Skip to contentSkip to conversation

fuse: on 64-bit store time in d_fsdata directly

PATCH Superseded Khazhismel Kumykov 2 participants 6 messages Aug 22, 2019

Subsystems (2)

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/3 fuse: on 64-bit store time in d_fsdata directly

Implements the optimization noted in f75fdf22b0a8 ("fuse: don't use ->d_time"), as the additional memory can be significant. (In particular, on SLAB configurations this 8-byte alloc becomes 32 bytes). Per-dentry, this can consume significant memory.

---
 fs/fuse/dir.c | 19 +++++++++++++++++++
 1 file changed, 19 insertions(+)
fs/fuse/dir.c +19 -0
index dd0f64f7bc06..f9c59a296568 100644
@@ -24,6 +24,18 @@ static void fuse_advise_use_readdirplus(struct inode *dir)
 	set_bit(FUSE_I_ADVISE_RDPLUS, &fi->state);
 }
 
+#if BITS_PER_LONG >= 64
+static inline void fuse_dentry_settime(struct dentry *entry, u64 time)
+{
+	entry->d_fsdata = (void *) time;
+}
+
+static inline u64 fuse_dentry_time(struct dentry *entry)
+{
+	return (u64)entry->d_fsdata;
+}
+
+#else
 union fuse_dentry {
 	u64 time;
 	struct rcu_head rcu;
@@ -38,6 +50,7 @@ static inline u64 fuse_dentry_time(struct dentry *entry)
 {
 	return ((union fuse_dentry *) entry->d_fsdata)->time;
 }
+#endif
 
 /*
  * FUSE caches dentries and attributes with separate timeout.  The
@@ -242,6 +255,7 @@ static int fuse_dentry_revalidate(struct dentry *entry, unsigned int flags)
 	goto out;
 }
 
+#if BITS_PER_LONG < 64
 static int fuse_dentry_init(struct dentry *dentry)
 {
 	dentry->d_fsdata = kzalloc(sizeof(union fuse_dentry), GFP_KERNEL);
@@ -254,16 +268,21 @@ static void fuse_dentry_release(struct dentry *dentry)
 
 	kfree_rcu(fd, rcu);
 }
+#endif
 
 const struct dentry_operations fuse_dentry_operations = {
 	.d_revalidate	= fuse_dentry_revalidate,
+#if BITS_PER_LONG < 64
 	.d_init		= fuse_dentry_init,
 	.d_release	= fuse_dentry_release,
+#endif
 };
 
 const struct dentry_operations fuse_root_dentry_operations = {
+#if BITS_PER_LONG < 64
 	.d_init		= fuse_dentry_init,
 	.d_release	= fuse_dentry_release,
+#endif
 };
 
 int fuse_valid_type(int m)
Signature · 1 line
2.23.0.187.g17f5b7556c-goog
#1 lore raw

PATCH 2/3 fuse: kmemcg account fs data

account per-file, dentry, and inode data

accounts the per-file reserved request, adding new fuse_request_alloc_account()

blockdev/superblock and temporary per-request data was left alone, as this usually isn't accounted

---
 fs/fuse/dev.c    | 6 ++++++
 fs/fuse/dir.c    | 3 ++-
 fs/fuse/file.c   | 4 ++--
 fs/fuse/fuse_i.h | 2 ++
 fs/fuse/inode.c  | 3 ++-
 5 files changed, 14 insertions(+), 4 deletions(-)
fs/fuse/dev.c +6 -0
index ea8237513dfa..a0d166a6596f 100644
@@ -97,6 +97,12 @@ struct fuse_req *fuse_request_alloc(unsigned npages)
 }
 EXPORT_SYMBOL_GPL(fuse_request_alloc);
 
