#include <linux/vmalloc.h>
#include "ipath_kernel.h"
-#include "ips_common.h"
#include "ipath_layer.h"
+#include "ipath_common.h"
static void ipath_update_pio_bufs(struct ipath_devdata *);
.id_table = ipath_pci_tbl,
};
-/*
- * This is where port 0's rcvhdrtail register is written back; we also
- * want nothing else sharing the cache line, so make it a cache line
- * in size. Used for all units.
- */
-volatile __le64 *ipath_port0_rcvhdrtail;
-dma_addr_t ipath_port0_rcvhdrtail_dma;
-static int port0_rcvhdrtail_refs;
static inline void read_bars(struct ipath_devdata *dd, struct pci_dev *dev,
u32 *bar0, u32 *bar1)
return nunits;
}
-static int init_port0_rcvhdrtail(struct pci_dev *pdev)
-{
- int ret;
-
- mutex_lock(&ipath_mutex);
-
- if (!ipath_port0_rcvhdrtail) {
- ipath_port0_rcvhdrtail =
- dma_alloc_coherent(&pdev->dev,
- IPATH_PORT0_RCVHDRTAIL_SIZE,
- &ipath_port0_rcvhdrtail_dma,
- GFP_KERNEL);
-
- if (!ipath_port0_rcvhdrtail) {
- ret = -ENOMEM;
- goto bail;
- }
- }
- port0_rcvhdrtail_refs++;
- ret = 0;
-
-bail:
- mutex_unlock(&ipath_mutex);
-
- return ret;
-}
-
-static void cleanup_port0_rcvhdrtail(struct pci_dev *pdev)
-{
- mutex_lock(&ipath_mutex);
-
- if (!--port0_rcvhdrtail_refs) {
- dma_free_coherent(&pdev->dev, IPATH_PORT0_RCVHDRTAIL_SIZE,
- (void *) ipath_port0_rcvhdrtail,
- ipath_port0_rcvhdrtail_dma);
- ipath_port0_rcvhdrtail = NULL;
- }
-
- mutex_unlock(&ipath_mutex);
-}
-
/*
* These next two routines are placeholders in case we don't have per-arch
* code for controlling write combining. If explicit control of write
u32 bar0 = 0, bar1 = 0;
u8 rev;
- ret = init_port0_rcvhdrtail(pdev);
- if (ret < 0) {
- printk(KERN_ERR IPATH_DRV_NAME
- ": Could not allocate port0_rcvhdrtail: error %d\n",
- -ret);
- goto bail;
- }
-
dd = ipath_alloc_devdata(pdev);
if (IS_ERR(dd)) {
ret = PTR_ERR(dd);
printk(KERN_ERR IPATH_DRV_NAME
": Could not allocate devdata: error %d\n", -ret);
- goto bail_rcvhdrtail;
+ goto bail;
}
ipath_cdbg(VERBOSE, "initializing unit #%u\n", dd->ipath_unit);
* by ipath_setup_htconfig.
*/
dd->ipath_flags = 0;
+ dd->ipath_lli_counter = 0;
+ dd->ipath_lli_errors = 0;
if (dd->ipath_f_bus(dd, pdev))
ipath_dev_err(dd, "Failed to setup config space; "
"continuing anyway\n");
/*
- * set up our interrupt handler; SA_SHIRQ probably not needed,
+ * set up our interrupt handler; IRQF_SHARED probably not needed,
* since MSI interrupts shouldn't be shared but won't hurt for now.
* check 0 irq after we return from chip-specific bus setup, since
* that can affect this due to setup
ipath_dev_err(dd, "irq is 0, BIOS error? Interrupts won't "
"work\n");
else {
- ret = request_irq(pdev->irq, ipath_intr, SA_SHIRQ,
+ ret = request_irq(pdev->irq, ipath_intr, IRQF_SHARED,
IPATH_DRV_NAME, dd);
if (ret) {
ipath_dev_err(dd, "Couldn't setup irq handler, "
bail_devdata:
ipath_free_devdata(pdev, dd);
-bail_rcvhdrtail:
- cleanup_port0_rcvhdrtail(pdev);
-
bail:
return ret;
}
pci_disable_device(pdev);
ipath_free_devdata(pdev, dd);
- cleanup_port0_rcvhdrtail(pdev);
}
/* general driver use */
u8 pad, *bthbytes;
struct sk_buff *skb, *nskb;
- if (dd->ipath_port0_skbs && hdr->sub_opcode == OPCODE_ENCAP) {
+ if (dd->ipath_port0_skbs &&
+ hdr->sub_opcode == IPATH_ITH4X_OPCODE_ENCAP) {
/*
* Allocate a new sk_buff to replace the one we give
* to the network stack.
