[11/13] proc: readdir /proc/*/task
diff mbox series

Message ID 20180827231503.26899-11-adobriyan@gmail.com
State New, archived
Headers show
Series
  • [01/13] seq_file: rewrite seq_puts() in terms of seq_write()
Related show

Commit Message

Alexey Dobriyan Aug. 27, 2018, 11:15 p.m. UTC
---
 fs/proc/base.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

Comments

Ahmed S. Darwish Aug. 28, 2018, 12:36 p.m. UTC | #1
On Tue, Aug 28, 2018 at 02:15:01AM +0300, Alexey Dobriyan wrote:
> ---
>  fs/proc/base.c | 8 ++++----
>  1 file changed, 4 insertions(+), 4 deletions(-)
>

Missing description and S-o-b. Further comments below..

> diff --git a/fs/proc/base.c b/fs/proc/base.c
> index 33f444721965..668e465c86b3 100644
> --- a/fs/proc/base.c
> +++ b/fs/proc/base.c
> @@ -3549,11 +3549,11 @@ static int proc_task_readdir(struct file *file, struct dir_context *ctx)
>  	for (task = first_tid(proc_pid(inode), tid, ctx->pos - 2, ns);
>  	     task;
>  	     task = next_tid(task), ctx->pos++) {
> -		char name[10 + 1];
> -		unsigned int len;
> +		char name[10], *p = name + sizeof(name);
> +

Multiple issues:

- len should be 11, as was in the original code
  (0xffffffff = 4294967295, 10 letters)

- while we're at it, let's use a constant for the '11' instead of
  mysterious magic numbers

- 'p' is clearly overflowing the stack here

>  		tid = task_pid_nr_ns(task, ns);
> -		len = snprintf(name, sizeof(name), "%u", tid);
> -		if (!proc_fill_cache(file, ctx, name, len,
> +		p = _print_integer_u32(p, tid);
> +		if (!proc_fill_cache(file, ctx, p, name + sizeof(name) - p,

You're replacing snprintf() code __that did proper len checking__
with code that does not. That's not good.

I can't see how the fourth proc_fill_cache() parameter, ``name +
sizeof(name)'' safely ever replace the original 'len' parameter.
It's a pointer value .. (!)

Overall this looks like a broken patch submitted by mistake.

Thanks,
Ahmed S. Darwish Aug. 28, 2018, 1:04 p.m. UTC | #2
On Tue, Aug 28, 2018 at 12:36:22PM +0000, Ahmed S. Darwish wrote:
> On Tue, Aug 28, 2018 at 02:15:01AM +0300, Alexey Dobriyan wrote:
> > ---
> >  fs/proc/base.c | 8 ++++----
> >  1 file changed, 4 insertions(+), 4 deletions(-)
> >
> 
> Missing description and S-o-b. Further comments below..
> 
> > diff --git a/fs/proc/base.c b/fs/proc/base.c
> > index 33f444721965..668e465c86b3 100644
> > --- a/fs/proc/base.c
> > +++ b/fs/proc/base.c
> > @@ -3549,11 +3549,11 @@ static int proc_task_readdir(struct file *file, struct dir_context *ctx)
> >  	for (task = first_tid(proc_pid(inode), tid, ctx->pos - 2, ns);
> >  	     task;
> >  	     task = next_tid(task), ctx->pos++) {
> > -		char name[10 + 1];
> > -		unsigned int len;
> > +		char name[10], *p = name + sizeof(name);
> > +
> 
> Multiple issues:
> 
> - len should be 11, as was in the original code
>   (0xffffffff = 4294967295, 10 letters)
> 
> - while we're at it, let's use a constant for the '11' instead of
>   mysterious magic numbers
> 
> - 'p' is clearly overflowing the stack here
>

See below:

> >  		tid = task_pid_nr_ns(task, ns);
> > -		len = snprintf(name, sizeof(name), "%u", tid);
> > -		if (!proc_fill_cache(file, ctx, name, len,
> > +		p = _print_integer_u32(p, tid);
> > +		if (!proc_fill_cache(file, ctx, p, name + sizeof(name) - p,
> 
> You're replacing snprintf() code __that did proper len checking__
> with code that does not. That's not good.
> 
> I can't see how the fourth proc_fill_cache() parameter, ``name +
> sizeof(name)'' safely ever replace the original 'len' parameter.
> It's a pointer value .. (!)
>

Ok, there's a "- p" in the end, so the length looks to be Ok.

