Re: [PATCH 22/24] mdrestore: Define mdrestore ops for v2 format

[Chandan Babu R <chandan.babu@oracle.com>]

On Tue, May 23, 2023 at 11:06:02 AM -0700, Darrick J. Wong wrote:
> On Tue, May 23, 2023 at 02:30:48PM +0530, Chandan Babu R wrote:
>> This commit adds functionality to restore metadump stored in v2 format.
>> 
>> Signed-off-by: Chandan Babu R <chandan.babu@oracle.com>
>> ---
>>  mdrestore/xfs_mdrestore.c | 209 +++++++++++++++++++++++++++++++++++---
>>  1 file changed, 194 insertions(+), 15 deletions(-)
>> 
>> diff --git a/mdrestore/xfs_mdrestore.c b/mdrestore/xfs_mdrestore.c
>> index 615ecdc77..9e06d37dc 100644
>> --- a/mdrestore/xfs_mdrestore.c
>> +++ b/mdrestore/xfs_mdrestore.c
>> @@ -11,7 +11,8 @@ struct mdrestore_ops {
>>  	int (*read_header)(void *header, FILE *mdfp);
>>  	void (*show_info)(void *header, const char *mdfile);
>>  	void (*restore)(void *header, FILE *mdfp, int data_fd,
>> -			bool is_target_file);
>> +			bool is_data_target_file, int log_fd,
>> +			bool is_log_target_file);
>>  };
>>  
>>  static struct mdrestore {
>> @@ -148,7 +149,9 @@ restore_v1(
>>  	void		*header,
>>  	FILE		*mdfp,
>>  	int		data_fd,
>> -	bool		is_target_file)
>> +	bool		is_data_target_file,
>> +	int		log_fd,
>> +	bool		is_log_target_file)
>>  {
>>  	struct xfs_metablock	*mbp = header;
>>  	struct xfs_metablock	*metablock;
>> @@ -203,7 +206,7 @@ restore_v1(
>>  
>>  	((struct xfs_dsb*)block_buffer)->sb_inprogress = 1;
>>  
>> -	verify_device_size(data_fd, is_target_file, sb.sb_dblocks,
>> +	verify_device_size(data_fd, is_data_target_file, sb.sb_dblocks,
>>  			sb.sb_blocksize);
>>  
>>  	bytes_read = 0;
>> @@ -264,6 +267,163 @@ static struct mdrestore_ops mdrestore_ops_v1 = {
>>  	.restore = restore_v1,
>>  };
>>  
>> +static int
>> +read_header_v2(
>> +	void				*header,
>> +	FILE				*mdfp)
>> +{
>> +	struct xfs_metadump_header	*xmh = header;
>> +
>> +	rewind(mdfp);
>
> Does rewind() work if @mdfp is a pipe?
>
> I suspect the best you can do is read the first 4 bytes in main, pick
> the read_header function from that, and have the read_header_v[12] read
> in the rest of the header from there.  I use a lot of:
>
> xfs_metadump -ago /dev/sda - | gzip > foo.md.gz
> gzip -d < foo.md.gz | xfs_mdrestore -g - /dev/sdb
>
> to store compressed metadumps for future reference.
>
> (Well ok I use xz or zstd, but you get the point.)
>

Thanks for pointing out the bug and suggesting the fix.

>> +
>> +	if (fread(xmh, sizeof(*xmh), 1, mdfp) != 1)
>> +		fatal("error reading from metadump file\n");
>> +	if (xmh->xmh_magic != cpu_to_be32(XFS_MD_MAGIC_V2))
>> +		return -1;
>> +
>> +	return 0;
>> +}
>> +
>> +static void
>> +show_info_v2(
>> +	void				*header,
>> +	const char			*mdfile)
>> +{
>> +	struct xfs_metadump_header	*xmh;
>> +	uint32_t			incompat_flags;
>> +
>> +	xmh = header;
>> +	incompat_flags = be32_to_cpu(xmh->xmh_incompat_flags);
>> +
>> +	printf("%s: %sobfuscated, %s log, %s metadata blocks\n",
>> +		mdfile,
>> +		incompat_flags & XFS_MD2_INCOMPAT_OBFUSCATED ? "":"not ",
>> +		incompat_flags & XFS_MD2_INCOMPAT_DIRTYLOG ? "dirty":"clean",
>> +		incompat_flags & XFS_MD2_INCOMPAT_FULLBLOCKS ? "full":"zeroed");
>> +}
>> +
>> +static void
>> +restore_v2(
>> +	void			*header,
>> +	FILE			*mdfp,
>> +	int			data_fd,
>> +	bool			is_data_target_file,
>> +	int			log_fd,
>> +	bool			is_log_target_file)
>> +{
>> +	struct xfs_sb		sb;
>> +	struct xfs_meta_extent	xme;
>> +	char			*block_buffer;
>> +	int64_t			bytes_read;
>> +	uint64_t		offset;
>> +	int			prev_len;
>> +	int			len;
>> +
>> +	if (fread(&xme, sizeof(xme), 1, mdfp) != 1)
>> +		fatal("error reading from metadump file\n");
>> +
>> +	len = be32_to_cpu(xme.xme_len);
>> +	len <<= BBSHIFT;
>
> Do we need to validate xme_addr==0 and xme_len==1 here?
>

