From 559fb51ba7e66fe298b8355fabde1275b7def35f Mon Sep 17 00:00:00 2001 From: Scott Bardone Date: Thu, 23 Jun 2005 01:40:19 -0400 Subject: Update Chelsio gige net driver. - Use extern prefix for functions required. - Removed a lot of wrappers, including t1_read/write_reg_4. - Removed various macros, using native kernel calls now. - Enumerated various #defines. - Removed a lot of shared code which is not currently used in "NIC only" mode. - Removed dead code. Documentation/networking/cxgb.txt: - Updated release notes for version 2.1.1 drivers/net/chelsio/ch_ethtool.h - removed file, no longer using ETHTOOL namespace. drivers/net/chelsio/common.h - moved code from osdep.h to common.h - added comment to #endif indicating which symbol it closes. drivers/net/chelsio/cphy.h - removed dead code. - added comment to #endif indicating which symbol it closes. drivers/net/chelsio/cxgb2.c - use DMA_{32,64}BIT_MASK in include/linux/dma-mapping.h. - removed unused code. - use printk message for link info resembling drivers/net/mii.c. - no longer using the MODULE_xxx namespace. - no longer using "pci_" namespace. - no longer using ETHTOOL namespace. drivers/net/chelsio/cxgb2.h - removed file, merged into common.h drivers/net/chelsio/elmer0.h - removed dead code. - added various enums. - added comment to #endif indicating which symbol it closes. drivers/net/chelsio/espi.c - removed various macros, using native kernel calls now. - removed a lot of wrappers, including t1_read/write_reg_4. drivers/net/chelsio/espi.h - added comment to #endif indicating which symbol it closes. drivers/net/chelsio/gmac.h - added comment to #endif indicating which symbol it closes. drivers/net/chelsio/mv88x201x.c - changes to sync with Chelsio TOT. drivers/net/chelsio/osdep.h - removed file, consolidation. osdep was used to translate wrapper functions since our code supports multiple OSs. removed wrappers. drivers/net/chelsio/pm3393.c - removed various macros, using native kernel calls now. - removed a lot of wrappers, including t1_read/write_reg_4. - removed unused code. drivers/net/chelsio/regs.h - added a few register entries for future and current feature support. - added comment to #endif indicating which symbol it closes. drivers/net/chelsio/sge.c - rewrote large portion of scatter-gather engine to stabilize performance. - using u8/u16/u32 kernel types instead of __u8/__u16/__u32 compiler types. drivers/net/chelsio/sge.h - rewrote large portion of scatter-gather engine to stabilize performance. - added comment to #endif indicating which symbol it closes. drivers/net/chelsio/subr.c - merged tp.c into subr.c - removed various macros, using native kernel calls now. - removed a lot of wrappers, including t1_read/write_reg_4. - removed unused code. drivers/net/chelsio/suni1x10gexp_regs.h - modified copyright and authorship of file. - added comment to #endif indicating which symbol it closes. drivers/net/chelsio/tp.c - removed file, merged into subr.c. drivers/net/chelsio/tp.h - removed file. include/linux/pci_ids.h - patched to include PCI_VENDOR_ID_CHELSIO 0x1425, removed define from our code. --- drivers/net/chelsio/sge.c | 1859 +++++++++++++++++++++++++-------------------- 1 file changed, 1046 insertions(+), 813 deletions(-) (limited to 'drivers/net/chelsio/sge.c') diff --git a/drivers/net/chelsio/sge.c b/drivers/net/chelsio/sge.c index bcf8b1e..53b41d9 100644 --- a/drivers/net/chelsio/sge.c +++ b/drivers/net/chelsio/sge.c @@ -1,8 +1,8 @@ /***************************************************************************** * * * File: sge.c * - * $Revision: 1.13 $ * - * $Date: 2005/03/23 07:41:27 $ * + * $Revision: 1.26 $ * + * $Date: 2005/06/21 18:29:48 $ * * Description: * * DMA engine. * * part of the Chelsio 10Gb Ethernet Driver. * @@ -58,59 +58,62 @@ #include "regs.h" #include "espi.h" + +#ifdef NETIF_F_TSO #include +#endif #define SGE_CMDQ_N 2 #define SGE_FREELQ_N 2 -#define SGE_CMDQ0_E_N 512 +#define SGE_CMDQ0_E_N 1024 #define SGE_CMDQ1_E_N 128 #define SGE_FREEL_SIZE 4096 #define SGE_JUMBO_FREEL_SIZE 512 #define SGE_FREEL_REFILL_THRESH 16 #define SGE_RESPQ_E_N 1024 -#define SGE_INTR_BUCKETSIZE 100 -#define SGE_INTR_LATBUCKETS 5 -#define SGE_INTR_MAXBUCKETS 11 -#define SGE_INTRTIMER0 1 -#define SGE_INTRTIMER1 50 -#define SGE_INTRTIMER_NRES 10000 -#define SGE_RX_COPY_THRESHOLD 256 +#define SGE_INTRTIMER_NRES 1000 +#define SGE_RX_COPY_THRES 256 #define SGE_RX_SM_BUF_SIZE 1536 -#define SGE_RESPQ_REPLENISH_THRES ((3 * SGE_RESPQ_E_N) / 4) +# define SGE_RX_DROP_THRES 2 + +#define SGE_RESPQ_REPLENISH_THRES (SGE_RESPQ_E_N / 4) + +/* + * Period of the TX buffer reclaim timer. This timer does not need to run + * frequently as TX buffers are usually reclaimed by new TX packets. + */ +#define TX_RECLAIM_PERIOD (HZ / 4) -#define SGE_RX_OFFSET 2 #ifndef NET_IP_ALIGN -# define NET_IP_ALIGN SGE_RX_OFFSET +# define NET_IP_ALIGN 2 #endif +#define M_CMD_LEN 0x7fffffff +#define V_CMD_LEN(v) (v) +#define G_CMD_LEN(v) ((v) & M_CMD_LEN) +#define V_CMD_GEN1(v) ((v) << 31) +#define V_CMD_GEN2(v) (v) +#define F_CMD_DATAVALID (1 << 1) +#define F_CMD_SOP (1 << 2) +#define V_CMD_EOP(v) ((v) << 3) + /* - * Memory Mapped HW Command, Freelist and Response Queue Descriptors + * Command queue, receive buffer list, and response queue descriptors. */ #if defined(__BIG_ENDIAN_BITFIELD) struct cmdQ_e { - u32 AddrLow; - u32 GenerationBit : 1; - u32 BufferLength : 31; - u32 RespQueueSelector : 4; - u32 ResponseTokens : 12; - u32 CmdId : 8; - u32 Reserved : 3; - u32 TokenValid : 1; - u32 Eop : 1; - u32 Sop : 1; - u32 DataValid : 1; - u32 GenerationBit2 : 1; - u32 AddrHigh; + u32 addr_lo; + u32 len_gen; + u32 flags; + u32 addr_hi; }; struct freelQ_e { - u32 AddrLow; - u32 GenerationBit : 1; - u32 BufferLength : 31; - u32 Reserved : 31; - u32 GenerationBit2 : 1; - u32 AddrHigh; + u32 addr_lo; + u32 len_gen; + u32 gen2; + u32 addr_hi; }; struct respQ_e { @@ -128,31 +131,19 @@ struct respQ_e { u32 GenerationBit : 1; u32 BufferLength; }; - #elif defined(__LITTLE_ENDIAN_BITFIELD) struct cmdQ_e { - u32 BufferLength : 31; - u32 GenerationBit : 1; - u32 AddrLow; - u32 AddrHigh; - u32 GenerationBit2 : 1; - u32 DataValid : 1; - u32 Sop : 1; - u32 Eop : 1; - u32 TokenValid : 1; - u32 Reserved : 3; - u32 CmdId : 8; - u32 ResponseTokens : 12; - u32 RespQueueSelector : 4; + u32 len_gen; + u32 addr_lo; + u32 addr_hi; + u32 flags; }; struct freelQ_e { - u32 BufferLength : 31; - u32 GenerationBit : 1; - u32 AddrLow; - u32 AddrHigh; - u32 GenerationBit2 : 1; - u32 Reserved : 31; + u32 len_gen; + u32 addr_lo; + u32 addr_hi; + u32 gen2; }; struct respQ_e { @@ -179,7 +170,6 @@ struct cmdQ_ce { struct sk_buff *skb; DECLARE_PCI_UNMAP_ADDR(dma_addr); DECLARE_PCI_UNMAP_LEN(dma_len); - unsigned int single; }; struct freelQ_ce { @@ -189,44 +179,52 @@ struct freelQ_ce { }; /* - * SW Command, Freelist and Response Queue + * SW command, freelist and response rings */ struct cmdQ { - atomic_t asleep; /* HW DMA Fetch status */ - atomic_t credits; /* # available descriptors for TX */ - atomic_t pio_pidx; /* Variable updated on Doorbell */ - u16 entries_n; /* # descriptors for TX */ - u16 pidx; /* producer index (SW) */ - u16 cidx; /* consumer index (HW) */ - u8 genbit; /* current generation (=valid) bit */ - struct cmdQ_e *entries; /* HW command descriptor Q */ - struct cmdQ_ce *centries; /* SW command context descriptor Q */ - spinlock_t Qlock; /* Lock to protect cmdQ enqueuing */ - dma_addr_t dma_addr; /* DMA addr HW command descriptor Q */ + unsigned long status; /* HW DMA fetch status */ + unsigned int in_use; /* # of in-use command descriptors */ + unsigned int size; /* # of descriptors */ + unsigned int processed; /* total # of descs HW has processed */ + unsigned int cleaned; /* total # of descs SW has reclaimed */ + unsigned int stop_thres; /* SW TX queue suspend threshold */ + u16 pidx; /* producer index (SW) */ + u16 cidx; /* consumer index (HW) */ + u8 genbit; /* current generation (=valid) bit */ + u8 sop; /* is next entry start of packet? */ + struct cmdQ_e *entries; /* HW command descriptor Q */ + struct cmdQ_ce *centries; /* SW command context descriptor Q */ + spinlock_t lock; /* Lock to protect cmdQ enqueuing */ + dma_addr_t dma_addr; /* DMA addr HW command descriptor Q */ }; struct freelQ { - unsigned int credits; /* # of available RX buffers */ - unsigned int entries_n; /* free list capacity */ - u16 pidx; /* producer index (SW) */ - u16 cidx; /* consumer index (HW) */ + unsigned int credits; /* # of available RX buffers */ + unsigned int size; /* free list capacity */ + u16 pidx; /* producer index (SW) */ + u16 cidx; /* consumer index (HW) */ u16 rx_buffer_size; /* Buffer size on this free list */ u16 dma_offset; /* DMA offset to align IP headers */ - u8 genbit; /* current generation (=valid) bit */ - struct freelQ_e *entries; /* HW freelist descriptor Q */ - struct freelQ_ce *centries; /* SW freelist conext descriptor Q */ - dma_addr_t dma_addr; /* DMA addr HW freelist descriptor Q */ + u16 recycleq_idx; /* skb recycle q to use */ + u8 genbit; /* current generation (=valid) bit */ + struct freelQ_e *entries; /* HW freelist descriptor Q */ + struct freelQ_ce *centries; /* SW freelist context descriptor Q */ + dma_addr_t dma_addr; /* DMA addr HW freelist descriptor Q */ }; struct respQ { - u16 credits; /* # of available respQ descriptors */ - u16 credits_pend; /* # of not yet returned descriptors */ - u16 entries_n; /* # of response Q descriptors */ - u16 pidx; /* producer index (HW) */ - u16 cidx; /* consumer index (SW) */ - u8 genbit; /* current generation(=valid) bit */ + unsigned int credits; /* credits to be returned to SGE */ + unsigned int size; /* # of response Q descriptors */ + u16 cidx; /* consumer index (SW) */ + u8 genbit; /* current generation(=valid) bit */ struct respQ_e *entries; /* HW response descriptor Q */ - dma_addr_t dma_addr; /* DMA addr HW response descriptor Q */ + dma_addr_t dma_addr; /* DMA addr HW response descriptor Q */ +}; + +/* Bit flags for cmdQ.status */ +enum { + CMDQ_STAT_RUNNING = 1, /* fetch engine is running */ + CMDQ_STAT_LAST_PKT_DB = 2 /* last packet rung the doorbell */ }; /* @@ -239,134 +237,50 @@ struct respQ { */ struct sge { struct adapter *adapter; /* adapter backpointer */ - struct freelQ freelQ[SGE_FREELQ_N]; /* freelist Q(s) */ - struct respQ respQ; /* response Q instatiation */ + struct net_device *netdev; /* netdevice backpointer */ + struct freelQ freelQ[SGE_FREELQ_N]; /* buffer free lists */ + struct respQ respQ; /* response Q */ + unsigned long stopped_tx_queues; /* bitmap of suspended Tx queues */ unsigned int rx_pkt_pad; /* RX padding for L2 packets */ unsigned int jumbo_fl; /* jumbo freelist Q index */ - u32 intrtimer[SGE_INTR_MAXBUCKETS]; /* ! */ - u32 currIndex; /* current index into intrtimer[] */ - u32 intrtimer_nres; /* no resource interrupt timer value */ - u32 sge_control; /* shadow content of sge control reg */ - struct sge_intr_counts intr_cnt; - struct timer_list ptimer; - struct sk_buff *pskb; - u32 ptimeout; - struct cmdQ cmdQ[SGE_CMDQ_N] ____cacheline_aligned; /* command Q(s)*/ + unsigned int intrtimer_nres; /* no-resource interrupt timer */ + unsigned int fixed_intrtimer;/* non-adaptive interrupt timer */ + struct timer_list tx_reclaim_timer; /* reclaims TX buffers */ + struct timer_list espibug_timer; + unsigned int espibug_timeout; + struct sk_buff *espibug_skb; + u32 sge_control; /* shadow value of sge control reg */ + struct sge_intr_counts stats; + struct sge_port_stats port_stats[MAX_NPORTS]; + struct cmdQ cmdQ[SGE_CMDQ_N] ____cacheline_aligned_in_smp; }; -static unsigned int t1_sge_tx(struct sk_buff *skb, struct adapter *adapter, - unsigned int qid); - /* * PIO to indicate that memory mapped Q contains valid descriptor(s). */ -static inline void doorbell_pio(struct sge *sge, u32 val) +static inline void doorbell_pio(struct adapter *adapter, u32 val) { wmb(); - t1_write_reg_4(sge->adapter, A_SG_DOORBELL, val); -} - -/* - * Disables the DMA engine. - */ -void t1_sge_stop(struct sge *sge) -{ - t1_write_reg_4(sge->adapter, A_SG_CONTROL, 0); - t1_read_reg_4(sge->adapter, A_SG_CONTROL); /* flush write */ - if (is_T2(sge->adapter)) - del_timer_sync(&sge->ptimer); -} - -static u8 ch_mac_addr[ETH_ALEN] = {0x0, 0x7, 0x43, 0x0, 0x0, 0x0}; -static void t1_espi_workaround(void *data) -{ - struct adapter *adapter = (struct adapter *)data; - struct sge *sge = adapter->sge; - - if (netif_running(adapter->port[0].dev) && - atomic_read(&sge->cmdQ[0].asleep)) { - - u32 seop = t1_espi_get_mon(adapter, 0x930, 0); - - if ((seop & 0xfff0fff) == 0xfff && sge->pskb) { - struct sk_buff *skb = sge->pskb; - if (!skb->cb[0]) { - memcpy(skb->data+sizeof(struct cpl_tx_pkt), ch_mac_addr, ETH_ALEN); - memcpy(skb->data+skb->len-10, ch_mac_addr, ETH_ALEN); - - skb->cb[0] = 0xff; - } - t1_sge_tx(skb, adapter,0); - } - } - mod_timer(&adapter->sge->ptimer, jiffies + sge->ptimeout); -} - -/* - * Enables the DMA engine. - */ -void t1_sge_start(struct sge *sge) -{ - t1_write_reg_4(sge->adapter, A_SG_CONTROL, sge->sge_control); - t1_read_reg_4(sge->adapter, A_SG_CONTROL); /* flush write */ - if (is_T2(sge->adapter)) { - init_timer(&sge->ptimer); - sge->ptimer.function = (void *)&t1_espi_workaround; - sge->ptimer.data = (unsigned long)sge->adapter; - sge->ptimer.expires = jiffies + sge->ptimeout; - add_timer(&sge->ptimer); - } -} - -/* - * Creates a t1_sge structure and returns suggested resource parameters. - */ -struct sge * __devinit t1_sge_create(struct adapter *adapter, - struct sge_params *p) -{ - struct sge *sge = kmalloc(sizeof(*sge), GFP_KERNEL); - - if (!sge) - return NULL; - memset(sge, 0, sizeof(*sge)); - - if (is_T2(adapter)) - sge->ptimeout = 1; /* finest allowed */ - - sge->adapter = adapter; - sge->rx_pkt_pad = t1_is_T1B(adapter) ? 0 : SGE_RX_OFFSET; - sge->jumbo_fl = t1_is_T1B(adapter) ? 1 : 0; - - p->cmdQ_size[0] = SGE_CMDQ0_E_N; - p->cmdQ_size[1] = SGE_CMDQ1_E_N; - p->freelQ_size[!sge->jumbo_fl] = SGE_FREEL_SIZE; - p->freelQ_size[sge->jumbo_fl] = SGE_JUMBO_FREEL_SIZE; - p->rx_coalesce_usecs = SGE_INTRTIMER1; - p->last_rx_coalesce_raw = SGE_INTRTIMER1 * - (board_info(sge->adapter)->clock_core / 1000000); - p->default_rx_coalesce_usecs = SGE_INTRTIMER1; - p->coalesce_enable = 0; /* Turn off adaptive algorithm by default */ - p->sample_interval_usecs = 0; - return sge; + writel(val, adapter->regs + A_SG_DOORBELL); } /* * Frees all RX buffers on the freelist Q. The caller must make sure that * the SGE is turned off before calling this function. */ -static void free_freelQ_buffers(struct pci_dev *pdev, struct freelQ *Q) +static void free_freelQ_buffers(struct pci_dev *pdev, struct freelQ *q) { - unsigned int cidx = Q->cidx, credits = Q->credits; + unsigned int cidx = q->cidx; - while (credits--) { - struct freelQ_ce *ce = &Q->centries[cidx]; + while (q->credits--) { + struct freelQ_ce *ce = &q->centries[cidx]; pci_unmap_single(pdev, pci_unmap_addr(ce, dma_addr), pci_unmap_len(ce, dma_len), PCI_DMA_FROMDEVICE); dev_kfree_skb(ce->skb); ce->skb = NULL; - if (++cidx == Q->entries_n) + if (++cidx == q->size) cidx = 0; } } @@ -380,29 +294,29 @@ static void free_rx_resources(struct sge *sge) unsigned int size, i; if (sge->respQ.entries) { - size = sizeof(struct respQ_e) * sge->respQ.entries_n; + size = sizeof(struct respQ_e) * sge->respQ.size; pci_free_consistent(pdev, size, sge->respQ.entries, sge->respQ.dma_addr); } for (i = 0; i < SGE_FREELQ_N; i++) { - struct freelQ *Q = &sge->freelQ[i]; + struct freelQ *q = &sge->freelQ[i]; - if (Q->centries) { - free_freelQ_buffers(pdev, Q); - kfree(Q->centries); + if (q->centries) { + free_freelQ_buffers(pdev, q); + kfree(q->centries); } - if (Q->entries) { - size = sizeof(struct freelQ_e) * Q->entries_n; - pci_free_consistent(pdev, size, Q->entries, - Q->dma_addr); + if (q->entries) { + size = sizeof(struct freelQ_e) * q->size; + pci_free_consistent(pdev, size, q->entries, + q->dma_addr); } } } /* * Allocates basic RX resources, consisting of memory mapped freelist Qs and a - * response Q. + * response queue. */ static int alloc_rx_resources(struct sge *sge, struct sge_params *p) { @@ -410,21 +324,22 @@ static int alloc_rx_resources(struct sge *sge, struct sge_params *p) unsigned int size, i; for (i = 0; i < SGE_FREELQ_N; i++) { - struct freelQ *Q = &sge->freelQ[i]; - - Q->genbit = 1; - Q->entries_n = p->freelQ_size[i]; - Q->dma_offset = SGE_RX_OFFSET - sge->rx_pkt_pad; - size = sizeof(struct freelQ_e) * Q->entries_n; - Q->entries = (struct freelQ_e *) - pci_alloc_consistent(pdev, size, &Q->dma_addr); - if (!Q->entries) + struct freelQ *q = &sge->freelQ[i]; + + q->genbit = 1; + q->size = p->freelQ_size[i]; + q->dma_offset = sge->rx_pkt_pad ? 0 : NET_IP_ALIGN; + size = sizeof(struct freelQ_e) * q->size; + q->entries = (struct freelQ_e *) + pci_alloc_consistent(pdev, size, &q->dma_addr); + if (!q->entries) goto err_no_mem; - memset(Q->entries, 0, size); - Q->centries = kcalloc(Q->entries_n, sizeof(struct freelQ_ce), - GFP_KERNEL); - if (!Q->centries) + memset(q->entries, 0, size); + size = sizeof(struct freelQ_ce) * q->size; + q->centries = kmalloc(size, GFP_KERNEL); + if (!q->centries) goto err_no_mem; + memset(q->centries, 0, size); } /* @@ -440,10 +355,17 @@ static int alloc_rx_resources(struct sge *sge, struct sge_params *p) sge->freelQ[sge->jumbo_fl].rx_buffer_size = (16 * 1024) - SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); + /* + * Setup which skb recycle Q should be used when recycling buffers from + * each free list. + */ + sge->freelQ[!sge->jumbo_fl].recycleq_idx = 0; + sge->freelQ[sge->jumbo_fl].recycleq_idx = 1; + sge->respQ.genbit = 1; - sge->respQ.entries_n = SGE_RESPQ_E_N; - sge->respQ.credits = SGE_RESPQ_E_N; - size = sizeof(struct respQ_e) * sge->respQ.entries_n; + sge->respQ.size = SGE_RESPQ_E_N; + sge->respQ.credits = 0; + size = sizeof(struct respQ_e) * sge->respQ.size; sge->respQ.entries = (struct respQ_e *) pci_alloc_consistent(pdev, size, &sge->respQ.dma_addr); if (!sge->respQ.entries) @@ -457,48 +379,37 @@ err_no_mem: } /* - * Frees 'credits_pend' TX buffers and returns the credits to Q->credits. - * - * The adaptive algorithm receives the total size of the buffers freed - * accumulated in @*totpayload. No initialization of this argument here. - * + * Reclaims n TX descriptors and frees the buffers associated with them. */ -static void free_cmdQ_buffers(struct sge *sge, struct cmdQ *Q, - unsigned int credits_pend, unsigned int *totpayload) +static void free_cmdQ_buffers(struct sge *sge, struct cmdQ *q, unsigned int n) { + struct cmdQ_ce *ce; struct pci_dev *pdev = sge->adapter->pdev; - struct sk_buff *skb; - struct cmdQ_ce *ce, *cq = Q->centries; - unsigned int entries_n = Q->entries_n, cidx = Q->cidx, - i = credits_pend; - + unsigned int cidx = q->cidx; - ce = &cq[cidx]; - while (i--) { - if (ce->single) + q->in_use -= n; + ce = &q->centries[cidx]; + while (n--) { + if (q->sop) pci_unmap_single(pdev, pci_unmap_addr(ce, dma_addr), - pci_unmap_len(ce, dma_len), + pci_unmap_len(ce, dma_len), PCI_DMA_TODEVICE); else pci_unmap_page(pdev, pci_unmap_addr(ce, dma_addr), - pci_unmap_len(ce, dma_len), + pci_unmap_len(ce, dma_len), PCI_DMA_TODEVICE); - if (totpayload) - *totpayload += pci_unmap_len(ce, dma_len); - - skb = ce->skb; - if (skb) - dev_kfree_skb_irq(skb); - + q->sop = 0; + if (ce->skb) { + dev_kfree_skb(ce->skb); + q->sop = 1; + } ce++; - if (++cidx == entries_n) { + if (++cidx == q->size) { cidx = 0; - ce = cq; + ce = q->centries; } } - - Q->cidx = cidx; - atomic_add(credits_pend, &Q->credits); + q->cidx = cidx; } /* @@ -512,20 +423,17 @@ static void free_tx_resources(struct sge *sge) unsigned int size, i; for (i = 0; i < SGE_CMDQ_N; i++) { - struct cmdQ *Q = &sge->cmdQ[i]; + struct cmdQ *q = &sge->cmdQ[i]; - if (Q->centries) { - unsigned int pending = Q->entries_n - - atomic_read(&Q->credits); - - if (pending) - free_cmdQ_buffers(sge, Q, pending, NULL); - kfree(Q->centries); + if (q->centries) { + if (q->in_use) + free_cmdQ_buffers(sge, q, q->in_use); + kfree(q->centries); } - if (Q->entries) { - size = sizeof(struct cmdQ_e) * Q->entries_n; - pci_free_consistent(pdev, size, Q->entries, - Q->dma_addr); + if (q->entries) { + size = sizeof(struct cmdQ_e) * q->size; + pci_free_consistent(pdev, size, q->entries, + q->dma_addr); } } } @@ -539,25 +447,38 @@ static int alloc_tx_resources(struct sge *sge, struct sge_params *p) unsigned int size, i; for (i = 0; i < SGE_CMDQ_N; i++) { - struct cmdQ *Q = &sge->cmdQ[i]; - - Q->genbit = 1; - Q->entries_n = p->cmdQ_size[i]; - atomic_set(&Q->credits, Q->entries_n); - atomic_set(&Q->asleep, 1); - spin_lock_init(&Q->Qlock); - size = sizeof(struct cmdQ_e) * Q->entries_n; - Q->entries = (struct cmdQ_e *) - pci_alloc_consistent(pdev, size, &Q->dma_addr); - if (!Q->entries) + struct cmdQ *q = &sge->cmdQ[i]; + + q->genbit = 1; + q->sop = 1; + q->size = p->cmdQ_size[i]; + q->in_use = 0; + q->status = 0; + q->processed = q->cleaned = 0; + q->stop_thres = 0; + spin_lock_init(&q->lock); + size = sizeof(struct cmdQ_e) * q->size; + q->entries = (struct cmdQ_e *) + pci_alloc_consistent(pdev, size, &q->dma_addr); + if (!q->entries) goto err_no_mem; - memset(Q->entries, 0, size); - Q->centries = kcalloc(Q->entries_n, sizeof(struct cmdQ_ce), - GFP_KERNEL); - if (!Q->centries) + memset(q->entries, 0, size); + size = sizeof(struct cmdQ_ce) * q->size; + q->centries = kmalloc(size, GFP_KERNEL); + if (!q->centries) goto err_no_mem; + memset(q->centries, 0, size); } + /* + * CommandQ 0 handles Ethernet and TOE packets, while queue 1 is TOE + * only. For queue 0 set the stop threshold so we can handle one more + * packet from each port, plus reserve an additional 24 entries for + * Ethernet packets only. Queue 1 never suspends nor do we reserve + * space for Ethernet packets. + */ + sge->cmdQ[0].stop_thres = sge->adapter->params.nports * + (MAX_SKB_FRAGS + 1); return 0; err_no_mem: @@ -569,9 +490,9 @@ static inline void setup_ring_params(struct adapter *adapter, u64 addr, u32 size, int base_reg_lo, int base_reg_hi, int size_reg) { - t1_write_reg_4(adapter, base_reg_lo, (u32)addr); - t1_write_reg_4(adapter, base_reg_hi, addr >> 32); - t1_write_reg_4(adapter, size_reg, size); + writel((u32)addr, adapter->regs + base_reg_lo); + writel(addr >> 32, adapter->regs + base_reg_hi); + writel(size, adapter->regs + size_reg); } /* @@ -585,97 +506,52 @@ void t1_set_vlan_accel(struct adapter *adapter, int on_off) if (on_off) sge->sge_control |= F_VLAN_XTRACT; if (adapter->open_device_map) { - t1_write_reg_4(adapter, A_SG_CONTROL, sge->sge_control); - t1_read_reg_4(adapter, A_SG_CONTROL); /* flush */ + writel(sge->sge_control, adapter->regs + A_SG_CONTROL); + readl(adapter->regs + A_SG_CONTROL); /* flush */ } } /* - * Sets the interrupt latency timer when the adaptive Rx coalescing - * is turned off. Do nothing when it is turned on again. - * - * This routine relies on the fact that the caller has already set - * the adaptive policy in adapter->sge_params before calling it. -*/ -int t1_sge_set_coalesce_params(struct sge *sge, struct sge_params *p) -{ - if (!p->coalesce_enable) { - u32 newTimer = p->rx_coalesce_usecs * - (board_info(sge->adapter)->clock_core / 1000000); - - t1_write_reg_4(sge->adapter, A_SG_INTRTIMER, newTimer); - } - return 0; -} - -/* * Programs the various SGE registers. However, the engine is not yet enabled, * but sge->sge_control is setup and ready to go. */ static void configure_sge(struct sge *sge, struct sge_params *p) { struct adapter *ap = sge->adapter; - int i; - - t1_write_reg_4(ap, A_SG_CONTROL, 0); - setup_ring_params(ap, sge->cmdQ[0].dma_addr, sge->cmdQ[0].entries_n, + + writel(0, ap->regs + A_SG_CONTROL); + setup_ring_params(ap, sge->cmdQ[0].dma_addr, sge->cmdQ[0].size, A_SG_CMD0BASELWR, A_SG_CMD0BASEUPR, A_SG_CMD0SIZE); - setup_ring_params(ap, sge->cmdQ[1].dma_addr, sge->cmdQ[1].entries_n, + setup_ring_params(ap, sge->cmdQ[1].dma_addr, sge->cmdQ[1].size, A_SG_CMD1BASELWR, A_SG_CMD1BASEUPR, A_SG_CMD1SIZE); setup_ring_params(ap, sge->freelQ[0].dma_addr, - sge->freelQ[0].entries_n, A_SG_FL0BASELWR, + sge->freelQ[0].size, A_SG_FL0BASELWR, A_SG_FL0BASEUPR, A_SG_FL0SIZE); setup_ring_params(ap, sge->freelQ[1].dma_addr, - sge->freelQ[1].entries_n, A_SG_FL1BASELWR, + sge->freelQ[1].size, A_SG_FL1BASELWR, A_SG_FL1BASEUPR, A_SG_FL1SIZE); /* The threshold comparison uses <. */ - t1_write_reg_4(ap, A_SG_FLTHRESHOLD, SGE_RX_SM_BUF_SIZE + 1); + writel(SGE_RX_SM_BUF_SIZE + 1, ap->regs + A_SG_FLTHRESHOLD); - setup_ring_params(ap, sge->respQ.dma_addr, sge->respQ.entries_n, - A_SG_RSPBASELWR, A_SG_RSPBASEUPR, A_SG_RSPSIZE); - t1_write_reg_4(ap, A_SG_RSPQUEUECREDIT, (u32)sge->respQ.entries_n); + setup_ring_params(ap, sge->respQ.dma_addr, sge->respQ.size, + A_SG_RSPBASELWR, A_SG_RSPBASEUPR, A_SG_RSPSIZE); + writel((u32)sge->respQ.size - 1, ap->regs + A_SG_RSPQUEUECREDIT); sge->sge_control = F_CMDQ0_ENABLE | F_CMDQ1_ENABLE | F_FL0_ENABLE | F_FL1_ENABLE | F_CPL_ENABLE | F_RESPONSE_QUEUE_ENABLE | V_CMDQ_PRIORITY(2) | F_DISABLE_CMDQ1_GTS | F_ISCSI_COALESCE | + F_DISABLE_FL0_GTS | F_DISABLE_FL1_GTS | V_RX_PKT_OFFSET(sge->rx_pkt_pad); #if defined(__BIG_ENDIAN_BITFIELD) sge->sge_control |= F_ENABLE_BIG_ENDIAN; #endif - /* - * Initialize the SGE Interrupt Timer arrray: - * intrtimer[0] = (SGE_INTRTIMER0) usec - * intrtimer[0intrtimer[0] = board_info(sge->adapter)->clock_core / 1000000; - for (i = 1; i < SGE_INTR_LATBUCKETS; ++i) { - sge->intrtimer[i] = SGE_INTRTIMER0 + (2 * i); - sge->intrtimer[i] *= sge->intrtimer[0]; - } - for (i = SGE_INTR_LATBUCKETS; i < SGE_INTR_MAXBUCKETS - 1; ++i) { - sge->intrtimer[i] = (i - 3) * 6; - sge->intrtimer[i] *= sge->intrtimer[0]; - } - sge->intrtimer[SGE_INTR_MAXBUCKETS - 1] = - sge->intrtimer[0] * SGE_INTRTIMER1; - /* Initialize resource timer */ - sge->intrtimer_nres = sge->intrtimer[0] * SGE_INTRTIMER_NRES; - /* Finally finish initialization of intrtimer[0] */ - sge->intrtimer[0] *= SGE_INTRTIMER0; - /* Initialize for a throughput oriented workload */ - sge->currIndex = SGE_INTR_MAXBUCKETS - 1; - - if (p->coalesce_enable) - t1_write_reg_4(ap, A_SG_INTRTIMER, - sge->intrtimer[sge->currIndex]); - else - t1_sge_set_coalesce_params(sge, p); + /* Initialize no-resource timer */ + sge->intrtimer_nres = SGE_INTRTIMER_NRES * core_ticks_per_usec(ap); + + t1_sge_set_coalesce_params(sge, p); } /* @@ -684,31 +560,8 @@ static void configure_sge(struct sge *sge, struct sge_params *p) static inline unsigned int jumbo_payload_capacity(const struct sge *sge) { return sge->freelQ[sge->jumbo_fl].rx_buffer_size - - sizeof(struct cpl_rx_data) - SGE_RX_OFFSET + sge->rx_pkt_pad; -} - -/* - * Allocates both RX and TX resources and configures the SGE. However, - * the hardware is not enabled yet. - */ -int t1_sge_configure(struct sge *sge, struct sge_params *p) -{ - if (alloc_rx_resources(sge, p)) - return -ENOMEM; - if (alloc_tx_resources(sge, p)) { - free_rx_resources(sge); - return -ENOMEM; - } - configure_sge(sge, p); - - /* - * Now that we have sized the free lists calculate the payload - * capacity of the large buffers. Other parts of the driver use - * this to set the max offload coalescing size so that RX packets - * do not overflow our large buffers. - */ - p->large_buf_capacity = jumbo_payload_capacity(sge); - return 0; + sge->freelQ[sge->jumbo_fl].dma_offset - + sizeof(struct cpl_rx_data); } /* @@ -716,8 +569,9 @@ int t1_sge_configure(struct sge *sge, struct sge_params *p) */ void t1_sge_destroy(struct sge *sge) { - if (sge->pskb) - dev_kfree_skb(sge->pskb); + if (sge->espibug_skb) + kfree_skb(sge->espibug_skb); + free_tx_resources(sge); free_rx_resources(sge); kfree(sge); @@ -735,75 +589,75 @@ void t1_sge_destroy(struct sge *sge) * we specify a RX_OFFSET in order to make sure that the IP header is 4B * aligned. */ -static void refill_free_list(struct sge *sge, struct freelQ *Q) +static void refill_free_list(struct sge *sge, struct freelQ *q) { struct pci_dev *pdev = sge->adapter->pdev; - struct freelQ_ce *ce = &Q->centries[Q->pidx]; - struct