gh0s7/mckernel
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

hfi1/user_exp_rcv: rework main loop

Browse Source 
New loop now takes into account pages not physically contiguous.
Also some minor improvements, e.g. make the spin_lock used more locally,
reuse a group we had if we had one, etc.
This commit is contained in:
Dominique Martinet
2017-09-21 16:24:45 +09:00
committed by Balazs Gerofi
parent 0f8f88ca46
commit f5ced648ef
3 changed files with 178 additions and 176 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
kernel/chip.c
View File

@ -54,6 +54,15 @@
#include <hfi1/chip_registers.h>
#include <hfi1/chip.h>
//#define DEBUG_PRINT_CHIP
#ifdef DEBUG_PRINT_CHIP
#define dkprintf(...) kprintf(__VA_ARGS__)
#else
#define dkprintf(...) do { if(0) kprintf(__VA_ARGS__); } while (0)
#endif
/*
* index is the index into the receive array
*/
@ -85,6 +94,8 @@ void hfi1_put_tid(struct hfi1_devdata *dd, u32 index,
| (u64)order << RCV_ARRAY_RT_BUF_SIZE_SHIFT
| ((pa >> RT_ADDR_SHIFT) & RCV_ARRAY_RT_ADDR_MASK)
<< RCV_ARRAY_RT_ADDR_SHIFT;
dkprintf("type %d, index 0x%x, pa 0x%lx, bsize 0x%lx, reg 0x%llx\n",
type, index, pa, (unsigned long)order, reg);
writeq(reg, base + (index * 8));
if (type == PT_EAGER)

2
kernel/include/hfi1/chip.h
View File

@ -51,6 +51,8 @@
* This file contains all of the defines that is specific to the HFI chip
*/
#define MAX_EXPECTED_BUFFER (2048 * 1024)
void hfi1_put_tid(struct hfi1_devdata *dd, u32 index,
u32 type, unsigned long pa, u16 order);
void hfi1_clear_tids(struct hfi1_ctxtdata *rcd);

