Skip to contentSkip to conversation

Allow to disable shmem.o

PATCH In review Andi Kleen 4 participants 10 messages Aug 7, 2004 – Aug 8, 2004

Subsystems (5)
Memory management VFS MM core Page cache +1 more subsystems
lore
Share & tools
lore

Andi Kleen

author
#0 lore raw

PATCH 1/1 Allow to disable shmem.o

This patch allows to not compile mm/shmem.o in. It is now dependent on CONFIG_TMPFS, which is the main user. This allows shrinking the kernel a bit more when needed. Disabling TMPFS disables the mmap MAP_ANONYMOUS|MAP_SHARED functionality now. That's ok, because Linux 2.2 also didn't have it iirc. To make sure this doesn't happen by accident I made the TMPFS configuration dependent on CONFIG_EMBEDDED. This means a normal configuration enables TMPFS, which is ok since it is only minor additional overhead to shmem.o and also quite useful for shared memory.

Patch for 2.6.7, but applies with some offsets to 2.6.8rc3

-Andi

diff -u linux-2.6.7-boot64/drivers/char/mem.c-SHMEM linux-2.6.7-boot64/drivers/char/mem.c

include/linux/mm.h renamed from include/linux/mm.h-SHMEM +8 -0
diff -u linux-2.6.7-boot64/include/linux/mm.h-SHMEM linux-2.6.7-boot64/include/linux/mm.h
@@ -496,7 +496,15 @@
 					unsigned long addr);
 struct file *shmem_file_setup(char * name, loff_t size, unsigned long flags);
 void shmem_lock(struct file * file, int lock);
+
+#ifdef CONFIG_TMPFS
 int shmem_zero_setup(struct vm_area_struct *);
+#else
+static inline int shmem_zero_setup(struct vm_area_struct *vma)
+{
+	return -EINVAL;
+}
+#endif
 
 /*
  * Parameter block passed down to zap_pte_range in exceptional cases.
include/linux/swap.h renamed from include/linux/swap.h-SHMEM +7 -0
diff -u linux-2.6.7-boot64/include/linux/swap.h-SHMEM linux-2.6.7-boot64/include/linux/swap.h
@@ -177,8 +177,15 @@
 extern int vm_swappiness;
 
 #ifdef CONFIG_MMU
+#ifdef CONFIG_TMPFS
 /* linux/mm/shmem.c */
 extern int shmem_unuse(swp_entry_t entry, struct page *page);
+#else
+static inline int shmem_unuse(swp_entry_t entry, struct page *page)
+{
+	return 0;
+}
+#endif
 #endif /* CONFIG_MMU */
 
 extern void swap_unplug_io_fn(struct backing_dev_info *, struct page *);
fs/Kconfig renamed from fs/Kconfig-SHMEM +2 -1
diff -u linux-2.6.7-boot64/fs/Kconfig-SHMEM linux-2.6.7-boot64/fs/Kconfig
@@ -914,7 +914,8 @@
 	  extended attributes for file security labels, say N.
 
 config TMPFS
-	bool "Virtual memory file system support (former shm fs)"
+	bool "Virtual memory file system support (former shm fs)" if EMBEDDED
+	default y
 	help
 	  Tmpfs is a file system which keeps all files in virtual memory.
mm/Makefile renamed from mm/Makefile-SHMEM +2 -1
diff -u linux-2.6.7-boot64/mm/Makefile-SHMEM linux-2.6.7-boot64/mm/Makefile
@@ -5,13 +5,14 @@
 mmu-y			:= nommu.o
 mmu-$(CONFIG_MMU)	:= fremap.o highmem.o madvise.o memory.o mincore.o \
 			   mlock.o mmap.o mprotect.o mremap.o msync.o rmap.o \
-			   shmem.o vmalloc.o
+			   vmalloc.o
 
 obj-y			:= bootmem.o filemap.o mempool.o oom_kill.o fadvise.o \
 			   page_alloc.o page-writeback.o pdflush.o prio_tree.o \
 			   readahead.o slab.o swap.o truncate.o vmscan.o \
 			   $(mmu-y)
 
+obj-$(CONFIG_TMPFS)	+= shmem.o
 obj-$(CONFIG_SWAP)	+= page_io.o swap_state.o swapfile.o
 obj-$(CONFIG_HUGETLBFS)	+= hugetlb.o
 obj-$(CONFIG_NUMA) 	+= mempolicy.o
#1 lore raw

Please no, not this half-baked patch anyway.

