2 * Copyright(c) 2015, 2016 Intel Corporation.
4 * This file is provided under a dual BSD/GPLv2 license. When using or
5 * redistributing this file, you may do so under either license.
9 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of version 2 of the GNU General Public License as
11 * published by the Free Software Foundation.
13 * This program is distributed in the hope that it will be useful, but
14 * WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * General Public License for more details.
20 * Redistribution and use in source and binary forms, with or without
21 * modification, are permitted provided that the following conditions
24 * - Redistributions of source code must retain the above copyright
25 * notice, this list of conditions and the following disclaimer.
26 * - Redistributions in binary form must reproduce the above copyright
27 * notice, this list of conditions and the following disclaimer in
28 * the documentation and/or other materials provided with the
30 * - Neither the name of Intel Corporation nor the names of its
31 * contributors may be used to endorse or promote products derived
32 * from this software without specific prior written permission.
34 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
35 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
36 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
37 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
38 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
39 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
40 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
41 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
42 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
43 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
44 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
47 #include <linux/poll.h>
48 #include <linux/cdev.h>
49 #include <linux/vmalloc.h>
59 #include "user_sdma.h"
60 #include "user_exp_rcv.h"
66 #define pr_fmt(fmt) DRIVER_NAME ": " fmt
68 #define SEND_CTXT_HALT_TIMEOUT 1000 /* msecs */
71 * File operation functions
73 static int hfi1_file_open(struct inode *, struct file *);
74 static int hfi1_file_close(struct inode *, struct file *);
75 static ssize_t hfi1_write_iter(struct kiocb *, struct iov_iter *);
76 static unsigned int hfi1_poll(struct file *, struct poll_table_struct *);
77 static int hfi1_file_mmap(struct file *, struct vm_area_struct *);
79 static u64 kvirt_to_phys(void *);
80 static int assign_ctxt(struct file *, struct hfi1_user_info *);
81 static int init_subctxts(struct hfi1_ctxtdata *, const struct hfi1_user_info *);
82 static int user_init(struct file *);
83 static int get_ctxt_info(struct file *, void __user *, __u32);
84 static int get_base_info(struct file *, void __user *, __u32);
85 static int setup_ctxt(struct file *);
86 static int setup_subctxt(struct hfi1_ctxtdata *);
87 static int get_user_context(struct file *, struct hfi1_user_info *, int);
88 static int find_shared_ctxt(struct file *, const struct hfi1_user_info *);
89 static int allocate_ctxt(struct file *, struct hfi1_devdata *,
90 struct hfi1_user_info *);
91 static unsigned int poll_urgent(struct file *, struct poll_table_struct *);
92 static unsigned int poll_next(struct file *, struct poll_table_struct *);
93 static int user_event_ack(struct hfi1_ctxtdata *, int, unsigned long);
94 static int set_ctxt_pkey(struct hfi1_ctxtdata *, unsigned, u16);
95 static int manage_rcvq(struct hfi1_ctxtdata *, unsigned, int);
96 static int vma_fault(struct vm_area_struct *, struct vm_fault *);
97 static long hfi1_file_ioctl(struct file *fp, unsigned int cmd,
100 static const struct file_operations hfi1_file_ops = {
101 .owner = THIS_MODULE,
102 .write_iter = hfi1_write_iter,
103 .open = hfi1_file_open,
104 .release = hfi1_file_close,
105 .unlocked_ioctl = hfi1_file_ioctl,
107 .mmap = hfi1_file_mmap,
108 .llseek = noop_llseek,
111 static struct vm_operations_struct vm_ops = {
116 * Types of memories mapped into user processes' space
135 * Masks and offsets defining the mmap tokens
137 #define HFI1_MMAP_OFFSET_MASK 0xfffULL
138 #define HFI1_MMAP_OFFSET_SHIFT 0
139 #define HFI1_MMAP_SUBCTXT_MASK 0xfULL
140 #define HFI1_MMAP_SUBCTXT_SHIFT 12
141 #define HFI1_MMAP_CTXT_MASK 0xffULL
142 #define HFI1_MMAP_CTXT_SHIFT 16
143 #define HFI1_MMAP_TYPE_MASK 0xfULL
144 #define HFI1_MMAP_TYPE_SHIFT 24
145 #define HFI1_MMAP_MAGIC_MASK 0xffffffffULL
146 #define HFI1_MMAP_MAGIC_SHIFT 32
148 #define HFI1_MMAP_MAGIC 0xdabbad00
150 #define HFI1_MMAP_TOKEN_SET(field, val) \