freelQ_e *e = &Q->entries[Q->pidx]; - unsigned int dma_len = Q->rx_buffer_size - Q->dma_offset; + struct freelQ_ce *ce = &q->centries[q->pidx]; + struct freelQ_e *e = &q->entries[q->pidx]; + unsigned int dma_len = q->rx_buffer_size - q->dma_offset; - while (Q->credits < Q->entries_n) { - if (e->GenerationBit != Q->genbit) { - struct sk_buff *skb; - dma_addr_t mapping; + while (q->credits < q->size) { + struct sk_buff *skb; + dma_addr_t mapping; - skb = alloc_skb(Q->rx_buffer_size, GFP_ATOMIC); - if (!skb) - break; - if (Q->dma_offset) - skb_reserve(skb, Q->dma_offset); - mapping = pci_map_single(pdev, skb->data, dma_len, - PCI_DMA_FROMDEVICE); - ce->skb = skb; - pci_unmap_addr_set(ce, dma_addr, mapping); - pci_unmap_len_set(ce, dma_len, dma_len); - e->AddrLow = (u32)mapping; - e->AddrHigh = (u64)mapping >> 32; - e->BufferLength = dma_len; - e->GenerationBit = e->GenerationBit2 = Q->genbit; - } + skb = alloc_skb(q->rx_buffer_size, GFP_ATOMIC); + if (!skb) + break; + + skb_reserve(skb, q->dma_offset); + mapping = pci_map_single(pdev, skb->data, dma_len, + PCI_DMA_FROMDEVICE); + ce->skb = skb; + pci_unmap_addr_set(ce, dma_addr, mapping); + pci_unmap_len_set(ce, dma_len, dma_len); + e->addr_lo = (u32)mapping; + e->addr_hi = (u64)mapping >> 32; + e->len_gen = V_CMD_LEN(dma_len) | V_CMD_GEN1(q->genbit); + wmb(); + e->gen2 = V_CMD_GEN2(q->genbit); e++; ce++; - if (++Q->pidx == Q->entries_n) { - Q->pidx = 0; - Q->genbit ^= 1; - ce = Q->centries; - e = Q->entries; + if (++q->pidx == q->size) { + q->pidx = 0; + q->genbit ^= 1; + ce = q->centries; + e = q->entries; } - Q->credits++; + q->credits++; } } /* - * Calls refill_free_list for both freelist Qs. If we cannot - * fill at least 1/4 of both Qs, we go into 'few interrupt mode' in order - * to give the system time to free up resources. + * Calls refill_free_list for both free lists. If we cannot fill at least 1/4 + * of both rings, we go into 'few interrupt mode' in order to give the system + * time to free up resources. */ static void freelQs_empty(struct sge *sge) { - u32 irq_reg = t1_read_reg_4(sge->adapter, A_SG_INT_ENABLE); + struct adapter *adapter = sge->adapter; + u32 irq_reg = readl(adapter->regs + A_SG_INT_ENABLE); u32 irqholdoff_reg; refill_free_list(sge, &sge->freelQ[0]); refill_free_list(sge, &sge->freelQ[1]); - if (sge->freelQ[0].credits > (sge->freelQ[0].entries_n >> 2) && - sge->freelQ[1].credits > (sge->freelQ[1].entries_n >> 2)) { + if (sge->freelQ[0].credits > (sge->freelQ[0].size >> 2) && + sge->freelQ[1].credits > (sge->freelQ[1].size >> 2)) { irq_reg |= F_FL_EXHAUSTED; - irqholdoff_reg = sge->intrtimer[sge->currIndex]; + irqholdoff_reg = sge->fixed_intrtimer; } else { /* Clear the F_FL_EXHAUSTED interrupts for now */ irq_reg &= ~F_FL_EXHAUSTED; irqholdoff_reg = sge->intrtimer_nres; } - t1_write_reg_4(sge->adapter, A_SG_INTRTIMER, irqholdoff_reg); - t1_write_reg_4(sge->adapter, A_SG_INT_ENABLE, irq_reg); + writel(irqholdoff_reg, adapter->regs + A_SG_INTRTIMER); + writel(irq_reg, adapter->regs + A_SG_INT_ENABLE); /* We reenable the Qs to force a freelist GTS interrupt later */ - doorbell_pio(sge, F_FL0_ENABLE | F_FL1_ENABLE); + doorbell_pio(adapter, F_FL0_ENABLE | F_FL1_ENABLE); } #define SGE_PL_INTR_MASK (F_PL_INTR_SGE_ERR | F_PL_INTR_SGE_DATA) @@ -816,10 +670,10 @@ static void freelQs_empty(struct sge *sge) */ void t1_sge_intr_disable(struct sge *sge) { - u32 val = t1_read_reg_4(sge->adapter, A_PL_ENABLE); + u32 val = readl(sge->adapter->regs + A_PL_ENABLE); - t1_write_reg_4(sge->adapter, A_PL_ENABLE, val & ~SGE_PL_INTR_MASK); - t1_write_reg_4(sge->adapter, A_SG_INT_ENABLE, 0); + writel(val & ~SGE_PL_INTR_MASK, sge->adapter->regs + A_PL_ENABLE); + writel(0, sge->adapter->regs + A_SG_INT_ENABLE); } /* @@ -828,12 +682,12 @@ void t1_sge_intr_disable(struct sge *sge) void t1_sge_intr_enable(struct sge *sge) { u32 en = SGE_INT_ENABLE; - u32 val = t1_read_reg_4(sge->adapter, A_PL_ENABLE); + u32 val = readl(sge->adapter->regs + A_PL_ENABLE); if (sge->adapter->flags & TSO_CAPABLE) en &= ~F_PACKET_TOO_BIG; - t1_write_reg_4(sge->adapter, A_SG_INT_ENABLE, en); - t1_write_reg_4(sge->adapter, A_PL_ENABLE, val | SGE_PL_INTR_MASK); + writel(en, sge->adapter->regs + A_SG_INT_ENABLE); + writel(val | SGE_PL_INTR_MASK, sge->adapter->regs + A_PL_ENABLE); } /* @@ -841,8 +695,8 @@ void t1_sge_intr_enable(struct sge *sge) */ void t1_sge_intr_clear(struct sge *sge) { - t1_write_reg_4(sge->adapter, A_PL_CAUSE, SGE_PL_INTR_MASK); - t1_write_reg_4(sge->adapter, A_SG_INT_CAUSE, 0xffffffff); + writel(SGE_PL_INTR_MASK, sge->adapter->regs + A_PL_CAUSE); + writel(0xffffffff, sge->adapter->regs + A_SG_INT_CAUSE); } /* @@ -851,464 +705,673 @@ void t1_sge_intr_clear(struct sge *sge) int t1_sge_intr_error_handler(struct sge *sge) { struct adapter *adapter = sge->adapter; - u32 cause = t1_read_reg_4(adapter, A_SG_INT_CAUSE); + u32 cause = readl(adapter->regs + A_SG_INT_CAUSE); if (adapter->flags & TSO_CAPABLE) cause &= ~F_PACKET_TOO_BIG; if (cause & F_RESPQ_EXHAUSTED) - sge->intr_cnt.respQ_empty++; + sge->stats.respQ_empty++; if (cause & F_RESPQ_OVERFLOW) { - sge->intr_cnt.respQ_overflow++; + sge->stats.respQ_overflow++; CH_ALERT("%s: SGE response queue overflow\n", adapter->name); } if (cause & F_FL_EXHAUSTED) { - sge->intr_cnt.freelistQ_empty++; + sge->stats.freelistQ_empty++; freelQs_empty(sge); } if (cause & F_PACKET_TOO_BIG) { - sge->intr_cnt.pkt_too_big++; + sge->stats.pkt_too_big++; CH_ALERT("%s: SGE max packet size exceeded\n", adapter->name); } if (cause & F_PACKET_MISMATCH) { - sge->intr_cnt.pkt_mismatch++; + sge->stats.pkt_mismatch++; CH_ALERT("%s: SGE packet mismatch\n", adapter->name); } if (cause & SGE_INT_FATAL) t1_fatal_err(adapter); - t1_write_reg_4(adapter, A_SG_INT_CAUSE, cause); + writel(cause, adapter->regs + A_SG_INT_CAUSE); return 0; } -/* - * The following code is copied from 2.6, where the skb_pull is doing the - * right thing and only pulls ETH_HLEN. +const struct sge_intr_counts *t1_sge_get_intr_counts(struct sge *sge) +{ + return &sge->stats; +} + +const struct sge_port_stats *t1_sge_get_port_stats(struct sge *sge, int port) +{ + return &sge->port_stats[port]; +} + +/** + * recycle_fl_buf - recycle a free list buffer + * @fl: the free list + * @idx: index of buffer to recycle * - * Determine the packet's protocol ID. The rule here is that we - * assume 802.3 if the type field is short enough to be a length. - * This is normal practice and works for any 'now in use' protocol. + * Recycles the specified buffer on the given free list by adding it at + * the next available slot on the list. */ -static unsigned short sge_eth_type_trans(struct sk_buff *skb, - struct net_device *dev) +static void recycle_fl_buf(struct freelQ *fl, int idx) { - struct ethhdr *eth; - unsigned char *rawp; + struct freelQ_e *from = &fl->entries[idx]; + struct freelQ_e *to = &fl->entries[fl->pidx]; - skb->mac.raw = skb->data; - skb_pull(skb, ETH_HLEN); - eth = (struct ethhdr *)skb->mac.raw; + fl->centries[fl->pidx] = fl->centries[idx]; + to->addr_lo = from->addr_lo; + to->addr_hi = from->addr_hi; + to->len_gen = G_CMD_LEN(from->len_gen) | V_CMD_GEN1(fl->genbit); + wmb(); + to->gen2 = V_CMD_GEN2(fl->genbit); + fl->credits++; - if (*eth->h_dest&1) { - if(memcmp(eth->h_dest, dev->broadcast, ETH_ALEN) == 0) - skb->pkt_type = PACKET_BROADCAST; - else - skb->pkt_type = PACKET_MULTICAST; + if (++fl->pidx == fl->size) { + fl->pidx = 0; + fl->genbit ^= 1; } +} - /* - * This ALLMULTI check should be redundant by 1.4 - * so don't forget to remove it. - * - * Seems, you forgot to remove it. All silly devices - * seems to set IFF_PROMISC. - */ +/** + * get_packet - return the next ingress packet buffer + * @pdev: the PCI device that received the packet + * @fl: the SGE free list holding the packet + * @len: the actual packet length, excluding any SGE padding + * @dma_pad: padding at beginning of buffer left by SGE DMA + * @skb_pad: padding to be used if the packet is copied + * @copy_thres: length threshold under which a packet should be copied + * @drop_thres: # of remaining buffers before we start dropping packets + * + * Get the next packet from a free list and complete setup of the + * sk_buff. If the packet is small we make a copy and recycle the + * original buffer, otherwise we use the original buffer itself. If a + * positive drop threshold is supplied packets are dropped and their + * buffers recycled if (a) the number of remaining buffers is under the + * threshold and the packet is too big to copy, or (b) the packet should + * be copied but there is no memory for the copy. + */ +static inline struct sk_buff *get_packet(struct pci_dev *pdev, + struct freelQ *fl, unsigned int len, + int dma_pad, int skb_pad, + unsigned int copy_thres, + unsigned int drop_thres) +{ + struct sk_buff *skb; + struct freelQ_ce *ce = &fl->centries[fl->cidx]; + + if (len < copy_thres) { + skb = alloc_skb(len + skb_pad, GFP_ATOMIC); + if (likely(skb != NULL)) { + skb_reserve(skb, skb_pad); + skb_put(skb, len); + pci_dma_sync_single_for_cpu(pdev, + pci_unmap_addr(ce, dma_addr), + pci_unmap_len(ce, dma_len), + PCI_DMA_FROMDEVICE); + memcpy(skb->data, ce->skb->data + dma_pad, len); + pci_dma_sync_single_for_device(pdev, + pci_unmap_addr(ce, dma_addr), + pci_unmap_len(ce, dma_len), + PCI_DMA_FROMDEVICE); + } else if (!drop_thres) + goto