From: "Darrick J. Wong" <djwong@kernel.org>
To: djwong@kernel.org, cem@kernel.org
Cc: Christoph Hellwig <hch@lst.de>,
cmaiolino@redhat.com, linux-xfs@vger.kernel.org,
hch@infradead.org
Subject: [PATCH 092/111] libxfs: support in-memory buffer cache targets
Date: Mon, 15 Apr 2024 18:01:10 -0700 [thread overview]
Message-ID: <171322883539.211103.15663396557404367362.stgit@frogsfrogsfrogs> (raw)
In-Reply-To: <171322882240.211103.3776766269442402814.stgit@frogsfrogsfrogs>
From: Darrick J. Wong <djwong@kernel.org>
Allow the buffer cache to target in-memory files by connecting it to
xfiles. The next few patches will enable creating xfs_btrees in memory.
Unlike the kernel version of this patch, we use a partitioned xfile to
avoid overflowing the fd table instead of opening a separate memfd for
each target.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
---
libxfs/Makefile | 4 +
libxfs/buf_mem.c | 235 ++++++++++++++++++++++++++++++++++++++++++++++++++++
libxfs/buf_mem.h | 26 ++++++
libxfs/init.c | 4 +
libxfs/libxfs_io.h | 22 +++++
libxfs/rdwr.c | 41 ++++-----
6 files changed, 310 insertions(+), 22 deletions(-)
create mode 100644 libxfs/buf_mem.c
create mode 100644 libxfs/buf_mem.h
diff --git a/libxfs/Makefile b/libxfs/Makefile
index 43e8ae183229..8dc9a79059ed 100644
--- a/libxfs/Makefile
+++ b/libxfs/Makefile
@@ -26,6 +26,7 @@ HFILES = \
libxfs_priv.h \
linux-err.h \
topology.h \
+ buf_mem.h \
xfile.h \
xfs_ag_resv.h \
xfs_alloc.h \
@@ -58,7 +59,8 @@ HFILES = \
xfs_trans_space.h \
xfs_dir2_priv.h
-CFILES = cache.c \
+CFILES = buf_mem.c \
+ cache.c \
defer_item.c \
init.c \
kmem.c \
diff --git a/libxfs/buf_mem.c b/libxfs/buf_mem.c
new file mode 100644
index 000000000000..7c8fa1d2cdcd
--- /dev/null
+++ b/libxfs/buf_mem.c
@@ -0,0 +1,235 @@
+// SPDX-License-Identifier: GPL-2.0-or-later
+/*
+ * Copyright (c) 2023-2024 Oracle. All Rights Reserved.
+ * Author: Darrick J. Wong <djwong@kernel.org>
+ */
+#include "libxfs_priv.h"
+#include "libxfs.h"
+#include "libxfs/xfile.h"
+#include "libxfs/buf_mem.h"
+#include <sys/mman.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+
+/*
+ * Buffer Cache for In-Memory Files
+ * ================================
+ *
+ * Offline fsck wants to create ephemeral ordered recordsets. The existing
+ * btree infrastructure can do this, but we need the buffer cache to target
+ * memory instead of block devices.
+ *
+ * xfiles meet those requirements. Therefore, the xmbuf mechanism uses a
+ * partition on an xfile to store the staging data.
+ *
+ * xmbufs assume that the caller will handle all required concurrency
+ * management. The resulting xfs_buf objects are kept private to the xmbuf
+ * (they are not recycled to the LRU) because b_addr is mapped directly to the
+ * memfd file.
+ *
+ * The only supported block size is the system page size.
