2 * Copyright (c) 2006, 2007 QLogic Corporation. All rights reserved.
3 * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
5 * This software is available to you under a choice of one of two
6 * licenses. You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the
9 * OpenIB.org BSD license below:
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
19 * - Redistributions in binary form must reproduce the above
20 * copyright notice, this list of conditions and the following
21 * disclaimer in the documentation and/or other materials
22 * provided with the distribution.
24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34 #include <linux/spinlock.h>
35 #include <linux/idr.h>
36 #include <linux/pci.h>
38 #include <linux/delay.h>
39 #include <linux/netdevice.h>
40 #include <linux/vmalloc.h>
42 #include "ipath_kernel.h"
43 #include "ipath_verbs.h"
44 #include "ipath_common.h"
46 static void ipath_update_pio_bufs(struct ipath_devdata *);
48 const char *ipath_get_unit_name(int unit)
50 static char iname[16];
51 snprintf(iname, sizeof iname, "infinipath%u", unit);
55 #define DRIVER_LOAD_MSG "QLogic " IPATH_DRV_NAME " loaded: "
56 #define PFX IPATH_DRV_NAME ": "
59 * The size has to be longer than this string, so we can append
60 * board/chip information to it in the init code.
62 const char ib_ipath_version[] = IPATH_IDSTR "\n";
64 static struct idr unit_table;
65 DEFINE_SPINLOCK(ipath_devs_lock);
66 LIST_HEAD(ipath_dev_list);
68 wait_queue_head_t ipath_state_wait;
70 unsigned ipath_debug = __IPATH_INFO;
72 module_param_named(debug, ipath_debug, uint, S_IWUSR | S_IRUGO);
73 MODULE_PARM_DESC(debug, "mask for debug prints");
74 EXPORT_SYMBOL_GPL(ipath_debug);
76 unsigned ipath_mtu4096 = 1; /* max 4KB IB mtu by default, if supported */
77 module_param_named(mtu4096, ipath_mtu4096, uint, S_IRUGO);
78 MODULE_PARM_DESC(mtu4096, "enable MTU of 4096 bytes, if supported");
80 static unsigned ipath_hol_timeout_ms = 13000;
81 module_param_named(hol_timeout_ms, ipath_hol_timeout_ms, uint, S_IRUGO);
82 MODULE_PARM_DESC(hol_timeout_ms,
83 "duration of user app suspension after link failure");
85 unsigned ipath_linkrecovery = 1;
86 module_param_named(linkrecovery, ipath_linkrecovery, uint, S_IWUSR | S_IRUGO);
87 MODULE_PARM_DESC(linkrecovery, "enable workaround for link recovery issue");
89 MODULE_LICENSE("GPL");
90 MODULE_AUTHOR("QLogic <support@qlogic.com>");
91 MODULE_DESCRIPTION("QLogic InfiniPath driver");
93 const char *ipath_ibcstatus_str[] = {
100 "LState6", /* unused */
101 "LState7", /* unused */
107 "LState0xD", /* unused */
112 static void __devexit ipath_remove_one(struct pci_dev *);
113 static int __devinit ipath_init_one(struct pci_dev *,
114 const struct pci_device_id *);
116 /* Only needed for registration, nothing else needs this info */
117 #define PCI_VENDOR_ID_PATHSCALE 0x1fc1
118 #define PCI_DEVICE_ID_INFINIPATH_HT 0xd
119 #define PCI_DEVICE_ID_INFINIPATH_PE800 0x10
121 /* Number of seconds before our card status check... */
122 #define STATUS_TIMEOUT 60
124 static const struct pci_device_id ipath_pci_tbl[] = {
125 { PCI_DEVICE(PCI_VENDOR_ID_PATHSCALE, PCI_DEVICE_ID_INFINIPATH_HT) },
126 { PCI_DEVICE(PCI_VENDOR_ID_PATHSCALE, PCI_DEVICE_ID_INFINIPATH_PE800) },
130 MODULE_DEVICE_TABLE(pci, ipath_pci_tbl);
132 static struct pci_driver ipath_driver = {
133 .name = IPATH_DRV_NAME,
134 .probe = ipath_init_one,
135 .remove = __devexit_p(ipath_remove_one),
136 .id_table = ipath_pci_tbl,
138 .groups = ipath_driver_attr_groups,
142 static void ipath_check_status(struct work_struct *work)
144 struct ipath_devdata *dd = container_of(work, struct ipath_devdata,
148 * If we don't have any interrupts, let the user know and
149 * don't bother checking again.
151 if (dd->ipath_int_counter == 0)
152 dev_err(&dd->pcidev->dev, "No interrupts detected.\n");
155 static inline void read_bars(struct ipath_devdata *dd, struct pci_dev *dev,
156 u32 *bar0, u32 *bar1)
160 ret = pci_read_config_dword(dev, PCI_BASE_ADDRESS_0, bar0);
162 ipath_dev_err(dd, "failed to read bar0 before enable: "
165 ret = pci_read_config_dword(dev, PCI_BASE_ADDRESS_1, bar1);
167 ipath_dev_err(dd, "failed to read bar1 before enable: "
170 ipath_dbg("Read bar0 %x bar1 %x\n", *bar0, *bar1);
173 static void ipath_free_devdata(struct pci_dev *pdev,
174 struct ipath_devdata *dd)
178 pci_set_drvdata(pdev, NULL);
180 if (dd->ipath_unit != -1) {
181 spin_lock_irqsave(&ipath_devs_lock, flags);
182 idr_remove(&unit_table, dd->ipath_unit);
183 list_del(&dd->ipath_list);
184 spin_unlock_irqrestore(&ipath_devs_lock, flags);
189 static struct ipath_devdata *ipath_alloc_devdata(struct pci_dev *pdev)
192 struct ipath_devdata *dd;
195 if (!idr_pre_get(&unit_table, GFP_KERNEL)) {
196 dd = ERR_PTR(-ENOMEM);
200 dd = vmalloc(sizeof(*dd));
202 dd = ERR_PTR(-ENOMEM);
205 memset(dd, 0, sizeof(*dd));
208 spin_lock_irqsave(&ipath_devs_lock, flags);
210 ret = idr_get_new(&unit_table, dd, &dd->ipath_unit);
212 printk(KERN_ERR IPATH_DRV_NAME
213 ": Could not allocate unit ID: error %d\n", -ret);
214 ipath_free_devdata(pdev, dd);
220 pci_set_drvdata(pdev, dd);
222 INIT_DELAYED_WORK(&dd->status_work, ipath_check_status);
224 list_add(&dd->ipath_list, &ipath_dev_list);
227 spin_unlock_irqrestore(&ipath_devs_lock, flags);
233 static inline struct ipath_devdata *__ipath_lookup(int unit)
235 return idr_find(&unit_table, unit);
238 struct ipath_devdata *ipath_lookup(int unit)
240 struct ipath_devdata *dd;
243 spin_lock_irqsave(&ipath_devs_lock, flags);
244 dd = __ipath_lookup(unit);
245 spin_unlock_irqrestore(&ipath_devs_lock, flags);
250 int ipath_count_units(int *npresentp, int *nupp, int *maxportsp)
252 int nunits, npresent, nup;
253 struct ipath_devdata *dd;
257 nunits = npresent = nup = maxports = 0;
259 spin_lock_irqsave(&ipath_devs_lock, flags);
261 list_for_each_entry(dd, &ipath_dev_list, ipath_list) {
263 if ((dd->ipath_flags & IPATH_PRESENT) && dd->ipath_kregbase)
266 !(dd->ipath_flags & (IPATH_DISABLED | IPATH_LINKDOWN
269 if (dd->ipath_cfgports > maxports)
270 maxports = dd->ipath_cfgports;
273 spin_unlock_irqrestore(&ipath_devs_lock, flags);
276 *npresentp = npresent;
280 *maxportsp = maxports;
286 * These next two routines are placeholders in case we don't have per-arch
287 * code for controlling write combining. If explicit control of write
288 * combining is not available, performance will probably be awful.
291 int __attribute__((weak)) ipath_enable_wc(struct ipath_devdata *dd)
296 void __attribute__((weak)) ipath_disable_wc(struct ipath_devdata *dd)
301 * Perform a PIO buffer bandwidth write test, to verify proper system
302 * configuration. Even when all the setup calls work, occasionally
303 * BIOS or other issues can prevent write combining from working, or
304 * can cause other bandwidth problems to the chip.
306 * This test simply writes the same buffer over and over again, and
307 * measures close to the peak bandwidth to the chip (not testing
308 * data bandwidth to the wire). On chips that use an address-based
309 * trigger to send packets to the wire, this is easy. On chips that
310 * use a count to trigger, we want to make sure that the packet doesn't
311 * go out on the wire, or trigger flow control checks.
313 static void ipath_verify_pioperf(struct ipath_devdata *dd)
315 u32 pbnum, cnt, lcnt;
320 piobuf = ipath_getpiobuf(dd, 0, &pbnum);
322 dev_info(&dd->pcidev->dev,
323 "No PIObufs for checking perf, skipping\n");
328 * Enough to give us a reasonable test, less than piobuf size, and
329 * likely multiple of store buffer length.
335 dev_info(&dd->pcidev->dev,
336 "Couldn't get memory for checking PIO perf,"
341 preempt_disable(); /* we want reasonably accurate elapsed time */
342 msecs = 1 + jiffies_to_msecs(jiffies);
343 for (lcnt = 0; lcnt < 10000U; lcnt++) {
344 /* wait until we cross msec boundary */
345 if (jiffies_to_msecs(jiffies) >= msecs)
350 ipath_disable_armlaunch(dd);
352 writeq(0, piobuf); /* length 0, no dwords actually sent */
356 * this is only roughly accurate, since even with preempt we
357 * still take interrupts that could take a while. Running for
358 * >= 5 msec seems to get us "close enough" to accurate values
360 msecs = jiffies_to_msecs(jiffies);
361 for (emsecs = lcnt = 0; emsecs <= 5UL; lcnt++) {
362 __iowrite32_copy(piobuf + 64, addr, cnt >> 2);
363 emsecs = jiffies_to_msecs(jiffies) - msecs;
366 /* 1 GiB/sec, slightly over IB SDR line rate */
367 if (lcnt < (emsecs * 1024U))
369 "Performance problem: bandwidth to PIO buffers is "
371 lcnt / (u32) emsecs);
373 ipath_dbg("PIO buffer bandwidth %u MiB/sec is OK\n",
374 lcnt / (u32) emsecs);
381 /* disarm piobuf, so it's available again */
382 ipath_disarm_piobufs(dd, pbnum, 1);
383 ipath_enable_armlaunch(dd);
386 static int __devinit ipath_init_one(struct pci_dev *pdev,
387 const struct pci_device_id *ent)
390 struct ipath_devdata *dd;
391 unsigned long long addr;
392 u32 bar0 = 0, bar1 = 0;
394 dd = ipath_alloc_devdata(pdev);
397 printk(KERN_ERR IPATH_DRV_NAME
398 ": Could not allocate devdata: error %d\n", -ret);
402 ipath_cdbg(VERBOSE, "initializing unit #%u\n", dd->ipath_unit);
404 ret = pci_enable_device(pdev);
406 /* This can happen iff:
408 * We did a chip reset, and then failed to reprogram the
409 * BAR, or the chip reset due to an internal error. We then
410 * unloaded the driver and reloaded it.
