mbuf: move pool pointer in first half
Subsystems (3)
PATCH 1/1 mbuf: move pool pointer in first half
The mempool pointer in the mbuf struct is moved
from the second to the first half.
It should increase performance on most systems having 64-byte cache line, i.e. mbuf is split in two cache lines.
Due to this change, tx_offload is moved, so some vector data paths may need to be adjusted. Note: OCTEON TX2 check is removed temporarily!
Moving this field gives more space to dynfield1.
This is how the mbuf layout looks like (pahole-style):
word type name byte size
0 void * buf_addr; /* 0 + 8 */
1 rte_iova_t buf_iova /* 8 + 8 */
/* --- RTE_MARKER64 rearm_data; */
2 uint16_t data_off; /* 16 + 2 */
uint16_t refcnt; /* 18 + 2 */
uint16_t nb_segs; /* 20 + 2 */
uint16_t port; /* 22 + 2 */
3 uint64_t ol_flags; /* 24 + 8 */
/* --- RTE_MARKER rx_descriptor_fields1; */
4 uint32_t union packet_type; /* 32 + 4 */
uint32_t pkt_len; /* 36 + 4 */
5 uint16_t data_len; /* 40 + 2 */
uint16_t vlan_tci; /* 42 + 2 */
5.5 uint64_t union hash; /* 44 + 8 */
6.5 uint16_t vlan_tci_outer; /* 52 + 2 */
uint16_t buf_len; /* 54 + 2 */
7 struct rte_mempool * pool; /* 56 + 8 */
/* --- RTE_MARKER cacheline1; */
8 struct rte_mbuf * next; /* 64 + 8 */
9 uint64_t union tx_offload; /* 72 + 8 */
10 struct rte_mbuf_ext_shared_info * shinfo; /* 80 + 8 */
11 uint16_t priv_size; /* 88 + 2 */
uint16_t timesync; /* 90 + 2 */
11.5 uint32_t dynfield1[9]; /* 92 + 36 */16 /* --- END 128 */
--- doc/guides/rel_notes/deprecation.rst | 7 ------- drivers/net/octeontx2/otx2_ethdev.c | 2 -- lib/librte_kni/rte_kni_common.h | 3 ++- lib/librte_mbuf/rte_mbuf.h | 1 - lib/librte_mbuf/rte_mbuf_core.h | 5 ++--- lib/librte_mbuf/rte_mbuf_dyn.c | 1 - 6 files changed, 4 insertions(+), 15 deletions(-)
doc/guides/rel_notes/deprecation.rst +0 -7
@@ -81,13 +81,6 @@ Deprecation Notices us extending existing enum/define. One solution can be using a fixed size array instead of ``.*MAX.*`` value. -* mbuf: Some fields will be converted to dynamic API in DPDK 20.11 - in order to reserve more space for the dynamic fields, as explained in - `this presentation <https://www.youtube.com/watch?v=Ttl6MlhmzWY>`_. - As a consequence, the layout of the ``struct rte_mbuf`` will be re-arranged, - avoiding impact on vectorized implementation of the driver datapaths, - while evaluating performance gains of a better use of the first cache line. - * ethdev: The flow director API, including ``rte_eth_conf.fdir_conf`` field, and the related structures (``rte_fdir_*`` and ``rte_eth_fdir_*``), will be removed in DPDK 20.11.
drivers/net/octeontx2/otx2_ethdev.c +0 -2
@@ -748,8 +748,6 @@ nix_tx_offload_flags(struct rte_eth_dev *eth_dev) offsetof(struct rte_mbuf, buf_iova) + 16);RTE_BUILD_BUG_ON (offsetof(struct rte_mbuf , pkt_len) != offsetof(struct rte_mbuf , ol_flags) +12 ); -RTE_BUILD_BUG_ON (offsetof(struct rte_mbuf , tx_offload) != - offsetof(struct rte_mbuf , pool) +2 *sizeof (void *));if (conf &DEV_TX_OFFLOAD_VLAN_INSERT || conf &DEV_TX_OFFLOAD_QINQ_INSERT )
lib/librte_kni/rte_kni_common.h +2 -1
@@ -84,10 +84,11 @@ struct rte_kni_mbuf {char pad2[4 ];uint32_t pkt_len;/**< Total pkt len: sum of all segment data_len. */ uint16_t data_len;/**< Amount of data in segment buffer. */ +char pad3[14 ]; +void *pool;/* fields on second cache line */ __attribute__((__aligned__ (RTE_CACHE_LINE_MIN_SIZE ))) -void *pool;void *next;/**< Physical address of next mbuf in kernel. */ };
lib/librte_mbuf/rte_mbuf.h +0 -1
@@ -1107,7 +1107,6 @@ rte_pktmbuf_attach_extbuf(struct rte_mbuf *m, void *buf_addr,static inline void rte_mbuf_dynfield_copy (struct rte_mbuf *mdst,const struct rte_mbuf *msrc) { -memcpy (&mdst->dynfield0, msrc->dynfield0,sizeof (mdst->dynfield0));memcpy (&mdst->dynfield1, msrc->dynfield1,sizeof (mdst->dynfield1)); }
lib/librte_mbuf/rte_mbuf_core.h +2 -3
@@ -586,12 +586,11 @@ struct rte_mbuf {uint16_t buf_len;/**< Length of segment buffer. */ -uint64_t dynfield0[1 ];/**< Reserved for dynamic fields. */ +struct rte_mempool *pool;/**< Pool from which mbuf was allocated. */ /* second cache line - fields only used in slow path or on TX */ RTE_MARKER cacheline1 __rte_cache_min_aligned; -struct rte_mempool *pool;/**< Pool from which mbuf was allocated. */ struct rte_mbuf *next;/**< Next segment of scattered packet. */ /* fields to support TX offloads */
@@ -645,7 +644,7 @@ struct rte_mbuf {/** Timesync flags for use with IEEE1588. */ uint16_t timesync; -uint32_t dynfield1[7 ];/**< Reserved for dynamic fields. */ +uint32_t dynfield1[9 ];/**< Reserved for dynamic fields. */ } __rte_cache_aligned;/**
lib/librte_mbuf/rte_mbuf_dyn.c +0 -1
@@ -125,7 +125,6 @@ init_shared_mem(void)* rte_mbuf_dynfield_copy(). */ memset (shm,0 ,sizeof (*shm)); -mark_free (dynfield0);mark_free (dynfield1);/* init free_flags */
Signature · 1 line
2.28.0
Quoted 6 lines · Thomas Monjalon
The mempool pointer in the mbuf struct is moved
from the second to the first half.
It should increase performance on most systems having 64-byte cache line,i.e. mbuf is split in two cache lines.
But In any event, Tx needs to touch the pool to freeing back to the pool upon Tx completion. Right?
Not able to understand the motivation for moving it to the first 64B cache line? The gain varies from driver to driver. For example, a Typical ARM-based NPU does not need to
touch the pool in Rx and its been filled by HW. Whereas it needs to touch in Tx if the reference count is implemented.
Quoted 3 lines
Due to this change,
tx_offloadis moved, so some vector data paths may need to be adjusted. Note: OCTEON TX2 check is removed temporarily!
It will be breaking the Tx path, Please just don't remove the static assert without adjusting the code.
Quoted 144 lines
Moving this field gives more space to dynfield1.
This is how the mbuf layout looks like (pahole-style):
word type name byte size 0 void * buf_addr; /* 0 + 8 */ 1 rte_iova_t buf_iova /* 8 + 8 */ /* --- RTE_MARKER64 rearm_data; */ 2 uint16_t data_off; /* 16 + 2 */ uint16_t refcnt; /* 18 + 2 */ uint16_t nb_segs; /* 20 + 2 */ uint16_t port; /* 22 + 2 */ 3 uint64_t ol_flags; /* 24 + 8 */ /* --- RTE_MARKER rx_descriptor_fields1; */ 4 uint32_t union packet_type; /* 32 + 4 */ uint32_t pkt_len; /* 36 + 4 */ 5 uint16_t data_len; /* 40 + 2 */ uint16_t vlan_tci; /* 42 + 2 */ 5.5 uint64_t union hash; /* 44 + 8 */ 6.5 uint16_t vlan_tci_outer; /* 52 + 2 */ uint16_t buf_len; /* 54 + 2 */ 7 struct rte_mempool * pool; /* 56 + 8 */ /* --- RTE_MARKER cacheline1; */ 8 struct rte_mbuf * next; /* 64 + 8 */ 9 uint64_t union tx_offload; /* 72 + 8 */ 10 struct rte_mbuf_ext_shared_info * shinfo; /* 80 + 8 */ 11 uint16_t priv_size; /* 88 + 2 */ uint16_t timesync; /* 90 + 2 */ 11.5 uint32_t dynfield1[9]; /* 92 + 36 */16 /* --- END 128 */
Signed-off-by Thomas Monjalon <[email protected]>--- doc/guides/rel_notes/deprecation.rst | 7 ------- drivers/net/octeontx2/otx2_ethdev.c | 2 -- lib/librte_kni/rte_kni_common.h | 3 ++- lib/librte_mbuf/rte_mbuf.h | 1 - lib/librte_mbuf/rte_mbuf_core.h | 5 ++--- lib/librte_mbuf/rte_mbuf_dyn.c | 1 - 6 files changed, 4 insertions(+), 15 deletions(-)doc/guides/rel_notes/deprecation.rst +0
-7@@ -81,13 +81,6 @@ Deprecation Notices us extending existing enum/define. One solution can be using a fixed size array instead of ``.*MAX.*`` value. -* mbuf: Some fields will be converted to dynamic API in DPDK 20.11 - in order to reserve more space for the dynamic fields, as explained in - `this presentation <https://www.youtube.com/watch?v=Ttl6MlhmzWY>`_. - As a consequence, the layout of the ``struct rte_mbuf`` will be re-arranged, - avoiding impact on vectorized implementation of the driver datapaths, - while evaluating performance gains of a better use of the first cache line. - * ethdev: The flow director API, including ``rte_eth_conf.fdir_conf`` field, and the related structures (``rte_fdir_*`` and ``rte_eth_fdir_*``), will be removed in DPDK 20.11.drivers/net/octeontx2/otx2_ethdev.c +0
-2@@ -748,8 +748,6 @@ nix_tx_offload_flags(struct rte_eth_dev *eth_dev) offsetof(struct rte_mbuf, buf_iova) + 16);RTE_BUILD_BUG_ON (offsetof(struct rte_mbuf , pkt_len) != offsetof(struct rte_mbuf , ol_flags) +12 ); -RTE_BUILD_BUG_ON (offsetof(struct rte_mbuf , tx_offload) != - offsetof(struct rte_mbuf , pool) +2 *sizeof (void *));if (conf &DEV_TX_OFFLOAD_VLAN_INSERT || conf &DEV_TX_OFFLOAD_QINQ_INSERT )lib/librte_kni/rte_kni_common.h +2
-1@@ -84,10 +84,11 @@ struct rte_kni_mbuf {char pad2[4 ];uint32_t pkt_len;/**< Total pkt len: sum of all segment data_len. */ uint16_t data_len;/**< Amount of data in segment buffer. */ +char pad3[14 ]; +void *pool;/* fields on second cache line */ __attribute__((__aligned__ (RTE_CACHE_LINE_MIN_SIZE ))) -void *pool;void *next;/**< Physical address of next mbuf in kernel. */ };lib/librte_mbuf/rte_mbuf.h +0
-1@@ -1107,7 +1107,6 @@ rte_pktmbuf_attach_extbuf(struct rte_mbuf *m, void *buf_addr,static inline void rte_mbuf_dynfield_copy (struct rte_mbuf *mdst,const struct rte_mbuf *msrc) { -memcpy (&mdst->dynfield0, msrc->dynfield0,sizeof (mdst->dynfield0));memcpy (&mdst->dynfield1, msrc->dynfield1,sizeof (mdst->dynfield1)); }lib/librte_mbuf/rte_mbuf_core.h +2
-3@@ -586,12 +586,11 @@ struct rte_mbuf {uint16_t buf_len;/**< Length of segment buffer. */ -uint64_t dynfield0[1 ];/**< Reserved for dynamic fields. */ +struct rte_mempool *pool;/**< Pool from which mbuf was allocated. */ /* second cache line - fields only used in slow path or on TX */ RTE_MARKER cacheline1 __rte_cache_min_aligned; -struct rte_mempool *pool;/**< Pool from which mbuf was allocated. */ struct rte_mbuf *next;/**< Next segment of scattered packet. */ /* fields to support TX offloads */ @@ -645,7 +644,7 @@ struct rte_mbuf {/** Timesync flags for use with IEEE1588. */ uint16_t timesync; -uint32_t dynfield1[7 ];/**< Reserved for dynamic fields. */ +uint32_t dynfield1[9 ];/**< Reserved for dynamic fields. */ } __rte_cache_aligned;/** lib/librte_mbuf/rte_mbuf_dyn.c +0
-2@@ -125,7 +125,6 @@ init_shared_mem(void)* rte_mbuf_dynfield_copy(). */ memset (shm,0 ,sizeof (*shm)); -mark_free (dynfield0);mark_free (dynfield1);/* init free_flags */ --2.28.0
07/11/2020 18:12, Jerin Jacob:
Quoted 16 lines · Thomas Monjalon
Quoted 6 lines
The mempool pointer in the mbuf struct is moved
from the second to the first half.
It should increase performance on most systems having 64-byte cache line,i.e. mbuf is split in two cache lines.
But In any event, Tx needs to touch the pool to freeing back to the pool upon Tx completion. Right?
Not able to understand the motivation for moving it to the first 64B cache line? The gain varies from driver to driver. For example, a Typical ARM-based NPU does not need to
touch the pool in Rx and its been filled by HW. Whereas it needs to touch in Tx if the reference count is implemented.
Quoted 5 lines
Quoted 2 lines
Due to this change,
tx_offloadis moved, so some vector data paths may need to be adjusted. Note: OCTEON TX2 check is removed temporarily!It will be breaking the Tx path, Please just don't remove the static assert without adjusting the code.
Of course not.
I looked at the vector Tx path of OCTEON TX2,
it's close to be impossible to understand :)
Please help!
Quoted 17 lines · Thomas Monjalon
07/11/2020 18:12, Jerin Jacob:
Quoted 15 lines · Thomas Monjalon
Quoted 6 lines
The mempool pointer in the mbuf struct is moved
from the second to the first half.
It should increase performance on most systems having 64-byte cache line,i.e. mbuf is split in two cache lines.
But In any event, Tx needs to touch the pool to freeing back to the pool upon Tx completion. Right?
