-
Notifications
You must be signed in to change notification settings - Fork 1.7k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
vdev_disk: rewrite BIO filling machinery to avoid split pages #15588
Conversation
fbd8f55
to
12c58b4
Compare
Last push reworks a few things in response to review comments, and other things discovered while trying to answer them:
|
module/os/linux/zfs/vdev_disk.c
Outdated
zio_delay_interrupt(zio); | ||
|
||
/* Finish cleanup */ | ||
vbio_free(vbio); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Would we call it before zio_delay_interrupt(), we would not need explicit vbio_return_abd() call.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yeah, my thought was to get the zio into the pipeline at the earliest possible opportunity, but now I say it loud its a bit silly. I'll fix that up.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Tested doing the vbio cleanup before submitting the zio, and there's a small but measurable throughput drop on a couple of test workloads - about 1-2%. They're still ~2.6x over master so I won't be sad about losing them, but still.
I've rewritten a comment to try to make it clearer why its this way, but if you'd rather the simpler version, I'd be ok with that. Its robn@270b44c.
a58a95c
to
e4156eb
Compare
e4156eb
to
f27dc07
Compare
a2a44ad
to
b5e6138
Compare
I believe this is good to go, pending outcome from test results. @behlendorf If/when this lands, I'd like to also take it to 2.2, but I'm wary of introducing a change so fundamental on a stable series. I'm thinking about a module load parameter that toggles which implement is used, default to the old one. And then we could recommend folk struggling with LUKS (#15533), or seeing rejected SCSI IO (original case), or just wanting to try it out and maybe test the performance change, without risk. Assuming the implementation is sane, would you be interested in that? (Incidentally, I am dogfooding this on my daily driver now (on top of 2.2.2). Its a pretty boring laptop workload though, but its not nothing). |
I was thinking along the same lines. Assuming we could do this is a reasonable way I think it'd even make sense to keep both versions in the master branch for a while. Enabled by default in master branch and disabled by default in 2.2 until we have enough testing with it. |
b5e6138
to
c22ae5b
Compare
Alright, see what you think of that. Changes to previous:
I've done as much testing as I can on both modes. I managed to induce a split (and SCSI rejection with "classic"), none on "new", as expected. Performance on my sanity check workload remains 2.7x on the new. |
@robn - how safe is this to test on systems where i'd need to restore data if i blow things up? 😄 |
Top commit adds @behlendorf I'll be interested to know if that changes anything in your test case. |
@sempervictus I think it just works or doesn't - I can't think of any reason why a bug would cause any on-disk corruption. So worst case, you should be able to switch it over to |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I hope we drop zfs_vdev_disk_classic and the duplication pretty soon. I would not call old code "very stable", considering we know it has issues with #15452.
ac214be
to
d298fc4
Compare
Maybe "very stable" is too far, but at least its a known quantity that has served well for a long time. I would hope that the new code can get enough testing on 2.2 that we could remove classic for 2.3. |
ecd9b3b
to
f058215
Compare
This is just setting up for the next couple of commits, which will add a new IO function and a parameter to select it. Reviewed-by: Alexander Motin <[email protected]> Reviewed-by: Brian Behlendorf <[email protected]> Signed-off-by: Rob Norris <[email protected]> Sponsored-by: Klara, Inc. Sponsored-by: Wasabi Technology, Inc. Closes openzfs#15533 Closes openzfs#15588 (cherry picked from commit c4a13ba)
This commit tackles a number of issues in the way BIOs (`struct bio`) are constructed for submission to the Linux block layer. The kernel has a hard upper limit on the number of pages/segments that can be added to a BIO, as well as a separate limit for each device (related to its queue depth and other scheduling characteristics). ZFS counts the number of memory pages in the request ABD (`abd_nr_pages_off()`, and then uses that as the number of segments to put into the BIO, up to the hard upper limit. If it requires more than the limit, it will create multiple BIOs. Leaving aside the fact that page count method is wrong (see below), not limiting to the device segment max means that the device driver will need to split the BIO in half. This is alone is not necessarily a problem, but it interacts with another issue to cause a much larger problem. The kernel function to add a segment to a BIO (`bio_add_page()`) takes a `struct page` pointer, and offset+len within it. `struct page` can represent a run of contiguous memory pages (known as a "compound page"). In can be of arbitrary length. The ZFS functions that count ABD pages and load them into the BIO (`abd_nr_pages_off()`, `bio_map()` and `abd_bio_map_off()`) will never consider a page to be more than `PAGE_SIZE` (4K), even if the `struct page` is for multiple pages. In this case, it will load the same `struct page` into the BIO multiple times, with the offset adjusted each time. With a sufficiently large ABD, this can easily lead to the BIO being entirely filled much earlier than it could have been. This is also further contributes