Re: dirEntries removes entire branches of empty directories
On Thursday, 10 November 2022 at 21:27:28 UTC, Ali Çehreli wrote: However, ftw performs about twice as fast as dirEntries Yes, `dirEntries` isn't as fast as it could be. Here is a directory iterator which tries to strictly not do more work than what it must: https://github.com/CyberShadow/ae/blob/86b016fd258ebc26f0da3239a6332c4ebecd3215/sys/file.d#L178
Re: dirEntries removes entire branches of empty directories
On 11/14/22 14:41, kdevel wrote: > the ftw version gets the whole information from readdir alone. Created an enhancement request: https://issues.dlang.org/show_bug.cgi?id=23512 Ali
Re: dirEntries removes entire branches of empty directories
On Monday, 14 November 2022 at 21:05:01 UTC, kdevel wrote: [...] the runtimes are ftw : 98 ms, 470 ÎŒs, and 2 *beeep* dirEntries: 170 ms, 515 ÎŒs, and 2 *beeep* (to be continued) When I examine the process with strace it appears that the ftw version gets the whole information from readdir alone. The dirEntries version seems to call lstat on every file (in order to check that it is not a symlink) Breakpoint 1, 0xf7cc59d4 in lstat64 () from [...]gcc-12.1/lib/libgphobos.so.3 (gdb) bt #0 0xf7cc59d4 in lstat64 () from [...]gcc-12.1/lib/libgphobos.so.3 #1 0xf7a5269b in std.file.DirEntry._ensureLStatDone() () from [...]gcc-12.1/lib/libgphobos.so.3 #2 0xf7a5276a in std.file.DirEntry.linkAttributes() () from [...]gcc-12.1/lib/libgphobos.so.3 #3 0xf7a528c9 in std.file.DirIteratorImpl.mayStepIn() () from [...]gcc-12.1/lib/libgphobos.so.3 #4 0xf7a545ae in std.file.DirIteratorImpl() () from [...]gcc-12.1/lib/libgphobos.so.3 #5 0xf7a5466e in core.internal.lifetime() () from [...]gcc-12.1/lib/libgphobos.so.3 #6 0xf7a546ef in core.internal.lifetime() () from [...]gcc-12.1/lib/libgphobos.so.3 #7 0xf7a54726 in core.lifetime() () from [...]gcc-12.1/lib/libgphobos.so.3 #8 0xf7a54762 in std.typecons() () from [...]gcc-12.1/lib/libgphobos.so.3 #9 0xf7a547d6 in std.typecons() () from [...]gcc-12.1/lib/libgphobos.so.3 #10 0xf7a54811 in std.file.DirIterator.__ctor() () from [...]gcc-12.1/lib/libgphobos.so.3 #11 0xf7a54882 in std.file.dirEntries() () from [...]gcc-12.1/lib/libgphobos.so.3 #12 0x08088e25 in direntrybenchmark.directoryWalk_dirEntries() (root=..., dump=false) at direntrybenchmark.d:111 and after that an additional stat on the same file in order to check if it is a directory: Breakpoint 2, 0xf7cc5954 in stat64 () from [...]gcc-12.1/lib/libgphobos.so.3 (gdb) bt #0 0xf7cc5954 in stat64 () from [...]gcc-12.1/lib/libgphobos.so.3 #1 0xf7a527e1 in std.file.DirEntry._ensureStatOrLStatDone() () from [...]gcc-12.1/lib/libgphobos.so.3 #2 0xf7a5287a in std.file.DirEntry.isDir() () from [...]gcc-12.1/lib/libgphobos.so.3 #3 0x08088e6c in direntrybenchmark.directoryWalk_dirEntries() (root=..., dump=) at direntrybenchmark.d:112 #4 0x08089105 in __lambda7 (__capture=0xf7411000) at direntrybenchmark.d:158 #5 0x0809f8aa in benchmark (n=1, __capture=) at /md11/sda2-usr2l/gcc-12.1/lib/gcc/x86_64-pc-linux-gnu/12.1.0/include/d/std/datetime/stopwatch.d:421 #6 __foreachbody9 (__capture=0xf7411000, __applyArg0=..., __applyArg1=...) at direntrybenchmark.d:162 #7 0xf7c98bd9 in _aaApply2 () from [...]gcc-12.1/lib/libgphobos.so.3 #8 0x0808dee1 in direntrybenchmark.main_() (args=...) at direntrybenchmark.d:161