/* another ether packet received */
ipath_stats.sps_ether_rpkts++;
}
- else if (hdr->sub_opcode == OPCODE_LID_ARP)
+ else if (hdr->sub_opcode == IPATH_ITH4X_OPCODE_LID_ARP)
__ipath_layer_rcv_lid(dd, hdr);
}
const u32 rsize = dd->ipath_rcvhdrentsize; /* words */
const u32 maxcnt = dd->ipath_rcvhdrcnt * rsize; /* words */
u32 etail = -1, l, hdrqtail;
- struct ips_message_header *hdr;
- u32 eflags, i, etype, tlen, pkttot = 0;
+ struct ipath_message_header *hdr;
+ u32 eflags, i, etype, tlen, pkttot = 0, updegr=0, reloop=0;
static u64 totcalls; /* stats, may eventually remove */
char emsg[128];
if (test_and_set_bit(0, &dd->ipath_rcv_pending))
goto bail;
- if (dd->ipath_port0head ==
- (u32)le64_to_cpu(*dd->ipath_hdrqtailptr))
+ l = dd->ipath_port0head;
+ hdrqtail = (u32) le64_to_cpu(*dd->ipath_hdrqtailptr);
+ if (l == hdrqtail)
goto done;
-gotmore:
- /*
- * read only once at start. If in flood situation, this helps
- * performance slightly. If more arrive while we are processing,
- * we'll come back here and do them
- */
- hdrqtail = (u32)le64_to_cpu(*dd->ipath_hdrqtailptr);
-
- for (i = 0, l = dd->ipath_port0head; l != hdrqtail; i++) {
+reloop:
+ for (i = 0; l != hdrqtail; i++) {
u32 qp;
u8 *bthbytes;
rc = (u64 *) (dd->ipath_pd[0]->port_rcvhdrq + (l << 2));
- hdr = (struct ips_message_header *)&rc[1];
+ hdr = (struct ipath_message_header *)&rc[1];
/*
* could make a network order version of IPATH_KD_QP, and
* do the obvious shift before masking to speed this up.
qp = ntohl(hdr->bth[1]) & 0xffffff;
bthbytes = (u8 *) hdr->bth;
- eflags = ips_get_hdr_err_flags((__le32 *) rc);
- etype = ips_get_rcv_type((__le32 *) rc);
+ eflags = ipath_hdrget_err_flags((__le32 *) rc);
+ etype = ipath_hdrget_rcv_type((__le32 *) rc);
/* total length */
- tlen = ips_get_length_in_bytes((__le32 *) rc);
+ tlen = ipath_hdrget_length_in_bytes((__le32 *) rc);
ebuf = NULL;
if (etype != RCVHQ_RCV_TYPE_EXPECTED) {
/*
* set ebuf (so we try to copy data) unless the
* length requires it.
*/
- etail = ips_get_index((__le32 *) rc);
+ etail = ipath_hdrget_index((__le32 *) rc);
if (tlen > sizeof(*hdr) ||
etype == RCVHQ_RCV_TYPE_NON_KD)
ebuf = ipath_get_egrbuf(dd, etail, 0);
*/
if (etype != RCVHQ_RCV_TYPE_NON_KD && etype !=
- RCVHQ_RCV_TYPE_ERROR && ips_get_ipath_ver(
+ RCVHQ_RCV_TYPE_ERROR && ipath_hdrget_ipath_ver(
hdr->iph.ver_port_tid_offset) !=
IPS_PROTO_VERSION) {
ipath_cdbg(PKT, "Bad InfiniPath protocol version "
ipath_cdbg(PKT, "RHFerrs %x hdrqtail=%x typ=%u "
"tlen=%x opcode=%x egridx=%x: %s\n",
eflags, l, etype, tlen, bthbytes[0],
- ips_get_index((__le32 *) rc), emsg);
+ ipath_hdrget_index((__le32 *) rc), emsg);
+ /* Count local link integrity errors. */
+ if (eflags & (INFINIPATH_RHF_H_ICRCERR |
+ INFINIPATH_RHF_H_VCRCERR)) {
+ u8 n = (dd->ipath_ibcctrl >>
+ INFINIPATH_IBCC_PHYERRTHRESHOLD_SHIFT) &
+ INFINIPATH_IBCC_PHYERRTHRESHOLD_MASK;
+
+ if (++dd->ipath_lli_counter > n) {
+ dd->ipath_lli_counter = 0;
+ dd->ipath_lli_errors++;
+ }
+ }
} else if (etype == RCVHQ_RCV_TYPE_NON_KD) {
int ret = __ipath_verbs_rcv(dd, rc + 1,
ebuf, tlen);
ipath_cdbg(VERBOSE,
"received IB packet, "
"not SMA (QP=%x)\n", qp);
+ if (dd->ipath_lli_counter)
+ dd->ipath_lli_counter--;
+
} else if (etype == RCVHQ_RCV_TYPE_EAGER) {
if (qp == IPATH_KD_QP &&
bthbytes[0] == ipath_layer_rcv_opcode &&
l += rsize;
if (l >= maxcnt)
l = 0;
+ if (etype != RCVHQ_RCV_TYPE_EXPECTED)
+ updegr = 1;
/*
- * update for each packet, to help prevent overflows if we
- * have lots of packets.
