[7823] in linux-scsi channel archive
SCSI queueing status.
daemon@ATHENA.MIT.EDU (Eric Youngdale)
Thu Jan 13 12:59:04 2000
Message-ID: <007601bf5d81$3a0d55a0$2017a8c0@eric.home>
From: "Eric Youngdale" <eric@andante.org>
To: <linux-scsi@vger.rutgers.edu>
Date: Wed, 12 Jan 2000 23:46:59 -0500
MIME-Version: 1.0
Content-Type: multipart/mixed;
boundary="----=_NextPart_000_0072_01BF5D57.511A9DE0"
This is a multi-part message in MIME format.
------=_NextPart_000_0072_01BF5D57.511A9DE0
Content-Type: multipart/alternative;
boundary="----=_NextPart_001_0073_01BF5D57.511A9DE0"
------=_NextPart_001_0073_01BF5D57.511A9DE0
Content-Type: text/plain;
charset="iso-8859-1"
Content-Transfer-Encoding: quoted-printable
Just thought I would drop a note and tell people where things stand.
For the most part, things have been kind of quiet lately. I am =
trying to polish off a few remaining issues. I think I have gdth =
licked, and I think I have the host blocking feature added back again, =
and at this point I am waiting for the people testing. In the =
background I am starting to clean up a few other things - I have split =
part of scsi.c into a new file (and in the long run, there will probably =
be several others as well). Unfortunately scsi.c has served as sort of =
a repository for a bunch of loosely related things, and both readability =
and maintainability have suffered. I am appending the current patchset, =
in the event that anyone else wants to look at it.
I was talking with someone about how to implement automatic spinup =
of spun down drives, and I realized that there was a design flaw in the =
new queueing code. Basically as things stand currently we pre-allocate =
Scsi_Cmnd structures based upon the queue depth for the device. For a =
disk with a queue depth of 1 there will be only one Scsi_Cmnd structure =
dedicated to this device. When a user does something which would cause =
an ioctl to be generated, the Scsi_Cmnd structure is allocated in the =
ioctl code, and then passed via scsi_do_cmd() or scsi_wait_cmd() into =
the mid-layer. As it turns out, this will be inserted into the queue at =
the end.
The queue handler function currently only looks at the head of the =
queue. It keeps queueing things as long as resources are available to =
queue more commands. One of the resources that gets used up of course =
is the supply of Scsi_Cmnd structures. It is basically assumed that all =
of the structures that are allocated are essentially queued to the =
device, and if we run out, the queue handler can simply return without =
doing anything. The idea is that when one of the commands that is =
currently running completes that we would make another stab at queueing =
the next request at the head of the queue.
Consider a disk that is fairly active, and with a queue depth of 1. =
Imagine that a user issues an ioctl - this will allocate the only =
Scsi_Cmnd structure, and drop it at the tail of the queue. The =
scsi_request_fn() function will look at the head of the queue, and be =
unable to queue that request because there are no Scsi_Cmnd structures =
available, and return. Hence a sort of deadlock will arise.
My guess at the moment is that this situation isn't all that common =
- nobody has reported it yet, but it is only a matter of time before =
somebody stumbles across it. The reason it hasn't come up is that it =
would only appear for a disk or a cdrom that is fairly busy, has a small =
queue depth, and for which an ioctl is issued. The trick of course is =
to figure out how to fix it. =20
In the short term, I could easily add a simple hack to detect this =
situation and recover from it. No big deal here, and until a more =
architecturally satisfying solution arises, this will do.
In the longer term, there are several things that come to mind. =
First of all, upper level drivers probably don't actually need a =
Scsi_Cmnd structure for anything except for completion processing. As =
things currently stand, an upper level driver simply allocates the =
thing, and then calls scsi_do_cmd/scsi_wait_cmd without having done much =
to the Scsi_Cmnd structure at all. Once the command is complete, then =
the upper level driver usually releases it after looking at the status =
code. There are a handful of things that the upper level driver =
actually needs - usually just the completion code and in some cases the =
sense data in the event that something went wrong. I could come up with =
something analagous to scsi_do_cmd or scsi_wait_cmd that doesn't even =
take a Scsi_Cmnd structure, but at this point I don't know to what =
degree it is possible to replace scsi_do_cmd/scsi_wait_cmd. If a 100% =
replacement is feasible, then this would be a complete solution.
The second thought that comes to mind is that we really need to get =
away from pre-allocating Scsi_Cmnd structures for devices. There should =
be a general purpose pool allocated for the host instead (to prevent =
starvation, there might be one Scsi_Cmnd pre-allocated per device, and =
then the rest come from a general purpose pool, but that's just an =
idea). The general idea is that managing the queue depth would be done =
at a different level, so that it is less likely that we get into the =
trouble I described above. While we do have a known problem of =
overallocating Scsi_Cmnd structures, I don't like the idea of =
allocating consumables if they are not needed right away, and this would =
be more of a band-aid.
The last possibility is that for the case of things like ioctl it is =
*possible* that we could allocate an additional Scsi_Cmnd from the =
general kernel memory pool and hook it in. In general you don't want to =
call kmalloc() anywhere that is in the code path during I/O operations =
that might take place during swapping, but ioctl generally won't fall in =
this category. I am not that wild about this one.
As things stand I am still thinking about it. A quickie hack type =
of workaround certainly can be done to prevent any user-visible =
symptoms, but at this point I am trying to decide what would make the =
most sense from an architectural point of view. I won't bother to code =
up the workaround unless people start to experience this problem in the =
field. Unless I am hit with a brilliant flash of insight, I will =
probably play with the first possibility first and see how much trouble =
it gets me into.
-Eric
------=_NextPart_001_0073_01BF5D57.511A9DE0
Content-Type: text/html;
charset="iso-8859-1"
Content-Transfer-Encoding: quoted-printable
<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.0 Transitional//EN">
<HTML><HEAD>
<META content=3D"text/html; charset=3Diso-8859-1" =
http-equiv=3DContent-Type>
<META content=3D"MSHTML 5.00.2919.6307" name=3DGENERATOR>
<STYLE></STYLE>
</HEAD>
<BODY bgColor=3D#ffffff>
<DIV><FONT face=3DArial size=3D2> Just thought I would =
drop a note=20
and tell people where things stand.</FONT></DIV>
<DIV> </DIV>
<DIV><FONT face=3DArial size=3D2> For the most part, =
things have=20
been kind of quiet lately. I am trying to polish off a few =
remaining=20
issues. I think I have gdth licked, and I think I have the host =
blocking=20
feature added back again, and at this point I am waiting for the people=20
testing. In the background I am starting to clean up a few =
other=20
things - I have split part of scsi.c into a new file (and in the long =
run, there=20
will probably be several others as well). Unfortunately scsi.c has =
served=20
as sort of a repository for a bunch of loosely related things, and =
both=20
readability and maintainability have suffered. I am appending the =
current=20
patchset, in the event that anyone else wants to look at =
it.</FONT></DIV>
<DIV> </DIV>
<DIV><FONT face=3DArial size=3D2> I was talking with =
someone about=20
how to implement automatic spinup of spun down drives, and I realized =
that there=20
was a design flaw in the new queueing code. Basically as things =
stand=20
currently we pre-allocate Scsi_Cmnd structures based upon the queue =
depth for=20
the device. For a disk with a queue depth of 1 there =
will be=20
only one Scsi_Cmnd structure dedicated to this device. When a user =
does=20
something which would cause an ioctl to be generated, the Scsi_Cmnd =
structure is=20
allocated in the ioctl code, and then passed via scsi_do_cmd() or=20
scsi_wait_cmd() into the mid-layer. As it turns out, this will be =
inserted=20
into the queue at the end.</FONT></DIV>
<DIV> </DIV>
<DIV><FONT face=3DArial size=3D2> The queue handler =
function=20
currently only looks at the head of the queue. It keeps queueing =
things as=20
long as resources are available to queue more commands. One of the =
resources that gets used up of course is the supply of Scsi_Cmnd=20
structures. It is basically assumed that all of the structures =
that are=20
allocated are essentially queued to the device, and if we run out, the =
queue=20
handler can simply return without doing anything. The idea is that =
when=20
one of the commands that is currently running completes that we =
would make=20
another stab at queueing the next request at the head of the =
queue.</FONT></DIV>
<DIV> </DIV>
<DIV><FONT face=3DArial size=3D2> Consider a disk that =
is fairly=20
active, and with a queue depth of 1. Imagine that a user issues an =
ioctl -=20
this will allocate the only Scsi_Cmnd structure, and drop it at the tail =
of the=20
queue. The scsi_request_fn() function will look at the head of the =
queue,=20
and be unable to queue that request because there are no Scsi_Cmnd=20
structures available, and return. Hence a sort of deadlock will=20
arise.</FONT></DIV>
<DIV> </DIV>
<DIV><FONT face=3DArial size=3D2> My guess at the =
moment is that=20
this situation isn't all that common - nobody has reported it yet, but =
it is=20
only a matter of time before somebody stumbles across it. =
The reason=20
it hasn't come up is that it would only appear for a disk or a cdrom =
that is=20
fairly busy, has a small queue depth, and for which an ioctl is =
issued. =20
The trick of course is to figure out how to fix it. </FONT></DIV>
<DIV> </DIV>
<DIV><FONT face=3DArial size=3D2> In the short term, I =
could=20
easily add a simple hack to detect this situation and recover from =
it. No=20
big deal here, and until a more architecturally satisfying solution =
arises, this=20
will do.</FONT></DIV>
<DIV> </DIV>
<DIV><FONT face=3DArial size=3D2> In the longer term, =
there=20
are several things that come to mind. First of all, upper =
level=20
drivers probably don't actually need a Scsi_Cmnd structure for anything =
except=20
for completion processing. As things currently stand, an upper =
level=20
driver simply allocates the thing, and then calls =
scsi_do_cmd/scsi_wait_cmd=20
without having done much to the Scsi_Cmnd structure at all. Once =
the=20
command is complete, then the upper level driver usually releases =
it after=20
looking at the status code. There are a handful of things that the =
upper=20
level driver actually needs - usually just the completion code and in =
some cases=20
the sense data in the event that something went wrong.</FONT><FONT =
face=3DArial=20
size=3D2> I could come up with something analagous to scsi_do_cmd =
or=20
scsi_wait_cmd that doesn't even take a Scsi_Cmnd structure, but at this =
point I=20
don't know to what degree it is possible to replace=20
scsi_do_cmd/scsi_wait_cmd. If a 100% replacement is feasible, then =
this=20
would be a complete solution.</FONT></DIV>
<DIV> </DIV>
<DIV><FONT face=3DArial size=3D2> The second thought =
that comes to=20
mind is that we really need to get away from pre-allocating Scsi_Cmnd =
structures=20
for devices. There should be a general purpose pool allocated for =
the host=20
instead (to prevent starvation, there might be one Scsi_Cmnd =
pre-allocated per=20
device, and then the rest come from a general purpose pool, but that's =
just an=20
idea). The general idea is that managing the queue depth =
would be=20
done at a different level, so that it is less likely that we get into =
the=20
trouble I described above. While we do have a known problem of=20
overallocating Scsi_Cmnd structures, I don't like the idea of =
allocating=20
consumables if they are not needed right away, and this would be more of =
a=20
band-aid.</FONT></DIV>
<DIV><FONT face=3DArial size=3D2></FONT> </DIV>
<DIV><FONT face=3DArial size=3D2> The last =
possibility is=20
that for the case of things like ioctl it is *possible* that we could =
allocate=20
an additional Scsi_Cmnd from the general kernel memory pool and hook it=20
in. In general you don't want to call kmalloc() anywhere that is =
in the=20
code path during I/O operations that might take place during swapping, =
but ioctl=20
generally won't fall in this category. I am not that wild about =
this=20
one.</FONT></DIV>
<DIV> </DIV>
<DIV><FONT face=3DArial size=3D2> As things stand I am =
still=20
thinking about it. A quickie hack type of workaround =
certainly=20
can be done to prevent any user-visible symptoms, but at this point I am =
trying=20
to decide what would make the most sense from an architectural point of=20
view. I won't bother to code up the workaround unless people start =
to=20
experience this problem in the field. Unless I am hit with a =
brilliant flash of insight, I will probably play with the first =
possibility=20
first and see how much trouble it gets me into.</FONT></DIV>
<DIV> </DIV>
<DIV><FONT face=3DArial size=3D2>-Eric</FONT></DIV>
<DIV> </DIV></BODY></HTML>
------=_NextPart_001_0073_01BF5D57.511A9DE0--
------=_NextPart_000_0072_01BF5D57.511A9DE0
Content-Type: application/octet-stream;
name="linux39a.diff"
Content-Transfer-Encoding: quoted-printable
Content-Disposition: attachment;
filename="linux39a.diff"
Index: linux/Documentation/Configure.help
diff -u linux/Documentation/Configure.help:1.1.1.3 =
linux/Documentation/Configure.help:1.2
--- linux/Documentation/Configure.help:1.1.1.3 Fri Jan 7 22:33:08 2000
+++ linux/Documentation/Configure.help Mon Jan 10 21:45:08 2000
@@ -4103,6 +4103,17 @@
so most people can say N here and should in fact do so, because it
is safer.