Re: dirEntries removes entire branches of empty directories
On Friday, 11 November 2022 at 16:00:12 UTC, Ali Çehreli wrote: On 11/11/22 05:13, kdevel wrote: > dmd -O compiled patched (see below!) version applied to /usr/bin on my > desktop > yields: > > ftw : 363 ms, 750 ÎŒs, and 5 [*] > dirEntries: 18 secs, 831 ms, 738 ÎŒs, and 3 [*] Great. I did not use -O with my test. It may have to do something with the performance of the hard disk. It has to do with the large number of symlinks. When I use dirEntries(root, SpanMode.depth, false) the runtime is dramatically reduced and with entries ~= DirectoryEntry(entry, entry.size); the runtimes are ftw : 98 ms, 470 ÎŒs, and 2 *beeep* dirEntries: 170 ms, 515 ÎŒs, and 2 *beeep* (to be continued)
Re: dirEntries removes entire branches of empty directories
On 11/11/22 08:00, Ali Çehreli wrote: > It may have to do something with the performance of the hard disk. I meant "the reason you got a much better improvement" may have to do something with the performance differences of your hard disk and mine. Ali
Re: dirEntries removes entire branches of empty directories
On 11/11/22 05:13, kdevel wrote:
> dmd -O compiled patched (see below!) version applied to /usr/bin on my
> desktop
> yields:
>
> ftw : 363 ms, 750 ÎŒs, and 5 [*]
> dirEntries: 18 secs, 831 ms, 738 ÎŒs, and 3 [*]
Great. I did not use -O with my test. It may have to do something with
the performance of the hard disk.
ftw wins big time. Being just a D binding of a C library function, its
compilation should be quick too.
>> entries ~= DirectoryEntry(entry, entry.getSize);
>> }
>
> strace reports that entry.getSize invokes stat on the file a second
> time. Isn't
> the stat buf saved in the entry?
That's my bad. entry.size is the cached version of the file size.
> This also gives rise for a complication with symlinks pointing to the
> directory
> which contain them:
>
> $ pwd
> /tmp/k/sub
> $ ln -s . foo
> $ ../direntrybenchmark .
>
std.file.FileException@8[...]/linux/bin64/../../src/phobos/std/file.d(1150):
./foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo:
Too many levels of symbolic links
So, ftw does not have that problem? Perhaps because of its default
symlink behavior? There is also the more capable nftw, where the caller
can specify some flags. And yes, there it is:
FTW_PHYS
If set, do not follow symbolic links. (This is what you want.)
If not set, symbolic links are followed, but no file is reported
twice.
If FTW_PHYS is not set, but FTW_DEPTH is set, then the function
fn() is never called for a directory that would be a descendant
of itself.
> -const dir = buildNormalizedPath("/home/ali/dlang");
> +const dir = buildNormalizedPath(args[1]);
That one, and I had switched the arguments on the following call. One
more example where string interpolation would be useful:
writefln!"Ignoring type %s file: %s\n(See 'man nftw')b"(
path, typeflag);
I meant the arguments in the reverse order there.
OT: And there is a 'b' character at the end of that format string which
almost certainly appeared when I botched a Ctrl-b command in my editor. :)
Ali
Re: dirEntries removes entire branches of empty directories
On Thursday, 10 November 2022 at 21:27:28 UTC, Ali Çehreli wrote:
On 11/9/22 12:06, Ali Çehreli wrote:
> I am using its sibling 'ftw'
Now that we know that dirEntries works properly, I decided not
to use ftw.
