Skip to contentSkip to conversation

pstore/blk: Use the normal block device I/O path

PATCH In review Kees Cook 3 participants 5 messages Jun 14, 2021 – Jun 15, 2021

Subsystems (3)
lore
Share & tools
lore

Kees Cook

author
#0 lore raw

PATCH 1/1 pstore/blk: Use the normal block device I/O path

Stop poking into block layer internals and just open the block device file an use kernel_read and kernel_write on it. Note that this means the transformation from name_to_dev_t can't be used anymore when pstore_blk is loaded as a module: a full filesystem device path name must be used instead.

---
This reworks hch's proposal from
https://lore.kernel.org/lkml/[email protected]/

---
 fs/pstore/blk.c | 262 ++++++++++++++++--------------------------------
 1 file changed, 84 insertions(+), 178 deletions(-)
fs/pstore/blk.c +84 -178
index 4bb8a344957a..35458445cbd4 100644
@@ -8,15 +8,16 @@
 
 #include <linux/kernel.h>
 #include <linux/module.h>
-#include "../../block/blk.h"
 #include <linux/blkdev.h>
 #include <linux/string.h>
 #include <linux/of.h>
 #include <linux/of_address.h>
 #include <linux/platform_device.h>
 #include <linux/pstore_blk.h>
+#include <linux/fs.h>
+#include <linux/file.h>
+#include <linux/init_syscalls.h>
 #include <linux/mount.h>
-#include <linux/uio.h>
 
 static long kmsg_size = CONFIG_PSTORE_BLK_KMSG_SIZE;
 module_param(kmsg_size, long, 0400);
@@ -60,23 +61,25 @@ MODULE_PARM_DESC(best_effort, "use best effort to write (i.e. do not require sto
  *
  * Usually, this will be a partition of a block device.
  *
- * blkdev accepts the following variants:
- * 1) <hex_major><hex_minor> device number in hexadecimal representation,
- *    with no leading 0x, for example b302.
- * 2) /dev/<disk_name> represents the device number of disk
- * 3) /dev/<disk_name><decimal> represents the device number
+ * blkdev accepts the following variants, when built as a module:
+ * 1) /dev/<disk_name> represents the device number of disk
+ * 2) /dev/<disk_name><decimal> represents the device number
  *    of partition - device number of disk plus the partition number
- * 4) /dev/<disk_name>p<decimal> - same as the above, that form is
+ * 3) /dev/<disk_name>p<decimal> - same as the above, that form is
  *    used when disk name of partitioned disk ends on a digit.
- * 5) PARTUUID=00112233-4455-6677-8899-AABBCCDDEEFF representing the
+ *
+ * blkdev accepts the following variants when built into the kernel:
+ * 1) <hex_major><hex_minor> device number in hexadecimal representation,
+ *    with no leading 0x, for example b302.
+ * 2) PARTUUID=00112233-4455-6677-8899-AABBCCDDEEFF representing the
  *    unique id of a partition if the partition table provides it.
  *    The UUID may be either an EFI/GPT UUID, or refer to an MSDOS
  *    partition using the format SSSSSSSS-PP, where SSSSSSSS is a zero-
  *    filled hex representation of the 32-bit "NT disk signature", and PP
  *    is a zero-filled hex representation of the 1-based partition number.
- * 6) PARTUUID=<UUID>/PARTNROFF=<int> to select a partition in relation to
+ * 3) PARTUUID=<UUID>/PARTNROFF=<int> to select a partition in relation to
  *    a partition with a known unique id.
- * 7) <major>:<minor> major and minor number of the device separated by
+ * 4) <major>:<minor> major and minor number of the device separated by
  *    a colon.
  */
 static char blkdev[80] = CONFIG_PSTORE_BLK_BLKDEV;
@@ -88,15 +91,9 @@ MODULE_PARM_DESC(blkdev, "block device for pstore storage");
  * during the register/unregister functions.
  */
 static DEFINE_MUTEX(pstore_blk_lock);
-static struct block_device *psblk_bdev;
+static struct file *psblk_file;
 static struct pstore_zone_info *pstore_zone_info;
 
