[SSSD] [PATCH 0/4]: Actual memory cache implementation
Jan Zelený
jzeleny at redhat.com
Mon Mar 19 09:21:12 UTC 2012
> On Fri, 2012-03-16 at 14:25 +0100, Jan Zelený wrote:
> > > On Thu, 2012-01-12 at 11:49 -0500, Simo Sorce wrote:
> > > > On Tue, 2012-01-10 at 14:33 -0500, Simo Sorce wrote:
> > > > > On Tue, 2012-01-10 at 10:15 -0500, Simo Sorce wrote:
> > > > > > > Sure, we can talk about it. I'm looking at it from the users'
> > > > > > > perspectives, who I think would generally expect (and be
> > > > > > > alright
> > > > > >
> > > > > > with)
> > > > > >
> > > > > > > the fast cache being emptied on service restart. Since we still
> > > > > > > have
> > > > > >
> > > > > > the
> > > > > >
> > > > > > > not-quite-as-fast persistent LDB cache, I think the gain isn't
> > > > > > > worth
> > > > > >
> > > > > > the
> > > > > >
> > > > > > > user confusion.
> > > > > >
> > > > > > Ok, I think I can understand this, it will also simplify the code
> > > > > > I guess so I'll change the patch.
> > > > >
> > > > > Attached new set of patches.
> > > > >
> > > > > I changed the init code to always create a new cache file on
> > > > > restart, incidentally this also got rid of the stat() call, so
> > > > > that problem has been put to rest permanently :)
> > > > >
> > > > > I also changed all functions to use errno_t as the return type in
> > > > > their declaration when they return a errno code.
> > > > >
> > > > > I think all the suggestions seen in the thread so far have been
> > > > > implemented as requested at this point, with the exception of the
> > > > > part where I should implement and use new options as that required
> > > > > additional code instead of mere refactoring. I will add those
> > > > > features in a later rebase.
> > > >
> > > > New revision.
> > > >
> > > > I removed bootid from the cache file header, given we recreate the
> > > > file every time sssd restarts it was useles.
> > > >
> > > > I replaced it with a seed, and made the seed used in the hash table
> > > > randomly regenerated every time the hash table is recreated.
> > > >
> > > > This is used to avoid collision based attacks [1].
> > > >
> > > > Simo.
> > > >
> > > > [1] https://lwn.net/Articles/474912/
> > >
> > > I started looking at these last night. I have yet to complete a full
> > > review, but I'm attaching a set of patches rebased atop the current
> > > master so several of us can start reviewing.
> >
> > Ok, I finished the review. I am fine with these patches in general, I
> > have just a few comments and/or questions:
> >
> > Why only one hash table for both name and uidkey? I think using two would
> > make things a little bit more simple and fast (of course a tradeoff
> > would be slightly higher memory consumption).
>
> I wanted to limit memory consumption, it is important especially in
> 32bit apps that have a much more limited virtual memory (remember that
> this is mmapped in any application).
True, I haven't considered the mmapping by any application.
> > responder/nss/nsssrv_mmap_cache.c:220
> > - this condition is unnecessarry because of the condition on line 208
>
> Can you post the actual code snippest, if I understand what you mean I
> think I have line numbers off by 1 ... but also do not see the
> unnecessary condition. cur is adavanced since the test in line 208/209
> so it is not testing the same thing afaics.
I hope I read the code correctly, this is by far the most complex part of all
your patches.
First you have this condition:
if (mcc->free_table[t] == 0xff) {
cur = ((cur + 8) & ~7);
if (cur >= tot_slots) {
cur = 0;
}
continue;
}
Right after that you have following cycle with condition right after it:
for (t = ((cur + 8) & ~7) ; cur < t; cur++) {
MC_PROBE_BIT(mcc->free_table, cur, used);
if (!used) break;
}
if (cur == t) {
if (cur >= tot_slots) {
cur = 0;
}
continue;
}
Considering the condition if (mcc->free_table[t] == 0xff), the cur will never
equal to t. That would imply that there was no bit set to zero in mcc-
>free_table[t], which is already ruled out by the previous condition. Instead
of this condition, I'd simply place a comment above the for cycle: "There is a
zero bit in the byte, find it". I think it would improve readability of the
code.
> > responder/nss/nsssrv_mmap_cache.c:242
> > - the assignment is redundant
>
> I guess it's t, yeah that's a leftover ..
>
> > responder/nss/nsssrv_mmap_cache.c:275
> > - why this condition? I think the following while will cover it
>
> slot is used to dereference data from the table, if it accesses beyond
> the table it will cause a segfault.
I know that, my point was that from my understanding the slot variable can
either contain valid number or MC_INVALID_VAL, nothing else. Of course if I
missed anything or you are just being defensive just in case, I withdraw my
comment.
> > responder/nss/nsssrv_mmap_cache.c:516
> > - sizeof(uint32_t) -> sizeof(h.status)
>
> yes it is better to not assume.
>
> > responder/nss/nsssrv_mmap_cache.c:517
> > - sizeof(uint32_t) -> sizeof(h.status)
>
> same as above, yes.
>
> > responder/nss/nsssrv_mmap_cache.c:561
> > - shouldn't the umask better be 0133?
>
> I don't think we care for explicitly removing the executable bit, the
> default umask doesn't set it. What matters here is that it is nor
> writable.
Ok
> > close vs. munmap
> > Just out of curiosity: does it matter which should be first? I found two
> > places
>
> > in the code where the order is different:
> no it shouldn't matter.
>
> > sss_client/nss_mc_common.c:146
> > responder/nss/nsssrv_mmap_cache.c:717
>
> I changed the second one to do the munmap first
>
> > That's all from my side. As I've said, my comments are rather minor
> > things and suggestions, so in general I give these patches a green
> > light.
>
> Attached new patches with some of the picks fixed.
Thanks
Jan
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 836 bytes
Desc: This is a digitally signed message part.
URL: <https://lists.fedorahosted.org/pipermail/sssd-devel/attachments/20120319/f677c38e/attachment.sig>
More information about the sssd-devel
mailing list