Not able to understand the motivation for moving it to the first 64B cache line? The gain varies from driver to driver. For example, a Typical ARM-based NPU does not need to
touch the pool in Rx and its been filled by HW. Whereas it needs to touch in Tx if the reference count is implemented.
See below.
Quoted 11 lines
Quoted 6 lines
Quoted 3 lines
Due to this change,
tx_offloadis moved, so some vector data paths may need to be adjusted. Note: OCTEON TX2 check is removed temporarily!It will be breaking the Tx path, Please just don't remove the static assert without adjusting the code.
Of course not.
I looked at the vector Tx path of OCTEON TX2,
it's close to be impossible to understand :)
Please help!
Off course. Could you check the above section any share the rationale for this change
and where it helps and how much it helps?
Quoted 2 lines
07/11/2020 20:05, Jerin Jacob:
Quoted 27 lines · Thomas Monjalon
Quoted 16 lines
07/11/2020 18:12, Jerin Jacob:
Quoted 15 lines · Thomas Monjalon
Quoted 6 lines
The mempool pointer in the mbuf struct is moved
from the second to the first half.
It should increase performance on most systems having 64-byte cache line,i.e. mbuf is split in two cache lines.
But In any event, Tx needs to touch the pool to freeing back to the pool upon Tx completion. Right?
Not able to understand the motivation for moving it to the first 64B cache line? The gain varies from driver to driver. For example, a Typical ARM-based NPU does not need to
touch the pool in Rx and its been filled by HW. Whereas it needs to touch in Tx if the reference count is implemented.See below.
Quoted 7 lines
Quoted 6 lines
Quoted 3 lines
Due to this change,
tx_offloadis moved, so some vector data paths may need to be adjusted. Note: OCTEON TX2 check is removed temporarily!It will be breaking the Tx path, Please just don't remove the static assert without adjusting the code.
Quoted 8 lines
Quoted 4 lines
Of course not.
I looked at the vector Tx path of OCTEON TX2,
it's close to be impossible to understand :)
Please help!Off course. Could you check the above section any share the rationale for this change
and where it helps and how much it helps?
It has been concluded in the techboard meeting you were part of. I don't understand why we restart this discussion again. I won't have the energy to restart this process myself. If you don't want to apply the techboard decision, then please do the necessary to request another quick decision.
Quoted 38 lines · Thomas Monjalon
07/11/2020 20:05, Jerin Jacob:
Quoted 27 lines · Thomas Monjalon
Quoted 16 lines
07/11/2020 18:12, Jerin Jacob:
Quoted 15 lines · Thomas Monjalon
Quoted 6 lines
The mempool pointer in the mbuf struct is moved
from the second to the first half.
It should increase performance on most systems having 64-byte cache line,i.e. mbuf is split in two cache lines.
But In any event, Tx needs to touch the pool to freeing back to the pool upon Tx completion. Right?
Not able to understand the motivation for moving it to the first 64B cache line? The gain varies from driver to driver. For example, a Typical ARM-based NPU does not need to
touch the pool in Rx and its been filled by HW. Whereas it needs to touch in Tx if the reference count is implemented.See below.
Quoted 7 lines
Quoted 6 lines
Quoted 3 lines
Due to this change,
tx_offloadis moved, so some vector data paths may need to be adjusted. Note: OCTEON TX2 check is removed temporarily!It will be breaking the Tx path, Please just don't remove the static assert without adjusting the code.
Quoted 8 lines
Quoted 4 lines
Of course not.
I looked at the vector Tx path of OCTEON TX2,
it's close to be impossible to understand :)
Please help!Off course. Could you check the above section any share the rationale for this change
and where it helps and how much it helps?
Quoted 5 lines
It has been concluded in the techboard meeting you were part of. I don't understand why we restart this discussion again. I won't have the energy to restart this process myself. If you don't want to apply the techboard decision, then please do the necessary to request another quick decision.
Yes. Initially, I thought it is OK as we have 128B CL, After looking into Thomas's change, I realized
it is not good for ARM64 64B catchlines based NPU as
- A Typical ARM-based NPU does not need to touch the pool in Rx and its been filled by HW. Whereas it needs to
touch in Tx if the reference count is implemented.
- Also it will be effecting exiting vector routines
I request to reconsider the tech board decision.
Quoted 3 lines
09/11/2020 06:18, Jerin Jacob:
Quoted 44 lines · Thomas Monjalon
Quoted 37 lines
07/11/2020 20:05, Jerin Jacob:
Quoted 27 lines · Thomas Monjalon
Quoted 16 lines
07/11/2020 18:12, Jerin Jacob:
Quoted 15 lines · Thomas Monjalon
Quoted 6 lines
The mempool pointer in the mbuf struct is moved
from the second to the first half.
It should increase performance on most systems having 64-byte cache line,i.e. mbuf is split in two cache lines.
But In any event, Tx needs to touch the pool to freeing back to the pool upon Tx completion. Right?
Not able to understand the motivation for moving it to the first 64B cache line? The gain varies from driver to driver. For example, a Typical ARM-based NPU does not need to
touch the pool in Rx and its been filled by HW. Whereas it needs to touch in Tx if the reference count is implemented.See below.
Quoted 7 lines
Quoted 6 lines
Quoted 3 lines
Due to this change,
tx_offloadis moved, so some vector data paths may need to be adjusted. Note: OCTEON TX2 check is removed temporarily!It will be breaking the Tx path, Please just don't remove the static assert without adjusting the code.
Quoted 8 lines
Quoted 4 lines
Of course not.
I looked at the vector Tx path of OCTEON TX2,
it's close to be impossible to understand :)
Please help!Off course. Could you check the above section any share the rationale for this change
and where it helps and how much it helps?Quoted 5 lines
It has been concluded in the techboard meeting you were part of. I don't understand why we restart this discussion again. I won't have the energy to restart this process myself. If you don't want to apply the techboard decision, then please do the necessary to request another quick decision.
Quoted 6 lines
Yes. Initially, I thought it is OK as we have 128B CL, After looking into Thomas's change, I realized
it is not good for ARM64 64B catchlines based NPU as
- A Typical ARM-based NPU does not need to touch the pool in Rx and its been filled by HW. Whereas it needs to
touch in Tx if the reference count is implemented.
Small note: It is not true for all Arm platforms.
Quoted 3 lines
- Also it will be effecting exiting vector routines
I request to reconsider the tech board decision.
Quoted 46 lines · Thomas Monjalon
09/11/2020 06:18, Jerin Jacob:
Quoted 44 lines · Thomas Monjalon
Quoted 42 lines
07/11/2020 20:05, Jerin Jacob:
Quoted 27 lines · Thomas Monjalon
Quoted 16 lines
07/11/2020 18:12, Jerin Jacob:
Quoted 15 lines · Thomas Monjalon
Quoted 6 lines
The mempool pointer in the mbuf struct is moved
from the second to the first half.
It should increase performance on most systems having 64-byte cache line,i.e. mbuf is split in two cache lines.
But In any event, Tx needs to touch the pool to freeing back to the pool upon Tx completion. Right?
Not able to understand the motivation for moving it to the first 64B cache line? The gain varies from driver to driver. For example, a Typical ARM-based NPU does not need to
touch the pool in Rx and its been filled by HW. Whereas it needs to touch in Tx if the reference count is implemented.See below.
Quoted 7 lines
Quoted 6 lines
Quoted 3 lines
Due to this change,
tx_offloadis moved, so some vector data paths may need to be adjusted. Note: OCTEON TX2 check is removed temporarily!It will be breaking the Tx path, Please just don't remove the static assert without adjusting the code.
Quoted 8 lines
Quoted 4 lines
Of course not.
I looked at the vector Tx path of OCTEON TX2,
it's close to be impossible to understand :)
Please help!Off course. Could you check the above section any share the rationale for this change
and where it helps and how much it helps?It has been concluded in the techboard meeting you were part of. I don't understand why we restart this discussion again. I won't have the energy to restart this process myself. If you don't want to apply the techboard decision, then please do the necessary to request another quick decision.
Quoted 8 lines
Quoted 6 lines
Yes. Initially, I thought it is OK as we have 128B CL, After looking into Thomas's change, I realized
it is not good for ARM64 64B catchlines based NPU as
- A Typical ARM-based NPU does not need to touch the pool in Rx and its been filled by HW. Whereas it needs to
touch in Tx if the reference count is implemented.Small note: It is not true for all Arm platforms.
Yes. Generally, it is not specific to Arm platform. Any platform that has HW accelerated mempool.
Quoted 7 lines
Quoted 4 lines
- Also it will be effecting exiting vector routines
I request to reconsider the tech board decision.
Quoted 54 lines · Jerin Jacob
Quoted 53 lines · Thomas Monjalon
09/11/2020 06:18, Jerin Jacob:
Quoted 50 lines · Thomas Monjalon
Quoted 37 lines
07/11/2020 20:05, Jerin Jacob:
Quoted 27 lines · Thomas Monjalon
Quoted 16 lines
07/11/2020 18:12, Jerin Jacob:
Quoted 15 lines · Thomas Monjalon
Quoted 6 lines
The mempool pointer in the mbuf struct is moved
from the second to the first half.
It should increase performance on most systems having 64-byte cache line,i.e. mbuf is split in two cache lines.
But In any event, Tx needs to touch the pool to freeing back to the pool upon Tx completion. Right?
Not able to understand the motivation for moving it to the first 64B cache line? The gain varies from driver to driver. For example, a Typical ARM-based NPU does not need to
touch the pool in Rx and its been filled by HW. Whereas it needs to touch in Tx if the reference count is implemented.See below.
Quoted 7 lines
Quoted 6 lines
Quoted 3 lines
Due to this change,
tx_offloadis moved, so some vector data paths may need to be adjusted. Note: OCTEON TX2 check is removed temporarily!It will be breaking the Tx path, Please just don't remove the static assert without adjusting the code.
Quoted 8 lines
Quoted 4 lines
Of course not.
I looked at the vector Tx path of OCTEON TX2,
it's close to be impossible to understand :)
Please help!Off course. Could you check the above section any share the rationale for this change
and where it helps and how much it helps?Quoted 5 lines
It has been concluded in the techboard meeting you were part of. I don't understand why we restart this discussion again. I won't have the energy to restart this process myself. If you don't want to apply the techboard decision, then please do the necessary to request another quick decision.
Yes. Initially, I thought it is OK as we have 128B CL, After looking into Thomas's change, I realized
it is not good for ARM64 64B catchlines based NPU as
- A Typical ARM-based NPU does not need to touch the pool in Rx and its been filled by HW. Whereas it needs to
touch in Tx if the reference count is implemented.
Quoted 5 lines
Quoted 1 line
Small note: It is not true for all Arm platforms.
Yes. Generally, it is not specific to Arm platform. Any platform that has HW accelerated mempool.
Hi Jerin, Thomas,
For platforms without hw mempool support too, the same holds that the pool pointer should never need to be touched on RX. However, as we discussed on the tech board, moving the pointer may help a little some cases, and especially less optimized drivers, and there seemed to be no downside to it. Jerin, can you please clarify why moving the pool pointer would cause a problem? Is it going to cause things to be slower on some systems for some reasons?
Regards,
/Bruce
Quoted 64 lines · Bruce Richardson
Quoted 54 lines · Jerin Jacob
Quoted 53 lines · Thomas Monjalon
09/11/2020 06:18, Jerin Jacob:
Quoted 50 lines · Thomas Monjalon
Quoted 37 lines
07/11/2020 20:05, Jerin Jacob:
Quoted 27 lines · Thomas Monjalon
Quoted 16 lines
07/11/2020 18:12, Jerin Jacob:
Quoted 15 lines · Thomas Monjalon
Quoted 6 lines
The mempool pointer in the mbuf struct is moved
from the second to the first half.
It should increase performance on most systems having 64-byte cache line,i.e. mbuf is split in two cache lines.
But In any event, Tx needs to touch the pool to freeing back to the pool upon Tx completion. Right?
Not able to understand the motivation for moving it to the first 64B cache line? The gain varies from driver to driver. For example, a Typical ARM-based NPU does not need to
touch the pool in Rx and its been filled by HW. Whereas it needs to touch in Tx if the reference count is implemented.See below.
Quoted 7 lines
Quoted 6 lines
Quoted 3 lines
Due to this change,
tx_offloadis moved, so some vector data paths may need to be adjusted. Note: OCTEON TX2 check is removed temporarily!It will be breaking the Tx path, Please just don't remove the static assert without adjusting the code.
Quoted 8 lines
Quoted 4 lines
Of course not.
I looked at the vector Tx path of OCTEON TX2,
it's close to be impossible to understand :)
Please help!Off course. Could you check the above section any share the rationale for this change
and where it helps and how much it helps?Quoted 5 lines
It has been concluded in the techboard meeting you were part of. I don't understand why we restart this discussion again. I won't have the energy to restart this process myself. If you don't want to apply the techboard decision, then please do the necessary to request another quick decision.
Yes. Initially, I thought it is OK as we have 128B CL, After looking into Thomas's change, I realized
it is not good for ARM64 64B catchlines based NPU as
- A Typical ARM-based NPU does not need to touch the pool in Rx and its been filled by HW. Whereas it needs to
touch in Tx if the reference count is implemented.Quoted 5 lines
Quoted 1 line
Small note: It is not true for all Arm platforms.
Yes. Generally, it is not specific to Arm platform. Any platform that has HW accelerated mempool.
Hi Jerin, Thomas,
Quoted 7 lines
For platforms without hw mempool support too, the same holds that the pool pointer should never need to be touched on RX. However, as we discussed on the tech board, moving the pointer may help a little some cases, and especially less optimized drivers, and there seemed to be no downside to it. Jerin, can you please clarify why moving the pool pointer would cause a problem? Is it going to cause things to be slower on some systems for some reasons?
I overlooked on that part, No specific issue in moving first cache line.
Hi @Thomas Monjalon
Any specific reason why you removed the static assert from octeontx2. I am not able
to compilation issue with that static assert.
The current vector driver assumes pool and tx offload needs to 2 DWORDS apart, Which is the case before and after your change.
Please remove that static assert change, No issue from my side on this patch.
In general, it is too much effort to re-verify and measure performance impact with
all the cases after RC2, I hope this will last mbuf change in this release.
Quoted 3 lines
Regards,
/Bruce
09/11/2020 13:01, Jerin Jacob:
Quoted 3 lines
Hi @Thomas Monjalon
Any specific reason why you removed the static assert from octeontx2.
I have a build failure when cross-compiling for octeontx2.
Quoted 1 line
I am not able to compilation issue with that static assert.
