Back to home page

OSCL-LXR
 
 

     

0001 // SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause
0002 /*
0003  * Copyright(c) 2020 Cornelis Networks, Inc.
0004  * Copyright(c) 2015-2020 Intel Corporation.
0005  */
0006 
0007 #include <linux/poll.h>
0008 #include <linux/cdev.h>
0009 #include <linux/vmalloc.h>
0010 #include <linux/io.h>
0011 #include <linux/sched/mm.h>
0012 #include <linux/bitmap.h>
0013 
0014 #include <rdma/ib.h>
0015 
0016 #include "hfi.h"
0017 #include "pio.h"
0018 #include "device.h"
0019 #include "common.h"
0020 #include "trace.h"
0021 #include "mmu_rb.h"
0022 #include "user_sdma.h"
0023 #include "user_exp_rcv.h"
0024 #include "aspm.h"
0025 
0026 #undef pr_fmt
0027 #define pr_fmt(fmt) DRIVER_NAME ": " fmt
0028 
0029 #define SEND_CTXT_HALT_TIMEOUT 1000 /* msecs */
0030 
0031 /*
0032  * File operation functions
0033  */
0034 static int hfi1_file_open(struct inode *inode, struct file *fp);
0035 static int hfi1_file_close(struct inode *inode, struct file *fp);
0036 static ssize_t hfi1_write_iter(struct kiocb *kiocb, struct iov_iter *from);
0037 static __poll_t hfi1_poll(struct file *fp, struct poll_table_struct *pt);
0038 static int hfi1_file_mmap(struct file *fp, struct vm_area_struct *vma);
0039 
0040 static u64 kvirt_to_phys(void *addr);
0041 static int assign_ctxt(struct hfi1_filedata *fd, unsigned long arg, u32 len);
0042 static void init_subctxts(struct hfi1_ctxtdata *uctxt,
0043               const struct hfi1_user_info *uinfo);
0044 static int init_user_ctxt(struct hfi1_filedata *fd,
0045               struct hfi1_ctxtdata *uctxt);
0046 static void user_init(struct hfi1_ctxtdata *uctxt);
0047 static int get_ctxt_info(struct hfi1_filedata *fd, unsigned long arg, u32 len);
0048 static int get_base_info(struct hfi1_filedata *fd, unsigned long arg, u32 len);
0049 static int user_exp_rcv_setup(struct hfi1_filedata *fd, unsigned long arg,
0050                   u32 len);
0051 static int user_exp_rcv_clear(struct hfi1_filedata *fd, unsigned long arg,
0052                   u32 len);
0053 static int user_exp_rcv_invalid(struct hfi1_filedata *fd, unsigned long arg,
0054                 u32 len);
0055 static int setup_base_ctxt(struct hfi1_filedata *fd,
0056                struct hfi1_ctxtdata *uctxt);
0057 static int setup_subctxt(struct hfi1_ctxtdata *uctxt);
0058 
0059 static int find_sub_ctxt(struct hfi1_filedata *fd,
0060              const struct hfi1_user_info *uinfo);
0061 static int allocate_ctxt(struct hfi1_filedata *fd, struct hfi1_devdata *dd,
0062              struct hfi1_user_info *uinfo,
0063              struct hfi1_ctxtdata **cd);
0064 static void deallocate_ctxt(struct hfi1_ctxtdata *uctxt);
0065 static __poll_t poll_urgent(struct file *fp, struct poll_table_struct *pt);
0066 static __poll_t poll_next(struct file *fp, struct poll_table_struct *pt);
0067 static int user_event_ack(struct hfi1_ctxtdata *uctxt, u16 subctxt,
0068               unsigned long arg);
0069 static int set_ctxt_pkey(struct hfi1_ctxtdata *uctxt, unsigned long arg);
0070 static int ctxt_reset(struct hfi1_ctxtdata *uctxt);
0071 static int manage_rcvq(struct hfi1_ctxtdata *uctxt, u16 subctxt,
0072                unsigned long arg);
0073 static vm_fault_t vma_fault(struct vm_fault *vmf);
0074 static long hfi1_file_ioctl(struct file *fp, unsigned int cmd,
0075                 unsigned long arg);
0076 
0077 static const struct file_operations hfi1_file_ops = {
0078     .owner = THIS_MODULE,
0079     .write_iter = hfi1_write_iter,
0080     .open = hfi1_file_open,
0081     .release = hfi1_file_close,
0082     .unlocked_ioctl = hfi1_file_ioctl,
0083     .poll = hfi1_poll,
0084     .mmap = hfi1_file_mmap,
0085     .llseek = noop_llseek,
0086 };
0087 
0088 static const struct vm_operations_struct vm_ops = {
0089     .fault = vma_fault,
0090 };
0091 
0092 /*
0093  * Types of memories mapped into user processes' space
0094  */
0095 enum mmap_types {
0096     PIO_BUFS = 1,
0097     PIO_BUFS_SOP,
0098     PIO_CRED,
0099     RCV_HDRQ,
0100     RCV_EGRBUF,
0101     UREGS,
0102     EVENTS,
0103     STATUS,
0104     RTAIL,
0105     SUBCTXT_UREGS,
0106     SUBCTXT_RCV_HDRQ,
0107     SUBCTXT_EGRBUF,
0108     SDMA_COMP
0109 };
0110 
0111 /*
0112  * Masks and offsets defining the mmap tokens
0113  */
0114 #define HFI1_MMAP_OFFSET_MASK   0xfffULL
0115 #define HFI1_MMAP_OFFSET_SHIFT  0
0116 #define HFI1_MMAP_SUBCTXT_MASK  0xfULL
0117 #define HFI1_MMAP_SUBCTXT_SHIFT 12
0118 #define HFI1_MMAP_CTXT_MASK     0xffULL
0119 #define HFI1_MMAP_CTXT_SHIFT    16
0120 #define HFI1_MMAP_TYPE_MASK     0xfULL
0121 #define HFI1_MMAP_TYPE_SHIFT    24
0122 #define HFI1_MMAP_MAGIC_MASK    0xffffffffULL
0123 #define HFI1_MMAP_MAGIC_SHIFT   32
0124 
0125 #define HFI1_MMAP_MAGIC         0xdabbad00
0126 
0127 #define HFI1_MMAP_TOKEN_SET(field, val) \
0128     (((val) & HFI1_MMAP_##field##_MASK) << HFI1_MMAP_##field##_SHIFT)
0129 #define HFI1_MMAP_TOKEN_GET(field, token) \
0130     (((token) >> HFI1_MMAP_##field##_SHIFT) & HFI1_MMAP_##field##_MASK)
0131 #define HFI1_MMAP_TOKEN(type, ctxt, subctxt, addr)   \
0132     (HFI1_MMAP_TOKEN_SET(MAGIC, HFI1_MMAP_MAGIC) | \
0133     HFI1_MMAP_TOKEN_SET(TYPE, type) | \
0134     HFI1_MMAP_TOKEN_SET(CTXT, ctxt) | \
0135     HFI1_MMAP_TOKEN_SET(SUBCTXT, subctxt) | \
0136     HFI1_MMAP_TOKEN_SET(OFFSET, (offset_in_page(addr))))
0137 
0138 #define dbg(fmt, ...)               \
0139     pr_info(fmt, ##__VA_ARGS__)
0140 
0141 static inline int is_valid_mmap(u64 token)
0142 {
0143     return (HFI1_MMAP_TOKEN_GET(MAGIC, token) == HFI1_MMAP_MAGIC);
0144 }
0145 
0146 static int hfi1_file_open(struct inode *inode, struct file *fp)
0147 {
0148     struct hfi1_filedata *fd;
0149     struct hfi1_devdata *dd = container_of(inode->i_cdev,
0150                            struct hfi1_devdata,
0151                            user_cdev);
0152 
0153     if (!((dd->flags & HFI1_PRESENT) && dd->kregbase1))
0154         return -EINVAL;
0155 
0156     if (!refcount_inc_not_zero(&dd->user_refcount))
0157         return -ENXIO;
0158 
0159     /* The real work is performed later in assign_ctxt() */
0160 
0161     fd = kzalloc(sizeof(*fd), GFP_KERNEL);
0162 
0163     if (!fd || init_srcu_struct(&fd->pq_srcu))
0164         goto nomem;
0165     spin_lock_init(&fd->pq_rcu_lock);
0166     spin_lock_init(&fd->tid_lock);
0167     spin_lock_init(&fd->invalid_lock);
0168     fd->rec_cpu_num = -1; /* no cpu affinity by default */
0169     fd->dd = dd;
0170     fp->private_data = fd;
0171     return 0;
0172 nomem:
0173     kfree(fd);
0174     fp->private_data = NULL;
0175     if (refcount_dec_and_test(&dd->user_refcount))
0176         complete(&dd->user_comp);
0177     return -ENOMEM;
0178 }
0179 
0180 static long hfi1_file_ioctl(struct file *fp, unsigned int cmd,
0181                 unsigned long arg)
0182 {
0183     struct hfi1_filedata *fd = fp->private_data;