use_orig_buf; - else if (1 /*dev->flags&IFF_PROMISC*/) - { - if(memcmp(eth->h_dest,dev->dev_addr, ETH_ALEN)) - skb->pkt_type=PACKET_OTHERHOST; + recycle_fl_buf(fl, fl->cidx); + return skb; } - if (ntohs(eth->h_proto) >= 1536) - return eth->h_proto; - - rawp = skb->data; + if (fl->credits < drop_thres) { + recycle_fl_buf(fl, fl->cidx); + return NULL; + } - /* - * This is a magic hack to spot IPX packets. Older Novell breaks - * the protocol design and runs IPX over 802.3 without an 802.2 LLC - * layer. We look for FFFF which isn't a used 802.2 SSAP/DSAP. This - * won't work for fault tolerant netware but does for the rest. - */ - if (*(unsigned short *)rawp == 0xFFFF) - return htons(ETH_P_802_3); +use_orig_buf: + pci_unmap_single(pdev, pci_unmap_addr(ce, dma_addr), + pci_unmap_len(ce, dma_len), PCI_DMA_FROMDEVICE); + skb = ce->skb; + skb_reserve(skb, dma_pad); + skb_put(skb, len); + return skb; +} - /* - * Real 802.2 LLC - */ - return htons(ETH_P_802_2); +/** + * unexpected_offload - handle an unexpected offload packet + * @adapter: the adapter + * @fl: the free list that received the packet + * + * Called when we receive an unexpected offload packet (e.g., the TOE + * function is disabled or the card is a NIC). Prints a message and + * recycles the buffer. + */ +static void unexpected_offload(struct adapter *adapter, struct freelQ *fl) +{ + struct freelQ_ce *ce = &fl->centries[fl->cidx]; + struct sk_buff *skb = ce->skb; + + pci_dma_sync_single_for_cpu(adapter->pdev, pci_unmap_addr(ce, dma_addr), + pci_unmap_len(ce, dma_len), PCI_DMA_FROMDEVICE); + CH_ERR("%s: unexpected offload packet, cmd %u\n", + adapter->name, *skb->data); + recycle_fl_buf(fl, fl->cidx); } /* - * Prepare the received buffer and pass it up the stack. If it is small enough - * and allocation doesn't fail, we use a new sk_buff and copy the content. + * Write the command descriptors to transmit the given skb starting at + * descriptor pidx with the given generation. */ -static unsigned int t1_sge_rx(struct sge *sge, struct freelQ *Q, - unsigned int len, unsigned int offload) +static inline void write_tx_descs(struct adapter *adapter, struct sk_buff *skb, + unsigned int pidx, unsigned int gen, + struct cmdQ *q) { - struct sk_buff *skb; - struct adapter *adapter = sge->adapter; - struct freelQ_ce *ce = &Q->centries[Q->cidx]; + dma_addr_t mapping; + struct cmdQ_e *e, *e1; + struct cmdQ_ce *ce; + unsigned int i, flags, nfrags = skb_shinfo(skb)->nr_frags; + + mapping = pci_map_single(adapter->pdev, skb->data, + skb->len - skb->data_len, PCI_DMA_TODEVICE); + ce = &q->centries[pidx]; + ce->skb = NULL; + pci_unmap_addr_set(ce, dma_addr, mapping); + pci_unmap_len_set(ce, dma_len, skb->len - skb->data_len); - if (len <= SGE_RX_COPY_THRESHOLD && - (skb = alloc_skb(len + NET_IP_ALIGN, GFP_ATOMIC))) { - struct freelQ_e *e; - char *src = ce->skb->data; + flags = F_CMD_DATAVALID | F_CMD_SOP | V_CMD_EOP(nfrags == 0) | + V_CMD_GEN2(gen); + e = &q->entries[pidx]; + e->addr_lo = (u32)mapping; + e->addr_hi = (u64)mapping >> 32; + e->len_gen = V_CMD_LEN(skb->len - skb->data_len) | V_CMD_GEN1(gen); + for (e1 = e, i = 0; nfrags--; i++) { + skb_frag_t *frag = &skb_shinfo(skb)->frags[i]; - pci_dma_sync_single_for_cpu(adapter->pdev, - pci_unmap_addr(ce, dma_addr), - pci_unmap_len(ce, dma_len), - PCI_DMA_FROMDEVICE); - if (!offload) { - skb_reserve(skb, NET_IP_ALIGN); - src += sge->rx_pkt_pad; + ce++; + e1++; + if (++pidx == q->size) { + pidx = 0; + gen ^= 1; + ce = q->centries; + e1 = q->entries; } - memcpy(skb->data, src, len); - /* Reuse the entry. */ - e = &Q->entries[Q->cidx]; - e->GenerationBit ^= 1; - e->GenerationBit2 ^= 1; - } else { - pci_unmap_single(adapter->pdev, pci_unmap_addr(ce, dma_addr), - pci_unmap_len(ce, dma_len), - PCI_DMA_FROMDEVICE); - skb = ce->skb; - if (!offload && sge->rx_pkt_pad) - __skb_pull(skb, sge->rx_pkt_pad); + mapping = pci_map_page(adapter->pdev, frag->page, + frag->page_offset, frag->size, + PCI_DMA_TODEVICE); + ce->skb = NULL; + pci_unmap_addr_set(ce, dma_addr, mapping); + pci_unmap_len_set(ce, dma_len, frag->size); + + e1->addr_lo = (u32)mapping; + e1->addr_hi = (u64)mapping >> 32; + e1->len_gen = V_CMD_LEN(frag->size) | V_CMD_GEN1(gen); + e1->flags = F_CMD_DATAVALID | V_CMD_EOP(nfrags == 0) | + V_CMD_GEN2(gen); } - skb_put(skb, len); + ce->skb = skb; + wmb(); + e->flags = flags; +} +/* + * Clean up completed Tx buffers. + */ +static inline void reclaim_completed_tx(struct sge *sge, struct cmdQ *q) +{ + unsigned int reclaim = q->processed - q->cleaned; - if (unlikely(offload)) { - { - printk(KERN_ERR - "%s: unexpected offloaded packet, cmd %u\n", - adapter->name, *skb->data); - dev_kfree_skb_any(skb); - } - } else { - struct cpl_rx_pkt *p = (struct cpl_rx_pkt *)skb->data; - - skb_pull(skb, sizeof(*p)); - skb->dev = adapter->port[p->iff].dev; - skb->dev->last_rx = jiffies; - skb->protocol = sge_eth_type_trans(skb, skb->dev); - if ((adapter->flags & RX_CSUM_ENABLED) && p->csum == 0xffff && - skb->protocol == htons(ETH_P_IP) && - (skb->data[9] == IPPROTO_TCP || - skb->data[9] == IPPROTO_UDP)) - skb->ip_summed = CHECKSUM_UNNECESSARY; - else - skb->ip_summed = CHECKSUM_NONE; - if (adapter->vlan_grp && p->vlan_valid) - vlan_hwaccel_rx(skb, adapter->vlan_grp, - ntohs(p->vlan)); - else - netif_rx(skb); + if (reclaim) { + free_cmdQ_buffers(sge, q, reclaim); + q->cleaned += reclaim; } +} - if (++Q->cidx == Q->entries_n) - Q->cidx = 0; +#ifndef SET_ETHTOOL_OPS +# define __netif_rx_complete(dev) netif_rx_complete(dev) +#endif - if (unlikely(--Q->credits < Q->entries_n - SGE_FREEL_REFILL_THRESH)) - refill_free_list(sge, Q); - return 1; +/* + * We cannot use the standard netif_rx_schedule_prep() because we have multiple + * ports plus the TOE all multiplexing onto a single response queue, therefore + * accepting new responses cannot depend on the state of any particular port. + * So define our own equivalent that omits the netif_running() test. + */ +static inline int napi_schedule_prep(struct net_device *dev) +{ + return !test_and_set_bit(__LINK_STATE_RX_SCHED, &dev->state); } -/* - * Adaptive interrupt timer logic to keep the CPU utilization to - * manageable levels. Basically, as the Average Packet Size (APS) - * gets higher, the interrupt latency setting gets longer. Every - * SGE_INTR_BUCKETSIZE (of 100B) causes a bump of 2usec to the - * base value of SGE_INTRTIMER0. At large values of payload the - * latency hits the ceiling value of SGE_INTRTIMER1 stored at - * index SGE_INTR_MAXBUCKETS-1 in sge->intrtimer[]. +/** + * sge_rx - process an ingress ethernet packet + * @sge: the sge structure + * @fl: the free list that contains the packet buffer + * @len: the packet length * - * sge->currIndex caches the last index to save unneeded PIOs. + * Process an ingress ethernet pakcet and deliver it to the stack. */ -static inline void update_intr_timer(struct sge *sge, unsigned int avg_payload) +static int sge_rx(struct sge *sge, struct freelQ *fl, unsigned int len) { - unsigned int newIndex; + struct sk_buff *skb; + struct cpl_rx_pkt *p; + struct adapter *adapter = sge->adapter; - newIndex = avg_payload / SGE_INTR_BUCKETSIZE; - if (newIndex > SGE_INTR_MAXBUCKETS - 1) { - newIndex = SGE_INTR_MAXBUCKETS - 1; - } - /* Save a PIO with this check....maybe */ - if (newIndex != sge->currIndex) { - t1_write_reg_4(sge->adapter, A_SG_INTRTIMER, - sge->intrtimer[newIndex]); - sge->currIndex = newIndex; - sge->adapter->params.sge.last_rx_coalesce_raw = - sge->intrtimer[newIndex]; + sge->stats.ethernet_pkts++; + skb = get_packet(adapter->pdev, fl, len - sge->rx_pkt_pad, + sge->rx_pkt_pad, 2, SGE_RX_COPY_THRES, + SGE_RX_DROP_THRES); + if (!skb) { + sge->port_stats[0].rx_drops++; /* charge only port 0 for now */ + return 0; } + + p = (struct cpl_rx_pkt *)skb->data; + skb_pull(skb, sizeof(*p)); + skb->dev = adapter->port[p->iff].dev; + skb->dev->last_rx = jiffies; + skb->protocol = eth_type_trans(skb, skb->dev); + if ((adapter->flags & RX_CSUM_ENABLED) && p->csum == 0xffff && + skb->protocol == htons(ETH_P_IP) && + (skb->data[9] == IPPROTO_TCP || skb->data[9] == IPPROTO_UDP)) { + sge->port_stats[p->iff].rx_cso_good++; + skb->ip_summed = CHECKSUM_UNNECESSARY; + } else + skb->ip_summed = CHECKSUM_NONE; + + if (unlikely(adapter->vlan_grp && p->vlan_valid)) { + sge->port_stats[p->iff].vlan_xtract++; + if (adapter->params.sge.polling) + vlan_hwaccel_receive_skb(skb, adapter->vlan_grp, + ntohs(p->vlan)); + else + vlan_hwaccel_rx(skb, adapter->vlan_grp, + ntohs(p->vlan)); + } else if (adapter->params.sge.polling) + netif_receive_skb(skb); + else + netif_rx(skb); + return 0; } /* - * Returns true if command queue q_num has enough available descriptors that + * Returns true if a command queue has enough available descriptors that * we can resume Tx operation after temporarily disabling its packet queue. */ -static inline int enough_free_Tx_descs(struct sge *sge, int q_num) +static inline int enough_free_Tx_descs(const struct cmdQ *q) { - return atomic_read(&sge->cmdQ[q_num].credits) > - (sge->cmdQ[q_num].entries_n >> 2); + unsigned int r = q->processed - q->cleaned; + + return q->in_use - r < (q->size >> 1); } /* - * Main interrupt handler, optimized assuming that we took a 'DATA' - * interrupt. - * - * 1. Clear the interrupt - * 2. Loop while we find valid descriptors and process them; accumulate - * information that can be processed after the loop - * 3. Tell the SGE at which index we stopped processing descriptors - * 4. Bookkeeping; free TX buffers, ring doorbell if there are any - * outstanding TX buffers waiting, replenish RX buffers, potentially - * reenable upper layers if they were turned off due to lack of TX - * resources which are available again. - * 5. If we took an interrupt, but no valid respQ descriptors was found we - * let the slow_intr_handler run and do error handling. + * Called when sufficient space has become available in the SGE command queues + * after the Tx packet schedulers have been suspended to restart the Tx path. */ -irqreturn_t t1_interrupt(int irq, void *cookie, struct pt_regs *regs) +static void restart_tx_queues(struct sge *sge) { - struct net_device *netdev; - struct adapter *adapter = cookie; - struct sge *sge = adapter->sge; - struct respQ *Q = &sge->respQ; - unsigned int credits = Q->credits, flags = 0, ret = 0; - unsigned int tot_rxpayload = 0, tot_txpayload = 0, n_rx = 0, n_tx = 0; - unsigned int credits_pend[SGE_CMDQ_N] = { 0, 0 }; + struct adapter *adap = sge->adapter; - struct respQ_e *e = &Q->entries[Q->cidx]; - prefetch(e); + if (enough_free_Tx_descs(&sge->cmdQ[0])) { + int i; + + for_each_port(adap, i) { + struct net_device *nd = adap->port[i].dev; + + if (test_and_clear_bit(nd->if_port, + &sge->stopped_tx_queues) && + netif_running(nd)) { + sge->stats.cmdQ_restarted[3]++; + netif_wake_queue(nd); + } + } + } +} + +/* + * update_tx_info is called from the interrupt handler/NAPI to return cmdQ0 + * information. + */ +static unsigned int update_tx_info(struct adapter *adapter, + unsigned int flags, + unsigned int pr0) +{ + struct sge *sge = adapter->sge; + struct cmdQ *cmdq = &sge->cmdQ[0]; - t1_write_reg_4(adapter, A_PL_CAUSE, F_PL_INTR_SGE_DATA); + cmdq->processed += pr0; + if (flags & F_CMDQ0_ENABLE) { + clear_bit(CMDQ_STAT_RUNNING, &cmdq->status); + + if (cmdq->cleaned + cmdq->in_use != cmdq->processed && + !test_and_set_bit(CMDQ_STAT_LAST_PKT_DB, &cmdq->status)) { + set_bit(CMDQ_STAT_RUNNING, &cmdq->status); + writel(F_CMDQ0_ENABLE, adapter->regs + A_SG_DOORBELL); + } + flags &= ~F_CMDQ0_ENABLE; + } + + if (unlikely(sge->stopped_tx_queues != 0)) + restart_tx_queues(sge); - while (e->GenerationBit == Q->genbit) { - if (--credits < SGE_RESPQ_REPLENISH_THRES) { - u32 n = Q->entries_n - credits - 1; + return flags; +} - t1_write_reg_4(adapter, A_SG_RSPQUEUECREDIT, n); - credits += n; +/* + * Process SGE responses, up to the supplied budget. Returns the number of + * responses processed. A negative budget is effectively unlimited. + */ +static int process_responses(struct adapter *adapter, int budget) +{ + struct sge *sge = adapter->sge; + struct respQ *q = &sge->respQ; + struct respQ_e *e = &q->entries[q->cidx]; + int budget_left = budget; + unsigned int flags = 0; + unsigned int cmdq_processed[SGE_CMDQ_N] = {0, 0}; + + + while (likely(budget_left && e->GenerationBit == q->genbit)) { + flags |= e->Qsleeping; + + cmdq_processed[0] += e->Cmdq0CreditReturn; + cmdq_processed[1] += e->Cmdq1CreditReturn; + + /* We batch updates to the TX side to avoid cacheline + * ping-pong of TX state information on MP where the sender + * might run on a different CPU than this function... + */ + if (unlikely(flags & F_CMDQ0_ENABLE || cmdq_processed[0] > 64)) { + flags = update_tx_info(adapter, flags, cmdq_processed[0]); + cmdq_processed[0] = 0; + } + if (unlikely(cmdq_processed[1] > 16)) { + sge->cmdQ[1].processed += cmdq_processed[1]; + cmdq_processed[1] = 0; } if (likely(e->DataValid)) { - if (!e->Sop || !e->Eop) + struct freelQ *fl = &sge->freelQ[e->FreelistQid]; + + if (unlikely(!e->Sop || !e->Eop)) BUG(); - t1_sge_rx(sge, &sge->freelQ[e->FreelistQid], - e->BufferLength, e->Offload); - tot_rxpayload += e->BufferLength; - ++n_rx; - } - flags |= e->Qsleeping; - credits_pend[0] += e->Cmdq0CreditReturn; - credits_pend[1] += e->Cmdq1CreditReturn; + if (unlikely(e->Offload)) + unexpected_offload(adapter, fl); + else + sge_rx(sge, fl, e->BufferLength); + + /* + * Note: this depends on each packet consuming a + * single free-list buffer; cf. the BUG above. + */ + if (++fl->cidx == fl->size) + fl->cidx = 0; + if (unlikely(--fl->credits < + fl->size - SGE_FREEL_REFILL_THRESH)) + refill_free_list(sge, fl); + } else + sge->stats.pure_rsps++; -#ifdef CONFIG_SMP - /* - * If enough cmdQ0 buffers have finished DMAing free them so - * anyone that may be waiting for their release can continue. - * We do this only on MP systems to allow other CPUs to proceed - * promptly. UP systems can wait for the free_cmdQ_buffers() - * calls after this loop as the sole CPU is currently busy in - * this loop. - */ - if (unlikely(credits_pend[0] > SGE_FREEL_REFILL_THRESH)) { - free_cmdQ_buffers(sge, &sge->cmdQ[0], credits_pend[0], - &tot_txpayload); - n_tx += credits_pend[0]; - credits_pend[0] = 0; - } -#endif - ret++; e++; - if (unlikely(++Q->cidx == Q->entries_n)) { - Q->cidx = 0; - Q->genbit ^= 1; - e = Q->entries; + if (unlikely(++q->cidx == q->size)) { + q->cidx = 0; + q->genbit ^= 1; + e = q->entries; + } + prefetch(e); + + if (++q->credits > SGE_RESPQ_REPLENISH_THRES) { + writel(q->credits, adapter->regs + A_SG_RSPQUEUECREDIT); + q->credits = 0; } + --budget_left; } - Q->credits = credits; - t1_write_reg_4(adapter, A_SG_SLEEPING, Q->cidx); + flags = update_tx_info(adapter, flags, cmdq_processed[0]); + sge->cmdQ[1].processed += cmdq_processed[1]; - if (credits_pend[0]) - free_cmdQ_buffers(sge, &sge->cmdQ[0], credits_pend[0], &tot_txpayload); - if (credits_pend[1]) - free_cmdQ_buffers(sge, &sge->cmdQ[1], credits_pend[1], &tot_txpayload); + budget -= budget_left; + return budget; +} - /* Do any coalescing and interrupt latency timer adjustments */ - if (adapter->params.sge.coalesce_enable) { - unsigned int avg_txpayload = 0, avg_rxpayload = 0; +/* + * A simpler version of process_responses() that handles only pure (i.e., + * non data-carrying) responses. Such respones are too light-weight to justify + * calling a softirq when using NAPI, so we handle them specially in hard + * interrupt context. The function is called with a pointer to a response, + * which the caller must ensure is a valid pure response. Returns 1 if it + * encounters a valid data-carrying response, 0 otherwise. + */ +static int process_pure_responses(struct adapter *adapter, struct respQ_e *e) +{ + struct sge *sge = adapter->sge; + struct respQ *q = &sge->respQ; + unsigned int flags = 0; + unsigned int cmdq_processed[SGE_CMDQ_N] = {0, 0}; - n_tx += credits_pend[0] + credits_pend[1]; + do { + flags |= e->Qsleeping; - /* - * Choose larger avg. payload size to increase - * throughput and reduce [CPU util., intr/s.] - * - * Throughput behavior favored in mixed-mode. - */ - if (n_tx) - avg_txpayload = tot_txpayload/n_tx; - if (n_rx) - avg_rxpayload = tot_rxpayload/n_rx; - - if (n_tx && avg_txpayload > avg_rxpayload){ - update_intr_timer(sge, avg_txpayload); - } else if (n_rx) { - update_intr_timer(sge, avg_rxpayload); + cmdq_processed[0] += e->Cmdq0CreditReturn; + cmdq_processed[1] += e->Cmdq1CreditReturn; + + e++; + if (unlikely(++q->cidx == q->size)) { + q->cidx = 0; + q->genbit ^= 1; + e = q->entries; } - } - - if (flags & F_CMDQ0_ENABLE) { - struct cmdQ *cmdQ = &sge->cmdQ[0]; + prefetch(e); - atomic_set(&cmdQ->asleep, 1); - if (atomic_read(&cmdQ->pio_pidx) != cmdQ->pidx) { - doorbell_pio(sge, F_CMDQ0_ENABLE); - atomic_set(&cmdQ->pio_pidx, cmdQ->pidx); + if (++q->credits > SGE_RESPQ_REPLENISH_THRES) { + writel(q->credits, adapter->regs + A_SG_RSPQUEUECREDIT); + q->credits = 0; } - } - if (unlikely(flags & (F_FL0_ENABLE | F_FL1_ENABLE))) - freelQs_empty(sge); + sge->stats.pure_rsps++; + } while (e->GenerationBit == q->genbit && !e->DataValid); - netdev = adapter->port[0].dev; - if (unlikely(netif_queue_stopped(netdev) && netif_carrier_ok(netdev) && - enough_free_Tx_descs(sge, 0) && - enough_free_Tx_descs(sge, 1))) { - netif_wake_queue(netdev); - } - if (unlikely(!ret)) - ret = t1_slow_intr_handler(adapter); + flags = update_tx_info(adapter, flags, cmdq_processed[0]); + sge->cmdQ[1].processed += cmdq_processed[1]; - return IRQ_RETVAL(ret != 0); + return e->GenerationBit == q->genbit; } /* - * Enqueues the sk_buff onto the cmdQ[qid] and has hardware fetch it. - * - * The code figures out how many entries the sk_buff will require in the - * cmdQ and updates the cmdQ data structure with the state once the enqueue - * has complete. Then, it doesn't access the global structure anymore, but - * uses the corresponding fields on the stack. In conjuction with a spinlock - * around that code, we can make the function reentrant without holding the - * lock when we actually enqueue (which might be expensive, especially on - * architectures with IO MMUs). + * Handler for new data events when using NAPI. This does not need any locking + * or protection from interrupts as data interrupts are off at this point and + * other adapter interrupts do not interfere. */ -static unsigned int t1_sge_tx(struct sk_buff *skb, struct adapter *adapter, - unsigned int qid) +static int t1_poll(struct net_device *dev, int *budget) { - struct sge *sge = adapter->sge; - struct cmdQ *Q = &sge->cmdQ[qid]; - struct cmdQ_e *e; - struct cmdQ_ce *ce; - dma_addr_t mapping; - unsigned int