Message ID | 20220628131032.213986-4-thuth@redhat.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | s390-ccw bios fixes and clean-ups | expand |
On Tue, 2022-06-28 at 15:10 +0200, Thomas Huth wrote: > The logic of trying an final ISO or ECKD boot on virtio-block devices > is > very weird: Since the geometry hardly ever matches in > virtio_disk_is_scsi(), > virtio_blk_setup_device() always sets a "guessed" disk geometry via > virtio_assume_scsi() (which is certainly also wrong in a lot of > cases). > > zipl_load_vblk() then sees that there's been a > "virtio_guessed_disk_nature" > and tries to fix up the geometry again via virtio_assume_iso9660() > before > always trying to do ipl_iso_el_torito(). That's a very brain-twisting > way of attempting to boot from ISO images, which won't work anymore > after > the following patches that will clean up the virtio_assume_scsi() > mess > (and thus get rid of the "virtio_guessed_disk_nature" here). > > Let's try a better approach instead: ISO files always have a magic > string "CD001" at offset 0x8001 (see e.g. the ECMA-119 specification) > which we can use to decide whether we should try to boot in ISO 9660 > mode (which we should also try if we see a sector size of 2048). > > And if we were not able to boot in ISO mode here, the final boot > attempt > before panicking is to boot in ECKD mode. Since this is our last boot > attempt anyway, simply always assume the ECKD geometry here (if the > sector > size was not 4096 yet), so that we also do not depend on the guessed > disk > geometry from virtio_blk_setup_device() here anymore. > > Signed-off-by: Thomas Huth <thuth@redhat.com> > --- > pc-bios/s390-ccw/bootmap.c | 25 +++++++++++++++++++++---- > 1 file changed, 21 insertions(+), 4 deletions(-) > > diff --git a/pc-bios/s390-ccw/bootmap.c b/pc-bios/s390-ccw/bootmap.c > index 56411ab3b6..3181b05382 100644 > --- a/pc-bios/s390-ccw/bootmap.c > +++ b/pc-bios/s390-ccw/bootmap.c > @@ -780,18 +780,35 @@ static void ipl_iso_el_torito(void) > } > } > > +/** > + * Detect whether we're trying to boot from an .ISO image. > + * These always have a signature string "CD001" at offset 0x8001. > + */ > +static bool has_iso_signature(void) > +{ > + if (virtio_read(0x8000 / virtio_get_block_size(), sec)) { Certainly unlikely, but virtio_get_block_size is able to return zero. > + return false; > + } > + > + return !memcmp("CD001", &sec[1], 5); > +} > + > /******************************************************************* > **** > * Bus specific IPL sequences > */ > > static void zipl_load_vblk(void) > { > - if (virtio_guessed_disk_nature()) { > - virtio_assume_iso9660(); > + int blksize = virtio_get_block_size(); > + > + if (blksize == VIRTIO_ISO_BLOCK_SIZE || has_iso_signature()) { > + if (blksize != VIRTIO_ISO_BLOCK_SIZE) { > + virtio_assume_iso9660(); > + } > + ipl_iso_el_torito(); > } > - ipl_iso_el_torito(); > > - if (virtio_guessed_disk_nature()) { > + if (blksize != VIRTIO_DASD_BLOCK_SIZE) { > sclp_print("Using guessed DASD geometry.\n"); > virtio_assume_eckd(); > }
On 01/07/2022 17.36, Eric Farman wrote: > On Tue, 2022-06-28 at 15:10 +0200, Thomas Huth wrote: >> The logic of trying an final ISO or ECKD boot on virtio-block devices >> is >> very weird: Since the geometry hardly ever matches in >> virtio_disk_is_scsi(), >> virtio_blk_setup_device() always sets a "guessed" disk geometry via >> virtio_assume_scsi() (which is certainly also wrong in a lot of >> cases). >> >> zipl_load_vblk() then sees that there's been a >> "virtio_guessed_disk_nature" >> and tries to fix up the geometry again via virtio_assume_iso9660() >> before >> always trying to do ipl_iso_el_torito(). That's a very brain-twisting >> way of attempting to boot from ISO images, which won't work anymore >> after >> the following patches that will clean up the virtio_assume_scsi() >> mess >> (and thus get rid of the "virtio_guessed_disk_nature" here). >> >> Let's try a better approach instead: ISO files always have a magic >> string "CD001" at