libfuse introduced [1] a special `/dev/fd/N` syntax for the mountpoint:
It means that a privileged parent process:

 * Opened /dev/fuse
 * Called mount() on a real mountpoint directory
 * Inherited the fd to /dev/fuse to us
 * Informs us about the fd number via /dev/fd/N

This functionality is used to allow FUSE mounts inside containers
that have neither root permissions nor suid binaries [2], and
for the --drop_privileges flag of mount.fuse3 [4]

Tested with singularity and gocryptfs and actually works [3].

v2: Added doccomment for NewServer.
v3: Added specific error message on Server.Unmount().
v4: Moved mount details to package comment

[1] https://github.com/libfuse/libfuse/commit/64e11073b9347fcf9c6d1eea143763ba9e946f70
[2] https://github.com/rfjakob/gocryptfs/issues/590
[3] $ singularity run --fusemount "host:gocryptfs --extpass echo --extpass test /tmp/a /mnt" docker://ubuntu
    INFO:    Using cached SIF image
    Reading password from extpass program "echo", arguments: ["test"]
    Decrypting master key
    bash: /home/jakob/.cargo/env: No such file or directory
    bash: /home/jakob/.cargo/env: No such file or directory
    bash: /home/jakob/.cargo/env: No such file or directory
    Singularity> Filesystem mounted and ready.
[4] man mount.fuse3

Change-Id: Ibcc2464b0ef1e3d236207981b487fd9a7d94c910

Update the package comment to refer to the fs package
instead of nodefs and pathfs, which are deprecated.

Change-Id: I317e53152b63b982298a7de29120191488dcd362

This makes the function available to testing, which will
be used in the next commit.

Change-Id: I2df7aaf50748f209964da6f14764e81237d49027

We enable FUSE_CAP_ASYNC_READ per default, which means that
the kernel can (and does) submit multiple concurrent out-of-order
read requests to service userspace reads and kernel readahead.

For some backing storages, like Amazon Cloud Drive, out-of-order
reads are expensive.

gocryptfs has implemented a delay-based workaround with its
`-serialize_reads` flag for this case
(see https://github.com/rfjakob/gocryptfs/issues/92 for details).

Not enabling FUSE_CAP_ASYNC_READ makes the kernel do this for us,
as verified by adding debug output to gocryptfs, so expose it as
a mount flag in MountOptions.

Fixes: https://github.com/hanwen/go-fuse/issues/140
Graphs-at: https://github.com/hanwen/go-fuse/issues/395
Related: https://github.com/rfjakob/gocryptfs/issues/92
Change-Id: I10f947d71e1453989c4a9b66fbb0407f7163994f

The root directory always get nodeid 1, and the v2 api
originally had "nodeid == inode number".

But since 1d0096e5,
inode number and node id are independent, and we can allow
inode number 1.

Why does this matter? MacOS actually uses inode number 1 on ExFAT
mounts, triggering a panic on overlay filesystems like gocryptfs:

    panic: using reserved ID 1 for inode number

Fixes: https://github.com/rfjakob/gocryptfs/issues/585
Change-Id: I5e3476da65400d32efdb3dac1fe4901c64c00a89

Drops an obsolete entry pointing to go-fuse v1.0.0.

Change-Id: Iaf8e880d6f39e63b593a7929b8b4ff649c1d7807

The file type is not a bitmap, and looking at

	S_IFBLK                          = 0x6000
	S_IFCHR                          = 0x2000
	S_IFDIR                          = 0x4000
	S_IFIFO                          = 0x1000
	S_IFLNK                          = 0xa000
	S_IFREG                          = 0x8000
	S_IFSOCK                         = 0xc000

and confirmed by the added test, we used to classify
block devices and sockets as directories:

 --- FAIL: TestInodeIsDir (0.00s)
     inode_test.go:25: wrong result for case struct { mode uint32; dir bool }{mode:0x6000, dir:false}
     inode_test.go:25: wrong result for case struct { mode uint32; dir bool }{mode:0xc000, dir:false}

The check is fixed now and the test passes.