+ * update head regs on last packet, and every 16 packets.
+ * Reduce bus traffic, while still trying to prevent
+ * rcvhdrq overflows, for when the queue is nearly full
*/
- (void)ipath_write_ureg(dd, ur_rcvhdrhead,
- dd->ipath_rhdrhead_intr_off | l, 0);
- if (etype != RCVHQ_RCV_TYPE_EXPECTED)
- (void)ipath_write_ureg(dd, ur_rcvegrindexhead,
- etail, 0);
+ if (l == hdrqtail || (i && !(i&0xf))) {
+ u64 lval;
+ if (l == hdrqtail) /* PE-800 interrupt only on last */
+ lval = dd->ipath_rhdrhead_intr_off | l;
+ else
+ lval = l;
+ (void)ipath_write_ureg(dd, ur_rcvhdrhead, lval, 0);
+ if (updegr) {
+ (void)ipath_write_ureg(dd, ur_rcvegrindexhead,
+ etail, 0);
+ updegr = 0;
+ }
+ }
+ }
+
+ if (!dd->ipath_rhdrhead_intr_off && !reloop) {
+ /* HT-400 workaround; we can have a race clearing chip
+ * interrupt with another interrupt about to be delivered,
+ * and can clear it before it is delivered on the GPIO
+ * workaround. By doing the extra check here for the
+ * in-memory tail register updating while we were doing
+ * earlier packets, we "almost" guarantee we have covered
+ * that case.
+ */
+ u32 hqtail = (u32)le64_to_cpu(*dd->ipath_hdrqtailptr);
+ if (hqtail != hdrqtail) {
+ hdrqtail = hqtail;
+ reloop = 1; /* loop 1 extra time at most */
+ goto reloop;
+ }
}
pkttot += i;
dd->ipath_port0head = l;
- if (hdrqtail != (u32)le64_to_cpu(*dd->ipath_hdrqtailptr))
- /* more arrived while we handled first batch */
- goto gotmore;
-
if (pkttot > ipath_stats.sps_maxpkts_call)
ipath_stats.sps_maxpkts_call = pkttot;
ipath_stats.sps_port0pkts += pkttot;
* @dd: the infinipath device
* @pd: the port data
*
- * this *must* be physically contiguous memory, and for now,
- * that limits it to what kmalloc can do.
+ * this must be contiguous memory (from an i/o perspective), and must be
+ * DMA'able (which means for some systems, it will go through an IOMMU,
+ * or be forced into a low address range).
*/
int ipath_create_rcvhdrq(struct ipath_devdata *dd,
struct ipath_portdata *pd)
{
- int ret = 0, amt;
+ int ret = 0;
- amt = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
- sizeof(u32), PAGE_SIZE);
if (!pd->port_rcvhdrq) {
- /*
- * not using REPEAT isn't viable; at 128KB, we can easily
- * fail this. The problem with REPEAT is we can block here
- * "forever". There isn't an inbetween, unfortunately. We
- * could reduce the risk by never freeing the rcvhdrq except
- * at unload, but even then, the first time a port is used,
- * we could delay for some time...