offset 0x8001 (see e.g. the ECMA-119 specification) >> which we can use to decide whether we should try to boot in ISO 9660 >> mode (which we should also try if we see a sector size of 2048). >> >> And if we were not able to boot in ISO mode here, the final boot >> attempt >> before panicking is to boot in ECKD mode. Since this is our last boot >> attempt anyway, simply always assume the ECKD geometry here (if the >> sector >> size was not 4096 yet), so that we also do not depend on the guessed >> disk >> geometry from virtio_blk_setup_device() here anymore. >> >> Signed-off-by: Thomas Huth <thuth@redhat.com> >> --- >> pc-bios/s390-ccw/bootmap.c | 25 +++++++++++++++++++++---- >> 1 file changed, 21 insertions(+), 4 deletions(-) >> >> diff --git a/pc-bios/s390-ccw/bootmap.c b/pc-bios/s390-ccw/bootmap.c >> index 56411ab3b6..3181b05382 100644 >> --- a/pc-bios/s390-ccw/bootmap.c >> +++ b/pc-bios/s390-ccw/bootmap.c >> @@ -780,18 +780,35 @@ static void ipl_iso_el_torito(void) >> } >> } >> >> +/** >> + * Detect whether we're trying to boot from an .ISO image. >> + * These always have a signature string "CD001" at offset 0x8001. >> + */ >> +static bool has_iso_signature(void) >> +{ >> + if (virtio_read(0x8000 / virtio_get_block_size(), sec)) { > > Certainly unlikely, but virtio_get_block_size is able to return zero. Right ... I guess I'd better add a check for this here. Thanks, Thomas
diff --git a/pc-bios/s390-ccw/bootmap.c b/pc-bios/s390-ccw/bootmap.c index 56411ab3b6..3181b05382 100644 --- a/pc-bios/s390-ccw/bootmap.c +++ b/pc-bios/s390-ccw/bootmap.c @@ -780,18 +780,35 @@ static void ipl_iso_el_torito(void) } } +/** + * Detect whether we're trying to boot from an .ISO image. + * These always have a signature string "CD001" at offset 0x8001. + */ +static bool has_iso_signature(void) +{ + if (virtio_read(0x8000 / virtio_get_block_size(), sec)) { + return false; + } + + return !memcmp("CD001", &sec[1], 5); +} + /*********************************************************************** * Bus specific IPL sequences */ static void zipl_load_vblk(void) { - if (virtio_guessed_disk_nature()) { - virtio_assume_iso9660(); + int blksize = virtio_get_block_size(); + + if (blksize == VIRTIO_ISO_BLOCK_SIZE || has_iso_signature()) { + if (blksize != VIRTIO_ISO_BLOCK_SIZE) { + virtio_assume_iso9660(); + } + ipl_iso_el_torito(); } - ipl_iso_el_torito(); - if (virtio_guessed_disk_nature()) { + if (blksize != VIRTIO_DASD_BLOCK_SIZE) { sclp_print("Using guessed DASD geometry.\n"); virtio_assume_eckd(); }
The logic of trying an final ISO or ECKD boot on virtio-block devices is very weird: Since the geometry hardly ever matches in virtio_disk_is_scsi(), virtio_blk_setup_device() always sets a "guessed" disk geometry via virtio_assume_scsi() (which is certainly also wrong in a lot of cases). zipl_load_vblk() then sees that there's been a "virtio_guessed_disk_nature" and tries to fix up the geometry again via virtio_assume_iso9660() before always trying to do ipl_iso_el_torito(). That's a very brain-twisting way of attempting to boot from ISO images, which won't work anymore after the following patches that will clean up the virtio_assume_scsi() mess (and thus get rid of the "virtio_guessed_disk_nature" here). Let's try a better approach instead: ISO files always have a magic string "CD001" at offset 0x8001 (see e.g. the ECMA-119 specification) which we can use to decide whether we should try to boot in ISO 9660 mode (which we should also try if we see a sector size of 2048). And if we were not able to boot in ISO mode here, the final boot attempt before panicking is to boot in ECKD mode. Since this is our last boot attempt anyway, simply always assume the ECKD geometry here (if the sector size was not 4096 yet), so that we also do not depend on the guessed disk geometry from virtio_blk_setup_device() here anymore. Signed-off-by: Thomas Huth <thuth@redhat.com> --- pc-bios/s390-ccw/bootmap.c | 25 +++++++++++++++++++++---- 1 file changed, 21 insertions(+), 4 deletions(-)