[RFC PATCH] iwlwifi: Memory allocation optimization.

[RFC PATCH] iwlwifi: Memory allocation optimization.

[Tomas Winkler]

From: Gregory Greenman <gregory.greenman@intel.com>

This patch optimizes memory allocation. The cmd member of
iwl_tx_queue was allocated previously as a continuous block
of memory. This patch allocates separate memory chunks for each command
and maps/unmaps these chunks in the run time.


Signed-off-by: Gregory Greenman <gregory.greenman@intel.com>
Signed-off-by: Tomas Winkler <tomas.winkler@intel.com>
---
This patch should help a bit the memory allocation problems seen on 64 bit platforms
This patch still not went through exesive testing so it's not to be applied yet into 
a production code 

 drivers/net/wireless/iwlwifi/iwl-5000.c |    6 +-
 drivers/net/wireless/iwlwifi/iwl-dev.h  |    3 +-
 drivers/net/wireless/iwlwifi/iwl-hcmd.c |    2 +-
 drivers/net/wireless/iwlwifi/iwl-tx.c   |   75 ++++++++++++++++++++++---------
 4 files changed, 58 insertions(+), 28 deletions(-)

diff --git a/drivers/net/wireless/iwlwifi/iwl-5000.c b/drivers/net/wireless/iwlwifi/iwl-5000.c
index 717db0d..3cf2a69 100644
--- a/drivers/net/wireless/iwlwifi/iwl-5000.c
+++ b/drivers/net/wireless/iwlwifi/iwl-5000.c
@@ -914,8 +914,8 @@ static void iwl5000_txq_update_byte_cnt_tbl(struct iwl_priv *priv,
 	len = byte_cnt + IWL_TX_CRC_SIZE + IWL_TX_DELIMITER_SIZE;
 
 	if (txq_id != IWL_CMD_QUEUE_NUM) {
-		sta = txq->cmd[txq->q.write_ptr].cmd.tx.sta_id;
-		sec_ctl = txq->cmd[txq->q.write_ptr].cmd.tx.sec_ctl;
+		sta = txq->cmd[txq->q.write_ptr]->cmd.tx.sta_id;
+		sec_ctl = txq->cmd[txq->q.write_ptr]->cmd.tx.sec_ctl;
 
 		switch (sec_ctl & TX_CMD_SEC_MSK) {
 		case TX_CMD_SEC_CCM:
@@ -954,7 +954,7 @@ static void iwl5000_txq_inval_byte_cnt_tbl(struct iwl_priv *priv,
 	u8 sta = 0;
 
 	if (txq_id != IWL_CMD_QUEUE_NUM)
-		sta = txq->cmd[txq->q.read_ptr].cmd.tx.sta_id;
+		sta = txq->cmd[txq->q.read_ptr]->cmd.tx.sta_id;
 
 	shared_data->queues_byte_cnt_tbls[txq_id].tfd_offset[txq->q.read_ptr].
 					val = cpu_to_le16(1 | (sta << 12));
diff --git a/drivers/net/wireless/iwlwifi/iwl-dev.h b/drivers/net/wireless/iwlwifi/iwl-dev.h
index d1289cf..8a5f981 100644
--- a/drivers/net/wireless/iwlwifi/iwl-dev.h
+++ b/drivers/net/wireless/iwlwifi/iwl-dev.h
@@ -138,8 +138,7 @@ struct iwl_tx_info {
 struct iwl_tx_queue {
 	struct iwl_queue q;
 	struct iwl_tfd_frame *bd;
-	struct iwl_cmd *cmd;
-	dma_addr_t dma_addr_cmd;
+	struct iwl_cmd *cmd[TFD_TX_CMD_SLOTS];
 	struct iwl_tx_info *txb;
 	int need_update;
 	int sched_retry;
diff --git a/drivers/net/wireless/iwlwifi/iwl-hcmd.c b/drivers/net/wireless/iwlwifi/iwl-hcmd.c
index 8fa991b..6512834 100644
--- a/drivers/net/wireless/iwlwifi/iwl-hcmd.c
+++ b/drivers/net/wireless/iwlwifi/iwl-hcmd.c
@@ -228,7 +228,7 @@ cancel:
 		 * TX cmd queue. Otherwise in case the cmd comes
 		 * in later, it will possibly set an invalid
 		 * address (cmd->meta.source). */
-		qcmd = &priv->txq[IWL_CMD_QUEUE_NUM].cmd[cmd_idx];
+		qcmd = priv->txq[IWL_CMD_QUEUE_NUM].cmd[cmd_idx];
 		qcmd->meta.flags &= ~CMD_WANT_SKB;
 	}
 fail:
diff --git a/drivers/net/wireless/iwlwifi/iwl-tx.c b/drivers/net/wireless/iwlwifi/iwl-tx.c
index 032641d..f5f7cd3 100644
--- a/drivers/net/wireless/iwlwifi/iwl-tx.c
+++ b/drivers/net/wireless/iwlwifi/iwl-tx.c
@@ -208,11 +208,12 @@ EXPORT_SYMBOL(iwl_txq_update_write_ptr);
  * Free all buffers.
  * 0-fill, but do not free "txq" descriptor structure.
  */
-static void iwl_tx_queue_free(struct iwl_priv *priv, struct iwl_tx_queue *txq)
+static void iwl_tx_queue_free(struct iwl_priv *priv, int txq_id)
 {
+	struct iwl_tx_queue *txq = &priv->txq[txq_id];
 	struct iwl_queue *q = &txq->q;
 	struct pci_dev *dev = priv->pci_dev;
-	int len;
+	int i, slots_num, len;
 
