Message-ID: <CAPj211vOqGBQXK0W0yYYsUzV0DJvvEd0bWF=3Fzj5E3BLkzAcA@mail.gmail.com>
Date:   Tue, 29 Nov 2022 00:06:05 +0100
From:   "Fabio M. De Francesco" <fmdefrancesco@...il.com>
To:     Al Viro <viro@...iv.linux.org.uk>
Cc:     Evgeniy Dushistov <dushistov@...l.ru>,
        Christian Brauner <brauner@...nel.org>,
        linux-kernel@...r.kernel.org, bpf@...r.kernel.org,
        "Venkataramanan, Anirudh" <anirudh.venkataramanan@...el.com>,
        Ira Weiny <ira.weiny@...el.com>
Subject: Re: [RESEND PATCH v4] fs/ufs: Replace kmap() with kmap_local_page()

On domenica 27 novembre 2022 17:15:28 CET Al Viro wrote:
> On Sat, Nov 26, 2022 at 03:12:16PM +0100, Fabio M. De Francesco wrote:
> > Al, thanks for taking the time to answer in such detail although I can
> > believe that, to spot the mistakes in my patch and then write your email,
> > it probably took you 10 minutes or so :-)
>
> FWIW, it's generally useful to ask yourself if patch can be made less noisy
-
> are there any regular parts that could be carved out of it, how much is due
to
> calling conventions changes and can those be done differently, etc.
>
> If it can be decomposed into simpler parts, it becomes easier to review (and
> that includes debugging - when bisect points to a commit and you need to
find
> what's wrong with it, splitting that change into several steps is often
> a useful approach; if nothing else, you get a chance to bisect that and
> narrow the area to look into, but often enough you'll spot the bug while
> carving the change up).

About a year ago, Greg K-H asked me (and another developer who was working
with me) to split a small batch of 5 patches into more than 20. He explained
pretty much the same reasons why we should split patches into smaller parts,
easier to understand and review. It's also much easier for the developers
themselves to prepare and debug before committing.

I thought I got the message, but I fell into the same trap again and sent one
big patch (two if we're even talking about fs/sysv) which is not so easy to
review. I guess that's why they've been set aside and not commented on for a
long time.

>
> Below is what I'd actually done when reviewing that thing (and it took more
> than 10 minutes):

I was just kidding when I was talking about those "10 minutes". At the same
time I wanted to underline how the things that are difficult for me are much
easier for you to analyze and solve.

>
> What happens in this patch?

What follows is a great lesson about how to (1) make sense of what this change
is doing (2) how it is doing it (3) how we can optimize (especially how to
remove redundancy and unnecessary complexity.

The lessons which follow are great. Furthermore, I'm also having the
opportunity to follow your mental process. That will also help very much with
this task and also in future.

I see no need to reply to each of the suggestions you provided with such
clarity and huge amount of details.

As I already said I'll rework trying to do my best, although it will take some
time because of several personal reasons that are of no interests to the
readers.

I'm sincerely grateful to you for spending so much time (forget those 10
minutes, please) to help me.

