From: Manish Kurup <manish.kurup@broadcom.com>
To: dev@dpdk.org
Cc: ajit.khaparde@broadcom.com, Farah Smith <farah.smith@broadcom.com>
Subject: [PATCH 39/54] net/bnxt: thor2 truflow memory manager bug
Date: Mon, 29 Sep 2025 20:35:49 -0400 [thread overview]
Message-ID: <20250930003604.87108-40-manish.kurup@broadcom.com> (raw)
In-Reply-To: <20250930003604.87108-1-manish.kurup@broadcom.com>
From: Farah Smith <farah.smith@broadcom.com>
A performance optimization was made based upon the belief that blocks
in the memory manager always had non-full blocks prior to full blocks.
But this is not the case. The performance optimization stopped looking
for space in the block list as soon as the first full block was found
in the list. This fix adds code so that whenever a block is full, it is
moved to the end of the list so that the search can stop upon reaching
the first full block. Remove current_blk_idx. Adjust max_records based
upon max contiguous.
Signed-off-by: Farah Smith <farah.smith@broadcom.com>
Reviewed-by: Manish Kurup <manish.kurup@broadcom.com>
---
drivers/net/bnxt/hcapi/cfa_v3/mm/cfa_mm.c | 83 ++++++++++++-------
.../net/bnxt/hcapi/cfa_v3/mm/cfa_mm_priv.h | 2 +-
2 files changed, 56 insertions(+), 29 deletions(-)
diff --git a/drivers/net/bnxt/hcapi/cfa_v3/mm/cfa_mm.c b/drivers/net/bnxt/hcapi/cfa_v3/mm/cfa_mm.c
index 05528dd3e4..6e21d513ac 100644
--- a/drivers/net/bnxt/hcapi/cfa_v3/mm/cfa_mm.c
+++ b/drivers/net/bnxt/hcapi/cfa_v3/mm/cfa_mm.c
@@ -48,6 +48,10 @@ int cfa_mm_query(struct cfa_mm_query_parms *parms)
max_records = parms->max_records;
max_contig_records = (uint16_t)parms->max_contig_records;
+ /* Align to max_contig_records */
+ max_records = (max_records + (max_contig_records - 1)) &
+ ~(max_contig_records - 1);
+
if (unlikely(!(CFA_CHECK_BOUNDS(max_records, 1, CFA_MM_MAX_RECORDS) &&
IS_POWER_2(max_contig_records) &&
CFA_CHECK_BOUNDS(max_contig_records, 1,
@@ -79,6 +83,10 @@ int cfa_mm_open(void *cmm, struct cfa_mm_open_parms *parms)
max_records = parms->max_records;
max_contig_records = (uint16_t)parms->max_contig_records;
+ /* Align to max_contig_records */
+ max_records = (max_records + (max_contig_records - 1)) &
+ ~(max_contig_records - 1);
+
if (unlikely(!(CFA_CHECK_BOUNDS(max_records, 1, CFA_MM_MAX_RECORDS) &&
IS_POWER_2(max_contig_records) &&
CFA_CHECK_BOUNDS(max_contig_records, 1,
@@ -115,11 +123,11 @@ int cfa_mm_open(void *cmm, struct cfa_mm_open_parms *parms)
context->blk_bmap_tbl = (uint8_t *)(context->blk_tbl + num_blocks);
context->blk_list_tbl[0].first_blk_idx = 0;
- context->blk_list_tbl[0].current_blk_idx = 0;
+ context->blk_list_tbl[0].last_blk_idx = 0;
for (i = 1; i < num_lists; i++) {
context->blk_list_tbl[i].first_blk_idx = CFA_MM_INVALID32;
- context->blk_list_tbl[i].current_blk_idx = CFA_MM_INVALID32;