301
kernel/user_exp_rcv.c
View File

@ -51,8 +51,15 @@
#include <hfi1/chip.h>
#include <hfi1/user_exp_rcv.h>
static int program_rcvarray(struct hfi1_filedata *, uintptr_t, u16, struct tid_group *,
u32 *);
//#define DEBUG_PRINT_USER_EXP_RCV
#ifdef DEBUG_PRINT_USER_EXP_RCV
#define dkprintf(...) kprintf(__VA_ARGS__)
#else
#define dkprintf(...) do { if(0) kprintf(__VA_ARGS__); } while (0)
#endif
static int program_rcvarray(struct hfi1_filedata *, uintptr_t, size_t, u32 *);
static int set_rcvarray_entry(struct hfi1_filedata *, uintptr_t,
u32, struct tid_group *,
u16);
@ -70,65 +77,18 @@ struct tid_rb_node {
/*
* RcvArray entry allocation for Expected Receives is done by the
* following algorithm:
*
* The context keeps 3 lists of groups of RcvArray entries:
* 1. List of empty groups - tid_group_list
* This list is created during user context creation and
* contains elements which describe sets (of 8) of empty
* RcvArray entries.
* 2. List of partially used groups - tid_used_list
* This list contains sets of RcvArray entries which are
* not completely used up. Another mapping request could
* use some of all of the remaining entries.
* 3. List of full groups - tid_full_list
* This is the list where sets that are completely used
* up go.
*
* An attempt to optimize the usage of RcvArray entries is
* made by finding all sets of physically contiguous pages in a
* user's buffer.
* These physically contiguous sets are further split into
* sizes supported by the receive engine of the HFI. The
* resulting sets of pages are stored in struct tid_pageset,
* which describes the sets as:
* * .count - number of pages in this set
* * .idx - starting index into struct page ** array
* of this set
*
* From this point on, the algorithm deals with the page sets
* described above. The number of pagesets is divided by the
* RcvArray group size to produce the number of full groups
* needed.
*
* Groups from the 3 lists are manipulated using the following
* rules:
* 1. For each set of 8 pagesets, a complete group from
* tid_group_list is taken, programmed, and moved to
* the tid_full_list list.
* 2. For all remaining pagesets:
* 2.1 If the tid_used_list is empty and the tid_group_list
* is empty, stop processing pageset and return only
* what has been programmed up to this point.
* 2.2 If the tid_used_list is empty and the tid_group_list
* is not empty, move a group from tid_group_list to
* tid_used_list.
* 2.3 For each group is tid_used_group, program as much as
* can fit into the group. If the group becomes fully
* used, move it to tid_full_list.
*/
int hfi1_user_exp_rcv_setup(struct hfi1_filedata *fd, struct hfi1_tid_info *tinfo)
{
int ret = -EFAULT;
struct hfi1_ctxtdata *uctxt = fd->uctxt;
uintptr_t vaddr = tinfo->vaddr;
u32 tid[20]; /* at most 20 requests with this algorithm */
uintptr_t vaddr_prev, vaddr, vaddr_end;
u32 *tidlist;
u16 tididx = 0;
s16 order;
u32 npages;
struct process_vm *vm = cpu_local_var(current)->vm;
size_t base_pgsize;
size_t base_pgsize, len = 0;
pte_t *ptep;
u64 phys;
u64 phys_start = 0, phys_prev = 0, phys;
if (!tinfo->length)
return -EINVAL;
@ -138,87 +98,107 @@ int hfi1_user_exp_rcv_setup(struct hfi1_filedata *fd, struct hfi1_tid_info *tinf
return -EINVAL;
}
tidlist = kmalloc(sizeof(*tidlist)*uctxt->expected_count,
IHK_MC_AP_NOWAIT);
if (!tidlist)
return -ENOMEM;
/* Verify that access is OK for the user buffer */
// TODO: iterate over vm memory ranges for write access
// return -EFAULT;
/* Simplified design: vaddr to vaddr + tinfo->length is contiguous
* for us, but program_rcvarray only deals with powers of two
* -> we need as many requests as there are bits set in length
*
* Note that we only work with multiples of 4k, so round up and shift
*/
npages = (tinfo->length + 4095) >> 12;
vaddr_end = tinfo->vaddr + tinfo->length;
dkprintf("setup start: 0x%llx, length: %zu\n", tinfo->vaddr,
tinfo->length);
for (vaddr = tinfo->vaddr; vaddr < vaddr_end; vaddr += base_pgsize) {
ptep = ihk_mc_pt_lookup_pte(vm->address_space->page_table,
(void*)vaddr, 0, 0, &base_pgsize, 0);
(void*)vaddr, 0, 0, &base_pgsize,
0);
if (unlikely(!ptep || !pte_is_present(ptep))) {
kprintf("%s: ERRROR: no valid PTE for 0x%lx\n",
