| Message ID | 20230920222231.686275-1-dhowells@redhat.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | iov_iter: Convert the iterator macros into inline funcs | expand |
From: David Howells > Sent: 20 September 2023 23:22 ... > (8) Move the copy-and-csum code to net/ where it can be in proximity with > the code that uses it. This eliminates the code if CONFIG_NET=n and > allows for the slim possibility of it being inlined. > > (9) Fold memcpy_and_csum() in to its two users. > > (10) Move csum_and_copy_from_iter_full() out of line and merge in > csum_and_copy_from_iter() since the former is the only caller of the > latter. I thought that the real idea behind these was to do the checksum at the same time as the copy to avoid loading the data into the L1 data-cache twice - especially for long buffers. I wonder how often there are multiple iov[] that actually make it better than just check summing the linear buffer? I had a feeling that check summing of udp data was done during copy_to/from_user, but the code can't be the copy-and-csum here for that because it is missing support form odd-length buffers. Intel x86 desktop chips can easily checksum at 8 bytes/clock (But probably not with the current code!). (I've got ~12 bytes/clock using adox and adcx but that loop is entirely horrid and it would need run-time patching. Especially since I think some AMD cpu execute them very slowly.) OTOH 'rep movs[bq]' copy will copy 16 bytes/clock (32 if the destination is 32 byte aligned - it pretty much won't be). So you'd need a csum-and-copy loop that did 16 bytes every three clocks to get the same throughput for long buffers. In principle splitting the 'adc memory' into two instructions is the same number of u-ops - but I'm sure I've tried to do that and failed and the extra memory write can happen in parallel with everything else. So I don't think you'll get 16 bytes in two clocks - but you might get it is three. OTOH for a cpu where memcpy is code loop summing the data in the copy loop is likely to be a gain. But I suspect doing the checksum and copy at the same time got 'all to complicated' to actually implement fully. With most modern ethernet chips checksumming receive pacakets does it really get used enough for the additional complexity? David - Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)
David Laight <David.Laight@ACULAB.COM> wrote: > > (8) Move the copy-and-csum code to net/ where it can be in proximity with > > the code that uses it. This eliminates the code if CONFIG_NET=n and > > allows for the slim possibility of it being inlined. > > > > (9) Fold memcpy_and_csum() in to its two users. > > > > (10) Move csum_and_copy_from_iter_full() out of line and merge in > > csum_and_copy_from_iter() since the former is the only caller of the > > latter. > > I thought that the real idea behind these was to do the checksum > at the same time as the copy to avoid loading the data into the L1 > data-cache twice - especially for long buffers. > I wonder how often there are multiple iov[] that actually make > it better than just check summing the linear buffer? It also reduces the overhead for finding the data to checksum in the case the packet gets split since we're doing the checksumming as we copy - but with a linear buffer, that's negligible. > I had a feeling that check summing of udp data was done during > copy_to/from_user, but the code can't be the copy-and-csum here > for that because it is missing support form odd-length buffers. Is there a bug there? > Intel x86 desktop chips can easily checksum at 8 bytes/clock > (But probably not with the current code!). > (I've got ~12 bytes/clock using adox and adcx but that loop > is entirely horrid and it would need run-time patching. > Especially since I think some AMD cpu execute them very slowly.) > > OTOH 'rep movs[bq]' copy will copy 16 bytes/clock (32 if the > destination is 32 byte aligned - it pretty much won't be). > > So you'd need a csum-and-copy loop that did 16 bytes every > three clocks to get the same throughput for long buffers. > In principle splitting the 'adc memory' into two instructions > is the same number of u-ops - but I'm sure I've tried to do > that and failed and the extra memory write can happen in > parallel with everything else. > So I don't think you'll get 16 bytes in two clocks - but you > might get it is three. > > OTOH for a cpu where memcpy is code loop summing the data in > the copy loop is likely to be a gain. > > But I suspect doing the checksum and copy at the same time > got 'all to complicated' to actually implement fully. > With most modern ethernet chips checksumming receive pacakets > does it really get used enough for the additional complexity? You may be right. That's more a question for the networking folks than for me. It's entirely possible that the checksumming code is just not used on modern systems these days. Maybe Willem can comment since he's the UDP maintainer? David