to the problem caused by the incorrect segment limit calculation, as its much easier to go past the device limit, and so require a split. Again, this is not a problem on its own. The logic for "never submit more than `PAGE_SIZE`" is actually a little more subtle. It will actually never submit a buffer that crosses a 4K page boundary. In practice, this is fine, as most ABDs are scattered, that is a list of complete 4K pages, and so are loaded in as such. Linear ABDs are typically allocated from slabs, and for small sizes they are frequently not aligned to page boundaries. For example, a 12K allocation can span four pages, eg: -- 4K -- -- 4K -- -- 4K -- -- 4K -- | | | | | :## ######## ######## ######: [1K, 4K, 4K, 3K] Such an allocation would be loaded into a BIO as you see: [1K, 4K, 4K, 3K] This tends not to be a problem in practice, because even if the BIO were filled and needed to be split, each half would still have either a start or end aligned to the logical block size of the device (assuming 4K at least). --- In ideal circumstances, these shortcomings don't cause any particular problems. Its when they start to interact with other ZFS features that things get interesting. Aggregation will create a "gang" ABD, which is simply a list of other ABDs. Iterating over a gang ABD is just iterating over each ABD within it in turn. Because the segments are simply loaded in order, we can end up with uneven segments either side of the "gap" between the two ABDs. For example, two 12K ABDs might be aggregated and then loaded as: [1K, 4K, 4K, 3K, 2K, 4K, 4K, 2K] Should a split occur, each individual BIO can end up either having an start or end offset that is not aligned to the logical block size, which some drivers (eg SCSI) will reject. However, this tends not to happen because the default aggregation limit usually keeps the BIO small enough to not require more than one split, and most pages are actually full 4K pages, so hitting an uneven gap is very rare anyway. If the pool is under particular memory pressure, then an IO can be broken down into a "gang block", a 512-byte block composed of a header and up to three block pointers. Each points to a fragment of the original write, or in turn, another gang block, breaking the original data up over and over until space can be found in the pool for each of them. Each gang header is a separate 512-byte memory allocation from a slab, that needs to be written down to disk. When the gang header is added to the BIO, its a single 512-byte segment. Pulling all this together, consider a large aggregated write of gang blocks. This results a BIO containing lots of 512-byte segments. Given our tendency to overfill the BIO, a split is likely, and most possible split points will yield a pair of BIOs that are misaligned. Drivers that care, like the SCSI driver, will reject them. --- This commit is a substantial refactor and rewrite of much of `vdev_disk` to sort all this out. `vdev_bio_max_segs()` now returns the ideal maximum size for the device, if available. There's also a tuneable `zfs_vdev_disk_max_segs` to override this, to assist with testing. We scan the ABD up front to count the number of pages within it, and to confirm that if we submitted all those pages to one or more BIOs, it could be split at any point with creating a misaligned BIO. If the pages in the BIO are not usable (as in any of the above situations), the ABD is linearised, and then checked again. This is the same technique used in `vdev_geom` on FreeBSD, adjusted for Linux's variable page size and allocator quirks. `vbio_t` is a cleanup and enhancement of the old `dio_request_t`. The idea is simply that it can hold all the state needed to create, submit and return multiple BIOs, including all the refcounts, the ABD copy if it was needed, and so on. Apart from what I hope is a clearer interface, the major difference is that because we know how many BIOs we'll need up front, we don't need the old overflow logic that would grow the BIO array, throw away all the old work and restart. We can get it right from the start. Reviewed-by: Alexander Motin <[email protected]> Reviewed-by: Brian Behlendorf <[email protected]> Signed-off-by: Rob Norris <[email protected]> Sponsored-by: Klara, Inc. Sponsored-by: Wasabi Technology, Inc. Closes openzfs#15533 Closes openzfs#15588 (cherry picked from commit 06a1960)
This makes the submission method selectable at module load time via the `zfs_vdev_disk_classic` parameter, allowing this change to be backported to 2.2 safely, and disabled in favour of the "classic" submission method if new problems come up. Reviewed-by: Alexander Motin <[email protected]> Reviewed-by: Brian Behlendorf <[email protected]> Signed-off-by: Rob Norris <[email protected]> Sponsored-by: Klara, Inc. Sponsored-by: Wasabi Technology, Inc. Closes openzfs#15533 Closes openzfs#15588 (cherry picked from commit df2169d)
Simplifies our code a lot, so we don't have to wait for each and reassemble them. Reviewed-by: Alexander Motin <[email protected]> Reviewed-by: Brian Behlendorf <[email protected]> Signed-off-by: Rob Norris <[email protected]> Sponsored-by: Klara, Inc. Sponsored-by: Wasabi Technology, Inc. Closes openzfs#15533 Closes openzfs#15588 (cherry picked from commit 72fd834)