151 (((val) & HFI1_MMAP_##field##_MASK) << HFI1_MMAP_##field##_SHIFT)
152 #define HFI1_MMAP_TOKEN_GET(field, token) \
153 (((token) >> HFI1_MMAP_##field##_SHIFT) & HFI1_MMAP_##field##_MASK)
154 #define HFI1_MMAP_TOKEN(type, ctxt, subctxt, addr) \
155 (HFI1_MMAP_TOKEN_SET(MAGIC, HFI1_MMAP_MAGIC) | \
156 HFI1_MMAP_TOKEN_SET(TYPE, type) | \
157 HFI1_MMAP_TOKEN_SET(CTXT, ctxt) | \
158 HFI1_MMAP_TOKEN_SET(SUBCTXT, subctxt) | \
159 HFI1_MMAP_TOKEN_SET(OFFSET, (offset_in_page(addr))))
161 #define dbg(fmt, ...) \
162 pr_info(fmt, ##__VA_ARGS__)
164 static inline int is_valid_mmap(u64 token)
166 return (HFI1_MMAP_TOKEN_GET(MAGIC, token) == HFI1_MMAP_MAGIC);
169 static int hfi1_file_open(struct inode *inode, struct file *fp)
171 struct hfi1_devdata *dd = container_of(inode->i_cdev,
175 /* Just take a ref now. Not all opens result in a context assign */
176 kobject_get(&dd->kobj);
178 /* The real work is performed later in assign_ctxt() */
179 fp->private_data = kzalloc(sizeof(struct hfi1_filedata), GFP_KERNEL);
180 if (fp->private_data) /* no cpu affinity by default */
181 ((struct hfi1_filedata *)fp->private_data)->rec_cpu_num = -1;
182 return fp->private_data ? 0 : -ENOMEM;
185 static long hfi1_file_ioctl(struct file *fp, unsigned int cmd,
188 struct hfi1_filedata *fd = fp->private_data;
189 struct hfi1_ctxtdata *uctxt = fd->uctxt;
190 struct hfi1_user_info uinfo;
191 struct hfi1_tid_info tinfo;
195 unsigned long ul_uval = 0;
198 hfi1_cdbg(IOCTL, "IOCTL recv: 0x%x", cmd);
199 if (cmd != HFI1_IOCTL_ASSIGN_CTXT &&
200 cmd != HFI1_IOCTL_GET_VERS &&
205 case HFI1_IOCTL_ASSIGN_CTXT:
206 if (copy_from_user(&uinfo,
207 (struct hfi1_user_info __user *)arg,
211 ret = assign_ctxt(fp, &uinfo);
219 case HFI1_IOCTL_CTXT_INFO:
220 ret = get_ctxt_info(fp, (void __user *)(unsigned long)arg,
221 sizeof(struct hfi1_ctxt_info));
223 case HFI1_IOCTL_USER_INFO:
224 ret = get_base_info(fp, (void __user *)(unsigned long)arg,
225 sizeof(struct hfi1_base_info));
227 case HFI1_IOCTL_CREDIT_UPD:
228 if (uctxt && uctxt->sc)
229 sc_return_credits(uctxt->sc);
232 case HFI1_IOCTL_TID_UPDATE:
233 if (copy_from_user(&tinfo,
234 (struct hfi11_tid_info __user *)arg,
238 ret = hfi1_user_exp_rcv_setup(fp, &tinfo);
241 * Copy the number of tidlist entries we used
242 * and the length of the buffer we registered.
243 * These fields are adjacent in the structure so
244 * we can copy them at the same time.
246 addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
247 if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
248 sizeof(tinfo.tidcnt) +
249 sizeof(tinfo.length)))
254 case HFI1_IOCTL_TID_FREE:
255 if (copy_from_user(&tinfo,
256 (struct hfi11_tid_info __user *)arg,
260 ret = hfi1_user_exp_rcv_clear(fp, &tinfo);
263 addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
264 if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
265 sizeof(tinfo.tidcnt)))
269 case HFI1_IOCTL_TID_INVAL_READ:
270 if (copy_from_user(&tinfo,
271 (struct hfi11_tid_info __user *)arg,
275 ret = hfi1_user_exp_rcv_invalid(fp, &tinfo);
278 addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
279 if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
280 sizeof(tinfo.tidcnt)))
284 case HFI1_IOCTL_RECV_CTRL:
285 ret = get_user(uval, (int __user *)arg);
288 ret = manage_rcvq(uctxt, fd->subctxt, uval);
291 case HFI1_IOCTL_POLL_TYPE:
292 ret = get_user(uval, (int __user *)arg);
295 uctxt->poll_type = (typeof(uctxt->poll_type))uval;
298 case HFI1_IOCTL_ACK_EVENT:
299 ret = get_user(ul_uval, (unsigned long __user *)arg);
302 ret = user_event_ack(uctxt, fd->subctxt, ul_uval);
305 case HFI1_IOCTL_SET_PKEY:
306 ret = get_user(uval16, (u16 __user *)arg);
309 if (HFI1_CAP_IS_USET(PKEY_CHECK))
310 ret = set_ctxt_pkey(uctxt, fd->subctxt, uval16);
315 case HFI1_IOCTL_CTXT_RESET: {
316 struct send_context *sc;
317 struct hfi1_devdata *dd;
319 if (!uctxt || !uctxt->dd || !uctxt->sc)
323 * There is no protection here. User level has to
324 * guarantee that no one will be writing to the send
325 * context while it is being re-initialized.
326 * If user level breaks that guarantee, it will break
327 * it's own context and no one else's.
332 * Wait until the interrupt handler has marked the
333 * context as halted or frozen. Report error if we time
336 wait_event_interruptible_timeout(
337 sc->halt_wait, (sc->flags & SCF_HALTED),
338 msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
339 if (!(sc->flags & SCF_HALTED))
343 * If the send context was halted due to a Freeze,
344 * wait until the device has been "unfrozen" before
345 * resetting the context.
