefile
mm/shmem_quota.c [new file with mode: 0644]

index 18d034e..8218a71 100644 (file)
@@ -233,6 +233,18 @@ config TMPFS_INODE64
 
          If unsure, say N.
 
+config TMPFS_QUOTA
+       bool "Tmpfs quota support"
+       depends on TMPFS
+       select QUOTA
+       help
+         Quota support allows to set per user and group limits for tmpfs
+         usage.  Say Y to enable quota support. Once enabled you can control
+         user and group quota enforcement with quota, usrquota and grpquota
+         mount options.
+
+         If unsure, say N.
+
 config ARCH_SUPPORTS_HUGETLBFS
        def_bool n
 
index 9029abd..7abfaf7 100644 (file)
 
 /* inode in-kernel data */
 
+#ifdef CONFIG_TMPFS_QUOTA
+#define SHMEM_MAXQUOTAS 2
+#endif
+
 struct shmem_inode_info {
        spinlock_t              lock;
        unsigned int            seals;          /* shmem seals */
@@ -172,4 +176,12 @@ extern int shmem_mfill_atomic_pte(pmd_t *dst_pmd,
 #endif /* CONFIG_SHMEM */
 #endif /* CONFIG_USERFAULTFD */
 
+/*
+ * Used space is stored as unsigned 64-bit value in bytes but
+ * quota core supports only signed 64-bit values so use that
+ * as a limit
+ */
+#define SHMEM_QUOTA_MAX_SPC_LIMIT 0x7fffffffffffffffLL /* 2^63-1 */
+#define SHMEM_QUOTA_MAX_INO_LIMIT 0x7fffffffffffffffLL
+
 #endif
index f17c963..5209010 100644 (file)
@@ -77,6 +77,7 @@
 #define        QFMT_VFS_V0 2
 #define QFMT_OCFS2 3
 #define        QFMT_VFS_V1 4
+#define        QFMT_SHMEM 5
 
 /* Size of block in which space limits are passed through the quota
  * interface */
index 678530a..d4ee209 100644 (file)
@@ -51,7 +51,7 @@ obj-y                 := filemap.o mempool.o oom_kill.o fadvise.o \
                           readahead.o swap.o truncate.o vmscan.o shmem.o \
                           util.o mmzone.o vmstat.o backing-dev.o \
                           mm_init.o percpu.o slab_common.o \
-                          compaction.o show_mem.o\
+                          compaction.o show_mem.o shmem_quota.o\
                           interval_tree.o list_lru.o workingset.o \
                           debug.o gup.o mmap_lock.o $(mmu-y)
 
