On 21/11/2018 20:15, Andy Lutomirski wrote:
On Nov 21, 2018, at 9:34 AM, Igor Stoppa <igor.stoppa@xxxxxxxxx> wrote:
[...]
There might be other reasons for replicating the mm_struct.
If I understand correctly how the text patching works, it happens sequentially, because of the text_mutex used by arch_jump_label_transform
Which might be fine for this specific case, but I think I shouldn't introduce a global mutex, when it comes to data.
Most likely, if two or more cores want to perform a write rare operation, there is no correlation between them, they could proceed in parallel. And if there really is, then the user of the API should introduce own locking, for that specific case.
Text patching uses the same VA for different physical addresses, so it need a mutex to avoid conflicts. I think that, for rare writes, you should just map each rare-writable address at a *different* VA. You’ll still need a mutex (mmap_sem) to synchronize allocation and freeing of rare-writable ranges, but that shouldn’t have much contention.
I have studied the code involved with Nadav's patchset.
I am perplexed about these sentences you wrote.
More to the point (to the best of my understanding):
poking_init()
-------------
1. it gets one random poking address and ensures to have at least 2
consecutive PTEs from the same PMD
2. it then proceeds to map/unmap an address from the first of the 2
consecutive PTEs, so that, later on, there will be no need to
allocate pages, which might fail, if poking from atomic context.
3. at this point, the page tables are populated, for the address that
was obtained at point 1, and this is ok, because the address is fixed
write_rare
----------
4. it can happen on any available core / thread at any time, therefore
each of them needs a different address
5. CPUs support hotplug, but from what I have read, I might get away
with having up to nr_cpus different addresses (determined at init)
and I would follow the same technique used by Nadav, of forcing the
mapping of 1 or 2 (1 could be enough, I have to loop anyway, at
some point) pages at each address, to ensure the population of the
page tables
so far, so good, but ...
6. the addresses used by each CPU are fixed
7. I do not understand the reference you make to
"allocation and freeing of rare-writable ranges", because if I
treat the range as such, then there is a risk that I need to
populate more entries in the page table, which would have problems
with the atomic context, unless write_rare from atomic is ruled
out. If write_rare from atomic can be avoided, then I can also have
one-use randomized addresses at each write-rare operation, instead
of fixed ones, like in point 6.
and, apologies for being dense: the following is still not clear to me:
8. you wrote:
> You’ll still need a mutex (mmap_sem) to synchronize allocation
> and freeing of rare-writable ranges, but that shouldn’t have much
> contention.
What causes the contention? It's the fact that the various cores
are using the same mm, if I understood correctly.
However, if there was one mm for each core, wouldn't that make it
unnecessary to have any mutex?
I feel there must be some obvious reason why multiple mms are not a
good idea, yet I cannot grasp it :-(
switch_mm_irqs_off() seems to have lots of references to
"this_cpu_something"; if there is any optimization from having the
same next across multiple cores, I'm missing it
[...]
I would either refactor it or create a new function to handle the write. The main thing that Nadav is adding that I think you’ll want to use is the infrastructure for temporarily switching mms from a non-kernel-thread context.
yes
[...]
You’ll still want Nadav’s code for setting up the mm in the first place, though.
yes
--
thanks, igor