+ context->blk_list_tbl[i].last_blk_idx = CFA_MM_INVALID32;
}
for (i = 0; i < num_blocks; i++) {
@@ -162,6 +170,7 @@ int cfa_mm_close(void *cmm)
return 0;
}
+/* Allocate a block idx from the free list */
static uint32_t cfa_mm_blk_alloc(struct cfa_mm *context)
{
uint32_t blk_idx;
@@ -179,8 +188,6 @@ static uint32_t cfa_mm_blk_alloc(struct cfa_mm *context)
free_list->first_blk_idx =
context->blk_tbl[free_list->first_blk_idx].next_blk_idx;
- free_list->current_blk_idx = free_list->first_blk_idx;
-
if (free_list->first_blk_idx != CFA_MM_INVALID32) {
context->blk_tbl[free_list->first_blk_idx].prev_blk_idx =
CFA_MM_INVALID32;
@@ -192,6 +199,7 @@ static uint32_t cfa_mm_blk_alloc(struct cfa_mm *context)
return blk_idx;
}
+/* Return a block index to the free list */
static void cfa_mm_blk_free(struct cfa_mm *context, uint32_t blk_idx)
{
struct cfa_mm_blk_list *free_list = context->blk_list_tbl;
@@ -208,16 +216,17 @@ static void cfa_mm_blk_free(struct cfa_mm *context, uint32_t blk_idx)
}
free_list->first_blk_idx = blk_idx;
- free_list->current_blk_idx = blk_idx;
}
+/* insert at the top of a non-free list */
static void cfa_mm_blk_insert(struct cfa_mm *context,
struct cfa_mm_blk_list *blk_list,
uint32_t blk_idx)
{
+ /* there are no entries in the list so init all to this one */
if (blk_list->first_blk_idx == CFA_MM_INVALID32) {
blk_list->first_blk_idx = blk_idx;
- blk_list->current_blk_idx = blk_idx;
+ blk_list->last_blk_idx = blk_idx;
} else {
struct cfa_mm_blk *blk_info = &context->blk_tbl[blk_idx];
@@ -226,10 +235,29 @@ static void cfa_mm_blk_insert(struct cfa_mm *context,
context->blk_tbl[blk_list->first_blk_idx].prev_blk_idx =
blk_idx;
blk_list->first_blk_idx = blk_idx;
- blk_list->current_blk_idx = blk_idx;
}
}
+/* insert at the bottom of a non-free list */
+static void cfa_mm_blk_insert_last(struct cfa_mm *context,
+ struct cfa_mm_blk_list *blk_list,
+ uint32_t blk_idx)
+{
+ if (blk_list->last_blk_idx == CFA_MM_INVALID32) {
+ blk_list->first_blk_idx = blk_idx;
+ blk_list->last_blk_idx = blk_idx;
+ } else {
+ struct cfa_mm_blk *blk_info = &context->blk_tbl[blk_idx];
+
+ blk_info->prev_blk_idx = blk_list->last_blk_idx;
+ blk_info->next_blk_idx = CFA_MM_INVALID32;
+ context->blk_tbl[blk_list->last_blk_idx].next_blk_idx =
+ blk_idx;
+ blk_list->last_blk_idx = blk_idx;
+ }
+}
+
+/* delete from anywhere in the list */
static void cfa_mm_blk_delete(struct cfa_mm *context,
struct cfa_mm_blk_list *blk_list,
uint32_t blk_idx)
@@ -239,15 +267,20 @@ static void cfa_mm_blk_delete(struct cfa_mm *context,
if (blk_list->first_blk_idx == CFA_MM_INVALID32)
return;
+ if (blk_list->last_blk_idx == blk_idx) {
+ blk_list->last_blk_idx = blk_info->prev_blk_idx;
+ if (blk_list->last_blk_idx != CFA_MM_INVALID32) {
+ context->blk_tbl[blk_list->last_blk_idx].next_blk_idx =
+ CFA_MM_INVALID32;
