Message ID | 1492123127-6266-6-git-send-email-jgunthorpe@obsidianresearch.com (mailing list archive) |
---|---|
State | Changes Requested |
Headers | show |
On 4/14/2017 1:38 AM, Jason Gunthorpe wrote: > This pattern is common in a couple of drivers, and needs the s390 > syscall. > > The common version properly handles 32 bit and prevents reordering of > the stores, which is the stated reason for this to exist. It is also > slightly more optimized, since it assumes a non-zero transfer length. > > Signed-off-by: Jason Gunthorpe <jgunthorpe@obsidianresearch.com> > --- > providers/mlx4/mmio.h | 43 ----------------------------------------- > providers/mlx4/qp.c | 5 ++--- > providers/mlx5/qp.c | 17 ++++++----------- > util/mmio.h | 53 ++++++++++++++++++++++++++++++++++++++++++++++++++- > 4 files changed, 60 insertions(+), 58 deletions(-) > delete mode 100644 providers/mlx4/mmio.h > > diff --git a/providers/mlx4/mmio.h b/providers/mlx4/mmio.h > deleted file mode 100644 > index 9821e85224dcfd..00000000000000 > diff --git a/providers/mlx4/qp.c b/providers/mlx4/qp.c > index 423f59533de68d..e7f10b9f1524d5 100644 > --- a/providers/mlx4/qp.c > +++ b/providers/mlx4/qp.c > @@ -43,7 +43,6 @@ > > #include "mlx4.h" > #include "wqe.h" > -#include "mmio.h" > > static const uint32_t mlx4_ib_opcode[] = { > [IBV_WR_SEND] = MLX4_OPCODE_SEND, > @@ -481,8 +480,8 @@ out: > */ > mmio_wc_spinlock(&ctx->bf_lock); > > - mlx4_bf_copy(ctx->bf_page + ctx->bf_offset, (unsigned long *) ctrl, > - align(size * 16, 64)); > + mmio_memcpy_x64(ctx->bf_page + ctx->bf_offset, ctrl, > + align(size * 16, 64)); > /* Flush before toggling bf_offset to be latency oriented */ > mmio_flush_writes(); > > diff --git a/providers/mlx5/qp.c b/providers/mlx5/qp.c > index 7f67a0b61b221f..c4789bf0d909a4 100644 > --- a/providers/mlx5/qp.c > +++ b/providers/mlx5/qp.c > @@ -239,19 +239,14 @@ static void set_data_ptr_seg_atomic(struct mlx5_wqe_data_seg *dseg, > static void mlx5_bf_copy(unsigned long long *dst, unsigned long long *src, > unsigned bytecnt, struct mlx5_qp *qp) > { > - while (bytecnt > 0) { > - *dst++ = *src++; > - *dst++ = *src++; > - *dst++ = *src++; > - *dst++ = *src++; > - *dst++ = *src++; > - *dst++ = *src++; > - *dst++ = *src++; > - *dst++ = *src++; > - bytecnt -= 8 * sizeof(unsigned long long); > + do { > + mmio_memcpy_x64(dst, src, 64); > + bytecnt -= 64; > + dst += 8; > + src += 8; It looks like the above +=8 is wrong in 32 bit systems, agree ? > if (unlikely(src == qp->sq.qend)) > src = qp->sq_start; > - } > + } while (bytecnt > 0); > } > > static uint32_t send_ieth(struct ibv_send_wr *wr) > diff --git a/util/mmio.h b/util/mmio.h > index 0b89f5fcbe000e..1d45d6d6364d4e 100644 > --- a/util/mmio.h > +++ b/util/mmio.h > @@ -56,6 +56,7 @@ > #include <linux/types.h> > #include <stdatomic.h> > #include <stdint.h> > +#include <stddef.h> > #include <endian.h> > > #include <config.h> > @@ -158,7 +159,6 @@ static inline uint8_t mmio_read8(const void *addr) > return atomic_load_explicit((_Atomic(uint32_t) *)addr, > memory_order_relaxed); > } > - > #endif /* __s390x__ */ > > MAKE_WRITE(mmio_write16, 16) > @@ -200,6 +200,57 @@ __le64 mmio_read64_le(const void *addr); > return le##_SZ_##toh(_NAME_##_le(addr)); \ > } > > +/* This strictly guarantees the order of TLP generation for the memory copy to > + be in ascending address order. > +*/ > +#ifdef __s390x__ > +static inline void mmio_memcpy_x64(void *dest, const void *src, size_t bytecnt) > +{ > + s390_mmio_write(addr, src, bytecnt); > +} > +#else > + > +/* Transfer is some multiple of 64 bytes */ > +static inline void mmio_memcpy_x64(void *dest, const void *src, size_t bytecnt) > +{ > + uintptr_t *dst_p = dest; > + > + /* Caller must guarantee: > + assert(bytecnt != 0); > + assert((bytecnt % 64) == 0); > + assert(((uintptr_t)dest) % __alignof__(*dst) == 0); > + assert(((uintptr_t)src) % __alignof__(*dst) == 0); > + */ > + > + /* Use the native word size for the copy */ > + if (sizeof(*dst_p) == 8) { We expect this 'if' to