Re: [lng-odp] [PATCHv5] helper : Fix UDP checksum computation
Hi, I sent a new patch(V6) making the code more understandable. Thanks, Grig -Original Message- From: Ilya Maximets [mailto:i.maxim...@samsung.com] Sent: Wednesday, December 09, 2015 2:52 PM To: Grigore Ion-B17953; lng-odp@lists.linaro.org Subject: Re: [PATCHv5] helper : Fix UDP checksum computation On 09.12.2015 15:26, Ion Grigore wrote: > See inline comments. > > -Original Message- > From: Ilya Maximets [mailto:i.maxim...@samsung.com] > Sent: Wednesday, December 09, 2015 12:04 PM > To: Grigore Ion-B17953 ; > lng-odp@lists.linaro.org > Subject: Re: [PATCHv5] helper : Fix UDP checksum computation > > Some comments below. > > On 09.12.2015 11:59, ion.grig...@freescale.com wrote: >> From: Grigore Ion >> >> This patch fixes the following problems: >> - checksum computation for LE platforms >> - checksum is computed in the CPU endianness. The returned result >> must be converted to the BE ordering when it is used to update the >> UDP checksum in a packet. >> - checksum computation for packets having the UDP length not a >> multiple of 2 >> - fixes the UDP checksum associated validation test >> >> Signed-off-by: Grigore Ion >> --- >> v5: >> - Checksum in CPU endianness fix added (Ilya Maximets) >> v4: >> - Verify checksum in CPU endianness in the associated test (Ilya >> Maximets) >> v3: >> - fix the UDP checksum computation in the associated test (Maxim >> Uvarov) >> v2: >> - patch updated to the last master (Maxim Uvarov) >> v1: >> - Move variables declaration on top of block. (Maxim Uvarov) >> - Check patch with checkpatch script. (Maxim Uvarov) >> - L3 header presence is tested twice. (Alexandru Badicioiu) >> - Remove unnecessary check for L3 header presence. (Bill Fischofer) >> - Modify check of odp_packet_l4_offset() return. (Bill Fischofer) >> >> helper/include/odp/helper/udp.h | 55 >> + >> helper/test/odp_chksum.c| 4 +-- >> 2 files changed, 25 insertions(+), 34 deletions(-) >> >> diff --git a/helper/include/odp/helper/udp.h >> b/helper/include/odp/helper/udp.h index 06c439b..9e7256d 100644 >> --- a/helper/include/odp/helper/udp.h >> +++ b/helper/include/odp/helper/udp.h >> @@ -4,7 +4,6 @@ >> * SPDX-License-Identifier: BSD-3-Clause >> */ >> >> - >> /** >> * @file >> * >> @@ -22,7 +21,6 @@ extern "C" { >> #include >> #include >> >> - >> /** @addtogroup odph_header ODPH HEADER >> * @{ >> */ >> @@ -44,46 +42,39 @@ typedef struct ODP_PACKED { >> * This function uses odp packet to calc checksum >> * >> * @param pkt calculate chksum for pkt >> - * @return checksum value >> + * @return checksum value in CPU endianness >> */ >> static inline uint16_t odph_ipv4_udp_chksum(odp_packet_t pkt) { >> -uint32_t sum = 0; >> -odph_udphdr_t *udph; >> -odph_ipv4hdr_t *iph; >> -uint16_t udplen; >> -uint8_t *buf; >> - >> -if (!odp_packet_l3_offset(pkt)) >> -return 0; >> +odph_ipv4hdr_t *iph; >> +odph_udphdr_t *udph; >> +uint32_tsum; >> +uint16_tudplen, *buf; >> >> -if (!odp_packet_l4_offset(pkt)) >> +if (odp_packet_l4_offset(pkt) == ODP_PACKET_OFFSET_INVALID) >> return 0; >> - >> iph = (odph_ipv4hdr_t *)odp_packet_l3_ptr(pkt, NULL); >> udph = (odph_udphdr_t *)odp_packet_l4_ptr(pkt, NULL); >> -udplen = odp_be_to_cpu_16(udph->length); >> - >> -/* 32-bit sum of all 16-bit words covered by UDP chksum */ >> +/* 32-bit sum of UDP pseudo-header */ >> sum = (iph->src_addr & 0x) + (iph->src_addr >> 16) + >> - (iph->dst_addr & 0x) + (iph->dst_addr >> 16) + >> - (uint16_t)iph->proto + udplen; >> -for (buf = (uint8_t *)udph; udplen > 1; udplen -= 2) { >> -sum += ((*buf << 8) + *(buf + 1)); >> -buf += 2; >> -} >> - >> -/* Fold