0184     struct hfi1_ctxtdata *uctxt = fd->uctxt;
0185     int ret = 0;
0186     int uval = 0;
0187 
0188     hfi1_cdbg(IOCTL, "IOCTL recv: 0x%x", cmd);
0189     if (cmd != HFI1_IOCTL_ASSIGN_CTXT &&
0190         cmd != HFI1_IOCTL_GET_VERS &&
0191         !uctxt)
0192         return -EINVAL;
0193 
0194     switch (cmd) {
0195     case HFI1_IOCTL_ASSIGN_CTXT:
0196         ret = assign_ctxt(fd, arg, _IOC_SIZE(cmd));
0197         break;
0198 
0199     case HFI1_IOCTL_CTXT_INFO:
0200         ret = get_ctxt_info(fd, arg, _IOC_SIZE(cmd));
0201         break;
0202 
0203     case HFI1_IOCTL_USER_INFO:
0204         ret = get_base_info(fd, arg, _IOC_SIZE(cmd));
0205         break;
0206 
0207     case HFI1_IOCTL_CREDIT_UPD:
0208         if (uctxt)
0209             sc_return_credits(uctxt->sc);
0210         break;
0211 
0212     case HFI1_IOCTL_TID_UPDATE:
0213         ret = user_exp_rcv_setup(fd, arg, _IOC_SIZE(cmd));
0214         break;
0215 
0216     case HFI1_IOCTL_TID_FREE:
0217         ret = user_exp_rcv_clear(fd, arg, _IOC_SIZE(cmd));
0218         break;
0219 
0220     case HFI1_IOCTL_TID_INVAL_READ:
0221         ret = user_exp_rcv_invalid(fd, arg, _IOC_SIZE(cmd));
0222         break;
0223 
0224     case HFI1_IOCTL_RECV_CTRL:
0225         ret = manage_rcvq(uctxt, fd->subctxt, arg);
0226         break;
0227 
0228     case HFI1_IOCTL_POLL_TYPE:
0229         if (get_user(uval, (int __user *)arg))
0230             return -EFAULT;
0231         uctxt->poll_type = (typeof(uctxt->poll_type))uval;
0232         break;
0233 
0234     case HFI1_IOCTL_ACK_EVENT:
0235         ret = user_event_ack(uctxt, fd->subctxt, arg);
0236         break;
0237 
0238     case HFI1_IOCTL_SET_PKEY:
0239         ret = set_ctxt_pkey(uctxt, arg);
0240         break;
0241 
0242     case HFI1_IOCTL_CTXT_RESET:
0243         ret = ctxt_reset(uctxt);
0244         break;
0245 
0246     case HFI1_IOCTL_GET_VERS:
0247         uval = HFI1_USER_SWVERSION;
0248         if (put_user(uval, (int __user *)arg))
0249             return -EFAULT;
0250         break;
0251 
0252     default:
0253         return -EINVAL;
0254     }
0255 
0256     return ret;
0257 }
0258 
0259 static ssize_t hfi1_write_iter(struct kiocb *kiocb, struct iov_iter *from)
0260 {
0261     struct hfi1_filedata *fd = kiocb->ki_filp->private_data;
0262     struct hfi1_user_sdma_pkt_q *pq;
0263     struct hfi1_user_sdma_comp_q *cq = fd->cq;
0264     int done = 0, reqs = 0;
0265     unsigned long dim = from->nr_segs;
0266     int idx;
0267 
0268     if (!HFI1_CAP_IS_KSET(SDMA))
0269         return -EINVAL;
0270     idx = srcu_read_lock(&fd->pq_srcu);
0271     pq = srcu_dereference(fd->pq, &fd->pq_srcu);
0272     if (!cq || !pq) {
0273         srcu_read_unlock(&fd->pq_srcu, idx);
0274         return -EIO;
0275     }
0276 
0277     if (!iter_is_iovec(from) || !dim) {
0278         srcu_read_unlock(&fd->pq_srcu, idx);
0279         return -EINVAL;
0280     }
0281 
0282     trace_hfi1_sdma_request(fd->dd, fd->uctxt->ctxt, fd->subctxt, dim);
0283 
0284     if (atomic_read(&pq->n_reqs) == pq->n_max_reqs) {
0285         srcu_read_unlock(&fd->pq_srcu, idx);
0286         return -ENOSPC;
0287     }
0288 
0289     while (dim) {
0290         int ret;
0291         unsigned long count = 0;
0292 
0293         ret = hfi1_user_sdma_process_request(
0294             fd, (struct iovec *)(from->iov + done),
0295             dim, &count);
0296         if (ret) {
0297             reqs = ret;
0298             break;
0299         }
0300         dim -= count;
0301         done += count;
0302         reqs++;
0303     }
0304 
0305     srcu_read_unlock(&fd->pq_srcu, idx);
0306     return reqs;
0307 }
0308 
0309 static int hfi1_file_mmap(struct file *fp, struct vm_area_struct *vma)
0310 {
0311     struct hfi1_filedata *fd = fp->private_data;
0312     struct hfi1_ctxtdata *uctxt = fd->uctxt;
0313     struct hfi1_devdata *dd;
0314     unsigned long flags;
0315     u64 token = vma->vm_pgoff << PAGE_SHIFT,
0316         memaddr = 0;
0317     void *memvirt = NULL;
0318     u8 subctxt, mapio = 0, vmf = 0, type;
0319     ssize_t memlen = 0;
0320     int ret = 0;
0321     u16 ctxt;
0322 
0323     if (!is_valid_mmap(token) || !uctxt ||
0324         !(vma->vm_flags & VM_SHARED)) {
0325         ret = -EINVAL;
0326         goto done;
0327     }
0328     dd = uctxt->dd;
0329     ctxt = HFI1_MMAP_TOKEN_GET(CTXT, token);
0330     subctxt = HFI1_MMAP_TOKEN_GET(SUBCTXT, token);
0331     type = HFI1_MMAP_TOKEN_GET(TYPE, token);
0332     if (ctxt != uctxt->ctxt || subctxt != fd->subctxt) {
0333         ret = -EINVAL;
0334         goto done;
0335     }
0336 
0337     flags = vma->vm_flags;
0338 
0339     switch (type) {
0340     case PIO_BUFS:
0341     case PIO_BUFS_SOP:
0342         memaddr = ((dd->physaddr + TXE_PIO_SEND) +
0343                 /* chip pio base */
0344                (uctxt->sc->hw_context * BIT(16))) +
0345                 /* 64K PIO space / ctxt */
0346             (type == PIO_BUFS_SOP ?
0347                 (TXE_PIO_SIZE / 2) : 0); /* sop? */
0348         /*
0349          * Map only the amount allocated to the context, not the
0350          * entire available context's PIO space.
0351          */
0352         memlen = PAGE_ALIGN(uctxt->sc->credits * PIO_BLOCK_SIZE);
0353         flags &= ~VM_MAYREAD;
0354         flags |= VM_DONTCOPY | VM_DONTEXPAND;
0355         vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
0356         mapio = 1;
0357         break;
0358     case PIO_CRED:
0359         if (flags & VM_WRITE) {
0360             ret = -EPERM;
0361             goto done;
0362         }
0363         /*
0364          * The credit return location for this context could be on the
0365          * second or third page allocated for credit returns (if number
0366          * of enabled contexts > 64 and 128 respectively).
0367          */
0368         memvirt = dd->cr_base[uctxt->numa_id].va;
0369         memaddr = virt_to_phys(memvirt) +
0370             (((u64)uctxt->sc->hw_free -
0371               (u64)dd->cr_base[uctxt->numa_id].va) & PAGE_MASK);
0372         memlen = PAGE_SIZE;
0373         flags &= ~VM_MAYWRITE;
0374         flags |= VM_DONTCOPY | VM_DONTEXPAND;
0375         /*
0376          * The driver has already allocated memory for credit
0377          * returns and programmed it into the chip. Has that
0378          * memory been flagged as non-cached?
0379          */
0380         /* vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); */
0381         mapio = 1;
0382         break;
0383     case RCV_HDRQ:
0384         memlen = rcvhdrq_size(uctxt);
0385         memvirt = uctxt->rcvhdrq;
0386         break;
0387     case RCV_EGRBUF: {
0388         unsigned long addr;
0389         int i;
0390         /*
0391          * The RcvEgr buffer need to be handled differently
0392          * as multiple non-contiguous pages need to be mapped
0393          * into the user process.