On Fri, Sep 22, 2023 at 2:01 PM David Howells <dhowells@redhat.com> wrote: > > David Laight <David.Laight@ACULAB.COM> wrote: > > > > (8) Move the copy-and-csum code to net/ where it can be in proximity with > > > the code that uses it. This eliminates the code if CONFIG_NET=n and > > > allows for the slim possibility of it being inlined. > > > > > > (9) Fold memcpy_and_csum() in to its two users. > > > > > > (10) Move csum_and_copy_from_iter_full() out of line and merge in > > > csum_and_copy_from_iter() since the former is the only caller of the > > > latter. > > > > I thought that the real idea behind these was to do the checksum > > at the same time as the copy to avoid loading the data into the L1 > > data-cache twice - especially for long buffers. > > I wonder how often there are multiple iov[] that actually make > > it better than just check summing the linear buffer? > > It also reduces the overhead for finding the data to checksum in the case the > packet gets split since we're doing the checksumming as we copy - but with a > linear buffer, that's negligible. > > > I had a feeling that check summing of udp data was done during > > copy_to/from_user, but the code can't be the copy-and-csum here > > for that because it is missing support form odd-length buffers. > > Is there a bug there? > > > Intel x86 desktop chips can easily checksum at 8 bytes/clock > > (But probably not with the current code!). > > (I've got ~12 bytes/clock using adox and adcx but that loop > > is entirely horrid and it would need run-time patching. > > Especially since I think some AMD cpu execute them very slowly.) > > > > OTOH 'rep movs[bq]' copy will copy 16 bytes/clock (32 if the > > destination is 32 byte aligned - it pretty much won't be). > > > > So you'd need a csum-and-copy loop that did 16 bytes every > > three clocks to get the same throughput for long buffers. > > In principle splitting the 'adc memory' into two instructions > > is the same number of u-ops - but I'm sure I've tried to do > > that and failed and the extra memory write can happen in > > parallel with everything else. > > So I don't think you'll get 16 bytes in two clocks - but you > > might get it is three. > > > > OTOH for a cpu where memcpy is code loop summing the data in > > the copy loop is likely to be a gain. > > > > But I suspect doing the checksum and copy at the same time > > got 'all to complicated' to actually implement fully. > > With most modern ethernet chips checksumming receive pacakets > > does it really get used enough for the additional complexity? > > You may be right. That's more a question for the networking folks than for > me. It's entirely possible that the checksumming code is just not used on > modern systems these days. > > Maybe Willem can comment since he's the UDP maintainer? Perhaps these days it is more relevant to embedded systems than high end servers.
From: Willem de Bruijn > Sent: 23 September 2023 07:59 > > On Fri, Sep 22, 2023 at 2:01 PM David Howells <dhowells@redhat.com> wrote: > > > > David Laight <David.Laight@ACULAB.COM> wrote: > > > > > > (8) Move the copy-and-csum code to net/ where it can be in proximity with > > > > the code that uses it. This eliminates the code if CONFIG_NET=n and > > > > allows for the slim possibility of it being inlined. > > > > > > > > (9) Fold memcpy_and_csum() in to its two users. > > > > > > > > (10) Move csum_and_copy_from_iter_full() out of line and merge in > > > > csum_and_copy_from_iter() since the former is the only caller of the > > > > latter. > > > > > > I thought that the real idea behind these was to do the checksum > > > at the same time as the copy to avoid loading the data into the L1 > > > data-cache twice - especially for long buffers. > > > I wonder how often there are multiple iov[] that actually make > > > it better than just check summing the linear buffer? > > > > It also reduces the overhead for finding the data to checksum in the case the > > packet gets split since we're doing the checksumming as we copy - but with a > > linear buffer, that's negligible. > > > > > I had a feeling that check summing of udp data was done during > > > copy_to/from_user, but the code can't be the copy-and-csum here > > > for that because it is missing support form odd-length buffers. > > > > Is there a bug there? No, I misread the code - i shouldn't scan patches when I'd got a viral head code... ... > > You may be right. That's more a question for the networking folks than for > > me. It's entirely possible that the checksumming code is just not used on > > modern systems these days. > > > > Maybe Willem can comment since he's the UDP maintainer? > > Perhaps these days it is more relevant to embedded systems than high > end servers. The checksum and copy are done together. I probably missed it because the function isn't passed the old checksum (which it can pretty much process for free). Instead the caller is adding it afterwards - which involves and extra explicit csum_add(). The x86-x84 ip checksum loops are all horrid though. The unrolling in them is so 1990's. With the out-of-order pipeline the memory accesses tend to take care of themselves. Not to mention that a whole raft of (now oldish) cpu take two clocks to execute 'adc'. David - Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)