-struct bdev_info {
-	dev_t devt;
-	sector_t nr_sects;
-	sector_t start_sect;
-};
-
 #define check_size(name, alignsize) ({				\
 	long _##name_ = (name);					\
 	_##name_ = _##name_ <= 0 ? 0 : (_##name_ * 1024);	\
@@ -114,8 +111,19 @@ static int __register_pstore_device(struct pstore_device_info *dev)
 
 	lockdep_assert_held(&pstore_blk_lock);
 
-	if (!dev || !dev->total_size || !dev->read || !dev->write)
+	if (!dev || !dev->total_size || !dev->read || !dev->write) {
+		if (!dev)
+			pr_err("NULL device info\n");
+		else {
+			if (!dev->total_size)
+				pr_err("zero sized device\n");
+			if (!dev->read)
+				pr_err("no read handler for device\n");
+			if (!dev->write)
+				pr_err("no write handler for device\n");
+		}
 		return -EINVAL;
+	}
 
 	/* someone already registered before */
 	if (pstore_zone_info)
@@ -205,203 +213,101 @@ void unregister_pstore_device(struct pstore_device_info *dev)
 }
 EXPORT_SYMBOL_GPL(unregister_pstore_device);
 
-/**
- * psblk_get_bdev() - open block device
- *
- * @holder:	Exclusive holder identifier
- * @info:	Information about bdev to fill in
- *
- * Return: pointer to block device on success and others on error.
- *
- * On success, the returned block_device has reference count of one.
- */
-static struct block_device *psblk_get_bdev(void *holder,
-					   struct bdev_info *info)
-{
-	struct block_device *bdev = ERR_PTR(-ENODEV);
-	fmode_t mode = FMODE_READ | FMODE_WRITE;
-	sector_t nr_sects;
-
-	lockdep_assert_held(&pstore_blk_lock);
-
-	if (pstore_zone_info)
-		return ERR_PTR(-EBUSY);
-
-	if (!blkdev[0])
-		return ERR_PTR(-ENODEV);
-
-	if (holder)
-		mode |= FMODE_EXCL;
-	bdev = blkdev_get_by_path(blkdev, mode, holder);
-	if (IS_ERR(bdev)) {
-		dev_t devt;
-
-		devt = name_to_dev_t(blkdev);
-		if (devt == 0)
-			return ERR_PTR(-ENODEV);
-		bdev = blkdev_get_by_dev(devt, mode, holder);
-		if (IS_ERR(bdev))
-			return bdev;
-	}
-
-	nr_sects = bdev_nr_sectors(bdev);
-	if (!nr_sects) {
-		pr_err("not enough space for '%s'\n", blkdev);
-		blkdev_put(bdev, mode);
-		return ERR_PTR(-ENOSPC);
-	}
-
-	if (info) {
-		info->devt = bdev->bd_dev;
-		info->nr_sects = nr_sects;
-		info->start_sect = get_start_sect(bdev);
-	}
-
-	return bdev;
-}
-
-static void psblk_put_bdev(struct block_device *bdev, void *holder)
-{
-	fmode_t mode = FMODE_READ | FMODE_WRITE;
-
-	lockdep_assert_held(&pstore_blk_lock);
-
-	if (!bdev)
-		return;
-
-	if (holder)
-		mode |= FMODE_EXCL;
-	blkdev_put(bdev, mode);
-}
-
 static ssize_t psblk_generic_blk_read(char *buf, size_t bytes, loff_t pos)
 {
-	struct block_device *bdev = psblk_bdev;
-	struct file file;
-	struct kiocb kiocb;
-	struct iov_iter iter;
-	struct kvec iov = {.iov_base = buf, .iov_len = bytes};
-
-	if (!bdev)
-		return -ENODEV;
-
-	memset(&file, 0, sizeof(struct file));
-	file.f_mapping = bdev->bd_inode->i_mapping;
-	file.f_flags = O_DSYNC | __O_SYNC | O_NOATIME;
-	file.f_inode = bdev->bd_inode;
-	file_ra_state_init(&file.f_ra, file.f_mapping);
-
-	init_sync_kiocb(&kiocb, &file);
-	kiocb.ki_pos = pos;
-	iov_iter_kvec(&iter, READ, &iov, 1, bytes);
-
-	return generic_file_read_iter(&kiocb, &iter);
+	return kernel_read(psblk_file, buf, bytes, &pos);
 }
 