Quoted 3 lines · Andi Kleen

This patch allows to not compile mm/shmem.o in. It is now dependent on CONFIG_TMPFS, which is the main user.

Well, one of its users, yes.

Quoted 1 line

This allows shrinking the kernel a bit more when needed.

Good.

Quoted 3 lines

Disabling TMPFS disables the mmap MAP_ANONYMOUS|MAP_SHARED functionality now. That's ok, because Linux 2.2 also didn't have it iirc.

That's not an argument for throwing out support for POSIX/SuS functionality which 2.4 has supported all along. But never mind, embedded may not need it, and your defaults are the right way round.

Quoted 5 lines

To make sure this doesn't happen by
accident I made the TMPFS configuration dependent on CONFIG_EMBEDDED. This means a normal configuration enables TMPFS, which is ok since it is only minor additional overhead to shmem.o and also quite useful for shared memory.

At the least you need to add a dependency on CONFIG_SYSVIPC. And you shouldn't disable standard functionality without saying so in the Kconfig help.

You're completely changing the meaning and configurability of CONFIG_TMPFS. Perhaps it's not a useful config option any more (too often needed "y", too little saved by "n"), and yours is more useful. But if you want to go this way, you must change its name and Documentation and #ifdefs too.

I do agree that mm/shmem.o is larger (did someone at the back say "bloated"?) than we'd wish, that embedded in particular would like it smaller, and embedded might have no need to support SysV IPC and Posix SHM and shared anonymous and tmpfs.

But I prefer Matt's tiny tiny-shmem.c which does support all those, using ramfs instead, and says in Kconfig what it's doing. But perhaps you're wanting to avoid ramfs too?

A lot (but less than I once imagined) of the overhead in mm/shmem.o does come from its support for swap. I've not been eager to pepper it with #ifdef CONFIG_SWAPs; but if there's a real demand for that I could give it a try (while abstracting away the #ifdefs as much as possible).

Hugh

#2 lore raw
Quoted 4 lines · Hugh Dickins

But I prefer Matt's tiny tiny-shmem.c which does support all those, using ramfs instead, and says in Kconfig what it's doing. But perhaps you're wanting to avoid ramfs too?

Ramfs is hard to avoid as it's used internally for / at boot and so on. My patch for comparison. Comments appreciated.

configurable tiny shmem/tmpfs support

 tiny-mpm/fs/ramfs/inode.c   |    8 +-
 tiny-mpm/include/linux/mm.h |   10 +++
 tiny-mpm/init/Kconfig       |   14 ++++
 tiny-mpm/mm/Makefile        |    4 +
 tiny-mpm/mm/filemap.c       |    2 
 tiny-mpm/mm/tiny-shmem.c    |  128 ++++++++++++++++++++++++++++++++++++++++++++
 6 files changed, 159 insertions(+), 7 deletions(-)
mm/Makefile +4 -1
Index: tiny/mm/Makefile
===================================================================
@@ -5,7 +5,7 @@
 mmu-y			:= nommu.o
 mmu-$(CONFIG_MMU)	:= fremap.o highmem.o madvise.o memory.o mincore.o \
 			   mlock.o mmap.o mprotect.o mremap.o msync.o rmap.o \
-			   shmem.o vmalloc.o
+			   vmalloc.o
 
 obj-y			:= bootmem.o filemap.o mempool.o oom_kill.o fadvise.o \
 			   page_alloc.o page-writeback.o pdflush.o prio_tree.o \

@@ -16,3 +16,6 @@
 obj-$(CONFIG_SWAP)	+= page_io.o swap_state.o swapfile.o
 obj-$(CONFIG_HUGETLBFS)	+= hugetlb.o
 obj-$(CONFIG_NUMA) 	+= mempolicy.o
+obj-$(CONFIG_SHMEM) += shmem.o
+obj-$(CONFIG_TINY_SHMEM) += tiny-shmem.o
+
mm/tiny-shmem.c +128 -0