347 if (sc->flags & SCF_FROZEN) {
348 wait_event_interruptible_timeout(
350 !(ACCESS_ONCE(dd->flags) & HFI1_FROZEN),
351 msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
352 if (dd->flags & HFI1_FROZEN)
355 if (dd->flags & HFI1_FORCED_FREEZE)
357 * Don't allow context reset if we are into
364 hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_ENB,
367 ret = sc_restart(sc);
370 sc_return_credits(sc);
374 case HFI1_IOCTL_GET_VERS:
375 uval = HFI1_USER_SWVERSION;
376 if (put_user(uval, (int __user *)arg))
387 static ssize_t hfi1_write_iter(struct kiocb *kiocb, struct iov_iter *from)
389 struct hfi1_filedata *fd = kiocb->ki_filp->private_data;
390 struct hfi1_user_sdma_pkt_q *pq = fd->pq;
391 struct hfi1_user_sdma_comp_q *cq = fd->cq;
392 int ret = 0, done = 0, reqs = 0;
393 unsigned long dim = from->nr_segs;
400 if (!iter_is_iovec(from) || !dim) {
405 hfi1_cdbg(SDMA, "SDMA request from %u:%u (%lu)",
406 fd->uctxt->ctxt, fd->subctxt, dim);
408 if (atomic_read(&pq->n_reqs) == pq->n_max_reqs) {
414 unsigned long count = 0;
416 ret = hfi1_user_sdma_process_request(
417 kiocb->ki_filp, (struct iovec *)(from->iov + done),
426 return ret ? ret : reqs;
429 static int hfi1_file_mmap(struct file *fp, struct vm_area_struct *vma)
431 struct hfi1_filedata *fd = fp->private_data;
432 struct hfi1_ctxtdata *uctxt = fd->uctxt;
433 struct hfi1_devdata *dd;
434 unsigned long flags, pfn;
435 u64 token = vma->vm_pgoff << PAGE_SHIFT,
437 u8 subctxt, mapio = 0, vmf = 0, type;
442 if (!is_valid_mmap(token) || !uctxt ||
443 !(vma->vm_flags & VM_SHARED)) {
448 ctxt = HFI1_MMAP_TOKEN_GET(CTXT, token);
449 subctxt = HFI1_MMAP_TOKEN_GET(SUBCTXT, token);
450 type = HFI1_MMAP_TOKEN_GET(TYPE, token);
451 if (ctxt != uctxt->ctxt || subctxt != fd->subctxt) {
456 flags = vma->vm_flags;
461 memaddr = ((dd->physaddr + TXE_PIO_SEND) +
463 (uctxt->sc->hw_context * BIT(16))) +
464 /* 64K PIO space / ctxt */
465 (type == PIO_BUFS_SOP ?
466 (TXE_PIO_SIZE / 2) : 0); /* sop? */
468 * Map only the amount allocated to the context, not the
469 * entire available context's PIO space.
471 memlen = PAGE_ALIGN(uctxt->sc->credits * PIO_BLOCK_SIZE);
472 flags &= ~VM_MAYREAD;
473 flags |= VM_DONTCOPY | VM_DONTEXPAND;
474 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
478 if (flags & VM_WRITE) {
483 * The credit return location for this context could be on the
484 * second or third page allocated for credit returns (if number
485 * of enabled contexts > 64 and 128 respectively).
487 memaddr = dd->cr_base[uctxt->numa_id].pa +
488 (((u64)uctxt->sc->hw_free -
489 (u64)dd->cr_base[uctxt->numa_id].va) & PAGE_MASK);
491 flags &= ~VM_MAYWRITE;
492 flags |= VM_DONTCOPY | VM_DONTEXPAND;
494 * The driver has already allocated memory for credit
495 * returns and programmed it into the chip. Has that
496 * memory been flagged as non-cached?
498 /* vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); */
502 memaddr = uctxt->rcvhdrq_phys;
503 memlen = uctxt->rcvhdrq_size;
509 * The RcvEgr buffer need to be handled differently
510 * as multiple non-contiguous pages need to be mapped
511 * into the user process.
513 memlen = uctxt->egrbufs.size;
514 if ((vma->vm_end - vma->vm_start) != memlen) {
515 dd_dev_err(dd, "Eager buffer map size invalid (%lu != %lu)\n",
516 (vma->vm_end - vma->vm_start), memlen);
520 if (vma->vm_flags & VM_WRITE) {
524 vma->vm_flags &= ~VM_MAYWRITE;
525 addr = vma->vm_start;
526 for (i = 0 ; i < uctxt->egrbufs.numbufs; i++) {
527 ret = remap_pfn_range(
529 uctxt->egrbufs.buffers[i].phys >> PAGE_SHIFT,
530 uctxt->egrbufs.buffers[i].len,
534 addr += uctxt->egrbufs.buffers[i].len;
541 * Map only the page that contains this context's user
544 memaddr = (unsigned long)
545 (dd->physaddr + RXE_PER_CONTEXT_USER)
546 + (uctxt->ctxt * RXE_PER_CONTEXT_SIZE);
548 * TidFlow table is on the same page as the rest of the
552 flags |= VM_DONTCOPY | VM_DONTEXPAND;
553 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
558 * Use the page where this context's flags are. User level
559 * knows where it's own bitmap is within the page.
561 memaddr = (unsigned long)(dd->events +
562 ((uctxt->ctxt - dd->first_user_ctxt) *
563 HFI1_MAX_SHARED_CTXTS)) & PAGE_MASK;
566 * v3.7 removes VM_RESERVED but the effect is kept by
569 flags |= VM_IO | VM_DONTEXPAND;
573 memaddr = kvirt_to_phys((void *)dd->status);
575 flags |= VM_IO | VM_DONTEXPAND;
578 if (!HFI1_CAP_IS_USET(DMA_RTAIL)) {
580 * If the memory allocation failed, the context alloc
581 * also would have failed, so we would never get here
586 if (flags & VM_WRITE) {
590 memaddr = uctxt->rcvhdrqtailaddr_phys;
592 flags &= ~VM_MAYWRITE;
595 memaddr = (u64)uctxt->subctxt_uregbase;
597 flags |= VM_IO | VM_DONTEXPAND;
600 case SUBCTXT_RCV_HDRQ:
601 memaddr = (u64)uctxt->subctxt_rcvhdr_base;
602 memlen = uctxt->rcvhdrq_size * uctxt->subctxt_cnt;
603 flags |= VM_IO | VM_DONTEXPAND;
607 memaddr = (u64)uctxt->subctxt_rcvegrbuf;
608 memlen = uctxt->egrbufs.size * uctxt->subctxt_cnt;
609 flags |= VM_IO | VM_DONTEXPAND;
610 flags &= ~VM_MAYWRITE;
614 struct hfi1_user_sdma_comp_q *cq = fd->cq;
620 memaddr = (u64)cq->comps;
621 memlen = PAGE_ALIGN(sizeof(*cq->comps) * cq->nentries);
622 flags |= VM_IO | VM_DONTEXPAND;
631 if ((vma->vm_end - vma->vm_start) != memlen) {
632 hfi1_cdbg(PROC, "%u:%u Memory size mismatch %lu:%lu",
633 uctxt->ctxt, fd->subctxt,
634 (vma->vm_end - vma->vm_start), memlen);
639 vma->vm_flags = flags;
641 "%u:%u type:%u io/vf:%d/%d, addr:0x%llx, len:%lu(%lu), flags:0x%lx\n",
642 ctxt, subctxt, type, mapio, vmf, memaddr, memlen,
643 vma->vm_end - vma->vm_start, vma->vm_flags);
644 pfn = (unsigned long)(memaddr >> PAGE_SHIFT);
647 vma->vm_ops = &vm_ops;
650 ret = io_remap_pfn_range(vma, vma->vm_start, pfn, memlen,
653 ret = remap_pfn_range(vma, vma->vm_start, pfn, memlen,
661 * Local (non-chip) user memory is not mapped right away but as it is
662 * accessed by the user-level code.
