aboutsummaryrefslogtreecommitdiff
path: root/drivers/infiniband/hw/ipath/ipath_file_ops.c
diff options
context:
space:
mode:
authorBryan O'Sullivan <bos@pathscale.com>2006-07-01 04:36:03 -0700
committerLinus Torvalds <torvalds@g5.osdl.org>2006-07-01 09:56:00 -0700
commitf37bda92461313ad3bbfbf5660adc849c69718bf (patch)
treea5fe4737ca6b8fcbe2cf9b58466d6340ee12fe56 /drivers/infiniband/hw/ipath/ipath_file_ops.c
parent06993ca6bc46419027b45198a58447f4f05c14f6 (diff)
[PATCH] IB/ipath: memory management cleanups
Made in-memory rcvhdrq tail update be in dma_alloc'ed memory, not random user or special kernel (needed for ppc, also "just the right thing to do"). Some cleanups to make unexpected link transitions less likely to produce complaints about packet errors, and also to not leave SMA packets stuck and unable to go out. A few other random debug and comment cleanups. Always init rcvhdrq head/tail registers to 0, to avoid race conditions (should have been that way some time ago). Signed-off-by: Dave Olson <dave.olson@qlogic.com> Signed-off-by: Bryan O'Sullivan <bryan.osullivan@qlogic.com> Cc: "Michael S. Tsirkin" <mst@mellanox.co.il> Cc: Roland Dreier <rolandd@cisco.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Diffstat (limited to 'drivers/infiniband/hw/ipath/ipath_file_ops.c')
-rw-r--r--drivers/infiniband/hw/ipath/ipath_file_ops.c290
1 files changed, 115 insertions, 175 deletions
diff --git a/drivers/infiniband/hw/ipath/ipath_file_ops.c b/drivers/infiniband/hw/ipath/ipath_file_ops.c
index 03689dbe1a9..e89d3a17acd 100644
--- a/drivers/infiniband/hw/ipath/ipath_file_ops.c
+++ b/drivers/infiniband/hw/ipath/ipath_file_ops.c
@@ -123,6 +123,7 @@ static int ipath_get_base_info(struct ipath_portdata *pd,
* on to yet another method of dealing with this
*/
kinfo->spi_rcvhdr_base = (u64) pd->port_rcvhdrq_phys;
+ kinfo->spi_rcvhdr_tailaddr = (u64)pd->port_rcvhdrqtailaddr_phys;
kinfo->spi_rcv_egrbufs = (u64) pd->port_rcvegr_phys;
kinfo->spi_pioavailaddr = (u64) dd->ipath_pioavailregs_phys;
kinfo->spi_status = (u64) kinfo->spi_pioavailaddr +
@@ -785,11 +786,12 @@ static int ipath_create_user_egr(struct ipath_portdata *pd)
bail_rcvegrbuf_phys:
for (e = 0; e < pd->port_rcvegrbuf_chunks &&
- pd->port_rcvegrbuf[e]; e++)
+ pd->port_rcvegrbuf[e]; e++) {
dma_free_coherent(&dd->pcidev->dev, size,
pd->port_rcvegrbuf[e],
pd->port_rcvegrbuf_phys[e]);
+ }
vfree(pd->port_rcvegrbuf_phys);
pd->port_rcvegrbuf_phys = NULL;
bail_rcvegrbuf:
@@ -804,10 +806,7 @@ static int ipath_do_user_init(struct ipath_portdata *pd,
{
int ret = 0;
struct ipath_devdata *dd = pd->port_dd;
- u64 physaddr, uaddr, off, atmp;
- struct page *pagep;
u32 head32;
- u64 head;
/* for now, if major version is different, bail */
if ((uinfo->spu_userversion >> 16) != IPATH_USER_SWMAJOR) {
@@ -832,54 +831,6 @@ static int ipath_do_user_init(struct ipath_portdata *pd,
/* for now we do nothing with rcvhdrcnt: uinfo->spu_rcvhdrcnt */
- /* set up for the rcvhdr Q tail register writeback to user memory */
- if (!uinfo->spu_rcvhdraddr ||
- !access_ok(VERIFY_WRITE, (u64 __user *) (unsigned long)
- uinfo->spu_rcvhdraddr, sizeof(u64))) {
- ipath_dbg("Port %d rcvhdrtail addr %llx not valid\n",