128 lines not on this page · raw message

mm/filemap.c +1 -1
Index: tiny/mm/filemap.c
===================================================================
@@ -1410,7 +1410,7 @@
 	return 0;
 }
 
-static struct vm_operations_struct generic_file_vm_ops = {
+struct vm_operations_struct generic_file_vm_ops = {
 	.nopage		= filemap_nopage,
 	.populate	= filemap_populate,
 };
include/linux/mm.h +10 -1
Index: tiny/include/linux/mm.h
===================================================================
@@ -489,14 +489,23 @@
 
 extern void show_free_areas(void);
 
+#ifdef CONFIG_SHMEM
 struct page *shmem_nopage(struct vm_area_struct * vma,
 			unsigned long address, int *type);
 int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *new);
 struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
 					unsigned long addr);
-struct file *shmem_file_setup(char * name, loff_t size, unsigned long flags);
 void shmem_lock(struct file * file, int lock);
+#else
+#define shmem_nopage filemap_nopage
+#define shmem_lock(a, b) /* always in memory, no need to lock */
+#define shmem_set_policy(a, b) (0)
+#define shmem_get_policy(a, b) (NULL)
+#endif
+
 int shmem_zero_setup(struct vm_area_struct *);
+struct file *shmem_file_setup(char * name, loff_t size, unsigned long flags);
+
 
 /*
  * Parameter block passed down to zap_pte_range in exceptional cases.
fs/ramfs/inode.c +4 -4
Index: tiny/fs/ramfs/inode.c
===================================================================
@@ -39,7 +39,7 @@
 
 static struct super_operations ramfs_ops;
 static struct address_space_operations ramfs_aops;
-static struct file_operations ramfs_file_operations;
+extern struct file_operations ramfs_file_operations;
 static struct inode_operations ramfs_file_inode_operations;
 static struct inode_operations ramfs_dir_inode_operations;
 

@@ -48,7 +48,7 @@
 	.memory_backed	= 1,	/* Does not contribute to dirty memory */
 };
 
-static struct inode *ramfs_get_inode(struct super_block *sb, int mode, dev_t dev)
+struct inode *ramfs_get_inode(struct super_block *sb, int mode, dev_t dev)
 {
 	struct inode * inode = new_inode(sb);
 

@@ -146,7 +146,7 @@
 	.commit_write	= simple_commit_write
 };
 
-static struct file_operations ramfs_file_operations = {
+struct file_operations ramfs_file_operations = {
 	.read		= generic_file_read,
 	.write		= generic_file_write,
 	.mmap		= generic_file_mmap,

@@ -199,7 +199,7 @@
 	return 0;
 }
 
-static struct super_block *ramfs_get_sb(struct file_system_type *fs_type,
+struct super_block *ramfs_get_sb(struct file_system_type *fs_type,
 	int flags, const char *dev_name, void *data)
 {
 	return get_sb_nodev(fs_type, flags, data, ramfs_fill_super);
init/Kconfig +14 -0
Index: tiny/init/Kconfig
===================================================================
@@ -461,8 +461,22 @@
 	help
 	  Disabling this moves operations on semaphores out of line.
 
+config SHMEM
+	default y
+	bool "Use full shmem filesystem" if EMBEDDED && MMU
+	help
+	  The shmem is an internal filesystem used to manage shared memory.
+	  It is backed by swap and manages resource limits. It is also exported
+	  to userspace as tmpfs if TMPFS is enabled. Disabling this
+	  option replaces shmem and tmpfs with the much simpler ramfs code,
+	  which may be appropriate on small systems without swap.
+
 endmenu		# General setup
 
+config TINY_SHMEM
+	default !SHMEM
+	bool
+
 config CRC32_CALC
 	default !CRC32_TABLES
 	bool
Signature · 1 line
Mathematics is the supreme nostalgia of our time.
author
#3 lore raw

mm/tiny-shmem.c line 27

Quoted 5 lines · Matt Mackall
mm/tiny-shmem.c +5 -0
+extern struct inode *ramfs_get_inode(struct super_block *sb, int mode, dev_t dev);
+extern struct super_block *ramfs_get_sb(struct file_system_type *fs_type,
+	 int flags, const char *dev_name, void *data);
+extern struct file_operations ramfs_file_operations;
+extern struct vm_operations_struct generic_file_vm_ops;

