Re: [**EXTERNAL**] [PATCH 03/30] net/ice/base: remove unnecessary control queue array

["Gudimetla, Leela Sankar" <lgudimet@ciena.com>]

[-- Attachment #1: Type: text/plain, Size: 19700 bytes --]

Hi @Qiming Yang<mailto:qiming.yang@intel.com>, @Jacob Keller<mailto:jacob.e.keller@intel.com>

Good day!
I am seeing  a crash in ice/base running DPDK-2111-stable. We use VPP.
While looking at DPDK-emails, I see some relevant changes here.
Can you please take a look at the back-trace and suggest/point if this has been found and fixed ?

(gdb) bt
#0  __pthread_kill_implementation (threadid=<optimized out>, signo=signo@entry=6, no_tid=no_tid@entry=0) at pthread_kill.c:44
#1  0x00007fb6f9983773 in __pthread_kill_internal (signo=6, threadid=<optimized out>) at pthread_kill.c:78
#2  0x00007fb6f9938876 in __GI_raise (sig=sig@entry=6) at ../sysdeps/posix/raise.c:26
#3  0x00007fb6f99237e3 in __GI_abort () at abort.c:79
#4  0x00000000004083fa in os_exit (code=code@entry=1) at /usr/src/debug/vpp/21.01+gitAUTOINC+fdd8bd2f89-r0/git/src/vpp/vnet/main.c:434
#5  0x00007fb6f9c6b8d1 in unix_signal_handler (signum=11, si=<optimized out>, uc=<optimized out>) at /usr/src/debug/vpp/21.01+gitAUTOINC+fdd8bd2f89-r0/git/src/vlib/unix/main.c:190
#6  <signal handler called>
#7  0x00007fb6b37e5850 in ice_sq_send_cmd_nolock (hw=0xac0318700, cq=0xac031a1c0, desc=0x7fb6b643c5e0, buf=0xac2a78200, buf_size=6, cd=0x0) at ../git/drivers/net/ice/base/ice_controlq.c:1058
#8  0x00007fb6b37f428a in ice_sq_send_cmd (hw=0xac0318700, cq=0xac031a1c0, desc=0x7fb6b643c5e0, buf=0xac2a78200, buf_size=6, cd=0x0) at ../git/drivers/net/ice/base/ice_controlq.c:1126
#9  0x00007fb6b3824ddb in ice_sq_send_cmd_retry (hw=0xac0318700, cq=0xac031a1c0, desc=0x7fb6b643c5e0, buf=0xac2a78200, buf_size=6, cd=0x0) at ../git/drivers/net/ice/base/ice_common.c:1729
#10 0x00007fb6b38335ee in ice_aq_send_cmd (hw=0xac0318700, desc=0x7fb6b643c5e0, buf=0xac2a78200, buf_size=6, cd=0x0) at ../git/drivers/net/ice/base/ice_common.c:1788
#11 0x00007fb6b3833fca in ice_aq_alloc_free_res (hw=0xac0318700, num_entries=1, buf=0xac2a78200, buf_size=6, opc=ice_aqc_opc_alloc_res, cd=0x0) at ../git/drivers/net/ice/base/ice_common.c:2124
#12 0x00007fb6b38340a1 in ice_alloc_hw_res (hw=0xac0318700, type=96, num=1, btm=false, res=0x7fb6b643f348) at ../git/drivers/net/ice/base/ice_common.c:2154
#13 0x00007fb6b39e25e8 in ice_alloc_prof_id (hw=0xac0318700, blk=ICE_BLK_RSS, prof_id=0x7fb6b643f39f "") at ../git/drivers/net/ice/base/ice_flex_pipe.c:3309
#14 0x00007fb6b3a03659 in ice_add_prof (hw=0xac0318700, blk=ICE_BLK_RSS, id=17179875328, ptypes=0xac2a790fc "", attr=0x0, attr_cnt=0, es=0xac2a78fd0, masks=0xac2a7909a, fd_swap=true)
