Re: [PATCH v9 03/18] kunit: test: add string_stream a std::stream like string builder

[Brendan Higgins]

On Mon, Jul 15, 2019 at 1:43 PM Stephen Boyd <sboyd@xxxxxxxxxx> wrote:
>
> Quoting Brendan Higgins (2019-07-12 01:17:29)
> > diff --git a/include/kunit/string-stream.h b/include/kunit/string-stream.h
> > new file mode 100644
> > index 0000000000000..0552a05781afe
> > --- /dev/null
> > +++ b/include/kunit/string-stream.h
> > @@ -0,0 +1,49 @@
> > +/* SPDX-License-Identifier: GPL-2.0 */
> > +/*
> > + * C++ stream style string builder used in KUnit for building messages.
> > + *
> > + * Copyright (C) 2019, Google LLC.
> > + * Author: Brendan Higgins <brendanhiggins@xxxxxxxxxx>
> > + */
> > +
> > +#ifndef _KUNIT_STRING_STREAM_H
> > +#define _KUNIT_STRING_STREAM_H
> > +
> > +#include <linux/types.h>
> > +#include <linux/spinlock.h>
> > +#include <linux/kref.h>
>
> What is this include for? I'd expect to see linux/list.h instead.

Sorry about that. I used to reference count this before I made it a
kunit managed resource.

> > +#include <stdarg.h>
> > +
> > +struct string_stream_fragment {
> > +       struct list_head node;
> > +       char *fragment;
> > +};
> > +
> > +struct string_stream {
> > +       size_t length;
> > +       struct list_head fragments;
> > +       /* length and fragments are protected by this lock */
> > +       spinlock_t lock;
> > +};
> > +
> > diff --git a/kunit/string-stream.c b/kunit/string-stream.c
> > new file mode 100644
> > index 0000000000000..0463a92dad74b
> > --- /dev/null
> > +++ b/kunit/string-stream.c
> > @@ -0,0 +1,147 @@
> > +// SPDX-License-Identifier: GPL-2.0
> > +/*
> > + * C++ stream style string builder used in KUnit for building messages.
> > + *
> > + * Copyright (C) 2019, Google LLC.
> > + * Author: Brendan Higgins <brendanhiggins@xxxxxxxxxx>
> > + */
> > +
> > +#include <linux/list.h>
> > +#include <linux/slab.h>
> > +#include <kunit/string-stream.h>
> > +#include <kunit/test.h>
> > +
> > +int string_stream_vadd(struct string_stream *stream,
> > +                      const char *fmt,
> > +                      va_list args)
> > +{
> > +       struct string_stream_fragment *frag_container;
> > +       int len;
> > +       va_list args_for_counting;
> > +       unsigned long flags;
> > +
> > +       /* Make a copy because `vsnprintf` could change it */
> > +       va_copy(args_for_counting, args);
> > +
> > +       /* Need space for null byte. */
> > +       len = vsnprintf(NULL, 0, fmt, args_for_counting) + 1;
> > +
> > +       va_end(args_for_counting);
> > +
> > +       frag_container = kmalloc(sizeof(*frag_container), GFP_KERNEL);
>
> This is confusing in that it allocates with GFP_KERNEL but then grabs a
> spinlock to add and remove from the fragment list. Is it ever going to
> be called from a place where it can't sleep? If so, the GFP_KERNEL needs
> to be changed. Otherwise, maybe a mutex would work better to protect
> access to the fragment list.

Right, using a mutex here would be fine. Sorry, I meant to filter for
my usage of them after you asked me to remove them in 01, but
evidently I forgot to do so. Sorry, will fix.

> I also wonder if it would be better to just have a big slop buffer of a
> 4K page or something so that we almost never have to allocate anything
> with a string_stream and we can just rely on a reader consuming data
> while writers are writing. That might work out better, but I don't quite
> understand the use case for the string stream.

That makes sense, but might that also waste memory since we will
almost never need that much memory?

> > +       if (!frag_container)
> > +               return -ENOMEM;
> > +
> > +       frag_container->fragment = kmalloc(len, GFP_KERNEL);
> > +       if (!frag_container->fragment) {
> > +               kfree(frag_container);
> > +               return -ENOMEM;
> > +       }
> > +
> > +       len = vsnprintf(frag_container->fragment, len, fmt, args);
> > +       spin_lock_irqsave(&stream->lock, flags);
> > +       stream->length += len;
> > +       list_add_tail(&frag_container->node, &stream->fragments);
> > +       spin_unlock_irqrestore(&stream->lock, flags);
> > +
> > +       return 0;
> > +}
> > +
> [...]
> > +
> > +bool string_stream_is_empty(struct string_stream *stream)
> > +{
> > +       bool is_empty;
> > +       unsigned long flags;
> > +
> > +       spin_lock_irqsave(&stream->lock, flags);
>
> I'm not sure what benefit grabbing the lock is having here. If the list
> isn't empty after this is called then the race isn't resolved by
> grabbing and releasing the lock. The function is returning stale data in
> that case.

Good point, I didn't realize list_empty was protected by READ_ONCE. Will fix.

> > +       is_empty = list_empty(&stream->fragments);
> > +       spin_unlock_irqrestore(&stream->lock, flags);
> > +
> > +       return is_empty;
> > +}
> > +
> > +static int string_stream_init(struct kunit_resource *res, void *context)
> > +{
> > +       struct string_stream *stream;
> > +
> > +       stream = kzalloc(sizeof(*stream), GFP_KERNEL);
> > +       if (!stream)
> > +               return -ENOMEM;
> > +
> > +       res->allocation = stream;
> > +       INIT_LIST_HEAD(&stream->fragments);
> > +       spin_lock_init(&stream->lock);
> > +
> > +       return 0;
> > +}
> > +
> > +static void string_stream_free(struct kunit_resource *res)
> > +{
> > +       struct string_stream *stream = res->allocation;
> > +
> > +       string_stream_clear(stream);
> > +       kfree(stream);
> > +}
> > +
> > +struct string_stream *alloc_string_stream(struct kunit *test)
> > +{
> > +       struct kunit_resource *res;
> > +
> > +       res = kunit_alloc_resource(test,
> > +                                  string_stream_init,
> > +                                  string_stream_free,
> > +                                  NULL);
> > +
> > +       if (!res)
> > +               return NULL;
> > +
> > +       return res->allocation;
>
> Maybe kunit_alloc_resource() should just return res->allocation, or
> NULL, so that these functions can be simplified to 'return
> kunit_alloc_resource()'? Does the caller ever care to do anything with
> struct kunit_resource anyway?

Another good point. I think originally I thought it might, but now
with the mandatory init function, the user has to provide a function
where they can do the init work. They might as well do it there. Will
fix.