This should be all in header files.

mm/tiny-shmem.c line 64

Quoted 11 lines

+ */

mm/tiny-shmem.c +10 -0
+struct file *shmem_file_setup(char *name, loff_t size, unsigned long flags)
+{
+	int error;
+	struct file *file;
+	struct inode *inode;
+	struct dentry *dentry, *root;
+	struct qstr this;
+
+	if (IS_ERR(shm_mnt))
+		return (void *)shm_mnt;

Why this strange cast?

Quoted 7 lines
mm/tiny-shmem.c +7 -0
+
+	error = -ENOMEM;
+	this.name = name;
+	this.len = strlen(name);
+	this.hash = 0; /* will go */
+	root = shm_mnt->mnt_root;
+	dentry = d_alloc(root, &this);

mm/tiny-shmem.c line 81

Quoted 10 lines
mm/tiny-shmem.c +10 -0
+	if (!dentry)
+		goto put_memory;
+
+	error = -ENFILE;
+	file = get_empty_filp();
+	if (!file)
+		goto put_dentry;
+
+	error = -ENOSPC;
+	inode = ramfs_get_inode(root->d_sb, S_IFREG | S_IRWXUGO, 0);

Hmm, won't this allow everybody else to open it in /proc/pid/fd/ ? (existing shmem.c seems to use it too, but it looks a bit bogus)

-AndI

#4 lore raw
Quoted 8 lines · Andi Kleen

mm/tiny-shmem.c line 27

Quoted 5 lines · Matt Mackall
mm/tiny-shmem.c +5 -0
+extern struct inode *ramfs_get_inode(struct super_block *sb, int mode, dev_t dev);
+extern struct super_block *ramfs_get_sb(struct file_system_type *fs_type,
+	 int flags, const char *dev_name, void *data);
+extern struct file_operations ramfs_file_operations;
+extern struct vm_operations_struct generic_file_vm_ops;

This should be all in header files.

Yes.

Quoted 4 lines

mm/tiny-shmem.c line 64

Quoted 2 lines
mm/tiny-shmem.c +2 -0
+	if (IS_ERR(shm_mnt))
+		return (void *)shm_mnt;

Why this strange cast?

Matt copied from shmem.c, so blame me. Strange cast because it's one of those IS_ERR things, an -errno hiding inside a pointer, and it's being propagated from an initialization error on shm_mnt to a runtime error on any struct file for that mount. I wrote (void *) because it seemed to express the ambiguity better than pretending struct file *.

Quoted 4 lines

mm/tiny-shmem.c line 81

Quoted 1 line
mm/tiny-shmem.c +1 -0
+	inode = ramfs_get_inode(root->d_sb, S_IFREG | S_IRWXUGO, 0);

Hmm, won't this allow everybody else to open it in /proc/pid/fd/ ? (existing shmem.c seems to use it too, but it looks a bit bogus)

But there's no fd associated with it?

Hugh

#5 lore raw
In reply to Matt Mackall
Quoted 7 lines · Matt Mackall
Quoted 4 lines · Hugh Dickins

But I prefer Matt's tiny tiny-shmem.c which does support all those, using ramfs instead, and says in Kconfig what it's doing. But perhaps you're wanting to avoid ramfs too?

Ramfs is hard to avoid as it's used internally for / at boot and so on.

Ah, right.

Quoted 1 line

My patch for comparison. Comments appreciated.

Looks pretty good, but would need rediffing against mainline or -mm if heading that way: the init/Kconfig patch is peculiar to your -tiny.

The comment at the head of tiny-shmem.c:

Quoted 4 lines

+ * This is intended for small system where the benefits of the full
+ * shmem code (swap-backed and resource-limited) are outweighed by their
+ * complexity. On systems without swap and not using userspace /dev/shm,
+ * this code should be effectively equivalent, but much lighter weight.

Very fair statement. But isn't it actually supporting /dev/shm fine?

