On Mon, 27 May 2024 at 12:20, Linus Torvalds
<torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>
> With just a couple of helpers it would be mostly a cleanup.
Looking around, I think we could even take four bits. Without any
debugging, 'struct file' is 160 bytes on 32-bit, it would not hurt at
all to just force a 16-byte alignment.
In fact, in practice it is aligned more than that - we already use
SLAB_HWCACHE_ALIGN, which in most cases means that it's 64-byte
aligned.
But to be safe, we should specify the 16 bytes in the
kmem_cache_create() call, and we should just make this all very
explicit:
--- a/fs/file_table.c
+++ b/fs/file_table.c
@@ -512,7 +512,7 @@ EXPORT_SYMBOL(__fput_sync);
void __init files_init(void)
{
- filp_cachep = kmem_cache_create("filp", sizeof(struct file), 0,
+ filp_cachep = kmem_cache_create("filp", sizeof(struct file), 16,
SLAB_TYPESAFE_BY_RCU | SLAB_HWCACHE_ALIGN |
SLAB_PANIC | SLAB_ACCOUNT, NULL);
percpu_counter_init(&nr_files, 0, GFP_KERNEL);
--- a/include/linux/fs.h
+++ b/include/linux/fs.h
@@ -1025,7 +1025,7 @@ struct file {
errseq_t f_wb_err;
errseq_t f_sb_err; /* for syncfs */
} __randomize_layout
- __attribute__((aligned(4))); /* lest something weird
decides that 2 is OK */
+ __attribute__((aligned(16))); /* Up to four tag bits */
struct file_handle {
__u32 handle_bytes;
and while four tag bits isn't something to waste, it looks pretty
reasonable here.
Linus