+ */
+
+/* Figure out the xfile buffer cache block size here */
+unsigned int XMBUF_BLOCKSIZE;
+unsigned int XMBUF_BLOCKSHIFT;
+
+void
+xmbuf_libinit(void)
+{
+ long ret = sysconf(_SC_PAGESIZE);
+
+ /* If we don't find a power-of-two page size, go with 4k. */
+ if (ret < 0 || !is_power_of_2(ret))
+ ret = 4096;
+
+ XMBUF_BLOCKSIZE = ret;
+ XMBUF_BLOCKSHIFT = libxfs_highbit32(XMBUF_BLOCKSIZE);
+}
+
+/* Allocate a new cache node (aka a xfs_buf) */
+static struct cache_node *
+xmbuf_cache_alloc(
+ cache_key_t key)
+{
+ struct xfs_bufkey *bufkey = (struct xfs_bufkey *)key;
+ struct xfs_buf *bp;
+ int error;
+
+ bp = kmem_cache_zalloc(xfs_buf_cache, 0);
+ if (!bp)
+ return NULL;
+
+ bp->b_cache_key = bufkey->blkno;
+ bp->b_length = bufkey->bblen;
+ bp->b_target = bufkey->buftarg;
+ bp->b_mount = bufkey->buftarg->bt_mount;
+
+ pthread_mutex_init(&bp->b_lock, NULL);
+ INIT_LIST_HEAD(&bp->b_li_list);
+ bp->b_maps = &bp->__b_map;
+
+ bp->b_nmaps = 1;
+ bp->b_maps[0].bm_bn = bufkey->blkno;
+ bp->b_maps[0].bm_len = bp->b_length;
+
+ error = xmbuf_map_page(bp);
+ if (error) {
+ fprintf(stderr,
+ _("%s: %s can't mmap %u bytes at xfile offset %llu: %s\n"),
+ progname, __FUNCTION__, BBTOB(bp->b_length),
+ (unsigned long long)BBTOB(bufkey->blkno),
+ strerror(error));
+
+ kmem_cache_free(xfs_buf_cache, bp);
+ return NULL;
+ }
+
+ return &bp->b_node;
+}
+
+/* Flush a buffer to disk before purging the cache node */
+static int
+xmbuf_cache_flush(
+ struct cache_node *node)
+{
+ /* direct mapped buffers do not need writing */
+ return 0;
+}
+
+/* Release resources, free the buffer. */
+static void
+xmbuf_cache_relse(
+ struct cache_node *node)
+{
+ struct xfs_buf *bp;
+
+ bp = container_of(node, struct xfs_buf, b_node);
+ xmbuf_unmap_page(bp);
+ kmem_cache_free(xfs_buf_cache, bp);
+}
+
+/* Release a bunch of buffers */
+static unsigned int
+xmbuf_cache_bulkrelse(
+ struct cache *cache,
+ struct list_head *list)
+{
+ struct cache_node *cn, *n;
+ int count = 0;
+
+ if (list_empty(list))
+ return 0;
+
+ list_for_each_entry_safe(cn, n, list, cn_mru) {
+ xmbuf_cache_relse(cn);
+ count++;
+ }
+
+ return count;
+}
+
+static struct cache_operations xmbuf_bcache_operations = {
+ .hash = libxfs_bhash,
+ .alloc = xmbuf_cache_alloc,
+ .flush = xmbuf_cache_flush,
+ .relse = xmbuf_cache_relse,
+ .compare = libxfs_bcompare,
+ .bulkrelse = xmbuf_cache_bulkrelse
+};
+
+/*
+ * Allocate a buffer cache target for a memory-backed file and set up the
+ * buffer target.
+ */
+int
+xmbuf_alloc(
+ struct xfs_mount *mp,
+ const char *descr,
+ unsigned long long maxpos,
+ struct xfs_buftarg **btpp)
+{
+ struct xfs_buftarg *btp;
+ struct xfile *xfile;
+ struct cache *cache;
+ int error;
+
+ btp = kzalloc(sizeof(*btp), GFP_KERNEL);
+ if (!btp)
+ return -ENOMEM;
+
+ error = xfile_create(descr, maxpos, &xfile);
+ if (error)
+ goto out_btp;
+
+ cache = cache_init(0, LIBXFS_BHASHSIZE(NULL), &xmbuf_bcache_operations);
+ if (!cache) {
+ error = -ENOMEM;
+ goto out_xfile;
+ }
+
+ /* Initialize buffer target */
+ btp->bt_mount = mp;
+ btp->bt_bdev = (dev_t)-1;
+ btp->bt_bdev_fd = -1;
+ btp->bt_xfile = xfile;
+ btp->bcache = cache;
+
+ error = pthread_mutex_init(&btp->lock, NULL);
+ if (error)
+ goto out_cache;
+
+ *btpp = btp;
+ return 0;
+
+out_cache:
+ cache_destroy(cache);
+out_xfile:
+ xfile_destroy(xfile);
+out_btp:
+ kfree(btp);
+ return error;
+}
+
+/* Free a buffer cache target for a memory-backed file. */
+void
+xmbuf_free(
+ struct xfs_buftarg *btp)
+{
+ ASSERT(xfs_buftarg_is_mem(btp));
+
+ cache_destroy(btp->bcache);
+ pthread_mutex_destroy(&btp->lock);
+ xfile_destroy(btp->bt_xfile);
+ kfree(btp);
+}
+
+/* Directly map a memfd page into the buffer cache. */
+int
+xmbuf_map_page(
+ struct xfs_buf *bp)
+{
+ struct xfile *xfile = bp->b_target->bt_xfile;
+ void *p;
+ loff_t pos;
+
+ pos = xfile->partition_pos + BBTOB(xfs_buf_daddr(bp));
+ p = mmap(NULL, BBTOB(bp->b_length), PROT_READ | PROT_WRITE, MAP_SHARED,
+ xfile->fcb->fd, pos);
+ if (p == MAP_FAILED)
+ return -errno;
+
+ bp->b_addr = p;
+ bp->b_flags |= LIBXFS_B_UPTODATE | LIBXFS_B_UNCHECKED;
+ bp->b_error = 0;
+ return 0;
+}
+
+/* Unmap a memfd page that was mapped into the buffer cache. */
+void
+xmbuf_unmap_page(
+ struct xfs_buf *bp)
+{
+ munmap(bp->b_addr, BBTOB(bp->b_length));
+ bp->b_addr = NULL;
+}
diff --git a/libxfs/buf_mem.h b/libxfs/buf_mem.h
new file mode 100644
index 000000000000..d2be2c4240b6
--- /dev/null
+++ b/libxfs/buf_mem.h
@@ -0,0 +1,26 @@
+// SPDX-License-Identifier: GPL-2.0-or-later
+/*
+ * Copyright (c) 2023-2024 Oracle. All Rights Reserved.