Before 4.5 (specifically, torvalds/linux@ddc58f2), head and tail pages in a compound page were refcounted separately. This means that using the head page without taking a reference to it could see it cleaned up later before we're finished with it. Specifically, bio_add_page() would take a reference, and drop its reference after the bio completion callback returns. If the zio is executed immediately from the completion callback, this is usually ok, as any data is referenced through the tail page referenced by the ABD, and so becomes "live" that way. If there's a delay in zio execution (high load, error injection), then the head page can be freed, along with any dirty flags or other indicators that the underlying memory is used. Later, when the zio completes and that memory is accessed, its either unmapped and an unhandled fault takes down the entire system, or it is mapped and we end up messing around in someone else's memory. Both of these are very bad. The solution on these older kernels is to take a reference to the head page when we use it, and release it when we're done. There's not really a sensible way under our current structure to do this; the "best" would be to keep a list of head page references in the ABD, and release them when the ABD is freed. Since this additional overhead is totally unnecessary on 4.5+, where head and tail pages share refcounts, I've opted to simply not use the compound head in ABD page iteration there. This is theoretically less efficient (though cleaning up head page references would add overhead), but its safe, and we still get the other benefits of not mapping pages before adding them to a bio and not mis-splitting pages. There doesn't appear to be an obvious symbol name or config option we can match on to discover this behaviour in configure (and the mm/page APIs have changed a lot since then anyway), so I've gone with a simple version check. Reviewed-by: Alexander Motin <[email protected]> Reviewed-by: Brian Behlendorf <[email protected]> Signed-off-by: Rob Norris <[email protected]> Sponsored-by: Klara, Inc. Sponsored-by: Wasabi Technology, Inc. Closes openzfs#15533 Closes openzfs#15588 (cherry picked from commit c6be6ce)
Before 5.4 we have to do a little math. Reviewed-by: Alexander Motin <[email protected]> Reviewed-by: Brian Behlendorf <[email protected]> Signed-off-by: Rob Norris <[email protected]> Sponsored-by: Klara, Inc. Sponsored-by: Wasabi Technology, Inc. Closes #15533 Closes #15588 (cherry picked from commit df04efe)
The regular ABD iterators yield data buffers, so they have to map and unmap pages into kernel memory. If the caller only wants to count chunks, or can use page pointers directly, then the map/unmap is just unnecessary overhead. This adds adb_iterate_page_func, which yields unmapped struct page instead. Reviewed-by: Alexander Motin <[email protected]> Reviewed-by: Brian Behlendorf <[email protected]> Signed-off-by: Rob Norris <[email protected]> Sponsored-by: Klara, Inc. Sponsored-by: Wasabi Technology, Inc. Closes #15533 Closes #15588 (cherry picked from commit 390b448)
This is just renaming the existing functions we're about to replace and grouping them together to make the next commits easier to follow. Reviewed-by: Alexander Motin <[email protected]> Reviewed-by: Brian Behlendorf <[email protected]> Signed-off-by: Rob Norris <[email protected]> Sponsored-by: Klara, Inc. Sponsored-by: Wasabi Technology, Inc. Closes #15533 Closes #15588 (cherry picked from commit f3b85d7)
Light reshuffle to make it a bit more linear to read and get rid of a bunch of args that aren't needed in all cases. Reviewed-by: Alexander Motin <[email protected]> Reviewed-by: Brian Behlendorf <[email protected]> Signed-off-by: Rob Norris <[email protected]> Sponsored-by: Klara, Inc. Sponsored-by: Wasabi Technology, Inc. Closes #15533 Closes #15588 (cherry picked from commit 867178a)
This is just setting up for the next couple of commits, which will add a new IO function and a parameter to select it. Reviewed-by: Alexander Motin <[email protected]> Reviewed-by: Brian Behlendorf <[email protected]> Signed-off-by: Rob Norris <[email protected]> Sponsored-by: Klara, Inc. Sponsored-by: Wasabi Technology, Inc. Closes #15533 Closes #15588 (cherry picked from commit c4a13ba)
This commit tackles a number of issues in the way BIOs (`struct bio`) are constructed for submission to the Linux block layer. The kernel has a hard upper limit on the number of pages/segments that can be added to a BIO, as well as a separate limit for each device (related to its queue depth and other scheduling characteristics). ZFS counts the number of memory pages in the request ABD (`abd_nr_pages_off()`, and then uses that as the number of segments to put into the BIO, up to the hard upper limit. If it requires more than the limit, it will create multiple BIOs. Leaving aside the fact that page count method is wrong (see below), not limiting to the device segment max means that the device driver will need to split the BIO in half. This is alone is not necessarily a problem, but it interacts with another issue to cause a much larger problem. The kernel function to add a segment to a BIO (`bio_add_page()`) takes a `struct page` pointer, and offset+len within it. `struct page` can represent a run of contiguous memory pages (known as a "compound page"). In can be of arbitrary length. The ZFS functions that count ABD pages and load them into the BIO (`abd_nr_pages_off()`, `bio_map()` and `abd_bio_map_off()`) will never consider a page to be more than `PAGE_SIZE` (4K), even if the `struct page` is for multiple pages. In this case, it will load the same `struct page` into the BIO multiple times, with the offset