Hi Jens, Could you take these patches into the block tree? Note that it's on top of the kunit patchset though it could be separate. The patches convert the iov_iter iteration macros to be inline functions. (1) Convert iter->user_backed to user_backed_iter() in the sound PCM driver. (2) Convert iter->user_backed to user_backed_iter() in a couple of infiniband drivers. (3) Renumber the type enum so that the ITER_* constants match the order in iterate_and_advance*(). (4) Since (3) puts UBUF and IOVEC at 0 and 1, change user_backed_iter() to just use the type value and get rid of the extra flag. (5) Convert the iov_iter iteration macros to always-inline functions to make the code easier to follow. It uses function pointers, but they get optimised away. The priv2 argument likewise gets optimised away if unused. The functions are placed into a header file so that bespoke iterators can be created elsewhere. For instance, rbd has an optimisation that requires it to scan to the buffer it is given to see if it is all zeros. It would be nice if this could use iterate_and_advance() - but that's currently buried inside lib/iov_iter.c. Whilst we could provide a generic iteration function that takes a pair of function pointers instead, it does add around 16% overhead in the framework, presumably from the combination of function pointers and various mitigations. Further, if it is known that just a particular iterator-type is in play, just that iteration function can be used. (6) Move the check for ->copy_mc to _copy_from_iter() and copy_page_from_iter_atomic() rather than in memcpy_from_iter_mc() where it gets repeated for every segment. Instead, we check once and invoke a side function that can use iterate_bvec() rather than iterate_and_advance() and supply a different step function. (7) Provide a cut-down iteration function that only handles kernel-backed iterators (ie. BVEC, KVEC, XARRAY and DISCARD) for situations where we know that we won't see UBUF/IOVEC. This crops up in a number of filesystems where we decant a UBUF/IOVEC iter into a series of BVEC iters. (8) Move the copy-and-csum code to net/ where it can be in proximity with the code that uses it. This eliminates the code if CONFIG_NET=n and allows for the slim possibility of it being inlined. (9) Fold memcpy_and_csum() in to its two users. (10) Move csum_and_copy_from_iter_full() out of line and merge in csum_and_copy_from_iter() since the former is the only caller of the latter. (11) Move hash_and_copy_to_iter() to net/ where it can be with its only caller. Anyway, the changes in compiled function size either side of patches (5)+(6) on x86_64 look like: __copy_from_iter_mc new 0xe8 _copy_from_iter inc 0x360 -> 0x3a7 +0x47 _copy_from_iter_flushcache inc 0x34c -> 0x38e +0x42 _copy_from_iter_nocache inc 0x354 -> 0x378 +0x24 _copy_mc_to_iter inc 0x396 -> 0x3f1 +0x5b _copy_to_iter inc 0x33b -> 0x385 +0x4a copy_page_from_iter_atomic.part.0 inc 0x393 -> 0x3e0 +0x4d copy_page_to_iter_nofault.part.0 inc 0x3de -> 0x3eb +0xd copyin del 0x30 copyout del 0x2d copyout_mc del 0x2b csum_and_copy_from_iter inc 0x3db -> 0x41d +0x42 csum_and_copy_to_iter inc 0x45d -> 0x48d +0x30 iov_iter_zero inc 0x34a -> 0x36a +0x20 memcpy_from_iter.isra.0 del 0x1f Note that there's a noticeable expansion on some of the main functions because a number of the helpers get inlined instead of being called. In terms of benchmarking patch (5)+(6), three runs without them: iov_kunit_benchmark_ubuf: avg 3955 uS, stddev 169 uS iov_kunit_benchmark_ubuf: avg 4122 uS, stddev 1292 uS iov_kunit_benchmark_ubuf: avg 4451 uS, stddev 1362 uS iov_kunit_benchmark_iovec: avg 6607 uS, stddev 22 uS iov_kunit_benchmark_iovec: avg 6608 uS, stddev 19 uS iov_kunit_benchmark_iovec: avg 6609 uS, stddev 24 uS iov_kunit_benchmark_bvec: avg 3166 uS, stddev 11 uS iov_kunit_benchmark_bvec: avg 3167 uS, stddev 13 uS iov_kunit_benchmark_bvec: avg 3170 uS, stddev 16 uS iov_kunit_benchmark_bvec_split: avg 3394 uS, stddev 12 uS iov_kunit_benchmark_bvec_split: avg 3394 uS, stddev 20 uS iov_kunit_benchmark_bvec_split: avg 3395 uS, stddev 19 uS iov_kunit_benchmark_kvec: avg 2672 uS, stddev 12 