0394          */
0395         memlen = uctxt->egrbufs.size;
0396         if ((vma->vm_end - vma->vm_start) != memlen) {
0397             dd_dev_err(dd, "Eager buffer map size invalid (%lu != %lu)\n",
0398                    (vma->vm_end - vma->vm_start), memlen);
0399             ret = -EINVAL;
0400             goto done;
0401         }
0402         if (vma->vm_flags & VM_WRITE) {
0403             ret = -EPERM;
0404             goto done;
0405         }
0406         vma->vm_flags &= ~VM_MAYWRITE;
0407         addr = vma->vm_start;
0408         for (i = 0 ; i < uctxt->egrbufs.numbufs; i++) {
0409             memlen = uctxt->egrbufs.buffers[i].len;
0410             memvirt = uctxt->egrbufs.buffers[i].addr;
0411             ret = remap_pfn_range(
0412                 vma, addr,
0413                 /*
0414                  * virt_to_pfn() does the same, but
0415                  * it's not available on x86_64
0416                  * when CONFIG_MMU is enabled.
0417                  */
0418                 PFN_DOWN(__pa(memvirt)),
0419                 memlen,
0420                 vma->vm_page_prot);
0421             if (ret < 0)
0422                 goto done;
0423             addr += memlen;
0424         }
0425         ret = 0;
0426         goto done;
0427     }
0428     case UREGS:
0429         /*
0430          * Map only the page that contains this context's user
0431          * registers.
0432          */
0433         memaddr = (unsigned long)
0434             (dd->physaddr + RXE_PER_CONTEXT_USER)
0435             + (uctxt->ctxt * RXE_PER_CONTEXT_SIZE);
0436         /*
0437          * TidFlow table is on the same page as the rest of the
0438          * user registers.
0439          */
0440         memlen = PAGE_SIZE;
0441         flags |= VM_DONTCOPY | VM_DONTEXPAND;
0442         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
0443         mapio = 1;
0444         break;
0445     case EVENTS:
0446         /*
0447          * Use the page where this context's flags are. User level
0448          * knows where it's own bitmap is within the page.
0449          */
0450         memaddr = (unsigned long)
0451             (dd->events + uctxt_offset(uctxt)) & PAGE_MASK;
0452         memlen = PAGE_SIZE;
0453         /*
0454          * v3.7 removes VM_RESERVED but the effect is kept by
0455          * using VM_IO.
0456          */
0457         flags |= VM_IO | VM_DONTEXPAND;
0458         vmf = 1;
0459         break;
0460     case STATUS:
0461         if (flags & VM_WRITE) {
0462             ret = -EPERM;
0463             goto done;
0464         }
0465         memaddr = kvirt_to_phys((void *)dd->status);
0466         memlen = PAGE_SIZE;
0467         flags |= VM_IO | VM_DONTEXPAND;
0468         break;
0469     case RTAIL:
0470         if (!HFI1_CAP_IS_USET(DMA_RTAIL)) {
0471             /*
0472              * If the memory allocation failed, the context alloc
0473              * also would have failed, so we would never get here
0474              */
0475             ret = -EINVAL;
0476             goto done;
0477         }
0478         if ((flags & VM_WRITE) || !hfi1_rcvhdrtail_kvaddr(uctxt)) {
0479             ret = -EPERM;
0480             goto done;
0481         }
0482         memlen = PAGE_SIZE;
0483         memvirt = (void *)hfi1_rcvhdrtail_kvaddr(uctxt);
0484         flags &= ~VM_MAYWRITE;
0485         break;
0486     case SUBCTXT_UREGS:
0487         memaddr = (u64)uctxt->subctxt_uregbase;
0488         memlen = PAGE_SIZE;
0489         flags |= VM_IO | VM_DONTEXPAND;
0490         vmf = 1;
0491         break;
0492     case SUBCTXT_RCV_HDRQ:
0493         memaddr = (u64)uctxt->subctxt_rcvhdr_base;
0494         memlen = rcvhdrq_size(uctxt) * uctxt->subctxt_cnt;
0495         flags |= VM_IO | VM_DONTEXPAND;
0496         vmf = 1;
0497         break;
0498     case SUBCTXT_EGRBUF:
0499         memaddr = (u64)uctxt->subctxt_rcvegrbuf;
0500         memlen = uctxt->egrbufs.size * uctxt->subctxt_cnt;
0501         flags |= VM_IO | VM_DONTEXPAND;
0502         flags &= ~VM_MAYWRITE;
0503         vmf = 1;
0504         break;
0505     case SDMA_COMP: {
0506         struct hfi1_user_sdma_comp_q *cq = fd->cq;
0507 
0508         if (!cq) {
0509             ret = -EFAULT;
0510             goto done;
0511         }
0512         memaddr = (u64)cq->comps;
0513         memlen = PAGE_ALIGN(sizeof(*cq->comps) * cq->nentries);
0514         flags |= VM_IO | VM_DONTEXPAND;
0515         vmf = 1;
0516         break;
0517     }
0518     default:
0519         ret = -EINVAL;
0520         break;
0521     }
0522 
0523     if ((vma->vm_end - vma->vm_start) != memlen) {
0524         hfi1_cdbg(PROC, "%u:%u Memory size mismatch %lu:%lu",
0525               uctxt->ctxt, fd->subctxt,
0526               (vma->vm_end - vma->vm_start), memlen);
0527         ret = -EINVAL;
0528         goto done;
0529     }
0530 
0531     vma->vm_flags = flags;
0532     hfi1_cdbg(PROC,
0533           "%u:%u type:%u io/vf:%d/%d, addr:0x%llx, len:%lu(%lu), flags:0x%lx\n",
0534             ctxt, subctxt, type, mapio, vmf, memaddr, memlen,
0535             vma->vm_end - vma->vm_start, vma->vm_flags);
0536     if (vmf) {
0537         vma->vm_pgoff = PFN_DOWN(memaddr);
0538         vma->vm_ops = &vm_ops;
0539         ret = 0;
0540     } else if (mapio) {
0541         ret = io_remap_pfn_range(vma, vma->vm_start,
0542                      PFN_DOWN(memaddr),
0543                      memlen,
0544                      vma->vm_page_prot);
0545     } else if (memvirt) {
0546         ret = remap_pfn_range(vma, vma->vm_start,
0547                       PFN_DOWN(__pa(memvirt)),
0548                       memlen,
0549                       vma->vm_page_prot);
0550     } else {
0551         ret = remap_pfn_range(vma, vma->vm_start,
0552                       PFN_DOWN(memaddr),
0553                       memlen,
0554                       vma->vm_page_prot);
0555     }
0556 done:
0557     return ret;
0558 }
0559 
0560 /*
0561  * Local (non-chip) user memory is not mapped right away but as it is
0562  * accessed by the user-level code.
0563  */
0564 static vm_fault_t vma_fault(struct vm_fault *vmf)
0565 {
0566     struct page *page;
0567 
0568     page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
0569     if (!page)
0570         return VM_FAULT_SIGBUS;
0571 
0572     get_page(page);
0573     vmf->page = page;
0574 
0575     return 0;
0576 }
0577 
0578 static __poll_t hfi1_poll(struct file *fp, struct poll_table_struct *pt)
0579 {
0580     struct hfi1_ctxtdata *uctxt;
0581     __poll_t pollflag;
0582 
0583     uctxt = ((struct hfi1_filedata *)fp->private_data)->uctxt;
0584     if (!uctxt)
0585         pollflag = EPOLLERR;
0586     else if (uctxt->poll_type == HFI1_POLL_TYPE_URGENT)
0587         pollflag = poll_urgent(fp, pt);
0588     else  if (uctxt->poll_type == HFI1_POLL_TYPE_ANYRCV)
0589         pollflag = poll_next(fp, pt);
0590     else /* invalid */
0591         pollflag = EPOLLERR;
0592 
0593     return pollflag;
0594 }
0595 
0596 static int hfi1_file_close(struct inode *inode, struct file *fp)
0597 {
0598     struct hfi1_filedata *fdata = fp->private_data;
0599     struct hfi1_ctxtdata *uctxt = fdata->uctxt;
0600     struct hfi1_devdata *dd = container_of(inode->i_cdev,
0601                            struct hfi1_devdata,
0602                            user_cdev);
0603     unsigned long flags, *ev;
0604 
0605     fp->private_data = NULL;
0606 
0607     if (!uctxt)
0608         goto done;
0609 
0610     hfi1_cdbg(PROC, "closing ctxt %u:%u", uctxt->ctxt, fdata->subctxt);
0611 
0612     flush_wc();
0613     /* drain user sdma queue */
0614     hfi1_user_sdma_free_queues(fdata, uctxt);
0615 
0616     /* release the cpu */
0617     hfi1_put_proc_affinity(fdata->rec_cpu_num);
0618 
0619     /* clean up rcv side */
0620     hfi1_user_exp_rcv_free(fdata);
0621 
0622     /*
0623      * fdata->uctxt is used in the above cleanup.  It is not ready to be
0624      * removed until here.