adjusted each time. With a sufficiently large ABD, this can easily lead to the BIO being entirely filled much earlier than it could have been. This is also further contributes to the problem caused by the incorrect segment limit calculation, as its much easier to go past the device limit, and so require a split. Again, this is not a problem on its own. The logic for "never submit more than `PAGE_SIZE`" is actually a little more subtle. It will actually never submit a buffer that crosses a 4K page boundary. In practice, this is fine, as most ABDs are scattered, that is a list of complete 4K pages, and so are loaded in as such. Linear ABDs are typically allocated from slabs, and for small sizes they are frequently not aligned to page boundaries. For example, a 12K allocation can span four pages, eg: -- 4K -- -- 4K -- -- 4K -- -- 4K -- | | | | | :## ######## ######## ######: [1K, 4K, 4K, 3K] Such an allocation would be loaded into a BIO as you see: [1K, 4K, 4K, 3K] This tends not to be a problem in practice, because even if the BIO were filled and needed to be split, each half would still have either a start or end aligned to the logical block size of the device (assuming 4K at least). --- In ideal circumstances, these shortcomings don't cause any particular problems. Its when they start to interact with other ZFS features that things get interesting. Aggregation will create a "gang" ABD, which is simply a list of other ABDs. Iterating over a gang ABD is just iterating over each ABD within it in turn. Because the segments are simply loaded in order, we can end up with uneven segments either side of the "gap" between the two ABDs. For example, two 12K ABDs might be aggregated and then loaded as: [1K, 4K, 4K, 3K, 2K, 4K, 4K, 2K] Should a split occur, each individual BIO can end up either having an start or end offset that is not aligned to the logical block size, which some drivers (eg SCSI) will reject. However, this tends not to happen because the default aggregation limit usually keeps the BIO small enough to not require more than one split, and most pages are actually full 4K pages, so hitting an uneven gap is very rare anyway. If the pool is under particular memory pressure, then an IO can be broken down into a "gang block", a 512-byte block composed of a header and up to three block pointers. Each points to a fragment of the original write, or in turn, another gang block, breaking the original data up over and over until space can be found in the pool for each of them. Each gang header is a separate 512-byte memory allocation from a slab, that needs to be written down to disk. When the gang header is added to the BIO, its a single 512-byte segment. Pulling all this together, consider a large aggregated write of gang blocks. This results a BIO containing lots of 512-byte segments. Given our tendency to overfill the BIO, a split is likely, and most possible split points will yield a pair of BIOs that are misaligned. Drivers that care, like the SCSI driver, will reject them. --- This commit is a substantial refactor and rewrite of much of `vdev_disk` to sort all this out. `vdev_bio_max_segs()` now returns the ideal maximum size for the device, if available. There's also a tuneable `zfs_vdev_disk_max_segs` to override this, to assist with testing. We scan the ABD up front to count the number of pages within it, and to confirm that if we submitted all those pages to one or more BIOs, it could be split at any point with creating a misaligned BIO. If the pages in the BIO are not usable (as in any of the above situations), the ABD is linearised, and then checked again. This is the same technique used in `vdev_geom` on FreeBSD, adjusted for Linux's variable page size and allocator quirks. `vbio_t` is a cleanup and enhancement of the old `dio_request_t`. The idea is simply that it can hold all the state needed to create, submit and return multiple BIOs, including all the refcounts, the ABD copy if it was needed, and so on. Apart from what I hope is a clearer interface, the major difference is that because we know how many BIOs we'll need up front, we don't need the old overflow logic that would grow the BIO array, throw away all the old work and restart. We can get it right from the start. Reviewed-by: Alexander Motin <[email protected]> Reviewed-by: Brian Behlendorf <[email protected]> Signed-off-by: Rob Norris <[email protected]> Sponsored-by: Klara, Inc. Sponsored-by: Wasabi Technology, Inc. Closes #15533 Closes #15588 (cherry picked from commit 06a1960)
This makes the submission method selectable at module load time via the `zfs_vdev_disk_classic` parameter, allowing this change to be backported to 2.2 safely, and disabled in favour of the "classic" submission method if new problems come up. Reviewed-by: Alexander Motin <[email protected]> Reviewed-by: Brian Behlendorf <[email protected]> Signed-off-by: Rob Norris <[email protected]> Sponsored-by: Klara, Inc. Sponsored-by: Wasabi Technology, Inc. Closes #15533 Closes #15588 (cherry picked from commit df2169d)
Simplifies our code a lot, so we don't have to wait for each and reassemble them. Reviewed-by: Alexander Motin <[email protected]> Reviewed-by: Brian Behlendorf <[email protected]> Signed-off-by: Rob Norris <[email protected]> Sponsored-by: Klara, Inc. Sponsored-by: Wasabi Technology, Inc. Closes #15533 Closes #15588 (cherry picked from commit 72fd834)