credits, pidx, genbit; + struct adapter *adapter = dev->priv; + int effective_budget = min(*budget, dev->quota); + + int work_done = process_responses(adapter, effective_budget); + *budget -= work_done; + dev->quota -= work_done; - unsigned int count = 1 + skb_shinfo(skb)->nr_frags; + if (work_done >= effective_budget) + return 1; + + __netif_rx_complete(dev); /* - * Coming from the timer + * Because we don't atomically flush the following write it is + * possible that in very rare cases it can reach the device in a way + * that races with a new response being written plus an error interrupt + * causing the NAPI interrupt handler below to return unhandled status + * to the OS. To protect against this would require flushing the write + * and doing both the write and the flush with interrupts off. Way too + * expensive and unjustifiable given the rarity of the race. */ - if ((skb == sge->pskb)) { - /* - * Quit if any cmdQ activities - */ - if (!spin_trylock(&Q->Qlock)) - return 0; - if (atomic_read(&Q->credits) != Q->entries_n) { - spin_unlock(&Q->Qlock); - return 0; - } - } - else - spin_lock(&Q->Qlock); - - genbit = Q->genbit; - pidx = Q->pidx; - credits = atomic_read(&Q->credits); - - credits -= count; - atomic_sub(count, &Q->credits); - Q->pidx += count; - if (Q->pidx >= Q->entries_n) { - Q->pidx -= Q->entries_n; - Q->genbit ^= 1; - } + writel(adapter->sge->respQ.cidx, adapter->regs + A_SG_SLEEPING); + return 0; +} - if (unlikely(credits < (MAX_SKB_FRAGS + 1))) { - sge->intr_cnt.cmdQ_full[qid]++; - netif_stop_queue(adapter->port[0].dev); - } - spin_unlock(&Q->Qlock); +/* + * Returns true if the device is already scheduled for polling. + */ +static inline int napi_is_scheduled(struct net_device *dev) +{ + return test_bit(__LINK_STATE_RX_SCHED, &dev->state); +} - mapping = pci_map_single(adapter->pdev, skb->data, - skb->len - skb->data_len, PCI_DMA_TODEVICE); - ce = &Q->centries[pidx]; - ce->skb = NULL; - pci_unmap_addr_set(ce, dma_addr, mapping); - pci_unmap_len_set(ce, dma_len, skb->len - skb->data_len); - ce->single = 1; +/* + * NAPI version of the main interrupt handler. + */ +static irqreturn_t t1_interrupt_napi(int irq, void *data, struct pt_regs *regs) +{ + int handled; + struct adapter *adapter = data; + struct sge *sge = adapter->sge; + struct respQ *q = &adapter->sge->respQ; - e = &Q->entries[pidx]; - e->Sop = 1; - e->DataValid = 1; - e->BufferLength = skb->len - skb->data_len; - e->AddrHigh = (u64)mapping >> 32; - e->AddrLow = (u32)mapping; + /* + * Clear the SGE_DATA interrupt first thing. Normally the NAPI + * handler has control of the response queue and the interrupt handler + * can look at the queue reliably only once it knows NAPI is off. + * We can't wait that long to clear the SGE_DATA interrupt because we + * could race with t1_poll rearming the SGE interrupt, so we need to + * clear the interrupt speculatively and really early on. + */ + writel(F_PL_INTR_SGE_DATA, adapter->regs + A_PL_CAUSE); + + spin_lock(&adapter->async_lock); + if (!napi_is_scheduled(sge->netdev)) { + struct respQ_e *e = &q->entries[q->cidx]; + + if (e->GenerationBit == q->genbit) { + if (e->DataValid || + process_pure_responses(adapter, e)) { + if (likely(napi_schedule_prep(sge->netdev))) + __netif_rx_schedule(sge->netdev); + else + printk(KERN_CRIT + "NAPI schedule failure!\n"); + } else + writel(q->cidx, adapter->regs + A_SG_SLEEPING); + handled = 1; + goto unlock; + } else + writel(q->cidx, adapter->regs + A_SG_SLEEPING); + } else + if (readl(adapter->regs + A_PL_CAUSE) & F_PL_INTR_SGE_DATA) + printk(KERN_ERR "data interrupt while NAPI running\n"); + + handled = t1_slow_intr_handler(adapter); + if (!handled) + sge->stats.unhandled_irqs++; + unlock: + spin_unlock(&adapter->async_lock); + return IRQ_RETVAL(handled != 0); +} - if (--count > 0) { - unsigned int i; +/* + * Main interrupt handler, optimized assuming that we took a 'DATA' + * interrupt. + * + * 1. Clear the interrupt + * 2. Loop while we find valid descriptors and process them; accumulate + * information that can be processed after the loop + * 3. Tell the SGE at which index we stopped processing descriptors + * 4. Bookkeeping; free TX buffers, ring doorbell if there are any + * outstanding TX buffers waiting, replenish RX buffers, potentially + * reenable upper layers if they were turned off due to lack of TX + * resources which are available again. + * 5. If we took an interrupt, but no valid respQ descriptors was found we + * let the slow_intr_handler run and do error handling. + */ +static irqreturn_t t1_interrupt(int irq, void *cookie, struct pt_regs *regs) +{ + int work_done; + struct respQ_e *e; + struct adapter *adapter = cookie; + struct respQ *Q = &adapter->sge->respQ; - e->Eop = 0; - wmb(); - e->GenerationBit = e->GenerationBit2 = genbit; + spin_lock(&adapter->async_lock); + e = &Q->entries[Q->cidx]; + prefetch(e); - for (i = 0; i < count; i++) { - skb_frag_t *frag = &skb_shinfo(skb)->frags[i]; + writel(F_PL_INTR_SGE_DATA, adapter->regs + A_PL_CAUSE); - ce++; e++; - if (++pidx == Q->entries_n) { - pidx = 0; - genbit ^= 1; - ce = Q->centries; - e = Q->entries; - } + if (likely(e->GenerationBit == Q->genbit)) + work_done = process_responses(adapter, -1); + else + work_done = t1_slow_intr_handler(adapter); - mapping = pci_map_page(adapter->pdev, frag->page, - frag->page_offset, - frag->size, - PCI_DMA_TODEVICE); - ce->skb = NULL; - pci_unmap_addr_set(ce, dma_addr, mapping); - pci_unmap_len_set(ce, dma_len, frag->size); - ce->single = 0; - - e->Sop = 0; - e->DataValid = 1; - e->BufferLength = frag->size; - e->AddrHigh = (u64)mapping >> 32; - e->AddrLow = (u32)mapping; - - if (i < count - 1) { - e->Eop = 0; - wmb(); - e->GenerationBit = e->GenerationBit2 = genbit; - } + /* + * The unconditional clearing of the PL_CAUSE above may have raced + * with DMA completion and the corresponding generation of a response + * to cause us to miss the resulting data interrupt. The next write + * is also unconditional to recover the missed interrupt and render + * this race harmless. + */ + writel(Q->cidx, adapter->regs + A_SG_SLEEPING); + + if (!work_done) + adapter->sge->stats.unhandled_irqs++; + spin_unlock(&adapter->async_lock); + return IRQ_RETVAL(work_done != 0); +} + +intr_handler_t t1_select_intr_handler(adapter_t *adapter) +{ + return adapter->params.sge.polling ? t1_interrupt_napi : t1_interrupt; +} + +/* + * Enqueues the sk_buff onto the cmdQ[qid] and has hardware fetch it. + * + * The code figures out how many entries the sk_buff will require in the + * cmdQ and updates the cmdQ data structure with the state once the enqueue + * has complete. Then, it doesn't access the global structure anymore, but + * uses the corresponding fields on the stack. In conjuction with a spinlock + * around that code, we can make the function reentrant without holding the + * lock when we actually enqueue (which might be expensive, especially on + * architectures with IO MMUs). + * + * This runs with softirqs disabled. + */ +unsigned int t1_sge_tx(struct sk_buff *skb, struct adapter *adapter, + unsigned int qid, struct net_device *dev) +{ + struct sge *sge = adapter->sge; + struct cmdQ *q = &sge->cmdQ[qid]; + unsigned int credits, pidx, genbit, count; + + spin_lock(&q->lock); + reclaim_completed_tx(sge, q); + + pidx = q->pidx; + credits = q->size - q->in_use; + count = 1 + skb_shinfo(skb)->nr_frags; + + { /* Ethernet packet */ + if (unlikely(credits < count)) { + netif_stop_queue(dev); + set_bit(dev->if_port, &sge->stopped_tx_queues); + sge->stats.cmdQ_full[3]++; + spin_unlock(&q->lock); + CH_ERR("%s: Tx ring full while queue awake!\n", + adapter->name); + return 1; } + if (unlikely(credits - count < q->stop_thres)) { + sge->stats.cmdQ_full[3]++; + netif_stop_queue(dev); + set_bit(dev->if_port, &sge->stopped_tx_queues); + } + } + q->in_use += count; + genbit = q->genbit; + q->pidx += count; + if (q->pidx >= q->size) { + q->pidx -= q->size; + q->genbit ^= 1; } + spin_unlock(&q->lock); - if (skb != sge->pskb) - ce->skb = skb; - e->Eop = 1; - wmb(); - e->GenerationBit = e->GenerationBit2 = genbit; + write_tx_descs(adapter, skb, pidx, genbit, q); /* * We always ring the doorbell for cmdQ1. For cmdQ0, we only ring @@ -1317,12 +1380,14 @@ static unsigned int t1_sge_tx(struct sk_buff *skb, struct adapter *adapter, * then the interrupt handler will detect the outstanding TX packet * and ring the doorbell for us. */ - if (qid) { - doorbell_pio(sge, F_CMDQ1_ENABLE); - } else if (atomic_read(&Q->asleep)) { - atomic_set(&Q->asleep, 0); - doorbell_pio(sge, F_CMDQ0_ENABLE); - atomic_set(&Q->pio_pidx, Q->pidx); + if (qid) + doorbell_pio(adapter, F_CMDQ1_ENABLE); + else { + clear_bit(CMDQ_STAT_LAST_PKT_DB, &q->status); + if (test_and_set_bit(CMDQ_STAT_RUNNING, &q->status) == 0) { + set_bit(CMDQ_STAT_LAST_PKT_DB, &q->status); + writel(F_CMDQ0_ENABLE, adapter->regs + A_SG_DOORBELL); + } } return 0; } @@ -1330,37 +1395,35 @@ static unsigned int t1_sge_tx(struct sk_buff *skb, struct adapter *adapter, #define MK_ETH_TYPE_MSS(type, mss) (((mss) & 0x3FFF) | ((type) << 14)) /* + * eth_hdr_len - return the length of an Ethernet header + * @data: pointer to the start of the Ethernet header + * + * Returns the length of an Ethernet header, including optional VLAN tag. + */ +static inline int eth_hdr_len(const void *data) +{ + const