libata previously depended upon waits in prereset to get resets after
hotplug right for both spin up and device ready wait. This was
necessary both for reliablity and speed as reset was likely to fail if
initiated too early and each try usually took more than 30secs to
fail. Previous patches fixed the reliability part by fixing status
and SCR handling in resets. This patch remedies the speed part by
improving reset sequencing.
Prereset waiting timeout is adjusted to 10s because spinup wait is
replaced by reset sequencing and !BSY wait is not as important as
before. During boot or module loading where the drive is already
fully spun up, !BSY wait succeeds immediately, so 10s should be enough
in most cases. It matters after hotplugging or other error
conditions, but in those cases, !BSY wait in prereset simply can't be
relied upon due to the varied and weird behaviors ATA controllers and
devices show.
Reset is now driven by ata_eh_reset_timeouts[] table which contains
timeouts for each reset try. The first reset can be softreset but the
following ones are always hardreset if available. Each timeout
defines deadline for the reset try. If a reset try fails, reset is
retried with the next timeout till the end of the timeout table is
reached. If a reset try fails before the timeout with error, libata
waits till the deadline of the failed try before retrying.
IOW, the timeout table defines timetable of reset tries such that the
n'th try always begins at least after the sum of all previous timeouts
has passed. The current timetable defines 4 tries and takes around 1
minute.
@0 : First try. This should succeed most of the time during boot.
@10 : 10s is enough to spin up most consumer harddrives. Give it
another shot.
@20 : 20s should spin up > 99% of working drives. This has 30s
timeout for retarded devices needing long idleness post reset.
@55 : Final try with 5s timeout just in case.
The above timetable is trade off between not annoying the device too
much with frequent resets and taking reasonable amount of time in most
cases. Some controllers may do better with shorter timeouts while
others may fare better with longer but we just can't rely upon LLD
writers to test each controller with wide variety of devices using
various scenarios. We need default behavior which reasonably fits
most cases.
I've tested the above timetable on a dozen SATA controllers and a few
PATA controllers with about a dozen different drives from all major
vendors and 4 different ODDs from three different vendors for both
boot and hotplug (if available) cases.
Boot probing is not affected unless the device is broken in which
cases new code gives up on the port after a minute rather than five or
nine minutes. When hotplugging, most devices get detected on the
first or second try. Multi-platter drives with long spin up time
which sometimes took > 40 secs with the original code, now usually
comes up during the second try and at least right after the third try
@20.
Signed-off-by: Tejun Heo <htejun@gmail.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Add @deadline to prereset and reset methods and make them honor it.
ata_wait_ready() which directly takes @deadline is implemented to be
used as the wait function. This patch is in preparation for EH timing
improvements.
* ata_wait_ready() never does busy sleep. It's only used from EH and
no wait in EH is that urgent. This function also prints 'be
patient' message automatically after 5 secs of waiting if more than
3 secs is remaining till deadline.
* ata_bus_post_reset() now fails with error code if any of its wait
fails. This is important because earlier reset tries will have
shorter timeout than the spec requires. If a device fails to
respond before the short timeout, reset should be retried with
longer timeout rather than silently ignoring the device.
There are three behavior differences.
1. Timeout is applied to both devices at once, not separately. This
is more consistent with what the spec says.
2. When a device passes devchk but fails to become ready before
deadline. Previouly, post_reset would just succeed and let
device classification remove the device. New code fails the
reset thus causing reset retry. After a few times, EH will give
up disabling the port.
3. When slave device passes devchk but fails to become accessible
(TF-wise) after reset. Original code disables dev1 after 30s
timeout and continues as if the device doesn't exist, while the
patched code fails reset. When this happens, new code fails
reset on whole port rather than proceeding with only the primary
device.
If the failing device is suffering transient problems, new code
retries reset which is a better behavior. If the failing device is
actually broken, the net effect is identical to it, but not to the
other device sharing the channel. In the previous code, reset would
have succeeded after 30s thus detecting the working one. In the new
code, reset fails and whole port gets disabled. IMO, it's a
pathological case anyway (broken device sharing bus with working
one) and doesn't really matter.
* ata_bus_softreset() is changed to return error code from
ata_bus_post_reset(). It used to return 0 unconditionally.
* Spin up waiting is to be removed and not converted to honor
deadline.
* To be on the safe side, deadline is set to 40s for the time being.