And your shmem_file_setup hasn't quite kept up with the times:
shmat oopses in vma_link because shmem_file_setup is lacking a

	file->f_mapping = inode->i_mapping;

Plus Andi rightly advises externs in a header file.

Hugh

author
#6 lore raw
In reply to Hugh Dickins
Quoted 9 lines · Hugh Dickins
Quoted 5 lines

To make sure this doesn't happen by
accident I made the TMPFS configuration dependent on CONFIG_EMBEDDED. This means a normal configuration enables TMPFS, which is ok since it is only minor additional overhead to shmem.o and also quite useful for shared memory.

At the least you need to add a dependency on CONFIG_SYSVIPC. And you shouldn't disable standard functionality without saying so in the Kconfig help.

Hmm, right. Appended is a new patch that adds a select TMPFS to SYSVIPC. i suspect the standard functionality is already covered by CONFIG_EMBEDDED.

Quoted 5 lines

You're completely changing the meaning and configurability of CONFIG_TMPFS. Perhaps it's not a useful config option any more (too often needed "y", too little saved by "n"), and yours is more useful. But if you want to go this way, you must change its name and Documentation and #ifdefs too.

Perhaps, but calling it CONFIG_TMPFS is not too bad.

Quoted 8 lines

I do agree that mm/shmem.o is larger (did someone at the back say "bloated"?) than we'd wish, that embedded in particular would like it smaller, and embedded might have no need to support SysV IPC and Posix SHM and shared anonymous and tmpfs.

But I prefer Matt's tiny tiny-shmem.c which does support all those, using ramfs instead, and says in Kconfig what it's doing. But perhaps you're wanting to avoid ramfs too?

Makes sense, adding this may be an additional useful step. For the beginning I think my patch is fine though.

RAMFS cannot be turned off right now. I tried.

Quoted 6 lines

A lot (but less than I once imagined) of the overhead in mm/shmem.o does come from its support for swap. I've not been eager to pepper it with #ifdef CONFIG_SWAPs; but if there's a real demand for that I could give it a try (while abstracting away the #ifdefs as much as possible).

I personally don't want to turn off CONFIG_SWAP. I doubt my architecture would even compile without that.

Andrew, please consider merging this version.

-Andi

-------------------------------------------------------

Allow to disable shmem.o by disabling CONFIG_TMPFS. This will disable MAP_ANONYMOUS|MAP_SHARED. To make sure this doesn't happen by mistake make tmpfs dependent on CONFIG_EMBEDDED now.

diff -u linux-2.6.7-boot64/drivers/char/mem.c-SHMEM linux-2.6.7-boot64/drivers/char/mem.c

include/linux/mm.h renamed from include/linux/mm.h-SHMEM +8 -0
diff -u linux-2.6.7-boot64/include/linux/mm.h-SHMEM linux-2.6.7-boot64/include/linux/mm.h
@@ -496,7 +496,15 @@
 					unsigned long addr);
 struct file *shmem_file_setup(char * name, loff_t size, unsigned long flags);
 void shmem_lock(struct file * file, int lock);
+
+#ifdef CONFIG_TMPFS
 int shmem_zero_setup(struct vm_area_struct *);
+#else
+static inline int shmem_zero_setup(struct vm_area_struct *vma)
+{
+	return -EINVAL;
+}
+#endif
 
 /*
  * Parameter block passed down to zap_pte_range in exceptional cases.
include/linux/swap.h renamed from include/linux/swap.h-SHMEM +7 -0
diff -u linux-2.6.7-boot64/include/linux/swap.h-SHMEM linux-2.6.7-boot64/include/linux/swap.h
@@ -177,8 +177,15 @@
 extern int vm_swappiness;
 
 #ifdef CONFIG_MMU
+#ifdef CONFIG_TMPFS
 /* linux/mm/shmem.c */
 extern int shmem_unuse(swp_entry_t entry, struct page *page);
+#else
+static inline int shmem_unuse(swp_entry_t entry, struct page *page)
+{
+	return 0;
+}
+#endif
 #endif /* CONFIG_MMU */
 
 extern void swap_unplug_io_fn(struct backing_dev_info *, struct page *);
fs/Kconfig renamed from fs/Kconfig-SHMEM +2 -1
diff -u linux-2.6.7-boot64/fs/Kconfig-SHMEM linux-2.6.7-boot64/fs/Kconfig
@@ -914,7 +914,8 @@
 	  extended attributes for file security labels, say N.
 