412 * Both reset cases set the BAR back to initial state. For
413 * the latter case, the AER sticky error bit at offset 0x718
414 * should be set, but the Linux kernel doesn't yet know
415 * about that, it appears. If the original BAR was retained
416 * in the kernel data structures, this may be OK.
418 ipath_dev_err(dd, "enable unit %d failed: error %d\n",
419 dd->ipath_unit, -ret);
422 addr = pci_resource_start(pdev, 0);
423 len = pci_resource_len(pdev, 0);
424 ipath_cdbg(VERBOSE, "regbase (0) %llx len %d pdev->irq %d, vend %x/%x "
425 "driver_data %lx\n", addr, len, pdev->irq, ent->vendor,
426 ent->device, ent->driver_data);
428 read_bars(dd, pdev, &bar0, &bar1);
430 if (!bar1 && !(bar0 & ~0xf)) {
432 dev_info(&pdev->dev, "BAR is 0 (probable RESET), "
433 "rewriting as %llx\n", addr);
434 ret = pci_write_config_dword(
435 pdev, PCI_BASE_ADDRESS_0, addr);
437 ipath_dev_err(dd, "rewrite of BAR0 "
438 "failed: err %d\n", -ret);
441 ret = pci_write_config_dword(
442 pdev, PCI_BASE_ADDRESS_1, addr >> 32);
444 ipath_dev_err(dd, "rewrite of BAR1 "
445 "failed: err %d\n", -ret);
449 ipath_dev_err(dd, "BAR is 0 (probable RESET), "
450 "not usable until reboot\n");
456 ret = pci_request_regions(pdev, IPATH_DRV_NAME);
458 dev_info(&pdev->dev, "pci_request_regions unit %u fails: "
459 "err %d\n", dd->ipath_unit, -ret);
463 ret = pci_set_dma_mask(pdev, DMA_64BIT_MASK);
466 * if the 64 bit setup fails, try 32 bit. Some systems
467 * do not setup 64 bit maps on systems with 2GB or less
470 ret = pci_set_dma_mask(pdev, DMA_32BIT_MASK);
473 "Unable to set DMA mask for unit %u: %d\n",
474 dd->ipath_unit, ret);
478 ipath_dbg("No 64bit DMA mask, used 32 bit mask\n");
479 ret = pci_set_consistent_dma_mask(pdev, DMA_32BIT_MASK);
482 "Unable to set DMA consistent mask "
484 dd->ipath_unit, ret);
489 ret = pci_set_consistent_dma_mask(pdev, DMA_64BIT_MASK);
492 "Unable to set DMA consistent mask "
494 dd->ipath_unit, ret);
497 pci_set_master(pdev);
500 * Save BARs to rewrite after device reset. Save all 64 bits of
503 dd->ipath_pcibar0 = addr;
504 dd->ipath_pcibar1 = addr >> 32;
505 dd->ipath_deviceid = ent->device; /* save for later use */
506 dd->ipath_vendorid = ent->vendor;
508 /* setup the chip-specific functions, as early as possible. */
509 switch (ent->device) {
510 case PCI_DEVICE_ID_INFINIPATH_HT:
512 ipath_init_iba6110_funcs(dd);
515 ipath_dev_err(dd, "QLogic HT device 0x%x cannot work if "
516 "CONFIG_HT_IRQ is not enabled\n", ent->device);
519 case PCI_DEVICE_ID_INFINIPATH_PE800:
520 #ifdef CONFIG_PCI_MSI
521 ipath_init_iba6120_funcs(dd);
524 ipath_dev_err(dd, "QLogic PCIE device 0x%x cannot work if "
525 "CONFIG_PCI_MSI is not enabled\n", ent->device);
529 ipath_dev_err(dd, "Found unknown QLogic deviceid 0x%x, "
530 "failing\n", ent->device);
534 for (j = 0; j < 6; j++) {
535 if (!pdev->resource[j].start)
537 ipath_cdbg(VERBOSE, "BAR %d start %llx, end %llx, len %llx\n",
538 j, (unsigned long long)pdev->resource[j].start,
539 (unsigned long long)pdev->resource[j].end,
540 (unsigned long long)pci_resource_len(pdev, j));
544 ipath_dev_err(dd, "No valid address in BAR 0!\n");
549 dd->ipath_pcirev = pdev->revision;
551 #if defined(__powerpc__)
552 /* There isn't a generic way to specify writethrough mappings */
553 dd->ipath_kregbase = __ioremap(addr, len,
554 (_PAGE_NO_CACHE|_PAGE_WRITETHRU));
556 dd->ipath_kregbase = ioremap_nocache(addr, len);
559 if (!dd->ipath_kregbase) {
560 ipath_dbg("Unable to map io addr %llx to kvirt, failing\n",
565 dd->ipath_kregend = (u64 __iomem *)
566 ((void __iomem *)dd->ipath_kregbase + len);
567 dd->ipath_physaddr = addr; /* used for io_remap, etc. */
569 ipath_cdbg(VERBOSE, "mapped io addr %llx to kregbase %p\n",
570 addr, dd->ipath_kregbase);
573 * clear ipath_flags here instead of in ipath_init_chip as it is set
574 * by ipath_setup_htconfig.
577 dd->ipath_lli_counter = 0;
578 dd->ipath_lli_errors = 0;
580 if (dd->ipath_f_bus(dd, pdev))
581 ipath_dev_err(dd, "Failed to setup config space; "
582 "continuing anyway\n");
585 * set up our interrupt handler; IRQF_SHARED probably not needed,
586 * since MSI interrupts shouldn't be shared but won't hurt for now.
587 * check 0 irq after we return from chip-specific bus setup, since
588 * that can affect this due to setup
591 ipath_dev_err(dd, "irq is 0, BIOS error? Interrupts won't "
594 ret = request_irq(dd->ipath_irq, ipath_intr, IRQF_SHARED,
597 ipath_dev_err(dd, "Couldn't setup irq handler, "
598 "irq=%d: %d\n", dd->ipath_irq, ret);
603 ret = ipath_init_chip(dd, 0); /* do the chip-specific init */
607 ret = ipath_enable_wc(dd);
610 ipath_dev_err(dd, "Write combining not enabled "
611 "(err %d): performance may be poor\n",
616 ipath_verify_pioperf(dd);
618 ipath_device_create_group(&pdev->dev, dd);
619 ipathfs_add_device(dd);
622 ipath_register_ib_device(dd);
624 /* Check that card status in STATUS_TIMEOUT seconds. */
625 schedule_delayed_work(&dd->status_work, HZ * STATUS_TIMEOUT);
630 if (pdev->irq) free_irq(pdev->irq, dd);
633 iounmap((volatile void __iomem *) dd->ipath_kregbase);
636 pci_release_regions(pdev);
639 pci_disable_device(pdev);
642 ipath_free_devdata(pdev, dd);
648 static void __devexit cleanup_device(struct ipath_devdata *dd)
652 if (*dd->ipath_statusp & IPATH_STATUS_CHIP_PRESENT) {
653 /* can't do anything more with chip; needs re-init */
654 *dd->ipath_statusp &= ~IPATH_STATUS_CHIP_PRESENT;
655 if (dd->ipath_kregbase) {
657 * if we haven't already cleaned up before these are
658 * to ensure any register reads/writes "fail" until
661 dd->ipath_kregbase = NULL;
662 dd->ipath_uregbase = 0;
663 dd->ipath_sregbase = 0;
664 dd->ipath_cregbase = 0;
665 dd->ipath_kregsize = 0;
667 ipath_disable_wc(dd);
670 if (dd->ipath_pioavailregs_dma) {
671 dma_free_coherent(&dd->pcidev->dev, PAGE_SIZE,
672 (void *) dd->ipath_pioavailregs_dma,
673 dd->ipath_pioavailregs_phys);
674 dd->ipath_pioavailregs_dma = NULL;
676 if (dd->ipath_dummy_hdrq) {
677 dma_free_coherent(&dd->pcidev->dev,
678 dd->ipath_pd[0]->port_rcvhdrq_size,
679 dd->ipath_dummy_hdrq, dd->ipath_dummy_hdrq_phys);
680 dd->ipath_dummy_hdrq = NULL;
683 if (dd->ipath_pageshadow) {
684 struct page **tmpp = dd->ipath_pageshadow;
685 dma_addr_t *tmpd = dd->ipath_physshadow;
688 ipath_cdbg(VERBOSE, "Unlocking any expTID pages still "
690 for (port = 0; port < dd->ipath_cfgports; port++) {
691 int port_tidbase = port * dd->ipath_rcvtidcnt;
692 int maxtid = port_tidbase + dd->ipath_rcvtidcnt;
693 for (i = port_tidbase; i < maxtid; i++) {
696 pci_unmap_page(dd->pcidev, tmpd[i],
697 PAGE_SIZE, PCI_DMA_FROMDEVICE);
698 ipath_release_user_pages(&tmpp[i], 1);
704 ipath_stats.sps_pageunlocks += cnt;
705 ipath_cdbg(VERBOSE, "There were still %u expTID "
706 "entries locked\n", cnt);
708 if (ipath_stats.sps_pagelocks ||
709 ipath_stats.sps_pageunlocks)
710 ipath_cdbg(VERBOSE, "%llu pages locked, %llu "
711 "unlocked via ipath_m{un}lock\n",
713 ipath_stats.sps_pagelocks,
715 ipath_stats.sps_pageunlocks);
717 ipath_cdbg(VERBOSE, "Free shadow page tid array at %p\n",
718 dd->ipath_pageshadow);
719 tmpp = dd->ipath_pageshadow;
720 dd->ipath_pageshadow = NULL;
725 * free any resources still in use (usually just kernel ports)
726 * at unload; we do for portcnt, not cfgports, because cfgports
727 * could have changed while we were loaded.