- */
+ dma_addr_t phys_hdrqtail;
gfp_t gfp_flags = GFP_USER | __GFP_COMP;
+ int amt = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
+ sizeof(u32), PAGE_SIZE);
pd->port_rcvhdrq = dma_alloc_coherent(
&dd->pcidev->dev, amt, &pd->port_rcvhdrq_phys,
ret = -ENOMEM;
goto bail;
}
+ pd->port_rcvhdrtail_kvaddr = dma_alloc_coherent(
+ &dd->pcidev->dev, PAGE_SIZE, &phys_hdrqtail, GFP_KERNEL);
+ if (!pd->port_rcvhdrtail_kvaddr) {
+ ipath_dev_err(dd, "attempt to allocate 1 page "
+ "for port %u rcvhdrqtailaddr failed\n",
+ pd->port_port);
+ ret = -ENOMEM;
+ goto bail;
+ }
+ pd->port_rcvhdrqtailaddr_phys = phys_hdrqtail;
pd->port_rcvhdrq_size = amt;
(unsigned long) pd->port_rcvhdrq_phys,
(unsigned long) pd->port_rcvhdrq_size,
pd->port_port);
- } else {
- /*
- * clear for security, sanity, and/or debugging, each
- * time we reuse
- */
- memset(pd->port_rcvhdrq, 0, amt);
+
+ ipath_cdbg(VERBOSE, "port %d hdrtailaddr, %llx physical\n",
+ pd->port_port,
+ (unsigned long long) phys_hdrqtail);
}
+ else
+ ipath_cdbg(VERBOSE, "reuse port %d rcvhdrq @%p %llx phys; "
+ "hdrtailaddr@%p %llx physical\n",
+ pd->port_port, pd->port_rcvhdrq,
+ pd->port_rcvhdrq_phys, pd->port_rcvhdrtail_kvaddr,
+ (unsigned long long)pd->port_rcvhdrqtailaddr_phys);
+
+ /* clear for security and sanity on each use */
+ memset(pd->port_rcvhdrq, 0, pd->port_rcvhdrq_size);
+ memset((void *)pd->port_rcvhdrtail_kvaddr, 0, PAGE_SIZE);
/*
* tell chip each time we init it, even if we are re-using previous
- * memory (we zero it at process close)
+ * memory (we zero the register at process close)
*/
- ipath_cdbg(VERBOSE, "writing port %d rcvhdraddr as %lx\n",
- pd->port_port, (unsigned long) pd->port_rcvhdrq_phys);
+ ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdrtailaddr,
+ pd->port_port, pd->port_rcvhdrqtailaddr_phys);
ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdraddr,
pd->port_port, pd->port_rcvhdrq_phys);
[INFINIPATH_IBCC_LINKCMD_ARMED] = "ARMED",
[INFINIPATH_IBCC_LINKCMD_ACTIVE] = "ACTIVE"
};
+ int linkcmd = (which >> INFINIPATH_IBCC_LINKCMD_SHIFT) &
+ INFINIPATH_IBCC_LINKCMD_MASK;
+
ipath_cdbg(SMA, "Trying to move unit %u to %s, current ltstate "
"is %s\n", dd->ipath_unit,
- what[(which >> INFINIPATH_IBCC_LINKCMD_SHIFT) &
- INFINIPATH_IBCC_LINKCMD_MASK],
+ what[linkcmd],
ipath_ibcstatus_str[
(ipath_read_kreg64
(dd, dd->ipath_kregs->kr_ibcstatus) >>
INFINIPATH_IBCS_LINKTRAININGSTATE_SHIFT) &
INFINIPATH_IBCS_LINKTRAININGSTATE_MASK]);
+ /* flush all queued sends when going to DOWN or INIT, to be sure that
+ * they don't block SMA and other MAD packets */
+ if (!linkcmd || linkcmd == INFINIPATH_IBCC_LINKCMD_INIT) {
+ ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
+ INFINIPATH_S_ABORT);
+ ipath_disarm_piobufs(dd, dd->ipath_lastport_piobuf,
+ (unsigned)(dd->ipath_piobcnt2k +
+ dd->ipath_piobcnt4k) -
+ dd->ipath_lastport_piobuf);
+ }
ipath_write_kreg(dd, dd->ipath_kregs->kr_ibcctrl,
dd->ipath_ibcctrl | which);
/**
* ipath_free_pddata - free a port's allocated data
* @dd: the infinipath device
- * @port: the port
- * @freehdrq: free the port data structure if true
+ * @pd: the portdata structure
*
- * when closing, free up any allocated data for a port, if the
- * reference count goes to zero
- * Note: this also optionally frees the portdata itself!
- * Any changes here have to be matched up with the reinit case
- * of ipath_init_chip(), which calls this routine on reinit after reset.
+ * free up any allocated data for a port
+ * This should not touch anything that would affect a simultaneous
+ * re-allocation of port data, because it is called after ipath_mutex
+ * is released (and can be called from reinit as well).
+ * It should never change any chip state, or global driver state.
+ * (The only exception to global state is freeing the port0 port0_skbs.)