664 static int vma_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
668 page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
670 return VM_FAULT_SIGBUS;
678 static unsigned int hfi1_poll(struct file *fp, struct poll_table_struct *pt)
680 struct hfi1_ctxtdata *uctxt;
683 uctxt = ((struct hfi1_filedata *)fp->private_data)->uctxt;
686 else if (uctxt->poll_type == HFI1_POLL_TYPE_URGENT)
687 pollflag = poll_urgent(fp, pt);
688 else if (uctxt->poll_type == HFI1_POLL_TYPE_ANYRCV)
689 pollflag = poll_next(fp, pt);
696 static int hfi1_file_close(struct inode *inode, struct file *fp)
698 struct hfi1_filedata *fdata = fp->private_data;
699 struct hfi1_ctxtdata *uctxt = fdata->uctxt;
700 struct hfi1_devdata *dd = container_of(inode->i_cdev,
703 unsigned long flags, *ev;
705 fp->private_data = NULL;
710 hfi1_cdbg(PROC, "freeing ctxt %u:%u", uctxt->ctxt, fdata->subctxt);
711 mutex_lock(&hfi1_mutex);
714 /* drain user sdma queue */
715 hfi1_user_sdma_free_queues(fdata);
717 /* release the cpu */
718 hfi1_put_proc_affinity(dd, fdata->rec_cpu_num);
721 * Clear any left over, unhandled events so the next process that
722 * gets this context doesn't get confused.
724 ev = dd->events + ((uctxt->ctxt - dd->first_user_ctxt) *
725 HFI1_MAX_SHARED_CTXTS) + fdata->subctxt;
729 uctxt->active_slaves &= ~(1 << fdata->subctxt);
730 uctxt->subpid[fdata->subctxt] = 0;
731 mutex_unlock(&hfi1_mutex);
735 spin_lock_irqsave(&dd->uctxt_lock, flags);
737 * Disable receive context and interrupt available, reset all
738 * RcvCtxtCtrl bits to default values.
740 hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_DIS |
741 HFI1_RCVCTRL_TIDFLOW_DIS |
742 HFI1_RCVCTRL_INTRAVAIL_DIS |
743 HFI1_RCVCTRL_TAILUPD_DIS |
744 HFI1_RCVCTRL_ONE_PKT_EGR_DIS |
745 HFI1_RCVCTRL_NO_RHQ_DROP_DIS |
746 HFI1_RCVCTRL_NO_EGR_DROP_DIS, uctxt->ctxt);
747 /* Clear the context's J_KEY */
748 hfi1_clear_ctxt_jkey(dd, uctxt->ctxt);
750 * Reset context integrity checks to default.
751 * (writes to CSRs probably belong in chip.c)
753 write_kctxt_csr(dd, uctxt->sc->hw_context, SEND_CTXT_CHECK_ENABLE,
754 hfi1_pkt_default_send_ctxt_mask(dd, uctxt->sc->type));
755 sc_disable(uctxt->sc);
757 spin_unlock_irqrestore(&dd->uctxt_lock, flags);
759 dd->rcd[uctxt->ctxt] = NULL;
761 hfi1_user_exp_rcv_free(fdata);
762 hfi1_clear_ctxt_pkey(dd, uctxt->ctxt);
764 uctxt->rcvwait_to = 0;
765 uctxt->piowait_to = 0;
766 uctxt->rcvnowait = 0;
767 uctxt->pionowait = 0;
768 uctxt->event_flags = 0;
770 hfi1_stats.sps_ctxts--;
771 if (++dd->freectxts == dd->num_user_contexts)
773 mutex_unlock(&hfi1_mutex);
774 hfi1_free_ctxtdata(dd, uctxt);
776 kobject_put(&dd->kobj);
782 * Convert kernel *virtual* addresses to physical addresses.
783 * This is used to vmalloc'ed addresses.
785 static u64 kvirt_to_phys(void *addr)
790 page = vmalloc_to_page(addr);
792 paddr = page_to_pfn(page) << PAGE_SHIFT;
797 static int assign_ctxt(struct file *fp, struct hfi1_user_info *uinfo)
799 int i_minor, ret = 0;
800 unsigned int swmajor, swminor;
802 swmajor = uinfo->userversion >> 16;
803 if (swmajor != HFI1_USER_SWMAJOR) {
808 swminor = uinfo->userversion & 0xffff;
810 mutex_lock(&hfi1_mutex);
811 /* First, lets check if we need to setup a shared context? */
812 if (uinfo->subctxt_cnt) {
813 struct hfi1_filedata *fd = fp->private_data;
815 ret = find_shared_ctxt(fp, uinfo);
819 fd->rec_cpu_num = hfi1_get_proc_affinity(
820 fd->uctxt->dd, fd->uctxt->numa_id);
824 * We execute the following block if we couldn't find a
825 * shared context or if context sharing is not required.