0625      */
0626     fdata->uctxt = NULL;
0627     hfi1_rcd_put(uctxt);
0628 
0629     /*
0630      * Clear any left over, unhandled events so the next process that
0631      * gets this context doesn't get confused.
0632      */
0633     ev = dd->events + uctxt_offset(uctxt) + fdata->subctxt;
0634     *ev = 0;
0635 
0636     spin_lock_irqsave(&dd->uctxt_lock, flags);
0637     __clear_bit(fdata->subctxt, uctxt->in_use_ctxts);
0638     if (!bitmap_empty(uctxt->in_use_ctxts, HFI1_MAX_SHARED_CTXTS)) {
0639         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
0640         goto done;
0641     }
0642     spin_unlock_irqrestore(&dd->uctxt_lock, flags);
0643 
0644     /*
0645      * Disable receive context and interrupt available, reset all
0646      * RcvCtxtCtrl bits to default values.
0647      */
0648     hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_DIS |
0649              HFI1_RCVCTRL_TIDFLOW_DIS |
0650              HFI1_RCVCTRL_INTRAVAIL_DIS |
0651              HFI1_RCVCTRL_TAILUPD_DIS |
0652              HFI1_RCVCTRL_ONE_PKT_EGR_DIS |
0653              HFI1_RCVCTRL_NO_RHQ_DROP_DIS |
0654              HFI1_RCVCTRL_NO_EGR_DROP_DIS |
0655              HFI1_RCVCTRL_URGENT_DIS, uctxt);
0656     /* Clear the context's J_KEY */
0657     hfi1_clear_ctxt_jkey(dd, uctxt);
0658     /*
0659      * If a send context is allocated, reset context integrity
0660      * checks to default and disable the send context.
0661      */
0662     if (uctxt->sc) {
0663         sc_disable(uctxt->sc);
0664         set_pio_integrity(uctxt->sc);
0665     }
0666 
0667     hfi1_free_ctxt_rcv_groups(uctxt);
0668     hfi1_clear_ctxt_pkey(dd, uctxt);
0669 
0670     uctxt->event_flags = 0;
0671 
0672     deallocate_ctxt(uctxt);
0673 done:
0674 
0675     if (refcount_dec_and_test(&dd->user_refcount))
0676         complete(&dd->user_comp);
0677 
0678     cleanup_srcu_struct(&fdata->pq_srcu);
0679     kfree(fdata);
0680     return 0;
0681 }
0682 
0683 /*
0684  * Convert kernel *virtual* addresses to physical addresses.
0685  * This is used to vmalloc'ed addresses.
0686  */
0687 static u64 kvirt_to_phys(void *addr)
0688 {
0689     struct page *page;
0690     u64 paddr = 0;
0691 
0692     page = vmalloc_to_page(addr);
0693     if (page)
0694         paddr = page_to_pfn(page) << PAGE_SHIFT;
0695 
0696     return paddr;
0697 }
0698 
0699 /**
0700  * complete_subctxt - complete sub-context info
0701  * @fd: valid filedata pointer
0702  *
0703  * Sub-context info can only be set up after the base context
0704  * has been completed.  This is indicated by the clearing of the
0705  * HFI1_CTXT_BASE_UINIT bit.
0706  *
0707  * Wait for the bit to be cleared, and then complete the subcontext
0708  * initialization.
0709  *
0710  */
0711 static int complete_subctxt(struct hfi1_filedata *fd)
0712 {
0713     int ret;
0714     unsigned long flags;
0715 
0716     /*
0717      * sub-context info can only be set up after the base context
0718      * has been completed.
0719      */
0720     ret = wait_event_interruptible(
0721         fd->uctxt->wait,
0722         !test_bit(HFI1_CTXT_BASE_UNINIT, &fd->uctxt->event_flags));
0723 
0724     if (test_bit(HFI1_CTXT_BASE_FAILED, &fd->uctxt->event_flags))
0725         ret = -ENOMEM;
0726 
0727     /* Finish the sub-context init */
0728     if (!ret) {
0729         fd->rec_cpu_num = hfi1_get_proc_affinity(fd->uctxt->numa_id);
0730         ret = init_user_ctxt(fd, fd->uctxt);
0731     }
0732 
0733     if (ret) {
0734         spin_lock_irqsave(&fd->dd->uctxt_lock, flags);
0735         __clear_bit(fd->subctxt, fd->uctxt->in_use_ctxts);
0736         spin_unlock_irqrestore(&fd->dd->uctxt_lock, flags);
0737         hfi1_rcd_put(fd->uctxt);
0738         fd->uctxt = NULL;
0739     }
0740 
0741     return ret;
0742 }
0743 
0744 static int assign_ctxt(struct hfi1_filedata *fd, unsigned long arg, u32 len)
0745 {
0746     int ret;
0747     unsigned int swmajor;
0748     struct hfi1_ctxtdata *uctxt = NULL;
0749     struct hfi1_user_info uinfo;
0750 
0751     if (fd->uctxt)
0752         return -EINVAL;
0753 
0754     if (sizeof(uinfo) != len)
0755         return -EINVAL;
0756 
0757     if (copy_from_user(&uinfo, (void __user *)arg, sizeof(uinfo)))
0758         return -EFAULT;
0759 
0760     swmajor = uinfo.userversion >> 16;
0761     if (swmajor != HFI1_USER_SWMAJOR)
0762         return -ENODEV;
0763 
0764     if (uinfo.subctxt_cnt > HFI1_MAX_SHARED_CTXTS)
0765         return -EINVAL;
0766 
0767     /*
0768      * Acquire the mutex to protect against multiple creations of what
0769      * could be a shared base context.
0770      */
0771     mutex_lock(&hfi1_mutex);
0772     /*
0773      * Get a sub context if available  (fd->uctxt will be set).
0774      * ret < 0 error, 0 no context, 1 sub-context found
0775      */
0776     ret = find_sub_ctxt(fd, &uinfo);
0777 
0778     /*
0779      * Allocate a base context if context sharing is not required or a
0780      * sub context wasn't found.
0781      */
0782     if (!ret)
0783         ret = allocate_ctxt(fd, fd->dd, &uinfo, &uctxt);
0784 
0785     mutex_unlock(&hfi1_mutex);
0786 
0787     /* Depending on the context type, finish the appropriate init */
0788     switch (ret) {
0789     case 0:
0790         ret = setup_base_ctxt(fd, uctxt);
0791         if (ret)
0792             deallocate_ctxt(uctxt);
0793         break;
0794     case 1:
0795         ret = complete_subctxt(fd);
0796         break;
0797     default:
0798         break;
0799     }
0800 
0801     return ret;
0802 }
0803 
0804 /**
0805  * match_ctxt - match context
0806  * @fd: valid filedata pointer
0807  * @uinfo: user info to compare base context with
0808  * @uctxt: context to compare uinfo to.
0809  *
0810  * Compare the given context with the given information to see if it
0811  * can be used for a sub context.
0812  */
0813 static int match_ctxt(struct hfi1_filedata *fd,
0814               const struct hfi1_user_info *uinfo,
0815               struct hfi1_ctxtdata *uctxt)
0816 {
0817     struct hfi1_devdata *dd = fd->dd;
0818     unsigned long flags;
0819     u16 subctxt;
0820 
0821     /* Skip dynamically allocated kernel contexts */
0822     if (uctxt->sc && (uctxt->sc->type == SC_KERNEL))
0823         return 0;
0824 
0825     /* Skip ctxt if it doesn't match the requested one */
0826     if (memcmp(uctxt->uuid, uinfo->uuid, sizeof(uctxt->uuid)) ||
0827         uctxt->jkey != generate_jkey(current_uid()) ||
0828         uctxt->subctxt_id != uinfo->subctxt_id ||
0829         uctxt->subctxt_cnt != uinfo->subctxt_cnt)
0830         return 0;
0831 
0832     /* Verify the sharing process matches the base */
0833     if (uctxt->userversion != uinfo->userversion)
0834         return -EINVAL;
0835 
0836     /* Find an unused sub context */
0837     spin_lock_irqsave(&dd->uctxt_lock, flags);
0838     if (bitmap_empty(uctxt->in_use_ctxts, HFI1_MAX_SHARED_CTXTS)) {
0839         /* context is being closed, do not use */
0840         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
0841         return 0;
0842     }
0843 
0844     subctxt = find_first_zero_bit(uctxt->in_use_ctxts,
0845                       HFI1_MAX_SHARED_CTXTS);
0846     if (subctxt >= uctxt->subctxt_cnt) {
0847         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
0848         return -EBUSY;
0849     }
0850 
0851     fd->subctxt = subctxt;
0852     __set_bit(fd->subctxt, uctxt->in_use_ctxts);
0853     spin_unlock_irqrestore(&dd->uctxt_lock, flags);
0854 
0855     fd->uctxt = uctxt;
0856     hfi1_rcd_get(uctxt);
0857 
0858     return 1;
0859 }
0860 
0861 /**
0862  * find_sub_ctxt - fund sub-context
0863  * @fd: valid filedata pointer
0864  * @uinfo: matching info to use to find a possible context to share.