729 for (port = 0; port < dd->ipath_portcnt; port++) {
730 struct ipath_portdata *pd = dd->ipath_pd[port];
731 dd->ipath_pd[port] = NULL;
732 ipath_free_pddata(dd, pd);
736 * debuggability, in case some cleanup path tries to use it
742 static void __devexit ipath_remove_one(struct pci_dev *pdev)
744 struct ipath_devdata *dd = pci_get_drvdata(pdev);
746 ipath_cdbg(VERBOSE, "removing, pdev=%p, dd=%p\n", pdev, dd);
749 * disable the IB link early, to be sure no new packets arrive, which
750 * complicates the shutdown process
752 ipath_shutdown_device(dd);
754 cancel_delayed_work(&dd->status_work);
755 flush_scheduled_work();
758 ipath_unregister_ib_device(dd->verbs_dev);
760 ipath_diag_remove(dd);
761 ipath_user_remove(dd);
762 ipathfs_remove_device(dd);
763 ipath_device_remove_group(&pdev->dev, dd);
765 ipath_cdbg(VERBOSE, "Releasing pci memory regions, dd %p, "
766 "unit %u\n", dd, (u32) dd->ipath_unit);
771 * turn off rcv, send, and interrupts for all ports, all drivers
772 * should also hard reset the chip here?
773 * free up port 0 (kernel) rcvhdr, egr bufs, and eventually tid bufs
774 * for all versions of the driver, if they were allocated
777 ipath_cdbg(VERBOSE, "unit %u free irq %d\n",
778 dd->ipath_unit, dd->ipath_irq);
779 dd->ipath_f_free_irq(dd);
781 ipath_dbg("irq is 0, not doing free_irq "
782 "for unit %u\n", dd->ipath_unit);
784 * we check for NULL here, because it's outside
785 * the kregbase check, and we need to call it
786 * after the free_irq. Thus it's possible that
787 * the function pointers were never initialized.
789 if (dd->ipath_f_cleanup)
790 /* clean up chip-specific stuff */
791 dd->ipath_f_cleanup(dd);
793 ipath_cdbg(VERBOSE, "Unmapping kregbase %p\n", dd->ipath_kregbase);
794 iounmap((volatile void __iomem *) dd->ipath_kregbase);
795 pci_release_regions(pdev);
796 ipath_cdbg(VERBOSE, "calling pci_disable_device\n");
797 pci_disable_device(pdev);
799 ipath_free_devdata(pdev, dd);
802 /* general driver use */
803 DEFINE_MUTEX(ipath_mutex);
805 static DEFINE_SPINLOCK(ipath_pioavail_lock);
808 * ipath_disarm_piobufs - cancel a range of PIO buffers
809 * @dd: the infinipath device
810 * @first: the first PIO buffer to cancel
811 * @cnt: the number of PIO buffers to cancel
813 * cancel a range of PIO buffers, used when they might be armed, but
814 * not triggered. Used at init to ensure buffer state, and also user
815 * process close, in case it died while writing to a PIO buffer
818 void ipath_disarm_piobufs(struct ipath_devdata *dd, unsigned first,
821 unsigned i, last = first + cnt;
824 ipath_cdbg(PKT, "disarm %u PIObufs first=%u\n", cnt, first);
825 for (i = first; i < last; i++) {
826 spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
828 * The disarm-related bits are write-only, so it
829 * is ok to OR them in with our copy of sendctrl
830 * while we hold the lock.
832 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
833 dd->ipath_sendctrl | INFINIPATH_S_DISARM |
834 (i << INFINIPATH_S_DISARMPIOBUF_SHIFT));
835 /* can't disarm bufs back-to-back per iba7220 spec */
836 ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
837 spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
839 /* on some older chips, update may not happen after cancel */
840 ipath_force_pio_avail_update(dd);
844 * ipath_wait_linkstate - wait for an IB link state change to occur
845 * @dd: the infinipath device
846 * @state: the state to wait for
847 * @msecs: the number of milliseconds to wait
849 * wait up to msecs milliseconds for IB link state change to occur for
850 * now, take the easy polling route. Currently used only by
851 * ipath_set_linkstate. Returns 0 if state reached, otherwise
852 * -ETIMEDOUT state can have multiple states set, for any of several
855 int ipath_wait_linkstate(struct ipath_devdata *dd, u32 state, int msecs)
857 dd->ipath_state_wanted = state;
858 wait_event_interruptible_timeout(ipath_state_wait,
859 (dd->ipath_flags & state),
860 msecs_to_jiffies(msecs));
861 dd->ipath_state_wanted = 0;
863 if (!(dd->ipath_flags & state)) {
865 ipath_cdbg(VERBOSE, "Didn't reach linkstate %s within %u"
867 /* test INIT ahead of DOWN, both can be set */
868 (state & IPATH_LINKINIT) ? "INIT" :
869 ((state & IPATH_LINKDOWN) ? "DOWN" :
870 ((state & IPATH_LINKARMED) ? "ARM" : "ACTIVE")),
872 val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_ibcstatus);
873 ipath_cdbg(VERBOSE, "ibcc=%llx ibcstatus=%llx (%s)\n",
874 (unsigned long long) ipath_read_kreg64(
875 dd, dd->ipath_kregs->kr_ibcctrl),
876 (unsigned long long) val,
877 ipath_ibcstatus_str[val & 0xf]);
879 return (dd->ipath_flags & state) ? 0 : -ETIMEDOUT;
883 * Decode the error status into strings, deciding whether to always
884 * print * it or not depending on "normal packet errors" vs everything
885 * else. Return 1 if "real" errors, otherwise 0 if only packet
886 * errors, so caller can decide what to print with the string.
888 int ipath_decode_err(char *buf, size_t blen, ipath_err_t err)
892 if (err & INFINIPATH_E_PKTERRS) {
893 if (!(err & ~INFINIPATH_E_PKTERRS))
894 iserr = 0; // if only packet errors.
895 if (ipath_debug & __IPATH_ERRPKTDBG) {
896 if (err & INFINIPATH_E_REBP)
897 strlcat(buf, "EBP ", blen);
898 if (err & INFINIPATH_E_RVCRC)
899 strlcat(buf, "VCRC ", blen);
900 if (err & INFINIPATH_E_RICRC) {
901 strlcat(buf, "CRC ", blen);
902 // clear for check below, so only once
903 err &= INFINIPATH_E_RICRC;
905 if (err & INFINIPATH_E_RSHORTPKTLEN)
906 strlcat(buf, "rshortpktlen ", blen);
907 if (err & INFINIPATH_E_SDROPPEDDATAPKT)
908 strlcat(buf, "sdroppeddatapkt ", blen);
909 if (err & INFINIPATH_E_SPKTLEN)
910 strlcat(buf, "spktlen ", blen);
912 if ((err & INFINIPATH_E_RICRC) &&
913 !(err&(INFINIPATH_E_RVCRC|INFINIPATH_E_REBP)))
914 strlcat(buf, "CRC ", blen);
918 if (err & INFINIPATH_E_RHDRLEN)
919 strlcat(buf, "rhdrlen ", blen);
920 if (err & INFINIPATH_E_RBADTID)
921 strlcat(buf, "rbadtid ", blen);
922 if (err & INFINIPATH_E_RBADVERSION)
923 strlcat(buf, "rbadversion ", blen);
924 if (err & INFINIPATH_E_RHDR)
925 strlcat(buf, "rhdr ", blen);
926 if (err & INFINIPATH_E_RLONGPKTLEN)
927 strlcat(buf, "rlongpktlen ", blen);
928 if (err & INFINIPATH_E_RMAXPKTLEN)
929 strlcat(buf, "rmaxpktlen ", blen);
930 if (err & INFINIPATH_E_RMINPKTLEN)
931 strlcat(buf, "rminpktlen ", blen);
932 if (err & INFINIPATH_E_SMINPKTLEN)
933 strlcat(buf, "sminpktlen ", blen);
934 if (err & INFINIPATH_E_RFORMATERR)
935 strlcat(buf, "rformaterr ", blen);
936 if (err & INFINIPATH_E_RUNSUPVL)
937 strlcat(buf, "runsupvl ", blen);
938 if (err & INFINIPATH_E_RUNEXPCHAR)
939 strlcat(buf, "runexpchar ", blen);
940 if (err & INFINIPATH_E_RIBFLOW)
941 strlcat(buf, "ribflow ", blen);
942 if (err & INFINIPATH_E_SUNDERRUN)
943 strlcat(buf, "sunderrun ", blen);
944 if (err & INFINIPATH_E_SPIOARMLAUNCH)
945 strlcat(buf, "spioarmlaunch ", blen);
946 if (err & INFINIPATH_E_SUNEXPERRPKTNUM)
947 strlcat(buf, "sunexperrpktnum ", blen);
948 if (err & INFINIPATH_E_SDROPPEDSMPPKT)
949 strlcat(buf, "sdroppedsmppkt ", blen);
950 if (err & INFINIPATH_E_SMAXPKTLEN)
951 strlcat(buf, "smaxpktlen ", blen);
952 if (err & INFINIPATH_E_SUNSUPVL)
953 strlcat(buf, "sunsupVL ", blen);
954 if (err & INFINIPATH_E_INVALIDADDR)
955 strlcat(buf, "invalidaddr ", blen);
956 if (err & INFINIPATH_E_RRCVEGRFULL)
957 strlcat(buf, "rcvegrfull ", blen);
958 if (err & INFINIPATH_E_RRCVHDRFULL)
959 strlcat(buf, "rcvhdrfull ", blen);
960 if (err & INFINIPATH_E_IBSTATUSCHANGED)
961 strlcat(buf, "ibcstatuschg ", blen);
962 if (err & INFINIPATH_E_RIBLOSTLINK)
963 strlcat(buf, "riblostlink ", blen);
964 if (err & INFINIPATH_E_HARDWARE)
965 strlcat(buf, "hardware ", blen);
966 if (err & INFINIPATH_E_RESET)
967 strlcat(buf, "reset ", blen);
973 * get_rhf_errstring - decode RHF errors
974 * @err: the err number
975 * @msg: the output buffer
976 * @len: the length of the output buffer
978 * only used one place now, may want more later
980 static void get_rhf_errstring(u32 err, char *msg, size_t len)
982 /* if no errors, and so don't need to check what's first */
985 if (err & INFINIPATH_RHF_H_ICRCERR)
986 strlcat(msg, "icrcerr ", len);
987 if (err & INFINIPATH_RHF_H_VCRCERR)
988 strlcat(msg, "vcrcerr ", len);
989 if (err & INFINIPATH_RHF_H_PARITYERR)
990 strlcat(msg, "parityerr ", len);
991 if (err & INFINIPATH_RHF_H_LENERR)
992 strlcat(msg, "lenerr ", len);
993 if (err & INFINIPATH_RHF_H_MTUERR)
994 strlcat(msg, "mtuerr ", len);
995 if (err & INFINIPATH_RHF_H_IHDRERR)
996 /* infinipath hdr checksum error */
997 strlcat(msg, "ipathhdrerr ", len);
998 if (err & INFINIPATH_RHF_H_TIDERR)
999 strlcat(msg, "tiderr ", len);
1000 if (err & INFINIPATH_RHF_H_MKERR)
1001 /* bad port, offset, etc. */
1002 strlcat(msg, "invalid ipathhdr ", len);
1003 if (err & INFINIPATH_RHF_H_IBERR)
1004 strlcat(msg, "iberr ", len);
1005 if (err & INFINIPATH_RHF_L_SWA)
1006 strlcat(msg, "swA ", len);
1007 if (err & INFINIPATH_RHF_L_SWB)
1008 strlcat(msg, "swB ", len);
1012 * ipath_get_egrbuf - get an eager buffer
1013 * @dd: the infinipath device
1014 * @bufnum: the eager buffer to get
1016 * must only be called if ipath_pd[port] is known to be allocated
1018 static inline void *ipath_get_egrbuf(struct ipath_devdata *dd, u32 bufnum)
1020 return dd->ipath_port0_skbinfo ?