- pd->port_port,
- (unsigned long long) uinfo->spu_rcvhdraddr);
- ret = -EINVAL;
- goto done;
- }
-
- off = offset_in_page(uinfo->spu_rcvhdraddr);
- uaddr = PAGE_MASK & (unsigned long) uinfo->spu_rcvhdraddr;
- ret = ipath_get_user_pages_nocopy(uaddr, &pagep);
- if (ret) {
- dev_info(&dd->pcidev->dev, "Failed to lookup and lock "
- "address %llx for rcvhdrtail: errno %d\n",
- (unsigned long long) uinfo->spu_rcvhdraddr, -ret);
- goto done;
- }
- ipath_stats.sps_pagelocks++;
- pd->port_rcvhdrtail_uaddr = uaddr;
- pd->port_rcvhdrtail_pagep = pagep;
- pd->port_rcvhdrtail_kvaddr =
- page_address(pagep);
- pd->port_rcvhdrtail_kvaddr += off;
- physaddr = page_to_phys(pagep) + off;
- ipath_cdbg(VERBOSE, "port %d user addr %llx hdrtailaddr, %llx "
- "physical (off=%llx)\n",
- pd->port_port,
- (unsigned long long) uinfo->spu_rcvhdraddr,
- (unsigned long long) physaddr, (unsigned long long) off);
- ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdrtailaddr,
- pd->port_port, physaddr);
- atmp = ipath_read_kreg64_port(dd,
- dd->ipath_kregs->kr_rcvhdrtailaddr,
- pd->port_port);
- if (physaddr != atmp) {
- ipath_dev_err(dd,
- "Catastrophic software error, "
- "RcvHdrTailAddr%u written as %llx, "
- "read back as %llx\n", pd->port_port,
- (unsigned long long) physaddr,
- (unsigned long long) atmp);
- ret = -EINVAL;
- goto done;
- }
-
/* for right now, kernel piobufs are at end, so port 1 is at 0 */
pd->port_piobufs = dd->ipath_piobufbase +
dd->ipath_pbufsport * (pd->port_port -
@@ -898,26 +849,18 @@ static int ipath_do_user_init(struct ipath_portdata *pd,
ret = ipath_create_user_egr(pd);
if (ret)
goto done;
- /* enable receives now */
- /* atomically set enable bit for this port */
- set_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
- &dd->ipath_rcvctrl);
/*
- * set the head registers for this port to the current values
+ * set the eager head register for this port to the current values
* of the tail pointers, since we don't know if they were
* updated on last use of the port.
*/
- head32 = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
- head = (u64) head32;
- ipath_write_ureg(dd, ur_rcvhdrhead, head, pd->port_port);
head32 = ipath_read_ureg32(dd, ur_rcvegrindextail, pd->port_port);
ipath_write_ureg(dd, ur_rcvegrindexhead, head32, pd->port_port);
dd->ipath_lastegrheads[pd->port_port] = -1;
dd->ipath_lastrcvhdrqtails[pd->port_port] = -1;
- ipath_cdbg(VERBOSE, "Wrote port%d head %llx, egrhead %x from "
- "tail regs\n", pd->port_port,
- (unsigned long long) head, head32);
+ ipath_cdbg(VERBOSE, "Wrote port%d egrhead %x from tail regs\n",
+ pd->port_port, head32);
pd->port_tidcursor = 0; /* start at beginning after open */
/*
* now enable the port; the tail registers will be written to memory
@@ -926,24 +869,76 @@ static int ipath_do_user_init(struct ipath_portdata *pd,
* transition from 0 to 1, so clear it first, then set it as part of
* enabling the port. This will (very briefly) affect any other
* open ports, but it shouldn't be long enough to be an issue.
+ * We explictly set the in-memory copy to 0 beforehand, so we don't
+ * have to wait to be sure the DMA update has happened.
