net/ark: fix index arithmetic optimization bug
PATCH 1/1 net/ark: fix index arithmetic optimization bug
fix for compiler optimizer error using int32_t. (a - b) > 0 can behave differently under optimization at values near max and min bounds.
This patch replaces int32_t with uint32_t except for necessary casts.
--- drivers/net/ark/ark_ethdev_tx.c | 19 ++++++++++--------- 1 file changed, 10 insertions(+), 9 deletions(-)
drivers/net/ark/ark_ethdev_tx.c +10 -9
@@ -39,8 +39,8 @@ struct ark_tx_queue {uint32_t queue_mask;/* 3 indexes to the paired data rings. */ -int32_t prod_index;/* where to put the next one */ -int32_t free_index;/* mbuf has been freed */ +uint32_t prod_index;/* where to put the next one */ +uint32_t free_index;/* mbuf has been freed */ /* The queue Id is used to identify the HW Q */ uint16_t phys_qid;
@@ -49,7 +49,7 @@ struct ark_tx_queue {/* next cache line - fields written by device */ RTE_MARKER cacheline1 __rte_cache_min_aligned; -volatile int32_t cons_index;/* hw is done, can be freed */ +volatile uint32_t cons_index;/* hw is done, can be freed */ } __rte_cache_aligned;/* Forward declarations */
@@ -108,7 +108,7 @@ eth_ark_xmit_pkts(void *vtxq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)uint32_t user_meta[5 ];int stat; -int32_t prod_index_limit; +uint32_t prod_index_limit;uint16_t nb;uint8_t user_len =0 ;const uint32_t min_pkt_len =ARK_MIN_TX_PKTLEN ;
@@ -124,7 +124,7 @@ eth_ark_xmit_pkts(void *vtxq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts) prod_index_limit = queue->queue_size + queue->free_index -4 ;for (nb =0 ; - (nb < nb_pkts) && (prod_index_limit - queue->prod_index) >0 ; + (nb < nb_pkts) && (int32_t )(prod_index_limit - queue->prod_index) >0 ; ++nb) { mbuf = tx_pkts[nb];
@@ -194,13 +194,13 @@ eth_ark_tx_jumbo(struct ark_tx_queue *queue, struct rte_mbuf *mbuf, uint32_t *user_meta, uint8_t meta_cnt) {struct rte_mbuf *next; -int32_t free_queue_space; +uint32_t free_queue_space;uint8_t flags =ARK_DDM_SOP ; free_queue_space = queue->queue_mask - (queue->prod_index - queue->free_index);/* We need up to 4 mbufs for first header and 2 for subsequent ones */ -if (unlikely (free_queue_space < (2 + (2 * mbuf->nb_segs)))) +if (unlikely (free_queue_space < (2U + (2U * mbuf->nb_segs))))return -1 ;while (mbuf !=NULL ) {
@@ -392,10 +392,11 @@ free_completed_tx(struct ark_tx_queue *queue) {struct rte_mbuf *mbuf;union ark_tx_meta *meta; -int32_t top_index; +uint32_t top_index; top_index = queue->cons_index;/* read once */ -while ((top_index - queue->free_index) >0 ) { + +while ((int32_t )(top_index - queue->free_index) >0 ) { meta = &queue->meta_q[queue->free_index & queue->queue_mask];if (likely ((meta->flags &ARK_DDM_SOP ) !=0 )) { mbuf = queue->bufs[queue->free_index &
Signature · 1 line
2.34.1
PATCH v2 1/1 net/ark: fix index arithmetic optimization bug
fix for compiler optimizer error using int32_t. (a - b) > 0 can behave differently under optimization at values near max and min bounds.
This patch replaces int32_t with uint32_t except for necessary casts.
