On Fri, Feb 21, 2020 at 12:18 PM Vivek Goyal <vgoyal@xxxxxxxxxx> wrote: > > On Fri, Feb 21, 2020 at 01:32:48PM -0500, Jeff Moyer wrote: > > Vivek Goyal <vgoyal@xxxxxxxxxx> writes: > > > > > On Thu, Feb 20, 2020 at 04:35:17PM -0500, Jeff Moyer wrote: > > >> Vivek Goyal <vgoyal@xxxxxxxxxx> writes: > > >> > > >> > Currently pmem_clear_poison() expects offset and len to be sector aligned. > > >> > Atleast that seems to be the assumption with which code has been written. > > >> > It is called only from pmem_do_bvec() which is called only from pmem_rw_page() > > >> > and pmem_make_request() which will only passe sector aligned offset and len. > > >> > > > >> > Soon we want use this function from dax_zero_page_range() code path which > > >> > can try to zero arbitrary range of memory with-in a page. So update this > > >> > function to assume that offset and length can be arbitrary and do the > > >> > necessary alignments as needed. > > >> > > >> What caller will try to zero a range that is smaller than a sector? > > > > > > Hi Jeff, > > > > > > New dax zeroing interface (dax_zero_page_range()) can technically pass > > > a range which is less than a sector. Or which is bigger than a sector > > > but start and end are not aligned on sector boundaries. > > > > Sure, but who will call it with misaligned ranges? > > create a file foo.txt of size 4K and then truncate it. > > "truncate -s 23 foo.txt". Filesystems try to zero the bytes from 24 to > 4095. > > I have also written a test for this. > > https://github.com/rhvgoyal/misc/blob/master/pmem-tests/iomap-range-test.sh#L102 > > > > > > At this point of time, all I care about is that case of an arbitrary > > > range is handeled well. So if a caller passes a range in, we figure > > > out subrange which is sector aligned in terms of start and end, and > > > clear poison on those sectors and ignore rest of the range. And > > > this itself will be an improvement over current behavior where > > > nothing is cleared if I/O is not sector aligned. > > > > I don't think this makes sense. The caller needs to know about the > > blast radius of errors. This is why I asked for a concrete example. > > It might make more sense, for example, to return an error if not all of > > the errors could be cleared. > > > > >> > nvdimm_clear_poison() seems to assume offset and len to be aligned to > > >> > clear_err_unit boundary. But this is currently internal detail and is > > >> > not exported for others to use. So for now, continue to align offset and > > >> > length to SECTOR_SIZE boundary. Improving it further and to align it > > >> > to clear_err_unit boundary is a TODO item for future. > > >> > > >> When there is a poisoned range of persistent memory, it is recorded by > > >> the badblocks infrastructure, which currently operates on sectors. So, > > >> no matter what the error unit is for the hardware, we currently can't > > >> record/report to userspace anything smaller than a sector, and so that > > >> is what we expect when clearing errors. > > >> > > >> Continuing on for completeness, we will currently not map a page with > > >> badblocks into a process' address space. So, let's say you have 256 > > >> bytes of bad pmem, we will tell you we've lost 512 bytes, and even if > > >> you access a valid mmap()d address in the same page as the poisoned > > >> memory, you will get a segfault. > > >> > > >> Userspace can fix up the error by calling write(2) and friends to > > >> provide new data, or by punching a hole and writing new data to the hole > > >> (which may result in getting a new block, or reallocating the old block > > >> and zeroing it, which will clear the error). > > > > > > Fair enough. I do not need poison clearing at finer granularity. It might > > > be needed once dev_dax path wants to clear poison. Not sure how exactly > > > that works. > > > > It doesn't. :) > > > > >> > + /* > > >> > + * Callers can pass arbitrary offset and len. But nvdimm_clear_poison() > > >> > + * expects memory offset and length to meet certain alignment > > >> > + * restrction (clear_err_unit). Currently nvdimm does not export > > >> ^^^^^^^^^^^^^^^^^^^^^^ > > >> > + * required alignment. So align offset and length to sector boundary > > >> > > >> What is "nvdimm" in that sentence? Because the nvdimm most certainly > > >> does export the required alignment. Perhaps you meant libnvdimm? > > > > > > I meant nvdimm_clear_poison() function in