There is no issue when compiling for x86.
Quoted 2 lines
The current vector driver assumes pool and tx offload needs to 2 DWORDS apart, Which is the case before and after your change.
You're right, pool and tx_offload are moved together. The only difference is passing the cache line frontier.
Quoted 1 line
Please remove that static assert change, No issue from my side on this patch.
I cannot remove it without fixing something else, maybe an issue in cache line alignment?
Quoted 3 lines
In general, it is too much effort to re-verify and measure performance impact with
all the cases after RC2, I hope this will last mbuf change in this release.
Yes it is the last change to mbuf layout, sorry for pushing all these changes so late.
Quoted 7 lines · Thomas Monjalon
09/11/2020 13:01, Jerin Jacob:
Quoted 3 lines
Hi @Thomas Monjalon
Any specific reason why you removed the static assert from octeontx2.
I have a build failure when cross-compiling for octeontx2.
I am trying the below command, I am not able to see any issue meson build --cross-file config/arm/arm64_octeontx2_linux_gcc
Are you facing the issue with 32bit? Could you share the steps to reproduce and gcc version?
Quoted 8 lines
Quoted 2 lines
I am not able to compilation issue with that static assert.
There is no issue when compiling for x86.
Quoted 2 lines
The current vector driver assumes pool and tx offload needs to 2 DWORDS apart, Which is the case before and after your change.
Quoted 7 lines
You're right, pool and
tx_offloadare moved together. The only difference is passing the cache line frontier.Quoted 1 line
Please remove that static assert change, No issue from my side on this patch.
I cannot remove it without fixing something else, maybe an issue in cache line alignment?
Can try the below fix. If the issue is seen with 32bit.
(quoted diff) +0 -0
b/drivers/net/octeontx2/otx2_ethdev.c
index 6cebbe677..66a4d429d 100644
drivers/net/octeontx2/otx2_ethdev.c +1 -1
@@ -749,7 +749,7 @@ nix_tx_offload_flags(struct rte_eth_dev *eth_dev)RTE_BUILD_BUG_ON (offsetof(struct rte_mbuf , pkt_len) != offsetof(struct rte_mbuf , ol_flags) +12 );RTE_BUILD_BUG_ON (offsetof(struct rte_mbuf , tx_offload) != - offsetof(struct rte_mbuf , pool) +2 *sizeof (void *)); + offsetof(struct rte_mbuf, pool) + 2 *
sizeof(uint64_t));
Quoted 10 lines
Quoted 4 lines
In general, it is too much effort to re-verify and measure performance impact with
all the cases after RC2, I hope this will last mbuf change in this release.Yes it is the last change to mbuf layout, sorry for pushing all these changes so late.
09/11/2020 14:35, Jerin Jacob:
Quoted 8 lines · Thomas Monjalon
Quoted 7 lines
09/11/2020 13:01, Jerin Jacob:
Quoted 3 lines
Hi @Thomas Monjalon
Any specific reason why you removed the static assert from octeontx2.
I have a build failure when cross-compiling for octeontx2.
I was wrong here.
Quoted 5 lines
I am trying the below command, I am not able to see any issue meson build --cross-file config/arm/arm64_octeontx2_linux_gcc
Are you facing the issue with 32bit? Could you share the steps to reproduce and gcc version?
Oh you're right, the issue was with 32-bit build, sorry for the confusion.
drivers/net/octeontx2/otx2_ethdev.c nix_tx_offload_flags(struct rte_eth_dev *eth_dev) · line 752
Quoted 8 lines
drivers/net/octeontx2/otx2_ethdev.c +1
-1@@ -749,7 +749,7 @@ nix_tx_offload_flags(struct rte_eth_dev *eth_dev)RTE_BUILD_BUG_ON (offsetof(struct rte_mbuf , pkt_len) != offsetof(struct rte_mbuf , ol_flags) +12 );RTE_BUILD_BUG_ON (offsetof(struct rte_mbuf , tx_offload) != - offsetof(struct rte_mbuf , pool) +2 *sizeof (void *)); + offsetof(struct rte_mbuf , pool) +2 *sizeof (uint64_t ));
The actual "fix" isoffsetof(struct rte_mbuf, pool) + sizeof(uint64_t) + sizeof(void *)
I don't understand the octeontx2 vector code. Please check what is the impact of this offset change. BTW, is 32-bit build really supported with octeontx2?
Quoted 13 lines · Thomas Monjalon
09/11/2020 14:35, Jerin Jacob:
Quoted 8 lines · Thomas Monjalon
Quoted 7 lines
09/11/2020 13:01, Jerin Jacob:
Quoted 3 lines
Hi @Thomas Monjalon
Any specific reason why you removed the static assert from octeontx2.
I have a build failure when cross-compiling for octeontx2.
I was wrong here.
Quoted 7 lines
Quoted 5 lines
I am trying the below command, I am not able to see any issue meson build --cross-file config/arm/arm64_octeontx2_linux_gcc
Are you facing the issue with 32bit? Could you share the steps to reproduce and gcc version?
Oh you're right, the issue was with 32-bit build,
Thanks
Quoted 15 lines
sorry for the confusion.
drivers/net/octeontx2/otx2_ethdev.c nix_tx_offload_flags(struct rte_eth_dev *eth_dev) · line 752
Quoted 8 lines
drivers/net/octeontx2/otx2_ethdev.c +1
-1@@ -749,7 +749,7 @@ nix_tx_offload_flags(struct rte_eth_dev *eth_dev)RTE_BUILD_BUG_ON (offsetof(struct rte_mbuf , pkt_len) != offsetof(struct rte_mbuf , ol_flags) +12 );RTE_BUILD_BUG_ON (offsetof(struct rte_mbuf , tx_offload) != - offsetof(struct rte_mbuf , pool) +2 *sizeof (void *)); + offsetof(struct rte_mbuf , pool) +2 *sizeof (uint64_t ));The actual "fix" is
offsetof(struct rte_mbuf, pool)+sizeof(uint64_t)+sizeof(void *)
Quoted 2 lines
I don't understand the octeontx2 vector code. Please check what is the impact of this offset change.
Tested the changes, No issue seen. All the expectation of vector code is expressed with RTE_BUILD_BUG_ON.
Quoted 1 line
BTW, is 32-bit build really supported with octeontx2?
No. I think, keeping assert as "sizeof(void *)"(Same as now) and remove build support for 32bit works too for octeontx2.
We will add it when really required.
Quoted 2 lines
09/11/2020 15:08, Jerin Jacob:
Quoted 27 lines · Thomas Monjalon
Quoted 5 lines
09/11/2020 14:35, Jerin Jacob:
Quoted 2 lines
Are you facing the issue with 32bit? Could you share the steps to reproduce and gcc version?
Oh you're right, the issue was with 32-bit build,
Thanks
Quoted 15 lines
sorry for the confusion.
drivers/net/octeontx2/otx2_ethdev.c nix_tx_offload_flags(struct rte_eth_dev *eth_dev) · line 752
Quoted 8 lines
drivers/net/octeontx2/otx2_ethdev.c +1
-1@@ -749,7 +749,7 @@ nix_tx_offload_flags(struct rte_eth_dev *eth_dev)RTE_BUILD_BUG_ON (offsetof(struct rte_mbuf , pkt_len) != offsetof(struct rte_mbuf , ol_flags) +12 );RTE_BUILD_BUG_ON (offsetof(struct rte_mbuf , tx_offload) != - offsetof(struct rte_mbuf , pool) +2 *sizeof (void *)); + offsetof(struct rte_mbuf , pool) +2 *sizeof (uint64_t ));The actual "fix" is
offsetof(struct rte_mbuf, pool)+sizeof(uint64_t)+sizeof(void *)Quoted 2 lines
I don't understand the octeontx2 vector code. Please check what is the impact of this offset change.
Quoted 7 lines
Tested the changes, No issue seen. All the expectation of vector code is expressed with
RTE_BUILD_BUG_ON.Quoted 1 line
BTW, is 32-bit build really supported with octeontx2?
No. I think, keeping assert as "
sizeof(void *)"(Same as now) and remove build support for 32bit works too for octeontx2.
OK, I think it's better than tweaking RTE_BUILD_BUG_ON for something not really supported.
Quoted 1 line
We will add it when really required.
Then I'm going to send a patch to disable octeontx2 drivers on 32-bit.
Note there is another build issue with octeontx2 drivers on CentOS/RHEL 7 with Arm GGC 4.8.
Quoted 37 lines · Thomas Monjalon
09/11/2020 15:08, Jerin Jacob:
Quoted 27 lines · Thomas Monjalon
Quoted 5 lines
09/11/2020 14:35, Jerin Jacob:
Quoted 2 lines
Are you facing the issue with 32bit? Could you share the steps to reproduce and gcc version?
Oh you're right, the issue was with 32-bit build,
Thanks
Quoted 15 lines
sorry for the confusion.
drivers/net/octeontx2/otx2_ethdev.c nix_tx_offload_flags(struct rte_eth_dev *eth_dev) · line 752
Quoted 8 lines
drivers/net/octeontx2/otx2_ethdev.c +1
-1@@ -749,7 +749,7 @@ nix_tx_offload_flags(struct rte_eth_dev *eth_dev)RTE_BUILD_BUG_ON (offsetof(struct rte_mbuf , pkt_len) != offsetof(struct rte_mbuf , ol_flags) +12 );RTE_BUILD_BUG_ON (offsetof(struct rte_mbuf , tx_offload) != - offsetof(struct rte_mbuf , pool) +2 *sizeof (void *)); + offsetof(struct rte_mbuf , pool) +2 *sizeof (uint64_t ));The actual "fix" is
offsetof(struct rte_mbuf, pool)+sizeof(uint64_t)+sizeof(void *)Quoted 2 lines
I don't understand the octeontx2 vector code. Please check what is the impact of this offset change.
Quoted 7 lines
Tested the changes, No issue seen. All the expectation of vector code is expressed with
RTE_BUILD_BUG_ON.Quoted 1 line
BTW, is 32-bit build really supported with octeontx2?
No. I think, keeping assert as "
sizeof(void *)"(Same as now) and remove build support for 32bit works too for octeontx2.
Quoted 6 lines
OK, I think it's better than tweaking
RTE_BUILD_BUG_ONfor something not really supported.Quoted 1 line
We will add it when really required.
Then I'm going to send a patch to disable octeontx2 drivers on 32-bit.
OK.
Quoted 3 lines
Note there is another build issue with octeontx2 drivers on CentOS/RHEL 7 with Arm GGC 4.8.
It is a segfault from the compiler. In Makefile, we have skipped building < 4.9, not sure what needs to be done
for meson.
Quoted 2 lines
+CC techboard
Quoted 59 lines
From: Jerin Jacob [mailto:[email protected]]
Sent: Monday, November 9, 2020 6:18 AMQuoted 3 lines · Thomas Monjalon
07/11/2020 20:05, Jerin Jacob:
Quoted 1 line
On Sun, Nov 8, 2020 at 12:09 AM Thomas Monjalon
Quoted 2 lines · <[email protected]> wrote:
Quoted 2 lines
Quoted 2 lines
07/11/2020 18:12, Jerin Jacob:
Quoted 1 line
On Sat, Nov 7, 2020 at 10:04 PM Thomas Monjalon
Quoted 4 lines · <[email protected]> wrote:
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
The mempool pointer in the mbuf struct is moved from the second to the first half.
It should increase performance on most systems having 64-bytecache line,
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 2 lines
i.e. mbuf is split in two cache lines.
But In any event, Tx needs to touch the pool to freeing back to
the
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
pool upon Tx completion. Right?
Not able to understand the motivation for moving it to thefirst 64B cache line?
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
The gain varies from driver to driver. For example, a Typical ARM-based NPU does not need to
touch the pool in Rx and its been filled by HW. Whereas itneeds to
Quoted 6 lines
Quoted 6 lines
Quoted 1 line
Quoted 1 line
touch in Tx if the reference count is implemented.
See below.
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Due to this change,
tx_offloadis moved, so some vector datapaths
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
may need to be adjusted. Note: OCTEON TX2 check is removed
temporarily!
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
It will be breaking the Tx path, Please just don't remove the
static
Quoted 8 lines
Quoted 8 lines
Quoted 6 lines
Quoted 1 line
assert without adjusting the code.
Of course not.
I looked at the vector Tx path of OCTEON TX2,
it's close to be impossible to understand :)
Please help!Off course. Could you check the above section any share the
rationale
Quoted 8 lines
Quoted 2 lines
for this change
and where it helps and how much it helps?It has been concluded in the techboard meeting you were part of. I don't understand why we restart this discussion again. I won't have the energy to restart this process myself. If you don't want to apply the techboard decision, then please do the necessary to request another quick decision.
Quoted 6 lines
Yes. Initially, I thought it is OK as we have 128B CL, After looking into Thomas's change, I realized
it is not good for ARM64 64B catchlines based NPU as
- A Typical ARM-based NPU does not need to touch the pool in Rx and its been filled by HW. Whereas it needs to
touch in Tx if the reference count is implemented.
Jerin, I don't understand what the problem is here...
Since RX doesn't touch m->pool, it shouldn't matter for RX which cache line m->pool resides in. I get that.
You are saying that TX needs to touch m->pool if the reference count is implemented. I get that. But I don't understand why it is worse having m->pool in the first cache line than in the second cache line; can you please clarify?
Quoted 1 line
- Also it will be effecting exiting vector routines
That is unavoidable if we move something from the second to the first cache line.
It may require some rework on the vector routines, but it shouldn't be too difficult for whoever wrote these vector routines.
Quoted 2 lines
I request to reconsider the tech board decision.
I was on the techboard meeting as an observer (or whatever the correct term would be for non-members), and this is my impression of the decision on the meeting:
The techboard clearly decided not to move any dynamic fields in the first cache line, on the grounds that if we move them away again in a later version, DPDK users utilizing a dynamic field in the first cache line might experience a performance drop at that later time. And this will be a very bad user experience, causing grief and complaints. To me, this seemed like a firm decision, based on solid arguments.
Then the techboard discussed which other field to move to the freed up space in the first cache line. There were no performance reports showing any improvements by moving the any of the suggested fields (m->pool, m->next, m->tx_offload), and there was a performance report showing no improvements by moving m->next in a test case with large segmented packets. The techboard decided to move m->pool as originally suggested. To me, this seemed like a somewhat random choice between A, B and C, on the grounds that moving one of them is probably better than moving none of them.
The techboard made its decision based on the information available at that time.
Unfortunately, I do not have the resources to test the performance improvement by moving m->next to the first cache line instead of m->pool and utilizing the DEV_TX_OFFLOAD_MBUF_FAST_FREE flag mentioned by Konstantin.