+ }
+ }
+
if (blk_list->first_blk_idx == blk_idx) {
blk_list->first_blk_idx = blk_info->next_blk_idx;
if (blk_list->first_blk_idx != CFA_MM_INVALID32) {
context->blk_tbl[blk_list->first_blk_idx].prev_blk_idx =
CFA_MM_INVALID32;
}
- if (blk_list->current_blk_idx == blk_idx)
- blk_list->current_blk_idx = blk_list->first_blk_idx;
-
return;
}
@@ -260,20 +293,6 @@ static void cfa_mm_blk_delete(struct cfa_mm *context,
context->blk_tbl[blk_info->next_blk_idx].prev_blk_idx =
blk_info->prev_blk_idx;
}
-
- if (blk_list->current_blk_idx == blk_idx) {
- if (blk_info->next_blk_idx != CFA_MM_INVALID32) {
- blk_list->current_blk_idx = blk_info->next_blk_idx;
- } else {
- if (blk_info->prev_blk_idx != CFA_MM_INVALID32) {
- blk_list->current_blk_idx =
- blk_info->prev_blk_idx;
- } else {
- blk_list->current_blk_idx =
- blk_list->first_blk_idx;
- }
- }
- }
}
/* Returns true if the bit in the bitmap is set to 'val' else returns false */
@@ -413,12 +432,19 @@ int cfa_mm_alloc(void *cmm, struct cfa_mm_alloc_parms *parms)
blk_info->num_contig_records = num_records;
} else {
- blk_idx = blk_list->current_blk_idx;
+ blk_idx = blk_list->first_blk_idx;
blk_info = &context->blk_tbl[blk_idx];
}
while (blk_info->num_free_records < num_records) {
- if (blk_info->next_blk_idx == CFA_MM_INVALID32 || !blk_info->num_free_records) {
+ /*
+ * All non-full entries precede full entries so
+ * upon seeing the first full entry, allocate
+ * new block as this means all following records
+ * are full.
+ */
+ if (blk_info->next_blk_idx == CFA_MM_INVALID32 ||
+ !blk_info->num_free_records) {
blk_idx = cfa_mm_blk_alloc(context);
if (unlikely(blk_idx == CFA_MM_INVALID32)) {
ret = -ENOMEM;
@@ -433,8 +459,6 @@ int cfa_mm_alloc(void *cmm, struct cfa_mm_alloc_parms *parms)
} else {
blk_idx = blk_info->next_blk_idx;
blk_info = &context->blk_tbl[blk_idx];
-
- blk_list->current_blk_idx = blk_idx;
}
}
@@ -459,6 +483,9 @@ int cfa_mm_alloc(void *cmm, struct cfa_mm_alloc_parms *parms)
blk_info->num_free_records -= num_records;
if (!blk_info->num_free_records) {
+ /* move block to the end of the list if it is full */
+ cfa_mm_blk_delete(context, blk_list, blk_idx);
+ cfa_mm_blk_insert_last(context, blk_list, blk_idx);
blk_info->first_free_record = context->records_per_block;
} else {
cnt = NUM_ALIGN_UNITS(context->records_per_block,
diff --git a/drivers/net/bnxt/hcapi/cfa_v3/mm/cfa_mm_priv.h b/drivers/net/bnxt/hcapi/cfa_v3/mm/cfa_mm_priv.h
index 2a646217ec..dbac1c3cf2 100644
--- a/drivers/net/bnxt/hcapi/cfa_v3/mm/cfa_mm_priv.h
+++ b/drivers/net/bnxt/hcapi/cfa_v3/mm/cfa_mm_priv.h
@@ -47,7 +47,7 @@ struct cfa_mm_blk_list {
/* Index of the first block in the list */
uint32_t first_blk_idx;
/* Index of the current block having free records */
- uint32_t current_blk_idx;
+ uint32_t last_blk_idx;
};
/**
--
