| Message ID | 20200319123006.37578-3-wqu@suse.com |
|---|---|
| State | Superseded |
| Headers | show |
| Series | uboot: fs/btrfs: Fix read error on LZO compressed extents | expand |
Hi Qu, On 19/03/2020 13:30, Qu Wenruo wrote: > [BUG] > For certain btrfs files with compressed file extent, uboot will fail to > load it: > > btrfs_read_extent_reg: disk_bytenr=14229504 disk_len=73728 offset=0 nr_bytes=131 > 072 > decompress_lzo: tot_len=70770 > decompress_lzo: in_len=1389 > decompress_lzo: in_len=2400 > decompress_lzo: in_len=3002 > decompress_lzo: in_len=1379 > decompress_lzo: in_len=88539136 > decompress_lzo: header error, in_len=88539136 clen=65534 tot_len=62580 > > NOTE: except the last line, all other lines are debug output. > > [CAUSE] > Btrfs lzo compression uses its own format to record compressed size > (segmant header, LE32). > > However to make decompression easier, we never put such segment header > across page boundary. > > In above case, the xxd dump of the lzo compressed data looks like this: > > 00001fe0: 4cdc 02fc 0bfd 02c0 dc02 0d13 0100 0001 L............... > 00001ff0: 0000 0008 0300 0000 0000 0011 0000|0000 ................ > 00002000: 4705 0000 0001 cc02 0000 0000 0000 1e01 G............... > > '|' is the "expected" segment header start position. > > But in that page, there are only 2 bytes left, can't contain the 4 bytes > segment header. > > So btrfs compression will skip that 2 bytes, put the segment header in > next page directly. > > Uboot doesn't have such check, and read the header with 2 bytes offset, > result 0x05470000 (88539136), other than the expected result > 0x00000547 (1351), resulting above error. > > [FIX] > Follow the btrfs-progs restore implementation, by introducing tot_in to > record total processed bytes (including headers), and do proper page > boundary skip to fix it. > > Signed-off-by: Qu Wenruo <wqu at suse.com> > --- > fs/btrfs/compression.c | 20 ++++++++++++++++++++ > 1 file changed, 20 insertions(+) > > diff --git a/fs/btrfs/compression.c b/fs/btrfs/compression.c > index 4ef44ce11485..2a6ac8bb1029 100644 > --- a/fs/btrfs/compression.c > +++ b/fs/btrfs/compression.c > @@ -9,6 +9,7 @@ > #include <malloc.h> > #include <linux/lzo.h> > #include <linux/zstd.h> > +#include <linux/compat.h> > #include <u-boot/zlib.h> > #include <asm/unaligned.h> > > @@ -17,6 +18,7 @@ > static u32 decompress_lzo(const u8 *cbuf, u32 clen, u8 *dbuf, u32 dlen) > { > u32 tot_len, in_len, res; > + u32 tot_in = 0; > size_t out_len; > int ret; > > @@ -27,6 +29,7 @@ static u32 decompress_lzo(const u8 *cbuf, u32 clen, u8 *dbuf, u32 dlen) > cbuf += LZO_LEN; > clen -= LZO_LEN; > tot_len -= LZO_LEN; > + tot_in += LZO_LEN; > > if (tot_len == 0 && dlen) > return -1; > @@ -36,6 +39,9 @@ static u32 decompress_lzo(const u8 *cbuf, u32 clen, u8 *dbuf, u32 dlen) > res = 0; > > while (tot_len > LZO_LEN) { > + size_t mod_page; > + size_t rem_page; > + > in_len = le32_to_cpu(get_unaligned((u32 *)cbuf)); > cbuf += LZO_LEN; > clen -= LZO_LEN; > @@ -44,6 +50,7 @@ static u32 decompress_lzo(const u8 *cbuf, u32 clen, u8 *dbuf, u32 dlen) > return -1; > > tot_len -= (LZO_LEN + in_len); > + tot_in += (LZO_LEN + in_len); > > out_len = dlen; > ret = lzo1x_decompress_safe(cbuf, in_len, dbuf, &out_len); > @@ -56,6 +63,19 @@ static u32 decompress_lzo(const u8 *cbuf, u32 clen, u8 *dbuf, u32 dlen) > dlen -= out_len; > > res += out_len; > + > + /* > + * If the 4 bytes header does not fit to the rest of the page we > + * have to move to next one, or we read some garbage. > + */ > + mod_page = tot_in % PAGE_SIZE; in U-Boot we use 4K page sizes, but the OS could use another page size (16K or 64k). Would we need to adapt that code to reflect which page size is used on the medium we want to access? Regards, Matthias > + rem_page = PAGE_SIZE - mod_page; > + if (rem_page < LZO_LEN) { > + cbuf += rem_page; > + tot_in += rem_page; > + clen -= rem_page; > + tot_len -= rem_page; > + } > } > > return res; >
On Thu, Mar 19, 2020 at 02:33:28PM +0100, Matthias Brugger wrote: > > dlen -= out_len; > > > > res += out_len; > > + > > + /* > > + * If the 4 bytes header does not fit to the rest of the page we > > + * have to move to next one, or we read some garbage. > > + */ > > + mod_page = tot_in % PAGE_SIZE; > > in U-Boot we use 4K page sizes, but the OS could use another page size (16K or > 64k). Would we need to adapt that code to reflect which page size is used on the > medium we want to access? Yes, it is the 'sectorsize' as it's set up in fs_info or it's equivalent in uboot. For kernel the page size == sectorsize is kind of implicit and verified at mount time.
On 19/03/2020 14:56, David Sterba wrote: > On Thu, Mar 19, 2020 at 02:33:28PM +0100, Matthias Brugger wrote: >>> dlen -= out_len; >>> >>> res += out_len; >>> + >>> + /* >>> + * If the 4 bytes header does not fit to the rest of the page we >>> + * have to move to next one, or we read some garbage. >>> + */ >>> + mod_page = tot_in % PAGE_SIZE; >> >> in U-Boot we use 4K page sizes, but the OS could use another page size (16K or >> 64k). Would we need to adapt that code to reflect which page size is used on the >> medium we want to access? > > Yes, it is the 'sectorsize' as it's set up in fs_info or it's equivalent > in uboot. For kernel the page size == sectorsize is kind of implicit and > verified at mount time. > Does this mean we would need to add a Kconfig option to set the sectorsize in U-Boot? Regards, Matthias
On Thu, Mar 19, 2020 at 03:34:12PM +0100, Matthias Brugger wrote: > > > On 19/03/2020 14:56, David Sterba wrote: > > On Thu, Mar 19, 2020 at 02:33:28PM +0100, Matthias Brugger wrote: > >>> dlen -= out_len; > >>> > >>> res += out_len; > >>> + > >>> + /* > >>> + * If the 4 bytes header does not fit to the rest of the page we > >>> + * have to move to next one, or we read some garbage. > >>> + */ > >>> + mod_page = tot_in % PAGE_SIZE; > >> > >> in U-Boot we use 4K page sizes, but the OS could use another page size (16K or > >> 64k). Would we need to adapt that code to reflect which page size is used on the > >> medium we want to access? > > > > Yes, it is the 'sectorsize' as it's set up in fs_info or it's equivalent > > in uboot. For kernel the page size == sectorsize is kind of implicit and > > verified at mount time. > > > > Does this mean we would need to add a Kconfig option to set the sectorsize in > U-Boot? No, the value depends on the filesystem so it can't be a config option. What I mean is btrfs_super_block::sectorsize, where the superblock is btrfs_info::sb.