+struct fuse_req *fuse_request_alloc_account(unsigned int npages)
+{
+	return __fuse_request_alloc(npages, GFP_KERNEL_ACCOUNT);
+}
+EXPORT_SYMBOL_GPL(fuse_request_alloc_account);
+
 struct fuse_req *fuse_request_alloc_nofs(unsigned npages)
 {
 	return __fuse_request_alloc(npages, GFP_NOFS);
fs/fuse/dir.c +2 -1
index f9c59a296568..2013e1222de7 100644
@@ -258,7 +258,8 @@ static int fuse_dentry_revalidate(struct dentry *entry, unsigned int flags)
 #if BITS_PER_LONG < 64
 static int fuse_dentry_init(struct dentry *dentry)
 {
-	dentry->d_fsdata = kzalloc(sizeof(union fuse_dentry), GFP_KERNEL);
+	dentry->d_fsdata = kzalloc(sizeof(union fuse_dentry),
+				   GFP_KERNEL_ACCOUNT | __GFP_RECLAIMABLE);
 
 	return dentry->d_fsdata ? 0 : -ENOMEM;
 }
fs/fuse/file.c +2 -2
index 5ae2828beb00..c584ad7478b3 100644
@@ -45,12 +45,12 @@ struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
 {
 	struct fuse_file *ff;
 
-	ff = kzalloc(sizeof(struct fuse_file), GFP_KERNEL);
+	ff = kzalloc(sizeof(struct fuse_file), GFP_KERNEL_ACCOUNT);
 	if (unlikely(!ff))
 		return NULL;
 
 	ff->fc = fc;
-	ff->reserved_req = fuse_request_alloc(0);
+	ff->reserved_req = fuse_request_alloc_account(0);
 	if (unlikely(!ff->reserved_req)) {
 		kfree(ff);
 		return NULL;
fs/fuse/fuse_i.h +2 -0
index 24dbca777775..08161b2d9b08 100644
@@ -904,6 +904,8 @@ void __exit fuse_ctl_cleanup(void);
  */
 struct fuse_req *fuse_request_alloc(unsigned npages);
 
+struct fuse_req *fuse_request_alloc_account(unsigned int npages);
+
 struct fuse_req *fuse_request_alloc_nofs(unsigned npages);
 
 bool fuse_req_realloc_pages(struct fuse_conn *fc, struct fuse_req *req,
fs/fuse/inode.c +2 -1
index 4bb885b0f032..eab44ddc68b9 100644
@@ -76,7 +76,8 @@ struct fuse_mount_data {
 
 struct fuse_forget_link *fuse_alloc_forget(void)
 {
-	return kzalloc(sizeof(struct fuse_forget_link), GFP_KERNEL);
+	return kzalloc(sizeof(struct fuse_forget_link),
+		       GFP_KERNEL_ACCOUNT | __GFP_RECLAIMABLE);
 }
 
 static struct inode *fuse_alloc_inode(struct super_block *sb)
Signature · 1 line
2.23.0.187.g17f5b7556c-goog
#2 lore raw

PATCH 3/3 fuse: pass gfp flags to fuse_request_alloc

In reply to Khazhismel Kumykov

Instead of having a helper per flag

---
 fs/fuse/dev.c    | 22 +++-------------------
 fs/fuse/file.c   |  6 +++---
 fs/fuse/fuse_i.h |  6 +-----
 fs/fuse/inode.c  |  4 ++--
 4 files changed, 9 insertions(+), 29 deletions(-)
fs/fuse/dev.c +3 -19
index a0d166a6596f..c957620ce7ba 100644
@@ -66,7 +66,7 @@ static struct page **fuse_req_pages_alloc(unsigned int npages, gfp_t flags,
 	return pages;
 }
 
-static struct fuse_req *__fuse_request_alloc(unsigned npages, gfp_t flags)
+struct fuse_req *fuse_request_alloc(unsigned int npages, gfp_t flags)
 {
 	struct fuse_req *req = kmem_cache_zalloc(fuse_req_cachep, flags);
 	if (req) {
@@ -90,24 +90,8 @@ static struct fuse_req *__fuse_request_alloc(unsigned npages, gfp_t flags)
 	}
 	return req;
 }
-
-struct fuse_req *fuse_request_alloc(unsigned npages)
-{
-	return __fuse_request_alloc(npages, GFP_KERNEL);
-}
 EXPORT_SYMBOL_GPL(fuse_request_alloc);
 
-struct fuse_req *fuse_request_alloc_account(unsigned int npages)
-{
-	return __fuse_request_alloc(npages, GFP_KERNEL_ACCOUNT);
-}
-EXPORT_SYMBOL_GPL(fuse_request_alloc_account);
-
-struct fuse_req *fuse_request_alloc_nofs(unsigned npages)
-{
-	return __fuse_request_alloc(npages, GFP_NOFS);
-}
-
 static void fuse_req_pages_free(struct fuse_req *req)
 {
 	if (req->pages != req->inline_pages)
@@ -207,7 +191,7 @@ static struct fuse_req *__fuse_get_req(struct fuse_conn *fc, unsigned npages,
 	if (fc->conn_error)
 		goto out;
 