    at ../git/drivers/net/ice/base/ice_flex_pipe.c:5028
#15 0x00007fb6b3a20210 in ice_flow_add_prof_sync (hw=0xac0318700, blk=ICE_BLK_RSS, dir=ICE_FLOW_RX, prof_id=17179875328, segs=0xac2a79200, segs_cnt=1 '\001', acts=0x0, acts_cnt=0 '\000',
    prof=0x7fb6b6444f10) at ../git/drivers/net/ice/base/ice_flow.c:2245
#16 0x00007fb6b3a20f9a in ice_flow_add_prof (hw=0xac0318700, blk=ICE_BLK_RSS, dir=ICE_FLOW_RX, prof_id=17179875328, segs=0xac2a79200, segs_cnt=1 '\001', acts=0x0, acts_cnt=0 '\000',
    prof=0x7fb6b6444f10) at ../git/drivers/net/ice/base/ice_flow.c:2646
#17 0x00007fb6b3a483af in ice_add_rss_cfg_sync (hw=0xac0318700, vsi_handle=0, cfg=0x7fb6b6444f60) at ../git/drivers/net/ice/base/ice_flow.c:4276
#18 0x00007fb6b3a48503 in ice_add_rss_cfg (hw=0xac0318700, vsi_handle=0, cfg=0x7fb6b6444fe0) at ../git/drivers/net/ice/base/ice_flow.c:4329
#19 0x00007fb6b3c18577 in ice_add_rss_cfg_wrap (pf=0xac031b718, vsi_id=0, cfg=0x7fb6b6444fe0) at ../git/drivers/net/ice/ice_ethdev.c:2998
#20 0x00007fb6b3c186cc in ice_rss_hash_set (pf=0xac031b718, rss_hf=12220) at ../git/drivers/net/ice/ice_ethdev.c:3038
#21 0x00007fb6b3c2eba9 in ice_init_rss (pf=0xac031b718) at ../git/drivers/net/ice/ice_ethdev.c:3289
#22 0x00007fb6b3c2eccd in ice_dev_configure (dev=0x7fb6b8dc04c0 <rte_eth_devices>) at ../git/drivers/net/ice/ice_ethdev.c:3323
#23 0x00007fb6b8ca82a3 in rte_eth_dev_configure (port_id=0, nb_rx_q=2, nb_tx_q=3, dev_conf=0x7fb6bbde3820) at ../git/lib/ethdev/rte_ethdev.c:1633
#24 0x00007fb6b8fb7fef in dpdk_device_setup (xd=xd@entry=0x7fb6bbde37c0) at /usr/src/debug/vpp/21.01+gitAUTOINC+fdd8bd2f89-r0/git/src/plugins/dpdk/device/common.c:92
#25 0x00007fb6b8fcbe37 in dpdk_lib_init (dm=0x7fb6b901d260 <dpdk_main>) at /usr/src/debug/vpp/21.01+gitAUTOINC+fdd8bd2f89-r0/git/src/plugins/dpdk/device/init.c:759
#26 0x00007fb6b8fd0936 in dpdk_process (vm=0x4aef40 <vlib_global_main>, rt=<optimized out>, f=<optimized out>)
    at /usr/src/debug/vpp/21.01+gitAUTOINC+fdd8bd2f89-r0/git/src/plugins/dpdk/device/init.c:1684
#27 0x00007fb6f9c1c7b7 in vlib_process_bootstrap (_a=<optimized out>) at /usr/src/debug/vpp/21.01+gitAUTOINC+fdd8bd2f89-r0/git/src/vlib/main.c:1477
#28 0x00007fb6f9b24ea8 in clib_calljmp () at /usr/src/debug/vpp/21.01+gitAUTOINC+fdd8bd2f89-r0/git/src/vppinfra/longjmp.S:123
#29 0x00007fb6b8656d10 in ?? ()
#30 0x00007fb6f9c2075c in vlib_process_startup (f=0x0, p=0x7fb6b966e080, vm=0x4aef40 <vlib_global_main>) at /usr/src/debug/vpp/21.01+gitAUTOINC+fdd8bd2f89-r0/git/src/vlib/main.c:1502
#31 dispatch_process (vm=0x4aef40 <vlib_global_main>, p=0x7fb6b966e080, last_time_stamp=<optimized out>, f=0x0) at /usr/src/debug/vpp/21.01+gitAUTOINC+fdd8bd2f89-r0/git/src/vlib/main.c:1558
#32 0x0000000000000000 in ?? ()