If no new information comes to light, we cannot expect the techboard to change a decision it has already made.
In any case, I am grateful for the joint effort put into nurturing the mbuf, and especially Thomas' unrelenting hard work in this area!
Med venlig hilsen / kind regards
- Morten Brørup
Quoted 81 lines · Morten Brørup
+CC techboard
Quoted 59 lines
From: Jerin Jacob [mailto:[email protected]]
Sent: Monday, November 9, 2020 6:18 AMQuoted 3 lines · Thomas Monjalon
07/11/2020 20:05, Jerin Jacob:
Quoted 1 line
On Sun, Nov 8, 2020 at 12:09 AM Thomas Monjalon
Quoted 2 lines · <[email protected]> wrote:
Quoted 2 lines
Quoted 2 lines
07/11/2020 18:12, Jerin Jacob:
Quoted 1 line
On Sat, Nov 7, 2020 at 10:04 PM Thomas Monjalon
Quoted 4 lines · <[email protected]> wrote:
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
The mempool pointer in the mbuf struct is moved from the second to the first half.
It should increase performance on most systems having 64-bytecache line,
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 2 lines
i.e. mbuf is split in two cache lines.
But In any event, Tx needs to touch the pool to freeing back to
the
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
pool upon Tx completion. Right?
Not able to understand the motivation for moving it to thefirst 64B cache line?
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
The gain varies from driver to driver. For example, a Typical ARM-based NPU does not need to
touch the pool in Rx and its been filled by HW. Whereas itneeds to
Quoted 6 lines
Quoted 6 lines
Quoted 1 line
Quoted 1 line
touch in Tx if the reference count is implemented.
See below.
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Due to this change,
tx_offloadis moved, so some vector datapaths
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
may need to be adjusted. Note: OCTEON TX2 check is removed
temporarily!
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
It will be breaking the Tx path, Please just don't remove the
static
Quoted 8 lines
Quoted 8 lines
Quoted 6 lines
Quoted 1 line
assert without adjusting the code.
Of course not.
I looked at the vector Tx path of OCTEON TX2,
it's close to be impossible to understand :)
Please help!Off course. Could you check the above section any share the
rationale
Quoted 8 lines
Quoted 2 lines
for this change
and where it helps and how much it helps?It has been concluded in the techboard meeting you were part of. I don't understand why we restart this discussion again. I won't have the energy to restart this process myself. If you don't want to apply the techboard decision, then please do the necessary to request another quick decision.
Quoted 6 lines
Yes. Initially, I thought it is OK as we have 128B CL, After looking into Thomas's change, I realized
it is not good for ARM64 64B catchlines based NPU as
- A Typical ARM-based NPU does not need to touch the pool in Rx and its been filled by HW. Whereas it needs to
touch in Tx if the reference count is implemented.Jerin, I don't understand what the problem is here...
Since RX doesn't touch
m->pool, it shouldn't matter for RX which cache linem->poolresides in. I get that.You are saying that TX needs to touch
m->poolif the reference count is implemented. I get that. But I don't understand why it is worse havingm->poolin the first cache line than in the second cache line; can you please clarify?Quoted 1 line
- Also it will be effecting exiting vector routines
That is unavoidable if we move something from the second to the first cache line.
It may require some rework on the vector routines, but it shouldn't be too difficult for whoever wrote these vector routines.
Quoted 1 line
Quoted 8 lines
Quoted 1 line
I request to reconsider the tech board decision.
I was on the techboard meeting as an observer (or whatever the correct term would be for non-members), and this is my impression of the decision on the meeting:
The techboard clearly decided not to move any dynamic fields in the first cache line, on the grounds that if we move them away again in a later version, DPDK users utilizing a dynamic field in the first cache line might experience a performance drop at that later time. And this will be a very bad user experience, causing grief and complaints. To me, this seemed like a firm decision, based on solid arguments.
Then the techboard discussed which other field to move to the freed up space in the first cache line. There were no performance reports showing any improvements by moving the any of the suggested fields (
m->pool,m->next,m->tx_offload), and there was a performance report showing no improvements by movingm->nextin a test case with large segmented packets. The techboard decided to movem->poolas originally suggested. To me, this seemed like a somewhat random choice between A, B and C, on the grounds that moving one of them is probably better than moving none of them.
This largely tallies with what I remember of the discussion too.
I'd also add though that the choice between the next pointer and the pool pointer came down to the fact that the next pointer is only used for chained, multi-segment, packets - which also tend to be larger packets - while the pool pointer is of relevance to all packets, big and small, single and multi-segment.
/Bruce
Quoted 126 lines
From: dev [mailto:[email protected]] On Behalf Of Bruce Richardson
Sent: Monday, November 9, 2020 11:06 AMQuoted 6 lines · Morten Brørup
+CC techboard
Quoted 4 lines
From: Jerin Jacob [mailto:[email protected]]
Sent: Monday, November 9, 2020 6:18 AMOn Sun, Nov 8, 2020 at 2:03 AM Thomas Monjalon
Quoted 4 lines
Quoted 4 lines
Quoted 3 lines · wrote:
07/11/2020 20:05, Jerin Jacob:
Quoted 1 line
On Sun, Nov 8, 2020 at 12:09 AM Thomas Monjalon
Quoted 8 lines
Quoted 8 lines
Quoted 2 lines · <[email protected]> wrote:
Quoted 2 lines
Quoted 2 lines
07/11/2020 18:12, Jerin Jacob:
Quoted 1 line
On Sat, Nov 7, 2020 at 10:04 PM Thomas Monjalon
Quoted 4 lines · <[email protected]> wrote:
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
The mempool pointer in the mbuf struct is moved from the second to the first half.
It should increase performance on most systems having 64-byte
Quoted 5 lines
Quoted 5 lines
cache line,
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 2 lines
i.e. mbuf is split in two cache lines.
But In any event, Tx needs to touch the pool to freeing
back to
Quoted 5 lines
Quoted 5 lines
the
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
pool upon Tx completion. Right?
Not able to understand the motivation for moving it to thefirst 64B cache line?
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
The gain varies from driver to driver. For example, a
Typical
Quoted 9 lines
Quoted 9 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
ARM-based NPU does not need to
touch the pool in Rx and its been filled by HW. Whereas itneeds to
Quoted 6 lines
Quoted 6 lines
Quoted 1 line
Quoted 1 line
touch in Tx if the reference count is implemented.
See below.
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Due to this change,
tx_offloadis moved, so some vectordata
Quoted 2 lines
Quoted 2 lines
paths
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
may need to be adjusted. Note: OCTEON TX2 check is
removed
Quoted 3 lines
Quoted 3 lines
temporarily!
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
It will be breaking the Tx path, Please just don't remove
the
Quoted 13 lines
Quoted 13 lines
static
Quoted 8 lines
Quoted 8 lines
Quoted 6 lines
Quoted 1 line
assert without adjusting the code.
Of course not.
I looked at the vector Tx path of OCTEON TX2,
it's close to be impossible to understand :)
Please help!Off course. Could you check the above section any share the
rationale
Quoted 3 lines
Quoted 2 lines
for this change
and where it helps and how much it helps?Quoted 7 lines
Quoted 7 lines
Quoted 5 lines
It has been concluded in the techboard meeting you were part of. I don't understand why we restart this discussion again. I won't have the energy to restart this process myself. If you don't want to apply the techboard decision, then please do the necessary to request another quick decision.
Yes. Initially, I thought it is OK as we have 128B CL, After
looking
Quoted 3 lines
Quoted 3 lines
into Thomas's change, I realized
it is not good for ARM64 64B catchlines based NPU as
- A Typical ARM-based NPU does not need to touch the pool in Rxand
Quoted 6 lines
Quoted 2 lines
its been filled by HW. Whereas it needs to touch in Tx if the reference count is implemented.
Jerin, I don't understand what the problem is here...
Since RX doesn't touch
m->pool, it shouldn't matter for RX whichcache line
m->poolresides in. I get that.Quoted 2 lines
You are saying that TX needs to touch
m->poolif the reference countis implemented. I get that. But I don't understand why it is worse having
m->poolin the first cache line than in the second cache line; can you please clarify?Quoted 4 lines
Quoted 2 lines
- Also it will be effecting exiting vector routines
That is unavoidable if we move something from the second to the first
cache line.
Quoted 2 lines
It may require some rework on the vector routines, but it shouldn't
be too difficult for whoever wrote these vector routines.
Quoted 5 lines
Quoted 3 lines
I request to reconsider the tech board decision.
I was on the techboard meeting as an observer (or whatever the
correct term would be for non-members), and this is my impression of the decision on the meeting:
Quoted 2 lines
The techboard clearly decided not to move any dynamic fields in the
first cache line, on the grounds that if we move them away again in a later version, DPDK users utilizing a dynamic field in the first cache line might experience a performance drop at that later time. And this will be a very bad user experience, causing grief and complaints. To me, this seemed like a firm decision, based on solid arguments.
Quoted 2 lines
Then the techboard discussed which other field to move to the freed
up space in the first cache line. There were no performance reports showing any improvements by moving the any of the suggested fields (m-
Quoted 1 line
pool,
m->next,m->tx_offload), and there was a performance reportshowing no improvements by moving
m->nextin a test case with large segmented packets. The techboard decided to movem->poolas originally suggested. To me, this seemed like a somewhat random choice between A, B and C, on the grounds that moving one of them is probably better than moving none of them.Quoted 1 line
This largely tallies with what I remember of the discussion too.
Quoted 7 lines
I'd also add though that the choice between the next pointer and the pool
pointer came down to the fact that the next pointer is only used for chained, multi-segment, packets - which also tend to be larger packets -
while the pool pointer is of relevance to all packets, big and small, single and multi-segment.
I wish that was the case, but I am not so sure...
It is true that m->next is NULL for non-segmented packets.
However, m->next is read in the likely path of rte_pktmbuf_prefree_seg(), which is called from e.g. rte_pktmbuf_free_seg(), and again from rte_pktmbuf_free(). And I assume that these functions are quite commonly used.
Not being a PMD developer, I might be wrong about this assumption.
Quoted 2 lines
/Bruce
Quoted 134 lines · On Mon, 9 Nov 2020 11:21:02 +0100 Morten Brørup <[email protected]>…
Quoted 126 lines
From: dev [mailto:[email protected]] On Behalf Of Bruce Richardson
Sent: Monday, November 9, 2020 11:06 AMQuoted 6 lines · Morten Brørup
+CC techboard
Quoted 4 lines
From: Jerin Jacob [mailto:[email protected]]
Sent: Monday, November 9, 2020 6:18 AMOn Sun, Nov 8, 2020 at 2:03 AM Thomas Monjalon
Quoted 4 lines
Quoted 4 lines
Quoted 3 lines · wrote:
07/11/2020 20:05, Jerin Jacob:
Quoted 1 line
On Sun, Nov 8, 2020 at 12:09 AM Thomas Monjalon
Quoted 8 lines
Quoted 8 lines
Quoted 2 lines · <[email protected]> wrote:
Quoted 2 lines
Quoted 2 lines
07/11/2020 18:12, Jerin Jacob:
Quoted 1 line
On Sat, Nov 7, 2020 at 10:04 PM Thomas Monjalon
Quoted 4 lines · <[email protected]> wrote:
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
The mempool pointer in the mbuf struct is moved from the second to the first half.
It should increase performance on most systems having 64-byte
Quoted 5 lines
Quoted 5 lines
cache line,
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 2 lines
i.e. mbuf is split in two cache lines.
But In any event, Tx needs to touch the pool to freeing
back to
Quoted 5 lines
Quoted 5 lines
the
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
pool upon Tx completion. Right?
Not able to understand the motivation for moving it to thefirst 64B cache line?
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
The gain varies from driver to driver. For example, a
Typical
Quoted 9 lines
Quoted 9 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
ARM-based NPU does not need to
touch the pool in Rx and its been filled by HW. Whereas itneeds to
Quoted 6 lines
Quoted 6 lines
Quoted 1 line
Quoted 1 line
touch in Tx if the reference count is implemented.
See below.
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Due to this change,
tx_offloadis moved, so some vectordata
Quoted 2 lines
Quoted 2 lines
paths
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
may need to be adjusted. Note: OCTEON TX2 check is
removed
Quoted 3 lines
Quoted 3 lines
temporarily!
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
It will be breaking the Tx path, Please just don't remove
the
Quoted 13 lines
Quoted 13 lines
static
Quoted 8 lines
Quoted 8 lines
Quoted 6 lines
Quoted 1 line
assert without adjusting the code.
Of course not.
I looked at the vector Tx path of OCTEON TX2,
it's close to be impossible to understand :)
Please help!Off course. Could you check the above section any share the
rationale
Quoted 3 lines
Quoted 2 lines
for this change
and where it helps and how much it helps?Quoted 7 lines
Quoted 7 lines
Quoted 5 lines
It has been concluded in the techboard meeting you were part of. I don't understand why we restart this discussion again. I won't have the energy to restart this process myself. If you don't want to apply the techboard decision, then please do the necessary to request another quick decision.
Yes. Initially, I thought it is OK as we have 128B CL, After
looking
Quoted 3 lines
Quoted 3 lines
into Thomas's change, I realized
it is not good for ARM64 64B catchlines based NPU as
- A Typical ARM-based NPU does not need to touch the pool in Rxand
Quoted 6 lines
Quoted 2 lines
its been filled by HW. Whereas it needs to touch in Tx if the reference count is implemented.
Jerin, I don't understand what the problem is here...
Since RX doesn't touch
m->pool, it shouldn't matter for RX whichcache line
m->poolresides in. I get that.Quoted 2 lines
You are saying that TX needs to touch
m->poolif the reference countis implemented. I get that. But I don't understand why it is worse having
m->poolin the first cache line than in the second cache line; can you please clarify?Quoted 4 lines
Quoted 2 lines
- Also it will be effecting exiting vector routines
That is unavoidable if we move something from the second to the first
cache line.
Quoted 2 lines
It may require some rework on the vector routines, but it shouldn't
be too difficult for whoever wrote these vector routines.
Quoted 5 lines
Quoted 3 lines
I request to reconsider the tech board decision.
I was on the techboard meeting as an observer (or whatever the
correct term would be for non-members), and this is my impression of the decision on the meeting:
Quoted 2 lines
The techboard clearly decided not to move any dynamic fields in the
first cache line, on the grounds that if we move them away again in a later version, DPDK users utilizing a dynamic field in the first cache line might experience a performance drop at that later time. And this will be a very bad user experience, causing grief and complaints. To me, this seemed like a firm decision, based on solid arguments.
