* y/nodefs-cancel:
  Y nodefs: Propagate context to File methods
  fuse: dump WRITE argument in debug logs
  Make pollHack work on read-only mounts
  example: loopback: implement -ro flag
  fs: add TestRoMount
  posixtest: add DirSeek test, Go port of xfstests generic/257
  fs: bridge: implement simple directory seeking
  fs: cleanup after TestForget

So that e.g. File.Read could be canceled.

This is already so in fs/ package. However WCFS in Wendelin.core still
uses older nodefs package, and for now it was faster to teach old
nodefs.File about cancel propagation.

References:

wendelin.core@b17aeb8c
wendelin.core@f05271b1
wendelin.core@5ba816da
go123@7ad867a3
go123@0bdac628
go123@d2dc6c09

Was proposed to upstream in FYI mode:

https://github.com/hanwen/go-fuse/pull/343
https://review.gerrithub.io/c/hanwen/go-fuse/+/488787

Change-Id: Ia9f49f04d87ad3d42f1e5b14c62fe9b425f404a9

This provides feedback to file system authors about what we receive
from the kernel. This can help figure out which system is to blame for
data corruption.

Change-Id: I7cd7f340bf705c9972c95cfc6d3d8612f1ce8b34

Use LOOKUP & OPEN instead of CREATE so it also
works on read-only mounts.

go-fuse/fs$ go test -run TestRoMount -v
=== RUN   TestRoMount
20:31:07.892736 rx 2: INIT i0 {7.31 Ra 0x20000 NO_OPEN_SUPPORT,ASYNC_READ,POSIX_LOCKS,EXPORT_SUPPORT,AUTO_INVAL_DATA,SPLICE_MOVE,SPLICE_READ,HANDLE_KILLPRIV,ABORT_ERROR,EXPLICIT_INVAL_DATA,BIG_WRITES,SPLICE_WRITE,READDIRPLUS,WRITEBACK_CACHE,IOCTL_DIR,READDIRPLUS_AUTO,ASYNC_DIO,PARALLEL_DIROPS,ATOMIC_O_TRUNC,DONT_MASK,FLOCK_LOCKS,POSIX_ACL,MAX_PAGES,CACHE_SYMLINKS,NO_OPENDIR_SUPPORT}
20:31:07.892799 tx 2:     OK, {7.28 Ra 0x20000 BIG_WRITES,READDIRPLUS,PARALLEL_DIROPS,ASYNC_READ,AUTO_INVAL_DATA,NO_OPEN_SUPPORT 0/0 Wr 0x10000 Tg 0x0}
20:31:07.892967 rx 4: LOOKUP i1 [".go-fuse-epoll-hack"] 20b
20:31:07.893002 tx 4:     OK, {i18446744073709551615 g0 tE=0s tA=0s {M0100644 SZ=0 L=1 0:0 B0*0 i0:18446744073709551615 A 0.000000 M 0.000000 C 0.000000}}
20:31:07.893030 rx 6: OPEN i18446744073709551615 {O_RDONLY,0x8000}
20:31:07.893037 tx 6:     OK, {Fh 18446744073709551615 }
20:31:07.893053 rx 8: POLL i18446744073709551615
20:31:07.893059 tx 8:     38=function not implemented
20:31:07.893074 rx 10: FLUSH i18446744073709551615 {Fh 18446744073709551615}
20:31:07.893079 tx 10:     34=numerical result out of range
20:31:07.893131 rx 12: RELEASE i18446744073709551615 {Fh 18446744073709551615 0x8000  L0}
20:31:07.893153 tx 12:     34=numerical result out of range
20:31:07.896705 received ENODEV (unmount request), thread exiting
20:31:07.896719 received ENODEV (unmount request), thread exiting
20:31:07.896888 received ENODEV (unmount request), thread exiting

Change-Id: I6de861bdccf03184a4ca3deeefb1cfdc930aabe3
--- PASS: TestRoMount (0.01s)
PASS
ok  	github.com/hanwen/go-fuse/v2/fs	0.015s

Change-Id: Ic66ca153f5b739e5e081988fdc48d51cde5470f0

Currently fails like this

	18:13:57.008805 Mount fail: read-only file system

because pollHack requires write access.

This will be fixed in pollHack in a later commit.

Change-Id: I37b459339eda2e8270ca8094455db3ddbf764911

Also add a test if "ro" works directly in the fs
package.

