Re: [PATCH v7 1/8] fs: introduce kernel_pread_file* support

[Scott Branden]

Hi Matthew,

On 2020-06-09 6:21 a.m., Matthew Wilcox wrote:
> On Mon, Jun 08, 2020 at 03:29:22PM -0700, Scott Branden wrote:
> > Hi Matthew,
> >
> > I am requesting the experts in the filesystem subsystem to come to a
> > consensus here.
> > This is not my area of expertise at all but every time I have addressed all
> > of the
> > outstanding concerns someone else comes along and raises another one.
> I appreciate it's frustrating for you, but this is the nature of
> patch review.  I haven't even read the first five or so submissions.
> I can see them in my inbox and they look like long threads.  I'm not
> particularly inclined to read them.  I happened to read v6, and reacted
> to the API being ugly.
Thanks for the review. Yes, I do see the enum being ugly now
and have removed it in v8 of the patch. Hopefully it addresses
your concerns. More comments below.
> > Please see me comments below.
> >
> > On 2020-06-06 8:52 a.m., Matthew Wilcox wrote:
> > > On Fri, Jun 05, 2020 at 10:04:51PM -0700, Scott Branden wrote:
> > > > -int kernel_read_file(struct file *file, void **buf, loff_t *size,
> > > > -		     loff_t max_size, enum kernel_read_file_id id)
> > > > -{
> > > > -	loff_t i_size, pos;
> > Please note that how checkpatch generated the diff here. The code
> > modifications
> > below are for a new function kernel_pread_file, they do not modify the
> > existing API
> > kernel_read_file. kernel_read_file requests the ENTIRE file is read. So we
> > need to be
> > able to differentiate whether it is ok to read just a portion of the file or
> > not.
> You've gone about this in entirely the wrong way though.  This enum to
> read the entire file or a partial is just bad design.
Your point on the enum is valid.
I've removed it from design. Hopefully it is cleaner now.
> > > > +int kernel_pread_file(struct file *file, void **buf, loff_t *size,
> > > > +		      loff_t pos, loff_t max_size,
> > > > +		      enum kernel_pread_opt opt,
> > > > +		      enum kernel_read_file_id id)
> > So, to share common code a new kernel_pread_opt needed to be added in order
> > to specify whether
> > it was ok to read a partial file or not, and provide an offset into the file
> > where to begin reading.
> > The meaning of parameters doesn't change in the bonkers API. max_size still
> > means max size, etc.
> > These options are needed so common code can be shared with kernel_read_file
> > api.
> Does pread() in userspace take seven parameters?  No.  It takes four.
> What you're doing is taking all the complexity of all of the interfaces
> and stuffing it all down into the bottom function instead of handling
> some of the complexity in the wrapper functions.  For example, you
> could support the functionality of 'max_size' in kernel_read_file()
> and leave it out of the kernel_pread_file() interface.
I have removed the enum necessary in the kernel pread call now,
so it is down to 6.
The other 2 parameters are necessary as they are in kernel read.

max_size makes no sense to remove - it serves the same purpose
as in userspace pread and read functions. To specify the max size
to read.
> > > I think what we actually want is:
> > >
> > > ssize_t vmap_file_range(struct file *, loff_t start, loff_t end, void **bufp);
> > > void vunmap_file_range(struct file *, void *buf);
> > >
> > > If end > i_size, limit the allocation to i_size.  Returns the number
> > > of bytes allocated, or a negative errno.  Writes the pointer allocated
> > > to *bufp.  Internally, it should use the page cache to read in the pages
> > > (taking appropriate reference counts).  Then it maps them using vmap()
> > > instead of copying them to a private vmalloc() array.
> > > kernel_read_file() can be converted to use this API.  The users will
> > > need to be changed to call kernel_read_end(struct file *file, void *buf)
> > > instead of vfree() so it can call allow_write_access() for them.
> > >
> > > vmap_file_range() has a lot of potential uses.  I'm surprised we don't
> > > have it already, to be honest.
> > Such a change sounds like it could be done in a later patch series.
> > It's an incomplete solution. It would work for some of the needed
> > operations but not others.
> > For kernel_read_file, I don't see how in your new API it indicates if the
> > end of the file was reached or not.
> That's the point.  It doesn't.  If a caller needs that, then they can
> figure that out themselves.
No, they can't. The caller only calls kernel_read_file once and expects
the whole file to be read. The kernel_read_file doesn't work like 
userspace.
There is no tracking like userspace of where in the file you read?
> > Also, please note that buffers may be preallocated and shouldn't be freed
> > by the kernel in some cases and
> > allocated and freed by the kernel in others.
> You're trying to build the swiss army knife of functions.  Swiss army
> knives are useful, but they're no good for carving a steak.
Hopefully I'm carving steak now.
> > I would like the experts here to decide on what needs to be done so we can
> > move forward
> > and get kernel_pread_file support added soon.
> You know, you haven't even said _why_ you want this.  The cover letter
> just says "I want this", and doesn't say why it's needed.
Cover letter updated.

Thanks,
Scott