 	if (q->n_bd == 0)
 		return;
@@ -227,7 +228,12 @@ static void iwl_tx_queue_free(struct iwl_priv *priv, struct iwl_tx_queue *txq)
 		len += IWL_MAX_SCAN_SIZE;
 
 	/* De-alloc array of command/tx buffers */
-	pci_free_consistent(dev, len, txq->cmd, txq->dma_addr_cmd);
+	slots_num = (txq_id == IWL_CMD_QUEUE_NUM) ?
+			TFD_CMD_SLOTS : TFD_TX_CMD_SLOTS;
+	for (i = 0; i < slots_num; i++)
+		kfree(txq->cmd[i]);
+	if (txq_id == IWL_CMD_QUEUE_NUM)
+		kfree(txq->cmd[slots_num]);
 
 	/* De-alloc circular buffer of TFDs */
 	if (txq->q.n_bd)
@@ -400,8 +406,7 @@ static int iwl_tx_queue_init(struct iwl_priv *priv,
 			     struct iwl_tx_queue *txq,
 			     int slots_num, u32 txq_id)
 {
-	struct pci_dev *dev = priv->pci_dev;
-	int len;
+	int i, len;
 	int rc = 0;
 
 	/*
@@ -412,17 +417,24 @@ static int iwl_tx_queue_init(struct iwl_priv *priv,
 	 * For normal Tx queues (all other queues), no super-size command
 	 * space is needed.
 	 */
-	len = sizeof(struct iwl_cmd) * slots_num;
-	if (txq_id == IWL_CMD_QUEUE_NUM)
-		len +=  IWL_MAX_SCAN_SIZE;
-	txq->cmd = pci_alloc_consistent(dev, len, &txq->dma_addr_cmd);
-	if (!txq->cmd)
-		return -ENOMEM;
+	for (i = 0; i <= slots_num; i++) {
+		if (i == slots_num)
+			if (txq_id == IWL_CMD_QUEUE_NUM)
+				len = IWL_MAX_SCAN_SIZE;
+			else
+				continue;
+		else
+			len = sizeof(struct iwl_cmd);
+		txq->cmd[i] = kmalloc(len, GFP_KERNEL | GFP_DMA);
+		if (!txq->cmd[i])
+			return -ENOMEM;
+	}
 
 	/* Alloc driver data array and TFD circular buffer */
 	rc = iwl_tx_queue_alloc(priv, txq, txq_id);
 	if (rc) {
-		pci_free_consistent(dev, len, txq->cmd, txq->dma_addr_cmd);
+		for (i = 0; i < slots_num; i++)
+			kfree(txq->cmd[i]);
 
 		return -ENOMEM;
 	}
@@ -451,7 +463,7 @@ void iwl_hw_txq_ctx_free(struct iwl_priv *priv)
 
 	/* Tx queues */
 	for (txq_id = 0; txq_id < priv->hw_params.max_txq_num; txq_id++)
-		iwl_tx_queue_free(priv, &priv->txq[txq_id]);
+		iwl_tx_queue_free(priv, txq_id);
 
 	/* Keep-warm buffer */
 	iwl_kw_free(priv);
@@ -865,7 +877,7 @@ int iwl_tx_skb(struct iwl_priv *priv, struct sk_buff *skb)
 	txq->txb[q->write_ptr].skb[0] = skb;
 
 	/* Set up first empty entry in queue's array of Tx/cmd buffers */
-	out_cmd = &txq->cmd[idx];
+	out_cmd = txq->cmd[idx];
 	tx_cmd = &out_cmd->cmd.tx;
 	memset(&out_cmd->hdr, 0, sizeof(out_cmd->hdr));
 	memset(tx_cmd, 0, sizeof(struct iwl_tx_cmd));
@@ -905,8 +917,9 @@ int iwl_tx_skb(struct iwl_priv *priv, struct sk_buff *skb)
 