Currently fails like this (will be fixed later):

	--- FAIL: TestRoMount (0.01s)
	    simple_test.go:116: read-only file syste

Change-Id: I8194360b1d092268784b9c2b17de274b86633aa2

Regression test for https://github.com/hanwen/go-fuse/issues/344
"[v2 API] xfstests generic/257 fails (getdents seek trouble)"

Change-Id: Icefab07f6762f8e56de525713facf992797d4435

"seeking" works by skipping forward to the requested offset,
first rewinding to zero if seeking back.

The performance is obviously bad, however, seeking
inside a directory does not seem to be too common.

I think the solution is good enough until we see a
use case where this causes performance issues.

Fixes https://github.com/hanwen/go-fuse/issues/344

Change-Id: Ia7f7fbffaf932c16919eafede6b70f4aff245c25

* y/nodefs-cancel:
  Y nodefs: Propagate context to File methods
  fs: add a test verifying that forget drops nodes
  nodefs: Allow for several Lookup requests to be served simultaneously
  tests: fs: TestOpenDirectIO: move to /var/tmp
  Travis CI: disable output buffering using "go test -p 1"
  go.mod: upgrade to Go 1.13
  tests: fix data race in TestDeletedInodePath
  Travis CI: upgrade to Ubuntu 18.04

So that e.g. File.Read could be canceled.

This is already so in fs/ package. However WCFS in Wendelin.core still
uses older nodefs package, and for now it was faster to teach old
nodefs.File about cancel propagation.

References:

kirr/wendelin.core@b17aeb8c
kirr/wendelin.core@f05271b1
kirr/wendelin.core@5ba816da
kirr/go123@7ad867a3
kirr/go123@0bdac628
kirr/go123@d2dc6c09

Change-Id: Ia9f49f04d87ad3d42f1e5b14c62fe9b425f404a9

Change-Id: Ie2aeb012346b2fbcc32de09ea05b39069291d803

Change-Id: Ife141f38f2b7598223935963e2916b0b9a609cca

Commit d0fca860 introduced big lookup lock with the idea to "make sure
we don't return forgotten nodes to the kernel". However this way it also
started to prevent several Lookup handlers to be running simultaneously,
which can deadlock if e.g. Lookup handler somehow synchronizes with
other thread which caused the lookup:

https://github.com/hanwen/go-fuse/commit/d0fca860#commitcomment-37772099

On the surface the fix is easy: if we need to prevent lookups running in
parallel to forget, we can turn lookupLock into shared one wrt Lookup,
but exclusive wrt Forget.

However a more correct fix would be to review nodefs locking completely:
there is already per-filesystem treeLock RWMutex, which is _already_
used to synchronize for forgets and lookups. From this point of view
lookupLock is unneeded and the correct fix that d0fca860 should have been
doing is to correct the scope of treeLock to be used more precisely.

However that would be a more intrusive change, and given that nodefs is
considered deprecated today, imho it is better to proceed with making
lookupLock a shared one.

The test that is added deadlocks without provided fix.

I suggest not to reject this patch based on rationale that "nodefs is
deprecated" as there still are real filesystems that use nodefs.

I had not reviewed v2 go-fuse/fs locking yet.

Change-Id: I18e01457f474dea31dc17186dfe6db582c2e6337

* y/lookup-parallel:
  nodefs: Allow for several Lookup requests to be served simultaneously

Commit d0fca860 introduced big lookup lock with the idea to "make sure
we don't return forgotten nodes to the kernel". However this way it also
started to prevent several Lookup handlers to be running simultaneously,
which can deadlock if e.g. Lookup handler somehow synchronizes with
other thread which caused the lookup:

https://github.com/hanwen/go-fuse/commit/d0fca860#commitcomment-37772099

On the surface the fix is easy: if we need to prevent lookups running in
parallel to forget, we can turn lookupLock into shared one wrt Lookup,
but exclusive wrt Forget.

However a more correct fix would be to review nodefs locking completely:
there is already per-filesystem treeLock RWMutex, which is _already_
used to synchronize for forgets and lookups. From this point of view
lookupLock is unneeded and the correct fix that d0fca860 should have been
doing is to correct the scope of treeLock to be used more precisely.

However that would be a more intrusive change, and given that nodefs is
considered deprecated today, imho it is better to proceed with making
lookupLock a shared one.

The test that is added deadlocks without provided fix.

