Noticed an interesting bug with test generic/586 where the file size can be reset smaller (shrinking the file) due to being sent too late (after the last write, instead of before it). Trying it to Samba e.g. if I use default mount parms it works, I see write of 1MB at 1MB, then filesize shrinks to 1MB, then write of 1MB against at 1MB so file size up at its expected value (2MB) at end of test, but when mounting with cifsacl (perhaps due to lease breaks) I see different ordering of the final write and set file size so the final set file size shrinks it to 1MB. The test is intended to test: "Race an appending aio dio write to the second block of a file while simultaneously fallocating to the first block." so it is plausible that we are missing a lock somewhere. Any ideas? -- Thanks, Steve