=20
+Enable host blocking to help with buggy DMA chipsets
+CONFIG_SCSI_HOST_BLOCK
+ Some ISA DMA chipsets are buggy in the sense that if there is more
+ than one ISA DMA busmaster active at the same time that the system
+ becomes unstable. In order to ensure stability, the host block
+ feature was added which ensures that only one such card has active
+ commands at one time. You only need this if you have more than
+ one ISA busmaster on your system - this does not apply to PCI
+ hosts, and it certainly doesn't apply if you only have one
+ SCSI host adapter.
+
Verbose SCSI error reporting (kernel size +=3D12K)
CONFIG_SCSI_CONSTANTS
The error messages regarding your SCSI hardware will be easier to
Index: linux/drivers/scsi/Config.in
diff -u linux/drivers/scsi/Config.in:1.1.1.3 =
linux/drivers/scsi/Config.in:1.2
--- linux/drivers/scsi/Config.in:1.1.1.3 Sat Dec 25 12:58:42 1999
+++ linux/drivers/scsi/Config.in Mon Jan 10 21:44:27 2000
@@ -18,6 +18,7 @@
=20
bool ' Verbose SCSI error reporting (kernel size +=3D12K)' =
CONFIG_SCSI_CONSTANTS
bool ' SCSI logging facility' CONFIG_SCSI_LOGGING
+bool ' Enable host blocking to help with buggy DMA chipsets' =
CONFIG_SCSI_HOST_BLOCK
=20
mainmenu_option next_comment
comment 'SCSI low-level drivers'
Index: linux/drivers/scsi/Makefile
diff -u linux/drivers/scsi/Makefile:1.1.1.4 =
linux/drivers/scsi/Makefile:1.2
--- linux/drivers/scsi/Makefile:1.1.1.4 Thu Jan 6 01:53:07 2000
+++ linux/drivers/scsi/Makefile Mon Jan 10 21:44:27 2000
@@ -41,7 +41,7 @@
endif
L_OBJS +=3D scsi_n_syms.o hosts.o scsi_ioctl.o constants.o scsicam.o
L_OBJS +=3D scsi_error.o scsi_obsolete.o scsi_queue.o scsi_lib.o=20
- L_OBJS +=3D scsi_merge.o scsi_proc.o
+ L_OBJS +=3D scsi_merge.o scsi_proc.o scsi_dma.o
else
ifeq ($(CONFIG_SCSI),m)
MIX_OBJS +=3D scsi_syms.o
@@ -722,10 +722,10 @@
=20
scsi_mod.o: $(MIX_OBJS) hosts.o scsi.o scsi_ioctl.o constants.o \
scsicam.o scsi_proc.o scsi_error.o scsi_obsolete.o \
- scsi_queue.o scsi_lib.o scsi_merge.o
+ scsi_queue.o scsi_lib.o scsi_merge.o scsi_dma.o
$(LD) $(LD_RFLAG) -r -o $@ $(MIX_OBJS) hosts.o scsi.o scsi_ioctl.o \
constants.o scsicam.o scsi_proc.o scsi_merge.o \
- scsi_error.o scsi_obsolete.o scsi_queue.o scsi_lib.o
+ scsi_error.o scsi_obsolete.o scsi_queue.o scsi_lib.o scsi_dma.o
=20
sr_mod.o: sr.o sr_ioctl.o sr_vendor.o
$(LD) $(LD_RFLAG) -r -o $@ sr.o sr_ioctl.o sr_vendor.o
Index: linux/drivers/scsi/gdth.c
diff -u linux/drivers/scsi/gdth.c:1.1.1.1 linux/drivers/scsi/gdth.c:1.2
--- linux/drivers/scsi/gdth.c:1.1.1.1 Mon Jan 3 14:27:56 2000
+++ linux/drivers/scsi/gdth.c Mon Jan 3 18:31:13 2000
@@ -3157,7 +3157,7 @@
NUMDATA(shp)->busnum=3D 0;
=20
ha->pccb =3D CMDDATA(shp);
- ha->pscratch =3D scsi_init_malloc(GDTH_SCRATCH, GFP_ATOMIC =
| GFP_DMA);
+ ha->pscratch =3D (void *) __get_free_pages(GFP_ATOMIC | =
GFP_DMA, GDTH_SCRATCH_ORD);
ha->scratch_busy =3D FALSE;
ha->req_first =3D NULL;
ha->tid_cnt =3D MAX_HDRIVES;
@@ -3172,7 +3172,7 @@
--gdth_ctr_count;
--gdth_ctr_vcount;
if (ha->pscratch !=3D NULL)
- scsi_init_free((void *)ha->pscratch, GDTH_SCRATCH);
+ free_pages((unsigned long)ha->pscratch, =
GDTH_SCRATCH_ORD);
free_irq(ha->irq,NULL);
scsi_unregister(shp);
continue;
@@ -3223,7 +3223,7 @@
NUMDATA(shp)->hanum));
=20
ha->pccb =3D CMDDATA(shp);
- ha->pscratch =3D scsi_init_malloc(GDTH_SCRATCH, GFP_ATOMIC =
| GFP_DMA);
+ ha->pscratch =3D (void *) __get_free_pages(GFP_ATOMIC | =
GFP_DMA, GDTH_SCRATCH_ORD);
ha->scratch_busy =3D FALSE;
ha->req_first =3D NULL;
ha->tid_cnt =3D MAX_HDRIVES;
@@ -3238,7 +3238,7 @@
--gdth_ctr_count;
--gdth_ctr_vcount;
if (ha->pscratch !=3D NULL)
- scsi_init_free((void *)ha->pscratch, GDTH_SCRATCH);
+ free_pages((unsigned long)ha->pscratch, =
GDTH_SCRATCH_ORD);
free_irq(ha->irq,NULL);
scsi_unregister(shp);
continue;
@@ -3293,7 +3293,7 @@
NUMDATA(shp)->busnum=3D 0;
=20
ha->pccb =3D CMDDATA(shp);
- ha->pscratch =3D scsi_init_malloc(GDTH_SCRATCH, GFP_ATOMIC =
| GFP_DMA);
+ ha->pscratch =3D (void *) __get_free_pages(GFP_ATOMIC | =
GFP_DMA, GDTH_SCRATCH_ORD);
ha->scratch_busy =3D FALSE;
ha->req_first =3D NULL;
ha->tid_cnt =3D pcistr[ctr].device_id >=3D 0x200 ? MAXID : =
MAX_HDRIVES;
@@ -3308,7 +3308,7 @@
--gdth_ctr_count;
--gdth_ctr_vcount;
if (ha->pscratch !=3D NULL)
- scsi_init_free((void *)ha->pscratch, GDTH_SCRATCH);
+ free_pages((unsigned long)ha->pscratch, =
GDTH_SCRATCH_ORD);
free_irq(ha->irq,NULL);
scsi_unregister(shp);
continue;
@@ -3359,7 +3359,7 @@
if (shp->dma_channel !=3D 0xff) {
free_dma(shp->dma_channel);
}
- scsi_init_free((void *)ha->pscratch, GDTH_SCRATCH);
+ free_pages((unsigned long)ha->pscratch, GDTH_SCRATCH_ORD);
gdth_ctr_released++;
TRACE2(("gdth_release(): HA %d of %d\n",=20
gdth_ctr_released, gdth_ctr_count));
@@ -3561,22 +3561,19 @@
{
int i;
gdth_ha_str *ha;
- Scsi_Cmnd scp;
- Scsi_Device sdev;
+ Scsi_Cmnd * scp;
+ Scsi_Device * sdev;
gdth_cmd_str gdtcmd;
=20
TRACE2(("gdth_flush() hanum %d\n",hanum));
ha =3D HADATA(gdth_ctr_tab[hanum]);
- memset(&sdev,0,sizeof(Scsi_Device));
- memset(&scp, 0,sizeof(Scsi_Cmnd));
- sdev.host =3D gdth_ctr_tab[hanum];
- sdev.id =3D sdev.host->this_id;
- scp.cmd_len =3D 12;
- scp.host =3D gdth_ctr_tab[hanum];
- scp.target =3D sdev.host->this_id;
- scp.device =3D &sdev;
- scp.use_sg =3D 0;
=20
+ sdev =3D scsi_get_host_dev(gdth_ctr_tab[hanum]);
+ scp =3D scsi_allocate_device(sdev, 1, FALSE);
+
+ scp->cmd_len =3D 12;
+ scp->use_sg =3D 0;
+
for (i =3D 0; i < MAX_HDRIVES; ++i) {
if (ha->hdr[i].present) {
gdtcmd.BoardNode =3D LOCALBOARD;
@@ -3586,9 +3583,11 @@
gdtcmd.u.cache.BlockNo =3D 1;
gdtcmd.u.cache.sg_canz =3D 0;
TRACE2(("gdth_flush(): flush ha %d drive %d\n", hanum, i));
- gdth_do_cmd(&scp, &gdtcmd, 30);
+ gdth_do_cmd(scp, &gdtcmd, 30);
}
}
+ scsi_release_command(scp);
+ scsi_free_host_dev(sdev);
}
=20
/* shutdown routine */
@@ -3596,8 +3595,8 @@
{
int hanum;
#ifndef __alpha__
- Scsi_Cmnd scp;
- Scsi_Device sdev;
+ Scsi_Cmnd * scp;
+ Scsi_Device * sdev;
gdth_cmd_str gdtcmd;
#endif
=20
@@ -3610,23 +3609,21 @@
=20
#ifndef __alpha__
/* controller reset */
- memset(&sdev,0,sizeof(Scsi_Device));
- memset(&scp, 0,sizeof(Scsi_Cmnd));
- sdev.host =3D gdth_ctr_tab[hanum];
- sdev.id =3D sdev.host->this_id;
- scp.cmd_len =3D 12;
- scp.host =3D gdth_ctr_tab[hanum];
- scp.target =3D sdev.host->this_id;
- scp.device =3D &sdev;
- scp.use_sg =3D 0;
+ sdev =3D scsi_get_host_dev(gdth_ctr_tab[hanum]);
+ scp =3D scsi_allocate_device(sdev, 1, FALSE);
+ scp->cmd_len =3D 12;
+ scp->use_sg =3D 0;
=20
gdtcmd.BoardNode =3D LOCALBOARD;
gdtcmd.Service =3D CACHESERVICE;
gdtcmd.OpCode =3D GDT_RESET;
TRACE2(("gdth_halt(): reset controller %d\n", hanum));
- gdth_do_cmd(&scp, &gdtcmd, 10);
+ gdth_do_cmd(scp, &gdtcmd, 10);
+ scsi_release_command(scp);
+ scsi_free_host_dev(sdev);
#endif
}
+
printk("Done.\n");
=20
#ifdef GDTH_STATISTICS
Index: linux/drivers/scsi/gdth.h
diff -u linux/drivers/scsi/gdth.h:1.1.1.1 linux/drivers/scsi/gdth.h:1.2
--- linux/drivers/scsi/gdth.h:1.1.1.1 Mon Jan 3 14:27:56 2000
+++ linux/drivers/scsi/gdth.h Mon Jan 3 18:31:14 2000
@@ -126,7 +126,8 @@
#endif
=20
/* limits */
-#define GDTH_SCRATCH 4096 /* 4KB scratch buffer =
*/
+#define GDTH_SCRATCH PAGE_SIZE /* 4KB scratch =
buffer */
+#define GDTH_SCRATCH_ORD 0 /* order 0 means 1 page =
*/
#define GDTH_MAXCMDS 124
#define GDTH_MAXC_P_L 16 /* max. cmds per lun */
#define GDTH_MAX_RAW 2 /* max. cmds per raw =
device */
Index: linux/drivers/scsi/gdth_proc.c
diff -u linux/drivers/scsi/gdth_proc.c:1.1.1.1 =
linux/drivers/scsi/gdth_proc.c:1.3
--- linux/drivers/scsi/gdth_proc.c:1.1.1.1 Mon Jan 3 14:45:15 2000
+++ linux/drivers/scsi/gdth_proc.c Mon Jan 10 21:44:27 2000
@@ -31,22 +31,17 @@
static int gdth_set_info(char *buffer,int length,int vh,int hanum,int =
busnum)
{
int ret_val;
- Scsi_Cmnd scp;