0865  *
0866  * The hfi1_mutex must be held when this function is called.  It is
0867  * necessary to ensure serialized creation of shared contexts.
0868  *
0869  * Return:
0870  *    0      No sub-context found
0871  *    1      Subcontext found and allocated
0872  *    errno  EINVAL (incorrect parameters)
0873  *           EBUSY (all sub contexts in use)
0874  */
0875 static int find_sub_ctxt(struct hfi1_filedata *fd,
0876              const struct hfi1_user_info *uinfo)
0877 {
0878     struct hfi1_ctxtdata *uctxt;
0879     struct hfi1_devdata *dd = fd->dd;
0880     u16 i;
0881     int ret;
0882 
0883     if (!uinfo->subctxt_cnt)
0884         return 0;
0885 
0886     for (i = dd->first_dyn_alloc_ctxt; i < dd->num_rcv_contexts; i++) {
0887         uctxt = hfi1_rcd_get_by_index(dd, i);
0888         if (uctxt) {
0889             ret = match_ctxt(fd, uinfo, uctxt);
0890             hfi1_rcd_put(uctxt);
0891             /* value of != 0 will return */
0892             if (ret)
0893                 return ret;
0894         }
0895     }
0896 
0897     return 0;
0898 }
0899 
0900 static int allocate_ctxt(struct hfi1_filedata *fd, struct hfi1_devdata *dd,
0901              struct hfi1_user_info *uinfo,
0902              struct hfi1_ctxtdata **rcd)
0903 {
0904     struct hfi1_ctxtdata *uctxt;
0905     int ret, numa;
0906 
0907     if (dd->flags & HFI1_FROZEN) {
0908         /*
0909          * Pick an error that is unique from all other errors
0910          * that are returned so the user process knows that
0911          * it tried to allocate while the SPC was frozen.  It
0912          * it should be able to retry with success in a short
0913          * while.
0914          */
0915         return -EIO;
0916     }
0917 
0918     if (!dd->freectxts)
0919         return -EBUSY;
0920 
0921     /*
0922      * If we don't have a NUMA node requested, preference is towards
0923      * device NUMA node.
0924      */
0925     fd->rec_cpu_num = hfi1_get_proc_affinity(dd->node);
0926     if (fd->rec_cpu_num != -1)
0927         numa = cpu_to_node(fd->rec_cpu_num);
0928     else
0929         numa = numa_node_id();
0930     ret = hfi1_create_ctxtdata(dd->pport, numa, &uctxt);
0931     if (ret < 0) {
0932         dd_dev_err(dd, "user ctxtdata allocation failed\n");
0933         return ret;
0934     }
0935     hfi1_cdbg(PROC, "[%u:%u] pid %u assigned to CPU %d (NUMA %u)",
0936           uctxt->ctxt, fd->subctxt, current->pid, fd->rec_cpu_num,
0937           uctxt->numa_id);
0938 
0939     /*
0940      * Allocate and enable a PIO send context.
0941      */
0942     uctxt->sc = sc_alloc(dd, SC_USER, uctxt->rcvhdrqentsize, dd->node);
0943     if (!uctxt->sc) {
0944         ret = -ENOMEM;
0945         goto ctxdata_free;
0946     }
0947     hfi1_cdbg(PROC, "allocated send context %u(%u)\n", uctxt->sc->sw_index,
0948           uctxt->sc->hw_context);
0949     ret = sc_enable(uctxt->sc);
0950     if (ret)
0951         goto ctxdata_free;
0952 
0953     /*
0954      * Setup sub context information if the user-level has requested
0955      * sub contexts.
0956      * This has to be done here so the rest of the sub-contexts find the
0957      * proper base context.
0958      * NOTE: _set_bit() can be used here because the context creation is
0959      * protected by the mutex (rather than the spin_lock), and will be the
0960      * very first instance of this context.
0961      */
0962     __set_bit(0, uctxt->in_use_ctxts);
0963     if (uinfo->subctxt_cnt)
0964         init_subctxts(uctxt, uinfo);
0965     uctxt->userversion = uinfo->userversion;
0966     uctxt->flags = hfi1_cap_mask; /* save current flag state */
0967     init_waitqueue_head(&uctxt->wait);
0968     strlcpy(uctxt->comm, current->comm, sizeof(uctxt->comm));
0969     memcpy(uctxt->uuid, uinfo->uuid, sizeof(uctxt->uuid));
0970     uctxt->jkey = generate_jkey(current_uid());
0971     hfi1_stats.sps_ctxts++;
0972     /*
0973      * Disable ASPM when there are open user/PSM contexts to avoid
0974      * issues with ASPM L1 exit latency
0975      */
0976     if (dd->freectxts-- == dd->num_user_contexts)
0977         aspm_disable_all(dd);
0978 
0979     *rcd = uctxt;
0980 
0981     return 0;
0982 
0983 ctxdata_free:
0984     hfi1_free_ctxt(uctxt);
0985     return ret;
0986 }
0987 
0988 static void deallocate_ctxt(struct hfi1_ctxtdata *uctxt)
0989 {
0990     mutex_lock(&hfi1_mutex);
0991     hfi1_stats.sps_ctxts--;
0992     if (++uctxt->dd->freectxts == uctxt->dd->num_user_contexts)
0993         aspm_enable_all(uctxt->dd);
0994     mutex_unlock(&hfi1_mutex);
0995 
0996     hfi1_free_ctxt(uctxt);
0997 }
0998 
0999 static void init_subctxts(struct hfi1_ctxtdata *uctxt,
1000               const struct hfi1_user_info *uinfo)
1001 {
1002     uctxt->subctxt_cnt = uinfo->subctxt_cnt;
1003     uctxt->subctxt_id = uinfo->subctxt_id;
1004     set_bit(HFI1_CTXT_BASE_UNINIT, &uctxt->event_flags);
1005 }
1006 
1007 static int setup_subctxt(struct hfi1_ctxtdata *uctxt)
1008 {
1009     int ret = 0;
1010     u16 num_subctxts = uctxt->subctxt_cnt;
1011 
1012     uctxt->subctxt_uregbase = vmalloc_user(PAGE_SIZE);
1013     if (!uctxt->subctxt_uregbase)
1014         return -ENOMEM;
1015 
1016     /* We can take the size of the RcvHdr Queue from the master */
1017     uctxt->subctxt_rcvhdr_base = vmalloc_user(rcvhdrq_size(uctxt) *
1018                           num_subctxts);
1019     if (!uctxt->subctxt_rcvhdr_base) {
1020         ret = -ENOMEM;
1021         goto bail_ureg;
1022     }
1023 
1024     uctxt->subctxt_rcvegrbuf = vmalloc_user(uctxt->egrbufs.size *
1025                         num_subctxts);
1026     if (!uctxt->subctxt_rcvegrbuf) {
1027         ret = -ENOMEM;
1028         goto bail_rhdr;
1029     }
1030 
1031     return 0;
1032 
1033 bail_rhdr:
1034     vfree(uctxt->subctxt_rcvhdr_base);
1035     uctxt->subctxt_rcvhdr_base = NULL;
1036 bail_ureg:
1037     vfree(uctxt->subctxt_uregbase);
1038     uctxt->subctxt_uregbase = NULL;
1039 
1040     return ret;
1041 }
1042 
1043 static void user_init(struct hfi1_ctxtdata *uctxt)
1044 {
1045     unsigned int rcvctrl_ops = 0;
1046 
1047     /* initialize poll variables... */
1048     uctxt->urgent = 0;
1049     uctxt->urgent_poll = 0;
1050 
1051     /*
1052      * Now enable the ctxt for receive.
1053      * For chips that are set to DMA the tail register to memory
1054      * when they change (and when the update bit transitions from
1055      * 0 to 1.  So for those chips, we turn it off and then back on.