1021 (void *) dd->ipath_port0_skbinfo[bufnum].skb->data : NULL;
1025 * ipath_alloc_skb - allocate an skb and buffer with possible constraints
1026 * @dd: the infinipath device
1027 * @gfp_mask: the sk_buff SFP mask
1029 struct sk_buff *ipath_alloc_skb(struct ipath_devdata *dd,
1032 struct sk_buff *skb;
1036 * Only fully supported way to handle this is to allocate lots
1037 * extra, align as needed, and then do skb_reserve(). That wastes
1038 * a lot of memory... I'll have to hack this into infinipath_copy
1043 * We need 2 extra bytes for ipath_ether data sent in the
1044 * key header. In order to keep everything dword aligned,
1045 * we'll reserve 4 bytes.
1047 len = dd->ipath_ibmaxlen + 4;
1049 if (dd->ipath_flags & IPATH_4BYTE_TID) {
1050 /* We need a 2KB multiple alignment, and there is no way
1051 * to do it except to allocate extra and then skb_reserve
1052 * enough to bring it up to the right alignment.
1057 skb = __dev_alloc_skb(len, gfp_mask);
1059 ipath_dev_err(dd, "Failed to allocate skbuff, length %u\n",
1064 skb_reserve(skb, 4);
1066 if (dd->ipath_flags & IPATH_4BYTE_TID) {
1067 u32 una = (unsigned long)skb->data & 2047;
1069 skb_reserve(skb, 2048 - una);
1076 static void ipath_rcv_hdrerr(struct ipath_devdata *dd,
1083 struct ipath_message_header *hdr;
1085 get_rhf_errstring(eflags, emsg, sizeof emsg);
1086 hdr = (struct ipath_message_header *)&rc[1];
1087 ipath_cdbg(PKT, "RHFerrs %x hdrqtail=%x typ=%u "
1088 "tlen=%x opcode=%x egridx=%x: %s\n",
1090 ipath_hdrget_rcv_type((__le32 *) rc),
1091 ipath_hdrget_length_in_bytes((__le32 *) rc),
1092 be32_to_cpu(hdr->bth[0]) >> 24,
1095 /* Count local link integrity errors. */
1096 if (eflags & (INFINIPATH_RHF_H_ICRCERR | INFINIPATH_RHF_H_VCRCERR)) {
1097 u8 n = (dd->ipath_ibcctrl >>
1098 INFINIPATH_IBCC_PHYERRTHRESHOLD_SHIFT) &
1099 INFINIPATH_IBCC_PHYERRTHRESHOLD_MASK;
1101 if (++dd->ipath_lli_counter > n) {
1102 dd->ipath_lli_counter = 0;
1103 dd->ipath_lli_errors++;
1109 * ipath_kreceive - receive a packet
1110 * @pd: the infinipath port
1112 * called from interrupt handler for errors or receive interrupt
1114 void ipath_kreceive(struct ipath_portdata *pd)
1117 struct ipath_devdata *dd = pd->port_dd;
1119 const u32 rsize = dd->ipath_rcvhdrentsize; /* words */
1120 const u32 maxcnt = dd->ipath_rcvhdrcnt * rsize; /* words */
1121 u32 etail = -1, l, hdrqtail;
1122 struct ipath_message_header *hdr;
1123 u32 eflags, i, etype, tlen, pkttot = 0, updegr=0, reloop=0;
1124 static u64 totcalls; /* stats, may eventually remove */
1126 if (!dd->ipath_hdrqtailptr) {
1128 "hdrqtailptr not set, can't do receives\n");
1133 hdrqtail = ipath_get_rcvhdrtail(pd);
1138 for (i = 0; l != hdrqtail; i++) {
1142 rc = (u64 *) (pd->port_rcvhdrq + (l << 2));
1143 hdr = (struct ipath_message_header *)&rc[1];
1145 * could make a network order version of IPATH_KD_QP, and
1146 * do the obvious shift before masking to speed this up.
1148 qp = ntohl(hdr->bth[1]) & 0xffffff;
1149 bthbytes = (u8 *) hdr->bth;
1151 eflags = ipath_hdrget_err_flags((__le32 *) rc);
1152 etype = ipath_hdrget_rcv_type((__le32 *) rc);
1154 tlen = ipath_hdrget_length_in_bytes((__le32 *) rc);
1156 if (etype != RCVHQ_RCV_TYPE_EXPECTED) {
1158 * it turns out that the chips uses an eager buffer
1159 * for all non-expected packets, whether it "needs"
1160 * one or not. So always get the index, but don't
1161 * set ebuf (so we try to copy data) unless the
1162 * length requires it.
1164 etail = ipath_hdrget_index((__le32 *) rc);
1165 if (tlen > sizeof(*hdr) ||
1166 etype == RCVHQ_RCV_TYPE_NON_KD)
1167 ebuf = ipath_get_egrbuf(dd, etail);
1171 * both tiderr and ipathhdrerr are set for all plain IB
1172 * packets; only ipathhdrerr should be set.
1175 if (etype != RCVHQ_RCV_TYPE_NON_KD && etype !=
1176 RCVHQ_RCV_TYPE_ERROR && ipath_hdrget_ipath_ver(
1177 hdr->iph.ver_port_tid_offset) !=
1178 IPS_PROTO_VERSION) {
1179 ipath_cdbg(PKT, "Bad InfiniPath protocol version "
1183 if (unlikely(eflags))
1184 ipath_rcv_hdrerr(dd, eflags, l, etail, rc);
1185 else if (etype == RCVHQ_RCV_TYPE_NON_KD) {
1186 ipath_ib_rcv(dd->verbs_dev, rc + 1, ebuf, tlen);
1187 if (dd->ipath_lli_counter)
1188 dd->ipath_lli_counter--;
1189 ipath_cdbg(PKT, "typ %x, opcode %x (eager, "
1190 "qp=%x), len %x; ignored\n",
1191 etype, bthbytes[0], qp, tlen);
1193 else if (etype == RCVHQ_RCV_TYPE_EAGER)
1194 ipath_cdbg(PKT, "typ %x, opcode %x (eager, "
1195 "qp=%x), len %x; ignored\n",
1196 etype, bthbytes[0], qp, tlen);
1197 else if (etype == RCVHQ_RCV_TYPE_EXPECTED)
1198 ipath_dbg("Bug: Expected TID, opcode %x; ignored\n",
1199 be32_to_cpu(hdr->bth[0]) & 0xff);
1202 * error packet, type of error unknown.
1203 * Probably type 3, but we don't know, so don't
1204 * even try to print the opcode, etc.
1206 ipath_dbg("Error Pkt, but no eflags! egrbuf %x, "
1207 "len %x\nhdrq@%lx;hdrq+%x rhf: %llx; "
1208 "hdr %llx %llx %llx %llx %llx\n",
1209 etail, tlen, (unsigned long) rc, l,
1210 (unsigned long long) rc[0],
1211 (unsigned long long) rc[1],
1212 (unsigned long long) rc[2],
1213 (unsigned long long) rc[3],
1214 (unsigned long long) rc[4],
1215 (unsigned long long) rc[5]);
1220 if (etype != RCVHQ_RCV_TYPE_EXPECTED)
1223 * update head regs on last packet, and every 16 packets.
1224 * Reduce bus traffic, while still trying to prevent
1225 * rcvhdrq overflows, for when the queue is nearly full
1227 if (l == hdrqtail || (i && !(i&0xf))) {
1230 /* request IBA6120 interrupt only on last */
1231 lval = dd->ipath_rhdrhead_intr_off | l;
1234 ipath_write_ureg(dd, ur_rcvhdrhead, lval, 0);
1236 ipath_write_ureg(dd, ur_rcvegrindexhead,
1243 if (!dd->ipath_rhdrhead_intr_off && !reloop) {
1244 /* IBA6110 workaround; we can have a race clearing chip
1245 * interrupt with another interrupt about to be delivered,
1246 * and can clear it before it is delivered on the GPIO
1247 * workaround. By doing the extra check here for the
1248 * in-memory tail register updating while we were doing
1249 * earlier packets, we "almost" guarantee we have covered
1252 u32 hqtail = ipath_get_rcvhdrtail(pd);
1253 if (hqtail != hdrqtail) {
1255 reloop = 1; /* loop 1 extra time at most */
1264 if (pkttot > ipath_stats.sps_maxpkts_call)
1265 ipath_stats.sps_maxpkts_call = pkttot;
1266 ipath_stats.sps_port0pkts += pkttot;
1267 ipath_stats.sps_avgpkts_call =
1268 ipath_stats.sps_port0pkts / ++totcalls;
1274 * ipath_update_pio_bufs - update shadow copy of the PIO availability map
1275 * @dd: the infinipath device
1277 * called whenever our local copy indicates we have run out of send buffers
1278 * NOTE: This can be called from interrupt context by some code
1279 * and from non-interrupt context by ipath_getpiobuf().
1282 static void ipath_update_pio_bufs(struct ipath_devdata *dd)
1284 unsigned long flags;
1286 const unsigned piobregs = (unsigned)dd->ipath_pioavregs;
1288 /* If the generation (check) bits have changed, then we update the
1289 * busy bit for the corresponding PIO buffer. This algorithm will
1290 * modify positions to the value they already have in some cases
1291 * (i.e., no change), but it's faster than changing only the bits
1292 * that have changed.
1294 * We would like to do this atomicly, to avoid spinlocks in the
1295 * critical send path, but that's not really possible, given the
1296 * type of changes, and that this routine could be called on
1297 * multiple cpu's simultaneously, so we lock in this routine only,
1298 * to avoid conflicting updates; all we change is the shadow, and
1299 * it's a single 64 bit memory location, so by definition the update
1300 * is atomic in terms of what other cpu's can see in testing the
1301 * bits. The spin_lock overhead isn't too bad, since it only
1302 * happens when all buffers are in use, so only cpu overhead, not
1303 * latency or bandwidth is affected.