- Scsi_Device sdev;
+ Scsi_Cmnd * scp;
+ Scsi_Device * sdev;
gdth_iowr_str *piowr;
=20
TRACE2(("gdth_set_info() ha %d bus %d\n",hanum,busnum));
piowr =3D (gdth_iowr_str *)buffer;
=20
- memset(&sdev,0,sizeof(Scsi_Device));
- memset(&scp, 0,sizeof(Scsi_Cmnd));
- sdev.host =3D gdth_ctr_vtab[vh];
- sdev.id =3D sdev.host->this_id;
- scp.cmd_len =3D 12;
- scp.host =3D gdth_ctr_vtab[vh];
- scp.target =3D sdev.host->this_id;
- scp.device =3D &sdev;
- scp.use_sg =3D 0;
+ sdev =3D scsi_get_host_dev(gdth_ctr_vtab[vh]);
+ scp =3D scsi_allocate_device(sdev, 1, FALSE);
+ scp->cmd_len =3D 12;
+ scp->use_sg =3D 0;
=20
if (length >=3D 4) {
if (strncmp(buffer,"gdth",4) =3D=3D 0) {
@@ -62,10 +57,14 @@
} else {
ret_val =3D -EINVAL;
}
+
+ scsi_release_command(scp);
+ scsi_free_host_dev(sdev);
+
return ret_val;
}
=20
-static int gdth_set_asc_info(char *buffer,int length,int =
hanum,Scsi_Cmnd scp)
+static int gdth_set_asc_info(char *buffer,int length,int =
hanum,Scsi_Cmnd * scp)
{
int orig_length, drive, wb_mode;
int i, found;
@@ -105,7 +104,7 @@
gdtcmd.u.cache.DeviceNo =3D i;
gdtcmd.u.cache.BlockNo =3D 1;
gdtcmd.u.cache.sg_canz =3D 0;
- gdth_do_cmd(&scp, &gdtcmd, 30);
+ gdth_do_cmd(scp, &gdtcmd, 30);
}
}
if (!found)
@@ -158,7 +157,7 @@
gdtcmd.u.ioctl.subfunc =3D CACHE_CONFIG;
gdtcmd.u.ioctl.channel =3D INVALID_CHANNEL;
pcpar->write_back =3D wb_mode=3D=3D1 ? 0:1;
- gdth_do_cmd(&scp, &gdtcmd, 30);
+ gdth_do_cmd(scp, &gdtcmd, 30);
gdth_ioctl_free(hanum);
printk("Done.\n");
return(orig_length);
@@ -168,7 +167,7 @@
return(-EINVAL);
}
=20
-static int gdth_set_bin_info(char *buffer,int length,int =
hanum,Scsi_Cmnd scp)
+static int gdth_set_bin_info(char *buffer,int length,int =
hanum,Scsi_Cmnd * scp)
{
unchar i, j;
gdth_ha_str *ha;
@@ -241,8 +240,8 @@
*ppadd2 =3D virt_to_bus(piord->iu.general.data+add_size);
}
/* do IOCTL */
- gdth_do_cmd(&scp, pcmd, piowr->timeout);
- piord->status =3D (ulong32)scp.SCp.Message;
+ gdth_do_cmd(scp, pcmd, piowr->timeout);
+ piord->status =3D (ulong32)scp->SCp.Message;
break;
=20
case GDTIOCTL_DRVERS:
@@ -401,8 +400,8 @@
=20
gdth_cmd_str gdtcmd;
gdth_evt_str estr;
- Scsi_Cmnd scp;
- Scsi_Device sdev;
+ Scsi_Cmnd * scp;
+ Scsi_Device *sdev;
char hrec[161];
struct timeval tv;
=20
@@ -417,15 +416,10 @@
ha =3D HADATA(gdth_ctr_tab[hanum]);
id =3D length;
=20
- memset(&sdev,0,sizeof(Scsi_Device));
- memset(&scp, 0,sizeof(Scsi_Cmnd));
- sdev.host =3D gdth_ctr_vtab[vh];
- sdev.id =3D sdev.host->this_id;
- scp.cmd_len =3D 12;
- scp.host =3D gdth_ctr_vtab[vh];
- scp.target =3D sdev.host->this_id;
- scp.device =3D &sdev;
- scp.use_sg =3D 0;
+ sdev =3D scsi_get_host_dev(gdth_ctr_vtab[vh]);
+ scp =3D scsi_allocate_device(sdev, 1, FALSE);
+ scp->cmd_len =3D 12;
+ scp->use_sg =3D 0;
=20
/* look for buffer ID in length */
if (id > 1) {
@@ -531,11 +525,11 @@
sizeof(pds->list[0]);
if (pds->entries > cnt)
pds->entries =3D cnt;
- gdth_do_cmd(&scp, &gdtcmd, 30);
- if (scp.SCp.Message !=3D S_OK)=20
+ gdth_do_cmd(scp, &gdtcmd, 30);
+ if (scp->SCp.Message !=3D S_OK)=20
pds->count =3D 0;
TRACE2(("pdr_statistics() entries %d status %d\n",
- pds->count, scp.SCp.Message));
+ pds->count, scp->SCp.Message));
=20
/* other IOCTLs must fit into area GDTH_SCRATCH/4 */
for (j =3D 0; j < ha->raw[i].pdev_cnt; ++j) {
@@ -551,8 +545,8 @@
gdtcmd.u.ioctl.subfunc =3D SCSI_DR_INFO | =
L_CTRL_PATTERN;
gdtcmd.u.ioctl.channel =3D=20
ha->raw[i].address | ha->raw[i].id_list[j];
- gdth_do_cmd(&scp, &gdtcmd, 30);
- if (scp.SCp.Message =3D=3D S_OK) {
+ gdth_do_cmd(scp, &gdtcmd, 30);
+ if (scp->SCp.Message =3D=3D S_OK) {
strncpy(hrec,pdi->vendor,8);
strncpy(hrec+8,pdi->product,16);
strncpy(hrec+24,pdi->revision,4);
@@ -602,8 +596,8 @@
gdtcmd.u.ioctl.channel =3D=20
ha->raw[i].address | ha->raw[i].id_list[j];
pdef->sddc_type =3D 0x08;
- gdth_do_cmd(&scp, &gdtcmd, 30);
- if (scp.SCp.Message =3D=3D S_OK) {
+ gdth_do_cmd(scp, &gdtcmd, 30);
+ if (scp->SCp.Message =3D=3D S_OK) {
size =3D sprintf(buffer+len,
" Grown Defects:\t%d\n",
pdef->sddc_cnt);
@@ -649,8 +643,8 @@
gdtcmd.u.ioctl.param_size =3D =
sizeof(gdth_cdrinfo_str);
gdtcmd.u.ioctl.subfunc =3D CACHE_DRV_INFO;
gdtcmd.u.ioctl.channel =3D drv_no;
- gdth_do_cmd(&scp, &gdtcmd, 30);
- if (scp.SCp.Message !=3D S_OK)
+ gdth_do_cmd(scp, &gdtcmd, 30);
+ if (scp->SCp.Message !=3D S_OK)
break;
pcdi->ld_dtype >>=3D 16;
j++;
@@ -746,8 +740,8 @@
gdtcmd.u.ioctl.param_size =3D =
sizeof(gdth_arrayinf_str);
gdtcmd.u.ioctl.subfunc =3D ARRAY_INFO | =
LA_CTRL_PATTERN;
gdtcmd.u.ioctl.channel =3D i;
- gdth_do_cmd(&scp, &gdtcmd, 30);
- if (scp.SCp.Message =3D=3D S_OK) {
+ gdth_do_cmd(scp, &gdtcmd, 30);
+ if (scp->SCp.Message =3D=3D S_OK) {
if (pai->ai_state =3D=3D 0)
strcpy(hrec, "idle");
else if (pai->ai_state =3D=3D 2)
@@ -821,8 +815,8 @@
gdtcmd.u.ioctl.channel =3D i;
phg->entries =3D MAX_HDRIVES;
phg->offset =3D GDTOFFSOF(gdth_hget_str, entry[0]);=20
- gdth_do_cmd(&scp, &gdtcmd, 30);
- if (scp.SCp.Message !=3D S_OK) {
+ gdth_do_cmd(scp, &gdtcmd, 30);
+ if (scp->SCp.Message !=3D S_OK) {
ha->hdr[i].ldr_no =3D i;
ha->hdr[i].rw_attribs =3D 0;
ha->hdr[i].start_sec =3D 0;
@@ -837,7 +831,7 @@
}
}
TRACE2(("host_get entries %d status %d\n",
- phg->entries, scp.SCp.Message));
+ phg->entries, scp->SCp.Message));
}
gdth_ioctl_free(hanum);
=20
@@ -915,6 +909,10 @@
}
=20
stop_output:
+
+ scsi_release_command(scp);
+ scsi_free_host_dev(sdev);
+
*start =3D buffer +(offset-begin);
len -=3D (offset-begin);
if (len > length)
Index: linux/drivers/scsi/gdth_proc.h
diff -u linux/drivers/scsi/gdth_proc.h:1.1.1.1 =
linux/drivers/scsi/gdth_proc.h:1.2
--- linux/drivers/scsi/gdth_proc.h:1.1.1.1 Mon Jan 3 15:22:51 2000
+++ linux/drivers/scsi/gdth_proc.h Mon Jan 3 18:31:14 2000
@@ -6,8 +6,8 @@
*/
=20
static int gdth_set_info(char *buffer,int length,int vh,int hanum,int =
busnum);
-static int gdth_set_asc_info(char *buffer,int length,int =
hanum,Scsi_Cmnd scp);
-static int gdth_set_bin_info(char *buffer,int length,int =
hanum,Scsi_Cmnd scp);
+static int gdth_set_asc_info(char *buffer,int length,int =
hanum,Scsi_Cmnd * scp);
+static int gdth_set_bin_info(char *buffer,int length,int =
hanum,Scsi_Cmnd * scp);
static int gdth_get_info(char *buffer,char **start,off_t offset,
int length,int vh,int hanum,int busnum);
=20
Index: linux/drivers/scsi/hosts.c
diff -u linux/drivers/scsi/hosts.c:1.1.1.2 =
linux/drivers/scsi/hosts.c:1.2
--- linux/drivers/scsi/hosts.c:1.1.1.2 Sat Dec 18 18:33:45 1999
+++ linux/drivers/scsi/hosts.c Mon Jan 10 21:44:27 2000
@@ -869,7 +869,8 @@
printk ("scsi : %d host%s.\n", next_scsi_host,
(next_scsi_host =3D=3D 1) ? "" : "s");
=20
- =20
+ scsi_make_blocked_list();
+ =20
/* Now attach the high level drivers */
#ifdef CONFIG_BLK_DEV_SD
scsi_register_device(&sd_template);
Index: linux/drivers/scsi/hosts.h
diff -u linux/drivers/scsi/hosts.h:1.1.1.4 =
linux/drivers/scsi/hosts.h:1.4
--- linux/drivers/scsi/hosts.h:1.1.1.4 Fri Jan 7 22:33:08 2000
+++ linux/drivers/scsi/hosts.h Mon Jan 10 21:44:27 2000
@@ -334,6 +334,13 @@
unsigned int max_lun;
unsigned int max_channel;
=20
+ /*
+ * Pointer to a circularly linked list - this indicates the hosts
+ * that should be locked out of performing I/O while we have an =
active
+ * command on this host.
+ */
+ struct Scsi_Host * block;
+ unsigned wish_block:1;
=20
/* These parameters should be set by the detect routine */
unsigned long base;
Index: linux/drivers/scsi/scsi.c
diff -u linux/drivers/scsi/scsi.c:1.1.1.7 linux/drivers/scsi/scsi.c:1.12
--- linux/drivers/scsi/scsi.c:1.1.1.7 Fri Jan 7 22:33:08 2000
+++ linux/drivers/scsi/scsi.c Mon Jan 10 22:17:27 2000
@@ -86,21 +86,6 @@
* Definitions and constants.
*/
=20
-/*
- * PAGE_SIZE must be a multiple of the sector size (512). True
- * for all reasonably recent architectures (even the VAX...).