Before 4.5 (specifically, torvalds/linux@ddc58f2), head and tail pages in a compound page were refcounted separately. This means that using the head page without taking a reference to it could see it cleaned up later before we're finished with it. Specifically, bio_add_page() would take a reference, and drop its reference after the bio completion callback returns. If the zio is executed immediately from the completion callback, this is usually ok, as any data is referenced through the tail page referenced by the ABD, and so becomes "live" that way. If there's a delay in zio execution (high load, error injection), then the head page can be freed, along with any dirty flags or other indicators that the underlying memory is used. Later, when the zio completes and that memory is accessed, its either unmapped and an unhandled fault takes down the entire system, or it is mapped and we end up messing around in someone else's memory. Both of these are very bad. The solution on these older kernels is to take a reference to the head page when we use it, and release it when we're done. There's not really a sensible way under our current structure to do this; the "best" would be to keep a list of head page references in the ABD, and release them when the ABD is freed. Since this additional overhead is totally unnecessary on 4.5+, where head and tail pages share refcounts, I've opted to simply not use the compound head in ABD page iteration there. This is theoretically less efficient (though cleaning up head page references would add overhead), but its safe, and we still get the other benefits of not mapping pages before adding them to a bio and not mis-splitting pages. There doesn't appear to be an obvious symbol name or config option we can match on to discover this behaviour in configure (and the mm/page APIs have changed a lot since then anyway), so I've gone with a simple version check. Reviewed-by: Alexander Motin <[email protected]> Reviewed-by: Brian Behlendorf <[email protected]> Signed-off-by: Rob Norris <[email protected]> Sponsored-by: Klara, Inc. Sponsored-by: Wasabi Technology, Inc. Closes #15533 Closes #15588 (cherry picked from commit c6be6ce)
Simplifies our code a lot, so we don't have to wait for each and reassemble them. Reviewed-by: Alexander Motin <[email protected]> Reviewed-by: Brian Behlendorf <[email protected]> Signed-off-by: Rob Norris <[email protected]> Sponsored-by: Klara, Inc. Sponsored-by: Wasabi Technology, Inc. Closes openzfs#15533 Closes openzfs#15588 (cherry picked from commit 72fd834)
Before 4.5 (specifically, torvalds/linux@ddc58f2), head and tail pages in a compound page were refcounted separately. This means that using the head page without taking a reference to it could see it cleaned up later before we're finished with it. Specifically, bio_add_page() would take a reference, and drop its reference after the bio completion callback returns. If the zio is executed immediately from the completion callback, this is usually ok, as any data is referenced through the tail page referenced by the ABD, and so becomes "live" that way. If there's a delay in zio execution (high load, error injection), then the head page can be freed, along with any dirty flags or other indicators that the underlying memory is used. Later, when the zio completes and that memory is accessed, its either unmapped and an unhandled fault takes down the entire system, or it is mapped and we end up messing around in someone else's memory. Both of these are very bad. The solution on these older kernels is to take a reference to the head page when we use it, and release it when we're done. There's not really a sensible way under our current structure to do this; the "best" would be to keep a list of head page references in the ABD, and release them when the ABD is freed. Since this additional overhead is totally unnecessary on 4.5+, where head and tail pages share refcounts, I've opted to simply not use the compound head in ABD page iteration there. This is theoretically less efficient (though cleaning up head page references would add overhead), but its safe, and we still get the other benefits of not mapping pages before adding them to a bio and not mis-splitting pages. There doesn't appear to be an obvious symbol name or config option we can match on to discover this behaviour in configure (and the mm/page APIs have changed a lot since then anyway), so I've gone with a simple version check. Reviewed-by: Alexander Motin <[email protected]> Reviewed-by: Brian Behlendorf <[email protected]> Signed-off-by: Rob Norris <[email protected]> Sponsored-by: Klara, Inc. Sponsored-by: Wasabi Technology, Inc. Closes openzfs#15533 Closes openzfs#15588 (cherry picked from commit c6be6ce)
Before 5.4 we have to do a little math. Reviewed-by: Alexander Motin <[email protected]> Reviewed-by: Brian Behlendorf <[email protected]> Signed-off-by: Rob Norris <[email protected]> Sponsored-by: Klara, Inc. Sponsored-by: Wasabi Technology, Inc. Closes openzfs#15533 Closes openzfs#15588
The regular ABD iterators yield data buffers, so they have to map and unmap pages into kernel memory. If the caller only wants to count chunks, or can use page pointers directly, then the map/unmap is just unnecessary overhead. This adds adb_iterate_page_func, which yields unmapped struct page instead. Reviewed-by: Alexander Motin <[email protected]> Reviewed-by: Brian Behlendorf <[email protected]> Signed-off-by: Rob Norris <[email protected]> Sponsored-by: Klara, Inc. Sponsored-by: Wasabi Technology, Inc. Closes openzfs#15533 Closes openzfs#15588