Thanks,
Leela sankar

From: Qiming Yang <qiming.yang@intel.com>
Date: Thursday, April 27, 2023 at 1:38 AM
To: dev@dpdk.org <dev@dpdk.org>
Cc: qi.z.zhang@intel.com <qi.z.zhang@intel.com>, Qiming Yang <qiming.yang@intel.com>, Jacob Keller <jacob.e.keller@intel.com>
Subject: [**EXTERNAL**] [PATCH 03/30] net/ice/base: remove unnecessary control queue array
The driver allocates a cmd_buf array in addition to the desc_buf array.
This array stores an ice_sq_cd command details structure for each entry in
the control queue ring. And improve debug print for control queue messages.

Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Signed-off-by: Qiming Yang <qiming.yang@intel.com>
---
 drivers/net/ice/base/ice_controlq.c | 107 +++++++++++++++-------------
 drivers/net/ice/base/ice_controlq.h |   3 -
 2 files changed, 56 insertions(+), 54 deletions(-)

diff --git a/drivers/net/ice/base/ice_controlq.c b/drivers/net/ice/base/ice_controlq.c
index c34407b48c..acd6ad249b 100644
--- a/drivers/net/ice/base/ice_controlq.c
+++ b/drivers/net/ice/base/ice_controlq.c
@@ -101,13 +101,6 @@ ice_alloc_ctrlq_sq_ring(struct ice_hw *hw, struct ice_ctl_q_info *cq)
         if (!cq->sq.desc_buf.va)
                 return ICE_ERR_NO_MEMORY;

-       cq->sq.cmd_buf = ice_calloc(hw, cq->num_sq_entries,
-                                   sizeof(struct ice_sq_cd));
-       if (!cq->sq.cmd_buf) {
-               ice_free_dma_mem(hw, &cq->sq.desc_buf);
-               return ICE_ERR_NO_MEMORY;
-       }
-
         return ICE_SUCCESS;
 }

@@ -176,7 +169,7 @@ ice_alloc_rq_bufs(struct ice_hw *hw, struct ice_ctl_q_info *cq)
                 if (cq->rq_buf_size > ICE_AQ_LG_BUF)
                         desc->flags |= CPU_TO_LE16(ICE_AQ_FLAG_LB);
                 desc->opcode = 0;
-               /* This is in accordance with Admin queue design, there is no
+               /* This is in accordance with control queue design, there is no
                  * register for buffer size configuration
                  */
                 desc->datalen = CPU_TO_LE16(bi->size);
@@ -309,9 +302,6 @@ do {                                                                        \
                                 ice_free_dma_mem((hw),                  \
                                         &(qi)->ring.r.ring##_bi[i]);    \
         }                                                               \
-       /* free the buffer info list */                                 \
-       if ((qi)->ring.cmd_buf)                                         \
-               ice_free(hw, (qi)->ring.cmd_buf);                       \
         /* free DMA head */                                             \
         ice_free(hw, (qi)->ring.dma_head);                              \
 } while (0)
@@ -379,11 +369,11 @@ static enum ice_status ice_init_sq(struct ice_hw *hw, struct ice_ctl_q_info *cq)
 }

 /**
- * ice_init_rq - initialize ARQ
+ * ice_init_rq - initialize receive side of a control queue
  * @hw: pointer to the hardware structure
  * @cq: pointer to the specific Control queue
  *
- * The main initialization routine for the Admin Receive (Event) Queue.
+ * The main initialization routine for Receive side of a control queue.
  * Prior to calling this function, the driver *MUST* set the following fields
  * in the cq->structure:
  *     - cq->num_rq_entries
@@ -441,7 +431,7 @@ static enum ice_status ice_init_rq(struct ice_hw *hw, struct ice_ctl_q_info *cq)
 }

 /**
- * ice_shutdown_sq - shutdown the Control ATQ
+ * ice_shutdown_sq - shutdown the transmit side of a control queue
  * @hw: pointer to the hardware structure
  * @cq: pointer to the specific Control queue
  *
@@ -461,7 +451,7 @@ ice_shutdown_sq(struct ice_hw *hw, struct ice_ctl_q_info *cq)
                 goto shutdown_sq_out;
         }