1305 if (!dd->ipath_pioavailregs_dma) {
1306 ipath_dbg("Update shadow pioavail, but regs_dma NULL!\n");
1309 if (ipath_debug & __IPATH_VERBDBG) {
1310 /* only if packet debug and verbose */
1311 volatile __le64 *dma = dd->ipath_pioavailregs_dma;
1312 unsigned long *shadow = dd->ipath_pioavailshadow;
1314 ipath_cdbg(PKT, "Refill avail, dma0=%llx shad0=%lx, "
1315 "d1=%llx s1=%lx, d2=%llx s2=%lx, d3=%llx "
1317 (unsigned long long) le64_to_cpu(dma[0]),
1319 (unsigned long long) le64_to_cpu(dma[1]),
1321 (unsigned long long) le64_to_cpu(dma[2]),
1323 (unsigned long long) le64_to_cpu(dma[3]),
1327 PKT, "2nd group, dma4=%llx shad4=%lx, "
1328 "d5=%llx s5=%lx, d6=%llx s6=%lx, "
1330 (unsigned long long) le64_to_cpu(dma[4]),
1332 (unsigned long long) le64_to_cpu(dma[5]),
1334 (unsigned long long) le64_to_cpu(dma[6]),
1336 (unsigned long long) le64_to_cpu(dma[7]),
1339 spin_lock_irqsave(&ipath_pioavail_lock, flags);
1340 for (i = 0; i < piobregs; i++) {
1341 u64 pchbusy, pchg, piov, pnew;
1343 * Chip Errata: bug 6641; even and odd qwords>3 are swapped
1345 if (i > 3 && (dd->ipath_flags & IPATH_SWAP_PIOBUFS))
1346 piov = le64_to_cpu(dd->ipath_pioavailregs_dma[i ^ 1]);
1348 piov = le64_to_cpu(dd->ipath_pioavailregs_dma[i]);
1349 pchg = dd->ipath_pioavailkernel[i] &
1350 ~(dd->ipath_pioavailshadow[i] ^ piov);
1351 pchbusy = pchg << INFINIPATH_SENDPIOAVAIL_BUSY_SHIFT;
1352 if (pchg && (pchbusy & dd->ipath_pioavailshadow[i])) {
1353 pnew = dd->ipath_pioavailshadow[i] & ~pchbusy;
1354 pnew |= piov & pchbusy;
1355 dd->ipath_pioavailshadow[i] = pnew;
1358 spin_unlock_irqrestore(&ipath_pioavail_lock, flags);
1362 * ipath_setrcvhdrsize - set the receive header size
1363 * @dd: the infinipath device
1364 * @rhdrsize: the receive header size
1366 * called from user init code, and also layered driver init
1368 int ipath_setrcvhdrsize(struct ipath_devdata *dd, unsigned rhdrsize)
1372 if (dd->ipath_flags & IPATH_RCVHDRSZ_SET) {
1373 if (dd->ipath_rcvhdrsize != rhdrsize) {
1374 dev_info(&dd->pcidev->dev,
1375 "Error: can't set protocol header "
1376 "size %u, already %u\n",
1377 rhdrsize, dd->ipath_rcvhdrsize);
1380 ipath_cdbg(VERBOSE, "Reuse same protocol header "
1381 "size %u\n", dd->ipath_rcvhdrsize);
1382 } else if (rhdrsize > (dd->ipath_rcvhdrentsize -
1383 (sizeof(u64) / sizeof(u32)))) {
1384 ipath_dbg("Error: can't set protocol header size %u "
1385 "(> max %u)\n", rhdrsize,
1386 dd->ipath_rcvhdrentsize -
1387 (u32) (sizeof(u64) / sizeof(u32)));
1390 dd->ipath_flags |= IPATH_RCVHDRSZ_SET;
1391 dd->ipath_rcvhdrsize = rhdrsize;
1392 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvhdrsize,
1393 dd->ipath_rcvhdrsize);
1394 ipath_cdbg(VERBOSE, "Set protocol header size to %u\n",
1395 dd->ipath_rcvhdrsize);
1401 * debugging code and stats updates if no pio buffers available.
1403 static noinline void no_pio_bufs(struct ipath_devdata *dd)
1405 unsigned long *shadow = dd->ipath_pioavailshadow;
1406 __le64 *dma = (__le64 *)dd->ipath_pioavailregs_dma;
1408 dd->ipath_upd_pio_shadow = 1;
1411 * not atomic, but if we lose a stat count in a while, that's OK
1413 ipath_stats.sps_nopiobufs++;
1414 if (!(++dd->ipath_consec_nopiobuf % 100000)) {
1415 ipath_dbg("%u pio sends with no bufavail; dmacopy: "
1416 "%llx %llx %llx %llx; shadow: %lx %lx %lx %lx\n",
1417 dd->ipath_consec_nopiobuf,
1418 (unsigned long long) le64_to_cpu(dma[0]),
1419 (unsigned long long) le64_to_cpu(dma[1]),
1420 (unsigned long long) le64_to_cpu(dma[2]),
1421 (unsigned long long) le64_to_cpu(dma[3]),
1422 shadow[0], shadow[1], shadow[2], shadow[3]);
1424 * 4 buffers per byte, 4 registers above, cover rest
1427 if ((dd->ipath_piobcnt2k + dd->ipath_piobcnt4k) >
1428 (sizeof(shadow[0]) * 4 * 4))
1429 ipath_dbg("2nd group: dmacopy: %llx %llx "
1430 "%llx %llx; shadow: %lx %lx %lx %lx\n",
1431 (unsigned long long)le64_to_cpu(dma[4]),
1432 (unsigned long long)le64_to_cpu(dma[5]),
1433 (unsigned long long)le64_to_cpu(dma[6]),
1434 (unsigned long long)le64_to_cpu(dma[7]),
1435 shadow[4], shadow[5], shadow[6],
1441 * common code for normal driver pio buffer allocation, and reserved
1444 * do appropriate marking as busy, etc.
1445 * returns buffer number if one found (>=0), negative number is error.
1447 static u32 __iomem *ipath_getpiobuf_range(struct ipath_devdata *dd,
1448 u32 *pbufnum, u32 first, u32 last, u32 firsti)
1450 int i, j, updated = 0;
1452 unsigned long flags;
1453 unsigned long *shadow = dd->ipath_pioavailshadow;
1456 piobcnt = last - first;
1457 if (dd->ipath_upd_pio_shadow) {
1459 * Minor optimization. If we had no buffers on last call,
1460 * start out by doing the update; continue and do scan even
1461 * if no buffers were updated, to be paranoid
1463 ipath_update_pio_bufs(dd);
1470 * while test_and_set_bit() is atomic, we do that and then the
1471 * change_bit(), and the pair is not. See if this is the cause
1472 * of the remaining armlaunch errors.
1474 spin_lock_irqsave(&ipath_pioavail_lock, flags);
1475 for (j = 0; j < piobcnt; j++, i++) {
1478 if (__test_and_set_bit((2 * i) + 1, shadow))
1480 /* flip generation bit */
1481 __change_bit(2 * i, shadow);
1484 spin_unlock_irqrestore(&ipath_pioavail_lock, flags);
1489 * first time through; shadow exhausted, but may be
1490 * buffers available, try an update and then rescan.
1492 ipath_update_pio_bufs(dd);
1496 } else if (updated == 1 && piobcnt <=
1497 ((dd->ipath_sendctrl
1498 >> INFINIPATH_S_UPDTHRESH_SHIFT) &
1499 INFINIPATH_S_UPDTHRESH_MASK)) {
1501 * for chips supporting and using the update
1502 * threshold we need to force an update of the
1503 * in-memory copy if the count is less than the
1504 * thershold, then check one more time.
1506 ipath_force_pio_avail_update(dd);
1507 ipath_update_pio_bufs(dd);
1516 if (i < dd->ipath_piobcnt2k)
1517 buf = (u32 __iomem *) (dd->ipath_pio2kbase +
1518 i * dd->ipath_palign);
1520 buf = (u32 __iomem *)
1521 (dd->ipath_pio4kbase +
1522 (i - dd->ipath_piobcnt2k) * dd->ipath_4kalign);
1531 * ipath_getpiobuf - find an available pio buffer
1532 * @dd: the infinipath device
1533 * @plen: the size of the PIO buffer needed in 32-bit words
1534 * @pbufnum: the buffer number is placed here
1536 u32 __iomem *ipath_getpiobuf(struct ipath_devdata *dd, u32 plen, u32 *pbufnum)
1542 if (plen + 1 >= IPATH_SMALLBUF_DWORDS) {
1543 first = dd->ipath_piobcnt2k;
1544 lasti = dd->ipath_lastpioindexl;
1547 lasti = dd->ipath_lastpioindex;
1549 nbufs = dd->ipath_piobcnt2k + dd->ipath_piobcnt4k;
1550 buf = ipath_getpiobuf_range(dd, &pnum, first, nbufs, lasti);
1554 * Set next starting place. It's just an optimization,
1555 * it doesn't matter who wins on this, so no locking
1557 if (plen + 1 >= IPATH_SMALLBUF_DWORDS)
1558 dd->ipath_lastpioindexl = pnum + 1;
1560 dd->ipath_lastpioindex = pnum + 1;
1561 if (dd->ipath_upd_pio_shadow)
1562 dd->ipath_upd_pio_shadow = 0;
1563 if (dd->ipath_consec_nopiobuf)
1564 dd->ipath_consec_nopiobuf = 0;
1565 ipath_cdbg(VERBOSE, "Return piobuf%u %uk @ %p\n",
1566 pnum, (pnum < dd->ipath_piobcnt2k) ? 2 : 4, buf);
1575 * ipath_chg_pioavailkernel - change which send buffers are available for kernel
1576 * @dd: the infinipath device
1577 * @start: the starting send buffer number
1578 * @len: the number of send buffers
1579 * @avail: true if the buffers are available for kernel use, false otherwise
1581 void ipath_chg_pioavailkernel(struct ipath_devdata *dd, unsigned start,
1582 unsigned len, int avail)
1584 unsigned long flags;
1587 /* There are two bits per send buffer (busy and generation) */
1592 /* Set or clear the generation bits. */
1593 spin_lock_irqsave(&ipath_pioavail_lock, flags);
1594 while (start < end) {
1596 __clear_bit(start + INFINIPATH_SENDPIOAVAIL_BUSY_SHIFT,
1597 dd->ipath_pioavailshadow);
1598 __set_bit(start, dd->ipath_pioavailkernel);
1600 __set_bit(start + INFINIPATH_SENDPIOAVAIL_BUSY_SHIFT,
1601 dd->ipath_pioavailshadow);
1602 __clear_bit(start, dd->ipath_pioavailkernel);
1606 spin_unlock_irqrestore(&ipath_pioavail_lock, flags);
1610 * ipath_create_rcvhdrq - create a receive header queue
1611 * @dd: the infinipath device
1612 * @pd: the port data
1614 * this must be contiguous memory (from an i/o perspective), and must be
1615 * DMA'able (which means for some systems, it will go through an IOMMU,
1616 * or be forced into a low address range).