- */
-#define SECTOR_SIZE 512
-#define SECTORS_PER_PAGE (PAGE_SIZE/SECTOR_SIZE)
-
-#if SECTORS_PER_PAGE <=3D 8
-typedef unsigned char FreeSectorBitmap;
-#elif SECTORS_PER_PAGE <=3D 32
-typedef unsigned int FreeSectorBitmap;
-#else
-#error You lose.
-#endif
-
#define MIN_RESET_DELAY (2*HZ)
=20
/* Do not call reset on error if we just did a reset within 15 sec. */
@@ -139,12 +124,6 @@
static unsigned long serial_number =3D 0;
static Scsi_Cmnd *scsi_bh_queue_head =3D NULL;
static Scsi_Cmnd *scsi_bh_queue_tail =3D NULL;
-static FreeSectorBitmap *dma_malloc_freelist =3D NULL;
-static int need_isa_bounce_buffers;
-static unsigned int dma_sectors =3D 0;
-unsigned int scsi_dma_free_sectors =3D 0;
-unsigned int scsi_need_isa_buffer =3D 0;
-static unsigned char **dma_malloc_pages =3D NULL;
=20
/*
* Note - the initial logging level can be set here to log events at =
boot time.
@@ -173,7 +152,6 @@
/*=20
* Function prototypes.
*/
-static void resize_dma_pool(void);
static void print_inquiry(unsigned char *data);
extern void scsi_times_out(Scsi_Cmnd * SCpnt);
static int scan_scsis_single(int channel, int dev, int lun, int =
*max_scsi_dev,
@@ -290,6 +268,47 @@
{NULL, NULL, NULL}
};
=20
+
+/*
+ * Function: scsi_get_request_handler()
+ *
+ * Purpose: Selects queue handler function for a device.
+ *
+ * Arguments: SDpnt - device for which we need a handler function.
+ *
+ * Returns: Nothing
+ *
+ * Lock status: No locking assumed or required.
+ *
+ * Notes: Most devices will end up using scsi_request_fn for the
+ * handler function (at least as things are done now).
+ * The "block" feature basically ensures that only one of
+ * the blocked hosts is active at one time, mainly to work =
around
+ * buggy DMA chipsets where the memory gets starved.
+ * For this case, we have a special handler function, =
which
+ * does some checks and ultimately calls scsi_request_fn.
+ *
+ * As a future enhancement, it might be worthwhile to add =
support
+ * for stacked handlers - there might get to be too many =
permutations
+ * otherwise. Then again, we might just have one handler =
that does
+ * all of the special cases (a little bit slower), and =
those devices
+ * that don't need the special case code would directly =
call=20
+ * scsi_request_fn.
+ *
+ * As it stands, I can think of a number of special cases =
that
+ * we might need to handle. This would not only include =
the blocked
+ * case, but single_lun (for changers), and any special =
handling
+ * we might need for a spun-down disk to spin it back up =
again.
+ */
+static request_fn_proc * scsi_get_request_handler(Scsi_Device * SDpnt, =
struct Scsi_Host * SHpnt) {
+#ifdef CONFIG_SCSI_HOST_BLOCK
+ if( SHpnt->wish_block ) {
+ return scsi_blocked_request_fn;
+ }
+#endif
+ return scsi_request_fn;
+}
+
static int get_device_flags(unsigned char *response_data)
{
int i =3D 0;
@@ -314,7 +333,6 @@
return 0;
}
=20
-
static void scan_scsis_done(Scsi_Cmnd * SCpnt)
{
=20
@@ -437,7 +455,7 @@
* the queue actually represents. We could look it up, but it
* is pointless work.
*/
- blk_init_queue(&SDpnt->request_queue, scsi_request_fn);
+ blk_init_queue(&SDpnt->request_queue, =
scsi_get_request_handler(SDpnt, shpnt));
blk_queue_headactive(&SDpnt->request_queue, 0);
SDpnt->request_queue.queuedata =3D (void *) SDpnt;
/* Make sure we have something that is valid for DMA purposes */
@@ -520,7 +538,7 @@
}
}
}
- resize_dma_pool();
+ scsi_resize_dma_pool();
=20
for (sdtpnt =3D scsi_devicelist; sdtpnt; sdtpnt =3D sdtpnt->next) {
if (sdtpnt->finish && sdtpnt->nr_dev) {
@@ -882,7 +900,7 @@
* the queue actually represents. We could look it up, but it
* is pointless work.
*/
- blk_init_queue(&SDpnt->request_queue, scsi_request_fn);
+ blk_init_queue(&SDpnt->request_queue, scsi_get_request_handler(SDpnt, =
shpnt));
blk_queue_headactive(&SDpnt->request_queue, 0);
SDpnt->request_queue.queuedata =3D (void *) SDpnt;
SDpnt->host =3D shpnt;
@@ -990,11 +1008,6 @@
static spinlock_t device_request_lock =3D SPIN_LOCK_UNLOCKED;
=20
/*
- * Used for access to internal allocator used for DMA safe buffers.
- */
-static spinlock_t allocator_request_lock =3D SPIN_LOCK_UNLOCKED;
-
-/*
* Used to protect insertion into and removal from the queue of
* commands to be processed by the bottom half handler.
*/
@@ -1810,127 +1823,6 @@
static void scsi_unregister_host(Scsi_Host_Template *);
#endif
=20
-/*
- * Function: scsi_malloc
- *
- * Purpose: Allocate memory from the DMA-safe pool.
- *
- * Arguments: len - amount of memory we need.
- *
- * Lock status: No locks assumed to be held. This function is =
SMP-safe.
- *
- * Returns: Pointer to memory block.
- *
- * Notes: Prior to the new queue code, this function was not =
SMP-safe.
- * This function can only allocate in units of sectors
- * (i.e. 512 bytes).
- *
- * We cannot use the normal system allocator becuase we =
need
- * to be able to guarantee that we can process a complete =
disk
- * I/O request without touching the system allocator. =
Think
- * about it - if the system were heavily swapping, and =
tried to
- * write out a block of memory to disk, and the SCSI code =
needed
- * to allocate more memory in order to be able to write =
the
- * data to disk, you would wedge the system.
- */
-void *scsi_malloc(unsigned int len)
-{
- unsigned int nbits, mask;
- unsigned long flags;
-
- int i, j;
- if (len % SECTOR_SIZE !=3D 0 || len > PAGE_SIZE)
- return NULL;
-
- nbits =3D len >> 9;
- mask =3D (1 << nbits) - 1;
-
- spin_lock_irqsave(&allocator_request_lock, flags);
-
- for (i =3D 0; i < dma_sectors / SECTORS_PER_PAGE; i++)
- for (j =3D 0; j <=3D SECTORS_PER_PAGE - nbits; j++) {
- if ((dma_malloc_freelist[i] & (mask << j)) =3D=3D 0) {
- dma_malloc_freelist[i] |=3D (mask << j);
- scsi_dma_free_sectors -=3D nbits;
-#ifdef DEBUG
- SCSI_LOG_MLQUEUE(3, printk("SMalloc: %d %p [From:%p]\n", len, =
dma_malloc_pages[i] + (j << 9)));
- printk("SMalloc: %d %p [From:%p]\n", len, dma_malloc_pages[i] + (j =
<< 9));
-#endif
- spin_unlock_irqrestore(&allocator_request_lock, flags);
- return (void *) ((unsigned long) dma_malloc_pages[i] + (j << 9));
- }
- }
- spin_unlock_irqrestore(&allocator_request_lock, flags);
- return NULL; /* Nope. No more */
-}
-
-/*
- * Function: scsi_free
- *
- * Purpose: Free memory into the DMA-safe pool.
- *
- * Arguments: ptr - data block we are freeing.
- * len - size of block we are freeing.
- *
- * Lock status: No locks assumed to be held. This function is =
SMP-safe.
- *
- * Returns: Nothing
- *
- * Notes: This function *must* only be used to free memory
- * allocated from scsi_malloc().
- *
- * Prior to the new queue code, this function was not =
SMP-safe.
- * This function can only allocate in units of sectors
- * (i.e. 512 bytes).
- */
-int scsi_free(void *obj, unsigned int len)
-{
- unsigned int page, sector, nbits, mask;
- unsigned long flags;
-
-#ifdef DEBUG
- unsigned long ret =3D 0;
-
-#ifdef __mips__
- __asm__ __volatile__("move\t%0,$31":"=3Dr"(ret));
-#else
- ret =3D __builtin_return_address(0);
-#endif
- printk("scsi_free %p %d\n", obj, len);
- SCSI_LOG_MLQUEUE(3, printk("SFree: %p %d\n", obj, len));
-#endif
-
- spin_lock_irqsave(&allocator_request_lock, flags);
-
- for (page =3D 0; page < dma_sectors / SECTORS_PER_PAGE; page++) {
- unsigned long page_addr =3D (unsigned long) dma_malloc_pages[page];
- if ((unsigned long) obj >=3D page_addr &&
- (unsigned long) obj < page_addr + PAGE_SIZE) {
- sector =3D (((unsigned long) obj) - page_addr) >> 9;
-
- nbits =3D len >> 9;
- mask =3D (1 << nbits) - 1;
-
- if ((mask << sector) >=3D (1 << SECTORS_PER_PAGE))
- panic("scsi_free:Bad memory alignment");
-
- if ((dma_malloc_freelist[page] &
- (mask << sector)) !=3D (mask << sector)) {
-#ifdef DEBUG
- printk("scsi_free(obj=3D%p, len=3D%d) called from %08lx\n",
- obj, len, ret);
-#endif
- panic("scsi_free:Trying to free unused memory");
- }
- scsi_dma_free_sectors +=3D nbits;
- dma_malloc_freelist[page] &=3D ~(mask << sector);
- spin_unlock_irqrestore(&allocator_request_lock, flags);
- return 0;
- }
- }
- panic("scsi_free:Bad offset");
-}
-
=20
int scsi_loadable_module_flag; /* Set after we scan builtin drivers */
=20
@@ -2114,7 +2006,7 @@
/*
* This should build the DMA pool.
*/
- resize_dma_pool();
+ scsi_resize_dma_pool();
=20
/*
* OK, now we finish the initialization by doing spin-up, read
@@ -2465,217 +2357,6 @@
}
#endif
=20
-/*
- * Function: resize_dma_pool
- *
- * Purpose: Ensure that the DMA pool is sufficiently large to be
- * able to guarantee that we can always process I/O =
requests
- * without calling the system allocator.
- *
- * Arguments: None.
- *
- * Lock status: No locks assumed to be held. This function is =
SMP-safe.
- *
- * Returns: Nothing
- *
- * Notes: Prior to the new queue code, this function was not =
SMP-safe.
- * Go through the device list and recompute the most =
appropriate
- * size for the dma pool. Then grab more memory (as =
required).
- */
-static void resize_dma_pool(void)
-{
- int i, k;
- unsigned long size;
- unsigned long flags;
- struct Scsi_Host *shpnt;
- struct Scsi_Host *host =3D NULL;
- Scsi_Device *SDpnt;
- FreeSectorBitmap *new_dma_malloc_freelist =3D NULL;
- unsigned int new_dma_sectors =3D 0;
- unsigned int new_need_isa_buffer =3D 0;
- unsigned char **new_dma_malloc_pages =3D NULL;
- int out_of_space =3D 0;
-
- spin_lock_irqsave(&allocator_request_lock, flags);
-
- if (!scsi_hostlist) {
- /*
- * Free up the DMA pool.
- */
- if (scsi_dma_free_sectors !=3D dma_sectors)
- panic("SCSI DMA pool memory leak %d %d\n", scsi_dma_free_sectors, =
dma_sectors);
-
- for (i =3D 0; i < dma_sectors / SECTORS_PER_PAGE; i++)
- scsi_init_free(dma_malloc_pages[i], PAGE_SIZE);
- if (dma_malloc_pages)
- scsi_init_free((char *) dma_malloc_pages,
- (dma_sectors / SECTORS_PER_PAGE) * =
sizeof(*dma_malloc_pages));
- dma_malloc_pages =3D NULL;
- if (dma_malloc_freelist)
- scsi_init_free((char *) dma_malloc_freelist,
- (dma_sectors / SECTORS_PER_PAGE) * =
sizeof(*dma_malloc_freelist));
- dma_malloc_freelist =3D NULL;
- dma_sectors =3D 0;
- scsi_dma_free_sectors =3D 0;
- spin_unlock_irqrestore(&allocator_request_lock, flags);
- return;
- }
- /* Next, check to see if we need to extend the DMA buffer pool */
-
- new_dma_sectors =3D 2 * SECTORS_PER_PAGE; /* Base value we use */
-
- if (__pa(high_memory) - 1 > ISA_DMA_THRESHOLD)
- need_isa_bounce_buffers =3D 1;
- else
- need_isa_bounce_buffers =3D 0;
-
- if (scsi_devicelist)
- for (shpnt =3D scsi_hostlist; shpnt; shpnt =3D shpnt->next)
- new_dma_sectors +=3D SECTORS_PER_PAGE; /* Increment for each host */
-
- for (host =3D scsi_hostlist; host; host =3D host->next) {
- for (SDpnt =3D host->host_queue; SDpnt; SDpnt =3D SDpnt->next) {
- /*
- * sd and sr drivers allocate scatterlists.
