| Message ID | 20190402032957.26249-2-tobin@kernel.org (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | slob: Fix list_head bug during allocation | expand |
On Tue, 2 Apr 2019 14:29:57 +1100 "Tobin C. Harding" <tobin@kernel.org> wrote: > Currently we call (indirectly) list_del() then we manually try to combat > the fact that the list may be in an undefined state by getting 'prev' > and 'next' pointers in a somewhat contrived manner. It is hard to > verify that this works for all initial states of the list. Clearly the > author (me) got it wrong the first time because the 0day kernel testing > robot managed to crash the kernel thanks to this code. > > All this is done in order to do an optimisation aimed at preventing > fragmentation at the start of a slab. We can just skip this > optimisation any time the list is put into an undefined state since this > only occurs when an allocation completely fills the slab and in this > case the optimisation is unnecessary since we have not fragmented the slab > by this allocation. > > Change the page pointer passed to slob_alloc_page() to be a double > pointer so that we can set it to NULL to indicate that the page was > removed from the list. Skip the optimisation if the page was removed. > > Found thanks to the kernel test robot, email subject: > > 340d3d6178 ("mm/slob.c: respect list_head abstraction layer"): kernel BUG at lib/list_debug.c:31! > It's regrettable that this fixes slob-respect-list_head-abstraction-layer.patch but doesn't apply to that patch - slob-use-slab_list-instead-of-lru.patch gets in the way. So we end up with a patch series which introduces a bug and later fixes it. I guess we can live with that but if the need comes to respin this series, please do simply fix slob-respect-list_head-abstraction-layer.patch so we get a clean series.
On Mon, Apr 01, 2019 at 09:41:28PM -0700, Andrew Morton wrote: > On Tue, 2 Apr 2019 14:29:57 +1100 "Tobin C. Harding" <tobin@kernel.org> wrote: > > > Currently we call (indirectly) list_del() then we manually try to combat > > the fact that the list may be in an undefined state by getting 'prev' > > and 'next' pointers in a somewhat contrived manner. It is hard to > > verify that this works for all initial states of the list. Clearly the > > author (me) got it wrong the first time because the 0day kernel testing > > robot managed to crash the kernel thanks to this code. > > > > All this is done in order to do an optimisation aimed at preventing > > fragmentation at the start of a slab. We can just skip this > > optimisation any time the list is put into an undefined state since this > > only occurs when an allocation completely fills the slab and in this > > case the optimisation is unnecessary since we have not fragmented the slab > > by this allocation. > > > > Change the page pointer passed to slob_alloc_page() to be a double > > pointer so that we can set it to NULL to indicate that the page was > > removed from the list. Skip the optimisation if the page was removed. > > > > Found thanks to the kernel test robot, email subject: > > > > 340d3d6178 ("mm/slob.c: respect list_head abstraction layer"): kernel BUG at lib/list_debug.c:31! > > > > It's regrettable that this fixes > slob-respect-list_head-abstraction-layer.patch but doesn't apply to > that patch - slob-use-slab_list-instead-of-lru.patch gets in the way. > So we end up with a patch series which introduces a bug and later > fixes it. Yes I thought that also. Do you rebase the mm tree? Did you apply this right after slob-use-slab_list-instead-of-lru or to the current tip? If it is applied to the tip does this effect the ability to later bisect in between these two commits (if the need arises for some unrelated reason)? > I guess we can live with that but if the need comes to respin this > series, please do simply fix > slob-respect-list_head-abstraction-layer.patch so we get a clean > series. If its not too much work for you to apply the new series I'll do another version just to get this right. Tobin.
On Wed, 3 Apr 2019 06:05:38 +1100 "Tobin C. Harding" <me@tobin.cc> wrote: > > It's regrettable that this fixes > > slob-respect-list_head-abstraction-layer.patch but doesn't apply to > > that patch - slob-use-slab_list-instead-of-lru.patch gets in the way. > > So we end up with a patch series which introduces a bug and later > > fixes it. > > Yes I thought that also. Do you rebase the mm tree? Did you apply this > right after slob-use-slab_list-instead-of-lru or to the current tip? After slob-use-slab_list-instead-of-lru.patch > If > it is applied to the tip does this effect the ability to later bisect in > between these two commits (if the need arises for some unrelated reason)? There is a bisection hole but it is short and the bug is hardish to hit. > > I guess we can live with that but if the need comes to respin this > > series, please do simply fix > > slob-respect-list_head-abstraction-layer.patch so we get a clean > > series. > > If its not too much work for you to apply the new series I'll do another > version just to get this right. I guess that would be best, thanks.