828 i_minor = iminor(file_inode(fp)) - HFI1_USER_MINOR_BASE;
829 ret = get_user_context(fp, uinfo, i_minor);
832 mutex_unlock(&hfi1_mutex);
837 static int get_user_context(struct file *fp, struct hfi1_user_info *uinfo,
840 struct hfi1_devdata *dd = NULL;
841 int devmax, npresent, nup;
843 devmax = hfi1_count_units(&npresent, &nup);
850 dd = hfi1_lookup(devno);
853 else if (!dd->freectxts)
856 return allocate_ctxt(fp, dd, uinfo);
859 static int find_shared_ctxt(struct file *fp,
860 const struct hfi1_user_info *uinfo)
864 struct hfi1_filedata *fd = fp->private_data;
866 devmax = hfi1_count_units(NULL, NULL);
868 for (ndev = 0; ndev < devmax; ndev++) {
869 struct hfi1_devdata *dd = hfi1_lookup(ndev);
871 if (!(dd && (dd->flags & HFI1_PRESENT) && dd->kregbase))
873 for (i = dd->first_user_ctxt; i < dd->num_rcv_contexts; i++) {
874 struct hfi1_ctxtdata *uctxt = dd->rcd[i];
876 /* Skip ctxts which are not yet open */
877 if (!uctxt || !uctxt->cnt)
879 /* Skip ctxt if it doesn't match the requested one */
880 if (memcmp(uctxt->uuid, uinfo->uuid,
881 sizeof(uctxt->uuid)) ||
882 uctxt->jkey != generate_jkey(current_uid()) ||
883 uctxt->subctxt_id != uinfo->subctxt_id ||
884 uctxt->subctxt_cnt != uinfo->subctxt_cnt)
887 /* Verify the sharing process matches the master */
888 if (uctxt->userversion != uinfo->userversion ||
889 uctxt->cnt >= uctxt->subctxt_cnt) {
894 fd->subctxt = uctxt->cnt++;
895 uctxt->subpid[fd->subctxt] = current->pid;
896 uctxt->active_slaves |= 1 << fd->subctxt;
906 static int allocate_ctxt(struct file *fp, struct hfi1_devdata *dd,
907 struct hfi1_user_info *uinfo)
909 struct hfi1_filedata *fd = fp->private_data;
910 struct hfi1_ctxtdata *uctxt;
914 if (dd->flags & HFI1_FROZEN) {
916 * Pick an error that is unique from all other errors
917 * that are returned so the user process knows that
918 * it tried to allocate while the SPC was frozen. It
919 * it should be able to retry with success in a short
925 for (ctxt = dd->first_user_ctxt; ctxt < dd->num_rcv_contexts; ctxt++)
929 if (ctxt == dd->num_rcv_contexts)
932 fd->rec_cpu_num = hfi1_get_proc_affinity(dd, -1);
933 if (fd->rec_cpu_num != -1)
934 numa = cpu_to_node(fd->rec_cpu_num);
936 numa = numa_node_id();
937 uctxt = hfi1_create_ctxtdata(dd->pport, ctxt, numa);
940 "Unable to allocate ctxtdata memory, failing open\n");
943 hfi1_cdbg(PROC, "[%u:%u] pid %u assigned to CPU %d (NUMA %u)",
944 uctxt->ctxt, fd->subctxt, current->pid, fd->rec_cpu_num,
948 * Allocate and enable a PIO send context.
950 uctxt->sc = sc_alloc(dd, SC_USER, uctxt->rcvhdrqentsize,
955 hfi1_cdbg(PROC, "allocated send context %u(%u)\n", uctxt->sc->sw_index,
956 uctxt->sc->hw_context);
957 ret = sc_enable(uctxt->sc);
961 * Setup shared context resources if the user-level has requested
962 * shared contexts and this is the 'master' process.