However, ftw performs about twice as fast as dirEntries
(despite some common code in the implementation below).
dmd -O compiled patched (see below!) version applied to /usr/bin
on my desktop
yields:
ftw : 363 ms, 750 ÎŒs, and 5 [*]
dirEntries: 18 secs, 831 ms, 738 ÎŒs, and 3 [*]
(* = offending units removed)
[...]
foreach (entry; dirEntries(root, SpanMode.depth)) {
if (entry.isDir) {
dirs ~= entry;
} else {
entries ~= DirectoryEntry(entry, entry.getSize);
}
strace reports that entry.getSize invokes stat on the file a
second time. Isn't
the stat buf saved in the entry?
This also gives rise for a complication with symlinks pointing to
the directory
which contain them:
$ pwd
/tmp/k/sub
$ ln -s . foo
$ ../direntrybenchmark .
std.file.FileException@8[...]/linux/bin64/../../src/phobos/std/file.d(1150): ./foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo: Too many levels of symbolic links
[...]
[...]
if (args.length != 2) {
stderr.writefln!"Please provide the directory to
walk:\n\n %s \n"
(args[0].baseName);
return 1;
}
const dir = buildNormalizedPath("/home/ali/dlang");
diff --git a/direntrybenchmark.d b/direntrybenchmark.d
index 661df51..a9a5616 100644
--- a/direntrybenchmark.d
+++ b/direntrybenchmark.d
@@ -102,8 +102,9 @@ WalkResult directoryWalk_dirEntries(string
root) {
if (entry.isDir) {
dirs ~= entry;
-} else {
-entries ~= DirectoryEntry(entry, entry.getSize);
+}
+else {
+entries ~= DirectoryEntry(entry, 0);
}
}
@@ -133,7 +134,7 @@ int main(string[] args) {
return 1;
}
-const dir = buildNormalizedPath("/home/ali/dlang");
+const dir = buildNormalizedPath(args[1]);
auto timings = benchmark!({ directoryWalk_ftw(dir); },
{ directoryWalk_dirEntries(dir);
})(10);
Re: dirEntries removes entire branches of empty directories
On 11/9/22 12:06, Ali Çehreli wrote:
> I am using its sibling 'ftw'
Now that we know that dirEntries works properly, I decided not to use ftw.
However, ftw performs about twice as fast as dirEntries (despite some
common code in the implementation below). I am leaving it here in case
somebody finds it useful. (Why don't I put it on github then; ok, some
day I will.)
import core.sys.posix.sys.stat;
import std.algorithm;
import std.exception;
import std.file;
import std.path;
import std.range;
import std.string;
// The Posix "file tree walker" function
extern (C)
int ftw(const char *dirpath,
int function (const char *fpath, const stat_t *sb, int
typeflag) fn,
int nopenfd);
enum TypeFlag {
FTW_F, // regular file
FTW_D, // directory
// See 'man nftw' or /usr/include/ftw.h for the other values
}
struct DirectoryEntry {
string name;
ulong size;
}
struct WalkResult {
DirectoryEntry[] entries;
string[] emptyDirs;
}
WalkResult directoryWalk_ftw(string root) {
WalkResult impl_() {
// These have to be 'static' because ftw() does not allow us to
pass a
// context. And that's why this function must only be called from a
// synchronized block.
static DirectoryEntry[] entries;
static string[] dirs;
entries.length = 0;
entries.assumeSafeAppend();
dirs.length = 0;
dirs.assumeSafeAppend();
// This is the callback that ftw() uses.