-	req = fuse_request_alloc(npages);
+	req = fuse_request_alloc(npages, GFP_KERNEL);
 	err = -ENOMEM;
 	if (!req) {
 		if (for_background)
@@ -316,7 +300,7 @@ struct fuse_req *fuse_get_req_nofail_nopages(struct fuse_conn *fc,
 	wait_event(fc->blocked_waitq, fc->initialized);
 	/* Matches smp_wmb() in fuse_set_initialized() */
 	smp_rmb();
-	req = fuse_request_alloc(0);
+	req = fuse_request_alloc(0, GFP_KERNEL);
 	if (!req)
 		req = get_reserved_req(fc, file);
 
fs/fuse/file.c +3 -3
index c584ad7478b3..ae8c8016bb8e 100644
@@ -50,7 +50,7 @@ struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
 		return NULL;
 
 	ff->fc = fc;
-	ff->reserved_req = fuse_request_alloc_account(0);
+	ff->reserved_req = fuse_request_alloc(0, GFP_KERNEL_ACCOUNT);
 	if (unlikely(!ff->reserved_req)) {
 		kfree(ff);
 		return NULL;
@@ -1703,7 +1703,7 @@ static int fuse_writepage_locked(struct page *page)
 
 	set_page_writeback(page);
 
-	req = fuse_request_alloc_nofs(1);
+	req = fuse_request_alloc(1, GFP_NOFS);
 	if (!req)
 		goto err;
 
@@ -1923,7 +1923,7 @@ static int fuse_writepages_fill(struct page *page,
 		struct fuse_inode *fi = get_fuse_inode(inode);
 
 		err = -ENOMEM;
-		req = fuse_request_alloc_nofs(FUSE_REQ_INLINE_PAGES);
+		req = fuse_request_alloc(FUSE_REQ_INLINE_PAGES, GFP_NOFS);
 		if (!req) {
 			__free_page(tmp_page);
 			goto out_unlock;
fs/fuse/fuse_i.h +1 -5
index 08161b2d9b08..8080a51096e9 100644
@@ -902,11 +902,7 @@ void __exit fuse_ctl_cleanup(void);
 /**
  * Allocate a request
  */
-struct fuse_req *fuse_request_alloc(unsigned npages);
-
-struct fuse_req *fuse_request_alloc_account(unsigned int npages);
-
-struct fuse_req *fuse_request_alloc_nofs(unsigned npages);
+struct fuse_req *fuse_request_alloc(unsigned int npages, gfp_t flags);
 
 bool fuse_req_realloc_pages(struct fuse_conn *fc, struct fuse_req *req,
 			    gfp_t flags);
fs/fuse/inode.c +2 -2
index eab44ddc68b9..ad92e93eaddd 100644
@@ -1178,13 +1178,13 @@ static int fuse_fill_super(struct super_block *sb, void *data, int silent)
 	/* Root dentry doesn't have .d_revalidate */
 	sb->s_d_op = &fuse_dentry_operations;
 
-	init_req = fuse_request_alloc(0);
+	init_req = fuse_request_alloc(0, GFP_KERNEL);
 	if (!init_req)
 		goto err_put_root;
 	__set_bit(FR_BACKGROUND, &init_req->flags);
 
 	if (is_bdev) {
-		fc->destroy_req = fuse_request_alloc(0);
+		fc->destroy_req = fuse_request_alloc(0, GFP_KERNEL);
 		if (!fc->destroy_req)
 			goto err_free_init_req;
 	}
Signature · 1 line
2.23.0.187.g17f5b7556c-goog
#3 lore raw

Re: [PATCH 3/3] fuse: pass gfp flags to fuse_request_alloc

Quoted 4 lines · Khazhismel Kumykov

Instead of having a helper per flag

I think it would be better to re-order the patch 2 and 3 of this series. There will be less code churn.