1056      * This will (very briefly) affect any other open ctxts, but the
1057      * duration is very short, and therefore isn't an issue.  We
1058      * explicitly set the in-memory tail copy to 0 beforehand, so we
1059      * don't have to wait to be sure the DMA update has happened
1060      * (chip resets head/tail to 0 on transition to enable).
1061      */
1062     if (hfi1_rcvhdrtail_kvaddr(uctxt))
1063         clear_rcvhdrtail(uctxt);
1064 
1065     /* Setup J_KEY before enabling the context */
1066     hfi1_set_ctxt_jkey(uctxt->dd, uctxt, uctxt->jkey);
1067 
1068     rcvctrl_ops = HFI1_RCVCTRL_CTXT_ENB;
1069     rcvctrl_ops |= HFI1_RCVCTRL_URGENT_ENB;
1070     if (HFI1_CAP_UGET_MASK(uctxt->flags, HDRSUPP))
1071         rcvctrl_ops |= HFI1_RCVCTRL_TIDFLOW_ENB;
1072     /*
1073      * Ignore the bit in the flags for now until proper
1074      * support for multiple packet per rcv array entry is
1075      * added.
1076      */
1077     if (!HFI1_CAP_UGET_MASK(uctxt->flags, MULTI_PKT_EGR))
1078         rcvctrl_ops |= HFI1_RCVCTRL_ONE_PKT_EGR_ENB;
1079     if (HFI1_CAP_UGET_MASK(uctxt->flags, NODROP_EGR_FULL))
1080         rcvctrl_ops |= HFI1_RCVCTRL_NO_EGR_DROP_ENB;
1081     if (HFI1_CAP_UGET_MASK(uctxt->flags, NODROP_RHQ_FULL))
1082         rcvctrl_ops |= HFI1_RCVCTRL_NO_RHQ_DROP_ENB;
1083     /*
1084      * The RcvCtxtCtrl.TailUpd bit has to be explicitly written.
1085      * We can't rely on the correct value to be set from prior
1086      * uses of the chip or ctxt. Therefore, add the rcvctrl op
1087      * for both cases.
1088      */
1089     if (HFI1_CAP_UGET_MASK(uctxt->flags, DMA_RTAIL))
1090         rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_ENB;
1091     else
1092         rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_DIS;
1093     hfi1_rcvctrl(uctxt->dd, rcvctrl_ops, uctxt);
1094 }
1095 
1096 static int get_ctxt_info(struct hfi1_filedata *fd, unsigned long arg, u32 len)
1097 {
1098     struct hfi1_ctxt_info cinfo;
1099     struct hfi1_ctxtdata *uctxt = fd->uctxt;
1100 
1101     if (sizeof(cinfo) != len)
1102         return -EINVAL;
1103 
1104     memset(&cinfo, 0, sizeof(cinfo));
1105     cinfo.runtime_flags = (((uctxt->flags >> HFI1_CAP_MISC_SHIFT) &
1106                 HFI1_CAP_MISC_MASK) << HFI1_CAP_USER_SHIFT) |
1107             HFI1_CAP_UGET_MASK(uctxt->flags, MASK) |
1108             HFI1_CAP_KGET_MASK(uctxt->flags, K2U);
1109     /* adjust flag if this fd is not able to cache */
1110     if (!fd->use_mn)
1111         cinfo.runtime_flags |= HFI1_CAP_TID_UNMAP; /* no caching */
1112 
1113     cinfo.num_active = hfi1_count_active_units();
1114     cinfo.unit = uctxt->dd->unit;
1115     cinfo.ctxt = uctxt->ctxt;
1116     cinfo.subctxt = fd->subctxt;
1117     cinfo.rcvtids = roundup(uctxt->egrbufs.alloced,
1118                 uctxt->dd->rcv_entries.group_size) +
1119         uctxt->expected_count;
1120     cinfo.credits = uctxt->sc->credits;
1121     cinfo.numa_node = uctxt->numa_id;
1122     cinfo.rec_cpu = fd->rec_cpu_num;
1123     cinfo.send_ctxt = uctxt->sc->hw_context;
1124 
1125     cinfo.egrtids = uctxt->egrbufs.alloced;
1126     cinfo.rcvhdrq_cnt = get_hdrq_cnt(uctxt);
1127     cinfo.rcvhdrq_entsize = get_hdrqentsize(uctxt) << 2;
1128     cinfo.sdma_ring_size = fd->cq->nentries;
1129     cinfo.rcvegr_size = uctxt->egrbufs.rcvtid_size;
1130 
1131     trace_hfi1_ctxt_info(uctxt->dd, uctxt->ctxt, fd->subctxt, &cinfo);
1132     if (copy_to_user((void __user *)arg, &cinfo, len))
1133         return -EFAULT;
1134 
1135     return 0;
1136 }
1137 
1138 static int init_user_ctxt(struct hfi1_filedata *fd,
1139               struct hfi1_ctxtdata *uctxt)
1140 {
1141     int ret;
1142 
1143     ret = hfi1_user_sdma_alloc_queues(uctxt, fd);
1144     if (ret)
1145         return ret;
1146 
1147     ret = hfi1_user_exp_rcv_init(fd, uctxt);
1148     if (ret)
1149         hfi1_user_sdma_free_queues(fd, uctxt);
1150 
1151     return ret;
1152 }
1153 
1154 static int setup_base_ctxt(struct hfi1_filedata *fd,
1155                struct hfi1_ctxtdata *uctxt)
1156 {
1157     struct hfi1_devdata *dd = uctxt->dd;
1158     int ret = 0;
1159 
1160     hfi1_init_ctxt(uctxt->sc);
1161 
1162     /* Now allocate the RcvHdr queue and eager buffers. */
1163     ret = hfi1_create_rcvhdrq(dd, uctxt);
1164     if (ret)
1165         goto done;
1166 
1167     ret = hfi1_setup_eagerbufs(uctxt);
1168     if (ret)
1169         goto done;
1170 
1171     /* If sub-contexts are enabled, do the appropriate setup */
1172     if (uctxt->subctxt_cnt)
1173         ret = setup_subctxt(uctxt);
1174     if (ret)
1175         goto done;
1176 
1177     ret = hfi1_alloc_ctxt_rcv_groups(uctxt);
1178     if (ret)
1179         goto done;
1180 
1181     ret = init_user_ctxt(fd, uctxt);
1182     if (ret) {
1183         hfi1_free_ctxt_rcv_groups(uctxt);
1184         goto done;
1185     }
1186 
1187     user_init(uctxt);
1188 
1189     /* Now that the context is set up, the fd can get a reference. */
1190     fd->uctxt = uctxt;
1191     hfi1_rcd_get(uctxt);
1192 
1193 done:
1194     if (uctxt->subctxt_cnt) {
1195         /*
1196          * On error, set the failed bit so sub-contexts will clean up
1197          * correctly.
1198          */
1199         if (ret)
1200             set_bit(HFI1_CTXT_BASE_FAILED, &uctxt->event_flags);
1201 
1202         /*
1203          * Base context is done (successfully or not), notify anybody
1204          * using a sub-context that is waiting for this completion.
1205          */
1206         clear_bit(HFI1_CTXT_BASE_UNINIT, &uctxt->event_flags);
1207         wake_up(&uctxt->wait);
1208     }
1209 
1210     return ret;
1211 }
1212 
1213 static int get_base_info(struct hfi1_filedata *fd, unsigned long arg, u32 len)
1214 {
1215     struct hfi1_base_info binfo;
1216     struct hfi1_ctxtdata *uctxt = fd->uctxt;
1217     struct hfi1_devdata *dd = uctxt->dd;
1218     unsigned offset;
1219 
1220     trace_hfi1_uctxtdata(uctxt->dd, uctxt, fd->subctxt);
1221 
1222     if (sizeof(binfo) != len)
1223         return -EINVAL;
1224 
1225     memset(&binfo, 0, sizeof(binfo));
1226     binfo.hw_version = dd->revision;
1227     binfo.sw_version = HFI1_USER_SWVERSION;
1228     binfo.bthqp = RVT_KDETH_QP_PREFIX;
1229     binfo.jkey = uctxt->jkey;
1230     /*
1231      * If more than 64 contexts are enabled the allocated credit
1232      * return will span two or three contiguous pages. Since we only
1233      * map the page containing the context's credit return address,
1234      * we need to calculate the offset in the proper page.