Change-Id: I6490992d1fecc8a6bea7c2c4b2f1bca765d03d4c

This

	GOMAXPROCS=1 go test -v -count=1 -run TestDeleteNotify ./fuse/test

deadlocked reliably, apparently in the chdir() from cmd.Start().

We can do without the chdir() by using an absolte path.
This avoids the issue.

Fixes https://github.com/hanwen/go-fuse/issues/261

Change-Id: Ia9c8f15819c125c5bf94b085fa4c5f2977a6789a

Maps do not free all memory when elements get deleted
( https://github.com/golang/go/issues/20135 ).

As a workaround, we recreate our two big maps (kernelNodeIds & stableAttrs)
when they have shrunk dramatically (100 x smaller).

Benchmarkung with loopback (go version go1.16.2 linux/amd64) shows:

Before this change:

Step               VmRSS (kiB)
-----              -----------
Fresh mount          4000
ls -R 500k files   271100
after drop_cache   336448
wait ~ 3 minutes   101588

After:

Step               VmRSS (kiB)
-----              -----------
Fresh mount          4012
ls -R 500k files   271100
after drop_cache    31528

Results for gocryptfs are similar.

Has survived xfstests via gocryptfs.

Fixes: https://github.com/rfjakob/gocryptfs/issues/569
Change-Id: Idcae1ab953270516735839a034d586717647b8db

inoMap & nodeidMap have been renamed a long time ago,
but this comment has been missed.

Update it for the new names.

Change-Id: I5e1d86f5e8ac3b3219b324da09d3f628b3f0fb80

If writing the INTERRUPT response returns ENOENT it means that the
referred request is no longer known by the kernel. This is a normal if
the referred request already has completed. We thus suppress the scary
warning unless debugging is enabled.

Fixes #375.

Change-Id: I3266dbd8ad53a42db9e0e746e4734f284746b76c

Enabling the logger shows this helpful warning

  warning: Inode.Path: n3 is orphaned, replacing segment with ".go-fuse.5577006791947779410/deleted"

for the problem reported at https://github.com/hanwen/go-fuse/issues/398 ,
and it generally makes sense to have complete logs for tests.

Enable the logger.

Change-Id: I4602f100a31c6e50bc48251a877b97ce3013ff24

Enables kernel ACL enforcement.

Change-Id: I7fb5f6f8b350ad430da18e0173df942b98da7c02

With

	CAP_POSIX_ACL           = (1 << 20)

the ACLs actually work.

Change-Id: Ic91e9f8d4196c811d4c566455b0e8c4a7871b105

Introduce the new struct inodeParents that wraps a map
and one special slot for the most recent parent. Unit tests
included.

Because the map is lazily initialized, we should save some
memory on the common single-parent case (= file with no hard links)
compared to before.

Benchmarking with gocryptfs shows no discernible change
in performance. fsstress testing with gocryptfs shows no issues.

TestStaleHardlinks the previous commit passes now.

Change-Id: I8d69093abc906addde751a9e70dbd78a3a61371a

This test shows a problem in selecting a random path
for a node when there is more than one.

The failure looks like this:

  18:41:50.796468 rx 136: LOOKUP n1 ["link0"] 6b
  18:41:50.796489 tx 136:     OK, {n2 g1 tE=0s tA=0s {M0100600 SZ=0 L=1 1026:1026 B0*4096 i0:269663 A 1616348510.793212 M 1616348510.793212 C 1616348510.795212}}
  18:41:50.796535 rx 138: OPEN n2 {O_RDONLY,0x8000}
  18:41:50.796557 tx 138:     2=no such file or directory, {Fh 0 }

The LOOKUP succeeds (because the file "link0" is there).
But the OPEN fails because go-fuse chooses another (stale) path.

I will try to make the behavoir more robust against changes
behind our back, but this patch only adds a test to show the
problem.

Change-Id: I39b31ba717ddaaad7dda6ecd86707c75cd25102e

Change-Id: Ide563371

Release isn't synchronized, so a quick [close,unlink] sequence may
still fail without this delay.

Change-Id: I64170fca

This is a more complete emulation of Windows FS semantics.