I suggest not to reject this patch based on rationale that "nodefs is
deprecated" as there still are real filesystems that use nodefs.

I had not reviewed v2 go-fuse/fs locking yet.

Change-Id: I18e01457f474dea31dc17186dfe6db582c2e6337

* y/nodefs-cancel:
  Y nodefs: Propagate context to File methods

* master: (60 commits)
  fuse: increment loops WaitGroup on Server creation
  loopback: enable diagnostics logging
  fs: detect Inode.Path() hitting an orphaned inode
  tests: make RenameOpenDir more sensitive
  fs: add Options.Logger, and use throughout
  fs: fix FileAllocater type assertion
  loopback: leave file permissions on "000" files as-is
  example/loopback: make -allow-other imply default_permissions
  fuse: implement GETATTR for pollHack
  loopback: preserve owner when running as root
  tests: posixtest: add FstatDeleted
  loopback: use Lgetxattr and friends
  fs: remove typeChangeDebug code
  fs: don't log error failing MvChild
  posixtest: fix some lint errors
  fs: fix lint errors
  fs: failure to unmount on cleanup is a fatal error
  fs: plug fd leak in bridge_test.go
  fs: don't enforce dirent ordering.
  internal: skip test with other group if user is only 1 group
  ...

This used to create a hidden folder in the user's home
directory to avoid /tmp, which is on tmpfs on modern
distros.

Use /var/tmp instead, which is also a real filesystem,
and avoid cluttering the home dir.

Change-Id: Ifa0a1dbdf7a2c4b08e53d20e28f287207ff8ef47

Per default, "go test ./..." tests multiple packages in
parallel.

From "go help build":

	-p n
		the number of programs, such as build commands or
		test binaries, that can be run in parallel.
		The default is the number of CPUs available.

This seems rather innocent, but a consequence of tests
running in parallel is that the all the output of each test
is buffered until it completes. This prevents multiple tests
interleaving their output.

This also means that when a test hangs, we get no output at
all, which is pretty bad.

Disable parallelization and trade test time for better
debugability of hung tests.

Change-Id: I4c670b65baa2df3bef70bce622b830530a316ee7

This gets added on every build and every test run,
dirtying the git tree again and again.

Give in and let Go 1.13 have it's entry in go.mod.

Change-Id: I13679e54fca09ab818bc0167dff3f1ad851e763e

Running `go test ./... -race` showed a data race
at `rootNode.deleted`. This race cannot really happen,
but the race detector does not know that as the calls
go through FUSE an appear to not be synchronized.

Use atomic loads and stores to make the race detector
happy.

Change-Id: I26bd3e7d8efdd5b967fdb360eb17d7f71a8005c8

So that e.g. File.Read could be canceled.

This is already so in fs/ package. However WCFS in Wendelin.core still
uses older nodefs package, and for now it was faster to teach old
nodefs.File about cancel propagation.

References:

kirr/wendelin.core@b17aeb8c
kirr/wendelin.core@f05271b1
kirr/wendelin.core@5ba816da
kirr/go123@7ad867a3
kirr/go123@0bdac628
kirr/go123@d2dc6c09

If "dist" is not specified, Travis uses Ubuntu 16.04
with kernel 4.15.0-1028-gcp. This kernel misses fixes
to FUSE which cause hangs in TestParallelDiropsHang.

Specifying "dist: bionic" makes Travis use kernel
5.0.0-1026-gcp.

Fixes https://github.com/hanwen/go-fuse/issues/281
Change-Id: Ic0f9987726eb11c608ec0cf42b73e089ec1cd2db

Change-Id: If158380a909a6bf78212c75c8dddc57e8c955448

The server.Serve routine can be called both inline and as
goroutine. In the latter case, it is a synchronization error for Serve
to call serve.loops.Add() by itself, leading to a detected race.

Change-Id: I36f24bd36d1ae77d71e7d69a54ebdf5dbee9bd62

The loopback example is usually used to debug and develop.
Enable the new diagnostics logging facility, unless -q (quiet)
is passed.

Change-Id: I9ae5214fc33616656832a219fc1470f421786b8c

Some ".deleted" logic was in place, but did not work
properly when the Inode was in the root directory.

Fix that by introducing the `found` variable and
add a test that verifies that it works.

Also, return `.go-fuse.$RANDOM/deleted` to make it
very unlikely that the placeholder name matches an
actual file or directory. Introducing the `/deleted`
subdir makes sure operations creating a file or
directory fail reliably with ENOENT.