Quoted 2 lines
Then the techboard discussed which other field to move to the freed
up space in the first cache line. There were no performance reports showing any improvements by moving the any of the suggested fields (m-
Quoted 1 line
pool,
m->next,m->tx_offload), and there was a performance reportshowing no improvements by moving
m->nextin a test case with large segmented packets. The techboard decided to movem->poolas originally suggested. To me, this seemed like a somewhat random choice between A, B and C, on the grounds that moving one of them is probably better than moving none of them.Quoted 1 line
This largely tallies with what I remember of the discussion too.
Quoted 7 lines
I'd also add though that the choice between the next pointer and the pool
pointer came down to the fact that the next pointer is only used for chained, multi-segment, packets - which also tend to be larger packets -
while the pool pointer is of relevance to all packets, big and small, single and multi-segment.
Quoted 3 lines
I wish that was the case, but I am not so sure...
It is true that
m->nextis NULL for non-segmented packets.
Yes m->next is NULL for non-segmented packets.
Do we need to modify rte_mbuf_check() to enforce these checks?
Quoted 166 lines
From: Stephen Hemminger [mailto:[email protected]]
Sent: Monday, November 9, 2020 7:05 PMQuoted 1 line · On Mon, 9 Nov 2020 11:21:02 +0100 Morten Brørup <[email protected]>…
Quoted 1 line
From: dev [mailto:[email protected]] On Behalf Of Bruce
Richardson
Quoted 14 lines
Quoted 14 lines
Sent: Monday, November 9, 2020 11:06 AM
Quoted 6 lines · Morten Brørup
+CC techboard
Quoted 4 lines
From: Jerin Jacob [mailto:[email protected]]
Sent: Monday, November 9, 2020 6:18 AMOn Sun, Nov 8, 2020 at 2:03 AM Thomas Monjalon
Quoted 4 lines
Quoted 4 lines
Quoted 3 lines · wrote:
07/11/2020 20:05, Jerin Jacob:
Quoted 1 line
On Sun, Nov 8, 2020 at 12:09 AM Thomas Monjalon
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 2 lines · <[email protected]> wrote:
Quoted 2 lines
Quoted 2 lines
07/11/2020 18:12, Jerin Jacob:
Quoted 1 line
On Sat, Nov 7, 2020 at 10:04 PM Thomas Monjalon
Quoted 4 lines · <[email protected]> wrote:
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
The mempool pointer in the mbuf struct is moved from the second to the first half.
It should increase performance on most systems having64-
Quoted 10 lines
Quoted 10 lines
byte
Quoted 5 lines
Quoted 5 lines
cache line,
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 2 lines
i.e. mbuf is split in two cache lines.
But In any event, Tx needs to touch the pool to freeing
back to
Quoted 3 lines
Quoted 3 lines
the
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
pool upon Tx completion. Right?
Not able to understand the motivation for moving it tothe
Quoted 5 lines
Quoted 5 lines
Quoted 2 lines
Quoted 2 lines
first 64B cache line?
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
The gain varies from driver to driver. For example, a
Typical
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
ARM-based NPU does not need to
touch the pool in Rx and its been filled by HW. Whereasit
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
needs to
Quoted 6 lines
Quoted 6 lines
Quoted 1 line
Quoted 1 line
touch in Tx if the reference count is implemented.
See below.
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Due to this change,
tx_offloadis moved, so somevector
Quoted 7 lines
Quoted 7 lines
data
Quoted 2 lines
Quoted 2 lines
paths
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
may need to be adjusted. Note: OCTEON TX2 check is
removed
Quoted 3 lines
Quoted 3 lines
temporarily!
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
It will be breaking the Tx path, Please just don't
remove
Quoted 9 lines
Quoted 9 lines
the
Quoted 8 lines
Quoted 8 lines
static
Quoted 7 lines
Quoted 7 lines
Quoted 6 lines
Quoted 1 line
assert without adjusting the code.
Of course not.
I looked at the vector Tx path of OCTEON TX2,
it's close to be impossible to understand :)
Please help!Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 1 line
Quoted 1 line
Off course. Could you check the above section any share the
rationale
Quoted 4 lines
Quoted 2 lines
for this change
and where it helps and how much it helps?It has been concluded in the techboard meeting you were part
of.
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
I don't understand why we restart this discussion again. I won't have the energy to restart this process myself. If you don't want to apply the techboard decision, then
please
Quoted 7 lines
Quoted 7 lines
Quoted 3 lines
Quoted 3 lines
Quoted 1 line
do the necessary to request another quick decision.
Yes. Initially, I thought it is OK as we have 128B CL, After
looking
Quoted 3 lines
Quoted 3 lines
into Thomas's change, I realized
it is not good for ARM64 64B catchlines based NPU as
- A Typical ARM-based NPU does not need to touch the pool inRx
Quoted 10 lines
Quoted 10 lines
and
Quoted 6 lines
Quoted 2 lines
its been filled by HW. Whereas it needs to touch in Tx if the reference count is implemented.
Jerin, I don't understand what the problem is here...
Since RX doesn't touch
m->pool, it shouldn't matter for RX whichcache line
m->poolresides in. I get that.Quoted 2 lines
You are saying that TX needs to touch
m->poolif the referencecount
Quoted 2 lines
Quoted 2 lines
is implemented. I get that. But I don't understand why it is worse having
m->poolin the first cache line than in the second cacheline;
Quoted 5 lines
Quoted 5 lines
can you please clarify?
Quoted 4 lines
Quoted 2 lines
- Also it will be effecting exiting vector routines
That is unavoidable if we move something from the second to the
first
Quoted 3 lines
Quoted 3 lines
cache line.
Quoted 2 lines
It may require some rework on the vector routines, but it
shouldn't
Quoted 7 lines
Quoted 7 lines
be too difficult for whoever wrote these vector routines.
Quoted 5 lines
Quoted 3 lines
I request to reconsider the tech board decision.
I was on the techboard meeting as an observer (or whatever the
correct term would be for non-members), and this is my impression
of
Quoted 3 lines
Quoted 3 lines
the decision on the meeting:
Quoted 2 lines
The techboard clearly decided not to move any dynamic fields in
the
Quoted 1 line
Quoted 1 line
first cache line, on the grounds that if we move them away again in
a
Quoted 1 line
Quoted 1 line
later version, DPDK users utilizing a dynamic field in the first
cache
Quoted 1 line
Quoted 1 line
line might experience a performance drop at that later time. And
this
Quoted 1 line
Quoted 1 line
will be a very bad user experience, causing grief and complaints.
To
Quoted 3 lines
Quoted 3 lines
me, this seemed like a firm decision, based on solid arguments.
Quoted 2 lines
Then the techboard discussed which other field to move to the
freed
Quoted 2 lines
Quoted 2 lines
up space in the first cache line. There were no performance reports showing any improvements by moving the any of the suggested fields
(m-
Quoted 3 lines
Quoted 3 lines
Quoted 1 line
pool,
m->next,m->tx_offload), and there was a performance reportshowing no improvements by moving
m->nextin a test case with large segmented packets. The techboard decided to movem->poolasoriginally
Quoted 1 line
Quoted 1 line
suggested. To me, this seemed like a somewhat random choice between
A,
Quoted 1 line
Quoted 1 line
B and C, on the grounds that moving one of them is probably better
than
Quoted 6 lines
Quoted 6 lines
moving none of them.
Quoted 1 line
This largely tallies with what I remember of the discussion too.
I'd also add though that the choice between the next pointer and
the
Quoted 2 lines
Quoted 2 lines
pool
pointer came down to the fact that the next pointer is only usedfor
Quoted 1 line
Quoted 1 line
chained, multi-segment, packets - which also tend to be larger
packets
Quoted 2 lines
Quoted 2 lines
-
while the pool pointer is of relevance to all packets, big andsmall,
Quoted 8 lines
Quoted 5 lines
Quoted 1 line
single and multi-segment.
I wish that was the case, but I am not so sure...
It is true that
m->nextis NULL for non-segmented packets.Yes
m->nextis NULL for non-segmented packets.
Do we need to modifyrte_mbuf_check()to enforce these checks?
I went through rte_mbuf_check() in my head, and it already enforces this check:
When called with is_header set, the function proceeds to count the number of segments by following the m->next chain (while m->next is not NULL), and checks that the count matches m->nb_segs. The loop in the function (used when called with is_header set) is implemented to also check that m->nb_segs == 1 when m->next == NULL.
Quoted 7 lines
From: Thomas Monjalon [mailto:[email protected]]
Sent: Saturday, November 7, 2020 4:53 PMThe mempool pointer in the mbuf struct is moved
from the second to the first half.
It should increase performance on most systems having 64-byte cache line, i.e. mbuf is split in two cache lines.
Perhaps with the #define DEV_TX_OFFLOAD_MBUF_FAST_FREE mentioned by Konstantin, it might be better moving m->next instead, at least for applications with the opportunity to set this flag (e.g. applications with only one mbuf pool).
Unfortunately, the information about the DEV_TX_OFFLOAD_MBUF_FAST_FREE flag came to light after the techboard meeting, and I don't have any benchmarks to support this suggestion, so I guess it's hard to change the techboard's decision to move the pool pointer.
Quoted 3 lines
Due to this change,
tx_offloadis moved, so some vector data paths may need to be adjusted. Note: OCTEON TX2 check is removed temporarily!
Regardless if m->pool or m->next is moved, it will not affect any issues related to m->tx_offload moving to a new location in the second cache line.
Quoted 156 lines
Moving this field gives more space to dynfield1.
This is how the mbuf layout looks like (pahole-style):
word type name byte size 0 void * buf_addr; /* 0 + 8 */ 1 rte_iova_t buf_iova /* 8 + 8 */ /* --- RTE_MARKER64 rearm_data; */ 2 uint16_t data_off; /* 16 + 2 */ uint16_t refcnt; /* 18 + 2 */ uint16_t nb_segs; /* 20 + 2 */ uint16_t port; /* 22 + 2 */ 3 uint64_t ol_flags; /* 24 + 8 */ /* --- RTE_MARKER rx_descriptor_fields1; */ 4 uint32_t union packet_type; /* 32 + 4 */ uint32_t pkt_len; /* 36 + 4 */ 5 uint16_t data_len; /* 40 + 2 */ uint16_t vlan_tci; /* 42 + 2 */ 5.5 uint64_t union hash; /* 44 + 8 */ 6.5 uint16_t vlan_tci_outer; /* 52 + 2 */ uint16_t buf_len; /* 54 + 2 */ 7 struct rte_mempool * pool; /* 56 + 8 */ /* --- RTE_MARKER cacheline1; */ 8 struct rte_mbuf * next; /* 64 + 8 */ 9 uint64_t union tx_offload; /* 72 + 8 */ 10 struct rte_mbuf_ext_shared_info * shinfo; /* 80 + 8 */ 11 uint16_t priv_size; /* 88 + 2 */ uint16_t timesync; /* 90 + 2 */ 11.5 uint32_t dynfield1[9]; /* 92 + 36 */16 /* --- END 128 */
Signed-off-by Thomas Monjalon <[email protected]>--- doc/guides/rel_notes/deprecation.rst | 7 ------- drivers/net/octeontx2/otx2_ethdev.c | 2 -- lib/librte_kni/rte_kni_common.h | 3 ++- lib/librte_mbuf/rte_mbuf.h | 1 - lib/librte_mbuf/rte_mbuf_core.h | 5 ++--- lib/librte_mbuf/rte_mbuf_dyn.c | 1 - 6 files changed, 4 insertions(+), 15 deletions(-)doc/guides/rel_notes/deprecation.rst +0
-0@@ -81,13 +81,6 @@ Deprecation Notices us extending existing enum/define. One solution can be using a fixed size array instead of ``.*MAX.*``value.
doc/guides/rel_notes/deprecation.rst +0
-4-* mbuf: Some fields will be converted to dynamic API in DPDK 20.11 - in order to reserve more space for the dynamic fields, as explained in - `this presentation <https://www.youtube.com/watch?v=Ttl6MlhmzWY>`_. - As a consequence, the layout of the ``struct rte_mbuf`` will be re-arranged,
doc/guides/rel_notes/deprecation.rst +0
-2- avoiding impact on vectorized implementation of the driver datapaths, - while evaluating performance gains of a better use of the first cacheline.
-
* ethdev: The flow director API, including ``rte_eth_conf.fdir_conf``
field,
and the related structures (``rte_fdir_*`` and ``rte_eth_fdir_*``), will be removed in DPDK 20.11.drivers/net/octeontx2/otx2_ethdev.c +0
-2@@ -748,8 +748,6 @@ nix_tx_offload_flags(struct rte_eth_dev *eth_dev) offsetof(struct rte_mbuf, buf_iova) + 16);RTE_BUILD_BUG_ON (offsetof(struct rte_mbuf , pkt_len) != offsetof(struct rte_mbuf , ol_flags) +12 ); -RTE_BUILD_BUG_ON (offsetof(struct rte_mbuf , tx_offload) != - offsetof(struct rte_mbuf , pool) +2 *sizeof (void *));if (conf &DEV_TX_OFFLOAD_VLAN_INSERT || conf &DEV_TX_OFFLOAD_QINQ_INSERT )lib/librte_kni/rte_kni_common.h +0
-0@@ -84,10 +84,11 @@ struct rte_kni_mbuf {char pad2[4 ];uint32_t pkt_len;/**< Total pkt len: sum of all segment
data_len. */lib/librte_kni/rte_kni_common.h +2
-1uint16_t data_len;/**< Amount of data in segment buffer. */ +char pad3[14 ]; +void *pool;/* fields on second cache line */ __attribute__((__aligned__ (RTE_CACHE_LINE_MIN_SIZE ))) -void *pool;void *next;/**< Physical address of next mbuf in kernel. */
};lib/librte_mbuf/rte_mbuf.h +0
-1@@ -1107,7 +1107,6 @@ rte_pktmbuf_attach_extbuf(struct rte_mbuf *m, void *buf_addr, static inline voidrte_mbuf_dynfield_copy (struct rte_mbuf *mdst,const struct rte_mbuf *msrc) { -memcpy (&mdst->dynfield0, msrc->dynfield0,sizeof (mdst->dynfield0));memcpy (&mdst->dynfield1, msrc->dynfield1,sizeof (mdst->dynfield1)); }lib/librte_mbuf/rte_mbuf_core.h +2
-3@@ -586,12 +586,11 @@ struct rte_mbuf {uint16_t buf_len;/**< Length of segment buffer. */ -uint64_t dynfield0[1 ];/**< Reserved for dynamic fields. */ +struct rte_mempool *pool;/**< Pool from which mbuf was allocated. */ /* second cache line - fields only used in slow path or on TX */ RTE_MARKER cacheline1 __rte_cache_min_aligned; -struct rte_mempool *pool;/**< Pool from which mbuf was allocated. */ struct rte_mbuf *next;/**< Next segment of scattered packet. */ /* fields to support TX offloads */ @@ -645,7 +644,7 @@ struct rte_mbuf {/** Timesync flags for use with IEEE1588. */ uint16_t timesync; -uint32_t dynfield1[7 ];/**< Reserved for dynamic fields. */ +uint32_t dynfield1[9 ];/**< Reserved for dynamic fields. */ } __rte_cache_aligned;/** lib/librte_mbuf/rte_mbuf_dyn.c +0
-2@@ -125,7 +125,6 @@ init_shared_mem(void)* rte_mbuf_dynfield_copy(). */ memset (shm,0 ,sizeof (*shm)); -mark_free (dynfield0);mark_free (dynfield1);/* init free_flags */ --2.28.0
Quoted 12 lines · Morten Brørup
Quoted 7 lines
From: Thomas Monjalon [mailto:[email protected]]
Sent: Saturday, November 7, 2020 4:53 PMThe mempool pointer in the mbuf struct is moved
from the second to the first half.