---
v2:
* update patch to apply to dpdk-next-net
--- drivers/net/ark/ark_ethdev_tx.c | 19 ++++++++++--------- 1 file changed, 10 insertions(+), 9 deletions(-)
drivers/net/ark/ark_ethdev_tx.c +10 -9
@@ -39,8 +39,8 @@ struct __rte_cache_aligned ark_tx_queue {uint32_t queue_mask;/* 3 indexes to the paired data rings. */ -int32_t prod_index;/* where to put the next one */ -int32_t free_index;/* mbuf has been freed */ +uint32_t prod_index;/* where to put the next one */ +uint32_t free_index;/* mbuf has been freed */ /* The queue Id is used to identify the HW Q */ uint16_t phys_qid;
@@ -49,7 +49,7 @@ struct __rte_cache_aligned ark_tx_queue {/* next cache line - fields written by device */ alignas(RTE_CACHE_LINE_MIN_SIZE )RTE_MARKER cacheline1; -volatile int32_t cons_index;/* hw is done, can be freed */ +volatile uint32_t cons_index;/* hw is done, can be freed */ };/* Forward declarations */
@@ -108,7 +108,7 @@ eth_ark_xmit_pkts(void *vtxq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)uint32_t user_meta[5 ];int stat; -int32_t prod_index_limit; +uint32_t prod_index_limit;uint16_t nb;uint8_t user_len =0 ;const uint32_t min_pkt_len =ARK_MIN_TX_PKTLEN ;
@@ -124,7 +124,7 @@ eth_ark_xmit_pkts(void *vtxq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts) prod_index_limit = queue->queue_size + queue->free_index -4 ;for (nb =0 ; - (nb < nb_pkts) && (prod_index_limit - queue->prod_index) >0 ; + (nb < nb_pkts) && (int32_t )(prod_index_limit - queue->prod_index) >0 ; ++nb) { mbuf = tx_pkts[nb];
@@ -194,13 +194,13 @@ eth_ark_tx_jumbo(struct ark_tx_queue *queue, struct rte_mbuf *mbuf, uint32_t *user_meta, uint8_t meta_cnt) {struct rte_mbuf *next; -int32_t free_queue_space; +uint32_t free_queue_space;uint8_t flags =ARK_DDM_SOP ; free_queue_space = queue->queue_mask - (queue->prod_index - queue->free_index);/* We need up to 4 mbufs for first header and 2 for subsequent ones */ -if (unlikely (free_queue_space < (2 + (2 * mbuf->nb_segs)))) +if (unlikely (free_queue_space < (2U + (2U * mbuf->nb_segs))))return -1 ;while (mbuf !=NULL ) {
@@ -392,10 +392,11 @@ free_completed_tx(struct ark_tx_queue *queue) {struct rte_mbuf *mbuf;union ark_tx_meta *meta; -int32_t top_index; +uint32_t top_index; top_index = queue->cons_index;/* read once */ -while ((top_index - queue->free_index) >0 ) { + +while ((int32_t )(top_index - queue->free_index) >0 ) { meta = &queue->meta_q[queue->free_index & queue->queue_mask];if (likely ((meta->flags &ARK_DDM_SOP ) !=0 )) { mbuf = queue->bufs[queue->free_index &
Signature · 1 line
2.34.1
Quoted 4 lines · Ed Czeck
fix for compiler optimizer error using
int32_t. (a - b) > 0 can behave differently under optimization at values near max and min bounds.
Hi Ed,
Is this compiler optimization error, or can it be related to the undefined behavior of signed integer overflow?
Although it is undefined I guess compilers wrap value to INT_MIN.
Just to understand issue better, so that we can apply learning to other code base, can you please share values that create unexpected result?