be dropped at compile time to prevent performance penalty comparing the original code, correct ? > + const __be64 *src_p = src; > + > + do { > + /* Do 64 bytes at a time */ > + mmio_write64_be(dst_p++, *src_p++); > + mmio_write64_be(dst_p++, *src_p++); > + mmio_write64_be(dst_p++, *src_p++); > + mmio_write64_be(dst_p++, *src_p++); > + mmio_write64_be(dst_p++, *src_p++); > + mmio_write64_be(dst_p++, *src_p++); > + mmio_write64_be(dst_p++, *src_p++); > + mmio_write64_be(dst_p++, *src_p++); > + > + bytecnt -= 8 * sizeof(*dst_p); > + } while (bytecnt > 0); > + } else if (sizeof(*dst_p) == 4) { > + const __be32 *src_p = src; > + > + do { > + mmio_write32_be(dst_p++, *src_p++); > + mmio_write32_be(dst_p++, *src_p++); > + bytecnt -= 2 * sizeof(*dst_p); Any reason not to write at least 64 bytes here before checking byte count and looping again ? > + } while (bytecnt > 0); > + } > +} > +#endif > + > MAKE_WRITE(mmio_write16, 16) > MAKE_WRITE(mmio_write32, 32) > MAKE_WRITE(mmio_write64, 64) > -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Tue, Apr 18, 2017 at 07:22:07PM +0300, Yishai Hadas wrote: > >@@ -239,19 +239,14 @@ static void set_data_ptr_seg_atomic(struct mlx5_wqe_data_seg *dseg, > > static void mlx5_bf_copy(unsigned long long *dst, unsigned long long *src, > > unsigned bytecnt, struct mlx5_qp *qp) > > { > >- while (bytecnt > 0) { > >- *dst++ = *src++; > >- *dst++ = *src++; > >- *dst++ = *src++; > >- *dst++ = *src++; > >- *dst++ = *src++; > >- *dst++ = *src++; > >- *dst++ = *src++; > >- *dst++ = *src++; > >- bytecnt -= 8 * sizeof(unsigned long long); > >+ do { > >+ mmio_memcpy_x64(dst, src, 64); > >+ bytecnt -= 64; > >+ dst += 8; > >+ src += 8; > > It looks like the above +=8 is wrong in 32 bit systems, agree ? Hurm. On 32 bit systems 'unsigned long long' will still be 64 bit, so the above is OK. The above original is buggy on 32 bit because it is not guarenteed to generate stores strictly in increasing address order. I think the author's intent was to have used 'uintptr_t *'. I will change the arguments to be 'uint64_t *' for clarity. > >+ /* Use the native word size for the copy */ > >+ if (sizeof(*dst_p) == 8) { > > We expect this 'if' to be dropped at compile time to prevent performance > penalty comparing the original code, correct ? Yes. The entire mmio_memcpy_x64 expands to a bunch of movs with no branches as the transfer size is constant as well. The overall mlx5_bf_copy looses one branch because of the transformation to do/while > >+ } while (bytecnt > 0); > >+ } else if (sizeof(*dst_p) == 4) { > >+ const __be32 *src_p = src; > >+ > >+ do { > >+ mmio_write32_be(dst_p++, *src_p++); > >+ mmio_write32_be(dst_p++, *src_p++); > >+ bytecnt -= 2 * sizeof(*dst_p); > > Any reason not to write at least 64 bytes here before checking byte count > and looping again ? icache size? I debated doing that, but the consensus of the existing implementations seems to be against it.. We could do a 32 byte unwind which is probably a similar icache footprint? What would you like? Jason -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 4/18/2017 9:27 PM, Jason Gunthorpe wrote: >>> + } else if (sizeof(*dst_p) == 4) { >>> + const __be32 *src_p = src; >>> + >>> + do { >>> + mmio_write32_be(dst_p++, *src_p++); >>> + mmio_write32_be(dst_p++, *src_p++); >>> + bytecnt -= 2 * sizeof(*dst_p); >> >> Any reason not to write at least 64 bytes here before checking byte count >> and looping again ? > > icache size? I debated doing that, but the consensus of the existing > implementations seems to be against it.. > > We could do a 32 byte unwind which is probably a similar icache > footprint? > > What would you like? I'm fine with leaving the code as is following the existing implementation. -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/providers/mlx4/mmio.h b/providers/mlx4/mmio.h deleted file mode 100644 index 9821e85224dcfd..00000000000000 diff --git a/providers/mlx4/qp.c b/providers/mlx4/qp.c index 423f59533de68d..e7f10b9f1524d5 