*/
-void ipath_free_pddata(struct ipath_devdata *dd, u32 port, int freehdrq)
+void ipath_free_pddata(struct ipath_devdata *dd, struct ipath_portdata *pd)
{
- struct ipath_portdata *pd = dd->ipath_pd[port];
-
if (!pd)
return;
- if (freehdrq)
- /*
- * only clear and free portdata if we are going to also
- * release the hdrq, otherwise we leak the hdrq on each
- * open/close cycle
- */
- dd->ipath_pd[port] = NULL;
- if (freehdrq && pd->port_rcvhdrq) {
+
+ if (pd->port_rcvhdrq) {
ipath_cdbg(VERBOSE, "free closed port %d rcvhdrq @ %p "
"(size=%lu)\n", pd->port_port, pd->port_rcvhdrq,
(unsigned long) pd->port_rcvhdrq_size);
dma_free_coherent(&dd->pcidev->dev, pd->port_rcvhdrq_size,
pd->port_rcvhdrq, pd->port_rcvhdrq_phys);
pd->port_rcvhdrq = NULL;
+ if (pd->port_rcvhdrtail_kvaddr) {
+ dma_free_coherent(&dd->pcidev->dev, PAGE_SIZE,
+ (void *)pd->port_rcvhdrtail_kvaddr,
+ pd->port_rcvhdrqtailaddr_phys);
+ pd->port_rcvhdrtail_kvaddr = NULL;
+ }
}
- if (port && pd->port_rcvegrbuf) {
- /* always free this */
- if (pd->port_rcvegrbuf) {
- unsigned e;
-
- for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) {
- void *base = pd->port_rcvegrbuf[e];
- size_t size = pd->port_rcvegrbuf_size;
-
- ipath_cdbg(VERBOSE, "egrbuf free(%p, %lu), "
- "chunk %u/%u\n", base,
- (unsigned long) size,
- e, pd->port_rcvegrbuf_chunks);
- dma_free_coherent(
- &dd->pcidev->dev, size, base,
- pd->port_rcvegrbuf_phys[e]);
- }
- vfree(pd->port_rcvegrbuf);
- pd->port_rcvegrbuf = NULL;
- vfree(pd->port_rcvegrbuf_phys);
- pd->port_rcvegrbuf_phys = NULL;
+ if (pd->port_port && pd->port_rcvegrbuf) {
+ unsigned e;
+
+ for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) {
+ void *base = pd->port_rcvegrbuf[e];
+ size_t size = pd->port_rcvegrbuf_size;
+
+ ipath_cdbg(VERBOSE, "egrbuf free(%p, %lu), "
+ "chunk %u/%u\n", base,
+ (unsigned long) size,
+ e, pd->port_rcvegrbuf_chunks);
+ dma_free_coherent(&dd->pcidev->dev, size,
+ base, pd->port_rcvegrbuf_phys[e]);
}
+ vfree(pd->port_rcvegrbuf);
+ pd->port_rcvegrbuf = NULL;
+ vfree(pd->port_rcvegrbuf_phys);
+ pd->port_rcvegrbuf_phys = NULL;
pd->port_rcvegrbuf_chunks = 0;
- } else if (port == 0 && dd->ipath_port0_skbs) {
+ } else if (pd->port_port == 0 && dd->ipath_port0_skbs) {
unsigned e;
struct sk_buff **skbs = dd->ipath_port0_skbs;
dev_kfree_skb(skbs[e]);
vfree(skbs);
}
- if (freehdrq) {
- kfree(pd->port_tid_pg_list);
- kfree(pd);
- }
+ kfree(pd->port_tid_pg_list);
+ kfree(pd);
}
static int __init infinipath_init(void)
dd->ipath_pioavailregs_phys);
dd->ipath_pioavailregs_dma = NULL;
}
+ if (dd->ipath_dummy_hdrq) {
+ dma_free_coherent(&dd->pcidev->dev,
+ dd->ipath_pd[0]->port_rcvhdrq_size,
+ dd->ipath_dummy_hdrq, dd->ipath_dummy_hdrq_phys);
+ dd->ipath_dummy_hdrq = NULL;
+ }
if (dd->ipath_pageshadow) {
struct page **tmpp = dd->ipath_pageshadow;
/*
* free any resources still in use (usually just kernel ports)
- * at unload
+ * at unload; we do for portcnt, not cfgports, because cfgports
+ * could have changed while we were loaded.
*/
- for (port = 0; port < dd->ipath_cfgports; port++)
- ipath_free_pddata(dd, port, 1);
+ for (port = 0; port < dd->ipath_portcnt; port++) {
+ struct ipath_portdata *pd = dd->ipath_pd[port];
+ dd->ipath_pd[port] = NULL;
+ ipath_free_pddata(dd, pd);
+ }
kfree(dd->ipath_pd);
/*
* debuggability, in case some cleanup path tries to use it
projects / linux-2.6-omap-h63xx.git / blobdiff