Quoted 134 lines
---
 fs/fuse/dev.c    | 22 +++-------------------
 fs/fuse/file.c   |  6 +++---
 fs/fuse/fuse_i.h |  6 +-----
 fs/fuse/inode.c  |  4 ++--
 4 files changed, 9 insertions(+), 29 deletions(-)
fs/fuse/dev.c +3 -19
index a0d166a6596f..c957620ce7ba 100644
@@ -66,7 +66,7 @@ static struct page **fuse_req_pages_alloc(unsigned int npages, gfp_t flags,
        return pages;
 }
 
-static struct fuse_req *__fuse_request_alloc(unsigned npages, gfp_t flags)
+struct fuse_req *fuse_request_alloc(unsigned int npages, gfp_t flags)
 {
        struct fuse_req *req = kmem_cache_zalloc(fuse_req_cachep, flags);
        if (req) {
@@ -90,24 +90,8 @@ static struct fuse_req *__fuse_request_alloc(unsigned npages, gfp_t flags)
        }
        return req;
 }
-
-struct fuse_req *fuse_request_alloc(unsigned npages)
-{
-       return __fuse_request_alloc(npages, GFP_KERNEL);
-}
 EXPORT_SYMBOL_GPL(fuse_request_alloc);
 
-struct fuse_req *fuse_request_alloc_account(unsigned int npages)
-{
-       return __fuse_request_alloc(npages, GFP_KERNEL_ACCOUNT);
-}
-EXPORT_SYMBOL_GPL(fuse_request_alloc_account);
-
-struct fuse_req *fuse_request_alloc_nofs(unsigned npages)
-{
-       return __fuse_request_alloc(npages, GFP_NOFS);
-}
-
 static void fuse_req_pages_free(struct fuse_req *req)
 {
        if (req->pages != req->inline_pages)
@@ -207,7 +191,7 @@ static struct fuse_req *__fuse_get_req(struct fuse_conn *fc, unsigned npages,
        if (fc->conn_error)
                goto out;
 
-       req = fuse_request_alloc(npages);
+       req = fuse_request_alloc(npages, GFP_KERNEL);
        err = -ENOMEM;
        if (!req) {
                if (for_background)
@@ -316,7 +300,7 @@ struct fuse_req *fuse_get_req_nofail_nopages(struct fuse_conn *fc,
        wait_event(fc->blocked_waitq, fc->initialized);
        /* Matches smp_wmb() in fuse_set_initialized() */
        smp_rmb();
-       req = fuse_request_alloc(0);
+       req = fuse_request_alloc(0, GFP_KERNEL);
        if (!req)
                req = get_reserved_req(fc, file);
 
fs/fuse/file.c +3 -3
index c584ad7478b3..ae8c8016bb8e 100644
@@ -50,7 +50,7 @@ struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
                return NULL;
 
        ff->fc = fc;
-       ff->reserved_req = fuse_request_alloc_account(0);
+       ff->reserved_req = fuse_request_alloc(0, GFP_KERNEL_ACCOUNT);
        if (unlikely(!ff->reserved_req)) {
                kfree(ff);
                return NULL;
@@ -1703,7 +1703,7 @@ static int fuse_writepage_locked(struct page *page)
 
        set_page_writeback(page);
 
-       req = fuse_request_alloc_nofs(1);
+       req = fuse_request_alloc(1, GFP_NOFS);
        if (!req)
                goto err;
 
@@ -1923,7 +1923,7 @@ static int fuse_writepages_fill(struct page *page,
                struct fuse_inode *fi = get_fuse_inode(inode);
 
                err = -ENOMEM;
-               req = fuse_request_alloc_nofs(FUSE_REQ_INLINE_PAGES);
+               req = fuse_request_alloc(FUSE_REQ_INLINE_PAGES, GFP_NOFS);
                if (!req) {
                        __free_page(tmp_page);
                        goto out_unlock;
fs/fuse/fuse_i.h +1 -5
index 08161b2d9b08..8080a51096e9 100644
@@ -902,11 +902,7 @@ void __exit fuse_ctl_cleanup(void);
 /**
  * Allocate a request
  */
-struct fuse_req *fuse_request_alloc(unsigned npages);
-
-struct fuse_req *fuse_request_alloc_account(unsigned int npages);
-
-struct fuse_req *fuse_request_alloc_nofs(unsigned npages);
+struct fuse_req *fuse_request_alloc(unsigned int npages, gfp_t flags);
 
 bool fuse_req_realloc_pages(struct fuse_conn *fc, struct fuse_req *req,
                            gfp_t flags);
fs/fuse/inode.c +2 -3
index eab44ddc68b9..ad92e93eaddd 100644
@@ -1178,13 +1178,13 @@ static int fuse_fill_super(struct super_block *sb, void *data, int silent)
        /* Root dentry doesn't have .d_revalidate */
        sb->s_d_op = &fuse_dentry_operations;
 
