On 7/22/19 12:52 PM, Walter Wu wrote: > On Thu, 2019-07-18 at 19:11 +0300, Andrey Ryabinin wrote: >> >> On 7/15/19 6:06 AM, Walter Wu wrote: >>> On Fri, 2019-07-12 at 13:52 +0300, Andrey Ryabinin wrote: >>>> >>>> On 7/11/19 1:06 PM, Walter Wu wrote: >>>>> On Wed, 2019-07-10 at 21:24 +0300, Andrey Ryabinin wrote: >>>>>> >>>>>> On 7/9/19 5:53 AM, Walter Wu wrote: >>>>>>> On Mon, 2019-07-08 at 19:33 +0300, Andrey Ryabinin wrote: >>>>>>>> >>>>>>>> On 7/5/19 4:34 PM, Dmitry Vyukov wrote: >>>>>>>>> On Mon, Jul 1, 2019 at 11:56 AM Walter Wu <walter-zh.wu@xxxxxxxxxxxx> wrote: >>>>>> >>>>>>>>> >>>>>>>>> Sorry for delays. I am overwhelm by some urgent work. I afraid to >>>>>>>>> promise any dates because the next week I am on a conference, then >>>>>>>>> again a backlog and an intern starting... >>>>>>>>> >>>>>>>>> Andrey, do you still have concerns re this patch? This change allows >>>>>>>>> to print the free stack. >>>>>>>> >>>>>>>> I 'm not sure that quarantine is a best way to do that. Quarantine is made to delay freeing, but we don't that here. >>>>>>>> If we want to remember more free stacks wouldn't be easier simply to remember more stacks in object itself? >>>>>>>> Same for previously used tags for better use-after-free identification. >>>>>>>> >>>>>>> >>>>>>> Hi Andrey, >>>>>>> >>>>>>> We ever tried to use object itself to determine use-after-free >>>>>>> identification, but tag-based KASAN immediately released the pointer >>>>>>> after call kfree(), the original object will be used by another >>>>>>> pointer, if we use object itself to determine use-after-free issue, then >>>>>>> it has many false negative cases. so we create a lite quarantine(ring >>>>>>> buffers) to record recent free stacks in order to avoid those false >>>>>>> negative situations. >>>>>> >>>>>> I'm telling that *more* than one free stack and also tags per object can be stored. >>>>>> If object reused we would still have information about n-last usages of the object. >>>>>> It seems like much easier and more efficient solution than patch you proposing. >>>>>> >>>>> To make the object reused, we must ensure that no other pointers uses it >>>>> after kfree() release the pointer. >>>>> Scenario: >>>>> 1). The object reused information is valid when no another pointer uses >>>>> it. >>>>> 2). The object reused information is invalid when another pointer uses >>>>> it. >>>>> Do you mean that the object reused is scenario 1) ? >>>>> If yes, maybe we can change the calling quarantine_put() location. It >>>>> will be fully use that quarantine, but at scenario 2) it looks like to >>>>> need this patch. >>>>> If no, maybe i miss your meaning, would you tell me how to use invalid >>>>> object information? or? >>>>> >>>> >>>> >>>> KASAN keeps information about object with the object, right after payload in the kasan_alloc_meta struct. >>>> This information is always valid as long as slab page allocated. Currently it keeps only one last free stacktrace. >>>> It could be extended to record more free stacktraces and also record previously used tags which will allow you >>>> to identify use-after-free and extract right free stacktrace. >>> >>> Thanks for your explanation. >>> >>> For extend slub object, if one record is 9B (sizeof(u8)+ sizeof(struct >>> kasan_track)) and add five records into slub object, every slub object >>> may add 45B usage after the system runs longer. >>> Slub object number is easy more than 1,000,000(maybe it may be more >>> bigger), then the extending object memory usage should be 45MB, and >>> unfortunately it is no limit. The memory usage is more bigger than our >>> patch. >> >> No, it's not necessarily more. >> And there are other aspects to consider such as performance, how simple reliable the code is. >> >>> >>> We hope tag-based KASAN advantage is smaller memory usage. If it’s >>> possible, we should spend less memory in order to identify >>> use-after-free. Would you accept our patch after fine tune it? >> >> Sure, if you manage to fix issues and demonstrate that performance penalty of your >> patch is close to zero. > > > I remember that there are already the lists which you concern. Maybe we > can try to solve those problems one by one. > > 1. deadlock issue? cause by kmalloc() after kfree()? smp_call_on_cpu() > 2. decrease allocation fail, to modify GFP_NOWAIT flag to GFP_KERNEL? No, this is not gonna work. Ideally we shouldn't have any allocations there. It's not reliable and it hurts performance. > 3. check whether slim 48 bytes (sizeof (qlist_object) + > sizeof(kasan_alloc_meta)) and additional unique stacktrace in > stackdepot? > 4. duplicate struct 'kasan_track' information in two different places > Yup. > Would you have any other concern? or? > It would be nice to see some performance numbers. Something that uses slab allocations a lot, e.g. netperf STREAM_STREAM test.