Quoted 48 lines
This patch replaces
int32_twithuint32_texcept for necessary casts.Fixes 9ee9e0d3b85e ("net/ark: update to reflect FPGA updates")Signed-off-by Ed Czeck <[email protected]>---
v2:
* update patch to apply to dpdk-next-net--- drivers/net/ark/ark_ethdev_tx.c | 19 ++++++++++--------- 1 file changed, 10 insertions(+), 9 deletions(-)drivers/net/ark/ark_ethdev_tx.c +4
-4@@ -39,8 +39,8 @@ struct __rte_cache_aligned ark_tx_queue {uint32_t queue_mask;/* 3 indexes to the paired data rings. */ -int32_t prod_index;/* where to put the next one */ -int32_t free_index;/* mbuf has been freed */ +uint32_t prod_index;/* where to put the next one */ +uint32_t free_index;/* mbuf has been freed */ /* The queue Id is used to identify the HW Q */ uint16_t phys_qid;@@ -49,7 +49,7 @@ struct __rte_cache_aligned ark_tx_queue {/* next cache line - fields written by device */ alignas(RTE_CACHE_LINE_MIN_SIZE )RTE_MARKER cacheline1; -volatile int32_t cons_index;/* hw is done, can be freed */ +volatile uint32_t cons_index;/* hw is done, can be freed */ };/* Forward declarations */ @@ -108,7 +108,7 @@ eth_ark_xmit_pkts(void *vtxq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)uint32_t user_meta[5 ];int stat; -int32_t prod_index_limit; +uint32_t prod_index_limit;uint16_t nb;uint8_t user_len =0 ;const uint32_t min_pkt_len =ARK_MIN_TX_PKTLEN ;
drivers/net/ark/ark_ethdev_tx.c eth_ark_xmit_pkts(void *vtxq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts) · line 128
Quoted 6 lines
drivers/net/ark/ark_ethdev_tx.c +1
-1@@ -124,7 +124,7 @@ eth_ark_xmit_pkts(void *vtxq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts) prod_index_limit = queue->queue_size + queue->free_index -4 ; for (nb = 0; - (nb < nb_pkts) && (prod_index_limit - queue->prod_index) > 0; + (nb < nb_pkts) && (int32_t)(prod_index_limit - queue->prod_index) > 0;
I don't know possible ranges of the values, but in above case if the result of "prod_index_limit - queue->prod_index", is bigger than INT_MAX but smaller than UINT_MAX, although value is positive casting it to 'int' will make is negative and "> 0" check will be unexpected.
Quoted 33 lines
drivers/net/ark/ark_ethdev_tx.c +5
-4++nb) { mbuf = tx_pkts[nb];@@ -194,13 +194,13 @@ eth_ark_tx_jumbo(struct ark_tx_queue *queue, struct rte_mbuf *mbuf, uint32_t *user_meta, uint8_t meta_cnt) {struct rte_mbuf *next; -int32_t free_queue_space; +uint32_t free_queue_space;uint8_t flags =ARK_DDM_SOP ; free_queue_space = queue->queue_mask - (queue->prod_index - queue->free_index);/* We need up to 4 mbufs for first header and 2 for subsequent ones */ -if (unlikely (free_queue_space < (2 + (2 * mbuf->nb_segs)))) +if (unlikely (free_queue_space < (2U + (2U * mbuf->nb_segs))))return -1 ;while (mbuf !=NULL ) {@@ -392,10 +392,11 @@ free_completed_tx(struct ark_tx_queue *queue) {struct rte_mbuf *mbuf;union ark_tx_meta *meta; -int32_t top_index; +uint32_t top_index; top_index = queue->cons_index;/* read once */ -while ((top_index - queue->free_index) >0 ) { + +while ((int32_t )(top_index - queue->free_index) >0 ) { meta = &queue->meta_q[queue->free_index & queue->queue_mask];if (likely ((meta->flags &ARK_DDM_SOP ) !=0 )) { mbuf = queue->bufs[queue->free_index &
PATCH v3 1/1 net/ark: fix index arithmetic bug
In reply to Ed CzeckBehavior for signed integer overflow is not defined which can causes undesired behavior at values near max and min bounds.
The used of unsigned is defined as to use modulo arithmetic which is the desired behavior.
This patch replaces int32_t with uint32_t except for necessary casts.
---
v3:
* Clarify commit message. int_32_t overflow is undefined, while uint32_t is defined as modulo.
