[PATCH proxmox-backup] tape: workaround non-standard drive behavior for rewind

[PATCH proxmox-backup] tape: workaround non-standard drive behavior for rewind

[Dominik Csapak]

When the media label needs to be read, e.g. when a tape is inserted in
the drive freshly, a REWIND (01h) command is sent to the drive with the IMMED
bit set to 0, which means that the command should only return after the
rewind is finished.

For some drives, it seems that this is not the case and that some report
a 'rewind operation in progress' (ASC 0x00, ASCQ 0x1A) even after the
rewind command should be finished.

To workaround that, use the READ POSITION (34h) long form command
(reusing code from `position()`) to determine the current position and
wait until we reach the beginning of tape (position 0). If we encounter
the above mentioned error (or a 'becoming ready' error) during that
command, ignore it and try again.

Use the default scsi timeout (10minutes) for waiting, which should be
plenty (a rewind usually takes ~2 minutes).

Do this unconditionally, since for drives that behave as expected, the
one additional command after rewinding does not hurt, but it should fix
the issue for other drives.

While touching this code, also fix the typo 'detecthed'.

Signed-off-by: Dominik Csapak <d.csapak@proxmox.com>
---
 pbs-tape/src/sg_tape.rs  | 84 ++++++++++++++++++++++++++++++++++++----
 pbs-tape/src/sgutils2.rs | 14 +++++++
 2 files changed, 90 insertions(+), 8 deletions(-)

diff --git a/pbs-tape/src/sg_tape.rs b/pbs-tape/src/sg_tape.rs
index 04914e077..6df007b2d 100644
--- a/pbs-tape/src/sg_tape.rs
+++ b/pbs-tape/src/sg_tape.rs
@@ -41,7 +41,8 @@ use crate::{
     sgutils2::{
         InquiryInfo, ModeBlockDescriptor, ModeParameterHeader, ScsiError, SenseInfo, SgRaw,
         alloc_page_aligned_buffer, scsi_cmd_mode_select6, scsi_cmd_mode_select10, scsi_inquiry,
-        scsi_mode_sense, scsi_request_sense,
+        scsi_mode_sense, scsi_request_sense, sense_err_is_becoming_ready,
+        sense_err_is_rewind_in_progress,
     },
 };
 
@@ -358,9 +359,62 @@ impl SgTape {
             .do_command(&cmd)
             .map_err(|err| format_err!("rewind failed - {err}"))?;
 
+        // Some drives return from the non-immediate REWIND before it has
+        // actually finished (as if IMMED had been set), so the next command
+        // fails with "rewind operation in progress". Wait for the drive to
+        // reach beginning-of-tape before returning.
+        self.wait_until_rewound()?;
+
         Ok(())
     }
 
+    /// Wait until a rewind reached beginning-of-tape.
+    ///
+    /// Polls READ POSITION and busy-waits while the drive reports that the
+    /// rewind is still in progress (or that it is not ready yet). Returns as
+    /// soon as the drive reports beginning-of-tape, which happens on the first
+    /// iteration for a well-behaved drive.
+    fn wait_until_rewound(&mut self) -> Result<(), Error> {
+        log::debug!("waiting for drive to finish rewind");
+        let start = SystemTime::now();
+        let max_wait = std::time::Duration::new(Self::SCSI_TAPE_DEFAULT_TIMEOUT as u64, 0);
+
+        let mut last_msg: Option<String> = None;
+
+        loop {
+            match self.read_position_raw() {
+                Ok(data) => {
+                    let page = Self::decode_read_position(&data)?;
+                    let object_number = page.logical_object_number;
+                    if object_number == 0 {
+                        // beginning-of-tape reached, the rewind is done
+                        log::debug!("finished waiting for rewind");
+                        return Ok(());
+                    }
+                    // position readable, but not at beginning-of-tape yet
+                    log::debug!("rewind not done yet, at object position {object_number}");
+                }
+                Err(ScsiError::Sense(sense))
+                    if sense_err_is_rewind_in_progress(&sense)
+                        || sense_err_is_becoming_ready(&sense) =>
+                {
+                    let msg = sense.to_string();
+                    if last_msg.as_ref() != Some(&msg) {
+                        log::info!("waiting for rewind to complete - {msg}");
+                        last_msg = Some(msg);
+                    }
+                }
+                Err(err) => bail!("wait for rewind to complete failed - {err}"),
+            }
+
+            if start.elapsed()? > max_wait {
+                bail!("wait for rewind to complete failed - timeout");
+            }
+
+            std::thread::sleep(std::time::Duration::new(1, 0));
+        }
+    }
+
     #[allow(clippy::unusual_byte_groupings)]
     pub fn locate_file(&mut self, position: u64) -> Result<(), Error> {
         if position == 0 {
@@ -440,9 +494,11 @@ impl SgTape {
         Ok(())
     }
 
