Files
Shashank Mohan Jain 94defd41d8 errseq: don't let errseq_check_and_advance() hide later errors
errseq_check_and_advance() reads the errseq_t, sets ERRSEQ_SEEN with a
cmpxchg() and then advances *since to the value it computed, ignoring
whether the cmpxchg() succeeded. When it fails because errseq_set()
recorded a different error in the meantime, *since ends up holding a
value that was never stored in the errseq_t.

errseq_set() only bumps the counter when ERRSEQ_SEEN is set. So as long
as nobody has seen the new error, recording the original errno again
recreates exactly the old value, and once another subscriber marks it
seen, it is equal to the stale *since. The next check through that
cursor then reports nothing, although -ENOSPC was recorded after the
value that check reported, and -EIO was recorded again after it
returned:

  cursor f                writeback              cursor g
  --------                ---------              --------
                          set(-EIO)
  check_and_advance(&f)
    old = [c, EIO]
                          set(-ENOSPC)
                          -> [c, ENOSPC]
    cmpxchg() fails
    f = [c, EIO, SEEN]
    returns -EIO
                          set(-EIO)
                          -> [c, EIO]
                          (no bump: unseen)
                                                 check_and_advance(&g)
                                                 -> [c, EIO, SEEN]
                                                 returns -EIO
  check_and_advance(&f)
    [c, EIO, SEEN] == f
    returns 0             <- -ENOSPC and the second -EIO are lost

For file->f_wb_err this means that fsync() on one descriptor can return
0 although writeback failed after the previous fsync() on that
descriptor returned, when writeback errors race with fsync() on another
descriptor of the same file. The same applies to syncfs() through
sb->s_wb_err, which every file on the filesystem shares, and to the
ext4 and jbd2 cursors on the block device mapping. The kernel-doc
promises "Negative errno if one has been stored, or 0 if no new error
has occurred".

Retry with the value found when the cmpxchg() fails, so that *since is
only ever advanced to a value that was actually stored with ERRSEQ_SEEN
set. Every later errseq_set() then has to bump the counter, and the
error is reported. The retry loop terminates because each iteration
needs a concurrent update, like the loop in errseq_set().

A TLA+ model of errseq_set() and errseq_check_and_advance(), checked
with the TLC model checker, finds the lost error with the current code;
with this change TLC checks that an error recorded after a check
returned is always reported by the next check (exhaustively for two
subscribers with three checks each, and one writer recording four
errors or two writers recording two each). The KUnit race test added
in the next patch misses the error in 1-6% of 2 million rounds on
4-CPU UML before this change (depending on host load), and in none
after it.

Fixes: 84cbadadc6 ("lib: add errseq_t type and infrastructure for handling it")
Cc: [email protected]
Signed-off-by: Shashank Mohan Jain <[email protected]>
Link: https://patch.msgid.link/[email protected]
Reviewed-by: Jan Kara <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Christian Brauner (Amutable) <[email protected]>
2026-11-05 18:21:53 +01:00

