Re: [PATCH v7 2/8] fs: Add standard casefolding support

From: Eric Biggers
Date: Wed Feb 12 2020 - 01:57:50 EST


On Tue, Feb 11, 2020 at 10:34:40PM -0800, Eric Biggers wrote:
> On Mon, Feb 10, 2020 at 11:42:07PM +0000, Al Viro wrote:
> > On Mon, Feb 10, 2020 at 03:11:13PM -0800, Daniel Rosenberg wrote:
> > > On Fri, Feb 7, 2020 at 6:12 PM Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
> > > >
> > > > On Fri, Feb 07, 2020 at 05:35:46PM -0800, Daniel Rosenberg wrote:
> > > >
> > > >
> > > > Again, is that safe in case when the contents of the string str points to
> > > > keeps changing under you?
> > >
> > > I'm not sure what you mean. I thought it was safe to use the str and
> > > len passed into d_compare. Even if it gets changed under RCU
> > > conditions I thought there was some code to ensure that the name/len
> > > pair passed in is consistent, and any other inconsistencies would get
> > > caught by d_seq later. Are there unsafe code paths that can follow?
> >
> > If you ever fetch the same byte twice, you might see different values.
> > You need a fairly careful use of READ_ONCE() or equivalents to make
> > sure that you don't get screwed over by that.
> >
> > Sure, ->d_seq mismatch will throw the result out, but you need to make
> > sure you won't oops/step on uninitialized memory/etc. in process.
> >
> > It's not impossible to get right, but it's not trivial and you need all
> > code working with that much more careful than normal for string handling.
>
> It looks like this is a real problem, not just a "theoretical" data race.
> For example, see:
>
> utf8ncursor():
> /* The first byte of s may not be an utf8 continuation. */
> if (len > 0 && (*s & 0xC0) == 0x80)
> return -1;
>
> and then utf8byte():
> } else if ((*u8c->s & 0xC0) == 0x80) {
> /* This is a continuation of the current character. */
> if (!u8c->p)
> u8c->len--;
> return (unsigned char)*u8c->s++;
>
> The first byte of the string is checked in two different functions, so it's very
> likely to be loaded twice. In between, it could change from a non-continuation
> byte to a continuation byte. That would cause the string length to be
> decremented from 0 to UINT_MAX. Then utf8_strncasecmp() would run beyond the
> bounds of the string until something happened to mismatch.
>
> That's just an example that I found right away; there are probably more.
>
> IMO, this needs to be fixed before anyone can actually use the ext4 and f2fs
> casefolding stuff.
>
> I don't know the best solution. One option is to fix fs/unicode/ to handle
> concurrently modified strings. Another could be to see what it would take to
> serialize lookups and renames for casefolded directories...
>

Or (just throwing another idea out there) the dentry's name could be copied to a
temporary buffer in ->d_compare(). The simplest version would be:

u8 _name[NAME_MAX];

memcpy(_name, name, len);
name = _name;

Though, 255 bytes is a bit large for a stack buffer (so for long names it may
need kmalloc with GFP_ATOMIC), and technically it would need a special version
of memcpy() to be guaranteed safe from compiler optimizations (though I expect
this would work in practice).

Alternatively, take_dentry_name_snapshot() kind of does this already, except
that it takes a dentry and not a (name, len) pair.

- Eric