It should increase performance on most systems having 64-byte cache line, i.e. mbuf is split in two cache lines.Perhaps with the #define
DEV_TX_OFFLOAD_MBUF_FAST_FREEmentioned by Konstantin, it might be better movingm->nextinstead, at least for applications with the opportunity to set this flag (e.g. applications with only one mbuf pool).Unfortunately, the information about the
DEV_TX_OFFLOAD_MBUF_FAST_FREEflag came to light after the techboard meeting, and I don't have any benchmarks to support this suggestion, so I guess it's hard to change the techboard's decision to move the pool pointer.
The thing with FAST_FREE is that it doesn't just indicate that there is only one mbuf pool, but rather that none of the packets are chained mbufs either. Therefore, the flag applies equally to both next and pool pointers for optimization. [And the flag should perhaps be two flags!]
/Bruce
Quoted 25 lines
From: Bruce Richardson [mailto:[email protected]]
Sent: Monday, November 9, 2020 11:09 AMQuoted 6 lines · Morten Brørup
Quoted 6 lines
From: Thomas Monjalon [mailto:[email protected]]
Sent: Saturday, November 7, 2020 4:53 PMThe mempool pointer in the mbuf struct is moved
from the second to the first half.
It should increase performance on most systems having 64-byte cacheline,
Quoted 3 lines
Quoted 1 line
i.e. mbuf is split in two cache lines.
Perhaps with the #define
DEV_TX_OFFLOAD_MBUF_FAST_FREEmentioned byKonstantin, it might be better moving
m->nextinstead, at least for applications with the opportunity to set this flag (e.g. applications with only one mbuf pool).Quoted 2 lines
Unfortunately, the information about the
DEV_TX_OFFLOAD_MBUF_FAST_FREEflag came to light after the techboard meeting, and I don't have any benchmarks to support this suggestion, so I guess it's hard to change the techboard's decision to move the pool pointer.Quoted 1 line
Quoted 6 lines
The thing with
FAST_FREEis that it doesn't just indicate that there is only one mbuf pool, but rather that none of the packets are chained mbufs
either. Therefore, the flag applies equally to both next and pool pointers
for optimization. [And the flag should perhaps be two flags!]
I guess we could offer the same optimization to applications by adding a rte_mbuf_raw_free_bulk() function to the mbuf library. I will put that on my ToDo-list. :-)
Quoted 14 lines · Morten Brørup
Quoted 9 lines
Quoted 7 lines
From: Thomas Monjalon [mailto:[email protected]]
Sent: Saturday, November 7, 2020 4:53 PMThe mempool pointer in the mbuf struct is moved
from the second to the first half.
It should increase performance on most systems having 64-byte cache line, i.e. mbuf is split in two cache lines.Perhaps with the #define
DEV_TX_OFFLOAD_MBUF_FAST_FREEmentioned by Konstantin, it might be better movingm->nextinstead, atleast for applications with the opportunity to set this flag (e.g. applications with only one mbuf pool).
Quoted 1 line
Quoted 7 lines
Quoted 1 line
Unfortunately, the information about the
DEV_TX_OFFLOAD_MBUF_FAST_FREEflag came to light after the techboard meeting, and Idon't have any benchmarks to support this suggestion, so I guess it's hard to change the techboard's decision to move the pool pointer.
Quoted 1 line
The thing with
FAST_FREEis that it doesn't just indicate that there is only one mbuf pool, but rather that none of the packets are chained mbufs either.
Hmm, wonder where such assumption comes from?
There is nothing about forbidding chained mbufs in rte_ethdev.h right now:
#define DEV_TX_OFFLOAD_MBUF_FAST_FREE 0x00010000
/**< Device supports optimization for fast release of mbufs.
* When set application must guarantee that per-queue all mbufs comes from
* the same mempool and has refcnt = 1.
*/
Quoted 3 lines
Therefore, the flag applies equally to both next and pool pointers for optimization. [And the flag should perhaps be two flags!]
I think what Morten means here:
For FAST_FREE PMD can store pool pointer somewhere in its queue metadata and doesn't need to read it from the mbuf.
So if we move next to first line - mbuf_free code-path for FAST_FREE case can avoid touching second cache line in the mbuf.
Quoted 22 lines · Konstantin
Quoted 14 lines · Morten Brørup
Quoted 9 lines
Quoted 7 lines
From: Thomas Monjalon [mailto:[email protected]]
Sent: Saturday, November 7, 2020 4:53 PMThe mempool pointer in the mbuf struct is moved
from the second to the first half.
It should increase performance on most systems having 64-byte cache line, i.e. mbuf is split in two cache lines.Perhaps with the #define
DEV_TX_OFFLOAD_MBUF_FAST_FREEmentioned by Konstantin, it might be better movingm->nextinstead, atleast for applications with the opportunity to set this flag (e.g. applications with only one mbuf pool).
Quoted 1 line
Quoted 7 lines
Quoted 1 line
Unfortunately, the information about the
DEV_TX_OFFLOAD_MBUF_FAST_FREEflag came to light after the techboard meeting, and Idon't have any benchmarks to support this suggestion, so I guess it's hard to change the techboard's decision to move the pool pointer.
Quoted 1 line
The thing with
FAST_FREEis that it doesn't just indicate that there is only one mbuf pool, but rather that none of the packets are chained mbufs either.
Quoted 8 lines
Hmm, wonder where such assumption comes from?
There is nothing about forbidding chained mbufs inrte_ethdev.hright now:
#defineDEV_TX_OFFLOAD_MBUF_FAST_FREE0x00010000
/**< Device supports optimization for fast release of mbufs.
* When set application must guarantee that per-queue all mbufs comes from
* the same mempool and has refcnt = 1.
*/
I should have checked before mailing. I was mixing up the refcnt = 1 bit with chained mbuf support.
/Bruce
Quoted 31 lines
Quoted 11 lines
Quoted 9 lines
Quoted 7 lines
From: Thomas Monjalon [mailto:[email protected]]
Sent: Saturday, November 7, 2020 4:53 PMThe mempool pointer in the mbuf struct is moved
from the second to the first half.
It should increase performance on most systems having 64-byte cache line, i.e. mbuf is split in two cache lines.Perhaps with the #define
DEV_TX_OFFLOAD_MBUF_FAST_FREEmentioned by Konstantin, it might be better movingm->nextinstead, atleast for applications with the opportunity to set this flag (e.g. applications with only one mbuf pool).
Quoted 1 line
Quoted 7 lines
Quoted 1 line
Unfortunately, the information about the
DEV_TX_OFFLOAD_MBUF_FAST_FREEflag came to light after the techboard meeting, and Idon't have any benchmarks to support this suggestion, so I guess it's hard to change the techboard's decision to move the pool pointer.
Quoted 1 line
The thing with
FAST_FREEis that it doesn't just indicate that there is only one mbuf pool, but rather that none of the packets are chained mbufs either.Hmm, wonder where such assumption comes from?
There is nothing about forbidding chained mbufs inrte_ethdev.hright now:
#defineDEV_TX_OFFLOAD_MBUF_FAST_FREE0x00010000
/**< Device supports optimization for fast release of mbufs.
* When set application must guarantee that per-queue all mbufs comes from
* the same mempool and has refcnt = 1.
*/Quoted 3 lines
Therefore, the flag applies equally to both next and pool pointers for optimization. [And the flag should perhaps be two flags!]
Quoted 5 lines
I think what Morten means here:
ForFAST_FREEPMD can store pool pointer somewhere in its queue metadata and doesn't need to read it from the mbuf.
So if we move next to first line -mbuf_freecode-path forFAST_FREEcase can avoid touching second cache line in the mbuf.
Just as a different thought:
if we move 'next' to the first cache line, can we get rid of 'nb_segs' field completely?
v2 cover move mbuf pool pointer
In reply to Thomas MonjalonAfter some discussions, it seems the best option is still to use the remaining space in the first half to move the pool pointer.
The fields move has an impact on octeontx2 in 32-bit build.
v2: disable octeontx2 drivers in 32-bit build.
Thomas Monjalon (2):
drivers: disable OCTEON TX2 in 32-bit build
mbuf: move pool pointer in first half
doc/guides/rel_notes/deprecation.rst | 7 ------- drivers/common/octeontx2/meson.build | 18 ++++++------------ drivers/crypto/octeontx2/meson.build | 17 +++-------------- drivers/event/octeontx2/meson.build | 18 ++++++------------ drivers/mempool/octeontx2/meson.build | 18 ++++++------------ drivers/net/octeontx2/meson.build | 11 ++++++----- drivers/regex/octeontx2/meson.build | 22 +++------------------- lib/librte_kni/rte_kni_common.h | 3 ++- lib/librte_mbuf/rte_mbuf.h | 1 - lib/librte_mbuf/rte_mbuf_core.h | 5 ++--- lib/librte_mbuf/rte_mbuf_dyn.c | 1 - 11 files changed, 34 insertions(+), 87 deletions(-)
Signature · 1 line
2.28.0
PATCH v2 1/2 drivers: disable OCTEON TX2 in 32-bit build
The drivers for OCTEON TX2 are not supported in 32-bit mode.
--- drivers/common/octeontx2/meson.build | 18 ++++++------------ drivers/crypto/octeontx2/meson.build | 17 +++-------------- drivers/event/octeontx2/meson.build | 18 ++++++------------ drivers/mempool/octeontx2/meson.build | 18 ++++++------------ drivers/net/octeontx2/meson.build | 11 ++++++----- drivers/regex/octeontx2/meson.build | 22 +++------------------- 6 files changed, 30 insertions(+), 74 deletions(-)
drivers/common/octeontx2/meson.build +6 -12
@@ -2,6 +2,12 @@ # Copyright(C) 2019 Marvell International Ltd. # +if not dpdk_conf.get('RTE_ARCH_64') + build = false + reason = 'only supported on 64-bit' + subdir_done() +endif + sources= files('otx2_dev.c', 'otx2_irq.c', 'otx2_mbox.c',
@@ -9,18 +15,6 @@ sources= files('otx2_dev.c', 'otx2_sec_idev.c', ) -extra_flags = [] -# This integrated controller runs only on a arm64 machine, remove 32bit warnings -if not dpdk_conf.get('RTE_ARCH_64') - extra_flags += ['-Wno-int-to-pointer-cast', '-Wno-pointer-to-int-cast'] -endif - -foreach flag: extra_flags - if cc.has_argument(flag) - cflags += flag - endif -endforeach - deps = ['eal', 'pci', 'ethdev', 'kvargs'] includes += include_directories('../../common/octeontx2', '../../mempool/octeontx2', '../../bus/pci')
drivers/crypto/octeontx2/meson.build +3 -14
@@ -1,9 +1,10 @@ # SPDX-License-Identifier: BSD-3-Clause # Copyright (C) 2019 Marvell International Ltd. -if not is_linux +if not is_linux or not dpdk_conf.get('RTE_ARCH_64') build = false - reason = 'only supported on Linux' + reason = 'only supported on 64-bit Linux' + subdir_done() endif deps += ['bus_pci']
@@ -21,18 +22,6 @@ sources = files('otx2_cryptodev.c', 'otx2_cryptodev_ops.c', 'otx2_cryptodev_sec.c') -extra_flags = [] -# This integrated controller runs only on a arm64 machine, remove 32bit warnings -if not dpdk_conf.get('RTE_ARCH_64') - extra_flags += ['-Wno-int-to-pointer-cast', '-Wno-pointer-to-int-cast'] -endif - -foreach flag: extra_flags - if cc.has_argument(flag) - cflags += flag - endif -endforeach - includes += include_directories('../../common/cpt') includes += include_directories('../../common/octeontx2') includes += include_directories('../../crypto/octeontx2')
drivers/event/octeontx2/meson.build +6 -12
@@ -2,6 +2,12 @@ # Copyright(C) 2019 Marvell International Ltd. # +if not dpdk_conf.get('RTE_ARCH_64') + build = false + reason = 'only supported on 64-bit' + subdir_done() +endif + sources = files('otx2_worker.c', 'otx2_worker_dual.c', 'otx2_evdev.c',
@@ -13,18 +19,6 @@ sources = files('otx2_worker.c', 'otx2_tim_worker.c' ) -extra_flags = [] -# This integrated controller runs only on a arm64 machine, remove 32bit warnings -if not dpdk_conf.get('RTE_ARCH_64') - extra_flags += ['-Wno-int-to-pointer-cast', '-Wno-pointer-to-int-cast'] -endif - -foreach flag: extra_flags - if cc.has_argument(flag) - cflags += flag - endif -endforeach - deps += ['bus_pci', 'common_octeontx2', 'crypto_octeontx2', 'mempool_octeontx2', 'net_octeontx2'] includes += include_directories('../../crypto/octeontx2')
drivers/mempool/octeontx2/meson.build +6 -12
@@ -5,6 +5,12 @@ if is_windows build = false reason = 'not supported on Windows' + subdir_done() +endif +if not dpdk_conf.get('RTE_ARCH_64') + build = false + reason = 'only supported on 64-bit' + subdir_done() endif sources = files('otx2_mempool_ops.c',
@@ -13,16 +19,4 @@ sources = files('otx2_mempool_ops.c', 'otx2_mempool_debug.c' ) -extra_flags = [] -# This integrated controller runs only on a arm64 machine, remove 32bit warnings -if not dpdk_conf.get('RTE_ARCH_64') - extra_flags += ['-Wno-int-to-pointer-cast', '-Wno-pointer-to-int-cast'] -endif - -foreach flag: extra_flags - if cc.has_argument(flag) - cflags += flag - endif -endforeach - deps += ['eal', 'mbuf', 'kvargs', 'bus_pci', 'common_octeontx2', 'mempool']
drivers/net/octeontx2/meson.build +6 -5
@@ -2,6 +2,12 @@ # Copyright(C) 2019 Marvell International Ltd. # +if not dpdk_conf.get('RTE_ARCH_64') + build = false + reason = 'only supported on 64-bit' + subdir_done() +endif + sources = files('otx2_rx.c', 'otx2_tx.c', 'otx2_tm.c',
@@ -29,11 +35,6 @@ deps += ['bus_pci', 'cryptodev', 'eventdev', 'security'] deps += ['common_octeontx2', 'mempool_octeontx2'] extra_flags = ['-flax-vector-conversions'] -# This integrated controller runs only on a arm64 machine, remove 32bit warnings -if not dpdk_conf.get('RTE_ARCH_64') - extra_flags += ['-Wno-int-to-pointer-cast', '-Wno-pointer-to-int-cast'] -endif - foreach flag: extra_flags if cc.has_argument(flag) cflags += flag
drivers/regex/octeontx2/meson.build +3 -19
@@ -2,9 +2,10 @@ # Copyright(C) 2020 Marvell International Ltd. # -if not is_linux +if not is_linux or not dpdk_conf.get('RTE_ARCH_64') build = false - reason = 'only supported on Linux' + reason = 'only supported on 64-bit Linux' + subdir_done() endif lib = cc.find_library('librxp_compiler', required: false)
@@ -21,23 +22,6 @@ sources = files('otx2_regexdev.c', 'otx2_regexdev_compiler.c' ) -extra_flags = [] -# This integrated controller runs only on a arm64 machine, remove 32bit warnings -if not dpdk_conf.get('RTE_ARCH_64') - extra_flags += ['-Wno-int-to-pointer-cast', '-Wno-pointer-to-int-cast'] -endif - -# for clang 32-bit compiles we need libatomic for 64-bit atomic ops -if cc.get_id() == 'clang' and dpdk_conf.get('RTE_ARCH_64') == false - ext_deps += cc.find_library('atomic') -endif - -foreach flag: extra_flags - if cc.has_argument(flag) - cflags += flag - endif -endforeach - fmt_name = 'octeontx2_regex' deps += ['bus_pci', 'common_octeontx2', 'regexdev']
Signature · 1 line
2.28.0
Re: [dpdk-dev] [PATCH v2 1/2] drivers: disable OCTEON TX2 in 32-bit build
Quoted 5 lines · Thomas Monjalon
The drivers for OCTEON TX2 are not supported in 32-bit mode.