 	/* Physical address of this Tx command's header (not MAC header!),
 	 * within command buffer array. */
-	txcmd_phys = txq->dma_addr_cmd + sizeof(struct iwl_cmd) * idx +
-		     offsetof(struct iwl_cmd, hdr);
+	txcmd_phys = pci_map_single(priv->pci_dev, out_cmd,
+				sizeof(struct iwl_cmd), PCI_DMA_TODEVICE);
+	txcmd_phys += offsetof(struct iwl_cmd, hdr);
 
 	/* Add buffer containing Tx command and MAC(!) header to TFD's
 	 * first entry */
@@ -1010,7 +1023,7 @@ int iwl_enqueue_hcmd(struct iwl_priv *priv, struct iwl_host_cmd *cmd)
 	u32 idx;
 	u16 fix_size;
 	dma_addr_t phys_addr;
-	int ret;
+	int len, ret;
 	unsigned long flags;
 
 	cmd->len = priv->cfg->ops->utils->get_hcmd_size(cmd->id, cmd->len);
@@ -1040,7 +1053,7 @@ int iwl_enqueue_hcmd(struct iwl_priv *priv, struct iwl_host_cmd *cmd)
 	control_flags = (u32 *) tfd;
 
 	idx = get_cmd_index(q, q->write_ptr, cmd->meta.flags & CMD_SIZE_HUGE);
-	out_cmd = &txq->cmd[idx];
+	out_cmd = txq->cmd[idx];
 
 	out_cmd->hdr.cmd = cmd->id;
 	memcpy(&out_cmd->meta, &cmd->meta, sizeof(cmd->meta));
@@ -1054,9 +1067,11 @@ int iwl_enqueue_hcmd(struct iwl_priv *priv, struct iwl_host_cmd *cmd)
 			INDEX_TO_SEQ(q->write_ptr));
 	if (out_cmd->meta.flags & CMD_SIZE_HUGE)
 		out_cmd->hdr.sequence |= cpu_to_le16(SEQ_HUGE_FRAME);
-
-	phys_addr = txq->dma_addr_cmd + sizeof(txq->cmd[0]) * idx +
-			offsetof(struct iwl_cmd, hdr);
+	len = (idx == TFD_CMD_SLOTS) ?
+			IWL_MAX_SCAN_SIZE : sizeof(struct iwl_cmd);
+	phys_addr = pci_map_single(priv->pci_dev, out_cmd, len,
+						PCI_DMA_TODEVICE);
+	phys_addr += offsetof(struct iwl_cmd, hdr);
 	iwl_hw_txq_attach_buf_to_tfd(priv, tfd, phys_addr, fix_size);
 
 	IWL_DEBUG_HC("Sending command %s (#%x), seq: 0x%04X, "
@@ -1121,6 +1136,9 @@ static void iwl_hcmd_queue_reclaim(struct iwl_priv *priv, int txq_id, int index)
 {
 	struct iwl_tx_queue *txq = &priv->txq[txq_id];
 	struct iwl_queue *q = &txq->q;
+	struct iwl_tfd_frame *bd = &txq->bd[index];
+	dma_addr_t dma_addr;
+	int is_odd, buf_len;
 	int nfreed = 0;
 
 	if ((index >= q->n_bd) || (iwl_queue_used(q, index) == 0)) {
@@ -1138,6 +1156,19 @@ static void iwl_hcmd_queue_reclaim(struct iwl_priv *priv, int txq_id, int index)
 					q->write_ptr, q->read_ptr);
 			queue_work(priv->workqueue, &priv->restart);
 		}
+		is_odd = (index/2) & 0x1;
+		if (is_odd) {
+			dma_addr = IWL_GET_BITS(bd->pa[index], tb2_addr_lo16) |
+					(IWL_GET_BITS(bd->pa[index],
+							tb2_addr_hi20) << 16);
+			buf_len = IWL_GET_BITS(bd->pa[index], tb2_len);
+		} else {
+			dma_addr = le32_to_cpu(bd->pa[index].tb1_addr);
+			buf_len = IWL_GET_BITS(bd->pa[index], tb1_len);
+		}
+
+		pci_unmap_single(priv->pci_dev, dma_addr, buf_len,
+				 PCI_DMA_TODEVICE);
 		nfreed++;
 	}
 }
@@ -1169,7 +1200,7 @@ void iwl_tx_cmd_complete(struct iwl_priv *priv, struct iwl_rx_mem_buffer *rxb)
 	BUG_ON(txq_id != IWL_CMD_QUEUE_NUM);
 
 	cmd_index = get_cmd_index(&priv->txq[IWL_CMD_QUEUE_NUM].q, index, huge);
-	cmd = &priv->txq[IWL_CMD_QUEUE_NUM].cmd[cmd_index];
+	cmd = priv->txq[IWL_CMD_QUEUE_NUM].cmd[cmd_index];
 