-       /* Stop firmware AdminQ processing */
+       /* Stop processing of the control queue */
         wr32(hw, cq->sq.head, 0);
         wr32(hw, cq->sq.tail, 0);
         wr32(hw, cq->sq.len, 0);
@@ -834,7 +824,7 @@ void ice_destroy_all_ctrlq(struct ice_hw *hw)
 }

 /**
- * ice_clean_sq - cleans Admin send queue (ATQ)
+ * ice_clean_sq - cleans send side of a control queue
  * @hw: pointer to the hardware structure
  * @cq: pointer to the specific Control queue
  *
@@ -844,21 +834,17 @@ static u16 ice_clean_sq(struct ice_hw *hw, struct ice_ctl_q_info *cq)
 {
         struct ice_ctl_q_ring *sq = &cq->sq;
         u16 ntc = sq->next_to_clean;
-       struct ice_sq_cd *details;
         struct ice_aq_desc *desc;

         desc = ICE_CTL_Q_DESC(*sq, ntc);
-       details = ICE_CTL_Q_DETAILS(*sq, ntc);

         while (rd32(hw, cq->sq.head) != ntc) {
                 ice_debug(hw, ICE_DBG_AQ_MSG, "ntc %d head %d.\n", ntc, rd32(hw, cq->sq.head));
                 ice_memset(desc, 0, sizeof(*desc), ICE_DMA_MEM);
-               ice_memset(details, 0, sizeof(*details), ICE_NONDMA_MEM);
                 ntc++;
                 if (ntc == sq->count)
                         ntc = 0;
                 desc = ICE_CTL_Q_DESC(*sq, ntc);
-               details = ICE_CTL_Q_DETAILS(*sq, ntc);
         }

         sq->next_to_clean = ntc;
@@ -866,16 +852,42 @@ static u16 ice_clean_sq(struct ice_hw *hw, struct ice_ctl_q_info *cq)
         return ICE_CTL_Q_DESC_UNUSED(sq);
 }

+/**
+ * ice_ctl_q_str - Convert control queue type to string
+ * @qtype: the control queue type
+ *
+ * Returns: A string name for the given control queue type.
+ */
+static const char *ice_ctl_q_str(enum ice_ctl_q qtype)
+{
+       switch (qtype) {
+       case ICE_CTL_Q_UNKNOWN:
+               return "Unknown CQ";
+       case ICE_CTL_Q_ADMIN:
+               return "AQ";
+       case ICE_CTL_Q_MAILBOX:
+               return "MBXQ";
+       case ICE_CTL_Q_SB:
+               return "SBQ";
+       default:
+               return "Unrecognized CQ";
+       }
+}
+
 /**
  * ice_debug_cq
  * @hw: pointer to the hardware structure
+ * @cq: pointer to the specific Control queue
  * @desc: pointer to control queue descriptor
  * @buf: pointer to command buffer
  * @buf_len: max length of buf
+ * @response: true if this is the writeback response
  *
  * Dumps debug log about control command with descriptor contents.
  */
-static void ice_debug_cq(struct ice_hw *hw, void *desc, void *buf, u16 buf_len)
+static void
+ice_debug_cq(struct ice_hw *hw, struct ice_ctl_q_info *cq,
+            void *desc, void *buf, u16 buf_len, bool response)
 {
         struct ice_aq_desc *cq_desc = (struct ice_aq_desc *)desc;
         u16 datalen, flags;
@@ -889,7 +901,8 @@ static void ice_debug_cq(struct ice_hw *hw, void *desc, void *buf, u16 buf_len)
         datalen = LE16_TO_CPU(cq_desc->datalen);
         flags = LE16_TO_CPU(cq_desc->flags);

