[Qemu-devel] [PATCH 9/9] scsi-block: always use SG_IO for MMC devices

[Kevin Wolf <kwolf@redhat.com>]

From: Paolo Bonzini <pbonzini@redhat.com>

CD burning messes up the state of the host page cache and host block
device.  Just pass all operations down to the device, even though that
might have slightly worse performance.  Everything else just is not
reliable in combination with burning.

Reported-by: Thomas Schmitt <scdbackup@gmx.net>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
---
 hw/scsi-disk.c |   16 ++++++++++++++--
 1 files changed, 14 insertions(+), 2 deletions(-)

diff --git a/hw/scsi-disk.c b/hw/scsi-disk.c
index e3690ec..673948c 100644
--- a/hw/scsi-disk.c
+++ b/hw/scsi-disk.c
@@ -1703,8 +1703,20 @@ static SCSIRequest *scsi_block_new_request(SCSIDevice *d, uint32_t tag,
     case WRITE_VERIFY_10:
     case WRITE_VERIFY_12:
     case WRITE_VERIFY_16:
-        return scsi_req_alloc(&scsi_disk_reqops, &s->qdev, tag, lun,
-                              hba_private);
+        /* MMC writing cannot be done via pread/pwrite, because it sometimes
+         * involves writing beyond the maximum LBA or to negative LBA (lead-in).
+         * And once you do these writes, reading from the block device is
+         * unreliable, too.  It is even possible that reads deliver random data
+         * from the host page cache (this is probably a Linux bug).
+         *
+         * We might use scsi_disk_reqops as long as no writing commands are
+         * seen, but performance usually isn't paramount on optical media.  So,
+         * just make scsi-block operate the same as scsi-generic for them.
+         */
+        if (s->qdev.type != TYPE_ROM) {
+            return scsi_req_alloc(&scsi_disk_reqops, &s->qdev, tag, lun,
+                                  hba_private);
+        }
     }
 
     return scsi_req_alloc(&scsi_generic_req_ops, &s->qdev, tag, lun,
-- 
1.7.6.4