- * sr drivers may allocate for each command 1x2048 or 2x1024 extra
- * buffers for 2k sector size and 1k fs.
- * sg driver allocates buffers < 4k.
- * st driver does not need buffers from the dma pool.
- * estimate 4k buffer/command for devices of unknown type (should =
panic).
- */
- if (SDpnt->type =3D=3D TYPE_WORM || SDpnt->type =3D=3D TYPE_ROM ||
- SDpnt->type =3D=3D TYPE_DISK || SDpnt->type =3D=3D TYPE_MOD) {
- new_dma_sectors +=3D ((host->sg_tablesize *
- sizeof(struct scatterlist) + 511) >> 9) *
- SDpnt->queue_depth;
- if (SDpnt->type =3D=3D TYPE_WORM || SDpnt->type =3D=3D TYPE_ROM)
- new_dma_sectors +=3D (2048 >> 9) * SDpnt->queue_depth;
- } else if (SDpnt->type =3D=3D TYPE_SCANNER ||
- SDpnt->type =3D=3D TYPE_PROCESSOR ||
- SDpnt->type =3D=3D TYPE_MEDIUM_CHANGER ||
- SDpnt->type =3D=3D TYPE_ENCLOSURE) {
- new_dma_sectors +=3D (4096 >> 9) * SDpnt->queue_depth;
- } else {
- if (SDpnt->type !=3D TYPE_TAPE) {
- printk("resize_dma_pool: unknown device type %d\n", SDpnt->type);
- new_dma_sectors +=3D (4096 >> 9) * SDpnt->queue_depth;
- }
- }
-
- if (host->unchecked_isa_dma &&
- need_isa_bounce_buffers &&
- SDpnt->type !=3D TYPE_TAPE) {
- new_dma_sectors +=3D (PAGE_SIZE >> 9) * host->sg_tablesize *
- SDpnt->queue_depth;
- new_need_isa_buffer++;
- }
- }
- }
-
-#ifdef DEBUG_INIT
- printk("resize_dma_pool: needed dma sectors =3D %d\n", =
new_dma_sectors);
-#endif
-
- /* limit DMA memory to 32MB: */
- new_dma_sectors =3D (new_dma_sectors + 15) & 0xfff0;
-
- /*
- * We never shrink the buffers - this leads to
- * race conditions that I would rather not even think
- * about right now.
- */
-#if 0 /* Why do this? No gain and risks out_of_space */
- if (new_dma_sectors < dma_sectors)
- new_dma_sectors =3D dma_sectors;
-#endif
- if (new_dma_sectors <=3D dma_sectors) {
- spin_unlock_irqrestore(&allocator_request_lock, flags);
- return; /* best to quit while we are in front */
- }
-
- for (k =3D 0; k < 20; ++k) { /* just in case */
- out_of_space =3D 0;
- size =3D (new_dma_sectors / SECTORS_PER_PAGE) *
- sizeof(FreeSectorBitmap);
- new_dma_malloc_freelist =3D (FreeSectorBitmap *)
- scsi_init_malloc(size, GFP_ATOMIC);
- if (new_dma_malloc_freelist) {
- size =3D (new_dma_sectors / SECTORS_PER_PAGE) *
- sizeof(*new_dma_malloc_pages);
- new_dma_malloc_pages =3D (unsigned char **)
- scsi_init_malloc(size, GFP_ATOMIC);
- if (!new_dma_malloc_pages) {
- size =3D (new_dma_sectors / SECTORS_PER_PAGE) *
- sizeof(FreeSectorBitmap);
- scsi_init_free((char *) new_dma_malloc_freelist, size);
- out_of_space =3D 1;
- }
- } else
- out_of_space =3D 1;
-
- if ((!out_of_space) && (new_dma_sectors > dma_sectors)) {
- for (i =3D dma_sectors / SECTORS_PER_PAGE;
- i < new_dma_sectors / SECTORS_PER_PAGE; i++) {
- new_dma_malloc_pages[i] =3D (unsigned char *)
- scsi_init_malloc(PAGE_SIZE, GFP_ATOMIC | GFP_DMA);
- if (!new_dma_malloc_pages[i])
- break;
- }
- if (i !=3D new_dma_sectors / SECTORS_PER_PAGE) { /* clean up */
- int k =3D i;
-
- out_of_space =3D 1;
- for (i =3D 0; i < k; ++i)
- scsi_init_free(new_dma_malloc_pages[i], PAGE_SIZE);
- }
- }
- if (out_of_space) { /* try scaling down new_dma_sectors request */
- printk("scsi::resize_dma_pool: WARNING, dma_sectors=3D%u, "
- "wanted=3D%u, scaling\n", dma_sectors, new_dma_sectors);
- if (new_dma_sectors < (8 * SECTORS_PER_PAGE))
- break; /* pretty well hopeless ... */
- new_dma_sectors =3D (new_dma_sectors * 3) / 4;
- new_dma_sectors =3D (new_dma_sectors + 15) & 0xfff0;
- if (new_dma_sectors <=3D dma_sectors)
- break; /* stick with what we have got */
- } else
- break; /* found space ... */
- } /* end of for loop */
- if (out_of_space) {
- spin_unlock_irqrestore(&allocator_request_lock, flags);
- scsi_need_isa_buffer =3D new_need_isa_buffer; /* some useful info */
- printk(" WARNING, not enough memory, pool not expanded\n");
- return;
- }
- /* When we dick with the actual DMA list, we need to
- * protect things
- */
- if (dma_malloc_freelist) {
- size =3D (dma_sectors / SECTORS_PER_PAGE) * sizeof(FreeSectorBitmap);
- memcpy(new_dma_malloc_freelist, dma_malloc_freelist, size);
- scsi_init_free((char *) dma_malloc_freelist, size);
- }
- dma_malloc_freelist =3D new_dma_malloc_freelist;
-
- if (dma_malloc_pages) {
- size =3D (dma_sectors / SECTORS_PER_PAGE) * =
sizeof(*dma_malloc_pages);
- memcpy(new_dma_malloc_pages, dma_malloc_pages, size);
- scsi_init_free((char *) dma_malloc_pages, size);
- }
- scsi_dma_free_sectors +=3D new_dma_sectors - dma_sectors;
- dma_malloc_pages =3D new_dma_malloc_pages;
- dma_sectors =3D new_dma_sectors;
- scsi_need_isa_buffer =3D new_need_isa_buffer;
-
- spin_unlock_irqrestore(&allocator_request_lock, flags);
-
-#ifdef DEBUG_INIT
- printk("resize_dma_pool: dma free sectors =3D %d\n", =
scsi_dma_free_sectors);
- printk("resize_dma_pool: dma sectors =3D %d\n", dma_sectors);
- printk("resize_dma_pool: need isa buffers =3D %d\n", =
scsi_need_isa_buffer);
-#endif
-}
-
#ifdef CONFIG_MODULES /* a big #ifdef block... */
=20
/*
@@ -2771,6 +2452,8 @@
printk("scsi : %d host%s.\n", next_scsi_host,
(next_scsi_host =3D=3D 1) ? "" : "s");
=20
+ scsi_make_blocked_list();
+
/* The next step is to call scan_scsis here. This generates the
* Scsi_Devices entries
*/
@@ -2809,7 +2492,7 @@
* Now that we have all of the devices, resize the DMA pool,
* as required. */
if (!out_of_space)
- resize_dma_pool();
+ scsi_resize_dma_pool();
=20
=20
/* This does any final handling that is required. */
@@ -3027,7 +2710,7 @@
* do the right thing and free everything.
*/
if (!scsi_hosts)
- resize_dma_pool();
+ scsi_resize_dma_pool();
=20
printk("scsi : %d host%s.\n", next_scsi_host,
(next_scsi_host =3D=3D 1) ? "" : "s");
@@ -3039,6 +2722,7 @@
(scsi_memory_upper_value - scsi_init_memory_start) / 1024);
#endif
=20
+ scsi_make_blocked_list();
=20
/* There were some hosts that were loaded at boot time, so we cannot
do any more than this */
@@ -3122,7 +2806,7 @@
if (tpnt->finish && tpnt->nr_dev)
(*tpnt->finish) ();
if (!out_of_space)
- resize_dma_pool();
+ scsi_resize_dma_pool();
MOD_INC_USE_COUNT;
=20
if (out_of_space) {
@@ -3372,39 +3056,11 @@
=20
scsi_loadable_module_flag =3D 1;
=20
- dma_sectors =3D PAGE_SIZE / SECTOR_SIZE;
- scsi_dma_free_sectors =3D dma_sectors;
- /*
- * Set up a minimal DMA buffer list - this will be used during =
scan_scsis
- * in some cases.
- */
+ if( scsi_init_minimal_dma_pool() =3D=3D 0 )
+ {
+ return 1;
+ }
=20
- /* One bit per sector to indicate free/busy */
- size =3D (dma_sectors / SECTORS_PER_PAGE) * sizeof(FreeSectorBitmap);
- dma_malloc_freelist =3D (FreeSectorBitmap *)
- scsi_init_malloc(size, GFP_ATOMIC);
- if (dma_malloc_freelist) {
- /* One pointer per page for the page list */
- dma_malloc_pages =3D (unsigned char **) scsi_init_malloc(
- (dma_sectors / SECTORS_PER_PAGE) * =
sizeof(*dma_malloc_pages),
- GFP_ATOMIC);
- if (dma_malloc_pages) {
- dma_malloc_pages[0] =3D (unsigned char *)
- scsi_init_malloc(PAGE_SIZE, GFP_ATOMIC | GFP_DMA);
- if (dma_malloc_pages[0])
- has_space =3D 1;
- }
- }
- if (!has_space) {
- if (dma_malloc_freelist) {
- scsi_init_free((char *) dma_malloc_freelist, size);
- if (dma_malloc_pages)
- scsi_init_free((char *) dma_malloc_pages,
- (dma_sectors / SECTORS_PER_PAGE) * =
sizeof(*dma_malloc_pages));
- }
- printk("scsi::init_module: failed, out of memory\n");
- return 1;
- }
/*
* This is where the processing takes place for most everything
* when commands are completed.
@@ -3427,7 +3083,7 @@
/*
* Free up the DMA pool.
*/
- resize_dma_pool();
+ scsi_resize_dma_pool();
=20
}
=20
@@ -3481,7 +3137,7 @@
=20
SDpnt->device_queue =3D SCpnt;
=20
- blk_init_queue(&SDpnt->request_queue, scsi_request_fn);
+ blk_init_queue(&SDpnt->request_queue, =
scsi_get_request_handler(SDpnt, SDpnt->host));
blk_queue_headactive(&SDpnt->request_queue, 0);
SDpnt->request_queue.queuedata =3D (void *) SDpnt;
=20
@@ -3509,7 +3165,7 @@
*/
void scsi_free_host_dev(Scsi_Device * SDpnt)
{
- if( SDpnt->id !=3D SDpnt->host->this_id )
+ if( (unsigned char) SDpnt->id !=3D (unsigned char) =
SDpnt->host->this_id )
{
panic("Attempt to delete wrong device\n");
}
Index: linux/drivers/scsi/scsi.h
diff -u linux/drivers/scsi/scsi.h:1.1.1.5 linux/drivers/scsi/scsi.h:1.6
--- linux/drivers/scsi/scsi.h:1.1.1.5 Fri Jan 7 22:33:08 2000
+++ linux/drivers/scsi/scsi.h Mon Jan 10 21:44:27 2000
@@ -365,90 +365,121 @@
* Initializes all SCSI devices. This scans all scsi busses.