 	/* Input error checking is done when commands are added to queue. */
 	if (cmd->meta.flags & CMD_WANT_SKB) {
-- 
1.5.4.1

---------------------------------------------------------------------
Intel Israel (74) Limited

This e-mail and any attachments may contain confidential material for
the sole use of the intended recipient(s). Any review or distribution
by others is strictly prohibited. If you are not the intended
recipient, please contact the sender and delete all copies.

Re: [RFC PATCH] iwlwifi: Memory allocation optimization.

[Johannes Berg]

[-- Attachment #1: Type: text/plain, Size: 748 bytes --]

On Thu, 2008-07-03 at 19:41 +0300, Tomas Winkler wrote:
> From: Gregory Greenman <gregory.greenman@intel.com>
> 
> This patch optimizes memory allocation. The cmd member of
> iwl_tx_queue was allocated previously as a continuous block
> of memory. This patch allocates separate memory chunks for each command
> and maps/unmaps these chunks in the run time.

You still allocate a lot of unused memory. Why don't you, at least
temporarily until you've figured out how to do gather-IO [1], reduce the
maximum number of DMA fragments to 2 so you need 1/10th the memory? You
can always bump it back up after you do the necessary mac80211 changes.

johannes

[1] and I still maintain that you will not be able to, but that's
irrelevant.

[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 836 bytes --]

Re: [RFC PATCH] iwlwifi: Memory allocation optimization.

[Tomas Winkler]

On Sat, Jul 5, 2008 at 7:30 PM, Johannes Berg <johannes@sipsolutions.net> wrote:
> On Thu, 2008-07-03 at 19:41 +0300, Tomas Winkler wrote:
>> From: Gregory Greenman <gregory.greenman@intel.com>
>>
>> This patch optimizes memory allocation. The cmd member of
>> iwl_tx_queue was allocated previously as a continuous block
>> of memory. This patch allocates separate memory chunks for each command
>> and maps/unmaps these chunks in the run time.
>
> You still allocate a lot of unused memory. Why don't you, at least
> temporarily until you've figured out how to do gather-IO [1], reduce the
> maximum number of DMA fragments to 2 so you need 1/10th the memory? You
> can always bump it back up after you do the necessary mac80211 changes.

We have to improve more places in allocation and we hopefully had
mapped all of them.
This patch fixes allocation that was in order 5 in( (ATOMCI | DMA) and
was most urgent.
The section you've suggested is I think on 3rd or 4th place. Don't
worry we get there.
What is more important now that we get some feedback on this fix.
Thanks
Tomas


> johannes
>
> [1] and I still maintain that you will not be able to, but that's
> irrelevant.
>

Re: [RFC PATCH] iwlwifi: Memory allocation optimization.

[Johannes Berg]

[-- Attachment #1: Type: text/plain, Size: 1138 bytes --]


> We have to improve more places in allocation and we hopefully had
> mapped all of them.
> This patch fixes allocation that was in order 5 in( (ATOMCI | DMA) and
> was most urgent.
> The section you've suggested is I think on 3rd or 4th place. Don't
> worry we get there.
> What is more important now that we get some feedback on this fix.

Fair enough. I can't really comment on that patch. You might want to
have somebody do some spring cleaning in iwlwifi, I just found this
comment I inserted there during some refactoring:

4150c572 (Johannes Berg      2007-09-17 01:29:23 -0400 6941) /*
4150c572 (Johannes Berg      2007-09-17 01:29:23 -0400 6942)  * very dubious code was here; the probe filtering flag is never set:
4150c572 (Johannes Berg      2007-09-17 01:29:23 -0400 6943)  *
b481de9c (Zhu Yi             2007-09-25 17:54:57 -0700 6944)    if (unlikely(test_bit(STATUS_SCANNING, &priv->status)) &&
b481de9c (Zhu Yi             2007-09-25 17:54:57 -0700 6945)        !(priv->hw->flags & IEEE80211_HW_NO_PROBE_FILTERING)) {
4150c572 (Johannes Berg      2007-09-17 01:29:23 -0400 6946)  */

:)

johannes

[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 836 bytes --]