Nonetheless, the whole patch series is introducing funny code
like:

+/*
+ * Print an integer in decimal.
+ * "p" initially points PAST THE END OF THE BUFFER!
+ *
+ * DO NOT USE THESE FUNCTIONS!
+ *
+ * Do not copy these functions.
+ * Do not document these functions.
+ * Do not move these functions to lib/ or elsewhere.
+ * Do not export these functions to modules.
+ * Do not tell anyone about these functions.
+ */
+noinline
+char *_print_integer_u32(char *p, u32 x)
+{
+       do {
+               *--p = '0' + (x % 10);
+               x /= 10;
+       } while (x != 0);
+       return p;
+}

And thus the code using these functions is throwing invalid
past-the-stack pointers and strings with no NULL terminators
like there's no tomorrow...

IMHO It's an accident waiting to happen to sprinkle pointers
like that everywhere. Are we really in a super hot path to
justify all that?

/me confused
Alexey Dobriyan Aug. 28, 2018, 7:31 p.m. UTC | #3
On Tue, Aug 28, 2018 at 12:36:22PM +0000, Ahmed S. Darwish wrote:
> On Tue, Aug 28, 2018 at 02:15:01AM +0300, Alexey Dobriyan wrote:
> > ---
> >  fs/proc/base.c | 8 ++++----
> >  1 file changed, 4 insertions(+), 4 deletions(-)
> >
> 
> Missing description and S-o-b. Further comments below..
> 
> > diff --git a/fs/proc/base.c b/fs/proc/base.c
> > index 33f444721965..668e465c86b3 100644
> > --- a/fs/proc/base.c
> > +++ b/fs/proc/base.c
> > @@ -3549,11 +3549,11 @@ static int proc_task_readdir(struct file *file, struct dir_context *ctx)
> >  	for (task = first_tid(proc_pid(inode), tid, ctx->pos - 2, ns);
> >  	     task;
> >  	     task = next_tid(task), ctx->pos++) {
> > -		char name[10 + 1];
> > -		unsigned int len;
> > +		char name[10], *p = name + sizeof(name);
> > +
> 
> Multiple issues:
> 
> - len should be 11, as was in the original code
>   (0xffffffff = 4294967295, 10 letters)

len should be 10.

> - while we're at it, let's use a constant for the '11' instead of
>   mysterious magic numbers
> 
> - 'p' is clearly overflowing the stack here
> 
> >  		tid = task_pid_nr_ns(task, ns);
> > -		len = snprintf(name, sizeof(name), "%u", tid);
> > -		if (!proc_fill_cache(file, ctx, name, len,
> > +		p = _print_integer_u32(p, tid);
> > +		if (!proc_fill_cache(file, ctx, p, name + sizeof(name) - p,
> 
> You're replacing snprintf() code __that did proper len checking__
> with code that does not. That's not good.

Yes, the whole point of the patch is to skip length checking.