2.39.5 (Apple Git-154)
next prev parent reply other threads:[~2025-09-30 7:12 UTC|newest]
Thread overview: 55+ messages / expand[flat|nested] mbox.gz Atom feed top
2025-09-30 0:35 [PATCH 00/54] bnxt patchset Manish Kurup
2025-09-30 0:35 ` [PATCH 01/54] net/bnxt/tf_ulp: add bnxt app data for 25.11 Manish Kurup
2025-09-30 0:35 ` [PATCH 02/54] net/bnxt: fix a NULL pointer dereference in bnxt_rep funcs Manish Kurup
2025-09-30 0:35 ` [PATCH 03/54] net/bnxt: enable vector mode processing Manish Kurup
2025-09-30 0:35 ` [PATCH 04/54] net/bnxt/tf_ulp: add meter stats support for Thor2 Manish Kurup
2025-09-30 0:35 ` [PATCH 05/54] net/bnxt/tf_core: dynamic UPAR support for THOR2 Manish Kurup
2025-09-30 0:35 ` [PATCH 06/54] net/bnxt/tf_core: fix the miscalculation of the lkup table pool Manish Kurup
2025-09-30 0:35 ` [PATCH 07/54] net/bnxt/tf_core: thor2 TF table scope sizing adjustments Manish Kurup
2025-09-30 0:35 ` [PATCH 08/54] net/bnxt/tf_ulp: add support for global identifiers Manish Kurup
2025-09-30 0:35 ` [PATCH 09/54] net/bnxt/tf_core: add support for multi instance Manish Kurup
2025-09-30 0:35 ` [PATCH 10/54] net/bnxt/tf_core: fix table scope free Manish Kurup
2025-09-30 0:35 ` [PATCH 11/54] net/bnxt/tf_core: fix vfr clean up and stats lockup Manish Kurup
2025-09-30 0:35 ` [PATCH 12/54] net/bnxt/tf_ulp: add support for special vxlan Manish Kurup
2025-09-30 0:35 ` [PATCH 13/54] net/bnxt/tf_ulp: increase shared pool size to 32 Manish Kurup
2025-09-30 0:35 ` [PATCH 14/54] next/bnxt/tf_ulp: truflow fixes for meter and mac_addr cache Manish Kurup
2025-09-30 0:35 ` [PATCH 15/54] net/bnxt/tf_ulp: add support for tcam priority update Manish Kurup
2025-09-30 0:35 ` [PATCH 16/54] net/bnxt/tf_ulp: hot upgrade support Manish Kurup
2025-09-30 0:35 ` [PATCH 17/54] net/bnxt/tf_core: tcam manager logical id free Manish Kurup
2025-09-30 0:35 ` [PATCH 18/54] net/bnxt/tf_ulp: fix stats counter memory initialization Manish Kurup
2025-09-30 0:35 ` [PATCH 19/54] net/bnxt: fix max VFs count for thor2 Manish Kurup
2025-09-30 0:35 ` [PATCH 20/54] net/bnxt/tf_ulp: ovs-dpdk packet drop observed with thor2 Manish Kurup
2025-09-30 0:35 ` [PATCH 21/54] net/bnxt/tf_ulp: fix seg fault when devargs argument missing Manish Kurup
2025-09-30 0:35 ` [PATCH 22/54] net/bnxt: fix default rss config Manish Kurup
2025-09-30 0:35 ` [PATCH 23/54] net/bnxt/tf_ulp: enable support for global index table Manish Kurup
2025-09-30 0:35 ` [PATCH 24/54] net/bnxt/tf_core: fix build failure with flow scale option Manish Kurup
2025-09-30 0:35 ` [PATCH 25/54] net/bnxt: truflow remove redundant code for mpc init Manish Kurup
2025-09-30 0:35 ` [PATCH 26/54] net/bnxt/tf_ulp: optimize template enums Manish Kurup
2025-09-30 0:35 ` [PATCH 27/54] net/bnxt/tf_core: thor2 hot upgrade ungraceful quit crash Manish Kurup