*/
+ *pd->port_rcvhdrtail_kvaddr = 0ULL;
+ set_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
+ &dd->ipath_rcvctrl);
ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
dd->ipath_rcvctrl & ~INFINIPATH_R_TAILUPD);
ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
dd->ipath_rcvctrl);
-
done:
return ret;
}
+
+/* common code for the mappings on dma_alloc_coherent mem */
+static int ipath_mmap_mem(struct vm_area_struct *vma,
+ struct ipath_portdata *pd, unsigned len,
+ int write_ok, dma_addr_t addr, char *what)
+{
+ struct ipath_devdata *dd = pd->port_dd;
+ unsigned pfn = (unsigned long)addr >> PAGE_SHIFT;
+ int ret;
+
+ if ((vma->vm_end - vma->vm_start) > len) {
+ dev_info(&dd->pcidev->dev,
+ "FAIL on %s: len %lx > %x\n", what,
+ vma->vm_end - vma->vm_start, len);
+ ret = -EFAULT;
+ goto bail;
+ }
+
+ if (!write_ok) {
+ if (vma->vm_flags & VM_WRITE) {
+ dev_info(&dd->pcidev->dev,
+ "%s must be mapped readonly\n", what);
+ ret = -EPERM;
+ goto bail;
+ }
+
+ /* don't allow them to later change with mprotect */
+ vma->vm_flags &= ~VM_MAYWRITE;
+ }
+
+ ret = remap_pfn_range(vma, vma->vm_start, pfn,
+ len, vma->vm_page_prot);
+ if (ret)
+ dev_info(&dd->pcidev->dev,
+ "%s port%u mmap of %lx, %x bytes r%c failed: %d\n",
+ what, pd->port_port, (unsigned long)addr, len,
+ write_ok?'w':'o', ret);
+ else
+ ipath_cdbg(VERBOSE, "%s port%u mmaped %lx, %x bytes r%c\n",
+ what, pd->port_port, (unsigned long)addr, len,
+ write_ok?'w':'o');
+bail:
+ return ret;
+}
+
static int mmap_ureg(struct vm_area_struct *vma, struct ipath_devdata *dd,
u64 ureg)
{
unsigned long phys;
int ret;
- /* it's the real hardware, so io_remap works */
-
+ /*
+ * This is real hardware, so use io_remap. This is the mechanism
+ * for the user process to update the head registers for their port
+ * in the chip.
+ */
if ((vma->vm_end - vma->vm_start) > PAGE_SIZE) {
dev_info(&dd->pcidev->dev, "FAIL mmap userreg: reqlen "
"%lx > PAGE\n", vma->vm_end - vma->vm_start);
@@ -969,10 +964,11 @@ static int mmap_piobufs(struct vm_area_struct *vma,
int ret;
/*
- * When we map the PIO buffers, we want to map them as writeonly, no
- * read possible.
+ * When we map the PIO buffers in the chip, we want to map them as
+ * writeonly, no read possible. This prevents access to previous
+ * process data, and catches users who might try to read the i/o
+ * space due to a bug.
*/
-
if ((vma->vm_end - vma->vm_start) >
(dd->ipath_pbufsport * dd->ipath_palign)) {
dev_info(&dd->pcidev->dev, "FAIL mmap piobufs: "
@@ -983,11 +979,10 @@ static int mmap_piobufs(struct vm_area_struct *vma,
}
phys = dd->ipath_physaddr + pd->port_piobufs;
+
/*
- * Do *NOT* mark this as non-cached (PWT bit), or we don't get the
+ * Don't mark this as non-cached, or we don't get the
* write combining behavior we want on the PIO buffers!