Again, thank you very much,

Fabio

>       1) switch from kmap/kunmap to kmap_local_page/kunmap_local; that's
> in ufs_get_page() and ufs_put_page().  OK, so ufs_put_page() needs to be
> supplied with the virtual address in addition to page.  And callers of
> ufs_get_page() need the virtual address to access, not just the page.
>
>       2) pages fed to ufs_put_page() ultimately come from earlier
> ufs_get_page(), so presumably virtual address could be propagated along
> the same path.  OK...
>
>       3) change in ufs_inode_by_name() seems to fit the above -
> ufs_find_entry() used to return page, now it returns page and address;
> that address is fed to ufs_put_page().
>       So what has ufs_find_entry() become?  2 "in" arguments (directory
> inode and entry name), 2 "out" arguments (page and virtual address of the
> first byte in that page) and return value - pointer to directory entry.
> So basically we are passing directory and name and get a triple -
>       * pointer to directory entry
>       * page reference that need to be held so we could access that
> directory entry
>       * pointer to the first byte of the page that contains directory
> entry...
>       Looks redundant - the 3rd component is simply the 1st one rounded
> down to page size...
>
>       And it seems that at least some of the affected functions are
> in the same situation - they are already working in terms of <pointer to
> directory entry, pointer to page> pairs, and with those adding the
> virtual address is redundant.  Let's see...
>
>       4) functions with changed calling conventions are
>               ufs_put_page()
>               ufs_set_link()
>               ufs_check_page()
>               ufs_get_page()
>               ufs_dotdot()
>               ufs_find_entry()
>               ufs_delete_entry()
>       Out of those, ufs_put_page() and ufs_get_page() are consumer and
> producer of the virtual address, and ufs_check_page() needs the address
> as well, but other four all seem to be in the same boat -
> pair <pointer to directory entry, page> becomes a triple
> <pointer to directory entry, page, pointer to the first byte of page>
> and that's redundant.
>       ufs_set_link() and ufs_delete_entry() have that triple passed
> to them; ufs_find_entry() and ufs_dotdot() return such triple to the
> caller.
>       Looks like we could replace e.g.
>       de = ufs_dotdot(inode, &page, &addr);
>       if (de) {
>               ...
>               ufs_put_page(page, addr);
>       }
> with
>       de = ufs_dotdot(inode, &page);
>       if (de) {
>               addr = <round de down>
>               ...
>               ufs_put_page(page, addr);
>       }
>       That would return the call to what it used to be; whether it's a win
> or not, depends upon the things we use addr for.  Let's check...
>
>       5) ufs_set_link(inode, de, page, addr, ...) always gets de, page and
> addr coming from the same ufs_find_entry() or ufs_dotdot() call.  I.e.
> in all cases addr is rounded-down de; could as well find that value inside
> ufs_set_link(), returning the callers to original state.
>
>       6) ufs_delete_entry(dir, de, page, addr) - same story; de, page and
> addr come from the same call of ufs_find_entry(), so we can reconstruct
> addr by de inside ufs_delete_entry(); no need to pass it down, returning
> those callers to original state.
>
>       7) ufs_find_entry() - 4 callers, address is passed to
ufs_put_page(),
> ufs_delete_entry(), ufs_set_link().  The last two don't really need that
> argument, so we are left with "need to round down for passing to
> ufs_put_page()".
>
>       8) ufs_dotdot() - address is passed to ufs_put_page() or
ufs_set_link().
> The latter doesn't need that argument, so we are again left with "need to
> round down for passing to ufs_put_page()".
>
>       9) So ufs_put_page() gets rounded-down pointer to directory entry
here...
> Might be worth a helper that would take a page + address and do
ufs_put_page()
> for page and rounded down address...  But do we even need that to be
separate
> from ufs_put_page()?  After all, the addresses we are currently passing to
> ufs_put_page() are page-aligned, so rounding any of them down to page size
is
> a no-op.  Might as well change ufs_put_page() from "takes page reference and
> virtual address of beginning of the page" to "takes page reference and
> virtual address anywhere within the page"; if nothing else, it would reduce
> the clutter in fs/ufs/namei.c parts and it might be useful in fs/ufs/dir.c
> part as well.  OK, the intermediate plan:
>       * ufs_put_page() rounds the address down to page size
>       * ufs_set_link() and ufs_delete_entry() do not get page_addr passed
> to them; reconstruct it as rounded-down directory entry pointer.  Calling
> conventions are back to original.
>       * ufs_find_entry() and ufs_dotdot() do not bother returning
page_addr.
> Calling conventions are back to original.
>       * callers of ufs_put_page() in fs/ufs/namei.c give it directory
entry
> instead of page_addr.  Same for ufs_inode_by_name(), that got us started in
> that direction.
>
>       10) ufs_set_link() uses page_addr for two things - finding location
> in file (page_offset() + difference between de and page_addr) and feeding it
> to ufs_put_page().  The latter can be given de, as we do in other places;
> the former would become "difference between de and de rounded down to page
> size"... wait a sec - that's just the address modulo page size, aka.
> offset_in_page(de).  OK, we don't even need to reconstruct page_addr in
> there.  Win...
>
>       11) for ufs_get_page() we want the virtual address returned.
Looking
> through the callers, though, shows that calling conventions are clumsy -
> we have page = ufs_get_page(dir, n, &page_addr) followed by casting
> page_addr to various pointer types, assigning it to various pointer
> variables or, in one case, even ufs_get_page(dir, n, (void
> **)&char_pointer_var) Ouch...  We have 2 "in" arguments (directory and page
> number) and
> 2 values to return to caller - page reference and pointer to beginning of
> the page.  The type of the first component is always the same (struct page
*),
> but the second... depending upon the caller, it may be void *, char * or