1618 int ipath_create_rcvhdrq(struct ipath_devdata *dd,
1619 struct ipath_portdata *pd)
1623 if (!pd->port_rcvhdrq) {
1624 dma_addr_t phys_hdrqtail;
1625 gfp_t gfp_flags = GFP_USER | __GFP_COMP;
1626 int amt = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
1627 sizeof(u32), PAGE_SIZE);
1629 pd->port_rcvhdrq = dma_alloc_coherent(
1630 &dd->pcidev->dev, amt, &pd->port_rcvhdrq_phys,
1633 if (!pd->port_rcvhdrq) {
1634 ipath_dev_err(dd, "attempt to allocate %d bytes "
1635 "for port %u rcvhdrq failed\n",
1636 amt, pd->port_port);
1640 pd->port_rcvhdrtail_kvaddr = dma_alloc_coherent(
1641 &dd->pcidev->dev, PAGE_SIZE, &phys_hdrqtail, GFP_KERNEL);
1642 if (!pd->port_rcvhdrtail_kvaddr) {
1643 ipath_dev_err(dd, "attempt to allocate 1 page "
1644 "for port %u rcvhdrqtailaddr failed\n",
1647 dma_free_coherent(&dd->pcidev->dev, amt,
1648 pd->port_rcvhdrq, pd->port_rcvhdrq_phys);
1649 pd->port_rcvhdrq = NULL;
1652 pd->port_rcvhdrqtailaddr_phys = phys_hdrqtail;
1654 pd->port_rcvhdrq_size = amt;
1656 ipath_cdbg(VERBOSE, "%d pages at %p (phys %lx) size=%lu "
1657 "for port %u rcvhdr Q\n",
1658 amt >> PAGE_SHIFT, pd->port_rcvhdrq,
1659 (unsigned long) pd->port_rcvhdrq_phys,
1660 (unsigned long) pd->port_rcvhdrq_size,
1663 ipath_cdbg(VERBOSE, "port %d hdrtailaddr, %llx physical\n",
1665 (unsigned long long) phys_hdrqtail);
1668 ipath_cdbg(VERBOSE, "reuse port %d rcvhdrq @%p %llx phys; "
1669 "hdrtailaddr@%p %llx physical\n",
1670 pd->port_port, pd->port_rcvhdrq,
1671 (unsigned long long) pd->port_rcvhdrq_phys,
1672 pd->port_rcvhdrtail_kvaddr, (unsigned long long)
1673 pd->port_rcvhdrqtailaddr_phys);
1675 /* clear for security and sanity on each use */
1676 memset(pd->port_rcvhdrq, 0, pd->port_rcvhdrq_size);
1677 if (pd->port_rcvhdrtail_kvaddr)
1678 memset(pd->port_rcvhdrtail_kvaddr, 0, PAGE_SIZE);
1681 * tell chip each time we init it, even if we are re-using previous
1682 * memory (we zero the register at process close)
1684 ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdrtailaddr,
1685 pd->port_port, pd->port_rcvhdrqtailaddr_phys);
1686 ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdraddr,
1687 pd->port_port, pd->port_rcvhdrq_phys);
1696 * Flush all sends that might be in the ready to send state, as well as any
1697 * that are in the process of being sent. Used whenever we need to be
1698 * sure the send side is idle. Cleans up all buffer state by canceling
1699 * all pio buffers, and issuing an abort, which cleans up anything in the
1700 * launch fifo. The cancel is superfluous on some chip versions, but
1701 * it's safer to always do it.
1702 * PIOAvail bits are updated by the chip as if normal send had happened.
1704 void ipath_cancel_sends(struct ipath_devdata *dd, int restore_sendctrl)
1706 ipath_dbg("Cancelling all in-progress send buffers\n");
1707 dd->ipath_lastcancel = jiffies+HZ/2; /* skip armlaunch errs a bit */
1709 * the abort bit is auto-clearing. We read scratch to be sure
1710 * that cancels and the abort have taken effect in the chip.
1712 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
1713 INFINIPATH_S_ABORT);
1714 ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
1715 ipath_disarm_piobufs(dd, 0,
1716 (unsigned)(dd->ipath_piobcnt2k + dd->ipath_piobcnt4k));
1717 if (restore_sendctrl) /* else done by caller later */
1718 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
1719 dd->ipath_sendctrl);
1721 /* and again, be sure all have hit the chip */
1722 ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
1726 * Force an update of in-memory copy of the pioavail registers, when
1727 * needed for any of a variety of reasons. We read the scratch register
1728 * to make it highly likely that the update will have happened by the
1729 * time we return. If already off (as in cancel_sends above), this
1730 * routine is a nop, on the assumption that the caller will "do the
1733 void ipath_force_pio_avail_update(struct ipath_devdata *dd)
1735 unsigned long flags;
1737 spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
1738 if (dd->ipath_sendctrl & INFINIPATH_S_PIOBUFAVAILUPD) {
1739 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
1740 dd->ipath_sendctrl & ~INFINIPATH_S_PIOBUFAVAILUPD);
1741 ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
1742 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
1743 dd->ipath_sendctrl);
1744 ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
1746 spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
1749 static void ipath_set_ib_lstate(struct ipath_devdata *dd, int linkcmd,
1753 static const char *what[4] = {
1755 [INFINIPATH_IBCC_LINKCMD_DOWN] = "DOWN",
1756 [INFINIPATH_IBCC_LINKCMD_ARMED] = "ARMED",
1757 [INFINIPATH_IBCC_LINKCMD_ACTIVE] = "ACTIVE"
1760 if (linitcmd == INFINIPATH_IBCC_LINKINITCMD_DISABLE) {
1762 * If we are told to disable, note that so link-recovery
1763 * code does not attempt to bring us back up.
1766 dd->ipath_flags |= IPATH_IB_LINK_DISABLED;
1768 } else if (linitcmd) {
1770 * Any other linkinitcmd will lead to LINKDOWN and then
1771 * to INIT (if all is well), so clear flag to let
1772 * link-recovery code attempt to bring us back up.
1775 dd->ipath_flags &= ~IPATH_IB_LINK_DISABLED;
1779 mod_wd = (linkcmd << dd->ibcc_lc_shift) |
1780 (linitcmd << INFINIPATH_IBCC_LINKINITCMD_SHIFT);
1782 "Moving unit %u to %s (initcmd=0x%x), current ltstate is %s\n",
1783 dd->ipath_unit, what[linkcmd], linitcmd,
1784 ipath_ibcstatus_str[ipath_ib_linktrstate(dd,
1785 ipath_read_kreg64(dd, dd->ipath_kregs->kr_ibcstatus))]);
1787 ipath_write_kreg(dd, dd->ipath_kregs->kr_ibcctrl,
1788 dd->ipath_ibcctrl | mod_wd);
1789 /* read from chip so write is flushed */
1790 (void) ipath_read_kreg64(dd, dd->ipath_kregs->kr_ibcstatus);
1793 int ipath_set_linkstate(struct ipath_devdata *dd, u8 newstate)
1799 case IPATH_IB_LINKDOWN_ONLY:
1800 ipath_set_ib_lstate(dd, INFINIPATH_IBCC_LINKCMD_DOWN, 0);
1805 case IPATH_IB_LINKDOWN:
1806 ipath_set_ib_lstate(dd, INFINIPATH_IBCC_LINKCMD_DOWN,
1807 INFINIPATH_IBCC_LINKINITCMD_POLL);
1812 case IPATH_IB_LINKDOWN_SLEEP:
1813 ipath_set_ib_lstate(dd, INFINIPATH_IBCC_LINKCMD_DOWN,
1814 INFINIPATH_IBCC_LINKINITCMD_SLEEP);
1819 case IPATH_IB_LINKDOWN_DISABLE:
1820 ipath_set_ib_lstate(dd, INFINIPATH_IBCC_LINKCMD_DOWN,
1821 INFINIPATH_IBCC_LINKINITCMD_DISABLE);
1826 case IPATH_IB_LINKARM:
1827 if (dd->ipath_flags & IPATH_LINKARMED) {
1831 if (!(dd->ipath_flags &
1832 (IPATH_LINKINIT | IPATH_LINKACTIVE))) {
1836 ipath_set_ib_lstate(dd, INFINIPATH_IBCC_LINKCMD_ARMED, 0);
1839 * Since the port can transition to ACTIVE by receiving
1840 * a non VL 15 packet, wait for either state.
1842 lstate = IPATH_LINKARMED | IPATH_LINKACTIVE;
1845 case IPATH_IB_LINKACTIVE:
1846 if (dd->ipath_flags & IPATH_LINKACTIVE) {
1850 if (!(dd->ipath_flags & IPATH_LINKARMED)) {
1854 ipath_set_ib_lstate(dd, INFINIPATH_IBCC_LINKCMD_ACTIVE, 0);
1855 lstate = IPATH_LINKACTIVE;
1858 case IPATH_IB_LINK_LOOPBACK:
1859 dev_info(&dd->pcidev->dev, "Enabling IB local loopback\n");
1860 dd->ipath_ibcctrl |= INFINIPATH_IBCC_LOOPBACK;
1861 ipath_write_kreg(dd, dd->ipath_kregs->kr_ibcctrl,
1864 goto bail; // no state change to wait for
1866 case IPATH_IB_LINK_EXTERNAL:
1867 dev_info(&dd->pcidev->dev, "Disabling IB local loopback (normal)\n");
1868 dd->ipath_ibcctrl &= ~INFINIPATH_IBCC_LOOPBACK;
1869 ipath_write_kreg(dd, dd->ipath_kregs->kr_ibcctrl,
1872 goto bail; // no state change to wait for
1875 ipath_dbg("Invalid linkstate 0x%x requested\n", newstate);
1879 ret = ipath_wait_linkstate(dd, lstate, 2000);
1886 * ipath_set_mtu - set the MTU
1887 * @dd: the infinipath device
1890 * we can handle "any" incoming size, the issue here is whether we
1891 * need to restrict our outgoing size. For now, we don't do any
1892 * sanity checking on this, and we don't deal with what happens to
1893 * programs that are already running when the size changes.
1894 * NOTE: changing the MTU will usually cause the IBC to go back to
1895 * link initialize (IPATH_IBSTATE_INIT) state...
1897 int ipath_set_mtu(struct ipath_devdata *dd, u16 arg)
1904 * mtu is IB data payload max. It's the largest power of 2 less
1905 * than piosize (or even larger, since it only really controls the
1906 * largest we can receive; we can send the max of the mtu and
1907 * piosize). We check that it's one of the valid IB sizes.