extern (C)
int handler(const char *fpath, const stat_t *sb, int typeflag) {
const path = fpath.fromStringz.idup;
switch (typeflag) {
case TypeFlag.FTW_F:
entries ~= DirectoryEntry(path, sb.st_size);
break;
case TypeFlag.FTW_D:
dirs ~= path;
break;
default:
import std.stdio;
writefln!"Ignoring type %s file: %s\n(See 'man nftw')b"(
path, typeflag);
break;
}
return 0;
}
// The tree walk will be faster up-to this "search depth" (See
'man nftw')
enum nopenfd = 32;
const ret = ftw(root.toStringz, , nopenfd);
enforce(ret == 0,
format!"Failed walking the directory tree at %s; error:
%s"(
root, ret));
string[] nonEmptyDirs = chain(entries.map!(e => e.name),
dirs)
.map!dirName
.array
.sort
.uniq
.array;
sort(dirs);
string[] emptyDirs = setDifference(dirs, nonEmptyDirs)
.array;
return WalkResult(entries.dup, emptyDirs);
}
synchronized {
return impl_();
}
}
WalkResult directoryWalk_dirEntries(string root) {
DirectoryEntry[] entries;
string[] dirs;
foreach (entry; dirEntries(root, SpanMode.depth)) {
if (entry.isDir) {
dirs ~= entry;
} else {
entries ~= DirectoryEntry(entry, entry.getSize);
}
}
string[] nonEmptyDirs = chain(entries.map!(e => e.name),
dirs)
.map!dirName
.array
.sort
.uniq
.array;
sort(dirs);
string[] emptyDirs = setDifference(dirs, nonEmptyDirs)
.array;
return WalkResult(entries.dup, emptyDirs);
}
int main(string[] args) {
import std.datetime.stopwatch;
import std.stdio;
import std.path;
if (args.length != 2) {
stderr.writefln!"Please provide the directory to walk:\n\n %s
\n"
(args[0].baseName);
return 1;
}
const dir = buildNormalizedPath("/home/ali/dlang");
auto timings = benchmark!({ directoryWalk_ftw(dir); },
{ directoryWalk_dirEntries(dir); })(10);
writefln!("ftw : %s\n" ~
"dirEntries: %s")(timings[0], timings[1]);
return 0;
}
Ali
Re: dirEntries removes entire branches of empty directories
On Thu, Nov 10, 2022 at 07:07:33PM +, Imperatorn via Digitalmars-d-learn wrote: > On Thursday, 10 November 2022 at 16:34:53 UTC, Ali Çehreli wrote: > > On 11/9/22 11:30, Vladimir Panteleev wrote: > > > On Wednesday, 9 November 2022 at 19:05:58 UTC, Ali Çehreli > > > wrote: > > >> Running the program shows no output; 'a' is not visited as a > > >> directory entry. > > > > > > That's not what happens for me: > > > > Does not happen for me today either. (?) I must have confused myself > > both with my actual program and with a trivial isolated program that I > > had written to test it. > > > > Unless others have seen the same behavior yesterday there is no bug here > > today. :p > > > > Ali > > "walks away with a confused look on his face" > > Oh, did you run the program on Wednesday? Fool! I think it was because yesterday MSFT stock dipped, but today it rose by 15, so Windows is working properly again. :-P T -- "You are a very disagreeable person." "NO."
Re: dirEntries removes entire branches of empty directories
On Thursday, 10 November 2022 at 16:34:53 UTC, Ali Çehreli wrote: On 11/9/22 11:30, Vladimir Panteleev wrote: > On Wednesday, 9 November 2022 at 19:05:58 UTC, Ali Çehreli wrote: >> Running the program shows no output; 'a' is not visited as a directory >> entry. > > That's not what happens for me: Does not happen for me today either. (?) I must have confused myself both with my actual program and with a trivial isolated program that I had written to test it. Unless others have seen the same behavior yesterday there is no bug here today. :p Ali "walks away with a confused look on his face" Oh, did you run the program on Wednesday? Fool!