struct ethhdr *e = data; + + return e->h_proto == htons(ETH_P_8021Q) ? VLAN_ETH_HLEN : ETH_HLEN; +} + +/* * Adds the CPL header to the sk_buff and passes it to t1_sge_tx. */ int t1_start_xmit(struct sk_buff *skb, struct net_device *dev) { struct adapter *adapter = dev->priv; + struct sge_port_stats *st = &adapter->sge->port_stats[dev->if_port]; + struct sge *sge = adapter->sge; struct cpl_tx_pkt *cpl; - struct ethhdr *eth; - size_t max_len; - - /* - * We are using a non-standard hard_header_len and some kernel - * components, such as pktgen, do not handle it right. Complain - * when this happens but try to fix things up. - */ - if (unlikely(skb_headroom(skb) < dev->hard_header_len - ETH_HLEN)) { - struct sk_buff *orig_skb = skb; - - if (net_ratelimit()) - printk(KERN_ERR - "%s: Tx packet has inadequate headroom\n", - dev->name); - skb = skb_realloc_headroom(skb, sizeof(struct cpl_tx_pkt_lso)); - dev_kfree_skb_any(orig_skb); - if (!skb) - return -ENOMEM; - } +#ifdef NETIF_F_TSO if (skb_shinfo(skb)->tso_size) { int eth_type; struct cpl_tx_pkt_lso *hdr; + st->tso++; + eth_type = skb->nh.raw - skb->data == ETH_HLEN ? CPL_ETH_II : CPL_ETH_II_VLAN; @@ -1373,40 +1436,72 @@ int t1_start_xmit(struct sk_buff *skb, struct net_device *dev) skb_shinfo(skb)->tso_size)); hdr->len = htonl(skb->len - sizeof(*hdr)); cpl = (struct cpl_tx_pkt *)hdr; + sge->stats.tx_lso_pkts++; } else +#endif { /* - * An Ethernet packet must have at least space for - * the DIX Ethernet header and be no greater than - * the device set MTU. Otherwise trash the packet. + * Packets shorter than ETH_HLEN can break the MAC, drop them + * early. Also, we may get oversized packets because some + * parts of the kernel don't handle our unusual hard_header_len + * right, drop those too. */ - if (skb->len < ETH_HLEN) - goto t1_start_xmit_fail2; - eth = (struct ethhdr *)skb->data; - if (eth->h_proto == htons(ETH_P_8021Q)) - max_len = dev->mtu + VLAN_ETH_HLEN; - else - max_len = dev->mtu + ETH_HLEN; - if (skb->len > max_len) - goto t1_start_xmit_fail2; + if (unlikely(skb->len < ETH_HLEN || + skb->len > dev->mtu + eth_hdr_len(skb->data))) { + dev_kfree_skb_any(skb); + return NET_XMIT_SUCCESS; + } + + /* + * We are using a non-standard hard_header_len and some kernel + * components, such as pktgen, do not handle it right. + * Complain when this happens but try to fix things up. + */ + if (unlikely(skb_headroom(skb) < + dev->hard_header_len - ETH_HLEN)) { + struct sk_buff *orig_skb = skb; + + if (net_ratelimit()) + printk(KERN_ERR "%s: inadequate headroom in " + "Tx packet\n", dev->name); + skb = skb_realloc_headroom(skb, sizeof(*cpl)); + dev_kfree_skb_any(orig_skb); + if (!skb) + return -ENOMEM; + } if (!(adapter->flags & UDP_CSUM_CAPABLE) && skb->ip_summed == CHECKSUM_HW && - skb->nh.iph->protocol == IPPROTO_UDP && - skb_checksum_help(skb, 0)) - goto t1_start_xmit_fail3; - + skb->nh.iph->protocol == IPPROTO_UDP) + if (unlikely(skb_checksum_help(skb, 0))) { + dev_kfree_skb_any(skb); + return -ENOMEM; + } - if (!adapter->sge->pskb) { + /* Hmmm, assuming to catch the gratious arp... and we'll use + * it to flush out stuck espi packets... + */ + if (unlikely(!adapter->sge->espibug_skb)) { if (skb->protocol == htons(ETH_P_ARP) && - skb->nh.arph->ar_op == htons(ARPOP_REQUEST)) - adapter->sge->pskb = skb; + skb->nh.arph->ar_op == htons(ARPOP_REQUEST)) { + adapter->sge->espibug_skb = skb; + /* We want to re-use this skb later. We + * simply bump the reference count and it + * will not be freed... + */ + skb = skb_get(skb); + } } - cpl = (struct cpl_tx_pkt *)skb_push(skb, sizeof(*cpl)); + + cpl = (struct cpl_tx_pkt *)__skb_push(skb, sizeof(*cpl)); cpl->opcode = CPL_TX_PKT; cpl->ip_csum_dis = 1; /* SW calculates IP csum */ cpl->l4_csum_dis = skb->ip_summed == CHECKSUM_HW ? 0 : 1; /* the length field isn't used so don't bother setting it */ + + st->tx_cso += (skb->ip_summed == CHECKSUM_HW); + sge->stats.tx_do_cksum += (skb->ip_summed == CHECKSUM_HW); + sge->stats.tx_reg_pkts++; } cpl->iff = dev->if_port; @@ -1414,38 +1509,176 @@ int t1_start_xmit(struct sk_buff *skb, struct net_device *dev) if (adapter->vlan_grp && vlan_tx_tag_present(skb)) { cpl->vlan_valid = 1; cpl->vlan = htons(vlan_tx_tag_get(skb)); + st->vlan_insert++; } else #endif cpl->vlan_valid = 0; dev->trans_start = jiffies; - return t1_sge_tx(skb, adapter, 0); + return t1_sge_tx(skb, adapter, 0, dev); +} -t1_start_xmit_fail3: - printk(KERN_INFO "%s: Unable to complete checksum\n", dev->name); - goto t1_start_xmit_fail1; +/* + * Callback for the Tx buffer reclaim timer. Runs with softirqs disabled. + */ +static void sge_tx_reclaim_cb(unsigned long data) +{ + int i; + struct sge *sge = (struct sge *)data; + + for (i = 0; i < SGE_CMDQ_N; ++i) { + struct cmdQ *q = &sge->cmdQ[i]; + + if (!spin_trylock(&q->lock)) + continue; -t1_start_xmit_fail2: - printk(KERN_INFO "%s: Invalid packet length %d, dropping\n", - dev->name, skb->len); + reclaim_completed_tx(sge, q); + if (i == 0 && q->in_use) /* flush pending credits */ + writel(F_CMDQ0_ENABLE, + sge->adapter->regs + A_SG_DOORBELL); -t1_start_xmit_fail1: - dev_kfree_skb_any(skb); + spin_unlock(&q->lock); + } + mod_timer(&sge->tx_reclaim_timer, jiffies + TX_RECLAIM_PERIOD); +} + +/* + * Propagate changes of the SGE coalescing parameters to the HW. + */ +int t1_sge_set_coalesce_params(struct sge *sge, struct sge_params *p) +{ + sge->netdev->poll = t1_poll; + sge->fixed_intrtimer = p->rx_coalesce_usecs * + core_ticks_per_usec(sge->adapter); + writel(sge->fixed_intrtimer, sge->adapter->regs + A_SG_INTRTIMER); return 0; } -void t1_sge_set_ptimeout(adapter_t *adapter, u32 val) +/* + * Allocates both RX and TX resources and configures the SGE. However, + * the hardware is not enabled yet. + */ +int t1_sge_configure(struct sge *sge, struct sge_params *p) { - struct sge *sge = adapter->sge; + if (alloc_rx_resources(sge, p)) + return -ENOMEM; + if (alloc_tx_resources(sge, p)) { + free_rx_resources(sge); + return -ENOMEM; + } + configure_sge(sge, p); + + /* + * Now that we have sized the free lists calculate the payload + * capacity of the large buffers. Other parts of the driver use + * this to set the max offload coalescing size so that RX packets + * do not overflow our large buffers. + */ + p->large_buf_capacity = jumbo_payload_capacity(sge); + return 0; +} - if (is_T2(adapter)) - sge->ptimeout = max((u32)((HZ * val) / 1000), (u32)1); +/* + * Disables the DMA engine. + */ +void t1_sge_stop(struct sge *sge) +{ + writel(0, sge->adapter->regs + A_SG_CONTROL); + (void) readl(sge->adapter->regs + A_SG_CONTROL); /* flush */ + if (is_T2(sge->adapter)) + del_timer_sync(&sge->espibug_timer); + del_timer_sync(&sge->tx_reclaim_timer); } -u32 t1_sge_get_ptimeout(adapter_t *adapter) +/* + * Enables the DMA engine. + */ +void t1_sge_start(struct sge *sge) { + refill_free_list(sge, &sge->freelQ[0]); + refill_free_list(sge, &sge->freelQ[1]); + + writel(sge->sge_control, sge->adapter->regs + A_SG_CONTROL); + doorbell_pio(sge->adapter, F_FL0_ENABLE | F_FL1_ENABLE); + (void) readl(sge->adapter->regs + A_SG_CONTROL); /* flush */ + + mod_timer(&sge->tx_reclaim_timer, jiffies + TX_RECLAIM_PERIOD); + + if (is_T2(sge->adapter)) + mod_timer(&sge->espibug_timer, jiffies + sge->espibug_timeout); +} + +/* + * Callback for the T2 ESPI 'stuck packet feature' workaorund + */ +static void espibug_workaround(void *data) +{ + struct adapter *adapter = (struct adapter *)data; struct sge *sge = adapter->sge; - return (is_T2(adapter) ? ((sge->ptimeout * 1000) / HZ) : 0); + if (netif_running(adapter->port[0].dev)) { + struct sk_buff *skb = sge->espibug_skb; + + u32 seop = t1_espi_get_mon(adapter, 0x930, 0); + + if ((seop & 0xfff0fff) == 0xfff && skb) { + if (!skb->cb[0]) { + u8 ch_mac_addr[ETH_ALEN] = + {0x0, 0x7, 0x43, 0x0, 0x0, 0x0}; + memcpy(skb->data + sizeof(struct cpl_tx_pkt), + ch_mac_addr, ETH_ALEN); + memcpy(skb->data + skb->len - 10, ch_mac_addr, + ETH_ALEN); + skb->cb[0] = 0xff; + } + + /* bump the reference count to avoid freeing of the + * skb once the DMA has completed. + */ + skb = skb_get(skb); + t1_sge_tx(skb, adapter, 0, adapter->port[0].dev); + } + } + mod_timer(&sge->espibug_timer, jiffies + sge->espibug_timeout); } +/* + * Creates a t1_sge structure and returns suggested resource parameters. + */ +struct sge * __devinit t1_sge_create(struct adapter *adapter, + struct sge_params *p) +{ + struct sge *sge = kmalloc(sizeof(*sge), GFP_KERNEL); + + if (!sge) + return NULL; + memset(sge, 0, sizeof(*sge)); + + sge->adapter = adapter; + sge->netdev = adapter->port[0].dev; + sge->rx_pkt_pad = t1_is_T1B(adapter) ? 0 : 2; + sge->jumbo_fl = t1_is_T1B(adapter) ? 1 : 0; + + init_timer(&sge->tx_reclaim_timer); + sge->tx_reclaim_timer.data = (unsigned long)sge; + sge->tx_reclaim_timer.function = sge_tx_reclaim_cb; + + if (is_T2(sge->adapter)) { + init_timer(&sge->espibug_timer); + sge->espibug_timer.function = (void *)&espibug_workaround; + sge->espibug_timer.data = (unsigned long)sge->adapter; + sge->espibug_timeout = 1; + } + + + p->cmdQ_size[0] = SGE_CMDQ0_E_N; + p->cmdQ_size[1] = SGE_CMDQ1_E_N; + p->freelQ_size[!sge->jumbo_fl] = SGE_FREEL_SIZE; + p->freelQ_size[sge->jumbo_fl] = SGE_JUMBO_FREEL_SIZE; + p->rx_coalesce_usecs = 50; + p->coalesce_enable = 0; + p->sample_interval_usecs = 0; + p->polling = 0; + + return sge; +} -- cgit v1.1