* Additional comments in code.
v2:
* update patch to apply to dpdk-next-net
--- drivers/net/ark/ark_ethdev_tx.c | 24 +++++++++++++++--------- 1 file changed, 15 insertions(+), 9 deletions(-)
drivers/net/ark/ark_ethdev_tx.c +15 -9
@@ -39,8 +39,8 @@ struct __rte_cache_aligned ark_tx_queue {uint32_t queue_mask;/* 3 indexes to the paired data rings. */ -int32_t prod_index;/* where to put the next one */ -int32_t free_index;/* mbuf has been freed */ +uint32_t prod_index;/* where to put the next one */ +uint32_t free_index;/* mbuf has been freed */ /* The queue Id is used to identify the HW Q */ uint16_t phys_qid;
@@ -49,7 +49,7 @@ struct __rte_cache_aligned ark_tx_queue {/* next cache line - fields written by device */ alignas(RTE_CACHE_LINE_MIN_SIZE )RTE_MARKER cacheline1; -volatile int32_t cons_index;/* hw is done, can be freed */ +volatile uint32_t cons_index;/* hw is done, can be freed */ };/* Forward declarations */
@@ -108,7 +108,7 @@ eth_ark_xmit_pkts(void *vtxq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)uint32_t user_meta[5 ];int stat; -int32_t prod_index_limit; +uint32_t prod_index_limit;uint16_t nb;uint8_t user_len =0 ;const uint32_t min_pkt_len =ARK_MIN_TX_PKTLEN ;
@@ -123,8 +123,13 @@ eth_ark_xmit_pkts(void *vtxq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)/* leave 4 elements mpu data */ prod_index_limit = queue->queue_size + queue->free_index -4 ; +/* Populate the buffer bringing prod_index upto or slightly beyond +* prod_index_limit. Prod_index will increament by 2 or more each +* iteration. Note: indexes are uint32_t, cast to (signed) int32_t +* to catch the slight overage case; e.g. (200 - 201) +*/ for (nb =0 ; - (nb < nb_pkts) && (prod_index_limit - queue->prod_index) >0 ; + (nb < nb_pkts) && (int32_t )(prod_index_limit - queue->prod_index) >0 ; ++nb) { mbuf = tx_pkts[nb];
@@ -194,13 +199,13 @@ eth_ark_tx_jumbo(struct ark_tx_queue *queue, struct rte_mbuf *mbuf, uint32_t *user_meta, uint8_t meta_cnt) {struct rte_mbuf *next; -int32_t free_queue_space; +uint32_t free_queue_space;uint8_t flags =ARK_DDM_SOP ; free_queue_space = queue->queue_mask - (queue->prod_index - queue->free_index);/* We need up to 4 mbufs for first header and 2 for subsequent ones */ -if (unlikely (free_queue_space < (2 + (2 * mbuf->nb_segs)))) +if (unlikely (free_queue_space < (2U + (2U * mbuf->nb_segs))))return -1 ;while (mbuf !=NULL ) {
@@ -392,10 +397,11 @@ free_completed_tx(struct ark_tx_queue *queue) {struct rte_mbuf *mbuf;union ark_tx_meta *meta; -int32_t top_index; +uint32_t top_index; top_index = queue->cons_index;/* read once */ -while ((top_index - queue->free_index) >0 ) { + +while ((int32_t )(top_index - queue->free_index) >0 ) { meta = &queue->meta_q[queue->free_index & queue->queue_mask];if (likely ((meta->flags &ARK_DDM_SOP ) !=0 )) { mbuf = queue->bufs[queue->free_index &
Signature · 1 line
2.34.1
PATCH v4 1/1 net/ark: fix index arithmetic bug
In reply to Ed CzeckBehavior for signed integer overflow is not defined which can causes undesired behavior at values near max and min bounds.
The used of unsigned is defined as to use modulo arithmetic which is the desired behavior.
This patch replaces int32_t with uint32_t except for necessary casts.
---
v4:
* fix typo in comment
v3:
* Clarify commit message. int_32_t overflow is undefined, while uint32_t is defined as modulo.