 static ssize_t psblk_generic_blk_write(const char *buf, size_t bytes,
 		loff_t pos)
 {
-	struct block_device *bdev = psblk_bdev;
-	struct iov_iter iter;
-	struct kiocb kiocb;
-	struct file file;
-	ssize_t ret;
-	struct kvec iov = {.iov_base = (void *)buf, .iov_len = bytes};
-
-	if (!bdev)
-		return -ENODEV;
-
 	/* Console/Ftrace backend may handle buffer until flush dirty zones */
 	if (in_interrupt() || irqs_disabled())
 		return -EBUSY;
-
-	memset(&file, 0, sizeof(struct file));
-	file.f_mapping = bdev->bd_inode->i_mapping;
-	file.f_flags = O_DSYNC | __O_SYNC | O_NOATIME;
-	file.f_inode = bdev->bd_inode;
-
-	init_sync_kiocb(&kiocb, &file);
-	kiocb.ki_pos = pos;
-	iov_iter_kvec(&iter, WRITE, &iov, 1, bytes);
-
-	inode_lock(bdev->bd_inode);
-	ret = generic_write_checks(&kiocb, &iter);
-	if (ret > 0)
-		ret = generic_perform_write(&file, &iter, pos);
-	inode_unlock(bdev->bd_inode);
-
-	if (likely(ret > 0)) {
-		const struct file_operations f_op = {.fsync = blkdev_fsync};
-
-		file.f_op = &f_op;
-		kiocb.ki_pos += ret;
-		ret = generic_write_sync(&kiocb, ret);
-	}
-	return ret;
+	return kernel_write(psblk_file, buf, bytes, &pos);
 }
 
 /*
  * This takes its configuration only from the module parameters now.
- * See psblk_get_bdev() and blkdev.
  */
 static int __register_pstore_blk(void)
 {
-	char bdev_name[BDEVNAME_SIZE];
-	struct block_device *bdev;
-	struct pstore_device_info dev;
-	struct bdev_info binfo;
-	void *holder = blkdev;
+	struct pstore_device_info dev = {
+		.read = psblk_generic_blk_read,
+		.write = psblk_generic_blk_write,
+	};
 	int ret = -ENODEV;
 
 	lockdep_assert_held(&pstore_blk_lock);
 
-	/* hold bdev exclusively */
-	memset(&binfo, 0, sizeof(binfo));
-	bdev = psblk_get_bdev(holder, &binfo);
-	if (IS_ERR(bdev)) {
-		pr_err("failed to open '%s'!\n", blkdev);
-		return PTR_ERR(bdev);
+	if (!__is_defined(MODULE)) {
+		/*
+		 * During early boot the real root file system hasn't been
+		 * mounted yet, and no device nodes are present yet. Use the
+		 * same scheme to find the device that we use for mounting
+		 * the root file system.
+		 */
+		static const char devname[] = "/dev/pstore-blk";
+		dev_t dev = name_to_dev_t(blkdev);
+
+		if (!dev) {
+			pr_err("failed to resolve '%s'!\n", blkdev);
+			goto err;
+		}
+
+		init_unlink(devname);
+		init_mknod(devname, S_IFBLK | 0600, new_encode_dev(dev));
+		strscpy(blkdev, devname, sizeof(blkdev));
 	}
 
-	/* only allow driver matching the @blkdev */
-	if (!binfo.devt) {
-		pr_debug("no major\n");
-		ret = -ENODEV;
-		goto err_put_bdev;
+	psblk_file = filp_open(blkdev, O_RDWR | O_DSYNC | O_NOATIME | O_EXCL, 0);
+	if (IS_ERR(psblk_file)) {
+		ret = PTR_ERR(psblk_file);
+		pr_err("failed to open '%s': %d!\n", blkdev, ret);
+		goto err;
 	}
 
-	/* psblk_bdev must be assigned before register to pstore/blk */
-	psblk_bdev = bdev;
+	if (!S_ISBLK(file_inode(psblk_file)->i_mode)) {
+		pr_err("'%s' is not block device!\n", blkdev);
+		goto err_fput;
+	}
 