drivers/nvdimm/bus.c. Whatever > > > it is called. It first queries alignement required (clear_err_unit) and > > > then makes sure range passed in meets that alignment requirement. > > > > My point was your comment is misleading. > > > > >> We could potentially support clearing less than a sector, but I'd have > > >> to understand the use cases better before offerring implementation > > >> suggestions. > > > > > > I don't need clearing less than a secotr. Once somebody needs it they > > > can implement it. All I am doing is making sure current logic is not > > > broken when dax_zero_page_range() starts using this logic and passes > > > an arbitrary range. We need to make sure we internally align I/O > > > > An arbitrary range is the same thing as less than a sector. :) Do you > > know of an instance where the range will not be sector-aligned and sized? > > > > > and carve out an aligned sub-range and pass that subrange to > > > nvdimm_clear_poison(). > > > > And what happens to the rest? The caller is left to trip over the > > errors? That sounds pretty terrible. I really think there needs to be > > an explicit contract here. > > Ok, I think is is the contentious bit. Current interface > (__dax_zero_page_range()) either clears the poison (if I/O is aligned to > sector) or expects page to be free of poison. > > So in above example, of "truncate -s 23 foo.txt", currently I get an error > because range being zeroed is not sector aligned. So > __dax_zero_page_range() falls back to calling direct_access(). Which > fails because there are poisoned sectors in the page. > > With my patches, dax_zero_page_range(), clears the poison from sector 1 to > 7 but leaves sector 0 untouched and just writes zeroes from byte 0 to 511 > and returns success. > > So question is, is this better behavior or worse behavior. If sector 0 > was poisoned, it will continue to remain poisoned and caller will come > to know about it on next read and then it should try to truncate file > to length 0 or unlink file or restore that file to get rid of poison. > > IOW, if a partial block is being zeroed and if it is poisoned, caller > will not be return an error and poison will not be cleared and memory > will be zeroed. What do we expect in such cases. > > Do we expect an interface where if there are any bad blocks in the range > being zeroed, then they all should be cleared (and hence all I/O should > be aligned) otherwise error is returned. If yes, I could make that > change. This does not strike me as a good idea because it's a false security compared to the latent poison case. If the writes to an unknown poisoned location would otherwise succeed via a different I/O path (dax), it's an unsymmetric surprise to start returning errors just because you wrote zeroes as a side effect of truncate. > Downside of current interface is that it will clear as many blocks as > possible in the given range and leave starting and end blocks poisoned > (if it is unaligned) and not return error. That means a reader will > get error on these blocks again and they will have to try to clear it > again. I think what you have described in your truncate example is an improvement on what we have currently because x86 does not communicate write errors. Specifically, writing zeros via dax from userspace over unknown poison behaves the same as writing unaligned zeros over known poison. In both cases it's a best effort that always succeeds (no cpu exception), and may inadvertently clear poison as a side-effect. Otherwise, an error-block-aligned hole punch is the only way to trigger the kernel to try to clear known poison when the full block is reallocated. On movdir64b capable cpus the error clearing unit becomes 64-bytes rather than 256-bytes because that allows a cacheline to be written without triggering a line fill read. So the error clearing granularity gets better over time, but unfortunately not synchronous detection in the I/O path. I think a better way to improve poison handling is the long standing idea to integrate the badblock tracking into the filesystem directly. That way driver notifications of poison can be ingested into the filesystem and notifications sent on filenames rather than the current TOCTOU mess of trying to do a reverse lookup of badblock numbers to files. If the application can efficiently list and be notified of poison it can mitigate it immediately rather than trying to rely on write side effects.