*/
=20
-extern int scsi_dev_init(void);
-
-
-
-void *scsi_malloc(unsigned int);
-int scsi_free(void *, unsigned int);
extern unsigned int scsi_logging_level; /* What do we log? */
extern unsigned int scsi_dma_free_sectors; /* How much room do we have =
left */
extern unsigned int scsi_need_isa_buffer; /* True if some devices need =
indirection
* buffers */
-extern void scsi_make_blocked_list(void);
extern volatile int in_scan_scsis;
extern const unsigned char scsi_command_size[8];
=20
+
/*
* These are the error handling functions defined in scsi_error.c
*/
+extern void scsi_times_out(Scsi_Cmnd * SCpnt);
extern void scsi_add_timer(Scsi_Cmnd * SCset, int timeout,
void (*complete) (Scsi_Cmnd *));
-extern void scsi_done(Scsi_Cmnd * SCpnt);
extern int scsi_delete_timer(Scsi_Cmnd * SCset);
extern void scsi_error_handler(void *host);
-extern int scsi_retry_command(Scsi_Cmnd *);
-extern void scsi_finish_command(Scsi_Cmnd *);
extern int scsi_sense_valid(Scsi_Cmnd *);
extern int scsi_decide_disposition(Scsi_Cmnd * SCpnt);
extern int scsi_block_when_processing_errors(Scsi_Device *);
extern void scsi_sleep(int);
+
+/*
+ * Prototypes for functions in scsicam.c
+ */
extern int scsi_partsize(struct buffer_head *bh, unsigned long =
capacity,
unsigned int *cyls, unsigned int *hds,
unsigned int *secs);
=20
/*
+ * Prototypes for functions in scsi_dma.c
+ */
+void scsi_resize_dma_pool(void);
+int scsi_init_minimal_dma_pool(void);
+void *scsi_malloc(unsigned int);
+int scsi_free(void *, unsigned int);
+
+/*
* Prototypes for functions in scsi_merge.c
*/
extern void recount_segments(Scsi_Cmnd * SCpnt);
+extern void initialize_merge_fn(Scsi_Device * SDpnt);
=20
/*
+ * Prototypes for functions in scsi_queue.c
+ */
+extern int scsi_mlqueue_insert(Scsi_Cmnd * cmd, int reason);
+
+/*
* Prototypes for functions in scsi_lib.c
*/
-extern void initialize_merge_fn(Scsi_Device * SDpnt);
-extern void scsi_request_fn(request_queue_t * q);
+extern void scsi_blocked_request_fn(request_queue_t * q);
+extern Scsi_Cmnd *scsi_end_request(Scsi_Cmnd * SCpnt, int uptodate,
+ int sectors);
+extern struct Scsi_Device_Template *scsi_get_request_dev(struct request =
*);
+extern int scsi_init_cmd_errh(Scsi_Cmnd * SCpnt);
+extern int scsi_insert_special_cmd(Scsi_Cmnd * SCpnt, int);
+extern void scsi_io_completion(Scsi_Cmnd * SCpnt, int good_sectors,
+ int block_sectors);
+extern void scsi_make_blocked_list(void);
extern void scsi_queue_next_request(request_queue_t * q, Scsi_Cmnd * =
SCpnt);
+extern void scsi_request_fn(request_queue_t * q);
=20
-extern int scsi_insert_special_cmd(Scsi_Cmnd * SCpnt, int);
-extern int scsi_dispatch_cmd(Scsi_Cmnd * SCpnt);
=20
/*
* Prototypes for functions in scsi.c
*/
-
-/*
- * scsi_abort aborts the current command that is executing on host =
host.
- * The error code, if non zero is returned in the host byte, otherwise =
- * DID_ABORT is returned in the hostbyte.
- */
-
+extern int scsi_dispatch_cmd(Scsi_Cmnd * SCpnt);
+extern void scsi_bottom_half_handler(void);
+extern void scsi_build_commandblocks(Scsi_Device * SDpnt);
+extern void scsi_done(Scsi_Cmnd * SCpnt);
+extern void scsi_finish_command(Scsi_Cmnd *);
+extern int scsi_retry_command(Scsi_Cmnd *);
+extern Scsi_Cmnd *scsi_allocate_device(Scsi_Device *, int, int);
+extern void scsi_release_command(Scsi_Cmnd *);
extern void scsi_do_cmd(Scsi_Cmnd *, const void *cmnd,
void *buffer, unsigned bufflen,
void (*done) (struct scsi_cmnd *),
int timeout, int retries);
-
extern void scsi_wait_cmd(Scsi_Cmnd *, const void *cmnd,
void *buffer, unsigned bufflen,
void (*done) (struct scsi_cmnd *),
int timeout, int retries);
+extern int scsi_dev_init(void);
=20
-extern Scsi_Cmnd *scsi_allocate_device(Scsi_Device *, int, int);
-
-extern void scsi_release_command(Scsi_Cmnd *);
=20
+/*
+ * Prototypes for functions/data in hosts.c
+ */
extern int max_scsi_hosts;
=20
+/*
+ * Prototypes for functions in scsi_proc.c
+ */
extern void proc_print_scsidevice(Scsi_Device *, char *, int *, int);
extern struct proc_dir_entry *proc_scsi;
=20
+/*
+ * Prototypes for functions in constants.c
+ */
extern void print_command(unsigned char *);
extern void print_sense(const char *, Scsi_Cmnd *);
extern void print_driverbyte(int scsiresult);
extern void print_hostbyte(int scsiresult);
+extern void print_status (int status);
=20
/*
* The scsi_device struct contains what we know about each given scsi
* device.
+ *
+ * FIXME(eric) - one of the great regrets that I have is that I failed =
to define
+ * these structure elements as something like sdev_foo instead of foo. =
This would
+ * make it so much easier to grep through sources and so forth. I =
propose that
+ * all new elements that get added to these structures follow this =
convention.
+ * As time goes on and as people have the stomach for it, it should be =
possible to=20
+ * go back and retrofit at least some of the elements here with with =
the prefix.
*/
=20
struct scsi_device {
@@ -538,6 +569,14 @@
} Scsi_Pointer;
=20
=20
+/*
+ * FIXME(eric) - one of the great regrets that I have is that I failed =
to define
+ * these structure elements as something like sc_foo instead of foo. =
This would
+ * make it so much easier to grep through sources and so forth. I =
propose that
+ * all new elements that get added to these structures follow this =
convention.
+ * As time goes on and as people have the stomach for it, it should be =
possible to=20
+ * go back and retrofit at least some of the elements here with with =
the prefix.
+ */
struct scsi_cmnd {
/* private: */
/*
@@ -680,16 +719,6 @@
*/
#define SCSI_MLQUEUE_HOST_BUSY 0x1055
#define SCSI_MLQUEUE_DEVICE_BUSY 0x1056
-
-extern int scsi_mlqueue_insert(Scsi_Cmnd * cmd, int reason);
-
-extern Scsi_Cmnd *scsi_end_request(Scsi_Cmnd * SCpnt, int uptodate,
- int sectors);
-
-extern void scsi_io_completion(Scsi_Cmnd * SCpnt, int good_sectors,
- int block_sectors);
-
-extern struct Scsi_Device_Template *scsi_get_request_dev(struct request =
*);
=20
#define SCSI_SLEEP(QUEUE, CONDITION) { \
if (CONDITION) { \
Index: linux/drivers/scsi/scsi_dma.c
diff -u /dev/null linux/drivers/scsi/scsi_dma.c:1.3
--- /dev/null Mon Jan 10 22:17:54 2000
+++ linux/drivers/scsi/scsi_dma.c Sat Jan 8 21:19:18 2000
@@ -0,0 +1,442 @@
+/*
+ * scsi_dma.c Copyright (C) 2000 Eric Youngdale
+ *
+ * mid-level SCSI DMA bounce buffer allocator
+ *
+ */
+
+#define __NO_VERSION__
+#include <linux/config.h>
+#include <linux/module.h>
+#include <linux/blk.h>
+
+
+#include "scsi.h"
+#include "hosts.h"
+#include "constants.h"
+
+#ifdef CONFIG_KMOD
+#include <linux/kmod.h>
+#endif
+
+/*
+ * PAGE_SIZE must be a multiple of the sector size (512). True
+ * for all reasonably recent architectures (even the VAX...).
+ */
+#define SECTOR_SIZE 512
+#define SECTORS_PER_PAGE (PAGE_SIZE/SECTOR_SIZE)
+
+#if SECTORS_PER_PAGE <=3D 8
+typedef unsigned char FreeSectorBitmap;
+#elif SECTORS_PER_PAGE <=3D 32
+typedef unsigned int FreeSectorBitmap;
+#else
+#error You lose.
+#endif
+
+/*
+ * Used for access to internal allocator used for DMA safe buffers.
+ */
+static spinlock_t allocator_request_lock =3D SPIN_LOCK_UNLOCKED;
+
+static FreeSectorBitmap *dma_malloc_freelist =3D NULL;
+static int need_isa_bounce_buffers;
+static unsigned int dma_sectors =3D 0;
+unsigned int scsi_dma_free_sectors =3D 0;
+unsigned int scsi_need_isa_buffer =3D 0;
+static unsigned char **dma_malloc_pages =3D NULL;
+
+/*
+ * Function: scsi_malloc
+ *
+ * Purpose: Allocate memory from the DMA-safe pool.
+ *
+ * Arguments: len - amount of memory we need.
+ *
+ * Lock status: No locks assumed to be held. This function is =
SMP-safe.
+ *
+ * Returns: Pointer to memory block.
+ *
+ * Notes: Prior to the new queue code, this function was not =
SMP-safe.
+ * This function can only allocate in units of sectors
+ * (i.e. 512 bytes).
+ *
+ * We cannot use the normal system allocator becuase we =
need
+ * to be able to guarantee that we can process a complete =
disk
+ * I/O request without touching the system allocator. =
Think
+ * about it - if the system were heavily swapping, and =
tried to
+ * write out a block of memory to disk, and the SCSI code =
needed
+ * to allocate more memory in order to be able to write =
the
+ * data to disk, you would wedge the system.
+ */
+void *scsi_malloc(unsigned int len)
+{
+ unsigned int nbits, mask;
+ unsigned long flags;
+
+ int i, j;
+ if (len % SECTOR_SIZE !=3D 0 || len > PAGE_SIZE)
+ return NULL;
+
+ nbits =3D len >> 9;
+ mask =3D (1 << nbits) - 1;
+
+ spin_lock_irqsave(&allocator_request_lock, flags);
+
+ for (i =3D 0; i < dma_sectors / SECTORS_PER_PAGE; i++)
+ for (j =3D 0; j <=3D SECTORS_PER_PAGE - nbits; j++) {
+ if ((dma_malloc_freelist[i] & (mask << j)) =3D=3D 0) {
+ dma_malloc_freelist[i] |=3D (mask << j);
+ scsi_dma_free_sectors -=3D nbits;
+#ifdef DEBUG
+ SCSI_LOG_MLQUEUE(3, printk("SMalloc: %d %p [From:%p]\n", len, =
dma_malloc_pages[i] + (j << 9)));
+ printk("SMalloc: %d %p [From:%p]\n", len, dma_malloc_pages[i] + (j =
<< 9));
+#endif
+ spin_unlock_irqrestore(&allocator_request_lock, flags);
+ return (void *) ((unsigned long) dma_malloc_pages[i] + (j << 9));
+ }
+ }
+ spin_unlock_irqrestore(&allocator_request_lock, flags);
+ return NULL; /* Nope. No more */
+}
+
+/*
+ * Function: scsi_free
+ *
+ * Purpose: Free memory into the DMA-safe pool.
+ *
+ * Arguments: ptr - data block we are freeing.
+ * len - size of block we are freeing.
+ *
+ * Lock status: No locks assumed to be held. This function is =
SMP-safe.
+ *
+ * Returns: Nothing
+ *
+ * Notes: This function *must* only be used to free memory
+ * allocated from scsi_malloc().
+ *
+ * Prior to the new queue code, this function was not =
SMP-safe.
+ * This function can only allocate in units of sectors
+ * (i.e. 512 bytes).