__FUNCTION__, vaddr);
return -EFAULT;
ret = -EFAULT;
break;
}
phys = pte_get_phys(ptep);
for (order = 19; order >= 0; order--)
{
struct tid_group *grp;
if (!(npages & (1 << order)))
continue;
spin_lock(&fd->tid_lock);
if (!uctxt->tid_used_list.count) {
if (!uctxt->tid_group_list.count) {
goto unlock;
}
grp = tid_group_pop(&uctxt->tid_group_list);
} else {
grp = tid_group_pop(&uctxt->tid_used_list);
}
ret = program_rcvarray(fd, phys, order, grp, tid + (tididx++));
if (ret < 0) {
hfi1_cdbg(TID,
"Failed to program RcvArray entries %d",
if (!phys_prev) {
/* first pass */
phys_start = phys;
} else if (len == MAX_EXPECTED_BUFFER ||
phys - phys_prev != vaddr - vaddr_prev) {
/*
* Two ways to get here:
* - Segment has reached max size;
* - Found a new segment;
*
* Register what we have.
*/
ret = program_rcvarray(fd, phys_start, len,
tidlist + tididx);
if (ret <= 0) {
kprintf("Failed to program RcvArray entries: %d\n",
ret);
ret = -EFAULT;
} else if (WARN_ON(ret == 0)) {
}
tididx += ret;
phys_start = phys;
len = 0;
}
/* Corner case #1: base_pgsize is too big (requested length) */
if (vaddr + base_pgsize > vaddr_end) {
base_pgsize = vaddr_end - vaddr;
}
/* Corner case #2: base_pgsize is too big (max request size).
* This will result in an extra virt to phys lookup,
* but should be rare in practice
*/
if (len + base_pgsize > MAX_EXPECTED_BUFFER) {
size_t extra = len + base_pgsize - MAX_EXPECTED_BUFFER;
phys -= extra;
vaddr -= extra;
base_pgsize -= extra;
}
phys_prev = phys;
vaddr_prev = vaddr;
len += base_pgsize;
dkprintf("phys 0x%llx, vaddr 0x%llx, len %zu, base_pgsize %zu\n",
phys, vaddr, len, base_pgsize);
}
/* Register whatever is left */
ret = program_rcvarray(fd, phys_start, len,
tidlist + tididx);
if (ret <= 0) {
kprintf("Failed to program RcvArray entries: %d\n",
ret);
ret = -EFAULT;
}
tididx += ret;
if (grp->used == grp->size)
tid_group_add_tail(grp, &uctxt->tid_full_list);
else
tid_group_add_tail(grp, &uctxt->tid_used_list);
unlock:
spin_unlock(&fd->tid_lock);
phys += 1 << (order+12);
if (ret < 0)
break;
}
if (ret > 0) {
// TODO: can we use spin_lock with kernel locks?
spin_lock(&fd->tid_lock);
fd->tid_used += tididx;
spin_unlock(&fd->tid_lock);
tinfo->tidcnt = tididx;
if (copy_to_user((void __user *)(unsigned long)tinfo->tidlist,
tid, sizeof(tid)*tididx)) {
tidlist, sizeof(*tidlist)*tididx)) {
/*
* On failure to copy to the user level, we need to undo
* everything done so far so we don't leak resources.
*/
tinfo->tidlist = (unsigned long)&tid;
tinfo->tidlist = (unsigned long)&tidlist;
hfi1_user_exp_rcv_clear(fd, tinfo);
tinfo->tidlist = 0;
ret = -EFAULT;
}
}
kfree(tidlist);
return ret > 0 ? 0 : ret;
}
@ -239,17 +219,16 @@ int hfi1_user_exp_rcv_clear(struct hfi1_filedata *fd, struct hfi1_tid_info *tinf
goto done;
}
spin_lock(&fd->tid_lock);
dkprintf("Clear called, cnt %d\n", tinfo->tidcnt);
for (tididx = 0; tididx < tinfo->tidcnt; tididx++) {
ret = unprogram_rcvarray(fd, tidinfo[tididx], NULL);
if (ret) {
hfi1_cdbg(TID, "Failed to unprogram rcv array %d",
kprintf("Failed to unprogram rcv array %d\n",
ret);
break;
}
}
fd->tid_used -= tididx;
spin_unlock(&fd->tid_lock);
tinfo->tidcnt = tididx;
done:
kfree(tidinfo);
@ -258,55 +237,52 @@ done:
/**
* program_rcvarray() - program an RcvArray group with receive buffers
* @fd: file data
* @vaddr: starting user virtual address
* @grp: RcvArray group
* @sets: array of struct tid_pageset holding information on physically
* contiguous chunks from the user buffer
* @start: starting index into sets array
* @count: number of struct tid_pageset's to program
* @pages: an array of struct page * for the user buffer
* @ptid: information about the programmed RcvArray entries is to be encoded.
* @tididx: starting offset into tidlist
*
* This function will program up to 'count' number of RcvArray entries from the
* group 'grp'. To make best use of write-combining writes, the function will
* perform writes to the unused RcvArray entries which will be ignored by the
* HW. Each RcvArray entry will be programmed with a physically contiguous
* buffer chunk from the user's virtual buffer.
*
* Return:
* -EINVAL if the requested count is larger than the size of the group,