- * vma->vm_page_prot =
- * pgprot_noncached(vma->vm_page_prot);
*/
if (vma->vm_flags & VM_READ) {
@@ -999,8 +994,7 @@ static int mmap_piobufs(struct vm_area_struct *vma,
}
/* don't allow them to later change to readable with mprotect */
-
- vma->vm_flags &= ~VM_MAYWRITE;
+ vma->vm_flags &= ~VM_MAYREAD;
vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT,
@@ -1019,11 +1013,6 @@ static int mmap_rcvegrbufs(struct vm_area_struct *vma,
dma_addr_t *phys;
int ret;
- if (!pd->port_rcvegrbuf) {
- ret = -EFAULT;
- goto bail;
- }
-
size = pd->port_rcvegrbuf_size;
total_size = pd->port_rcvegrbuf_chunks * size;
if ((vma->vm_end - vma->vm_start) > total_size) {
@@ -1041,13 +1030,12 @@ static int mmap_rcvegrbufs(struct vm_area_struct *vma,
ret = -EPERM;
goto bail;
}
+ /* don't allow them to later change to writeable with mprotect */
+ vma->vm_flags &= ~VM_MAYWRITE;
start = vma->vm_start;
phys = pd->port_rcvegrbuf_phys;
- /* don't allow them to later change to writeable with mprotect */
- vma->vm_flags &= ~VM_MAYWRITE;
-
for (i = 0; i < pd->port_rcvegrbuf_chunks; i++, start += size) {
ret = remap_pfn_range(vma, start, phys[i] >> PAGE_SHIFT,
size, vma->vm_page_prot);
@@ -1060,78 +1048,6 @@ bail:
return ret;
}
-static int mmap_rcvhdrq(struct vm_area_struct *vma,
- struct ipath_portdata *pd)
-{
- struct ipath_devdata *dd = pd->port_dd;
- size_t total_size;
- int ret;
-
- /*
- * kmalloc'ed memory, physically contiguous; this is from
- * spi_rcvhdr_base; we allow user to map read-write so they can
- * write hdrq entries to allow protocol code to directly poll
- * whether a hdrq entry has been written.
- */
- total_size = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
- sizeof(u32), PAGE_SIZE);
- if ((vma->vm_end - vma->vm_start) > total_size) {
- dev_info(&dd->pcidev->dev,
- "FAIL on rcvhdrq: reqlen %lx > actual %lx\n",
- vma->vm_end - vma->vm_start,
- (unsigned long) total_size);
- ret = -EFAULT;
- goto bail;
- }
-
- ret = remap_pfn_range(vma, vma->vm_start,
- pd->port_rcvhdrq_phys >> PAGE_SHIFT,
- vma->vm_end - vma->vm_start,
- vma->vm_page_prot);
-bail:
- return ret;
-}
-
-static int mmap_pioavailregs(struct vm_area_struct *vma,
- struct ipath_portdata *pd)
-{
- struct ipath_devdata *dd = pd->port_dd;
- int ret;
-
- /*
- * when we map the PIO bufferavail registers, we want to map them as
- * readonly, no write possible.
- *
- * kmalloc'ed memory, physically contiguous, one page only, readonly
- */
-
- if ((vma->vm_end - vma->vm_start) > PAGE_SIZE) {
- dev_info(&dd->pcidev->dev, "FAIL on pioavailregs_dma: "
- "reqlen %lx > actual %lx\n",
- vma->vm_end - vma->vm_start,
- (unsigned long) PAGE_SIZE);
- ret = -EFAULT;
- goto bail;
- }
-
- if (vma->vm_flags & VM_WRITE) {
- dev_info(&dd->pcidev->dev,
- "Can't map pioavailregs as writable (flags=%lx)\n",
- vma->vm_flags);
- ret = -EPERM;
- goto bail;
- }
-
- /* don't allow them to later change with mprotect */
- vma->vm_flags &= ~VM_MAYWRITE;
-
- ret = remap_pfn_range(vma, vma->vm_start,
- dd->ipath_pioavailregs_phys >> PAGE_SHIFT,
- PAGE_SIZE, vma->vm_page_prot);
-bail:
- return ret;
-}
-
/**
* ipath_mmap - mmap various structures into user space
* @fp: the file pointer
@@ -1151,6 +1067,7 @@ static int ipath_mmap(struct file *fp, struct vm_area_struct *vma)
pd = port_fp(fp);
dd = pd->port_dd;
+
/*
* This is the ipath_do_user_init() code, mapping the shared buffers
* into the user process. The address referred to by vm_pgoff is the
@@ -1160,29 +1077,59 @@ static int ipath_mmap(struct file *fp, struct vm_area_struct *vma)
pgaddr = vma->vm_pgoff << PAGE_SHIFT;
/*
- * note that ureg does *NOT* have the kregvirt as part of it, to be
- * sure that for 32 bit programs, we don't end up trying to map a >
- * 44 address. Has to match ipath_get_base_info() code that sets
- * __spi_uregbase
+ * Must fit in 40 bits for our hardware; some checked elsewhere,
+ * but we'll be paranoid. Check for 0 is mostly in case one of the
+ * allocations failed, but user called mmap anyway. We want to catch
+ * that before it can match.