+ */
+int scsi_free(void *obj, unsigned int len)
+{
+ unsigned int page, sector, nbits, mask;
+ unsigned long flags;
+
+#ifdef DEBUG
+ unsigned long ret =3D 0;
+
+#ifdef __mips__
+ __asm__ __volatile__("move\t%0,$31":"=3Dr"(ret));
+#else
+ ret =3D __builtin_return_address(0);
+#endif
+ printk("scsi_free %p %d\n", obj, len);
+ SCSI_LOG_MLQUEUE(3, printk("SFree: %p %d\n", obj, len));
+#endif
+
+ spin_lock_irqsave(&allocator_request_lock, flags);
+
+ for (page =3D 0; page < dma_sectors / SECTORS_PER_PAGE; page++) {
+ unsigned long page_addr =3D (unsigned long) dma_malloc_pages[page];
+ if ((unsigned long) obj >=3D page_addr &&
+ (unsigned long) obj < page_addr + PAGE_SIZE) {
+ sector =3D (((unsigned long) obj) - page_addr) >> 9;
+
+ nbits =3D len >> 9;
+ mask =3D (1 << nbits) - 1;
+
+ if ((mask << sector) >=3D (1 << SECTORS_PER_PAGE))
+ panic("scsi_free:Bad memory alignment");
+
+ if ((dma_malloc_freelist[page] &
+ (mask << sector)) !=3D (mask << sector)) {
+#ifdef DEBUG
+ printk("scsi_free(obj=3D%p, len=3D%d) called from %08lx\n",
+ obj, len, ret);
+#endif
+ panic("scsi_free:Trying to free unused memory");
+ }
+ scsi_dma_free_sectors +=3D nbits;
+ dma_malloc_freelist[page] &=3D ~(mask << sector);
+ spin_unlock_irqrestore(&allocator_request_lock, flags);
+ return 0;
+ }
+ }
+ panic("scsi_free:Bad offset");
+}
+
+
+/*
+ * Function: scsi_resize_dma_pool
+ *
+ * Purpose: Ensure that the DMA pool is sufficiently large to be
+ * able to guarantee that we can always process I/O =
requests
+ * without calling the system allocator.
+ *
+ * Arguments: None.
+ *
+ * Lock status: No locks assumed to be held. This function is =
SMP-safe.
+ *
+ * Returns: Nothing
+ *
+ * Notes: Prior to the new queue code, this function was not =
SMP-safe.
+ * Go through the device list and recompute the most =
appropriate
+ * size for the dma pool. Then grab more memory (as =
required).
+ */
+void scsi_resize_dma_pool(void)
+{
+ int i, k;
+ unsigned long size;
+ unsigned long flags;
+ struct Scsi_Host *shpnt;
+ struct Scsi_Host *host =3D NULL;
+ Scsi_Device *SDpnt;
+ FreeSectorBitmap *new_dma_malloc_freelist =3D NULL;
+ unsigned int new_dma_sectors =3D 0;
+ unsigned int new_need_isa_buffer =3D 0;
+ unsigned char **new_dma_malloc_pages =3D NULL;
+ int out_of_space =3D 0;
+
+ spin_lock_irqsave(&allocator_request_lock, flags);
+
+ if (!scsi_hostlist) {
+ /*
+ * Free up the DMA pool.
+ */
+ if (scsi_dma_free_sectors !=3D dma_sectors)
+ panic("SCSI DMA pool memory leak %d %d\n", scsi_dma_free_sectors, =
dma_sectors);
+
+ for (i =3D 0; i < dma_sectors / SECTORS_PER_PAGE; i++)
+ free_pages((unsigned long) dma_malloc_pages[i], 0);
+ if (dma_malloc_pages)
+ kfree((char *) dma_malloc_pages);
+ dma_malloc_pages =3D NULL;
+ if (dma_malloc_freelist)
+ kfree((char *) dma_malloc_freelist);
+ dma_malloc_freelist =3D NULL;
+ dma_sectors =3D 0;
+ scsi_dma_free_sectors =3D 0;
+ spin_unlock_irqrestore(&allocator_request_lock, flags);
+ return;
+ }
+ /* Next, check to see if we need to extend the DMA buffer pool */
+
+ new_dma_sectors =3D 2 * SECTORS_PER_PAGE; /* Base value we use */
+
+ if (__pa(high_memory) - 1 > ISA_DMA_THRESHOLD)
+ need_isa_bounce_buffers =3D 1;
+ else
+ need_isa_bounce_buffers =3D 0;
+
+ if (scsi_devicelist)
+ for (shpnt =3D scsi_hostlist; shpnt; shpnt =3D shpnt->next)
+ new_dma_sectors +=3D SECTORS_PER_PAGE; /* Increment for each host */
+
+ for (host =3D scsi_hostlist; host; host =3D host->next) {
+ for (SDpnt =3D host->host_queue; SDpnt; SDpnt =3D SDpnt->next) {
+ /*
+ * sd and sr drivers allocate scatterlists.
+ * sr drivers may allocate for each command 1x2048 or 2x1024 extra
+ * buffers for 2k sector size and 1k fs.
+ * sg driver allocates buffers < 4k.
+ * st driver does not need buffers from the dma pool.
+ * estimate 4k buffer/command for devices of unknown type (should =
panic).
+ */
+ if (SDpnt->type =3D=3D TYPE_WORM || SDpnt->type =3D=3D TYPE_ROM ||
+ SDpnt->type =3D=3D TYPE_DISK || SDpnt->type =3D=3D TYPE_MOD) {
+ new_dma_sectors +=3D ((host->sg_tablesize *
+ sizeof(struct scatterlist) + 511) >> 9) *
+ SDpnt->queue_depth;
+ if (SDpnt->type =3D=3D TYPE_WORM || SDpnt->type =3D=3D TYPE_ROM)
+ new_dma_sectors +=3D (2048 >> 9) * SDpnt->queue_depth;
+ } else if (SDpnt->type =3D=3D TYPE_SCANNER ||
+ SDpnt->type =3D=3D TYPE_PROCESSOR ||
+ SDpnt->type =3D=3D TYPE_MEDIUM_CHANGER ||
+ SDpnt->type =3D=3D TYPE_ENCLOSURE) {
+ new_dma_sectors +=3D (4096 >> 9) * SDpnt->queue_depth;
+ } else {
+ if (SDpnt->type !=3D TYPE_TAPE) {
+ printk("resize_dma_pool: unknown device type %d\n", SDpnt->type);
+ new_dma_sectors +=3D (4096 >> 9) * SDpnt->queue_depth;
+ }
+ }
+
+ if (host->unchecked_isa_dma &&
+ need_isa_bounce_buffers &&
+ SDpnt->type !=3D TYPE_TAPE) {
+ new_dma_sectors +=3D (PAGE_SIZE >> 9) * host->sg_tablesize *
+ SDpnt->queue_depth;
+ new_need_isa_buffer++;
+ }
+ }
+ }
+
+#ifdef DEBUG_INIT
+ printk("resize_dma_pool: needed dma sectors =3D %d\n", =
new_dma_sectors);
+#endif
+
+ /* limit DMA memory to 32MB: */
+ new_dma_sectors =3D (new_dma_sectors + 15) & 0xfff0;
+
+ /*
+ * We never shrink the buffers - this leads to
+ * race conditions that I would rather not even think
+ * about right now.
+ */
+#if 0 /* Why do this? No gain and risks out_of_space */
+ if (new_dma_sectors < dma_sectors)
+ new_dma_sectors =3D dma_sectors;
+#endif
+ if (new_dma_sectors <=3D dma_sectors) {
+ spin_unlock_irqrestore(&allocator_request_lock, flags);
+ return; /* best to quit while we are in front */
+ }
+
+ for (k =3D 0; k < 20; ++k) { /* just in case */
+ out_of_space =3D 0;
+ size =3D (new_dma_sectors / SECTORS_PER_PAGE) *
+ sizeof(FreeSectorBitmap);
+ new_dma_malloc_freelist =3D (FreeSectorBitmap *)
+ kmalloc(size, GFP_ATOMIC);
+ if (new_dma_malloc_freelist) {
+ memset(new_dma_malloc_freelist, 0, size);
+ size =3D (new_dma_sectors / SECTORS_PER_PAGE) *
+ sizeof(*new_dma_malloc_pages);
+ new_dma_malloc_pages =3D (unsigned char **)
+ kmalloc(size, GFP_ATOMIC);
+ if (!new_dma_malloc_pages) {
+ size =3D (new_dma_sectors / SECTORS_PER_PAGE) *
+ sizeof(FreeSectorBitmap);
+ kfree((char *) new_dma_malloc_freelist);
+ out_of_space =3D 1;
+ } else {
+ memset(new_dma_malloc_pages, 0, size);
+ }
+ } else
+ out_of_space =3D 1;
+
+ if ((!out_of_space) && (new_dma_sectors > dma_sectors)) {
+ for (i =3D dma_sectors / SECTORS_PER_PAGE;
+ i < new_dma_sectors / SECTORS_PER_PAGE; i++) {
+ new_dma_malloc_pages[i] =3D (unsigned char *)
+ __get_free_pages(GFP_ATOMIC | GFP_DMA, 0);
+ if (!new_dma_malloc_pages[i])
+ break;
+ }
+ if (i !=3D new_dma_sectors / SECTORS_PER_PAGE) { /* clean up */
+ int k =3D i;
+
+ out_of_space =3D 1;
+ for (i =3D 0; i < k; ++i)
+ free_pages((unsigned long) new_dma_malloc_pages[i], 0);
+ }
+ }
+ if (out_of_space) { /* try scaling down new_dma_sectors request */
+ printk("scsi::resize_dma_pool: WARNING, dma_sectors=3D%u, "
+ "wanted=3D%u, scaling\n", dma_sectors, new_dma_sectors);
+ if (new_dma_sectors < (8 * SECTORS_PER_PAGE))
+ break; /* pretty well hopeless ... */
+ new_dma_sectors =3D (new_dma_sectors * 3) / 4;
+ new_dma_sectors =3D (new_dma_sectors + 15) & 0xfff0;
+ if (new_dma_sectors <=3D dma_sectors)
+ break; /* stick with what we have got */
+ } else
+ break; /* found space ... */
+ } /* end of for loop */
+ if (out_of_space) {
+ spin_unlock_irqrestore(&allocator_request_lock, flags);
+ scsi_need_isa_buffer =3D new_need_isa_buffer; /* some useful info */
+ printk(" WARNING, not enough memory, pool not expanded\n");
+ return;
+ }
+ /* When we dick with the actual DMA list, we need to
+ * protect things
+ */
+ if (dma_malloc_freelist) {
+ size =3D (dma_sectors / SECTORS_PER_PAGE) * sizeof(FreeSectorBitmap);
+ memcpy(new_dma_malloc_freelist, dma_malloc_freelist, size);
+ kfree((char *) dma_malloc_freelist);
+ }
+ dma_malloc_freelist =3D new_dma_malloc_freelist;
+
+ if (dma_malloc_pages) {
+ size =3D (dma_sectors / SECTORS_PER_PAGE) * =
sizeof(*dma_malloc_pages);
+ memcpy(new_dma_malloc_pages, dma_malloc_pages, size);
+ kfree((char *) dma_malloc_pages);
+ }
+ scsi_dma_free_sectors +=3D new_dma_sectors - dma_sectors;
+ dma_malloc_pages =3D new_dma_malloc_pages;
+ dma_sectors =3D new_dma_sectors;
+ scsi_need_isa_buffer =3D new_need_isa_buffer;
+
+ spin_unlock_irqrestore(&allocator_request_lock, flags);
+
+#ifdef DEBUG_INIT
+ printk("resize_dma_pool: dma free sectors =3D %d\n", =
scsi_dma_free_sectors);
+ printk("resize_dma_pool: dma sectors =3D %d\n", dma_sectors);
+ printk("resize_dma_pool: need isa buffers =3D %d\n", =
scsi_need_isa_buffer);
+#endif
+}
+
+/*
+ * Function: scsi_init_minimal_dma_pool
+ *
+ * Purpose: Allocate a minimal (1-page) DMA pool.
+ *
+ * Arguments: None.
+ *
+ * Lock status: No locks assumed to be held. This function is =
SMP-safe.
+ *
+ * Returns: Nothing
+ *
+ * Notes: =20
+ */
+int scsi_init_minimal_dma_pool(void)
+{
+ unsigned long size;
+ unsigned long flags;
+ int has_space =3D 0;
+
+ spin_lock_irqsave(&allocator_request_lock, flags);
+
+ dma_sectors =3D PAGE_SIZE / SECTOR_SIZE;
+ scsi_dma_free_sectors =3D dma_sectors;
+ /*
+ * Set up a minimal DMA buffer list - this will be used during =
scan_scsis
+ * in some cases.