* Additional comments in code.
v2:
* update patch to apply to dpdk-next-net
--- drivers/net/ark/ark_ethdev_tx.c | 24 +++++++++++++++--------- 1 file changed, 15 insertions(+), 9 deletions(-)
drivers/net/ark/ark_ethdev_tx.c +15 -9
@@ -39,8 +39,8 @@ struct __rte_cache_aligned ark_tx_queue {uint32_t queue_mask;/* 3 indexes to the paired data rings. */ -int32_t prod_index;/* where to put the next one */ -int32_t free_index;/* mbuf has been freed */ +uint32_t prod_index;/* where to put the next one */ +uint32_t free_index;/* mbuf has been freed */ /* The queue Id is used to identify the HW Q */ uint16_t phys_qid;
@@ -49,7 +49,7 @@ struct __rte_cache_aligned ark_tx_queue {/* next cache line - fields written by device */ alignas(RTE_CACHE_LINE_MIN_SIZE )RTE_MARKER cacheline1; -volatile int32_t cons_index;/* hw is done, can be freed */ +volatile uint32_t cons_index;/* hw is done, can be freed */ };/* Forward declarations */
@@ -108,7 +108,7 @@ eth_ark_xmit_pkts(void *vtxq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)uint32_t user_meta[5 ];int stat; -int32_t prod_index_limit; +uint32_t prod_index_limit;uint16_t nb;uint8_t user_len =0 ;const uint32_t min_pkt_len =ARK_MIN_TX_PKTLEN ;
@@ -123,8 +123,13 @@ eth_ark_xmit_pkts(void *vtxq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)/* leave 4 elements mpu data */ prod_index_limit = queue->queue_size + queue->free_index -4 ; +/* Populate the buffer bringing prod_index up to or slightly beyond +* prod_index_limit. Prod_index will increament by 2 or more each +* iteration. Note: indexes are uint32_t, cast to (signed) int32_t +* to catch the slight overage case; e.g. (200 - 201) +*/ for (nb =0 ; - (nb < nb_pkts) && (prod_index_limit - queue->prod_index) >0 ; + (nb < nb_pkts) && (int32_t )(prod_index_limit - queue->prod_index) >0 ; ++nb) { mbuf = tx_pkts[nb];
@@ -194,13 +199,13 @@ eth_ark_tx_jumbo(struct ark_tx_queue *queue, struct rte_mbuf *mbuf, uint32_t *user_meta, uint8_t meta_cnt) {struct rte_mbuf *next; -int32_t free_queue_space; +uint32_t free_queue_space;uint8_t flags =ARK_DDM_SOP ; free_queue_space = queue->queue_mask - (queue->prod_index - queue->free_index);/* We need up to 4 mbufs for first header and 2 for subsequent ones */ -if (unlikely (free_queue_space < (2 + (2 * mbuf->nb_segs)))) +if (unlikely (free_queue_space < (2U + (2U * mbuf->nb_segs))))return -1 ;while (mbuf !=NULL ) {
@@ -392,10 +397,11 @@ free_completed_tx(struct ark_tx_queue *queue) {struct rte_mbuf *mbuf;union ark_tx_meta *meta; -int32_t top_index; +uint32_t top_index; top_index = queue->cons_index;/* read once */ -while ((top_index - queue->free_index) >0 ) { + +while ((int32_t )(top_index - queue->free_index) >0 ) { meta = &queue->meta_q[queue->free_index & queue->queue_mask];if (likely ((meta->flags &ARK_DDM_SOP ) !=0 )) { mbuf = queue->bufs[queue->free_index &
Signature · 1 line
2.34.1
PATCH v5 1/1 net/ark: fix index arithmetic bug
In reply to Ed CzeckBehavior for signed integer overflow is not defined which can causes undesired behavior at values near max and min bounds.
The used of unsigned is defined as to use modulo arithmetic which is the desired behavior.
This patch replaces int32_t with uint32_t except for necessary casts.
---
v5:
* fix typo in comment
v4:
* fix typo in comment
v3:
* Clarify commit message. int_32_t overflow is undefined, while uint32_t is defined as modulo.