Re: dirEntries removes entire branches of empty directories
On 11/9/22 11:30, Vladimir Panteleev wrote: > On Wednesday, 9 November 2022 at 19:05:58 UTC, Ali Çehreli wrote: >> Running the program shows no output; 'a' is not visited as a directory >> entry. > > That's not what happens for me: Does not happen for me today either. (?) I must have confused myself both with my actual program and with a trivial isolated program that I had written to test it. Unless others have seen the same behavior yesterday there is no bug here today. :p Ali "walks away with a confused look on his face"
Re: dirEntries removes entire branches of empty directories
On Wednesday, 9 November 2022 at 19:59:57 UTC, Ali Çehreli wrote: On 11/9/22 11:48, Imperatorn wrote: > That's not the behaviour I get in Windows. Windows users deserve it! :p (At least it is better in this case. :) ) > When I create the subdirectory, I see it even if it's empty struct DirIteratorImpl has different implementations for Windows, etc. Ali Anyway, it's definitely a bug in that implementation
Re: dirEntries removes entire branches of empty directories
On Wednesday, 9 November 2022 at 20:06:15 UTC, Ali Çehreli wrote: On 11/9/22 11:05, Ali Çehreli wrote: It was pretty easy to use but there is a quality issue there: They failed to support a 'void*' context for the user! You can walk the tree but can't put the results into your local context! Boo!
Re: dirEntries removes entire branches of empty directories
On 11/9/22 11:05, Ali Çehreli wrote: > Can you think of a workaround to achieve that? Me, me, me! :) I've learned about the Posix function 'nftw' (but I am using its sibling 'ftw'). It was pretty easy to use but there is a quality issue there: They failed to support a 'void*' context for the user! You can walk the tree but can't put the results into your local context! Boo! I guess it was designed by someone who is happy with global variables. :) At least D makes it easy to guard access to module variables with 'synchronized', shared, etc. Ali
Re: dirEntries removes entire branches of empty directories
On 11/9/22 11:48, Imperatorn wrote: > That's not the behaviour I get in Windows. Windows users deserve it! :p (At least it is better in this case. :) ) > When I create the subdirectory, I see it even if it's empty struct DirIteratorImpl has different implementations for Windows, etc. Ali
Re: dirEntries removes entire branches of empty directories
On Wednesday, 9 November 2022 at 19:05:58 UTC, Ali Çehreli wrote: In case it matters, the file system is ext4. 1) Create a directory: [...] That's not the behaviour I get in Windows. When I create the subdirectory, I see it even if it's empty
Re: dirEntries removes entire branches of empty directories
On Wednesday, 9 November 2022 at 19:05:58 UTC, Ali Çehreli wrote:
In case it matters, the file system is ext4.
My code runs in tmp (tmpfs).
2) Make a sub-directory:
mkdir deleteme/a
Running the program shows no output; 'a' is not visited as a
directory entry.
Was say strace/ltrace?
```didi.d
import std.stdio;
import std.file;
void main (string [] args)
{
auto de = dirEntries (args[1], SpanMode.breadth);
foreach (e; de)
writeln(e.name);
}
```
```
$ mkdir -p deleteme/a
$ dmd didi
$ ./didi deleteme
deleteme/a
Do you think this is buggy behavior for dirEntries?
Sure.
Re: dirEntries removes entire branches of empty directories
On Wednesday, 9 November 2022 at 19:05:58 UTC, Ali Çehreli wrote:
Running the program shows no output; 'a' is not visited as a
directory entry.
That's not what happens for me:
```d
import std.exception;
import std.file;
import std.path;
import std.stdio;
void ls()
{
foreach (e; dirEntries(absolutePath("./deleteme"),
SpanMode.breadth)) {
writeln(e.name);
}
}
void main()
{
"./deleteme".rmdirRecurse.collectException;
"./deleteme".mkdir();
writeln("empty");
ls();
writeln("only a directory");
mkdir("./deleteme/a");
ls();
writeln("directory and file");
std.file.write("./deleteme/a/x", "");
ls();
}
```
Locally and on run.dlang.io I get:
```
empty
only a directory
/sandbox/./deleteme/a
directory and file
/sandbox/./deleteme/a
/sandbox/./deleteme/a/x
```