+ */
+
+ /* One bit per sector to indicate free/busy */
+ size =3D (dma_sectors / SECTORS_PER_PAGE) * sizeof(FreeSectorBitmap);
+ dma_malloc_freelist =3D (FreeSectorBitmap *)
+ kmalloc(size, GFP_ATOMIC);
+ if (dma_malloc_freelist) {
+ memset(dma_malloc_freelist, 0, size);
+ /* One pointer per page for the page list */
+ dma_malloc_pages =3D (unsigned char **) kmalloc(
+ (dma_sectors / SECTORS_PER_PAGE) * =
sizeof(*dma_malloc_pages),
+ GFP_ATOMIC);
+ if (dma_malloc_pages) {
+ memset(dma_malloc_pages, 0, size);
+ dma_malloc_pages[0] =3D (unsigned char *)
+ __get_free_pages(GFP_ATOMIC | GFP_DMA, 0);
+ if (dma_malloc_pages[0])
+ has_space =3D 1;
+ }
+ }
+ if (!has_space) {
+ if (dma_malloc_freelist) {
+ kfree((char *) dma_malloc_freelist);
+ if (dma_malloc_pages)
+ kfree((char *) dma_malloc_pages);
+ }
+ spin_unlock_irqrestore(&allocator_request_lock, flags);
+ printk("scsi::init_module: failed, out of memory\n");
+ return 1;
+ }
+
+ spin_unlock_irqrestore(&allocator_request_lock, flags);
+ return 0;
+}
Index: linux/drivers/scsi/scsi_lib.c
diff -u linux/drivers/scsi/scsi_lib.c:1.1.1.5 =
linux/drivers/scsi/scsi_lib.c:1.8
--- linux/drivers/scsi/scsi_lib.c:1.1.1.5 Fri Jan 7 22:33:08 2000
+++ linux/drivers/scsi/scsi_lib.c Mon Jan 10 21:44:27 2000
@@ -51,6 +51,13 @@
*/
=20
/*
+ * For hosts that request single-file access to the ISA bus, this is a =
pointer to
+ * the currently active host.
+ */
+volatile struct Scsi_Host *host_active =3D NULL;
+
+
+/*
* Function: scsi_insert_special_cmd()
*
* Purpose: Insert pre-formed command into request queue.
@@ -202,6 +209,23 @@
* If SCpnt is NULL, it means that the previous command
* was completely finished, and we should simply start
* a new command, if possible.
+ *
+ * This is where a lot of special case code has begun to
+ * accumulate. It doesn't really affect readability or
+ * anything, but it might be considered architecturally
+ * inelegant. If more of these special cases start to
+ * accumulate, I am thinking along the lines of implementing
+ * an atexit() like technology that gets run when commands
+ * complete. I am not convinced that it is worth the
+ * added overhead, however. Right now as things stand,
+ * there are simple conditional checks, and most hosts
+ * would skip past.
+ *
+ * Another possible solution would be to tailor different
+ * handler functions, sort of like what we did in scsi_merge.c.
+ * This is probably a better solution, but the number of different
+ * permutations grows as 2**N, and if too many more special cases
+ * get added, we start to get screwed.
*/
void scsi_queue_next_request(request_queue_t * q, Scsi_Cmnd * SCpnt)
{
@@ -287,6 +311,48 @@
SHpnt->some_device_starved =3D 0;
}
}
+
+ /*
+ * This is the code to deal with blocked hosts. The idea is that if =
the current host is blocked,
+ * yet the current host is inactive (as we have completed all requests =
outstanding, and the
+ * current host was the "owner", then we walk through the list and =
search for a new owner,
+ * and queue up some commands for those devices.
+ */
+#ifdef CONFIG_SCSI_HOST_BLOCK
+ if( SDpnt->host->block !=3D NULL
+ && host_active !=3D NULL
+ && SDpnt->host =3D=3D host_active
+ && host_active->host_busy =3D=3D 0 ) {
+ /*
+ * No host currently active. Look for someone who is idle and who =
has requests.
+ */
+ host_active =3D NULL;
+ =09
+ for(SHpnt =3D SDpnt->host->block; SHpnt !=3D SDpnt->host; SHpnt =3D =
SHpnt->block) {
+ for (SDpnt =3D SHpnt->host_queue; SDpnt; SDpnt =3D SDpnt->next) {
+ request_queue_t *q;
+ if ((SHpnt->can_queue > 0 && (SHpnt->host_busy >=3D =
SHpnt->can_queue))
+ || (SHpnt->host_blocked)) {
+ break;
+ }
+ if (SDpnt->device_blocked || !SDpnt->starved) {
+ continue;
+ }
+ q =3D &SDpnt->request_queue;
+ q->request_fn(q);
+ all_clear =3D 0;
+ }
+ if (SDpnt =3D=3D NULL && all_clear) {
+ SHpnt->some_device_starved =3D 0;
+ }
+ if( host_active !=3D NULL )
+ {
+ break;
+ }
+ }
+ }
+#endif
+
spin_unlock_irqrestore(&io_request_lock, flags);
}
=20
@@ -732,6 +798,56 @@
}
=20
/*
+ * Function: scsi_blocked_request_fn()
+ *
+ * Purpose: A request function wrapper for SCSI hosts that have =
blocking enabled.
+ *
+ * Arguments: q - Pointer to actual queue.
+ *
+ * Returns: Nothing
+ *
+ * Lock status: IO request lock assumed to be held when called.
+ *
+ * Notes:
+ */
+void scsi_blocked_request_fn(request_queue_t * q)
+{
+ Scsi_Device *SDpnt;
+ struct Scsi_Host *SHpnt;
+
+ ASSERT_LOCK(&io_request_lock, 1);
+
+ SDpnt =3D (Scsi_Device *) q->queuedata;
+ if (!SDpnt) {
+ panic("Missing device");
+ }
+ SHpnt =3D SDpnt->host;
+
+ /*
+ * If this host is currently blocked, then ignore the request for now.
+ */
+ if( host_active !=3D NULL && host_active !=3D SHpnt ) {
+ return;
+ }
+
+ if( host_active =3D=3D NULL ) {
+ host_active =3D SHpnt;
+ }
+
+ /*
+ * At this point, call the normal request function.
+ */
+ scsi_request_fn(q);
+
+ /*
+ * If the host isn't active, take it back again.
+ */
+ if( host_active =3D=3D SHpnt && SHpnt->host_busy =3D=3D 0 ) {
+ host_active =3D NULL;
+ }
+}
+
+/*
* Function: scsi_request_fn()
*
* Purpose: Generic version of request function for SCSI hosts.
@@ -985,3 +1101,91 @@
spin_lock_irq(&io_request_lock);
}
}
+
+/*
+ * Function: scsi_make_blocked_list
+ *
+ * Purpose: Build linked list of hosts that require blocking.
+ *
+ * Arguments: None.
+ *
+ * Returns: Nothing
+ *
+ * Notes: Blocking is sort of a hack that is used to prevent more =
than one
+ * host adapter from being active at one time. This is =
used in cases
+ * where the ISA bus becomes unreliable if you have more =
than one
+ * host adapter really pumping data through.
+ *
+ * We spent a lot of time examining the problem, and I =
*believe* that
+ * the problem is bus related as opposed to being a driver =
bug.
+ *
+ * The blocked list is used as part of the synchronization =
object
+ * that we use to ensure that only one host is active at =
one time.
+ * I (ERY) would like to make this go away someday, but =
this would
+ * require that we have a recursive mutex object.
+ *
+ * Note2: Now I wish I remember what I meant by that, because we now =
have
+ * reader-writer locks...
+ */
+
+void scsi_make_blocked_list(void)
+{
+#ifdef CONFIG_SCSI_HOST_BLOCK
+ int block_count =3D 0, index;
+ struct Scsi_Host *sh[128], *shpnt;
+
+ /*
+ * Create a circular linked list from the scsi hosts which have
+ * the "wish_block" field in the Scsi_Host structure set.
+ * The blocked list should include all the scsi hosts using ISA DMA.
+ * In some systems, using two dma channels simultaneously causes
+ * unpredictable results.
+ * Among the scsi hosts in the blocked list, only one host at a time
+ * is allowed to have active commands queued. The transition from
+ * one active host to the next one is allowed only when host_busy =
=3D=3D 0
+ * for the active host (which implies host_busy =3D=3D 0 for all the =
hosts
+ * in the list). Moreover for block devices the transition to a new
+ * active host is allowed only when a request is completed, since a
+ * block device request can be divided into multiple scsi commands
+ * (when there are few sg lists or clustering is disabled).
+ *
+ * (DB, 4 Feb 1995)
+ */
+
+
+ host_active =3D NULL;
+
+ for (shpnt =3D scsi_hostlist; shpnt; shpnt =3D shpnt->next) {
+
+#if 0
+ /*
+ * Is this is a candidate for the blocked list?
+ * Useful to put into the blocked list all the hosts whose driver
+ * does not know about the host->block feature.
+ */
+ if (shpnt->unchecked_isa_dma)
+ shpnt->wish_block =3D 1;
+#endif
+
+ if (shpnt->wish_block)
+ sh[block_count++] =3D shpnt;
+ }
+
+ if (block_count =3D=3D 1)
+ sh[0]->block =3D NULL;
+
+ else if (block_count > 1) {
+
+ for (index =3D 0; index < block_count - 1; index++) {
+ sh[index]->block =3D sh[index + 1];
+ printk("scsi%d : added to blocked host list.\n",
+ sh[index]->host_no);
+ }
+
+ sh[block_count - 1]->block =3D sh[0];
+ printk("scsi%d : added to blocked host list.\n",
+ sh[index]->host_no);
+ }
+#endif
+}
+
Index: linux/include/linux/isapnp.h
diff -u linux/include/linux/isapnp.h:1.1.1.1 =
linux/include/linux/isapnp.h:1.2
--- linux/include/linux/isapnp.h:1.1.1.1 Sat Jan 8 12:51:32 2000
+++ linux/include/linux/isapnp.h Sat Jan 8 13:07:09 2000
@@ -173,21 +173,21 @@
extern inline void isapnp_write_byte(unsigned char idx, unsigned char =
val) { ; }
extern inline void isapnp_write_word(unsigned char idx, unsigned short =
val) { ; }
extern inline void isapnp_write_dword(unsigned char idx, unsigned int =
val) { ; }
-extern void isapnp_wake(unsigned char csn) { ; }
-extern void isapnp_device(unsigned char device) { ; }
-extern void isapnp_activate(unsigned char device) { ; }
-extern void isapnp_deactivate(unsigned char device) { ; }
+extern inline void isapnp_wake(unsigned char csn) { ; }
+extern inline void isapnp_device(unsigned char device) { ; }
+extern inline void isapnp_activate(unsigned char device) { ; }
+extern inline void isapnp_deactivate(unsigned char device) { ; }
/* manager */
-extern struct pci_bus *isapnp_find_card(unsigned short vendor,
- unsigned short device,
- struct pci_bus *from) { return NULL; }
-extern struct pci_dev *isapnp_find_dev(struct pci_bus *card,
- unsigned short vendor,
- unsigned short function,
+extern inline struct pci_bus *isapnp_find_card(unsigned short vendor,
+ unsigned short device,
+ struct pci_bus *from) { return NULL; }
+extern inline struct pci_dev *isapnp_find_dev(struct pci_bus *card,
+ unsigned short vendor,
+ unsigned short function,
struct pci_dev *from) { return NULL; }
-extern void isapnp_resource_change(struct resource *resource,
- unsigned long start,
- unsigned long size) { ; }
+extern inline void isapnp_resource_change(struct resource *resource,
+ unsigned long start,
+ unsigned long size) { ; }
=20
#endif /* CONFIG_ISAPNP */
=20
Index: linux/include/linux/proc_fs.h
diff -u linux/include/linux/proc_fs.h:1.1.1.1 =
linux/include/linux/proc_fs.h:1.2
--- linux/include/linux/proc_fs.h:1.1.1.1 Sat Jan 8 12:51:31 2000
+++ linux/include/linux/proc_fs.h Sat Jan 8 13:07:09 2000
@@ -181,11 +181,11 @@
mode_t mode, struct proc_dir_entry *parent) { return NULL; }
=20
extern inline void remove_proc_entry(const char *name, struct =
proc_dir_entry *parent) {};
-extern inline proc_dir_entry *proc_symlink(const char *name,
+extern inline struct proc_dir_entry *proc_symlink(const char *name,
struct proc_dir_entry *parent,char *dest) {return NULL;}
-extern inline proc_dir_entry *proc_mknod(const char *name,mode_t mode,
+extern inline struct proc_dir_entry *proc_mknod(const char *name,mode_t =
mode,
struct proc_dir_entry *parent,kdev_t rdev) {return NULL;}
-extern struct proc_dir_entry *proc_mkdir(const char *name,
+extern inline struct proc_dir_entry *proc_mkdir(const char *name,
struct proc_dir_entry *parent) {return NULL;}
=20
extern inline struct proc_dir_entry *create_proc_read_entry(const char =
*name,
------=_NextPart_000_0072_01BF5D57.511A9DE0--
-
To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to majordomo@vger.rutgers.edu