Signed-off-by: Tejun Heo <htejun@gmail.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Separate ATA_EHI_DID_RESET into ATA_EHI_DID_SOFTRESET and
ATA_EHI_DID_HARDRESET. ATA_EHI_DID_RESET is redefined as OR of the
two flags. This patch doesn't introduce any behavior change. This
will be used later to determine whether _SDD is necessary or not.
Signed-off-by: Tejun Heo <htejun@gmail.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Signed-off-by: Alan Cox <alan@redhat.com>
Add support for ignoring the BIOS HPA result (off by default) and setting
the disk to the full available size unless already frozen.
Tested with various platforms/disks and confirmed to work with the
Macintosh (which broke earlier) and ata_piix (breakage due to the LBA48
readback that Tejun fixed).
For normal users this brings us, I believe, to feature parity with old IDE
(and of course more featured in some areas too).
Signed-off-by: Jeff Garzik <jeff@garzik.org>
All drivers are converted to new init model. Kill probe_ent,
ata_device_add() and ata_pci_init_native_mode().
Signed-off-by: Tejun Heo <htejun@gmail.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
These will be used to convert LLDs to new init model.
* Add irq_handler field to port_info. In new init model, requesting
IRQ is LLD's responsibility and libata doesn't need to know about
irq_handler. Most LLDs can simply register their irq_handler but
some need different irq_handler depending on specific chip. The
added port_info->irq_handler field can be used by LLDs to select
the matching IRQ handler in such cases.
* Add ata_dummy_port_info.
* Implement ata_pci_prepare_native_host(), a helper to alloc ATA host,
acquire all resources and init the host in one go.
Signed-off-by: Tejun Heo <htejun@gmail.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Convert native PCI host handling to alloc-init-register model. New
function ata_pci_init_native_host() follows the new init model and
replaces ata_pci_init_native_mode(). As there are remaining LLD
users, the old function isn't removed yet.
ata_pci_init_one() is reimplemented using the new function and now
fully converted to new init model.
Signed-off-by: Tejun Heo <htejun@gmail.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Implement ata_host_alloc_pinfo() and ata_host_register(). These helpers
will be used in the following patches to adopt new init model.
Signed-off-by: Tejun Heo <htejun@gmail.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Reorganize ata_host_alloc() and its subroutines into the following
three functions.
* ata_host_alloc() : allocates host and its ports. shost is not
registered automatically.
* ata_scsi_add_hosts() : allocates and adds shosts associated with an
ATA host. Used by ata_host_register().
* ata_host_register() : takes a fully initialized ata_host structure
and registers it to libata layer and probes it.
Only ata_host_alloc() and ata_host_register() are exported.
ata_device_add() is rewritten using the above functions. This patch
does not introduce any observable behavior change. Things worth
mentioning.
* print_id is assigned at registration time and LLDs are allowed to
overallocate ports and reduce host->n_ports during initialization.
ata_host_register() will throw away unused ports automatically.
* All SCSI host initialization stuff now resides in
ata_scsi_add_hosts() in libata-scsi.c, where it should be.
* ipr is now the only user of ata_host_init(). Either kill it by
converting ipr to use ata_host_alloc() and friends or rename and
move it to libata-scsi.c
Signed-off-by: Tejun Heo <htejun@gmail.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Separate out ata_host_start() from ata_device_add(). ata_host_start()
calls ->port_start on each port if available and freezes the port.
Signed-off-by: Tejun Heo <htejun@gmail.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Don't embed ap inside shost. Allocate it separately and point it back
from shosts's hostdata. This makes port allocation more flexible and
allows regular ATA and SAS share host alloc/init paths.
Signed-off-by: Tejun Heo <htejun@gmail.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
The READ/WRITE LONG commands are theoretically obsolete,
but the majority of drives in existance still implement them.
The WRITE_LONG and WRITE_LONG_ONCE commands are of particular
interest for fault injection testing -- eg. creating "media errors"
at specific locations on a disk.
The fussy bit is that these commands require a non-standard
sector size, usually 520 bytes instead of 512.
This patch adds support to libata for READ/WRITE LONG commands
issued via SG_IO/ATA_16.
Signed-off-by: Mark Lord <mlord@pobox.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
With Tejun having added adev->ap some time ago we can get rid of the
almost unused port being passed to mode filters. And while we are
doing filters, lets turn on the !IORDY filter as well.
Signed-off-by: Alan Cox <alan@redhat.com>
With some hand massaging from
Signed-off-by: Jeff Garzik <jeff@garzik.org>
This splits set_mode into do_set_mode and the wrapper so that a driver can
call the standard method inside its own. This in theory also obsoletes
->post_set_mode().
