On 6/23/21 7:26 AM, Coly Li wrote:

On 6/22/21 6:39 PM, Hannes Reinecke wrote:

quoted

On 6/15/21 7:49 AM, Coly Li wrote:

quoted

From: Jianpeng Ma <redacted>

This patch define the prototype data structures in memory and
initializes the nvm pages allocator.

The nvm address space which is managed by this allocator can consist of
many nvm namespaces, and some namespaces can compose into one nvm set,
like cache set. For this initial implementation, only one set can be
supported.

The users of this nvm pages allocator need to call register_namespace()
to register the nvdimm device (like /dev/pmemX) into this allocator as
the instance of struct nvm_namespace.

Reported-by: Randy Dunlap <redacted>
Signed-off-by: Jianpeng Ma <redacted>
Co-developed-by: Qiaowei Ren <redacted>
Signed-off-by: Qiaowei Ren <redacted>
Signed-off-by: Coly Li <redacted>
---
  drivers/md/bcache/Kconfig     |  10 ++
  drivers/md/bcache/Makefile    |   1 +
  drivers/md/bcache/nvm-pages.c | 295 ++++++++++++++++++++++++++++++++++
  drivers/md/bcache/nvm-pages.h |  74 +++++++++
  drivers/md/bcache/super.c     |   3 +
  5 files changed, 383 insertions(+)
  create mode 100644 drivers/md/bcache/nvm-pages.c
  create mode 100644 drivers/md/bcache/nvm-pages.h

diff --git a/drivers/md/bcache/Kconfig b/drivers/md/bcache/Kconfig
index d1ca4d059c20..a69f6c0e0507 100644
--- a/drivers/md/bcache/Kconfig
+++ b/drivers/md/bcache/Kconfig

@@ -35,3 +35,13 @@ config BCACHE_ASYNC_REGISTRATION
  	device path into this file will returns immediately and the real
  	registration work is handled in kernel work queue in asynchronous
  	way.
+
+config BCACHE_NVM_PAGES
+	bool "NVDIMM support for bcache (EXPERIMENTAL)"
+	depends on BCACHE
+	depends on 64BIT
+	depends on LIBNVDIMM
+	depends on DAX
+	help
+	  Allocate/release NV-memory pages for bcache and provide allocated pages
+	  for each requestor after system reboot.

diff --git a/drivers/md/bcache/Makefile b/drivers/md/bcache/Makefile
index 5b87e59676b8..2397bb7c7ffd 100644
--- a/drivers/md/bcache/Makefile
+++ b/drivers/md/bcache/Makefile

@@ -5,3 +5,4 @@ obj-$(CONFIG_BCACHE)	+= bcache.o
  bcache-y		:= alloc.o bset.o btree.o closure.o debug.o extents.o\
  	io.o journal.o movinggc.o request.o stats.o super.o sysfs.o trace.o\
  	util.o writeback.o features.o
+bcache-$(CONFIG_BCACHE_NVM_PAGES) += nvm-pages.o

diff --git a/drivers/md/bcache/nvm-pages.c b/drivers/md/bcache/nvm-pages.c
new file mode 100644
index 000000000000..18fdadbc502f
--- /dev/null
+++ b/drivers/md/bcache/nvm-pages.c

@@ -0,0 +1,295 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * Nvdimm page-buddy allocator
+ *
+ * Copyright (c) 2021, Intel Corporation.
+ * Copyright (c) 2021, Qiaowei Ren <qiaowei.ren@intel.com>.
+ * Copyright (c) 2021, Jianpeng Ma <jianpeng.ma@intel.com>.
+ */
+
+#if defined(CONFIG_BCACHE_NVM_PAGES)
+

No need for this 'if' statement as it'll be excluded by the Makefile
anyway if the config option isn't set.

Such if is necessary because stub routines are defined when
CONFIG_BCACHE_NVM_PAGES is not defined, e.g.

426 +#else
427 +
428 +static inline struct bch_nvm_namespace
*bch_register_namespace(const char *dev_path)
429 +{
430 +       return NULL;
431 +}
432 +static inline int bch_nvm_init(void)
433 +{
434 +       return 0;
435 +}
436 +static inline void bch_nvm_exit(void) { }
437 +
438 +#endif /* CONFIG_BCACHE_NVM_PAGES */

But then these stubs should be defined in the header file, not here.

[ .. ]

quoted

quoted

+			rc = -ENOMEM;
+			goto unlock;
+		}
+	}
+
+	only_set->nss[ns->sb->this_namespace_nr] = ns;
+
+	/* Firstly attach */

Initial attach?

Will fix in next post.

quoted

quoted

+	if ((unsigned long)ns->sb->owner_list_head == BCH_NVM_PAGES_OWNER_LIST_HEAD_OFFSET) {
+		struct bch_nvm_pages_owner_head *sys_owner_head;
+		struct bch_nvm_pgalloc_recs *sys_pgalloc_recs;
+
+		ns->sb->owner_list_head = ns->kaddr + BCH_NVM_PAGES_OWNER_LIST_HEAD_OFFSET;
+		sys_pgalloc_recs = ns->kaddr + BCH_NVM_PAGES_SYS_RECS_HEAD_OFFSET;
+
+		sys_owner_head = &(ns->sb->owner_list_head->heads[0]);
+		sys_owner_head->recs[0] = sys_pgalloc_recs;
+		ns->sb->csum = csum_set(ns->sb);
+

Hmm. You are trying to pick up the 'list_head' structure from NVM, right?