-       ice_debug(hw, ICE_DBG_AQ_DESC, "CQ CMD: opcode 0x%04X, flags 0x%04X, datalen 0x%04X, retval 0x%04X\n",
+       ice_debug(hw, ICE_DBG_AQ_DESC, "%s %s: opcode 0x%04X, flags 0x%04X, datalen 0x%04X, retval 0x%04X\n",
+                 ice_ctl_q_str(cq->qtype), response ? "Response" : "Command",
                   LE16_TO_CPU(cq_desc->opcode), flags, datalen,
                   LE16_TO_CPU(cq_desc->retval));
         ice_debug(hw, ICE_DBG_AQ_DESC, "\tcookie (h,l) 0x%08X 0x%08X\n",
@@ -914,23 +927,23 @@ static void ice_debug_cq(struct ice_hw *hw, void *desc, void *buf, u16 buf_len)
 }

 /**
- * ice_sq_done - check if FW has processed the Admin Send Queue (ATQ)
+ * ice_sq_done - check if the last send on a control queue has completed
  * @hw: pointer to the HW struct
  * @cq: pointer to the specific Control queue
  *
- * Returns true if the firmware has processed all descriptors on the
- * admin send queue. Returns false if there are still requests pending.
+ * Returns: true if all the descriptors on the send side of a control queue
+ *          are finished processing, false otherwise.
  */
 static bool ice_sq_done(struct ice_hw *hw, struct ice_ctl_q_info *cq)
 {
-       /* AQ designers suggest use of head for better
+       /* control queue designers suggest use of head for better
          * timing reliability than DD bit
          */
         return rd32(hw, cq->sq.head) == cq->sq.next_to_use;
 }

 /**
- * ice_sq_send_cmd_nolock - send command to Control Queue (ATQ)
+ * ice_sq_send_cmd_nolock - send command to a control queue
  * @hw: pointer to the HW struct
  * @cq: pointer to the specific Control queue
  * @desc: prefilled descriptor describing the command (non DMA mem)
@@ -938,8 +951,9 @@ static bool ice_sq_done(struct ice_hw *hw, struct ice_ctl_q_info *cq)
  * @buf_size: size of buffer for indirect commands (or 0 for direct commands)
  * @cd: pointer to command details structure
  *
- * This is the main send command routine for the ATQ. It runs the queue,
- * cleans the queue, etc.
+ * This is the main send command routine for a control queue. It prepares the
+ * command into a descriptor, bumps the send queue tail, waits for the command
+ * to complete, captures status and data for the command, etc.
  */
 enum ice_status
 ice_sq_send_cmd_nolock(struct ice_hw *hw, struct ice_ctl_q_info *cq,
@@ -950,7 +964,6 @@ ice_sq_send_cmd_nolock(struct ice_hw *hw, struct ice_ctl_q_info *cq,
         struct ice_aq_desc *desc_on_ring;
         bool cmd_completed = false;
         enum ice_status status = ICE_SUCCESS;
-       struct ice_sq_cd *details;
         u32 total_delay = 0;
         u16 retval = 0;
         u32 val = 0;
@@ -993,12 +1006,6 @@ ice_sq_send_cmd_nolock(struct ice_hw *hw, struct ice_ctl_q_info *cq,
                 goto sq_send_command_error;
         }

-       details = ICE_CTL_Q_DETAILS(cq->sq, cq->sq.next_to_use);
-       if (cd)
-               *details = *cd;
-       else
-               ice_memset(details, 0, sizeof(*details), ICE_NONDMA_MEM);
-
         /* Call clean and check queue available function to reclaim the
          * descriptors that were processed by FW/MBX; the function returns the
          * number of desc available. The clean function called here could be
@@ -1035,8 +1042,7 @@ ice_sq_send_cmd_nolock(struct ice_hw *hw, struct ice_ctl_q_info *cq,

         /* Debug desc and buffer */
         ice_debug(hw, ICE_DBG_AQ_DESC, "ATQ: Control Send queue desc and buffer:\n");
-
-       ice_debug_cq(hw, (void *)desc_on_ring, buf, buf_size);
+       ice_debug_cq(hw, cq, (void *)desc_on_ring, buf, buf_size, false);

         (cq->sq.next_to_use)++;
         if (cq->sq.next_to_use == cq->sq.count)
@@ -1084,13 +1090,12 @@ ice_sq_send_cmd_nolock(struct ice_hw *hw, struct ice_ctl_q_info *cq,
         }

         ice_debug(hw, ICE_DBG_AQ_MSG, "ATQ: desc and buffer writeback:\n");
-
-       ice_debug_cq(hw, (void *)desc, buf, buf_size);
+       ice_debug_cq(hw, cq, (void *)desc, buf, buf_size, true);