On 2020/3/20 ??12:28, David Sterba wrote: > On Thu, Mar 19, 2020 at 03:34:12PM +0100, Matthias Brugger wrote: >> >> >> On 19/03/2020 14:56, David Sterba wrote: >>> On Thu, Mar 19, 2020 at 02:33:28PM +0100, Matthias Brugger wrote: >>>>> dlen -= out_len; >>>>> >>>>> res += out_len; >>>>> + >>>>> + /* >>>>> + * If the 4 bytes header does not fit to the rest of the page we >>>>> + * have to move to next one, or we read some garbage. >>>>> + */ >>>>> + mod_page = tot_in % PAGE_SIZE; >>>> >>>> in U-Boot we use 4K page sizes, but the OS could use another page size (16K or >>>> 64k). Would we need to adapt that code to reflect which page size is used on the >>>> medium we want to access? >>> >>> Yes, it is the 'sectorsize' as it's set up in fs_info or it's equivalent >>> in uboot. For kernel the page size == sectorsize is kind of implicit and >>> verified at mount time. >>> >> >> Does this mean we would need to add a Kconfig option to set the sectorsize in >> U-Boot? > > No, the value depends on the filesystem so it can't be a config option. > What I mean is btrfs_super_block::sectorsize, where the superblock is > btrfs_info::sb. > Sorry for the delayed reply. (Stupid filter setup). Currently most Uboot boards should use the same page size setup for its kernel, and most btrfs uses 4K sector size. So for Uboot it should be no problem. Although the best practice is to read the fs_info::sectorsize as David mentioned, but the code base doesn't allow us to do that yet. So I'm going to backport the read part code from btrfs-progs in the near-future, and completely solve it, making it sector size independent. Would this plan looks sound? Or we need to wait for the full re-implementation? Thanks, Qu -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.asc Type: application/pgp-signature Size: 488 bytes Desc: OpenPGP digital signature URL: <https://lists.denx.de/pipermail/u-boot/attachments/20200324/2e2c9232/attachment.sig>
On 2020/3/20 ??12:28, David Sterba wrote: > On Thu, Mar 19, 2020 at 03:34:12PM +0100, Matthias Brugger wrote: >> >> >> On 19/03/2020 14:56, David Sterba wrote: >>> On Thu, Mar 19, 2020 at 02:33:28PM +0100, Matthias Brugger wrote: >>>>> dlen -= out_len; >>>>> >>>>> res += out_len; >>>>> + >>>>> + /* >>>>> + * If the 4 bytes header does not fit to the rest of the page we >>>>> + * have to move to next one, or we read some garbage. >>>>> + */ >>>>> + mod_page = tot_in % PAGE_SIZE; >>>> >>>> in U-Boot we use 4K page sizes, but the OS could use another page size (16K or >>>> 64k). Would we need to adapt that code to reflect which page size is used on the >>>> medium we want to access? >>> >>> Yes, it is the 'sectorsize' as it's set up in fs_info or it's equivalent >>> in uboot. For kernel the page size == sectorsize is kind of implicit and >>> verified at mount time. >>> >> >> Does this mean we would need to add a Kconfig option to set the sectorsize in >> U-Boot? > > No, the value depends on the filesystem so it can't be a config option. > What I mean is btrfs_super_block::sectorsize, where the superblock is > btrfs_info::sb. > Sorry for the delayed reply. (Stupid filter setup). Currently most Uboot boards should use the same page size setup for its kernel, and most btrfs uses 4K sector size. So for Uboot it should be no problem. Although the best practice is to read the fs_info::sectorsize as David mentioned, but the code base doesn't allow us to do that yet. So I'm going to backport the read part code from btrfs-progs in the near-future, and completely solve it, making it sector size independent. Would this plan look OK to you? Or we need to wait for the full re-implementation? Thanks, Qu -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.asc Type: application/pgp-signature Size: 488 bytes Desc: OpenPGP digital signature URL: <https://lists.denx.de/pipermail/u-boot/attachments/20200324/22e935bd/attachment.sig>
On Tue, 24 Mar 2020 19:03:30 +0800 Qu Wenruo <quwenruo.btrfs at gmx.com> wrote: > Sorry for the delayed reply. (Stupid filter setup). > > Currently most Uboot boards should use the same page size setup for its > kernel, and most btrfs uses 4K sector size. > > So for Uboot it should be no problem. > > Although the best practice is to read the fs_info::sectorsize as David > mentioned, but the code base doesn't allow us to do that yet. > > So I'm going to backport the read part code from btrfs-progs in the > near-future, and completely solve it, making it sector size independent. > > Would this plan looks sound? Or we need to wait for the full > re-implementation? > > Thanks, > Qu > The situation is Linux is such that btrfs sectorsize must be same as PAGE_SIZE, otherwise the Linux btrfs driver won't work. AFAIK there are only few architectures where PAGE_SIZE is not 4 KiB. btrfs filesystems created there cannot be mounted on systems with PAGE_SIZE = 4 KiB. I don't know if U-Boot is used on non 4KiB PAGE_SIZE boards. If it is, it should be solved, but I would check that before complicating the code.