Tests now look like this:

  $ go test ./fs -run TestPosix/RenameOpenDir -count 1  -v
  === RUN   TestPosix
  === RUN   TestPosix/RenameOpenDir
  [...]
  --- PASS: TestPosix (0.01s)
      --- SKIP: TestPosix/RenameOpenDir (0.01s)
          test.go:383: Fstat failed: no such file or directory. Known limitation - see https://github.com/hanwen/go-fuse/issues/55
  PASS
  ok    github.com/hanwen/go-fuse/v2/fs 0.014s

Change-Id: I2eb6fd48a11df543c9b7daf62647cb9d8a892568

The test seemed to pass because the inode number is overridden
in rawBridge.getattr, but looking at the permissions shows that
the wrong directory is stat()ed:

  $ go test ./fs -run TestPosix/RenameOpenDir -count 1  -v
  [...]
  17:49:46.454077 received ENODEV (unmount request), thread exiting
  17:49:46.454343 received ENODEV (unmount request), thread exiting
  --- PASS: TestPosix (0.01s)
      --- SKIP: TestPosix/RenameOpenDir (0.01s)
          test.go:392: got permissions 0755, want 0700. Known limitation - see https://github.com/hanwen/go-fuse/issues/55
  PASS
  ok    github.com/hanwen/go-fuse/v2/fs 0.016s

Also, add a log message whenever the inode number is overridden,
this should (probably) not happen during normal operation. And
it actually only happens once in the test suite (in RenameOpenDir):

  $ go test ./... -count 1 -v 2>&1 | grep "overriding ino"
  14:48:44.143694 warning: rawBridge.getattr: overriding ino 188663 with 186314

See https://github.com/hanwen/go-fuse/issues/55

Change-Id: I8b2ddb84c35a3b28b4f5e032e7113f8d484a5981

Change-Id: I4fabf222a306e5d3abdfda28422b046cd75c9a8c

The type assertion was meant to operate on `f.file`,
not on `n.ops` again.

Fixes xfstests generic/228.

Change-Id: I8f8ca0cead91a512a6d7826dd7e7e5d1887ae627

Set `opts.NullPermissions = true` to stop us from falsifying
file permissions on "000" files.

Fixes xfstests generic/088.

Change-Id: Ibabbdc97b1ae2531ca093bae6bb441ae15d3238e

With -allow-other, other users can access the mountpoint.
However, without default_permissions, nobody checks file
permissions.

Make loopback behave more like a regular filesystem by
enabling default_permissions if -allow-other is passed.

Fixes xfstests generic/087:

$ sudo ./check-loopback generic/087
fuse-xfstests gocryptfs-2018-08-18/67408ac7
Thu 26 Dec 2019 03:08:54 PM UTC

loopback is /usr/local/bin/loopback
FSTYP         -- fuse.loopback
PLATFORM      -- Linux/x86_64 brikett 5.3.12-300.fc31.x86_64
MKFS_OPTIONS  -- /var/tmp/fuse-xfstests/check-loopback/scratchdev
MOUNT_OPTIONS -- -o context=system_u:object_r:root_t:s0 /var/tmp/fuse-xfstests/check-loopback/scratchdev /var/tmp/fuse-xfstests/check-loopback/scratchdir

generic/087	 1s
Ran: generic/087
Passed all 1 tests

Runtime was 1 seconds, exit code 0

Change-Id: Ifa8ac38a3775c3cacf20d37638a95887c9c24f93

When the default_permissions option is passed to the kernel,
it issues a extra GETATTR on the .go-fuse-epoll-hack node.

Rejecting that with EIO makes `syscall.Creat` in `pollHack` fail,
ultimately erroring out during mount.