1909 if (arg != 256 && arg != 512 && arg != 1024 && arg != 2048 &&
1910 (arg != 4096 || !ipath_mtu4096)) {
1911 ipath_dbg("Trying to set invalid mtu %u, failing\n", arg);
1915 if (dd->ipath_ibmtu == arg) {
1916 ret = 0; /* same as current */
1920 piosize = dd->ipath_ibmaxlen;
1921 dd->ipath_ibmtu = arg;
1923 if (arg >= (piosize - IPATH_PIO_MAXIBHDR)) {
1924 /* Only if it's not the initial value (or reset to it) */
1925 if (piosize != dd->ipath_init_ibmaxlen) {
1926 if (arg > piosize && arg <= dd->ipath_init_ibmaxlen)
1927 piosize = dd->ipath_init_ibmaxlen;
1928 dd->ipath_ibmaxlen = piosize;
1931 } else if ((arg + IPATH_PIO_MAXIBHDR) != dd->ipath_ibmaxlen) {
1932 piosize = arg + IPATH_PIO_MAXIBHDR;
1933 ipath_cdbg(VERBOSE, "ibmaxlen was 0x%x, setting to 0x%x "
1934 "(mtu 0x%x)\n", dd->ipath_ibmaxlen, piosize,
1936 dd->ipath_ibmaxlen = piosize;
1941 u64 ibc = dd->ipath_ibcctrl, ibdw;
1943 * update our housekeeping variables, and set IBC max
1944 * size, same as init code; max IBC is max we allow in
1945 * buffer, less the qword pbc, plus 1 for ICRC, in dwords
1947 dd->ipath_ibmaxlen = piosize - 2 * sizeof(u32);
1948 ibdw = (dd->ipath_ibmaxlen >> 2) + 1;
1949 ibc &= ~(INFINIPATH_IBCC_MAXPKTLEN_MASK <<
1950 dd->ibcc_mpl_shift);
1951 ibc |= ibdw << dd->ibcc_mpl_shift;
1952 dd->ipath_ibcctrl = ibc;
1953 ipath_write_kreg(dd, dd->ipath_kregs->kr_ibcctrl,
1955 dd->ipath_f_tidtemplate(dd);
1964 int ipath_set_lid(struct ipath_devdata *dd, u32 arg, u8 lmc)
1966 dd->ipath_lid = arg;
1967 dd->ipath_lmc = lmc;
1974 * ipath_write_kreg_port - write a device's per-port 64-bit kernel register
1975 * @dd: the infinipath device
1976 * @regno: the register number to write
1977 * @port: the port containing the register
1978 * @value: the value to write
1980 * Registers that vary with the chip implementation constants (port)
1983 void ipath_write_kreg_port(const struct ipath_devdata *dd, ipath_kreg regno,
1984 unsigned port, u64 value)
1988 if (port < dd->ipath_portcnt &&
1989 (regno == dd->ipath_kregs->kr_rcvhdraddr ||
1990 regno == dd->ipath_kregs->kr_rcvhdrtailaddr))
1991 where = regno + port;
1995 ipath_write_kreg(dd, where, value);
1999 * Following deal with the "obviously simple" task of overriding the state
2000 * of the LEDS, which normally indicate link physical and logical status.
2001 * The complications arise in dealing with different hardware mappings
2002 * and the board-dependent routine being called from interrupts.
2003 * and then there's the requirement to _flash_ them.
2005 #define LED_OVER_FREQ_SHIFT 8
2006 #define LED_OVER_FREQ_MASK (0xFF<<LED_OVER_FREQ_SHIFT)
2007 /* Below is "non-zero" to force override, but both actual LEDs are off */
2008 #define LED_OVER_BOTH_OFF (8)
2010 static void ipath_run_led_override(unsigned long opaque)
2012 struct ipath_devdata *dd = (struct ipath_devdata *)opaque;
2015 u64 lstate, ltstate, val;
2017 if (!(dd->ipath_flags & IPATH_INITTED))
2020 pidx = dd->ipath_led_override_phase++ & 1;
2021 dd->ipath_led_override = dd->ipath_led_override_vals[pidx];
2022 timeoff = dd->ipath_led_override_timeoff;
2025 * below potentially restores the LED values per current status,
2026 * should also possibly setup the traffic-blink register,
2027 * but leave that to per-chip functions.
2029 val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_ibcstatus);
2030 ltstate = (val >> INFINIPATH_IBCS_LINKTRAININGSTATE_SHIFT) &
2032 lstate = (val >> dd->ibcs_ls_shift) & INFINIPATH_IBCS_LINKSTATE_MASK;
2034 dd->ipath_f_setextled(dd, lstate, ltstate);
2035 mod_timer(&dd->ipath_led_override_timer, jiffies + timeoff);
2038 void ipath_set_led_override(struct ipath_devdata *dd, unsigned int val)
2042 if (!(dd->ipath_flags & IPATH_INITTED))
2045 /* First check if we are blinking. If not, use 1HZ polling */
2047 freq = (val & LED_OVER_FREQ_MASK) >> LED_OVER_FREQ_SHIFT;
2050 /* For blink, set each phase from one nybble of val */
2051 dd->ipath_led_override_vals[0] = val & 0xF;
2052 dd->ipath_led_override_vals[1] = (val >> 4) & 0xF;
2053 timeoff = (HZ << 4)/freq;
2055 /* Non-blink set both phases the same. */
2056 dd->ipath_led_override_vals[0] = val & 0xF;
2057 dd->ipath_led_override_vals[1] = val & 0xF;
2059 dd->ipath_led_override_timeoff = timeoff;
2062 * If the timer has not already been started, do so. Use a "quick"
2063 * timeout so the function will be called soon, to look at our request.
2065 if (atomic_inc_return(&dd->ipath_led_override_timer_active) == 1) {
2066 /* Need to start timer */
2067 init_timer(&dd->ipath_led_override_timer);
2068 dd->ipath_led_override_timer.function =
2069 ipath_run_led_override;
2070 dd->ipath_led_override_timer.data = (unsigned long) dd;
2071 dd->ipath_led_override_timer.expires = jiffies + 1;
2072 add_timer(&dd->ipath_led_override_timer);
2074 atomic_dec(&dd->ipath_led_override_timer_active);
2079 * ipath_shutdown_device - shut down a device
2080 * @dd: the infinipath device
2082 * This is called to make the device quiet when we are about to
2083 * unload the driver, and also when the device is administratively
2084 * disabled. It does not free any data structures.
2085 * Everything it does has to be setup again by ipath_init_chip(dd,1)
2087 void ipath_shutdown_device(struct ipath_devdata *dd)
2089 unsigned long flags;
2091 ipath_dbg("Shutting down the device\n");
2093 ipath_hol_up(dd); /* make sure user processes aren't suspended */
2095 dd->ipath_flags |= IPATH_LINKUNK;
2096 dd->ipath_flags &= ~(IPATH_INITTED | IPATH_LINKDOWN |
2097 IPATH_LINKINIT | IPATH_LINKARMED |
2099 *dd->ipath_statusp &= ~(IPATH_STATUS_IB_CONF |
2100 IPATH_STATUS_IB_READY);
2102 /* mask interrupts, but not errors */
2103 ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask, 0ULL);
2105 dd->ipath_rcvctrl = 0;
2106 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
2110 * gracefully stop all sends allowing any in progress to trickle out
2113 spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
2114 dd->ipath_sendctrl = 0;
2115 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, dd->ipath_sendctrl);
2117 ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
2118 spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
2121 * enough for anything that's going to trickle out to have actually
2126 ipath_set_ib_lstate(dd, 0, INFINIPATH_IBCC_LINKINITCMD_DISABLE);
2127 ipath_cancel_sends(dd, 0);
2129 signal_ib_event(dd, IB_EVENT_PORT_ERR);
2132 dd->ipath_control &= ~INFINIPATH_C_LINKENABLE;
2133 ipath_write_kreg(dd, dd->ipath_kregs->kr_control,
2134 dd->ipath_control | INFINIPATH_C_FREEZEMODE);
2137 * clear SerdesEnable and turn the leds off; do this here because
2138 * we are unloading, so don't count on interrupts to move along
2139 * Turn the LEDs off explictly for the same reason.
2141 dd->ipath_f_quiet_serdes(dd);
2143 /* stop all the timers that might still be running */
2144 del_timer_sync(&dd->ipath_hol_timer);
2145 if (dd->ipath_stats_timer_active) {
2146 del_timer_sync(&dd->ipath_stats_timer);
2147 dd->ipath_stats_timer_active = 0;
2151 * clear all interrupts and errors, so that the next time the driver
2152 * is loaded or device is enabled, we know that whatever is set
2153 * happened while we were unloaded
2155 ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear,
2156 ~0ULL & ~INFINIPATH_HWE_MEMBISTFAILED);
2157 ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear, -1LL);
2158 ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, -1LL);
2160 ipath_cdbg(VERBOSE, "Flush time and errors to EEPROM\n");
2161 ipath_update_eeprom_log(dd);
2165 * ipath_free_pddata - free a port's allocated data
2166 * @dd: the infinipath device
2167 * @pd: the portdata structure
2169 * free up any allocated data for a port
2170 * This should not touch anything that would affect a simultaneous
2171 * re-allocation of port data, because it is called after ipath_mutex
2172 * is released (and can be called from reinit as well).
2173 * It should never change any chip state, or global driver state.
2174 * (The only exception to global state is freeing the port0 port0_skbs.)