Suggested-by Jerin Jacob <[email protected]>Signed-off-by Thomas Monjalon <[email protected]>
Quoted 224 lines
--- drivers/common/octeontx2/meson.build | 18 ++++++------------ drivers/crypto/octeontx2/meson.build | 17 +++-------------- drivers/event/octeontx2/meson.build | 18 ++++++------------ drivers/mempool/octeontx2/meson.build | 18 ++++++------------ drivers/net/octeontx2/meson.build | 11 ++++++----- drivers/regex/octeontx2/meson.build | 22 +++------------------- 6 files changed, 30 insertions(+), 74 deletions(-)drivers/common/octeontx2/meson.build +6
-12@@ -2,6 +2,12 @@ # Copyright(C) 2019 Marvell International Ltd. # +if not dpdk_conf.get('RTE_ARCH_64') + build = false + reason = 'only supported on 64-bit' + subdir_done() +endif + sources= files('otx2_dev.c', 'otx2_irq.c', 'otx2_mbox.c',@@ -9,18 +15,6 @@ sources= files('otx2_dev.c', 'otx2_sec_idev.c', ) -extra_flags = [] -# This integrated controller runs only on a arm64 machine, remove 32bit warnings -if not dpdk_conf.get('RTE_ARCH_64') - extra_flags += ['-Wno-int-to-pointer-cast', '-Wno-pointer-to-int-cast'] -endif - -foreach flag: extra_flags - if cc.has_argument(flag) - cflags += flag - endif -endforeach - deps = ['eal', 'pci', 'ethdev', 'kvargs'] includes += include_directories('../../common/octeontx2', '../../mempool/octeontx2', '../../bus/pci')drivers/crypto/octeontx2/meson.build +3
-14@@ -1,9 +1,10 @@ # SPDX-License-Identifier: BSD-3-Clause # Copyright (C) 2019 Marvell International Ltd. -if not is_linux +if not is_linux or not dpdk_conf.get('RTE_ARCH_64') build = false - reason = 'only supported on Linux' + reason = 'only supported on 64-bit Linux' + subdir_done() endif deps += ['bus_pci']@@ -21,18 +22,6 @@ sources = files('otx2_cryptodev.c', 'otx2_cryptodev_ops.c', 'otx2_cryptodev_sec.c') -extra_flags = [] -# This integrated controller runs only on a arm64 machine, remove 32bit warnings -if not dpdk_conf.get('RTE_ARCH_64') - extra_flags += ['-Wno-int-to-pointer-cast', '-Wno-pointer-to-int-cast'] -endif - -foreach flag: extra_flags - if cc.has_argument(flag) - cflags += flag - endif -endforeach - includes += include_directories('../../common/cpt') includes += include_directories('../../common/octeontx2') includes += include_directories('../../crypto/octeontx2')drivers/event/octeontx2/meson.build +6
-12@@ -2,6 +2,12 @@ # Copyright(C) 2019 Marvell International Ltd. # +if not dpdk_conf.get('RTE_ARCH_64') + build = false + reason = 'only supported on 64-bit' + subdir_done() +endif + sources = files('otx2_worker.c', 'otx2_worker_dual.c', 'otx2_evdev.c',@@ -13,18 +19,6 @@ sources = files('otx2_worker.c', 'otx2_tim_worker.c' ) -extra_flags = [] -# This integrated controller runs only on a arm64 machine, remove 32bit warnings -if not dpdk_conf.get('RTE_ARCH_64') - extra_flags += ['-Wno-int-to-pointer-cast', '-Wno-pointer-to-int-cast'] -endif - -foreach flag: extra_flags - if cc.has_argument(flag) - cflags += flag - endif -endforeach - deps += ['bus_pci', 'common_octeontx2', 'crypto_octeontx2', 'mempool_octeontx2', 'net_octeontx2'] includes += include_directories('../../crypto/octeontx2')drivers/mempool/octeontx2/meson.build +6
-12@@ -5,6 +5,12 @@ if is_windows build = false reason = 'not supported on Windows' + subdir_done() +endif +if not dpdk_conf.get('RTE_ARCH_64') + build = false + reason = 'only supported on 64-bit' + subdir_done() endif sources = files('otx2_mempool_ops.c',@@ -13,16 +19,4 @@ sources = files('otx2_mempool_ops.c', 'otx2_mempool_debug.c' ) -extra_flags = [] -# This integrated controller runs only on a arm64 machine, remove 32bit warnings -if not dpdk_conf.get('RTE_ARCH_64') - extra_flags += ['-Wno-int-to-pointer-cast', '-Wno-pointer-to-int-cast'] -endif - -foreach flag: extra_flags - if cc.has_argument(flag) - cflags += flag - endif -endforeach - deps += ['eal', 'mbuf', 'kvargs', 'bus_pci', 'common_octeontx2', 'mempool']drivers/net/octeontx2/meson.build +6
-5@@ -2,6 +2,12 @@ # Copyright(C) 2019 Marvell International Ltd. # +if not dpdk_conf.get('RTE_ARCH_64') + build = false + reason = 'only supported on 64-bit' + subdir_done() +endif + sources = files('otx2_rx.c', 'otx2_tx.c', 'otx2_tm.c',@@ -29,11 +35,6 @@ deps += ['bus_pci', 'cryptodev', 'eventdev', 'security'] deps += ['common_octeontx2', 'mempool_octeontx2'] extra_flags = ['-flax-vector-conversions'] -# This integrated controller runs only on a arm64 machine, remove 32bit warnings -if not dpdk_conf.get('RTE_ARCH_64') - extra_flags += ['-Wno-int-to-pointer-cast', '-Wno-pointer-to-int-cast'] -endif - foreach flag: extra_flags if cc.has_argument(flag) cflags += flagdrivers/regex/octeontx2/meson.build +3
-20@@ -2,9 +2,10 @@ # Copyright(C) 2020 Marvell International Ltd. # -if not is_linux +if not is_linux or not dpdk_conf.get('RTE_ARCH_64') build = false - reason = 'only supported on Linux' + reason = 'only supported on 64-bit Linux' + subdir_done() endif lib = cc.find_library('librxp_compiler', required: false)@@ -21,23 +22,6 @@ sources = files('otx2_regexdev.c', 'otx2_regexdev_compiler.c' ) -extra_flags = [] -# This integrated controller runs only on a arm64 machine, remove 32bit warnings -if not dpdk_conf.get('RTE_ARCH_64') - extra_flags += ['-Wno-int-to-pointer-cast', '-Wno-pointer-to-int-cast'] -endif - -# for clang 32-bit compiles we need libatomic for 64-bit atomic ops -if cc.get_id() == 'clang' and dpdk_conf.get('RTE_ARCH_64') == false - ext_deps += cc.find_library('atomic') -endif - -foreach flag: extra_flags - if cc.has_argument(flag) - cflags += flag - endif -endforeach - fmt_name = 'octeontx2_regex' deps += ['bus_pci', 'common_octeontx2', 'regexdev'] --2.28.0
PATCH v2 2/2 mbuf: move pool pointer in first half
In reply to Thomas MonjalonAccording to the Technical Board decision
(http://mails.dpdk.org/archives/dev/2020-November/191859.html),
the mempool pointer in the mbuf struct is moved
from the second to the first half.
It may increase performance in some cases
on systems having 64-byte cache line, i.e. mbuf split in two cache lines.
Due to this change, tx_offload is moved. Hopefully no vector data path is impacted.
Moving this field gives more space to dynfield1 while dropping the temporary dynfield0.
This is how the mbuf layout looks like (pahole-style):
word type name byte size
0 void * buf_addr; /* 0 + 8 */
1 rte_iova_t buf_iova /* 8 + 8 */
/* --- RTE_MARKER64 rearm_data; */
2 uint16_t data_off; /* 16 + 2 */
uint16_t refcnt; /* 18 + 2 */
uint16_t nb_segs; /* 20 + 2 */
uint16_t port; /* 22 + 2 */
3 uint64_t ol_flags; /* 24 + 8 */
/* --- RTE_MARKER rx_descriptor_fields1; */
4 uint32_t union packet_type; /* 32 + 4 */
uint32_t pkt_len; /* 36 + 4 */
5 uint16_t data_len; /* 40 + 2 */
uint16_t vlan_tci; /* 42 + 2 */
5.5 uint64_t union hash; /* 44 + 8 */
6.5 uint16_t vlan_tci_outer; /* 52 + 2 */
uint16_t buf_len; /* 54 + 2 */
7 struct rte_mempool * pool; /* 56 + 8 */
/* --- RTE_MARKER cacheline1; */
8 struct rte_mbuf * next; /* 64 + 8 */
9 uint64_t union tx_offload; /* 72 + 8 */
10 struct rte_mbuf_ext_shared_info * shinfo; /* 80 + 8 */
11 uint16_t priv_size; /* 88 + 2 */
uint16_t timesync; /* 90 + 2 */
11.5 uint32_t dynfield1[9]; /* 92 + 36 */16 /* --- END 128 */
--- doc/guides/rel_notes/deprecation.rst | 7 ------- lib/librte_kni/rte_kni_common.h | 3 ++- lib/librte_mbuf/rte_mbuf.h | 1 - lib/librte_mbuf/rte_mbuf_core.h | 5 ++--- lib/librte_mbuf/rte_mbuf_dyn.c | 1 - 5 files changed, 4 insertions(+), 13 deletions(-)
doc/guides/rel_notes/deprecation.rst +0 -7
@@ -81,13 +81,6 @@ Deprecation Notices us extending existing enum/define. One solution can be using a fixed size array instead of ``.*MAX.*`` value. -* mbuf: Some fields will be converted to dynamic API in DPDK 20.11 - in order to reserve more space for the dynamic fields, as explained in - `this presentation <https://www.youtube.com/watch?v=Ttl6MlhmzWY>`_. - As a consequence, the layout of the ``struct rte_mbuf`` will be re-arranged, - avoiding impact on vectorized implementation of the driver datapaths, - while evaluating performance gains of a better use of the first cache line. - * ethdev: The flow director API, including ``rte_eth_conf.fdir_conf`` field, and the related structures (``rte_fdir_*`` and ``rte_eth_fdir_*``), will be removed in DPDK 20.11.
lib/librte_kni/rte_kni_common.h +2 -1
@@ -84,10 +84,11 @@ struct rte_kni_mbuf {char pad2[4 ];uint32_t pkt_len;/**< Total pkt len: sum of all segment data_len. */ uint16_t data_len;/**< Amount of data in segment buffer. */ +char pad3[14 ]; +void *pool;/* fields on second cache line */ __attribute__((__aligned__ (RTE_CACHE_LINE_MIN_SIZE ))) -void *pool;void *next;/**< Physical address of next mbuf in kernel. */ };
lib/librte_mbuf/rte_mbuf.h +0 -1
@@ -1107,7 +1107,6 @@ rte_pktmbuf_attach_extbuf(struct rte_mbuf *m, void *buf_addr,static inline void rte_mbuf_dynfield_copy (struct rte_mbuf *mdst,const struct rte_mbuf *msrc) { -memcpy (&mdst->dynfield0, msrc->dynfield0,sizeof (mdst->dynfield0));memcpy (&mdst->dynfield1, msrc->dynfield1,sizeof (mdst->dynfield1)); }
lib/librte_mbuf/rte_mbuf_core.h +2 -3
@@ -586,12 +586,11 @@ struct rte_mbuf {uint16_t buf_len;/**< Length of segment buffer. */ -uint64_t dynfield0[1 ];/**< Reserved for dynamic fields. */ +struct rte_mempool *pool;/**< Pool from which mbuf was allocated. */ /* second cache line - fields only used in slow path or on TX */ RTE_MARKER cacheline1 __rte_cache_min_aligned; -struct rte_mempool *pool;/**< Pool from which mbuf was allocated. */ struct rte_mbuf *next;/**< Next segment of scattered packet. */ /* fields to support TX offloads */
@@ -645,7 +644,7 @@ struct rte_mbuf {/** Timesync flags for use with IEEE1588. */ uint16_t timesync; -uint32_t dynfield1[7 ];/**< Reserved for dynamic fields. */ +uint32_t dynfield1[9 ];/**< Reserved for dynamic fields. */ } __rte_cache_aligned;/**
lib/librte_mbuf/rte_mbuf_dyn.c +0 -1
@@ -125,7 +125,6 @@ init_shared_mem(void)* rte_mbuf_dynfield_copy(). */ memset (shm,0 ,sizeof (*shm)); -mark_free (dynfield0);mark_free (dynfield1);/* init free_flags */
Signature · 1 line
2.28.0
Quoted 11 lines
From: Thomas Monjalon [mailto:[email protected]]
Sent: Monday, November 9, 2020 10:30 PMAccording to the Technical Board decision
(http://mails.dpdk.org/archives/dev/2020-November/191859.html),
the mempool pointer in the mbuf struct is moved
from the second to the first half.