 config TMPFS
-	bool "Virtual memory file system support (former shm fs)"
+	bool "Virtual memory file system support (former shm fs)" if EMBEDDED
+	default y
 	help
 	  Tmpfs is a file system which keeps all files in virtual memory.
mm/Makefile renamed from mm/Makefile-SHMEM +2 -1
diff -u linux-2.6.7-boot64/mm/Makefile-SHMEM linux-2.6.7-boot64/mm/Makefile
@@ -5,13 +5,14 @@
 mmu-y			:= nommu.o
 mmu-$(CONFIG_MMU)	:= fremap.o highmem.o madvise.o memory.o mincore.o \
 			   mlock.o mmap.o mprotect.o mremap.o msync.o rmap.o \
-			   shmem.o vmalloc.o
+			   vmalloc.o
 
 obj-y			:= bootmem.o filemap.o mempool.o oom_kill.o fadvise.o \
 			   page_alloc.o page-writeback.o pdflush.o prio_tree.o \
 			   readahead.o slab.o swap.o truncate.o vmscan.o \
 			   $(mmu-y)
 
+obj-$(CONFIG_TMPFS)	+= shmem.o
 obj-$(CONFIG_SWAP)	+= page_io.o swap_state.o swapfile.o
 obj-$(CONFIG_HUGETLBFS)	+= hugetlb.o
 obj-$(CONFIG_NUMA) 	+= mempolicy.o

diff -u linux-2.6.7-boot64/mm/mmap.c-SHMEM linux-2.6.7-boot64/mm/mmap.c diff -u linux-2.6.7-boot64/mm/swapfile.c-SHMEM linux-2.6.7-boot64/mm/swapfile.c

init/Kconfig renamed from init/Kconfig-SHMEM +1 -0
diff -u linux-2.6.7-boot64/init/Kconfig-SHMEM linux-2.6.7-boot64/init/Kconfig
@@ -77,6 +77,7 @@
 
 config SYSVIPC
 	bool "System V IPC"
+	select TMPFS
 	---help---
 	  Inter Process Communication is a suite of library functions and
 	  system calls which let processes (running programs) synchronize and
#7 lore raw
Quoted 1 line · Andi Kleen

i suspect the standard functionality is already covered by CONFIG_EMBEDDED.

It would still be appropriate to say what functionality is lost by configuring out, rather than leaving everyone to guess.

Quoted 7 lines
Quoted 5 lines

You're completely changing the meaning and configurability of CONFIG_TMPFS. Perhaps it's not a useful config option any more (too often needed "y", too little saved by "n"), and yours is more useful. But if you want to go this way, you must change its name and Documentation and #ifdefs too.

Perhaps, but calling it CONFIG_TMPFS is not too bad.

Okay, so I come along to clean up after you. Yes, we could work it that way. But somehow I still prefer Matt's patch, which offers a lot more.

Hugh

#8 lore raw
Quoted 2 lines · Hugh Dickins

But somehow I still prefer Matt's patch, which offers a lot more.

Me too - it's sneaky.

I suspect it's more SMp scalable than shmem.c too?

#9 lore raw
Quoted 7 lines · Andrew Morton
Quoted 2 lines · Hugh Dickins

But somehow I still prefer Matt's patch, which offers a lot more.

Me too - it's sneaky.

I suspect it's more SMp scalable than shmem.c too?

I bet it is (unless it falls into generic filepage overheads which shmem is skating past e.g. attempts to readahead).

Though I hope not notably more so by the time I've finished reworking shmem_getpage (was on the right track yesterday, but other issues have intervened today).

They should both come down to find_get_page (find_lock_page in shmem.c's case, only an issue when contention on same page), alloc_page, add_to_page_cache: contention on tree_lock.

Its main deficiency (aside from the lack of swap use, which in some contexts is a plus) is its lack of resource limiting.

Hugh