Change-Id: If9dec3ebeac808e1cc4f6d234290867c50e87a30

This avoids a panic.

Change-Id: I5c4cc0d875ecdbc3b07c2322b5a5a5e6fa493709

Previously maximum number of requests readers was hard-coded to 2.
Setting it to max(2, min(16, GOMAXPROCS)) improves the throughput on
systems with more cores available. Going beyond 16 readers seems to hurt
performance even if there are more cores available.

Benchmark GoFuseRead can be used for demonstrating the effects of this
variable. The benchmark reads 32 streams in parallel from a dummy
filesystem (read operations immediately return zeros). Example results
from an i7-8550U (8 cores):

| Max readers | Total throughput |
| ----------: | ---------------: |
|           2 |       13217 MB/s |
|           4 |       19202 MB/s |
|           8 |       19973 MB/s |
|          16 |       18994 MB/s |

On a 96 core system:

| Max readers | Total throughput |
| ----------: | ---------------: |
|           2 |       11490 MB/s |
|           4 |       16129 MB/s |
|           8 |       24263 MB/s |
|          16 |       29568 MB/s |
|          32 |       28262 MB/s |

Note that improvements won't be as dramatic for real filesytem
implementations. In benchmarks for a filesystem doing real work I see a
30-40% improvement (8.3 -> 11.4 GB/s) on the 96 core system.

Also tweaked some of the other benchmarks so they don't leave behind
mountpoints.

Fixes #388.

Change-Id: Ibff17d7fc92195f078a9ccff818a31f3a58873f2

This test mmap()s a file, which means reading the
memory can block, triggering file I/O and FUSE requests.
With GOMAXPROCS=1, we don't have a free thread to handle
those requests, deadlocking the test:

	GOMAXPROCS=1 go test -v -count=1 -run TestCacheControl ./fuse/test

Use mlock() to fix that, also add a warning to the
"fuse" package docs.

Fixes https://github.com/hanwen/go-fuse/issues/261

Change-Id: Ia32fb1826c2be602cf4f0f6cf73956af797c940c

Make LoopbackNode and LoopbackRoot public, and add an example of how
to extend them.

Fixes #364.

Change-Id: Idcce1bf3765e93ece3bcf9d14de8a957b8c6b63a

Closes #379.

Change-Id: Ieb821fe7e68a3b7822306ad620dcdf834b5dfc12

Document LoopbackNode, LoopbackRoot. Use loopbackRoot.newNode() for
creating new nodes

Change-Id: I625c7706

Change-Id: I19d52f14f63f1481b2b1aa9dcd46c91fe4e1ad99

This leaves only one type to implement for the loopback file system,
which will simplify extending the loopback functionality.

Change-Id: I8906755f3bdf2ea63557d5bb25f50651521364c7

This provides a test for GitHub PR #378, ie. the previous commit

Change-Id: I9473722e

Change-Id: I8845cac5c18d2640b46971564f94f7ff1741a0ee

On OSX, syscall.O_TRUNC on the Create call apparently triggers a
SETATTR call with size 0, which must be responded to correctly.

Change-Id: Icb99891eae99ced6ffbd5fb1dadf42ffad687534

Some of the error codes due to the poll hack get returned back to the
OSX mount process, which leads to failing mounts.

Change-Id: I5b10509c3b8173343caf3f67c0e12a25ddfae32b

syscall.Dirent declares a 256-byte "name" field. In reality, this
field is only as long as necessary to transmit the name, so more
entries can be packed in a page.

When casting the result to syscall.Dirent, the result might run over
the end of the byte buffer, which the race detector doesn't like.

Address this by using our own flavor of syscall.Dirent. All linux
flavors use the same structure layout, except for the amount of
padding at the end.

Change-Id: I618f7ef9e167894e1ab25b2113dcd20d7b76ceb4

nodeid and inode number used to be identical, which is
why both could be marked as "iXXX". This is no longer the
case, so make nodeids distinguishable.

