options: add support more POSIX errnos with negative support#2001
options: add support more POSIX errnos with negative support#2001minwooim wants to merge 3 commits intoaxboe:masterfrom
Conversation
Added more error numbers(errno) after ERANGE to support various errno string to options like `--ignore_error=ETIMEDOUT`. unvme-cli libunvmed ioengine returns ETIMEDOUT if a command is timed out. To mask this situation with `--ignore_error=` option, errno after ERANGE should be supported in `str2errr()`. Signed-off-by: Minwoo Im <minwoo.im@samsung.com>
Some of ioengines (e.g., io_uring_cmd) returns negative errno to represent system error as negative errno instead of positive values for NVMe-specific error status. To masking expected situations with --ignore_error= option, added support for negative value of errno syntax. Signed-off-by: Minwoo Im <minwoo.im@samsung.com>
IO_U_F_DEVICE_ERROR has been introduced in Commit ebe67b6 ("io_uring: Add IO_U_F_DEVICE_ERROR to identify error types") to parse two different types of errors can happen: (1) device error(e.g., NVMe-specific error status code) and (2) system error(e.g., EINVAL). `--ignore_error=` option let user mask expected situation, but if user wants to mask system error rather than device error, no way to represent the system error except for such as -EINVAL. Even if user put -EINVAL to the option, it will be checked as a positive value since @io_u->error will be set with abs() in io_uring_cmd ioengine. This patch flips the given @io_u->error positive value to a negative value if the given @io_u->flags represents system error by IO_U_F_DEVICE_ERROR. Signed-off-by: Minwoo Im <minwoo.im@samsung.com>
vincentkfu
left a comment
vincentkfu
left a comment
There was a problem hiding this comment.
This change concerns me because it changes the default behavior for error handling. Most of the time io_u->error will be a system error since only io_uring_cmd sets IO_U_F_DEVICE_ERROR. So this will flip the sign of the error in most cases. This will break jobs that depend on fio's current behavior.
Why can't you just use --ignore_error=EINVAL to mask EINVAL instead of relying on this patch and using --ignore_error=-EINVAL?
|
The reason I wanted to mask with -EINVAL was exactly what you mentioned: with io_uring_cmd, the NVMe CQ entry’s status code can currently be masked through io_u->error. But you’re right — flipping the error-handling sign for all other cases just to accommodate a single ioengine shouldn’t happen.
The problem is that if IO_U_F_DEVICE_ERROR is set and io_u->error contains the value 22 (0x16), we can’t tell whether that represents EINVAL (22) or an NVMe SGL Offset Invalid error (sct=0x0, sc=0x16). |
|
Would you like to split this PR into two pieces? One patch adding more errno values can go in with no problems, but a second one to distinguish between errno values and device errors will require more thought. |
Will do that, @vincentkfu , Thanks for the review. |
2 participants
Patch series to support masking expected system error with
--ignore_error=.