963 * This has to be done here so the rest of the sub-contexts find the
966 if (uinfo->subctxt_cnt && !fd->subctxt) {
967 ret = init_subctxts(uctxt, uinfo);
969 * On error, we don't need to disable and de-allocate the
970 * send context because it will be done during file close
975 uctxt->userversion = uinfo->userversion;
976 uctxt->pid = current->pid;
977 uctxt->flags = HFI1_CAP_UGET(MASK);
978 init_waitqueue_head(&uctxt->wait);
979 strlcpy(uctxt->comm, current->comm, sizeof(uctxt->comm));
980 memcpy(uctxt->uuid, uinfo->uuid, sizeof(uctxt->uuid));
981 uctxt->jkey = generate_jkey(current_uid());
982 INIT_LIST_HEAD(&uctxt->sdma_queues);
983 spin_lock_init(&uctxt->sdma_qlock);
984 hfi1_stats.sps_ctxts++;
986 * Disable ASPM when there are open user/PSM contexts to avoid
987 * issues with ASPM L1 exit latency
989 if (dd->freectxts-- == dd->num_user_contexts)
990 aspm_disable_all(dd);
996 static int init_subctxts(struct hfi1_ctxtdata *uctxt,
997 const struct hfi1_user_info *uinfo)
999 unsigned num_subctxts;
1001 num_subctxts = uinfo->subctxt_cnt;
1002 if (num_subctxts > HFI1_MAX_SHARED_CTXTS)
1005 uctxt->subctxt_cnt = uinfo->subctxt_cnt;
1006 uctxt->subctxt_id = uinfo->subctxt_id;
1007 uctxt->active_slaves = 1;
1008 uctxt->redirect_seq_cnt = 1;
1009 set_bit(HFI1_CTXT_MASTER_UNINIT, &uctxt->event_flags);
1014 static int setup_subctxt(struct hfi1_ctxtdata *uctxt)
1017 unsigned num_subctxts = uctxt->subctxt_cnt;
1019 uctxt->subctxt_uregbase = vmalloc_user(PAGE_SIZE);
1020 if (!uctxt->subctxt_uregbase) {
1024 /* We can take the size of the RcvHdr Queue from the master */
1025 uctxt->subctxt_rcvhdr_base = vmalloc_user(uctxt->rcvhdrq_size *
1027 if (!uctxt->subctxt_rcvhdr_base) {
1032 uctxt->subctxt_rcvegrbuf = vmalloc_user(uctxt->egrbufs.size *
1034 if (!uctxt->subctxt_rcvegrbuf) {
1040 vfree(uctxt->subctxt_rcvhdr_base);
1042 vfree(uctxt->subctxt_uregbase);
1043 uctxt->subctxt_uregbase = NULL;
1048 static int user_init(struct file *fp)
1050 unsigned int rcvctrl_ops = 0;
1051 struct hfi1_filedata *fd = fp->private_data;
1052 struct hfi1_ctxtdata *uctxt = fd->uctxt;
1054 /* make sure that the context has already been setup */
1055 if (!test_bit(HFI1_CTXT_SETUP_DONE, &uctxt->event_flags))
1058 /* initialize poll variables... */
1060 uctxt->urgent_poll = 0;
1063 * Now enable the ctxt for receive.
1064 * For chips that are set to DMA the tail register to memory
1065 * when they change (and when the update bit transitions from
1066 * 0 to 1. So for those chips, we turn it off and then back on.
1067 * This will (very briefly) affect any other open ctxts, but the
1068 * duration is very short, and therefore isn't an issue. We
1069 * explicitly set the in-memory tail copy to 0 beforehand, so we
1070 * don't have to wait to be sure the DMA update has happened
1071 * (chip resets head/tail to 0 on transition to enable).
1073 if (uctxt->rcvhdrtail_kvaddr)
1074 clear_rcvhdrtail(uctxt);
1076 /* Setup J_KEY before enabling the context */
1077 hfi1_set_ctxt_jkey(uctxt->dd, uctxt->ctxt, uctxt->jkey);
1079 rcvctrl_ops = HFI1_RCVCTRL_CTXT_ENB;
1080 if (HFI1_CAP_KGET_MASK(uctxt->flags, HDRSUPP))
1081 rcvctrl_ops |= HFI1_RCVCTRL_TIDFLOW_ENB;
1083 * Ignore the bit in the flags for now until proper
1084 * support for multiple packet per rcv array entry is
1087 if (!HFI1_CAP_KGET_MASK(uctxt->flags, MULTI_PKT_EGR))
1088 rcvctrl_ops |= HFI1_RCVCTRL_ONE_PKT_EGR_ENB;
1089 if (HFI1_CAP_KGET_MASK(uctxt->flags, NODROP_EGR_FULL))
1090 rcvctrl_ops |= HFI1_RCVCTRL_NO_EGR_DROP_ENB;
1091 if (HFI1_CAP_KGET_MASK(uctxt->flags, NODROP_RHQ_FULL))
1092 rcvctrl_ops |= HFI1_RCVCTRL_NO_RHQ_DROP_ENB;
1094 * The RcvCtxtCtrl.TailUpd bit has to be explicitly written.
1095 * We can't rely on the correct value to be set from prior
1096 * uses of the chip or ctxt. Therefore, add the rcvctrl op
1099 if (HFI1_CAP_KGET_MASK(uctxt->flags, DMA_RTAIL))
1100 rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_ENB;
1102 rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_DIS;
1103 hfi1_rcvctrl(uctxt->dd, rcvctrl_ops, uctxt->ctxt);
1105 /* Notify any waiting slaves */
1106 if (uctxt->subctxt_cnt) {
1107 clear_bit(HFI1_CTXT_MASTER_UNINIT, &uctxt->event_flags);
1108 wake_up(&uctxt->wait);
1114 static int get_ctxt_info(struct file *fp, void __user *ubase, __u32 len)
1116 struct hfi1_ctxt_info cinfo;
1117 struct hfi1_filedata *fd = fp->private_data;
1118 struct hfi1_ctxtdata *uctxt = fd->uctxt;
1121 memset(&cinfo, 0, sizeof(cinfo));
1122 ret = hfi1_get_base_kinfo(uctxt, &cinfo);
1125 cinfo.num_active = hfi1_count_active_units();
1126 cinfo.unit = uctxt->dd->unit;
1127 cinfo.ctxt = uctxt->ctxt;
1128 cinfo.subctxt = fd->subctxt;
1129 cinfo.rcvtids = roundup(uctxt->egrbufs.alloced,
1130 uctxt->dd->rcv_entries.group_size) +
1131 uctxt->expected_count;
1132 cinfo.credits = uctxt->sc->credits;
1133 cinfo.numa_node = uctxt->numa_id;
1134 cinfo.rec_cpu = fd->rec_cpu_num;
1135 cinfo.send_ctxt = uctxt->sc->hw_context;
1137 cinfo.egrtids = uctxt->egrbufs.alloced;
1138 cinfo.rcvhdrq_cnt = uctxt->rcvhdrq_cnt;
1139 cinfo.rcvhdrq_entsize = uctxt->rcvhdrqentsize << 2;
1140 cinfo.sdma_ring_size = fd->cq->nentries;
1141 cinfo.rcvegr_size = uctxt->egrbufs.rcvtid_size;
1143 trace_hfi1_ctxt_info(uctxt->dd, uctxt->ctxt, fd->subctxt, cinfo);
1144 if (copy_to_user(ubase, &cinfo, sizeof(cinfo)))
1150 static int setup_ctxt(struct file *fp)
1152 struct hfi1_filedata *fd = fp->private_data;
1153 struct hfi1_ctxtdata *uctxt = fd->uctxt;
1154 struct hfi1_devdata *dd = uctxt->dd;
1158 * Context should be set up only once, including allocation and
1159 * programming of eager buffers. This is done if context sharing
1160 * is not requested or by the master process.