Signed-off-by: Alan Cox <alan@redhat.com>
Cc: Jeff Garzik <jeff@garzik.org>
Cc: Tejun Heo <htejun@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
2.6.21-rc has horrible problems with libata and PATA cable types (and
thus speeds). This occurs because Tejun fixed a pile of other bugs and
we now do cable detect enforcement for drive side detection properly.
Unfortunately we don't do the process around cable detection right. Tejun
identified the problem and pointed to the right Annex in the spec, this patch
implements the rest of the needed changes.
We add a ->cable_detect() method called after the identify
sequence which allows a host to do host side detection at this point
should it wish, or to modify the results of the drive side identify.
This separate ->cable_detect method also cleans up a lot of code because
many drivers have their own error_handler methods which really just set
the cable type.
If there is no ->cable_detect method the cable type is left alone so a
driver setting it earlier (eg because it has the SATA flags set or
because it uses the old error_handler approach) will still do the right
thing (or at least the same thing) as before.
This patch simply adds the cable_detect method and helpers it doesn't use
them but other follow up patches will (ie Adrian please don't submit
patches to unexport them ;))
Signed-off-by: Alan Cox <alan@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
It used to be impossible to get from ata_device to ata_port but that is
no longer true. Various methods have been cleaned up over time but
dev_config still takes both and most users don't need both anyway. Tidy
this one up
Signed-off-by: Alan Cox <alan@redhat.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
patch 4/4:
Limit ATAPI DMA to R/W commands only for TORiSAN DRD-N216 DVD-ROM drives
(http://bugzilla.kernel.org/show_bug.cgi?id=6710)
Signed-off-by: Albert Lee <albertcc@tw.ibm.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
patch 3/4:
The TORiSAN drive locks up when max sector == 256.
Limit max sector to 128 for the TORiSAN DRD-N216 drives.
(http://bugzilla.kernel.org/show_bug.cgi?id=6710)
Signed-off-by: Albert Lee <albertcc@tw.ibm.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
patch 1/4:
Reorder HSM_ST_FIRST, such that the task state transition is easier decoded with human eyes.
Signed-off-by: Albert Lee <albertcc@tw.ibm.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Conditionalize all PM related stuff in libata core layer using
CONFIG_PM.
Signed-off-by: Tejun Heo <htejun@gmail.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
The initial simplex handling code is fooled if you suspend and resume.
This also causes problems with some single channel controllers which
claim to be simplex.
The fix is fairly simple, instead of keeping a flag to remember if we
gave away the simplex channel we remember the actual owner. As the owner
is always part of the host_set we don't even need a refcount.
Knowing the owner also means we can reassign simplex DMA channels in
future hotplug code etc if we need to
Signed-off-by: Alan Cox <alan@redhat.com>
(and a signed-off for the patch I sent before while I remember)
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Also export dev_disable as this is needed by drivers doing slave decode
filtering, which will follow shortly
Signed-off-by: Alan Cox <alan@redhat.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
ata_port has two different id fields - id and port_no. id is
system-wide 1-based unique id for the port while port_no is 0-based
host-wide port number. The former is primarily used to identify the
ATA port to the user in printk messages while the latter is used in
various places in libata core and LLDs to index the port inside the
host.
The two fields feel quite similar and sometimes ap->id is used in
place of ap->port_no, which is very difficult to spot. This patch
renames ap->id to ap->print_id to reduce the possibility of such bugs.
Some printk messages are adjusted such that id string (ata%u[.%u])
isn't printed twice and/or to use ata_*_printk() instead of hardcoded
id format.
Signed-off-by: Tejun Heo <htejun@gmail.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
The current EH speed down code is more of a proof that the EH
framework is capable of adjusting transfer speed in response to error.
This patch puts some intelligence into EH speed down sequence. The
rules are..
* If there have been more than three timeout, HSM violation or
unclassified DEV errors for known supported commands during last 10
mins, NCQ is turned off.
* If there have been more than three timeout or HSM violation for known
supported command, transfer mode is slowed down. If DMA is active,
it is first slowered by one grade (e.g. UDMA133->100). If that
doesn't help, it's slowered to 40c limit (UDMA33). If PIO is
active, it's slowered by one grade first. If that doesn't help,
PIO0 is forced. Note that this rule does not change transfer mode.
DMA is never degraded into PIO by this rule.
* If there have been more than ten ATA bus, timeout, HSM violation or
unclassified device errors for known supported commands && speeding
down DMA mode didn't help, the device is forced into PIO mode. Note
that this rule is considered only for PATA devices and is pretty
difficult to trigger.