216 lines
6.8 KiB
C

// SPDX-License-Identifier: GPL-2.0
#include <linux/err.h>
#include <linux/bug.h>
#include <linux/atomic.h>
#include <linux/errseq.h>
#include <linux/log2.h>
/*
* An errseq_t is a way of recording errors in one place, and allowing any
* number of "subscribers" to tell whether it has changed since a previous
* point where it was sampled.
*
* It's implemented as an unsigned 32-bit value. The low order bits are
* designated to hold an error code (between 0 and -MAX_ERRNO). The upper bits
* are used as a counter. This is done with atomics instead of locking so that
* these functions can be called from any context.
*
* The general idea is for consumers to sample an errseq_t value. That value
* can later be used to tell whether any new errors have occurred since that
* sampling was done.
*
* Note that there is a risk of collisions if new errors are being recorded
* frequently, since we have so few bits to use as a counter.
*
* To mitigate this, one bit is used as a flag to tell whether the value has
* been sampled since a new value was recorded. That allows us to avoid bumping
* the counter if no one has sampled it since the last time an error was
* recorded.
*
* A new errseq_t should always be zeroed out. A errseq_t value of all zeroes
* is the special (but common) case where there has never been an error. An all
* zero value thus serves as the "epoch" if one wishes to know whether there
* has ever been an error set since it was first initialized.
*/
/* The low bits are designated for error code (max of MAX_ERRNO) */
#define ERRSEQ_SHIFT (ilog2(MAX_ERRNO) + 1)
/* This bit is used as a flag to indicate whether the value has been seen */
#define ERRSEQ_SEEN (1 << ERRSEQ_SHIFT)
/* Leverage macro ERRSEQ_SEEN to define errno mask macro here */
#define ERRNO_MASK (ERRSEQ_SEEN - 1)
/* The lowest bit of the counter */
#define ERRSEQ_CTR_INC (1 << (ERRSEQ_SHIFT + 1))
/**
* errseq_set - set a errseq_t for later reporting
* @eseq: errseq_t field that should be set
* @err: error to set (must be between -1 and -MAX_ERRNO)
*
* This function sets the error in @eseq, and increments the sequence counter
* if the last sequence was sampled at some point in the past.
*
* Any error set will always overwrite an existing error.
*
* Return: The previous value, primarily for debugging purposes. The
* return value should not be used as a previously sampled value in later
* calls as it will not have the SEEN flag set.
*/
errseq_t errseq_set(errseq_t *eseq, int err)
{
errseq_t cur, old;
/*
* Ensure the error code actually fits where we want it to go. If it
* doesn't then just throw a warning and don't record anything. We
* also don't accept zero here as that would effectively clear a
* previous error.
*/
old = READ_ONCE(*eseq);
if (WARN(unlikely(err == 0 || (unsigned int)-err > MAX_ERRNO),
"err = %d\n", err))
return old;
for (;;) {
errseq_t new;
/* Clear out error bits and set new error */
new = (old & ~(ERRNO_MASK | ERRSEQ_SEEN)) | -err;
/* Only increment if someone has looked at it */
if (old & ERRSEQ_SEEN)
new += ERRSEQ_CTR_INC;
/* If there would be no change, then call it done */
if (new == old) {
cur = new;
break;
}
/* Try to swap the new value into place */
cur = cmpxchg(eseq, old, new);
/*
* Call it success if we did the swap or someone else beat us
* to it for the same value.
*/
if (likely(cur == old || cur == new))
break;
/* Raced with an update, try again */
old = cur;
}
return cur;
}
EXPORT_SYMBOL(errseq_set);
/**
* errseq_sample() - Grab current errseq_t value.
* @eseq: Pointer to errseq_t to be sampled.
*
* This function allows callers to initialise their errseq_t variable.
* If the error has been "seen", new callers will not see an old error.
* If there is an unseen error in @eseq, the caller of this function will
* see it the next time it checks for an error.
*
* Context: Any context.
* Return: The current errseq value.
*/
errseq_t errseq_sample(errseq_t *eseq)
{
errseq_t old = READ_ONCE(*eseq);
/* If nobody has seen this error yet, then we can be the first. */
if (!(old & ERRSEQ_SEEN))
old = 0;
return old;
}
EXPORT_SYMBOL(errseq_sample);
/**
* errseq_check() - Has an error occurred since a particular sample point?
* @eseq: Pointer to errseq_t value to be checked.
* @since: Previously-sampled errseq_t from which to check.
*
* Grab the value that eseq points to, and see if it has changed @since
* the given value was sampled. The @since value is not advanced, so there
* is no need to mark the value as seen.
*
* Return: The latest error set in the errseq_t or 0 if it hasn't changed.
*/
int errseq_check(errseq_t *eseq, errseq_t since)
{
errseq_t cur = READ_ONCE(*eseq);
if (likely(cur == since))
return 0;
return -(cur & ERRNO_MASK);
}
EXPORT_SYMBOL(errseq_check);
/**
* errseq_check_and_advance() - Check an errseq_t and advance to current value.
* @eseq: Pointer to value being checked and reported.
* @since: Pointer to previously-sampled errseq_t to check against and advance.
*
* Grab the eseq value, and see whether it matches the value that @since
* points to. If it does, then just return 0.
*
* If it doesn't, then the value has changed. Set the "seen" flag, and try to
* swap it into place as the new eseq value. If the swap fails because the
* value changed, retry with the new value. Then, set that value as the new
* "since" value, and return whatever the error portion is set to.
*
* Note that no locking is provided here for concurrent updates to the "since"
* value. The caller must provide that if necessary. Because of this, callers
* may want to do a lockless errseq_check before taking the lock and calling
* this.
*
* Return: Negative errno if one has been stored, or 0 if no new error has
* occurred.
*/
int errseq_check_and_advance(errseq_t *eseq, errseq_t *since)
{
int err = 0;
errseq_t old, new;
/*
* Most callers will want to use the inline wrapper to check this,
* so that the common case of no error is handled without needing
* to take the lock that protects the "since" value.
*/
old = READ_ONCE(*eseq);
while (old != *since) {
/*
* Set the flag and try to swap it into place if it has
* changed.
*
* If the swap fails, a writer or another reader changed the
* value under us, so retry with the new value. "since" must
* only ever be set to a value that was stored with the SEEN
* flag: errseq_set() does not bump the counter while the
* flag is clear, so an unstored value could come back later
* and hide the errors that were recorded in the meantime.
*/
new = old | ERRSEQ_SEEN;
if (new != old) {
errseq_t cur = cmpxchg(eseq, old, new);
if (cur != old) {
old = cur;
continue;
}
}
*since = new;
err = -(new & ERRNO_MASK);
break;
}
return err;
}
EXPORT_SYMBOL(errseq_check_and_advance);