This is just renaming the existing functions we're about to replace and grouping them together to make the next commits easier to follow. Reviewed-by: Alexander Motin <[email protected]> Reviewed-by: Brian Behlendorf <[email protected]> Signed-off-by: Rob Norris <[email protected]> Sponsored-by: Klara, Inc. Sponsored-by: Wasabi Technology, Inc. Closes openzfs#15533 Closes openzfs#15588
Light reshuffle to make it a bit more linear to read and get rid of a bunch of args that aren't needed in all cases. Reviewed-by: Alexander Motin <[email protected]> Reviewed-by: Brian Behlendorf <[email protected]> Signed-off-by: Rob Norris <[email protected]> Sponsored-by: Klara, Inc. Sponsored-by: Wasabi Technology, Inc. Closes openzfs#15533 Closes openzfs#15588
This is just setting up for the next couple of commits, which will add a new IO function and a parameter to select it. Reviewed-by: Alexander Motin <[email protected]> Reviewed-by: Brian Behlendorf <[email protected]> Signed-off-by: Rob Norris <[email protected]> Sponsored-by: Klara, Inc. Sponsored-by: Wasabi Technology, Inc. Closes openzfs#15533 Closes openzfs#15588
This commit tackles a number of issues in the way BIOs (`struct bio`) are constructed for submission to the Linux block layer. The kernel has a hard upper limit on the number of pages/segments that can be added to a BIO, as well as a separate limit for each device (related to its queue depth and other scheduling characteristics). ZFS counts the number of memory pages in the request ABD (`abd_nr_pages_off()`, and then uses that as the number of segments to put into the BIO, up to the hard upper limit. If it requires more than the limit, it will create multiple BIOs. Leaving aside the fact that page count method is wrong (see below), not limiting to the device segment max means that the device driver will need to split the BIO in half. This is alone is not necessarily a problem, but it interacts with another issue to cause a much larger problem. The kernel function to add a segment to a BIO (`bio_add_page()`) takes a `struct page` pointer, and offset+len within it. `struct page` can represent a run of contiguous memory pages (known as a "compound page"). In can be of arbitrary length. The ZFS functions that count ABD pages and load them into the BIO (`abd_nr_pages_off()`, `bio_map()` and `abd_bio_map_off()`) will never consider a page to be more than `PAGE_SIZE` (4K), even if the `struct page` is for multiple pages. In this case, it will load the same `struct page` into the BIO multiple times, with the offset adjusted each time. With a sufficiently large ABD, this can easily lead to the BIO being entirely filled much earlier than it could have been. This is also further contributes to the problem caused by the incorrect segment limit calculation, as its much easier to go past the device limit, and so require a split. Again, this is not a problem on its own. The logic for "never submit more than `PAGE_SIZE`" is actually a little more subtle. It will actually never submit a buffer that crosses a 4K page boundary. In practice, this is fine, as most ABDs are scattered, that is a list of complete 4K pages, and so are loaded in as such. Linear ABDs are typically allocated from slabs, and for small sizes they are frequently not aligned to page boundaries. For example, a 12K allocation can span four pages, eg: -- 4K -- -- 4K -- -- 4K -- -- 4K -- | | | | | :## ######## ######## ######: [1K, 4K, 4K, 3K] Such an allocation would be loaded into a BIO as you see: [1K, 4K, 4K, 3K] This tends not to be a problem in practice, because even if the BIO were filled and needed to be split, each half would still have either a start or end aligned to the logical block size of the device (assuming 4K at least). --- In ideal circumstances, these shortcomings don't cause any particular problems. Its when they start to interact with other ZFS features that things get interesting. Aggregation will create a "gang" ABD, which is simply a list of other ABDs. Iterating over a gang ABD is just iterating over each ABD within it in turn. Because the segments are simply loaded in order, we can end up with uneven segments either side of the "gap" between the two ABDs. For example, two 12K ABDs might be aggregated and then loaded as: [1K, 4K, 4K, 3K, 2K, 4K, 4K, 2K] Should a split occur, each individual BIO can end up either having an start or end offset that is not aligned to the logical block size, which some drivers (eg SCSI) will reject. However, this tends not to happen because the default aggregation limit usually keeps the BIO small enough to not require more than one split, and most pages are actually full 4K pages, so hitting an uneven gap is very rare anyway. If the pool is under particular memory pressure, then an IO can be broken down into a "gang block", a 512-byte block composed of a header and up to three block pointers. Each points to a fragment of the original write, or in turn, another gang block, breaking the original data up over and over until space can be found in the pool for each of them. Each gang header is a separate 512-byte memory allocation from a slab, that needs to be written down to disk. When the gang header is added to the BIO, its a single 512-byte segment. Pulling all this together, consider a large aggregated write of gang blocks. This results a BIO containing lots of 512-byte segments. Given our tendency to overfill the BIO, a split is likely, and most possible split points