One error can only trigger one rule at a time. After a rule is
triggered, error history is cleared such that the next speed down
happens only after some number of errors are accumulated. This makes
sense because now speed down is done in bigger stride.
Signed-off-by: Tejun Heo <htejun@gmail.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
This is the patch for PATA controller of Celleb.
This driver uses the managed iomap (devres).
Because this driver needs special taskfile accesses, there is
a copy of ata_std_softreset(). ata_dev_try_classify() is exported
so that it can be used in this function.
Signed-off-by: Kou Ishizaki <kou.ishizaki@toshiba.co.jp>
Signed-off-by: Akira Iguchi <akira2.iguchi@toshiba.co.jp>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
_GTF is an acpi method that is used to reinitialize the drive. It returns
a task file containing ata commands that are sent back to the drive to restore
it to boot up defaults.
Signed-off-by: Kristen Carlson Accardi <kristen.c.accardi@intel.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
(cherry picked from 9c69cab24b51a89664f4c0dfaf8a436d32117624 commit)
This is the first preparation to doing the !IORDY cases properly. Further
diffs will then add the needed logic to do it right.
Signed-off-by: Alan Cox <alan@redhat.com>
Cc: Jeff Garzik <jeff@garzik.org>
Cc: Tejun Heo <htejun@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
The ATA_ENABLE_PATA define was never meant to be permanent, and in
recent kernels, it's already been unconditionally enabled. Remove.
Signed-off-by: Jeff Garzik <jeff@garzik.org>
This patch is against the libata core and headers.
Two IRQ calls are added in ata_port_operations.
- irq_on() is used to enable interrupts.
- irq_ack() is used to acknowledge a device interrupt.
In most drivers, ata_irq_on() and ata_irq_ack() are used for
irq_on and irq_ack respectively.
In some drivers (ex: ahci, sata_sil24) which cannot use them
as is, ata_dummy_irq_on() and ata_dummy_irq_ack() are used.
Signed-off-by: Kou Ishizaki <kou.ishizaki@toshiba.co.jp>
Signed-off-by: Akira Iguchi <akira2.iguchi@toshiba.co.jp>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Convert libata core layer and LLDs to use iomap.
* managed iomap is used. Pointer to pcim_iomap_table() is cached at
host->iomap and used through out LLDs. This basically replaces
host->mmio_base.
* if possible, pcim_iomap_regions() is used
Most iomap operation conversions are taken from Jeff Garzik
<jgarzik@pobox.com>'s iomap branch.
Signed-off-by: Tejun Heo <htejun@gmail.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Now that all LLDs are converted to use devres, default stop callbacks
are unused. Remove them.
Signed-off-by: Tejun Heo <htejun@gmail.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Update libata core layer to use devres.
* ata_device_add() acquires all resources in managed mode.
* ata_host is allocated as devres associated with ata_host_release.
* Port attached status is handled as devres associated with
ata_host_attach_release().
* Initialization failure and host removal is handedl by releasing
devres group.
* Except for ata_scsi_release() removal, LLD interface remains the
same. Some functions use hacky is_managed test to support both
managed and unmanaged devices. These will go away once all LLDs are
updated to use devres.
Signed-off-by: Tejun Heo <htejun@gmail.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Implement ata_host_detach() which calls ata_port_detach() for each
port in the host and export it. ata_port_detach() is now internal and
thus un-exported. ata_host_detach() will be used as the 'deregister
from libata layer' function after devres conversion.
Signed-off-by: Tejun Heo <htejun@gmail.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
libata used two separate sets of variables to record request size and
current offset for ATA and ATAPI. This is confusing and fragile.
This patch replaces qc->nsect/cursect with qc->nbytes/curbytes and
kills them. Also, ata_pio_sector() is updated to use bytes for
qc->cursg_ofs instead of sectors. The field used to be used in bytes
for ATAPI and in sectors for ATA.
Signed-off-by: Tejun Heo <htejun@gmail.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Handle pci_enable_device() failure while resuming. This patch kills
the "ignoring return value of 'pci_enable_device'" warning message and
propagates __must_check through ata_pci_device_do_resume().
Signed-off-by: Tejun Heo <htejun@gmail.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Updated diff which doesn't move the comment as per Jeff's request and
corrects the docs as per report on l/k
Signed-off-by: Alan Cox <alan@redhat.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
> Looks like you should use ata_busy_wait() here, rather than reproducing
> the same code again.