+ * Author: Darrick J. Wong <djwong@kernel.org>
+ */
+#ifndef __XFS_BUF_MEM_H__
+#define __XFS_BUF_MEM_H__
+
+extern unsigned int XMBUF_BLOCKSIZE;
+extern unsigned int XMBUF_BLOCKSHIFT;
+
+void xmbuf_libinit(void);
+
+static inline bool xfs_buftarg_is_mem(const struct xfs_buftarg *target)
+{
+ return target->bt_xfile != NULL;
+}
+
+int xmbuf_alloc(struct xfs_mount *mp, const char *descr,
+ unsigned long long maxpos, struct xfs_buftarg **btpp);
+void xmbuf_free(struct xfs_buftarg *btp);
+
+int xmbuf_map_page(struct xfs_buf *bp);
+void xmbuf_unmap_page(struct xfs_buf *bp);
+
+#endif /* __XFS_BUF_MEM_H__ */
diff --git a/libxfs/init.c b/libxfs/init.c
index f002dc93cd56..f5cd85655cf0 100644
--- a/libxfs/init.c
+++ b/libxfs/init.c
@@ -22,6 +22,8 @@
#include "xfs_rmap_btree.h"
#include "xfs_refcount_btree.h"
#include "libfrog/platform.h"
+#include "libxfs/xfile.h"
+#include "libxfs/buf_mem.h"
#include "xfs_format.h"
#include "xfs_da_format.h"
@@ -253,6 +255,7 @@ int
libxfs_init(struct libxfs_init *a)
{
xfs_check_ondisk_structs();
+ xmbuf_libinit();
rcu_init();
rcu_register_thread();
radix_tree_init();
@@ -463,6 +466,7 @@ libxfs_buftarg_alloc(
btp->bt_mount = mp;
btp->bt_bdev = dev->dev;
btp->bt_bdev_fd = dev->fd;
+ btp->bt_xfile = NULL;
btp->flags = 0;
if (write_fails) {
btp->writes_left = write_fails;
diff --git a/libxfs/libxfs_io.h b/libxfs/libxfs_io.h
index 7877e17685b8..ae3c4a9484c7 100644
--- a/libxfs/libxfs_io.h
+++ b/libxfs/libxfs_io.h
@@ -27,6 +27,7 @@ struct xfs_buftarg {
unsigned long writes_left;
dev_t bt_bdev;
int bt_bdev_fd;
+ struct xfile *bt_xfile;
unsigned int flags;
struct cache *bcache; /* buffer cache */
};
@@ -58,6 +59,27 @@ xfs_buftarg_trip_write(
void libxfs_buftarg_init(struct xfs_mount *mp, struct libxfs_init *xi);
int libxfs_blkdev_issue_flush(struct xfs_buftarg *btp);
+/*
+ * The bufkey is used to pass the new buffer information to the cache object
+ * allocation routine. Because discontiguous buffers need to pass different
+ * information, we need fields to pass that information. However, because the
+ * blkno and bblen is needed for the initial cache entry lookup (i.e. for
+ * bcompare) the fact that the map/nmaps is non-null to switch to discontiguous
+ * buffer initialisation instead of a contiguous buffer.