Example:

23:47:17.637736 rx 9156: MKDIR n28 {0777 (00)} ["d90X"] 5b
23:47:17.637801 tx 9156:     OK, {n112 g1 tE=1s tA=1s {M040755 SZ=4096 L=2 1026:1026 B8*4096 i0:2769795 A 1601761637.637385 M 1601761637.637385 C 1601761637.637385}}

v4: also update orphaned inode warning

Change-Id: Ib29dabfd4cf6692cf4118191f8dc73529562426f

Using the inode numbers as the nodeid causes problems
when the fs reuses inode numbers. This is the case with
any overlay filesystem that is backed by ext4 like the
loopback example or gocryptfs.

We already had the expSleep() and re-add hack,

	7090b027 fs: wait out apparent inode type change
	68f70527 fs: addNewChild(): handle concurrent FORGETs

to mitigate some of the problems (at the risk of deadlocking
forever), but I could not find a way to work around the following
case uncovered by fsstress:

The kernel expects a fresh nodeid from MKDIR (see #372 for a
debug log).

This is now guaranteed by passing the O_EXCL to addNewChild().
However, this also means that the hard link detection must
happen in addNewChild() as opposed to newInodeUnlocked()
before. The expSleep and re-add hacks are no longer needed
and have been dropped.

This survived 24 hours (42587 iterations) of fsstress testing.
Tested was the loopback example running on top of ext4 on
Linux 5.8.10.

Fixes https://github.com/hanwen/go-fuse/issues/372 .

v2: Rename inoMap -> stableAttrs, nodeidMap -> kernelNodeIds
    acc. to feedback from Han-Wen.
v3: initialize Inode.nodeId early
v4: drop leftover file stash.txt

Change-Id: Ibb36a886f15d48727daa10b9717ea88e45a6b8af

With rename and link, it hangs within 20ms, so we can reduce the
runtime.

Also add deadlock detection & statistics printout.

A good run looks like this:

	$ TMPDIR=/var/tmp go test -run TestFsstress -v
	=== RUN   TestFsstress
	    simple_test.go:601: Operation statistics:
	    simple_test.go:604:      creat:    33 ok, 132978 fail,  0 hung
	    simple_test.go:604:       link:    85 ok,  37196 fail,  0 hung
	    simple_test.go:604:      mkdir:    60 ok, 155273 fail,  0 hung
	    simple_test.go:604:     mkfifo:    45 ok, 161937 fail,  0 hung
	    simple_test.go:604: mknod_fifo:    55 ok, 154459 fail,  0 hung
	    simple_test.go:604:  mknod_reg:    77 ok, 180614 fail,  0 hung
	    simple_test.go:604: mknod_sock:    83 ok, 139479 fail,  0 hung
	    simple_test.go:604:    readdir:    13 ok,      0 fail,  0 hung
	    simple_test.go:604:     remove:   180 ok,     40 fail,  0 hung
	    simple_test.go:604:     rename:   646 ok,    137 fail,  0 hung
	    simple_test.go:604:      rmdir:    48 ok,  48342 fail,  0 hung
	    simple_test.go:604:    symlink:    30 ok, 119954 fail,  0 hung
	    simple_test.go:604:     unlink:   150 ok,   6121 fail,  0 hung

	--- PASS: TestFsstress (1.02s)
	PASS
	ok  	github.com/hanwen/go-fuse/v2/fs	1.022s

A bad run like this:

	go-fuse/fs$ TMPDIR=/var/tmp go test -run TestFsstress -v
	=== RUN   TestFsstress
	    simple_test.go:592: timeout waiting for goroutines to exit (deadlocked?)
	    simple_test.go:601: Operation statistics:
	    simple_test.go:604:      creat:     0 ok,     44 fail, 10 hung
	    simple_test.go:604:       link:     0 ok,      6 fail, 10 hung
	    simple_test.go:604:      mkdir:     1 ok,    283 fail, 10 hung
	    simple_test.go:604:     mkfifo:     1 ok,    140 fail, 10 hung
	    simple_test.go:604: mknod_fifo:     1 ok,   1246 fail, 10 hung
	    simple_test.go:604:  mknod_reg:     0 ok,     19 fail, 10 hung
	    simple_test.go:604: mknod_sock:     0 ok,    110 fail, 10 hung
	    simple_test.go:604:    readdir:     1 ok,      0 fail,  1 hung
	    simple_test.go:604:     remove:     1 ok,      0 fail, 10 hung
	    simple_test.go:604:     rename:     0 ok,     12 fail, 10 hung
	    simple_test.go:604:      rmdir:     1 ok,    346 fail, 10 hung
	    simple_test.go:604:    symlink:     1 ok,   1456 fail, 10 hung
	    simple_test.go:604:     unlink:     2 ok,     15 fail, 10 hung
	    simple_test.go:51: /usr/bin/fusermount: failed to unmount /var/tmp/TestFsstress302646914/mnt: Device or resource busy
	         (code exit status 1)