100644 --- a/providers/mlx4/qp.c +++ b/providers/mlx4/qp.c @@ -43,7 +43,6 @@ #include "mlx4.h" #include "wqe.h" -#include "mmio.h" static const uint32_t mlx4_ib_opcode[] = { [IBV_WR_SEND] = MLX4_OPCODE_SEND, @@ -481,8 +480,8 @@ out: */ mmio_wc_spinlock(&ctx->bf_lock); - mlx4_bf_copy(ctx->bf_page + ctx->bf_offset, (unsigned long *) ctrl, - align(size * 16, 64)); + mmio_memcpy_x64(ctx->bf_page + ctx->bf_offset, ctrl, + align(size * 16, 64)); /* Flush before toggling bf_offset to be latency oriented */ mmio_flush_writes(); diff --git a/providers/mlx5/qp.c b/providers/mlx5/qp.c index 7f67a0b61b221f..c4789bf0d909a4 100644 --- a/providers/mlx5/qp.c +++ b/providers/mlx5/qp.c @@ -239,19 +239,14 @@ static void set_data_ptr_seg_atomic(struct mlx5_wqe_data_seg *dseg, static void mlx5_bf_copy(unsigned long long *dst, unsigned long long *src, unsigned bytecnt, struct mlx5_qp *qp) { - while (bytecnt > 0) { - *dst++ = *src++; - *dst++ = *src++; - *dst++ = *src++; - *dst++ = *src++; - *dst++ = *src++; - *dst++ = *src++; - *dst++ = *src++; - *dst++ = *src++; - bytecnt -= 8 * sizeof(unsigned long long); + do { + mmio_memcpy_x64(dst, src, 64); + bytecnt -= 64; + dst += 8; + src += 8; if (unlikely(src == qp->sq.qend)) src = qp->sq_start; - } + } while (bytecnt > 0); } static uint32_t send_ieth(struct ibv_send_wr *wr) diff --git a/util/mmio.h b/util/mmio.h index 0b89f5fcbe000e..1d45d6d6364d4e 100644 --- a/util/mmio.h +++ b/util/mmio.h @@ -56,6 +56,7 @@ #include <linux/types.h> #include <stdatomic.h> #include <stdint.h> +#include <stddef.h> #include <endian.h> #include <config.h> @@ -158,7 +159,6 @@ static inline uint8_t mmio_read8(const void *addr) return atomic_load_explicit((_Atomic(uint32_t) *)addr, memory_order_relaxed); } - #endif /* __s390x__ */ MAKE_WRITE(mmio_write16, 16) @@ -200,6 +200,57 @@ __le64 mmio_read64_le(const void *addr); return le##_SZ_##toh(_NAME_##_le(addr)); \ } +/* This strictly guarantees the order of TLP generation for the memory copy to + be in ascending address order. +*/ +#ifdef __s390x__ +static inline void mmio_memcpy_x64(void *dest, const void *src, size_t bytecnt) +{ + s390_mmio_write(addr, src, bytecnt); +} +#else + +/* Transfer is some multiple of 64 bytes */ +static inline void mmio_memcpy_x64(void *dest, const void *src, size_t bytecnt) +{ + uintptr_t *dst_p = dest; + + /* Caller must guarantee: + assert(bytecnt != 0); + assert((bytecnt % 64) == 0); + assert(((uintptr_t)dest) % __alignof__(*dst) == 0); + assert(((uintptr_t)src) % __alignof__(*dst) == 0); + */ + + /* Use the native word size for the copy */ + if (sizeof(*dst_p) == 8) { + const __be64 *src_p = src; + + do { + /* Do 64 bytes at a time */ + mmio_write64_be(dst_p++, *src_p++); + mmio_write64_be(dst_p++, *src_p++); + mmio_write64_be(dst_p++, *src_p++); + mmio_write64_be(dst_p++, *src_p++); + mmio_write64_be(dst_p++, *src_p++); + mmio_write64_be(dst_p++, *src_p++); + mmio_write64_be(dst_p++, *src_p++); + mmio_write64_be(dst_p++, *src_p++); + + bytecnt -= 8 * sizeof(*dst_p); + } while (bytecnt > 0); + } else if (sizeof(*dst_p) == 4) { + const __be32 *src_p = src; + + do { + mmio_write32_be(dst_p++, *src_p++); + mmio_write32_be(dst_p++, *src_p++); + bytecnt -= 2 * sizeof(*dst_p); + } while (bytecnt > 0); + } +} +#endif + MAKE_WRITE(mmio_write16, 16) MAKE_WRITE(mmio_write32, 32) MAKE_WRITE(mmio_write64, 64)
This pattern is common in a couple of drivers, and needs the s390 syscall. The common version properly handles 32 bit and prevents reordering of the stores, which is the stated reason for this to exist. It is also slightly more optimized, since it assumes a non-zero transfer length. Signed-off-by: Jason Gunthorpe <jgunthorpe@obsidianresearch.com> --- providers/mlx4/mmio.h | 43 ----------------------------------------- providers/mlx4/qp.c | 5 ++--- providers/mlx5/qp.c | 17 ++++++----------- util/mmio.h | 53 ++++++++++++++++++++++++++++++++++++++++++++++++++- 4 files changed, 60 insertions(+), 58 deletions(-) delete mode 100644 providers/mlx4/mmio.h