diff --git a/mm/shmem_quota.c b/mm/shmem_quota.c
new file mode 100644 (file)
index 0000000..e92b8ec
--- /dev/null
@@ -0,0 +1,318 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * In memory quota format relies on quota infrastructure to store dquot
+ * information for us. While conventional quota formats for file systems
+ * with persistent storage can load quota information into dquot from the
+ * storage on-demand and hence quota dquot shrinker can free any dquot
+ * that is not currently being used, it must be avoided here. Otherwise we
+ * can lose valuable information, user provided limits, because there is
+ * no persistent storage to load the information from afterwards.
+ *
+ * One information that in-memory quota format needs to keep track of is
+ * a sorted list of ids for each quota type. This is done by utilizing
+ * an rb tree which root is stored in mem_dqinfo->dqi_priv for each quota
+ * type.
+ *
+ * This format can be used to support quota on file system without persistent
+ * storage such as tmpfs.
+ *
+ * Author:     Lukas Czerner <lczerner@redhat.com>
+ *             Carlos Maiolino <cmaiolino@redhat.com>
+ *
+ * Copyright (C) 2023 Red Hat, Inc.
+ */
+#include <linux/errno.h>
+#include <linux/fs.h>
+#include <linux/mount.h>
+#include <linux/kernel.h>
+#include <linux/init.h>
+#include <linux/module.h>
+#include <linux/slab.h>
+#include <linux/rbtree.h>
+#include <linux/shmem_fs.h>
+
+#include <linux/quotaops.h>
+#include <linux/quota.h>
+
+#ifdef CONFIG_TMPFS_QUOTA
+
+/*
+ * The following constants define the amount of time given a user
+ * before the soft limits are treated as hard limits (usually resulting
+ * in an allocation failure). The timer is started when the user crosses
+ * their soft limit, it is reset when they go below their soft limit.
+ */
+#define SHMEM_MAX_IQ_TIME 604800       /* (7*24*60*60) 1 week */
+#define SHMEM_MAX_DQ_TIME 604800       /* (7*24*60*60) 1 week */
+
+struct quota_id {
+       struct rb_node  node;
+       qid_t           id;
+       qsize_t         bhardlimit;
+       qsize_t         bsoftlimit;
+       qsize_t         ihardlimit;
+       qsize_t         isoftlimit;
+};
+
+static int shmem_check_quota_file(struct super_block *sb, int type)
+{
+       /* There is no real quota file, nothing to do */
+       return 1;
+}
+
+/*
+ * There is no real quota file. Just allocate rb_root for quota ids and
+ * set limits
+ */
+static int shmem_read_file_info(struct super_block *sb, int type)
+{
+       struct quota_info *dqopt = sb_dqopt(sb);
+       struct mem_dqinfo *info = &dqopt->info[type];
+
+       info->dqi_priv = kzalloc(sizeof(struct rb_root), GFP_NOFS);
+       if (!info->dqi_priv)
+               return -ENOMEM;
+
+       info->dqi_max_spc_limit = SHMEM_QUOTA_MAX_SPC_LIMIT;
+       info->dqi_max_ino_limit = SHMEM_QUOTA_MAX_INO_LIMIT;
+
+       info->dqi_bgrace = SHMEM_MAX_DQ_TIME;
+       info->dqi_igrace = SHMEM_MAX_IQ_TIME;
+       info->dqi_flags = 0;
+
+       return 0;
+}
+
+static int shmem_write_file_info(struct super_block *sb, int type)
+{
+       /* There is no real quota file, nothing to do */
+       return 0;
+}
+
+/*
+ * Free all the quota_id entries in the rb tree and rb_root.
+ */
+static int shmem_free_file_info(struct super_block *sb, int type)
+{
+       struct mem_dqinfo *info = &sb_dqopt(sb)->info[type];
+       struct rb_root *root = info->dqi_priv;
+       struct quota_id *entry;
+       struct rb_node *node;
+
+       info->dqi_priv = NULL;
+       node = rb_first(root);
+       while (node) {
+               entry = rb_entry(node, struct quota_id, node);
+               node = rb_next(&entry->node);
+
+               rb_erase(&entry->node, root);
+               kfree(entry);
+       }
+
+       kfree(root);
+       return 0;
+}
+
+static int shmem_get_next_id(struct super_block *sb, struct kqid *qid)
+{
+       struct mem_dqinfo *info = sb_dqinfo(sb, qid->type);
+       struct rb_node *node = ((struct rb_root *)info->dqi_priv)->rb_node;
+       qid_t id = from_kqid(&init_user_ns, *qid);
+       struct quota_info *dqopt = sb_dqopt(sb);
+       struct quota_id *entry = NULL;
+       int ret = 0;
+
+       if (!sb_has_quota_active(sb, qid->type))
+               return -ESRCH;
+
+       down_read(&dqopt->dqio_sem);
+       while (node) {
+               entry = rb_entry(node, struct quota_id, node);
+
+               if (id < entry->id)
+                       node = node->rb_left;
+               else if (id > entry->id)
+                       node = node->rb_right;
+               else
+                       goto got_next_id;
+       }
+
+       if (!entry) {
+               ret = -ENOENT;
+               goto out_unlock;
+       }
+
+       if (id > entry->id) {
+               node = rb_next(&entry->node);
+               if (!node) {
+                       ret = -ENOENT;
+                       goto out_unlock;
+               }
+               entry = rb_entry(node, struct quota_id, node);
+       }
+
+got_next_id:
+       *qid = make_kqid(&init_user_ns, qid->type, entry->id);
+out_unlock:
+       up_read(&dqopt->dqio_sem);
+       return ret;
+}
+
+/*
+ * Load dquot with limits from existing entry, or create the new entry if
+ * it does not exist.
+ */
+static int shmem_acquire_dquot(struct dquot *dquot)
+{