* Additional comments in code.
v2:
* update patch to apply to dpdk-next-net
--- drivers/net/ark/ark_ethdev_tx.c | 24 +++++++++++++++--------- 1 file changed, 15 insertions(+), 9 deletions(-)
drivers/net/ark/ark_ethdev_tx.c +15 -9
@@ -39,8 +39,8 @@ struct __rte_cache_aligned ark_tx_queue {uint32_t queue_mask;/* 3 indexes to the paired data rings. */ -int32_t prod_index;/* where to put the next one */ -int32_t free_index;/* mbuf has been freed */ +uint32_t prod_index;/* where to put the next one */ +uint32_t free_index;/* mbuf has been freed */ /* The queue Id is used to identify the HW Q */ uint16_t phys_qid;
@@ -49,7 +49,7 @@ struct __rte_cache_aligned ark_tx_queue {/* next cache line - fields written by device */ alignas(RTE_CACHE_LINE_MIN_SIZE )RTE_MARKER cacheline1; -volatile int32_t cons_index;/* hw is done, can be freed */ +volatile uint32_t cons_index;/* hw is done, can be freed */ };/* Forward declarations */
@@ -108,7 +108,7 @@ eth_ark_xmit_pkts(void *vtxq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)uint32_t user_meta[5 ];int stat; -int32_t prod_index_limit; +uint32_t prod_index_limit;uint16_t nb;uint8_t user_len =0 ;const uint32_t min_pkt_len =ARK_MIN_TX_PKTLEN ;
@@ -123,8 +123,13 @@ eth_ark_xmit_pkts(void *vtxq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)/* leave 4 elements mpu data */ prod_index_limit = queue->queue_size + queue->free_index -4 ; +/* Populate the buffer bringing prod_index up to or slightly beyond +* prod_index_limit. Prod_index will increment by 2 or more each +* iteration. Note: indexes are uint32_t, cast to (signed) int32_t +* to catch the slight overage case; e.g. (200 - 201) +*/ for (nb =0 ; - (nb < nb_pkts) && (prod_index_limit - queue->prod_index) >0 ; + (nb < nb_pkts) && (int32_t )(prod_index_limit - queue->prod_index) >0 ; ++nb) { mbuf = tx_pkts[nb];
@@ -194,13 +199,13 @@ eth_ark_tx_jumbo(struct ark_tx_queue *queue, struct rte_mbuf *mbuf, uint32_t *user_meta, uint8_t meta_cnt) {struct rte_mbuf *next; -int32_t free_queue_space; +uint32_t free_queue_space;uint8_t flags =ARK_DDM_SOP ; free_queue_space = queue->queue_mask - (queue->prod_index - queue->free_index);/* We need up to 4 mbufs for first header and 2 for subsequent ones */ -if (unlikely (free_queue_space < (2 + (2 * mbuf->nb_segs)))) +if (unlikely (free_queue_space < (2U + (2U * mbuf->nb_segs))))return -1 ;while (mbuf !=NULL ) {
@@ -392,10 +397,11 @@ free_completed_tx(struct ark_tx_queue *queue) {struct rte_mbuf *mbuf;union ark_tx_meta *meta; -int32_t top_index; +uint32_t top_index; top_index = queue->cons_index;/* read once */ -while ((top_index - queue->free_index) >0 ) { + +while ((int32_t )(top_index - queue->free_index) >0 ) { meta = &queue->meta_q[queue->free_index & queue->queue_mask];if (likely ((meta->flags &ARK_DDM_SOP ) !=0 )) { mbuf = queue->bufs[queue->free_index &
Signature · 1 line
2.34.1
Re: [PATCH v5] net/ark: fix index arithmetic bug
Quoted 8 lines · Ed Czeck
Behavior for signed integer overflow is not defined which can causes undesired behavior at values near max and min bounds.
The used of unsigned is defined as to use modulo arithmetic which is the desired behavior.
This patch replacesint32_twithuint32_texcept for necessary casts.
Quoted 5 lines
Fixes 9ee9e0d3b85e ("net/ark: update to reflect FPGA updates")Signed-off-by Ed Czeck <[email protected]>
Applied to dpdk-next-net/main, thanks.