> struct ufs_dir_entry *.  Passing it out via "out" argument is painful -
> return value is much more suitable for that; there we can return void * and
> have the assignment in callers convert it as needed.  Page reference, OTOH,
> is well-suited for passing out via "out" argument - same type in all cases.
>       So it looks like we want addr = ufs_get_page(dir, n, &page);
> as calling conventions.  Callers need to check IS_ERR(addr) instead of
> IS_ERR(page), but that's about it.  *AND* they might be able to cut
> down on assignments - just store the return value directly in variable
> they would copy it into.
>
>       12) ufs_dotdot() needs not bother with returning page_address.
> With new calling conventions for ufs_get_page() it doesn't need
> 'page' or 'page_addr' at all - we want the value of
>       ufs_get_page(dir, 0, p)
> but we only want it converted to struct ufs_dir_entry *.  And we already
> have a variable of just that type declared right there, so we can do
> simply
>       struct ufs_dir_entry *de = ufs_get_page(dir, 0, p);
> The rest becomes
>       if (!IS_ERR(de)) {
>               de = ufs_next_entry(dir->i_sb, de);
>       } else {
>               de = NULL;
>       }
>       return de;
> or, better yet,
>       if (!IS_ERR(de))
>               return ufs_next_entry(dir->i_sb, de);
>       else
>               return NULL;
> Makes sense - ".." is the second entry in directory, so we ask for the
> first page, treat its beginning as entry pointer and return the next
> entry...  And yes, it looks like ufs_get_page() callers are happier
> with such calling conventions.
>
>       13) ufs_find_entry() - getting rid of res_page_addr argument
> is obvious, so's adjusting the calling conventions for ufs_get_page().
> Surprisingly subtle question is whether we can get rid of 'page_addr'
> there; now that we don't store it in *res_page_addr the only thing
> we use it for is ufs_put_page().  With ufs_get_page() calling
> conventions change we'd get
>       page_addr = ufs_get_page(dir, n, &page);
>       if (!IS_ERR(page_addr)) {
>               kaddr = page_addr;
>               de = (struct ufs_dir_entry *) kaddr;
>               kaddr += ufs_last_byte(dir, n) - reclen;
>               while ((char *) de <= kaddr) {
>                      if (ufs_match(sb, namelen, name, de))
>                               goto found;
>                       de = ufs_next_entry(sb, de);
>               }
>               ufs_put_page(page, page_addr);
>       }
> We have a couple of other pointers around - kaddr and de.  kaddr is
> set to page_addr + something, then de iterates through the entries
> starting at page_addr until it gets past kaddr, looking for the match.
> So we get to the end of loop if we have *not* found a match within
> this page, and de is not guaranteed to point within the same page when
> we leave the loop.  Not suitable for passing to ufs_put_page(), then.
> kaddr, OTOH, would better be pointing into the same page - we use
> de <= kaddr as permission to dereference de, and we'd better not run
> out of page.  A look at ufs_last_byte() shows that its return value
> is always no greater than PAGE_SIZE, so kaddr is guaranteed to point
> within the same page.  Good.  So we can turn that into
>       kaddr = ufs_get_page(dir, n, &page);
>       if (!IS_ERR(kaddr)) {
>               de = (struct ufs_dir_entry *) kaddr;
>               kaddr += ufs_last_byte(dir, n) - reclen;
>               ....
>               ufs_put_page(page, kaddr);
>       }
> and page_addr bites the dust.
>
>       14) ufs_add_link() - similar adjustment to new calling conventions
> for ufs_get_page().  Uses of page_addr: fed to ufs_put_page() (same as
> in ufs_find_entry() kaddr is guaranteed to point into the same page and
> thus can be used instead) and calculation of position in directory, same
> as we'd seen in ufs_set_link().  The latter becomes page_offset(page) +
> offset_in_page(de), killing page_addr off.  BTW, we get
>               kaddr = ufs_get_page(dir, n, &page);
>               err = PTR_ERR(kaddr);
>               if (IS_ERR(kaddr))
>                       goto out;
> with out: being just 'return err;', which suggests
>               kaddr = ufs_get_page(dir, n, &page);
>               if (IS_ERR(kaddr))
>                       return ERR_PTR(kaddr);
> instead (and that was the only goto out; so the label can be removed).
> The value stored in err in case !IS_ERR(kaddr) is (thankfully) never
> used - would've been a bug otherwise.  So this is an equivalent
> transformation.
>
>       15) ufs_readdir() - trivial modifications to deal with
ufs_get_page()
> calling conventions change; an ugly cast goes away.
>
>       16) ufs_delete_entry() - kaddr argument is dropped, we make it a
local
> variable initialized with rounded down dir.  Incidentally, 'dir - kaddr' can
> be replaced with offset_in_page(dir) in there and similar for (char *)pde -
> kaddr.
>
>       17) ufs_empty_dir() - we adjust to ufs_get_page() calling
conventions
> change, page_addr can be dropped same as in ufs_find_entry().  Need to lift
> kaddr out of loop to deal with not_empty: part (hadn't noticed when writing
> original reply, but compiler would've caught that)
>
>       18) fs/ufs/namei.c parts do shrink down to pretty much just 'pass
> the right directory entry to ufs_put_page()' (and most of ufs.h changes
> disapper).  ufs_get_page() calling conventions change is actually the bulk
> of what remains of this patch after that massage.  Other than that we have
> several 'use offset_in_page()' local changes and extra argument of
> ufs_put_page().  Hmm...  offset_in_page() actually counts as a valid (and
> trivially safe) optimization, independent of anything kmap-related.
> Worth carving out as the first step in series.  Come to think of that,
> ufs_get_page() calling conventions change *also* can be done without
> the kmap_local_page() switchover; we would do kmap() and return
> page_address(page). Callers would be happier with such calling conventions,
> and carving that part out would reduce the patch to just the 'switch
> ufs_get_page() to kmap_local_page(), switch ufs_put_page() to kunmap_local()
> and supply it with a pointer into the page in question'.  Which is much
> smaller and easier to review.

Powered by blists - more mailing lists