-       init_req = fuse_request_alloc(0);
+       init_req = fuse_request_alloc(0, GFP_KERNEL);
        if (!init_req)
                goto err_put_root;
        __set_bit(FR_BACKGROUND, &init_req->flags);
 
        if (is_bdev) {
-               fc->destroy_req = fuse_request_alloc(0);
+               fc->destroy_req = fuse_request_alloc(0, GFP_KERNEL);
                if (!fc->destroy_req)
                        goto err_free_init_req;
        }
--

2.23.0.187.g17f5b7556c-goog

#4 lore raw

Re: [PATCH 3/3] fuse: pass gfp flags to fuse_request_alloc

Quoted 9 lines · Shakeel Butt
Quoted 4 lines · Khazhismel Kumykov

Instead of having a helper per flag

I think it would be better to re-order the patch 2 and 3 of this series. There will be less code churn.

That makes sense to me, I'll follow up with a v2

Reviewed-by on v1 1/3
#5 lore raw
In reply to Khazhismel Kumykov
Quoted 7 lines · Khazhismel Kumykov

Implements the optimization noted in f75fdf22b0a8 ("fuse: don't use ->d_time"), as the additional memory can be significant. (In particular, on SLAB configurations this 8-byte alloc becomes 32 bytes). Per-dentry, this can consume significant memory.

Actually we are seeing this in production where a job creating a lot of fuse files cause a lot of extra system level overhead.

Quoted 68 lines
---
 fs/fuse/dir.c | 19 +++++++++++++++++++
 1 file changed, 19 insertions(+)
fs/fuse/dir.c +19 -1
index dd0f64f7bc06..f9c59a296568 100644
@@ -24,6 +24,18 @@ static void fuse_advise_use_readdirplus(struct inode *dir)
        set_bit(FUSE_I_ADVISE_RDPLUS, &fi->state);
 }
 
+#if BITS_PER_LONG >= 64
+static inline void fuse_dentry_settime(struct dentry *entry, u64 time)
+{
+       entry->d_fsdata = (void *) time;
+}
+
+static inline u64 fuse_dentry_time(struct dentry *entry)
+{
+       return (u64)entry->d_fsdata;
+}
+
+#else
 union fuse_dentry {
        u64 time;
        struct rcu_head rcu;
@@ -38,6 +50,7 @@ static inline u64 fuse_dentry_time(struct dentry *entry)
 {
        return ((union fuse_dentry *) entry->d_fsdata)->time;
 }
+#endif
 
 /*
  * FUSE caches dentries and attributes with separate timeout.  The
@@ -242,6 +255,7 @@ static int fuse_dentry_revalidate(struct dentry *entry, unsigned int flags)
        goto out;
 }
 
+#if BITS_PER_LONG < 64
 static int fuse_dentry_init(struct dentry *dentry)
 {
        dentry->d_fsdata = kzalloc(sizeof(union fuse_dentry), GFP_KERNEL);
@@ -254,16 +268,21 @@ static void fuse_dentry_release(struct dentry *dentry)
 
        kfree_rcu(fd, rcu);
 }
+#endif
 
 const struct dentry_operations fuse_dentry_operations = {
        .d_revalidate   = fuse_dentry_revalidate,
+#if BITS_PER_LONG < 64
        .d_init         = fuse_dentry_init,
        .d_release      = fuse_dentry_release,
+#endif
 };
 
 const struct dentry_operations fuse_root_dentry_operations = {
+#if BITS_PER_LONG < 64
        .d_init         = fuse_dentry_init,
        .d_release      = fuse_dentry_release,
+#endif
 };
 
 int fuse_valid_type(int m)
--

2.23.0.187.g17f5b7556c-goog