sum to 16 bits: add carrier to result */ >> -while (sum >> 16) >> -sum = (sum & 0x) + (sum >> 16); >> - >> +(iph->dst_addr & 0x) + (iph->dst_addr >> 16) + >> +odp_be_to_cpu_16(iph->proto) + udph->length; > > This line is also strange. You're using odp_be_to_cpu_16 for 8-bit digit. > At this point, iph->proto firstly will be extended to u16 in CPU endianness > and after that odp_be_to_cpu_16() will return it to BE. You will have right > position of all bytes (8 zero bits + 8 bits of iph->proto), but the way how > this done is wrong. > > [IGR] Here the proto field is added as the LSB in a 16-bit BE value. Do you > think the following code clearer ? > > uint16 val; > ... > val = iph->proto; > sum += odp_be_to_cpu_16(val); No, because 'val' is in CPU endianness and you're using odp_be_to_cpu_16() for variable that is already in CPU endianness. > >> +udplen = odp_be_to_cpu_16(udph->length); >> +buf =
Re: [lng-odp] [PATCHv5] helper : Fix UDP checksum computation
Some comments below. On 09.12.2015 11:59, ion.grig...@freescale.com wrote: > From: Grigore Ion> > This patch fixes the following problems: > - checksum computation for LE platforms > - checksum is computed in the CPU endianness. The returned result > must be converted to the BE ordering when it is used to update > the UDP checksum in a packet. > - checksum computation for packets having the UDP length not a > multiple of 2 > - fixes the UDP checksum associated validation test > > Signed-off-by: Grigore Ion > --- > v5: > - Checksum in CPU endianness fix added (Ilya Maximets) > v4: > - Verify checksum in CPU endianness in the associated test > (Ilya Maximets) > v3: > - fix the UDP checksum computation in the associated test > (Maxim Uvarov) > v2: > - patch updated to the last master (Maxim Uvarov) > v1: > - Move variables declaration on top of block. (Maxim Uvarov) > - Check patch with checkpatch script. (Maxim Uvarov) > - L3 header presence is tested twice. (Alexandru Badicioiu) > - Remove unnecessary check for L3 header presence. (Bill Fischofer) > - Modify check of odp_packet_l4_offset() return. (Bill Fischofer) > > helper/include/odp/helper/udp.h | 55 > + > helper/test/odp_chksum.c| 4 +-- > 2 files changed, 25 insertions(+), 34 deletions(-) > > diff --git a/helper/include/odp/helper/udp.h b/helper/include/odp/helper/udp.h > index 06c439b..9e7256d 100644 > --- a/helper/include/odp/helper/udp.h > +++ b/helper/include/odp/helper/udp.h > @@ -4,7 +4,6 @@ > * SPDX-License-Identifier: BSD-3-Clause > */ > > - > /** > * @file > * > @@ -22,7 +21,6 @@ extern "C" { > #include > #include > > - > /** @addtogroup odph_header ODPH HEADER > * @{ > */ > @@ -44,46 +42,39 @@ typedef struct ODP_PACKED { > * This function uses odp packet to calc checksum > * > * @param pkt calculate chksum for pkt > - * @return checksum value > + * @return checksum value in CPU endianness > */ > static inline uint16_t odph_ipv4_udp_chksum(odp_packet_t pkt) > { > - uint32_t sum = 0; > - odph_udphdr_t *udph; > - odph_ipv4hdr_t *iph; > - uint16_t udplen; > - uint8_t *buf; > - > - if (!odp_packet_l3_offset(pkt)) > - return 0; > + odph_ipv4hdr_t *iph; > + odph_udphdr_t *udph; > + uint32_tsum; > + uint16_tudplen, *buf; > > - if (!odp_packet_l4_offset(pkt)) > + if (odp_packet_l4_offset(pkt) == ODP_PACKET_OFFSET_INVALID) > return 0; > - > iph = (odph_ipv4hdr_t *)odp_packet_l3_ptr(pkt, NULL); > udph = (odph_udphdr_t *)odp_packet_l4_ptr(pkt, NULL); > - udplen = odp_be_to_cpu_16(udph->length); > - > - /* 32-bit sum of all 16-bit words covered by UDP chksum */ > + /* 32-bit sum of UDP pseudo-header */ > sum = (iph->src_addr & 0x) + (iph->src_addr >> 16) + > - (iph->dst_addr & 0x) + (iph->dst_addr >> 16) + > - (uint16_t)iph->proto + udplen; > - for (buf = (uint8_t *)udph; udplen > 1; udplen -= 2) { > - sum += ((*buf << 8) + *(buf + 1)); > - buf += 2; > - } > - > - /* Fold sum to 16 bits: add carrier to result */ > - while (sum >> 16) > - sum = (sum & 0x) + (sum >> 16); > - > + (iph->dst_addr & 0x) + (iph->dst_addr >> 16) + > + odp_be_to_cpu_16(iph->proto) + udph->length; This line is also strange. You're using odp_be_to_cpu_16 for 8-bit digit. At this point, iph->proto firstly will be extended to u16 in CPU endianness and after that odp_be_to_cpu_16() will return it to BE. You will have right position of all bytes (8 zero bits + 8 bits of iph->proto), but the way how this done is wrong. > + udplen = odp_be_to_cpu_16(udph->length); > + buf = (uint16_t *)((void *)udph); > + /* 32-bit sum of UDP header (checksum field cleared) and UDP data */ > + for ( ; udplen > 1; udplen -= 2) > + sum += *buf++; > + /* Length is not a multiple of 2 bytes */ > + if (udplen) > + sum += odp_be_to_cpu_16(*((uint8_t *)buf) << 8); Same at this point. You're using odp_be_to_cpu_16() to perform opposite operation. May be, better to use byte arrays here to avoid that strange endianness and compiler aware operations. > + /* Fold sum to 16 bits */ > + sum = (sum & 0x) + (sum >> 16); > + /* Add carrier (0/1) to result */ > + sum += (sum >> 16); > /* 1's complement */ > sum = ~sum; > - > - /* set computation result */ > - sum = (sum == 0x0) ? 0x : sum; > - > - return sum; > + /* Set computation result in CPU endianness */ > + return (sum == 0x0) ? 0x : odp_be_to_cpu_16(sum); > } > > /** @internal Compile time assert */ > diff --git a/helper/test/odp_chksum.c b/helper/test/odp_chksum.c > index 1d417a8..152018a 100644 > --- a/helper/test/odp_chksum.c > +++
Re: [lng-odp] [PATCHv5] helper : Fix UDP checksum computation
On 09.12.2015 15:26, Ion Grigore wrote: > See inline comments. > > -Original Message- > From: Ilya Maximets [mailto:i.maxim...@samsung.com] > Sent: Wednesday, December 09, 2015 12:04 PM > To: Grigore Ion-B17953; lng-odp@lists.linaro.org > Subject: Re: [PATCHv5] helper : Fix UDP checksum computation > > Some comments below. > > On 09.12.2015 11:59, ion.grig...@freescale.com wrote: >> From: Grigore Ion >> >> This patch fixes the following problems: >> - checksum computation for LE platforms >> - checksum is computed in the CPU endianness. The returned result must >> be converted to the BE ordering when it is used to update the UDP >> checksum in a packet. >> - checksum computation for packets having the UDP length not a >> multiple of 2 >> - fixes the UDP checksum associated validation test >> >> Signed-off-by: Grigore Ion >> --- >> v5: >> - Checksum in CPU endianness fix added (Ilya Maximets) >> v4: >> - Verify checksum in CPU endianness in the associated test (Ilya >> Maximets) >> v3: >> - fix the UDP checksum computation in the associated test (Maxim >> Uvarov) >> v2: >> - patch updated to the last master (Maxim Uvarov) >> v1: >> - Move variables declaration on top of block. (Maxim Uvarov) >> - Check patch with checkpatch script. (Maxim Uvarov) >> - L3 header presence is tested twice. (Alexandru Badicioiu) >> - Remove unnecessary check for L3 header presence. (Bill Fischofer) >> - Modify check of odp_packet_l4_offset() return. (Bill Fischofer) >> >> helper/include/odp/helper/udp.h | 55 >> + >> helper/test/odp_chksum.c| 4 +-- >> 2 files changed, 25 insertions(+), 34 deletions(-) >> >> diff --git a/helper/include/odp/helper/udp.h >> b/helper/include/odp/helper/udp.h index 06c439b..9e7256d 100644 >> --- a/helper/include/odp/helper/udp.h >> +++ b/helper/include/odp/helper/udp.h >> @@ -4,7 +4,6 @@ >> * SPDX-License-Identifier: BSD-3-Clause >> */ >> >> - >> /** >> * @file >> * >> @@ -22,7 +21,6 @@ extern "C" { >> #include >> #include >> >> - >> /** @addtogroup odph_header ODPH HEADER >> * @{ >> */ >> @@ -44,46 +42,39 @@ typedef struct ODP_PACKED { >> * This function uses odp packet to calc checksum >> * >> * @param pkt calculate chksum for pkt >> - * @return checksum value >> + * @return checksum value in CPU endianness >> */ >> static inline uint16_t odph_ipv4_udp_chksum(odp_packet_t pkt) { >> -uint32_t sum = 0; >> -odph_udphdr_t *udph; >> -odph_ipv4hdr_t *iph; >> -uint16_t udplen; >> -uint8_t *buf; >> - >> -if (!odp_packet_l3_offset(pkt)) >> -return 0; >> +odph_ipv4hdr_t *iph; >> +odph_udphdr_t *udph; >> +uint32_tsum; >> +uint16_tudplen, *buf; >> >> -if (!odp_packet_l4_offset(pkt)) >> +if (odp_packet_l4_offset(pkt) == ODP_PACKET_OFFSET_INVALID) >> return 0; >> - >> iph = (odph_ipv4hdr_t *)odp_packet_l3_ptr(pkt, NULL); >> udph = (odph_udphdr_t *)odp_packet_l4_ptr(pkt, NULL); >> -udplen = odp_be_to_cpu_16(udph->length); >> - >> -/* 32-bit sum of all 16-bit words covered by UDP chksum */ >> +/* 32-bit sum of UDP pseudo-header */ >> sum = (iph->src_addr & 0x) + (iph->src_addr >> 16) + >> - (iph->dst_addr & 0x) + (iph->dst_addr >> 16) + >> - (uint16_t)iph->proto + udplen; >> -for (buf = (uint8_t *)udph; udplen > 1; udplen -= 2) { >> -sum += ((*buf << 8) + *(buf + 1)); >> -buf += 2; >> -} >> - >> -/* Fold sum to 16 bits: add carrier to result */ >> -while (sum >> 16) >> -sum = (sum & 0x) + (sum >> 16); >> - >> +(iph->dst_addr & 0x) + (iph->dst_addr >> 16) + >> +odp_be_to_cpu_16(iph->proto) + udph->length; > > This line is also strange. You're using odp_be_to_cpu_16 for 8-bit digit. > At this point, iph->proto firstly will be extended to u16 in CPU endianness > and after that odp_be_to_cpu_16() will return it to BE. You will have right > position of all bytes (8 zero bits + 8 bits of iph->proto), but the way how > this done is wrong. > > [IGR] Here the proto field is added as the LSB in a 16-bit BE value. Do you > think the following code clearer ? > > uint16 val; > ... > val = iph->proto; > sum += odp_be_to_cpu_16(val); No, because 'val' is in CPU endianness and you're using odp_be_to_cpu_16() for variable that is already in CPU endianness. > >> +udplen = odp_be_to_cpu_16(udph->length); >> +buf = (uint16_t *)((void *)udph); >> +/* 32-bit sum of UDP header (checksum field cleared) and UDP data */ >> +for ( ; udplen > 1; udplen -= 2) >> +sum += *buf++; >> +/* Length is not a multiple of 2 bytes */ >> +if (udplen) >> +sum += odp_be_to_cpu_16(*((uint8_t *)buf) << 8); > > Same at this point. You're
Re: [lng-odp] [PATCHv5] helper : Fix UDP checksum computation
See inline comments. -Original Message- From: Ilya Maximets [mailto:i.maxim...@samsung.com] Sent: Wednesday, December 09, 2015 12:04 PM To: Grigore Ion-B17953; lng-odp@lists.linaro.org Subject: Re: [PATCHv5] helper : Fix UDP checksum computation Some comments below. On 09.12.2015 11:59, ion.grig...