2025-09-30 0:35 ` [PATCH 28/54] net/bnxt/tf_ulp: support MPLS packets Manish Kurup
2025-09-30 0:35 ` [PATCH 29/54] net/bnxt/tf_core: add backing store debug to dpdk Manish Kurup
2025-09-30 0:35 ` [PATCH 30/54] net/bnxt/tf_core: truflow global table scope Manish Kurup
2025-09-30 0:35 ` [PATCH 31/54] net/bnxt/tf_ulp: ulp parser support to handle gre key Manish Kurup
2025-09-30 0:35 ` [PATCH 32/54] net/bnxt/tf_core: handle out of order MPC completions Manish Kurup
2025-09-30 0:35 ` [PATCH 33/54] net/bnxt/tf_ulp: socket direct enable Manish Kurup
2025-09-30 0:35 ` [PATCH 34/54] net/bnxt: fix adding udp_tunnel_port Manish Kurup
2025-09-30 0:35 ` [PATCH 35/54] net/bnxt/tf_ulp: add non vfr mode capability Manish Kurup
2025-09-30 0:35 ` [PATCH 36/54] net/bnxt: avoid iova range check when external memory is used Manish Kurup
2025-09-30 0:35 ` [PATCH 37/54] net/bnxt: avoid potential segfault in VFR handling Manish Kurup
2025-09-30 0:35 ` [PATCH 38/54] net/bnxt/tf_ulp: change rte_mem_virt2iova to rte_mem_virt2phys Manish Kurup
2025-09-30 0:35 ` Manish Kurup [this message]
2025-09-30 0:35 ` [PATCH 40/54] net/bnxt: fix stats collection when rx queue is not set Manish Kurup
2025-09-30 0:35 ` [PATCH 41/54] net/bnxt: fix rss configuration when set to none Manish Kurup
2025-09-30 0:35 ` [PATCH 42/54] net/bnxt: packet drop after port stop and start Manish Kurup
2025-09-30 0:35 ` [PATCH 43/54] net/bnxt/tf_core: fix truflow crash on memory allocation failure Manish Kurup
2025-09-30 0:35 ` [PATCH 44/54] net/bnxt: truflow remove RTE devarg processing for mpc=1 Manish Kurup
2025-09-30 0:35 ` [PATCH 45/54] net/bnxt: add meson build options for TruFlow Manish Kurup
2025-09-30 0:35 ` [PATCH 46/54] net/bnxt: truflow HSI struct fixes Manish Kurup
2025-09-30 0:35 ` [PATCH 47/54] net/bnxt/tf_ulp: truflow add pf action handler Manish Kurup
2025-09-30 0:35 ` [PATCH 48/54] net/bnxt/tf_ulp: add support for unicast only feature Manish Kurup
2025-09-30 0:35 ` [PATCH 49/54] net/bnxt/tf_core: remove excessive debug logging Manish Kurup
2025-09-30 0:36 ` [PATCH 50/54] net/bnxt/tf_core: fix truflow PF init failure on sriov disabled Manish Kurup
2025-09-30 0:36 ` [PATCH 51/54] net/bnxt/tf_ulp: fixes to enable TF functionality Manish Kurup
2025-09-30 0:36 ` [PATCH 52/54] net/bnxt/tf_ulp: add feature bit rx miss handling Manish Kurup
2025-09-30 0:36 ` [PATCH 53/54] net/bnxt: add support for truflow promiscuous mode Manish Kurup
2025-09-30 0:36 ` [PATCH 54/54] net/bnxt/tf_ulp: remove Truflow DEBUG code Manish Kurup
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20250930003604.87108-40-manish.kurup@broadcom.com \
--to=manish.kurup@broadcom.com \
--cc=ajit.khaparde@broadcom.com \
--cc=dev@dpdk.org \
--cc=farah.smith@broadcom.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).