will yield a pair of BIOs that are misaligned. Drivers that care, like the SCSI driver, will reject them. --- This commit is a substantial refactor and rewrite of much of `vdev_disk` to sort all this out. `vdev_bio_max_segs()` now returns the ideal maximum size for the device, if available. There's also a tuneable `zfs_vdev_disk_max_segs` to override this, to assist with testing. We scan the ABD up front to count the number of pages within it, and to confirm that if we submitted all those pages to one or more BIOs, it could be split at any point with creating a misaligned BIO. If the pages in the BIO are not usable (as in any of the above situations), the ABD is linearised, and then checked again. This is the same technique used in `vdev_geom` on FreeBSD, adjusted for Linux's variable page size and allocator quirks. `vbio_t` is a cleanup and enhancement of the old `dio_request_t`. The idea is simply that it can hold all the state needed to create, submit and return multiple BIOs, including all the refcounts, the ABD copy if it was needed, and so on. Apart from what I hope is a clearer interface, the major difference is that because we know how many BIOs we'll need up front, we don't need the old overflow logic that would grow the BIO array, throw away all the old work and restart. We can get it right from the start. Reviewed-by: Alexander Motin <[email protected]> Reviewed-by: Brian Behlendorf <[email protected]> Signed-off-by: Rob Norris <[email protected]> Sponsored-by: Klara, Inc. Sponsored-by: Wasabi Technology, Inc. Closes openzfs#15533 Closes openzfs#15588
This makes the submission method selectable at module load time via the `zfs_vdev_disk_classic` parameter, allowing this change to be backported to 2.2 safely, and disabled in favour of the "classic" submission method if new problems come up. Reviewed-by: Alexander Motin <[email protected]> Reviewed-by: Brian Behlendorf <[email protected]> Signed-off-by: Rob Norris <[email protected]> Sponsored-by: Klara, Inc. Sponsored-by: Wasabi Technology, Inc. Closes openzfs#15533 Closes openzfs#15588
Simplifies our code a lot, so we don't have to wait for each and reassemble them. Reviewed-by: Alexander Motin <[email protected]> Reviewed-by: Brian Behlendorf <[email protected]> Signed-off-by: Rob Norris <[email protected]> Sponsored-by: Klara, Inc. Sponsored-by: Wasabi Technology, Inc. Closes openzfs#15533 Closes openzfs#15588
Before 4.5 (specifically, torvalds/linux@ddc58f2), head and tail pages in a compound page were refcounted separately. This means that using the head page without taking a reference to it could see it cleaned up later before we're finished with it. Specifically, bio_add_page() would take a reference, and drop its reference after the bio completion callback returns. If the zio is executed immediately from the completion callback, this is usually ok, as any data is referenced through the tail page referenced by the ABD, and so becomes "live" that way. If there's a delay in zio execution (high load, error injection), then the head page can be freed, along with any dirty flags or other indicators that the underlying memory is used. Later, when the zio completes and that memory is accessed, its either unmapped and an unhandled fault takes down the entire system, or it is mapped and we end up messing around in someone else's memory. Both of these are very bad. The solution on these older kernels is to take a reference to the head page when we use it, and release it when we're done. There's not really a sensible way under our current structure to do this; the "best" would be to keep a list of head page references in the ABD, and release them when the ABD is freed. Since this additional overhead is totally unnecessary on 4.5+, where head and tail pages share refcounts, I've opted to simply not use the compound head in ABD page iteration there. This is theoretically less efficient (though cleaning up head page references would add overhead), but its safe, and we still get the other benefits of not mapping pages before adding them to a bio and not mis-splitting pages. There doesn't appear to be an obvious symbol name or config option we can match on to discover this behaviour in configure (and the mm/page APIs have changed a lot since then anyway), so I've gone with a simple version check. Reviewed-by: Alexander Motin <[email protected]> Reviewed-by: Brian Behlendorf <[email protected]> Signed-off-by: Rob Norris <[email protected]> Sponsored-by: Klara, Inc. Sponsored-by: Wasabi Technology, Inc. Closes openzfs#15533 Closes openzfs#15588
Motivation and Context
This PR tackles a number of issues in the way BIOs (
struct bio
) are constructed for submission to the Linux block layer.BIO segment limits are set incorrectly
The kernel has a hard upper limit on the number of pages/segments that can be added to a BIO, as well as a separate limit for each device (related to its queue depth and other scheduling characteristics).
ZFS counts the number of memory pages in the request ABD (
abd_nr_pages_off()
, and then uses that as the number of segments to put into the BIO, up to the hard upper limit. If it requires more than the limit, it will create multiple BIOs.Leaving aside the fact that page count method is wrong (see below), not limiting to the device segment max means that the device driver will need to split the BIO in half. This is alone is not necessarily a problem, but it interacts with another issue to cause a much larger problem.