         /* save writeback AQ if requested */
-       if (details->wb_desc)
-               ice_memcpy(details->wb_desc, desc_on_ring,
-                          sizeof(*details->wb_desc), ICE_DMA_TO_NONDMA);
+       if (cd && cd->wb_desc)
+               ice_memcpy(cd->wb_desc, desc_on_ring,
+                          sizeof(*cd->wb_desc), ICE_DMA_TO_NONDMA);

         /* update the error if time out occurred */
         if (!cmd_completed) {
@@ -1109,7 +1114,7 @@ ice_sq_send_cmd_nolock(struct ice_hw *hw, struct ice_ctl_q_info *cq,
 }

 /**
- * ice_sq_send_cmd - send command to Control Queue (ATQ)
+ * ice_sq_send_cmd - send command to a control queue
  * @hw: pointer to the HW struct
  * @cq: pointer to the specific Control queue
  * @desc: prefilled descriptor describing the command
@@ -1117,8 +1122,9 @@ ice_sq_send_cmd_nolock(struct ice_hw *hw, struct ice_ctl_q_info *cq,
  * @buf_size: size of buffer for indirect commands (or 0 for direct commands)
  * @cd: pointer to command details structure
  *
- * This is the main send command routine for the ATQ. It runs the queue,
- * cleans the queue, etc.
+ * Main command for the transmit side of a control queue. It puts the command
+ * on the queue, bumps the tail, waits for processing of the command, captures
+ * command status and results, etc.
  */
 enum ice_status
 ice_sq_send_cmd(struct ice_hw *hw, struct ice_ctl_q_info *cq,
@@ -1160,9 +1166,9 @@ void ice_fill_dflt_direct_cmd_desc(struct ice_aq_desc *desc, u16 opcode)
  * @e: event info from the receive descriptor, includes any buffers
  * @pending: number of events that could be left to process
  *
- * This function cleans one Admin Receive Queue element and returns
- * the contents through e. It can also return how many events are
- * left to process through 'pending'.
+ * Clean one element from the receive side of a control queue. On return 'e'
+ * contains contents of the message, and 'pending' contains the number of
+ * events left to process.
  */
 enum ice_status
 ice_clean_rq_elem(struct ice_hw *hw, struct ice_ctl_q_info *cq,
@@ -1218,8 +1224,7 @@ ice_clean_rq_elem(struct ice_hw *hw, struct ice_ctl_q_info *cq,
                            e->msg_len, ICE_DMA_TO_NONDMA);

         ice_debug(hw, ICE_DBG_AQ_DESC, "ARQ: desc and buffer:\n");
-
-       ice_debug_cq(hw, (void *)desc, e->msg_buf, cq->rq_buf_size);
+       ice_debug_cq(hw, cq, (void *)desc, e->msg_buf, cq->rq_buf_size, true);

         /* Restore the original datalen and buffer address in the desc,
          * FW updates datalen to indicate the event message size
diff --git a/drivers/net/ice/base/ice_controlq.h b/drivers/net/ice/base/ice_controlq.h
index 986604ec3c..5c60469693 100644
--- a/drivers/net/ice/base/ice_controlq.h
+++ b/drivers/net/ice/base/ice_controlq.h
@@ -43,7 +43,6 @@ enum ice_ctl_q {
 struct ice_ctl_q_ring {
         void *dma_head;                 /* Virtual address to DMA head */
         struct ice_dma_mem desc_buf;    /* descriptor ring memory */
-       void *cmd_buf;                  /* command buffer memory */

         union {
                 struct ice_dma_mem *sq_bi;
@@ -73,8 +72,6 @@ struct ice_sq_cd {
         struct ice_aq_desc *wb_desc;
 };

-#define ICE_CTL_Q_DETAILS(R, i) (&(((struct ice_sq_cd *)((R).cmd_buf))[i]))
-
 /* rq event information */
 struct ice_rq_event_info {
         struct ice_aq_desc desc;
--
2.25.1

[-- Attachment #2: Type: text/html, Size: 38034 bytes --]