+ */
+struct xfs_bufkey {
+ struct xfs_buftarg *buftarg;
+ xfs_daddr_t blkno;
+ unsigned int bblen;
+ struct xfs_buf_map *map;
+ int nmaps;
+};
+
+/* for buf_mem.c only: */
+unsigned int libxfs_bhash(cache_key_t key, unsigned int hashsize,
+ unsigned int hashshift);
+int libxfs_bcompare(struct cache_node *node, cache_key_t key);
+
#define LIBXFS_BBTOOFF64(bbs) (((xfs_off_t)(bbs)) << BBSHIFT)
#define XB_PAGES 2
diff --git a/libxfs/rdwr.c b/libxfs/rdwr.c
index cf986a7e7820..50760cd866e3 100644
--- a/libxfs/rdwr.c
+++ b/libxfs/rdwr.c
@@ -18,7 +18,8 @@
#include "xfs_inode.h"
#include "xfs_trans.h"
#include "libfrog/platform.h"
-
+#include "libxfs/xfile.h"
+#include "libxfs/buf_mem.h"
#include "libxfs.h"
static void libxfs_brelse(struct cache_node *node);
@@ -69,6 +70,9 @@ libxfs_device_zero(struct xfs_buftarg *btp, xfs_daddr_t start, uint len)
char *z;
int error;
+ if (xfs_buftarg_is_mem(btp))
+ return -EOPNOTSUPP;
+
start_offset = LIBXFS_BBTOOFF64(start);
/* try to use special zeroing methods, fall back to writes if needed */
@@ -167,26 +171,10 @@ static struct cache_mru xfs_buf_freelist =
{{&xfs_buf_freelist.cm_list, &xfs_buf_freelist.cm_list},
0, PTHREAD_MUTEX_INITIALIZER };
-/*
- * The bufkey is used to pass the new buffer information to the cache object
- * allocation routine. Because discontiguous buffers need to pass different
- * information, we need fields to pass that information. However, because the
- * blkno and bblen is needed for the initial cache entry lookup (i.e. for
- * bcompare) the fact that the map/nmaps is non-null to switch to discontiguous
- * buffer initialisation instead of a contiguous buffer.
- */
-struct xfs_bufkey {
- struct xfs_buftarg *buftarg;
- xfs_daddr_t blkno;
- unsigned int bblen;
- struct xfs_buf_map *map;
- int nmaps;
-};
-
/* 2^63 + 2^61 - 2^57 + 2^54 - 2^51 - 2^18 + 1 */
#define GOLDEN_RATIO_PRIME 0x9e37fffffffc0001UL
#define CACHE_LINE_SIZE 64
-static unsigned int
+unsigned int
libxfs_bhash(cache_key_t key, unsigned int hashsize, unsigned int hashshift)
{
uint64_t hashval = ((struct xfs_bufkey *)key)->blkno;
@@ -197,7 +185,7 @@ libxfs_bhash(cache_key_t key, unsigned int hashsize, unsigned int hashshift)
return tmp % hashsize;
}
-static int
+int
libxfs_bcompare(
struct cache_node *node,
cache_key_t key)
@@ -231,6 +219,8 @@ static void
__initbuf(struct xfs_buf *bp, struct xfs_buftarg *btp, xfs_daddr_t bno,
unsigned int bytes)
{
+ ASSERT(!xfs_buftarg_is_mem(btp));
+
bp->b_flags = 0;
bp->b_cache_key = bno;
bp->b_length = BTOBB(bytes);
@@ -577,7 +567,6 @@ libxfs_balloc(
return &bp->b_node;
}
-
static int
__read_buf(int fd, void *buf, int len, off_t offset, int flags)
{
@@ -607,6 +596,9 @@ libxfs_readbufr(struct xfs_buftarg *btp, xfs_daddr_t blkno, struct xfs_buf *bp,
ASSERT(len <= bp->b_length);
+ if (xfs_buftarg_is_mem(btp))
+ return 0;
+
error = __read_buf(fd, bp->b_addr, bytes, LIBXFS_BBTOOFF64(blkno), flags);
if (!error &&
bp->b_target == btp &&
@@ -639,6 +631,9 @@ libxfs_readbufr_map(struct xfs_buftarg *btp, struct xfs_buf *bp, int flags)
void *buf;
int i;
+ if (xfs_buftarg_is_mem(btp))
+ return 0;
+
buf = bp->b_addr;
for (i = 0; i < bp->b_nmaps; i++) {
off_t offset = LIBXFS_BBTOOFF64(bp->b_maps[i].bm_bn);
@@ -857,7 +852,9 @@ libxfs_bwrite(
}
}
- if (!(bp->b_flags & LIBXFS_B_DISCONTIG)) {