$ loopback -allow-other -debug b a
16:06:39.576856 rx 2: INIT i0 {7.31 Ra 0x20000 NO_OPEN_SUPPORT,PARALLEL_DIROPS,ABORT_ERROR,EXPLICIT_INVAL_DATA,SPLICE_WRITE,READDIRPLUS,ASYNC_DIO,DONT_MASK,NO_OPENDIR_SUPPORT,AUTO_INVAL_DATA,READDIRPLUS_AUTO,POSIX_ACL,HANDLE_KILLPRIV,MAX_PAGES,ATOMIC_O_TRUNC,EXPORT_SUPPORT,SPLICE_MOVE,BIG_WRITES,SPLICE_READ,FLOCK_LOCKS,IOCTL_DIR,WRITEBACK_CACHE,ASYNC_READ,POSIX_LOCKS,CACHE_SYMLINKS}
16:06:39.576999 tx 2:     OK, {7.28 Ra 0x20000 AUTO_INVAL_DATA,BIG_WRITES,ASYNC_READ,NO_OPEN_SUPPORT,PARALLEL_DIROPS,READDIRPLUS 0/0 Wr 0x10000 Tg 0x0}
16:06:39.578670 rx 4: GETATTR i1 {Fh 0}
16:06:39.578717 rx 6: GETATTR i1 {Fh 0}
16:06:39.578735 tx 4:     OK, {tA=1s {M040755 SZ=40 L=2 1026:1026 B0*4096 i0:1 A 1577370827.990394 M 1577370827.990394 C 1577370827.990394}}
16:06:39.578765 tx 6:     OK, {tA=1s {M040755 SZ=40 L=2 1026:1026 B0*4096 i0:1 A 1577370827.990394 M 1577370827.990394 C 1577370827.990394}}
16:06:39.579028 rx 8: LOOKUP i1 [".go-fuse-epoll-hack"] 20b
16:06:39.579053 tx 8:     2=no such file or directory, {i0 g0 tE=0s tA=0s {M00 SZ=0 L=0 0:0 B0*0 i0:0 A 0.000000 M 0.000000 C 0.000000}}
16:06:39.579087 rx 10: CREATE i1 {0100100 [WRONLY,TRUNC,CREAT,0x8000] (022)} [".go-fuse-epoll-hack"] 20b
16:06:39.579113 tx 10:     OK, {i18446744073709551615 g0 {M0100644 SZ=0 L=1 0:0 B0*0 i0:18446744073709551615 A 0.000000 M 0.000000 C 0.000000} &{18446744073709551615 0 0}}
16:06:39.579199 rx 14: GETATTR i1 {Fh 0}
16:06:39.579205 rx 12: GETATTR i18446744073709551615 {Fh 0}
16:06:39.579216 tx 14:     OK, {tA=1s {M040755 SZ=40 L=2 1026:1026 B0*4096 i0:1 A 1577370827.990394 M 1577370827.990394 C 1577370827.990394}}
16:06:39.579237 rx 16: LOOKUP i1 [".go-fuse-epoll-hack"] 20b
16:06:39.579242 tx 12:     5=input/output error, {tA=0s {M00 SZ=0 L=0 0:0 B0*0 i0:0 A 0.000000 M 0.000000 C 0.000000}}
16:06:39.579247 tx 16:     2=no such file or directory, {i0 g0 tE=0s tA=0s {M00 SZ=0 L=0 0:0 B0*0 i0:0 A 0.000000 M 0.000000 C 0.000000}}
16:06:39.579270 Mount fail: input/output error
16:06:39.579271 rx 20: LOOKUP i1 [".Trash"] 7b

Change-Id: I20024baf2e8f386b637abbd236b188bfdfa8579f

If we are running as root have caller information from
the context, try to set the owner of newly created files
and directories.

Lays the foundation for fixing xfstests generic/087
(xfstests always run as root).

Change-Id: Ib8d768153f3ec82ce572021a433a6398560efd44

Excercises the fd-finding logic in rawBridge.GetAttr.
No issues found, tests pass, logic works fine.

Change-Id: I49731b8ec5b41344d58409c8ca615f466bd95f29

Stop following symlinks when working with extended
attributes, and add a test for it.

Problem found by xfstests generic/062.

Change-Id: I67f94451322cdfebdcbcc3af21679ccd4e2800d7

Change-Id: I87f5ff47daeb64412ff877038619d6cdf5e6ec92

Document MvChild failure condition.

Change-Id: I928335eb79c73bdedb14c9f7544c3b15e3862af8

Change-Id: I15d31de59b4c34caff132bdc7d09ad520d0e68cb

Change-Id: I3bce0eee02e0f8d1a2d02c4a4f065bb023179ab1

Any subsequent attempt to remove the mountdir will fail if the unmount
does not succeed.

Use syscall.Rmdir rather than os.Remove. We know the mount point is a
directory, and Go's cleverness can only complicate matters here.

Change-Id: I62cfad63f34af17b78c3184884a2972673f631a3