It may increase performance in some cases
on systems having 64-byte cache line, i.e. mbuf split in two cache lines.
Quoted 1 line
Due to this change,
tx_offloadis moved.
A minor correction: All fields after m->pool are moved up 8 bytes, not only m->tx_offload.
Quoted 59 lines
Hopefully no vector data path is impacted.
Moving this field gives more space to dynfield1 while dropping the temporary dynfield0.
This is how the mbuf layout looks like (pahole-style):
word type name byte size 0 void * buf_addr; /* 0 + 8 */ 1 rte_iova_t buf_iova /* 8 + 8 */ /* --- RTE_MARKER64 rearm_data;*/
2uint16_tdata_off; /* 16 + 2
*/
uint16_trefcnt; /* 18 + 2
*/
uint16_tnb_segs; /* 20 + 2
*/
uint16_tport; /* 22 + 2
*/
3uint64_tol_flags; /* 24 + 8
*//* --- RTE_MARKER rx_descriptor_fields1;*/
4uint32_tunionpacket_type; /* 32 + 4
*/
uint32_tpkt_len; /* 36 + 4
*/
5uint16_tdata_len; /* 40 + 2
*/
uint16_tvlan_tci; /* 42 + 2
*/
5.5uint64_tunion hash; /* 44 + 8
*/
6.5uint16_tvlan_tci_outer; /* 52 + 2
*/
uint16_tbuf_len; /* 54 + 2
*/
7 structrte_mempool* pool; /* 56 + 8
*//* --- RTE_MARKER cacheline1;*/
8 structrte_mbuf* next; /* 64 + 8
*/
9uint64_tuniontx_offload; /* 72 + 8
*/
10 structrte_mbuf_ext_shared_info* shinfo; /* 80 + 8
*/
11uint16_tpriv_size; /* 88 + 2
*/
uint16_ttimesync; /* 90 + 2
*/
11.5uint32_tdynfield1[9]; /* 92 + 36
*/
16 /* --- END 128
*/
Quoted 1 line
Signed-off-by Thomas Monjalon <[email protected]>
The minor detail doesn’t prevent...
10/11/2020 11:05, Morten Brørup:
Quoted 11 lines
Quoted 11 lines
From: Thomas Monjalon [mailto:[email protected]]
Sent: Monday, November 9, 2020 10:30 PMAccording to the Technical Board decision
(http://mails.dpdk.org/archives/dev/2020-November/191859.html),
the mempool pointer in the mbuf struct is moved
from the second to the first half.
It may increase performance in some cases
on systems having 64-byte cache line, i.e. mbuf split in two cache lines.
Quoted 3 lines
Quoted 1 line
Due to this change,
tx_offloadis moved.A minor correction: All fields after
m->poolare moved up 8 bytes, not onlym->tx_offload.
Yes I will improve the message before merging, thanks.
Quoted 67 lines
Quoted 60 lines
Hopefully no vector data path is impacted.
Moving this field gives more space to dynfield1 while dropping the temporary dynfield0.
This is how the mbuf layout looks like (pahole-style):
word type name byte size 0 void * buf_addr; /* 0 + 8 */ 1 rte_iova_t buf_iova /* 8 + 8 */ /* --- RTE_MARKER64 rearm_data;*/
2uint16_tdata_off; /* 16 + 2
*/
uint16_trefcnt; /* 18 + 2
*/
uint16_tnb_segs; /* 20 + 2
*/
uint16_tport; /* 22 + 2
*/
3uint64_tol_flags; /* 24 + 8
*//* --- RTE_MARKER rx_descriptor_fields1;*/
4uint32_tunionpacket_type; /* 32 + 4
*/
uint32_tpkt_len; /* 36 + 4
*/
5uint16_tdata_len; /* 40 + 2
*/
uint16_tvlan_tci; /* 42 + 2
*/
5.5uint64_tunion hash; /* 44 + 8
*/
6.5uint16_tvlan_tci_outer; /* 52 + 2
*/
uint16_tbuf_len; /* 54 + 2
*/
7 structrte_mempool* pool; /* 56 + 8
*//* --- RTE_MARKER cacheline1;*/
8 structrte_mbuf* next; /* 64 + 8
*/
9uint64_tuniontx_offload; /* 72 + 8
*/
10 structrte_mbuf_ext_shared_info* shinfo; /* 80 + 8
*/
11uint16_tpriv_size; /* 88 + 2
*/
uint16_ttimesync; /* 90 + 2
*/
11.5uint32_tdynfield1[9]; /* 92 + 36
*/
16 /* --- END 128
*/Quoted 1 line
Signed-off-by Thomas Monjalon <[email protected]>The minor detail doesn’t prevent...
Acked-by Morten Brørup <[email protected]>
Quoted 40 lines · Thomas Monjalon
According to the Technical Board decision
(http://mails.dpdk.org/archives/dev/2020-November/191859.html),
the mempool pointer in the mbuf struct is moved
from the second to the first half.
It may increase performance in some cases
on systems having 64-byte cache line, i.e. mbuf split in two cache lines.Due to this change,
tx_offloadis moved. Hopefully no vector data path is impacted.Moving this field gives more space to dynfield1 while dropping the temporary dynfield0.
This is how the mbuf layout looks like (pahole-style):
word type name byte size 0 void * buf_addr; /* 0 + 8 */ 1 rte_iova_t buf_iova /* 8 + 8 */ /* --- RTE_MARKER64 rearm_data; */ 2 uint16_t data_off; /* 16 + 2 */ uint16_t refcnt; /* 18 + 2 */ uint16_t nb_segs; /* 20 + 2 */ uint16_t port; /* 22 + 2 */ 3 uint64_t ol_flags; /* 24 + 8 */ /* --- RTE_MARKER rx_descriptor_fields1; */ 4 uint32_t union packet_type; /* 32 + 4 */ uint32_t pkt_len; /* 36 + 4 */ 5 uint16_t data_len; /* 40 + 2 */ uint16_t vlan_tci; /* 42 + 2 */ 5.5 uint64_t union hash; /* 44 + 8 */ 6.5 uint16_t vlan_tci_outer; /* 52 + 2 */ uint16_t buf_len; /* 54 + 2 */ 7 struct rte_mempool * pool; /* 56 + 8 */ /* --- RTE_MARKER cacheline1; */ 8 struct rte_mbuf * next; /* 64 + 8 */ 9 uint64_t union tx_offload; /* 72 + 8 */ 10 struct rte_mbuf_ext_shared_info * shinfo; /* 80 + 8 */ 11 uint16_t priv_size; /* 88 + 2 */ uint16_t timesync; /* 90 + 2 */ 11.5 uint32_t dynfield1[9]; /* 92 + 36 */
Quoted 3 lines
16 /* --- END 128 */
Signed-off-by Thomas Monjalon <[email protected]>
Jerin Jacob
Acked-byQuoted 44 lines · Olivier Matz
Quoted 42 lines · Thomas Monjalon
According to the Technical Board decision
(http://mails.dpdk.org/archives/dev/2020-November/191859.html),
the mempool pointer in the mbuf struct is moved
from the second to the first half.
It may increase performance in some cases
on systems having 64-byte cache line, i.e. mbuf split in two cache lines.Due to this change,
tx_offloadis moved. Hopefully no vector data path is impacted.Moving this field gives more space to dynfield1 while dropping the temporary dynfield0.
This is how the mbuf layout looks like (pahole-style):
word type name byte size 0 void * buf_addr; /* 0 + 8 */ 1 rte_iova_t buf_iova /* 8 + 8 */ /* --- RTE_MARKER64 rearm_data; */ 2 uint16_t data_off; /* 16 + 2 */ uint16_t refcnt; /* 18 + 2 */ uint16_t nb_segs; /* 20 + 2 */ uint16_t port; /* 22 + 2 */ 3 uint64_t ol_flags; /* 24 + 8 */ /* --- RTE_MARKER rx_descriptor_fields1; */ 4 uint32_t union packet_type; /* 32 + 4 */ uint32_t pkt_len; /* 36 + 4 */ 5 uint16_t data_len; /* 40 + 2 */ uint16_t vlan_tci; /* 42 + 2 */ 5.5 uint64_t union hash; /* 44 + 8 */ 6.5 uint16_t vlan_tci_outer; /* 52 + 2 */ uint16_t buf_len; /* 54 + 2 */ 7 struct rte_mempool * pool; /* 56 + 8 */ /* --- RTE_MARKER cacheline1; */ 8 struct rte_mbuf * next; /* 64 + 8 */ 9 uint64_t union tx_offload; /* 72 + 8 */ 10 struct rte_mbuf_ext_shared_info * shinfo; /* 80 + 8 */ 11 uint16_t priv_size; /* 88 + 2 */ uint16_t timesync; /* 90 + 2 */ 11.5 uint32_t dynfield1[9]; /* 92 + 36 */16 /* --- END 128 */
Quoted 3 lines
Quoted 1 line
Signed-off-by Thomas Monjalon <[email protected]>Acked-by Olivier Matz <[email protected]>
Quoted 42 lines
Quoted 42 lines
Quoted 42 lines
According to the Technical Board decision
(http://mails.dpdk.org/archives/dev/2020-November/191859.html),
the mempool pointer in the mbuf struct is moved
from the second to the first half.
It may increase performance in some cases
on systems having 64-byte cache line, i.e. mbuf split in two cache lines.Due to this change,
tx_offloadis moved. Hopefully no vector data path is impacted.Moving this field gives more space to dynfield1 while dropping the temporary dynfield0.
This is how the mbuf layout looks like (pahole-style):
word type name byte size 0 void * buf_addr; /* 0 + 8 */ 1 rte_iova_t buf_iova /* 8 + 8 */ /* --- RTE_MARKER64 rearm_data; */ 2 uint16_t data_off; /* 16 + 2 */ uint16_t refcnt; /* 18 + 2 */ uint16_t nb_segs; /* 20 + 2 */ uint16_t port; /* 22 + 2 */ 3 uint64_t ol_flags; /* 24 + 8 */ /* --- RTE_MARKER rx_descriptor_fields1; */ 4 uint32_t union packet_type; /* 32 + 4 */ uint32_t pkt_len; /* 36 + 4 */ 5 uint16_t data_len; /* 40 + 2 */ uint16_t vlan_tci; /* 42 + 2 */ 5.5 uint64_t union hash; /* 44 + 8 */ 6.5 uint16_t vlan_tci_outer; /* 52 + 2 */ uint16_t buf_len; /* 54 + 2 */ 7 struct rte_mempool * pool; /* 56 + 8 */ /* --- RTE_MARKER cacheline1; */ 8 struct rte_mbuf * next; /* 64 + 8 */ 9 uint64_t union tx_offload; /* 72 + 8 */ 10 struct rte_mbuf_ext_shared_info * shinfo; /* 80 + 8 */ 11 uint16_t priv_size; /* 88 + 2 */ uint16_t timesync; /* 90 + 2 */ 11.5 uint32_t dynfield1[9]; /* 92 + 36 */16 /* --- END 128 */
Quoted 5 lines
Quoted 3 lines
Quoted 1 line
Signed-off-by Thomas Monjalon <[email protected]>Acked-by Olivier Matz <[email protected]>Acked-by Jerin Jacob <[email protected]>
Series applied
Stephen Hemminger
Acked-byQuoted 43 lines · On Tue, 10 Nov 2020 17:25:09 +0100 Olivier Matz <[email protected]> wrote:
Quoted 42 lines · Thomas Monjalon
According to the Technical Board decision
(http://mails.dpdk.org/archives/dev/2020-November/191859.html),
the mempool pointer in the mbuf struct is moved
from the second to the first half.
It may increase performance in some cases
on systems having 64-byte cache line, i.e. mbuf split in two cache lines.Due to this change,
tx_offloadis moved. Hopefully no vector data path is impacted.Moving this field gives more space to dynfield1 while dropping the temporary dynfield0.
This is how the mbuf layout looks like (pahole-style):
word type name byte size 0 void * buf_addr; /* 0 + 8 */ 1 rte_iova_t buf_iova /* 8 + 8 */ /* --- RTE_MARKER64 rearm_data; */ 2 uint16_t data_off; /* 16 + 2 */ uint16_t refcnt; /* 18 + 2 */ uint16_t nb_segs; /* 20 + 2 */ uint16_t port; /* 22 + 2 */ 3 uint64_t ol_flags; /* 24 + 8 */ /* --- RTE_MARKER rx_descriptor_fields1; */ 4 uint32_t union packet_type; /* 32 + 4 */ uint32_t pkt_len; /* 36 + 4 */ 5 uint16_t data_len; /* 40 + 2 */ uint16_t vlan_tci; /* 42 + 2 */ 5.5 uint64_t union hash; /* 44 + 8 */ 6.5 uint16_t vlan_tci_outer; /* 52 + 2 */ uint16_t buf_len; /* 54 + 2 */ 7 struct rte_mempool * pool; /* 56 + 8 */ /* --- RTE_MARKER cacheline1; */ 8 struct rte_mbuf * next; /* 64 + 8 */ 9 uint64_t union tx_offload; /* 72 + 8 */ 10 struct rte_mbuf_ext_shared_info * shinfo; /* 80 + 8 */ 11 uint16_t priv_size; /* 88 + 2 */ uint16_t timesync; /* 90 + 2 */ 11.5 uint32_t dynfield1[9]; /* 92 + 36 */16 /* --- END 128 */
Quoted 3 lines
Quoted 1 line
Signed-off-by Thomas Monjalon <[email protected]>Acked-by Olivier Matz <[email protected]>
Thanks this looks good.