2176 void ipath_free_pddata(struct ipath_devdata *dd, struct ipath_portdata *pd)
2181 if (pd->port_rcvhdrq) {
2182 ipath_cdbg(VERBOSE, "free closed port %d rcvhdrq @ %p "
2183 "(size=%lu)\n", pd->port_port, pd->port_rcvhdrq,
2184 (unsigned long) pd->port_rcvhdrq_size);
2185 dma_free_coherent(&dd->pcidev->dev, pd->port_rcvhdrq_size,
2186 pd->port_rcvhdrq, pd->port_rcvhdrq_phys);
2187 pd->port_rcvhdrq = NULL;
2188 if (pd->port_rcvhdrtail_kvaddr) {
2189 dma_free_coherent(&dd->pcidev->dev, PAGE_SIZE,
2190 pd->port_rcvhdrtail_kvaddr,
2191 pd->port_rcvhdrqtailaddr_phys);
2192 pd->port_rcvhdrtail_kvaddr = NULL;
2195 if (pd->port_port && pd->port_rcvegrbuf) {
2198 for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) {
2199 void *base = pd->port_rcvegrbuf[e];
2200 size_t size = pd->port_rcvegrbuf_size;
2202 ipath_cdbg(VERBOSE, "egrbuf free(%p, %lu), "
2203 "chunk %u/%u\n", base,
2204 (unsigned long) size,
2205 e, pd->port_rcvegrbuf_chunks);
2206 dma_free_coherent(&dd->pcidev->dev, size,
2207 base, pd->port_rcvegrbuf_phys[e]);
2209 kfree(pd->port_rcvegrbuf);
2210 pd->port_rcvegrbuf = NULL;
2211 kfree(pd->port_rcvegrbuf_phys);
2212 pd->port_rcvegrbuf_phys = NULL;
2213 pd->port_rcvegrbuf_chunks = 0;
2214 } else if (pd->port_port == 0 && dd->ipath_port0_skbinfo) {
2216 struct ipath_skbinfo *skbinfo = dd->ipath_port0_skbinfo;
2218 dd->ipath_port0_skbinfo = NULL;
2219 ipath_cdbg(VERBOSE, "free closed port %d "
2220 "ipath_port0_skbinfo @ %p\n", pd->port_port,
2222 for (e = 0; e < dd->ipath_rcvegrcnt; e++)
2223 if (skbinfo[e].skb) {
2224 pci_unmap_single(dd->pcidev, skbinfo[e].phys,
2226 PCI_DMA_FROMDEVICE);
2227 dev_kfree_skb(skbinfo[e].skb);
2231 kfree(pd->port_tid_pg_list);
2232 vfree(pd->subport_uregbase);
2233 vfree(pd->subport_rcvegrbuf);
2234 vfree(pd->subport_rcvhdr_base);
2238 static int __init infinipath_init(void)
2242 if (ipath_debug & __IPATH_DBG)
2243 printk(KERN_INFO DRIVER_LOAD_MSG "%s", ib_ipath_version);
2246 * These must be called before the driver is registered with
2247 * the PCI subsystem.
2249 idr_init(&unit_table);
2250 if (!idr_pre_get(&unit_table, GFP_KERNEL)) {
2255 ret = pci_register_driver(&ipath_driver);
2257 printk(KERN_ERR IPATH_DRV_NAME
2258 ": Unable to register driver: error %d\n", -ret);
2262 ret = ipath_init_ipathfs();
2264 printk(KERN_ERR IPATH_DRV_NAME ": Unable to create "
2265 "ipathfs: error %d\n", -ret);
2272 pci_unregister_driver(&ipath_driver);
2275 idr_destroy(&unit_table);
2281 static void __exit infinipath_cleanup(void)
2283 ipath_exit_ipathfs();
2285 ipath_cdbg(VERBOSE, "Unregistering pci driver\n");
2286 pci_unregister_driver(&ipath_driver);
2288 idr_destroy(&unit_table);
2292 * ipath_reset_device - reset the chip if possible
2293 * @unit: the device to reset
2295 * Whether or not reset is successful, we attempt to re-initialize the chip
2296 * (that is, much like a driver unload/reload). We clear the INITTED flag
2297 * so that the various entry points will fail until we reinitialize. For
2298 * now, we only allow this if no user ports are open that use chip resources
2300 int ipath_reset_device(int unit)
2303 struct ipath_devdata *dd = ipath_lookup(unit);
2310 if (atomic_read(&dd->ipath_led_override_timer_active)) {
2311 /* Need to stop LED timer, _then_ shut off LEDs */
2312 del_timer_sync(&dd->ipath_led_override_timer);
2313 atomic_set(&dd->ipath_led_override_timer_active, 0);
2316 /* Shut off LEDs after we are sure timer is not running */
2317 dd->ipath_led_override = LED_OVER_BOTH_OFF;
2318 dd->ipath_f_setextled(dd, 0, 0);
2320 dev_info(&dd->pcidev->dev, "Reset on unit %u requested\n", unit);
2322 if (!dd->ipath_kregbase || !(dd->ipath_flags & IPATH_PRESENT)) {
2323 dev_info(&dd->pcidev->dev, "Invalid unit number %u or "
2324 "not initialized or not present\n", unit);
2330 for (i = 1; i < dd->ipath_cfgports; i++) {
2331 if (dd->ipath_pd[i] && dd->ipath_pd[i]->port_cnt) {
2332 ipath_dbg("unit %u port %d is in use "
2333 "(PID %u cmd %s), can't reset\n",
2335 dd->ipath_pd[i]->port_pid,
2336 dd->ipath_pd[i]->port_comm);
2342 dd->ipath_flags &= ~IPATH_INITTED;
2343 ret = dd->ipath_f_reset(dd);
2345 ipath_dbg("reset was not successful\n");
2346 ipath_dbg("Trying to reinitialize unit %u after reset attempt\n",
2348 ret = ipath_init_chip(dd, 1);
2350 ipath_dev_err(dd, "Reinitialize unit %u after "
2351 "reset failed with %d\n", unit, ret);
2353 dev_info(&dd->pcidev->dev, "Reinitialized unit %u after "
2354 "resetting\n", unit);
2361 * send a signal to all the processes that have the driver open
2362 * through the normal interfaces (i.e., everything other than diags
2363 * interface). Returns number of signalled processes.
2365 static int ipath_signal_procs(struct ipath_devdata *dd, int sig)
2367 int i, sub, any = 0;
2372 for (i = 1; i < dd->ipath_cfgports; i++) {
2373 if (!dd->ipath_pd[i] || !dd->ipath_pd[i]->port_cnt ||
2374 !dd->ipath_pd[i]->port_pid)
2376 pid = dd->ipath_pd[i]->port_pid;
2377 dev_info(&dd->pcidev->dev, "context %d in use "
2378 "(PID %u), sending signal %d\n",
2380 kill_proc(pid, sig, 1);
2382 for (sub = 0; sub < INFINIPATH_MAX_SUBPORT; sub++) {
2383 pid = dd->ipath_pd[i]->port_subpid[sub];
2386 dev_info(&dd->pcidev->dev, "sub-context "
2387 "%d:%d in use (PID %u), sending "
2388 "signal %d\n", i, sub, pid, sig);
2389 kill_proc(pid, sig, 1);
2396 static void ipath_hol_signal_down(struct ipath_devdata *dd)
2398 if (ipath_signal_procs(dd, SIGSTOP))
2399 ipath_dbg("Stopped some processes\n");
2400 ipath_cancel_sends(dd, 1);
2404 static void ipath_hol_signal_up(struct ipath_devdata *dd)
2406 if (ipath_signal_procs(dd, SIGCONT))
2407 ipath_dbg("Continued some processes\n");
2411 * link is down, stop any users processes, and flush pending sends
2412 * to prevent HoL blocking, then start the HoL timer that
2413 * periodically continues, then stop procs, so they can detect
2414 * link down if they want, and do something about it.
2415 * Timer may already be running, so use __mod_timer, not add_timer.
2417 void ipath_hol_down(struct ipath_devdata *dd)
2419 dd->ipath_hol_state = IPATH_HOL_DOWN;
2420 ipath_hol_signal_down(dd);
2421 dd->ipath_hol_next = IPATH_HOL_DOWNCONT;
2422 dd->ipath_hol_timer.expires = jiffies +
2423 msecs_to_jiffies(ipath_hol_timeout_ms);
2424 __mod_timer(&dd->ipath_hol_timer, dd->ipath_hol_timer.expires);
2428 * link is up, continue any user processes, and ensure timer
2429 * is a nop, if running. Let timer keep running, if set; it
2430 * will nop when it sees the link is up
2432 void ipath_hol_up(struct ipath_devdata *dd)
2434 ipath_hol_signal_up(dd);
2435 dd->ipath_hol_state = IPATH_HOL_UP;
2439 * toggle the running/not running state of user proceses
2440 * to prevent HoL blocking on chip resources, but still allow
2441 * user processes to do link down special case handling.
2442 * Should only be called via the timer
2444 void ipath_hol_event(unsigned long opaque)
2446 struct ipath_devdata *dd = (struct ipath_devdata *)opaque;
2448 if (dd->ipath_hol_next == IPATH_HOL_DOWNSTOP
2449 && dd->ipath_hol_state != IPATH_HOL_UP) {
2450 dd->ipath_hol_next = IPATH_HOL_DOWNCONT;
2451 ipath_dbg("Stopping processes\n");
2452 ipath_hol_signal_down(dd);
2453 } else { /* may do "extra" if also in ipath_hol_up() */
2454 dd->ipath_hol_next = IPATH_HOL_DOWNSTOP;
2455 ipath_dbg("Continuing processes\n");
2456 ipath_hol_signal_up(dd);
2458 if (dd->ipath_hol_state == IPATH_HOL_UP)
2459 ipath_dbg("link's up, don't resched timer\n");
2461 dd->ipath_hol_timer.expires = jiffies +
2462 msecs_to_jiffies(ipath_hol_timeout_ms);
2463 __mod_timer(&dd->ipath_hol_timer,
2464 dd->ipath_hol_timer.expires);
2468 int ipath_set_rx_pol_inv(struct ipath_devdata *dd, u8 new_pol_inv)
2471 if ( new_pol_inv > INFINIPATH_XGXS_RX_POL_MASK ) {
2474 if ( dd->ipath_rx_pol_inv != new_pol_inv ) {
2475 dd->ipath_rx_pol_inv = new_pol_inv;
2476 val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_xgxsconfig);
2477 val &= ~(INFINIPATH_XGXS_RX_POL_MASK <<
2478 INFINIPATH_XGXS_RX_POL_SHIFT);
2479 val |= ((u64)dd->ipath_rx_pol_inv) <<
2480 INFINIPATH_XGXS_RX_POL_SHIFT;
2481 ipath_write_kreg(dd, dd->ipath_kregs->kr_xgxsconfig, val);
2487 * Disable and enable the armlaunch error. Used for PIO bandwidth testing on
2488 * the 7220, which is count-based, rather than trigger-based. Safe for the
2489 * driver check, since it's at init. Not completely safe when used for
2490 * user-mode checking, since some error checking can be lost, but not
2491 * particularly risky, and only has problematic side-effects in the face of
2492 * very buggy user code. There is no reference counting, but that's also
2493 * fine, given the intended use.
2495 void ipath_enable_armlaunch(struct ipath_devdata *dd)
2497 dd->ipath_lasterror &= ~INFINIPATH_E_SPIOARMLAUNCH;
2498 ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear,
2499 INFINIPATH_E_SPIOARMLAUNCH);
2500 dd->ipath_errormask |= INFINIPATH_E_SPIOARMLAUNCH;
2501 ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask,
2502 dd->ipath_errormask);
2505 void ipath_disable_armlaunch(struct ipath_devdata *dd)
2507 /* so don't re-enable if already set */
2508 dd->ipath_maskederrs &= ~INFINIPATH_E_SPIOARMLAUNCH;
2509 dd->ipath_errormask &= ~INFINIPATH_E_SPIOARMLAUNCH;
2510 ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask,
2511 dd->ipath_errormask);
2514 module_init(infinipath_init);
2515 module_exit(infinipath_cleanup);