Added FUSE support #55

Merged
Elara6331 merged 65 commits from yannickulrich/itd:fuse into master 2023-03-25 22:23:52 +00:00
2 changed files with 46 additions and 33 deletions
Showing only changes of commit 87c78566c1 - Show all commits

35
fuse.go
View File

@ -12,7 +12,7 @@ import (
func startFuse(ctx context.Context, dev *infinitime.Device) error {
yannickulrich marked this conversation as resolved Outdated

This function should be called startFUSE instead to adhere to Go naming conventions

This function should be called `startFUSE` instead to adhere to Go naming conventions

Thanks, I'm relatively new to Go, sorry for these issues

Done in 673383f

Thanks, I'm relatively new to Go, sorry for these issues Done in 673383f
// This is where we'll mount the FS
os.Mkdir(k.String("fuse.mountpoint"), 0755)
yannickulrich marked this conversation as resolved Outdated

Use os.MkdirAll here instead so that it creates parent directories, and handle the error (just return it).

Use `os.MkdirAll` here instead so that it creates parent directories, and handle the error (just return it).

This is actually a bit more complicated than that. If the mountpoint already exists, fuse will crash. We also can't delete the mountpoint beforehand (rm: cannot remove '/tmp/itd/mnt': Transport endpoint is not connected). The best way to solve this should be calling the unmount function. How would you suggest going about doing this?

This is actually a bit more complicated than that. If the mountpoint already exists, fuse will crash. We also can't delete the mountpoint beforehand (`rm: cannot remove '/tmp/itd/mnt': Transport endpoint is not connected`). The best way to solve this should be calling the [unmount function](https://pkg.go.dev/github.com/hanwen/go-fuse/v2@v2.2.0/fuse#Server.Unmount). How would you suggest going about doing this?

Yeah, this one is going to be a bit more complicated. The FUSE library does have a different unmount function that you could call before trying to mount the fs, but it's not exported, so we'll need to do a small hack to get access to it anyway. In the fusefs package, add a file called unmount.go with the following contents:

unmount.go
package fusefs

import (
    _ "unsafe"

    "github.com/hanwen/go-fuse/v2/fuse"
)

func Unmount(mountPoint string) error {
    return unmount(mountPoint, &fuse.MountOptions{DirectMount: false})
}

// Unfortunately, the FUSE library does not export its unmount function,
// so this is required until that changes
//go:linkname unmount github.com/hanwen/go-fuse/v2/fuse.unmount
func unmount(mountPoint string, opts *fuse.MountOptions) error

Then, you can simply call that function at the top of startFUSE like so:

// Ignore the error because nothing might be mounted on the mountpoint
_ = fusefs.Unmount(k.String("fuse.mountpoint"))

I'll file an issue with the fuse library to export that function once this is merged.

The best way to solve this should be calling the unmount function.

I agree that this should be done. However, this doesn't solve the problem completely because if ITD panics for whatever reason, it won't get run and then the user will get a confusing message. Let me see what the best way would be to call this function on shutdown.

Yeah, this one is going to be a bit more complicated. The FUSE library does have [a different unmount function](https://github.com/hanwen/go-fuse/blob/0f728ba15b38579efefc3dc47821882ca18ffea7/fuse/mount_linux.go#L133) that you could call before trying to mount the fs, but it's not exported, so we'll need to do a small hack to get access to it anyway. In the `fusefs` package, add a file called `unmount.go` with the following contents: <details> <summary><code>unmount.go</code></summary> ```go package fusefs import ( _ "unsafe" "github.com/hanwen/go-fuse/v2/fuse" ) func Unmount(mountPoint string) error { return unmount(mountPoint, &fuse.MountOptions{DirectMount: false}) } // Unfortunately, the FUSE library does not export its unmount function, // so this is required until that changes //go:linkname unmount github.com/hanwen/go-fuse/v2/fuse.unmount func unmount(mountPoint string, opts *fuse.MountOptions) error ``` </details> Then, you can simply call that function at the top of `startFUSE` like so: ```go // Ignore the error because nothing might be mounted on the mountpoint _ = fusefs.Unmount(k.String("fuse.mountpoint")) ``` I'll file an issue with the fuse library to export that function once this is merged. > The best way to solve this should be calling the unmount function. I agree that this should be done. However, this doesn't solve the problem completely because if ITD panics for whatever reason, it won't get run and then the user will get a confusing message. Let me see what the best way would be to call this function on shutdown.