Yes, I will add the check.

>> +
>> +	block_buffer = calloc(1, len);
>> +	if (block_buffer == NULL)
>> +		fatal("memory allocation failure\n");
>> +
>> +	if (fread(block_buffer, len, 1, mdfp) != 1)
>> +		fatal("error reading from metadump file\n");
>> +
>> +	libxfs_sb_from_disk(&sb, (struct xfs_dsb *)block_buffer);
>> +
>> +	if (sb.sb_magicnum != XFS_SB_MAGIC)
>> +		fatal("bad magic number for primary superblock\n");
>> +
>> +	if (sb.sb_logstart == 0 && log_fd == -1)
>> +		fatal("External Log device is required\n");
>> +
>> +	((struct xfs_dsb *)block_buffer)->sb_inprogress = 1;
>> +
>> +	verify_device_size(data_fd, is_data_target_file, sb.sb_dblocks,
>> +			sb.sb_blocksize);
>> +
>> +	if (sb.sb_logstart == 0)
>> +		verify_device_size(log_fd, is_log_target_file, sb.sb_logblocks,
>> +				sb.sb_blocksize);
>> +
>> +	bytes_read = 0;
>> +
>> +	do {
>> +		int fd;
>> +
>> +		if (mdrestore.show_progress &&
>> +			(bytes_read & ((1 << 20) - 1)) == 0)
>> +			print_progress("%lld MB read", bytes_read >> 20);
>
> Doesn't this miss a progress report if a metadata extent bumps
> bytes_read across a MB boundary without actually landing on it?  Say
> you've written 1020K, and the next xfs_meta_extent is 8k long.
>
> 	if (metadump.show_progress) {
> 		static int64_t	mb_read;
> 		int64_t		mb_now = bytes_read >> 20;
>
> 		if (mb_now != mb_read) {
> 			print_progress("%lld MB read", mb_now);
> 			mb_read = mb_now;
> 		}
> 	}
>

I will include the above suggestion in the patchset.

>> +
>> +		offset = be64_to_cpu(xme.xme_addr) & XME_ADDR_DEVICE_MASK;
>> +		offset <<= BBSHIFT;
>
> offset = BBTOB(be64_to_cpu() ... ); ?

Yes, this is much better than what I had written.

>
> Also, I'd have thought that XME_ADDR_DEVICE_MASK is what you use to
> decode the device, not what you use to decode the address within a
> device.
>
>> +
>> +		if (be64_to_cpu(xme.xme_addr) & XME_ADDR_DATA_DEVICE)
>> +			fd = data_fd;
>> +		else if (be64_to_cpu(xme.xme_addr) & XME_ADDR_LOG_DEVICE)
>> +			fd = log_fd;
>> +		else
>> +			ASSERT(0);
>
> If you instead defined the constants like this:
>
> #define XME_ADDR_DEVICE_SHIFT	54
> #define XME_ADDR_DEVICE_MASK	((1ULL << XME_ADDR_DEVICE_SHIFT) - 1)

You probably meant to define XME_ADDR_DADDR_MASK.

> #define XME_ADDR_DATA_DEVICE	(0 << XME_ADDR_DEVICE_SHIFT)
> #define XME_ADDR_LOG_DEVICE	(1 << XME_ADDR_DEVICE_SHIFT)
> #define XME_ADDR_RT_DEVICE	(2 << XME_ADDR_DEVICE_SHIFT)
>
> #define XME_ADDR_DEVICE_MASK	(3 << XME_ADDR_DEVICE_SHIFT)
>
> Then the above becomes:
>
> 	offset = BBTOB(be64_to_cpu(xme.xme_addr) & XME_ADDR_DADDR_MASK);
>
> 	switch (be64_to_cpu(xme.xme_addr) & XME_ADDR_DEVICE_MASK) {
> 	case XME_ADDR_DATA_DEVICE:
> 		fd = data_fd;
> 		break;
> 	...
> 	}

Yes, this looks much better. Thanks for your suggestion.