BIOs are filled inefficiently
The kernel function to add a segment to a BIO (
bio_add_page()
) takes astruct page
pointer, and offset+len within it.struct page
represents a run of contiguous memory pages (known as a "compound page"). In can be of arbitrary length.The ZFS functions that count ABD pages and load them into the BIO (
abd_nr_pages_off()
,bio_map()
andabd_bio_map_off()
) will never consider a page to be more thanPAGE_SIZE
(4K), even if thestruct page
is for multiple pages. In this case, it will load the samestruct page
into the BIO multiple times, with the offset adjusted each time.With a sufficiently large ABD, this can easily lead to the BIO being entirely filled much earlier than it could have been. This is also further contributes to the problem caused by the incorrect segment limit calculation, as its much easier to go past the device limit, and so require a split.
Again, this is not a problem on its own.
Incomplete pages are submitted to BIOs
The logic for "never submit more than
PAGE_SIZE
" is actually a little more subtle. It will actually never submit a buffer that crosses a 4K page boundary.In practice, this is fine, as most ABDs are scattered, that is a list of complete 4K pages, and so are loaded in as such.
Linear ABDs are typically allocated from slabs, and for small sizes they are frequently not aligned to page boundaries. For example, a 12K allocation can span four pages, eg:
Such an allocation would be loaded into a BIO as you see:
This tends not to be a problem in practice, because even if the BIO were filled and needed to be split, each half would still have either a start or end aligned to the logical block size of the device (assuming 4K at least).
In ideal circumstances, these shortcomings don't cause any particular problems. Its when they start to interact with other ZFS features that things get interesting.
Aggregation
Aggregation will create a "gang" ABD, which is simply a list of other ABDs. Iterating over a gang ABD is just iterating over each ABD within it in turn.
Because the segments are simply loaded in order, we can end up with uneven segments either side of the "gap" between the two ABDs. For example, two 12K ABDs might be aggregated and then loaded as:
Should a split occur, each individual BIO can end up either having an start or end offset that is not aligned to the logical block size, which some drivers (eg SCSI) will reject. However, this tends not to happen because the default aggregation limit usually keeps the BIO small enough to not require more than one split, and most pages are actually full 4K pages, so hitting an uneven gap is very rare anyway.
Gang blocks
If the pool is under particular memory pressure, then an IO can be broken down into a "gang block", a 512-byte block composed of a header and up to three block pointers. Each points to a fragment of the original write, or in turn, another gang block, breaking the original data up over and over until space can be found in the pool for each of them.
Each gang header is a separate 512-byte memory allocation from a slab, that needs to be written down to disk. When the gang header is added to the BIO, its a single 512-byte segment.
Aggregation with gang blocks
Pulling all this together, consider a large aggregated write of gang blocks. This results a BIO containing lots of 512-byte segments. Given our tendency to overfill the BIO, a split is likely, and most possible split points will yield a pair of BIOs that are misaligned. Drivers that care, like the SCSI driver, will reject them.
Description
This commit is a substantial refactor and rewrite of much of
vdev_disk
to sort all this out.Configure maximum segment size for device
vdev_bio_max_segs()
now returns the ideal maximum size for the device, if available. There's also a tuneablevdev_disk_max_segs
to override this, to assist with testing.ABDs checked up front for page count and alignment
We scan the ABD up front to count the number of pages within it, and to confirm that if we submitted all those pages to one or more BIOs, it could be split at any point with creating a misaligned BIO.
If that wouldn't be possible (as in any of the above situations), the ABD is linearised, and then checked again. This is the same technique used in
vdev_geom
on FreeBSD, adjusted for Linux's variable page size and allocator quirks.In the end, a count of segments is produced, and this is combined with the max seg count to determine how many BIOs will be needed.
Virtual block IO object
vbio_t
is a cleanup and enhancement of the olddio_request_t
. The idea is simply that it can hold all the state needed to create, submit and return multiple BIOs, including all the refcounts, the ABD copy if it was needed, and so on. Apart from what I hope is a clearer interface, the major difference is that because we know how many BIOs we'll need up front, we don't need the old overflow logic that would grow the BIO array, throw away all the old work and restart. We can get it right from the start.Other cleanups
Lots of cleanup, particularly through
vdev_disk_io_start
to make it feel a bit more like current ZFS code. Loads of comments too.Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
How Has This Been Tested?
Full test suite run passed.
Various customer workloads have been tried (in QA/test environments), including a heavy write workload on 7x raidz3-12 pools. and are not showing any signs of problems, either from within ZFS or from block devices or elsewhere.
(Surprisingly, heavy async write loads are getting ~2.7x throughput, though sync write loads are a more modest ~1.1x. I had expected some improvement, as we're not spending so much time breaking down the ABDs, but I didn't imagine the overheads could be that high).
On stock ZFS the splitting issue is complicated to reproduce, as we need to:
The last one means the SCSI driver, unless there’s another I don’t know about.
The first three can be demonstrated by forcing all allocations to be gang blocks.
Without this PR, this test should yield
request not aligned to the logical block size
errors almost immediately:Types of changes
Checklist:
Signed-off-by
.