No, this is not READ, it's WRITE onto NVDIMM.

sys_owner_head points to NVDIMM, since ns->sb->owner_list_head is updated,
the checksum of ns->sb should be updated to new value onto the NVDIMM.
This is what the above line does.

Ah, right.

quoted

In doing so, don't you need to validate the structure (eg by checking
the checksum) before doing so to ensure that the contents are valid?

The check sum checking for READ is done in read_nvdimm_meta_super() in
following lines,
198 +       r = -EINVAL;
199 +       expected_csum = csum_set(sb);
200 +       if (expected_csum != sb->csum) {
201 +               pr_info("csum is not match with expected one\n");
202 +               goto put_page;
203 +       }

Once thing to note is, currently all NVDIMM update is not power failure
considered. This is the next big task to do after the first small code
base merged.

quoted

quoted

+		sys_pgalloc_recs->owner = sys_owner_head;
+	} else
+		BUG_ON(ns->sb->owner_list_head !=
+			(ns->kaddr + BCH_NVM_PAGES_OWNER_LIST_HEAD_OFFSET));
+
+unlock:
+	mutex_unlock(&only_set->lock);
+	return rc;
+}
+
+static int read_nvdimm_meta_super(struct block_device *bdev,
+			      struct bch_nvm_namespace *ns)
+{
+	struct page *page;
+	struct bch_nvm_pages_sb *sb;
+	int r = 0;
+	uint64_t expected_csum = 0;
+
+	page = read_cache_page_gfp(bdev->bd_inode->i_mapping,
+			BCH_NVM_PAGES_SB_OFFSET >> PAGE_SHIFT, GFP_KERNEL);
+
+	if (IS_ERR(page))
+		return -EIO;
+
+	sb = (struct bch_nvm_pages_sb *)(page_address(page) +
+					offset_in_page(BCH_NVM_PAGES_SB_OFFSET));
+	r = -EINVAL;
+	expected_csum = csum_set(sb);
+	if (expected_csum != sb->csum) {
+		pr_info("csum is not match with expected one\n");
+		goto put_page;
+	}
+
+	if (memcmp(sb->magic, bch_nvm_pages_magic, 16)) {
+		pr_info("invalid bch_nvm_pages_magic\n");
+		goto put_page;
+	}
+
+	if (sb->total_namespaces_nr != 1) {
+		pr_info("currently only support one nvm device\n");
+		goto put_page;
+	}
+
+	if (sb->sb_offset != BCH_NVM_PAGES_SB_OFFSET) {
+		pr_info("invalid superblock offset\n");
+		goto put_page;
+	}
+
+	r = 0;
+	/* temporary use for DAX API */
+	ns->page_size = sb->page_size;
+	ns->pages_total = sb->pages_total;
+
+put_page:
+	put_page(page);
+	return r;
+}
+
+struct bch_nvm_namespace *bch_register_namespace(const char *dev_path)
+{
+	struct bch_nvm_namespace *ns;
+	int err;
+	pgoff_t pgoff;
+	char buf[BDEVNAME_SIZE];
+	struct block_device *bdev;
+	int id;
+	char *path = NULL;
+
+	path = kstrndup(dev_path, 512, GFP_KERNEL);
+	if (!path) {
+		pr_err("kstrndup failed\n");
+		return ERR_PTR(-ENOMEM);
+	}
+
+	bdev = blkdev_get_by_path(strim(path),
+				  FMODE_READ|FMODE_WRITE|FMODE_EXEC,
+				  only_set);
+	if (IS_ERR(bdev)) {
+		pr_info("get %s error: %ld\n", dev_path, PTR_ERR(bdev));
+		kfree(path);
+		return ERR_PTR(PTR_ERR(bdev));
+	}
+
+	err = -ENOMEM;
+	ns = kzalloc(sizeof(struct bch_nvm_namespace), GFP_KERNEL);
+	if (!ns)
+		goto bdput;
+
+	err = -EIO;
+	if (read_nvdimm_meta_super(bdev, ns)) {
+		pr_info("%s read nvdimm meta super block failed.\n",
+			bdevname(bdev, buf));
+		goto free_ns;
+	}
+
+	err = -EOPNOTSUPP;
+	if (!bdev_dax_supported(bdev, ns->page_size)) {
+		pr_info("%s don't support DAX\n", bdevname(bdev, buf));
+		goto free_ns;
+	}
+
+	err = -EINVAL;
+	if (bdev_dax_pgoff(bdev, 0, ns->page_size, &pgoff)) {
+		pr_info("invalid offset of %s\n", bdevname(bdev, buf));
+		goto free_ns;
+	}
+
+	err = -ENOMEM;
+	ns->dax_dev = fs_dax_get_by_bdev(bdev);
+	if (!ns->dax_dev) {
+		pr_info("can't by dax device by %s\n", bdevname(bdev, buf));
+		goto free_ns;
+	}
+
+	err = -EINVAL;
+	id = dax_read_lock();
+	if (dax_direct_access(ns->dax_dev, pgoff, ns->pages_total,
+			      &ns->kaddr, &ns->start_pfn) <= 0) {
+		pr_info("dax_direct_access error\n");
+		dax_read_unlock(id);
+		goto free_ns;
+	}
+	dax_read_unlock(id);
+
+	ns->sb = ns->kaddr + BCH_NVM_PAGES_SB_OFFSET;
+