Cool, thank you! I have implemented this 🙂

Cool, thank you! I have implemented this 🙂
root := &ITNode{kind: 0}
root := fusefs.BuildRootNode(dev)
server, err := fs.Mount(k.String("fuse.mountpoint"), root, &fs.Options{
MountOptions: fuse.MountOptions{
// Set to true to see how the file system works.
@ -32,43 +32,12 @@ func startFuse(ctx context.Context, dev *infinitime.Device) error {
Str("target", k.String("fuse.mountpoint")).
Send()
properties[0] = ITProperty{"heartrate", 2,
func(ctx context.Context) (<-chan []byte, error) {
ans, err := dev.WatchHeartRate(ctx)
return converterU8(ctx, ans), err
}}
properties[1] = ITProperty{"battery", 3,
func(ctx context.Context) (<-chan []byte, error) {
ans, err := dev.WatchBatteryLevel(ctx)
return converterU8(ctx, ans), err
}}
properties[2] = ITProperty{"motion", 4,
func(ctx context.Context) (<-chan []byte, error) {
ans, err := dev.WatchMotion(ctx)
return converterMotionValues(ctx, ans), err
}}
properties[3] = ITProperty{"stepcount", 5,
func(ctx context.Context) (<-chan []byte, error) {
ans, err := dev.WatchStepCount(ctx)
return converterU32(ctx, ans), err
}}
properties[4] = ITProperty{"version", 6,
func(ctx context.Context) (<-chan []byte, error) {
ans, err := dev.Version()
return converter1String(ctx, ans), err
}}
properties[5] = ITProperty{"address", 7,
func(ctx context.Context) (<-chan []byte, error) {
ans := dev.Address()
return converter1String(ctx, ans), nil
}}
fusefs.BuildProperties(dev)
myfs, err = dev.FS()
if err != nil {
log.Warn("Error getting BLE filesystem").Err(err).Send()
return err
}
inodemap = make(map[string]uint64)
// Wait until unmount before exiting
go server.Serve()

View File

@ -89,7 +89,51 @@ type ITNode struct {
path string
}
func BuildRootNode(dev *infinitime.Device) *ITNode {
inodemap = make(map[string]uint64)
myfs, _ = dev.FS()
return &ITNode{kind: 0}
}
var properties = make([]ITProperty, 6)
yannickulrich marked this conversation as resolved
Review

In my opinion, "device" is unclear. I think it should be called "info" instead, that seems clearer.

Another thing is that you seem to be providing continuous sensor data by never sending an EOF, which means files will be read forever. Files generally don't work like that, so it leads to consequences. Since the syscall for reading from the file will never return until new information is received, a program that reads from this file will not be able to be killed, even by SIGKILL. You can see this if you try to cat the file. It can also break your system's shell depending on how it's configured. Instead, I think you should just provide single data points, and people will just have to use itctl if they want continuous data, because files aren't meant to work this way.

In my opinion, "device" is unclear. I think it should be called "info" instead, that seems clearer. Another thing is that you seem to be providing continuous sensor data by never sending an EOF, which means files will be read forever. Files generally don't work like that, so it leads to consequences. Since the syscall for reading from the file will never return until new information is received, a program that reads from this file will not be able to be killed, even by `SIGKILL`. You can see this if you try to `cat` the file. It can also break your system's shell depending on how it's configured. Instead, I think you should just provide single data points, and people will just have to use `itctl` if they want continuous data, because files aren't meant to work this way.
Review

Happy to change that.

Regarding the continuous data stream, isn't this how device files work under Linux? If I run cat /dev/hidraw0 I get a live stream of data. I was hoping to get the same behaviour. I'm happy to try and fix any bug but think this is probably the best way of doing this. Alternatively, we can have to files, a live stream file (eg. /tmp/itd/mnt/live/motion) and a one-shot file (eg. /tmp/itd/mnt/info/motion). Would you be opposed to this?