It waits in 10uS chunks while 1uS chunks were used in the workaround.
Could indeed do that once I know the fix is right. While I'm at it the
ata_busy_wait kerneldoc is borked so here's a fix
Signed-off-by: Alan Cox <alan@redhat.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Some uli controllers have stuck SIMPLEX bit which can't be cleared
with ata_pci_clear_simplex(), but the controller is capable of doing
DMAs on both channels simultaneously. Implement ATA_FLAG_IGN_SIMPLEX
which makes libata ignore the simplex bit and use it in sata_uli.
Signed-off-by: Tejun Heo <htejun@gmail.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
When set_mode() changed ->set_mode didn't adapt. This makes the needed
changes and removes the relevant FIXME case.
Signed-off-by: Alan Cox <alan@redhat.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Initialize qc->pad_len for each new command. This ensures
that pad_len is not set to a stale value for zero data
length commands.
Signed-off-by: Brian King <brking@linux.vnet.ibm.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Fixup the inialization of qc->n_elem. It currently gets
initialized to 1 for commands that do not transfer any data.
Fix this by initializing n_elem to 0 and only setting to 1
in ata_scsi_qc_new when there is data to transfer. This fixes
some problems seen with SATA devices attached to ipr adapters.
Signed-off-by: Brian King <brking@linux.vnet.ibm.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
libata didn't used to init qc->dma_dir to any specific value on qc
initialization and command translation path didn't set qc->dma_dir if
the command doesn't need data transfer. This made non-data commands
to have random qc->dma_dir.
This usually doesn't cause problem because LLDs usually check
qc->protocol first and look at qc->dma_dir iff the command needs data
transfer but this doesn't hold for all LLDs.
It might be worthwhile to rename qc->dma_dir to qc->data_dir as we use
the field to tag data direction for both PIO and DMA protocols.
This problem has been spotted by James Bottomley.
Signed-off-by: Tejun Heo <htejun@gmail.com>
Cc: James Bottomley <James.Bottomley@SteelEye.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Conflicts:
drivers/ata/libata-scsi.c
include/linux/libata.h
Futher merge of Linus's head and compilation fixups.
Signed-Off-By: David Howells <dhowells@redhat.com>
libata switched to IRQ-driven IDENTIFY when IRQ-driven PIO was
introduced. This has caused a lot of problems including device
misdetection and phantom device.
ATA_FLAG_DETECT_POLLING was added recently to selectively use polling
IDENTIFY on problemetic drivers but many controllers and devices are
affected by this problem and trying to adding ATA_FLAG_DETECT_POLLING
for each such case is diffcult and not very rewarding.
This patch makes libata always use polling IDENTIFY. This is
consistent with libata's original behavior and drivers/ide's behavior.
Signed-off-by: Tejun Heo <htejun@gmail.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
This patch implements ATA_FLAG_SETXFER_POLLING and use in pata_via.
If this flag is set, transfer mode setting performed by polling not by
interrupt. This should help those controllers which raise interrupt
before the command is actually complete on SETXFER.
Rationale for this approach.
* uses existing facility and relatively simple
* no busy sleep in the interrupt handler
* updating drivers is easy
While at it, kill now unused flag ATA_FLAG_SRST in pata_via.
Signed-off-by: Tejun Heo <htejun@gmail.com>
HSM_ST_UNKNOWN is not used anywhere. Its value is zero and supposed
to serve sanity check purpose but HSM_ST_IDLE is used for that
purpose. This unused state causes confusion. After a port is
initialized but before the first command is executed, the idle hsm
state is UNKNOWN. However, once a command has completed, the idle hsm
state is IDLE. This defeats sanity check in ata_pio_task() for the
first command.
This patch removes HSM_ST_UNKNOWN and consequently make HSM_ST_IDLE
the default state.
Signed-off-by: Tejun Heo <htejun@gmail.com>
On some controllers (ICHs in piix mode), there is *NO* reliable way to
determine device presence other than issuing IDENTIFY and see how the
transaction proceeds by watching the TF status register.
libata acted this way before irq-pio and phantom devices caused very
little problem but now that IDENTIFY is performed using IRQ drive PIO,
such phantom devices now result in multiple 30sec timeouts during
boot.
This patch implements ATA_FLAG_DETECT_POLLING. If a LLD sets this
flag, libata core issues the initial IDENTIFY in polling mode and if
the initial data transfer fails w/ HSM violation, the port is
considered to be empty thus replicating the old libata and IDE
behavior.
Signed-off-by: Jeff Garzik <jeff@garzik.org>