1162 if (!uctxt->subctxt_cnt || !fd->subctxt) {
1163 ret = hfi1_init_ctxt(uctxt->sc);
1167 /* Now allocate the RcvHdr queue and eager buffers. */
1168 ret = hfi1_create_rcvhdrq(dd, uctxt);
1171 ret = hfi1_setup_eagerbufs(uctxt);
1174 if (uctxt->subctxt_cnt && !fd->subctxt) {
1175 ret = setup_subctxt(uctxt);
1180 ret = wait_event_interruptible(uctxt->wait, !test_bit(
1181 HFI1_CTXT_MASTER_UNINIT,
1182 &uctxt->event_flags));
1187 ret = hfi1_user_sdma_alloc_queues(uctxt, fp);
1191 * Expected receive has to be setup for all processes (including
1192 * shared contexts). However, it has to be done after the master
1193 * context has been fully configured as it depends on the
1194 * eager/expected split of the RcvArray entries.
1195 * Setting it up here ensures that the subcontexts will be waiting
1196 * (due to the above wait_event_interruptible() until the master
1199 ret = hfi1_user_exp_rcv_init(fp);
1203 set_bit(HFI1_CTXT_SETUP_DONE, &uctxt->event_flags);
1208 static int get_base_info(struct file *fp, void __user *ubase, __u32 len)
1210 struct hfi1_base_info binfo;
1211 struct hfi1_filedata *fd = fp->private_data;
1212 struct hfi1_ctxtdata *uctxt = fd->uctxt;
1213 struct hfi1_devdata *dd = uctxt->dd;
1218 trace_hfi1_uctxtdata(uctxt->dd, uctxt);
1220 memset(&binfo, 0, sizeof(binfo));
1221 binfo.hw_version = dd->revision;
1222 binfo.sw_version = HFI1_KERN_SWVERSION;
1223 binfo.bthqp = kdeth_qp;
1224 binfo.jkey = uctxt->jkey;
1226 * If more than 64 contexts are enabled the allocated credit
1227 * return will span two or three contiguous pages. Since we only
1228 * map the page containing the context's credit return address,
1229 * we need to calculate the offset in the proper page.
1231 offset = ((u64)uctxt->sc->hw_free -
1232 (u64)dd->cr_base[uctxt->numa_id].va) % PAGE_SIZE;
1233 binfo.sc_credits_addr = HFI1_MMAP_TOKEN(PIO_CRED, uctxt->ctxt,
1234 fd->subctxt, offset);
1235 binfo.pio_bufbase = HFI1_MMAP_TOKEN(PIO_BUFS, uctxt->ctxt,
1237 uctxt->sc->base_addr);
1238 binfo.pio_bufbase_sop = HFI1_MMAP_TOKEN(PIO_BUFS_SOP,
1241 uctxt->sc->base_addr);
1242 binfo.rcvhdr_bufbase = HFI1_MMAP_TOKEN(RCV_HDRQ, uctxt->ctxt,
1245 binfo.rcvegr_bufbase = HFI1_MMAP_TOKEN(RCV_EGRBUF, uctxt->ctxt,
1247 uctxt->egrbufs.rcvtids[0].phys);
1248 binfo.sdma_comp_bufbase = HFI1_MMAP_TOKEN(SDMA_COMP, uctxt->ctxt,
1252 * (RXE_PER_CONTEXT_USER + (ctxt * RXE_PER_CONTEXT_SIZE))
1254 binfo.user_regbase = HFI1_MMAP_TOKEN(UREGS, uctxt->ctxt,
1256 offset = offset_in_page((((uctxt->ctxt - dd->first_user_ctxt) *
1257 HFI1_MAX_SHARED_CTXTS) + fd->subctxt) *
1258 sizeof(*dd->events));
1259 binfo.events_bufbase = HFI1_MMAP_TOKEN(EVENTS, uctxt->ctxt,
1262 binfo.status_bufbase = HFI1_MMAP_TOKEN(STATUS, uctxt->ctxt,
1265 if (HFI1_CAP_IS_USET(DMA_RTAIL))
1266 binfo.rcvhdrtail_base = HFI1_MMAP_TOKEN(RTAIL, uctxt->ctxt,
1268 if (uctxt->subctxt_cnt) {
1269 binfo.subctxt_uregbase = HFI1_MMAP_TOKEN(SUBCTXT_UREGS,
1272 binfo.subctxt_rcvhdrbuf = HFI1_MMAP_TOKEN(SUBCTXT_RCV_HDRQ,
1275 binfo.subctxt_rcvegrbuf = HFI1_MMAP_TOKEN(SUBCTXT_EGRBUF,
1279 sz = (len < sizeof(binfo)) ? len : sizeof(binfo);
1280 if (copy_to_user(ubase, &binfo, sz))
1285 static unsigned int poll_urgent(struct file *fp,
1286 struct poll_table_struct *pt)
1288 struct hfi1_filedata *fd = fp->private_data;
1289 struct hfi1_ctxtdata *uctxt = fd->uctxt;
1290 struct hfi1_devdata *dd = uctxt->dd;
1293 poll_wait(fp, &uctxt->wait, pt);
1295 spin_lock_irq(&dd->uctxt_lock);
1296 if (uctxt->urgent != uctxt->urgent_poll) {
1297 pollflag = POLLIN | POLLRDNORM;
1298 uctxt->urgent_poll = uctxt->urgent;
1301 set_bit(HFI1_CTXT_WAITING_URG, &uctxt->event_flags);
1303 spin_unlock_irq(&dd->uctxt_lock);
1308 static unsigned int poll_next(struct file *fp,
1309 struct poll_table_struct *pt)
1311 struct hfi1_filedata *fd = fp->private_data;
1312 struct hfi1_ctxtdata *uctxt = fd->uctxt;
1313 struct hfi1_devdata *dd = uctxt->dd;
1316 poll_wait(fp, &uctxt->wait, pt);
1318 spin_lock_irq(&dd->uctxt_lock);
1319 if (hdrqempty(uctxt)) {
1320 set_bit(HFI1_CTXT_WAITING_RCV, &uctxt->event_flags);
1321 hfi1_rcvctrl(dd, HFI1_RCVCTRL_INTRAVAIL_ENB, uctxt->ctxt);
1324 pollflag = POLLIN | POLLRDNORM;
1326 spin_unlock_irq(&dd->uctxt_lock);
1332 * Find all user contexts in use, and set the specified bit in their
1334 * See also find_ctxt() for a similar use, that is specific to send buffers.