-	memset(&dev, 0, sizeof(dev));
-	dev.total_size = binfo.nr_sects << SECTOR_SHIFT;
-	dev.read = psblk_generic_blk_read;
-	dev.write = psblk_generic_blk_write;
+	if (!psblk_file->f_mapping)
+		pr_err("missing f_mapping\n");
+	else if (!psblk_file->f_mapping->host)
+		pr_err("missing host\n");
+	else if (!I_BDEV(psblk_file->f_mapping->host))
+		pr_err("missing I_BDEV\n");
+	else if (!I_BDEV(psblk_file->f_mapping->host)->bd_inode)
+		pr_err("missing bd_inode\n");
+	else
+		dev.total_size = i_size_read(I_BDEV(psblk_file->f_mapping->host)->bd_inode);
 
 	ret = __register_pstore_device(&dev);
 	if (ret)
-		goto err_put_bdev;
+		goto err_fput;
 
-	bdevname(bdev, bdev_name);
-	pr_info("attached %s (no dedicated panic_write!)\n", bdev_name);
+	pr_info("attached %s (%zu) (no dedicated panic_write!)\n",
+		blkdev, dev.total_size);
 	return 0;
 
-err_put_bdev:
-	psblk_bdev = NULL;
-	psblk_put_bdev(bdev, holder);
+err_fput:
+	fput(psblk_file);
+err:
+	psblk_file = NULL;
+
 	return ret;
 }
 
-static void __unregister_pstore_blk(unsigned int major)
+static void __unregister_pstore_blk(struct file *device)
 {
 	struct pstore_device_info dev = { .read = psblk_generic_blk_read };
-	void *holder = blkdev;
 
 	lockdep_assert_held(&pstore_blk_lock);
-	if (psblk_bdev && MAJOR(psblk_bdev->bd_dev) == major) {
+	if (psblk_file && psblk_file == device) {
 		__unregister_pstore_device(&dev);
-		psblk_put_bdev(psblk_bdev, holder);
-		psblk_bdev = NULL;
+		fput(psblk_file);
+		psblk_file = NULL;
 	}
 }
 
@@ -435,8 +341,8 @@ late_initcall(pstore_blk_init);
 static void __exit pstore_blk_exit(void)
 {
 	mutex_lock(&pstore_blk_lock);
-	if (psblk_bdev)
-		__unregister_pstore_blk(MAJOR(psblk_bdev->bd_dev));
+	if (psblk_file)
+		__unregister_pstore_blk(psblk_file);
 	else {
 		struct pstore_device_info dev = { };
 
Signature · 1 line
2.25.1
maintainer
#1 lore raw

fs/pstore/blk.c void unregister_pstore_device(struct pstore_device_info *dev)

Quoted 9 lines · Kees Cook

static ssize_t psblk_generic_blk_write(const char *buf, size_t bytes,

fs/pstore/blk.c +1 -0
 		loff_t pos)
 {
 
 	/* Console/Ftrace backend may handle buffer until flush dirty zones */
 	if (in_interrupt() || irqs_disabled())
 		return -EBUSY;
 
+	return kernel_write(psblk_file, buf, bytes, &pos);

In which locking environments could that be called? The checks above look like that thing could be called from just about any context; could that happen when the caller is holding a page locked?

IOW, what are those checks really trying to do?

author
#2 lore raw
Quoted 10 lines · Al Viro
Quoted 7 lines · Kees Cook

static ssize_t psblk_generic_blk_write(const char *buf, size_t bytes,

fs/pstore/blk.c +0 -0
 		loff_t pos)
 {
 
 	/* Console/Ftrace backend may handle buffer until flush dirty zones */
 	if (in_interrupt() || irqs_disabled())
 		return -EBUSY;
Quoted 5 lines

fs/pstore/blk.c void unregister_pstore_device(struct pstore_device_info *dev) · line 227

Quoted 1 line
fs/pstore/blk.c +1 -0
+	return kernel_write(psblk_file, buf, bytes, &pos);

In which locking environments could that be called? The checks above look like that thing could be called from just about any context; could that happen when the caller is holding a page locked?