* -ENOMEM or -EFAULT on error from set_rcvarray_entry(), or
* number of RcvArray entries programmed.
*/
static int program_rcvarray(struct hfi1_filedata *fd, uintptr_t phys,
u16 order,
struct tid_group *grp,
u32 *ptid)
size_t len, u32 *ptid)
{
struct hfi1_ctxtdata *uctxt = fd->uctxt;
struct hfi1_devdata *dd = uctxt->dd;
u16 idx;
u16 idx = 0;
s16 order;
u32 tidinfo = 0, rcventry;
int ret = -ENOMEM, count = 0;
struct tid_group *grp = NULL;
/* lock is taken at loop edges */
spin_lock(&fd->tid_lock);
while (len > 0) {
if (!grp) {
if (!uctxt->tid_used_list.count) {
if (!uctxt->tid_group_list.count) {
spin_unlock(&fd->tid_lock);
/* return what we have so far */
return count ? count : -ENOMEM;
}
grp = tid_group_pop(&uctxt->tid_group_list);
} else {
grp = tid_group_pop(&uctxt->tid_used_list);
}
}
/* Find the first unused entry in the group */
for (idx = 0; idx < grp->size; idx++) {
for (; idx < grp->size; idx++) {
if (!(grp->map & (1 << idx))) {
break;
}
}
spin_unlock(&fd->tid_lock);
int ret = 0;
/*
* If this entry in the group is used, move to the next one.
* If we go past the end of the group, exit the loop.
*/
/* order is power of two of 4k (2^12) pages */
order = fls(len) - 13;
if (order < 0)
order = 0;
dkprintf("len %u, order %u\n", len, order);
rcv_array_wc_fill(dd, grp->base + idx);
rcventry = grp->base + idx;
ret = set_rcvarray_entry(fd, phys, rcventry, grp,
order);
if (ret)
@ -314,11 +290,29 @@ static int program_rcvarray(struct hfi1_filedata *fd, uintptr_t phys,
tidinfo = rcventry2tidinfo(rcventry - uctxt->expected_base) |
EXP_TID_SET(LEN, 1 << order);
*ptid = tidinfo;
ptid[count++] = tidinfo;
len -= 1 << (order + 12);
phys += 1 << (order + 12);
spin_lock(&fd->tid_lock);
grp->used++;
grp->map |= 1 << idx++;
return 1;
/* optimization: keep same group if possible. */
if (grp->used < grp->size && len > 0)
continue;
if (grp->used == grp->size)
tid_group_add_tail(grp, &uctxt->tid_full_list);
else
tid_group_add_tail(grp, &uctxt->tid_used_list);
idx = 0;
grp = NULL;
}
spin_unlock(&fd->tid_lock);
return count;
}
static int set_rcvarray_entry(struct hfi1_filedata *fd, uintptr_t phys,
@ -337,7 +331,8 @@ static int set_rcvarray_entry(struct hfi1_filedata *fd, uintptr_t phys,
if (!node)
return -ENOMEM;
kprintf("Registering rcventry %d, phys 0x%p, len %u\n", rcventry, phys, 1 << (order+12));
dkprintf("Registering rcventry %d, phys 0x%p, len %u\n", rcventry,
phys, 1 << (order+12));
node->phys = phys;
node->len = 1 << (order+12);
@ -370,19 +365,24 @@ static int unprogram_rcvarray(struct hfi1_filedata *fd, u32 tidinfo,
return -EINVAL;
}
if (tidctrl == 0x3)
if (tidctrl == 0x3) {
kprintf("tidctrl = 3 for rcventry %d\n",
tididx + 2 + uctxt->expected_base);
return -EINVAL;
}
rcventry = tididx + (tidctrl - 1);
node = fd->entry_to_rb[rcventry];
if (!node || node->rcventry != (uctxt->expected_base + rcventry))
if (!node || node->rcventry != (uctxt->expected_base + rcventry)) {
kprintf("bad entry %d\n", rcventry);
return -EBADF;
}
if (grp)
*grp = node->grp;
kprintf("Clearing rcventry %d, phys 0x%p, len %u\n", node->rcventry,
dkprintf("Clearing rcventry %d, phys 0x%p, len %u\n", node->rcventry,
node->phys, node->len);
fd->entry_to_rb[rcventry] = NULL;
@ -396,11 +396,6 @@ static void clear_tid_node(struct hfi1_filedata *fd, struct tid_rb_node *node)
struct hfi1_ctxtdata *uctxt = fd->uctxt;
struct hfi1_devdata *dd = uctxt->dd;
#if 0
trace_hfi1_exp_tid_unreg(uctxt->ctxt, fd->subctxt, node->rcventry,
node->npages, node->mmu.addr, node->phys,
node->dma_addr);
#endif
hfi1_put_tid(dd, node->rcventry, PT_INVALID, 0, 0);
/*
@ -409,14 +404,7 @@ static void clear_tid_node(struct hfi1_filedata *fd, struct tid_rb_node *node)
*/
flush_wc();
#if 0
pci_unmap_single(dd->pcidev, node->dma_addr, node->mmu.len,
PCI_DMA_FROMDEVICE);
hfi1_release_user_pages(fd->mm, node->pages, node->npages, true);
fd->tid_n_pinned -= node->npages;
#endif
spin_lock(&fd->tid_lock);
node->grp->used--;
node->grp->map &= ~(1 << (node->rcventry - node->grp->base));
@ -426,5 +414,6 @@ static void clear_tid_node(struct hfi1_filedata *fd, struct tid_rb_node *node)
else if (!node->grp->used)
tid_group_move(node->grp, &uctxt->tid_used_list,
&uctxt->tid_group_list);
spin_unlock(&fd->tid_lock);
kfree(node);
}