> I can't see how the fourth proc_fill_cache() parameter, ``name +
> sizeof(name)'' safely ever replace the original 'len' parameter.
> It's a pointer value .. (!)
> 
> Overall this looks like a broken patch submitted by mistake.
Alexey Dobriyan Aug. 28, 2018, 7:35 p.m. UTC | #4
On Tue, Aug 28, 2018 at 01:04:40PM +0000, Ahmed S. Darwish wrote:
> On Tue, Aug 28, 2018 at 12:36:22PM +0000, Ahmed S. Darwish wrote:
> > On Tue, Aug 28, 2018 at 02:15:01AM +0300, Alexey Dobriyan wrote:
> > > ---
> > >  fs/proc/base.c | 8 ++++----
> > >  1 file changed, 4 insertions(+), 4 deletions(-)
> > >
> > 
> > Missing description and S-o-b. Further comments below..
> > 
> > > diff --git a/fs/proc/base.c b/fs/proc/base.c
> > > index 33f444721965..668e465c86b3 100644
> > > --- a/fs/proc/base.c
> > > +++ b/fs/proc/base.c
> > > @@ -3549,11 +3549,11 @@ static int proc_task_readdir(struct file *file, struct dir_context *ctx)
> > >  	for (task = first_tid(proc_pid(inode), tid, ctx->pos - 2, ns);
> > >  	     task;
> > >  	     task = next_tid(task), ctx->pos++) {
> > > -		char name[10 + 1];
> > > -		unsigned int len;
> > > +		char name[10], *p = name + sizeof(name);
> > > +
> > 
> > Multiple issues:
> > 
> > - len should be 11, as was in the original code
> >   (0xffffffff = 4294967295, 10 letters)
> > 
> > - while we're at it, let's use a constant for the '11' instead of
> >   mysterious magic numbers
> > 
> > - 'p' is clearly overflowing the stack here
> >
> 
> See below:
> 
> > >  		tid = task_pid_nr_ns(task, ns);
> > > -		len = snprintf(name, sizeof(name), "%u", tid);
> > > -		if (!proc_fill_cache(file, ctx, name, len,
> > > +		p = _print_integer_u32(p, tid);
> > > +		if (!proc_fill_cache(file, ctx, p, name + sizeof(name) - p,
> > 
> > You're replacing snprintf() code __that did proper len checking__
> > with code that does not. That's not good.
> > 
> > I can't see how the fourth proc_fill_cache() parameter, ``name +
> > sizeof(name)'' safely ever replace the original 'len' parameter.
> > It's a pointer value .. (!)
> >
> 
> Ok, there's a "- p" in the end, so the length looks to be Ok.
> 
> Nonetheless, the whole patch series is introducing funny code
> like:
> 
> +/*
> + * Print an integer in decimal.
> + * "p" initially points PAST THE END OF THE BUFFER!
> + *
> + * DO NOT USE THESE FUNCTIONS!
> + *
> + * Do not copy these functions.
> + * Do not document these functions.
> + * Do not move these functions to lib/ or elsewhere.
> + * Do not export these functions to modules.
> + * Do not tell anyone about these functions.
> + */
> +noinline
> +char *_print_integer_u32(char *p, u32 x)
> +{
> +       do {
> +               *--p = '0' + (x % 10);
> +               x /= 10;
> +       } while (x != 0);
> +       return p;
> +}
> 
> And thus the code using these functions is throwing invalid
> past-the-stack pointers and strings with no NULL terminators
> like there's no tomorrow...
> 
> IMHO It's an accident waiting to happen to sprinkle pointers
> like that everywhere.

It is not if people will be prohibited from moving this code to lib/ and
"improving" it by adding more parameters.

> Are we really in a super hot path to justify all that?

/proc is very slow, try profiling just about anything involving /proc.
Andrew Morton Aug. 29, 2018, 11:43 p.m. UTC | #5
On Tue, 28 Aug 2018 22:35:02 +0300 Alexey Dobriyan <adobriyan@gmail.com> wrote:

> > Are we really in a super hot path to justify all that?
> 
> /proc is very slow, try profiling just about anything involving /proc.

So how much does this patchset help?  Some timing measurements would
really help things along, if they show goodness.
Alexey Dobriyan Aug. 30, 2018, 1:20 p.m. UTC | #6
On Wed, Aug 29, 2018 at 04:43:30PM -0700, Andrew Morton wrote:
> On Tue, 28 Aug 2018 22:35:02 +0300 Alexey Dobriyan <adobriyan@gmail.com> wrote:
> 
> > > Are we really in a super hot path to justify all that?
> > 
> > /proc is very slow, try profiling just about anything involving /proc.
> 
> So how much does this patchset help?  Some timing measurements would
> really help things along, if they show goodness.

Benchmarking results are in individual patches. They are ~1-7-20%
better overall. However those are microbenchmarks.

Patch
diff mbox series

diff --git a/fs/proc/base.c b/fs/proc/base.c
index 33f444721965..668e465c86b3 100644
--- a/fs/proc/base.c
+++ b/fs/proc/base.c
@@ -3549,11 +3549,11 @@  static int proc_task_readdir(struct file *file, struct dir_context *ctx)
 	for (task = first_tid(proc_pid(inode), tid, ctx->pos - 2, ns);
 	     task;
 	     task = next_tid(task), ctx->pos++) {
-		char name[10 + 1];
-		unsigned int len;
+		char name[10], *p = name + sizeof(name);
+
 		tid = task_pid_nr_ns(task, ns);
-		len = snprintf(name, sizeof(name), "%u", tid);
-		if (!proc_fill_cache(file, ctx, name, len,
+		p = _print_integer_u32(p, tid);
+		if (!proc_fill_cache(file, ctx, p, name + sizeof(name) - p,
 				proc_task_instantiate, task, NULL)) {
 			/* returning this tgid failed, save it as the first
 			 * pid for the next readir call */