1235      */
1236     offset = ((u64)uctxt->sc->hw_free -
1237           (u64)dd->cr_base[uctxt->numa_id].va) % PAGE_SIZE;
1238     binfo.sc_credits_addr = HFI1_MMAP_TOKEN(PIO_CRED, uctxt->ctxt,
1239                         fd->subctxt, offset);
1240     binfo.pio_bufbase = HFI1_MMAP_TOKEN(PIO_BUFS, uctxt->ctxt,
1241                         fd->subctxt,
1242                         uctxt->sc->base_addr);
1243     binfo.pio_bufbase_sop = HFI1_MMAP_TOKEN(PIO_BUFS_SOP,
1244                         uctxt->ctxt,
1245                         fd->subctxt,
1246                         uctxt->sc->base_addr);
1247     binfo.rcvhdr_bufbase = HFI1_MMAP_TOKEN(RCV_HDRQ, uctxt->ctxt,
1248                            fd->subctxt,
1249                            uctxt->rcvhdrq);
1250     binfo.rcvegr_bufbase = HFI1_MMAP_TOKEN(RCV_EGRBUF, uctxt->ctxt,
1251                            fd->subctxt,
1252                            uctxt->egrbufs.rcvtids[0].dma);
1253     binfo.sdma_comp_bufbase = HFI1_MMAP_TOKEN(SDMA_COMP, uctxt->ctxt,
1254                           fd->subctxt, 0);
1255     /*
1256      * user regs are at
1257      * (RXE_PER_CONTEXT_USER + (ctxt * RXE_PER_CONTEXT_SIZE))
1258      */
1259     binfo.user_regbase = HFI1_MMAP_TOKEN(UREGS, uctxt->ctxt,
1260                          fd->subctxt, 0);
1261     offset = offset_in_page((uctxt_offset(uctxt) + fd->subctxt) *
1262                 sizeof(*dd->events));
1263     binfo.events_bufbase = HFI1_MMAP_TOKEN(EVENTS, uctxt->ctxt,
1264                            fd->subctxt,
1265                            offset);
1266     binfo.status_bufbase = HFI1_MMAP_TOKEN(STATUS, uctxt->ctxt,
1267                            fd->subctxt,
1268                            dd->status);
1269     if (HFI1_CAP_IS_USET(DMA_RTAIL))
1270         binfo.rcvhdrtail_base = HFI1_MMAP_TOKEN(RTAIL, uctxt->ctxt,
1271                             fd->subctxt, 0);
1272     if (uctxt->subctxt_cnt) {
1273         binfo.subctxt_uregbase = HFI1_MMAP_TOKEN(SUBCTXT_UREGS,
1274                              uctxt->ctxt,
1275                              fd->subctxt, 0);
1276         binfo.subctxt_rcvhdrbuf = HFI1_MMAP_TOKEN(SUBCTXT_RCV_HDRQ,
1277                               uctxt->ctxt,
1278                               fd->subctxt, 0);
1279         binfo.subctxt_rcvegrbuf = HFI1_MMAP_TOKEN(SUBCTXT_EGRBUF,
1280                               uctxt->ctxt,
1281                               fd->subctxt, 0);
1282     }
1283 
1284     if (copy_to_user((void __user *)arg, &binfo, len))
1285         return -EFAULT;
1286 
1287     return 0;
1288 }
1289 
1290 /**
1291  * user_exp_rcv_setup - Set up the given tid rcv list
1292  * @fd: file data of the current driver instance
1293  * @arg: ioctl argumnent for user space information
1294  * @len: length of data structure associated with ioctl command
1295  *
1296  * Wrapper to validate ioctl information before doing _rcv_setup.
1297  *
1298  */
1299 static int user_exp_rcv_setup(struct hfi1_filedata *fd, unsigned long arg,
1300                   u32 len)
1301 {
1302     int ret;
1303     unsigned long addr;
1304     struct hfi1_tid_info tinfo;
1305 
1306     if (sizeof(tinfo) != len)
1307         return -EINVAL;
1308 
1309     if (copy_from_user(&tinfo, (void __user *)arg, (sizeof(tinfo))))
1310         return -EFAULT;
1311 
1312     ret = hfi1_user_exp_rcv_setup(fd, &tinfo);
1313     if (!ret) {
1314         /*
1315          * Copy the number of tidlist entries we used
1316          * and the length of the buffer we registered.
1317          */
1318         addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
1319         if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
1320                  sizeof(tinfo.tidcnt)))
1321             return -EFAULT;
1322 
1323         addr = arg + offsetof(struct hfi1_tid_info, length);
1324         if (copy_to_user((void __user *)addr, &tinfo.length,
1325                  sizeof(tinfo.length)))
1326             ret = -EFAULT;
1327     }
1328 
1329     return ret;
1330 }
1331 
1332 /**
1333  * user_exp_rcv_clear - Clear the given tid rcv list
1334  * @fd: file data of the current driver instance
1335  * @arg: ioctl argumnent for user space information
1336  * @len: length of data structure associated with ioctl command
1337  *
1338  * The hfi1_user_exp_rcv_clear() can be called from the error path.  Because
1339  * of this, we need to use this wrapper to copy the user space information
1340  * before doing the clear.
1341  */
1342 static int user_exp_rcv_clear(struct hfi1_filedata *fd, unsigned long arg,
1343                   u32 len)
1344 {
1345     int ret;
1346     unsigned long addr;
1347     struct hfi1_tid_info tinfo;
1348 
1349     if (sizeof(tinfo) != len)
1350         return -EINVAL;
1351 
1352     if (copy_from_user(&tinfo, (void __user *)arg, (sizeof(tinfo))))
1353         return -EFAULT;
1354 
1355     ret = hfi1_user_exp_rcv_clear(fd, &tinfo);
1356     if (!ret) {
1357         addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
1358         if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
1359                  sizeof(tinfo.tidcnt)))
1360             return -EFAULT;
1361     }
1362 
1363     return ret;
1364 }
1365 
1366 /**
1367  * user_exp_rcv_invalid - Invalidate the given tid rcv list
1368  * @fd: file data of the current driver instance
1369  * @arg: ioctl argumnent for user space information
1370  * @len: length of data structure associated with ioctl command
1371  *
1372  * Wrapper to validate ioctl information before doing _rcv_invalid.
1373  *
1374  */
1375 static int user_exp_rcv_invalid(struct hfi1_filedata *fd, unsigned long arg,
1376                 u32 len)
1377 {
1378     int ret;
1379     unsigned long addr;
1380     struct hfi1_tid_info tinfo;
1381 
1382     if (sizeof(tinfo) != len)
1383         return -EINVAL;
1384 
1385     if (!fd->invalid_tids)
1386         return -EINVAL;
1387 
1388     if (copy_from_user(&tinfo, (void __user *)arg, (sizeof(tinfo))))
1389         return -EFAULT;
1390 
1391     ret = hfi1_user_exp_rcv_invalid(fd, &tinfo);
1392     if (ret)
1393         return ret;
1394 
1395     addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
1396     if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
1397              sizeof(tinfo.tidcnt)))
1398         ret = -EFAULT;
1399 
1400     return ret;
1401 }
1402 
1403 static __poll_t poll_urgent(struct file *fp,
1404                 struct poll_table_struct *pt)
1405 {
1406     struct hfi1_filedata *fd = fp->private_data;
1407     struct hfi1_ctxtdata *uctxt = fd->uctxt;
1408     struct hfi1_devdata *dd = uctxt->dd;
1409     __poll_t pollflag;
1410 
1411     poll_wait(fp, &uctxt->wait, pt);
1412 
1413     spin_lock_irq(&dd->uctxt_lock);
1414     if (uctxt->urgent != uctxt->urgent_poll) {
1415         pollflag = EPOLLIN | EPOLLRDNORM;
1416         uctxt->urgent_poll = uctxt->urgent;
1417     } else {
1418         pollflag = 0;
1419         set_bit(HFI1_CTXT_WAITING_URG, &uctxt->event_flags);
1420     }
1421     spin_unlock_irq(&dd->uctxt_lock);
1422 
1423     return pollflag;
1424 }
1425 
1426 static __poll_t poll_next(struct file *fp,
1427                   struct poll_table_struct *pt)
1428 {
1429     struct hfi1_filedata *fd = fp->private_data;
1430     struct hfi1_ctxtdata *uctxt = fd->uctxt;
1431     struct hfi1_devdata *dd = uctxt->dd;
1432     __poll_t pollflag;
1433 
1434     poll_wait(fp, &uctxt->wait, pt);
1435 
1436     spin_lock_irq(&dd->uctxt_lock);
1437     if (hdrqempty(uctxt)) {
1438         set_bit(HFI1_CTXT_WAITING_RCV, &uctxt->event_flags);
1439         hfi1_rcvctrl(dd, HFI1_RCVCTRL_INTRAVAIL_ENB, uctxt);
1440         pollflag = 0;
1441     } else {
1442         pollflag = EPOLLIN | EPOLLRDNORM;
1443     }
1444     spin_unlock_irq(&dd->uctxt_lock);
1445 
1446     return pollflag;
1447 }
1448 
1449 /*
1450  * Find all user contexts in use, and set the specified bit in their
1451  * event mask.