diff --git a/fs/btrfs/compression.c b/fs/btrfs/compression.c index 4ef44ce11485..2a6ac8bb1029 100644 --- a/fs/btrfs/compression.c +++ b/fs/btrfs/compression.c @@ -9,6 +9,7 @@ #include <malloc.h> #include <linux/lzo.h> #include <linux/zstd.h> +#include <linux/compat.h> #include <u-boot/zlib.h> #include <asm/unaligned.h> @@ -17,6 +18,7 @@ static u32 decompress_lzo(const u8 *cbuf, u32 clen, u8 *dbuf, u32 dlen) { u32 tot_len, in_len, res; + u32 tot_in = 0; size_t out_len; int ret; @@ -27,6 +29,7 @@ static u32 decompress_lzo(const u8 *cbuf, u32 clen, u8 *dbuf, u32 dlen) cbuf += LZO_LEN; clen -= LZO_LEN; tot_len -= LZO_LEN; + tot_in += LZO_LEN; if (tot_len == 0 && dlen) return -1; @@ -36,6 +39,9 @@ static u32 decompress_lzo(const u8 *cbuf, u32 clen, u8 *dbuf, u32 dlen) res = 0; while (tot_len > LZO_LEN) { + size_t mod_page; + size_t rem_page; + in_len = le32_to_cpu(get_unaligned((u32 *)cbuf)); cbuf += LZO_LEN; clen -= LZO_LEN; @@ -44,6 +50,7 @@ static u32 decompress_lzo(const u8 *cbuf, u32 clen, u8 *dbuf, u32 dlen) return -1; tot_len -= (LZO_LEN + in_len); + tot_in += (LZO_LEN + in_len); out_len = dlen; ret = lzo1x_decompress_safe(cbuf, in_len, dbuf, &out_len); @@ -56,6 +63,19 @@ static u32 decompress_lzo(const u8 *cbuf, u32 clen, u8 *dbuf, u32 dlen) dlen -= out_len; res += out_len; + + /* + * If the 4 bytes header does not fit to the rest of the page we + * have to move to next one, or we read some garbage. + */ + mod_page = tot_in % PAGE_SIZE; + rem_page = PAGE_SIZE - mod_page; + if (rem_page < LZO_LEN) { + cbuf += rem_page; + tot_in += rem_page; + clen -= rem_page; + tot_len -= rem_page; + } } return res;
[BUG] For certain btrfs files with compressed file extent, uboot will fail to load it: btrfs_read_extent_reg: disk_bytenr=14229504 disk_len=73728 offset=0 nr_bytes=131 072 decompress_lzo: tot_len=70770 decompress_lzo: in_len=1389 decompress_lzo: in_len=2400 decompress_lzo: in_len=3002 decompress_lzo: in_len=1379 decompress_lzo: in_len=88539136 decompress_lzo: header error, in_len=88539136 clen=65534 tot_len=62580 NOTE: except the last line, all other lines are debug output. [CAUSE] Btrfs lzo compression uses its own format to record compressed size (segmant header, LE32). However to make decompression easier, we never put such segment header across page boundary. In above case, the xxd dump of the lzo compressed data looks like this: 00001fe0: 4cdc 02fc 0bfd 02c0 dc02 0d13 0100 0001 L............... 00001ff0: 0000 0008 0300 0000 0000 0011 0000|0000 ................ 00002000: 4705 0000 0001 cc02 0000 0000 0000 1e01 G............... '|' is the "expected" segment header start position. But in that page, there are only 2 bytes left, can't contain the 4 bytes segment header. So btrfs compression will skip that 2 bytes, put the segment header in next page directly. Uboot doesn't have such check, and read the header with 2 bytes offset, result 0x05470000 (88539136), other than the expected result 0x00000547 (1351), resulting above error. [FIX] Follow the btrfs-progs restore implementation, by introducing tot_in to record total processed bytes (including headers), and do proper page boundary skip to fix it. Signed-off-by: Qu Wenruo <wqu at suse.com> --- fs/btrfs/compression.c | 20 ++++++++++++++++++++ 1 file changed, 20 insertions(+)