1336 int hfi1_set_uevent_bits(struct hfi1_pportdata *ppd, const int evtbit)
1338 struct hfi1_ctxtdata *uctxt;
1339 struct hfi1_devdata *dd = ppd->dd;
1342 unsigned long flags;
1349 spin_lock_irqsave(&dd->uctxt_lock, flags);
1350 for (ctxt = dd->first_user_ctxt; ctxt < dd->num_rcv_contexts;
1352 uctxt = dd->rcd[ctxt];
1354 unsigned long *evs = dd->events +
1355 (uctxt->ctxt - dd->first_user_ctxt) *
1356 HFI1_MAX_SHARED_CTXTS;
1359 * subctxt_cnt is 0 if not shared, so do base
1360 * separately, first, then remaining subctxt, if any
1362 set_bit(evtbit, evs);
1363 for (i = 1; i < uctxt->subctxt_cnt; i++)
1364 set_bit(evtbit, evs + i);
1367 spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1373 * manage_rcvq - manage a context's receive queue
1374 * @uctxt: the context
1375 * @subctxt: the sub-context
1376 * @start_stop: action to carry out
1378 * start_stop == 0 disables receive on the context, for use in queue
1379 * overflow conditions. start_stop==1 re-enables, to be used to
1380 * re-init the software copy of the head register
1382 static int manage_rcvq(struct hfi1_ctxtdata *uctxt, unsigned subctxt,
1385 struct hfi1_devdata *dd = uctxt->dd;
1386 unsigned int rcvctrl_op;
1390 /* atomically clear receive enable ctxt. */
1393 * On enable, force in-memory copy of the tail register to
1394 * 0, so that protocol code doesn't have to worry about
1395 * whether or not the chip has yet updated the in-memory
1396 * copy or not on return from the system call. The chip
1397 * always resets it's tail register back to 0 on a
1398 * transition from disabled to enabled.
1400 if (uctxt->rcvhdrtail_kvaddr)
1401 clear_rcvhdrtail(uctxt);
1402 rcvctrl_op = HFI1_RCVCTRL_CTXT_ENB;
1404 rcvctrl_op = HFI1_RCVCTRL_CTXT_DIS;
1406 hfi1_rcvctrl(dd, rcvctrl_op, uctxt->ctxt);
1407 /* always; new head should be equal to new tail; see above */
1413 * clear the event notifier events for this context.
1414 * User process then performs actions appropriate to bit having been
1415 * set, if desired, and checks again in future.
1417 static int user_event_ack(struct hfi1_ctxtdata *uctxt, int subctxt,
1418 unsigned long events)
1421 struct hfi1_devdata *dd = uctxt->dd;
1427 evs = dd->events + ((uctxt->ctxt - dd->first_user_ctxt) *
1428 HFI1_MAX_SHARED_CTXTS) + subctxt;
1430 for (i = 0; i <= _HFI1_MAX_EVENT_BIT; i++) {
1431 if (!test_bit(i, &events))
1438 static int set_ctxt_pkey(struct hfi1_ctxtdata *uctxt, unsigned subctxt,
1441 int ret = -ENOENT, i, intable = 0;
1442 struct hfi1_pportdata *ppd = uctxt->ppd;
1443 struct hfi1_devdata *dd = uctxt->dd;
1445 if (pkey == LIM_MGMT_P_KEY || pkey == FULL_MGMT_P_KEY) {
1450 for (i = 0; i < ARRAY_SIZE(ppd->pkeys); i++)
1451 if (pkey == ppd->pkeys[i]) {
1457 ret = hfi1_set_ctxt_pkey(dd, uctxt->ctxt, pkey);
1462 static void user_remove(struct hfi1_devdata *dd)
1465 hfi1_cdev_cleanup(&dd->user_cdev, &dd->user_device);
1468 static int user_add(struct hfi1_devdata *dd)
1473 snprintf(name, sizeof(name), "%s_%d", class_name(), dd->unit);
1474 ret = hfi1_cdev_init(dd->unit, name, &hfi1_file_ops,
1475 &dd->user_cdev, &dd->user_device,
1484 * Create per-unit files in /dev
1486 int hfi1_device_create(struct hfi1_devdata *dd)
1488 return user_add(dd);
1492 * Remove per-unit files in /dev
1493 * void, core kernel returns no errors for this stuff
1495 void hfi1_device_remove(struct hfi1_devdata *dd)