uS iov_kunit_benchmark_kvec: avg 2672 uS, stddev 12 uS iov_kunit_benchmark_kvec: avg 2672 uS, stddev 9 uS iov_kunit_benchmark_xarray: avg 3719 uS, stddev 9 uS iov_kunit_benchmark_xarray: avg 3719 uS, stddev 9 uS iov_kunit_benchmark_xarray: avg 3721 uS, stddev 24 uS and three runs with them: iov_kunit_benchmark_ubuf: avg 4110 uS, stddev 1254 uS iov_kunit_benchmark_ubuf: avg 4141 uS, stddev 1411 uS iov_kunit_benchmark_ubuf: avg 4572 uS, stddev 1889 uS iov_kunit_benchmark_iovec: avg 6582 uS, stddev 27 uS iov_kunit_benchmark_iovec: avg 6585 uS, stddev 25 uS iov_kunit_benchmark_iovec: avg 6586 uS, stddev 48 uS iov_kunit_benchmark_bvec: avg 3175 uS, stddev 13 uS iov_kunit_benchmark_bvec: avg 3177 uS, stddev 12 uS iov_kunit_benchmark_bvec: avg 3178 uS, stddev 12 uS iov_kunit_benchmark_bvec_split: avg 3380 uS, stddev 20 uS iov_kunit_benchmark_bvec_split: avg 3384 uS, stddev 15 uS iov_kunit_benchmark_bvec_split: avg 3386 uS, stddev 25 uS iov_kunit_benchmark_kvec: avg 2671 uS, stddev 11 uS iov_kunit_benchmark_kvec: avg 2672 uS, stddev 12 uS iov_kunit_benchmark_kvec: avg 2677 uS, stddev 20 uS iov_kunit_benchmark_xarray: avg 3599 uS, stddev 20 uS iov_kunit_benchmark_xarray: avg 3603 uS, stddev 8 uS iov_kunit_benchmark_xarray: avg 3610 uS, stddev 16 uS I've pushed the patches here also: https://git.kernel.org/pub/scm/linux/kernel/git/dhowells/linux-fs.git/log/?h=iov-cleanup David Changes ======= ver #5) - Moved kunit test patches out to a separate set. - Moved "iter->count - progress" into individual iteration subfunctions. - Fix iterate_iovec() and to update iter->__iovec after subtracting it to calculate iter->nr_segs. - Merged the function inlining patch and the move-to-header patch. - Rearranged the patch order slightly to put the patches to move networking stuff to net/ last. - Went back to a simpler patch to special-case ->copy_mc in _copy_from_iter() and copy_page_from_iter_atomic() before we get to call iterate_and_advance() so as to reduce the number of checks for this. ver #4) - Fix iterate_bvec() and iterate_kvec() to update iter->bvec and iter->kvec after subtracting it to calculate iter->nr_segs. - Change iterate_xarray() to use start+progress rather than increasing start to reduce code size. - Added patches to move some iteration functions over to net/ as the files there can #include the iterator framework. - Added a patch to benchmark the iteration. - Tried an experimental patch to make copy_from_iter() and similar always catch MCE. ver #3) - Use min_t(size_t,) not min() to avoid a warning on Hexagon. - Inline all the step functions. - Added a patch to better handle copy_mc. ver #2) - Rebased on top of Willy's changes in linux-next. - Change the checksum argument to the iteration functions to be a general void* and use it to pass iter->copy_mc flag to memcpy_from_iter_mc() to avoid using a function pointer. - Arrange the end of the iterate_*() functions to look the same to give the optimiser the best chance. - Make iterate_and_advance() a wrapper around iterate_and_advance2(). - Adjust iterate_and_advance2() to use if-else-if-else-if-else rather than switch(), to put ITER_BVEC before KVEC and to mark UBUF and IOVEC as likely(). - Move "iter->count += progress" into iterate_and_advance2() from the iterate functions. - Mark a number of the iterator helpers with __always_inline. - Fix _copy_from_iter_flushcache() to use memcpy_from_iter_flushcache() not memcpy_from_iter(). Link: https://lore.kernel.org/r/3710261.1691764329@warthog.procyon.org.uk/ # v1 Link: https://lore.kernel.org/r/855.1692047347@warthog.procyon.org.uk/ # v2 Link: https://lore.kernel.org/r/20230816120741.534415-1-dhowells@redhat.com/ # v3 Link: https://lore.kernel.org/r/20230913165648.2570623-1-dhowells@redhat.com/ # v4 David Howells (11): sound: Fix snd_pcm_readv()/writev() to use iov access functions infiniband: Use user_backed_iter() to see if iterator is UBUF/IOVEC iov_iter: Renumber ITER_* constants iov_iter: Derive user-backedness from the iterator type iov_iter: Convert iterate*() to inline funcs iov_iter: Don't deal with iter->copy_mc in memcpy_from_iter_mc() iov_iter: Add a kernel-type iterator-only iteration function iov_iter, net: Move csum_and_copy_to/from_iter() to net/ iov_iter, net: Fold in csum_and_memcpy() iov_iter, net: Merge csum_and_copy_from_iter{,_full}() together iov_iter, net: Move hash_and_copy_to_iter() to net/ drivers/infiniband/hw/hfi1/file_ops.c | 2 +- drivers/infiniband/hw/qib/qib_file_ops.c | 2 +- include/linux/iov_iter.h | 305 ++++++++++++++++ include/linux/skbuff.h | 3 + include/linux/uio.h | 29 +- lib/iov_iter.c | 441 +++++++---------------- net/core/datagram.c | 75 +++- net/core/skbuff.c | 40 ++ sound/core/pcm_native.c | 4 +- 9 files changed, 569 insertions(+), 332 deletions(-) create mode 100644 include/linux/iov_iter.h