The contexts are determined by both each of the pstore "front ends":

PSTORE_FLAGS_DMESG:
static struct kmsg_dumper pstore_dumper = {
.dump = pstore_dump,
...
kmsg_dump_register(&pstore_dumper);

PSTORE_FLAGS_CONSOLE:
static struct console pstore_console = {
.write = pstore_console_write,
...
register_console(&pstore_console);

PSTORE_FLAGS_FTRACE:
static struct ftrace_ops pstore_ftrace_ops __read_mostly = {
.func = pstore_ftrace_call,
...
register_ftrace_function(&pstore_ftrace_ops);

PSTORE_FLAGS_PMSG:
static const struct file_operations pmsg_fops = {
...
.write = write_pmsg,
...
pmsg_major = register_chrdev(0, PMSG_NAME, &pmsg_fops);

and each of the pstore "back ends". (ram, EFI vars, block, etc.)

Quoted 1 line

IOW, what are those checks really trying to do?

Traditionally, the most restrictive case is kmsg_dump, but that's the whole point here of the "best effort" mode: if we can't safely make the call and no panic handler has been registered, we must skip the call.

e.g. the RAM pstore backend has all its buffers preallocated, and it'll just write directly into them. The handling here has gotten progressive weirder, as more back ends landed -- i.e. EFI var writing added some limits to the kind of locking pstore could do, etc.

It may turn out that the checks above aren't needed. I haven't tried it without, but I suspect it's for the kmsg_dump case.

-Kees

Signature · 1 line
Kees Cook
#3 lore raw
In reply to Kees Cook
Quoted 4 lines
fs/pstore/blk.c +3 -1
-	if (!dev || !dev->total_size || !dev->read || !dev->write)
+	if (!dev || !dev->total_size || !dev->read || !dev->write) {
+		if (!dev)
+			pr_err("NULL device info\n");

fs/pstore/blk.c static int __register_pstore_device(struct pstore_device_info *dev) · line 126

Quoted 10 lines
fs/pstore/blk.c +9 -0
+		else {
+			if (!dev->total_size)
+				pr_err("zero sized device\n");
+			if (!dev->read)
+				pr_err("no read handler for device\n");
+			if (!dev->write)
+				pr_err("no write handler for device\n");
+		}
 		return -EINVAL;
+	}

This is completely unrelated and should be a separate patch. And it also looks rather strange, I'd at very least split the dev check out and return early without the weird compound statement, but would probably handle each one separate. All assuming that we really need all these debug printks.

Quoted 4 lines

/*
* This takes its configuration only from the module parameters now. */
static int __register_pstore_blk(void)

This needs a __init annotation now.

fs/pstore/blk.c void unregister_pstore_device(struct pstore_device_info *dev) · line 238

Quoted 7 lines

{

fs/pstore/blk.c +4 -0
+	struct pstore_device_info dev = {
+		.read = psblk_generic_blk_read,
+		.write = psblk_generic_blk_write,
+	};

On-stack method tables are a little odd..

fs/pstore/blk.c void unregister_pstore_device(struct pstore_device_info *dev) · line 243

Quoted 1 line
fs/pstore/blk.c +1 -0
+	if (!__is_defined(MODULE)) {

This looks a little weird. Can we define a rapper for this in config.h that is a little more self-explanatory, e.g. in_module()?

fs/pstore/blk.c void unregister_pstore_device(struct pstore_device_info *dev) · line 276

Quoted 2 lines
fs/pstore/blk.c +2 -0
+	if (!psblk_file->f_mapping)
+		pr_err("missing f_mapping\n");

Can't ever be true.

fs/pstore/blk.c void unregister_pstore_device(struct pstore_device_info *dev) · line 278

Quoted 2 lines
fs/pstore/blk.c +2 -0
+	else if (!psblk_file->f_mapping->host)
+		pr_err("missing host\n");

Can't ever be true either.

fs/pstore/blk.c void unregister_pstore_device(struct pstore_device_info *dev) · line 282

Quoted 4 lines
fs/pstore/blk.c +4 -0
+	else if (!I_BDEV(psblk_file->f_mapping->host))
+		pr_err("missing I_BDEV\n");
+	else if (!I_BDEV(psblk_file->f_mapping->host)->bd_inode)
+		pr_err("missing bd_inode\n");