diff --git a/mm/slob.c b/mm/slob.c index 21af3fdb457a..c543da10df45 100644 --- a/mm/slob.c +++ b/mm/slob.c @@ -213,10 +213,18 @@ static void slob_free_pages(void *b, int order) } /* - * Allocate a slob block within a given slob_page sp. + * slob_page_alloc() - Allocate a slob block within a given slob_page sp. + * @spp: Page to look in, return parameter. + * @size: Size of the allocation. + * @align: Allocation alignment. + * + * Tries to find a chunk of memory at least @size within page. If the + * allocation fills up page then page is removed from list, in this case + * *spp will be set to %NULL to signal that list removal occurred. */ -static void *slob_page_alloc(struct page *sp, size_t size, int align) +static void *slob_page_alloc(struct page **spp, size_t size, int align) { + struct page *sp = *spp; slob_t *prev, *cur, *aligned = NULL; int delta = 0, units = SLOB_UNITS(size); @@ -254,8 +262,11 @@ static void *slob_page_alloc(struct page *sp, size_t size, int align) } sp->units -= units; - if (!sp->units) + if (!sp->units) { clear_slob_page_free(sp); + /* Signal that page was removed from list. */ + *spp = NULL; + } return cur; } if (slob_last(cur)) @@ -268,7 +279,7 @@ static void *slob_page_alloc(struct page *sp, size_t size, int align) */ static void *slob_alloc(size_t size, gfp_t gfp, int align, int node) { - struct page *sp, *prev, *next; + struct page *sp; struct list_head *slob_list; slob_t *b = NULL; unsigned long flags; @@ -283,6 +294,7 @@ static void *slob_alloc(size_t size, gfp_t gfp, int align, int node) spin_lock_irqsave(&slob_lock, flags); /* Iterate through each partially free page, try to find room */ list_for_each_entry(sp, slob_list, slab_list) { + struct page **spp = &sp; #ifdef CONFIG_NUMA /* * If there's a node specification, search for a partial @@ -295,27 +307,25 @@ static void *slob_alloc(size_t size, gfp_t gfp, int align, int node) if (sp->units < SLOB_UNITS(size)) continue; - /* - * Cache previous entry because slob_page_alloc() may - * remove sp from slob_list. - */ - prev = list_prev_entry(sp, slab_list); - /* Attempt to alloc */ - b = slob_page_alloc(sp, size, align); + b = slob_page_alloc(spp, size, align); if (!b) continue; - next = list_next_entry(prev, slab_list); /* This may or may not be sp */ - /* - * Improve fragment distribution and reduce our average - * search time by starting our next search here. (see - * Knuth vol 1, sec 2.5, pg 449) + * If slob_page_alloc() removed sp from the list then we + * cannot call list functions on sp. Just bail, don't + * worry about the optimisation below. */ - if (!list_is_first(&next->slab_list, slob_list)) - list_rotate_to_front(&next->slab_list, slob_list); - + if (*spp) { + /* + * Improve fragment distribution and reduce our average + * search time by starting our next search here. (see + * Knuth vol 1, sec 2.5, pg 449) + */ + if (!list_is_first(&sp->slab_list, slob_list)) + list_rotate_to_front(&sp->slab_list, slob_list); + } break; } spin_unlock_irqrestore(&slob_lock, flags); @@ -334,7 +344,7 @@ static void *slob_alloc(size_t size, gfp_t gfp, int align, int node) INIT_LIST_HEAD(&sp->slab_list); set_slob(b, SLOB_UNITS(PAGE_SIZE), b + SLOB_UNITS(PAGE_SIZE)); set_slob_page_free(sp, slob_list); - b = slob_page_alloc(sp, size, align); + b = slob_page_alloc(&sp, size, align); BUG_ON(!b); spin_unlock_irqrestore(&slob_lock, flags); }
Currently we call (indirectly) list_del() then we manually try to combat the fact that the list may be in an undefined state by getting 'prev' and 'next' pointers in a somewhat contrived manner. It is hard to verify that this works for all initial states of the list. Clearly the author (me) got it wrong the first time because the 0day kernel testing robot managed to crash the kernel thanks to this code. All this is done in order to do an optimisation aimed at preventing fragmentation at the start of a slab. We can just skip this optimisation any time the list is put into an undefined state since this only occurs when an allocation completely fills the slab and in this case the optimisation is unnecessary since we have not fragmented the slab by this allocation. Change the page pointer passed to slob_alloc_page() to be a double pointer so that we can set it to NULL to indicate that the page was removed from the list. Skip the optimisation if the page was removed. Found thanks to the kernel test robot, email subject: 340d3d6178 ("mm/slob.c: respect list_head abstraction layer"): kernel BUG at lib/list_debug.c:31! Reported-by: kernel test robot <lkp@intel.com> Signed-off-by: Tobin C. Harding <tobin@kernel.org> --- mm/slob.c | 50 ++++++++++++++++++++++++++++++-------------------- 1 file changed, 30 insertions(+), 20 deletions(-)