1452  * See also find_ctxt() for a similar use, that is specific to send buffers.
1453  */
1454 int hfi1_set_uevent_bits(struct hfi1_pportdata *ppd, const int evtbit)
1455 {
1456     struct hfi1_ctxtdata *uctxt;
1457     struct hfi1_devdata *dd = ppd->dd;
1458     u16 ctxt;
1459 
1460     if (!dd->events)
1461         return -EINVAL;
1462 
1463     for (ctxt = dd->first_dyn_alloc_ctxt; ctxt < dd->num_rcv_contexts;
1464          ctxt++) {
1465         uctxt = hfi1_rcd_get_by_index(dd, ctxt);
1466         if (uctxt) {
1467             unsigned long *evs;
1468             int i;
1469             /*
1470              * subctxt_cnt is 0 if not shared, so do base
1471              * separately, first, then remaining subctxt, if any
1472              */
1473             evs = dd->events + uctxt_offset(uctxt);
1474             set_bit(evtbit, evs);
1475             for (i = 1; i < uctxt->subctxt_cnt; i++)
1476                 set_bit(evtbit, evs + i);
1477             hfi1_rcd_put(uctxt);
1478         }
1479     }
1480 
1481     return 0;
1482 }
1483 
1484 /**
1485  * manage_rcvq - manage a context's receive queue
1486  * @uctxt: the context
1487  * @subctxt: the sub-context
1488  * @arg: start/stop action to carry out
1489  *
1490  * start_stop == 0 disables receive on the context, for use in queue
1491  * overflow conditions.  start_stop==1 re-enables, to be used to
1492  * re-init the software copy of the head register
1493  */
1494 static int manage_rcvq(struct hfi1_ctxtdata *uctxt, u16 subctxt,
1495                unsigned long arg)
1496 {
1497     struct hfi1_devdata *dd = uctxt->dd;
1498     unsigned int rcvctrl_op;
1499     int start_stop;
1500 
1501     if (subctxt)
1502         return 0;
1503 
1504     if (get_user(start_stop, (int __user *)arg))
1505         return -EFAULT;
1506 
1507     /* atomically clear receive enable ctxt. */
1508     if (start_stop) {
1509         /*
1510          * On enable, force in-memory copy of the tail register to
1511          * 0, so that protocol code doesn't have to worry about
1512          * whether or not the chip has yet updated the in-memory
1513          * copy or not on return from the system call. The chip
1514          * always resets it's tail register back to 0 on a
1515          * transition from disabled to enabled.
1516          */
1517         if (hfi1_rcvhdrtail_kvaddr(uctxt))
1518             clear_rcvhdrtail(uctxt);
1519         rcvctrl_op = HFI1_RCVCTRL_CTXT_ENB;
1520     } else {
1521         rcvctrl_op = HFI1_RCVCTRL_CTXT_DIS;
1522     }
1523     hfi1_rcvctrl(dd, rcvctrl_op, uctxt);
1524     /* always; new head should be equal to new tail; see above */
1525 
1526     return 0;
1527 }
1528 
1529 /*
1530  * clear the event notifier events for this context.
1531  * User process then performs actions appropriate to bit having been
1532  * set, if desired, and checks again in future.
1533  */
1534 static int user_event_ack(struct hfi1_ctxtdata *uctxt, u16 subctxt,
1535               unsigned long arg)
1536 {
1537     int i;
1538     struct hfi1_devdata *dd = uctxt->dd;
1539     unsigned long *evs;
1540     unsigned long events;
1541 
1542     if (!dd->events)
1543         return 0;
1544 
1545     if (get_user(events, (unsigned long __user *)arg))
1546         return -EFAULT;
1547 
1548     evs = dd->events + uctxt_offset(uctxt) + subctxt;
1549 
1550     for (i = 0; i <= _HFI1_MAX_EVENT_BIT; i++) {
1551         if (!test_bit(i, &events))
1552             continue;
1553         clear_bit(i, evs);
1554     }
1555     return 0;
1556 }
1557 
1558 static int set_ctxt_pkey(struct hfi1_ctxtdata *uctxt, unsigned long arg)
1559 {
1560     int i;
1561     struct hfi1_pportdata *ppd = uctxt->ppd;
1562     struct hfi1_devdata *dd = uctxt->dd;
1563     u16 pkey;
1564 
1565     if (!HFI1_CAP_IS_USET(PKEY_CHECK))
1566         return -EPERM;
1567 
1568     if (get_user(pkey, (u16 __user *)arg))
1569         return -EFAULT;
1570 
1571     if (pkey == LIM_MGMT_P_KEY || pkey == FULL_MGMT_P_KEY)
1572         return -EINVAL;
1573 
1574     for (i = 0; i < ARRAY_SIZE(ppd->pkeys); i++)
1575         if (pkey == ppd->pkeys[i])
1576             return hfi1_set_ctxt_pkey(dd, uctxt, pkey);
1577 
1578     return -ENOENT;
1579 }
1580 
1581 /**
1582  * ctxt_reset - Reset the user context
1583  * @uctxt: valid user context
1584  */
1585 static int ctxt_reset(struct hfi1_ctxtdata *uctxt)
1586 {
1587     struct send_context *sc;
1588     struct hfi1_devdata *dd;
1589     int ret = 0;
1590 
1591     if (!uctxt || !uctxt->dd || !uctxt->sc)
1592         return -EINVAL;
1593 
1594     /*
1595      * There is no protection here. User level has to guarantee that
1596      * no one will be writing to the send context while it is being
1597      * re-initialized.  If user level breaks that guarantee, it will
1598      * break it's own context and no one else's.
1599      */
1600     dd = uctxt->dd;
1601     sc = uctxt->sc;
1602 
1603     /*
1604      * Wait until the interrupt handler has marked the context as
1605      * halted or frozen. Report error if we time out.
1606      */
1607     wait_event_interruptible_timeout(
1608         sc->halt_wait, (sc->flags & SCF_HALTED),
1609         msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
1610     if (!(sc->flags & SCF_HALTED))
1611         return -ENOLCK;
1612 
1613     /*
1614      * If the send context was halted due to a Freeze, wait until the
1615      * device has been "unfrozen" before resetting the context.
1616      */
1617     if (sc->flags & SCF_FROZEN) {
1618         wait_event_interruptible_timeout(
1619             dd->event_queue,
1620             !(READ_ONCE(dd->flags) & HFI1_FROZEN),
1621             msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
1622         if (dd->flags & HFI1_FROZEN)
1623             return -ENOLCK;
1624 
1625         if (dd->flags & HFI1_FORCED_FREEZE)
1626             /*
1627              * Don't allow context reset if we are into
1628              * forced freeze
1629              */
1630             return -ENODEV;
1631 
1632         sc_disable(sc);
1633         ret = sc_enable(sc);
1634         hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_ENB, uctxt);
1635     } else {
1636         ret = sc_restart(sc);
1637     }
1638     if (!ret)
1639         sc_return_credits(sc);
1640 
1641     return ret;
1642 }
1643 
1644 static void user_remove(struct hfi1_devdata *dd)
1645 {
1646 
1647     hfi1_cdev_cleanup(&dd->user_cdev, &dd->user_device);
1648 }
1649 
1650 static int user_add(struct hfi1_devdata *dd)
1651 {
1652     char name[10];
1653     int ret;
1654 
1655     snprintf(name, sizeof(name), "%s_%d", class_name(), dd->unit);
1656     ret = hfi1_cdev_init(dd->unit, name, &hfi1_file_ops,
1657                  &dd->user_cdev, &dd->user_device,
1658                  true, &dd->verbs_dev.rdi.ibdev.dev.kobj);
1659     if (ret)
1660         user_remove(dd);
1661 
1662     return ret;
1663 }
1664 
1665 /*
1666  * Create per-unit files in /dev
1667  */
1668 int hfi1_device_create(struct hfi1_devdata *dd)
1669 {
1670     return user_add(dd);
1671 }
1672 
1673 /*
1674  * Remove per-unit files in /dev
1675  * void, core kernel returns no errors for this stuff
1676  */
1677 void hfi1_device_remove(struct hfi1_devdata *dd)
1678 {
1679     user_remove(dd);
1680 }
[ Source navigation ] [ Diff markup ] [ Identifier search ] [ general search ]

This page was automatically generated by the 2.1.0 LXR engine.

The LXR team
Valid CSS 2.1! Valid HTML 4.01!