I'm not 100% sure what you mean with the cat example. This works perfectly for me:

$ cat /tmp/itd/mnt/info/motion
22 -5 -1042
22 -5 -1042
22 -5 -1042
19 -5 -1041
20 -5 -1042
18 -6 -1038
19 0 -1041
24 0 -1043
23 -6 -1041
^C
$

In the itd log I also see the Done method of the watcher getting called. Could you perhaps elaborate on the problem you see?

Happy to change that. Regarding the continuous data stream, isn't this how device files work under Linux? If I run `cat /dev/hidraw0` I get a live stream of data. I was hoping to get the same behaviour. I'm happy to try and fix any bug but think this is probably the best way of doing this. Alternatively, we can have to files, a live stream file (eg. `/tmp/itd/mnt/live/motion`) and a one-shot file (eg. `/tmp/itd/mnt/info/motion`). Would you be opposed to this? I'm not 100% sure what you mean with the `cat` example. This works perfectly for me: ``` $ cat /tmp/itd/mnt/info/motion 22 -5 -1042 22 -5 -1042 22 -5 -1042 19 -5 -1041 20 -5 -1042 18 -6 -1038 19 0 -1041 24 0 -1043 23 -6 -1041 ^C $ ``` In the `itd` log I also see the [Done method of the watcher](https://gitea.arsenm.dev/Arsen6331/infinitime/src/branch/master/infinitime.go#L693-L697) getting called. Could you perhaps elaborate on the problem you see?
Review

Could you perhaps elaborate on the problem you see?

Try this with something that returns data less commonly, such as battery. Every time cat tries to read from the file, it executes a syscall. Since data isn't returned until ITD receives it from the watch, the syscall doesn't return until then either. What this means is that whatever program is reading the file will be executing kernel code, which prevents you from killing it.

Linux does this kind of stuff with a char device file. Unfortunately, FUSE doesn't provide a way to make a char device. There seems to be a workaround, but it only works for the root user, so it would not work in ITD. I did test that, and I can confirm it doesn't work.

> Could you perhaps elaborate on the problem you see? Try this with something that returns data less commonly, such as `battery`. Every time `cat` tries to read from the file, it executes a syscall. Since data isn't returned until ITD receives it from the watch, the syscall doesn't return until then either. What this means is that whatever program is reading the file will be executing kernel code, which prevents you from killing it. Linux does this kind of stuff with a char device file. Unfortunately, FUSE doesn't provide a way to make a char device. There seems to be a workaround, but it only works for the root user, so it would not work in ITD. I did test that, and I can confirm it doesn't work.
Review

Okay, fair enough. Sorry for doubting you. How do you feel about keeping a stream-type file for motion? I'm not 100% sure how much sense the instantaneous motion makes.. Otherwise, I'm going to remove the entire channel stuff.

Okay, fair enough. Sorry for doubting you. How do you feel about keeping a stream-type file for motion? I'm not 100% sure how much sense the instantaneous motion makes.. Otherwise, I'm going to remove the entire channel stuff.
Review

Yeah, instantaneous motion data doesn't make much sense, but the continuous motion data file also doesn't work all the time. InfiniTime will stop sending motion data if you have Raise To Wake and ShakeWake disabled and you put the watch to sleep with the button. This will cause programs to once again read forever without the ability to be killed, so I think it would be better to remove the continuous data entirely. I would love to have it, but if it breaks most programs, it's not worth it, especially since itctl can be used instead.