І_BDEV just does pointer arithmetics, so it can't ever return NULL. And there are no block device inodes without bd_inode either. And all of this is per definition present for open S_ISBLK inodes.

author
#4 lore raw
Quoted 12 lines · Christoph Hellwig
Quoted 12 lines
fs/pstore/blk.c +11 -1
-	if (!dev || !dev->total_size || !dev->read || !dev->write)
+	if (!dev || !dev->total_size || !dev->read || !dev->write) {
+		if (!dev)
+			pr_err("NULL device info\n");
+		else {
+			if (!dev->total_size)
+				pr_err("zero sized device\n");
+			if (!dev->read)
+				pr_err("no read handler for device\n");
+			if (!dev->write)
+				pr_err("no write handler for device\n");
+		}
Quoted 8 lines

fs/pstore/blk.c static int __register_pstore_device(struct pstore_device_info *dev) · line 126

Quoted 2 lines
fs/pstore/blk.c +1 -0
 		return -EINVAL;
+	}

This is completely unrelated and should be a separate patch. And it also looks rather strange, I'd at very least split the dev check out and return early without the weird compound statement, but would probably handle each one separate. All assuming that we really need all these debug printks.

Agreed -- I've moved it to a separate patch.

Quoted 6 lines
Quoted 4 lines

/*
* This takes its configuration only from the module parameters now. */
static int __register_pstore_blk(void)

This needs a __init annotation now.

Ah yes, good point. I've rearranged things to avoid this.

Quoted 7 lines

fs/pstore/blk.c void unregister_pstore_device(struct pstore_device_info *dev) · line 238

Quoted 5 lines

{

fs/pstore/blk.c +4 -0
+	struct pstore_device_info dev = {
+		.read = psblk_generic_blk_read,
+		.write = psblk_generic_blk_write,
+	};

On-stack method tables are a little odd..

struct pstore_device_info is mainly an argument passing structure, not an ops structure. There is some weird over-engineering here, which I'll fix up in a follow-up patch.

Quoted 4 lines

fs/pstore/blk.c void unregister_pstore_device(struct pstore_device_info *dev) · line 243

Quoted 1 line
fs/pstore/blk.c +1 -0
+	if (!__is_defined(MODULE)) {

This looks a little weird. Can we define a rapper for this in config.h that is a little more self-explanatory, e.g. in_module()?

I've adjusted this.

Quoted 11 lines

fs/pstore/blk.c void unregister_pstore_device(struct pstore_device_info *dev) · line 276

Quoted 3 lines
fs/pstore/blk.c +2 -0
+	if (!psblk_file->f_mapping)
+		pr_err("missing f_mapping\n");

Can't ever be true.

fs/pstore/blk.c void unregister_pstore_device(struct pstore_device_info *dev) · line 278

Quoted 2 lines
fs/pstore/blk.c +2 -0
+	else if (!psblk_file->f_mapping->host)
+		pr_err("missing host\n");

Can't ever be true either.

Quoted 8 lines

fs/pstore/blk.c void unregister_pstore_device(struct pstore_device_info *dev) · line 282

Quoted 4 lines
fs/pstore/blk.c +4 -0
+	else if (!I_BDEV(psblk_file->f_mapping->host))
+		pr_err("missing I_BDEV\n");
+	else if (!I_BDEV(psblk_file->f_mapping->host)->bd_inode)
+		pr_err("missing bd_inode\n");

І_BDEV just does pointer arithmetics, so it can't ever return NULL. And there are no block device inodes without bd_inode either. And all of this is per definition present for open S_ISBLK inodes.

Okay, good. There were a lot of dead-ends in here, and after the removal of i_bdev, it wasn't obvious which things were going to exist. I've removed all these checks.

One thing I noticed it that it seems we're missing a global helper for "->f_mapping->host", which I see repeated a lot. (There is a local helper bdev_file_inode().)

$ git grep '\bf_mapping->host\b' | wc -l
149

Thanks for the review! I'll send a v2 series.

Signature · 1 line
Kees Cook