You already read the superblock in read_nvdimm_meta_super(), right?
Wouldn't it be better to first do the 'dax_direct_access()' call, and
then check the superblock?
That way you'll ensure that dax_direct_access()' did the right thing;
with the current code you are using two different methods of accessing
the superblock, which theoretically can result in one method succeeding,
the other not ...

We have to do it. Because the mapping size of dax_direct_access() is
from ns->pages_total, it is stored on NVDIMM. Before calling
dax_direct_acess()
we need to make sure ns is an valid super block stored on NVDIMM, then
we can
trust value of ns->pages_total to do the DAX mapping.

Another method is firstly mapping a small fixed range (e.g. 1GB), and
check whether the super block on NVDIMM is valid. If yes, re-map the
whole space indicated by ns->pages_total. But the two-times accessing cannot
be avoided.

Oh, indeed, you are correct. Disregard my comments here.

quoted

quoted

+	err = -EINVAL;
+	/* Check magic again to make sure DAX mapping is correct */
+	if (memcmp(ns->sb->magic, bch_nvm_pages_magic, 16)) {
+		pr_info("invalid bch_nvm_pages_magic after DAX mapping\n");
+		goto free_ns;
+	}
+
+	err = attach_nvm_set(ns);
+	if (err < 0)
+		goto free_ns;
+
+	ns->page_size = ns->sb->page_size;
+	ns->pages_offset = ns->sb->pages_offset;
+	ns->pages_total = ns->sb->pages_total;
+	ns->free = 0;
+	ns->bdev = bdev;
+	ns->nvm_set = only_set;
+	mutex_init(&ns->lock);
+
+	if (ns->sb->this_namespace_nr == 0) {
+		pr_info("only first namespace contain owner info\n");
+		err = init_owner_info(ns);
+		if (err < 0) {
+			pr_info("init_owner_info met error %d\n", err);
+			only_set->nss[ns->sb->this_namespace_nr] = NULL;
+			goto free_ns;
+		}
+	}
+
+	kfree(path);
+	return ns;
+free_ns:
+	kfree(ns);
+bdput:
+	blkdev_put(bdev, FMODE_READ|FMODE_WRITE|FMODE_EXEC);
+	kfree(path);
+	return ERR_PTR(err);
+}
+EXPORT_SYMBOL_GPL(bch_register_namespace);
+
+int __init bch_nvm_init(void)
+{
+	only_set = kzalloc(sizeof(*only_set), GFP_KERNEL);
+	if (!only_set)
+		return -ENOMEM;
+
+	only_set->total_namespaces_nr = 0;
+	only_set->owner_list_head = NULL;
+	only_set->nss = NULL;
+
+	mutex_init(&only_set->lock);
+
+	pr_info("bcache nvm init\n");
+	return 0;
+}
+
+void bch_nvm_exit(void)
+{
+	release_nvm_set(only_set);
+	pr_info("bcache nvm exit\n");
+}
+
+#endif /* CONFIG_BCACHE_NVM_PAGES */

diff --git a/drivers/md/bcache/nvm-pages.h b/drivers/md/bcache/nvm-pages.h
new file mode 100644
index 000000000000..3e24c4dee7fd
--- /dev/null
+++ b/drivers/md/bcache/nvm-pages.h

@@ -0,0 +1,74 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+
+#ifndef _BCACHE_NVM_PAGES_H
+#define _BCACHE_NVM_PAGES_H
+
+#if defined(CONFIG_BCACHE_NVM_PAGES)
+#include <linux/bcache-nvm.h>
+#endif /* CONFIG_BCACHE_NVM_PAGES */
+

Hmm? What is that doing here?
Please move it into the source file.

This is temporary before the whole NVDIMM support for bcache is completed.

drivers/md/bcache/nvm-pages.h has to be included because there are still
stub
routines in this header. Such stub routines like,

+static inline struct bch_nvm_namespace *bch_register_namespace(const char *dev_path)
+{
+	return NULL;
+}

will be removed after the whole code completed and merged. So currently we have to
do this to make sure the NVDIMM related code won't be leaked out if such experimental
configure is not enabled.


Thanks for your review. The addressed issue will be fixed and updated in next post.

Cheers,

Hannes
-- 
Dr. Hannes Reinecke                Kernel Storage Architect
hare@suse.de                              +49 911 74053 688
SUSE Software Solutions GmbH, Maxfeldstr. 5, 90409 Nürnberg
HRB 36809 (AG Nürnberg), Geschäftsführer: Felix Imendörffer