Yeah, instantaneous motion data doesn't make much sense, but the continuous motion data file also doesn't work all the time. InfiniTime will stop sending motion data if you have Raise To Wake and ShakeWake disabled and you put the watch to sleep with the button. This will cause programs to once again read forever without the ability to be killed, so I think it would be better to remove the continuous data entirely. I would love to have it, but if it breaks most programs, it's not worth it, especially since `itctl` can be used instead.
func BuildProperties(dev *infinitime.Device) {
properties[0] = ITProperty{"heartrate", 2,
func(ctx context.Context) (<-chan []byte, error) {
ans, err := dev.WatchHeartRate(ctx)
return converterU8(ctx, ans), err
}}
properties[1] = ITProperty{"battery", 3,
func(ctx context.Context) (<-chan []byte, error) {
ans, err := dev.WatchBatteryLevel(ctx)
return converterU8(ctx, ans), err
}}
properties[2] = ITProperty{"motion", 4,
func(ctx context.Context) (<-chan []byte, error) {
ans, err := dev.WatchMotion(ctx)
return converterMotionValues(ctx, ans), err
}}
properties[3] = ITProperty{"stepcount", 5,
func(ctx context.Context) (<-chan []byte, error) {
ans, err := dev.WatchStepCount(ctx)
return converterU32(ctx, ans), err
}}
properties[4] = ITProperty{"version", 6,
yannickulrich marked this conversation as resolved Outdated

You should definitely handle this error

You should definitely handle this error

Done in a54ca7a

Done in a54ca7a
func(ctx context.Context) (<-chan []byte, error) {
ans, err := dev.Version()
return converter1String(ctx, ans), err
}}
properties[5] = ITProperty{"address", 7,
func(ctx context.Context) (<-chan []byte, error) {
ans := dev.Address()
Elara6331 marked this conversation as resolved Outdated

How would you recommend doing this? I suppose it could fail for all sorts of reasons such as

  • no such file or directory ENOENT
  • generic input/output error EIO
  • invalid argument EINVAL
  • connection aborted ECONNABORTED
  • connection refused ECONNREFUSED
  • connection reset ECONNRESET
  • is actually a file EISNAM

In other places such as when we open a file, we could have

  • is actually a folder EISDIR
  • file exists EEXIST

Again, I'm sorry but I'm not a Go expert and don't really now how to do this properly.. especially when dealing with FSError

How would you recommend doing this? I suppose it could fail for all sorts of reasons such as * no such file or directory `ENOENT` * generic input/output error `EIO` * invalid argument `EINVAL` * connection aborted `ECONNABORTED` * connection refused `ECONNREFUSED` * connection reset `ECONNRESET` * is actually a file `EISNAM` In other places such as when we open a file, we could have * is actually a folder `EISDIR` * file exists `EEXIST` Again, I'm sorry but I'm not a Go expert and don't really now how to do this properly.. especially when dealing with `FSError`

The err can be several different kinds of errors, and FSError is just one of them. It's actually a type I made. You can see it here: https://gitea.arsenm.dev/Arsen6331/infinitime/src/commit/512d48bc2469/blefs/error.go#L20. It contains an error code you can check to see what went wrong, and you can scroll down to see the meaning of each code.

In this case, I'd use a Go type switch to check which error type actually occurred and then check the code or do whatever else needs to be done. Maybe there could be a function like syscallErr() that takes an error and returns the proper syscall error?

If you don't feel comfortable doing that, I can merge this and then implement it myself and send you the commit so you can see how I did it, or you can just try it yourself, whatever you feel would be better.

The `err` can be several different kinds of errors, and `FSError` is just one of them. It's actually a type I made. You can see it here: https://gitea.arsenm.dev/Arsen6331/infinitime/src/commit/512d48bc2469/blefs/error.go#L20. It contains an error code you can check to see what went wrong, and you can scroll down to see the meaning of each code. In this case, I'd use a Go type switch to check which error type actually occurred and then check the code or do whatever else needs to be done. Maybe there could be a function like `syscallErr()` that takes an `error` and returns the proper syscall error? If you don't feel comfortable doing that, I can merge this and then implement it myself and send you the commit so you can see how I did it, or you can just try it yourself, whatever you feel would be better.

Something like in 4c59561a99? There are a few TODO where I'm not sure what the correct POSIX error would be and improvised. If you have a better idea, feel free to change them though

Something like in 4c59561a99? There are a few `TODO` where I'm not sure what the correct POSIX error would be and improvised. If you have a better idea, feel free to change them though

That looks good, thanks

That looks good, thanks
return converter1String(ctx, ans), nil
}}
}
var myfs *blefs.FS = nil;
var inodemap map[string]uint64 = nil;