+       struct mem_dqinfo *info = sb_dqinfo(dquot->dq_sb, dquot->dq_id.type);
+       struct rb_node **n = &((struct rb_root *)info->dqi_priv)->rb_node;
+       struct rb_node *parent = NULL, *new_node = NULL;
+       struct quota_id *new_entry, *entry;
+       qid_t id = from_kqid(&init_user_ns, dquot->dq_id);
+       struct quota_info *dqopt = sb_dqopt(dquot->dq_sb);
+       int ret = 0;
+
+       mutex_lock(&dquot->dq_lock);
+
+       down_write(&dqopt->dqio_sem);
+       while (*n) {
+               parent = *n;
+               entry = rb_entry(parent, struct quota_id, node);
+
+               if (id < entry->id)
+                       n = &(*n)->rb_left;
+               else if (id > entry->id)
+                       n = &(*n)->rb_right;
+               else
+                       goto found;
+       }
+
+       /* We don't have entry for this id yet, create it */
+       new_entry = kzalloc(sizeof(struct quota_id), GFP_NOFS);
+       if (!new_entry) {
+               ret = -ENOMEM;
+               goto out_unlock;
+       }
+
+       new_entry->id = id;
+       new_node = &new_entry->node;
+       rb_link_node(new_node, parent, n);
+       rb_insert_color(new_node, (struct rb_root *)info->dqi_priv);
+       entry = new_entry;
+
+found:
+       /* Load the stored limits from the tree */
+       spin_lock(&dquot->dq_dqb_lock);
+       dquot->dq_dqb.dqb_bhardlimit = entry->bhardlimit;
+       dquot->dq_dqb.dqb_bsoftlimit = entry->bsoftlimit;
+       dquot->dq_dqb.dqb_ihardlimit = entry->ihardlimit;
+       dquot->dq_dqb.dqb_isoftlimit = entry->isoftlimit;
+
+       if (!dquot->dq_dqb.dqb_bhardlimit &&
+           !dquot->dq_dqb.dqb_bsoftlimit &&
+           !dquot->dq_dqb.dqb_ihardlimit &&
+           !dquot->dq_dqb.dqb_isoftlimit)
+               set_bit(DQ_FAKE_B, &dquot->dq_flags);
+       spin_unlock(&dquot->dq_dqb_lock);
+
+       /* Make sure flags update is visible after dquot has been filled */
+       smp_mb__before_atomic();
+       set_bit(DQ_ACTIVE_B, &dquot->dq_flags);
+out_unlock:
+       up_write(&dqopt->dqio_sem);
+       mutex_unlock(&dquot->dq_lock);
+       return ret;
+}
+
+/*
+ * Store limits from dquot in the tree unless it's fake. If it is fake
+ * remove the id from the tree since there is no useful information in
+ * there.
+ */
+static int shmem_release_dquot(struct dquot *dquot)
+{
+       struct mem_dqinfo *info = sb_dqinfo(dquot->dq_sb, dquot->dq_id.type);
+       struct rb_node *node = ((struct rb_root *)info->dqi_priv)->rb_node;
+       qid_t id = from_kqid(&init_user_ns, dquot->dq_id);
+       struct quota_info *dqopt = sb_dqopt(dquot->dq_sb);
+       struct quota_id *entry = NULL;
+
+       mutex_lock(&dquot->dq_lock);
+       /* Check whether we are not racing with some other dqget() */
+       if (dquot_is_busy(dquot))
+               goto out_dqlock;
+
+       down_write(&dqopt->dqio_sem);
+       while (node) {
+               entry = rb_entry(node, struct quota_id, node);
+
+               if (id < entry->id)
+                       node = node->rb_left;
+               else if (id > entry->id)
+                       node = node->rb_right;
+               else
+                       goto found;
+       }
+
+       /* We should always find the entry in the rb tree */
+       WARN_ONCE(1, "quota id %u from dquot %p, not in rb tree!\n", id, dquot);
+       up_write(&dqopt->dqio_sem);
+       mutex_unlock(&dquot->dq_lock);
+       return -ENOENT;
+
+found:
+       if (test_bit(DQ_FAKE_B, &dquot->dq_flags)) {
+               /* Remove entry from the tree */
+               rb_erase(&entry->node, info->dqi_priv);
+               kfree(entry);
+       } else {
+               /* Store the limits in the tree */
+               spin_lock(&dquot->dq_dqb_lock);
+               entry->bhardlimit = dquot->dq_dqb.dqb_bhardlimit;
+               entry->bsoftlimit = dquot->dq_dqb.dqb_bsoftlimit;
+               entry->ihardlimit = dquot->dq_dqb.dqb_ihardlimit;
+               entry->isoftlimit = dquot->dq_dqb.dqb_isoftlimit;
+               spin_unlock(&dquot->dq_dqb_lock);
+       }
+
+       clear_bit(DQ_ACTIVE_B, &dquot->dq_flags);
+       up_write(&dqopt->dqio_sem);
+
+out_dqlock:
+       mutex_unlock(&dquot->dq_lock);
+       return 0;
+}
+
+static int shmem_mark_dquot_dirty(struct dquot *dquot)
+{
+       return 0;
+}
+
+static int shmem_dquot_write_info(struct super_block *sb, int type)
+{
+       return 0;
+}
+
+static const struct quota_format_ops shmem_format_ops = {
+       .check_quota_file       = shmem_check_quota_file,
+       .read_file_info         = shmem_read_file_info,
+       .write_file_info        = shmem_write_file_info,
+       .free_file_info         = shmem_free_file_info,
+};
+
+struct quota_format_type shmem_quota_format = {
+       .qf_fmt_id = QFMT_SHMEM,
+       .qf_ops = &shmem_format_ops,
+       .qf_owner = THIS_MODULE
+};
+
+const struct dquot_operations shmem_quota_operations = {
+       .acquire_dquot          = shmem_acquire_dquot,
+       .release_dquot          = shmem_release_dquot,
+       .alloc_dquot            = dquot_alloc,
+       .destroy_dquot          = dquot_destroy,
+       .write_info             = shmem_dquot_write_info,
+       .mark_dirty             = shmem_mark_dquot_dirty,
+       .get_next_id            = shmem_get_next_id,
+};
+#endif /* CONFIG_TMPFS_QUOTA */