*/
+ if (!pgaddr || pgaddr >= (1ULL<<40)) {
+ ipath_dev_err(dd, "Bad phys addr %llx, start %lx, end %lx\n",
+ (unsigned long long)pgaddr, vma->vm_start, vma->vm_end);
+ return -EINVAL;
+ }
+ /* just the offset of the port user registers, not physical addr */
ureg = dd->ipath_uregbase + dd->ipath_palign * pd->port_port;
- ipath_cdbg(MM, "pgaddr %llx vm_start=%lx len %lx port %u:%u\n",
+ ipath_cdbg(MM, "ushare: pgaddr %llx vm_start=%lx, vmlen %lx\n",
(unsigned long long) pgaddr, vma->vm_start,
- vma->vm_end - vma->vm_start, dd->ipath_unit,
- pd->port_port);
+ vma->vm_end - vma->vm_start);
- if (pgaddr == ureg)
+ if (vma->vm_start & (PAGE_SIZE-1)) {
+ ipath_dev_err(dd,
+ "vm_start not aligned: %lx, end=%lx phys %lx\n",
+ vma->vm_start, vma->vm_end, (unsigned long)pgaddr);
+ ret = -EINVAL;
+ }
+ else if (pgaddr == ureg)
ret = mmap_ureg(vma, dd, ureg);
else if (pgaddr == pd->port_piobufs)
ret = mmap_piobufs(vma, dd, pd);
else if (pgaddr == (u64) pd->port_rcvegr_phys)
ret = mmap_rcvegrbufs(vma, pd);
- else if (pgaddr == (u64) pd->port_rcvhdrq_phys)
- ret = mmap_rcvhdrq(vma, pd);
+ else if (pgaddr == (u64) pd->port_rcvhdrq_phys) {
+ /*
+ * The rcvhdrq itself; readonly except on HT-400 (so have
+ * to allow writable mapping), multiple pages, contiguous
+ * from an i/o perspective.
+ */
+ unsigned total_size =
+ ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize
+ * sizeof(u32), PAGE_SIZE);
+ ret = ipath_mmap_mem(vma, pd, total_size, 1,
+ pd->port_rcvhdrq_phys,
+ "rcvhdrq");
+ }
+ else if (pgaddr == (u64)pd->port_rcvhdrqtailaddr_phys)
+ /* in-memory copy of rcvhdrq tail register */
+ ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
+ pd->port_rcvhdrqtailaddr_phys,
+ "rcvhdrq tail");
else if (pgaddr == dd->ipath_pioavailregs_phys)
- ret = mmap_pioavailregs(vma, pd);
+ /* in-memory copy of pioavail registers */
+ ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
+ dd->ipath_pioavailregs_phys,
+ "pioavail registers");
else
ret = -EINVAL;
@@ -1539,14 +1486,6 @@ static int ipath_close(struct inode *in, struct file *fp)
}
if (dd->ipath_kregbase) {
- if (pd->port_rcvhdrtail_uaddr) {
- pd->port_rcvhdrtail_uaddr = 0;
- pd->port_rcvhdrtail_kvaddr = NULL;
- ipath_release_user_pages_on_close(
- &pd->port_rcvhdrtail_pagep, 1);
- pd->port_rcvhdrtail_pagep = NULL;
- ipath_stats.sps_pageunlocks++;
- }
ipath_write_kreg_port(
dd, dd->ipath_kregs->kr_rcvhdrtailaddr,
port, 0ULL);
@@ -1583,9 +1522,9 @@ static int ipath_close(struct inode *in, struct file *fp)
dd->ipath_f_clear_tids(dd, pd->port_port);
- ipath_free_pddata(dd, pd->port_port, 0);
-
+ dd->ipath_pd[pd->port_port] = NULL; /* before releasing mutex */
mutex_unlock(&ipath_mutex);
+ ipath_free_pddata(dd, pd); /* after releasing the mutex */
return ret;
}
@@ -1905,3 +1844,4 @@ void ipath_user_remove(struct ipath_devdata *dd)
bail:
return;
}
+