>> +
>> +		if (pwrite(fd, block_buffer, len, offset) < 0)
>> +			fatal("error writing to %s device at offset %llu: %s\n",
>> +				fd == data_fd ? "data": "log", offset,
>> +				strerror(errno));
>> +
>> +                if (fread(&xme, sizeof(xme), 1, mdfp) != 1) {
>> +			if (feof(mdfp))
>> +				break;
>> +			fatal("error reading from metadump file\n");
>> +		}
>> +
>> +		prev_len = len;
>> +		len = be32_to_cpu(xme.xme_len);
>> +		len <<= BBSHIFT;
>> +		if (len > prev_len) {
>> +			void *p;
>> +			p = realloc(block_buffer, len);
>
> Would it be preferable to declare an 8MB buffer and only copy contents
> in that granularity?  Technically speaking, xme_len == -1U would require
> us to allocate a 2TB buffer, wouldn't it?
>

Yes, I agree. I will make the required changes.

>> +			if (p == NULL) {
>> +				free(block_buffer);
>> +				fatal("memory allocation failure\n");
>> +			}
>> +			block_buffer = p;
>> +		}
>> +
>> +		if (fread(block_buffer, len, 1, mdfp) != 1)
>> +			fatal("error reading from metadump file\n");
>> +
>> +		bytes_read += len;
>> +	} while (1);
>> +
>> +	if (mdrestore.progress_since_warning)
>> +		putchar('\n');
>> +
>> +        memset(block_buffer, 0, sb.sb_sectsize);
>
> Tabs not spaces.
>
>> +	sb.sb_inprogress = 0;
>> +	libxfs_sb_to_disk((struct xfs_dsb *)block_buffer, &sb);
>> +	if (xfs_sb_version_hascrc(&sb)) {
>> +		xfs_update_cksum(block_buffer, sb.sb_sectsize,
>> +				offsetof(struct xfs_sb, sb_crc));
>> +	}
>> +
>> +	if (pwrite(data_fd, block_buffer, sb.sb_sectsize, 0) < 0)
>> +		fatal("error writing primary superblock: %s\n",
>> +			strerror(errno));
>> +
>> +	free(block_buffer);
>> +
>> +	return;
>> +}
>> +
>> +static struct mdrestore_ops mdrestore_ops_v2 = {
>> +	.read_header = read_header_v2,
>> +	.show_info = show_info_v2,
>> +	.restore = restore_v2,
>> +};
>> +
>>  static void
>>  usage(void)
>>  {
>> @@ -276,11 +436,16 @@ main(
>>  	int 		argc,
>>  	char 		**argv)
>>  {
>> -	FILE		*src_f;
>> -	int		dst_fd;
>> -	int		c;
>> -	bool		is_target_file;
>> -	struct xfs_metablock	mb;
>> +	struct xfs_metadump_header	xmh;
>> +	struct xfs_metablock		mb;
>
> Hmm...
>
>> +	FILE				*src_f;
>> +	char				*logdev = NULL;
>> +	void				*header;
>> +	int				data_dev_fd;
>> +	int				log_dev_fd;
>> +	int				c;
>> +	bool				is_data_dev_file;
>> +	bool				is_log_dev_file;
>>  
>>  	mdrestore.show_progress = 0;
>>  	mdrestore.show_info = 0;
>> @@ -327,13 +492,18 @@ main(
>>  			fatal("cannot open source dump file\n");
>>  	}
>>  
>> -	if (mdrestore_ops_v1.read_header(&mb, src_f) == 0)
>> +	if (mdrestore_ops_v1.read_header(&mb, src_f) == 0) {
>>  		mdrestore.mdrops = &mdrestore_ops_v1;
>> -	else
>> +		header = &mb;
>> +	} else if (mdrestore_ops_v2.read_header(&xmh, src_f) == 0) {
>> +		mdrestore.mdrops = &mdrestore_ops_v2;
>> +		header = &xmh;
>
> Perhaps define a union of both header formats, then pass that to
> ->read_header, ->show_info, and ->restore?

Sure. I will do that.

-- 
chandan