@freescale.com wrote: > From: Grigore Ion > > This patch fixes the following problems: > - checksum computation for LE platforms > - checksum is computed in the CPU endianness. The returned result must > be converted to the BE ordering when it is used to update the UDP > checksum in a packet. > - checksum computation for packets having the UDP length not a > multiple of 2 > - fixes the UDP checksum associated validation test > > Signed-off-by: Grigore Ion > --- > v5: > - Checksum in CPU endianness fix added (Ilya Maximets) > v4: > - Verify checksum in CPU endianness in the associated test (Ilya > Maximets) > v3: > - fix the UDP checksum computation in the associated test (Maxim > Uvarov) > v2: > - patch updated to the last master (Maxim Uvarov) > v1: > - Move variables declaration on top of block. (Maxim Uvarov) > - Check patch with checkpatch script. (Maxim Uvarov) > - L3 header presence is tested twice. (Alexandru Badicioiu) > - Remove unnecessary check for L3 header presence. (Bill Fischofer) > - Modify check of odp_packet_l4_offset() return. (Bill Fischofer) > > helper/include/odp/helper/udp.h | 55 > + > helper/test/odp_chksum.c| 4 +-- > 2 files changed, 25 insertions(+), 34 deletions(-) > > diff --git a/helper/include/odp/helper/udp.h > b/helper/include/odp/helper/udp.h index 06c439b..9e7256d 100644 > --- a/helper/include/odp/helper/udp.h > +++ b/helper/include/odp/helper/udp.h > @@ -4,7 +4,6 @@ > * SPDX-License-Identifier: BSD-3-Clause > */ > > - > /** > * @file > * > @@ -22,7 +21,6 @@ extern "C" { > #include > #include > > - > /** @addtogroup odph_header ODPH HEADER > * @{ > */ > @@ -44,46 +42,39 @@ typedef struct ODP_PACKED { > * This function uses odp packet to calc checksum > * > * @param pkt calculate chksum for pkt > - * @return checksum value > + * @return checksum value in CPU endianness > */ > static inline uint16_t odph_ipv4_udp_chksum(odp_packet_t pkt) { > - uint32_t sum = 0; > - odph_udphdr_t *udph; > - odph_ipv4hdr_t *iph; > - uint16_t udplen; > - uint8_t *buf; > - > - if (!odp_packet_l3_offset(pkt)) > - return 0; > + odph_ipv4hdr_t *iph; > + odph_udphdr_t *udph; > + uint32_tsum; > + uint16_tudplen, *buf; > > - if (!odp_packet_l4_offset(pkt)) > + if (odp_packet_l4_offset(pkt) == ODP_PACKET_OFFSET_INVALID) > return 0; > - > iph = (odph_ipv4hdr_t *)odp_packet_l3_ptr(pkt, NULL); > udph = (odph_udphdr_t *)odp_packet_l4_ptr(pkt, NULL); > - udplen = odp_be_to_cpu_16(udph->length); > - > - /* 32-bit sum of all 16-bit words covered by UDP chksum */ > + /* 32-bit sum of UDP pseudo-header */ > sum = (iph->src_addr & 0x) + (iph->src_addr >> 16) + > - (iph->dst_addr & 0x) + (iph->dst_addr >> 16) + > - (uint16_t)iph->proto + udplen; > - for (buf = (uint8_t *)udph; udplen > 1; udplen -= 2) { > - sum += ((*buf << 8) + *(buf + 1)); > - buf += 2; > - } > - > - /* Fold sum to 16 bits: add carrier to result */ > - while (sum >> 16) > - sum = (sum & 0x) + (sum >> 16); > - > + (iph->dst_addr & 0x) + (iph->dst_addr >> 16) + > + odp_be_to_cpu_16(iph->proto) + udph->length; This line is also strange. You're using odp_be_to_cpu_16 for 8-bit digit. At this point, iph->proto firstly will be extended to u16 in CPU endianness and after that odp_be_to_cpu_16() will return it to BE. You will have right position of all bytes (8 zero bits + 8 bits of iph->proto), but the way how this done is wrong. [IGR] Here the proto field is added as the LSB in a 16-bit BE value. Do you think the following code clearer ? uint16 val; ... val = iph->proto; sum += odp_be_to_cpu_16(val); > + udplen = odp_be_to_cpu_16(udph->length); > + buf = (uint16_t *)((void *)udph); > + /* 32-bit sum of UDP header (checksum field cleared) and UDP data */ > + for ( ; udplen > 1; udplen -= 2) > + sum += *buf++; > + /* Length is not a multiple of 2 bytes */ > + if (udplen) > + sum += odp_be_to_cpu_16(*((uint8_t *)buf) << 8); Same at this point. You're using odp_be_to_cpu_16() to perform opposite operation. May be, better to use byte arrays here to avoid that strange endianness and compiler aware operations. [IGR] Here the last byte is added as the MSB in a 16-bit BE value. Do you think the following code clearer ?