	--- FAIL: TestFsstress (2.30s)
	FAIL

	[hangs]

Change-Id: I9eb0ce02eb1df766be257d6a87de53c6ffa6d21e

TestFsstress is loosely modeled after xfstest's fsstress. It performs rapid
parallel removes / creates / readdirs. Coupled with inode reuse, this test
used to deadlock go-fuse quite quickly.

Change-Id: Ia3af099e08c74d3c160062c1ed414430b6c1e7c9

If this happens, it means our node accounting is messed up.
Abort the process by panicing instead of wrapping around
to UINT_MAX.

Change-Id: Id662ab03877d44a16785bfd164a96173a3b7519f

If kernel internally serializes lookup requests for a directory, the
test will deadlock, becaue it tries to issue two such request and pause
lookup handlers, untill both requests come in.

For example here is how this test hangs on Ubuntu Xenial 16.04:

kirr@kubu:~/t/go-fuse$ go test -count=1 -v -run TestNodeParallelLookup ./fuse/test/
=== RUN   TestNodeParallelLookup
18:24:38.494651 rx 1: INIT i0 {7.23 Ra 0x20000 POSIX_LOCKS,ATOMIC_O_TRUNC,EXPORT_SUPPORT,DONT_MASK,FLOCK_LOCKS,READDIRPLUS_AUTO,ASYNC_DIO,WRITEBACK_CACHE,SPLICE_WRITE,SPLICE_READ,IOCTL_DIR,AUTO_INVAL_DATA,READDIRPLUS,NO_OPEN_SUPPORT,ASYNC_READ,BIG_WRITES,SPLICE_MOVE}
18:24:38.494713 tx 1:     OK, {7.23 Ra 0x20000 ASYNC_READ,BIG_WRITES,AUTO_INVAL_DATA,READDIRPLUS,NO_OPEN_SUPPORT 0/0 Wr 0x10000 Tg 0x0}
18:24:38.495351 rx 2: LOOKUP i1 [".go-fuse-epoll-hack"] 20b
18:24:38.495467 tx 2:     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}}
18:24:38.495530 rx 3: OPEN i18446744073709551615 {O_RDONLY,0x8000}
18:24:38.495575 tx 3:     OK, {Fh 18446744073709551615 }
18:24:38.495624 rx 4: POLL i18446744073709551615
18:24:38.495659 tx 4:     38=function not implemented
18:24:38.495706 rx 5: FLUSH i18446744073709551615 {Fh 18446744073709551615}
18:24:38.495741 tx 5:     34=numerical result out of range
18:24:38.495812 rx 6: RELEASE i18446744073709551615 {Fh 18446744073709551615 0x8000  L0}
18:24:38.495849 tx 6:     34=numerical result out of range
18:24:38.495904 rx 7: LOOKUP i1 ["world"] 6b
18:24:38.495957 I: <- lookup "world"
< hang >

Amends: 17c0c400 (nodefs: Allow for several Lookup requests to be served simultaneously)
Updates: https://github.com/hanwen/go-fuse/issues/261

Change-Id: I5d0a76d1190af805351326c378f69cf8ff781a37

GOMAXPROCS=1 currently triggers tests hang:
https://github.com/hanwen/go-fuse/issues/261

Change-Id: I3dc98c8bbdaac7a4ce86639e405ba1a1cef4c295

Change-Id: I8291dceac66bd02bc8fbb845d1f125c4cf22f48e