+ if (xfs_buftarg_is_mem(bp->b_target)) {
+ bp->b_error = 0;
+ } else if (!(bp->b_flags & LIBXFS_B_DISCONTIG)) {
bp->b_error = __write_buf(fd, bp->b_addr, BBTOB(bp->b_length),
LIBXFS_BBTOOFF64(xfs_buf_daddr(bp)),
bp->b_flags);
@@ -917,6 +914,8 @@ libxfs_buf_prepare_mru(
xfs_perag_put(bp->b_pag);
bp->b_pag = NULL;
+ ASSERT(!xfs_buftarg_is_mem(btp));
+
if (!(bp->b_flags & LIBXFS_B_DIRTY))
return;
next prev parent reply other threads:[~2024-04-16 1:01 UTC|newest]
Thread overview: 36+ messages / expand[flat|nested] mbox.gz Atom feed top
2024-04-16 0:51 [PATCHBOMB v3] xfsprogs: everything headed towards 6.9 Darrick J. Wong
2024-04-16 0:57 ` [PATCHSET 1/4] xfsprogs: bug fixes for 6.8 Darrick J. Wong
2024-04-16 0:58 ` [PATCH 1/5] xfs_repair: double-check with shortform attr verifiers Darrick J. Wong
2024-04-16 0:59 ` [PATCH 2/5] xfs_db: improve number extraction in getbitval Darrick J. Wong
2024-04-16 4:53 ` Christoph Hellwig
2024-04-16 0:59 ` [PATCH 3/5] xfs_scrub: fix threadcount estimates for phase 6 Darrick J. Wong
2024-04-16 4:53 ` Christoph Hellwig
2024-04-16 0:59 ` [PATCH 4/5] xfs_scrub: don't fail while reporting media scan errors Darrick J. Wong
2024-04-16 0:59 ` [PATCH 5/5] xfs_io: add linux madvise advice codes Darrick J. Wong
2024-04-17 7:34 ` [PATCHSET 1/4] xfsprogs: bug fixes for 6.8 Carlos Maiolino
2024-04-17 15:30 ` Darrick J. Wong
2024-04-16 0:58 ` [PATCHSET 2/4] libxfs: sync with 6.9 Darrick J. Wong
2024-04-16 1:00 ` [PATCH 088/111] libxfs: teach buftargs to maintain their own buffer hashtable Darrick J. Wong
2024-04-16 1:00 ` [PATCH 089/111] libxfs: add xfile support Darrick J. Wong
2024-04-16 1:00 ` [PATCH 090/111] libxfs: partition memfd files to avoid using too many fds Darrick J. Wong
2024-04-16 4:55 ` Christoph Hellwig
2024-04-16 15:49 ` Darrick J. Wong
2024-04-16 16:29 ` Christoph Hellwig
2024-04-16 16:57 ` Darrick J. Wong
2024-04-16 18:47 ` Christoph Hellwig
2024-04-16 18:55 ` Darrick J. Wong
2024-04-24 17:20 ` [PATCH v3.1 " Darrick J. Wong
2024-04-16 1:00 ` [PATCH 091/111] xfs: teach buftargs to maintain their own buffer hashtable Darrick J. Wong
2024-04-16 1:01 ` Darrick J. Wong [this message]
2024-04-16 0:58 ` [PATCHSET v30.3 3/4] xfsprogs: bmap log intent cleanups Darrick J. Wong
2024-04-16 1:01 ` [PATCH 1/4] libxfs: remove kmem_alloc, kmem_zalloc, and kmem_free Darrick J. Wong
2024-04-16 4:55 ` Christoph Hellwig
2024-04-16 1:01 ` [PATCH 2/4] libxfs: add a bi_entry helper Darrick J. Wong
2024-04-16 4:55 ` Christoph Hellwig
2024-04-16 1:01 ` [PATCH 3/4] libxfs: reuse xfs_bmap_update_cancel_item Darrick J. Wong
2024-04-16 4:55 ` Christoph Hellwig
2024-04-16 1:02 ` [PATCH 4/4] libxfs: add a xattr_entry helper Darrick J. Wong
2024-04-16 4:56 ` Christoph Hellwig
2024-04-16 0:58 ` [PATCHSET v30.3 4/4] xfs_repair: minor fixes Darrick J. Wong
2024-04-16 1:02 ` [PATCH 1/1] xfs_repair: check num before bplist[num] Darrick J. Wong
2024-04-16 4:56 ` Christoph Hellwig
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=171322883539.211103.15663396557404367362.stgit@frogsfrogsfrogs \
--to=djwong@kernel.org \
--cc=cem@kernel.org \
--cc=cmaiolino@redhat.com \
--cc=hch@infradead.org \
--cc=hch@lst.de \
--cc=linux-xfs@vger.kernel.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.