From: Keno Fischer <keno@alumni.harvard.edu>

Signed-off-by: Keno Fischer <keno@juliacomputing.com>
---
 hw/9pfs/9p-local.c | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/hw/9pfs/9p-local.c b/hw/9pfs/9p-local.c
index c55ea25..3e358b7 100644
--- a/hw/9pfs/9p-local.c
+++ b/hw/9pfs/9p-local.c
@@ -669,6 +669,13 @@ static int local_mknod(FsContext *fs_ctx, V9fsPath *dir_path,
         return -1;
     }
 
+#ifdef CONFIG_DARWIN
+    /* Darwin doesn't have mknodat and it's unlikely to work anyway,
+       so let's just mark it as unsupported */
+    err = -1;
+    errno = EOPNOTSUPP;
+    goto out;
+#else
     if (fs_ctx->export_flags & V9FS_SM_MAPPED ||
         fs_ctx->export_flags & V9FS_SM_MAPPED_FILE) {
         err = mknodat(dirfd, name, fs_ctx->fmode | S_IFREG, 0);
@@ -699,6 +706,8 @@ static int local_mknod(FsContext *fs_ctx, V9fsPath *dir_path,
 
 err_end:
     unlinkat_preserve_errno(dirfd, name, 0);
+#endif
+
 out:
     close_preserve_errno(dirfd);
     return err;
-- 
2.8.1

On Sat, 26 May 2018 01:23:13 -0400
keno@juliacomputing.com wrote:

> From: Keno Fischer <keno@alumni.harvard.edu>
> 
> Signed-off-by: Keno Fischer <keno@juliacomputing.com>
> ---
>  hw/9pfs/9p-local.c | 9 +++++++++
>  1 file changed, 9 insertions(+)
> 
> diff --git a/hw/9pfs/9p-local.c b/hw/9pfs/9p-local.c
> index c55ea25..3e358b7 100644
> --- a/hw/9pfs/9p-local.c
> +++ b/hw/9pfs/9p-local.c
> @@ -669,6 +669,13 @@ static int local_mknod(FsContext *fs_ctx, V9fsPath *dir_path,
>          return -1;
>      }
>  
> +#ifdef CONFIG_DARWIN
> +    /* Darwin doesn't have mknodat and it's unlikely to work anyway,

What's unlikely to work ?

> +       so let's just mark it as unsupported */
> +    err = -1;
> +    errno = EOPNOTSUPP;
> +    goto out;
> +#else

Please introduce qemu_mknodat() with distinct implementations for linux
and darwin.

>      if (fs_ctx->export_flags & V9FS_SM_MAPPED ||
>          fs_ctx->export_flags & V9FS_SM_MAPPED_FILE) {
>          err = mknodat(dirfd, name, fs_ctx->fmode | S_IFREG, 0);
> @@ -699,6 +706,8 @@ static int local_mknod(FsContext *fs_ctx, V9fsPath *dir_path,
>  
>  err_end:
>      unlinkat_preserve_errno(dirfd, name, 0);
> +#endif
> +
>  out:
>      close_preserve_errno(dirfd);
>      return err;

>> +#ifdef CONFIG_DARWIN
>> +    /* Darwin doesn't have mknodat and it's unlikely to work anyway,
>
> What's unlikely to work ?
>

My concern was that allowing this would cause unexpected
behavior, since the device numbers will differ between OS X
and Linux. Though maybe this isn't the place to worry about
that.

On Thu, 31 May 2018 12:37:56 -0400
Keno Fischer <keno@juliacomputing.com> wrote:

> >> +#ifdef CONFIG_DARWIN
> >> +    /* Darwin doesn't have mknodat and it's unlikely to work anyway,  
> >
> > What's unlikely to work ?
> >  
> 
> My concern was that allowing this would cause unexpected
> behavior, since the device numbers will differ between OS X
> and Linux. Though maybe this isn't the place to worry about
> that.

The numbers may differ indeed but we don't really care since the
server never opens device files. This is just a directory entry.

>> My concern was that allowing this would cause unexpected
>> behavior, since the device numbers will differ between OS X
>> and Linux. Though maybe this isn't the place to worry about
>> that.
>
> The numbers may differ indeed but we don't really care since the
> server never opens device files. This is just a directory entry.

Ok, let me try to implement it. However, I don't think it is possible
to implement mknodat (or at least I can't think of how) on Darwin
directly. I could use regular mknod, but presumably, this is used
to avoid a race condition between creating the device and setting
the permissions. Can you think of a good way to resolve that?

On Thu, May 31, 2018 at 6:56 PM, Keno Fischer <keno@juliacomputing.com> wrote:
>>> My concern was that allowing this would cause unexpected
>>> behavior, since the device numbers will differ between OS X
>>> and Linux. Though maybe this isn't the place to worry about
>>> that.
>>
>> The numbers may differ indeed but we don't really care since the
>> server never opens device files. This is just a directory entry.
>
> Ok, let me try to implement it. However, I don't think it is possible
> to implement mknodat (or at least I can't think of how) on Darwin
> directly. I could use regular mknod, but presumably, this is used
> to avoid a race condition between creating the device and setting
> the permissions. Can you think of a good way to resolve that?

Would it work to fchdir in to the directory and use a cwd-relative
mknod, then fchdir back? Or do we need to maintain the cwd
while in this code?

On Thu, May 31, 2018 at 7:06 PM, Keno Fischer <keno@juliacomputing.com> wrote:
> On Thu, May 31, 2018 at 6:56 PM, Keno Fischer <keno@juliacomputing.com> wrote:
>>>> My concern was that allowing this would cause unexpected
>>>> behavior, since the device numbers will differ between OS X
>>>> and Linux. Though maybe this isn't the place to worry about
>>>> that.
>>>
>>> The numbers may differ indeed but we don't really care since the
>>> server never opens device files. This is just a directory entry.
>>
>> Ok, let me try to implement it. However, I don't think it is possible
>> to implement mknodat (or at least I can't think of how) on Darwin
>> directly. I could use regular mknod, but presumably, this is used
>> to avoid a race condition between creating the device and setting
>> the permissions. Can you think of a good way to resolve that?
>
> Would it work to fchdir in to the directory and use a cwd-relative
> mknod, then fchdir back? Or do we need to maintain the cwd
> while in this code?

Sorry for the triple-self-post here, but I took a look at the Darwin kernel
source and there's an unexposed (from the Darwin C library) syscall that
only changes the current thread's cwd. That seems like it should be safe,
so I'll go ahead and use that to implement mknodat. I'll also submit a
feature request to apple to implement mknodat.