-    pub fn position(&mut self) -> Result<ReadPositionLongPage, Error> {
-        let expected_size = std::mem::size_of::<ReadPositionLongPage>();
-
+    /// Send READ POSITION LONG FORM and return the raw response.
+    ///
+    /// Returns the raw `ScsiError` so callers can inspect the sense data, for
+    /// example to busy-wait while a rewind is still in progress.
+    fn read_position_raw(&mut self) -> Result<Vec<u8>, ScsiError> {
         let mut sg_raw = SgRaw::new(&mut self.file, 32)?;
         sg_raw.set_timeout(30); // use short timeout
         let mut cmd = Vec::new();
@@ -451,9 +507,11 @@ impl SgTape {
         // reference manual.
         cmd.extend([0x34, 0x06, 0, 0, 0, 0, 0, 0, 0, 0]); // READ POSITION LONG FORM
 
-        let data = sg_raw
-            .do_command(&cmd)
-            .map_err(|err| format_err!("read position failed - {}", err))?;
+        Ok(sg_raw.do_command(&cmd)?.to_vec())
+    }
+
+    fn decode_read_position(data: &[u8]) -> Result<ReadPositionLongPage, Error> {
+        let expected_size = std::mem::size_of::<ReadPositionLongPage>();
 
         let page = proxmox_lang::try_block!({
             if data.len() != expected_size {
@@ -472,8 +530,18 @@ impl SgTape {
         })
         .map_err(|err: Error| format_err!("decode position page failed - {err}"))?;
 
+        Ok(page)
+    }
+
+    pub fn position(&mut self) -> Result<ReadPositionLongPage, Error> {
+        let data = self
+            .read_position_raw()
+            .map_err(|err| format_err!("read position failed - {err}"))?;
+
+        let page = Self::decode_read_position(&data)?;
+
         if page.partition_number != 0 {
-            bail!("detecthed partitioned tape - not supported");
+            bail!("detected partitioned tape - not supported");
         }
 
         Ok(page)
diff --git a/pbs-tape/src/sgutils2.rs b/pbs-tape/src/sgutils2.rs
index 340616c73..b778e85a5 100644
--- a/pbs-tape/src/sgutils2.rs
+++ b/pbs-tape/src/sgutils2.rs
@@ -900,6 +900,20 @@ pub fn sense_err_is_invalid_command(err: &SenseInfo) -> bool {
     err.sense_key == SENSE_KEY_ILLEGAL_REQUEST && err.asc == 0x20 && err.ascq == 0x00
 }
 
+/// True if the given sense info is REWIND OPERATION IN PROGRESS, i.e. the
+/// drive is still rewinding.
+/// <https://www.t10.org/lists/asc-num.htm#ASC_00>
+pub fn sense_err_is_rewind_in_progress(err: &SenseInfo) -> bool {
+    err.asc == 0x00 && err.ascq == 0x1A
+}
+
+/// True if the given sense info is LOGICAL UNIT IS IN PROCESS OF BECOMING
+/// READY, i.e. the drive is not ready yet but expected to become ready.
+/// <https://www.t10.org/lists/asc-num.htm#ASC_04>
+pub fn sense_err_is_becoming_ready(err: &SenseInfo) -> bool {
+    err.sense_key == SENSE_KEY_NOT_READY && err.asc == 0x04 && err.ascq == 0x01
+}
+
 /// Run SCSI Mode Sense - try Mode Sense(10) first, fallback to Mode Sense(6)
 ///
 /// Warning: P needs to be repr(C, packed)]
-- 
2.47.3

applied: [PATCH proxmox-backup] tape: workaround non-standard drive behavior for rewind

[Fabian Grünbichler]

On Wed, 01 Jul 2026 16:26:55 +0200, Dominik Csapak wrote:
> When the media label needs to be read, e.g. when a tape is inserted in
> the drive freshly, a REWIND (01h) command is sent to the drive with the IMMED
> bit set to 0, which means that the command should only return after the
> rewind is finished.
> 
> For some drives, it seems that this is not the case and that some report
> a 'rewind operation in progress' (ASC 0x00, ASCQ 0x1A) even after the
> rewind command should be finished.
> 
> [...]

Applied, thanks!

[1/1] tape: workaround non-standard drive behavior for rewind
      commit: a85064d6a700d54adacb112566f7856832cbd375

Best regards,
-- 
Fabian Grünbichler <f.gruenbichler@proxmox.com>

Re: applied: [PATCH proxmox-backup] tape: workaround non-standard drive behavior for rewind

[Dominik Csapak]

btw. this patch probably should not have been applied in this form.
dietmar wrote an answer but this didn't reach the mailing list

the root cause is probably another process/client interfering
with the tape drive, not a direct issue with the drive itself.
(and we can't safeguard against that)

while it'd do not much harm, i think it best to revert this again
(for now)

sorry for the inconvenience